ECPPM 2022 - eWork and eBusiness in Architecture, Engineering and Construction 2022 1032406739, 9781032406732
ECPPM 2022 - eWork and eBusiness in Architecture, Engineering and Construction contains the papers presented at the 14th
539 92 28MB
English Pages 812 [813] Year 2023
Cover
Title Page
Copyright Page
Table of Contents
Preface
Committee Members
Organizing Committee
Sustainable and circular driven digitalisation: Life-cycle assessment
Sustainability assessment of a novel reusable and demountable steel-concrete composite floor system
1 INTRODUCTION
2 NOVEL REUSABLE STEEL-CONCRETEFLOOR
3 STRUCTURAL FUNCTIONAL UNIT FORCOMPARATIVE ANALYSIS
4 STRUCTURAL ANALYSIS AND BILL OFMATERIALS
5 LIFE CYCLE ASSESSMENT
6 CONCLUSIONS
ACKNOWLEDGMENTS
REFERENCES
Probabilistic life cycle analysis as a sustainability-focused design tool for industrialized construction
1 INTRODUCTION
2 FRAMEWORK FOR PROBABILISTIC LCAAS A BUILDING DESIGN TOOL
3 PROCESS FOR PROBABILISTIC LCA OFBUILDINGS
4 INDUSTRIALIZED CONSTRUCTION CASESTUDY
5 CONCLUSIONS
ACKNOWLEDGMENTS
REFERENCES
Integrating Level(s) LCA in BIM: A tool for estimating LCA and LCC impacts in a case study
1 INTRODUCTION
2 MATERIAL AND METHOD
3 RESULTS AND DISCUSSION
4 CONCLUSIONS
ACKNOWLEDGMENTS
REFERENCES
How can LCA inform early-stage design to meet Danish regulations? The sustainability opportunity metric
1 INTRODUCTION
2 BACKGROUND
3 RELATEDWORK
4 METHODOLOGY
5 FRAMEWORK AND CONCEPTUALIZATION
6 PROTOTYPE DEMONSTRATION TOOL
7 CONCLUSIONS
REFERENCES
Life cycle potentials and improvement opportunities as guidance for early-stage design decisions
1 INTRODUCTION
2 STATE OF THE ART
3 TERMINOLOGY
4 METHODS
5 USE CASE
6 RESULTS & DISCUSSION
7 CONCLUSION & OUTLOOK
ACKNOWLEDGEMENTS
REFERENCES
Structure and LCA-driven building design support in early phases using knowledge-based methods and domain knowledge
1 INTRODUCTION
2 KNOWLEDGE-BASED DESIGN
3 DESIGN SUPPORT CONCEPT & METHODS
4 CASE STUDY
5 DISCUSSION
6 OUTLOOK
7 SUMMARY
ACKNOWLEDGEMENTS
REFERENCES
Challenges and experiences with the reuse of products in building design
1 INTRODUCTION
2 METHODOLOGY
3 RESULTS
4 DISCUSSION AND CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
Evaluating existing digital platforms enabling the reuse of reclaimed building materials and components for circularity
1 INTRODUCTION
2 METHODS
3 RESULTS
4 DISCUSSION
5 CONTRIBUTION TO INDUSTRY
ACKNOWLEDGEMENT
REFERENCES
Semantic Material Bank: A web-based linked data approach for building decommissioning and material reuse
1 INTRODUCTION
2 BACKGROUND AND RELEVANTWORKS
3 METHODOLOGY
4 SEMANTIC MATERIAL BANK
5 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
NLP-based semantic model healing for calculating LCA in early building design stages
1 INTRODUCTION AND MOTIVATION
2 STATE OF THE ART
3 SEMANTIC MODEL HEALING
4 EXPERIMENTS & RESULTS
5 DISCUSSION & OUTLOOK
REFERENCES
Construction product identification and localization using RFID tags and GS1 data synchronization system
1 INTRODUCTION
2 PRODUCT INFORMATION CODING
3 DATA SYNCHRONIZATION
4 BARCODE VERSUS RFID SYSTEMSIN CONSTRUCTION
5 FINAL REMARKS
REFERENCES
What comes first when implementing data templates? Refurbishment case study
1 INTRODUCTION
2 CONTRIBUTIONS TO THE BODY OFKNOWLEDGE AND INDUSTRY
3 METHODOLOGY
4 CASE STUDY
5 FINDINGS AND DISCUSSION
6 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
Chaos and black boxes – Barriers to traceability of construction materials
1 INTRODUCTION
2 THEORETICAL FRAMEWORK
3 METHODS
4 FINDINGS
5 DISCUSSION AND CONCLUSIONS
REFERENCES
Evaluating four types of data parsing methods for machine learning integration from building information models
1 INTRODUCTION
2 METHODOLOGY
3 RESULTS
4 DISCUSSION AND CONCLUSION
REFERENCES
Digital supported collaboration: Multimodel
Extending ICDD implementation to a dynamic multimodel framework
1 INTRODUCTION
2 RELATEDWORK
3 METHODOLOGY
4 IMPLEMENTATION
5 CONCLUSION AND FUTUREWORK
ACKNOWLEDGEMENT
REFERENCES
Management of BIM-based digital twins in multimodels
1 INTRODUCTION
2 DIGITAL TWINS AND MULTIMODELS INCONSTRUCTION
3 MULTIMODEL BASED MANAGEMENT OFDIGITAL TWINS
4 CASE STUDY: BRIDGE ASSESSMENT
5 CONCLUSION AND OUTLOOK
REFERENCES
Enriching BIM-based construction schedules with semantics using BPMN and LBD
1 INTRODUCTION
2 RELATEDWORK
3 METHODOLOGY
4 RESULTS
5 DISCUSSION AND CONCLUSION
REFERENCES
A simulative framework to evaluate constructability through parameter optimization at early design stage
1 GENERAL BACKGROUND AND RESEARCHMOTIVATIONS
2 DESIGN OPTIMISATION THROUGH THEVISUAL PROGRAMMING APPROACHEXPLAINED
3 CONCLUSIONS
4 DISCUSSION AND FUTUREWORKS
REFERENCES
Automatic generation of work breakdown structures for evaluation of parallelizability of assembly sequences
1 INTRODUCTION
2 RELATEDWORK
3 PROPOSED METHODOLOGY
4 IMPLEMENTATION
5 CASE STUDIES
6 CONCLUSION AND OUTLOOK
REFERENCES
Construction process time optimization of a reinforced concrete reaction slab – Implementing the VDC methodology
1 INTRODUCTION
2 MATERIAL AND METHOD
3 RESULTS
4 DISCUSSION
5 CONCLUSIONS
REFERENCES
VDC framework proposal for curtain wall construction process optimization
1 INTRODUCTION
2 MATERIALS AND METHODS
3 RESULTS
4 RESULTS DISCUSSION
5 CONCLUSIONS AND FUTURE RESEARCH
REFERENCES
Using production metrics to compare and understand VDC elements practiced by general contractors in Norway
1 INTRODUCTION
2 BACKGROUND AND MOTIVATION
3 POINT OF DEPARTURE
4 RESEARCH METHODOLOGY
5 RESULTS
6 DISCUSSION, IMPLICATIONS AND FUTURERESEARCH
7 CONCLUSION
REFERENCES
APPENDIX A
Digital twin: Building performance
A-EYE tech: Framework to evaluate an AI construction visibility platform
1 INTRODUCTION
2 LITERATURE REVIEW
3 A-EYE PROCESS ANALYSIS
4 CONCLUSION
REFERENCES
Digital twin in healthcare facilities: Linking real-time air quality data to BIM
1 INTRODUCTION
2 DIGITAL TWIN TO SUPPORT SPACEMANAGEMENT AND FACILITYMAINTENANCE NEEDS IN HEALTHCAREFACILITIES
3 DIGITAL TWIN PROTOTYPE OF ANOPERATING ROOM: CASE STUDY
4 CREATING THE DIGITAL TWIN INPOWER BI
5 DISCUSSION AND CONCLUSION
REFERENCES
A novel BIM platform paradigm for the building erformance domain
1 INTRODUCTION
2 RESEARCH METHODOLOGY
3 BUILDING PERFORMANCE MODELINGWORKFLOW IN THE NOVEL PARADIGM
4 ASPECTS OF IMPLEMENTATION
5 DISCUSSION
6 CONCLUSION
ACKNOWLEDGMENT
REFERENCES
Towards European standardization of digitalization approaches for monitoring and safety of bridges and tunnels
1 INTRODUCTION
2 DATA REVIEW OF INTEROPERABILITY INMONITORING AND MAINTENANCE
3 REPORT OF THE PAN-EUROPEAN MEETINGWITH THE COMMUNITY OF PRACTICE
4 CONCLUDING REMARKS
ACKNOWLEDGEMENT
REFERENCES
TwinGen: Advanced technologies to automatically generate digital twins for operation and maintenance of existing bridges
1 INTRODUCTION
2 POINT CLOUD SEGMENTATION
3 MODEL FITTING
4 DAMAGE SEGMENTATION
5 DATA LINKING AND MODELLING
6 CONCLUSION AND OUTLOOK
REFERENCES
Digital Twin for AECOO – Framework proposal and use cases
1 INTRODUCTION
2 RESEARCH DESING AND METHODS
3 PREVIOUS RESEARCH
4 DEVELOPMENT
5 USE CASES
6 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
OpenSantCugat: A platform for municipalities to provide access to building data
1 INTRODUTION
2 OPEN DATA PLATFORMS
3 OPENSANTCUGAT: AN OPEN DATAPLATFORM FOR MUNICIPALITIES
4 APPLICATIONS
5 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
Digital management for unforeseen trigger events using ISO 19650
1 INTRODUCTION
2 METHOD
3 THEORY
4 RESULTS
5 DISCUSSION
6 CONLUSION
REFERENCES
A lean strategic FM service model based on the digital twin
1 INTRODUCTION
2 LITERATURE REVIEW
3 RESEARCH METHODOLOGY
4 RESULTS AND DISCUSSION
5 CONCLUSIONS AND FUTUREWORK
REFERENCES
A Digital Twin prototype for smart parking management
1 INTRODUCTION
2 METHODOLOGY
3 RESULTS AND DISCUSSION
4 CONCLUSION AND FUTUREWORK
REFERENCES
BIM integrated data management for the Operation and Maintenance (O&M) of railway
1 INTRODUCTION
2 DATA COLLECTION AND PREPROCESSING
3 DATA VISUALIZATION ANDORGANIZATION ON BIM PLATFORM
4 EXPECTED FUNCTIONS OF 3D BIMPLATFORM
5 CONCLUSION
REFERENCES
How to represent damage data in BIM Models? – A literature review
1 INTRODUCTION
2 RESEARCH METHODOLOGY
3 RESULTS
4 DISCUSSION
5 CONCLUSION
ACKNOWLEDGEMENT
REFERENCES
Correlation and comparison between digital twin and cyber physical systems
1 INTRODUCTION
2 METHODOLOGY
3 DIGITAL TWIN (DT) ORIGIN
4 CYBER PHYSICAL SYSTEM (CPS)
5 DIFFERENCES AND SIMILARITIESBETWEEN DT AND CPS
6 CONCLUSION
REFERENCES
HBIM application in historic town: A scoping literature review
1 INTRODUCTION
2 THEORETICAL FRAMEWORK
3 METHODS
4 RESULTS
5 DISCUSSION
6 CONCLUSIONS
REFERENCES
The use of HBIM and scanning in cultural heritage projects
1 INTRODUCTION
2 THEORETICAL LENS
3 METHOD
4 ANALYSIS
5 DISCUSSION
6 CONCLUSION
REFERENCES
Enabling technologies: Energy
High-order second-level space boundary surface calculation for building energy performance simulation models
1 INTRODUCTION
2 TRADITIONAL APPROACH
3 PRELIMINARIES
4 EXISTING ALGORITHM
5 APPLICATION EXAMPLE
6 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
Aspects of BIM-to-BEM information transfer: A tale of two workflows
1 INTRODUCTION AND BACKGROUND
2 APPROACH
3 RESULTS AND DISCUSSION
4 CONCLUSION
REFERENCES
Predicting annual heating energy use using BIM: A simplified method for early design phase
1 INTRODUCTION
2 METHOD
3 RESULTS
4 DISCUSSION
5 CONCLUSIONS
ACKNOWLEDGMENTS
REFERENCES
Predicting semantic building information (BIM) with Recurrent Neural Networks
1 INTRODUCTION
2 BACKGROUND AND RELATEDWORK
3 APPROACH
4 EVALUATION
5 CONCLUSION AND FUTUREWORK
ACKNOWLEDGMENTS
REFERENCES
International assessment of artificial intelligence applications in the AEC sector
1 INTRODUCTION
2 BACKGROUND
3 METHODOLOGICAL APPROACH
4 RESULTS
5 DISCUSSION
6 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
Predicting occupant evacuation times to improve building design
1 INTRODUCTION
2 BACKGROUND AND RELATEDWORK
3 METHODOLOGY
4 NEURAL NETWORK ARCHITECTURE
5 NEURAL NETWORK RESULTS ANDEVALUATION
6 CONCLUSION AND FUTUREWORK
ACKNOWLEDGMENTS
REFERENCES
Inferring interconnections of construction drawings for bridges using deep learning-based methods
1 INTRODUCTION
2 RELATEDWORKS
3 METHODOLOGY
4 VALIDATION AND TEST RESULTS
5 CONCLUSIONS
ACKNOWLEDGMENTS
REFERENCES
An ontology-supported case-based reasoning approach for damage assessment
1 INTRODUCTION
2 RELATEDWORK
3 PROCESS MODEL
4 KNOWLEDGE MODEL
5 EXEMPLARY APPLICATION
6 CONCLUSION
REFERENCES
Towards a semantic enriching framework for construction site images
1 INTRODUCTION
2 PAPER REVIEW
3 RESEARCH SCOPE AND METHODOLOGY
4 THE ONTOLOGY MODEL
5 CASE STUDY
6 CONCLUSION
REFERENCES
Machine learning and genetic algorithms for conformal geometries in design support systems
1 INTRODUCTION
2 MACHINE LEARNING (ML)
3 GENETIC ALGORITHMS (GA)
4 CONCLUSIONS
REFERENCES
Predicting the construction duration in the predesign phase with decision trees
1 INTRODUCTION
2 LITERATURE REVIEW
3 MATERIALS AND METHODS
4 RESULTS AND DISCUSSION
5 CONCLUSION
ACKNOWLEDGEMENT
REFERENCES
Tower crane layout planning through Generative Adversarial Network
1 INTRODUCTION
2 LITERATURE REVIEW
3 TOWER CRANE LAYOUT GENERATIONPROBLEM
4 TC-GAN
5 IMPLEMENTATION AND RESULTS
6 CONCLUSION
Consideration of detailing in the graph-based retrieval of design variants
1 INTRODUCTION
2 BACKGROUND
3 METHODOLOGY
4 DEMONSTRATION
5 CONCLUSION AND FUTUREWORK
REFERENCES
Retrieve information from construction documents with BERT and unsupervised learning
1 INTRODUCTION
2 RELATEDWORKS
3 METHODOLOGIES
4 CONCEPT AND IMPLEMENTATION
5 EVALUATION AND DISCUSSION
6 CONCLUSION
ACKNOWLEDGEMENTS
REFERENCES
Automated floorplan generation using mathematical optimization
1 INTRODUCTIONS
2 STATE OF THE ART
3 METHODOLOGY
4 APPLICATIONS
5 RESULTS AND DISCUSSIONS
6 CONCLUSIONS AND LIMITATIONS
ACKNOWLEDGEMENT
REFERENCES
Information extraction and NLP for the interpretation of building permits: An Italian case study
1 INTRODUCTION
2 RESEARCH METHODOLOGY
3 CASE OF STUDY
4 DISCUSSIONS
5 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
Process-based building permit review – A knowledge engineering approach
1 INTRODUCTION
2 BACKGROUND
3 RESEARCH METHODOLOGY
4 IMPLEMENTATION AND VALIDATION
5 RESULTS AND DISCUSSION
6 SUMMARY AND CONCLUSION
REFERENCES
Training on digitized building regulations for automated rule extraction
1 INTRODUCTION
2 INVESTIGATING LRML BUILDINGREGULATIONS
3 NORMALISING THE LRML CORPUS
4 EVALUATING THE LRML CORPUS
5 DISCUSSION
6 CONCLUSIONS AND FUTUREWORK
ACKNOWLEDGEMENTS
REFERENCES
Using RASE semantic mark-up for normative, definitive and descriptive knowledge
1 INTRODUCTION
2 METHOD
3 THREE KINDS OF OPERABLE KNOWLEDGE
4 RASE BASED SERVICES
5 CONCLUSION
REFERENCES
How practice is represented in BIM-based model checking research – A literature review and reflections
1 GENERAL INSTRUCTIONS
2 METHODOLOGY
3 RESULTS
4 DISCUSSION AND CONCLUSION
REFERENCES
Model healing: Toward a framework for building designs to achieve code compliance
1 INTRODUCTION
2 RELATEDWORK
3 THE MODEL ADAPTATION PROCESS
4 PROOF OF CONCEPT
5 CONCLUSION AND FUTUREWORKS
ACKNOWLEDGEMENTS
REFERENCES
A multi-representation method of building rules for automatic code compliance checking
1 INTRODUCTION
2 METHOD
3 EVALUATION OF EXISTINGREPRESENTATION METHODS
4 A CLASSIFICATION OF BUILDING RULES
5 PROPOSED RULE REPRESENTATIONMETHOD
6 DISCUSSION
7 CONCLUSION
REFERENCES
Development of augmented BIM models for built environment management
1 INTRODUCTION
2 BACKGROUND
3 METHODOLOGY
4 SYSTEM IMPLEMENTATION
5 CONCLUSIONS
REFERENCES
BIM-based disaster response: Facilitating indoor path planning for various agents
1 INTRODUCTION
2 STATE OF THE ART
3 METHODOLOGY
4 SIMULATION AND RESULTS
5 CONCLUSION
ACKNOWLEDGEMENT
REFERENCES
Suitability of three national cost breakdown structures for automated quantity take-off in road projects
1 INTRODUCTION
2 THEORETICAL BACKGROUND
3 METHODOLOGY
4 RESULTS AND DISCUSSION
5 CONCLUSION
REFERENCES
Modular robotic system for the construction industry
1 INTRODUCTION
2 STATE OF THE ART
3 METHOD
4 USE CASE
5 CONCLUSION
ACKNOWLEDGEMENT
REFERENCES
Designing a framework for seamless integration of open data services to support disaster response
1 INTRODUCTION
2 BACKGROUND AND RELATEDWORK
3 CONCEPT
4 CONCEPTUAL IMPLEMENTATION
5 CONCLUSION
REFERENCES
Practical experiences with 5D building information modeling – A systematic literature review
1 INTRODUCTION
2 METHODOLOGY
3 RESULTS AND DISSCUSSIONS
4 CONCLUSION
REFERENCES
Towards data mining on construction sites: Heterogeneous data acquisition and fusion
1 INTRODUCTION
2 BACKGROUND
3 PROPOSED SOLUTION
4 EXPERIMENTS
5 RESULTS
6 CONCLUSION
ACKNOWLEDGMENTS
REFERENCES
Multi-view KPConv for enhanced 3D point cloud semantic segmentation using multi-modal fusion with 2D images
1 INTRODUCTION
2 RELATEDWORKS
3 METHODOLOGY
4 RESULTS AND ANALYSIS
5 CONCLUSION AND FUTUREWORKS
REFERENCES
Occupancy assessment for lighting evaluation using digital twin technology
1 INTRODUCTION
2 RESEARCH METHOD
3 PROBLEM IDENTIFICATION ANDMOTIVATION
4 OBJECTIVE AND REQUIREMENTS OF THESOLUTION
5 DESIGN AND DEVELOPMENT
6 DEMONSTRATION
7 EVALUATION
8 CONCLUSION
REFERENCES
Digital twinning of bridges from point cloud data by deep learning and parametric models
1 INTRODUCTION
2 RELATED RESEARCH
3 OVERVIEW ON THE PROPOSED METHOD
4 SEMANTIC SEGMENTATION
5 PARAMETRIC MODELING
6 CONCLUSION
ACKNOWLEDGMENT
REFERENCES
A hybrid top-down, bottom-up approach for 3D space parsing using dense RGB point clouds
1 INTRODUCTION
2 RELATEDWORK
3 METHODOLOGY
4 RESULTS AND DISCUSSION
5 CONCLUSION
REFERENCES
Automated deterministic model-based indoor scan planning
1 INTRODUCTION
2 BACKGROUND
3 METHOD
4 EXPERIMENT AND RESULTS
5 CONCLUSION AND OUTLOOK
ACKNOWLEDGMENTS
REFERENCES
OGM2PGBM: Robust BIM-based 2D-LiDAR localization for lifelong indoor navigation
1 INTRODUCTION
2 THEORETICAL BACKGROUND
3 RELATED RESEARCH
4 METHODOLOGY
5 EXPERIMENTS
6 RESULTS AND ANALYSIS
7 CONCLUSIONS
8 FUTUREWORK
ACKNOWLEDGMENTS
REFERENCES
Leveraging text mining and network analysis for a semi-automated work order process analytics
1 INTRODUCTION & BACKGROUND
2 RESEARCH GAP
3 A BOTTOM-UP APPROACH
4 METHODOLOGY
5 EXAMPLE
6 DYNAMIC MODEL MAINTENANCE
7 SUMMARY AND CONCLUSION
REFERENCES
From theory to ontology: Representing people in building performance simulation models
1 INTRODUCTION
2 ASPECTS OF A HIGH-LEVEL BEHAVIORALTHEORY
3 TOWARD A THEORY-DRIVEN OCO
4 A CONCLUDING NOTE ON THEOPERATIONALIZATION OF THEONTOLOGY
REFERENCES
Semantic enrichment of object associations across federated BIM semantic graphs in a common data environment
1 INTRODUCTION
2 RELATEDWORK
3 CDE FOR SEMANTIC ENRICHMENT
4 ASSOCIATION SEMANTIC ENRICHMENT
5 DISCUSSION
6 CONCLUSIONS
ACKNOWLEDGEMENT
REFERENCES
An ontology-based expert system for quality inspection planning in the construction execution
1 INTRODUCTION
2 BACKGROUND
3 METHOD
4 IMPLEMENTATION OF INSPECTIONREASONING
5 IMPLEMENTATION AND EVALUATION
6 DISCUSSION AND CONCLUSION
REFERENCES
Human factors: Business model
Digitalisation initiative of O&G offshore projects contractual procedures
1 INTRODUCTION
2 BACKGROUND
3 METHOD
4 O&G USE CASE
5 CONCLUSION
ACKNOWLEDGEMENTS
REFERENCES
Can traditional delivery model still fit in BIM procurement? Case study of a New Zealand local government
1 INTRODUCTION
2 LITERATURE REVIEW
3 METHODOLOGY
4 CASE STUDY
5 FINDINGS
6 DISCUSSION
7 CONCLUSION
REFERENCES
Silos and transparency in construction industry materials supply chains
1 INTRODUCTION
2 THEORETICAL FRAMEWORK
3 METHODS
4 RESULTS
5 DISCUSSION AND CONCLUSION
REFERENCES
Designing the business model for the end-of-life phase
1 INTRODUCTION
2 LITERATURE BACKGROUND
3 RESEARCH DESIGN
4 FRAMEWORK FOR BiMEL DEVELOPMENT
5 CONCLUSION
ACKNOWLEDGEMENT
REFERENCES
Data sovereignty within the construction process
1 INTRODUCTION
2 METHODOLOGICAL APPROACH
3 ANALYSIS OF DATA SOVEREIGNTY INPOLITICAL STRATEGIES AND IN LAW
4 DISCUSSION: BALANCING THE INTERESTSWITHIN THE CONSTRUCTION PROCESS
5 IMPLEMENTING DATA SOVEREIGNTYWITHIN THE CONSTRUCTION PROCESS
6 CONCLUSION AND FUTURE RESEARCH
ACKNOWLEDGEMENTS
REFERENCES
Agile implementation for BIM education. Role of the human factor to create Scrum teams
1 INTRODUCTION
2 STATE OF THE ART
3 METHODS
4 RESULTS
5 DISCUSSIONS
6 CONTRIBUTIONS TO THE INDUSTRY
7 CONCLUSION
8 EDUCATION MANIFESTO PROPOSAL
REFERENCES
A framework for meta-disciplinary building analysis
1 INTRODUCTION
2 DISCIPLINARITY
3 META-DISCIPLINARY BUILDING DESIGN
4 ARCHITECTURE FRAMEWORK
5 METHODOLOGY
6 META-DISCIPLINARY OBJECTS
7 DISCIPLINARY NEEDS
8 COMPUTATIONAL METHODS
9 MAPPING METADISCIPLINARY OBJECTSTO COMPUTATIONAL APPROACHES
10 CONCLUSIONS
REFERENCES
Same same, but different – Or how construction informatics gets taught at universities in Norway and Sweden
1 INTRODUCTION
2 SCANDINAVIAN TRADITION
3 METHODOLOGY
4 RESULTS
5 DISCUSSION
6 CONCLUSION
REFERENCES
Architectural education based on integrated design and its effects on professional life
1 INTRODUCTION
2 INTEGRATED DESIGN
3 AN EDUCATION MODAL PROPOSAL:INTERDISCIPLINARY STUDIO 2 ANDINTEGRATED SYSTEM DESIGN COURSESIN IMU
4 ASSESMENT OF THE MODAL
5 CONCLUSION
REFERENCES
Legacy practices supporting BIM adoption in Portugal – Reflections from a large use case
1 BACKGROUND
2 CONTRIBUTION TO THE BODY OFKNOWLEDGE
3 METHODOLOGY
4 BIM AND LEGACY PRACTICES
5 PARQUE ESCOLAR EXAMPLE
6 LESSONS FROM PRONICIMPLEMENTATION
7 CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
Understanding “resistance to change” for BIM adoption and new research ways forward
1 INTRODUCTION
2 LITERATURE REVIEW
3 METHODOLOGY
4 RESULTS AND DISCUSSIONS
5 CONCLUSION
6 REFLECTIONS ON FURTHER RESEARCH
REFERENCES
BIM adoption in small-scale infrastructure projects – Investigation on the German railway sector
1 INTRODUCTION
2 LITERATURE REVIEW
3 METHODOLOGY
4 RESULTS AND DISCUSSIONS
5 CONCLUSION
REFERENCES
Information management: Benefits and challenges of mobilisation
1 INTRODUCTION
2 LITERATURE REVIEW
3 METHODOLOGY
4 FINDINGS
5 DISCUSSIONS
6 CONCLUSION, LIMITATIONS AND FUTURERECOMMENDATION
REFERENCES
Interoperability: Implementation
Systematic investigation of interoperability issues between BIM and BEM
1 INTRODUCTION
2 METHODOLOGY
3 CURRENT STATE OF BIM TO BEMINTEROPERABILITY
4 SYSTEMATIC INVESTIGATION OFINTEROPERABILITY
5 CONCLUSION
ACKNOWLEDGMENT
REFERENCES
Digital support for monitoring cast in-situ concrete
1 INTRODUCTION
2 MATERIAL AND METHOD
3 RESULTS
4 DISCUSSION AND CONCLUSIONS
ACKNOWLEDGEMENTS
REFERENCES
A framework of improved interoperability for VGI3D platform
1 INTRODUCTION
2 NEW FRAMEWORK FOR VGI3D PLATFORM
3 THE IMPROVED INTEROPERABILITY
4 CONCLUSION AND FUTUREWORK
ACKNOWLEDGEMENT
REFERENCES
GreenBIM – Fundamentals for the integration of building greening in openBIM projects
1 INTRODUCTION
2 METHODS
3 RESULTS
4 DISCUSSION
ACKNOWLEDGEMENTS
REFERENCES
A proposed IFC extension for timber construction buildings to enable acoustics simulation
1 INTRODUCTION
2 SOUND ANALYSIS IN TIMBERCONSTRUCTION
3 REQUIREMENTS FOR IFC IN TIMBERCONSTRUCTION
4 JUNCTIONS IN BIM MODEL
5 TECHNICAL MODEL FOR SOUNDANALYSIS
6 CONCLUSION
ACKNOWLEDGMENT
REFERENCES
Analysis and effective use of inverse relation attributes in IFC
1 INTRODUCTION
2 METHOD
3 RESULTS
4 DISCUSSION
5 CONCLUSION
REFERENCES
Extending the IFC-Standard for fire safety building permit
1 INSTRUCTION
2 BACKGROUND
3 BIM IN FIRE PROTECTION
4 DISCUSSION
5 CONCLUSION AND FUTUREWORK
REFERENCES
Conceptual design of a reference process for the transformation of unstructured object catalogs into classification hierarchies
1 INTRODUCTION
2 RELATEDWORK
3 METHODOLOGY
4 IMPLEMENTATION
5 DATA SET
6 RESULTS
7 CONCLUSION
REFERENCES
Research projects
A strategic roadmap for the development of digital platforms in construction: The DigiPLACE strategic roadmap
1 INTRODUCTION AND CONTEXT
2 OVERALL METHODOLOGY
3 DEVELOPMENT OF THE STRATEGICROADMAP SCENARIOS
4 STRATEGIC ROADMAP ANDIMPLEMENTATION ACTION PLAN
5 CONCLUSION
ACKNOWLEDGEMENT
REFERENCES
SmartBuilt4EU: Towards a strategic research and policy agenda for the European smart buildings community
1 INTRODUCTION
2 RESEARCH METHODS
3 TASK FORCE 1: END-USER ACCEPTANCEAND ATTRACTIVENESS OF SMARTSOLUTIONS
4 TASK FORCE 2: INTEROPERABILITY
5 TASK FORCE 3: SMART BUILDINGSPROVIDING FLEXIBILITY TO ELECTRICITYGRIDS
6 TASK FORCE 4: BUSINESS MODELS ANDSMART FINANCING
7 DISCUSSION AND CONCLUSIONS
ACKNOWLEDGMENT
REFERENCES
Author index
Recommend Papers
![Proceedings of SECON’21: Structural Engineering and Construction Management (Lecture Notes in Civil Engineering, 171) [1st ed. 2022]
3030803112, 9783030803117](https://ebin.pub/img/200x200/proceedings-of-secon21-structural-engineering-and-construction-management-lecture-notes-in-civil-engineering-171-1st-ed-2022-3030803112-9783030803117.jpg)
![Mechanical and electrical systems in architecture, engineering, and construction [5th ed]
9780135000045, 0135000041](https://ebin.pub/img/200x200/mechanical-and-electrical-systems-in-architecture-engineering-and-construction-5th-ed-9780135000045-0135000041.jpg)
![Industrial and Civil Construction 2022: Selected Papers (Lecture Notes in Civil Engineering, 436) [1st ed. 2024]
3031444310, 9783031444319](https://ebin.pub/img/200x200/industrial-and-civil-construction-2022-selected-papers-lecture-notes-in-civil-engineering-436-1st-ed-2024-3031444310-9783031444319.jpg)
![Advances in Product Design Engineering (Management and Industrial Engineering) [1st ed. 2022]
9783030981235, 9783030981242, 3030981231](https://ebin.pub/img/200x200/advances-in-product-design-engineering-management-and-industrial-engineering-1st-ed-2022-9783030981235-9783030981242-3030981231.jpg)
- Author / Uploaded
- Eilif Hjelseth
- Sujesh F. Sujan
- Raimar J. Scherer
- Similar Topics
- Science (general)
- International Conferences and Symposiums
File loading please wait...
Citation preview
ECPPM 2022 – EWORK AND EBUSINESS IN ARCHITECTURE, ENGINEERING AND CONSTRUCTION
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction contains the papers presented at the 14th European Conference on Product & Process Modelling (ECPPM 2022, Trondheim, Norway, 14–16 September 2022), and builds on a long-standing history of excellence in product and process modelling in the construction industry, which is currently known as Building Information Modelling (BIM). The following topics and applications are given attention: • • • • • • •
• •
Sustainable and Circular Driven Digitalisation: Data Driven Design and/or Decision Support Assessment and Documentation of Sustainability Information lifecycle Data Management: Collection, Processing and Presentation of Environmental Product Documentation (EPD) and Product Data Templates (PDT) Digital Enabled Collaboration: Integrated and Multi-Disciplinary Processes Virtual Design and Construction (VDC): Production Metrics, Integrated Concurrent Engineering, Lean Construction and Information Integration Automation of Processes: Automation of Design and Engineering Processes, Parametric Modelling and Robotic Process Automation Expert Systems: BIM based model and compliance checking Enabling Technologies: Machine Learning, Big Data, Artificial and Augmented Intelligence, Digital Twins, Semantic Technology Sensors and IoT Production with Autonomous Machinery, Robotics and Combinations of Existing and New Technical Solutions Frameworks for Implementation: International Information Management Series (ISO 19650), and Other International Standards (ISO), European (CEN) and National Standards, Digital Platforms and Ecosystems Human Factors in Digital Application: Digital Innovation, Economy of Digitalisation, Client, Organisational, Team and/or Individual Perspectives
Over the past 25 years, the biennial ECPPM conference proceedings series has provided researchers and practitioners with a unique platform to present and discuss the latest developments regarding emerging BIM technologies and complementary issues for their adoption in the AEC/FM industry.
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
PROCEEDINGS OF THE 14th EUROPEAN CONFERENCE ON PRODUCT AND PROCESS MODELLING (ECPPM 2022), TRONDHEIM, NORWAY, SEPTEMBER 14–16, 2022
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction
Edited by
Eilif Hjelseth Norwegian University of Science and Technology, Trondheim, Norway
Sujesh F. Sujan Norwegian University of Science and Technology, Trondheim, Norway
Raimar J. Scherer University of Technology, Dresden, Germany
First published 2023 by CRC Press/Balkema 4 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN e-mail: [email protected] www.routledge.com – www.taylorandfrancis.com CRC Press/Balkema is an imprint of the Taylor & Francis Group, an informa business © 2023 selection and editorial matter, Eilif Hjelseth, Sujesh F. Sujan & Raimar J. Scherer; individual chapters, the contributors The right of Eilif Hjelseth, Sujesh F. Sujan & Raimar J. Scherer to be identified as the authors of the editorial material, and of the authors for their individual chapters, has been asserted in accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988. All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form or by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying and recording, or in any information storage or retrieval system, without permission in writing from the publishers. Although all care is taken to ensure integrity and the quality of this publication and the information herein, no responsibility is assumed by the publishers nor the author for any damage to the property or persons as a result of operation or use of this publication and/or the information contained herein. Library of Congress Cataloging-in-Publication Data A catalog record has been requested for this book ISBN: 978-1-032-40673-2 (hbk) ISBN: 978-1-032-40674-9 (pbk) ISBN: 978-1-003-35422-2 (ebk) DOI: 10.1201/9781003354222 Typeset in Times New Roman by MPS Limited, Chennai, India
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Editor(s), ISBN 978-1-032-40673-2
Table of Contents
Preface Committee Members Organizing Committee
x xiii xv
Sustainable and circular driven digitalisation Life-cycle assessment Sustainability assessment of a novel reusable and demountable steel-concrete composite floor system J. Fodor, A. Akbarieh, M. Schäfer & F.N. Teferle
3
Probabilistic life cycle analysis as a sustainability-focused design tool for industrialized construction T. Hegarty & M. Lepech
11
Integrating Level(s) LCA in BIM: A tool for estimating LCA and LCC impacts in a case study M.T.H.A. Ferreira, J.D. Silvestre, A.A. Costa, H.B. & R.A. Bohne How can LCA inform early-stage design to meet Danish regulations? The sustainability opportunity metric A. Kamari & C. Schultz Life cycle potentials and improvement opportunities as guidance for early-stage design decisions J. Staudt, M. Margesin, C. Zong, F. Deghim, W. Lang, A. Zahedi, F. Petzold & P. Schneider-Marin Structure and LCA-driven building design support in early phases using knowledge-based methods and domain knowledge D. Steiner, M. Schnellenbach-Held, J. Staudt, M. Margesin, C. Zong & W. Lang
19
27 35
43
Processes Challenges and experiences with the reuse of products in building design A. Tomczak, M. Łuczkowski, E. Hjelseth & S.F. Sujan Evaluating existing digital platforms enabling the reuse of reclaimed building materials and components for circularity W. Wuyts, Y. Liu, X. Huang & L. Huang
53
60
Product data Semantic Material Bank: A web-based linked data approach for building decommissioning and material reuse A. Akbarieh, F.N. Teferle & J. O’Donnell NLP-based semantic model healing for calculating LCA in early building design stages K. Forth, J. Abualdenien & A. Borrmann Construction product identification and localization using RFID tags and GS1 data synchronization system A. Glema, M. Ma´ckowiak & Z. Rusinek What comes first when implementing data templates? Refurbishment case study P. Mˆeda, D. Calvetti, H. Sousa & E. Hjelseth
v
69 77
85 91
Chaos and black boxes – Barriers to traceability of construction materials K. Mohn & J. Lohne Evaluating four types of data parsing methods for machine learning integration from building information models F. Sellberg, J. Buthke, P.F. Sonne-Frederiksen & P.N. Gade
99
106
Digital supported collaboration Multimodel Extending ICDD implementation to a dynamic multimodel framework N. Al-Sadoon, P. Katranuschkov & R.J. Scherer
115
Management of BIM-based digital twins in multimodels M. Polter & R.J. Scherer
123
Processes Enriching BIM-based construction schedules with semantics using BPMN and LBD P. Hagedorn, K. Sigalov, L. Höltgen, M. Müller, T. Sola & M. König
133
A simulative framework to evaluate constructability through parameter optimization at early design stage F.L. Rossini & G. Novembri
141
Automatic generation of work breakdown structures for evaluation of parallelizability of assembly sequences J.M. Weber & M. König
147
Virtual design & construction Construction process time optimization of a reinforced concrete reaction slab – Implementing the VDC methodology M. Barcena, M.C. Borja & A.A. Del Savio VDC framework proposal for curtain wall construction process optimization V. Bustamante, J.P. Cedrón & A.A. Del Savio Using production metrics to compare and understand VDC elements practiced by general contractors in Norway T. Majumdar, M.A. Fischer, S.G. Rasmussen, K. Johannesdottír & E. Hjelseth
157 165
173
Digital twin Building performance A-EYE tech: Framework to evaluate an AI construction visibility platform A. Hassan, A. Hore & M. Mulville
181
Digital twin in healthcare facilities: Linking real-time air quality data to BIM A. Harode, W. Thabet & M. B DuLaney
188
A novel BIM platform paradigm for the building erformance domain D. Utkucu & R. Sacks
196
Towards European standardization of digitalization approaches for monitoring and safety of bridges and tunnels R. Sebastian, M. Weise, N. Mitsch, I. Giurgiu & A. Sánchez Rodríguez
204
Facilities management TwinGen: Advanced technologies to automatically generate digital twins for operation and maintenance of existing bridges S. Vilgertshofer, M.S. Mafipour, A. Borrmann, J. Martens, T. Blut, R. Becker, J. Blankenbach, A. Göbels, J. Beetz, F. Celik, B. Faltin & M. König
vi
213
Digital Twin for AECOO – Framework proposal and use cases D. Calvetti, P. Mˆeda, E. Hjelseth & S.F. Sujan
221
OpenSantCugat: A platform for municipalities to provide access to building data L. Madrazo, Á. Sicilia, E. Ortet & A. Calvo
229
Digital management for unforeseen trigger events using ISO 19650 B. Godager & K. Mohn
235
A lean strategic FM service model based on the digital twin A. Mbabu, J. Underwood & M. Munir
242
A Digital Twin prototype for smart parking management Y. Zou, F. Ye, A. Li, M. Munir, E. Hjelseth & S.F. Sujan
250
BIM integrated data management for the Operation and Maintenance (O&M) of railway R. Tao, Y. Pan, A. Lau & M.E. Mossefin
258
How to represent damage data in BIM Models? – A literature review W. Teclaw, K. Gradeci, N. Labonnote & E. Hjelseth
263
Correlation and comparison between digital twin and cyber physical systems A.U. Khan & L. Huang
268
Heritage HBIM application in historic town: A scoping literature review B.N. Prabowo, E. Hjelseth & A. Temeljotov-Salaj
277
The use of HBIM and scanning in cultural heritage projects C.N. Rolfsen, A.K. Lassen, S.M. Lein & E.Z. Zabrodina
285
Enabling technologies Energy High-order second-level space boundary surface calculation for building energy performance simulation models G.N. Lilis, K. Katsigarakis & D. Rovas
295
Aspects of BIM-to-BEM information transfer: A tale of two workflows O. Spielhaupter & A. Mahdavi
303
Predicting annual heating energy use using BIM: A simplified method for early design phase M.F. Stendahl & M.C. Dubois
311
Machine learning Predicting semantic building information (BIM) with Recurrent Neural Networks B. Mete, J. Bielski, C. Langenhan, F. Petzold, V. Eisenstadt & K.D. Althoff
321
International assessment of artificial intelligence applications in the AEC sector D. Cisterna, C. Lagos & S. Haghsheno
327
Predicting occupant evacuation times to improve building design J. Clever, J. Abualdenien, R.K. Dubey & A. Borrmann
335
Inferring interconnections of construction drawings for bridges using deep learning-based methods B. Faltin, P. Schönfelder & M. König
343
An ontology-supported case-based reasoning approach for damage assessment A.H. Hamdan & R.J. Scherer
351
Towards a semantic enriching framework for construction site images C. Zeng & T. Hartmann
358
Machine learning and genetic algorithms for conformal geometries in design support systems S.J.F. van Hassel, H. Hofmeyer, T. Ezendam & P. Pauwels
366
vii
Predicting the construction duration in the predesign phase with decision trees S. Lauble, S. Haghsheno & M. Franz
374
Tower crane layout planning through Generative Adversarial Network R. Li, H.L. Chi, Z. Peng & J. Chen
382
Consideration of detailing in the graph-based retrieval of design variants D. Napps & M. König
389
Retrieve information from construction documents with BERT and unsupervised learning M. Shi, T. Heinz & U. Rüppel
397
Automated floorplan generation using mathematical optimization Y. Zhong & A. Geiger
404
Model checking Information extraction and NLP for the interpretation of building permits: An Italian case study S. Comai, E. Agrawal, G. Malacarne, M. Sadak, S.M. Ventura & A.L.C. Ciribini
413
Process-based building permit review – A knowledge engineering approach J. Fauth, W. Müller & S. Seiß
421
Training on digitized building regulations for automated rule extraction S. Fuchs, J. Dimyadi, M. Witbrock & R. Amor
428
Using RASE semantic mark-up for normative, definitive and descriptive knowledge N. Nisbet & L. Ma
436
How practice is represented in BIM-based model checking research – A literature review and reflections P.N. Gade, D.H. Lauritzen, M. Andersen & E. Hjelseth
443
Model healing: Toward a framework for building designs to achieve code compliance J. Wu, R.K. Dubey, J. Abualdenien & A. Borrmann
450
A multi-representation method of building rules for automatic code compliance checking Z. Zhang, N. Nisbet, L. Ma & T. Broyd
458
Processes Development of augmented BIM models for built environment management L. Binni, B. Naticchia, A. Corneli & M. Prifti
469
BIM-based disaster response: Facilitating indoor path planning for various agents A. Dugstad, R.K. Dubey, J. Abualdenien & A. Borrmann
477
Suitability of three national cost breakdown structures for automated quantity take-off in road projects D. Fürstenberg, A. Nast & P. Mˆeda
485
Modular robotic system for the construction industry S. Han & K. Menzel
493
Designing a framework for seamless integration of open data services to support disaster response T.J. Huyeng, T. Bittner & U. Rüppel
500
Practical experiences with 5D building information modeling – A systematic literature review A. Nast & C. Koch
508
Towards data mining on construction sites: Heterogeneous data acquisition and fusion F. Pfitzner, A. Braun & A. Borrmann
516
Scanning Multi-view KPConv for enhanced 3D point cloud semantic segmentation using multi-modal fusion with 2D images C. Du, M.A. Vega, Y. Pan & A. Borrmann
viii
527
Occupancy assessment for lighting evaluation using digital twin technology P. Johansson, G. Fischl & K. Hammar
535
Digital twinning of bridges from point cloud data by deep learning and parametric models M.S. Mafipour, S. Vilgertshofer & A. Borrmann
543
A hybrid top-down, bottom-up approach for 3D space parsing using dense RGB point clouds M. Mehranfar, A. Braun & A. Borrmann
551
Automated deterministic model-based indoor scan planning F. Noichl & A. Borrmann
559
OGM2PGBM: Robust BIM-based 2D-LiDAR localization for lifelong indoor navigation M.A. Vega Torres, A. Braun & A. Borrmann
567
Semantic technology Leveraging text mining and network analysis for a semi-automated work order process analytics S. Sobhkhiz & T. El-Diraby
577
From theory to ontology: Representing people in building performance simulation models A. Mahdavi, D. Wolosiuk & C. Berger
583
Semantic enrichment of object associations across federated BIM semantic graphs in a common data environment B. Ouyang, Z. Wang & R. Sacks An ontology-based expert system for quality inspection planning in the construction execution S. Seiß
591 599
Human factors Business model Digitalisation initiative of O&G offshore projects contractual procedures D.L.M. Nascimento, A.B. Roeder, F.N.M. Araújo, C.L. Bechtold & D. Calvetti Can traditional delivery model still fit in BIM procurement? Case study of a New Zealand local government J.N. Jiang, T.F.P. Henning & Y. Zhou
609
615
Silos and transparency in construction industry materials supply chains J. Lohne & K. Mohn
622
Designing the business model for the end-of-life phase M. Sre´ckovi´c & G. Šibenik
630
Data sovereignty within the construction process B. Weber & M. Achenbach
637
Education Agile implementation for BIM education. Role of the human factor to create Scrum teams D. Delgado Vendrell & O. Liebana Carrasco
647
A framework for meta-disciplinary building analysis T. McGinley & T. Krijnen
654
Same same, but different – Or how construction informatics gets taught at universities in Norway and Sweden C. Merschbrock, E. Onstein, P.E. Danielsen & P. Johansson Architectural education based on integrated design and its effects on professional life Z. Yazıcıo˘glu & A. Dikba¸s
ix
662 669
Implementation Legacy practices supporting BIM adoption in Portugal – Reflections from a large use case P. Mˆeda, J. Teixeira, D. Calvetti, Y. Ribeiro, J. Moreira & H. Sousa
679
Understanding “resistance to change” for BIM adoption and new research ways forward A. Nast & A. Rekve
687
BIM adoption in small-scale infrastructure projects – Investigation on the German railway sector A. Nast & C. Koch Information management: Benefits and challenges of mobilisation J.T. Shukla & M. Bolpagni
695 703
Interoperability Implementation Systematic investigation of interoperability issues between BIM and BEM M. Afzal, K. Widding, E. Hjelseth & M. Hamdy
713
Digital support for monitoring cast in-situ concrete G. Kjellmark, G. Pe˜naloza & E. Hjelseth
721
A framework of improved interoperability for VGI3D platform G. Kong, E. Hjelseth, H. Fan & G. Lobaccaro
729
GreenBIM – Fundamentals for the integration of building greening in openBIM projects J. Murschetz, M. Monsberger, B. Knoll, R. Graf, A. Renkin, R. Dopheide, F. Schiefermair & J. Kräftner
736
Technical A proposed IFC extension for timber construction buildings to enable acoustics simulation C. Chˆateauvieux-Hellwig, J. Abualdenien & A. Borrmann
745
Analysis and effective use of inverse relation attributes in IFC H. Tauscher
753
Extending the IFC-Standard for fire safety building permit J. Walter & J. Díaz
760
Conceptual design of a reference process for the transformation of unstructured object catalogs into classification hierarchies S. Zentgraf & M. König
766
Research projects A strategic roadmap for the development of digital platforms in construction: The DigiPLACE strategic roadmap C. Mirarchi, A. Pavan, C. Gatto & S. Angotti
777
SmartBuilt4EU: Towards a strategic research and policy agenda for the European smart buildings community C. Coujard, K.L. Eloire, A. Zarli & A. David
785
Author index
795
x
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Editor(s), ISBN 978-1-032-40673-2
Preface
The European Conference on Product and Process Modelling (ECPPM) was held in Trondheim from 14th to 16th September 2022, signifying the first physical conference after 4 years due to the Covid-19 pandemic. During this period, the importance of digital solutions to keep society and the AEC/FM industry running was evident. The experiences show that digitalization is an enabler and an opportunity to transform the way we work and collaborate. The use of digital technology to communicate in virtual meetings or collaborate in real-time by use of web-based services in design and coordination software are some of the many examples. Without these digital solutions, both society and the AEC/FM industry would have lost significant productivity due to the Covid-19 pandemic. However, physical meetings and unstructured discussions have a unique capability to provide new perspectives and ideas in an organic way different to structured virtual meetings. The participation from researchers around the globe indicates the need for conferences with physical attendance that focus on relationship building and sharing knowledge organically. Afterall, humans are a social species that rely on physical interaction to thrive. The challenge for the future is to create a mix of virtual and physical that utilises the best of both approaches. Another effect the pandemic had on ECPPM 2022 was in travel uncertainty. This challenge was overcome by planning the submission deadline as close as possible to the conference, enabling late updates and submissions to be presented. Proceedings in printed book and online format were made available after the conference. Furthermore, all presentations are published on the conference website (https://www.ecppm2022.org) enabling post-conference activities and networking. ECPPM 2022 introduced a presentation template which focused on the most important elements for the listener, as opposed to the traditional structure of the paper. In addition, longer breaks between the sessions were introduced to enable networking digitally supported using a conference app which included information/bibliography about each researcher, their affiliation and abstract. ECPPM 2022 had Keynote speeches at the beginning and end of each day which was included to give all participants joint understanding and perspectives for further discussions. Thanks to our speakers including, Raimar Scherer, Arto Kiviniemi, Leif Granholm, Ardeshir Mahdavi and Léon van Berlo for their inspiring keynote presentations. The conference included a Research Design and Paper writing workshop in two parts led by Professor Mirosław Skibniewski. The workshop comprised of an excellent combination of presentations and discussions with participants on scientific writing and publication. Additionally, a workshop focusing on completed, ongoing or planned research projects was included with the intention of using the network effect to take advantage of the variation of researchers attending ECPPM. Whilst ECPPM is a scientific conference, the dominating purpose in the research projects is to support the AEC/FM industry both in the short and long-term. Likewise, the AEC/FM industry need to communicate their challenges to the research community. An “Industry Day” was included in the middle of the conference to serve this purpose. A special thanks to Gabrielle Bergh (Norwegian University of Life Sciences) who supported the organisation of the Industry day. The day was structured with three thematic tracks: – TheVDC (Virtual Design and Construction) track included a thematic speech by Prof. Martin Fischer, followed up by a workshop in the holistic implementation of VDC in entire organisations. The track concluded with best practice presentations from the industry. – The industry implementation track included presentations from Emma Hooper (Bond Bryan Digital) about ISO 19650 Information Management and Cathrine Mørch (Statsbygg - the Norwegian building authority) about innovation and implementation in industry. – The sustainability track was presented by John Hainsworth (Mott MacDonald) about sustainability-driven digital changes in the industry and Trine Dyrstad Pettersen (Byggevareindustrien) about the practical implementation of Environmental Product Declarations (EPD). These variations of conference activities indicate that ECPPM is in development and adapting to serve as a platform for multiple organisations (e.g. industry and academia) to contribute towards a scientific event encompassing digitalisation. Whilst ECPPM has a long history with a good foundation, ECPPM intends to
xi
evolve to remain the most influential series of conferences within digitalisation. ECPPM 2022 is an example of this evolution. The principal core in digitalisation is PRODUCT AND PROCESS MODELLING, where “and” is emphasized in this conference. This balance is not represented in the profile of papers, approximately 80% focus on technology-oriented product perspective. Process modelling represents a fundamental way to express current and new ways of working and collaborating. Until now, it has been difficult to justify positive economic returns of investments in digital solutions. Digital in principle should be a balanced integration of people, process and technology. Research on new processes connecting people and technology is limited, whilst definition of points of departure are based mainly on literature without direct connection to changes in practice. One way to improve impact and relevance is by integrating product and process modelling in research by defining the point of departure and implementation with industry. Engaging with industry through research can improve validity and relevance whilst encouraging multidisciplinary competency. Including industry in the ECPPM community can contribute to significant improvement in research outcomes and how this outcome is defined, explored, communicated and implemented. ECCPM recognised Digital Twin as one of the most discussed topics in the industry and was a dominating topic in the conference. The Digital Twin term has been used as a concept in multiple ways to enable communication between the built environment and a digital model. Digital twin challenges the traditional understanding of BIM and enables a more dynamic approach facilitating a lifecycle perspective as the default mindset. For the industry, this means many new services serving various stakeholders across the lifecycle. ECCPM 2022 has demonstrated the increased focus on sustainability, signifying the opportunity for the entire AEC/FM industry (as well as the rest of society) to improve use of digitalisation. Decisions must be based on facts including societal, economical, environmental data to preserve nature for future generations. The AEC/FM industry is responsible for approximately 40% of energy consumption, waste and material use globally. Even if the exact numbers can be variant based on other factors, there is no doubt about the significance of the role that the AEC/FM industry has to play to achieve various sustainability and circularity targets. At the moment, the systemic contributions from industry are limited, with pockets of excellent practice scattered across the sector. These pockets of excellence need to spread across industry. This means, research communities like ECPPM have a vital role to play in achieving and defining the processes and products used in industry to enable integrated solutions for community wide use cases. One way to bridge industry and academia is to support the development of international standards. Sustainability is the integration of societal, economic, environmental perspectives. The United Nations has established 17 sustainability goals. The AEC/FM industry has normally focused on the economical sustainability goals; 8, 9, 10 and 12. However, ECPPM have given special attention to maybe the most important; 17 – Partnership for the Goals. This goal is often under communicated or understood as just an intention to collaborate. In ECPPM 2022 we profile this as the need for integration of relevant structured information from multiple sources to enable versatile use. Without a joint understanding supported by relevant information, it is not possible to make balanced decisions. This is an interesting and motivating time in the domain but also demanding and complex, both related to the way that research communities such as ECPPM operate and produce research impact. A special thanks to the chairman and founder of the ECPPM conference Prof. Raimar Scherer. His long-time efforts were awarded (in person) an “honorary chairman” at the EAPPM assembly meeting. I would also like to give a special acknowledgement of my colleague Dr. Sujesh F. Sujan for his effort in planning and connection to the conference, as well for his excellent scientific contribution in the editorial team. Thanks to all members of the scientific committee whose review of both abstracts and paper with supporting the authors to increase the scientific quality of the papers. Last but not the least, thanks to all researchers for choosing ECPPM 2022 in Trondheim to network and present their recent research. ECPPM 2024 will be held in Dresden to commemorate the 30th anniversary of ECPPM series of conferences within product and process modelling, the core in all digitalisation initiatives. Eilif Hjelseth, Professor Conference Host and Chair
xii
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Editor(s), ISBN 978-1-032-40673-2
Committee Members
Raimar Scherer, Germany Jan Karlshøj, Denmark Symeon Cristodulou, Cyprus Secretary Ziga Turk, Slovenia Committee Members Robert Amor, New Zealand Ezio Arlati, Italy Vladimir Bazjanac, Croatia Jakob Beetz, Netherlands Adam Borkowski, Poland Jan Cervenka, Czech. Rep Symeon Crhistodulou, Cyprus Attila Dikbas, Turkey Djordje Djordjevic, Serbia Boyan Georgiev, Bulgaria Ricardo Gonçalves, Portugal Gudni Gudnason, Iceland Eilif Hjelseth, Norway Noemi Jimenez Redondo, Spain Jan Karlshøj, Denmark Tuomas Laine, Finland Ardeshir Mahdavi, Austria Karsten Menzel, Ireland Sergio Munoz, Spain Pieter Pauwels, Belgium Byron Protopsaltis, Greece Svetla Radeva, Bulgaria Yacine Rezgui, UK Dimitrios Rovas, Greece Vaidotas Sarka, Lithuania Raimar Scherer, Germany Ian Smith, Switzerland Rasso Steinmann, Germany Vänio Tarandi, Sweden Alain Zarli, France
Chairperson Vice Chairperson
xiii
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Editor(s), ISBN 978-1-032-40673-2
Organizing Committee
Hjelseth, Eilif, Norwegian University of Science and Technology, Trondheim, Norway (Scientific Chair) Sujan, Sujesh Francis, Norwegian University of Science and Technology, Trondheim, Norway (Conference Coordinator) Bergh, Gabrielle, Norwegian University of Life Sciences, Oslo, Norway (Industry Day Coordinator) Amundsen, Monica, Norwegian University of Science and Technology, NTNU VIDERE, Trondheim, Norway Almaas, Hege Elisabeth, Norwegian University of Science and Technology, NTNU VIDERE, Trondheim, Norway Scientific Committee Aksenova, Gulnaz, UCL, United Kingdom Amor, Robert, University of Auckland, New Zealand Beetz, Jakob, RWTH Aachen University, Germany Bergh, Gabrielle, Norwegian University of Life Sciences, Norway Bohne, Rolf André, Norwegian University of Science and Technology, Norway Boje, Calin, Luxembourg Institute of Science and Technology, Luxembourg Bolpagni, Marzia, Mace, United Kingdom Borrmann, Andre, Technical University of Munich, Germany Boton, Conrad, École de Technologie Supérieure, Canada Coates, Stephen, University of Salford, United Kingdom Drogemuller, Robin, Queensland University of Technology, Australia El-Diraby, Tamer, University of Toronto, Canada Ergen, Esin, Istanbul Technical University, Turkey Fauth, Judith, RIB Software, Germany Fürstenberg, David, COWI, Norway Godager, Bjørn Arild, Norwegian University of Science and Technology, Norway Graham, Matt, Mott MacDonald, United Kingdom Guerriero, Annie, Luxembourg Institute of Science and Technology, Luxembourg Halin, Gilles, University of Lorraine, France Hamdy, Mohamed, Norwegian University of Science and Technology, Norway Hartmann, Timo, Technische Universität Berlin, Germany Hassan, Tarek, Loughborough University, United Kingdom Hjelseth, Eilif, Norwegian University of Science and Technology, Norway Houck, Leif Daniel, Norwegian University of Life Sciences, Norway Johansen, Agnar, Norwegian University of Science and Technology, Norway Karlshøj, Jan, Technical University of Denmark, Denmark Katranuschkov, Peter, TU Dresden, Germany Kazi, Sami, VTT Technical Research Centre of Finland, Finland Kiani, Kaveh, University of Salford, United Kingdom Kifokeris, Dimosthenis, Chalmers University of Technology, Sweden Kiviniemi, Arto, University of Liverpool, Finland Klakegg, Ole Jonny, Norwegian University of Science and Technology, Norway Koenig, Markus, Ruhr-University Bochum, Deutschland Kubicki, Sylvain, Luxembourg Institute of Science and Technology, Luxembourg Kumar, Bimal, Northumbria University, United Kingdom Lædre, Ola, Norwegian University of Science and Technology, Norway Lohne, Jardar, Norwegian University of Science and Technology, Norway Makarfi, Usman Umar, University of Salford, United Kingdom Mastrolembo Ventura, Silvia, University of Brescia, Italy Mêda, Pedro, Ic – Instituto Da Construção, Portugal Mejlænder-Larsen, Øystein, Multiconsult, Norway Merschbrock, Christoph, Norwegian University of Science and Technology, Norway Mirarchi, Claudio, Politecnico di Milano, Italy
xv
Munir, Mustapha Yakubu, University of Salford, United Kingdom Munoz, Sergio, buildingSMART Spain, Spain Mutis, IVAN, Illinois Institute of Technology, United States Nast, Anna, Bauhaus University Weimar, Germany Nisbet, Nicholas, University College London and AEC3, United Kingdom Nørkjær Gade, Peter, University College Nordjylland, Denmark O’Donnell, James, University College Dublin, Ireland Onstein, Erling, Norwegian University of Science and Technology, Norway Pauwels, Pieter, Eindhoven University of Technology, Netherlands Petrova, Ekaterina, Eindhoven University of Technology, Netherlands Pinti, Lidia, Politecnico di Milano, Italy Rekve, Anders, Norwegian University of Science and Technology, Norway Rolfsen, Christian Nordahl, Oslo Metropolitan University, Norway Rossini, Francesco Livio, La Sapienza, Italy Rovas, Dimitrios, University College London, United Kingdom Roxin, Ana, University of Burgundy, France Santos, Eduardo Toledo, Universidade de São Paulo, Brazil Schapke, Sven-Eric, Thinkproject, Germany Scherer, Raimar Josef, TU Dresden, Germany Simeone, Davide, Webuild, Italy Smith, Ian, TU Munich, Germany Sujan, Sujesh Francis, NTNU/Mott MacDonald, United Kingdom Svidt, Kjeld, Aalborg University, Denmark Tarandi, Väino Kristjan, V Tarandi AB, Sweden Tauscher, Helga, HTW Dresden – University of Applied Sciences, Germany Temeljotov-Salaj, Alenka, Norwegian University of Science and Technology, Norway Turk, Žiga, University of Ljubljana, Slovenia van Berlo, Léon, buildingSMART International, United Kingdom Zarli, Alain, ECTP, France Zou, Yang, The University of Auckland, New Zealand
xvi
Sustainable and circular driven digitalisation Life-cycle assessment
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Sustainability assessment of a novel reusable and demountable steel-concrete composite floor system J. Fodor, A. Akbarieh, M. Schäfer & F.N. Teferle Department of Engineering, University of Luxembourg, Luxembourg, Grand Duchy of Luxembourg
ABSTRACT: Raw materials extraction, production of components, transportation and reverse logistics activities that run in the construction sector are constantly depleting the available global resources. Sustainability of the construction industry and its ability to adopt to the principles of circular economy is under question. This paper addresses these questions through the introduction of a novel reusable steel-concrete composite floor system. Its reuse potential is evaluated through comparative BIM-based Life Cycle Analysis with contemporary systems.
1 INTRODUCTION One of the negative contributors to climate change is unsustainable material extraction and consumption. Not only that it leads to resource depletion, but also the propagation of landfills all over the earth at the end of the Life Cycle (LC) of materials and components. The construction industry is among the highest material-intensive industries with a huge waste output. Global data shows that about 40% of energy and process-related carbon dioxide (CO2 ) emissions (IEA 2019), as well as 50% of material extraction (European Commission 2022) and 35% of all waste in the European Union, are attributed to the construction industry (Eurostat 2018). A glance at this data indicates that not only unsustainable construction is detrimental to the future trajectory of the planet, but also demolition and waste generation will put restraints on Earth’s resources. That is why deconstruction and Design for Deconstruction (DfD) strategies are promoted as means to eliminate unnecessary waste from construction activities. Deconstruction (or disassembly) as opposed to demolition is the act of non-destructive removal of construction components (ISO 2022). In this way, components could be reused in another project. Therefore, the LC is extended, which lowers the carbon footprint. In the same line, DfD is encouraged to be considered in the early design stage to design out waste. DfD is a way to design a product or asset in a way that disassembly is facilitated at the end of its useful life, with components or parts to be further reused, if not, recycled or recovered for energy, hence, diverted from the waste stream (ISO 2022). The explained methods are ways that the construction industry can circularize its consumption and progress from linear business models to circular economybased business models. The circular economy aims to keep products, components, or materials at their DOI 10.1201/9781003354222-1
highest quality in the value chain as long as possible (European Commission 2019), therefore creating subcycles for further reuse enabling the re-introduction of materials into the market while avoiding waste as much as possible. Having materials employed in the value chain can hugely impact resource extraction and manufacturing activities, which could lower carbon emissions and other negative ecological impacts. In this paper, a novel reusable and demountable steel-concrete composite floor system is presented that is designed for deconstruction and multiple reuses.The structural layout of novel design is presented in section 2. For the purpose of sustainability assessment of the novel demountable design, the scope and functional unit of the Life Cycle Assessment (LCA) are defined in the section 3. In order to set the scene for LCA comparative analysis, in section 4 the novel and contemporary floor systems are structurally analyzed and designed. In section 5, the LCA methodology and results are discussed, followed by conclusion of the key findings in section 6.
2 NOVEL REUSABLE STEEL-CONCRETE FLOOR 2.1 Floor layout The demountable floor system comprises modular one-way concrete solid slab modules connected with the secondary steel beams by applying the novel demountable, friction-based shear connector device in order to establish the shear connection between the steel section and the concrete slab (Fodor 2022). The one-way solid slab modules are represented in two types (simply supported and continuous) and are placed in horizontally offset layout (Figure 1) in
3
Figure 1. Layout of the demountable floor system. Figure 2. Layout of the shear connector device.
order to improve the inplane behaviour of the composite floor. The composite action is achieved only between the secondary beams and the modular concrete solid slab (the primary beams are steel beam elements) as due to the rail channel orientation it is only possible to provide the shear connection in one direction (the longitudinal direction of the secondary beam). The demountability of the shear connector allows for the demountability of the whole steel-concrete composite flooring system and reusability of its main components (modular slab elements, primary and secondary steel beams). The mechanical properties of the novel shear connector allow for predictability of the mechanical response of the steel-concrete composite floor (composite beam) in accordance with EN 1994-1-1 (CEN 2004) that governs the design of the steel-concrete composite structures.
Having the embedded rail channel as an element that transfers the connector shear force into the concrete slab, the concrete related failure modes are mitigated (Fodor 2022). At the failure load of the shear connector the concrete in the vicinity of the load introduction remains uncracked, hence the one-way concrete slab module might be reused. The addition of disc-springs in the connector assembly prevents pretension loss due to the possible embedment related contraction of the clamping package. This is especially important in the sliding part of the connector force-slip response. The connector device exhibits high initial stiffness, comparable shear capacity to the one of an equivalent headed stud connector and outstanding slip capacity (ductility) what allows for plastic design of the composite beam coupled with the uniform distribution of the longitudinal shear (equidistant arrangement of the shear connectors). In order to utilize efficiently the capacity of the shear connector in the longitudinal shear connection, the ductility requirement is a consequence of the modularity of the floor system as in order to have modular solid slab elements, the distance between connectors (embedded rails) in the direction of the beam span has to be equal. The high initial stiffness of the shear connector (Figure 4) allows that the composite beam response adheres to the complete interaction elastic behaviour in SLS load ranges if the full degree of shear connection is applied (η = 1.00). This way the stiffness of the composite beam may be estimated assuming complete interaction composite beam behaviour based on the idealized moment of inertia of the composite cross section to validate the serviceability limit states (SLS). At the ultimate limit state (ULS) concerning the bending capacity of the composite cross section and the design of the longitudinal shear connection, the high ductility of the shear connector (Figure 3) allows for uniform distribution of longitudinal shear coupled with the plastic capacity of the composite cross section Mpl,rd at the beam midspan.
2.2 Connector device The shear connector device is comprised of highstrength pretension bolt assembly (HV), steel rail plate that fits into the rail channel socket that is embedded into the concrete matrix of the one-way solid slab module, zinc hot-dip galvanized back plate, disc springs and large diameter washers that are necessary to accommodate the pretension force distribution between the nut and the clamping package (Figure 2). The ease of installation is guaranteed by the large construction tolerances in the direction of beam span (slotted bolthole) and in the transverse direction (rail channel socket). Applying the pretension force into the bolt assembly the pretension force is transformed into the clamping force in its entirety. Two distinctive faying surfaces may be identified. The first faying surface is formed between the bottom surface of the slab module and the steel section while the second faying surface is formed between the bottom surface of the top flange of the steel section and the backplate (Figure 2). The friction resistance, as a result of the pretension and the frictional properties of the faying surfaces, is responsible for the provision of the longitudinal shear connection along the length of the composite beam.
4
the best case recycled in their entirety. In this analysis it is assumed that all the comprising parts of these solutions are recycled as the end-of-life scenario. In the last case (III) the main structural components of the structural flooring system, one-way concrete solid slab modules, secondary and primary steel beams are suitable for further reuse. The goal of this analysis is to prove that after several reuse cycles of the critical structural components the environmental impact of the novel demountable composite steel-concrete flooring system is significantly lower than the one inflicted by employing the contemporary non-demountable solutions (I–II).
Figure 3. Force-slip behaviour of the shear connector.
3 STRUCTURAL FUNCTIONAL UNIT FOR COMPARATIVE ANALYSIS
4 STRUCTURAL ANALYSIS AND BILL OF MATERIALS
The scope of the LCA analysis (LCA functional unit) is focused on structural floor solutions of a typical office floor plan with an architectural gird composed by applying λ = 1.5m base length (SCI 2008) and considers only the floor structure while disregarding vertical elements and bracing systems.The total area of the floor section is A × B = 14.0 × 36.0 m (Figure 4). The designated floor plan represents a typical office space with a floor width (span) of A = 14.0 m (Neufert 2012). In accordance with EN 1991-1-1 (CEN 2002) the load area is designated accordingly as Category B and based on this categorization the characteristic values for the imposed loads (uniformly distributed area loads) were defined. For the defined functional unit three structural flooring systems are provided so that all fulfil SLS (serviceability) and ULS (ultimate) limit states in accordance with EN 1990 (CEN 2005). The first solution (I) is contemporary RC flat slab with drop panels that accommodate the introduction of the reaction force into the two-way solid slab (Figure 4). In the second case (II) it is contemporary steel-concrete composite floor system achieved applying COFRAPLUS steel sheeting, secondary composite beams with welded headed studs and primary steel beams. The last solution (III) is the aforementioned de mountable steel-concrete composite floor system (Figure 1). The first two cases (I–II) represent contemporary structural solutions where the elements are in
The structural analysis was performed using commercially available design software asAutodesk Robot and ABC beam design software of ArcelorMittal. The assumed characteristic load values were defined in accordance with EN 1991-1-1 (Table 1). Table 1.
Floor area loads.
Load type
Load intensity [kN/m2 ]
Imposed load Partitions Superimposed dead load
3.0 1.2 1.0
4.1 RC flat slab with drop panels The RC flat slab (solution I in Figure 4) is designed adopting the slab thickness of h = 250mm and the concrete class of C30/37 regarding the deformation criteria. The reinforcement material was designated as B500B. At the column support points square drop panels with dimensions of A × B = 1500 × 1500mm are provided with the total drop cap thickness of h1 = 400mm. In this manner the requirement for additional punching reinforcement besides the flexural one is avoided where the maximum punching shear ratio was 0.9.
Figure 4. Floorplan grid (Left), and Structural floor solutions (Right).
5
The maximum flexural reinforcement ratio of ρl = 0.42% is recorded in the top zone above the columns (ρl, min = 0.151%) in the dominant bearing direction. The amount of consumed reinforcement material per cubic meter of concrete (93.6 kg/m3 ) corresponds to the contemporary reinforcement consumption regarding flat slab systems (95 − 135 kg/m3 ) having in mind the application of drop panels. The total amount of required concrete and reinforcement material is provided in the following (Table 2). Table 2.
Material consumption – RC flat slab.
Material
Amount [t]
Concrete C30/37 Reinforcement B500B
4.3 Demountable composite floor The demountable and reusable composite floor system (solution III in Figure 4) was designed applying the same design methodology of EN 19994-1-1 as in the case of the contemporary composite system with the difference that the shear connection in this case was achieved using the aforementioned demountable shear connector device (Figure 1). The modular slab elements were designed as simply supported and continuous one-way solid RC slabs with the span of L = 3.0m. The thickness of the slab modules is h = 140mm as well and the concrete material is designated as C30/37. The obtained reinforcement ratio is slightly higher than the minimal (ρl,min = 0.151%). The composite floor is assembled in unpropped condition by connecting the HEA 400 S235 hot-dip zinc galvanized hot-rolled steel section with modular slab elements using the demountable shear connection. The maximum flexural utilization of the composite cross section at the midspan is 0.88 while the first eigen-frequency is higher than 3Hz (3.13Hz). Due to the presence of the embedded rail channels the concrete slabs remain undamaged even if the ultimate capacity of the shear connector is reached what allows for the later reuse of the composite floor main elements. The only component that has to be reproduced in each life-cycle are the pretension bolt assemblies (HV) as after pretension they have to be discarded in accordance with EN 1090-2 (CEN 2008). The total amount of required concrete, reinforcement and structural steel material for this solution is provided in the following table (Table 4).
331.42 11.8
4.2 Contemporary composite floor The contemporary composite floor system (solution II in Figure 4) is designed as long-span secondary beam system. The composite slab with the thickness of h = 140mm is achieved using COFRA PLUS 60 profiled steel sheeting with the sheet thickness of t = 0.88mm. The composite floor system is achieved by connecting the composite slab with the IPE 400 S355 hot rolled steel section by utilizing the Nelson headed studs (d = 22 mm,hs = 125 mm) in each rib of profiled steel sheeting. In order to decrease the resource demand in the construction phase, the composite beam is unpropped during the construction. The maximum utilization ratio of the composite beam by bending at the midspan is 0.96 while the first eigen-frequency for the quasi-permanent SLS load combination is almost 3Hz (2.72Hz). In order to mitigate the excessive deflection of the composite beam, the steel section is precambered before construction (wc = 90 mm). The primary beam was designed as a steel element that supports the secondary composite beams. The same IPE 400 S355 section was sufficient to fulfill the limit states. Table 3. floor.
Table 4. Material consumption – Reusable steel-concrete composite floor. Material Concrete C20/25 Reinforcement B500B Structural steel S235 Shear connector (Reuse) Shear connector (No reuse)
Amount [t] 175.53 3.15 34.21 0.22 0.49
Material consumption – Composite
Material Concrete C20/25 Reinforcement B500B Cofrpa Plus 60 deck Structural steel S355
5 LIFE CYCLE ASSESSMENT Amount [t]
Life Cycle Assessment (LCA) is a methodology to quantify the environmental impacts of a product or process in a standardized manner over its lifespan. Standardized environmental data enables professionals to make informed decisions regarding the degree of sustainability of a product, but more importantly, enables comparison of alternative products or solutions regarding sustainability. The LCA demonstrates the carbon output of various products and can be a good way to weight how other sustainable alternatives fair. For this reason, this study employs LCA methodology, firstly to assess the performance regarding
129.45 1.82 5.04 17.09
The total amount of required concrete, reinforcement and structural steel material is provided in the Table 3. The structural steel consumption (33.9kg/m2 ) corresponds to the upper limit of expected amount (25 − 35kg/m2 ).
6
The goal of this LCA is to primarily assess the environmental performance of the demountable composite floor because the motivation behind the structural design was higher environmental performance through multiple reuse cycles (here, 3 LC). Reuse is considered as a circular strategy because it lowers raw material consumption, which translates into lower raw material extraction. Moreover, reuse directly diminishes Construction and Demolition Waste (CDW) and landfilling. Not only materials are saved, but also less land will be occupied for landfilling purposes. Therefore, through LCA methodology one can demonstrate that the demountable composite floor is truly sustainable. In other words, it could be sustainable only if its environmental performance over multiple life LCs is better than a single LC. Another way through which higher sustainable performance of the demountable composite floor (section 4.3) can be proved is through comparison with conventional designs that are prevalent in the construction industry (section 4.1 and 4.2). These conventional floor systems are traditionally designed for one LC despite that some of the materials locked in the modules have the potential to be further reused instead of being demolished and landfilled. Therefore, they make a good case for comparing business-as-usual versus novel demountable structural designs.
of the LCA: the use and maintenance (B 1–7) stage is excluded from the scope due to irrelevance to the LCA goal. In other words, it is considered that the buildings that host the three different scenarios have the same category of use in order for the floor systems to be subjected to the same load category throughout their lifespan. This consideration simplifies the LCA calculation and helps to focus on the goal of the study. Production phase which includes stages A1, A2 and A3 is mandatory to be declared and it is considered in accordance with EN 15804. The stage A4 is in the scope regarding the construction stage. However, the design team estimated that the installation of the three floor systems will require the same amount of energy and labor. Therefore, A5 stage is considered out of the scope since it will be canceled out in the emission comparisons in the end. Moreover, it is not considered that the next service life is using materials that are recycled from the first service life. Lastly, no material loss during the reuse of elements is anticipated. When it comes to the End-of-Life phase (EoL), C1 and C2 stages are in the scope in addition to the stage D. Recycling after the floor systems fulfilled their intended function is considered for all scenarios. Taking D stage into account renders a more realistic circular LCA. Often D stage values are negative because of the positive impact of recycling on the sustainable performance of products. Another assumption was made based on the limitation of the available LC stages in the ÖKOBAUDAT database. For some materials, such as reinforcement rebars and meshes, A4 stage impacts were not available, however for those available they were lower than C2. Hence, for that unavailable A4s, the C2 impacts were applied although this might lead to a slight overestimation of the results. Notwithstanding, this overestimation is reflected in every scenario. Therefore, it does not invalidate the intended comparison. This LCA study relies on the publicly available data in the German-based ÖKOBAUDAT platform (ÖKOBAUDAT 2022) since the three structural designs are using market-available materials. Although the materials used in the three solutions are slightly different, a hypothesis is made that the energy and labor used for the product LC are similar. At the end of the required service life of the building it was considered that the floor system is completely disassembled regardless of whether the whole building is deconstructed. Building deconstruction is beyond the scope of this comparative evaluation of sustainability.
5.2 The scope and system boundary of LCA
5.3 Building information modelling (BIM)-based LCA
The LCA system boundary follows the modular division of the life cycle as suggested in EN15804. Figure 5 presents all LC stages. The ones colored in green are considered in the scope of this study. Since the goal is to demonstrate how multiple reuse cycles affect and lower the negative environmental impacts, an important assumption is made to outline the scope
After the goals, scope, system boundary and impact categories are decided, there are several ways to perform the LCA analysis. This study used a Building Information Modelling (BIM)-based approach in order to semi-automatically calculate the impacts. BIM is an object-oriented methodology to model, manage or modify the products and processes
sustainability of the novel reusable steel-concrete composite floor system that is explained in the section 1. Secondly, to compare how does this new reusable product fair with respect to other conventional designs as if it is maintained well, it can live infinitely through multiple reuses, while in conventional designs, the building and the floor systems are demolished and landfilled (or recycled at best). Globally, LCA is defined and explained in detail in ISO 14040 series, whereas in Europe, it is addressed by EN 15804+A2 (CEN 2021) that provides the baseline for Life cycle assessment in construction industry. The EN 15804+A2 compliant datasets for evaluation of the environmental performance are used in this project. Based on these standards, LCA has four main stages: (1) goal and scope definition, (2) life cycle inventory analysis (LCI), (3) life cycle impact assessment (LCIA), and (4) interpretation of the results. That said, environmental LCA requires several choices and assumptions that dictate the outcomes. 5.1 Goals of LCA
7
information in the construction industry. Topographic, geometric and semantic information can be modelled in BIM. This includes dimensions, volumes, types of materials and their properties, sustainability-related information, etc. Autodesk Revit is the main BIM authoring tool that was used in this study. All floor systems were firstly modelled in Autodesk Inventor. They were imported using rfa format, which is the native family template for Autodesk Revit. After being loaded in the Revit family editor, they were sent to a rvt file, which is Revit’s working environment. A Dynamo script was developed to include all the LC impacts from an Excel file and communicate it with Revit through shared parameters. This is a necessary step as Revit has no direct link to Excel files while the Dynamo enhances Revit’s capabilities by facilitating interaction with Excel files. Excel files contained the LC impact categories and values that were imported from the ÖKOBAUDAT platform. After the LCA calculation was completed, the results were automatically written back to Excel file via Dynamo. The Excel file was linked with Microsoft Power BI for further visualizations of the results.
Equation 1 shows one LC impact (I) for a conventional floor system (FS) and for the reusable steelconcrete composite floor system assuming one LC (50 years). A1−A3 C1−C2 A4 D + IFS + IFS + IFS IFS = IFS
(1)
One could argue as well that the actual one LC impact of the demountable composite floor is when the cumulative impact is divided by the number of LCs (n). The presented solution is designed in a way that it can be used for 3 or more LCs. Therefore, we can argue that the LC impact should be divided by 3. However, in the worst-case scenario, it is assumed that the floor will be used only for one LC. This is an important assumption because for this very reason, the new design is structurally stronger and uses more materials to keep the structural performance for one or more than one LC. With simple maintenance the same floor could be reused over and over again. Based on these assumptions, the LC impacts of the three different floor solutions are calculated for one LC span of 50 years based on Equation 1 for all environmental impact categories based on the EN 15804 + A2 standard (Figure 6). n A4 A1−A3 C1−C2 n D IFS = IFS + IFS /n (2) + IFS + IFS
5.4 LCA calculation In the first step, one full LC is calculated for every scenario (Figure 6). Each scenario starts with resource extraction and production (A1–3) and transportation of the products to the site (A4). Since the maintenance is not in the scope, the next stages are C1 and C2, that are deconstruction (or demolition) and deconstruction transportation. Since recycling is an almost established practice in the industry, category D (i.e., recycling potential) is considered for the two conventional floor solutions. The same consideration is made for the reusable steel-concrete composite floor system.
Equation 2 demonstrates normalized LC impact of demountable and reusable floor system. After manufacturing, transportation and installation (A1, A2, A3, A4), the floor system is disassembled from the building, it will be transported to another building (A4), which will be disassembled and transported to a recycling center (C1, C2). Use and maintenance is not in the scope, hence not in the formula above. The floor system will be recycled after its last life cycle (D).
Figure 6. One life cycle impacts of the three floor solutions over 50 years based on Equation 1.
8
Figures 7 and 8 illustrate the comparison of cumulative impacts for two and three LCs of the three floor solutions. Despite what is shown in Figure 6, for comparison purposes in Figures 7 and 8 the analysis relies on the streamlined LCA approach (Crawford 2011) in order to narrowly focus on the category of Global Warming Potential (GWP) based on the Kg CO2 -equivalent. Figure 8 shows the comparison of three LCs for all the 3 scenarios based on the total GWP. Each bar in Figure 8 displays the normalized LC impact, meaning that the associated LC impacts are divided by the number of LCs.
pieces, they must be designed for long-term load bearing without losing structural robustness during several use cycles. There is a message hidden within the LCA results of this study. For the demountable floor system, if the reuse promise is not fulfilled and the component does not go to the second or third LC, then the impact of one LC could be higher. Thus, in the future, mechanisms should be devised to ensure the proper reuse of the DfD-based designs. Otherwise, the sustainable impacts of reusability and circularity will not be realised.
5.5 LCA results and interpretation Not only the GWP of the reusable floor is higher than other solutions for one LC (Figure 6), but all the other environmental impact indicators are slightly higher as well. Nevertheless, the situation changes when the floor systems are compared for two LCs, i.e., 100 years (Figure 7), although only reported for GWP. The RC floor slab with drop panels seems to have a considerably higher impact than the other alternatives. The cumulative GWP of the demountable floor system shows no significant increase from one LC, although being significantly lower than the two others. GWP-total of one LC for demountable floor solution is 50470 kg CO2 -eq, while it is 52149.24 kg CO2 -eq for two LCs. Attributing this low increase in emissions to reusing modules instead of reproducing and remanufacturing them from scratch, one can say that even in one reuse, (two LCs) the positive impacts of reusable and demountable modules are becoming apparent. The same conclusion can be drawn for the impacts of three LCs; the reuse of the demountable structural floor system yields considerably lower environmental impacts. However, to rely on the reusability of the structural
Figure 8. Comparisons of GWP impacts of the three solutions based on Equation 2 (normalized impact to one life cycle).
Figure 7. Comparison of climate change impact of the three floor system alternatives for three different lifecycle (LC) periods (one LC: 50 years, two LCs: 100 years, three LCs: 150 years.
9
6 CONCLUSIONS
(ECON4SD) research project. The LCA is performed by the sixth work package group of the ECON4SD project, which is funded by the European Regional Development Fund (2014-2020), with grant agreement No 2017-02-015-15.
Unsustainable material extraction and consumption have led to an increase in global carbon emissions and raw material shortages. Construction projects often have huge environmental footprints, which could be reduced if designed with a proper strategy for disassembly and reassembly, which implies further reuse of structural elements upon non-destructive disassembly procedures. To this aim, this paper, firstly, describes a novel reusable and demountable structural floor system. Secondly, it performs an LCA study for the demountable floor system and later compares that with two conventional floor solutions, i.e., Reinforced concrete and Contemporary steel-concrete composite floors. The goal of the LCA was to assess the sustainable feasibility of the proposed reusable system through a comparative study. The LCA results confirmed that the reuse of structural components bypasses extraction and production life cycle stages, which drastically lower the influential environmental impacts, i.e., global warming potential. The results demonstrated that one lifecycle of the demountable floor system emits slightly more carbon because more materials are employed in the design to ensure a robust structural performance for longer service life in comparison with conventional designs. However, when the demountable floor system is used for more than one lifecycle, it is considerably more sustainable with respect to the conventional designs.
REFERENCES Cen, (2002). EN 1991-1-1: Actions on structures - Part 1-1: General actions - Densities, self-weight, imposed loads for buildings. Cen, (2004). EN (1994-1-1): Design of composite steel and concrete structures - Part 1-1: General rules and rules for buildings. Cen, (2005). EN (1990): Basis of structural design. Cen, (2008). EN 1090-2: Execution of steel structures and aluminium structures - Part 2: Technical requirements for steel structures. Cen, (2021). BS EN 15804:2012+A2:2019. Crawford, Robert, (2011). Life Cycle Assessment in the Built Environment. 1. European commission, (2019). The European Green Deal. European commission, (2022). A new circular economy action plan for a cleaner and more competitive Europe. Eurostat, (2018). Generation of waste by waste category. [online]. (2018). Available from: https://ec.europa.eu/ eurostat/databrowser/explore/all/all_themes Fodor, Jovan, (2022.) Investigation in reusable composite flooring systems in steel and concrete based on composite behaviour by friction. Iea, (2019). World Energy Outlook (2019). Iso, (2022). ISO 20887:(2020). Neufert, Ernst, (2012). Bauentwurfslehre. Springer. Neufert, E. (2012) Bauentwurfslehre. Springer. Ökobaudat (2022). Availableat: https://www.oekobaudat.de/ en.html (Accessed: 6 August 2022). Sci, (2008). Steel buildings in Europe: Multi-storey steel buildings Part 2: Concept Design.
ACKNOWLEDGMENTS The demountable steel-concrete composite flooring system is developed by the second work package group of the Eco-Construction for Sustainable Development
10
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Probabilistic life cycle analysis as a sustainability-focused design tool for industrialized construction T. Hegarty & M. Lepech Stanford University, Stanford, CA, USA
ABSTRACT: The Architecture, Engineering, and Construction (AEC) industry is undergoing transformation in two key areas: (1) increased focus on performance-based sustainability metrics to quantify and reduce carbon emissions produced by the built environment and (2) integration of industrialized construction (IC) approaches such as process automation. This research is the first application of probabilistic life cycle analysis (LCA) as a sustainability-focused design tool for an early-stage IC concept case study. To produce probabilistic LCA results, a Pedigree Matrix was used to quantify uncertainty in LCA inputs to a Monte Carlo simulation (MCS). The SIPMath Excel® plug-in is used to perform nearly instantaneous, low-cost MCS automatically in Excel®. This is the first use of Stochastic Information Packets (SIPs) as a medium for capturing and using uncertainty information in the LCA of a building. This probabilistic LCA approach improves sustainability metrics in a highly uncertain design space, relative to the initial IC design concept. Research findings show that the probabilistic LCA model informed the design team’s decision making by providing rapid, targeted, high-level feedback on decisions such as material choice. Preliminary results indicate a 12% reduction in predicted mean lifecycle carbon dioxide emissions achieved through the IC approach relative to an equivalent conventional construction approach. A rapid, low-cost approach to probabilistic LCA is valuable to industry as a data-driven design tool that informs continuous improvements to design for IC companies. The integration of rigorous sustainability metrics to quantify and improve the sustainability value proposition of IC versus traditional construction methods has potential to attract and retain customers.
1 INTRODUCTION Worldwide, the built environment is a major contributor to carbon emissions. Building construction and operations alone are associated with 36% of global carbon emissions. (GlobalABC 2018) In order to meet global and national carbon emission goals, such as those laid out in the Paris Climate Agreement and by President Biden’s climate plan in the United States, the Architecture, Engineering, and Construction (AEC) industry will need to innovate more rapidly than in the past. (IPCC 2018) The industrialization of construction has been viewed as a broad shift from delivering buildings as unique, one-off projects, to considering them repeatable products. (Bertram et al. 2019) In other industries, “mass customization” is used to describe approaches whereby products remain customizable and unique to various degrees, while the manufacturing process is one of mass production. These ideas are being extended to AEC outcomes (i.e., buildings) which can remain customizable and unique to various degrees, while the construction process itself is more akin to mass production. Although performance-based sustainability-focused design methods and IC approaches are gaining traction
DOI 10.1201/9781003354222-2
in some countries and regions, current practice in much of the AEC industry does not quantitatively consider environmental sustainability. (Roberts et al. 2020) Barriers that have been identified in the literature to sustainability-focused design of buildings include perceived unreliability of LCA results due to uncertainties (Hoxha et al. 2017), a lack of quantitative targets for a “sustainable” design, and the high cost associated with obtaining meaningful metrics for comparison of designs, which can be particularly prohibitive when buildings are designed as singular projects. An additional barrier is the lack of a probabilistic design approach that manages high uncertainty in the early design phase when opportunities for improvements are greatest. (Basbagill et al. 2013). Quantitatively accounting for uncertainties in LCA provides a more transparent understanding of the projected range of environmental impacts, provides more credibility to comparative results, and makes the interpretation of results more reliable and meaningful. (Basbagill et al. 2013; Guo & Murphy 2012). When uncertainty is quantified in LCAs, Monte Carlo Simulation (MCS) is the most commonly applied approach. (Groen et al. 2014). However, only about 20% of recent LCAs published between 2014 and 2018 quantified
11
uncertainty through any approach. (Bamber et al. 2020). This study introduces a framework for the probabilistic design of sustainable buildings that is based on a practical, yet rigorous, probabilistic LCA methodology. In Section 2, this paper introduces the framework for probabilistic LCA as a design tool for buildings, in the context of design targets. Section 3 presents the practical process for performing probabilistic LCA to guide the design of IC buildings, highlighting the novel integration of designer surveys to incorporate uncertainty inputs into the LCA via the use of a Pedigree Matrix and SIPmath for MCS. Section 4 discusses a case study of an early-stage IC concept (CarbonCondo) compared to its conventional construction counterpart. This case study is the first application of probabilistic LCA as a design tool for an IC concept, as well as the first use of probabilistic LCA to compare an IC concept to its conventional construction counterpart. It explores the hypothesis that providing timely probabilistic LCA results to a conceptual design team can effectively reduce predicted carbon emissions for the IC design concept, relative to both the initial concept and a conventional construction equivalent. Section 5 provides conclusions, limitations, and future work.
to the time of functional obsolescence (tfo ). Also, as shown in Figure 1a, the cumulative impact differences over time between two projects or construction methods (i.e., IC versus conventional construction as shown in Figure 1b) can be clearly illustrated, and compared at any time, t. Depending on individual project goals, the cumulative impact can be expressed using midpoint environmental indicators such as global warming potential (kg CO2 -equivalents), acidification potential (H+molequivalents), solid waste (kg), primary energy (MJ), etc. Environmental impact metrics are derived from widely accepted environmental impact assessment midpoint indicator protocols (e.g., TRACI in the US, ReCiPe in Europe).
Figure 1b. Life cycle stages associated with IC buildings (top) compared to conventional construction (bottom), with impacts at time of construction, ic and impacts at time of functional obsolescence, ifo , marked.
2 FRAMEWORK FOR PROBABILISTIC LCA AS A BUILDING DESIGN TOOL 2.1 Visualizing probabilistic LCA results
2.2 Calculating probability of meeting sustainable design targets
Probabilistic design of more sustainable buildings requires quantitative measurement of the environmental, social, or economic impacts of the building over its full life cycle (from initial material extraction up to end of life), as shown in Figure 1 (modified from Lepech et al. 2014). In Figure 1a, the accumulation of environmental impact throughout the life of a building is shown as a “probabilistic impact envelope” that accrues out into the future as resources are consumed and maintenance or repair actions are performed, up
Setting sustainability-focused targets at the design phase, along with introducing LCA design tools, can enable improved building designs with significantly reduced environmental indicators. (RussellSmith et al. 2015) Such targets are drawn from policy goals or customer preferences. For example, design targets can be adopted from the Intergovernmental Panel on Climate Change’s (IPCC) proposed reductions in global carbon emissions. According to the IPCC, limiting warming to 1.5◦ C will require reaching net zero carbon emissions globally around 2050. (IPCC 2018) In light of this, some corporate and private clients have set sustainability commitments that manifest in quantitative design targets, such as net zero carbon. In this way, policy can drive more sustainable building design via code-enforced mandates and by influencing building owners to demand higher sustainability performance from the AEC industry. When a sustainability target is set, the probability of successfully reducing impacts to meet that design target is computed for any time of interest, t, by calculating the proportion of the MC trials where the target is met, as shown in Equation 1.
Figure 1a. Idealized probabilistic impact envelopes (dashed lines) for cumulative impact from construction to functional obsolescence for status quo building design in black and alternative building design (e.g., IC) in gray.
ps (t) =
12
nmax 1 II(in (t) ≤ itarget (t))
nmax n=1
(1)
which is typically visualized via a process flow diagram. This life cycle inventory is then aggregated into a set of life cycle impact indicators through life cycle impact assessment (LCIA), and impact indicators in line with the needs of the project are selected. Following methods used by Basbagill et al. (Basbagill et al. 2013) and Russell-Smith et al. (RussellSmith et al. 2015), this framework recommends the LCA study is managed in a Microsoft Excel® spreadsheet or other LCA-specific software system capturing material quantities, transport distances, impact factors, energy use data, component and building lifespans, and any location-specific factors for each life cycle stage of the building or component. The bulk of the effort associated with LCA of buildings involves gathering all the aforementioned inputs for each building approach under consideration. Many of these inputs come from construction documents and quantity estimates. Other information is taken from manufacturer’s information, industry standards, databases, or discussions with industry partners. Impact data is taken from LCA databases (e.g., EcoInvent), published Environmental Product Declarations (EPDs), or literature, as applicable. Operational phase energy use can be determined from energy models constructed in Building Energy Model (BEM) software (e.g., EnergyPlus), unless specific operational energy consumption data is available from energy consultants or past building performance data.
where ps (t) is the probability of meeting the target reduction in environmental midpoint indicator at time t, in (t) is the cumulative impact at time t of the newly designed building for trial n, itarget (t) is the targeted impact level at time t, and nmax is the number of trials simulated. Last, II is an indicator function that has a value of 1 when the statement enclosed in the parentheses is true. In addition to comparing to design targets, it is often desirable to be able to compare a new design to a baseline or status quo design, which requires comparing two probability density functions (PDFs). A visual comparison of two PDFs at a time of interest, t, is illustrated in Figure 1a. ?The probability of successfully reducing impacts relative to a status quo design at time t is computed using Equation 2. ps (t) =
nmax 1 II(in (t) ≤ in,old (t))
nmax n=1
(2)
which is much the same as Eq. (1), only here itarget (t) is replaced with in,old (t), which is the cumulative impact at time t of the status quo building design for trial n. Note that this calculation assumes the two output distributions are independent of each other (i.e. have unique seeds) and contain the same number of trials. Such calculations allow designers to probabilistically evaluate whether an alternative building design option reduces impacts compared to the status quo approach or a design target. Using this framework, architects, engineers, and designers can assess how to best meet reduction targets at the lowest cost and can explicitly consider tradeoffs between confidence levels and cost consequences.
Step 2: Input parameter distribution At the early stages of design, little information is known about design parameters. Thus, characterization of uncertainty associated with design parameters is challenging. Working with project design professionals, probabilistic distributions can be determined for input parameters using a Pedigree Matrix approach to transform qualitative uncertainty information into quantitative distribution parameters, as discussed later in Section 3.2. Where possible, input probability density functions that rely on industry product data will be identified, but practically speaking, primary and secondary uncertainty data is rarely published or available from industry partners. In cases where limited uncertainty data (i.e., beyond a single point estimator, such as mean) is available, Pomponi et al. have demonstrated that using a maximum and minimum value to create a uniform probability density function is adequate for probabilistic LCA modeling. (Pomponi et al. 2017) Based on the Central Limit Theorem, Pomponi et al. found that after 10,000 Monte Carlo simulations it is difficult to distinguish between analyses using uniform distributions or Gaussian distributions for underlying input parameters.
3 PROCESS FOR PROBABILISTIC LCA OF BUILDINGS 3.1 Probabilistic LCA in excel® using SIPmath Generating useful probabilistic LCA results that can inform design decisions involves a four-step procedure. The first step is the preliminary, deterministic LCA. The second is the determination of input parameter distributions. The third is the Monte Carlo Simulation, which outputs the distribution of potential impacts. The final step is the comparison of the results to the target or status quo design via calculation of the probability of meeting sustainable design targets and visualization of the results of the probabilistic LCA. Step 1: Deterministic LCA Each LCA begins with a standard deterministic LCA, governed by ISO 14040 and ISO 14044 series standards. (Technical Committee ISO 2006) Following the ISO standards, after determining the scope and boundaries of the LCA, a life cycle inventory (LCI) is constructed to quantify all the processes, materials, and flows that take place within the system boundaries,
Step 3: Monte Carlo simulation via SIPmath Once the probability density functions for the LCA inputs are defined, SIPmath in Microsoft Excel® can be used to model the distributions as Stochastic Information Packets (SIPs) in place of the point value inputs
13
quantitative probabilistic distribution point estimators for the LCA inputs. A Pedigree Matrix uses common language descriptions of data quality to quantitatively assess how closely the data in an LCA database matches the realworld situation being modeled. (Ciroth et al. 2016; Greenhouse Gas Protocol 2011) The Pedigree Matrix translates simple qualitative information regarding data applicability into quantitative uncertainty factors, xi , which are then transformed into parameters that define probabilistic distributions. More information on the specifics of the Pedigree Matrix used can be found in Ciroth et al. (Ciroth et al. 2016) The process of determining the parameters for a distribution, specifically a lognormal distribution, is explained in depth for illustration. Lognormal distributions always yield positive real values, leading to widespread use in LCA modeling applications. The Pedigree Matrix can also be used to generate parameters for additional distributions (e.g., normal, triangular, uniform). (Muller et al. 2016) SIPmath can model various distributions, should the modeler predict that a different type of distribution would more closely capture the actual distribution of the data. To generate a lognormal distribution, SIPmath requires the two input parameters from the distribution: the 50th percentile and any other percentile between the 60th and the 99th. The average is input as the 50th percentile figure, and then the “maximum value,” or 97.5th percentile, can be computed from the 95% interval geometric standard deviation as shown in Equations 3 and 4. Equation 3 is used to calculate the 95% interval geometric standard deviation of the distribution around the single data point. (Ciroth et al. 2016) ⎛ ⎞ 2 n 1 x i ⎠ σg−95 = exp ⎝ · (3) ln n i=1 x¯ g
to the deterministic LCA. Probabilistic LCAs have been conducted by Savage and Thibault (Savage & Thibault 2014) and Zirps et al. (Zirps et al. 2020) in Excel® based on a SIPmath MCS approach. SIPmath probabilistic modeling performs computations using Stochastic Information Packets, in which uncertainty is modeled as an array of possible outcomes. (Probability Management 2018; S. L. Savage & Thibault 2014) Within SIPmath, uncertainties are represented as thousands of possible outcomes within an array. Such preprocessing of uncertain outcomes enables rapid probabilistic analysis of many uncertain variables simultaneously in the native Microsoft Excel® environment. The SIPmath Excel® plug-in is used to perform nearly instantaneous MCS in Excel® . This results in a distribution of potential total life cycle impacts, in the form of an output SIP. When a comparative LCA between an IC approach and its conventional construction counterpart is being performed, output distributions of total life cycle impacts are produced for both the IC approach and the baseline traditional construction approach. Setting a unique Start Variable ID and the same number of MCS trials is important to allow comparison of the output SIPs. (Savage et al. 2016) Step 4: Comparison and visualization of results Once the MCS is completed, it is possible to calculate the probability of successfully meeting the design targets, using the equations described in Section 2.2. If the design target has not yet been achieved, it is important to be able to identify opportunities for environmental impact reductions. Therefore, it may be desirable to break down the probabilistic results into more specific life cycle stages, as shown in Figure 1b, or to map the probabilistic envelope of potential impacts over the lifecycle, as shown in Figure 1a. Such visualizations provide insight into where the greatest impacts accrue, allowing for targeted design interventions to maximize sustainability improvements with the least effort or cost. Within a specific life cycle stage, impacts can also be shown for individual materials or processes, to identify specific design changes that might reduce impacts. Then the model can be updated with new inputs to test how changes alter the overall impact of the design.
with x¯ g = n ni=1 xi , the geometric mean of x, and σg−95 is the 95% interval geometric standard deviation of the distribution. Then the “maximum value,” or 97.5th percentile, can be computed from σg−95 , as shown in Equation 4. maxValue = σg2 × µg = σg−95 × µg
(4)
where µg is the original deterministic average input, and σg is the geometric standard deviation. The Pedigree Matrix-based survey enables ongoing collection of qualitative uncertainty inputs from designers, based on their answers to multiple-choice questions for each of the six indicator categories (reliability, completeness, temporal correlation, geographical correlation, further technological correlation, and sample size) for a given LCA input of interest. The answers to these multiple-choice questions are then transferred into an Excel® spreadsheet, which has been set up to automatically calculate the 97.5% figure for the lognormal distribution around that average input.
3.2 Integrating qualitative uncertainty via pedigree matrix-based surveys Elicitation of expert knowledge is frequently used in fields that grapple with high levels of uncertainty and little data, such as risk management. Especially during the preliminary building design phase, there is high uncertainty. Therefore, it is important to determine and integrate uncertainty by eliciting structured feedback from project experts. To systematically account for uncertainty in the LCA, a survey based on the Pedigree Matrix approach has been developed for project teams that converts qualitative, multiple-choice answers into
14
must achieve lower lifecycle carbon emissions than those associated with the US average condo. Wood framing was used in 91% of all multifamily buildings completed in 2020, (US Census Bureau 2020) and wooden condominium buildings are typically up to three stories high.
Given the number of inputs to an LCA, it is recommended that this approach target inputs that have the greatest bearing on the outcome of the LCA, such as the main construction materials. Where expert elicitation through surveys of the building designers is not practical due to project constraints, the probabilistic LCA practitioner can apply the Pedigree Matrix approach based on their own knowledge of the inputs, either to individual inputs or batches of similar inputs based on their judgement, with the former preferred due to greater accuracy. This survey-based approach has been piloted with the CarbonCondo case study, as described in Section 4.
Table 1. Inventory of conventional condo materials comprising more than 5% of the total building mass. Conventional Condo
4 INDUSTRIALIZED CONSTRUCTION CASE STUDY
Material
Quantity kg
concrete wood framing
362,294 75,270
stucco drywall
To demonstrate how the novel probabilistic LCA framework described in Sections 1 and 2 can serve as a valuable design tool for practitioners, and to compare IC approaches to their conventional construction counterparts, a case study is presented.
62,714 36,554
SimaPro Analogue Concrete, normal {GLO} Sawnwood, softwood, kiln dried, planed {RoW} Stucco {GLO} Gypsum plasterboard {GLO}
Meanwhile, the structural properties of the CNT composite structural system used in the CarbonCondo project enabled building to five stories, thus allowing a smaller physical footprint. Building square footage and unit number were standardized across both LCAs; the functional unit used for comparison in this study was a 20-unit condo building of 1,962 m2 located in Seattle. Table 2 provides an inventory the main materials involved in the construction of the CarbonCondo, showing materials comprising at least 5% of the total building mass. Information on proposed carbon-based material formats (including CNT and cFoam) was provided by industry collaborators at NanoComp and CFoam, enabling LCAs of these highly specialized materials that lack LCI data in existing literature or databases.
4.1 CarbonCondo versus conventional condo CarbonCondo is an Advanced Research Projects Agency–Energy (ARPA-E) funded collaboration across academic institutions including MIT and Stanford, as well as industry partners. This work compares a newly conceived carbon-based unitary material logic building approach (CarbonCondo) versus average construction in the United States. CarbonCondo looks to demonstrate a streamlined unitary-material logic to create high-quality buildings that use polymeric composites in place of traditional materials. Specifically, the project leverages natural gas pyrolysis carbon nanotube (CNT) sheets (Huntsman 2019) produced by project collaborator NanoComp and carbon foam insulation produced by project collaborator CFoam, a division ofTouchstone Research Laboratory. This type of construction is truly novel within the construction industry, demonstrating a high-performance structural building envelope, as seen in Figure 2. (Goulthorpe 2020)
Table 2. Inventory of CarbonCondo materials comprising more than 5% of the total building mass. CarbonCondo
Material cFoam CNT PET
Quantity kg 115,804 65,617 53,570
epoxy
40,886
SimaPro Analogue Not available* Not available* Polyethylene, low density, granulate {GLO} Epoxy resin, liquid {GLO}
*No material analogues available so individual LCAs were performed based on data provided by industry collaborators.
4.2 CarbonCondo: Probabilistic environmental impact model
2
Figure 2. Conceptual design of a 20-unit, 1,962 m CarbonCondo. (Goulthorpe 2020).
Table 3 provides a summary of the data sources for the probabilistic LCA for CarbonCondo. Composite Design Studio (CDS), a structural engineering and design firm, provided the structural design of
A critical minimum sustainable design target outlined by the project was that the CarbonCondo concept
15
Table 3. Summary of the data sources for the deterministic and probabilistic LCA inputs for CarbonCondo. CarbonCondo Embodied carbon
Operational carbon
Deterministic Material quantities, wastage (CDS, MIT) Transport distances (Seattle specific) Prefabrication energy use (MouldCAM, MIT) On-site assembly equipment use (MIT) Impact factors (SimaPro) Probabilistic Pedigree Matrixes integrating survey results Deterministic Energy use (Transsolar, Seattle specific) Energy impact factors (EPA, Seattle specific) Probabilistic Pedigree Matrixes integrating survey results Figure 3. Part of the “CarbonCondo LCA Uncertainty Response Form” used to gather input from project partners on their qualitative understanding of key LCA inputs. Survey ultiple choice questions were derived from the Pedigree Matrix.
the CarbonCondo and associated construction quantity takeoffs that were needed for the LCA, as well as predicted wastage factors. Project collaborators at MIT provided material quantity data for additional non-structural materials. Collaborators at MouldCAM and MIT contributed data on predicted pre-fabrication energy use as well as on the use of equipment in on-site assembly. Impact factors came both from SimaPro and from individual life cycle inventories based on consultations with manufacturers. Transsolar, a building energy consulting firm, created a detailed operational energy consumption model which provided energy use results for the CarbonCondo. Energy impact factors for Seattle were provided by the US Environmental Protection Agency (EPA). The Pedigree Matrix approach was used to individually convert these deterministic inputs to probabilistic inputs. Specifically, an uncertainty survey was developed to inform probabilistic LCA inputs, including how closely the current CNT material format (using methane as a precursor) may match the actual material format used in the CarbonCondo. Part of this survey is shown in Figure 3, where the multiple-choice selections are derived from the Pedigree Matrix, and there are five categories for assessing uncertainties, namely data reliability, data completeness, temporal correlation, geographic correlation, and technology readiness level correlation. Project partners mentioned above filled out this survey, which collected qualitative uncertainty information that was then integrated in an Excel® -based probabilistic LCA model for the CarbonCondo, using a SIPmath MCS with 10,000 trials.
(PNNL) energy models, and material wastage factors were from expert estimates in Estimating in Building Construction. Impact factors to consider the impacts of these materials from material extraction through on-site construction were taken from SimaPro.
Table 4. Summary of the data sources for the deterministic and probabilistic LCA inputs for the conventional condo. Conventional Condo Embodied carbon
Operational carbon
Deterministic Material quantities (RSMeans, PNNL) Material wastage factors (Estimating in Building Construction) Transport distances Impact factors (SimaPro) Probabilistic Pedigree Matrixes integrating survey results Deterministic Energy use (BPD, Seattle specific) Energy impact factors (EPA, Seattle specific) Probabilistic Energy use (BPD, Seattle specific)
The Pedigree Matrix approach was applied to inputs to move from deterministic to probabilistic results. For operational impacts, the Building Performance Database (BPD) was used to obtain real world data on existing buildings, which provided primary data on the full probabilistic distribution of actual energy use intensity for condo buildings in Seattle Figure 4 shows the distribution of annual site energy use of apartment units in the Seattle climate zone, compared with the predicted CarbonCondo energy use results. All this
4.3 Conventional condo: Probabilistic environmental impact model Table 4 provides a summary of the data sources for the probabilistic LCA for the US average conventional condo. The average building material quantities were derived from the RSMeans building cost estimator tool and the Pacific Northwest National Laboratory
16
5 CONCLUSIONS The AEC industry is being challenged to innovate rapidly to meet global demand for housing while reducing carbon emissions in line with targets such as those set by the IPCC. This paper presented a novel probabilistic LCA framework for the design of buildings that enables designers to balance sustainability indicators such as cost and carbon emissions. Specifically, this paper explains the first application of SIPmath to simplify probabilistic LCA for sustainability-focused building design, which could help accelerate its adoption by AEC professionals, as well as the first known use of Pedigree Matrix-based surveys for elicitation of uncertainty information from multiple project stakeholders. The CarbonCondo case study was presented to demonstrate this framework and process. The probabilistic LCA framework presented enables transparent LCA results that can provide earlystage feedback to designers to meet sustainability targets, by supporting informed decisions on material choices, material quantities, transportation options, construction methods, operational performance, maintenance and repair decisions, and end-of-life. The probabilistic LCA model, from the case study presented in Section 4, informed the CarbonCondo design team’s decision making by providing rapid, targeted feedback on the life cycle impacts of decisions such as choice of matrix material. Generating results that explicitly integrate uncertainty helps decision makers interpret the confidence level associated with LCA results, allowing informed tradeoffs between additional building design goals (e.g., costs, aesthetics). The proposed rapid, low-cost approach to probabilistic LCA is valuable to IC companies as a data-driven design tool that informs continuous improvements to design. The integration of rigorous sustainability metrics to quantify and improve the sustainability value proposition of IC versus traditional construction methods has potential to attract and retain customers, in light of the growing prevalence of personal and corporate sustainability commitments. Limitations of this work include the lack of a fully equivalent site-built alternative for the CarbonCondo, thus necessitating comparison to the US average condo, which is difficult to definitively define, and limitations in the scale of survey feedback obtained. Additionally, the case study focused on a single location and just one of many possible approaches to industrialized construction. Future work will expand the use of SIPmath and Pedigree Matrix-based surveys to compute the probabilities of successfully meeting environmental targets for additional IC approaches and constructed facilities. Applying this approach to additional IC case studies in additional contexts is expected to improve the reliability of the sustainability value proposition of IC. Additionally, future integration of the sustainable target value (STV) framework (Russell-Smith et al. 2015) can provide a more rigorous goal for reduction targets.
Figure 4. Comparison of conventional condo and predicted CarbonCondo annual site energy use for a single apartment unit. Data for real world apartment units in Seattle were obtained from the BPD to form the confidence intervals.
information was integrated in an Excel® -based probabilistic LCA model for the US average condo, using a SIPmath MCS with 10,000 trials. 4.4 Probabilistic impact assessment results A preliminary probabilistic LCA model was constructed for the CarbonCondo project and the US average condo. This probabilistic LCA approach improved sustainability metrics in a highly uncertain design space, relative to the initial IC design concept. The probabilistic LCA model informed the CarbonCondo design team’s decision making by providing rapid, targeted, high-level feedback on decisions such as matrix material choice. As shown in Figure 5, choosing PET over epoxy for the matrix material reduced the average embodied carbon impacts at the building scale. Probabilistic LCA also enables quantification of the probability that a given CarbonCondo concept successfully reduces the total lifecycle carbon emissions of the CarbonCondo relative to the US average condo. Preliminary results indicate a 12% reduction in predicted mean lifecycle carbon dioxide emissions achieved for the CarbonCondo relative to the US average condo.
Figure 5. Probability density functions of embodied carbon impacts for conventional condo compared to embodied carbon impacts for two CarbonCondo matrix material alternatives: PET and epoxy. PET shows the lowest average impacts.
17
ACKNOWLEDGMENTS
Hoxha, E., Habert, G., Lasvaux, S., Chevalier, J., & Le Roy, R. (2017). Influence of construction material uncertainties on residential building LCA reliability. Journal of Cleaner Production, 144, 33–47. Huntsman. (2019). Miralon Sheets/Tape. https://www. miralon.com/sheet/tape IPCC. (2018). Summary for Policymakers. In Press, 1(3), 374–381. Lepech, M. D., Geiker, M., & Stang, H. (2014). Probabilistic design and management of environmentally sustainable repair and rehabilitation of reinforced concrete structures. Cement and Concrete Composites, 47, 19–31. Muller, S., Lesage, P., Ciroth, A., Mutel, C., Weidema, B. P., & Samson, R. (2016). The application of the pedigree approach to the distributions foreseen in ecoinvent v3. International Journal of Life Cycle Assessment, 21(9), 1327–1337. Pomponi, F., D’Amico, B., & Moncaster, A. (2017). A Method to Facilitate Uncertainty Analysis in LCAs of Buildings. Energies, 10(4), 524. ProbabilityManagement. (2018). The Open SIPmathTM 2.0 Standard. https://www.probabilitymanagement.org/20standard Roberts, M., Allen, S., & Coley, D. (2020). Life cycle assessment in the building design process – A systematic literature review. Building and Environment, 185(August), 107274. Russell-Smith, S. V., Lepech, M. D., Fruchter, R., & Meyer, Y. B. (2015). Sustainable target value design: Integrating life cycle assessment and target value design to improve building energy and environmental performance. Journal of Cleaner Production, 88, 43–51. Savage, S. L., & Thibault, J. M. (2014). Towards a Simulation Network or the Medium Is the Monte Carlo. 4126–4133. Savage, S., Thibault, M., & Empey, D. (2016). SIPmathTM Modeler Tools for Excel. August, 1–57. Technical Committee ISO. (2006). Life Cycle Assessment — Principles and Framework. Iso 14040, 3, 28. US Census Bureau. (2020). Annual 2020 Characteristics of New Housing. Zirps, M., Lepech, M. D., Savage, S., Michel, A., Stang, H., & Geiker, M. (2020). Probabilistic Design of Sustainable Reinforced Concrete Infrastructure Repairs Using SIPmath. Frontiers in Built Environment, 6(May).
The authors thank Professor Mark Goulthorpe, and all the project partners on the ARPA-E grant for their contributions. This work was supported by Stanford’s Precourt Institute Seed Funding, the NSF GRFP, and the Stanford Graduate Fellowship (SGF).This research is partly funded by the US NSF. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF. REFERENCES Bamber, N., Turner, I., Arulnathan, V., Li, Y., Zargar Ershadi, S., Smart, A., & Pelletier, N. (2020). Comparing sources and analysis of uncertainty in consequential and attributional life cycle assessment: review of current practice and recommendations. International Journal of Life Cycle Assessment, 25(1), 168–180. Basbagill, J., Flager, F., Lepech, M. D., & Fischer, M. (2013). Application of life-cycle assessment to early stage building design for reduced embodied environmental impacts. Building and Environment, 60, 81–92. Bertram, N., Fuchs, S., Mischke, J., Palter, R., Strube, G., & Woetzel, J. (2019). Modular construction: From projects to products. In Capital Projects & Infrastructure (Issue June). Ciroth, A., Muller, S., Weidema, B., & Lesage, P. (2016). Empirically based uncertainty factors for the pedigree matrix in ecoinvent. International Journal of Life Cycle Assessment, 21(9), 1338–1348. GlobalABC. (2018). Global Alliance for Buildings and Construction, 2018 Global Status Report. 325. Goulthorpe, M. (2020). Conceptual Seattle CarbonCondo Design. ARPA-E. Greenhouse Gas Protocol. (2011). Quantitative Inventory Uncertainty. https://ghgprotocol.org/product-standard Groen, E. A., Heijungs, R., Bokkers, E. A. M., & de Boer, I. J. M. (2014). Methods for uncertainty propagation in life cycle assessment. Environmental Modelling and Software, 62, 316–325. Guo, M., & Murphy, R. J. (2012). LCA data quality: Sensitivity and uncertainty analysis. Science of the Total Environment, 435–436, 230–243.
18
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Integrating Level(s) LCA in BIM: A tool for estimating LCA and LCC impacts in a case study M.T.H.A. Ferreira & J.D. Silvestre CERIS, Instituto Superior Técnico, Universidade de Lisboa, Lisbon, Portugal
A.A. Costa BUILT Colab, Collaborative Laboratory for the Digital Built Environment, Oporto, Portugal
H.B. & R.A. Bohne Department of Civil and Environmental Engineering, Norwegian University of Science and Technology, Trondheim, Norway
ABSTRACT: This research aims to tackle green inefficiencies using Building Information Modeling (BIM), particularly by developing an innovative Level(s) LCA plugin to estimate the building’s environmental and cost impacts and thereby help decision-makers to account for the full Life Cycle (LC) of the building in the design phase. The plugin provides a visual interface that shows both quantitative and graphical results of the impacts. The plugin was used in a case study of new construction dwelling to estimate the LC impacts for walls. The main results show that the Level(s) LCA plugin is suitable to perform both environmental and economic analysis and can now be used in other design projects to anticipate and mitigate the impacts of the construction sector. The case study can be seen as a proof-of-concept that such an integration in BIM offers results of high relevance when in the search for ways to optimize the LC impact.
1 INTRODUCTION 1.1 European framework In line with the most ambitious targets established in the Paris Agreement in 2015, the EU Green Deal aims to reduce carbon emissions by 55% or more until 2030 and for Europe to be a climate-neutral continent by 2050 (EC 2020, 2021). Construction and retrofit of buildings cause substantial environmental impacts (EC 2019) due to their significant consumption of energy (40%) and materials and energy-related greenhouse gas emissions (36%) (EC 2010). The construction industry plays a major role in the decarbonization process (EC 2019). However, the insufficient connection between EU targets and available design tools for sustainability assessment is a major issue to achieve such important goals on climate issues. Our research aims to resolve this by aligning European initiatives with the integration of a new Life Cycle Assessment (LCA) plugin in Building Information Modeling (BIM). The environmental LCA and the economic Life Cycle Cost (LCC) are methodologies that are recognized and standardized at the European level (CEN 2011, 2012), which are being increasingly used in the construction sector by experts and researchers. DOI 10.1201/9781003354222-3
Gradually, the Life Cycle (LC) paradigm is penetrating the construction sector, and building designs are now commonly assessed by the sum of all impacts and costs during their LC. However, not often as an integrated part of BIM, where important design choices are often made. Level(s) is a common framework for the assessment of buildings’ sustainability across Europe, which has an LC approach. It supports measurement and improvement from design to end-of-life, covering both renovation and new construction (EC 2021). The Level(s) framework provides a common methodology for assessing the sustainability of buildings based on six macro-objectives. It contributes, therefore, to achieve EU and Member States policy goals in energy, material and water use, waste production, and indoor air quality, in an LC perspective (EC 2021). Intending to bring buildings into the Circular Economy, Level(s) comprises a set of indicators, scenario tools, a data collection tool, checklists, and rating systems that allow professionals and project actors to measure buildings’ performance (EC 2021). The research presented here aims to use sustainability indicators established in the Level(s) framework in a new plugin. From the Level(s) six macro-objectives, the ones necessary for an LCA analysis were selected:
19
2019), the following indicators were also included for I-Level(s) LCA: Abiotic Depletion Potential for fossil fuels (ADPf), Abiotic Depletion Potential for non-fossil resources, minerals and metals (ADPm), Acidification Potential (AP) accumulated exceedance, Eutrophication Potential (EP), a fraction of nutrients reaching freshwater end compartment (EPf), EP fraction of nutrients reaching marine end compartment (EPm), EP accumulated exceedance EP terrestrial (EPt), Depletion potential of the stratospheric ozone layer, Ozone Depletion Potential (ODP), Photochemical Ozone Creation Potential (POCP), and Water (user) deprivation potential, deprivation-weighted water consumption (WDP). To reach macro-objective 6 Optimized Life Cycle Cost and Value indicator, the initial costs indicator of a material or element was included for III- Level(s) Cost.
objective 1 Greenhouse gas emissions along building life cycle, covering LC GWP; and objective 6 Optimized LCC and value. For this reason, from an LCA perspective the plugin covers the part of Level(s) framework related to LCA and it is named “Level(s) LCA Plugin”. BIM methodology is largely used in design offices and practices, nevertheless the authors found that there is a research gap in the integration of environmental and economic assessments with BIM, having a lack of tools for BIM to perform and automatize those assessments aligned with EU directives (EC 2021). The aim of this research is to associate green ambitions with technology by using BIM software to quantify building environmental impacts and help supply chains to make their decisions through the full LC of the building, from early design stages to the end of life. This was done by developing a Level(s) LCA plugin software to perform Level(s) LCA and LCC calculations.
Table 1.
2 MATERIAL AND METHOD 2.1 Plugin framework and indicators (environmental and cost) The purpose of an LCA plugin is to give the user a set of options to perform environmental and economic calculations depending on the stage of the project where they are, and the level of information needed. The aim here is to give the user different functionalities that correspond to different ways to estimate the environmental and/or economic impacts given different levels of information and different environmental indicators, so that the user can choose the one that fits better the project objectives. Within the scope of LCA studies, it was necessary to develop an adequate Product Data Template (PDT) that defines and uniformizes the Level(s) framework parameters and put them together for further integration in BIM. This PDT includes the necessary parameters to enable designers to do holistic and dynamic Level(s) assessments. The definition of an adequate PDT (Martina, F. 2018), with structure for BIM objects, was necessary to correctly define the environmental and economic indicators that will be read as shared parameters in a BIM environment. Table 1 shows the structure of the PDT per functionality. The proposed Level(s) LCA plugin for BIM integrates different types of sustainability analysis in its functionalities, which are shown in Table 1, sorted for each of I- Level(s) LCA, II- Level(s) GWP, and IIILevel(s) LCC, respectively. For the II- Level(s) GWP the following indicators were considered to perform the analysis: GWP total (GWPt), GWP Fossil (GWPf), GWP Biogenic (GWPb), and GWP Land use and Land use change (GWPl). To go further on the cradle to grave LCA analysis as recommended by EU (Dott, N. 2020) on the standard EN15804:2012+A2:2019 (EN
Product data template for the Level(s) LCA plugin.
Indicator/ Parameter Name
ILevel(s) LCA
IILevel(s) GWP
IIILevel(s) LCC
ADPf ADPm AP EPf EPm EPt ODP POCP WDP
x x x x x x x x x
-
-
GWPt GWPf GWPb GWPl Initial Costs
x x x x -
x x x x -
x
Units MJ kg Sb eq mol H+ eq. kg PO4 eq. kg N eq. mol N eq. kg CFC-11 eq. kg C2 H4 m3 world eq. deprived kg CO2 eq. kg CO2 eq. kg CO2 eq. kg CO2 eq. €/m2 /yr
2.2 Plugin development Following the structure established in the PDT (Table 1), all the environmental and economic indicators were inputted into the plugin to be read as shared parameters and added to BIM Elements/ objects or materials. The plugin was programmed with the analytical models for calculation of the different indicators and functionalities (chapter 2.3). To perform the analysis, it was also necessary to collect environmental and economic data and information for each product to be analyzed. The data was collected from the environmental product declarations (EPDs) for all types of walls and materials contained in the model of the case study. After that, the information was updated to the BIM Model through the plugin and the shared parameters of BIM Objects or materials were filled (Figure 1). Figure 1 represents all the elements and materials of the model loaded through the plugin and the data from the database.
20
Figure 1.
Caption of the list of the BIM elements and materials from the model.
EIaM environmental impact (of category x) of material a.
2.3 Plugin functionalities and analytic models Mentioned indicators are read in the construction BIM elements/ objects, families, materials and can be analyzed directly within a BIM software. The core indicators were divided per category of analysis (Table 1). Each of the environmental and cost indicators will be converted into parameters to be read in BIM.
2.3.2 Level(s) GWP Level(s) GWP analysis takes into consideration four main environmental indicators: GWP total, GWPf, GWPb, and GWPl. The functional unit of measure is kg CO2 multiplied by the area of wall material applied following equation 1.
2.3.1 Level(s) LCA To perform a Level(s) LCA analysis in line with EU standards, specifically EN 15804+A2 (EN15804:2012+A2:2019), the parameters addressed in the plugin to this specific functionality were ADPf, ADPm, AP, EPf, EPm, EPt, ODP, POCP, WDP, GWP, GWPf, GWPb, and GWPl. Each one is expressed per functional unit or declared unit and then automatically multiplied per the quantity on the construction element by the plugin following equation (1) (Santos, R. 2019): EIxMC
=
i
(QaM xEIaM )
2.3.3 Level(s) cost For the Level(s) Cost, initial costs were taken into consideration in Euros multiplied either per m2 , m3 , or unit corresponding to the information contained in the EPD following equation 2 (Santos, R. 2019): C MC =
i
(QaM xACaM )
(2)
a=1
C MC costs resulting from the manufacturing and construction phase (covering A1–A5 modules); i, number of existing materials. QaM quantity of materials used in the construction. ACaM the acquisition cost of material a.
(1)
a=1
EIxMC environmental impact of category x resulting from the manufacturing and construction phase (covering the A1–A5 life cycle modules); i, j, k number of existing materials i, transportation j, and construction utilities k; QaM quantity of material a;
2.4 Building life cycle (LC) stages A building’s LC stages are divided into four modulesA, B, C and D (EN15804:2012+A2:2019). Table 2 shows
21
in BIM families and materials, and the information regarding the impacts will remain associated to each element after the analysis is completed. Figure 3 shows the plugin loading the environmental indicators with their functional unit of measurement for the Level(s) LCA analysis. Figure 4 shows the plugin loading the environmental indicators for Level(s) GWP analysis and, finally, Figure 5 shows the economic indicators for Level(s) cost analysis. After creating a set of parameters correspondent to the analysis chosen automatically through the plugin the next step is to load the values of each category of impact with the information collected from the Database of the Oekobaudat (oekobaudat.de) with the values from the EPD of each product/ wall material existing in the case study (Figure 1). For the Level(s) Cost analysis per unit for initial costs the data was collected per product from the database of the website Kuantokusta (kuantokusta.pt) which contains the cost of products in the Portuguese market. Then the plugin does the calculation using the information and quantities obtained from the BIM model and linking them with the information from the database. The results were organized per plugin functionality as following subchapters.
module A1 to A5 for the production and construction process stages, and Table 3 shows modules B1 to B7 that correspond to the use stage, with B1 to B5 related to building factors and B6 and B7 related to the operation energy and water consumptions of the building. Table 4 shows modules C1 to C4 that refer to the end-of-life stage of the building, and a final module D corresponds to the benefits and loads beyond the system boundary, potential for reuse, recycling, and energy recovery. Table 2.
LC stages, module A.
Raw Material Supply Transport Manufacturing Construction installation
Table 3.
Production
Construction
A1 A2 A3 –
– A4 – A5
LC stages B Modules/ use stage. Use stage
Use
B1
Maintenance Repair Replacement Refurbishment Operational Energy Use Operational Water Use
B2 B3 B4 B5 B6 B7
Table 4.
LC stages C/D modules/ end of life. End of Life stage
Deconstruction, demolition Transport Waste Processing Disposal Reuse, Recycling, or energy recovery potentials
C1 C2 C3 C4 D
Figure 2.
Caption of plugin functionalities
3 RESULTS AND DISCUSSION 3.1 Introduction of the case study
In this study we have so far only incorporated the product and construction stages in the plugin (Table 2). Stage A is also applied to the case study and the data collected from the EPDs to perform the impacts.
The case study used to test and validate the use of the plugin and its three main functionalities is an isolated house dwelling in Montemor, Alentejo, Portugal, designed by Atelier dos Remédios. This is a threebedroom house with three toilets, a kitchen and an annex for events. It is a housing scheme connected with nature in the green landscape of Alentejo. The plugin is used to explore wall type solutions for both interior and exterior walls (facades) considering traditional Portuguese construction materials having
2.5 Level(s) LCA plugin The first step is to choose the type of analysis that the user would like to perform (Figure 2); secondly the indicators of the corresponding analysis are loaded by the plugin to BIM shared parameters and visible
22
brick as the main material. The design option to be studied with the plugin is the environmental and economic impacts integrating the parameters mentioned in Table 1 for each functionality performed. The objective is to execute the three functionalities of the plugin (Table 1) for a full analysis of all the walls designed in the house and verify environmental and economic impacts. In Figure 6 it is possible to see the first window of the plugin wherein the first step of the three analyses one should be chosen. The report will give designers help in understanding the impact of their choices. The main result of this study is the validation of the plugin through the case study detailed in the following subchapters, including completing the analysis and extrapolating their results. 3.2 Level(s) LCA analysis
Figure 3. Caption of the list of environmental indicators for Level(s) LCA analysis.
First analysis performed with the plugin was I-Level(s) LCA analysis. After the analysis is selected the environmental indicators from PDT explicit in Table 1 are loaded through the plugin to the model and read in their BIM families and materials. Figures 7 and 8 show a schedule with the total results for Level(s) LCA analysis of the brick walls. The results were exported to both MS-Excel tables and schedules in Revit. The results can be analyzed externally and in the schedules inside Revit where the information is kept in the BIM model. Also, 3D visualizations were created for each environmental indicator using a graphical scale of colors where color red represents wall elements with the highest impact in the model, color blue medium impact, and color green the lowest impact, so the user can graphically understand the impacts of all choices made. In Figure 9, it is possible to see the 3D visualization for the impact category POCP (photochemical ozone creation potential), and clearly visualize that the exterior walls of the house envelope have the highest impact compared to interior walls. The same 3D visualization was generated for the remaining impact categories and different results were archived per indicator reflecting the quantitative results from Figures 7 and 8.
Figure 4. Caption of the list of environmental indicators for Level(s) GWP analysis.
3.3 Level(s) GWP analysis Regarding the second functionality, Level(s) GWP, Figure 10 shows the total analysis results for Level(s) GWP analysis of the case study. It is possible to check the impact of the walls and each material contained. The results can be exported to an MS-Excel file and to a schedule inside the BIM model. The case study is being used to test the walls’ impact, but it is likewise also possible to isolate any kind of elements such as floors, windows, doors, roof, or have the results for all the elements aggregated together. Beyond the analytical calculations automatically done by the plugin with the equations presented in chapter 2.3, a 3D scheme for each GWP indicator of analysis was automatically generated, including all the parameters necessary to make each analysis: GWP
Figure 5. Caption of the list of economic indicators for Level(s) Cost analysis.
23
Figure 6.
Caption of BIM environment with the plugin in use in the case study.
Figure 9. Caption of the 3D visualization of the total results of Level(s) LCA analysis of walls for POPC environmental indicator.
Figure 7. Caption of the list of the total results of Level(s) LCA analysis of Walls 1/2.
Figure 10. Caption of the list of the total results of Level(s) GWP analysis of Walls.
Figure 8. Caption of the list of the total results of Level(s) LCA analysis of Walls 2/2.
with higher impact than the interiors walls. For Level(s) GWP, the case study’s traditional construction brick walls have a GWP total of 1.95E5 kg CO2 , for a total area of walls of 422m2 , according to the data compiled
total, GWPf, GWPl, and GWPb, see Fig 11. With the graphical color scheme, we can easily conclude that the materials placed on the exterior walls are the ones
24
4 CONCLUSIONS Beyond a qualitative understanding of the life cycle environmental and economic impacts of a specific project the objective of this research is to give to building designers a tool necessary to test several options of design and materials choices in order to optimize their projects. We believe that the Level(s) LCA plugin has a great potential in offering a powerful design tool that can help reducing the impacts of the construction sector by making more direct environmental and cost impact estimates as an integrated part of BIM use practices. One of the constraints of the current practice is the lack of environmental data complying with the most updated standards of analysis, EN 15804+A2 (EN15804:2012+A2:2019), since most of it is still complying with EN 15804+A1 which has fewer environmental indicators and does not, for example, consider individually GWPf, GWPb, GWPl. For further analysis and the development of the plugin, the authors also hope that more EPDs will be available in the near future from Portuguese suppliers, in order to enrich the database and consequently allow for a more complete analysis in building design practices. For the Level(s) Cost further indicators should be considered, in addition to initial costs, as the labors hours to complete each task is large and the need for specialization is demanding. Regarding LC Stages, the current analysis in the plugin is focused on the product and construction stages (A) only; for future analysis and expansion of the plugin we are working on addressing further stages, such as use stage (B1-B7) end-of-life stage (C1-C4) and potential benefits from reuse, recovery and recycling (D). For a deeper analysis and subsequent research, all stages of the LC should be taken into consideration.
Figure 11. Caption of the 3D visualization of the total results of Level(s) GWP analysis of walls.
from the EPDs and the calculations automatically done by the plugin applying the analytical models explained in chapter 2.3.
3.4 Level(s) cost analysis As mentioned previously in chapter 2.1 (Table 1), the Level(s) Cost in the scope of this research takes into consideration the initial costs of each material or element. A schedule with the total analysis is generated and the possibility to export to MS-Excel is given to the user. Also, a 3D visualization is automatically generated, and a color scheme is applied as in the environmental analysis (Figure 12). From the 3D visualization of the impacts visible in Figure 12 the authors can conclude that the highest costs are in the façade and exterior elements of the building.
ACKNOWLEDGMENTS This work was supported through the CERIS, ISTUniversity of Lisbon and Fundação para a Ciência e Tecnologia (FCT) reference UI/BD/153398/2022. The authors would also like to thank the support of the Circular EcoBIM project, funded by EEA Grants within the Environment programme. The authors would also like to express their gratitude to BUILT CoLAB for their support on the technological development of the plugin. Finally, the authors would like to demonstrate their gratitude to the design office ‘Atelier dos Remédios’ and Prof. Dr. Francisco T. Bastos for providing a case study.
Figure 12. Caption of the 3D visualization of the total results of Level(s) Cost analysis of walls.
3.5 Validation The application and validation of the plugin in the case study, and the possibility of testing the plugin with different construction alternatives, is by the authors seen as a big improvement for the twin transition (green and digital) that is needed in the construction sector: green with the Level(s) LCA and Level(s) GWP analysis and digital through BIM methodology and software development, aligning the Level(s) framework and EU standards with BIM.
REFERENCES CEN (2011). Sustainability of construction works – Assessment of buildings – Part 2: Framework for the assessment of environmental performance. EN 15643-2. Brussels, Belgium, Comité Européen de Normalisation.
25
CEN (2012). Sustainability of construction works – Assessment of buildings – Part 4: Framework for the assessment of economic performance. EN 15643-4. Brussels, Belgium, Comité Européen de Normalisation. Dodd, N., Donatello, S., & Cordella, M. (2020). Level(s) indicator 1.2: Life cycle Global Warming Potential (GWP) User manual: overview, instructions and guidance (publication version 1.0). EC (2010.) Retrieved (March 11, 2020), from https://ec. europa.eu/info/law/better-regulation/have-your-say/initia tives/12910-Revision-of-the-Energy-Performance-of-Bui ldings-Directive-2010-31-EU EC (2019). Communication from the European Commission: The European Green Deal. COM (2019) 640, (12 December 2019). Brussels, Belgium. EC (2020). Communication from the European Commission: A Renovation Wave for Europe – Greening our buildings, creating jobs, improving lives. COM(2020) 662. Brussels, Belgium. EC (2021). Retrieved (March 11, 2021), from https://ec. europa.eu/energy/topics/energy-efficiency/energy-efficie nt-buildings/renovation-wave_en EC: LEVEL(S) Putting whole life carbon principles into practice. (2021) b, doi:10.2779/79139
EC: LEVEL(S) Putting circularity into practice. (2021) c, doi:10.2779 /19010 EN15804:2012+A22019, Sustainability of construction works – Environmental product declarations – Core rules for the product category of construction products. https://www.kuantokusta.pt Martina F. (2018). Development of a BIM-based product data template for sustainability, Master’s degree thesis in ARC I – Scuola di Architettura Urbanistica Ingegneria delle Costruzioni, PolitecnicodiMilano. http://hdl.handle.net/10589/143041 Santos, R.; Costa, A. A. & Silvestre, J. D.; Pyl, L., (2019). Integration of LCA and LCC analysis within a BIM-based environment. Automation in Construction. 103, 127–149, March, DOI: 10.1016/j.autcon.(2019.02.011) Santos, R.; Costa, A. A. & Silvestre, J. D.; Vandenbergh, T.; Pyl, L., (2020) a. BIM-based life cycle assessment and life cycle costing of an office building in Western Europe. Building and Environment. 169, 106568, February, DOI: 10.1061/9780784480847.007 https://www.oekobaudat.de/
26
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
How can LCA inform early-stage design to meet Danish regulations? The sustainability opportunity metric A. Kamari & C. Schultz DIGIT and Department of Civil and Architectural Engineering, Aarhus University, Aarhus, Denmark
ABSTRACT: The National Strategy for Sustainable Construction in Denmark (in 2021) has agreed to include new requirements in which the building environmental impact calculation will become mandatory. We present a general decision-support system framework that delivers LCA information on highly uncertain early-stage designs, aimed at architects and architect engineers. Our key contribution is that we define a new sustainability opportunity metric assigned to early-design BIM model components. Our metric identifies sustainability opportunity “hot spots” (opportunities and risks) based on rapidly envisioning and qualitatively evaluating tens to hundreds of “what if ” scenarios that account for early-stage design uncertainty. A prototype demonstrates two visualizations of our metric on two BIM models.
1 INTRODUCTION Integrating a life-cycle perspective into the building and construction sector is significantly gaining attention in EU states and worldwide (BPIE 2022). Likewise, the National Strategy for Sustainable Construction in Denmark (2021) has now agreed to include new requirements in which the building environmental impact calculation will become mandatory. The initiative will ensure the introduction of requirements for the calculation of environmental impact in 2023 for all new construction and the introduction of a limit value for new construction over 1000 m2 from 2023 and otherwise new construction from 2025. Then, the requirement is gradually reduced to reach 7.5 kg CO2 eq/m2 /year by 2029 (Boligministeriet 2021). Life Cycle Assessment – LCA (Hellweg & Canals 2014) is a popular approach to measure and address environmental impacts over the whole life cycle of a building. Many LCA tools (Bueno & Fabricio 2018) have been developed that could in principle help architects and engineers to fulfill the above upcoming regulations. However, practical problems abound: (a) they can be complex and time-consuming to use (b) they are difficult to use in the early stages of design due to the user’s lack of finalized design details and a lack of access to necessary LCA data. This paper presents a new sustainability opportunity metric that is tailored to give architects LCA feedback on their early-stage designs. The central idea of the metric is to evaluate “what if” scenarios that capture ways in which the design may change and estimate the change in Key Performance Indicator (KPI) scores relative to the current design (i.e. the “no change” scenario). These “what if” scenarios are recalculated DOI 10.1201/9781003354222-4
continuously with each design change, and statistics are gathered and summarized in KPI change profiles assigned to each building component. These profiles give architects a rapid indication of which components offer opportunities and risks for improving or degrading the sustainability of their work in progress design. While this approach bares similarities with other related areas such as sensitivity analysis, the major departure is due to the significant uncertainty during early-stage architectural design on critical details that can drastically impact KPI scores. This raises the following three core issues: – the results of high-fidelity simulations may not apply when those simulations are highly sensitive to aspects of a design that are likely to change in unpredictable ways, e.g. fine-grained ray tracing for daylight analysis; – the ways in which actions will impact a design may only be expressed in coarse terms, rather than precise changes e.g. the geometry of components is expected to change (small change, large change) but there is not enough information to express these “what if” scenarios with precise geometric shapes; a correspondingly qualitative spatial way of expressing changing geometry needs to be defined that is sufficiently concrete so that KPI simulations and calculations can still be employed; – from a workflow perspective, the design activity in the early stages is rather exploratory, and thus LCA feedback given to architects needs to be delivered in real time, i.e. in the order of a second or less; this rules out the use of high-fidelity simulations and calculations that may take tens of seconds to hours.
27
2.2 LCA in the Danish building industry
Thus, our key research question (RQ) is: how can existing research on building LCA simplification be incorporated into a workflow process that supports early-stage architectural design with respect to the three core issues stemming from design uncertainty? Drawing on existing research on LCA simplification, sensitivity analysis from systems engineering, and qualitative reasoning about physical systems (Bobrow 1984; Kuipers 1994) we present a novel framework (Section 5) for computing our sustainability opportunity metric. The central focus and contributions of this paper are: C1. establishing the overall framework in a modular and extensible way, tailored to address research question RQ (Sections 5 and 5.1). C2. the formal definition of the sustainability opportunity metric and deriving and interpreting the KPI change profiles within our very general framework (Sections 5.2). In addition, we have developed a first prototype that demonstrates the framework on two building projects, one developed by a student in a master level course called “Integrated Engineering Project”, at the Civil and Architecture Engineering Department at Aarhus University, and another one from Autodesk Revit Sample projects (Section 6).
Building sustainability was first introduced in 2014 in Denmark via a governmental building strategy (Danish government 2014), as one of the five focus areas for the future political work within the sector. The strategy describes the possibility to introduce a voluntary sustainability class as part of the Danish building regulation and addressed the need to develop harmonized LCA and LCC tools for buildings. The Danish LCAByg1 software tool (Birgisdottir & Rasmussen 2019) was conceptualized, developed, and operated by the Danish Building Research Institute (today known as BUILD). The tool as free-to-use was launched in 2015. LCAByg aims at calculating the environmental profile and the consumption of resources from the 5 phases and 17 stages. However, in LCAbyg, it is currently only possible to calculate a selection of the 17 modules, i.e. production and transport of construction products (A1-3), replacement of building parts (B4), operational energy (B6) and waste processing at the end-of-life stage (C3-4). The data in the LCAByg-software is based on the German database for construction products, ÖKOBAUDAT2 . These data are generic, i.e., an average data from a product group. On the other hand, product-specific data input is often available from a manufacturer.The former data is in the software’s database, whereas the latter comes through Environmental Product Declarations (EPDs) for specific product data. Product-specific data increases the reliability of the LCA calculation, and thus usage of generic data should be kept to a minimum if possible and only used if necessary. Thus, the development and use of EPDs and more product-specific data to further mainstream LCA usage is needed by creating a demand from the products. The specific data can be incorporated in the assessment giving a more robust calculation that reflects a more accurate outcome (Birgisdottir & Rasmussen 2016).
2 BACKGROUND LCA is regulated by ISO 14040 and has been adapted for buildings in the EN 15978 standard. The LCA’s four fundamental steps, according to ISO 14040, are: Goal and Scope definition, Inventory Analysis, Impact Assessment, and Interpretation. According to EN 15978 standards, a complete LCA analysis includes five phases and 17 stages, including: Product (A1A3), Construction Process (A4-A5), Usage (B1-B7), End of Life (C1-C4), and Benefits (D) stages. 2.1 LCA for early building design
3 RELATED WORK
LCA can be used in the early design phase with reliable success (Kamari et al. 2022; Kotula & Kamari 2021). It can be used to enable better early-stage decision-making by providing feedback on the environmental impacts of Building Information Modeling – BIM (Eastman et al. 2011; Kamari & Kirkegaard 2019) design choices (Basbagill et al. 2013). In this view, LCA has a higher potential to influence the overall environment impact of building projects through optimization on embodied carbon and energy (Zimmermann et al. 2019). Integrating LCA in the early design phases enables comparison of design decisions which can be very decisive for the later stages of the project, e.g. comparison and prioritization of design and construction scenarios, selection of the materials, significance of building elements etc.
Integrating environmental impact measures through LCA calculations in the early design stages entails enabling and empowering architects to conduct these analyses in their daily practices. According to Hollberg & Ruth (2016), three key challenges are: Challenge 1. Complex and time-consuming process: The reason is (a) buildings usually consist of different building elements, each consisting of many different materials, which makes the necessary assessment of all material quantities a laborious task, and (b) many buildings possess a very long lifespan with a use phase that can easily last hundreds of years. A building’s use 1 2
28
https://www.lcabyg.dk/ https://www.oekobaudat.de/en.html
geometric representations and the way of communicating and representing a design. There is often a non-intuitive and impractical user interface so that development is essential in its simplification. In this reagrd, Kamari et al. (2021) reflects that appropriate and effective visualisation techniques should be developed for architects that provide a good overview of the problem in question. At the same time data visualization should be functional, insightful and enlightening, in order for the architects to fully grasp and accept the evidence it depicts as useful inputs in the design process.
may change over time, introducing a high degree of uncertainty. The end-of-life scenario is also uncertain. Challenge 2. Lack of knowledge: Architects lack the knowledge and experience necessary to perform an LCA. Challenge 3. Requirements for optimization: Evaluating building design through LCA is not sufficient on its own, as it does nothing to improve the design. In order to minimize environmental impact, an optimization based on different design variants is needed. For example, as most buildings are unique designs, the parameters that influence their energy performance vary from building to building. To tackle the above challenges, simplified approaches to conducting an LCA are needed which incorporate the knowledge of LCA experts in a design tool and allow architects to focus on their major task of designing building concepts (Hollberg & Ruth 2016). Integrating LCA in architects’ practices can enable them to identify and assess the climate impact of specific elements in a building as the structural system, outer walls, floors, windows, etc., to give an overview of their different significance to the building and what may be most beneficial in a given situation (“hot spots”). A comparison of different materials can be made and their environmental profile of the building. This must be used with care since the materials and their relevant quantities may change depending on the usage and purpose of the material and its functional abilities. Furthermore, it can show the relationship between a material and its energy and divide it into operational energy during use and the embodied energy related to the material itself. This is highly useful when comparing two types of materials or building elements, e.g., outer wall elements, and how their energy usage differentiates to determine the most optimal solution in the given project (ibid.).
3.2 Early design LCA methodology Myriad studies exist that have already strived to address the simplified LCA methodogloy for early stage design (Meex et al. 2018; Santos et al. 2019; Wittstock et al. 2012; Zimmermann et al. 2019). In addition, instructions have been developed in existing sustainability certification systems, i.e., the German DGNB3 system. We exploit these tools and seek to generalise and standarise the process of KPI evaluator simplification in our framework. 4 METHODOLOGY We use the methodology illustrated in Figure 1, which is based on two needs: (a) user requirements for ‘architect-friendly’ LCA tools (i.e., usable by architects during the early stages of the design process), and (b) criteria for simplifying the LCA methodology and parametrizing the calculation method in order to make it more applicable in an LCA software tool during the early stages of the design process (according to Meex et al. 2018).
3.1 Design-oriented user requirements Various studies have been carried out on the development of building performance simulation tools for early design stages of architecture work (Purup & Petersen 2020a, b). It is often argued that there is a generalized process of disconnection between the design process and the architectural simulation process. Tools for building performance simulation require the input of a large amount of data (Mahdavi 2011) being a common challenge for architects (FernandezAntolin et al. 2020; Schlueter & Thesseling 2009), and therefore, limiting data entry is vital. Many of the input data are not available in the early design stages, thus, it is necessary to use default values and templates. Lam et al. (2004) discuss that complex performance simulation tools do not necessarily provide better support for decision making. So, for architects, simple energy simulation tools can offer more advantages than complex ones (Warren 2002). With a focus on environmental simulation tools, Naboni (2013) argues the need for simulation tools to adapt to new needs in architecture, such as their
Figure 1. Framework on requirements for design-oriented LCA tools for early design (adapted from Meex et al. 2018).
The focus in this paper is on our new sustainability metric, and so other stages in the overall workflow are only presented briefly, with reference to prior research that we are building on.A formal evaluation and refinement of the framework via stakeholder engagement studies is also out of scope of the present paper and is left as a topic for the next research stage. We leverage existing research on simplifying LCA calculations as follows. LCA stages included (according to EN 15978): A1-A3 stages that cover extraction and processing of raw material, transportation of raw materials to manufacturers of building products and 3
29
https://www.dgnb.de/en/
are formally represented as action trees in the Renovation Domain Model, NovaDM (Kamari et al. 2019; Schultz & Kamari 2021). We employ a previously developed n-ary coverage strategy (ibid.) to generate a sample set of the full design space: 1-ary coverage means that every action is used in at least one sampled design, 2-ary coverage means that every (non-mutually exclusive) pair of atomic actions appear together in at least one sampled design, and so on. Each sampled design corresponds to a “what if” scenario used in the sustainability opportunity metric.
manufacturing of building products. We employ the following calculation: EIxA1−A3 =
i
QaM × EIaM
(1)
a=1
where EIxA1−A3 is the environmental impact of category x resulting from product and construction stage (A1-A3); i is the number of existing materials; QaM is the quantity of material a; EIaM is the environmental impact of category x resulting from production a. For LCA data-input we use the material database from the Ökobaudat platform, which consists of 900 datasets in compliance with EN 15804. The results of LCAs will be presented in GWP (kg CO2 eq.).
5.1 Generating and applying atomic actions The atomic action generation process is configured according to the project at hand to generate sets of actions that reflect future potential changes to the design. Actions are represented in a qualitative (rather than numerically precise) way that can be fed into the coarsened KPI evaluators. Changes include adding new building components with default parameters, e.g. balconies or an HVAC system. Existing components can be removed, modified based on their attributes, or geometrically transformed in coarse ways according to an ontology of the type of component and the context of the building, e.g. the dimensions of windows, walls, slabs, etc. can be scaled up or down according to a set of pre-determined (knowledge-based) plausible changes. Our key innovation in handling geometric change is that the consequences of changing the geometry of a particular component on the full BIM model can be approximated by ignoring the requirement of spatial consistency, which is often sufficient for coarse KPI evaluation. For example, we can consider a “what if” scenario where a window to the exterior is enlarged somewhere close to the size of the embedding wall without needing to check whether it intersects a door on the same wall, or whether any other spatial inconsistency occurs. Qualitative estimates of the change in daylight, ambient illuminance, thermal comfort, cost, etc. can be obtained by applying the coarsened KPI formulas with the new geometry, even if the exact geometric details will differ (to ensure spatial consistency) if that design option is taken later. In the case of components that impact the topology of a BIM model (and spatial constraints in general (Schultz et al. 2017)), we take an approach analogous to Cuttle’s “first-bounce” simplification of a comprehensive ray trace described above: a geometric change to a component is recursively propagated, breadth first, to all topologically connected components, but each component is only updated once at most.4 Often this will leave the BIM model in a spatially inconsistent, physically unrealizable state where numerous
5 FRAMEWORK AND CONCEPTUALIZATION Our framework is illustrated in Figure 2: boxes represent artefacts (files, data); round boxes represent computation procedures (functions); the fat arrow represents function transformation; thin arrows represent dataflow; stars indicate the parts of our framework that are novel.
Figure 2.
Sustainability opportunity metric framework.
KPI evaluator simplification is performed once for each KPI evaluator, independently of the projects they are used in. Qualitative simplification is primarily a manual model development process in our framework, based on qualitative modelling (Kuipers 1994), in combination with automated surrogate modelling (Westermann & Evans 2019). As an example, an excellent case study of manual qualitative simplification for lighting design is by Cuttle (2015) who defines mean room surface existence as an approximation of the average eye illuminance in a room by assuming that the lumens are uniformly distributed, and is calculated using the first-bounce lumens, with corresponding simple equations (derived from more complex iterative ray tracing calculations) for first reflected flux and room absorption. We refer readers to (Schultz et al. 2009) for a detailed case study on qualitative lighting KPIs that combines research by Flynn (1973) and Cuttle. Dataflows occur continuously, triggered by each change to the BIM model. The set of individual (atomic) actions that can be performed on a BIM model
4
Standard methods are used to create topological connectivity graphs of BIM models e.g. (Khalili & Chua, 2015). Dynamic geometry (Kortenkamp, 1999) propagates spatial changes, with priority firing to ensure breadth-first propagation and a tabu list to ensure elements are only changed at most once.
30
of the design, and how are they calculated (formulas, simulations). As many design details are unavailable, the calculation method is typically an adapted or simplified version of a complete and comprehensive calculation method. Let LCA metric m be a function applied to a BIM model B that evaluates to a real number score k, denoted m(B) = k. For brevity in the following definitions (and without loss of generality) we assume that LCA metrics are minimizing metrics, i.e. a lower score is better. (Aspect 4) LCA metric sensitivity and thresholds: sensitivity refers to the relationship where a relatively small change in the design has a relatively large impact on the overall LCA metric score. This aspect in the framework determines which sets of actions are interpreted as a single change and specifies quantity intervals of the LCA metric scores that define (qualitatively) no change, small change, and large change. Each single change Ai is a set of non-mutually exclusive actions Ai ⊆ ASIG as defined by the user and the context. To clarify, a single change Ai may involve exactly one action, e.g. changing the geometry of a particular floor slab, or more than one action, e.g. replacing a set of windows (a set of remove element actions) with balconies (a set of add element actions). Formally we denote the set of all single changes as A ⊆ Powerset(ASIG ).5 In addition, critical LCA metric constant thresholds are defined, denoted K1 , ..., Kn , e.g. in the Danish FB23 regulations there is a mandate to achieve less that kg CO2 eq./m2 /year.
spatial constraints in the original design are violated. However, we argue (to be formally verified in a subsequent paper) that this “coarsened” BIM model is sufficient for estimating coarsened KPI values. This contrasts sound and complete spatial constraint satisfaction which often involves iterative cycles of change propagation to find a spatially consistent design. 5.2 Theoretical foundations In this subsection we present the sustainability opportunity metric in a formal way. A BIM model B=(E, R) consists of a set of elements E, and a set of relationships R. An action, a, is a change that can be applied to a BIM model resulting in a new BIM model B = (E , R ), denoted a(B) = B . Each action is associated with the element, e, that it acts on, denoted subject(a) = e, such that either e ∈ E, or e ∈ E . Given a set of actions A, subject(A) = {subject(a)|a ∈ A}. Without loss of generality, we intend that the order that a set of actions A are applied does not change the resulting BIM model, that is, action application is necessarily associative, ∀ai , aj ∈ A · ai (aj (B)) = aj (ai (B)). We denote the application of a set of actions A on BIM model B as apply (A, B), defined recursively: apply (A, B) = if (A = {}) then B else (apply(A/{a}, a(B)) such that a ∈ A). Utilizing our framework consists of four major interrelated aspects that determine how the atomic action generation process is configured, and which KPI evaluators are to be employed (refer to Figure 2). (Aspect 1) Design uncertainty: how uncertainty is defined and formally represented, and what the scope of uncertainty under consideration is. This aspect is based on the targeted stage of early design, in which many critical design details that have a direct and potentially significant impact on the LCA metric scores are yet to be finalized. E.g. geometry, existence of particular elements, materials of elements. (Aspect 2) Design changes under consideration: which actions can be performed on the current design that change it. Applying some combination of actions generates a candidate design solution, and the set of all solutions (at the given step in the design process) defines the current design space. E.g. geometric changes represented as qualitative changes in volume; a set of distinct design solutions (e.g. comparing a solution with windows and a solution with balconies). Let B be the given BIM model that represents the current design, referred to below as the reference design. Let ASIG = {a1 , . . . , an } be the set of n defined actions. Two actions ai , aj ∈ ASIG are mutually exclusive, denoted mutex(ai , aj ), if they cannot both be applied together, e.g. removing a window and then changing that window’s geometry. The full design space Dfull is the set of BIM models that results from applying some combination of non-mutually exclusive actions, Dfull = {B |A ⊆ ASIG ∧ (∀ai , aj ∈ A · ¬mutex(ai , aj )) ∧ apply(A,B)=B }. (Aspect 3) LCA scope and metrics: which life cycle stages are within scope (i.e. A, B, C, D), which LCA metrics are used to calculate the environmental impact
5.3 Sustainability Opportunity Metrics Based on decisions for the above four aspects, the tool delivers LCA design information to the architect by computing the following two sustainability opportunity metrics for each element in the reference design B given metric m, and the set of atomic actions A. Firstly, the relationship between two key landmark values is critical for determining how the architect can make the most effective use of the two sustainability opportunity metrics, namely the specified threshold K for metric m, and the reference design score m(B): if K < m(B) then the reference design is already exceeding the required threshold and thus the designer is likely to be seeking ways of reducing this score; if m(B) ≤ K then the designer is likely to want to determine (a) how robust this score is with respect to the inherent uncertainty in the early design stage i.e. how sensitive is the condition of satisfying threshold K to small changes in the design; (b) which parts of the design have “slack” in terms of metric m, such that the designer is permitted to explore more environmentally expensive actions for gains in other aspects of the design (e.g. aesthetics) while still remaining within the required threshold. 5
31
The powerset of set S is the set of all subsets of S.
Sustainability Opportunity Metric [Element impact relative to change]: this metric measures the potential for each element e under change to have a positive or negative impact of the LCA metric m by calculating a m-score profile for e. The algorithm for the procedure E_Score calculates single change scores for element e. In the following algorithm, “:=” denotes variable assignment, “[x]” denotes a list with element x, “[]” denotes the empty list, and “ˆ” denotes list concatenation. Algorithm E_Score Input: e, B, A, m Output: S (a list of LCA metric scores) S = [] for each A in A: if e ∈ subject(A) then S := S ˆ [m(apply(A,B))] return S Each element, e, is subsequently assigned a score profile for metric m by calculating statistics over the list of scores delivered by the E_Score, as summarized in Table 1, where Se = E_Score(e), Se [i] denotes the ith element of Se , and n denotes the number of elements in Se .
category in the EPD database e.g. fiberglass reinforced concrete is mapped to concrete. Material categories are: – Aluminum – Asphalt – Brick – Ceramic – Cement – Clay
– Plaster – Roof tile – Insulation – Sand – Steel – Wood
To account for geometric uncertainty, four volume factors are defined for each component type that define four qualitative cases: small increase, large increate, small decrease, large decrease. The geometrically modified components are: – Roof – Wall – Floor – Window – Door – Ceiling – Stair
– Railing – Ramp – Furniture – Arch. Column – Curtain System – Curtain Mullion – Curtain Panel
– Strl. Column – Strl. Foundation – Strl. Frame – Strl. Framing System – Strl. Truss – Strl. Stiffener – Strl. Connections
Aspect 2 design changes: the set of atomic actions are derived directly from material and geometric uncertainty. Four actions are defined for each component that each modify its volume by a factor, depending on the component type. For this preliminary demonstration we do not propagate geometric consequences. This action set is used to generate 820 and 1596 “what if” scenarios, respectively for each BIM model in the demonstration. Aspect 3 LCA scope and metrics: in this case we focus on lifecycle stages A1-A3 and use the simplified LCA GWP metric (Equation 1) from Section 4.
6 PROTOTYPE DEMONSTRATION TOOL We have developed an operational prototype as a preliminary demonstration of our framework using two BIM models, consisting of 205 and 399 building components, respectively. In setting up the prototype we follow the four configuration aspects in Section 3.2. Aspect 1 design uncertainty: To account for design uncertainty in material selection, BIM model materials are mapped to their next more general material Table 1.
– Concrete (cast in situ) – Concrete (precast) – Glass – Metal – Mineral wool – Plastic
Score profile of element e and the corresponding sustainability opportunity/risk interpretation.
Metric
Interpretation
Improving potential: m(B) – min(Se )
Potential (extreme) improvement afforded by at least one action involving element e. Potential (extreme) degradation afforded by at least one action involving element e. High range indicates that the LCA metric m is sensitive to relatively small (i.e. single atomic) changes associated with element e.
Degrading potential: max(Se ) − m(B) Range: max(Se ) – min(Se ) n
i=1 Se [i]
Mean: µ =
n n
Variance: γ =
Skewness:
i=1 (Se [i]
− µ)2
n
1 n i=1 (Se [i] n γ 3/2
− µ)3
Mean below m(B) (resp. above m(B)) indicates, on average, actions directly involving e have potential for improving (resp. degrading) overall sustainability of the design, and thus element e is a “hot spot” for sustainability opportunity (resp. risk). Indicates expectation that e will realize its mean score (µ), assuming each action is equally likely. Low variance indicates high likelihood (less focus needed by architect), high variance suggests care must be taken to ensure changes involving e improve rather than degrade LCA metric m (i.e. suggesting e is a “hot spot” for sustainability opportunity and risk). Building on variance, highly positive (resp. negative) skew indicates most actions involving e score on the lower (resp. upper) side of the distribution. Care must be taken to avoid (resp. seek out) the few degrading (resp. improving) actions in the long tail of the distribution.
32
Aspect 4 Sensitivity: we do not preconfigure threshold sensitivity values here, and instead use the results to explore alternative ways of visualizing our sustainability opportunity metrics. Our system executes a sequence of activities, as illustrated in Figure 3. The tool requires only one activity from the user, which is an initialization of the LCA tool by clicking on its plug-in icon in the Autodesk Revit software. Our developed Revit add-on automatically extracts the geometry with the properties needed for further evaluation. Next, the tool automatically links the BIM elements from the material take-off with the material database created in a Microsoft Excel spreadsheet (developed by the authors).
V2 Threshold cutoff : (Figure 5) a threshold percentage α is specified. All components that achieve a relative score above α are colored green / red. The user manipulates a slider that determines the value of α ranging from 0%-100%, so the user sees the cutoff threshold animated as they gradually adjust the slider to more readily identify sustainability opportunity “hot spots”.
Figure 5. Threshold cutoff with four α values in sustainability opportunity mode.
7 CONCLUSIONS The major contribution and focus in this paper is (a) the development of a general framework tailored to equipping early-stage building designers with LCA information, and (b) a detailed formal specification of our sustainability opportunity metric (see Table 1). We exploit prior research in renovation domain modelling for specifying atomic actions and automatically generating useful design sample sets based on coverage analysis. We are in the process of conducting a series of studies with MSc architecture engineering students, and workshops with key architecture industry experts, to assess whether our framework and tool implementations indeed facilitate meaningful LCA input in early-stage design. Moreover, we are further developing the LCA database, the calculation methods and its required assumptions to include more LCA stages in line with EN 15978 standards (i.e., B4: replacement of building parts, B6: operational energy, and C3-4: waste processing at the end-of-life stage).
Figure 3. The LCA tool workflow.
The whole process of assigning Revit materials to the material database takes less than a few seconds. Next, the LCA BIM-based tool calculates (a) the total GWP of each Revit element, and (b) the total GWP for the whole BIM project. The tool then records the results, which are used to calculate the LCA profile of individual elements. Next, the add-on automatically calculates the opportunity and risk co-efficient for each element. Next, the plug-in presents the result to the user, including overriding the view display style in Revit, and changing the elements’ color. In all visualization approaches we implement two separate modes: either only opportunity (green) or only risk (red). We subsequently implement two visualizations, V1 Gradient, and V2 Threshold. V1 Gradient: (Figure 4) All components are colored along a gradient from white to green/red according to their mean sustainability score relative to the global minimum and maximum means (across all components).
REFERENCES Baitz, M., Albrecht, S., Brauner, E., Broadbent, C., Castellan, G., Conrath, P., ... & Tikana, L. (2013). LCA’s theory and practice: like ebony and ivory living in perfect harmony? International Journal of Life Cycle Assessment, 18(1), 5–13. Basbagill, J., Flager, F., Lepech, M., & Fischer, M. (2013). Application of life-cycle assessment to early stage building design for reduced embodied environmental impacts. Building and Environment, 60, 81–92. Birgisdottir, H., & Rasmussen, F.N. (2016). Introduction to LCA of Buildings. Danish Transport and Construction Agency: Copenhagen, Denmark. Birgisdottir, H., & Rasmussen, F.N. (2019). Development of LCAbyg: A national Life Cycle Assessment tool for buildings in Denmark. In IOP Conference Series: Earth and Environmental Science (Vol. 290, No. 1, p. 012039). IOP Publishing.
Figure 4. Gradient visualization using sustainability opportunity mode (green, left) and sustainability risk mode (red, right).
33
Lam, K. P., Huang,Y. C., & Zhai, C. (2004). Energy modeling tools assessment for early design phase. Center for Building Performance and Diagnostics School of Architecture, Carnegie Mellon University: Pittsburgh, PA, USA. Lam, K. P., Wong, N. H., & Henry, F. (1999).A study of the use of performance-based simulation tools for building design and evaluation in Singapore. Architecture, 1, 11–13. Mahdavi, A. (2011). The human dimension of building performance simulation. In 12th International IBPSA Conference: Building Simulation (pp. 14–16). Meex, E., Hollberg, A., Knapen, E., Hildebrand, L., & Verbeeck, G. (2018). Requirements for applying LCAbased environmental impact assessment tools in the early stages of building design. Building and Environment, 133, 228–236. Naboni, E. (2013). Environmental simulation tools in architectural practice. In PLEA 2013, Munich, Germany. Purup, P. B., & Petersen, S. (2020a). Research framework for development of building performance simulation tools for early design stages. Automation in Construction, 109. Purup, P. B., & Petersen, S. (2020b). Requirement analysis for building performance simulation tools conformed to fit design practice. Automation in Construction, 116. Santos, R., Costa, A. A., Silvestre, J. D., & Pyl, L. (2019). Integration of LCA and LCC analysis within a BIM-based environment. Automation in Construction, 103, 127–149. Schlueter, A., & Thesseling, F. (2009). Building information model based energy/exergy performance assessment in early design stages. Automation in construction, 18(2), 153–163. Schultz, C., Bhatt, M., & Borrmann, A. (2017). Bridging qualitative spatial constraints and feature-based parametric modelling: Expressing visibility and movement constraints. Advanced Engineering Informatics, 31, 2–17. Schultz, C., Amor, R., Lobb, B., & Guesgen, H. W. (2009). Qualitative design support for engineering and architecture. Advanced Engineering Informatics, 23(1), 68–80. Schultz, C., & Kamari, A. (2021). Diversity in Renovation Design: Theoretical Foundations of the Renovation Domain Model. Journal of Computing in Civil Engineering, 35(4). Warren, P. (2002). Bringing simulation to application. FaberMaunsell Limited. Westermann, P., & Evins, R. (2019). Surrogate modelling for sustainable building design–A review. Energy and Buildings, 198, 170–186. Wittstock, B., Gantner, J., Saunders, K.L.T., Anderson, J., Carter, C., Gyetvai, Z., ... & Sjostrom, T.B.W.C. (2012). EeBGuide guidance document part B: buildings. Operational guidance for life cycle assessment studies of the Energy-Efficient Buildings Initiative, 1–360. Zimmermann, R. K., Kanafani, K., Rasmussen, F. N., & Birgisdottir, H. (2019). Early Design Stage Building LCA using the LCAbyg tool: Comparing Cases for Early Stage and Detailed LCA Approaches. In IOP Conference Series: Earth and Environmental Science (Vol. 323, No. 1). IOP Publishing. Zimmermann, R. K., Andersen, C. M. E., Kanafani, K., & Birgisdottir, H. (2021). Whole Life Carbon Assessment of 60 buildings: Possibilities to develop benchmark values for LCA of buildings.
Bobrow, D. G. (1984). Qualitative reasoning about physical systems: an introduction. Artificial intelligence, 24(1–3), 1–5. Boligministeriet, I. O. (2021). National strategi for bæredygtigt byggeri. Retrieved from Link. BPIE-Buildings Performance Institute Europe (2022). A lifecycle perspective on the building sector: Good practice in Europe. Retrieved from Link. Bueno, C., & Fabricio, M.M. (2018). Comparative analysis between a complete LCA study and results from a BIMLCA plug-in. Automation in construction, 90, 188–200. Cuttle, C. (2015). Lighting design: a perception-based approach. Routledge. Danish Government (2014). Vejen til et styrket byggeri i Danmark – Regeringens byggepolitiske strategi (Copenhagen). Eastman, C. M., Eastman, C., Teicholz, P., Sacks, R., & Liston, K. (2011). BIM handbook: A guide to building information modeling for owners, managers, designers, engineers and contractors. John Wiley & Sons. Fernandez-Antolin, M.M., del-Río, J.M., del Ama Gonzalo, F., & Gonzalez-Lezcano, R.A. (2020). The relationship between the use of building performance simulation tools by recent graduate architects and the deficiencies in architectural education. Energies, 13(5), 1134. Flynn, J. E., Spencer, T. J., Martyniuk, O., & Hendrick, C. (1973). Interim study of procedures for investigating the effect of light on impression and behavior. Journal of the Illuminating Engineering Society, 3(1), 87-94. Hellweg, S., & Canals, L.M. (2014). Emerging approaches, challenges and opportunities in life cycle assessment. Science, 344(6188), 1109–1113. Hollberg, A., & Ruth, J. (2016). LCA in architectural design—a parametric approach. The International Journal of Life Cycle Assessment, 21(7), 943–960. Kamari, A., Kotula, B.M., & Schultz, C. (2022). A BIMbased LCA tool for sustainable building design during the early design stage. Smart and Sustainable Built Environment, 22. Kamari, A., Jensen, S.R., Petersen, S., & Kirkegaard, P.H. (2021). Sustainability Key Performance Indicators’ (KPIs) assessment and visualization aimed at architects in (early) renovation design processes. Nordic Journal of Architectural Research, 33(2), 41–80. Kamari, A., Schultz, C., & Kirkegaard, P.H. (2019). Constraint-based Renovation Design Support through the Renovation Domain Model. Automation in Construction, 104, 265–280. Kamari, A., & Kirkegaard, P.H. (2019). Potential shift of Integrated Design (ID) through BIM in Sustainable Building Renovation. IOP publications, proceeding of Sustainable Built Environment (SBE) Cardiff 2019, Cardiff, UK. Khalili, A., & Chua, D. H. (2015). IFC-based graph data model for topological queries on building elements. Journal of Computing in Civil Engineering, 29(3), 04014046. Kortenkamp, U. H. (1999). Foundations of dynamic geometry. Doctoral dissertation, ETH Zurich. Kotula, B.M., & Kamari, A. (2021). Development of a BIMbased LCA Tool to Support Sustainable Building Design during the Early Design Stage. In ASA 2020 (the 54th International Conference of the Architectural Science Association),Auckland, New Zealand, (the paper won best research paper award). Kuipers, B. (1994). Qualitative reasoning: modeling and simulation with incomplete knowledge. MIT press.
34
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Life cycle potentials and improvement opportunities as guidance for early-stage design decisions J. Staudt, M. Margesin, C. Zong, F. Deghim & W. Lang Institute of Energy Efficient and Sustainable Design and Building, Munich, Germany
A. Zahedi & F. Petzold TUM School of Engineering and Design, Technical University of Munich, Munich, Germany
P. Schneider-Marin Department of Architecture and Technology, Faculty of Architecture and Design, Norwegian University of Science and Technology, Trondheim, Norway
ABSTRACT: Fundamental planning decisions made early in the design process have significant impact on the final building’s performance. Computer-assisted approaches can currently only help in a limited way during these critical and recurring phases of design variants generation, assessment, and further elaboration. We investigate the concept of potentials and how it can support a digitally-aided design process. In this project potential is defined as the development possibilities of design variants or variant branches taking into consideration different evaluation criteria. These potentials serve as a link between the early stages of design and prospective future outcomes. To demonstrate potentials and their implementation we use life-cycle assessment (LCA). A real-world case study demonstrates the process of designing and detailing with guidance from potentials and improvement opportunities via a graphically laid out selective decision tree. This method helps designers locate areas with the greatest impact, communicate them to stakeholders, and make more informed design decisions.
1 INTRODUCTION Considering life cycle emissions of buildings early in the design process offers the possibility of creating the most impactful environmental improvements (Brophy & Lewis 2011). However, in a conventional linear design process, life cycle assessment (LCA) is usually performed during later phases after most design and material decisions have already been made. Performing LCA during early phases when many decisions are still open would require a multi-stage design space exploration resulting in a decision tree (Hollberg 2017). Such guided design paths can also point planners towards the most significant levers towards more sustainable outcomes and provide information to inform these decisions and help to communicate them to clients and other stakeholders. Current design assistance tends to (over)simplify during early phases to present users with a manageable amount of information. This can misguide them since they will be presented with choices suggesting precise information which does not reflect the uncertainty during early phases and the range of possible outcomes. Communicating and visualizing the range of possible outcomes and guiding designers towards improved outcomes is the goal of this research. The concept of DOI 10.1201/9781003354222-5
potentials has not been sufficiently studied as an effective way of guiding the decision-making process with uncertain information by focusing on the areas of the greatest possible impact. Performance indicators commonly used during later phases usually require detailed information. Since many decisions have not been made yet, the inclusion of uncertainty via value ranges is important to communicate the impact of early phase design decisions appropriately. Uncertainty with regards to LCA has been studied previously (Guimarães et al. 2019) but is not commonly part of feedback given to designers. Especially during early design phases, the added uncertainty by open design decisions can result in very significant differences between initial estimates and final outcomes (Backer & Lepech 2009). This paper aims at explaining the concept of potentials and improvement opportunities in the context of model-based and knowledge-based design and how they can be used during early design phases to guide users towards improved outcomes. Providing information early to support design decisions supports effective interaction between the architect’s early rough designs and the more precise analytical methods used during later stages. We are using LCA as an example to illustrate the concept of potentials and its application
35
improvement opportunities. (Harter et al. 2019) developed a method to include design uncertainties in life cycle energy assessment. Based on this method, (Schneider-Marin et al. 2020) performed a sensitivity analysis identifying design uncertainties with a strong influence on greenhouse gas emissions. This research develops this approach further into design guidance. The term potentials is often used to describe the potentials of LCA as a method itself. It has also been used to refer to environmental potentials in building construction during early design stages (Röck et al. 2018). Here, the LCA results are mapped onto a BIM model and “improvement potentials” are represented as box plots for various element classes. This, however, does not provide structure and guidance to users and does not show potentials and improvement opportunities across different scales. Potential life cycle performance (PLCP) has been described as an approach to calculate a range of plausible solutions to “provide a forecast of the LCP that can potentially be achieved in later design stages.” (p. 153, Hollberg 2017). Levels of Development (LOD) have been established as a useful concept in BIM to express “the maturity of the design information provided by the model, which comprises both a specification of the geometric detail as well as the semantic information required.” (Abualdenien et al. 2020). While these LODs are useful to document the progress of a linear design process, they do not guide users in the design process. Decision trees have been described as an ideal way to provide possible solutions for decision making (Hollberg 2017) but have not actually been implemented and visualized to provide guidance to improve embedded emissions outcomes. The selection and decision levels described in this paper capture a non-linear development process where various decisions can be made asynchronously which better reflects design practice during early stages. To capture and document tacit design knowledge, we devised the design episode approach, which divides the final design into smaller chapters and parts, with each episode relating to a design challenge and offering a solution utilizing narrative and storytelling tactics (Zahedi et al. 2022). Each Design Episode is a container that summarizes and documents a specific section of the design using written descriptions and the selection of related construction elements and spaces.
in a digitally supported design process to improve the sustainability of the design. Via a selective decision tree, designers receive visualized information revealing improvement opportunities at the whole building-, building part-, and material-level, highlighting areas with greater opportunities across these levels. Visualizing potentials and opportunities provides design guidance and assists in identifying areas of effective improvement. In this paper, we focus on embodied emissions reduction as an important indicator of the overall life-cycle performance of a building. The method is applied to a real-world sample project showing a design episode of designing and detailing with guidance by potentials and improvement opportunities, with a focus on embodied emission reduction using LCA information and construction knowledge. The method can be extended to other aspects such as construction cost, structure, operational energy, etc.
2 STATE OF THE ART This research is part of the EarlyBIM project (Abualdenien et al. 2020) focusing on early design stages (project brief, concept design, design development, equivalent to the German scale of fees for services by architects and engineers: Phases 1-3). As has been pointed out previously, the impact on the final outcome of a project is highest during these early stages (Østergård et al. 2017), yet at the same time, many decisions made during these early stages lack the inclusion of technical information and domain knowledge. To support and include the suggestions given by diverse domain experts, during the early stages of design, Zahedi & Petzold introduced an adaptive minimized machine-readable protocol based on BIM (Zahedi & Petzold 2019a). Consequently, an online platform was introduced by Meng et. al. (Meng et al. 2020) for facilitating interactive and intuitive visualizations of the collaborative workflow based on the above mentioned communication protocol. Furthermore, (Zahedi & Petzold 2019b) proposed various supportive methods for visual representation and exploration of simulation results during the early stages of design. To visualize the vagueness and uncertainties involved in building models across different design stages, Abualdenien & Borrmann (Abualdenien & Borrmann 2020), introduced a variety of visualization tools, including intrinsic, extrinsic, animation, and walkthroughs, that have been designed to display the information uncertainties related to various building elements. Using a template-based methodology, Jaskula et al. (Jaskula et al. 2021) presented a design variants’ evaluation and exploration tool called Archi-guide. Tools for early stage LCA have been developed and are available (e.g. CALAA and One Click LCA® ) but they do not sufficiently capture uncertainties and do not guide users towards areas with the greatest
3 TERMINOLOGY To clarify the use of certain terms which are central to our research we are defining them here as follows: 3.1 Potentials The design process is characterized by the generation and evaluation of variants and their potentials in their further development. In the context of this research,
36
potential is understood as the development possibilities of variants or variant branches and their parameters in consideration of similar solutions, showing the range of possible outcomes including uncertainties. The potential itself does not have a normative character (neither good nor bad), it is a positive statement based on empirical evidence (Lipsey 1975), which delimits the range of possibilities for future actions. The potential is, so to speak, contained within a design at a certain stage of the design process and contains a whole range of outcomes that can be activated via (future) design choices. The potential by itself does not judge or improve the outcome, but the potentials of different variants can be used to compare and identify improvement opportunities and select the variant with the best potential outcome (in our case the lowest global warming potential (GWP)). The evaluation of potentials comprises the systematic examination of the current variant and its parameters and the comparison of multiple variants and variant branches with each other. The potential can also be used to compare possible outcomes to certain target criteria to be achieved from requirements specifications or reference projects. The range of possible outcomes decreases throughout the progressing design process, both because more information becomes available, which reduces the uncertainty, and because more decisions will have been made. This can go in both directions. Some decisions preclude later improvement or would require going back in the design process requiring additional time and often incurring additional cost, others limit the range towards improved outcomes preventing later decisions from having a negative impact beyond a certain range. The goal of this paper is to show how potentials can be used to guide designers toward improved outcomes (Figure 1).
Figure 2. Potential (range of possible outcomes) and improvement opportunity.
4 METHODS 4.1 Research framework The main purpose of this paper is to explain the concept of potentials and demonstrate how it can be used to guide the design process towards better outcomes. To conduct this research, we have used a mixed methods approach. We are using a design episode based on an architect’s design process in the early design stages. To show how potentials can be used we have selected embodied greenhouse gas (GHG) emissions as an exemplary indicator to support more sustainable construction. This allows us to combine the architects’ “chaotic” early design process with the more analytical approach of an engineering discipline.
4.2 Analysis tools and methods LCA was used to calculate the embodied emissions using a knowledge based method which has been developed previously within the same research group (Schneider-Marin et al., unpubl.). The resulting EarlyData tool is composed of three components. First, an expert interface allows expert users to define building parts for multiple building types with a large variety of possible material combinations and thickness ranges which reflect the many possible choices in the later design process. This expert input is provided once and does not need to be adjusted, unless non-standard building parts are desired. The environmental data is provided via a database that is populated based on data from the German database Ökobaudat (Bundesministerium des Innern, für Bau und Heimat [BMI] 2020). Second, a designer user interface allows users to select among the predefined building types. Third, there is a module that extracts the areas of the various building parts from a basic BIM model reflecting the geometric detail of an early-stage design model. Combined, these modules allow designers to obtain quick feedback regarding embedded emissions without having to tediously define buildups which usually happens at a later stage. The thickness ranges and multiple material options result in uncertainty ranges which reflect the many possible choices in the later design process. This paper shows how users can be guided in the process of making these choices in an informed manner using the concept of potentials.
Figure 1. Potential, assistance and design process: Improvement opportunities at design decision points.
3.2 Improvement opportunities The difference in potential between two design variants, or two branches of the design variants tree, in overall life cycle performance results, is what we call the improvement opportunity (range) (Figure 2).
37
construction types, users can define branches for each building part category, which are structured based on the German cost groups (Kostengruppen – KG), which divide a building into a series of building part categories / groups. On the next level down, users select a layer type that represents the various layers within a building part. And finally, users get to choose materials within a given layer type. Once the decision on the choice of material has been made, the user has completed one branch of the decision tree. In theory, users can define all branches and subbranches of a construction type variant, but that would require a lot of decisions which are generally not suitable this early in the design process. Instead, the concept of potentials and improvement opportunities helps users to identify which decisions have the highest impact and should be decided early on to improve the final outcome. The selective decision tree allows users to create partially defined variants where the most relevant and impactful decisions can be made, captured, and evaluated during these early design stages.
4.3 Design guidance by potentials To communicate these potentials at various scales we have developed a graphic visualization interface, presented here for material selection. Via a selective decision tree (Figure 3) showing a whole branch of a design variant the designer receives visualized information revealing improvement opportunities, highlighting areas with greater improvement opportunities. The structure we have devised contains the following levels: – Building Level – Building Part Level – Material Level At each one of these levels, users are shown a series of selections and decisions based on potentials. While box plots are useful to show a distribution of frequency of results, we opted to show the full range from minmax since we are not presenting a statistical analysis but a range of options which are all possible choices for the designer. Figure 3 shows an excerpt of the overall structure of this selective decision tree. On a selection level users select which Building Part (level 3) or Layer Type (level 5) they want to improve, on the decision level (level 2 and 4) they decide which option they pick.
Figure 3.
5
USE CASE
To demonstrate the process of designing with the guidance of potentials via a selective decision tree (Figures 3–5) we are looking at a design path based on a real-world building project, the Building.Lab in Regensburg, Germany. It is a four-story, 2030 m2, mixed-use office / residential building, which is in the process of being built. The focus of this use case is embodied emissions (GWP) based on the LCA methodology described above. While the focus of this use case is the concept of potentials and guidance via a selective decision tree, we will also refer to the actual results to better illustrate the process. For this investigation, we have established the following three construction types: Wood, Brick, and Reinforced Concrete. Using the Building.Lab project as a case study, we go from the building level down to the material level and show the potentials at each selection and decision level. The decision tree (Figures 4–5) shows the process for one branch from the largest to the smallest scale: 1. Building level: The highest level of the selective decision tree, the “trunk”, shows the overall potential at the building level. This includes the development possibilities of all variants and variant branches for the given building geometry. In the presented selective decision tree, the results show the GWP range in t CO2 equivalent. 2. Construction type – decision level: This level presents the potentials of functionally equivalent construction types at building level. Here the user decides on the construction type based on the potential with the greatest improvement opportunity. In the case presented the user chooses ‘Wood Construction’ based on the potential and the greatest improvement opportunity.
Selective decision tree.
The ranges of the potentials show the user where the greatest improvement opportunities lie. Users operate in two different ways. On the first type of levels, the selection levels, users select which components they want to improve based on the greatest improvement opportunity. On the other type of levels, the decision levels, they decide between various functionally equivalent options. For a given variant tree they must decide on one of these options based on the highest improvement opportunity. In the given example the main variants on the building level are structured as various construction types (e.g. Brick, Reinforced concrete, Wood). Within these
38
4.
5.
6.
7.
do so at a later point to further improve the variant. Building part type (wall) – decision level:After having selected a particular building part the user now decides on one of the presented, functionally equivalent variants based on the greatest improvement opportunity. In the presented case the user chooses type ‘Wood solid 1’ out of four predefined wall types. Layer type – selection level: Within the chosen building part (e.g. wall type) users select the layer type with the highest improvement opportunity, in this case, KG 335B (wall cladding). Material – decision level: On this final level users decide on the (functionally equivalent) material for the selected layer type based on the greatest improvement opportunity. In this case ‘Softwood lumber’. This will eliminate all other options and will significantly reduce the range of potential outcomes. Feedback: By choosing a wall type and the material decision under point 6 many other (worse) possibilities are already excluded. If further branches of the decision tree were to be selected and decided, the range of potential outcomes could be further reduced and improved. In the last step, the user sees the resulting overall building range for the optimized wood building. This variant can then be compared to the initial overall potential as well as other building variants which can be created (in the presented case these are fictional placeholders). Step 7 loops back into step 1 where users can further refine this variant or create a new variant for comparison. The info box to the right shows the share of design options selected at the various selection levels and of decisions made.
6 RESULTS & DISCUSSION
Figure 4.
The presented approach both illustrates the concept of potentials and demonstrates how it can be used to guide designers through a design episode towards improved outcomes. In the given example the greatest GHG emissions improvement opportunities related to construction type, building part, and material decisions are identified. On a material level, this identifies equivalent materials with high emissions saving opportunity, whereas, on a building part level, layer combinations of materials with similar properties and their respective potentials are shown. On a building level, overall GHG emissions are revealed. Alternative branches of the selective decision tree could support embodied emissions decisions based on: – Structure types: massive vs skeleton, etc. – Façade: window to wall ratios, etc. – Energy standards – etc. Alternative evaluation criteria could be life cycle cost or operational energy.
Selective decision tree (part 1).
3. Building part – selection level: On this level the user selects a building part with a high improvement opportunity (e.g. KG 330B = exterior walls above ground load-bearing). In the given example the user could have also chosen KG 350 (ceilings/ horizontal building parts) and might
39
Figure 5. Selective decision tree (part 2).
of variants and the consequences of certain decisions in the further development of those variants. To make relevant decisions regarding LCA (embodied GHG emissions) it is not necessary to fully detail a building, but key decisions should be made in an informed manner to improve outcomes during later stages and avoid costly and tedious redesigns. The selective decision tree offers a framework and visualization method to guide users in this decision-making process and capture the decisions made at a given stage.
The concept of potentials and improvement opportunities communicates uncertainties in a transparent way. It also enables designers to make informed choices of where to focus their attention on various levels of detail and across building parts. A further important point about the concept of potentials is that it enables the comparison and evaluation of whole branches of the design with different types and degrees of definition rather than only certain variants at specific decision points with each other. The user can observe tendencies in different branches
40
energy). This will allow users to determine the most cost-effective way to realize LCA potentials. Since early-stage design decisions are considered most impactful on the overall life cycle performance of buildings, a method to identify, visualize and realize these potentials is of paramount importance to bridge the gap between early design considerations and later planning stages. The selective decision tree presented in this paper shows particular promise to optimize and actualize an environmental improvement. Further research adding multi-disciplinary knowledge and multi-criterial evaluation and representation will contribute significantly towards holistic sustainable design with limited available natural and financial resources.
The detailed representation of results in the selective decision tree allows identifying faulty outliers. In the presented case, ’Stainless steel sheet’ shows a particularly high GWP range and high values. When this is the case, it can make sense to look into the underlying data. In this case, the unlikely replacement cycles of 25 years exacerbate the negative results but the discussion of replacement cycles goes beyond the scope of this paper.
7 CONCLUSION & OUTLOOK Visualizing potentials and opportunities via a selective decision tree helps designers to locate areas with the greatest improvement opportunity, communicate them to stakeholders, and make more informed design decisions early on when the impact of these decisions is the highest. Regarding the use-case in this paper, providing insight about the potentials of a certain group of materials or a combination of building parts, supports the architect to achieve low-emission designs. Visualized results based on expert knowledge give designers reasons and arguments for communication with clients and other stakeholders. A repository of captured design episodes/ selective decision trees including past user decisions could enhance this method further. Rather than having to manually select and decide, this kind of repository of captured episodes would offer pre-defined sets of previously made decisions. Over time these episodes can be grouped into archetypes with quantified potentials for reference. As a future step, an extensive study of built reference projects and their life cycle impact should be undertaken to create benchmark values for these potentials. Similar statistical databases exist for building costs and, to a certain extent, even for operational energy, but embodied emissions are not yet captured comprehensively. Such a catalogue of reference buildings could further help designers by showing possible design choices and the final outcome. The more extensive and populated with detailed reference projects this building catalogue will be, the more accurate and effective it would make the insight about the potentials of different design elements, which leads to better supporting the architect during the important early stages of design. To reduce the ranges of the potentials given, further work needs to be invested into including additional multi-disciplinary expert knowledge regarding aspects that will affect final outcomes. Especially impactful in this regard is structural input as well as fire-related material and construction properties (Steiner et al. 2022). By choosing structural types, users can significantly reduce the structure related potential and consequently make the overall potential more precise. To support the decision-making process in an even more meaningful way in the near future, a multicriteria evaluation and representation will be developed (e.g. GWP + life cycle cost (LCC) + operational
ACKNOWLEDGEMENTS The outlined work is part of the research unit 2363 “Evaluation of building design variants in early phases on the basis of adaptive detailing strategies” funded by the German Research Foundation (DFG). The authors are grateful to the DFG for its support. We thank Lang Hugger Rampp GmbH and the Bayerischer Bauindustrieverband e.V. for letting us use the Building. Lab project as a case study. REFERENCES Abualdenien, J., & Borrmann, A. (2020). Vagueness visualization in building models across different design stages. Advanced Engineering Informatics, 45, 101107. Abualdenien, J., Schneider-Marin, P., Zahedi, A., Harter, H., Exner, H., Steiner, D., Schnellenbach-Held, M. (2020). Consistent management and evaluation of building models in the early design stages. ITcon, 25, 212–232. https://doi.org/10.36680/j.itcon.2020.013 Backer, J. W., & Lepech, M. D. (2009). Treatment of Uncertainties in Life Cycle Assessment. In 10th international conference on structural safety and reliability, Osaka, Japan. Brophy, V., & Lewis, J. O. (Eds.) (2011). A green vitruvius: Principles and practice of sustainable architectural design (2nd ed.). Washington, D.C: Earthscan. Bundesministerium des Innern, für Bau und Heimat (2020). Ökobaudat. Retrieved from www.oekobaudat.deCALAA GmbH (2022). CAALA [Computer software]. Munich. Retrieved from https://www.caala.de/ Guimarães, G. D., Saade, M. R. M., Zara, O. O. C., & Silva, V. G. (2019). Scenario uncertainties assessment within whole building LCA. In IOP Conference Series: Materials and Engineering Science (323). Symposium conducted at the meeting of SBE. Harter, H., Singh, M. M., Schneider-Marin, P., Lang, W., & Geyer, P. (2019). Uncertainty Analysis of Life Cycle Energy Assessment in Early Stages of Design. Energy and Buildings, 109635. https://doi.org/10.1016/j.enbuild. 2019.109635 Hollberg, A. (2017). A parametric method for building design optimization based on Life Cycle Assessment. BauhausUniversität Weimar. https://doi.org/10.25643/BAUHAUSUNIVERSITAET.3800
41
Jaskula, K., Zahedi, A. [A.], & Petzold, F. [F.] (2021). Archi-guide. Architect-friendly visualization assistance tool to compare and evaluate BIM-based design variants in early design phases using template-based methodology. In ECPPM 2021-eWork and eBusiness in Architecture, Engineering and Construction (pp. 153–162). CRC Press. Lipsey, R. G. (1975). An introduction to positive economics. 4th ed. London: Weidenfeld and Nicolson. Meng, Z., Zahedi, A., & Petzold, F. (2020). Web-Based Communication Platform for Decision Making in Early Design Phases. In Proceedings of the 37th International Symposium on Automation and Robotics in Construction (ISARC). One Click LCA Ltd. (2022). OneClick LCA [Computer software]. Retrieved from https://www.oneclicklca.com/ Østergård, T., Jensen, R. L., & Maagaard, S. E. (2017). Early Building Design: Informed decision-making by exploring multidimensional design space using sensitivity analysis. Energy and Buildings, 142, 8–22. https://doi.org/10.1016/j.enbuild.2017.02.059 Röck, M., Hollberg, A., Habert, G., & Passer, A. (2018). LCA and BIM: Integrated Assessment and Visualization of Building Elements’ Embodied Impacts for Design Guidance in Early Stages. Procedia CIRP, 69, 218–223. https://doi.org/10.1016/j.procir.2017.11.087 Schneider-Marin, P., Harter, H., Tkachuk, K., & Lang, W. (2020). Uncertainty Analysis of Embedded Energy and Greenhouse Gas Emissions Using BIM in Early Design. Sustainability, 7(12).
Schneider-Marin, P., Stocker, T., Abele, O., Margesin, M., Staudt, J., Abualdenien, J., & Lang, W. (submitted for publication). EarlyData knowledge base for material decisions in building design. Advanced Engineering Informatics. Steiner, D., Staudt, J., Margesin, M., Zong Chujun, Schnellenbach-Held, M., & Lang, W. (accepted paper). Multidisciplinary building design support in early planning phases using knowledge-based methods and domain knowledge. ECPPM. Zahedi, A., Abualdenien, J., Petzold, F., & Borrmann, A. (2022) BIM-based design decisions documentation using design episodes, explanation tags, and constraints, ITcon Vol. 27, pg. 756-780, https://doi.org/10.36680/j.itcon. 2022.037. Zahedi, A., & Petzold, F. Revit add-in for documenting design decisions and rationale: A BIM-based tool to capture tacit design knowledge and support its reuse. In CAADRIA 2022 9-15 April 2022. Retrieved from https://caadria2022.org/wp-content/uploads/2022/04/761.pdf Zahedi, A., & Petzold, F. (2019a). Adaptive Minimized Communication Protocol based on BIM. In 2019 European Conference on Computing in Construction. Zahedi, A., & Petzold, F. (Eds.) (2019b). Interaction with analysis and simulation methods via minimized computerreadable BIM-based communication protocol. Retrieved from http://papers.cumincad.org/data/works/att/ecaadesig radi2019_140.pdf
42
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Structure and LCA-driven building design support in early phases using knowledge-based methods and domain knowledge D. Steiner & M. Schnellenbach-Held Institute for Structural Concrete, University of Duisburg-Essen, Essen, Germany
J. Staudt, M. Margesin, C. Zong & W. Lang Institute of Energy Efficient and Sustainable Design and Building, Technical University of Munich, Munich, Germany
ABSTRACT: Early planning decisions highly influence the resource-efficiency and environmental impact of a building design. Through the inclusion of design knowledge regarding the structural design and embedded emissions, the design process can be effectively supported. For the early provision of the multidisciplinary knowledge and a related design assistance, the application of knowledge-based methods is proposed. Based on this approach, an expertise regarding the structural design and material-related embedded emissions is provided in the form of easy-to-understand design information. An achievable design support is demonstrated with the aid of a case study on a real-life building design. In this context, the early inclusion of knowledge enables a promising multidisciplinary design assistance regarding the structural and environmental performance. Through extension of the involved knowledge, the application range of the design support can be enlarged. For this purpose, the application and inclusion of further knowledge sources are addressed in future research.
1 INTRODUCTION As the building industry is one of the largest consumers of material and energy, it is a focus area for a reduction of global greenhouse gas (GHG) emissions and the related environmental impact. For building designs, such demands are analyzed in sustainability assessments that usually involve life cycle analysis (LCA). Due to the ongoing trend of using renewable energy and nearly zero energy buildings, the embodied energy gains an increasing percentage of GHG emissions. This building performance is significantly influenced by assumptions and design decisions in the conceptual and preliminary design phases. To minimize the demand of embedded energy and GHG emissions, performance evaluations are needed early in the design process. Since many design decisions are still to be made in these early phases, uncertainty arises from the lack of information. An approach to cover this vagueness in the design process is the use of sensitivity analyses. They enable the quantification of decision impacts, the visualization of uncertainties and the identification of most influential design parameters (Schneider-Marin et al. 2020). Structural design decisions show a significant impact on the embedded emissions of a building, as the material used and the type of the load-bearing superstructure highly influence the required material quantities. For this reason, structural engineering expertise should be involved early in the design process to DOI 10.1201/9781003354222-6
increase the sustainability of building designs. For this purpose, a variety of feasible structural concepts can be incorporated to cover different materials and construction types.Additionally, a preliminary dimensioning of the included structural members allows an assessment of required material quantities. This in turn enables the quantification of the embedded GHG emissions for the structural variants and the impact on the global warming potential (GWP). The resulting comparability of structures, materials and construction types represents a highly valuable support for design decisions in early phases (Schneider-Marin et al. 2020; Zhang et al. 2018). For the early provision of the required design information, the application of knowledge-based methods is a promising approach (Wang et al. 2021). They enable the inclusion of common recommendations in early design phases that are usually based on rough calculations and on the engineering experience of qualified planners. Through the integration of easily understandable design knowledge, a determination of comprehensible design information (Zhang et al. 2018) and a computer-aided collaborative design process (Ungureanu & Hartmann 2017) are facilitated. To support influential design decisions in early phases, the provision of domain knowledge by a knowledge-based approach is proposed. Included assessments on the structure and embedded emissions allow the quantification of the GWP. Resulting information enables the analysis of important design
43
To facilitate an interactive planning, B. Steiner developed an integrated structural-architectural design approach (Steiner et al. 2017). It comprises a structural layout analysis using simplified computational models to determine structural stability and stabilization proposals. In the process, a structural grid is derived from architectural constraints to determine an initial structural member arrangement. Subsequently, the grid is extended until all members are stable. In an optimization process, members are then reduced to achieve optimized grids and superstructures. For the design of performant structural topologies from spatial designs, Boonstra developed an approach that is based on spatial zoning and structural design grammars (Boonstra et al. 2020). It includes a space division to high resolution parts and the recombination to conformal zones. Subsequently, design grammars are applied to determine structural member arrangements and stabilization around the zones and thus to generate the structural topologies.
parameters and the comparison of different structure types.
2 KNOWLEDGE-BASED DESIGN The application of knowledge for design tasks is often referred to as knowledge-based engineering that can be seen as a fusion of artificial intelligence (AI), computer-aided design (CAD) and object-oriented programming (OOP) (La Rocca 2012). In the field of AI, knowledge is information about designs, design procedures and relationships, that is processed with methods like knowledge-based systems (KBS), casebased reasoning (CBR) or machine learning (ML) (Burggräf et al. 2020). This enables a guidance for the design process through a support of design decisions even for complex systems and considering uncertainty (Wang et al. 2021). Ontology-based approaches facilitate multidisciplinary design procedures through capturing and applying real-life communication in natural language (Ungureanu & Hartmann 2017; Ungureanu 2021; Zhang et al. 2018). Constructability and buildability assessments aim for an efficient integration of building design knowledge to achieve multidisciplinary project objectives, whereby an early incorporation through knowledge-based methods is recommended (Fadoul et al. 2020).
2.2 Embedded emission prediction
Material-related embedded emissions are usually determined at a later stage of the design process, after major decisions on the design of a building have already been made and can only be altered with significant redesign effort. At the same time these decisions have a significant impact on the environmental performance of a building. Consequentially it 2.1 Structural engineering is of paramount importance to provide designers with information regarding embedded emissions during the Maher developed an early approach for the inclusion early phases of a design when these decisions with of experience, rules of thumb, intuition and experthe highest impact are taken. LCA during these early tise (Maher 1987). Using knowledge-based expert phases is subject to multiple uncertainties stemming systems, such expert knowledge is applied for strucfrom various sources, including data, calculation and tural preliminary design. For this purpose, four expert design uncertainties (J.W. Backer et al. 2009). Cursystems deliver designs and alternatives for building rently available early stage LCA tools such as CAALA planners. They provide a support for locating strucand One Click LCA®do not sufficiently capture these tural members, arranging members in a superstructure, uncertainties. Yet communicating these uncertainties locating the lateral stiffening system and finding strucis important to transparently show designers the qualtural solutions. In the process, architectural and spatial ity of the information they are using as the basis for requirements are considered as constraints. their decisions (P. Schneider-Marin et al. 2020). At the For early structural design purposes, a knowledgesame time, it is important to reduce these uncertainties based system featuring fuzzy knowledge bases was developed by Schnellenbach-Held et al. (Schnellenbach- to better reflect the impact of various design options. Since the structure has a high impact, reducing the Held et al. 2006). It includes a fusion of structural uncertainty of the structural design early on will have calculation knowledge and imprecise assumptions to a significant impact on the overall uncertainty. This support a cooperative and comprehensive design. The allows designers to make more informed choices. The system provides structural design recommendations, a aim of this research is to both appropriately represent preliminary specification of structural member dimenthe range of possible outcomes as well as to reduce sions and an option to perform sensitivity analyses for this range by including captured expert knowledge into design modifications. the design process. A knowledge-based method has Geyer developed a grammar-based approach to been developed in a previous stage of this research support multidisciplinary design optimization (Geyer (P. Schneider-Marin et al. unpubl.) which enables the 2008). It comprises rules for a decomposition of the evaluation of embedded emissions (GWP) at an early spatial design to structural member arrangements, stage when detailed information is not yet available. member refinements and design modifications. The In this paper a method is presented, which integrates resulting design tree enables a design from shape to structural knowledge to reduce the range of potential structure and is applied for an optimization of hall outcomes. structures.
44
structural variant. This allows the quantification of embedded emissions and the subsequent determination of design potentials. Thus, designers are informed at an early stage about the consequences of deciding for a certain structural archetype. Implementations of the knowledge-based system will facilitate the design process as well as a rapid prototyping of building models. For this purpose, knowledge is stored in databases – also known as knowledge bases – that contain rules for the design development. Using methods of object-oriented programming (OOP), model entities – like structural members and nodes – feature knowledge-access methods as well as get- and set-methods for interactions. Based on a predefined grid, structural nodes are inserted that represent connections between members. Based on the application of knowledge, they indicate possible member placements for archetypes, allow a determination of conditions for member specification and enable load distribution assessments.
2.3 Presented approach The approach presented in this contribution aims for the inclusion of transparent and easy to understand domain knowledge for an efficient multidisciplinary support of early building design phases. Using knowledge-based methods, conceptual and preliminary building designs are suggested and evaluated. In the process, the spatial design is considered as architectural constraints, for which different structural designs are proposed, and related embedded emissions are estimated using domain knowledge. Uncertainty with regard to design information and design decisions is taken into account and expressed as value ranges. These potentials (cf. Staudt et al. 2022) allow the quantification and comparison of embedded emissions of structural variants to support planning decisions during early design phases.
3 DESIGN SUPPORT CONCEPT & METHODS 3.2 Embedded emissions
To achieve the support of the conceptual and preliminary design phases, domain knowledge is provided early on through knowledge-based methods. They comprise a knowledge-based system for structural design suggestions and an expert graphical user interface (GUI) for embedded emission assessments which are explained in the following. As main advantages of the proposed knowledge-based approach, expensive design loops are avoided and planning decisions are guided towards efficient and sustainable designs, minimizing planning time and costs.
Based on a knowledge-based method developed previously (P. Schneider-Marin et al. unpubl.) embedded emissions (GWP) are evaluated at an early stage, when detailed information is not yet available. Since embedded emission information is only available on a material level, and accurate material quantities are not available during early phases, the method developed seeks to bridge that gap. This method uses restructured information based on the German material database Ökobaudat (2022) and incorporates construction knowledge via expert input.This initial rough expert input includes possible layer combinations, layer thickness ranges, and material choices to define a spectrum of possible building parts for various construction types. The designation of these building parts is based on the German cost groups (Kostengruppen/ KGs) (DIN 276 2018-12). The end-user is presented a range of possible embedded emission outcomes, which reflects the uncertainty at these early design stages. While the tool takes insulation thicknesses into consideration based on required U-values for a selected energy standard it does not yet include specific expert knowledge regarding fire protection and structure. Since the structure makes up for a significant portion of the embedded emissions of a building (SchneiderMarin 2020) and structural designs can vary greatly, the range of possible outcomes is influenced significantly by the uncertainty caused by a lack of structural information. This paper presents a method how structural knowledge can be used to reduce this range and generate more accurate embedded emission values for various options. For the initial rough approximation, the tool uses so called “two component layers” (P. Schneider-Marin et al. unpubl.) to capture layers which consist of multiple materials in the same plane. The same approach is taken for a concrete skeleton structure (Figure 1) which is presented in the following case study. This is based
3.1 Structural design For structural design suggestions, a rule-based knowledge-based system is applied. The included knowledge provides information about structural design concepts, the preliminary specification of structural member dimensions, the load distribution through the superstructure and rough reinforcement degrees of reinforced concrete (RC) members. Based on this knowledge, different structural designs are developed to specify possible conditions for embedded emission assessments. At the beginning of the process, the spatial design and usage as well as the structural grid are prespecified. For these constraints, different structural variants are suggested, following archetypes that represent structural concepts. They involve knowledge about the kind and arrangement of structural members, and can be referred to as design grammars. For the arranged members, missing preliminary dimensions are then assessed based on common rules of thumb involving uncertain design information. This is combined with a simplified evaluation of the load distribution through related knowledge. Based on the preliminary structural member specifications and knowledge about estimated rough reinforcement degrees, necessary total amounts of material are summed up for each
45
on the assumptions that the structure is incorporated in the loadbearing exterior walls (KG331) and/or interior walls (KG341). Since reinforced concrete itself consists of two components (concrete and steel), a two component material within a two component layer is applied. The material data for these materials is based on Ökobaudat data and the ranges within the two component material are based on expert input. Ultimately, this results in a multicomponent layer with uncertainty both in the makeup of the reinforced concrete and the structure to insulation ratios.
insulation, is determined from statistical data based on a limited number of reference buildings. A more extensive analysis will be necessary to obtain statistically reliable numbers. These ratios include both vertical columns as well as solid walls for lateral stiffening as well as concrete cores. Since the exact location of these walls and cores is often not determined yet during early design phases, the ratios capture both scenarios with internal cores and with cores located at the facade for both loadbearing exterior walls (KG331) and/or interior walls (KG341), resulting in significant uncertainty (see also Figure 2). Solid structural building parts, e.g. structural slabs are assigned thickness ranges based on expert knowledge. 3.3 Multidisciplinary design support In the presented approach, the architectural design presents constraints for the proposed structural design specifications. The estimation of embedded emissions then allows the determination of GWP design potentials for the suggested structures. Structural archetypes represent different types for the proposed superstructure that is required for structural member specifications. This allows the application of rules of thumb to assess missing member dimensions. Subsequently, the preliminary member designs enhance the estimation of embedded emissions considering different building
Figure 1. Buildings parts with uncertainty ranges (LCA methodology): Exterior wall buildup with multi-component layer type for concrete skeleton (yellow/orange, KG331 B); concrete ceiling (KG351).
For these skeleton structures a percentage range of each of the two components, reinforced concrete and
Figure 2. Merging process for skeleton and solid construction (left) LCA Methodology: Exterior walls (yellow, KG331), interior walls (green, KG341), ceilings (grey, KG351); Structural Input (right): columns, exterior walls (red), interior walls (blue), ceilings (purple); Merged Model (middle): Structural input (columns (red, blue), walls (red, blue), ceilings (purple) replace material quantities based on early LCA estimates (yellow/orange, green).
46
members of in-situ normal concrete (NC) are considered. For these early design constraints, different variants for structural concepts are proposed, preliminarily designed and embedded emissions are assessed using the knowledge-based methods and domain.
part constructions. With regard to the archetypes, the quantified GWPs reveal the influence of the structural design to support design decisions. 3.4 Merged model The structural analysis input shown in this paper (Chapter 3.1) allows to replace the rough estimations with much more accurate material quantities and resulting embedded emissions which will decrease uncertainties significantly, resulting in a significantly smaller range of potential GWP (see Figure 2). Since the LCA methodology presented in chapter 3.2 is based on building part areas in m2 , a conversion of the structural input given in t and m3 to areas in m2 is needed. For the vertical elements (columns and walls) this provides the percentage range for reinforced concrete. The remaining percentage is filled with insulation based on values given by the LCA methodology. For the horizontal structural building parts, i.e. the structural slabs, thickness ranges and reinforcement ranges from the LCA methodology are substituted by new ranges based on the structural input.The structural input here shows more variation (and improvement opportunity) based on multiple structural options for these elements. The remaining building parts will be obtained from the estimate ranges based on the LCA methodology. This includes the additional insulation which is necessary to cover the concrete skeleton to avoid thermal bridges. Ultimately, users will be provided with embedded emission feedback showing lifecycle potential and improvement opportunities (Staudt et al. 2022) for the overall building and by building part/ cost group.
4.1 Structural design knowledge For the different structural design concepts, archetypes are formalized that involve the specification of structural members and their arrangement in the grid. According to structural engineering knowledge, the required slab thickness depends on the slab type that is correlated to the slab supports. For these reasons, the archetype specification is based on common slab types that also allow an application of related rules of thumb. Thus, single-span slabs (uniaxial linesupported), dual-span slabs (biaxial line-supported) and flat slabs (biaxial point-supported) are considered for the case study and demonstration purposes. Furthermore, columns are characteristic vertical members in skeleton structures, and are typically placed at crossing axes of the grid. They feature point-supports of flat slabs or the support of downstand beams, that represent line-supports of slabs along axes of the grid. Based on this knowledge, the archetypes are specified (see Figure 4) and applied to the constraints.
4 CASE STUDY An application of the knowledge-based methods and included domain knowledge in early planning phases is demonstrated with an example based on a real-life building design – the Building.Lab in Regensburg by the architects Lang, Hugger, Rampp. The architectural design shows a four-story, 2030 m2 , mixed-use building with a U-shaped plan including a teaching area, a residential area and a hallway area (see Figure 3). Figure 4. Considered structural archetypes.
Load specifications are based on knowledge from experience and Eurocode 1 (DIN EN 1991-1-1/DE). In early design phases, the consideration of vertical loads is sufficient for the preliminary determination of structural member dimensions. Following common structural design experience, the staircases are also considered as lateral stiffening cores. As they usually consist of RC walls, related wall members are placed. The use of a raft foundation is also prespecified. After the structural member arrangement, further missing member dimensions have to be specified for the estimation of material quantities. For this purpose, common rules of thumb are applied that also involve
Figure 3. Early architectural design of the case study.
Regarding this case study and for demonstration purposes, RC skeleton structures with massive RC
47
structural variants, the achievable smaller variations in the total ranges represent design potentials. Additionally, the total ranges are the design potentials of the more general RC skeleton type. The early provision of such information about design consequences represents a decision support encouraging designs with a potentially lower GWP impact.
an uncertainty due to a lack of information like the exact concrete class. Being in accordance with typical design requests, this preliminary design of members relates to the slab thicknesses, the downstand beam heights, and the column cross-sectional areas. As the reinforcement of RC members shows a high impact on embedded emissions, reinforcement degrees are assessed. For this purpose, rough estimations (Table 1) are examples that can be seen as average structural expert assessments. Table 1.
Table 3.
Cost group
Estimated rough reinforcement degrees.
Structural member
Reinforcement degree
Flat slab Single-span slab Dual-span slab Column Downstand beam Wall (core) Foundation slab
180 kg/m3 130 kg/m3 130 kg/m3 250 kg/m3 250 kg/m3 150 kg/m3 180 kg/m3
Estimated material quantities. Flat Single “y” Single “x” Dual-span [Vc in m3 ] [Vc in m3 ] [Vc in m3 ] [Vc in m3 ] [ms in to] [ms in to] [ms in to] [ms in to]
KG 320
[283-283] [51-51] KG 330B [78-82] [13-14] KG 340A [106-113] [17-19] KG 350 [355-452] [64-81] KG 360A [131-164] [24-30] Total [953-1094] [168-194]
[283-283] [51-51] [78-82] [13-14] [106-112] [17-18] [390-418] [54-60] [123-145] [17-21] [980-1039] [152-163]
[283-283] [51-51] [78-80] [13-14] [106-111] [17-18] [295-320] [42-48] [111-120] [16-18] [872-914] [139-149]
[283-283] [51-51] [78-81] [13-14] [106-112] [17-18] [331-411] [50-66] [119-151] [18-24] [918-1039] [149-173]
4.2 Embedded emission knowledge Based on the LCA methodology described in Chapter 3.2, estimated ranges for the whole building and the building parts (KGs) are shown in.
To get a significantly better understanding of the embedded emission impact of the overall building, the rough estimates are merged that are based on the LCA methodology with the structural input based on the method described in Chapter 3.4. Table 3 shows the material quantities for the structural building parts resulting from the structural input structured by cost groups (KGs) to match the structure of the LCA data. Merging the structural input (Table 3) with the LCA values (Table 4) reveals, that the ranges for the load bearing elements have decreased significantly (Figure 5). The reduction at whole building level (left) appears small because the uncertainties of the remaining building parts remain and will need to be reduced in future steps. Looking at the structural layers a significant reduction becomes apparent (right).
Table 2. Estimated embedded emissions of overall building w/o structure input.
Cost group
Building parts
KG 320 KG 330B KG 340A KG 340B KG 350 KG 360A KG 334 Total
Foundation Exterior walls, columns Interior walls, columns Interior walls (non-bearing) Ceilings, downstand beams Roof slab, downstand beams Windows
Whole Building [GWP in t CO2 -eq] [32-448] [14-367] [5-170] [1-275] [74-855] [21-441] [43-414] [190-2971]
Table 4. Estimated embedded emissions of overall building merged with input from structural methodology.
These results show the high impact of the structure on the overall embedded emission results as well as a wide range (potential) of GWP results for the building parts, which include loadbearing components when no specific structural knowledge is included. 4.3 Results of the multidisciplinary support The application of domain knowledge through knowledge-based methods results in the determination of material quantities (see Table 3) and GWPs (see Table 4) for the archetypes. Regarding the underlying
48
Cost group
Flat
Single “y” Single “x” Dual-span [GWP in t CO2 -eq]
KG 320 KG 330B KG 340A KG 340B KG 350 KG 360A KG 334 Total
[87-408] [39-367] [34-159] [1-275] [114-798] [34-416] [43-414] [353-2838]
[87-408] [39-367] [34-159] [1-275] [114-773] [29-404] [43-414] [347-2800]
[87-408] [39-365] [34-157] [1-275] [88-737] [25-396] [43-414] [318-2752]
[87-408] [39-336] [34-159] [1-275] [101-776] [28-408] [43-414] [333-2806]
simplicity and generalization of the domain knowledge in contrast to complexity and individuality of building designs. For this purpose, the application of AI methods seems to be promising. A proposed approach comprises a knowledge-based system with fuzzy knowledge bases (Schnellenbach-Held & Steiner 2021). Knowledge-supported numerical studies allow for a fast acquisition of structural assessments featuring more design context and a higher precision than rules of thumb. Additionally, design aspects are covered that are predefined in the presented case study. Furthermore, a structural rating is included that is based on an informed possibility. For the integration of varying lateral stiffening and foundation designs as well as other difficult structural aspects, the use of expert assessments through a linguistic fuzzy approach is proposed to meet highly complex structural challenges. Future research will also cover the connection to further planning disciplines and design support methods. Examples are a graph-based method for the integration of case-based reasoning in the design process (Napps et al. 2021), an approach for operational energy assessments using trained artificial neural networks (Chen et al. 2021) as well as lifecycle cost estimates (LCC) and disassembly estimations for the inclusion of possible reuses of structures in LCA.
Figure 5. GWP range without structural input (black) and with structural input (colored).
5 DISCUSSION The combination of knowledge from different domains supports multidisciplinary design assessments. In the process, the provision of domain knowledge is a key aspect to complement missing design information. The achieved multidisciplinary knowledge-based determination of GWP ranges allows the comparison of different structure types. Thus, designers are informed early in the design process about consequences of their decisions. This represents a valuable design support that is achieved through the application of knowledge-based methods. In the continuing design process, the design development comprises geometry harmonization, e.g. alignments of slab thicknesses or column crosssections. Included decisions result in more precise design specifications and potential evaluations, but also require additional design information, like the concrete class that has to be integrated in the knowledge bases and the evaluation procedure. Additional structural types can be included through archetypes for conforming structural concepts and corresponding knowledge for the preliminary member dimensioning. Furthermore, the design of lateral stiffening cores and the raft foundation are considered as constraints in this case study. As these design aspects represent highly complex structural challenges, an inclusion of related design variations demands further developments to complement the archetypes.
7 SUMMARY Decisions on the structural design highly influence required material quantities and embedded emissions of buildings. In early design phases, information about such consequences represent a decision support in the design process. For this purpose, the use of knowledge-based methods is proposed to provide domain knowledge for a design assistance. It comprises early assessments on the structure and embedded emissions to enable determinations of required material quantities and GWP. These performance indicators are formalized as value ranges that represent potentials of structural design variants. To include structural variations, archetypes represent different structural concepts that are applied to architectural constraints. Resulting arrangements and connections of structural members allow the application of domain knowledge to quantify the design potentials. Thus, the impact of design variations on the GWP is assessed to support design decisions in early phases. The procedure is demonstrated in a case study of a real-life early building design. Results underline the influence of the structural concept and show the design support through knowledge-based assessments of decision consequences. Further structural concepts can be included through related archetypes and domain knowledge.To complement complex structural aspects and design context, supplementary knowledge and AI methods can be applied.
6 OUTLOOK Next to the structural in-situ RC skeleton variants considered in the case study, further types of structures will be taken into account. They comprise other structural materials (such as wood and steel) and construction types. Related member arrangements are included through corresponding archetypes. This also necessitates the incorporation of additional domain knowledge, further design aspects and dependencies. Here, a major challenge arises between
49
ACKNOWLEDGEMENTS
simulation-based complex engineering design. Structural and multidisciplinary optimization 57: 393–416. https://doi.org/10.1007/s00158-017-1739-8 Maher, M. L. (1987). Expert systems for structural design. Journal of computing in civil engineering 1(4): 1-9. http:// dx.doi.org/10.1061/(ASCE)0887-3801(1987)1:4(270) Napps, D. & Pawlowski, D. & König, M. (2021). BIM-based Variant Retrieval of Building Designs Using Case-based Reasoning and Pattern Matching. In: Proceedings of the 38th International Symposium on Automation and Robotics in Construction, 435. https://doi.org/10.22260/ isarc2021/0060 One Click LCA®software. (2022). Automate Building Life Cycle Assessment with One Click LCA. Available online: https://www.oneclicklca.com/construction/lifecycle-assessment-software/ [Accessed 22 March 2022] Schneider-Marin, P. & Harter, H. & Tkachuk, K. & Lang, W. (2020). Uncertainty Analysis of Embedded Energy and Greenhouse Gas Emissions Using BIM in Early Design Stages. Sustainability 2020, 12(7), 2633. https://doi.org/ 10.3390/su12072633 Schneider-Marin, P. & Stocker, T. & Abele, O. & Staudt, J. & Margesin, M. & Lang, W. (unpubl.) Re-structuring of ecodata to include functional criteria: using data enhanced with knowledge in early phases of design. Advanced Engineering Informatics. Submitted paper. Schnellenbach-Held, M. & Hartmann, M. & Pullmann, T. (2006). Knowledge based modeling in networked cooperative building design using elements of fuzzy logic. In: Proceedings of the XIth International Conference on Computing in Civil and Building Engineering, Montreal 2006. Schnellenbach-Held, M. & Steiner, D. (2021). Application of AI methods for the integration of structural engineering knowledge in early planning phases. In: Proceedings of the 28th EG-ICE international workshop on intelligent computing in engineering, July 2021, Berlin, Germany. Staudt, J. & Margesin, M. & Zong, C. & Lang, W. & Zahedi, A. & Petzold, F. & Schneider-Marin, P. (2022). Life cycle potentials and opportunities of design variants for early-stage design assistance. In: European conference on product and process modeling 2022. Accepted paper. Steiner, B. & Mousavian, E. & Mehdizadeh Saradj, F. & Wimmer, M. & Musialski, P. (2017). Integrated structuralarchitectural design for interactive planning. Computer graphics forum 36(8): 80–94. https://doi.org/10.1111/cgf. 12996 Ungureanu, L.-C. & Hartmann, T. (2017). Natural language controlled parametric design. In: Proceedings of the 24th EG-ICE international workshop, Nottingham, UK. Ungureanu, L.-C. (2021). A Design Recommender System: A Rule-based Approach to Exploit Natural Language Imprecision using Belief and FuzzyTheories. In: Proceedings of the 28th EG-ICE international workshop, Berlin, Germany. Wang, R. & Milisavljevic-Syed, J. & Guo, L. & Huang, Y. & Wang, G. (2021). Knowledge-Based Design Guidance System for Cloud-Based Decision Support in the Design of Complex Engineered Systems. Journal of mechanical design 143(7): 072001-1. https://doi.org/10.1115/1.405 0247 Zhang, J. & Li, H. & Zhao,Y. & Ren, G. (2018). An ontologybased approach supporting holistic structural design with the consideration of safety, environmental impact and cost. Advances in Engineering Software 115: 26–39. http://dx.doi.org/10.1016/j.advengsoft.2017.08.010
The outlined work is part of the research unit 2363 “Evaluation of building design variants in early phases on the basis of adaptive detailing strategies” funded by the German Research Foundation (DFG). The authors are grateful to the DFG for its support. We thank Lang Hugger Rampp GmbH and the Bayerischer Bauindustrieverband e.V. for letting us use the building.lab project as a case study. REFERENCES Abualdenien, J. & Borrmann, A. (2019). A metamodel approach for formal specification and consistent management of multi-LOD building models. Advanced Engineering Informatics 40: 135–153. https://doi.org/10.1016/j.aei.2019.04.003 Baker, J.W. & Lepech, M.D. (2009). Treatment of Uncertainties in Life Cycle Assessment. In: 10th international conference on structural safety and reliability, Osaka, Japan. Boonstra, S. & van der Blom, K & Hofmeyer, H & Emmerich, M. T.M. (2020). Conceptual structural system layouts via design response grammars and evolutionary algorithms. Automation in construction 116: 103009. https://doi.org/10.1016/j.autcon.2019.103009 Burggräf, P. & Wagner, J. & Weißer, T. (2020). Knowledgebased problem solving in physical product development – A methodological review. Expert systems with application: X (5): 100025. https://doi.org/10.1016/j.eswax. 2020.100025 CAALA. (2019). The CAALA software is a comprehensive tool for energetic pre-dimensioning and life cycle assessment. Available online: https://caala.de/features [Accessed 22 March 2022] Chen, X. & Singh, M.M. & Geyer, P. (2021). Componentbased machine learning for predicting representative time-series of energy performance in building design. In: Proceedings of the 28th EG-ICE international workshop on intelligent computing in engineering, July 2021, Berlin, Germany. DIN 276: 2018-12, Kosten im Bauwesen. Berlin. Beuth Verlag GmbH. https://dx.doi.org/10.31030/2873248 Fadoul A. & Tizani W. & Osorio-Sandoval C.A. (2021): A Knowledge-Based Model for Constructability Assessment of Buildings Design Using BIM. Proceedings of the 18th International Conference on Computing in Civil and Building Engineering. ICCCBE 2020. Springer, Cham. https://doi.org/10.1007/978-3-030-51295-8_13 Federal Ministry for Housing, Urban Development and Building. (2022). Ökobaudat. Available online: https:// www.oekobaudat.de/en.html [Accessed 2 February 2022] Geyer, P. (2007). Multidisciplinary grammars supporting design optimization of buildings. Research in engineering design 18: 197-216. https://doi.org/10.1007/s00163-0070038-6 La Rocca, G. (2012). Knowledge based engineering: Between AI and CAD. Review of a language-based technology to support engineering design. Advanced Engineering Informatics 26: 159-179. https://doi.org/10.1016/ j.aei.2012.02.002 Liu, H. & Ong, Y.S. & Cai, J. (2018). A survey of adaptive sampling for global metamodeling in support of
50
Processes
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Challenges and experiences with the reuse of products in building design A. Tomczak, M. Łuczkowski & E. Hjelseth Norwegian University of Science and Technology, Trondheim, Norway
S.F. Sujan Mott MacDonald, Sheffield, UK Norwegian University of Science and Technology, Trondheim, Norway
ABSTRACT: Over a third of greenhouse gases emitted in Europe are related to buildings, and a third of that is the upfront embodied emissions from construction and manufacturing building materials. One uncommon way of limiting those emissions is to reuse already preowned building components before downcycling them. The study comprises twelve interviews with mainly Norwegian industry representatives to assess their attitude towards titular reuse. In the paper, we describe how consecutive design with preowned elements differs from initial design and what challenges it entails. Among the interviewee’s reflections are the importance of design time and convenience, relative costs of consecutive design, standardisation, lack of long-term thinking among clients and designers, and challenges with managing information about potentially reusable elements. Presented research adds to the knowledge on reuse barriers and opportunities. It can be helpful for understanding designers’ needs and seeking solutions to popularise material reuse.
1 INTRODUCTION The challenges of anthropocentric climate change require humanity to reduce its negative influence on the environment. The EU Taxonomy lists the transition to a circular economy (CE) as a way to pursue this goal and one of the key environmental objectives (European Commission 2020). However, the concept of CE is not yet widely applied in the construction industry, which significantly impacts greenhouse gas emissions. Building materials’ embodied emissions account for about 11% of global emissions (IEA & UNEP 2018). Çetin et al. (2021), based on the previous work of Bocken et al. (2016), have grouped multiple circular practices into four categories: ‘regenerate’, ‘narrow’, ‘slow’, and ‘close’ resource loops. In this work, we focus on four circular strategies that belong to the ‘slow’ and ‘close’ categories. The first two are about the initial design of products: the design for disassembly (DfD), sometimes also called design for decommissioning or reversibility, and the design for a long life of products. The third strategy is to reuse whole building components without disrupting their integrity. We distinguish reuse from recycling materials that constitute a product, its downcycling into a product of lower quality, or upcycling – of higher (Iacovidou & Purnell 2016). The last strategy – urban mining – is about closing the resource loop by bringing resources back to the economic cycle that would otherwise be lost. The difference between reuse and urban mining is minor and depends on whether the products were prepared for multiple uses at the production stage or not. DOI 10.1201/9781003354222-7
Apart from a few pilot projects, the industry has not yet implemented the concept. Multiple studies on the barriers to the reuse and DfD of building components were found in the literature. Sigrid Nordby et al. have identified barriers to reuse in the Norwegian context, such as lack of economic driving forces and underdeveloped market, complexity and longer design and construction time, documentation uncertainty, presence of hazardous substances in products, and need for compliance with regulations (Sigrid Nordby et al. 2019). The study by Arup and Ellen MacArthur Foundation adds to that list the fragmented nature of the industry, insufficient knowledge dissemination, and unsupportive regulations (Acharya et al. 2018). Others also list the need for education and a proper typology system (Iacovidou & Purnell 2016), technical problems, such as unreliable quality and damages suffered while demolishing (Cruz-Rios & Grau 2020), as well as lack of strength grading rules for reused elements and their difficult and expensive deconstruction (Hradil et al. 2014). This paper is the first part of the ongoing Design Science Research (DSR), which aims to investigate the needs, propose and design a solution that improves the design process with preowned elements and DfD. In this part, we intend to assess the experiences and attitudes towards DfD and reuse among decision-makers in the European context. 2 METHODOLOGY The initial steps of DSR are identification and raising awareness of the problem to be later addressed
53
Brütting 2020): (1) designing optimised, materialefficient systems, (2) employing low-carbon materials, (3) ensuring long-term usefulness of elements. Interviewees were asked to reflect on those strategies to see how they approach the third, spotlighted in this paper. Then, the conversations were guided towards the circularity theme to grasp the interviewee’s understanding and experiences with the concept. We asked about their definition of CE; how they approach the end-of-life scenarios on projects; how much their designs allow for future reuse; whether they are aware of the standards, such as ISO20887; how they are storing information that might be useful for reuse. The second slide was a diagram presenting the overlap of typical linear workflow and the circular workflow developed based on the literature review. Figure 1 shows the diagram improved with feedback from the interviewees. The participants were shown the slide and asked what challenges they experienced and foresaw with such workflow or what modifications they would suggest. A hypothetical scenario was also described where the stock of preowned elements is readily available, and we asked participants about possible objections to using such building blocks. The results are grouped into eight selected topics that were repeated throughout the interviews and are important in the context of the paper.
by the designed artefact (Dresch et al. 2015). At this stage, they suggest collecting information that would help understand the problem, its context and possible causes. We conducted twelve in-depth interviews with industry representatives to obtain knowledge about the challenges. Each interview lasted approximately an hour, was performed in English and followed a semistructured protocol. The interviewees were selected through convenience sampling, seeking practitioners at the forefront of circular building practices which significantly influence building design decisions. The selection was made from the contacts recommended by other scientists and the practitioners themselves. Ten out of twelve interviewees are working in the Norwegian industry, and the remaining two are based in Denmark and Germany and were selected to broaden the sample. Participants’ occupations are listed in Table 1. We focused the interviews on structural elements because they represent the most significant share of a building’s materials and corresponding embodied emissions. Table 1.
List of interviewees.
ID
Role description
Country
A B C D E
Architect & computational designer Structural engineer & Project Manager Modular timber building manufacturer Structural engineer, architectural office Project Development Director at Real Estate company Project manager, engineering company BIM & VDC specialist, civil engineering company CTO at Material Passport service provider Circularity expert CEO at glulam timber manufacturing company Business developer, sustainable buildings design Sustainability advisor at a structural engineering design company
Denmark Germany Norway Norway Norway
F G H X J K L
3 RESULTS 3.1 View on sustainability and importance of reuse
Norway Norway
At the beginning of the interviews, all participants were asked about their definition of a sustainable building, and the answers were quite consistent. The predominant reply was that sustainable building is the one we already have provided it is optimally utilised.
Norway Norway Norway Norway
‘More sustainable building is the one used by more people for a longer time’ – X
Norway
According to the answers, the second-best option is adapting the existing buildings by refurbishment. When tearing a building down becomes inevitable, half of the participants named the reuse of building components a key sustainable practice. Over half of the interviewees brought up appropriate material selection as a sustainability consideration. A third of respondents mentioned social or economic aspects of sustainability (B, E, F, H). After that, we presented the slide with three strategies. The first – designing optimised, materialefficient systems – was most favoured, especially by structural engineers, because design optimisation aims to lower the volume of elements, yielding environmental and financial cost reductions (B, F, J, K). The second strategy – the low-carbon materials – is mainly attributed to using timber and performing Life Cycle Assessment (LCA) to measure their projects’
All interviews were recorded, transcribed and anonymised. During interviews, notes were made that later served as initial coding. After that, we thematically analysed the transcriptions per each question, looking for repeating patterns in the answers. Our approach was inductive in the sense that the collected responses determined research themes instead of starting with a hypothesis. In some cases, interviews differed from the protocol depending on the interviewee’s expertise. First, we asked about each person’s understanding of sustainability and how they approach it in their work. Then, the participants were shown two slides. The first slide contained three strategies for pursuing the reduction of greenhouse gas emission of structural elements, adapted from (Fivet &
54
environmental performance (B, K). Even though it is not a direct reduction strategy, it helps to evaluate the savings and make more informed decisions. The interviews often mentioned timber because of its regenerative nature and carbon storage capacity. Still, interestingly many commented that timber is not always the best choice and should be applied wisely (A, D-F, X, J).
that it is possible to design for even 200 years without much problem at the client’s request. A similar reflection was shared by the client representative (E). Some also admit that adaptability and flexibility are among their projects’ objectives (E, K, X).
‘I don’t believe that timber [alone] is the solution to save the world. […] But the production, engineering and also assembly of the building should be done in a [sustainable way], and when their lifetime is done, it should be disassembled easily and reused again in a different building’ – D
In two interviews (E, L), there was a remark that buildings older than 1970 are easier to readapt than those built later. They blame it on the popularisation of optimisation and seeking cheaper and faster solutions. L gave the example of concrete slabs that used to have typical reinforcement throughout, unlike nowadays, detailed reinforcement layouts.
‘People live differently than before. Buildings should be flexible’ – K
Furthermore, according to experts timber is usually more expensive than alternatives and has more mechanical limitations (B, J). Timber’s environmental advantage highly depends on the end-of-life (EOL) scenario, which is unknown ahead of time, as it can be reused, downcycled to chipboards, or burnt, releasing stored carbon (C). Another problem with wood is that economic consideration might undermine environmental advantages, for example, if the cheaper wood is imported from far away (I), as happens to be the case in Norway. J adds that some timber products are harmful to the environment if disposed of at landfills due to toxins and heavy metals. Representative of building owners (E) prioritised the third strategy – ensuring long-term usefulness – as the most important, even though they admit it is often ignored. Most designers said it is not even considered in the design process, only when the project is about renovating an existing building. Three people extended the proposed strategies with off-site fabrication due to its waste prevention benefits (C); person F raised the issue of proper building placement as a factor affecting the environmental impact, and person X highlighted the importance of adaptability.
3.3 The circular workflow Figure 1 shows the simplified Business Process Model and Notation (BPMN) diagram of the reuse workflow, emphasising the linear and circular workflows overlap and processes where reuse decision-making occurs. The initial diagram was developed based on the domain literature (Durmisevic 2006; De Wolf et al. 2020) and was later updated based on the feedback from participants. The diagram consists of a typical, linear workflow (nodes 1-6, 10-11) and an intermediate reuse loop (nodes 7-9). The first starts with the production of materials using virgin and potentially recycled ingredients (1-2) and is followed by new building design (3) and construction (4). Completed building (5) serves its purpose for as long as it is functional and fit for demand, technically safe, and economically profitable (Wilkinson et al. 2014). After that time, it is deconstructed (6). A linear approach is usually concluded with demolishing the building, which leads to significant waste production. The waste can be downcycled (10) by crushing, melting, chopping, or burning and recycled into a new product. Downcycling methods have in common that material loses its original qualities but also prevents waste from ending its life by being disposed of at a landfill (11). The alternative inner reuse loop is prolonging the flow of building products by repeating steps 4-6 and adding three new steps: 7 – quality assessment and verification if an element can be reused, 8 – the socalled Material Bank, and 9 – consecutive design – which is similar to initial design (3), but embeds elements that have already been used. The Material Bank (MB) concept is defined twofold, either as a reused material marketplace (RMM) – a dedicated physical storage place with a collection of used elements from multiple donor buildings or as ‘Buildings as Material Banks’ (BAMB) – a virtual collection of pieces that are still part of the existing building but could be utilised again if the need arises (Debacker & Manshoven 2016; Rose & Stegemann 2018). Therefore, we show an association between steps 4 and 7.
3.2 (Lack of) long-term thinking As introduced in the previous section, the EOL is rarely planned at the design stage (G, J), and efforts to prolong the building’s lifespan, if any, happen only when the building is already aged. The only time when EOL was a part of a design scope was when required by ‘cradle to grave’ LCA. Almost all interviewees confirmed that in their designs lifespan of 50 to 60 years is assumed, as recommended by the standards, even knowing that technical properties would allow for much longer periods. ‘Fifty or sixty years is a too conservative assumption. Buildings should be designed for at least one hundred years’ – X According to J, timber structures are technically good for hundreds of years. Person G acknowledged
55
Figure 1. BPMN diagram of the linear and circular workflows.
examples of windows that have much lower transmittance values and steel that used to be of lower grade than is now common. Structural engineers also complained about the difficulty of adequately evaluating used products’ mechanical properties, the need for additional testing like chloride test for concrete, and problems with imposed dimensions and connections (B, D, K, L). L exemplifies cross-laminated timber – CLT – often connected with long screws that are impossible to disassemble without partial destruction. Materials have different reusability. Steel framing with permanent connections can be trimmed and repurposed, as exemplified by (Fivet & Brütting 2020), which is almost unachievable with monolithically connected reinforced concrete (12). On the other side, concrete reuse yields higher emissions savings than steel compared to recycling (Hradil et al. 2014). According to its manufacturer, timber in Norway is mainly connected in a reversible manner, making it suitable for reuse, and can have higher benefits than steel reuse (K).
3.4 New versus used – interviewees’ experiences on initial and consecutive design Facing a choice between initial design from virgin materials (3.) and consecutive design with element reuse (9.), most asked interviewees (C, E, G, H, J) confirmed that the latter is more expensive today. This is the case even for discarded elements acquired free of charge. According to respondents, the cost is driven by a much longer, iterative design process, lack of available data, low scale, and more complex assembly and transport. Also, the higher the uncertainty, the more conservative assumptions must be made, leading to an increase in material quantities. A repeating theme was also mentioned about the relatively low cost of new products and waste disposal compared to salaries, which discourages more sustainable practices (H, X, J, L): ‘Today, the price for virgin material and price for landfill is just too low, so either you have to increase the taxes or have regulations so that circular design should be the obvious choice. […] Using local steel which is recycled is much better [and should not be more expensive] than virgin steel transported from the other side of the world’. – X
3.5 Not designing for disassembly Several interviewees already declare having practical experience with reusing discarded components in new designs, but only in the case of person C, the products were prepared for it a priori. That person referred to residential and school projects, where from the beginning, it was known that a building would need to be dismantled or relocated after a relatively short time of three to ten years. The modular technology with reversible joints is favoured in such temporary structures, allowing multiple reassemblies. Designing for disassembly (DfD) as described in ISO 20887 (International Standard 2020) is recommended by incentives such as the ‘EU action plan for the Circular Economy’ (European Commission 2015) and the EU Taxonomy (European Commission 2020). However, when shown the slide with a circular workflow, most people identified barriers with the third process – initial design – saying that DfD is not in the scope of a typical project. The decisions at this stage significantly affect the potential future reusability of building components. Other studies have also
‘Buying new is just so easy and cheap, we would need to have lots of material banks’ – J ‘Materials are too cheap and labour relatively expensive, so reuse is not profitable’ – L Another aspect favouring new products, which has not been found in existing literature but appeared in several interviews, is the convenience of ordering new products compared to used ones. ‘[new] are easy to order and deliver, as opposed to reuse’ – J Another person elaborated that quantity matters: ‘it’s easier [to reuse] when you have thirty of the same doors instead of just one’ – L Before the product’s first life ends, it might get outdated as technology progresses. Interviewees mention
56
steps, especially information from the initial design (3.) should be stored and passed to the future designer. However, when asked, the interviewees admitted that they neither had access to the data they wanted nor passed enough information themselves that could help reuse components from their designs.
identified a lack of DfD in architectural practice (CruzRios & Grau 2020). Many of our interviewees said they do not know the DfD term and related standards (A, C, X, H), and some were aware of it but did not apply DfD in projects (B, D, J). Surprisingly, some designers specify bolted prefabricates with easy connection access, but their disassembly is not considered at the design stage; instead, ease of fast assembly. Modular building manufacturer (C) mentioned examples of projects that were designed with relocation in mind, but the DfD concept and standards were not known to them. Despite that, most were able to name good practices that would enable easier reuse. To name some:
‘We need a knowledge base’ – F. ‘Good software is needed! (…) The less we know, the more conservative assumptions we make.’ – J As useful, they list the element’s dimensions, exact material properties, including structural and acoustic properties, conditions it was used in, structural assumptions and results of calculations, how to detach elements, so connection information, for timber: does it have moisture cracks, what glues were used (important for fire safety), for concrete: what is the reinforcement, and optionally also mixture design and load history (D, G, J). Many interviewees highlighted the essential role of digitalisation in enabling reuse, lowering its costs and reducing uncertainty (C, E, G, H, J, K, L). They acknowledge applying Building Information Modeling (BIM) on most projects, except for a few where 2-dimensional CAD technology is chosen to reduce design costs. One person said their handover data is being integrated into the Facility Management (FM) solution used at the operational phase of building life (E), aiming to achieve the complete Digital Twin. Person G criticised BIM for its inability to track decision-making and poor implementation of the common schema – IFC. They admit that useful information, such as concrete mix design and structural calculations, is often kept in external documents and not included in the BIM data structure. This is somewhat confirmed by the digital solution provider (H), who said that BIM is a valuable but not exhausting source of information for reuse. Residual value, environmental data, and easiness of detachment are among the information rarely found in BIM. A repetitive pattern in the interviews reflects the isolated nature of projects and related data and the lack of overarching information and knowledge sharing (C, G, L). Similar findings are seen in a report from the BAMB project (Peters et al. 2017), which mapped slow circularity adoption to lack of transparency, information availability and automated connectivity to designers’ tools and FM systems. According to the report, information on reusable products should replicate the offering of new, unused products to improve trust. To remedy that problem, person J envisioned a BIM-based recommender solution that would show the consequences of design decisions, and L described a platform that allows for logging and accessing information about products and materials available for reuse. Architect A explained the mobile application they are working on that aims to allow for documenting dismantled objects. Interviewee K, however, identifies
‘You should optimise connections for multiple reassemblies, not only [increased] forces’ – K ‘Do not overcomplicate projects and pay attention at early design’ – D 3.6 Agency to reuse Interviewed designers do not see a significant technological barrier to reuse and admit that if clients request, they can reuse building components, provided adequate remuneration exists. According to them and manufacturers, it is about client awareness and willingness to invest. But based on the opinion of the interviewed client: ‘[we do] energy efficiency not just for BREEAM – it saves us money! We are willing to pay more for reuse, as green-certified building brings profit. […] The more projects we do, the more we are aware and know what to ask for.’ – E The primary motivation for the designers was their ambition from their sustainability awareness. They describe that their role is often to raise clients’ awareness, guide, and challenge them to reach higher objectives (D, K). Some see motivation coming from certifications, mostly Norwegian BREEAM, that lead to financial profits in higher building value and rent (D, F). However, one also said that reuse is not rewarded enough in current certifications. Among other motivators were marketing, the company’s reputation, better insurance, and bank offerings (I). On the other hand, people’s desire to get something new when they build lowers the motivation to reuse. Building with used products also raises fear for structure’s durability. From the architect, we hear that popularising reuse will require changing the attitude of architects to favour old, with the limitations it entails, and accept this aesthetics (A). 3.7 Information availability A vital aspect brought in many of the interviews is the information availability, which is related to the uncertainty of product documentation described in previous research (Sigrid Nordby et al. 2019). This issue prevails in the consecutive design stage (9.). Still, it must be addressed almost in all preceding
57
with the reuse of preowned components than with consciously designing for easier decommission. The lack of DfD leads to entering the potential circular workflow with a significant handicap that makes the disassembly (6.) complicated and laborious. This situation might be dictated by a lack of awareness among those who order work and the market favouring shortterm economic value. However, some interviewees follow the DfD guidelines for practical reasons, such as offsite fabrication and modular architecture. Furthermore, we found that the circular strategies can contradict the prevailing design strategy. While both approaches aim at emission reduction, the circular assumes that the future use of a building and its components is uncertain; hence it should be flexible and versatile, and the typical objective of design is to optimise for the single assumed scenario. That justifies why, in the interviewees’ opinion, it is easier to adapt older buildings than modern, more optimised ones, despite improvements in technology. From a methodological point of view, it was helpful to show the workflow diagram to diminish the confusion about the topic. Without clearly explaining the subject, there was confusion in the answers resulting from conflicting definitions regarding recycling, material reuse, circular economy etc. The diagram also led to a slight discrepancy between responses, depending on whether the preowned elements should be transported to intermediate physical storage – RMM, or be only tracked in a virtual repository and delivered straight from decommissioned building donor – BAMB. This shows that the concept needs some clarification, but practitioners accept the general idea. Designers say they can reuse structural elements, and most have the required knowledge, but it takes more time and effort, leading to higher costs. That effect is even more significant in Norway, where labour-intensive reuse of whole components turns more expensive due to relatively high labour costs compared to materials. Relatively low cost of virgin materials and waste disposal compared to more environmentally friendly alternatives might be prone to change due to economic and regulatory considerations. The interesting aspect, from the DSR perspective, is the importance of convenience of the design and construction with preowned products. To improve the adoption of CE practices, an artefact should simplify the procedure and make the process faster and more convenient, as well as reduce uncertainty and associated risk by providing relevant information about second-hand products. Some would like to see a digital tool in the form of design assistance or knowledge base. This should be further studied in the next stage of the research. Optimistically for the future of reuse, the motivation to practice circularity comes both from the architects, engineers and clients, provided they are aware of the benefits. Also, the main listed barriers are not constrained by the technical properties of materials and can be reduced by proper incentives, regulations,
that problem lies not in software but in people who are hesitant to change. 3.8 Elements’ standardisation Finally, the aspect that some expect to be vital for widespread reuse adoption is the standardisation of elements, connections or whole building modules (C-E, K, L). The concept is supported by the ISO standard (International Standard 2020), which advocates using standard-sized building components assembled with typical connections. Peters et al. suggest it can also enable the exchangeability of individual products (Peters et al. 2017), and others exemplify how such standardisation might be integrated into the design (Brütting et al. 2021). ‘We should move from on-site craftsmanship to industrialisation.’ – E. Apart from contributing to the stock of repeatable elements, standardisation also leads to more accessible consecutive design, quality assurance, and industrialisation that can prevent waste production. According to L, designers do not realise the impact of prevailing mass customisation on the environment. The downside of standardisation is that it can limit design freedom and block innovation. Contractors might even perceive it as a threat to their occupation (Anastasiades et al. 2021). Manufacturer of modules – C – name transportation as an essential constraint on element’s dimensions. Structural designer D adds to that the number of variables that should be considered, including a variety of connections and possible internal forces. 4 DISCUSSION AND CONCLUSIONS From the interviews, we noticed high awareness of the problem. Practitioners show a positive attitude towards the reuse of building elements, preparatory DfD and supplementary urban mining and perceive them as sustainable and desired practices that reduce the negative impact of the construction industry on the environment. However, this view may not be representative as the selection of participants was done by convenience sampling based on their expertise in sustainable buildings. It would be worth enhancing the validation of the study by collecting data from a bigger group representing various countries to better understand the issue’s scale and scope. Furthermore, the findings from this qualitative study could serve as a basis for defining a quantitative survey. The group not represented in this research are the building users and health and safety experts who might have objections to occupying buildings from reused components. The designers have a good awareness of DfD practices, such as reversible joints and more repetitive elements, even though the term and standards are unknown to many. More designers have experience
58
Durmisevic, El. 2006. “Transformable Building Structures.” TU Delft. European Commission. 2015. “Closing the Loop – An EU Action Plan for the Circular Economy.” Communication from the Commission to the European Parliament, the Council, the Europena Economic and Social Committee and the Committee of the Regions. European Commission. 2020. “EU Taxonomy – 2020/852 on the Establishment of a Framework to Facilitate Sustainable Investment, and Amending Regulation (EU) 2019/ 2088.” Fivet, Corentin, and Jan Brütting. 2020. “Nothing Is Lost, Nothing Is Created, Is Reused Structural Design for a Circular Economy.” Structural Engineer 98(1): 74–81. Hradil, Petr et al. 2014. “Re-Use of Structural Elements Environmentally Efficient Recovery of Building Components.” VTT Technical Research Centre of Finland. Iacovidou, Eleni, and Phil Purnell. 2016. “Mining the Physical Infrastructure: Opportunities, Barriers and Interventions in Promoting Structural Components Reuse.” The Science of the total environment 557–558: 791–807. IEA & UNEP. 2018. 2018 Global Status Report Towards a Zero-Emission, Efficient and Resilient Buildings and Construction Sector. International Standard. 2020. “ISO 20887 Sustainability in Buildings and Civil Engineering Works — Design for Disassembly and Adaptability.” Peters, M, A Ribeiro, J Oseyran, and K Wang. 2017. BAMB Report Buildings as Material Banks and the Need for Innovative Business Models. Rose, Colin M., and Julia A. Stegemann. 2018. “Characterising Existing Buildings as Material Banks (E-BAMB) to Enable Component Reuse.” Proceedings of the Institution of Civil Engineers: Engineering Sustainability 172(3): 129–40. Sigrid Nordby, Anne, Katie Zemlick, Elmira Kalhor, and Bruce M Thomson. 2019. “Barriers and Opportunities to Reuse of Building Materials in the Norwegian Construction Sector.” IOP Conference Series: Earth and Environmental Science 225(1): 012061. Wilkinson, Sara J, Hilde Remøy, and Craig Langston. 2014. “Sustainable Building Adaptation: Innovations in Decision-Making.” In , 296. De Wolf, Catherine, Endrit Hoxha, and Corentin Fivet. 2020. “Comparison of Environmental Assessment Methods When Reusing Building Components: A Case Study.” Sustainable Cities and Society 61(102322).
raising awareness, change of procedures and digital support. ACKNOWLEDGEMENTS The authors would like to thank the people who agreed to the interviews and shared their experiences and opinions about the topic, as well as reviewers for their evaluation and suggestions. REFERENCES Acharya, Devni, Richard Boyd, and Olivia Finch. 2018. “From Principles to Practices: First Steps towards a Circular Built Environment.” : 14. Anastasiades, K. et al. 2021. “Standardisation: An Essential Enabler for the Circular Reuse of Construction Components? A Trajectory for a Cleaner European Construction Industry.” Journal of Cleaner Production 298: 126864. Bocken, Nancy M.P. P, Ingrid de Pauw, Conny Bakker, and Bram van der Grinten. 2016. “Product Design and Business Model Strategies for a Circular Economy.” Journal of Industrial and Production Engineering 33(5): 308–20. Brütting, Jan, Gennaro Senatore, and Corentin Fivet. 2021. “Design and Fabrication of a Reusable Kit of Parts for Diverse Structures.” Automation in Construction 125: 103614. Çetin, Sultan et al. 2021. “Circular Digital Built Environment: An Emerging Framework.” Sustainability (Switzerland) 13(11): 6348. Cruz-Rios, Fernanda, and David Grau. 2020. “Design for Disassembly: An Analysis of the Practice (or Lack Thereof) in the United States.” Construction Research Congress 2020: Project Management and Controls, Materials, and Contracts – Selected Papers from the Construction Research Congress 2020: 992–1000. Debacker, Wim, and Saskia Manshoven. 2016. “D1- Synthesis of the State-of- the-Art.” Bamb. Dresch, Aline, Daniel Pacheco Lacerda, and José Antônio Valle Antunes. 2015. “Design Science Research: A Method for Science and Technology Advancement.” In Design Science Research:A Method for Science andTechnology Advancement, Springer International Publishing, 47–126.
59
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Evaluating existing digital platforms enabling the reuse of reclaimed building materials and components for circularity W. Wuyts, Y. Liu, X. Huang & L. Huang Department of Civil Engineering and Manufacturing, Norwegian University of Science and Technology, Gjøvik, Norway
ABSTRACT: In Europe, several examples for digital marketplaces for reuse of reclaimed building materials and components could enable more circularity in the architecture, engineering and construction (AEC) sector. Not many research and perspectives on digital platforms for reuse offer a compass for AEC to evaluate existing platforms. In this explorative and reflective paper, we identified and evaluated around 20 markets in Europe and one in North America, which are representative for its local or regional market. We explored them through a set of pre-defined functions (e.g., basic functions, service for selling), and the access to information. Many markets are segmented, small-scale, prototype platforms and initiatives not yet replicated or scaled up for wider use. We argue to take a multiple user group perspective, as different stakeholders are needed to collaborate and exchange materials and information; this requires seeing suppliers of materials and information as possible user groups. We propose a set of questions and guidelines for evaluation, according to match requirements of multiple user groups and functions of the platform, the capacity of sharing risks, costs, benefits and profit, circularity targets and a regional perspective.
1 INTRODUCTION In response to circular economy (CE) policies and 21st-century challenges, the Architecture, Engineering and Construction (AEC) industry is (re)discovering the practice of reuse of building materials and components. The reuse of components and materials requires new infrastructures and planning, and different stakeholder management. One of the crucial activities is effectively matching the suppliers and users of reclaimed material and components through efficient information sharing (i.e., where, how much, what quality, etc.). Another important activity is the relocation and temporary storage of these materials and components in material banks, and the collection and provision of the required information about these materials (and the way they are transported) by the new users. One of the emerging solutions to address these new needs in the industry and their customers is digital platforms for the reuse of building materials and components to support the recirculation of reclaimed materials and so-called material banks where these materials are stored (e.g., Marin et al. 2020; Waldman and Cai 2019). Different researchers argued and illustrated how they can contribute to sustainability (see the discussions by De Reuver et al. 2018; Staab et al. 2022), including reuse of materials and other circularity practices (Çetin et al. 2021): digital platforms are answers to needs of information and other requirements and can have different functions (see Staab
60
et al. 2022). Some papers address mostly the side of the provision of data, for example by inserting Building Information Models (BIM) (e.g., Durmisevic et al. 2021). However, the BIM model is limiting, as it requires certain skills and experience; few industry representatives only use it. For other users of the information to make informed decisions to spearhead circularity and reuse strategies, like policymakers, the BIM is too difficult. A platform has to serve different user groups within the AEC industry and their customers as well as policymakers and other target audiences that can help society. Although digital platforms are recognised by the AEC industry to enable reuse (e.g., Knoth et al. 2022) and windows of opportunities are opening in the EU and beyond despite barriers to data sharing and lack of standardisation (Staab et al. 2022), not many research and perspectives on digital platforms for reuse offer a compass for AEC to decide which platform to use for which requirement and which user group it benefits. In Europe, several examples of digital marketplaces for the reuse of reclaimed building materials and components could enable more circularity in the architecture, engineering and construction (AEC) sector. Not many research and perspectives on digital platforms for reuse offer a compass forAEC to evaluate existing platforms. Therefore, the first objective of this paper is to understand the requirements of different user groups and see if there are matches with functions/solutions provided by the platform.
DOI 10.1201/9781003354222-8
be yet another possible reason. Addressing these challenges requires a stronger focus on understanding the industrial context and the stakeholders involved, which is also common for studies on digital platforms in general (De Reuver et al. 2018). Hence, we build further on the work by Çetin et al. (2021) which provided insights about constellations of actors in the circular built environment, as well on the work of Durmisevic et al. (2021) that highlighted the main user groups of digital platforms in general. In the winter and spring of 2022, we were conducting 1:1 interviews with AEC companies in North-West Europe, attending webinars and other networking events related to digitalization enabling the reuse of building materials and components and we identified these AEC companies and other target audiences that can benefit of these platforms. Mostly, we distinguish two main target groups (that both again can be divided into smaller groups): Users of Information (UoI) need various kinds of information depending on their stakeholder roles (e.g., designers, real estate developers, building owners, public authorities and urban miners). Users of Materials (UoM) are those stakeholders who buy the materials on these secondary materials market platforms (e.g., construction companies). The same typology is also for suppliers of the digital platform: suppliers of materials (SoM) and suppliers of information (SoI). Some stakeholders might act as UoI and UoM or SoM and SoI. However, we argue to see suppliers as possible user groups, as they would also want something in exchange for their supply of resources (information, materials).
In our explorative research, we reflect also on the systemic effects of this circular solution in order to present a critique from especially an interdisciplinary and regional approach to the circular economy on the state-of-art of current digital platforms enabling the reuse of reclaimed building materials and components for circularity. These contributions will help to propose a list of guiding questions that can support practitioners and academics to choose or improve an existing platform that serves the CE practice of reuse and will have more systemic effects and a bigger impact on the city or region in which this platform is located. They will help to reflect on the extent to which the requirements of multiple user groups and functions of the platform are matched, the capacity of sharing risks, costs, benefits and profit, circularity targets and a regional perspective.
2 METHODS This paper conducted desktop-based research which consists of a preparation study (critical literature review) and case study research.
2.1 Preparation: Combining critical literature review with experiential insights In the autumn and winter of 2021, we conducted an integrative literature review, where we read papers from different disciplines as well as website pages and news media without any predefined concept in our minds. A critical review spans often literature of different disciplines and does not follow strict rules as a systematic literature review, because first, there are not enough papers on the subject and because one of the aims is to propose some conceptual framework for digital platforms for the reuse of wood (Snyder 2019). This preparation enveloped debates and reflections with representatives of the AEC sector, other research institutes and policymakers who are working on the circular built environment theme. For understanding the needs or requirements of user groups, we conducted interviews with different stakeholders in Belgium and Norway (not published yet). This helped to define categories (functions, requirements). The limit is the risk of missing potential functions and requirements.
2.3 Compiling a list of existing platforms In winter 2021, we started to compile a list of existing initiatives. As many digital platforms are recently launched, and academic literature is barely available (see 2.1), we used Google to search for existing platforms using keywords like digital platform, reuse building material, reclaimed material, building, recycled material, deconstruction, demolition, and webshop. During the aforementioned interviews (2.1.), webinars and other networks we attended in autumn 2021-spring 2022, we identified additional platforms and ended in May 2022 with a list of 21 existing platforms, all in Europe, except one in North America (Table 1). These platforms are representative of their local or regional market. Some companies/platforms serve only (public and/or private) organizations (e.g., Loopfront), while others are available for DIY reuse activities (e.g., Genbyg, Rotordc). Noteworthy, we identified platforms that contribute to the recirculation of secondary materials, like Amazon and Finn.no, but we excluded them, because their main goal is to create profit without considering explicitly environmental sustainability and circularity targets and they are mostly created from a linear economy model.
2.2 Identification of (potential) user groups and their requirements As the circular economy includes principles of sharing materials and extended producer responsibility (EPR), concerns about ownership and governance of solutions, products and materials or even data in these secondary material exchange platforms, could
61
Table 1.
the customers have to arrange the pick-up by themselves. 4. Consulting service: The owner provides special services for customers, often based on their data insights and knowledge. 5. CO2 saved value: Some platforms calculate and provide the embodied CO2 emissions value in the reclaimed materials. 6. Replacement cost/price: Some platforms provide the cost to reproduce a virgin material which has a similar function of reclaimed materials. 7. Material passport: The platform provides a document for each product that describes defined characteristics of materials in the product, which helps to calculate resource recovery or reusability indexes. 8. Bid: Some platforms work with an action, where customers can bid for a reclaimed material or component. Table 2. provides an overview of which functions are provided by the sampled platforms.
Summary of the existing platforms.
Country*
Name platforms**
Amount
Belgium
Materialenbank, OPALIS, RotorDC, 2ehands Genbyg, rezip Restdo Buurman, insert, Madaster Loopfront, Resirquel Ccbuild, recurkultur, tiptapp Annibis, tvz, ricardo, tutti Globechain reusewood
4
Denmark Germany The Netherlands
Norway Sweden Switzerland United Kingdom United States
2 1 2
2 3 4 1 1
Table 2. The 8 functions for digital platforms.
* Country where the organization is registered ** Names of the organization, network of project
1 reusewood insert opalis ricardo btvz Anibis tutti 2dehands Resirquel buurman rotordc materialbank genbyg tiptapp globechain rezip restado Retur-cultur Loopfront madaster ccbuild
3 RESULTS Many markets are segmented, small-scale, prototype platforms and initiatives not yet replicated or scaled up for wider use. We observed that they do not always take a regional approach, i.e. they do not communicate on their websites about key actors they (want to) target in a specific city or region. However, based on interviews, we know that some projects like Materialenbank Leuven and loopfront work closely together with public authorities at the urban and regional level. Madaster is replicating their platform from the Netherlands to other national markets (e.g., Norway). In some cities (e.g., Oslo), more platforms are competing, which leads to observed questions in forums and debates if these platforms can interoperate with each other and about their Application Programming Interface (API), especially as some (potential) user groups observed that a platform is missing information that can be provided by another platform and an interaction between data and information exchange could benefit them. The desktop-based study created a list of basic functions (e.g., digital shop, match and query) which were identified on most platforms, and a list of advanced functions (e.g. CO2 saved values). Based on explorative research, we define 8 functions: 1. Basic functions: search functions which meet the basic requirement for the demanding party to find the targeted information 2. Physical store:The website promotes a shop where customers can visit and buy materials. 3. Retail service: The platform enables the opportunity to buy reclaimed materials. This does not always include a logistic service; in some cases,
× × × × × × × × × × × × × × × × × × × ×
2
3
4
5
6
7
×
×
×
×
×
×
×
× * × ×
× × × × ×
× × × × ×
× × ×
× ×
8
×
× × ×
×
* × × ×
× × ×
× × × ×
× ×
*free charge
The most frequent observed function, besides the basic functions, is the replacement cost. More surprisingly, were that many platforms did not offer a physical store or a retail service where an exchange of materials can be established, which should be the role of a reuse platform. Some platforms only enable exchange, or even only collection management tools, for data. Lastly, we did not identify many functions that inform how they contribute to circularity targets (except F5), like CO2 saved value. We assessed the level of access to the materials and information and divided the search function into open search and limited search. The latter means and often
62
Table 4. The business model, with 1. Non-profit, 2. Advertisement or charge members, 3. Retail, 4. Consulting service, 5. Third-party, which means they work as a broker when there is a conflict.
implies payment or part of an exclusive network (of for example a project or a consortium) Table 3. The search mode. Open search
Reusewood Insert Opalis Ricardo Btvz Anibis Tutti 2dehands Resirquel Buurman Rotordc Materialbank Genbyg Tiptapp Globechain Rezip Restado Returcultur Loopfront Madaster Ccbuild
× × × × × × × × × × × × × × × × ×
Only open for special members
1 Reusewood Insert Opalis Ricardo Btvz Anibis Tutti 2dehands Resirquel Buurman Rotordc Materialbank Genbyg Tiptapp Globechain Rezip Restado Returcultur Loopfront Madaster Ccbuild
×
×
× ×
× × ×
×
× ×
2
3
4
×
×
×
5
× × × × × × × × × × × × ×
×
× ×
×
×
× × ×
restado among a few, follow the business model to open the material bank to attract potential members. Some platforms can have extra income by selling their products in their physical shops, such as insert and genbyg. Some platforms focus on enterprises by supporting consulting services, such as madaster. Since the consulting service is a customized service, these platforms usually charge members to list their material bank and to have extra consulting services. Summarised, in this explorative study, 21 existing digital platforms for mapping and selling reclaimed building materials are investigated. We observed that the main purpose of these platforms is to build a connection between material demanders and material suppliers. In some cases, additional and advanced functions are provided that diversify the portfolio of the organization and can create income, like consultancy services. Unsurprisingly, all the cases have the basic functions, namely, match function, search function, query function, and material bank or inventory function. However, we will discuss that especially some advanced functions, like co2 estimation (and preferably with a spatial analysis) are the functions that provide the information to make informed decisions leading to the circularity targets and impacts.
Based on this exploration, we assume that many platforms work like brokers that build a connection between demanders and suppliers, which may make money from advertisement. However, some platforms make money from consulting services, whose material banks are not open for citizens, but only open for special members or customers. Usually, these platforms have some special functions that other ordinary brokers can not provide. For example, the Norwegian tool Loopfront can provide information on co2 saved value and cost saved value estimation based on its unique database by using special estimation technology (Loopfront 2022). From the abovementioned, it seems some platforms are successful with large inventories in the material bank, while some platforms have few inventories in the bank. “ccbuild” and “BTVZ” are two typical forms which have few materials in the inventory (ccbuild, 2022; BTVZ; 2022). On the contrary, “genbyg” is the biggest digital market for building reused material in Denmark (Genbyg 2022). This comparison implies that one of the key KPIs for the digital platform is the numbers and types of materials in inventory. The number of members or turnover can be complementary to the popularity of KPI. The results in Table 4 are made by assumptions and interpretations that are based on observing the websites and interpreting the presence and communication on their charging mode, the existence of physical workshop, the types of service, and their payment mode. Table 4 indicates that some platforms have a non-profit models, while others have profit-driven business models. Many platforms, namely insert, ricardo, anibis,
4 DISCUSSION 4.1 Which functions match which requirements of which user groups? (And why?) Creating digital marketplaces for enabling reuse is nontrivial as large amounts of data need to be collected,
63
Another example is information about risks and costs. We did not find any information about the risks for different user groups, only eventual costs for using this platform. We argue for a more benefit-sharing, cost-sharing risk-sharing and profit distribution model for these small-scale existing platforms. We anticipate that the competition between existing and new platforms will embrace this model. We should evaluate existing platforms according to how their technology shares risk, benefit, cost and profit among all relevant stakeholders and user groups.
analyzed, and represented in a way making it available for interested parties seeking to reuse materials. This implies collaboration between different user groups, which means that such platforms should be attractive for all these user groups, to give sometimes data, in exchange for another resource or service. When we analyze the platforms and the reactions in debates we observed and interviews we conducted, we see a multiuser perspective is often missing. In compliance with other explorative and design work (Durmisevic et al. 2021), we identified seven possible main user groups of such platforms: 1) architects, designers and construction companies; 2) producers and manufacturers of building components (e.g. windows, 3) building owners and real estate developers, 4) urban miners and specialized demolition companies, 5) certification bodies and policymakers, 6) building end-users, often participating in Do-it-Yourself (DIY) construction, 7) researchers.
4.2 Limitations and future action In Europe, several examples of such digital marketplaces for reuse exist. However, most of those are segmented, small-scale, prototype platforms and initiatives not yet scaled up for wider use. We identified and evaluated only 21 markets and our sample will not include all existing digital platforms enabling the reuse of building materials and components. Future actions should involve updating the list in order to evaluate and monitor the functions and how they contribute to the industry and society. Moreover, we assessed them via their websites, and we have only access to the information that is accessible to each website user. Interviews with platform developers and managers might provide more insights about their identified costs and risks (e.g. losing possible income revenue or market position by sharing information about the market) for the different user groups, shortcomings that surfaced after building the platform: for example, in our conversations with AEC industry, we observed that some platform developers share that most companies are not so willing to share data or use the tool because of (company) privacy, implementation of the tool also requires an intensive trajectory of facilitation, quality of the data that can be extracted can largely differ. We miss information about one of the big success indicators of digital platforms, namely the number of users, the user groups, and if the platform created network effects, which means the reuse of the materials or components (De Reuver et al. 2018) Lastly, we noticed that different platforms, that link their goals and targets with circularity, might envision the concept in different ways. Circularity is a contested concept which gets interpreted differently (Korhonen et al. 2018), and this might lead to different products, or platforms that deliver different solutions. Therefore, we encourage even more research that studies the values in the circularity targets and the choice of functions and missions of a digital platform for reuse and other circularity enabling tools. This implies also recognizing users within user groups that will not benefit of these enabling tools, because they do not have the digital literacy. Taking an intersectional environmental approach to circularity tools (e.g. Wuyts & Marin 2022) will lead to even more ideas for a socially just risk- and benefit-sharing distribution model. A way
Table 5. Which functions (F) might interest which user groups (U)?, with X: obvious, ?: arguable.
F1 F2 F3 F4 F5 F6 F7 F8
U1
U2
U3
U4
U5
U6
U7
X X X ? X X X
X
X X X ? X X X X
X X X ? X X X
X
X
X
X ? X X X
X
? X X X
? ?
X X X
Table 5 matches the functions of the user groups. Some functions are more interesting for multiple user groups. In the case of the producers and manufacturers of these building components, most services are not interestingly, unless F5, if it helps them to provide information for marketing of their sustainability impact. However, mostly they produce elements for future reuse, but they do not contribute to the current circularity and cannot benefit much from or contribute, to such a digital platform if there are no services that support reverse logistics. In our sample, the frequent functions (next to the basic functions) are cost calculation (F6), promotion of the physical store (F3) and retail service (F4), not always with a logistic service. They are interesting for a few user groups, but in interviews and relevant literature (e.g., Knoth et al. 2022) we identified more concerns and needs about reclaimed wood that are not addressed. For example, most of these functions do not provide the data that can help researchers, policymakers and certification bodies, these platforms seem to take mostly a single user group perspective towards for example architects and/or DIY construction-involved citizens.
64
2. Match requirements and user group needs
for taking an intersectional approach is by focusing on a territory or region which this digital platform serves or would serve. This is in compliance with new initiatives such as the Circular Cities and Regions Initiative (CCRI) and newly launched concepts such as the Circular Systemic Solutions (CSS) that call for systemic thinking at the urban and regional level (European Commission 2022). Therefore, we applaud future research and action looking into evaluations of impact and success to apply a regional approach to circular economy and enabling technologies (Bianchi et al. 2022; Marjanovic et al. 2022). Firstly, this means that we do argue for decentralized initiatives and acknowledging the specificities within the region (e.g., different economic structure, Technology Readiness Level, or even awareness of circularity, acceptance of digital solutions, existing infrastructure) and envision different indicators for the different regions. Secondly, a regional approach encompasses our vision that material and product flows should be more localised. We read in blogs of some of these platforms about the phenomenon of (only) high-value secondary wood being transported from a country like Belgium to Australia. We want to stimulate rather local/regional markets and circular activities and encourage evaluations that encourage local and shorter supply chains. In a later reflection, we realized that several platforms that offer a retail service do not provide a logistics service. Buyers of the material have to arrange the pickup themselves. This can create a barrier for buyers, which can imply that we miss information about transport impact and miss the opportunity to link the selling of materials with the provision of the most optimal routing regarding environmental and economic costs. Lastly, there is also a whole body of research reminding us of those digital technologies and solutions should not create the problems that the circular economy aims to solve (e.g., creation of e-waste, GHG emissions of data storage plants). In a conversation with an urban miner looking for a platform, there was a concern about the traceability of the material. They were afraid that virgin materials and components would also be offered on such platforms, as some platforms serving the linear economy (e.g., finn.no, amazon) also offer secondhand materials and components. Future research should investigate possible unintended effects and possible misuses.
– Which user groups benefit from the offered services? – Which user groups do not benefit? – For a deeper reflection on why some user groups do not benefit: which values are behind the design and further development of the platform and the functions they provide? – … 3. Circularity indicators – Which environmental impacts are measured and communicated? – How many materials are resold and integrated into society again? – How are these indicators measured? – Can we trace back from where these materials come from and validate that they are reclaimed and not virgin materials? – … 4. Regional perspective – Who are all the key actors in the region? Where are they located? – Are some key actors missing? – How are they connected with each other and the materials, spaces, infrastructure and other resources? – What is the technology readiness level in this region? – What is their market radius? – Do they offer only locally mined components and materials? – … Noteworthy, we do not argue that digital platforms should offer as many services as possible, but we argue to take a multiple user group perspective, as different stakeholders are needed to collaborate and exchange; this requires seeing suppliers of materials and information as possible usergroups. For policy makers that want to map missing tools in their territory to enhance circularity or reuse, this compass can help to see if more platforms which different functions and business models are needed that can benefit the overlooked citizens and other actors that can gain benefits of reclaimed building materials and components. We did not apply these questions to all the case studies in our sample. However, in future European and national projects we are developing, we are planning to improve the compass and guidelines in cooperation with different user groups in selected pilots by testing this compass and collecting and processing feedback.
5 CONTRIBUTION TO INDUSTRY We propose a first set of guiding questions and indicators that can be asked and serve as a compass for AEC industry and other user groups to see if this platform is the right fit for their vision, targets, mission and especially values: 1. Digital platform success indicators – How many users are registered? – How many different user groups? – How frequently do they use a service? – …
ACKNOWLEDGEMENT This explorative research is part of the Norwegian GreenPlatform project, circWOOD funded by the Norwegian Research Council (project nr 328698).
65
REFERENCES
products in Norway. Journal of Cleaner Production, p.130494. Korhonen, J., Nuur, C., Feldmann, A. and Birkie, S.E., 2018. Circular economy as an essentially contested concept. Journal of cleaner production, 175, pp. 544–552. Loopfront (2022) https://www.loopfront.com Marin, J., Alaerts, L. and Van Acker, K., 2020. A materials bank for circular leuven: How to monitor ‘messy’circular city transition projects. Sustainability, 12(24), p. 10351. Marjanovi´c, M., Wuyts, W., Marin, J. and Williams, J., 2022. Uncovering the Holistic Pathways to Circular Cities— The Case of Alberta, Canada. Highlights of Sustainability, 1(2), pp. 65–87. Materialenbank Leuven. 2022. https://materialenbankleuven. be/ Morseletto, P., 2020. Targets for a circular economy. Resources, Conservation and Recycling, 153, p. 104553. Opalis. 2022. Ww.opalis.eu Ricardo. 2022. https://www.ricardo.ch/en/ Resirquel. 2022. http://www.resirqel.no/ Restado. 2022. https://restado.de/ Returkultur. 2022. https://www.returkultur.se/ Reusewood. 2022. Reusewood.org RE-ZIP. 2022. https://re-zip.dk/en/home/ Snyder, H., 2019. Literature review as a research methodology: An overview and guidelines. Journal of Business Research, 104, pp.333–339. Staab, P., Pietrón, D. and Hofmann, F., 2022. Sustainable Digital Market Design: A Data-Based Approach to the Circular Economy. Tiptapp. 2022. https://www.tiptapp.com/en Tutti. 2022. www.tutti.ch Wuyts, W. and Marin, J., 2022. “Nobody” matters in circular landscapes. Local Environment, pp.1-18. 2ehands. 2022. https://www.2dehands.be/
Anibis. 20220. www.anibis.ch Bianchi, M., Cordella, M. and Menger, P., 2022. Regional monitoring frameworks for the circular economy: implications from a territorial perspective. European Planning Studies, pp. 1–19. BTVZ. 2022. Www.Btvz.ch Buurman Rotterdam. 2022. https://www.buurmanrotterdam. nl/webstore Cai, G. and Waldmann, D., 2019. A material and component bank to facilitate material recycling and component reuse for a sustainable construction: Concept and preliminary study. CleanTechnologies and Environmental Policy, 21(10), pp. 2015–2032. Ccbuild. 2022. https://www.ccbuild.se/sv/marknadsplats Çetin, S., De Wolf, C. and Bocken, N., 2021. Circular digital built environment: An emerging framework. Sustainability, 13(11), p. 6348. De Reuver, M., Sørensen, C. and Basole, R.C., 2018. The digital platform: a research agenda. Journal of Information Technology, 33(2), pp. 124–135. Durmisevic, E., Guerriero, A., Boje, C., Domange, B. and Bosch, G., 2021, October. Development of a conceptual digital deconstruction platform with integrated Reversible BIM to aid decision making and facilitate a circular economy. In Proc. of the Joint Conference CIB W78-LDAC (Vol. 2021, pp. 11–15). European Commission (EC), 2022. About the circular cities and regions. Accessed on 20 July 2022 via https://circularcities-and-regions.eu/ Genbyg. 2022. https://genbyg.dk/ Globechain. 2022. https://globechain.com Insert. 2022. https://www.insert.nl/ Knoth, K., Fufa, S.M. and Seilskjær, E., 2022. Barriers, success factors, and perspectives for the reuse of construction
66
Product data
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Semantic Material Bank: A web-based linked data approach for building decommissioning and material reuse A. Akbarieh & F.N. Teferle Department of Engineering, University of Luxemburg, Luxembourg, Luxembourg
J. O’Donnell School of Mechanical and Materials Engineering and UCD Energy Institute, University College Dublin, Dublin, Ireland
ABSTRACT: One of the barriers to circular construction is the lack of availability or visibility of reusable materials and components at the right time and place. Therefore, this paper suggests a digital solution based on identified key stakeholders’ information requirements and market motivations. This solution helps close the material loop between the decommissioning phase and the new construction phase through semantic technologybased information exchanges among stakeholders. The proposed ontologies are twofold: 1) a Decommissioning & Reuse Ontology (DOR) that enriches information models with circular and End-of-Life cycle information while 2) the Ontology for Environmental Product Declaration (OEPD) digitalising standardised and comparable sustainable information. Both ontologies are employed in the Semantic Material Bank (SMB) proof-of-concept: a BIM-compliant digital urban mining solution through which defined stakeholders can evaluate the availability and status of reusable and recyclable elements for future construction projects.
1 INTRODUCTION The construction industry is among the key sectors highlighted in the new Circular Economy Action Plan (European Commission 2020a) because of the high material and carbon intensity of construction processes and products. 50% of material extraction (European Commission 2020a) is attributed to the construction industry globally, while 35% of all waste is Construction and Demolition Waste (CDW) in the European Union (European Commission 2016). This high material consumption and waste output indicates high potential for circularity in the construction sector. Circularity is a concept borrowed from Circular Economy, which is a contemporary paradigm for keeping the materials in the value chain for as long as possible with the highest quality possible (European Commission 2019). New and optimised material cycles are required for products or processes to accommodate this circular vision by diverting material from waste streams and encouraging recycling or reuse, which are only two of many ways that circularity can be manifested (Kirchherr et al. 2017). Alarming data estimates the total building stock to double over the next forty years as a result of population growth, while nearly a third of the present building stock is anticipated to come down (Build Reuse 2022). In Europe, above 80% of the existing building stock will be still standing in 2050 (European Commission 2020b). These statistics imply more land and resource use, with fewer DOI 10.1201/9781003354222-9
chances of replenishment. A proactive strategy for overcoming resource scarcity is material and component reuse, which has multiple ecological benefits including demand reduction for new resources, lower waste input to landfill sites and carbon emission reduction thanks to the extension of the lifecycle of materials in the value chain. According to a survey by Grosvenor (2021), lack of (1) existence and (2) access to standardised and certified reusable construction material are barriers to supply visibility, while lack of certainty on (1) data quality, (2) product availability and (3) procurement procedures and (4) quality of second-hand materials are barriers to demand in circular construction. This implies that the access to reusable materials for integration into the market and the latter is not wellestablished. The same observation is reported in the academic literature by Akbarieh et al. (2020); architects and designers need to be able to design with reusable materials in their modern digital platforms in order for circular construction to be realised on a larger scale. To this end, the topic of marketplaces and material Banks (Cai & Waldmann 2019) came to prominence. An example of the former is Opalis.eu – website that connects second-hand material dealers and customers through showcasing which dealer offers what product category and where. Another type of solution is websites that demonstrate brokers of waste services. However, the materials or their properties are not directly accessible through these websites.
69
Furthermore, the available reusable material platforms redistribute building parts that are not structural nor load-bearing, mainly windows, tiles, timber elements or fireplaces. Material Bank, on the other hand, is a catalyst that enables the movement of materials from their first lifecycle to another lifecycle but in a different project. Changing the game plan from linear to circular is not possible without introducing new actors and roles. A material bank, therefore, is an actor – supported by legislative frameworks and authorities – who assesses and recertifies used materials and components to reduce any misperception of risk among reusable material users, hence, offering the reusable materials to the market (Cai & Waldmann 2019). One should not mistake material bank as yet another type of marketplace since it has a regulatory status and acts as the mediator between actors. Material banks should offer Building Information Modelling (BIM)-compliant information (Akbarieh et al. 2021). BIM-based material banks spur creativity and new design approaches, which is only possible through access to digital representation and information of reusable materials and their seamless integration into the information models. However, such a material bank does not exist to our knowledge. The existing solutions only accept importing models and analysing their materials and components without any BIM-compliant export functionality. The circular or sustainable analysis results do not go back in the primary models and are often exported in nonBIM formats such as Excel or PDF (Jayasinghe et al. 2021; Madaster 2022) Having BIM-based product information available through material banks creates digital flexibility for interweaving relevant information such as material passports (Honic et al. 2019), data templates and building logbooks (Mêda et al. 2021). A number of ad-hoc solutions approached circularity in the built environment by creating earlydesign deconstruction and reuse scores for better decision making through proprietary BIM authoring
tools (Akanbi et al. 2018). However, there is no OpenBIM solution for model enrichment and exchange of circular information between information models or material banks. Therefore, the present study focuses on the OpenBIM exchange of circular information necessary for closing the material loop through the transition of reusable materials from the deconstruction phase to the material bank and the new construction activity. An overview of the conceptual scope of the study is illustrated in Figure 1. To address the gap expressed previously, this study zeros in on Linked Building Data (LBD) approaches. The new Decommissioning & Reuse Ontology (DOR) and Ontology for Environmental Product Declaration (OEPD) are suggested, which are later used in the proof-of-concept of a BIM-compliant, web-based material bank. This Semantic Material Bank (SMB) relies not only on the ontologies developed through this research study but also on data from heterogeneous sources. It acts as an aggregator of existing LBD ontologies enriched by circular information. SMB aims to create an urban scale solution to give visibility to reusable materials/components and empower various stakeholders to check if, when, where, and in what conditions what is available. The remaining sections are structured as follows. Section 2 briefly reviews the related works done in circular construction and linked open data to establish the purpose and scope. Section 3 discusses the methodology, ontology development and initial results, which is followed by an explanation on the Semantic Material Bank in Section 4. Lastly, conclusions and the future pathway are elaborated upon in Section 5. 2 BACKGROUND AND RELEVANT WORKS Most of the current circular buildings are assembled in a “Do It Yourself” (DIY) manner, meaning that if the design team (or building developers) have
Figure 1. Concept Diagram of the key aspects of the study including concerned parties and expected applications that require a range of necessary information for either digital information exchanges or physical material exchanges at different life cycle stages.
70
two purposes (1) describing the concepts and (2) defining the relationships between the concepts, in any domain of interest. Prominent work is done to capture and describe the building topology (both geometric and geographical information) such as Building Topology Ontology (BOT) (Rasmussen et al. 2017) and Building Product Ontology (BPO) (Wagner & Rüppel 2019). At the material and lifecycle level, Digital Construction Building Materials (DICM) (Karlapudi & Valluru 2021) and Digital Construction Lifecycle (DICL) (Karlapudi et al. 2021) are developed. Meanwhile, a surge of recent ontologies has captured circular information. Circular Exchange Ontology (CEO) and Circular Materials and Activities Ontology (CAMO) (Sauter et al. 2019) are the first circular ontologies developed that are related to the construction’s universe of discourse, followed by two conceptually proposed ontologies, Material Passport Ontology (MPO) (Kedir et al. 2021) and Building Circularity Assessment Ontology (BCAO) (Morkunaite et al. 2021). Table 1 summarises the description of each ontology and its main contribution. This recent rise in contribution to LBD-based circular information signals the industry’s need for such interoperable solutions. Although these ontologies contribute to the digital, circular construction discourse, they do not focus on defining circularity, reusability or recyclability on an element level. To address this identified gap, we suggest the Decommissioning and Reuse Ontology (DOR), which is further supported by the Ontology of Environmental Products Declaration (OEPD). These ontologies aim to firstly, enrich information models with circular information and, secondly, to create a link between information model and urban scale solutions, here, material bank, to overcome the reusable material visibility and accessibility gap identified previously.
access to certain reusable materials, they will use them and reshape their building design around the available reusable products. Not every actor has access to such time-sensitive information within the logical distance around them, nor having the required labour and time necessary for finding this information. To provide access to reusable material information so that relevant stakeholders can directly use such data in their information models, this study employs Linked Data and Semantic Web technologies. Semantic web technologies are the perfect solution for safeguarding and exchanging circular information that can span a whole lifecycle of a building (i.e., 50 years) before arriving at the next stakeholder’s doorstep. This is owed to high interoperability and zero-loss information exchange as well as connection to other open, contextual information (e.g., geospatial, governmental, historical or weather data) on the web (Werbrouck et al. 2021). Linked data technology empowers automatic and machine-interpretable solutions. Additionally, this technology has a high potential for preventing information obsolesce, which is a common concern between circular construction stakeholders due to the lengthy time span mentioned before. Nevertheless, the linguistic agency remains the role of domain experts. For this very reason, we looked at the existing literature on the circular economy, especially material reuse, and industrial reports and standards. An extensive literature review identified research gaps and consisted of three parts: (1) examine the state of the art of circular construction, as published in Akbarieh et al. (2020); (2) review the nexus of linked data approaches and circular information exchange and (3) review international standards, Horizon 2020 programs and industry-based reports, included but not limited to BAMB Material Passport, Interreg’s FCRBE, Product Circularity Data Sheet (PCDS) (Luxembourg Ministry of Economy 2020) , Madaster documentation, ISO 20887, and ISO 21930. This captures the best way to scientifically define circular or reuse potentials for construction buildings and products.
2.2 Literature review and qualitative industrial review for circular construction There is no fixed definition of what a reusable material is in the literature. “Intention to reuse” is stated as one possible definition criteria although uncertain (FCRBE 2021). Researchers discuss reusability without distinguishing if the perceived reusability is in preor post-consumption. However, ISO 20887 sketches reusability as “the ability of a material, product, component, or system to be used in its original form more than once and maintain its value and functional qualities during recovery to accommodate reapplication for the same or any purpose (ISO 2020).” This ability could be attributed to a component if it is designed for deconstruction (implying reuse after deconstruction), which we can say is design-based or pre-consumption reusability. This type of reusability is a promise (or intention) that should be fulfilled, until the fulfilment, it will remain a potential. If a component is damaged during the deconstruction process, it might lose its reusable status. The other type of reusability is ascertained through assessment of material bank or deconstruction/waste auditors in the post-consumption phase, where the component is
2.1 Linked building data: ontology search and assessment for circular construction Semantic Web technologies and LBD can flip current practices of storing product data in various databases and file formats by providing a common vocabulary to create, read, update and retrieve information (Pauwels et al. 2017). Having a shared, reusable, formal set of vocabularies with some consensus helps with establishing a common understanding in the project as well as between participants. That is the rationale behind the use of LBD approaches in this paper for circular construction information exchanges. LBD- and semantic web-based BIM exchanges have gained momentum over the past couple of years and a cluster of different ontologies have been developed to represent different aspects of construction products, processes, and projects. An ontology serves
71
certainly available and recertified for second reuse. A majority of scientific studies have counted on preconsumption reusability potentials for defining the circular future, whereas in reality, the majority of building stock is already built, meaning that the available reusable materials or components in future will be post-consumption reusable products that come from the existing building stock. The concept of “Circular Potential” comes from BAMB’s material passports (Heinrich & Lang 2019) and is aptly named. Even if a product is designed for deconstruction, as long as the reusability is not realised, it will remain as a potential. Likewise, a product in the exiting building stock can have circular potential. This is a concept later used in the Methodology section. An inspirational circular product is the novel demountable floor system for multi-lifecycle usages designed by Fodor & Schäfer (2021). This design shows the complexity of modelling circular information for building products. It is used in this study as a real example for ontology development in the next section. Not only is this floor system Designed for Deconstruction (DfD) according to ISO 20887 (ISO 2020), but also for multiple reuse and full reassembly. Some parts will be reused, some parts will be recycled and replaced by new similar elements. This replacement in circular discourse is called remanufacturing (Kirchherr et al. 2017). This example shows that any data modelling for circular information should go beyond saying that a product is reusable and instead focus on distinguishing which parts are reusable, recyclable, or have other End-of-Life (EoL) scenarios. This example proves that circular information enrichment in information models or material banks must consider not only the inflow and outflow of raw or secondary materials but also the relationship of materials and connections to each other. For example, for this floor Table 1.
system to be picked up by another designer in the second lifecycle, all sub-parts should be available together and, simultaneously, structurally robust. This circular structural design shows that any ontological attempt for detailed circular enrichment of products should be drilled down to the most basic part of an element and do not suffice to tag a whole product as reusable. The important issues discussed in this section are applied in the development of technological contributions of this study. 3 METHODOLOGY This section focuses on the methodology deployed in this study, starting with the definition of scope and competency questions, followed by further explanations for some of the classes and relationships used within the two proposed ontologies. The approach taken in this research is Ontology 101, based on which classes and properties are mainly developed in the ontology editor Protegé (version 5.5.0). After initial ontology modelling, the ontologies were queried via the SPARQL Protocol and RDF Query Language (SPARQL). The data used are individuals created based on BOT ontology and further enriched with the proposed vocabularies. Additional assessment should be done for final approval of the proposed ontologies. To set the goal and scope for the ontology and based on the findings of the literature review, it is clear that the existing circular ontologies mostly focus on the circular performance of a product or building in the early design stage whereas we carefully select our goal to enrich building elements with circular information independent of the in/out flow type of constituent materials – something that BCAO
Overview of the exiting ontologies for construction product or circular information enrichment.
Abb.
Name
Source
Description and aim
BOT
Building Topology Ontology Building Product Ontology
Rasmussen et al. 2017
Models building level topology. It is well aligned with IfcOwl Ontology. Models non-geometric product parts and connections. It can be used beyond construction products. Models materials and material properties
BPO DICM DICL CEO CAMO MPO BCAO
Digital Construction Building Materials Digital Construction Lifecy-cle Circular Exchange Ontology Circular Materials and Activ-ities Ontology Material Passport Ontology Building Circularity Assess-ment Ontology
Wagner & Rüppel 2019 Karlapudi et al. 2021 Karlapudi et al. 2021 Sauter et al. 2019 Sauter et al. 2019 Kedir et al. 2021 Morkunaite et al. 2021
72
Models BLS or Building LifeCycle Stages based on BS EN 16310:2013 Models activities and in/out flows of materials in different circu-lar projects Models circular and procedural activities, textile materials and by-products. Models actors crucial for circular construction and lifecycle phases of buildings based on material passports Models early-design stage data for calculating Circular Indica-tors based on Ellen MacArthur Indicator
property DOR:hasCircularPotentialIndicator are suggested. The latter enables a wide range of users to create applications that require a scoring system for circularity measurement and is based on the BAMB material passport suggestions. DOR ontology includes other classes and relationship, which we do not focus on them in the scope of this paper. The classes and relationships of DOR are aligned with BPO ontology since it is possible to define if a product is an assembly and if different assembly sub-parts have recycling or reuse potential independent of each other and not on a product level as discussed in section 2.2 and work of Fodor & Schäfer (2021). BPO:Product is aligned with BOT: Element, which is then linked to BOT:Building that could be linked to other ontologies of geo-queries. Figure 2 elaborates the conceptual relationship between DOR and its alignment with primarily BPO and further BOT, DICM and BCAO ontologies. Furthermore, to enrich post-consumption elements with such information, the DOR ontology has actor classes such as Material Bank Agent who can recertify (especially structural elements) for reuse potential among other circular potential possibilities. While some inspections can be done visually on-site or off-site, or virtually through modern technologies such as reality capturing, some structural elements would need to be further tested via material bank (Akbarieh et al 2021). ISO 21930:2017 (ISO 2017), which describes Environmental Product Declaration (EPD) requirements for construction products and services, is used as a reference for OEPD ontology development. This ontology defines the necessary classes and relationships for all construction products and services although Product Category Rules (PCRs) are not modelled. EPD aims to facilitate the sustainable comparison of similar products in the market through standardised, verifiable and consistent data provided by the manufacturers. The classes are relationships addressed in this ontology help to make EPDs machine-readable on one hand and enable the discovery and communication of standardised LCA data over the web on the other hand. OEPD, partially shown in blue classes in Figure 2, can be used not only in the construction industry but also in other business sectors. However, the current use-case of the OEPD is only for the construction sector. It is a simple ontology. It supports DOR ontology by further enriching products with harmonised sustainable data. Another benefit of OEPD ontology could be the integration of carbon emissions and circular potential information of products in one place. This direction will be explored in future steps of this project. Finally, Figure 3 shows two code snippets for successfully querying individuals populated with DOR and OEPD ontologies. In this paper, only the concept and certain classes and relationships are discussed. Both ontologies have further space for improvements as well as getting aligned with other ontologies for broader applications.
ontology tackled already. Our objective is the enrichment of models before or after deconstruction with circular information for the purpose of exchange of circular information between actors and projects through a material bank. However, expressing how a building element is reusable is challenging. This is because the circular economy is a rapidly evolving domain and different experts have their own circular vision or “picture of the world.” Therefore, each actor has a perspective that could be considered truth, specifically because there is no universal method of measuring or quantifying circularity at the product or building level, nor a universally agreed-upon unit of measurement unlike the domain of energy and sustainability, which uses Kg CO2 -eq as an example. The suggested ontology in the next part acknowledges this uncertainty and multi-perspective circular future, based on which it suggests a set of classes and relationships that enable a trans-perspective dialogue between circular actors. This helps different actors reach a mutual understanding. Circularity and circular indicators are subjective and need the judgment of an inspector.Yet, we included them in our model to serve a wider variety of users, who wish to score building components. The following competency questions portray the core functionality requirements of the Decommissioning & Reuse Ontology (DOR): 1. What elements are reusable? 2. What elements are reusable at time XX and place YY (or in ZZ Km from my location)? 3. What elements have a circularity indicator lower than AA? 4. What structural products are recertified for reuse by the material bank? 5. Which buildings are supposed to be decommissioned at time BB? To add further standardised lifecycle emission information to the products existing in the building stock, one last competency question is posed for the development of the OEPD ontology: 6. What products have an EPD declared module of C1 (i.e., deconstruction) with a Global Warming Potential (GWP) lower than CC kg of CO2 -eq? To answer these competency questions, one must consider that even if a product is designed for deconstruction or reuse, as long as this reusability is not realised, it will remain as a potential as discussed in section 2.2. Therefore, it is reusable until it gets damaged in the deconstruction process. The same element can now be in the state of remanufacturing or recycling (or both, depending on the demand in the market and the supporting regulations). Moreover, different researchers have different approaches to how circularity should be declared, some are satisfied with having a percentage indicator for ranking EoL state of a products (Durmisevic et al. 2021), while others are only interested in using the circular products in their project with no need for any circular indicator. To capture all these circular perspectives, an object property DOR:hasCircularPotential as well as a data
73
Figure 2. An abridged conceptual model of the Decommissioning and Reuse Ontology (DOR) in green, and the Ontology for Environmental Product Declarations (OEPD) in blue.
4 SEMANTIC MATERIAL BANK In the last part of this study, a semantic web-based urban mining solution is developed that empowers large-scale reuse of materials and addresses the gaps identified in section 1. The aim of the Semantic Material Bank (SMB) proof-of-concept is to connect the circular information of materials from the decommissioning stages to early design stages in order to facilitate urban mining and close the material loop. Reportedly, a problem that hinders the uptake of reusable design is the lack of awareness of the existence of the reusable materials at the time of new design and within the logical vicinity. Also, a burden for designers is that they would need to search on their own to see if a building is being demolished/decommissioned in which case, what materials it contains and if they are suited for reuse. Even if the two previous issues are solved, there is no digital representation of the available reusable elements nor any recorded and trusted information concerning their quantity and quality. Through SMB, even before a building is deconstructed, one can see the potential availability of reusable materials. This is our vision of a technical urban mining solution. Besides, the influx of construction products and by-products in and out of any geographic region can be investigated through SMB, which will bring further transparency to the construction value chain. Calculating circularity scores are beyond the current scope of the SMB. Figure 4 demonstrates the overall architecture of this proof-of-concept, which currently serves three sets of stakeholders: (1) building owners/ developers/facility managers; (2) material bank authority and
Figure 3. Two queries for the proposed DOR (Query one), and the OEPD (Query two) ontologies.
agents; (3) building designers and investors. Each of these stakeholders interacts with one part of the material bank. In this proof-of-concept, the following technologies are used: JavaScript, NodeJS, ExpressJs,
74
to render the geometries of these elements along with their circular information in SMB and use them in their BIM-supported designs. 5 CONCLUSIONS The circular economy demands a socio-cultural shift in pre- and post-consumption of products by changing economic workflows and expectations. However, circularity is effective when done on a large scale in a given urban level. Despite the desire of construction stakeholders in using reusable and circular products, barriers such as lack of visibility, access, guaranteed quality and interoperable data exist. To enable the exchange of the circular information and overcome the barriers, this study uses semantic web technologies to enrich information models with reusable (in general, circular) metadata through which the data from one component in the first lifecycle can be used, enriched and reused in the second lifecycle. This facilitates the reusability of the associated physical material in the real world. The main contribution of this study is the development of two ontologies that support the transition of construction elements from the deconstruction phase to the material bank for recertification, and later to the new construction project. Additionally, a proofof-concept showcases how new circular sub-cycles benefit from a web-based linked data approach for managing reusability information and exposing them to building designers or investors. The Decommissioning and Reuse Ontology (DOR) is proposed in a trans-perspective manner to support a circular future with different circular information enrichment possibilities. Reusability has different meanings to different stakeholders. DOR captures this difference and builds a holistic solution around it. In the scope of this paper, only circular potentials were discussed. In future steps, the ontology will be further developed. The Ontology for Environmental Product Declaration (OEPD) aims to make Environmental Product Declarations (EPD) machine-readable on one hand and integrates standardised lifecycle information with construction products offered by manufacturers on the other hand. Semantic Material Bank (SMB) proof-of-concept is developed as an OpenBIM-based reusable material management and discovery platform. SMB is a digital urban mining solution with which stakeholders can evaluate the availability and status of reusable and recyclable elements at the urban level. An advantage of this proof-of-concept is that users can upload and query the models in SMB independent of the BIM-authoring tools. Moreover, they should be able to download such information from SMB in an LBDbased format and have a full feedback loop with circular information. Not only does SMB employs the proposed ontologies, but also can connect them to a broader range of available information over the web. In the next steps, alignments between DOR, BPO and BOT will be further improved, and the geographical and geometrical support will be considered.
Figure 4. Architecture of the Semantic Material Bank.
MongoDB, CSS and HTML and GraphDB. The proposed architecture provides digital material bank platforms that share the same infrastructure in order to be interoperable for maximising the circulation of reusable materials. Yet, it offers enough flexibility to adjust to the legal and socio-economic requirements of every urban area. This will pave the way for a federation of material banks with different owners and server instances. SMB must be populated with building information models enriched by DOR and OEPD ontologies. Building owners/facility managers ought to upload the model-generated triples to the triple store provided in the back-end knowledge graph. Then, material bank agents will have a management tab that enables them to enrich the existing graph-based models further with the associated results of structural/chemical/sustainable assessments and sort out materials into reusable, recyclable, remanufactured and waste (discarded). In this way, the responsibility of deciding if something is reusable in the current building stock can be regulated and the information would be certified and standardised. The next set of users, i.e., designers and building developers, visit SMB and search for reusable elements that match their design. It will be up to them how they decide to employ those elements in their design (Akbarieh et al. 2020). Currently, users can only retrieve circular information and check which elements are reusable, remanufacturing or recyclable. This search function (linked to SPARQL queries) could be an attractive feature for building developers, investors and policymakers to see the big picture of available materials on the urban level. Users do not need to directly interact with the queries as there are pre-defined queries that they can use to get the information they seek.The future ambition is to enable users
75
ACKNOWLEDGEMENTS
Honic, M., Kovacic, I., Rechberger, H. 2019. Concept for a BIM-based Material Passport for buildings. IOP Conference Series: Earth and Environmental Science; 225. ISO, 2020. ISO 20887:2020 Sustainability in buildings and civil engineering works — Design for disassembly and adaptability — Principles, requirements and guidance. ISO, 2017. ISO 21930:2017 Sustainability in buildings and civil engineering works — Core rules for environmental product declarations of construction products and services. Jayasinghe, B., Paulus, M., Waldmann, D. 2021. Environmental impact assessment of buildings based on building information modelling. CIB W78 & LDAC; Luxembourg, 11–15 October 2021. Karlapudi, J., Valluru, P. 2021. Digital Construction Materials. URL https://digitalconstruction.github.io/Materials/ v/0.5/ (accessed 6.30.22). Karlapudi, J., Valluru, P., Menzel, K. 2021. Ontology approach for Building Lifecycle data management. ASCE International Conference on Computing in Civil Engineering; Florida, 12–14 September 2021. Kedir, F., Bucher, D.F., Hall, D.M. 2021. A Proposed Material Passport Ontology to Enable Circularity for Industrialized Construction. European Conference on Computing in Construction; Dublin, 26–28 July 2021. Kirchherr, J., Reike, D., Hekkert, M. 2017. Conceptualizing the circular economy: An analysis of 114 definitions. Resources, Conservation & Recycling 127: 221–232. Luxembourg Ministry of Economy. 2020. Guidance for “light” Product Circularity DataSheet (v3.2). URL https://pcds.lu/ (accessed 6.30.22). Madaster, 2022. Platform. Madaster. URL https://madaster. com/platform/ (accessed 6.30.22). Mêda, P., Hjelseth, E., Calvetti, D., Sousa, H. 2021. Enabling circular construction information flows using data templates – conceptual frameworks based on waste audit action. European Conference on Computing in Construction; Dublin, 26–28 July 2021. Morkunaite, L., Al-Naber, F.H., Petrova, E., Svidt, K. 2021. An Open Data Platform for Early-Stage Building Circularity Assessment. CIB W78 & LDAC; Luxembourg, 11–15 October 2021. Pauwels, P., Zhang, S., Lee, Y.-C. 2017. Semantic web technologies in AEC industry: A literature overview. Automation in. Construction. 73: 145–165. Rasmussen, M.H., Pauwels, P., Karlshøj, J., Hviid, C. 2017. Proposing a Central AEC Ontology That Allows for Domain Specific Extensions; Lean and Computing in Construction Congress; Heraklion, 4–7 July 2017. Sauter, E., Lemmens, R., Pauwels, P. 2019. CEO & CAMO ontologies: a circulation medium for materials in the construction industry. International Symposium on Life-Cycle Civil Engineering (IALCCE); Ghent, 28–31 October 2018. Wagner, A., Rüppel, U. 2019. BPO: The Building Product Ontology for Assembled Products. Linked Data in Architecture and Construction; Lisbon, 19–21 June 2019. Werbrouck, J., Pauwels, P., Beetz, J., Mannens, E. 2021. Data Patterns for the Organisation of Federated Linked Building Data. CIBW78 & LDAC; Luxembourg, 11–15 October 2021.
A. Akbarieh is funded by the European Regional Development Fund (2014-2020), with grant agreement No 2017-02-015-15 for the Eco-Construction for Sustainable Development project. J. O’Donnell received financial support for his contributions in part by a research grant from the European Union’s Horizon 2020 research and innovation programme through the CBIM-ETN, which is funded under the Marie Skłodowska-Curie grant agreement No 860555.
REFERENCES Akanbi, L. A., Oyedele, L. O., Akinade, O. O., Ajayia, A. O., Davila Delgadoa, M., Bilal, M., Bello, S. A. 2018. Salvaging building materials in a circular economy: A BIM-based whole-life performance estimator. Resources, Conservation and Recycling 129: 175–186. Akbarieh, A., Jayasinghe, L.B., Waldmann, D., Teferle, F.N. 2020. BIM-Based End-of-Lifecycle Decision Making and Digital Deconstruction: Literature Review. Sustainability 12(7): 2670. Akbarieh, A., Schäfer, M., Waldmann, D., Teferle, F.N. 2021. Post-Urban Mining Automation and Digitalisation for a Closed-Loop Circular Construction. CIB W78 & LDAC; Luxembourg, 11-15 October 2021. Build Reuse, 2022. About Build Reuse. Build Reuse. URL https://www.buildreuse.org/about (accessed 6.30.22). Cai, G., Waldmann, D. 2019. A material and component bank to facilitate material recycling and component reuse for a sustainable construction: concept and preliminary study. Clean Technologies and Environmental Policy 21(10): 2015–2032. Durmisevic, E., Guerriero, A., Boje, C., Domange, B., Bosch, G. 2021. Development of a conceptual digital deconstruction platform with integrated Reversible BIM to aid decision making and facilitate a circular economy. CIB W78 & LDAC; Luxembourg, 11–15 October 2021. European Commission, 2020a. A new circular economy action plan for a cleaner and more competitive Europe. Brussels. European Commission, 2020b. A Renovation Wave for Europe – greening our buildings, creating jobs, improving lives. Brussels. European Commission, 2019. The European Green Deal (No. COM (2019) 640 final). Brussels. European Commission, 2016. EU Construction & Demolition Waste Management Protocol. Brussels. FCRBE, 2021. Product or waste? Criteria for reuse. Fodor, J., Schäfer, M. 2021. Behavior of downstand simply supported steel-concrete composite beam applying friction based demountable shear connection. Eurosteel; Sheffield, 1–3 September 2021. Grosvenor, 2021. Accelerating material re-use. Heinrich, M., Lang, W. 2019. Materials Passports – Best Practice. Munich: Technische Universität München.
76
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
NLP-based semantic model healing for calculating LCA in early building design stages K. Forth, J. Abualdenien & A. Borrmann Technical University of Munich, Munich, Germany
ABSTRACT: To limit the global warming, the environmental impacts of new buildings need to be quantified and optimized already in the early design stages. Semantically rich models, such as Building Information Modeling (BIM), facilitate deriving consistent and automated quantity take-offs of the relevant components for calculating whole building life cycle assessments (LCA). A particular challenge is that early-stage BIM models typically lack stringency in terms of component modeling and material classification. Hence, this paper presents a methodology for enriching knowledge and characteristics from the coarse information available at the early design stages, in a process denoted as semantic model healing. In more detail, the proposed method employs different Natural Language Processing (NLP) strategies to increase the performance of automatically matching materials defined in a BIM model to a knowledge database with environmental indicators of commonly used components, facilitating a seamless LCA in the early stages of design.
1 INTRODUCTION AND MOTIVATION In order to reach the international goals of the Paris Agreement and improve the ecological impacts of new buildings, life cycle assessments (LCA) are an established method to calculate several environmental indicators along the whole life cycle of buildings. According to the United Nations, manufacturing of materials for building construction cause 11% of the global energy-related carbon emissions (Abergel et al. 2017). Accordingly, a careful LCA of the different design options is required in order to identify the main drivers and optimize the building design accordingly. Up to now, LCA has been performed mainly manually, which is time-consuming, especially quantifying the building components and mapping them to LCA databases (Llatas et al. 2020). Building Information Models (BIM) combine geometry and semantics and thus facilitate deriving consistent and automated quantity take-offs of the relevant components for calculating whole building life cycle assessments (LCA). Additionally, using and enriching the semantic information of e.g. materials provides a great potential to completely automate the calculation process (Safari & AzariJafari 2021). However, in early design stages, essential decisions are taken that have a significant impact on the carbon footprint of the final building design. At the same time, the early design stages are characterized
DOI 10.1201/9781003354222-10
by high uncertainty due to the lack of information and knowledge, making a holistic and consistent LCA for supporting design decisions and optimizing performance challenging (Schneider-Marin et al. 2020). In more detail, in the “rough” BIM models of early design stages, materials are rather defined by material groups than by specific types (e.g. “concrete” rather than “concrete C20/25”), which leads to a range of possibilities for each material group. Furthermore, several materials or component layers might not intentionally be defined yet, which opens the opportunity to explore and compare different design options. This paper aims to answer the following research question to facilitate a reliable LCA in the early stages: Is automated semantic healing of “rough” BIM models possible that allows assigning correct element types and materials to the respective model elements?
2 STATE OF THE ART 2.1 LCA in early design stages using 3D models For a whole building LCA in early design stages, there are different established approaches. One approach is using benchmarks, which are derived from already finished projects and are then transferred to early design stages. Gantner et al. introduced this approach, based on several design stages, where
77
subtypes and enrich semantics of IFC model entities (Koo et al. 2021). Costa & Sicilia propose a methodology of transforming and mapping building data from BIM models using Semantic Web technologies for an automated and flexible exchange with other software, e.g. whole building energy simulation (Costa & Sicilia 2020). Wu et al. proposed a natural-language-based retrieval engine for BIM object database (Wu et al. 2019). Their use case is to mapping building components in BIM object databases with a higher accuracy than with keyword-based methods. Reitschmid proposed a mapping algorithm of IFC materials to the LCA database Ökobaudat based on tokenization of material names and a distinct mapping or via Levenshtein distance (Reitschmidt 2015). Nevertheless, no Natural Language Processing (NLP) model was used and also no integration to elementspecific mapping was proposed. Locatelli et al. investigated in their scientometric analysis the synergies between NLP and BIM (Locatelli et al. 2021). Beside the field of Automatic Compliance Checking, they identified also Information Retrieval from BIM models and Information Enrichment of BIM objects as a further field of relevant application. Nevertheless, an automated mapping of LCA and IFC data on element level has not been developed yet (Safari & AzariJafari 2021).
different input information are needed (Braune et al. 2018). The early design stages are subdivided into occasion and initialization phase, where building types and general systems are decided, and design and approval planning, where element systems are decided. Nevertheless, with this approach, optimization of LCA using different design options is difficult because they don’t include all later details but rather benchmarks. On the other hand, Hollberg suggested a parametric approach, based on LCA profiles for several construction types and using the Visual Programming Language Grasshopper with Rhino (Hollberg 2016), where optimizations can be automatically performed. However, the modell intup of the calculation depends only on geometric and doesn’t include semantic information, same as with benchmarks. 2.2 BIM-LCA integration Wastiels & Decuypere classified five different integration workflows for calculating LCA using BIM models (Wastiels & Decuypere 2019). While the first three still require manual work, mainly for mapping IFC element information to LCA profiles, the fourth workflow is based on plugins of LCA software in BIM authoring tools. The fifth workflow by Wastiels & Decuypere includes a BIM object enrichment with LCA profiles. Rezaei et al. developed a method based on Revit models to calculate LCA in early and detailed design stages (Rezaei et al. 2019). LCA profiles on element levels are detailed into layers and material options, but the mapping to match Revit and LCA database assemblies is carried out manually. Nevertheless, the LCA results are given in ranges, due to uncertainties in early design stages, and not as total result. Eleftheriadis et al. proposed an BIM-embedded LCA approach focusing on structural design alternatives in early design stages (Eleftheriadis et al. 2018). However, they do not consider all life-cycle modules (only A1-A3) and is also based on Autodesk Revit. Horn et al. proposed an integration approach based on open BIM using Industry Foundation Classes (IFC) as data format (Horn et al. 2020). With the help of Information Delivery Manuals (IDM) and Model View Definitions (MVD), LCA for several level of development of building design are realized, also for early design stages.
3 SEMANTIC MODEL HEALING The semantic model healing process is part of a bigger framework, which we previously proposed (Forth et al. 2021). In the paper at hand, the focus is on how NLP techniques help to heal the BIM model semantically for the use case of LCA in early design stages. Typically, design decisions are finally decided by the client and not the architect, hence, the proposed methodology is leveraging open BIM data models. The proposed healing process is based on NLP, using different strategies to increase the performance of mapping materials from a rough BIM model to a knowledge database with environmental indicators of commonly used components. The knowledge database contains all missing information for LCA and has different levels of detail for a range of several potential design options of components, elements, and materials, including their dependencies. The semantic model healing process happens, when the incomplete IFC element data are matched with the detailed LCA knowledge database (LKdb). First in section 3.1, the structure of the LKdb is introduced, before in section 3.2 the method for matching is described. Section 3.3 investigates multiple NLP techniques such as GermaNet, spaCy, or BERT, before the following section 4 evaluates the performance of state-of-the-art deep learning models.
2.3 BIM data extraction & mapping methods Extracting data from IFC models and map those to different data structures or ontologies of the chosen use case is a complex task. Several approaches for different use cases have been developed recently. Koo et al. explored the use of 3D geometric deep neural networks to distinguish BIM element
78
Figure 1. Structure of the LCA Knowledge Database as UML schema.
UUIDs according to German cost groups using DIN 276 (Stenzel 2020). All material options have a name and classification according to DIN 276, which is derived by the German name in Ökobaudat. They are called options, as they are the most detailed level of a design option for LCA. Further entries are the UUID, included Modules and the encoded NLP vectors of the name (spans and tokens). Each material option is related to a material category, which is also stored in Ökobaudat. There are three different levels of categories, but for LKdb only the last level of categories is used, as it groups the datasets of the relevant material options. The category level is extended with another external input describing the service life of building components (BBSR 2017). For these, the IDs are mapped once to the corresponding material category for each classification. Also, for the material categories, the name and the classification are the keys and the encoded NLP vectors of the name (spans and tokens) are also stored. In the next level, material categories and options are used for setting up element layers. Different elements can consist of the same material category or option. The element layer and the element have default maximum and minimum thicknesses and are also classified to the third level of the German cost group classification according to DIN 276.The element layers can have different mixtures ratios, as e.g. reinforced concrete consists of two material inputs: concrete and reinforcement steel. Each element layer has a unique material position, so that elements consist of one or more layers with different material position orders.
3.1 LCA Knowledge Database The aim of the LCA Knowledge Database is to store all detailed information of typical building elements including all relevant information for calculating a holistic LCA. After the matching of IFC materials to material options in the LKdb and selecting the most similar element, all relevant data are queried for calculating the LCA. As shown in Figure 1, the general structure of the proposed LKdb consists of three different levels: element, material category and material option. As the LCA database, Ökobaudat was chosen (BBSR 2021), because it consists of more than 1400 datasets specifically of building products and is the most used LCI database in Germany. The Ökobaudat datasets consist of a Universally Unique Identifier (UUID) and the relevant life cycle modules. All datasets from Ökobaudat consist of several environmental impact categories, such as Global Warming Potential (GWP), Acidification Potential (AP), Eutrophication Potential (EP), Ozone Depletion Potential (ODP), Photochemical Creation Potential (POCP), Primary Energy Renewable (PERE) and many more. As the quality of some datasets in Ökobaudat is lacking such as missing data of End-of-Life (EoL) module, generic EoL scenarios have to be mapped manually from Ökobaudat. In this case, for each material option two UUID are mapped, one for the LCA Modules A1-A3 and another for the missing Endof-Life scenario (C3/C4/D). Stenzel conducted this manual mapping as well as a classification of all
79
Figure 2. Material-specific similarity analysis of IFC elements and LCA Knowledge Database using GermaNET, spaCy and BERT.
as robust as possible. For this reason, if there is no IFC material available for specific elements, the element name itself will be evaluated for NLP similarity analysis. Furthermore, only the elements corresponding to the classification are filtered and compared. The material names from the IFC elements as well as from the LCA Knowledge Database can be encoded either as whole expressions/spans or be tokenized beforehand.
3.2 Method for matching materials For matching the elements and materials of an IFC model to the LCA Knowledge Database, we propose employing NLP techniques to measure “semantic similarity” (Forth et al. 2021). Measuring the semantic similarity between the IFC element’s material information and the material names of the database involves converting the text of every material type to a vector representation. A vector is a list of numerical values, where the combination of them represents the overall meaning. When comparing two material names, the similarity between vectors A and B can be measured using the cosine similarity, while n is the dimension of the vector: n cos (θ) =
n i=1
Ai Bi n 2
i=1
Ai
i=1
3.3 NLP techniques This section introduces the three NLP techniques GermaNET, spaCy and BERT. In the next section, the performance of these techniques is evaluated and compared for measuring the similarity between the different material types.
(1) Bi2
3.3.1 GermaNET GermaNET is a Lexical-Semantic Net specialized for the German language, also known as the German version of the Princeton WordNet (Hamp & Feldweg 1997; Henrich & Hinirchs 2010). GermaNET relates German nouns, verbs and adjectives semantically by grouping lexical units that express the same concept into synsets (set of synonyms) and by defining semantic relations between these synsets. It can be represented as a graph, whose nodes are synsets and edges represent the semantic relations (Navigli & Martelli 2019). Therefore, the similarity is not measured using cosine similarity, but graph-related shortest path similarity, which is equal to the inverse of the shortest path length between two synsets. There are other pathrelated similarity analyses such as Wu-Palmer similarity or Leacock-Chodorow similarity, which were not considered in this paper.
To compute the vectors for similarity analysis, this paper investigates multiple NLP techniques and evaluates the performance of state-of-the-art deep learning models such as GermaNet (Hamp & Feldweg 1997; Henrich & Hinirchs 2010), SpaCy (Honnibal & Montani 2017), or BERT (Devlin et al. 2018), which will be introduced in the following sections. Figure 2 shows the general workflow for matching IFC elements to the previously introduced LCA Knowledge Database. Generally, all IFC elements consist of an element name, a required classification and IFC materials, including their names. As the LCA knowledge database is based on the German ÖKOBAUDAT and the German cost group classification system according to DIN 276, it is also required as classification of the IFC elements. The matching method is developed with the aim to perform
80
3.3.2 spaCy SpaCy is a pre-trained neural network model which offers state of-the-art accuracy in multiple languages (Honnibal & Montani 2017). Its large German model (de_core_news_lg) includes 500k unique vectors in its corpus and represents every word or expression with a vector of 300 dimensions. As sources for training data, existing corpi were used such as e.g. TiGer Corpus (Brants et al. 2004).
the synsets need to be analyzed before analyzing the shortest path similarity.
3.3.3 BERT BERT stands for Bidirectional Encoder Representations from Transformers and was released by Google in 2018 (Devlin et al. 2018). Transformers-based pretrained models are currently state-of-the-art and are capable of solving a different set of tasks as they “can represent the characteristics of word usage such as syntax and how words are used in various contexts” (Locatelli et al. 2021). Nevertheless, BERT represents each word or expression with a vector of 768 dimensions, which is significantly higher compared to spaCy, making the similarity calculation more time-consuming.
Figure 3. Synset identification rate of material pairs with GermaNET.
After the tokenization of the IFC material names, material options and their related material categories, synsets were identified to calculate the shortest path similarity. Nevertheless, synsets could not be identified for every token set, so that not for all 59 pairs synsets could be identified. As shown in Figure 3, only for 20,3% of the material category tokens and 40,7% of the material option tokens, a pair of synsets could be identified.
4 EXPERIMENTS & RESULTS In the following sections, first the case study is shortly introduced. Afterwards, the performance results of three different NLP techniques based on a manual matching is compared. Last, one IFC element is chosen to be prototypically matched and the LCA calculation is conducted using the LKdb and compared to conventional workflow results. 4.1 Case study For comparing the three different NLP techniques and their performance of their workflows, a real-world office building was chosen as a case study. This realworld project guarantees that the material naming is not optimized but according to current industry standards, so that the matching performances are tested under realistic conditions. In total, the case study office model consists of 2110 individual elements, which are summed up to 133 unique elements when grouped by element type. Those consists of 59 unique IFC materials, which were manually matched to LCA material options and categories from the LKdb, as a ground truth.
Figure 4. Shortest path similarity of NLP material using GermaNET.
Nevertheless, the shortest path similarity of the identified pairs of synsets show promising results (Figure 4). The mean of the similarity of material option tokens is 88,9% and of the material category tokens even 95,2%, both with little deviation. However, including the little synset identification rate of both, material options and material categories from the LKdb, the total similarity is very low and not sufficient for being used in the proposed matching methodology.
4.2 Results The following results of each NLP technique performance are based on the 59 pairs of IFC materials and the matched LCA material options and categories from the LKdb based on Ökobaudat.
4.2.2 spaCy For the results of spaCy and BERT, the similarities of tokens and whole spans of the material options and material categories are compared according Figure 2. As shown in Figure 5, the ranges of the cosine similarity of all different comparisons differ a lot. Generally, the similarity of IFC materials to the material option spans have the worst performance with
4.2.1 GermaNET As the workflow of the GermaNET differs from the other two NLP techniques, the identification rate of
81
Figure 5. Cosine similarity of NLP materials using spaCy.
Figure 6. Cosine similarity of NLP materials using BERT.
the mean being at 16,3%. This means, that the similarity for most matched pairs using material option spans and spaCy is only very little. The tokenization improves the performance of matching the material performances up to a mean of 49,2%. Also, the spans of the material categories are much better (mean at 40,5%). The tokenization of the material categories improves the performance results up to 50,2%. As an additional, performance result, the maximum similarity of all options (material option spans and tokens, as well as material category spans and tokens) is calculated. Its mean is 63,0%, but also the quartile ranges improved, compared to all other ranges. In general, the results are not sufficient, but show a promising strategy of getting the maximum similarity of every option.
cosine similarity of material option spans and material category spans are improving the results. Furthermore, the tokenization of both material options and material categories, as well as choosing the maximum similarity of every calculated option improved the result ranges significantly. However, the deviations in ranges were substantial and are generally too low, so that a further consideration for implementation was not investigated. Finally, BERT showed the most promising results. Low deviations of the result ranges and high cosine similarity of all strategies lead to a further implementation of the matching approach. Nevertheless, due to its large vectors with 786 dimensions, the calculation time is significantly higher than with spaCy and needs to be considered for further optimization.
4.2.3 BERT When evaluating BERT, the same similarity results are calculated as previously shown with spaCy also using cosine similarity. Figure 6 is showing the results as ranges of the material option spans and tokens and material category spans and tokens. Generally, all result ranges differ much less compared to the results using spaCy, which means that for all pairs more satisfying performances can be reached using BERT. Additionally, all means are between 79,3% (material category spans) and 86,4% (material option tokens). Also, the strategy of getting the maximum similarity of every option is improving the general promising results (mean 87,6%). Also, the minimum values of each result ranges show that BERT, generally performs much better than spaCy.
4.4 Prototypical element matching and calculation of LCA results Next, the proposed matching methodology was prototypically tested using LKdb and BERT. The LKdb was filled with example elements and element layers, based on domain knowledge and the structured Ökobaudat. As a test element, the exterior wall “Basiswand: STB 250_außen” from the case study was chosen (cost group 331, single material “Ortbeton”, total area 415.32 m2 , layer thickness 20 cm). The final matching shows if the highest cosine similarity was derived from a material category or the material option. In this test case, it is the material category with a cosine similarity of 85,4%. Therefore, the matched element within the cost group 331 is “Stahlbeton”, so also the reinforcement steel is included beside the range of different concrete options. As a comparison for manual matching and manual calculation, the software eLCA is used (BBSR). Only specific datasets of the Ökobaudat can be used. Therefore, the assumed LCA dataset is “Transportbeton C20/25”. For simplicity, GWP [kg CO2-eq./a] is chosen as the comparing indicator with a lifespan of 50 years. The results are shown in Figure 7: Besides the necessary effort and knowledge of the manual matching, the accuracy of the results is
4.3 Summary Generally, all three NLP techniques could be applied to the case study. Although GermaNET showed promising results in the ranges of shortest path similarity, the identification rate of synsets was too low. Therefore, a further implementation to the proposed matching methodology was not pursued further in this paper. The second tested NLP technique, spaCy, showed that different strategies of calculating the
82
REFERENCES Abergel, Thibaut; Dean, Brian; Dulac, John (2017): Global Status Report 2017. Hg. v. United Nations Environment Programme. International Energy Agency (IEA) for the Global Alliance for Buildings and Construction (GABC). BBSR: eLCA. available online under https://www. bauteileditor.de/, last checked 29.12.2021. BBSR (2017): Nutzungsdauern von Bauteilen – Informationsportal Nachhaltiges Bauen. Hg. v. Bundesinstitut für Bau-, Stadt- und Raumforschung. Online verfügbar unter http://www.relaunch-nb.online-now.de/index.php?id=91 &L=0, zuletzt geprüft am 28.12.2021. BBSR (2021): ÖKOBAUDAT. available online under https://www.oekobaudat.de/datenbank/browser-oekobau dat.html, last checked 06.04.2021. Brants, Sabine; Dipper, Stefanie; Eisenberg, Peter; HansenSchirra, Silvia; König, Esther; Lezius, Wolfgang et al. (2004): TIGER: Linguistic Interpretation of a German Corpus. In: Res Lang Comput 2 (4), S. 597–620. DOI: 10.1007/s11168-004-7431-3. Braune, Anna; Ruiz Durán, Christine; Gantner, Johannes (2018): Leitfaden zum Einsatz der Ökobilanzierung. Costa, G.; Sicilia, A. (2020): Alternatives for facilitating automatic transformation of BIM data using semantic query languages. In: Automation in Construction 120, S. 103384. DOI: 10.1016/j.autcon.2020.103384. Devlin, Jacob; Chang, Ming-Wei; Lee, Kenton; Toutanova, Kristina (2018): BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Online verfügbar unter https://arxiv.org/pdf/1810.04805. Eleftheriadis, S.; Duffour, P.; Mumovic, D. (2018): BIM-embedded life cycle carbon assessment of RC buildings using optimised structural design alternatives. In: Energy and Buildings 173, S. 587–600. DOI: 10.1016/j.enbuild.2018.05.042. Forth, Kasimir; Abualdenien, Jimmy; Borrmann, André; Fellermann, Sabrina; Schunicht, Christian (2021): Design optimization approach comparing multicriterial variants using BIM in early design stages. In: Proceedings of 38th International Symposium on Automation and Robotics in Construction (ISARC 2021), S. 235–242. DOI: 10.22260/ISARC2021/0034. Hamp, Birgit; Feldweg, Helmut (1997): GermaNet – a Lexical-Semantic Net for German. In: In Proceedings of ACL workshop Automatic Information Extraction and Building of Lexical Semantic Resources for NLP Applications, S. 9–15. Henrich, Verena; Hinirchs, Erhard (2010): GernEdiT – The GermaNet Editing Tool. In: Proceedings of the Seventh Conference on International Language Resources and Evaluation (LREC 2010). Valletta, Malta, pp. 2228-2235. Hollberg, Alexander (2016): A parametric method for building design optimization based on Life Cycle Assessment. PhD thesis. Bauhaus-Universität Weimar. Honnibal, Matthew; Montani, Ines (2017): spaCy 2: Natural Language Understanding with Bloom Embeddings, Convolutional Neural Networks and Incremental Parsing. In: To appear 7(1), S. 411–420. Horn, Rafael; Ebertshäuser, Sebastian; Di Bari, Roberta; Jorgji, Olivia; Traunspurger, René; Both, Petra von (2020): The BIM2LCA Approach: An Industry Foundation Classes (IFC)-Based Interface to Integrate Life Cycle Assessment in Integral Planning. In: Sustainability 12(16), S. 6558. Koo, Bonsang; Jung, Raekyu; Yu, Youngsu (2021): Automatic classification of wall and door BIM element
Figure 7. LCA result (GWP) of test element comparing manual matching with BERT matching and LKdb.
different. While the result with manual matching and eLCA is a single value, the proposed methodology returns a material category, considering the uncertainty of material choice in the early design stages. Therefore, the LKdb returns a range of material options for LCA calculation. Because the matching does not take place on a material level, but on an element level, the reinforcement steel is getting included in the LKdb element of reinforced concrete, which gives more realistic results. This is the reason, why the range of results is more than 200 kg CO2-eq./a higher than the manually matched eLCA result and is therefore more correct. Accordingly, it can be stated that by using the LKdb and proposed matching methodology the inaccurate BIM model can be semantically healed for a more accurate LCA in early design stages. 5 DISCUSSION & OUTLOOK By semantically healing BIM models for LCA, the analysis of embodied carbon becomes holistically more consistent and more comparable for early design stages. Furthermore, the LCA knowledge database provides design options for optimizing the building performance according to LCA results. The limitations of this research are the chosen LCA database (Ökobaudat) and the correlating German language. Other NLP models of different languages might perform differently, as well other LCA databases might have less datasets. In a next step, the matching should be carried out on multiple case studies and verified with manually calculated LCA results. Furthermore, the performance shall be increased by checking domain specific abbreviations, as for example “STB” stands for “Stahlbeton”, (reinforced concrete) but could not be identified by existing NLP models.As in this paper, the focus was set to material matching for comparing the performance of several NLP models, in a next step the elementspecific matching shall be included in the performance analysis. Finally, to make the method more robust, commonly used elements for each classification with default values shall be defined in the LCA Knowledge database.
83
subtypes using 3D geometric deep neural networks. In: Advanced Engineering Informatics 47, S. 101200. DOI: 10.1016/j.aei.2020.101200. Llatas, Carmen; Soust-Verdaguer, Bernardette; Passer, Alexander (2020): Implementing Life Cycle Sustainability Assessment during design stages in Building Information Modelling: From systematic literature review to a methodological approach. In: Building and Environment 182, S. 107164. DOI: 10.1016/j.buildenv.2020.107164. Locatelli, Mirko; Seghezzi, Elena; Pellegrini, Laura; Tagliabue, Lavinia Chiara; Di Giuda, Giuseppe Martino (2021): Exploring Natural Language Processing in Construction and Integration with Building Information Modeling: A Scientometric Analysis. In: Buildings 11 (12), S. 583. DOI: 10.3390/buildings11120583. Navigli, Roberto; Martelli, Federico (2019): An overview of word and sense similarity. In: Nat. Lang. Eng. 25 (06), S. 693–714. DOI: 10.1017/S1351324919000305. Reitschmidt, Georg (2015): Ökobilanzierung auf Basis von Building Information Modeling. Entwicklung eines Instruments zur automatisierten Ökobilanzierung der Herstellungsphase von Bauwerken unter Nutzung der Ökobau.dat und Building Information Modeling. Masterthesis. Technische Hochschule Mittelhessen, Gießen. Rezaei, Farzaneh; Bulle, Cécile; Lesage, Pascal (2019): Integrating building information modeling and life cycle
assessment in the early and detailed building design stages. In: Building and Environment 153, S. 158–167. DOI: 10.1016/j.buildenv.2019.01.034. Safari, Kaveh; AzariJafari, Hessam (2021): Challenges and opportunities for integrating BIM and LCA: Methodological choices and framework development. In: Sustainable Cities and Society 67, S. 102728. DOI: 10.1016/j.scs.2021.102728. Schneider-Marin, Patricia; Harter, Hannes; Tkachuk, Konstantin; Lang, Werner (2020): Uncertainty Analysis of Embedded Energy and Greenhouse Gas Emissions Using BIM in Early Design Stages. In: Sustainability 12 (7), S. 2633. DOI: 10.3390/su12072633. Stenzel, Valérie (2020): Wissensdatenbank für Graue Energie und Treibhauspotenzial von Baustoffen. Masterthesis. München, Technische Universität München. Lehrstuhl für energieeffizientes und nachhaltiges Planen und Bauen. Wastiels, L.; Decuypere, R. (2019): Identification and comparison of LCA-BIM integration strategies. In: IOP Conf. Ser.: Earth Environ. Sci. 323, S. 12101. DOI: 10.1088/1755-1315/323/1/012101. Wu, Songfei; Shen, Qiyu; Deng, Yichuan; Cheng, Jack (2019): Natural-language-based intelligent retrieval engine for BIM object database. In: Computers in Industry 108, S. 73–88. DOI: 10.1016/j.compind.2019.02.016.
84
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Construction product identification and localization using RFID tags and GS1 data synchronization system A. Glema Pozna´n University of Technology, Poznan, Poland
M. Ma´ckowiak Deceuninck, Poland, Jasin, Poland
Z. Rusinek GS1 Poland, Poznan, Poland
ABSTRACT: The topic of this paper has been motivated by interest of new technologies in the area of civil engineering. New digital ways of design, oversight of manufacturing, construction and ways of upkeep and control of already completed structures are observed in fast developing industry branches. These methods which are gaining popularity in Poland are often already a standard in other countries where building information modeling BIM is present in engineering, facility management and state administration practice. Basic information necessary to understand conception of RFID technology usage in civil engineering are presented. The description of a GS1 organization is given. In the next chapter RFID technology is described, ways of current usage and possible ways of implementing it into civil engineering industry products. There is the comparison of RFID tagging technology to bar-/quad-coding marking technology as well. As a case study one can read about example of system design of prefabricated product identification and localization using RFID technology in order to improve and optimize prefabricate production inside product factory, transportation, construction works, handover and facility operation. So, the effects of using RFID are investigated for improvement of products management during its life-cycle exploitation. Keywords: RFID, Identification and Localization, Building Fabricated Products, Digitization, BIM, Building Life Cycle
1 INTRODUCTION The topic of this paper has been motivated by interest of new technologies in the area of civil engineering. Unusually fast digitalization of our world is happening in all of our lives and in all branches of industry and science including construction industry. New digital ways of design, oversight of manufacturing, construction and ways of upkeep and control of already completed structures are observed in fast developing industry branches. Application of RFID (Radio Frequency Identify Index) with GS1 identification system for construction industry products is the aim of this paper. The industrial examples are introduced from and for MatBet/Matdeco (Poznan, Poland) building product company.
lines and spaces for encoding information (Figure 1). That information is encoded horizontally from left to right. 1D barcodes holds a limited number of characters, typically 20-25. In order to add more numbers, the barcode must be longer. The most familiar 1D barcodes are UPC codes found in common on grocery and consumer items. A 1D barcode depends on database connectivity in order to be meaningful; after a scanner reads the numbers in the code, they must be linked back to product or pricing data, or other information.
2 PRODUCT INFORMATION CODING 2.1 1D Bar coding 1D bar code (also known as a linear code) is a visual black and white pattern, using variable-width DOI 10.1201/9781003354222-11
Figure 1. Garden pot bar coding tag.
85
Tags that do not have their own power supply are called passive tags. Once in the electromagnetic field at the resonant frequency of the receiving system, these tags store the received energy in a capacitor which is incorporated into the tag structure. A response containing the identification (code) of the tag is sent after the tag has accumulated sufficient energy. The transponder does not respond immediately, only after some time. After a response it remains idle for a certain amount of time which allows multiple tags to be read. They are also in reading range at the same time. The most commonly used frequency is 125kHz. This allows information to be read from a distance of less than 0.5m. More sophisticated systems that provide readings from one to several meters operate at frequencies of 13 MHz to 5.8 GHz. RFID technology can also be divided on the basis of the frequency used in the system: LF, HF, UHF: low, high, ultra-high frequency. RFID currently has many applications, of which the authors has chosen to list and describe only the most popular ones (Figure 3):
2.2 2D Quad coding 2D barcode uses patterns, shapes, and dots to encrypt information both horizontally and vertically. 2D barcode can encrypt more characters (around 2000) in the same amount of space as a 1D barcode (which only has 20-25). Types of 2D codes include QR code, PDF417, and Data Matrix. In addition to hold more data, 2D barcodes can encrypt images, website addresses, and other binary data, which means that the codes can work independently of a database. 2D barcodes can be used to mark very small items when a traditional barcode label would not fit: see surgical instruments or circuit boards inside an electronic device. When it comes to customer selection, 2D barcodes are often people’s preference due to the amount of information a 2D barcode can hold in comparison to a 1D. 2.3 RFID Technology Radio-frequency identification (RFID) technology is a technology for Radio Frequency Identification Systems, or Radio Frequency ID tag technology. It uses radio waves (Figure 2) emitted by a reader to transmit information and power an electronic circuit (tags – also called transponders) to identify an object upon returned signal. This technique makes it possible to read and, depending on the tags used, write information about an object onto an RFID tag. Depending on the needs and the used devices, it is possible to read the recorded data at a distance o of several tens of centimeters up to several meters from the reader antenna. It is possible to read single objects as well as multiple tags in the reading field. The basic system configuration consists of: (I) reading device or system of devices:
Figure 3. Receipt of goods by GTIN barcode.
•
reader containing a transmitter, a receiver and a decoder; • transmit and receive antenna or two antennas: a transmitter and a receiver;
1. RFID solutions in manufacturing – process automation and accurate recording of events enables precise production management, process tracking, recording parameters and counting materials and products. Thanks to the use of RFID one can significantly optimize the time and cost of product manufacturing and maintain constant supervision over individual production stages. 2. management of company assets and inventories – RFID ensures efficient administration of company assets, such as equipment, materials, tools or machines. RFID tags also enable very fast and very accurate inventory of assets or stock. 3. transport and logistics – RFID technology enables easy tracking of tagged goods at every stage of transport, to accurately monitor the entire supply chain. Additionally, in aviation, it supports identification of baggage. In addition, it enables continuous surveillance of shipments providing information on location, dispensing and delivery, which is important for pharmaceuticals. 4. packaging records – RFID makes it possible to account for returnable packaging, package movements and, in the event of loss, allows the identification and recovery of lost items.
(II) transponder or tag array which consists of: •
passive electronic circuit (design is often an integrated circuit without a housing), which size can be as small as 0.4 x 0.4 mm. Passive circuits have no power supply; • antennas.
Figure 2. Example of RFID tag structure.
The RFID reader sends out an electromagnetic wave by means of a transmitter, then a receiving antenna responds the electromagnetic waves. Waves are filtered and decoded by means, so that the tag responses can be read.
86
GS1 logistics label, code quality verification, traceability audit or LEI (Legal Entity Identifier) code registration.
5. work time recording – using RCP cards is a common standard today. It uses cards and short-range readers. 6. evacuation support – RFID can support a system for evacuating employees from the workplace. With chips hidden in badges and readers along the evacuation route, it is possible to track and control the progress of plant evacuation in real time. 7. smartphones and mobile devices – NFC standard module (HF standard), extremely popular, widely used in smartphones and mobile devices. This module is most commonly used for contactless payments, but can also be used for the transfer of information between devices. The ISO standards that standardize RFID technology are: 1. ISO/IEC 18000-1 – governs the general concepts of RFID architecture. 2. ISO/IEC 18000-2 – specifies the parameters to be used for the communication of RFID LF (low frequency). In addition, it specifies protocols, commands and methods for reading single tag signals at these frequencies. 3. ISO/IEC 18000-3 – specifies parameters to be used for RFID communication at 13.56 MHz. RFID at 13.56 MHz. In addition, it specifies protocols, commands and methods for reading single tag signals at these frequencies. 4. ISO/IEC 18000-4 – specifies the parameters to be used for RFID communication at 2.45 GHz. It specifies protocols, commands and methods for reading single tag signals at these frequencies. 5. ISO/IEC 18000-6 – the standard regulates the physical interaction of RFID devices. It establishes protocols, commands and collision avoidance measures for passive RFID systems that operate in the UHF range. 6. ISO/IEC 18000-1 – defines the radio interface supporting active RFID tags in the 433 MHz band. 7. ISO 14443 – a standard that regulates proximity cards operating with NFC technology. 8. ISO 15693 – regulates proximity cards operating on the principle of near-field inductive coupling.
3.2 Means of identification The GS1 standard is the ability to uniquely identify products, assets, locations and similar data. All GS1 identifiers can be represented by GS1 barcodes, Radio Frequency Identification (RFID) tags, which can be scanned or read. Frequency Identification, which can be scanned or read automatically. Follow 4 types of numbers are used to identify data: – Global Location Number (GLN) used to identify of physical locations and legal entities – companies and organizations, and is used for electronic document interchange (EDI); – GlobalTrade Item Number (GTIN) used to uniquely identify any product or service which need to be priced, ordered or invoiced at any point in the supply chain; – Global Individual Asset Number (GIAN) global individual asset identifier; – Serial Shipping Container Code (SSCC) serial shipping container code. 3.3 GS1 in the construction industry Today, the construction market shows even growing demand for full traceability of products throughout the supply chain, both in terms of compliance and sustainability. The serial GTIN and its integration with BIM is a logical and necessary part of this development. The use of GS1 product identification methods may facilitate management of the construction site, material supply, human resources, as well as management of a building during its operation. In GS1 Poland there are approximately 1700 participants with manufacturer classifications for the construction industry. They include: 34% – Furniture and wood products 24% – Lighting equipment and cables 16% – Plastic and rubber products 8% – Plant and machinery 6% – Metal products 5% – Chemical 5% – Minerals, oil and derivatives 2% – Glass and ceramic products The use of GS1 product identification methods may facilitate management of the construction site, material supply, human resources, as well as management of a building during its operation. For example, after scanning with a smartphone the barcode /quad code attached to delivery, the person responsible for receiving goods at the construction site can quickly mark which items have actually arrived on site, which are not accounted or which have been damaged in transport, so that a replacement must be ordered. This makes it considerably easier and quicker to deliver the missing/damaged parts to construction site, eliminating the possibility of quantity confusion when they are ordered by a man.
3 DATA SYNCHRONIZATION 3.1 What is GS1? GS1 is an international organization with headquarters in Brussels (Belgium) and Princeton (USA).This organization manages the GS1 system on a global scale. At the national level, the so-called national organizations are responsible for the GS1 administration. All users of the GS1 System are actively supported by the national organizations that inform, train and advise on the implementation of standards. The national organizations also create cooperation forums for users, allowing them to exchange experiences with each other and actively participate in the development of existing and future standards. Among other things, GS1 offers advice on the implementation of the
87
Ultimately, the GTIN code will contain 16 types of information. Building Information may contain: – GTIN; – target market; – brand; – name; – net content; – unit [SI]; – global product classification GPC; – fire resistance (EI 15/EI 30/EI 45); – wind load resistance (class 0; class 1); – resistance to burglary (P1A; P2A; P3A); – frame [frame width in mm]; – opening width [escape width in mm]; – color according to NCS 4055-R95B; – color according to RAL classic system; – global warming footprint [kg CO2/unit]; – guarantee period (number of months). This will enable you to quickly obtain information on any product marked with a code. If a replacement arrives at the construction site, it will be possible to quickly check its parameters and certificates, without a laborious search on the internet. During the lifespan of a building, GS1 codes integrated with BIM can significantly speed up the servicing of equipment process, as well as the replacement of damaged components of the building itself. The service technician, after identifying a defective part, scans the GTIN code, so he or she gets accurate information about the type of part, stock status, or where it can be obtained. The building manager, having identified a broken tile in a common part of the building, can quickly order an identical replacement by scanning the barcode or QR code placed on the component with the appropriate tool or smartphone as device in almost everyone’s pocket, even hand. The building object user can obtain all the information that the manufacturer has decided to make available using the aforementioned codes without searching for the element in the building model, which can be cumbersome in the case of landscaping elements or individual windows and doors (Figure 4).
Figure 4. Garden pot material properties and geometry data.
low-quality and low-cost components when sewer pipes are hidden deep underground. The control of internal logistics, the open yard warehouse and the external transport of precast concrete building components is another example where the management of a manufacturing company can find real benefits from use of RFID and GS1. 3.5 BIMStreamer Managing Product Content After creation of the product model, data collection and synchronization there is one more important element of product digitization. The product BIM model is ready for online uploading of the openBIM product data, parameters, documents, etc. We have established cooperation with IT data management company Sagiton (Figure 5) with general PIM (Product Information Modeling) portfolio after getting the BIM specialization. BIMStreamer is the dedicated building data management system, partly fulfilling functions for construction works (scanning, identification, data base storage, various versions, local language data management and finally mobile & web application development).
3.4 Garden pot identification with GS1 As part of the collaboration with MatDeco and GS1 Poland, the garden pot made of architectural concrete have got all possible identification tags. First the geometrical BIM model of product have been created. The simplest geometry has been considered. We put attention here not for quantity of geometrical and non-geometrical data, but as possible as to clear presentation of the whole process of introducing identification and product data flow. The research to date has opened up the first prospects for building product manufacturers, the possibility of controlling an in-house product, good in quality, but not the cheapest, is built into customer’s system in accordance with the contract, without the sub-supplier’s incompatible substitution of
4 BARCODE VERSUS RFID SYSTEMS IN CONSTRUCTION. Construction companies use RFID stickers to monitor the condition of equipment and it movement. In production, transport, delivery and inventory, they mainly
88
(ii)
Figure 5. Garden pot data in Product Management System.
use barcodes – often created in the manufacturer’s system custom made for his specific requirements. Such codes are incompatible with worldwide standards and their reading is dependent on the conditions prevailing on the shop floor, e.g., heavy dust, dirt from the building site, sunlight damage can make it impossible to read the code with a reader, as it requires visual contact with the code. Code labels also often peel off from the product and are lost. Replacing a predominantly barcode-based system with one mainly based on RFID brings many advantages. RFID tags in contradiction to barcodes can be read in all conditions – the RFID reader does not require line of sight with the tag. Also, the inclusion of tags inside the product makes it impossible to lose it – it would only be impossible to read if the tag is not placed inside the product or it is physically damaged which is easy to catch at the production stage. RFID tags are more resistant than barcode stickers, scanning takes place automatically, unlike barcodes, which must be scanned manually. In order to streamline the construction process, RFID can be used in many stages of the process, as well as in intermediate and associated industries, following the example of industries already using RFID in a developed way. See set of fabricator use: (i) prefabricated product marking – tag placement for greater automation of prefabrication plants, which, in the diminishing human resources status, will allow to optimize the prefabrication production process. The amount of data contained in the RFID chip is greater than on a barcode label, the RFID tag is also impossible to lose or to falsify easily. Management of production could be controlled and optimized from the earliest stages of material arrival at the production facility till all the way through to the shipping final product out off the factory. Marking each stage of production allows for easy recording of supplier data, production/delivery dates of components, current conditions in the plant, as well as other variables that can affect product quality. At a higher level of system expansion, it is also
(iii)
(iv) (v) (vi)
(vii) (viii)
(ix)
89
possible to automatically record tools and persons directly involved in the production, storage and transport of the component. The movement of a component around a facility can be easily controlled by gates that read the entry/exit of a component from storage yard, sensors on the storage areas can make it easier to control the space on the site and indicate the most optimal storage location of the component. employee control – in a prefabrication plant tag makes it possible to easily control work time, as well as the current location of an employee. In large plants, the support of evacuation plan is invaluable by automatically determining the position of occupants in real time. Thanks to RFID mats placed at workstations, it is easy to read and record in the system which employee was working on a particular item. This enables us to easily determine work performance and redirect people to the optimal positions on which they perform the best. transport and logistics – the RFID chip makes it easy to control the movement of goods. Thanks to the long-range stationary readers placed in the gates, every item entering or leaving through the plant gate is read and added to or subtracted from the stock on an ongoing basis. materials or components cannot leave the premises without the managers’ knowledge, minimizing theft and errors in logistics. recording of the exact date and time of transport also makes it easy to optimize further processes for which the product is is used. construction site traffic control – thanks to RFID tags affixed to the vehicles of participants in the construction process and issued to employees, construction site traffic can be easily controlled. Labels applied to the vehicle body allow the entrance to the construction site to be automated with a barrier operating the system. Employees entering through the pedestrian entrance, thanks to their personal RFID tags, can enter independently onto the construction site. identity checks, while at the same time recording the start and on leaving the site the end of their working time. work tools that are tagged with RFID, it is easy to check their current amount and quality. If they are lost, the RFID signal makes them easy to find even in difficult conditions. support of the guarantee process and protection against forgery is possible to support the guarantee process of an item throughout its life. Thanks to the fact that the tags cannot be counterfeited, it is also easy to determine whether an item is a product of our company or counterfeit of inferior quality for which we are not responsible for the quality and performance, which can also grounds for declaring the guarantee null and void due to failure to meet its conditions.
The long lifetime of the technology is also an advantage over barcode labels, which are easily torn off or destroyed – and sometimes this is required before installation. All these advantages of RFID outweigh the costs and time required for implementation. Since RFID tags in precast construction are not a highly popular technology, further research must be carried out into the readability of tags in different concrete compositions.
Thanks to a large tag chips memory, it is easy to check under what conditions the component was made, who worked on it, at what plant (assuming the system is designed to enter such data during production), its serial number, and to check if placement is in line with the design, thus eliminating possible manufacturing errors. (x) RFID chips can easily support building management and maintenance. Starting with the identification of defective components for quick replacement with ones of similar or better performance, assisting with building security, rescue evacuation in the event of an emergency, police, fire brigades or special purposes forces, to managing building assets of the building depending on the location of the occupants. (xi) demolition, as a final building life-cycle stage, does not need wide description. Design of structural destruction or recycling of materials and equipment is according to all previous construction data and processes storage and management.
REFERENCES Eastman C., Teicholz P., Sacks R., Liston K., 2011. BIM Handbook. Guide to Building Information Modeling for Owners, Managers, Designers, Engineers, and Contractors, Wiley, New Jersey. Glema A., 2017. Building Information Modeling BIM – Level of Digital Construction, Archives of Civil Engineering, vol. 63, no. 3, s. 39–51. Kasznia D., Magiera J., Wierzowiecki P., 2017, BIM w praktyce. Standardy, wdro˙z enie, case study, PWN, Warszawa. Ma´ckowiak M., 2021. Development of a digital MatDeco building product in BIMStreamer BIM content management system, BSc Thesis, PUT, Poznañ. Rusinek Z., 2020, How GS1 supports industry?, Web|Pozna´n. Siemens, 2020. DigiIndex – Digitalization Level in Poland. Matdeco (2018) Available at: https://matdeco.pl/ GSIPolska Available at: https://www.gs1pl.org BimStreamer (2022) Available at: https://bimstreamer.com RifdPolska (2022) Available at: https://www.rfidpolska.pl/ technologia-rfid-co-to-jest RifdPolska (2022) Available at: https://www.rfidpolska.pl/ kategoria-produktu/tagi-rfid
5 FINAL REMARKS The RFID tagging system has a number of advantages over the tagging system with labels. Ease of reading, the ability to quickly expand the system with new elements, and high resistance to the conditions in which the system is used make it an ideal replacement or supplement to a barcode labeling system. The introduction of such a system will considerably speed up and allow greater production optimization. The additional security against counterfeiting, support of claims and service processes also have their undeniable advantages.
90
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
What comes first when implementing data templates? Refurbishment case study P. Mêda & D. Calvetti CONSTRUCT/Gequaltec, Instituto da Construção, Faculty of Engineering, University of Porto, Portugal
H. Sousa CONSTRUCT/Gequaltec, Faculty of Engineering, University of Porto, Portugal
E. Hjelseth Department of Civil and Environmental Engineering, Norwegian University of Science and Technology, Trondheim, Norway
ABSTRACT: A built entity is composed of different construction products and its data digitalization is found to be crucial. However, this is a highly complex task due to stakeholders’ wide range of solutions and requirements. The ability to identify the most relevant Data Templates (DT) is key to set meaningful steps for the value-chain innovation. This work contribution is the development of a case study using a Portuguese social housing building representing the stock to be refurbished under EU Renovation Wave. Findings indicate that relevant construction and deconstruction activities address elements as windows, roofs and doors. When targeting deconstruction activities, hazardousness and waste analysis are crucial vectors for product digitalization. On the other hand, when targeting construction activities, key aspects to evaluate are the energy efficiency index and facility management. Products BIM dimensionality, detail and layers are crucial elements for both types of activities for DT cases implementation.
1 INTRODUCTION 1.1 Overview The new EU proposal for a Construction Products Regulation (CPR) laying down harmonized conditions for the marketing of construction products assumes as general objectives the achievement of a wellfunctioning single market for construction products and to provide a contribution to the objectives of the green and digital transition, particularly the modern, resource-efficient and competitive economy (Commission 2022). In order to accomplish part of the second goal, it is assumed that providing information in digital format is a key requirement and that Digital Product Passports (DPP) are the way to accomplish that by gathering them, “to the extent possible, in a EU construction products database or system”; Article 78 (Commission 2022). The DPP digital format assurance are the Data Templates (DT), the core topic of this research. It is not viable to think that, in the short term, there will be DPP for all products, systems and objects (Mêda et al. 2021a). This requires an incremental and continuous effort that must concentrate on the definition of these structures, on the collection and production of data and on the strategic decision of the DOI 10.1201/9781003354222-12
data to be disclosed, facing the legal requirements, performance requirements and others that are relevant for specific stakeholders throughout the construction lifecycle. In addition, this will introduce changes to the processes across the construction value-chain, meaning that stakeholders will need to adjust and become skilled for this reality, especially in what relates to the DPP digital format requirements (Akyazi et al. 2020). At this point DT jump in, providing in the first moment the data framework that fulfils all the requirements to deliver standard, interoperable, machinereadable and traceable data (Mêda et al. 2021a). Following this, it is suggested to use a Data Templates mind-set to identify the most relevant ones in construction projects. The specific context (construction project) where requirements are defined can lead to very different concerns and outcomes. All represent dimensions that can be measured, contributing to a practical roadmap for incremental implementation (based on what is more important, what sells more, what contributes more to a defined performance) and disclosing practical experiences based on real situations. At this point, two main concepts need to be related: DPP and DT. As mentioned, DPP has its roots in the CPR proposal.
91
2 CONTRIBUTIONS TO THE BODY OF KNOWLEDGE AND INDUSTRY
1.2 Data templates In accordance with ISO 23387:2020 definition, Data Templates are standardized, and interoperable metadata structures used to describe the characteristics of construction objects, systems or products (ISO 2020). Given this standard alignment with the digital processes and assumptions, DT are also information exchange enablers across the construction life cycle. Although the standard mind-set is purely digital, there is still a lot of fuzziness in the industry in what is the meaning and range of the concept. As so, the authors have been addressing DT also as Digital Data Templates (DDT), to highlight the “true digital” that is meant to underlie this concept (Mêda et al. 2021a). In this manuscript DT is the acronym to be used. Data Templates are the products’ “skeleton of information” that organize and support data for defined types. This data can assume very different hierarchies, but one of the key aspects is the data properties. The same standard defines properties as the “inherent or acquired feature of an item” (CoBuilder, n.d.). Following the definition and assumptions, DT are meant to provide the “true digital” framework for specific product types.These product types can be defined by a harmonized standard, for example, a sandwich panel in accordance with EN 14509 (DSCiBE 2020). Considering the example of a real project situation, the design team can set different solutions for sandwich panels in a building refurbishment project, varying on the color, shape, thermal insulation value, composition, and brand. These solutions are meant to be placed in different roof sections. Performing a transposing of the concepts under discussion would translate into several DPP, one for each design solution, all using the same Data Template as background to provide the common structure. The example of a building refurbishment project is not innocent given the short-term trends as it will be further detailed.
Several authors have focused their research on data requirements/organization topics for several goals, such as energy performance (Bassanino et al. 2016), LCA (Llatas et al. 2019), waste (Quiñones et al. 2021). Few others look from multi-dimension lenses. The present research sets Data Templates as background and uses different lenses that are found to be relevant when approaching building refurbishment projects. The focus is placed on the specific industry goals without forgetting the core practical aspect of construction that is the budget. Built on these assumptions, the research question is the following: What are the priority Data Templates to implement when the case is the building refurbishment for housing? Multiple analyses are performed using a specific refurbishment process as a case study to obtain wisdom related to DT. Following the case study presentation, where the refurbishment goals are highlighted, the analysis of the Bill of Quantities (BoQ) provides insights related to the different types of activities and their detail at several levels. Refurbishment actions are characterized by deconstruction and construction activities. Both are relevant for the research as they can prove, among others, that circular economy can be fostered by information circularity. Harmonization and analysis of these two types of activities are performed and discussed, leading to the priority of DT. Associated to practical implementation this means that is the lenses is placed on waste the priority DT will be some, the same or different ones if the focus is energy efficiency and the priority DT’s if it is to consider several lenses (cost, energy efficiency and waste, as example). This is found to be the most valuable lesson for the practical industry stakeholders and for the scientific community. In addition, relevant dimensions to consider and a framework for future research are presented. A brief overview of Barriers, Enablers, Drivers for change and, Opportunities relating to DT implementation sets the ground for conclusions (Owen 2013).
1.3 Renovation wave The Renovation Wave for Europe strategy targets the refurbishment of 35 million building units by 2030 (European Commission 2020). EU’s building stock accounts for 220 million units (85% of the total) with more than 20 years of in-use (European Union 2020). In Portugal, buildings’ low energy efficiency is an unsuitable reality. The Portuguese government is committed to challenging goals for the renovation of buildings of about 363 million square meters by 2030, 635 million/m2 by 2040 and 748 million/m2 by 2050 (Government 2021). To achieve this goal and circular target actions, the European Architects, Engineers, Contractors, Operators and Owners (AECOO) sector should increase performance on many chains as construction products, design, construction and facility management (Calvetti & Mêda 2021). Given the aim and requirements for the Renovation Wave, building refurbishment projects constitute good examples to make case studies in order to make evaluations and implement innovations that can accomplish the identified targets.
3 METHODOLOGY The research methodology uses mixed methods and a case study that is representative of the Portuguese housing buildings stock that suffered a refurbishment and also representative of the scope of works to be performed in the near future relating to the EU renovation wave strategy. Both characteristics allow scalability of the results for a wide range of future projects (Fellows & Liu 2022). Based on the BoQ and budget estimate of the winning bid (public procurement procedure design-bidbuild procurement route), all activities are analyzed and classified as deconstruction, construction and mix works. This aims to provide awareness of the activity types regarding their essence. Complementary, the
92
performing a follow-up of the construction phase with access to all documents and regular site visits.
activities were also clustered following systems and built elements logic. As mentioned, the Data Templates mind-set is always present as an example of the windows of the same brand but with different dimensions and are assumed to have the same DT in their origin. All these visions allow the identification of the items’ relevance, namely in terms of cost and quantity. The most valuable (from an economic viewpoint) group of refurbishment activities will constitute the in-depth analysis sample. The option to perform it for the three groups is due to the central role of Data Templates as previously demonstrated and that now we aim to prove in practice. The life-cycle vision based on DT requires a multi-dimension approach and specific lenses for construction and deconstruction. Refurbishment actions must consider and manage both activities in the same process. IDDS critical analysis of the obtained results is made supporting the discussion and conclusions related to this specific case study and strategic roadmaps for future actions, both in practice as in research (Owen 2013).
4.2 BoQ – analysis and sampling definition The Bill of Quantities is an essential element from the design phase that is used to build the competitors’ proposal, namely the offered budget. The BoQ and prices from the winning offer are part of the winning bid documents and constitute the main element for this analysis. This refurbishment project, corresponding roughly to a gross building area of 4,844.24 m2 , was awarded €1,999,000.03. The analysis of each activity type and its economic relevance is one of the first actions. Although understanding that these items and prices are composed of materials, human resources, and equipment in different percentages, this division is not considered for the purpose of the analysis and is a matter of simplicity. Refurbishment projects are composed by demolition/deconstruction activities and new construction activities. In this case other types were evidenced, namely activities that mix both actions (replacement), auxiliary works such as site, temporary housing for tenants, among others and maintenance/repair activities. Figure 1 resumes, for the present case study, the economic weight of these different types.
4 CASE STUDY 4.1 Presentation A social housing building complex under refurbishment located in Custóias, Matosinhos, a city north of Porto and within the Porto Metropolitan Area, was identified to be used as a case study. Built-in the late 1970s, this complex comprises four blocks corresponding to 7 buildings. All entrances have four floors and two apartments per floor. The owner is the Matosinhos Municipality, which constituted Ma-tosinhosHabit, EM, to ensure operations management and maintenance. Since construction, the complex suffered few interventions aside from reactive maintenance when needed. Most of the existing elements are in the original condition, although inside the houses, several changes were made by tenants. With most of the elements at the end of their life period, an extensive renovation project was launched to improve the apartment’s health, safety and habitability conditions, performance spaces and systems redefinition, and significantly increase thermal comfort and energy efficiency. The process was performed using Design-Bid-Build’s most traditional procurement route for public contracts in Portugal. The design was made in mid-2010, and the documentation is composed of PDF files, electronic spreadsheets, and CAD drawings. The intervention can be characterized as deep refurbishment where the structural elements, the façade, partition walls and some elements whose conservation state does not require replacement (floor coverings on common areas and stairs and handrails) were to be maintained. All other elements were set to be replaced and new systems to be built, adapting these constructions to present requirements and regulations. The option for this project was due to the possibility of having access to all the design documentation and
Figure 1. Refurbishment case study – Economic relevance of the different types of activities.
It is relevant to highlight that nearly 15% of the budget is assigned to auxiliary works. This is related to site-associated expenses and mainly to the provision of temporary houses (containers) for tenants. The project strategy was to empty one building at a time, this is eight houses, and move the tenants to these temporary facilities. Associated to this are the human resources to support the moving of all belongings. Most of the budget is addressed to new construction activities, but it can be observed that 13% is to “mixed works”, many times referred to in the BoQ as replacement actions. Starting with the construction activities, the items present different organizations depending on the system, design discipline, quantities, as well as specific assumptions from the design teams. For the purpose of defining the Data Templates and what should come first, it is relevant to approach these items and their economic relevance from a construction products and related harmonized standards point of view (it is
93
activity types that represent nearly 70% of the total refurbishment budget. Considering the Auxiliary work’s relevance, that increases even more, as these items constitute the core of the budget with direct implications for the building components. As it will be further detailed, the “mix works” will be divided, and the calculated relevance for Construction and Demolition will add to the respective parts.
relevant to highlight that this is related with the CPR). With this, it means that two items addressing exterior windows with different geometry but with the same composition (example, aluminum and glass), the applicable Data Templates are the same, and the economic relevance is the sum of both. Other situations might be presented where a single activity supports the execution of one or more systems, for example, power supply and lighting systems, including all cables and accessories. Although this is common practice in more traditional projects where contractors are able to propose an overall cost by looking at drawings when using BIM or to perform other types of assessments, this way of work does not fit the purpose. These different situations and as-sumptions have a direct impact on the analysis as it will be further detailed. Deconstruction activities evidence low economic representativeness, and similar to what was mentioned regarding electrical and lighting systems, there are several items that glue many building components, or parts do be removed. Partially, this explains why these activities represent only 4% of the budget. If the items were more detailed or individualized, the cost might increase. However, in practice, the deconstruction cost will increase due to the amount and economic relevance of the “mix works” as it will be following detailed. Regarding these, it is not surprising to check that the most cost relevant are related to building envelope elements such as façade doors and windows and roofing. These are the leading lowperforming components from an energy performance lens when approaching the case. The activity with higher economic relevance is related to interior openings and partitions changes, as well as changes in façade openings. The “mix works” follow a similar pattern to the previous ones to what is needed to add the situation of a single item composed of two parts, a deconstruction action followed by a construction action. The situations range from replacing single elements (doors, windows and façade elements, footers) and the replacement of entire systems (telecommunications, sanitary drainage, water distribution). Regarding the most relevant situation, “Interior doors replacement”, the activity is composed of door removal including all minor components and the application of a new door together with the same components. In this specific situation, wooden doors with metallic surroundings are being replaced by wooden doors with wooden surroundings. From a DT or circularity point of view, this could be an interesting situation as the activity is pointing to the same Data Template – Wooden Door. This will be explored further. However, approaching this topic from an economic, sustainability or digitalization lens, the activities must be split into deconstruction and construction actions with the inherent costs attached, as will be explained. Figure 2 summarizes the most relevant Construction, Deconstruction and Replacement “mix works”
Figure 2. Most relevant activities from the different types.
As it is visible, 16 items/activities are responsible for a significant part of the budget. The criteria for the Deconstruction activities were to use all existing and in terms of Construction items. The option was to select the top 10 that, for this specific case, matched with items above 2,50% of the budget. Other dimensions such as sustainability, asset management, waste or digital modelling for project development will be considered according to their applicability to the types of activities. However, the budget is regarded as a critical dimension providing the background for analysis and comparisons as well as the assumption for the DT mind-set. 4.3 Sample analysis 4.3.1 Mix works In order to work the information, it is necessary to split the BoQ activities in two actions, one related to deconstruction and the other to construction. This task is easy to perform in terms of the item description. However, in terms of budget, there is a need to identify the proportion in which the division should be performed. For this action, a cost database that gathers BoQ items and prices from budget estimates and winning bids from 100 projects was used (Sousa & Mêda 2012; Sousa H. et al. 2016). Medium values from all estimates and bids were consolidated to find the cost for each specific action, for example, interior wooden doors removal (un) and interior wooden doors application (un). Figure 3 presents the three main budget
94
perspective, this activity frames on the EWC 17 01 07 – mixtures of concrete, bricks, tiles and ceramics and regarding waste audit considering the Intermediate data for elements Inventory characterization it can be classified as non-hazard, not reusable. In terms of possible and recommended outlets, Figure 4 presents the considered range. Windows and Blinds and Interior Doors present a similar behavior in relation to modelling, as frequently these elements present high definition not just in dimensionality but also in detail and appearance. Considering Data Templates, enclosures can be the same for both and, in this case, can be regarded as wood elements. On the other hand, there will be specific requirements for windows, doors and blinds as different harmonized standards apply to them. From a waste perspective, the corresponding EWC relate to wood, glass, and plastic. Due to the age and underperformance, windows are considered non-hazard, not reusable, while interior doors can be reused in other buildings with openings with similar dimensions.
relevant “mix works” divided into de-construction and construction parts.
Figure 3. Relevance of the deconstruction and construction parts in replacement activities.
As mentioned, the economic relevance of each of these parts will add to the Deconstruction and Construction sample items and become part of those analyses. Data Templates are structures meant to forecast all relevant data related to products, systems, and objects characteristics. The increased circularity ambition leads to multiple requirements crossing digitalization, performance, environment, and waste. All the factors underlying these requirements can vary depending on the Owner’s objectives, and singularities of the case, among many others. When looking in detail, the activity types can use aspects within these factors in the same way or can be exclusively applicable to a specific type: deconstruction and construction. In the end, all this environment and the cost as the background is aimed to support the decision of what data templates must come first. The following sections will work on these aspects within the two samples, performing in-depth analysis for each activity or group, highlighting the aspects in discussion, and giving the items’ descriptions of the involved limitations or assumptions that had to be made. 4.3.2 Deconstruction The deconstruction sample represents 6,03% of the overall budget. Although this is a small part of the total cost, these items are highly relevant to support the ambition of having waste audits for all buildings to be refurbished. This process needs to become more digital, and wisdom related to the elements to be removed, their recycling potential/outlets as elimination (EL), energy recovery (ER), backfilling (BF), recycle (RCI), external reuse (RUE), internal reuse (RUI), the Hazardousness, the European Waste Codes (EWC) and weight is vital. In addition to cost and BIM modelling practices, both dimensions will be common to the different types of activities; for deconstruction items, in-depth analysis will be performed using aspects such as Outlets, EWC and Hazardousness (European Commission 2018). All of these are meant to provide insights for the prioritization of the Data Templates (Mêda et al. 2021b). One aspect that is common to all sample items is that they address specific building components where it is easy to understand their constitution/more relevant parts. Starting with masonry walls demolition, there are three main Data Templates to consider, relating to blocks, masonry mortar and rendering mortar. In terms of modelling, walls are often modelled using a single layer with accurate dimensionality. From a waste
Figure 4. Overview of the deconstruction items analysis.
Footers and Sills are also similar as they often are not modelled. Footers have high relevance due to the amount of material removed. From a Data Template perspective, these fit in wood elements and stone elements. The applicable EWC correspond to the same materials 17 02 01 – wood and 19 12 09 – minerals. Sills are assumed to be non-hazard and reusable because it is a common practice in this type of intervention. Due to the measurements, the option for footers is often recycling. Roof items constitute the last to be analyzed and comes in last due to the singularity in terms of hazard. The case situation is a material containing asbestos, and therefore it has other requirements. The corresponding EWC is 17 06 05 – construction materials containing asbestos. From a modelling point of view, often roofs follow simplified modelling as a single
95
these systems in similar situations (piping, accessories, equipment, among others). Cluster 3 is composed of streamlined items where each requires one up to two Data Templates. The activities related to Interior Doors, Kitchen Furniture and Ceramic Tiles for housing pavements and walls are characterized by their high definition in terms of modelling. As well, they are out of the scope of the owner’s maintenance and do not have relevant contributions to the energetic performance of the building. It is interesting, however, to observe that Interior Doors are also applicable in deconstruction activities. Cluster 2 is made of a single type of item that is paintings. This also constitutes a particular situation in this analysis environment as it is a single DataTemplate of a not modelled item and with no contribution to the energy performance. However, different colors can introduce different behavior. Given its application in all buildings, the parts that belong to common areas are to be maintained by the owner. Therefore the Facility Management situation is “partial”. The three items that constitute Cluster 1 account for more than 25% of the refurbishment budget. The common aspect of these activities is that they belong to the building envelope, having a high impact on the Energetic Performance (EP) of the building. All are also elements that the building’s owner should maintain. The modelling definition is associated with the BIM objects provided by manufacturers but also with the Data Templates, where windows might have one or two depending on the association and the roof; in this particular case, a sandwich panel and an insulation layer correspond to two DT’s. ETICS might be the most undefined situation given the layers and the absence of a single harmonized standard, although there is a European Technical Approval. Depending on the granularity, it can be a single DT or several ones corresponding to the different products that constitute this system. Figure 5 aims to summarize the construction activities in an in-depth analysis highlighting the aspects related to energy efficiency and facility management. As in Figure 4, the Y-axis provides the priority Data Templates.
layer. Figure 4 aims to summarize this analysis highlighting the aspects related to waste and the priority Data Templates associated with this type of activity. 4.3.3 Construction As previously observed, construction activities are the most relevant part of the refurbishment budget, and the top 10 sample accounts for more than 60%. Some reasons underlie this situation will be detailed. Regarding this type of activities, the dimensions to be analyzed are related to the refurbishment goals and the owner’s primary concerns as it will have to ensure part of the maintenance of the building. Specific parts are to be maintained by tenants. The main dimensions are the contribution to improved energetic performance (EP) and the owner’s requirement to maintain the elements (FM). Cost (budget) and modelling (BIM) are part of the common parts, and there was an effort to evaluate if the solutions are hazardous or not or if they follow Design for Disassembly (DfD) assumptions (Rios et al. 2015). To facilitate the writing regarding the results, the activities were clustered as presented in Figure 5.
Figure 5. Overview of the construction activities analysis.
5 FINDINGS AND DISCUSSION
Starting with Cluster 4, it contains items of the BoQ that address different systems. The activities have the particularity of gathering in a single description all system components that are needed for the refurbishment. This is often a solution that some design teams take when it is impossible to perform an extensive evaluation of the needs. However, this can lead to difficulties during the execution and, more relevant for the purpose of this work, prevents the evaluation of the components and inherent Data Templates as well as the associated economic relevance that is key to set priorities. Given the broadness of these activities, results are fuzzy and complex to perform the mentioned analysis as all hypotheses are present considering the parts of
It is interesting to perceive that in this type of intervention, refurbishment, a small group of activities is responsible for a large part of the budget. When looking at these items in detail and setting the lenses of modelling and Data Templates, the items translate into very different situations. There are some that are pretty clear and others that are fuzzy. In this second case, it is impossible to perform evaluations, and one of the findings is that future projects using “less traditional” processes will need to provide a more detailed Bill of Quantities. This assumption also applies to “mix works”, as it is not always possible to individualize these activities in deconstruction and construction parts.
96
the stakeholder’s level, namely, to perceive and understand the benefits of their role and for the industry value-chain. As mentioned, to tackle difficulties, interoperability must be streamlined as there will be always differences on data and granularity options for Data Templates due to regional aspects, both in terms of construction solutions and design and construction practices.
It was observed that there are items that address the same type of elements, meaning that a single Data Template can be used to provide data for the construction and support relevant data for the deconstruction. Activities relating to exterior windows and interior doors are the most relevant, and that result from the sample. The roof could be another element, but in this specific situation, they are from different types/construction solutions, fitting in the context of different standards and Data Templates. Figure 6 is proposed as a summary of this vision where the multidimension analysis using common and specific factors found in Data Templates the support and common source of data.
6 CONCLUSIONS The multi-dimension analysis reveals that activities can have different weights/relevance depending on the applied lenses. The Data Templates mind-set is meaningful for the construction life-cycle, and the use of different lenses either for the construction as for the deconstruction allows a clear identification of the relevance, contributing in one to the DT priority definition and roadmap for digitalization implementation. The case study revealed that the most relevant Data Templates following a cost perspective could be resumed to eleven. This group ranges also the most relevant DT when the lenses is energy performance improvement, meaning that several analyses at this level but also at acoustics, and sustainability, among others, will benefit from this. There are also Data Templates that are common to both types of activities. This supports the vision that DTs are circular information enablers, boosting circular economy processes. The abovementioned group of Data Templates represents almost 50% of the total budget, where the most relevant part is related to new construction. There are limitations related to DT and DPP uncertainties, namely the absence of harmonized standards for some types of products as well as the granularity to be practiced relating Data Templates definition. The different constructive solutions in the built stock might also limit this case study to some types of buildings or specific locations. The way the BOQ are structured can also introduce some limitations on the analysis, as highlighted. The analyzed parameters do not constitute de overall environment of requirements but were selected because they were considered to be the most meaningful for this refurbishment project goals. Future research aims to detail and discuss this case study and its assumptions further, by bringing other methodologies. Detailed analysis focusing on other dimensions is to be performed, seeking the identification of key performance indicators to support guidelines and roadmaps. As well, the development of a framework to provide a view of the different lenses of one specific activity or product/systems is to be proposed in a short time. On a more practical level, this case study provided a clear understanding that with a small group of Data Templates it is possible to set Digital Product Passports that are meaningful for refurbishment projects in very different aspects/dimensions. This is key for all involved stakeholders to understand where to start and what are the more important characteristics to
Figure 6. Information circularity framework supporting circular economy processes using data templates.
From the analysis results, eleven Data Templates are worth being set as a priority. These are related to façade metallic windows and glass, blinds, wood profiles, aluminum sheets, interior wooden doors, stone elements, ETICS, ceramic tiles, paints, and mortar. Most of the aspects, as mentioned earlier, constitute Barriers to DTs implementation but are not directly related to them. At this level, factors associated with the definition of Data Templates granularity, data collection, technological improvements to assure interoperability, traceability, and stakeholders’ awareness are identified as the most relevant. On the side of the Enablers, the development of case studies such as the presented here, the training of stakeholders and the investment programs betting on built stock improvements and innovations implementation will stem the industry to make this part of the common practices. Associated with this are the Drivers for change, where the EU Policy and the funding for research and development projects at the EU and National levels will serve as support, incentive, and test bed for the changes. A significant part of the Challenges is at
97
European Commission. (2018). Guidelines for the waste audits before demolition and renovation works of buildings. UE Construction and Demolition Waste Management. Ref. Ares(2018)2360227 – 03/05/2018. Brussels. Retrieved from https://ec.europa.eu/docsroom/documents/ 31521/attachments/1/translations/en/renditions/native European Commission. (2020). A Renovation Wave for Europe. COM(2020) 662 final. Brussels. Fellows, R., & Liu, A. (2022). Research methods for construction (Wiley-Blac). Oxford. Government, P. (2021). Portuguese Recovery and Resilience Plan. Retrieved 9 February 2022, from https://recuperar portugal.gov.pt/ Group, D. S. C. in the B. E. (DSCiBE) U. (2020). Digital Supply Chains – Data Driven Collaboration. Retrieved from https://cobuilder.com/en/the-digital-supply-chaindata-driven-collaboration/ ISO. EN ISO 23387 Data templates for construction works entities of data templates, and how to link the data templates to Industry Foundation Classes (IFC), Pub. L. No. prEN ISO 23387:2018, 18 (2020). Switzerland. Llatas, C., Angulo Fornos, R., Bizcocho, N., Cortés Albalá, I., Falcón Ganfornina, R., Galeana, I., …Soust-Verdaguer, B. (2019). Towards a Life Cycle Sustainability Assessment method for the quantification and reduction of impacts of buildings life cycle. In IOP Conference Series: Earth and Environmental Science (Vol. 323). https://doi.org/10.1088/1755-1315/323/1/012107 Mêda, P., Calvetti, D., Hjelseth, E., & Sousa, H. (2021). Incremental Digital Twin Conceptualisations Targeting Data-Driven Circular Construction. Buildings, 11(11), 554. https://doi.org/10.3390/buildings11110554 Owen, R. (2013). Integrated Design & Delivery Solutions (IDDS). Rotterdam. Quiñones, R., Llatas, C., Montes, M. V., & Cortés, I. (2021). A multiplatform bim-integrated construction waste quantification model during design phase. The case of the structural system in a spanish building. Recycling, 6(3). https://doi.org/10.3390/recycling6030062 Rios, F. C., Chong, W. K., & Grau, D. (2015). Design for Disassembly and Deconstruction – Challenges and Opportunities. Procedia Engineering, 118, 1296–1304. https://doi.org/10.1016/J.PROENG.2015.08.485 Sousa, H., & Mêda, P. (2012). Collaborative construction based on work breakdown structures. In eWork and eBusiness in Architecture, Engineering and Construction – Proceedings of the European Conference on Product and Process Modelling, 2012, ECPPM 2012. Sousa, H.; Moreira, J.; Mêda, P. (2016). ProNIC on the Schools Refurbishment program – Contributions for the construction process improvement. In REHABEND 2016 – Euro-American Congress on Construction Pathology, Rehabilitation Technology and Heritage Management. Burgos.
disclose.This provides a clear roadmap for streamlined and short-term implementation in this type of projects. DataTemplates are information circularity enablers, and the performed IDDS identifies clear paths to overcome the challenges of its implementation. The road to a circular economy must integrate performance, digitalization, environment, and waste. The knowledge of What Data Templates come first for specific types of construction will support improvements in all these dimensions contributing to improved industry performance and a smarter built environment making better use of technologies and resources usage. ACKNOWLEDGEMENTS This research was funded by the following: 1. European Economic Area (EEA) Financial Mechanism 2014-2021, Environment, Climate Change and Low Carbon Economy Programme: 13_Call#2 GrowingCircle 2. Base Funding of the CONSTRUCT – Instituto de I&D em Estruturas e Construções-funded by national funds through the FCT/MCTES (PIDDAC): UIDB/04708/2020. REFERENCES Akyazi, T., Alvarez, I., Alberdi, E., Oyarbide-Zubillaga, A., Goti, A., & Bayon, F. (2020). Skills Needs of the Civil Engineering Sector in the European Union Countries: Current Situation and Future Trends. Applied Sciences 2020, Vol. 10, Page 7226, 10(20), 7226. https://doi.org/10.3390/APP10207226 Bassanino, M., Fernando, T., Wu, K., Ghazimirsaeid, S., Klobut, K., Mäkeläinen, T., & Hukkalainen, M. (2016). A collaborative environment for energy-efficient buildings within the context of a neighborhood. EWork and EBusiness in Architecture, Engineering and Construction – Proceedings of the 11th European Conference on Product and Process Modelling, ECPPM 2016, 299–308. Calvetti, D., Mêda, P., Sousa, H. (2021). Tech enablers to the EU Renovation Wave: Framework-based on the Communication (2020) 662. In 38th CIB W78 conference on Information and Communication Technologies for AECO. Luxembourg: LIST. Retrieved from http://itc.scix.net/paper/w78-2021-paper-061 CoBuilder. (n.d.). What is a Product Data Template. Retrieved 19 March 2019, from https://cobuilder.com/en/ what-is-a-product-data-template/ Commission, E. (2022). EU Construction Products Regulation proposal. COM(2022) 144 final. Brussels.
98
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Chaos and black boxes – Barriers to traceability of construction materials K. Mohn Department of Manufacturing and Civil Engineering, NTNU, Trondheim, Norway
J. Lohne Department of Civil and Environmental Engineering, NTNU, Trondheim, Norway
ABSTRACT: This paper identifies barriers to (digital) traceability in the construction industry today, based on literature a review, eighteen in-depth interviews representatives working at strategic levels from owners/main stakeholders, various suppliers, research institutions, contractors and industry organizations, and four in-depth interviews were conducted with other industries, which have gone further with the digitization processes than the construction industry, to see which successful methods they have used to succeed with traceability for their purposes and which ICT technologies have been used. A main challenge identified within the material value chains of construction projects is that most trading processes are analogous and determined by the contracted contractors and their subcontractors. Increased traceability requires increased client involvement. Without knowing which products are used in the buildings, it will prove impossible to understand how the buildings should report according to environment and other. This requires the development of digital processes.
1 INTRODUCTION A commonplace position in the literature today, when looking for the reasons for unethical or illegal practices within the construction industry, is to put emphasis on the low technological entry point for actors within the industry, thereby making it easier for criminal actors to access the industry (for an overview, see Lohne et al. 2019). Such perspectives are evidently correct; yet they leave out the crucial relationship between industry characteristics and unsolicited behaviour (see for instance Lohne et al. 2017). Of particular interest within this context are phenomena falling under the heading of concealment. As Ambraseys and Bilham, (2011) maintain, the “assembly of a building, from the pouring of foundations to the final coat of paint, is a process of concealment, a circumstance ideally suited to the omission or dilution of expensive but essential structural components”. Concealment, thus, lies at the heart of the construction industry. It is hard to overestimate the gravity of the challenges inflicted by concealment in the construction industry. Still, debates on the subject, with corresponding theoretical contributions, only seem to arise in the wake of tragic events. It is, for instance, in response to the tragic fire at Grenfell Tower that the true properties of materials in use in buildings come under debate. Based on observations on how unsolicited materials aggravated this incident, Watson et al. (2019) outlined a framework for traceability of all built assets. Concretely, they propose a follow-up on the idea of a ‘golden thread’ preserving critical information about design intent and the as-constructed building in a DOI 10.1201/9781003354222-13
proposed Digital Record (DR). Similarly, it is on basis of the six-digit death tolls resulting from the Haiti 2010 earthquake that Amraseys and Bilham (2011) carry out their analysis of the relationship between corruption-riven countries and a lack of resiliens to natural catastrophes within the context of the built environment. The evidence of widespread corruption throughout the construction industry is ample, as mapped by e.g. Chan and Owusu (2017) or Monteiro et al. (2020). Of interest here is the correspondence between the concealed nature of the construction industry practices and the existence of corruption: “Corruption is by nature covert” and the “construction industry — currently worth US$7.5 trillion annually and expected to more than double in the next decade — is recognized as being the most corrupt segment of the global economy” (Ambraseys & Bilham 2011). Countering concealment at all levels of the construction industry – concerning materials, organisation and transactions – is therefore paramount for all concerned with improving ethical standards. The ambition of this paper is to outline the possibility conditions for traceability to counteract concealment. According to the Oxford English Dictionary, traceability is the capability to trace something. It can be defined as the ability to verify the history, location, or application of an item by means of documented recorded identification. In the supply chain, traceability may be both a regulatory and an ethical or environmental issue. Traceability is increasingly becoming a core criterion for sustainability efforts related to supply chains wherein knowing the producer, workers and other links
99
stands as a necessary factor that underlies credible claims of social, economic, or environmental. At the reverse side, without traceability, actual tracking of especially materials is hardly possible. A common theme in the construction management literature, well illustrated by the analysis of Bäckstrand and Fredriksson (2020), is the challenges involved in knowledge sharing within the construction industry. In fact, despite information sharing is playing a crucial role in construction, sharing proves difficult and rarely prioritized. The results are depressing. Following the analyses of Thunberg and Fredriksson 2018 and Thunberg and Persson (2014), the authors maintain that poor information sharing resulting in poor delivery performance leads to that not even 40% of deliveries are flawless (Bäckstrand & Fredriksson 2020). Based on the above, this paper addresses the following research questions, namely 1) What is the state of the art of materials traceability within the construction industry today, 2) What are the main challenges to traceability within the construction industry materials supply chains today and 3) What main avenues can be envisaged for addressing the challenges identified? 2 THEORETICAL FRAMEWORK Specific supply chains are determined by the role they are to fulfil. As such, it is crucial to recognize the specificities of such chains in the context of construction industry projects, which “faces wide fluctuating demand cycles, project-specific product demands, uncertain production conditions and has to combine a diverse range of specialist skills within geographically dispersed short-term project environments” (Dainty et al. 2001: 163). These fundamental challenges change little over time. What is changing, are the potential of the technological solutions implemented to address them. In the following sections we, firstly, present an outline of the challenges pertaining to traceability of construction materials through its trade processes according to descriptions in authoritative, overall industry reports, secondly, present some thematic trends from the contemporary scientific literature, thirdly, introduce a telling example on the subject of e-trade, before, fourthly, outlining in what we consider the knowledge gap within the area to be. 2.1 Authoritative industry reports Authoritative industry reports in form of white papers are a particularly valuable source of understanding for those wanting an understanding of the overall challenges to the industry covered. A series of industry reports have documented challenges within the trade processes. The Latham (1994) report is thorough in describing aspects of ameliorating trade processes, especially from the perspective of procurement; analyses of the modes and effects of innovative procurement routes are ample. Still, trade processes are here mainly
considered from a productivity perspective: e.g., “All the proposals [presented here] will have an effect on productivity.” (p. 80). Still, it points to the need to develop “a more structured, standardised and ethical approach to the procurement and management of subcontractors.” (p. 83) This move towards structured and standardised approaches to procurement is followed up in the Egan report (1998) and its call for innovative collaboration approaches, first and foremost forms of partnering and PPPs. Such approaches are found to be efficient to “tackle fragmentation”, implicating the potential to standardise procurement and management better than “traditional contract-based procurement and project management” (p. 9). The Egan report is, however, silent on the ethical implications of new procurement approaches. Saxon (2005) underlines among other aspects the importance of balancing investment cost and life cycle costs. The central idea governing the report from Saxon is the concept of value, professing that the construction industry and its clients should focus more on the value produced by assets over their life-time than purely on their price and cost (being described as the present situation). Central to the challenge in the proposed turn are trade processes, where“[d]evelopments in construction procurement methods such as the Private Finance Initiative (PFI) have introduced the construction industry to the need for a ‘whole life cost’ approach to the procurement” (p. 27) Wolstenholme (2009) maintain that reforming procurement within organisations is easier said than done. There is a strong tendency, for instance, that “[w]hile the leadership of public organisations may be committed to the idea of best value, their procurement teams often still want to achieve lowest price.” (p. 14). Equally, different actors within the process typically will have different interests, where for instance contractors generally “procure in order to pass risk down the supply chain, rather than to draw up opportunities to create value by working as an integrated team” (p. 22). Alternative practices do, however, exist. Wolstenholme (2009) reports from a study tour led by Constructing Excellence to Japan, identifying practices that could be applied elsewhere. Within the context of this paper, their identification of “[t]ransparent procurement, contract and payment processes” along with “standardised processes, cost databases and standard procedures” (p. 21) as a key point of learning. This call for transparency in procurement seems to fall within the general need for a “stronger ethical stance within the industry”, avoiding the image of the industry as “being excessively focused on the bottom line” (p. 17). The World Economic Forum (2016) lament the slow adaption to and adaption of new technologies within the construction industry, both for its consequences on productivity, and for the implications it has on unwarranted behaviour among actors. The report underlines strongly the consequences of concealment in trade processes in general and for procurement in particular.
100
Besides advocating strong prosecution practices, the report recommends “the issue of corruption on construction projects can only be resolved by creating a corruption-resilient procurement environment, by implementing fair and transparent procurement procedures” (p. 10). In fact, even though “[b]ribery and corruption exist in all industries”, in the construction industry procurement “collusions between government staff and bidders are particularly common, even in developed countries” (p. 47). 2.2 Some trends in the literature The literature search revealed ample evidence that the challenges identified above have been addressed, at different levels and from different perspectives. At a practical level, for instance, Thunberg and Fredriksson (2018), outline several “supply chain-related problems facing the construction industry, such as poor construction site logistics, lack of communication and trust. These problems can jeopardize construction projects through delays and cost overruns”. At a systems level, Briscoe and Dainty (2005) has pointed out that “the large number of supply chain partners and the significant level of fragmentation limit the levels of integration that are achievable”. Measures to address the challenges have varied correspondingly. There is for example, ample evidence to be found for the procurer in need for practical frameworks etc. Equally, a need for project overarching measures is commonly proclaimed, such as by Dubois et al. (2019), concluding that “there are possibilities to enhance efficiency […] by widening the scope of coordination beyond the individual construction site”. Concerning both quality and the potential to corrupt practices described above, questions of materials quality have been raised. Particularly interesting here is the works pertaining to illicit or false materials, such as described by Minchin et al. (2013); CII (2014); Engebø et al. (2016; 2017) and Kjesbu et al. (2017a; 2017b). Of particular interest within the context of this paper is papers addressing root causes to the challenges experienced. Thunberg and Fredriksson (2018), for instance identified information sharing between actors, e.g. from the main contractor to suppliers, leading low information availability among the suppliers, as a major hindrance to chain performance. In order to address such fundamental challenges, Bäckstrand & Fredriksson (2020) maintain that there “is a need to increase understanding of construction suppliers’ coordination needs and their present information availability”. Digitalisation is often pointed to as a potential panacea for the challenges experienced. A typical example of the efforts made following this path is the work of Yevu et al. (2021). 2.3 E-procurement barriers – Yevu et al. (2021) Yevu et al. (2021) conducted a comprehensive analysis of the literature on the barriers to e-procurement adoption in the construction industry and reveal the interrelationship patterns of the barriers in literature. As
such, it resembles the ambitions resulting in the present paper, by focusing research interests on barriers to technology adoption. Yevu et al. (2021) underline how “synthesized knowledge for proper understanding of complex interrelationships between barriers to e-procurement adoption […] is limited”. In addition to this, they underline how “the limited understanding and information on the interrelationships between these barriers in literature, may increase the tendency of ignoring critical questions and areas that might improve research and industrial practice.” (2021:1). Given that “ these barriers do not influence e-procurement adoption as standalone factors”, then approaches focusing “individual barriers may not be comprehensive in showing the complexity of how these barriers influence e-procurement” (2021:2). The main body of their review, then, consists in an assessment of different barriers identified – 34 in total – and their distribution among the publications examined. These were assembled into four categories, notably organizational, technological, cultural, and legal and security barriers. Further, a network analysis was employed to depict the complex interrelationships among the barriers. As such, the analysis presents a comprehensive image of barriers identified by the literature that hinders e-procurement, and, to a certain extent, their relationship. The analysis of Yevu et al. (2021), however, illustrate a lack of willingness to address root causes to the challenges experienced. 2.4 Knowledge gap Given the above, it is manifest that the degree of transparency in the construction industry is lacking. In effect, the actual workings of the system permits for concealment at multiple stages in several aspects, including concretely the actual materials used to erect assets, their installation at construction sites and the workings of the materials supply chains. Concealment, in short, concerns both products, their use and the processes governing both their usage and their procurement. The nature of this concealment, however, what is concealed and how it is concealed, is largely unknown. 3 METHODS 3.1 Literature review The research reported on in this article was underpinned by a literature review based on using Google Scholar and scrutiny of leading journals addressing the field of supply chains in the construction industry. In addition, the proceedings of the IGLC conferences were scrutinized. These publication channels were searched using key terms such as “transparency”, “corruption”, “traceability”, “procurement”, “trade processes” “systemic change”, “construction industry” (and synonyms) and “Norwegian”, alone or combined using Boolean operators. Identified articles were utilized in the search using snowballing techniques (backwards and forwards), according to Wohlin (2014). The
101
search was carried out from September 2021 to February 2022 and resulted in 30 journal articles, conference articles, books and reports considered directly relevant for the theoretical framework. 3.2 Interviews During 2020-2021, a series of 18 informal in-depth interviews was carried out with central actors within the Norwegian construction industry and actors in industries who have witnessed similar challenges to those of the construction industry. In addition, four indepth interviews were conducted with other industries, which have gone further with the digitization processes than the construction industry, to see which successful methods they have used to succeed with traceability for their purposes and which information technologies have been used. These interviewees were selected on basis of hierarchical seniority (belonging to the upper echelons in their respective compagnies or being self-employed in some capacity, typically as senior ad-visors); their work-life experience (with the intention to grasp recent developments in a historic light, going back at least the last 35-40 years) and relationship to the use of digital solutions within the industry (in different capacities, such as supply chains, design, modelling etc.). In sum, the ensemble of the interviewees was considered sufficient to provide a good cross-sectional idea of the latest developments in the industry, both concerning enablers and barriers. On a practical level, the interviews were all conducted face-to-face, and lasted between one and four hours. As a result of the sensitivity of the subject matter and the potentially easy identification of interviewees given their stature, confidentiality was underlined as key to all interviewees. Correspondingly, effort has been made to separate recognizable interviewees from the presented analysis. The results provided thus represent a mélange of the different opinions, without clear personal attribution possible. 4 FINDINGS 4.1 Salient characteristics of contemporary construction industry trade Actors including manufacturers, wholesalers and building warehouses, have also had for a long time their own agency agents and importers. These have had who have defined services, for which they have been recompensed. In the recent past, global online stores have also emerged, in which all levels from pure fraud to professional business are represented. 4.2 Mapping of main challenges in the Norwegian construction industry 2021 The following mapping of main challenges is based on in-depth interviews with Norwegian construction
industry professionals, as outlined in the methods section. In this section the challenges will be presented separately, whilst in the next section we outline them according to an analytic overview.The main challenges according to industry professionals are the following: – There are significant transports to the construction site of used products that should not be used in the building (formwork materials, scaffolding, products related to equipment used by the contractor, etc.) – There are no specific receiving routines for control of orders – There are no concrete records of products actually used (applied concrete, what built-in where of products, waste in production (waste management) – Today’s contract formats are not product specified. – Only about 7% of e-commerce is linked to the professional market, and largely applies to the purchase of workwear, protective equipment, and electrical tools. The rest is for private customers. – Price is often decided upon before product is specified, so that there will be reiterations before the ordering processes have ended. This entails a high risk of delivery of products ill suited – System orders such as fire walls, climate walls are often adjusted down to partial deliveries due to ongoing bargaining a single product leads to part of the system orders being removed. This also means that building owners do not get approved documented deliverances from orders – Products, which are built into buildings today are not traceable. This means that today we do not know what the buildings consist of. Typically, only the packaging is marked, not the product itself. – Traceability is the entrance ticket for industrialization processes so that the products’ stated qualifications and properties can be satisfied. – The trading processes consist of parallel process looks with all six different actors with a high degree of re-fights, bargaining, threats, etc. and undocumented products (often called fake products) abounding. This is effected by many players playing out against each other, both with “multiple competitions” for the same delivery, but with different approaches. This also causes the use of resources in the procurement processes to be unnecessarily high, which increases the price level of the products. – Price is often determined before product. This tempts one to deliver products of different quality and function than what is described, to avoid losing money directly. In this way, real products essentially become fake undocumented products – Building system requests are divided before delivery into individual items in these deliveries being withdrawn. The fact that this also means that, for example, a fire wall is no longer certified, due to it being divided from different supplier links, where no one is longer responsible for two, does not seem to affect the executor, and is kept hidden from the
102
owners and managers, who will have the buildings in use. – About 35% of deliveries to the construction site do not have reception checks from the executors, i.e. the contractors/craftsmen. – It is known that the packing slip is used as a documentation on the use of high-quality products, but these have been returned, and the recipient also receives some “hand money” for such returns, while at the same time ordering low-quality products, which are then used in the builders. This is outright corruption of widow-read variety. – The GS1 standards, which are in use are not complete, so also the information carriers for traceability are insufficient. – There are no legal requirements for the identity of products used in buildings to be judged. As such, the Norwegian Building authority (DiBK)’s cautious demands for this have been diluted by active actions by the executing actors. Moreover, the documentation requirements for “reused” products have been revoked, although it is very difficult to judge whether an undamaged product is a new or reused one. All these processes are today without the involvement of the real procurers. Owner, managers and tenants of buildings. These processes may also be unknown to many of these requesters.
although the requests and orders can be made by the owner, managers, architects, craftsmen, users etc. How this is done is thus neither visible nor transparent to owners and/or managers. Equally interesting, is that product documentation and the authenticity of it is based on trust. It is easiest to describe this as “black-box” purchasing, where “blackholes” hide the real realities. This is described in bullet points in the listing made in section 4.2. All documentation to owners and managers also stops at best on what has been delivered to the construction site, and nothing about what is really built into the buildings. 4.4 Fundamental possibility conditions for digital trade Further, based on the reported main challenges as reported by the industry professionals, the black box described can be deconstructed into three different main categories, namely challenges pertaining to transactions, materials and functions.
4.3 The black box: Figural representation of the main challenges Salient traits from the above observations can be presented schematically as in Figure 1:
Figure 2. The black box deconstructed – an illustration of fundamental conditions of possibility for digital trade solutions.
Figure 1. Simplified representation of present-day procurement routes as experienced in practice. Actual transactions are characterized by a lack of transparency.
In practice, all purchases are made in the silos of the contractors for new construction and rehabilitations,
Based on the above, the challenges falling into the black box can be attacked following the deconstruction into three parts. Firstly, there is a need for transparency concerning the transactions involved. Secondly, there need to be tracability sufficient for assuring actual tracking of materials into the built asset. Thirdly, there needs to be sufficient control over what functions product assemblies fill in built assets.
103
The challenges are so comprehensive that each of these must first be addressed and solved separately, and then seen in context. Requirements for solving challenges in these three parts include, firstly, requirements for electronic labelling of products, machine-readable data capture and machine-readable data recognition (barcodes, QR codes, RFID, etc.). This is absolutely necessary to achieve data sharing of product information. Without it, such processes will be far beyond financial limits. The ISO 19650 series can provide guidelines on how transactions and functions/roles can be handled, but this is a new framework standard, which will require more use to develop best practice.
5 DISCUSSION AND CONCLUSIONS In this paper, we set out to address firstly, the state of the art of materials traceability within the construction industry today, secondly, to identify the main challenges to traceability within the construction industry materials supply chains today, and, thirdly, to identify main avenues for addressing the challenges identified. According to the brief going-through in the theoretical framework section and on from what interviewees maintained during interviews, materials traceability in today’s construction industry is at best dubious. A main idea governing the work presented in this paper is that e-commerce in the form presented by Yevu et al. (2021) will work with e.g. workpants since 1) it’s okay with non-descript objects, and 2) it’s a simple transaction. Today’s construction industry, however, does not work according to this way of operating. The construction industry requires complex, project specific products filling functions whose requirements are rigidly described in regulations. Following this, the analysis presented outline that the most important challenges to traceability lies in the concealed nature of the trade processes employed. Procurement, for instance, takes place amongst several different actors, and at various stages along the construction process. As such, the very structure of present-day procurement routes constitute a major hindrance for the introduction of digital trade. This is a root cause to challenges, leaving digitalisation of trade processes (such as the e-procurement described by Yevu et al. (2021)) impossible to implement. A main avenue for sorting the challenges identified goes through addressing the multifarious nature of trade within the industry as it occurs today. A pre-condition for establishing any sort of functioning digital trade is the identification of fundamental conditions for trade solutions, as illustrated in Figure 2 and outlined in chapter 4.4. Rather than addressing single barriers independently, there is a need for going to the root causes of the challenges, foremost of which identified here as traceability. There is traceability so that the products can be traced through the trading processes, which is the key.
Addressing practices arising from a lack of transparency needs by setting such requirements for information management throughout the value chain. Of utmost importance for future research and practice is understanding how to include perspectives of the building in use into the materials value chains. Without concern for the asset in use, achieving sustainable construction will prove impossible. Based on the above, we maintain that there is a great need for future research. Three pathways stand out as of immediate concern, notably, firstly, to understand how to enable tracking on how to make products from raw materials, secondly, to understand tracking of orders for installation in buildings and, thirdly, to understand “washing processes”, such as to solve loose digital processes from raw materials to products in construction. REFERENCES
104
Bäckstrand, J., & Fredriksson, A. 2020. The role of supplier information availability for construction supply chain performance. Production Planning & Control, 1–12. Briscoe, G., & Dainty, A. 2005. Construction supply chain integration: an elusive goal? Supply Chain Management: An International Journal. Chan, A.P. and Owusu, E.K., 2017. Corruption forms in the construction industry: Literature review. Journal of Construction Engineering and Management, 143(8), p.04017057. CII (2014). Mitigating threats of counterfeit materials in the capital projects industry. CII Research Report No. 30711, Austin, Texas. Dainty, R J, Millett S J and Briscoe G H 2001. New perspec-tives on construction supply chain integration, Supply Chain Management: An International Journal, 6(4), 163–173. Dubois, A., Hulthén, K., & Sundquist, V. 2019. Organising logistics and transport activities in construction. The International Journal of Logistics Management. Egan, J. (1998). The Egan report-rethinking construction. Report of the construction industry task force to the deputy prime minister. London. Engebø, A., Lohne, J., Rønn, P. E., & Lædre, O. 2016. Counterfeit materials in the Norwegian AEC-industry. In 24th Annual Conference of the International Group for Lean Construction. Engebø, A., Kjesbu, N., Lædre, O., & Lohne, J. 2017. Perceived consequences of counterfeit, fraudulent and substandard construction materials. Procedia Engineering, 196, 343–350. Kjesbu, N. E., Engebø, A., Lædre, O., & Lohne, J. (2017a). Counterfeit, fraudulent and sub-standard materials: the case of steel in Norway. In 25th Annual Conference of the International Group for Lean Construction, Heraklion, Greece (pp. 805–812). Kjesbu, N. E., Engeb, A., Lædre, O., & Lohne, J. 2017b. Countering counterfeit, fraudulent and sub-standard materials in construction: Countermeasures to avoid the use of counterfeit, fraudulent and sub-standard steel materials in the Norwegian construction industry. In 2017 12th Interna-tional Scientific and Technical Conference on Computer Sciences and Information Technologies (CSIT) (Vol. 2, pp. 92–99). IEEE.
Latham, M., Egan, J., Harris, F., McCaffer, R., Cooke, B., Williams, P., Fryer, B., Walker, A., McGeorge, D., Palmer, A. and Womack, J., 1994. The Latham Report. Constructing the Team. Lohne, J., Drevland, F., Åsgård, T., Hoven, T., Lædre, O., & Ryum, T. 2017. Homines Oeconomici or moral saints? On the purpose of educating civil engineers. In Proceedings of the 9th Nordic Conference on Construction Economics and Organization (pp. 362–374). Lohne, J., Kjesbu, N. E., Engebø, A., Young, B., & Lædre, O. 2019. Scoping literature review of crime in the AEC industry. Journal of Construction Engineering and Management, 145(6), 03119002. Minchin Jr, R. E., Cui, S., Walters, R. C., Issa, R., & Pan, J. 2013. Sino-American opinions and perceptions of counterfeiting in the construction supply chain. Journal of Construction Engineering and Management, 139(1), 1–8. Monteiro, B. K., Masiero, G., & Souza, F. D. 2020. Corruption in the construction industry: a review of recent literature. International Journal of Construction Management, 1–9. Oxford English Dictionary (Online ed.). “Traceable”. Oxford University Press, consulted 120522 Saxon, R. 2005. Be Valuable: A guide to creating value in the built environment. Constructing Excellence, London Thunberg, M., & Persson, F. 2014. Using the SCOR model’s performance measurements to improve construction logistics. Production planning & control, 25(13-14), 1065–1078.
Thunberg, M., & Fredriksson, A. 2018. Bringing planning back into the picture–How can supply chain planning aid in dealing with supply chain-related problems in construction? Construction Management and Economics, 36(8), 425–442. Watson, R., Kassem, M., & Li, J. 2019. Traceability for built assets: proposed framework for a Digital Record. In Creative Construction Conference 2019 (pp. 496–501). Budapest University of Technology and Economics. Wohlin, C. 2014. Guidelines for snowballing in systematic literature studies and a replication in software engineering. In Proceedings of the 18th international conference on evaluation and assessment in software engineering (pp. 1–10). Wolstenholme, Andrew, Simon A. Austin, Malcolm Bairstow, Adrian Blumenthal, John Lorimer, Steve McGuckin, Sandi Rhys Jones et al. 2009. “Never waste a good crisis: a review of progress since Rethinking Construction and thoughts for our future.” World Economic Forum 2016. Industry Agenda – Shaping the Future of Construction: a breakthrough in mindset and technology. WEF, Geneva. Yevu, S. K., Yu, A. T. W., & Darko, A. 2021. Barriers to electronic procurement adoption in the construction industry: a systematic review and interrelationships. International Journal of Construction Management, 1–15.
105
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Evaluating four types of data parsing methods for machine learning integration from building information models F. Sellberg, J. Buthke & P.F. Sonne-Frederiksen Link Arkitektur, Stockholm, Aarhus, Copenhagen, Sweden, Denmark
P.N. Gade University College of Northern Denmark, Aalborg, Denmark
ABSTRACT: A method and structure for architectural datasets specifically designed for the analysis, sorting, and ultimately reusing of building elements is proposed. Four different methods of parsing data from real-life projects using their building information models (BIM) for integration into a machine learning (ML) model were evaluated. As ML integration is becoming more important in the Architectural Engineering and Construction (AEC) industry, we see an increasing demand for high quality datasets. Four different methods and file formats were benchmarked, focusing on read and write-speeds for converting architectural BIM into datasets to be used in ML. Our results show that the current way of storing our projects in Industry Foundation Classes (IFC) is not optimal for the development and integration of new Artificial Intelligence (AI) assisted tools. This paper provides alternative methods and storage solutions for both developing new datasets internally and also for future work in creating a common federated learning setting for the AEC industry.
1 INTRODUCTION The construction industry is becoming increasingly complex with more and changing requirements with regard to climate change and sustainable needs. These new requirements further necessitate that the information created and exchanged is of high quality with regard to reliability and consistency. Moreover, the information required to create and manage building designs is continually increasing and requires that it can be accessed and edited by that user quickly. A major problem related to managing information for building projects is the time spent by designers in managing information. A study by Flager et al. (2007) found that designers, in general, spend about half of their time managing information. Hereby, designers are not spending their time creating new design information or doing analysis. Instead, the designers are often burdened by merely moving information around to ensure consistency and quality, i.e., coordinating existing information. In order to reduce the time designers, spend on managing information, new methods like Artificial Intelligence (AI) can be applied to automate this process, making it more efficient and ensuring highquality information across the building projects (Song et al. 2018; Zabin et al. 2022). While AI is noticed in many larger construction companies, it is still considered a fringe technology that is slowly being implemented in the industry (Abioye et al. 2021; Molio 2020; Natonal BIM Standards 2020). One of the major barriers to implementation is that the technology, in
106
many aspects, is immature and requires much skill to apply. As Abioye et al. (2021) argue, new roles in the construction industry need to be formed in order to cater to utilizing the benefits of AI. Moreover, acquiring these skills is also a major barrier due to the general talent shortage of people with skills in AI. Kyicska & Tsiutsiura (2021) argue that in order to better make use of AI in the construction industry, there needs to be a better understanding of what the users need. This can be done by experimenting with AI in order to identify potential solutions of usingAI for relevant cases for the industry. Specifically, machine learning (ML), a concrete AI methodology, has been promising for various use-cases in the construction industry to solve various issues by using learning data to train algorithms to make predictions that can be used in making decisions. 1.1 Data model Open datasets for use in the AEC industry are scarce and often consist of projects in a very isolated scope of context. The few high-quality datasets that have been made are often made for a large urban scale. Lu et al. (2019) showed an approach using convolutional neural network where 10.000 images was used from a case named 5M-Building as a dataset to detect buildings in pictures. However, such an approach is difficult to implement in a more common architectural scale. To tackle this issue, we propose that datasets can be created from a resource that most firms already have, their existing projects and building information models (BIM). DOI 10.1201/9781003354222-14
In our methods of converting from a BIM, we evaluate what file storage is relevant and how they can be used for a ML context. In an article by Wang & Tang (2021) it was suggested to save BIM information based on IFC into databases for long term storage using Java language and MySQL database. Based on that they created a prototype called IFCParser. Creating the prototype, they found that it assisted engineers that weren’t knowledgeable about IFC to easier to get and store BIM information on their own servers to focus on solving problems. Withers (2022) discuss in that a global shift is happening where assets move from tangible to intangible. Through new ways of storage, a more easily sold and quantifiable asset can be created for architectural firms in the way of filtered project data. In order to find a more optimized and stable file format for both long term storage and fast integration to new development of tools with a special emphasis on development of Machine learning methods on large architectural data sets. Easing designer’s workload from managing complex data parsing from building information format to another. This can be done by automating this work by using ML. In this article we measured four different file formats with our key metrics for evaluation, which are file size, write-speeds from native BIM software formats to a new file format, and read-speeds when loading it into a ML model to showcase the potential advantages of automating data parsing across the different showcased filetypes. These insights can potentially highlight the different approaches to using ML for data parsing in the AEC industry to help alleviate current challenges. 2 METHODOLOGY Our prototype is developed and evaluated using information from real-life building projects in Sweden and Denmark. Such an investigation can give insights into practical applications of ML in improving the handling of building project information stored in BIM. The proposed method for evaluation consists of two steps; the first is exporting a BIM into four commonly used file formats. The file formats were picked from common architectural use or in data science. In this process, we are evaluating the write time and file size. As a second step, we are loading the new exported files into a ML model consisting of a recurrent neural network and evaluating the read time of our data set used in the training of the neural network for each storage solution. 2.1 File formats Our four file formats that we evaluate are: 2.1.1 Industry Foundation Classes (IFC4) IFC4 is the most used file-format for interoperability and long-term storage of projects in architectural practices today. IFC is an object-based file-format
for description of architectural data. The data structure is formatted to easily be read by a multitude of software through the creation of standardized element definitions. 2.1.2 Json (2020–12) Json is a lightweight open standard file format using attribute value pairs, often used in transmitting data in web applications. It is not the most widely used format for machine learning but was chosen for its high human readability, flexibility, and ease of implementation. 2.1.3 Speckle (v2) Speckle is an open-source cloud-solution for BIM. It is used to stream projects onto a cloud server for interoperability and can be used as a long-term cloud-storage. It uses a similar structure to an IFC; for this specific format, we only evaluate write and read-speeds with no size comparison since a cloud-based database has no project specific file size related to it. 2.1.4 Petastorm (0.11.2) Petastorm is an open-source data access library using Apache Parquet datasets originally used for real-time deep learning for self-driving vehicles. It uses many modern approaches to optimize high velocity data feeding, as described by Qiu & Sun (2015), into a ML model such as local caching and sharding. 2.2 Neural network prototype The ML model is dependent on the individual use case. Here, for this example, we chose to train the model to predict CO2 emissions of different building elements. For that, we are feeding the model with dimensional information (width and volume) and information about object name and material. This means that we are dealing with two different kinds of data. On the one hand, we have number values, and on the other, we have ‘strings’ (text). Because an ML model can only handle numerical values, the text first needs to be converted and processed. This is done by converting each character into a one-hot encoded vector and then parsing those vectors through a recurrent neural network (RNN) to get a single classification of that string (see Figure 1). This classification is then used together with the other dimensional values to calculate a final CO2 emission by parsing those values through a fully connected neural net. 2.3 Prototype evaluation We are evaluating the prototype according to the following variables: write-speeds from the BIM to a storing format; storage sizes of each format and how they can be stored; read-speeds from each format into a ML model; and test run each to see how they perform with respect to time. To evaluate our method, we chose five larger projects in Denmark and Sweden that are all in the process of being or have been constructed and have a focus
107
Figure 1. Neural network prototype with Recurrent Neural Networks (RNN). Table 1.
Selected architectural projects and data.
NAME
LOCATION
YEAR
FUNCTION
SQUAREMETERS
SOFTWARE
LIDL VIGGBY LOJOBACKEN FRIPLEJEHJEM STRANDBOLIGERNE TV BYEN
Sweden, Stockholm Sweden, Stockholm Denmark, Haderslev Denmark, Copenhagen Denmark, Copenhagen
2020 2022 2020 2021 2022
Commercial Residential Residential Residential Residential
4323 m2 8045 m2 6633 m2 3284 m2 10312 m2
Archicad Archicad Revit Revit Revit
on sustainability. Our selection consists of residential projects, as they contain more elements to sort our evaluation and have more complex interior structures (see Table 1 for more detailed project information).
3 RESULTS 3.1 Write-speeds Write-speed is the time (measured in seconds) that it takes to export from Revit into the specified file format (see Figure 2). For Json and Petastorm, only walls
were exported, as our ML model was being trained on predicting kg/CO2 for walls. For IFC and speckle, all elements were exported, as is the standard of the formats. To simplify the comparison of the different formats, a graph of write-speeds in relation to the output file size was created to see how fast each method breaks out the relevant information. Something to notice is the size of the file created for each relevant dataset. Petastorm is the smallest, with an average of 0.026 MB, and IFC is the largest, with an average of 101.962 MB. As speckle is using a cloud storage solution, it is excluded from the time/size comparison graphs.
Figure 2. Write-speeds and write-speeds by file size for the formats.
108
3.1.1 IFC4 IFC, with an average write time of 310.9 s, has the longest write time of all the file formats in the comparison. This is largely because in our evaluation method, we exported all elements in our projects. When looking at write time/size, we can see that it performs the best. Because it is a file format made for interoperability, optimized for exporting large and complex projects, as well as being largely supported by BIM software, this was to be expected. 3.1.2 Json (2020-12) Json, with an average write time of 77.8 s and an average file size of 0.249 mb, provides architects with a fast method to export smaller segments of relevant information in architectural projects. When comparing write time/size, Json is the slowest in filtering out and storing relevant elements from the BIM. One would expect slightly faster speeds, as the format is made and optimized towards data interchange, but as the format provides a more human-readable text than the other formats, some overhead in the files is created, which shows in the write-speeds. 3.1.3 Speckle (v2) Speckle has the fastest average write time of 74.7 s with no apparent file size because it uses a cloud solution for saving the files, which obfuscates the file sizes. Something to note here is that the largest project “Lojobacken” was not correctly exported, so the longest write times are missing. This shows the main weakness in the system, i.e., the BIM projects need to be of a high quality to be able to be uploaded to Speckle. Write times include time to upload the project to the cloud, so it is heavily dependent on network speeds. In a similar fashion to how we treat IFCs, we are evaluating write times of an entire project. With
write-speeds being significantly faster than an IFC export and keeping all relevant information for both analysis and interoperability, this format is very promising. 3.1.4 Petastorm (0.11.2) Petastorm results in write times that look very similar to the results from Json, as they share a similar code base for filtering and exporting from BIM. Something to notice here is the slightly faster write-speeds and the significantly smaller file sizes, as a parquet-based database is being utilized instead. Write time/size is slightly skewed, as file sizes are so small in relation to the amount of information in them. 3.2 Read-speeds Read-speed is the time (measured in seconds) that a ML model takes to read and import the data from the different formats (see Figure 3). For IFC and Speckle, a filtering of walls had to be made, while Json and Petastorm already were filtered to only include walls in exporting to the file formats. To help compare the different formats that use a slightly different method of importing, a graph of read-speeds in relation to the input file size was created. As Speckle is using a cloudbased storage solution, it was excluded from the read time/size results. 3.2.1 IFC4 The average read time of IFC is 7.0 s, which is very fast considering their large file sizes. As the file system is made to be able to quickly import large files between software, this could be expected. A conflicting relationship can be seen through the results in that the larger the number of unique elements in the file gets, the read-time becomes exponentially longer. For large scale projects or aggregated projects, this will become an issue as the read times increase dramatically.
Figure 3. Read-speeds and read-speeds by file size for the formats.
109
3.2.2 Json (2020-12) Json has an average read time of 0.004 s and very small file sizes, as it only has the already filtered information in them. The lowest read time/size can still be seen. No decreasing speeds in relation to an increase in the number of elements in the files can be seen. For larger projects, Json would be the preferred file format. 3.2.3 Speckle (v2) Speckle provides the slowest read times at an average of 24.8 s, while this largely depends on the network connection. A relation to write times can be made where Speckle performed significantly better than IFC exports. No file size exists to compare against, but the number of elements in both IFC and Speckle are the same, as shown in Figure 3. IFC is performing more than 4 times faster than Speckle in filtering and importing the elements. 3.2.4 Petastorm (0.11.2) Petastorm has an average read time of 0.096 s, which is slightly slower than Json. This is mostly because it must convert and load the data into a structured spark data frame. When comparing read-speeds to file sizes, we can see a decrease in the evaluation metric the larger the project is. This is because the slowest process in the method is in creation of the data frame itself which the larger the file gets becomes a smaller process. So, for larger projects, a faster value in regard to read-speeds/size can be expected. Many of the overhead functions provided by the format, such as real time updates and sharding, are not used but are functions that can be used to heavily improve dataset processing in very large data sets.
4 DISCUSSION AND CONCLUSION This research plays a crucial role, as it constitutes a part of the conceptual basis for a new way to build up large datasets optimized for ML algorithms to read and write faster. The digitalization of many processes in the AEC industry is increasing, but very large unstructured datasets are very common, such as the dataset on five million buildings (Lu et al. 2019). To be able to fully leverage those datasets, the industry requires fast ways to search through the datasets and extract the specific information that is needed. One field of application is the reuse of building parts between building projects, where specific reusable building parts could be found and matched between datasets (1. Representing the building to be erected 2. Existing building acting as material banks). This process could potentially support a more circular future in AEC. Four methods were developed for the export and import of five real life projects into four different file formats, which then were loaded into a ML model predicting kg/CO2 in wall elements. Here, we outline the
possible use case scenarios of each format for the AEC industry. IFC current usage scenarios of full project interoperability and long-term storage with 3D information might not be its best use cases. With long write times but fast read times, the format is better suited for importing information into the analysis of complete projects such as ML models trained on tagging untagged building elements. A problem with IFC files is that they include very detailed data for commercial data purposes, where a more anonymized approach is required. For development of new tools where data from multiple projects are required, another file format should be applied. Petastorm usage is great for very large, aggregated datasets for ML with a predetermined function. With each element in the data frame needing to be predefined and converted into a tensor, a large technical understanding is needed to set up the exported data to be able to be both exported and imported in a correct way. Database approaches are great for small file sizes; this further supports the use in very large datasets, which is not something that is common in the AEC industry. Being a newly and niche developed system, integration into common software and libraries is not fully extended, giving it a clear disadvantage for developing tools connected to the format. Json has several use cases; in a ML context, its best application would be in aggregating filtered parts of projects where its small file size and fast read and writespeeds can be utilized. Examples would be in predicting wall material compositions or CO2 emissions. Another clear use case is in tool development where its common use in the field is a large contributing factor. Large shares of existing software and libraries already provide integration so development time can be decreased. Because data are highly adaptable and easy to make anonymous, Json would be the preferred format for commercializing project data for selling/buying between firms. Speckle provides interesting use cases where a fast write-speed but slower read-speeds gives it good use cases ranging from interoperability to the creation of non-platform specific tools. For inoperability and long-term cloud storage, the format excels, where read times are not as important and the flexibility of the format is more prioritized. A note is that the format is dependent on externally developed connectors for importing and exporting data, so a more optimized filtering and reading can be developed to further decrease read times. It has a low barrier of entry in the use and development, though a higher technical understanding is needed when optimizing and developing its connectors. Hawkins (2020) proposed a method for finding the minimum viable model and how many ML projects start out without a viable Return on Investment. A similar discussion can be had on architectural projects,
110
where many firms will not be able to find enough data to support ML trainings. An architectural project usually contains a great deal of information in different elements; the problem is thus the lack of quantitative data on specific element parts. While we evaluate how the different formats perform in speeds/size to find a clear comparison number for evaluation, a real dataset with thousands of projects would perform very differently. This is something the entire industry will have to tackle, as the sheer number of projects needed to perform a prediction based on a single element would be beyond what one firm could muster. As the AEC industry is moving forward, there is another major problem that comes up, which is noncontextual datasets either from datasets not relevant to the current context or from computer generated datasets. This is an issue that has been growing in recent years. The AEC industry is extra vulnerable to noncontextualized data, as building laws and standards are high variable between countries and continents. As larger datasets become more varied and lacking in specific data, a high number of local datasets would have to be created. This can be solved through just further training a model, although a lack of labeled data in the AEC industry would make this difficult. One option to handle the lack of data could be to investigate methods of federated learning setting outlined in Kairouz & McMahan (2021) where each party just trains the model on the portion of data that they have without revealing it. Such approaches are being developed, for example, in the medical industry where patient data are highly confidential. The drawback though is that it does not incentivize those with substantial datasets to participate, as they do not stand to gain as much compared to their contribution. Furthermore, those methods rely on the honesty of all participants to not corrupt the model by feeding it wrong information. Another angle for future investigation would be to investigate other forms of data. For example, not all projects exist as BIM from which data can be extracted. Some only exist as drawings or in their final build form. Being able to extract information from other datatypes would open possibilities for different use cases and expand the pool of available data immensely. One such case would be regarding existing structures and their transformation when it comes to circular economy. Being able to process point cloud scans, a simple method of digitizing buildings efficiently would enable the harnessing of the information embedded within those as well as information that has been aggregated over the lifetime of the building. Opening future use cases, for example, when it comes to the repurposing and transformation of those buildings.
REFERENCES
111
Abioye, S.O., Oyedele, L.O., Akanbi, L., Ajayi, A., Davila Delgado, J.M., Bilal, M., Akinade, O.O., et al. (2021), “Artificial intelligence in the construction industry: A review of present status, opportunities and future challenges”, Journal of Building Engineering, Elsevier Ltd, Vol. 44 No. October, p. 103299. Flager, F. and Haymaker, J. (2007), “A comparison of multidisciplinary design, analysis and optimization processes in the building construction and aerospace industries”, 24th W78 Conference on Bringing ITC Knowledge to Work, pp. 625–630. Hawkins, J. (2020), “Minimum Viable Model Estimates for Machine Learning Projects”, pp. 37–46. Kairouz, P. and McMahan, B. (2021), “Advances and Open Problems in Federated Learning”, Advances and Open Problems in Federated Learning. Foundations and Trends,® in Machine Learning, pp. 1–210. Kyivska, K. and Tsiutsiura, S. (2021), “Implementation of artificial intelligence in the construction industry and analysis of existing technologies”, Technology Audit and Production Reserves, Vol. 2 No. 2(58), pp. 12–15. Lu, Z., Xu, T., Liu, K., Liu, Z., Zhou, F. and Liu, Q. (2019), “5M-Building: A large-scale high-resolution building dataset with CNN based detection analysis”, Proceedings – International Conference on Tools with Artificial Intelligence, ICTAI, IEEE, Portland, OR, USA, Vol. 2019-Novem, pp. 1385–1389. Molio. (2020), “Byggeriets Digitale Barometer”, No. September. Natonal BIM Standards. (2020), “10th Annual UK’s National Building Specification Report 2020”, NBS Enterprises Ltd., pp. 1–39. Qiu, J. and Sun, Y. (2015), “A Research on Machine Learning Methods for Big Data Processing”, No. June 2016, available at:https://doi.org/10.2991/icitmi-15.2015.155. Song, J., Kim, J. and Lee, J.K. (2018), “NLP and deep learning-based analysis of building regulations to support automated rule checking system”, ISARC 2018 – 35th International Symposium on Automation and Robotics in Construction and International AEC/FM Hackathon: The Future of Building Things, No. Isarc. Wang, R. and Tang, Y. (2021), “Research on Parsing and Storage of BIM Information Based on IFC Standard”, IOP Conference Series: Earth and Environmental Science, Vol. 643 No. 1, available at:https://doi.org/10.1088/17551315/643/1/012172. Withers, L.W. (2022), “The accelerated shift to intangible assets and how to protect them”, available at: https://global.lockton.com/gb/en/news-insights/theaccelerated-shift-to-intangible-assets-and-how-to-protectthem (accessed 7 April 2022). Zabin, A., González, V.A., Zou, Y. and Amor, R. (2022), “Applications of machine learning to BIM: A systematic literature review”, Advanced Engineering Informatics, Vol. 51 No. April 2021, available at:https://doi.org/10.1016/j.aei.2021.101474.
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
Digital supported collaboration Multimodel
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Extending ICDD implementation to a dynamic multimodel framework N. Al-Sadoon, P. Katranuschkov & R.J. Scherer Institute of Construction Informatics, Technische Universität Dresden, Germany
ABSTRACT: This paper targets the development of a dynamic multimodel framework that can facilitate BIM based fire protection and evacuation planning and training. It is inspired by the EU project BEST, which develops a novel real time hazard and evacuation simulation system and integrates three models, namely CFD simulation, crowd simulation and dynamic building model. The latter provides comprehensive information about the building, but this information is static and represents the building in an idealized state where no changes are happening over time. In current practice, the building model is input to simulation tools as a static model. We propose an ontological framework extending the standard ICDD implementation (ISO 21597) to a dynamic Multimodel Framework aiming to explicitly allocate multiple dynamic values to elements in the building model. This enables consideration of dynamically changeable building elements’ status at simulation runtime and hence real-time interoperability of the interlinked simulation components and modules.
1 INTRODUCTION Today, owing to the rapid developments in BIM, wide opportunities are provided for manipulating the building information in various domains of the construction industry. In the field of safety design and hazard/risk assessment systems, various advanced simulation tools have been developed and BIM implementation by safety design professionals has been substantially increased. However, it is still lagging behind other disciplines (Davidson & Gales 2021). The challenge remains as to how to use information technology more efficiently to improve evacuation management in emergency events. Current BIM based safety simulation research focuses on either human behavior or on fire propagation while considering only the static building model information. Hence, the simulation results are approximate and not very reliable (Scherer et al. 2018). To achieve precise and more reliable simulation outcomes the interaction between both fire and occupants with the building should be considered in dynamicity. The consideration of dynamic features will allow transferring the changed status of the building elements in the real time during the fire simulation. As such, providing dynamic information about the buildings, occupants and fire propagation is a key issue (Eftekharirad et al. 2019). The Building Information Model schema IFC (ISO 16739) contains detailed digital data of a built facility designed in a BIM-based software environment. It provides comprehensive geometric and semantic information about the building in addition to the information about the equipment and technical systems inside the building like furniture and HVAC DOI 10.1201/9781003354222-15
systems. The latest IFC model schema offers useful general information as required by fire simulation tools but when considering the fire specific properties, it still needs to be extended to describe the concrete simulation requirements. To overcome these two issues, we propose the development of a Multimodel Framework to firstly, extend the building elements semantic information required for dynamic hazard and evacuation simulation by a BIM Extender, and secondly, extend the standard Information Container for linked Document Delivery (ICDD) implementation according to ISO 21597 by providing a Multimodel Engine (MME) to enable assigning and manipulating multiple dynamic values for any element in the building model. The approach was first proposed for the ongoing EU project BEST (Al-Sadoon & Scherer 2021), which is a novel real-time hazard and evacuation simulation system for both training and safety assessment purposes. The project required dynamic space changes and interactive rescuer consideration to achieve real-time simulation, and hence, to provide for safe building design and safe evacuation plans. BEST comprises several modules: an advanced fire, toxic gas and CBRN dispersion Computational Fluid Dynamics (CFD) simulation module, an advanced crowd simulation module, a new real-time training module with dynamic rescue scenarios, a new Multimodel Framework with an ontology-based link approach, and a dynamic building information model (BIM) regarding open/closed doors, windows, HVAC and other relevant changes of the scenario space. The specific focus of this paper is on the development of the dynamic building model regarding both geometric and functional variations during simulation
115
of building components, HVAC, and automation systems behavior, and the reactive interaction with CFD and crowd simulations. The research objective was achieved through the following steps. Firstly, the structure of ICDD, standardized in ISO 21597, was analyzed to define potential implementation extension. Secondly, the Linkset ontology was extended by adding new classes and properties. At last, the extended framework was implemented and tested on exemplary pilot models. The extended ICDD framework was initially developed to enable dynamicity to a building model and achieve interactive simulation at run-time. However, it is possible to apply the approach to many other situations as well. The possibility of adding multiple values explicitly for an element opens the door for tracking the value evolution of building elements from concept design through construction up to facility management using one and the same evolving model. 2 RELATED WORK The fire protection industry gains many advantages from the technology progress in BIM. Today’s fire protection engineering includes not only active and passive suppression systems but considers the overall life safety of buildings and the occupants. In addition to the 3D representations of buildings, BIM provides a high level of semantic information necessary to enhance the simulation used for fire propagation and crowd evacuation (Scherer et al. 2018, Sun & Turkan 2019). It is not surprising that in the last decade BIM based hazard and evacuation simulation research has increased significantly both in the area of BIM based safety design and BIM based fire safety evacuation, as well as in dynamic BIM based fire safety management. 2.1 BIM based fire safety design Fire safety engineering is an important part of high-rise building design but it is often excluded from Building Information Modeling (BIM) and lags behind other disciplines (Davidson & Gales 2021). In this regard, Shams Abadi et al. (2021) proposed an additional metric for evaluating renovation projects’ construction plans. They combined the building’s spatial and physical properties with co-simulation of fire and occupants’ evacuation behavior to estimate the average and maximum required safe egress time for various construction sequencing alternatives. Then they used these parameters, alongside with cost and duration, as a third decision criterion to evaluate construction schedule alternatives. In terms of fire safety design assessment and considering the fact that simulation tools still utilize data from static building models, Wang et al. (2021) combined the simulation results of four different evacuation strategies with the architectural design drawings to reduce the design defects and thereby increase evacuation efficiency. Their research considers the limited time of crowd tolerance under the influence of various fire factors.The findings showed a
significant effect on the optimization of the structural design of large public buildings and provided some references for emergency evacuation. 2.2 BIM based fire safety evacuation Undoubtedly, decision making on the optimal evacuation and rescue plan is the most important consideration during hazardous events. Initial work showed that robust and reliable technology is available to track and identify people and assets and evaluate the precision of certain RFID-based localization techniques (Menzel et al. 2008; Manzoor & Menzel 2011). Further studies have shown that BIM has the potential to support fire information retrieval and escape route planning. For example, Rüppel et al. (2010) developed a real-time information query system applying an ultra-wide band, wireless local area network and radio frequency identification, and integrated it with BIM to support the routing function of the evacuation system. Gerges et al. (2021) identified agents’ locations using a BIM based platform to send evacuation instructions and then conduct evacuation simulations under various scenarios.A smart fire evacuation system was recently presented by Wehbe & Shahrour (2021). It uses AI technology, fire, and evacuation simulation tools to learn and predict the best evacuation routes for occupants during a fire via the BIM environment. BIM can also be integrated with other technologies to capture and present more fire-related information. Many studies have increasingly realized the importance of combining BIM and the Internet of Things as a viable solution to achieve real-time information updates (Tang et al. 2019). 2.3 Dynamic BIM based fire safety management From construction management it is well known that the realization of real-time data by integrating BIM and monitoring and sensing devices brings significant benefits to assist construction operations and management (Keller et al. 2004, 2006). Furthermore, it was demonstrated that the dynamic evaluation of contextual data avoids information overflow and assists domain experts to quickly access relevant information (Ahmed et al. 2010). Making BIM based models dynamic to provide for real-time building information enables rapid decision-making and timely responsiveness to emergencies (Tang et al. 2019). Early work in that regard demonstrated that the availability of dynamic monitoring data in combination with BIM data and the capability for their timely analysis assists effective, complex decision making (Allan & Menzel 2009). Moreover, the performance of fire emergency management could be strongly influenced by occupants’ different behavior decisions. Ma & Wu (2020) developed a BIM based fire emergency management system that considers the behavior decisions of building users. Based on these decisions the system plans optimal action routes and sends the occupants visual route guidance via SMS. Within this system all information on occupant and fire/smoke conditions can be retrieved from the building to update
116
the emergency status continuously. An App in occupants’ smart phones is updated with fire locations, fire safety equipment, the safest egress path, and/or location status. However, while BIM integrated with sensor data has been fully utilized in the field of fire monitoring, BIM platforms still lack the ability to provide an immersive environment for decision making during rescue operations (Chen et al. 2021). For firefighting and rescue training purposes, Chen et al. (2021) developed a novel framework by integrating BIM, IoT, and Virtual Reality/Augmented Reality technologies to improve building fire safety and rescue efficiency. The research outcomes have shown that using fire monitoring and pathfinding indication as external information in an immersive simulated fire environment can reduce the psychological pressure of trainees, reduce the travel distance, and improve the firefighting efficiency. Also, by exploiting BIM’s 3D geometric data and visualization Wang et al. (2015) developed a BIM based model comprising four different modules, namely evacuation assessment, escape route planning, safety education, and equipment maintenance. For the sector of health and safety, Teizer et al. (2013) and Li et al. (2015) have proposed training systems using BIM and sensor data. They developed algorithms for realtime location tracking data to be used for analyzing safety and productivity in parallel with real-time and post-event visualization by utilizing Virtual Reality environments. Interesting research that considers the impact of the changes in space usage and building layouts during building renovation projects regarding crews and occupants evacuation behavior under emergency has been conducted by Eftekharirad et al. (2019). The study proved that the evacuation time and probability of injuries and causalities under emergency conditions increased considerably during renovation works. More interestingly, the study showed the risks of developing construction plans that meet minimum fire safety regulations without considering the dynamism of interactions between occupants and physical spaces, especially in the event of a hazardous event. For safety design assessment, the findings of this study help raise awareness regarding the importance of considering occupant interaction with changeable building spaces during hazardous events.
separate model resources. In the standard, the contents of these links are specified in a Linkset ontology where only static linking is supported, whereas multiple value “dynamic linking” is not available. The proposed methodology here is to extend the Linkset ontology by adopting a “Linked List” approach to allow allocating multiple values for a particular element thereby enabling the use of various building elements status at simulation run-time in a dynamic manner. To detail the methodology, firstly the ICDD standard is briefly reviewed, then the concept of the “Linked List approach” used to extend the ICDD framework is outlined, and finally, the proposed extended ICDD Schema is presented. 3.1 ICDD structure The ICDD standardized in ISO 21597, is the result of a synergy of efforts that spanned over a decade. These efforts started in Germany with the emergence of the Multimodel approach, firstly developed in the Mefisto project (Fuchs et al. 2011; Scherer & Schapke 2011), and in the Netherlands in the COINS project that developed an interdisciplinary container for the exchange of information (Hoeber et al. 2015). The hierarchical structure of the ICDD comprises an index file and three folders: (1) the Ontology Resources folder, where the ontology files Container.rdf and Linkset.rdf are located, (2) the Payload Documents folder storing the internal linked models and (3) the Payload Triples folder where all the link datasets are stored (Figure 1).
3 METHODOLOGY In hazardous events both fire propagation and occupants’ evacuation are affected significantly by the building’s geometry and space distribution. Hence, both metrics have to be considered in safety design engineering (Onyenobi et al. 2006). Our approach, performed in the frames of the EU project BEST, focuses on extending the recently standardized ICDD framework (ISO 21597), which enables the handling of multiple data resources as a single information container and specifies relationships between the intermodel data using links between the elements in these
Figure 1. ICDD structure according to (ISO 21597, 2020).
117
The index file contains all the meta-data to describe the container and to specify the documents that make up its contents. 3.2 Linked list approach A linked list is a linear collection of data elements where each element points to the next via a link. It supports multiple operations such as add, delete, display or search for an element in the list. The most common types of Linked Lists are: – Simple Linked List – each node has a next link only to reference the next node; – Doubly Linked List – each node has a next and a previous link to reference both the next and previous nodes; – Circular Linked List – the last node has a next link to reference the first node and the first node has a previous link to reference the last node. In our approach, the concept of the Doubly Linked List is selected to facilitate the allocation and search of multiple values for a building element in the IFC model. It proved its efficiency to extend the Linkset ontology as it will be explained hereinafter. 3.3 The proposed extended ICDD schema As mentioned, to achieve the envisaged dynamicity for building elements, the Linked List data structure is used to extend the Linkset ontology, which is defined as an RDF(S)/OWL file providing the object classes and properties used to specify links between documents, models and their elements in the container. The Linkset ontology specifies different linkage capabilities to link, for example, between a single element in a model and a related document, one element in
one document/model and multiple related elements in other documents/models, or among a set of elements in one document/model and related elements in multiple documents/models. Each link element in a linkset is related to exactly one internal/external document where the element has only one static value. To link an element using the optional hasIdentifier property, the element should have an identifier. There are three mechanisms to identify an element in a document: a string-based identifier, a query, or a URL-based identifier. A standard ls:link has not less than two ls:LinkElement’s, each referencing ls:hasDocument and an optional ls:hasIdentifier. The proposed extended schema for the Linkset ontology, highlighted in green in Figure 2 below, is developed to support deep model-based linking, hence ls:hasIdentifier is not optional here. The ls:LinkElement is extended to include a third object property named ls:hasValues to reference the added new object class ls:listDynamic Values, which in turn consists of the three object properties that have been respectively named ls:hasDynamicValue, ls:hasfirstValue and ls:haslast Value. ls:hasDynamicValue references the second new added object class ls:DynamicValues, which also consists of three object properties that have been named ls:hasNextIndex, ls:hasPreIndex and ls:value. The new added objects, object properties and datatype for the linkset ontology are listed in Table 1 below. Unlike the standard IFC model, where a data attribute typically has only one value, the extension for the Linkset ontology provides for an externally maintained multimodel based linked list approach, which facilitates assigning multiple explicit values for a building element attribute. This allows for different application scenarios of tracking the variation of a value throughout a project’s lifecycle within one and the same building model.
Figure 2. Linkset ontology showing the standard schema and the proposed extension (in green).
118
Table 1.
Extended Linkset ontology definitions.
Named entity Object ls:listDynamicValues ls:DynamicValues
4.1 BIM Extender
Description
A class referencing to a list of dynamic values A class referencing to dynamic values
Object type ls:hasValues ls:hasDynamicValue ls:hasfirstValue ls:haslastValue ls:hasNextIndex ls:hasPreIndex
A relation from ls:LinkElement to an ls:listDynamicValues A relation from ls: listDynamicValues to an ls:DynamicValues. Referencing to the first value in a list of dynamic values Referencing to the last value in a list of dynamic values A relation from ls:DynamicValues to a next ls:DynamicValues A relation from ls:DynamicValues to a previous ls:DynamicValues
Datatype ls:value
A String containing the value.
4 IMPLEMENTATION To validate the proposed approach, a two-story university building in Prague was used as pilot case study. The approach is implemented in the form of two independent, interacting services, namely (1) a BIM Extender, used to enhance an IFC model via a separate linked property set model, and (2) a Multimodel Engine (MME), supporting the creation and manipulation of ICDD data based on the proposed extended ICDD schema and providing for the integration and use of dynamic BIM features. The implementation is based on Python, Owlready2 and SQLite. Integration with other platform components is achieved using a REST API with requests and responses formalized in JSON. Figure 3 below shows schematically the current use of the developed services on the BEST platform.
The BIM model contains a high level of semantic information that can enhance the input data for a simulation used for fire propagation and crowd evacuation (Sun & Turkan 2019). However, the available BIM object libraries are currently not mature enough to supply fire and smoke simulations with enough information (Davidson & Gales 2021). To respond to the deficiency of property parameters specific to fire events, we have developed a BIM Extender service (Al-Sadoon & Scherer 2021) to enrich the BIM data with behavior and interaction information and the related possible object states. In alignment with the overall multimodel approach the extended BIM data are thereby made available without disturbance to the content of the original BIM data. A front-end user interface allows creating an IFC Property set file to extend the semantic information of any building element, which is then exported as a json file and added as elementary model to the Multimodel Container. For example, for safety design assessment, the user can add dynamicity attributes for student chairs in a classroom by creating a property named “Removable Furniture” with values (yes/no) or adding “Breakable Wall” with values (yes/no) for the internal walls. Figure 4 shows a property set file created to add two property sets for ifc Door. The first property set is” Door Status” having the values “Closed, Open, Locked”, whereas the second property is ”Fire Door” having the values “Yes/No” used to provide an additional fire related attribute to reflect the door type.
Figure 4. BIM extender user interface.
4.2 Mutlimodel Engine (MME)
Figure 3. BEST platform components.
ISO 21597 provides specifications for handling multiple documents as one information delivery in a container, and specifications to describe means of linking among these documents. The ICDD framework, as a generic information container format, is used to store and exchange such linked data. In our implementation, the proposed extended version of the ICDD framework is applied to achieve the desired dynamicity.
119
A BIM model of a two-story university building was created as a pilot model for concept verification. It comprises lecture rooms, labs, offices, and other building utilities, all of them furnished and equipped with HVAC and sprinkler systems as illustrated in Figure 5. The BEST platform aimed to be provided as Software as a Service (SaaS) for training and safety assessment purposes. The services enable the end user to design a fire evacuation scenario by defining both agent and hazard entities, place them within the building model, start a simulation, pause it, examine and edit the scenario, resume the simulation, and finally make informed decisions on the basis of the obtained results. In this context, the MME role is to assign and manage explicit multi dynamic values of building elements based on the end user’s predefined property sets. The sequence of operations can thereby be described as follows: – Upload the building model; – Define additional property set(s) using the BIM Extender interface and link them to the building model; – Set up a fire and crowd scenario, i.e. set the fire type and location and the agents’ locations and categories and link them to the building model; – Create the ICDD and send all scenario information to the simulation services to start the co-simulation.
Figure 5. University building model.
During the simulation, the end user can temporally stop it and edit the scenario. For example, at some point the end user (a safety designer) can decide to change the door type from a normal door to a fire door, or to change the status of lab benches to be removable in order to assess how these property changes would affect the fire and crowd simulation. In another context, the end user (safety training team) can decide to change the property of a window to be a breakable window to be used for rescue and thereby define an additional possible escape path. Then the end user will resume the simulation to assess the effect of the change on the propagation of the fire and the crowd evacuation. The MME role in such scenarios is to provide the following functions: 1. Create the Multimodel Container as specified in the ontologies Container.rdf and Linkset.rfd; 2. Add elementary models that can be any kind of data sources to be saved either internally in the container or kept at their sources and referenced by URL. In the discussed pilot, the elementary models are the IFC building model and Property set files in json format. 3. Create links based on the object classes and properties provided by the Linkset ontology, whereby each link specifies interdependencies among two or more elements contained in the elementary models. The developed extended specification provides for multiple linkage capabilities. In the case pilot, links are created between buildings object and their related extended property sets. For example, the ifcDoor entities are linked to the property door status. Then based on the extended Linkset ontology, the defined values for this property are added as a linked list and used dynamically during the simulations. The proposed extension hasValues enables allocating multiple values to an element. Figure 6 illustrates
Figure 6. Querying door dynamic values.
120
an instance of the extended ICDD Framework. It shows a direct link with two link elements: the ifcDoor element from the IFC model and the door status element from the property set file with the explicit multiple values assigned to the door element that can be dynamically associated with the door element based on scenario development.
REFERENCES
5 CONCLUSION AND FUTURE WORK The ICDD Multimodel approach enables linking, storing and exchanging of heterogeneous model data. This research goes a step further to add a dynamicity feature to the information container. The developed approach was already successfully applied in the BEST project, which is a novel real-time hazard and evacuation simulation system for both training and safety assessment purposes. The project required dynamic space changes and interactive rescuer consideration to achieve realtime BIM-based co-simulation for various hazard scenarios. The proposed extended Multimodel Framework was applied in that project to assign multiple dynamic values to various elements in the building model, as needed for the dynamic co-simulation management of fire and pollutant gas scenarios. Providing BIM with dynamic features that can be changed during simulation runtime can contribute significantly to the BIM/AEC market, meeting the high demand for simulating complex hazards to provide for safe building design and safe evacuation planning. However, the current implementation of the approach in the above mentioned project is limited to adding a few dynamic properties for building elements only. There could be a storage and performance problem when a huge amount of data is involved, such as sensor data. This is an expected limitation and hence subject of further research targeting storage efficiency and improved data transfer mechanisms using database or block chain technology. Based on the dynamic environment enabled by the proposed extended ICDD Framework various further opportunities to track value changes of building elements over the project lifecycle can be envisaged for further research. Examples include cost monitoring or tracking changes between as-designed and as-built model realization. When multiple values are explicitly managed, building elements can be tracked from concept design to construction and up to facility management using one and the same evolving and adapting coordination model. ACKNOWLEDGEMENT This research was made possible with the funding support of the European Commission for the project BEST (EUROSTARS E! 114043) and the technical support of the project partners. This support is herewith gratefully acknowledged.
121
Ahmed, A., Menzel, K., Ploennigs, J. & Cahill, B. (2010). Multi-dimensional building performance data management for continuous commissioning. Advanced Engineering Informatics, 24(4), pp. 466–475, doi: 10.1016/j. aei.2010.06.007. Allan, L. & Menzel, K. (2009). Virtual Enterprises for Integrated Energy Service Provision. In: Camarinha-Matos, L. M., Paraskakis, I. & Afsarmanesh, H. (eds.) Leveraging Knowledge for Innovation in Collaborative Networks, PRO-VE 2009, IFIP Advances in Information and Communication Technology, vol. 307, Springer, BerlinHeidelberg, Germany, https://doi.org/10.1007/978-3-64204568-4_68 Al-Sadoon, N. & Scherer, R. J. (2021). IFC Semantic Extension for Dynamic Fire Safety Evacuation Simulation. In Proc. of the CIB W78 Conference, Vol. 2021, pp. 11–15. Chen, H., Hou, L., Zhang, G. K. & Moon, S. (2021). Development of BIM, IoT and AR/VR technologies for fire safety and upskilling. Automation in Construction, 125, pp. 1– 13, doi: https://doi.org/10.1016/j.autcon.2021.103631. Davidson, A. & Gales, J. (2021). BIM and Fire Safety Engineering – Overview of State of the Art. International Journal of High-Rise Buildings, 10(4), pp. 251–263, doi: https://doi.org/ 10.21022/IJHRB.2021.10.4.251 Eftekharirad, R., Hosny, A., Nik-Bakht, M. & Hammad, A. (2019). Planning Building Renovation Projects for Safe Evacuation Provisions – An Agent-Based Modelling Approach. In: Proc. 16th IBPSA, Rome, Italy, pp. 2294– 2301, doi: https://doi.org/10.26868/25222708 .2019. 210897 Fuchs, S., Kadolsky, M. & Scherer, R. J. (2011). Formal Description of a Generic Multi-Model. In: Proc. 20th IEEE Int.Workshop on EnablingTechnologies: Infrastructure for Collaborative Enterprises, pp. 205–210, IEEE Publ. Gerges, M., Demian, P. & Adamu, Z. (2021). Customising Evacuation Instructions for High-Rise Residential Occupants to Expedite Fire Egress: Results from Agent-Based Simulation. Fire, 4(21), 18 p., doi: https://doi.org/10.3390/fire4020021 Hoeber, J. G.,Alsem, D. M. & Willems, P. H. (2015).The management of information over the life-cycle of a construction project using open-standard BIM. In: Proc. 32nd CIB W78 Conf., Eindhoven, The Netherlands, pp. 295–301. ISO 16739-1 2018. Industry Foundation Classes (IFC) for data sharing in the construction and facility management industries – Part 1: Data schema. International Organization for Standardization, Geneva, Switzerland, ICS 25.040.40 Industrial process measurement and control, 1474 p. ISO 21597-1 2020. Information Container for linked document delivery – Exchange Specification – Part 1: Container. International Organization for Standardization, Geneva, Switzerland, ICS 35.240.67 IT applications in building and construction industry, 41 p. Keller M., Katranuschkov P. & Menzel K. (2004). Modelling collaborative processes for Virtual Organisations in the building industry. In: Proc. eWork and eBusiness in Architecture, Engineering and Construction, pp. 417–431. Keller M., Scherer R. J., Menzel K., Theling T., Vanderhaeghen D. & Loos P. (2006). Support of Collaborative Business Process Networks in AEC, Journal of Information Technology in Construction (ITcon), 11(34), pp. 449–465. Li, H., Lu, M., Chan, G. & Skitmore, M. (2015). Proactive training system for safe and efficient precast installation.
Automation in Construction, Vol. 49, Part A, pp. 163–174, doi: https://doi.org/10.1016/j.autcon.2014.10.010 Ma, G. & Wu, Z. (2020). BIM-based building fire emergency management: Combining building users’ behavior decisions. Automation in Construction, Vol. 109, Article 102975, doi: https://doi.org/10.1016/j.autcon.2019.102975. Manzoor, F. & Menzel, K. (2011). Indoor localisation for complex building designs using passive RFID technology. In: Proc. XXXth URSI General Assembly and Scientific Symposium, pp. 1–4, IEEE Publ. Menzel, K., Cong, Z. & Allan, L. (2008). Potentials for radio frequency identification in AEC/FM. Tsinghua Science and Technology, 13(S1), pp. 329–335. Onyenobi, T., Hudson, J. & Ormerod, M. (2006). Impact of building geometry on fire spread rate. In: Proc. 22nd Annual ARCOM Conference, Birmingham, UK, Sept. 4-6, pp. 739–744, doi: 10.1.1.1076.51208 Scherer, R. J. & Schapke, S. E. (2011). A distributed multimodel-based management information system for simulation and decision-making on construction projects. Advanced Engineering Informatics, 25(4), pp. 582–599. Scherer, R. J., Luu, N.T., Katranuschkov, P., Spieckermann, S., Habenicht, I., Protopsaltis, B. & Pappou, T. (2018). Towards a multimodel approach for simulation of crowd behaviour under fire and toxic gas expansion in buildings. In: Rabe, M., Juan, A. A., Mustafee, N., Skoogh, A., Jain, S. & Johansson, B. (eds.) Proc. 2018 Winter Simulation Conference (WSC), pp. 3965–3976, IEEE Publ. Shams Abadi, S. T., Moniri Tokmehdash, N., Hosny, A. & Nik-Bakht, M. (2021). BIM-based co-simulation of fire and occupants’ behavior for safe construction rehabilitation planning. Fire, 4(4) 67, https://doi.org/10.3390/ fire4040067
122
Sun, Q. & Turkan, Y. (2019). A BIM Based simulation framework for fire evacuation planning. Advances in Informatics and Computing in Civil and Construction Engineering, Springer, pp. 431–438. Tang, S., Shelden, D.R., Eastman, C.M., Pishdad-Bozorgi, P. & Gao, X. (2019). A review of building information modeling (BIM) and the internet of things (IoT) devices integration: Present status and future trends. Automation in Construction, 101, pp. 127–139, doi: https://doi.org/10.1016/j.autcon.2019.01.020 Teizer, J., Cheng, T. & Fang, Y. (2013). Location tracking and data visualization technology to advance construction ironworkers’ education and training in safety and productivity. Automation in Construction, 35, pp. 53–68, doi: https://doi.org/10.1016/j.autcon.2013.03.004 Rüppel, U., Stübbe, K. M. & Zwinger, U. (2010). Indoor navigation integration platform for firefighting purposes. In: Proc. 2010 International Conference on Indoor Positioning and Indoor Navigation, pp. 1–6, IEEE Publ. Wang, F., Xu, X., Chen, M., Nzige, J. & Chong, F. (2021). Simulation Research on Fire Evacuation of Large Public Buildings Based on Building Information Modeling. Complex System Modeling and Simulation, 1(2), pp. 122– 130, doi: 10.23919/CSMS.2021.0012 Wang, S. H., Wang, W. C., Wang, K. C. & Shih, S. Y. (2015). Applying building information modeling to support fire safety management. Automation in construction, 59, pp. 158–167, doi: https://doi.org/10.1016/j.autcon.2015. 02.001 Wehbe, R. & Shahrour, I. (2021). A BIM-Based Smart System for Fire Evacuation. Future Internet, 13(9), 221, doi: https://doi.org/10.3390/fi13090221
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Management of BIM-based digital twins in multimodels M. Polter & R.J. Scherer Technische Universität Dresden, Dresden, Germany
ABSTRACT: This paper presents the use of the Multimodel method for the management of digital twins (DT) in Building Information Modeling (BIM)-based virtual labs. The contents of the Multimodel are based on the software tools required for the specific application, whereby the DT is not a monolithic artifact, but a dynamic set of interlinked domain-specific technical models. The links are specified in machine-interpretable formats, such as the Resource Description Framework or the Web Ontology Language and stored in separate link models. Additional semantic information, that is not part of the technical models, can thus be generated by inference mechanisms. This approach enables the automation of complex processes in cyber-physical systems and increases the robustness against human input errors. A prototypical service implementation and evaluation in a bridge monitoring application underpins the feasibility and benefits of the developed concept.
1 INTRODUCTION Building information modelling has become an integral part of construction projects across all phases. It is characterized by integrated use of a large number of direct and indirect information about a structure, which have to be managed task oriented over the whole structure life cycle without neglecting its history. The digital twin approach enables the behavior of the object to be investigated and studied through simulation before it is actually manufactured, which is not possible with current Building Information Models, because they consist of static information which may be updated at some time instances. BIM labs, such as the iVEL (Baumgärtel et al. 2012), are characterized by the semantic description of an object in order to simulate and study its behaviour and relationships with it‘s complex context. A digital twin (DT) consists of instantaneously updated information reflecting the real-world model. This means real-time data that has to be collected from the construction site or during the building operational phase preferably real time. BIM forms the bridge between pure real-time data management in cyber physical systems (CPS) and DTs (Boje & Kubicki 2021). Building Information Model and DT should always be synchronized so that the progress can be compared with the planning in real-time. DTs have played an essential role in manufacturing for over a decade, but they are still rare in construction. This is because the integration of processes along the building life cycle and the sharing of data has always been critical due to the many stakeholders involved in the architecture, engineering, construction and operation sector (Corneli 2021). In addition, the application of semantic web technologies and Internet of Things (IoT) for live construction sites is still too less explored DOI 10.1201/9781003354222-16
(Boje & Kubicki 2021). However, digital twins are also important for the construction industry, but here they are mainly used for the optimized control of building systems, e.g. in HVAC systems. The specific form of a digital twin depends to a large extent on the application context. Different simulation tools use different information and data formats. Storing all this information in a single domain model would make it huge, difficult to maintain, error-prone, redundant and also contradictory. Furthermore, such a monolithic single model would be incompatible with available tools. So far there is no model format that efficiently represents these heterogeneous data and data formats and thus makes the DT suitable for more than just one or a few use cases. The Multimodel method describes a solution for information management in interdisciplinary construction information processes. Multimodels bundle heterogeneous technical models from different domains and allow their elements to be linked in external, ID-based link models. The technical models themselves are not changed but are loosely and temporarily linked by the superordinate structure of the Multimodel (Fuchs 2015). This paper describes a concept for the efficient management of arbitrarily expandable digital twins in construction projects using the Multimodel method. Information is added to and read from domain-specific technical models using existing tools. These are combined via link models to form a higher-level, functionally linked digital twin. A web-based Multi-model Engine for Multimodel management is being developed, which can be integrated into other applications via service APIs. The Multimodel Engine has been tested as a data management component of a workflow engine that automatically uses the different parts
123
of the DT as data source and data store in different use cases. This paper is structured as follows: Section 2 provides an overview of the current state of software support for digital twins and Multimodels. The main part of this paper is then devoted in section 3 to explaining a concept for the Multimodel based management of digital twins and its prototypical implementation in a software service. In section 4, a case study for the evaluation of the prototype is described. The paper concludes with a summary of the work done and an outlook on current research in section 5.
2 DIGITAL TWINS AND MULTIMODELS IN CONSTRUCTION The IoT enables increasing integration of objects from the physical world into the virtual world of software and services (Seiger et al. 2019). CPS today can automatically digitize the properties of physical objects and thus compare the status of both object representations in real time. With this digital twin, the behavior of its physical counterpart can be examined with software and predicted under changing conditions. Haag (2018) demonstrates a concept for the development and continuous updating of the digital twin of an industrially manufactured product throughout the product’s lifetime. In construction, the same goal is pursued for buildings, from the construction phase through to demolition. Unlike the mass production of industrial products, buildings are usually unique projects, and their DT is one of a kind. However, current software applications in the construction industry are predominantly specialized for certain scenarios and do not offer the possibility of adapting the methods for creating and updating the DT to different tasks. Solutions for the efficient, cross-process integration of different tools are therefore essential for complex construction projects. Corneli et al. (2021) examined the implementation of digital twins in different applications and identified recurring, fundamental problems. Commercial tools are usually tailored to a specific domain and are mutually incompatible in most cases. In addition to incompatible data formats, this is due to the differences in the semantic models due to the use of unrelated modeling theories. The authors have developed a framework for creating DT based on modelbased software engineering (MBSE), which enables the semantic integration of information across different layers. The framework manages data sources in the form of services from which a domain-specific DT is generated. Toolchains for various application domains such as energy management, resource planning and systems diagnostics are provided. Collected data is represented in open or de-facto standardized file formats to ensure interoperability between the different tools. This approach brings with it the advantage of high reusability of data converters since the data
formats for a domain are defined from the outset. However, the toolchains are accordingly limited to tools that support the standardized data formats. In contrast, the approach presented in this work does not focus on the digital twin as the central object of the application, but rather on the processes to be automated for the respective application. Since the corresponding software tools are usually predefined, more effort is required here for the semantic integration of the information. The Multimodel method with its link model concept is used in order to manage information not in several domain-specific DTs, but in an integrated multi digital twin (MDT). The data relevant for a use case is made available through views and filter expressions. Fuchs (2015) has developed a Multimodel query language (MMQL) for this purpose. It provides the technical framework for generating Multimodels and the contained links, as well as for defining filter expressions. Boje & Kubicki (2021) present an abstract process for designing and implementing DT and the associated physical infrastructure. This digital twin factory also contains a semantic layer that encapsulates a common semantic description of all components and data sources. This should enable the adaptation of the DT to specific requirements by exchanging, adapting, or expanding its components and data sources. However, the authors do not go into detail about how this semantic description of any real-world element can be realized in practice. Maintaining the data semantics when changing the underlying technical models is a major challenge in the evolution of a DT. Pauwels (2017) provides an approach by using the semantic web linked data. Domain-specific knowledge is modeled together with the semantics in machine-interpretable formats such as RDF and OWL, whereby the semantics are not lost when the model is changed. Corresponding ontologies have been designed and tested by the W3C Linked Building Data Community Group. However, these are more related to the operational phase of buildings (Boje & Kubicki 2021) and do not provide sufficient support for the construction phase. In summary, it can be stated that solutions for the integration of DT in software applications for the construction industry already exist. However, these are usually very specialized and not sustainable in terms of adaptability to specific requirements and changing application scenarios.
3 MULTIMODEL BASED MANAGEMENT OF DIGITAL TWINS This work focuses on the development of a concept for the integration of a MDT into Building Information Modeling based on Multimodels and the implementation in a corresponding service component called Multimodel Engine. In conventional application scenarios, a digital twin is updated directly by adjusting individual parameters using measured values from
124
individual sensors. In this work, the focus is on the continuous, indirect refinement of a MDT based on automated, system identification-based simulations involving various sensor data sources. This is intended to create a database on the basis of which task-specific services can be implemented for different stakeholders in different life cycle phases of a building. Our research and development in this area was motivated by the following initial situation: – Large amounts of data to be managed, distributed across many domain-specific technical models (logical) and servers (physical). – Existence of a large selection of tools for processing and obtaining information on the technical modelspecific data. – Stakeholders from different departments who have read and write access to the data managed in DT. The following goals are to be achieved with the development of a software component for the management of Multimodel-based DT: – Provision of task-oriented access to data with established tools. – Administration of the distributed data as a closed information space through loose coupling, without changing the technical models or specifying restrictions on their storage location. – Ensuring data integrity when using the DT collaboratively. – Scalability, depending on the amount of data to be managed and the number of stakeholders. Our developments are based on the Multimodel method according to Fuchs & Scherer (2017), which is briefly explained below. 3.1 Multimodel method The term Multimodel container describes the structure of Multimodels, i.e., the arrangement and grouping of technical models and metadata in a multi-model. Technical models describe a domain-specific representation of information in a closed container in the form of a file (e.g., thermal model, structural model, HVAC model or cost model). This can be in a standardized (e.g., IFC, XML) or in a tool-specific format (e.g., as an Autodesk Revit file, Cervenka Atena file). The ISO standard 21597-1 (ISO 2020) describes the structure of Multimodel containers. The term ICDD (Information Container for Linked Data Delivery) is used in the standard, which defines a concrete data format for Multimodel containers in addition to the structure. However, the Multimodel method does not specify a superordinate scheme for the representation of the information contained in a Multimodel, which makes transformation processes unnecessary and enables a neutral exchange of the linked technical models. These exist “equally” in the Multimodel container, i.e. there is no leading and no subordinate models (Fuchs 2015). This enables the administration of large amounts of data, whereby the focus can be
on different domain-specific models depending on the task. The linking of data in Multimodels offers informational added value compared to the isolated consideration of technical models. Data can be connected to each other across different levels of detail, e.g. at file level, model elements with entire files or individual elements of model A with individual elements of model B. The digital twin, which is composed of a loose coupling of technical models and metadata, is presented to the user or external applications as a single closed information space. 3.2 Multimodel engine Lim (2019) have designed a technology stack to ensure the productive, efficient, competitive use of DT in industrial production (Figure 1). Not all features of factory production can be mapped to construction projects, which are too complex to be able to implement an automated adjustment of the processes on the basis of collected data. The feedback of the information gathered with the help of the DT and corresponding adjustments in the real world are usually reflected by human decisions in construction. Another difference between automated industrial production processes and construction projects is the uniqueness of the latter. Although the basic structure of the processes in the execution of construction projects remains the same, a separate digital twin must be created for each project. Additional challenges arise from the involvement of numerous stakeholders in the overall process, each bringing in individual requirements and data formats. Collaborative and thus potentially competing access to the DT requires mechanisms to ensure integrity and version management. The question of rights management and data security in collaborative digital twins is currently still the subject of research
Figure 1. Technology Stack for Digital Twins (Lim 2019).
125
and is being investigated in the iECO (IECO 2022) research project. However, most of the concepts and technical aspects of Lim’s proposed technology stack are independent of the concrete application domain and therefore serve as a guideline for the development of the Multimodel Engine. The architecture is designed in such a way that it is compatible with the requirements of the GAIA-X infrastructure (Braud 2021), which is currently under development, or can be adapted to it depending on the development status. In the following, the concepts and technical basics of the Multimodel Engine are explained using the layer architecture proposed by Lim (2019). 3.2.1 Data acquisition and transmission The Multimodel Engine is designed as a serviceoriented architecture and communicates with other services via REST (Representational State Transfer) interfaces. It is implemented as a containerized, scalable web service component. This enables the integration of the Multimodel Engine into other applications or its combination with other services, for example with different data storage technologies. By encapsulatingitasaDockercontainer (https://www.docker.com), the service can be run and scaled in any web service infrastructure, regardless of the operating system used. Figure 2 shows the principle of integrating distributed data into the Multimodel Engine. Technical models and ontologies are not tied to a specific storage location and are referenced via URLs in accordance with the REST principle. In addition to saving bandwidth and local storage space, this also brings with it the advantage of data sovereignty, since the data remains physically in the possession of the owner. The integration of dynamic sensor data brings with it the challenge that these are usually not available as closed files but as streams (Faschingbauer 2011). In order to map both these streams and the properties of the corresponding sensors and metadata in a uniform data space, they are transformed into the standardized Sematic Sensor Network Ontology format (W3C 2017). Depending on the application, the sensor information is then distributed in one or more
Figure 2. Multimodel Engine data integration.
graphs in a graph database. Due to heterogeneous data sources, such as provider-specific sensor interfaces or MQTT brokers, it is not possible to implement a generally applicable interface for querying the sensor data. Corresponding plugins for receiving the streams in an Internet-enabled exchange protocol (e.g. TCP, UDP, SOAP, MQTT) and transformation into an SSN graph are in the draft stage. The goal is a Multimodel Engine that allows an application-dependent integration of any sensor data sources. Furthermore, pre-processing and filtering of large amounts of data via data buffers must be possible in order not to overload the system. 3.2.2 Data representation Multimodels differ from loose model collections in that additional semantic information is generated and provided, which enables high-level functional services such as filters or AI-based methods and the propagation of changes across model boundaries. This semantic information is represented through separate links between technical models or model elements that are stored in ontology-based link models. One of the basic principles of the Multimodel method consists in preserving the integrity of the technical models and maintaining their formats. This ensures tool support and data sovereignty and prevents conflicts in the link models. When persisting the Multimodel, two aspects are distinguished. The structure of the Multimodel is kept in a relational database system together with descriptive metadata, whereby the concrete database implementation can be exchanged depending on the technical requirements. Databases allow the use of powerful data query languages such as Structured Query Language (SQL) for information filtering. The content of the Multimodel, which includes technical models and other user data (e.g. sensor data ontologies), is represented by references (see Figures 3 and 4). If required, the Multimodel Engine provides the user with the Multimodel for export in different representations. For example, the structure can be XML or OWL-based or the entire Multimodel can be exported in the form of an ISO-21597-1-compliant ICDD. The latter requires that the owners of the technical models release them for export. Figure 3 shows an ICDD on the left, which was designed as a data exchange format for Multimodels between different systems, and on the right the mapping to a tablebased representation of the Multimodel. In contrast to reference-based administration, all data of the Multimodel is physically stored in the ICDD. This allows not only the decoupled exchange of business models and semantic information, but also the persistence of specific Multimodel states in the form of snapshots. The latter is only possible to a limited extent with an exclusively reference-based administration, since the system has no exclusive control over the specific technical models. Since these remain with the owner until they are physically required, neither the permanent upto-dateness of the reference nor the immutability of the model content can be enforced. Instead, it is assumed that, according to the REST principle, each model
126
version has its own unchangeable URL. If the current application context requires the management of a history of the Multimodel, this can be achieved with the regular generation of ICMM (Information Container Multimodel). If only the structure of the Multimodel, but not the content itself, is to be versioned, then export functions are available in various formats such as XML or RDF. 3.2.3 Microservices Microservices are independent processes or software services that communicate with each other via interfaces that are independent of a programming language. In modern software applications, microservices are combined into complex applications. The BIMgrid Workflow Engine (Polter et al. 2020) works according to this principle by orchestrating the functions provided by microservices in complex workflows. These microservices obtain the necessary data from the digital twin managed by the Multi-model Engine (consumer services) or add data to the digital twin (producer services). They can also act as consumers and producers at the same time in order to continuously supplement or refine the DT on the basis produced process data.
An essential prerequisite for corresponding workflows is the definition of the application context for the DT in order to identify data sources, coordinate data streams and integrate supporting data preparation mechanisms. One advantage of construction projects is the clear structuring into sub-processes, the chronological sequence of which is known in advance in the majority of cases. This means that it is usually possible to clearly define which process accesses the digital twin, when and how. If there is a risk of data inconsistencies due to possible competing access to the DT, this can be prevented by sequencing or prioritizing the relevant sub-processes. The application scenario is decomposed into concrete steps that are modeled using the Business Process Model and Notation (BPMN), assembled into a formalized workflow and then imported into the BIMgrid Workflow Engine. Simulation tools and any data preparation, documentation and evaluation tools can now be controlled automatically according to the workflow without direct user interaction, provided they have the appropriate API. The Multimodel serves as a data source and sink for the individual process steps, with the digital twin being continuously supplemented and refined during the workflow process. In the following section, a prototypical implementation of the developed concept will be demonstrated in an application example for building monitoring in the operational phase.
4 CASE STUDY: BRIDGE ASSESSMENT
Figure 3. Mapping of a digital twin to different Multimodel representations.
Figure 4. SPARQL query to select all structurally damaged construction components from a structure (Hamdan et al. 2019).
In times of increasing traffic and decreasing budgets of local authorities, bridges often have to be operated beyond their intended lifetime, which creates a high demand for assessment of bridge health, deterioration and increased traffic load behavior. To assess the current condition of a bridge and make predictions about its expected lifetime, many manual on-site work is still required today. For reasons of cost, characteristic values are not measured continuously but only on a random basis and are often replaced by assumptions. In the cyberBridge project (Polter & Scherer 2018, Katranuschkov et al. 2020), a cyber-physical bridge monitoring system was designed which, through continuous bridge monitoring, allows predictions about the decay and thus the remaining lifetime of a bridge at low cost. The predictions are based on system identification at the crack propagation level, with existing software tools encapsulated in services and combined into an integrated platform. Figure 5 shows the prognosis workflow for the condition of a bridge based on continuous, system identification-based parameter studies and measured sensor values. The data source for the individual workflow steps is a MDT of the bridge, stored in an Information Container Multimodel (Figure 6). Simulation tools are increasingly offering IFC import interfaces, so that in many cases the transfer of the native bridge model (Bridge.ifc) to the structural model
127
Figure 5. Workflow of the damage system identification-based bridge monitoring.
simulation. The calculated values of static forces and deformations of the individual model candidates are compared with the actual values measured by sensors and the best-fit model is thus identified. This model candidate best reflects the actual condition of the bridge and serves as the basis for simulating the spread of damage in the next step. The calculated predictions are used to adjust and refine the damage model, which then serves as input for the next workflow run. The data generated during the individual workflow runs is stored in the Information Container Multimodel (ICMM) in its own subfolder named Simulation_x (where x is the number of the workflow run) and thus forms a history of the bridge health status (Figure 6). This new method enabled the reduction of assumptions and interpolated values when determining the remaining lifetime of a bridge and provided more reliable forecasts through the continuous, automated refinement of the digital twin and regular comparison with the physical counterpart. In Polter et al. (2021) another application of a Multimodel-based DT in a system identification-based workflow for the dynamic adjustment of a construction process is described.
5 CONCLUSION AND OUTLOOK
Figure 6. Digital twin for damage system identification.
(Structural_model.inp) can already be automated. Based on the initial structural model and the damage parameter variations defined in the separate variation model (VariationModel.xml), variants of the structural model are created and then subjected to a statics
In order to be able to use Building Information Modeling profitably and cost-effectively over all life cycle phases of a building, an efficient integration of the information stored in technical models and obtained by sensors is necessary. The continuous maintenance of this digital twin of a building offers a variety of options for controlling, documenting, and predicting the behavior of the building under a wide variety of conditions. The use of the Multimodel method for organizing and semantically enriching the information is particularly suitable here, since Multimodels provide the originally heterogeneous data as a single, homogeneous information space without changing the underlying data structures. The concept presented in this work for the multimodel-based management of digital twins allows an increase in the automation of complex workflows in cyber-physical simulation systems. The concept was implemented as a software service prototype and tested in two use cases. The results show, that with the developed method processes for monitoring a structure and predicting its behavior can be carried out continuously and with minimal user interaction. Compared to conventional investigation methods, that are carried out on a random basis with interpolated values, the approach presented here is able to increase the accuracy of the forecasts with reduced human effort. The greatest advantage of Multimodel-based Digital Twins compared to loose collections of technical models is the provision of cross-model semantics through link models. In order to fully exploit this potential, meaningful links must be defined. In this
128
context, our research is currently devoted to automating the linking process by defining so-called link templates. These are intended to describe links between models or model elements on an abstract level in order to make them more independent of specific technical models and thus enable their reuse in other applications. REFERENCES Baumgärtel, K., Katranuschkov, P., Scherer, R. J. (2012). Design And Software Architecture Of A Cloud-Based Virtual Energy Laboratory For Energy-Efficient Design And Life Cycle Simulation. eWork and eBusiness in Architecture, Engineering and Construction: 9-16. Reykjavik: CRC Press. Boje, C., Kubicki, S., Zarli, A., Rezgul, Y. (2021). A Digital Twin factory for construction. eWork and eBusiness in Architecture, Engineering and Construction. Moscow: CRC Press. Braud, A., Fromentoux, G., Radier, B., Le Grand, O. (2021). The Road to European Digital Sovereignty With Gaia-X and IDSA. IEEE Network 35(2): 4–5. Corneli, A., Naticchia, B., Carbonari, A., Vaccarini, M. (2021). A Framework For Development and Integration of Digital Twins in Construction. eWork and eBusiness in Architecture, Engineering and Construction: 291–298. Moscow: CRC Press. Faschingbauer, G. (2011). Simulationsbasierte Systemidentifikation im Rahmen der baubegleitenden geotechnischen Überwachung. Institute of Construction Informatics, Faculty of Civil Engineering, Technische Universität Dresden. Fuchs, S. (2015). Erschließung domänenübergreifender Informationsräume mit Multimodellen: Access of CrossDomain Information Spaces Using Multi-Models. Institute of Construction Informatics, Faculty of Civil Engineering, Technische Universität Dresden. Fuchs, S., Scherer, R. J. (2017). Multimodels - Instant nD-Modeling Using Original Data. Automation in Construction 75: 22–32. Haag, S., Anderl, R.(2018). Digital Twin - Proof of Concept. Manufacturing Letters 15: 64–66. Haller, A., Janowicz, K., Cox, S., Phuoc, D. L., Taylor, K., Lefrançois, M. (2017). Semantic Sensor
129
Network Ontology. https://www.w3.org/TR/vocab-ssn/. Web Resource (last visited: May 2022). Hamdan, A. H., Bonduel, M. (2019)a. Damage Ontology. https://alhakam.github.io/dot/ (last visited 08.04.2022). Hamdan, A. H., Bonduel, M., Scherer, R. J. (2019) b. An Ontological Model For The Representation Of Damage To Constructions. CEUR Workshop Proceedings 2389(6): 64–7. IECO Consortium. 2022. https://ieco-gaiax.de/. Web Resource (last accessed: May 2022). International Organization for Standardization (ISO) (2020). Information Container For Linked Document Delivery – Exchange Specification – Part 1: Container. https://www. iso.org/obp/ui/#iso:std:iso:21597:-1:ed-1:v1:en. Web Resource (last visited: May 2022). Katranuschkov, P., Hamdan, A. H., Lin, F., Polter, M., Mansperger, T., Heise, I., Scherer, R. J. (2020). Ein Simulations-und wissensbasiertes Systemidentifikationsverfahren für Brücken. Technical Report. Lim, K. Y. H., Zheng, P., Chen, C. H. (2020). A state-ofthe-art survey of Digital Twin: Techniques, Engineering Product Lifecycle Management and Business Innovation Perspectives. Journal of Intelligent Manufacturing 31(6): 1313–1337. Pauwels, P., Zhang, S., Lee, Y. C. (2017). Semantic Web Technologies in AEC Industry: A Literature Overview. Automation in construction 73: 145–165. Polter, M., Scherer, R. J.(2018). Towards The Application Of The BIMgrid Framework For Aged Bridge Behavior Identification. Proceedings of the 12th European Conference on Product & Process Modelling: 63–168. Polter, M., Katranuschkov, P., Scherer, R. J. (2020). A Generic Workflow Engine For Iterative, Simulation-Based NonLinear System Identifications. In 2020 Winter, Simulation Conference: 2671–2682. IEEE. Polter, M., Katranuschkov, P., Scherer, R. J. (2021). A Cyber Physical System For Dynamic Production Adaptation. ECPPM (2021)–eWork and eBusiness in Architecture, Engineering and Construction: 283–290. Moscow: CRC Press. Rasmussen, M. H., Pauwels, P., Lefrançois, M, Schneider, G. F. (2021). Building Topology Ontology. https://w3c-lbdcg.github.io/bot/. Web Resource (last visited: April 2022). Seiger, R., Huber, S., Heisig, P., Aßmann, U. (2019). Toward A Framework For Self-Adaptive Workflows In CyberPhysical Systems. Software & Systems Modeling 18(2): 1117–1134.
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
Processes
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Enriching BIM-based construction schedules with semantics using BPMN and LBD P. Hagedorn, K. Sigalov, L. Höltgen, M. Müller, T. Sola & M. König Chair of Computing in Engineering, Department of Civil and Environmental Engineering, Ruhr-Universität Bochum, Bochum, Germany
ABSTRACT: BIM-based construction scheduling is becoming increasingly important in research and practice due to the availability of the appropriate tools. However, it is characterized by a low level of detail and a lack of semantics, which is why the contained knowledge is hardly used for further analyses. Linked Building Data (LBD) is an intensively studied topic in construction information management, while the potentials and applications of LBD in BIM-based scheduling are still little investigated. LBD reveals new possibilities for improving semantic relationships between ontologies and enables the creation of knowledge graphs of BIM-based schedules for in-depth analysis. This paper investigates the enrichment of BIM-based scheduling with semantic information through formalizing schedules with BPMN 2.0 and converting and integrating process models into an RDFbased data structure in ICDD information containers. The integrated process knowledge can be queried via the SPARQL query language, allowing for extensible analysis and supporting decision-making processes.
1 INTRODUCTION The introduction of the BIM methodology has led to sustainable performance improvements in the construction industry (NBS 2018). Digitalization of products and processes by using BIM in the early planning stages and by adding further layers like time schedule (4D BIM) and cost estimation could further improve the efficiency and the collaboration and foster a shift toward more data-driven decisions (McKinsey 2020). Nevertheless, a breakthrough in BIM-based scheduling and cost estimation and their widespread practical use has not yet been achieved. In a more general sense, existing IT tools lack the interaction and combination of knowledge with building models (Wang & Meng 2019). The scattering of data across numerous documents and systems impedes Knowledge Management (KM) and the exchange and retention of knowledge through different construction phases. According to Wang & Meng (2019), BIM is still mainly used on the information level and a transition to the knowledge level needs future development. The authors emphasize the advantages of BIM-supported KM and the possibility of using BIM to overcome the gaps in IT-based KM. One of the identified gaps is the barely supported knowledge sharing across different domains. Thus, future research should consider improving semantic relationships between existing ontologies. Linked Building Data (LBD) summarizes the application of Linked Data (LD) on building data and integrates knowledge between domain-specific ontologies. Meanwhile, there are
DOI 10.1201/9781003354222-17
numerous published ontologies, such as Building Topology Ontology (BOT) (Rasmussen et al. 2021), Digital Construction Ontologies (DiCon) (Zheng et al. 2021), and Construction Tasks Ontology (CTO) (Bonduel 2021), as well as standardized exchange formats (ISO 21597-1 2020), which can provide a rich semantic environment for 4D BIM. A further important issue that prevents intensive integration of 4D BIM into the practice is the insufficient formalization methods and representation of construction knowledge, which allows for the automatic generation and maintenance of knowledge bases (Amer et al. 2021). Thus, the automatic information extraction in a process-oriented manner from 4D BIM models would be beneficial. In this context, the representation of construction processes using an established modeling language such as Business Process Modeling and Notation (BPMN) is reasonable for capturing explicit process knowledge and would contribute to BIM-supported KM. BPMN 2.0 is a well-established modeling language for business processes, which describes a graphical and a semantic notation for business processes and procedures. The notation is standardized in ISO/IEC 19510 (2013) and is widely used in the architecture, engineering, construction, and operation (AECO) industry. Thus, it is possible to formalize the process knowledge and represent it in an understandable way for all stakeholders involved. It enables the integration of events, which is important for more advanced constraint-based modeling and simulation (Wu et al. 2010). Besides its graphical representation, BPMN is
133
serialized in XML-based format for storing and interchanging processes. Due to its machine readability, it can be linked to knowledge graphs, enabling automated semantic enrichment of construction processes. Further advantages of this industry standard are the possible automatic checking of process models, executability, and easy exchangeability between different IT systems. The basis of the Semantic Web (SW) and LD approaches is the Resource Description Framework (RDF) (Pauwels, Costin, & Rasmussen 2022). This paper investigates the linkage of BIM-based processes with semantic information employing BPMN 2.0 XML and its conversion into RDF-based ontology data and proposes a respective concept and framework implementation. For the enrichment with semantics, existing ontologies, such as the BOT and CTO, and the RDFbased representation of IFC are examined. The presented solution is the result of the combination of two independent research outcomes and was tested within the scope of a case study. Thus, the goal is not the development of a monolithic application. Rather, this paper constitutes the feasibility of the presented idea and outlines its potential. The automated transformation of BIM-based schedules into BPMN-compliant processes can help to formalize process knowledge and create knowledge bases for construction scheduling. The final use of SPARQL enables to query spatial, topological, and schedule-related information, gaining new knowledge and supporting the decisionmaking process. Thus, detailed, semantically enriched BIM-based construction processes offer enormous opportunities for reuse and further in-depth analysis. 2 RELATED WORK 2.1 Methods for construction scheduling Due to the complexity of the task and the timeconsuming and error-prone manual creation of construction schedules, the scientific interest in computer-aided solutions is correspondingly high. The research focus of the last years was on the automation of construction sequencing, development of simulation models and optimization methods, and the conceptualization of knowledge-based systems to integrate expert knowledge (Amer et al. 2021; Faghihi et al. 2015; Montazer et al. 2017). Schedule simulations offer a powerful solution for detailed investigations and analysis of diverse scenarios. Nevertheless, the creation of suitable simulation models involves an enormous input effort. That is seen as the main weakness of this method (Wu et al. 2010) and is a substantial obstacle for the practical application. Additional advantages of the BIM method opened up new potentials for construction scheduling and led to intensive research in this application area. However, the initial effort to create detailed 4D models is high, and the software support for this process is not satisfactory. The automated schedule generation supplemented with the use of process templates would
facilitate the preparation of the input data for the 4D BIM-based simulations (Sigalov & König 2017; Wu et al. 2010). The extensive examination of research on automated construction planning and scheduling over the last three decades from Amer et al. (2021) equally confirms the need for an automated mechanism for the generation of such sequencing templates. Other gaps in knowledge identified as a result of this analysis are, i.a., the following: the rigidity of the data schemas for construction knowledge; infeasibility of manual creation and maintenance of construction templates and knowledge bases; the lack of synergies between automated planning systems and construction optimization. According to Amer et al. (2021), addressing these gaps would contribute to the further development of automated planning systems and testing of real-life projects.
2.2 Linked building data Recent research on the data exchange in AEC focuses on the use of SW and LD approaches for leveraging seamless information delivery in the lifecycle of construction projects (Pauwels et al. 2022). RDF framework provides the basis for these approaches and allows for the modeling of data in the form of subjectpredicate-object-triples of resources that are uniquely identified by means of Uniform Resource Identifiers (URI). RDF triple records form a graph by connecting two resources via a predicate. In addition to RDF, there is further key vocabulary for modeling complex schemata, so-called ontologies, which is standardized as the Web Ontology Language (OWL) (see Pauwels et al. 2022). With the development of the duality between IFC STEP in the different schema versions and ifcOWL, the use of building models in the context of RDF is facilitated (Pauwels & Terkaj 2016). Thus, a conversion from IFC STEP data to ifcOWL data can be performed (Bonduel et al. 2018). Besides the ifcOWL representation of building data, more lightweight ontologies are developed such as the BOT (Rasmussen et al. 2021). The BOT is developed to represent the topology of buildings and structures. It allows other domainspecific ontologies to refer to this central data model. A conversion routine from IFC to BOT is provided by Bonduel et al. (2018) and implemented in the IFC-to-LBD-converter. Moreover, this converter not only considers the topology as BOT instances but also converts semantic data of the IFC elements into RDF-based data. One of the advantages using SW technologies is the application of the dedicated query language SPARQL for retrieving data from RDF graphs (Harris & Seaborne 2013). In SPARQL, queries can be realized via the definition of graph patterns. In the context of this paper, SELECT is used to retrieve data from an RDF graph and CONSTRUCT is used to generate additional triples based on a query. SPARQL-Generate is a specification that extends SPARQL to read from structured documents such as JSON or XML
134
using iterator functions and to generate RDF triples (Lefrançois et al. 2017). 2.3 SW and LBD for construction scheduling LBD is also increasingly adopted in the field of construction scheduling (Getuli 2020; Karlapudi et al. 2020; Soman & Molina-Solana 2022). A productprocess integration approach for building renovation has been examined by Karlapudi et al. (2020) leading to an ontology-based modeling of building lifecycle stages, activities and stakeholders. However, their approach does not consider existing construction schedules and their linkage to BIM models. Getuli (2020) proposed a framework of modeling schedule knowledge in a domain ontology and relate it to BIM data in ifcOWL serialization to provide a knowledge base for construction scheduling. To model the sequence, the beginning, the end, and the duration of activities, they utilized the Time Ontology, which is a recommendation candidate of the World Wide Web Consortium. This ontology could be used reasonably in the scheduling context, however the associated scheduling ontology of Getuli is not RDFserialized available online. The CTO is developed by Bonduel (2021) for modeling tasks related to construction elements or spatial elements, e.g., defined by the BOT ontology. The ontology defines different types of tasks for installment, removal, modification, repair or inspection. The ontology contains properties for describing tasks, reifying triples on task accomplishment, and relating tasks to BOT individuals. Additional information of tasks can be added to the respective CTO objects using the Provenance Data Model (PROV) in the ontology serialization, as provided by Lebo et al. (2013). Information containers can be used to maintain 4D BIM models in a common context. To provide a rich semantic environment for 4D BIM models, the Information Container for linked Document Delivery (ICDD) according to the ISO 21597-1 (2020) is used in this research. This specification of
information containers allows for creating containers and link documents and entities inside these documents using SW technology. The storage of 4D BIM models in ICDD containers is demonstrated in Hagedorn & König (2021) in a use case of contextual validation of 4D BIM models. In this container, an IFC model and an XML schedule are stored and the links between entities of both documents are defined in an RDF linkset in compliance with ISO 21597. 3 METHODOLOGY The approach for the semantic enrichment of 4D BIM models adopted in this work comprises three stages of Schedule preparation, BPMN mapping and generation, and Semantic linkage. Figure 1 depicts these stages and gives an overview of the software components used, the procedure performed and the data generated. A BIM-based construction schedule provides the data basis for the concept. For the extraction of process knowledge, it is reasonable to have a detailed IFC model (LOD 300–400) with the associated tasks at least on the building element level. However, to be able to not only specify the completion order of building elements, but also to formalize the detailed execution processes, it is necessary to have multiple tasks per element. Due to their high complexity, such detailed schedules are still rarely produced. Furthermore, 4D models created with the available tools do not have sufficient structure to be used as a basis for simulation models. For this purpose, a prototypical implementation is used, which has been presented in Hartmann et al. (2012) and Sigalov & König (2017). This Java application allows for the automatic generation of detailed construction schedules using a constraint-based approach. Another important aspect is that a direct transfer of the entire schedule into a BPMN diagram is not reasonable and has no practical use. For this purpose, the Work Breakdown Structure (WBS) of the schedule can be used to identify the work packages. In the case where no WBS is given, the prototype can decompose
Figure 1. Process overview of the proposed concept divided into preparation, BPMN generation and semantic linkage employing the respective technologies.
135
Table 1.
Mapping overview of the BIM-based schedule, existing data structure, BPMN, and used ontologies.
ICDD container
Java Prototype
BPMN
CTO, BOT, PROV
Process ID Name StartTime EndTime ProcessDuration StartEvents: Map EndEvents: Map
Task ID Name Property(StartTime) Property(EndTime) -
cto:Task instance URI: sbld:Activity_{ID} cto:hasSimpleTaskDescription prov:startedAtTime prov:endedAtTime -
StartEvent
cto:afterFinishedTask
XML (Schedule) Tasks on lowest WBS level Task UID Name Start Finish Duration -
PredecessorLink
Predecessors Successors
-
-
IntermediateThrowEvent EndEvent SequenceFlow: sourceRef targetRef ParallelGateway
Process
Process
cto:TaskContext
ID Name
ID Name
instance URI: sbld:Process_{ID} rdfs:label
Entity
DataObject
bot:Element
EndEvents: Map
DataObjectReference
cto:isSubjectOfTask
-
cto:resultsInAddedStatement
cto:afterFinishedTask -
Tasks on upper WBS levels
Task OutlineNumber OutlineLevel UID Name IFC (Building model)
IFC Element Linkset
Link
the schedule into a set of related processes, referred to as sub-schedules, according to the algorithm presented in Sigalov & König (2017). Construction schedules defined using this prototype can be directly mapped into BPMN processes. In practice, however, schedules are created with conventional software tools, which is why only this scenario is considered (Figure 1 ①). The extraction of process models from the BIM-based construction schedules created in an external application requires some preparation steps described in the section 3.1. After decomposition into a set of sub-schedules, each sub-schedule is serialized into BPMN 2.0 XML as a BPMN process. For this purpose, a mapping, described in Section 3.2, is defined, which maps the basic elements to the data structure of the prototype by the elements of the BPMN (Figure 1 ⑦). Based on the BPMN XML serialization (Figure 1 ⑧), BPMN processes are transformed into an RDF document using SPARQL-Generate (cf. Section 3.3). In the last steps (Figure 1 ⑨– ), the IFC model associated with the construction schedule is transformed into an RDF graph and is semantically linked with the RDF graph of all BPMN processes.
3.1 Schedule preparation For the demonstration of the proposed concept, 4D BIM models are created in a commercial software application DESITE BIM from thinkproject (Figure 1 ①). This tool offers the possibility to create or import XML-based schedules with essential data about construction tasks and link them to IFC models. Next, the exported 4D model is transferred into an ICDD container in the ICDD Platform (Figure 1 ②– ③) as presented in Höltgen et al. (2021). The ICDD container is used as an open data model to share and handover 4D BIM models in a standardized way. The container includes a linkset, which relates each task from the schedule to one or more building elements from the IFC model. The schema of the container can be used to import 4D BIM models into other applications. To read the generated and exported ICDD container (Figure 1 ④) and convert the relevant data into the existing data structure of the Java prototype (Figure 1 ⑤), the container file is extracted, so its content can be read and parsed. Then, each file is read out by using a custom-written parser. The information about
136
the tasks and their sequences are extracted from the XML file, and the links to the IFC model from the RDF linkset. The mapping into the respective data structure and some required extensions are outlined briefly in the Section 3.2. The Java prototype follows the constraint-based approach, whereby all processes are defined including relevant information about constraints to be fulfilled and the produced results. Process constraints, representing prerequisites (StartEvents) and results (EndEvents), are generically described by Entity in certain states (StateData) (Table 3), making it possible to capture structural (IFC Elements) or non-structural elements in the same way. Such data structure allows for the automatic generation of schedules and can serve as a basis for simulation models, as further detailed in (Sigalov & König 2017). The core entity of the XML schedule is the element Task, which is mapped into the corresponding core class Process with the associated information about the ID, name, start time, end time, and duration. Each Process is then supplemented with data about constraints. First, the information about EndEvents is obtained by reading out the IFC GUIDs linked with the task from the RDF linkset. Thus, the attributes for the structural entities are provided directly by the IFC elements. StateData, including the description of the state and progress, are defined additionally, by inheriting the process name as the state name and setting the progress to 100 %. The order of the tasks is given by the PredecessorLink in the XML file (Table 3).After EndEvents are generated for all processes, this link is used to get the information about all prerequisites to be fulfilled. In this way, StartEvents are created by referencing all EndEvents of the predecessors. After that, the schedule can be generated automatically by matching StartEvents of one process to corresponding EndEvents of another process, which allows for changes and adjustments at any time. Process refinement is also possible at any time, since the prototype supports the hierarchical structure of the schedule through the definition and assignment of subprocesses. Therefore, the defined WBS from the XML file can be easily transferred as well. After the import of the ICDD is performed in the outlined way, the initial schedule is decomposed into a set of sub-schedules (Figure 1 ⑥), using the algorithm presented in (Sigalov & König 2017). Decomposition is given either directly using sub-processes defined according to the WBS or is set up automatically, grouping processes that belong to the same structural element or a group of elements. After that, each sub-schedule is mapped into the BPMN structure (Figure 1 ⑦), as described below. 3.2 BPMN mapping and generation A predefined or detected sub-schedule corresponds to a BPMN Process (Table 3), defined through its ID and Name. Core elements of a BPMN process are activities, events, gateways, and flows. An atomic
activity within a BPMN process is a Task, thus, all Processes, belonging to a sub-schedule, are mapped to BPMN Tasks. Attributes, such as ID and Name, are mapped directly to the corresponding attributes of the Task. Whereas, StartTime and EndTime are realized via BPMN Property elements, which can be assigned to a task. In BPMN, three different types of events are distinguished: Start, Intermediate, and End. The associated events of the first Process (or several parallel first Processes) in a sub-schedule are used to create a StartEvent. Similarly, the events of the last Processes are mapped as EndEvents. The start and end events of all other processes in the sub-schedule are created as Intermediate ThrowEvents. To provide the information on the IFC elements referenced within events in the Java application, the included Entities are mapped into DataObject elements, which is the primary construct for modeling data within the BPMN processes. DataObjects can be reused in the same BPMN process several times through DataObjectReference. For representing the process flow, the produced tasks and events have to be connected using flows and gateways. In BPMN, a SequenceFlow is used to connect two items and to show the order of the activities. SequenceFlow requires referencing the source and target items, which are given through the lists of predecessors and successors of a Process. In case an item has more than one predecessor or successor, a BPMN Gateway element is needed. The current version of the prototype implementation does not consider alternative process executions, which means that only parallel processes can be found. Thus, only ParallelGateways are used, which are generated additionally in such a case with several predecessors or successors. The described data structure enables the BPMN conversion (see Figure 1 ⑧) and forms the basis for the XML serialization. 3.3 Semantic linkage For the semantic linkage of the geometric and alphanumeric information of building elements with the procedural information and process states from the schedule, the established BOT ontology and the CTO ontology are mainly used. While DiCon (Zheng et al. 2021) also allows to model activities related to building elements, the ontology framework is considered being too comprehensive for this approach, but could be easily adopted for this process and can exchange or supplement the CTO ontology. Generating the CTO RDF data from the BPMN from Figure 1 ⑧, SPARQLGenerate is utilized as depicted in Figure 1 ⑨. For XML documents, there are functions in SPARQLGenerate XPath that make it possible to traverse through the XML document and define variables. The class cto:Task from the CTO ontology is used to model tasks from the BPMN according to Figure 1 ⑨ and Table 3. The property cto:afterFinishedTask is employed to
137
describe the procedural relation between tasks and preconditions for task executions. The property cto:has SimpleTask Method Description labels the respective task with a name or work description. Additionally, the start date and end date of tasks from the schedule are modeled using prov:startedAtTime and prov:endedAt Time. These vocabularies from the established PROV ontology provide the time and provenance information of tasks. Additional terms and vocabularies used in this research and the generated instances are defined in the namespace . The status of processes transferred into the BPMN events (StartEvent, IntermediateThrowEvent, EndEvent) is converted into a blank node of type rdf:statement which is then attached to the task via the cto:resultsInAddedStatement property.This statement asserts a triple to the RDF dataset referencing the respective building element as an rdf:subject, the rdf:predicate sbld:in State and the state string from the event as an rdf:object. Thus, the element is labeled with the status of the process after finishing the activity using RDF reification. Moreover, the upper WBS level tasks are generated from the BPMN according to Table 3 and allow for the grouping of tasks under a specific cto:TaskContext. To make the IFC building model compatible with the CTO graph data and allow queries on this aggregated graph, it must first be converted into the RDF format. Therefore, the IFCtoLBD-converter presented by Bonduel et al. (2018) is used as shown in Figure 1 ➃. Especially, the individuals of bot:element are of interest as schedule tasks are linked to these building elements. For this scenario, only properties are converted into RDF on PROPS Level 1 (c.f. Bonduel et al. (2018)). The properties generated from the building model can be used to perform extended queries (Figure 1 ), as shown in the results section. The generated RDF data is further processed and merged according Figure 1 . The CTO ontology defines the predicate cto:isSubjectOfTask as the connecting element between a bot:element as subject (rdfs:domain) and a cto:Task as object (rdfs:range). To link the CTO graph with the LBD graph, new triples of this pattern were constructed using SPARQL CONSTRUCT queries. In a first step, the CTO graph and the LBD graph were merged using a MERGE operation into a single default graph. On this graph, the CONSTRUCT query shown in Listing 1 was evaluated.
Listing 1. Using SPARQL CONSTRUCT to set the relation between bot:Element and cto:Task.
In the query body, all BOT elements and their associated GUIDs are first stored in SPARQL variables (Listing 1, lines 7–8) via the path props: globalIdIfcRoot_attribute_simple. Similarly, in lines 10–11, all CTO tasks and their corresponding GUIDs are stored in two further SPARQL variables using the predicate sbld:guid. The filter condition finalizes the query by ensuring that only a link between a BOT element and a CTO task is constructed, if the strings of the GUIDs match. The resulting new triples were then added to the default graph using another MERGE operation. These transactions on RDF data models are performed using Apache Jena, which is an open-source Java framework for creating applications for the SW and LD. The queries are executed using the Apache Jena ARQ application API. 4 RESULTS This section demonstrates the usage of the generated RDF graph after processing a 4D BIM model with the proposed method. The 4D model in this use case employs the Duplex apartment (BSI 2020) IFC model. The respective XML schedule is created in the software DESITE BIM and contains 25 tasks for strip foundation and five Level 1 tasks, like strip foundation - east, as shown in Figure 2. In the proposed methodology, ICDD containers are used to transfer the 4D BIM model from this proprietary software into the Java prototype and to integrate the enriched semantic data into the original container after being processed. The information containers used in this paper are created on a web platform presented in Hagedorn et al. (2022). This platform enables to generate and export containers, to add and to manage documents and RDF data inside containers, to create linksets containing several link types as specified in ISO 21597-1 (2020), and to perform SPARQL queries on the container. The detailed system architecture is described in Hagedorn et al. (2022).
Figure 2. Excerpt from the original schedule including a task group for a set of building elements.
138
The exemplary 4D BIM model is imported into the ICDD platform as shown in Figure 4. The structure of the created 4D ICDD container can be seen in the Explorer component of the platform. The corresponding RDF linksets are generated and the corresponding IFC model and XML schedule are located in the folder Payload documents. Afterwards, the container is read in the Java application for creating BPMN 2.0 XML files from it. For this, the files extracted from the container are read and parsed based on the methodology and mapping presented in Section 3.1. The resulting XML file contains BPMN processes for each Level 1 task. A BPMN diagram of the process illustrated in Figure 2, generated from its XML representation, is shown in Figure 3. Based on the XML representation, the semantic linkage is performed, enabling definition of semantic queries in the ICDD platform. First, the IFC building model is converted into the RDF format. After the generation and merging of BOT and CTO datasets from the BPMN according to the methodology presented in Section 3.1, these datasets are imported into the container as a payload triple file (see Figure 4, Explorer component). To demonstrate the feasibility of this approach, a SPARQL SELECT exemplary query is defined in Listing 2. This query represents a recurring task from the construction site planning where it is needed to gain insights about the concrete to be ordered on a specific day in a specific zone or level of the construction site. Thus, this query employs the quantity of material of the building elements, the topology of the building, and the procedural information from the schedule. The query is being evaluated on the final graph in order to demonstrate its usefulness. Results of the query are shown in Figure 4. The resulting quantities of concrete
per day and level are indicated and the linked elements in the IFC file are highlighted in the viewer. Listing 2. Querying delivery dates of concrete and required concrete per building storey for the task name “concrete placement”.
5 DISCUSSION AND CONCLUSION The use of LBD to bridge the gap between process and product information is a recent research topic in AECO industry. The scope of this contribution is to show, that (1) knowledge about construction processes can be automatically extracted from 4D BIM and formalized in a proven and standardized manner through BPMN; (2) the generated BPMN 2.0 XML datasets can be easily converted into the RDF format; (3) the converted data can be combined using existing ontologies and analyzed via semantic querying. Given the generic nature of the data schema as well as the ICDD and the BPMN standards, all BIMbased processes, e.g., in the operation phase, can be formally represented in the same manner. For semantic enrichment, domain-specific ontologies have to be
Figure 3. Visualization of the generated BPMN diagram.
Figure 4. 4D BIM model use case: Duplex building in a 4D ICDD container combined with schedule and generated triples from the proposed concept.
139
utilized, and the corresponding data must be converted into the respective RDF instances. The transformation of BIM-based schedules into BPMN-compliant processes provides a flexible and common data schema for explicit modeling of construction knowledge. It presents an automated mechanism for the extraction of process-oriented knowledge and could facilitate the creation of knowledge bases. In this sense, supporting the extraction of process templates is another possible use case of the introduced approach (Sigalov & König 2017). For this purpose, the intermediate step of pattern recognition (Figure 1 ), which is also implemented in the Java prototype, could be switched in. In this case, the application would identify the repeating processes and export only one process per cluster as a BPMN diagram. Combining 4D BIM with semantic information opens the possibility to query structured data and perform various in-depth analyses, including spatial, topological, and schedule-related querying. Furthermore, semantically enriched schedules open up further opportunities for automatic pattern recognition as well, because topological querying enables, e.g., location-driven grouping of tasks. By integrating additional ontologies, this information can also be taken into account during pattern recognition. It could contribute to combining knowledge from different projects and domains and creating cross-company or open-access repositories for construction processes. REFERENCES Amer, F., Koh, H. Y., & Golparvar-Fard, M. (2021). Automated Methods and Systems for Construction Planning and Scheduling: Critical Review of Three Decades of Research. Journal of Construction Engineering and Management 147(7). Bonduel, M. (2021). A Framework for a Linked DataBased Heritage BIM. Doctoral Dissertation, KU Leuven, Belgium. Bonduel, M., Oraskari, J., Pauwels, P., Vergauwen, M., & Klein, R. (2018). The IFC to Linked Building Data Converter - Current Status. In Proceedings of the 6th Linked Data in Architecture and Construction Workshop (LDAC 2018), pp. 34–43. BSI (2020). DuplexApartment Test Files. URL https://github. com / buildingSMART / Sample -Test - Files / tree / master / IFC%202x3/Duplex%20Apartment. Last accessed: 21.04. 2022. Faghihi, V., Nejat, A., Reinschmidt, K. F., & Kang, J. H. (2015). Automation in construction scheduling: a review of the literature. International Journal of Advanced Manufacturing Technology 81(9-12), 1845–1856. Getuli, V. (2020). Ontologies for Knowledge modeling in construction planning. Florence: Firenze University Press. Hagedorn, P. & König, M. (2021). BPMN-related Ontology for Modeling the Construction Information Delivery of Linked Building Data. In Proceedings of the 9th Linked Data in Architecture and Construction Workshop, pp. 91– 102. Hagedorn, P., Liu, L., König, M., Hajdin, R., Blumenfeld, T., Stöckner, M., Billmaier, M., Grossauer, K., & Gavin, K. (2022). BIM-enabled InfrastructureAsset Management using Information Containers and Semantic Web. Forthcoming. Journal of Computing in Civil Engineering.
Harris, S. & Seaborne, A. (2013). SPARQL 1.1 Query Language: W3C Recommendation 21 March 2013. Hartmann, V., Beucke, K. E., Shapir, K., & König, M. (2012). Model-based Scheduling for Construction Planning. In 14th International Conference on Computing in Civil and Building Engineering, Moscow. Höltgen, L., Cleve, F., & Hagedorn, P. (2021). Implementation of an Open Web Interface for the Container-based Exchange of Linked Building Data. In Proceedings of the 32nd Forum Bauinformatik 2021, TU Darmstadt, Germany, pp. 174–181. ISO 21597-1 (2020). Information container for linked document delivery: Exchange specification - Part 1: Container. ISO/IEC 19510 (2013). Information technology - Object Management Group Business Process Model and Notation. Karlapudi, J., Menzel, K., Törmä, S., Hryshchenko,A., & Valluru, P. (2020). Enhancement of BIM Data Representation in Product-Process Modelling for Building Renovation. In Product Lifecycle Management Enabling Smart X, Volume 594 of IFIP Advances in Information and Communication Technology, pp. 738–752. Lebo, T., Sahoo, S., & McGuiness, D. (2013). The PROV Ontology. W3C Recommendation. URL https://www. w3.org/TR/prov-o/. Last accessed: 21.04.2022. Lefrançois, M., Zimmermann, A., & Bakerally, N. (2017). A SPARQL Extension for Generating RDF from Heterogeneous Formats. In The Semantic Web, Volume 10249 of Lecture Notes in Computer Science, pp. 35–50. McKinsey (2020). The next normal in construction: How disruption is reshaping the world’s largest ecosystem. McKinsey & Company. Montazer, M., Rebolj, D., & Heck, D. (2017). A Comparison Review of Automated Construction Scheduling Methods. In Proceedings of the Joint Conference on Computing in Construction (JC3), Heraklion, Greece, pp. 137–144. NBS (2018). National BIM Report No. 8. RIBA Enterprises Ltd. Pauwels, P., Costin, A., & Rasmussen, M. H. (2022). Knowledge Graphs and Linked Data for the Built Environment. In Industry 4.0 for the Built Environment, pp. 157–183. Pauwels, P. & Terkaj, W. (2016). EXPRESS to OWL for construction industry: Towards a recommendable and usable ifcOWL ontology. Automation in Construction 63, 100–133. Rasmussen, M. H., Lefrançois, M., Schneider, G. F., & Pauwels, P. (2021). BOT: The building topology ontology of the W3C linked building data group. Semantic Web 12(1), 143–161. Sigalov, K. & König, M. (2017). Recognition of process patterns for BIM-based construction schedules. Advanced Engineering Informatics 33, 456–472. Soman, R. K. & Molina-Solana, M. (2022). Automating look-ahead schedule generation for construction using linked-data based constraint checking and reinforcement learning. Automation in Construction 134. Wang, H. & Meng, X. (2019). Transformation from IT-based knowledge management into BIM-supported knowledge management: A literature review. Expert Systems with Applications 121, 170–187. Wu, I. C., Borrmann, A., Beißert, U., König, M., & Rank, E. (2010). Bridge construction schedule generation with pattern-based construction methods and constraint-based simulation. Advanced Engineering Informatics 24(4), 379–388. Zheng,Y., Törmä, S., & Seppänen, O. (2021). A shared ontology suite for digital construction workflow. Automation in Construction 132.
140
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
A simulative framework to evaluate constructability through parameter optimization at early design stage F.L. Rossini & G. Novembri Department of Civil, Constructional and Environmental Engineering – DICEA, Sapienza, University of Rome, Italy
ABSTRACT: Construction process is becoming progressively more complex in terms of time and cost because of the in-creasing number of parameters to satisfy, whilst persist a lack of collaboration among design and construction actors. Construction phases strategies rarely influence choices made in design phase, and they considered consequently 4D and 5D analysis outcomes and not objectives of the whole design process. So, the aim of the research is to develop a methodology capable of providing designers a real time simulation of optimised con-struction solution. To reach this result, a combined algorithm of structural solution, construction difficulties and performance is developed. This collaborative procedure can check in real time the feasibility of the design solution in structural and constructional terms. So, a framework of the Visual Programming linking among the different tools involved is presented, and finally an experiment on a real construction site is briefly analysed. The future works will improve the autonomy of the system by changing the input data collected by user with those elaborated from experiences through a machine learning approach.
1 GENERAL BACKGROUND AND RESEARCH MOTIVATIONS 1.1 About current productivity of construction industry The construction sector satisfies one of man’s basic needs, up to characterise the culture of a people or even an era. The data available prior to the recent pandemic show that the construction industry accounts for approximatively 6% of GDP (AA VV, 2018) and, in fact, represents one of the most important components of the nation’s Gross Domestic Product - GDP. Considering the development of emerging economies such as China and India, they valued the development of the sector by 2030 at 14.7% of world GDP. So, considering those data, we can state that the improvement of the sector’s productivity by only the 1% would generate savings of about various Billion per year. Subsequently, we must add the beneficial effects on employment that an increase in the sector’s productivity could trigger, in relation to the lack of productivity growth that affects the construction sector (Figure 1). According with (De Boeck et al. 2019), the quality and quantity of construction can affect the liveability and economic attractiveness of cities and the well-being perceived by the population, while also highlighting its important social role. Construction also plays a strategic role because this sector realises and guarantees fundamental services
DOI 10.1201/9781003354222-18
to the life and the development of a country such as infrastructure, telecommunications, health, and education. Despite its fundamental role, a serious lack of formalisation and optimisation of processes still characterizes the sector. We need to consider that while the automotive sector from 1995 to 2017 improved productivity of about 3.0%, the building construction industry registers now a loss of productivity estimated of about 2,2% (Oxford Economics 2022). So, this delay of innovations and related productivity issues affects many aspects of economics. From a purely econometric point of view, the factors that hamper the productivity growth linked to technical and legislative aspects are analytically determined, and mostly depends to the skills, technology adopted, distribution and dimension of construction companies (Hasan et al. 2018). It is interesting to note that this study also considers factors related to the legislative framework, which today appears to be backward regarding the development of digital techniques (Noardo et al. 2022). In this sense, the research line of reshape regulation is active today in relation to introducing digital techniques in the management of contracts such as blockchain technology (Kiu 2020). From a purely technical point of view, we can identify several factors that affect productivity as the low level of construction and procurement processes, the limited use of automation and supply chain
141
Figure 1. Historical trend (to 1970 to 2020) of the output per hour worked. A comparison between construction industry and market sector. As we can see productivity has changed little in the construction industry in the past 50 years (Office for National Statistics – UK, 2021).
optimisation systems, an inadequate level of worker’s technical skills, limited formalisation, standardisation and simulation of construction processes and a very poor use of project management and lean construction technologies (Vaux & Kirk 2018). Evidently, a key-challenge that will be determinant is the pervasive introduction of digital technologies into construction, considering the capability of digital processes to avoid human errors and to automatise procedures, as the Construction 4.0 aims to. (Schoenbeck et al. 2020). This process has been ongoing for many years, mostly for design tasks, but the introduction of hitech solutions was hampered by some typical features of the building sector. Considering the specific case of the Italian market, both of Design and Construction companies are mostly small or medium-sized, and there is an inner difficulty to take the wave of innovation, due to the impossibility to exploit economies of scale for the amortisation of investments in equipment, education, and training of personnel. Otherwise, also from the client point of view, the implementation of digitalisation in their organization needs a completely new organization of skills, tools, and procedures. 1.2 Current methods and tools One cause that limits the development of an innovative formalisation of techniques and procedures for the construction management is for the tendency to use simple tools, often improperly such as the Gantt chart, which is appreciated above all for its immediacy of reading. Despite its clarity and immediacy, this simple technique, which is widely used in the sector, is ineffective in analysing situations where activities overlap. The technique makes it possible to indicate how many and which activities interfere in a work phase but does not make it possible to identify where this occurs on the site (Novembri et al. 2017). Nowadays, this tendency makes it difficult to apply advanced project management techniques and,
specifically, the full use of the Lean construction approach which, although known since the end of the 21st century (Green 2002), is still applied to a limited extent in the construction sector. If it were possible to overcome the criticalities outlined above, what has been defined as ‘productivity boosters’ would be activated, which are now essential for the construction sector. The immediate consequence of the extensive application of the previously mentioned techniques would be the formalisation and optimisation of the processes, which would have a positive impact in terms of environmental protection, reduction of polluting emissions, raw materials used and a lower impact on the environment because of the reduction of execution times. It should be considered that the impact could be significant because the construction sector absorbs approximatively 50% of global steel production and handles 30% of global pollutant gas emissions (Craveiro et al. 2019). 1.3 Limits and motivation of research Very often the diseconomies of the construction phases are due to design choices that do not consider the construction aspects (Love et al. 2016). Unforeseen difficulties on the construction site led both to longer working times and to results that are sometimes different from what planned (Hyun et al. 2020). To overcome this, we can handle various protocols of digital design management that implies a high level of collaboration among the actors of the process (Mencarelli et al. 2020). This could imply a fragmentation of models, difficulties in data exchange, and coordination issues. At the same time, the construction sector needs the activation of ‘productivity boosters’ that requires an enhanced level of automation of processes for reducing human error, reworking, and decision bottleneck (Love et al. 2019). The complexity of architectural design requires an improved formalisation of processes, which now appears to be linked to the ability of actors to manage the available methods and tools. So, the research proposal is about the improvement of methodologies and tools available currently for the orchestration of processes in construction, carrying them in a simulative environment automatically managed by Visual Programming – VP applications. The aim is the creation of a database populated no longer only by the properties and geometries of the elements (ie BIM model) but inclusive of the preferences and automatic evaluation of proposals, based on criteria of choice imposed by the designer or automatically deduced from previous experience, to be elaborated in future works through a machine learning approach. For these reasons, we developed a simulative model which creates a whole series of alternatives based on the achievement of the same structural safety standards of the project but reducing the cost of the material and the costs associated with the time and difficulties for the construction phases.
142
The study proposes a way in defining this process of formalisation and management, using advanced ICT tools directly available by designers, thanks to the possibilities given by Visual Programming. In fact, as predicted by Negroponte, the Augmented Architect needs to create their own digital tools and libraries of objects for increasing its design capabilities (Negroponte 1975). Currently, this is possible thanks to the Application Program Interface - API and with the Visual programming applications directly imbedded in digital design tools of current use. In this study, therefore, it was important to analyse the state-of-the art and test the capabilities of Reactive Programming (Ghannad et al. 2019) to assess the feasibility of the project in terms of time and costs through a continuous and iterative exchange between tools of the so-called software constellation. 2 DESIGN OPTIMISATION THROUGH THE VISUAL PROGRAMMING APPROACH EXPLAINED We tested the effectiveness of this methodology by applying the exposed optimisation strategy in a building refurbishment process. The goal is to link the automation of structural analysis with the economic evaluation of constructability of the identified solution. The implementation path of this methodology comprises: - BIM modeling (min. LOD 200); - Definition of structural analysis, design parameters and components’ library; - Definition of Constructability scores; - Formalisation of the target algorithm. During those steps the use of the Autodesk™building suite improves the fluidity of exchange between
software. But, on the other hand, this implies to consider the existing limits for the interoperability among different digital environments.At the end, using Revit, Dynamo and Robot ® we can rely on a robust workflow that limits the loss of information due to interoperability lacks. In the case study, we achieved the optimisation of the installation of the new metal trusses of the roof, the most important task due in this contract. The first step is the creation of a BIM model where we can identify the areas involved in the installation of the trusses (Figure 2). Here the complexity of the roofs’ shapes helps in evaluating the reliability of the methodology, making possible to verify the constant geometric consistency between the auto-generated result and the acceptability of the solution. Once the surface to be covered and the section of the roof has been defined, it carried the automatic subdivision of the influence areas of the trusses. Then we proceed with the rules setting. One of these is the definition of the min. and max. distance between trusses (ie: trusses span), to avoid since the early design phases all the solutions that could be evidently not applied (eg: having one hundred trusses over a linear development of twenty meters or having only two of them), and finally could represent only an unnecessary computational load on the workflow. Afterward, it needs to set the structural axes for the solutions that comply with these rules, to prepare the export to the structural analysis tool. The trusses, so far defined only in its spatial dimensions, are further detailed through the definition of the ties and struts as standard UNI profiles (ie HEA/HEB/UPN etc) and by the weight. Here too, as above, is made an a priori imposition of the range of elements’ number, to focus the computational efforts on the variables that may have technical acceptability.
Figure 2. Visual Programming section: automatic definition of structural trusses typology, their number and pace.
143
Figure 3. An excerpt of whole series of possibilities. In the x-axis the solution subdivided by the number of trusses. In the y-axis by the number of ties and struts.
Once it moves results directly to Autodesk Robot. This export process is made through a Dynamo script that automatically activates the structural analysis tool, and the architectural model is uploaded within, where these geometries are interpreted as line and nodes. This is possible by defining a script in Python that allows, through API, the direct access to the structural analysis tool. Within the Robot environment, VP
is used to produce the needed steps for the correct dimensioning of metallic profiles, depending on the type of truss assessed and the number of the elements required. To ease the analysis, it dimensions only the rods, while for the nodes an increase in the overall weight of 20% calculated on the total weight of the rods is imposed. The graphical result exposes a whole series of possibilities (Figure 3), given both by a different thickening of the trusses and, at a deeper scale, of the ties and struts. It is possible to set an output in *.xls format which, besides graphically illustrating the proposed solutions, associates the weight of the profiles with a unit price. Here, the preferred parameter is the global weight of the structure, to ease the movement of materials and allow the installation of structures, avoiding the use of complex construction machines, and equipment. The system defines the geometries and the profiles used, then the automated analysis procedure invokes the construction parameter for assessing the difficulty of execution of the various solutions. To avoid the increase of time due to data processing lump-sum values of difficulty have been attributed, starting from a minimum difficulty for solutions involving a smaller number of trusses because the larger working area, and fewer movements at height to a maximum level of difficulty with a larger number of trusses and ties and struts, and a consequent small area to work in. In this case, just considering the weight of the structures, there is a clear advantage in installing six trusses and four internal fields, where a field is made by four uprights joined by four crossbars. If we were to compensate for the values purely related to weight, with the corrective factor given by the difficulty of execution, we would see that the most convenient choice is the construction of five trusses characterised by three fields, thus confirming how much better it is to reduce the work on site in favour of a slight increase in the material’s weight (Figure 4).
Figure 4. On the left, the selected structural model represented in Robot environment. On the right, part of the real realized trusses.
144
3 CONCLUSIONS The subsequent actual construction of the structures only confirmed what we had assumed in the design phase, guaranteeing a saving of approximatively 15% compared to what the structural designer stated in his design evaluations. This is because previous design evaluations do not automatically combine aspects of static safety with the convenience and feasibility in the construction site, referring these evaluations to the designer’s implicit knowledge, which is mostly subjective. Or, in the worst case, to a claim from the construction company or a design variation occurred during the working phases, and the related waste of time and money. Furthermore, in this specific project, an important outcome was reached also from a procedural and administrative point of view. It was essential to verify that the new trusses weighed less than those removed, to include the work in the case provided by structural regional law as ‘local intervention’. This condition avoided to take the way of the global structural retrofit of the building, for which it would not have been possible to cover the related cost. Considering the high automation of this procedure, these computationally and technically demanding checking become a pre-set rule in the definition of structural design alternatives.
One of the possible next steps could be to extend the methodology developed to the construction supply chain, which is as complex as it is varied, and to introduce logistics automation systems already used for Big E-commerce platforms. This step would further reduce the need for storage areas on the construction site, as it would provide an operational supply chain completely ‘on demand’. In this way, it would be able, in a finally industrialised sense, to interpret the Just-In-Time - JIT concept on a large scale. REFERENCES
4 DISCUSSION AND FUTURE WORKS From the analysis of the results, it emerged that by combining and extending the potential of commonly used tools, the resources of the building process can be optimised. Besides the optimization of resources, here we propose methodologies capable of offering design alternatives, in order not only to improve the construction of the project but also to introduce the capability of the Digital Design Support System to optimise the project by itself. So, following this line of development, we will automatically review the project continuously, till to reach a near optimum balance between prefigured goals and resource savings. This is a further confirmation that the future of the BIM will be not an evolution of capability of the BIM intended such as a dynamic database of building element but the level of Co-Creation, in which AI-driven tools are constantly matching building elements, context data and client requirements for producing design alternatives, listed by their rate of fulfilment of the imposed project requirements. For the architectural design realm, although it has made important advances from this point of view, the time for the formalisation of an algorithm of beauty is still far away. Evidently, once they will achieve such a complexity of artificial thinking, we would face a cognitive evolution in the broadest sense of the term, because it would mean that machines felt feelings, arbitrary preferences, and sensory experiences.
145
AA.VV. 2017 Reinventing Construction: A Route to Higher Productivity, McKinsey Global Institute. Craveiro, F., Duarte, J.P., Bartolo, H., Bartolo P.J. 2019. Additive Manufacturing As An Enabling Technology For digital construction: a perspective on Construction 4.0. Automation in Construction (103), pp. 251–267. De Boeck, S., Bassens, D., Ryckewaert, M. 2019. Making space for a more foundational economy: the case of the construction sector in Brussels, in Geoforum (105), pp. 67–77. Ghannad, P., Lee, Y.-C., Dimyadi, J., Solihin, W. 2019. Automated BIM Data Validation Integrating Open-Standard Schema With Visual Programming Language Advanced Engineering Informatics (40), pp. 14–28. Green, S.D. 2002. The Human Resource Management Implications of Lean Construction: Critical Perspective And Conceptual Chasms. Journal of Construction Research (03)1, pp. 147–165. Hasan A., Baroudi B., Elmualim A., Rameezdeen R. 2018, Factors Affecting Construction Productivity: A 30 Year Systematic Review. Engineering, Construction andArchitectural Management, Vol. 25 No. 7, pp. 916–937. Hyun H., Kim H., Lee H.-S., Park M., Lee J. 2020. Integrated Design Process for Modular Construction Projects to Reduce Rework. Sustainability 12(2), p. 530. Kiu, M.S. 2020. Exploring The Potentials of Blockchain Application in Construction Industry: A Systematic Review. International Journal of Construction Management, DOI: 10.1080/15623599.2020.1833436. Love P.E.D., Smith J. 2016. Toward Error Management in Construction: Moving Beyond A Zero Vision. ASCE Journal of Construction Engineering Management 142(11), p. 04016058. Love P.E.D., Smith J., Ackermann F., Irani Z. 2019. Making Sense of Rework And Its Unintended Consequence in Projects: The Emergence of Uncomfortable Knowledge. International Journal of Project Management 37(3), pp. 501–516. Mencarelli, L., Chen, Q., Pagot, A., Grossmann, I.E. 2020. A Review on Superstructure Optimization Approaches in Process Systems Engineering. Computers & Chemical Engineering (136) 8. Muizz O. Sanni-Anibire, Rosli Mohamad Zin & Sunday Olusanya Olatunji. 2020. Machine Learning Model For Delay Risk Assessment in Tall Building Projects, International Journal of Construction Management. Negroponte N. 1975. The Architecture Machine. ComputerAided Design 7 (3), pp. 190–195. Noardo F., Guler D., Fauth J., Malacarne G., Mastrolembo Ventura S., Azenha M., Olsson P.-O., Senger L. 2022. Unveiling The Actual Progress of Digital Building Permit: Getting Awareness Through A Critical State of The Art Review. Building and Environment (213), p. 108854.
Office of National Statistics (UK) 2021. https://www.ons. gov.uk/economy/economicoutputandproductivity/ productivitymeasures/articles/productivityintheconstructionindustryuk2021/2021-10-19. Last access: 15/06/2022. Schoenbeck P., Loefsjoegard M., Ansell, A. 2020. Quantitative Review of Construction 4.0 Technology Presence
146
in Construction Project Research. Buildings, 173(10), p. 10100173. Vaux, J.S. and Kirk, W.M. 2018. Relationship Conflict in Construction Management: Performance And Productivity Problem. Journal of Construction Engineering and Management, 144(6), p. 04018032.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Automatic generation of work breakdown structures for evaluation of parallelizability of assembly sequences J.M. Weber & M. König Chair of Computing in Engineering, Department of Civil and Environmental Engineering, Ruhr-Universität Bochum, Bochum, Germany
ABSTRACT: To enable the efficient construction of industrial plants, seamless logistics on the construction site must be ensured. It can be achieved through precise planning and execution of logistics and assembly. Nowadays, assembly sequences are visually represented in practice using 4D BIM (Building Information Modeling) concept. The assembly schedules on the other hand, are prepared by hand in an imprecise and time-consuming manner. Our previous research has shown that semi-automatic rule-based creation of step-by-step assembly sequences is possible for any IFC (Industry Foundation Classes) models with a high level of detail. Step-by-step assembly sequences are generated by categorizing all building elements in the building model, specifying assembly rules for these categories, and selecting an assembly strategy subsequently. However, a qualitative comparison of these generated assembly sequences is impossible without performing simulation and determining the effort values for each task. To address this issue, this paper presents a methodology for the automatic generation of partial Work Breakdown Structures (WBS) to evaluate the generated assembly sequences. On the basis of the assembly sequence, the assembly rules and the collision database (of slightly augmented bounding boxes of building elements) of a BIM, partial WBS are generated for each individual building element. From the WBS, the length of the longest path to the corresponding start element is retrieved for each building element. This allows building elements to be assigned to the appropriate assembly levels, indicating that all assembly processes that are located in the same level, and thus potentially parallelizable, can be identified.
1 INTRODUCTION Large-scale projects such as plant construction are complex tasks in which seamlessly running logistics is critical to success. The logistics on the construction site include the flow of materials and resources and the optimization of processes. Today, logistics is still mainly planned by hand, associated with a significant amount of personnel effort and is prone to errors. Up to now, digital planning tools have not supported the planning of construction site logistics of this kind, requiring considerable project experience for the planner. Precise scheduling avoids problems with the temporary storage of building elements, while it also ensures that machines and personnel are not idle because critical building elements are missing or damaged due to improper storage, and replacements must be ordered. In our previous study (Weber et al. 2020), we have shown that the automatic generation of step-by-step assembly sequences for construction projects is possible and that different strategies for the assembly of buildings can be applied (Weber & König 2022). However, until now, there has been no way to objectively evaluate these strategies without generating a time schedule.
DOI 10.1201/9781003354222-19
This leads to the research question of how different strategies for automatically generated assembly operations can be compared in terms of their efficiency without knowing the exact effort values for each specific assembly task. In our approach, we therefore present an evaluation of the generated step-bystep assembly sequence using automatically generated Work Breakdown Structures (WBS) and their comparison. WBS is defined in the Project Management Body of Knowledge (PMBOK) as a “deliverable oriented hierarchical decomposition of the work to be executed by the project team” (Project Management Institute 2013). WBS is a tree structure that divides the entire project hierarchically into smaller and more manageable phases, deliverables, and work packages. WBS is used to represent the structure of projects in early phases by breaking them down into manageable activities in levels as an activity diagram. In practice, the individual project phases may overlap for faster project execution (Berthaut et al. 2014). With the proposed methodology, WBS can be created for each building component and the entire construction. In addition to the familiar tree structure of WBS, this can also result in further links with subordinate building elements when building elements are connected to more than one building element in different WBS levels.
147
2 RELATED WORK WBS was introduced as early as 1962 by the Department of Defense (DoD) in collaboration with the National Aeronautics and Space Administration (NASA) as part of a Program Evaluation Review Technique (PERT) costing system guidance document (United States 1962). Kadet & Frank (1964) first mentioned WBS in connection with PERT in the context of its potential suitability for planning in construction. Hashemi Golpayegani & Emamizadeh (2007) used neural networks to design WBS for project planning of construction projects and proposed their own process with specific steps and procedures. Norman, Brotherton & Fried (2008) name different types of WBS: task-oriented, process-oriented, deliverable-oriented, or time-oriented. The authors criticized the fact that task-oriented WBS are descriptions of the project processes and not the result of the project. They saw it as a risk that the project team spends much energy on the refinement of the project processes, which does not necessarily lead to the desired results, as opposed to a deliverable-oriented WBS. The WBS in the present approach are component-oriented and thus close to the mentioned task-oriented WBS. However, the focus is not on a process optimization using the WBS itself but only on an objective evaluation of assembly sequences by means of the corresponding WBS. Over the past decade, several tools and methods have been developed for sequencing activities and generating WBS. In general, BIM (Building Information Modeling) models, or three-dimensional building models, are largely limited to established methods in civil engineering for WBS development (Liu, AlHussein & Lu 2015). Siami-Irdemoosa, Dindarloo & Sharifzadeh (2015) recognized few published methodologies and tools for developing WBS for structures other than apartment buildings and boiler manufacturing. Therefore, they developed a methodology based on hierarchical neural networks with higher generalizability, which can be used for complex underground projects. More generic approaches such as neural networks (Bai et al. 2009; Golpayegani & Parvaresh 2011; Siami-Irdemoosa, Dindarloo & Sharifzadeh 2015) and knowledge-based reasoning (Mikulakova et al. 2010) require extensive historical data for learning and reasoning. Li & Lu (2017) proposed an analytical network flow-based optimization methodology to automatically generate network diagrams and WBS for earthworks project planning to cope with these problems. Their objective was to reduce interference between work packages and thus between subcontractors. For this purpose, the authors first define an optimal earth flow network and divide it into subflows that can be parallelized without interference. The subflows are then divided into haul jobs. Sharon & Dori (2012) also reviewed the WBS method and criticized the absence of a direct and explicit representation of the product facet in the project plan. Therefore, they proposed a new version of WBS that was enhanced to include product-related information. The revised
WBS allows project managers to focus on completing the required deliverables rather than on the execution of processes, as these do not necessarily align with expected outcomes.
3 PROPOSED METHODOLOGY In this research, an extension for the thinkproject BIM management software DESITE md is developed in order to map building elements of BIM models into WBS based on their assembly sequence so that conclusions can be derived about the potential parallelizability of assembly processes in the assembly sequence on the basis of the resulting WBS layers. With this knowledge, objective decisions can be made regarding the selection of assembly strategies. Until now, the selection of the strategy for generated step-bystep assembly sequences, as presented in our previous work (Weber et al. 2020), was purely subjective in the absence of exact effort values. In order to be able to make an objective statement about the quality of the selected strategy independently of exact effort values, the focus is therefore placed on the parallelizability of assembly processes. This parallelizability is made clear by WBS, in which parallelizable assembly processes are located on the same level. All assembly processes in the same level are independent of each other but require the levels below. The developed methodology for generating WBS is shown in Figure 1. The left side of the figure shows the input variables. These consist of the assembly sequence, the ruleset, and the collisions database, which formed the basis for the generation of the assembly sequence (Weber et al. 2020). The assembly sequence embodies the step-by-step building instructions, i.e., a list of the IDs of the components in the specific building order. This sequence must contain at least all construction steps up to the building element for which the WBS is to be created. If the WBS is to be created for the entire building, all assembly steps must be available in the assembly schedule. The ruleset and the collisions database form the basis for generating the assembly sequence. The ruleset contains all the necessary assembly rules for the correct assembly sequence. The collisions database identifies neighboring components and contains information about collisions between the slightly augmented bounding boxes of individual building elements. Depending on the configuration of offsets, even small distances between building elements can be listed as collisions since experience shows that BIM models are rarely created without errors, i.e., without occasional gaps or overlaps between building elements. A loop function is shown in the middle of the figure, which is repeated for all neighboring building elements. Thus, the element selected for creating the WBS is first taken from the assembly sequence, and then the building elements required for its assembly are identified using the ruleset and the collisions database. For these
148
Figure 1. Methodology for automatic generation of WBS.
identified components, the components required for their assembly are identified again, and so on, until finally, all components required for assembly are listed in the WBS according to their interdependencies. The WBS levels result from the interdependencies of the building components. Components that are a prerequisite for the construction of another component are assigned a lower WBS level than the related higherlevel component. For example, component B, whose collision- and rule-compliant assembly is directly dependent on component A, is assigned the next higher WBS level, i.e., increased by 1. Starting with, for instance, the foundation, parallel WBS branches can be derived so that various independent components may be assigned the same WBS level and are therefore potentially parallelizable. In the implementation, a detailed example of the WBS leveling is discussed.
4 IMPLEMENTATION A simplified 2D model of a steel construction is used here to visualize the implementation better. The steel structure shown in Figure 2 consists of a foundation (F), two base plates (BP), two columns (CM), four connecting plates (CP), five beams (types BM, BA, and BC), two vertical beams (BV), one cantilever (CL) and 16 bolts (B). The rules belonging to the assembly are given in Equations (1) to (6). The prerequisite for the assembly of a BM is that BM collides with two CPs, as shown in Equation (1). According to Equation (2), the assembly of a CP requires one CM.According to Equation (3), a BV requires one CM and one BM. Equation (4) assumes the presence of the F for the assembly of a BP, and the assembly of a CM presupposes one BP according to Equation (5). Equation (6) shows the rule for constructing the fasteners B and the welds (W). The
prerequisite for their assembly is that they connect any two building elements. BMr = 2CP
(1)
CPr = CM
(2)
BVr = CM ∧ BM
(3)
BPr = F
(4)
CMr = BP
(5)
Wr = Br = 2
(6)
R1 = (Wr , Br , {BMr , CPr , BVr , BPr , CMr })
(7)
The rules from Equations (1) to (6) are combined in Equation (7) to form the ruleset R1. The tuple here shows the prioritization of the connection elements W and B. Based on the collision data and the ruleset, the step-by-step assembly sequence is shown in Figure 3. The dotted lines represent welded connections, and the dashed lines represent bolted connections. In addition to the designation of the components, the assigned assembly index is shown at the bottom right of each component field. This assembly index describes the position according to the building element in the stepby-step assembly sequence. Here it starts with the foundation F and the assembly index 1 up to BM2 with the assembly index 14. Since the specific assembly sequence of welded joints and bolts is irrelevant, these are not given an assembly index. Using (utilizing) the ruleset and the collision database, required neighboring building elements are identified. WBS can be generated by combining this information with the assembly sequence, e.g., for the BV1 component. As shown in Figure 4, BV1 is connected to the components CM1 and BM1. The presented methodology now
149
Figure 2. Example steel construction.
Figure 3. Dependencies of the building elements of the example steel construction.
requires identifying the rule-compliant path from the respective component (here BV1) to the start element (here F) and the components contained therein. For this purpose, components are introduced into the WBS along with a descending assembly index (in the direction of the arrow) and according to the assembly rules. Permissible connections are shown in black and impermissible connections in red. Irrelevant components
and connections not considered due to invalid connections are shown in grey. In the example, BV1 has an assembly index of 15, CM1 has an assembly index of 7, and BM1 has an assembly index of 13. Thus, the build indices of CM1 and BM1 are lower than those of BV1, and these components must be built before BV1. The ruleset confirms these dependencies. CM1 and BM1 are connected via CP1. The assembly index
150
Figure 4. Dependencies for the BV1 building element.
of CP1 is again lower than that of BM1. CM1 also has a connection with CP3, but its assembly index is 11 (11 > 7), therefore, this connection is invalid for the WBS. Since BM2 lost its connection to the path, it is no longer considered. This procedure is carried out until the foundation is reached. The relationships of the building elements to each other are documented as WBS, as shown in Figure 5. Starting
Figure 5. WBS for BV1.
from BV1, all assembly indices descend until the foundation with assembly index 1 is finally reached. For a better overview, duplicate building elements are combined into one building element. The generated WBS shown in Figure 6 can be divided into six levels, I to
Figure 6. Adjusted combined WBS for BV1 with WBS levels.
151
Figure 8. Steel construction with five PAG.
Figure 7. Steel construction with six PAG.
VI. All building elements on the same level are independent of each other and can therefore be assembled in parallel. 5 CASE STUDIES In this section, two case studies were carried out: a simplified steel construction and a bridge construction model. The steel construction is easier to understand, allows for a better understanding of the implementation, and is therefore explained in detail. 5.1 Steel construction Figure 7 shows a steel construction whose building elements have been placed in a WBS. For this purpose, all building elements were first assigned to the categories BM, BP, CM, F, and PF (Prefab Floor). Subsequently, four assembly rules (Eq. (8) to (11)) were defined and summarized in ruleset R2 (shown in Equation (12)) as a prioritized ruleset, which always prioritizes the assembly of the BPs. This ensures symmetrical assembly since the used strategy for the generation of the assembly sequence in this case study is only according to the ruleset R2. Equation (8) specifies that four BMs are required for the construction of each PF. In Figure 8, three BM are obscured by PFs, and the two PF included in the model rest on a total of five BM, i.e., they share one BM. Equation (9) defines the rule for building CMs, which require either one BP or one CM. BMs (Equation (10)) require either two CMs or two BMs. For the construction of BP (Equation (11)), only the presence of the foundation is required. PFr = 4BM
(8)
CMr2 = BP ∨ CM
(9)
BMr2 = 2CM ∨ 2BM
(10)
BPr2 = F
(11)
R2 = BPr2 , PFr , CMr2 , BMr2 , BPr2
(12)
In Figure 7, building elements of the same WBS level (i.e., the same parallelization level) are colored in the same color. The light green CM each forms a pre-assembly group (PAG) with a BP and are therefore each built in a single assembly step. The rule responsible for these PAGs is the rule of BP since it was defined as the connecting building element of the PAGs. This results in 20 assembly steps in five WBS levels for the entire structure. Figure 8 shows the same building information model whose assembly sequence was generated with the same ruleset and the same assembly strategy, but in contrast to Figure 7, one column does not form a PAG with the associated BP. Thus, it cannot be built on the same WBS level as the other columns but requires the assembly of the base plate right below it in the previous assembly step. This change also changes the parallelizability in the following assembly steps for building elements built on top of it, resulting in a WBS with an additional level, i.e., with six levels for 21 assembly steps. Due to the better parallelizability, the assembly sequence of the model that provides more PAG and thus has fewer WBS levels is therefore preferable in this example. 5.2 Bridge model For validation that the proposed methodology works with more complex models, the bridge model shown in Figure 10 is a model of a real construction in Hamburg, Germany, which was created as part of the IFC-Bridge project funded by the Federal Ministry for Digital and Transport. The ruleset for this model includes eleven rules, which are not discussed in detail here because their understanding would require a very detailed discussion of the building elements included in the model, and such would exceed the scope of this paper. For the bridge model, the By Storage Areas (StA) strategy (Weber & König 2022) was followed. This strategy gives preference to the assembly of building elements
152
6 CONCLUSION AND OUTLOOK
Figure 9. Partial view of the IFC-bridge model with coloring according to WBS on basis of the SET strategy.
Figure 10. Partial view of the IFC-bridge model with coloring according to WBS on basis of the StA strategy.
from storage areas, provided that they are in compliance with the rules. This strategy aims to utilize the storage areas only for a short period of time, so they can be used optimally when the storage area is limited. For this case study, the components were assigned to two storage areas. PAG is only used for the beams and their face plates in this model. The generated building instructions contain 465 assembly steps, which can be divided into twelve levels in the automatically generated WBS. For the validation, another WBS was generated for the same bridge model, which is shown in Figure 9. In contrast to the first WBS, shown by the coloring of the building elements according to the WBS in Figure 10, the assembly sequence is based on the Same element types (SET) strategy (Weber & König 2022). This strategy favors similar components for optimizing assembly processes since specialized assembly equipment can be used efficiently to reduce waiting times. The assembly sequence also consists of 465 assembly steps, which can be divided into only eleven steps in the automatically generated WBS due to the simplified parallelization possibility resulting from the less restrictive rules. Thus, the SET strategy is the preferable strategy in this case.
With the help of our approach, it is possible to generate WBS for arbitrary 3D building models based on the corresponding step-by-step assembly sequences and the building rules and collision databases necessary for their generation. By dividing the generated WBS into levels and comparing their number of levels with each other, objective decisions can be made about the choice of assembly strategies. The fewer levels the generated WBS has, the higher is the potential parallelization of the assembly steps described in it. The case studies have shown that a linking of building elements to PAG and different assembly strategies has a direct impact on the parallelizability of the assembly processes of the models presented, even though the deviation of the number of WBS levels for the different selected construction strategies was at most one level. This is probably because the models are primarily symmetrical and have a low degree of complexity. Nonetheless, the generated WBS gives a direct overview of the potentially parallelizable assembly processes over the entire assembly sequence. At this time, it is not yet possible to draw any general conclusions regarding the assembly strategies to be preferred on the basis of these two case studies alone. Further studies must determine whether there is an assembly strategy that can be preferred in principle over other assembly strategies. The challenge here is that not every strategy is suitable for every construction. An example is the Level by Level (LBL) (Weber & König 2022) strategy, which cannot be used or can only be used to a limited extent for constructions that have a gradient in relation to their foundation such as the bridge model presented in this paper. Furthermore, our approach assumes an idealized construction site with unlimited resources in terms of material flow, work equipment, and personnel. Detailed data must be available for the individual assembly processes to make a more realistic statement about the actual parallelizability of assembly processes. This must include, in addition to the delivery schedules, the quantity and qualities of the required workers, machines, and working space. In this way, existing capacities can be utilized, required capacities can be determined, and conflicts between different trades and individual workers due to overlapping work areas can be considered. REFERENCES
153
Bai, Yongjun; Zhao, Yong; Chen, Yang & Chen, Lu. (2009). Designing Domain Work Breakdown Structure (DWBS) Using Neural Networks. Advances in Neural Networks – ISNN 2009 (Lecture Notes in Computer Science), vol. 5553, 1146–1153. Berlin, Heidelberg: Springer Berlin Heidelberg. Berthaut, F.; Pellerin, R.; Perrier, N. & Hajji, A. (2014). Timecost Trade-offs in Resource-constraint Project Scheduling Problems With Overlapping Modes. International Journal of Project Organisation and Management.
Golpayegani, Seyed A. H. & Parvaresh, Fereshteh. (2011). The Logical Precedence Network Planning of Projects, Considering The Finish-to-Start (FS) Relations, Using Neural Networks. The International Journal of Advanced Manufacturing Technology 55(9-12). 1123–1133. Hashemi Golpayegani, S. A. & Emamizadeh, Bahram. (2007). Designing Work Breakdown Structures Using Modular Neural Networks. Decision Support Systems 44(1). 202–222. Kadet, Jordan & Frank, Bruce H. (1964). PERT for the engineer. IEEE Spectrum 1(11). 131–137. Li, Duanshun & Lu, Ming. (2017). Automated Generation of Work Breakdown Structure and Project Network Model for Earthworks Project Planning: A Flow NetworkBased Optimization Approach. Journal of Construction Engineering and Management 143(1). 04016086. Liu, Hexu; Al-Hussein, Mohamed & Lu, Ming. (2015). BIMbased Integrated Approach For Detailed Construction Scheduling Under Resource Constraints. Automation in Construction 53. Mikulakova, Eva; König, Markus; Tauscher, Eike & Beucke, Karl. (2010). Knowledge-based Schedule Generation and Evaluation. Advanced Engineering Informatics 24.
154
Norman, Eric S.; Brotherton, Shelly A. & Fried, Robert T. (2008). Work Breakdown Structures. Hoboken, NJ: John Wiley & Sons, Inc. Project Management Institute. (2013).A GuideToThe Project Management Body Of Knowledge: (PMBOK®guide); An American National Standard ANSI-PMI 99-001-2013, 5th edn. (Global standard). Newtown Square, Pa.: PMI. Sharon, Amira & Dori, Dov. (2012). A Model-Based Approach for Planning Work Breakdown Structures of Complex Systems Projects. IFAC Proceedings Volumes 45(6). 1083–1088. Siami-Irdemoosa, Elnaz; Dindarloo, Saeid R. & Sharifzadeh, Mostafa. (2015). Work breakdown structure (WBS) development for underground construction. Automation in Construction 58. 85–94. United States. (1962). DOD and NASA guide: PERT COST systems design. Washington, USA. Weber, Jan & König, Markus. (2022). Strategies for RuleBased Generated Assembly Sequences in Large-Scale Plant Construction. ASCE I3CE 2021. 655–662. Weber, Jan; Stolipin, Jana; Jessen, Ulrich; König, Markus & Wenzel, Sigrid. (2020). Rule-based Generation of Assembly Sequences for Simulation in Large-scale Plant Construction. ISARC 2020. 155–162.
Virtual design & construction
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Construction process time optimization of a reinforced concrete reaction slab – Implementing the VDC methodology M. Barcena, M.C. Borja & A.A. Del Savio Civil Engineering Department, Universidad de Lima, Lima, Peru
ABSTRACT: The Architecture, Engineering, and Construction (AEC) industry has shown low productivity levels, mainly due to the fragmentation between the agents involved. Given this, the industry is modernizing and implementing collaborative methodologies to improve the development of projects. One of them is the VDC (Virtual Design and Construction) methodology. VDC allows us to improve design, construction, operation, and maintenance management, changing paradigms within traditional processes. This research shows how implementing VDC can help to optimize the construction process time of a reinforced concrete reaction slab in a civil engineering laboratory. As a main result, an optimized process was developed, generating a 44-day reduction in the construction time of the reaction slab.
1 INTRODUCTION The construction sector’s contribution in all countries is crucial for economic and social development (Yagual-Velástegui et al. 2018). It is an engine of the economy, reacts immediately to the behavior of the country’s growth, is a great generator of employment, and has an important private and public investment (Palomino et al. 2017). Furthermore, through construction, the population’s needs related to the development of infrastructure projects and housing solutions are met, becoming a permanent source of work with the intensive use of labor and generating significant activity indirectly in other sectors of a country’s economy (Botero Botero & Álvarez Villa 2004). However, despite its importance, the construction industry is one of the sectors with the lowest degree of development, becoming an activity characterized by great deficiencies and lack of effectiveness (Botero Botero & Álvarez Villa 2004). Because of these deficiencies, there is a need to apply new technologies and construction project management methodologies to plan and control in advance the restrictions that may be generated within this industry. Franz and Messner (2019) highlighted the main benefits generated by, for instance, BIM. Through a series of surveys, it was determined that 41% of the interviewees considered that the main benefit is the reduction of errors or omissions, and 35% considered that their main benefit is the collaboration between the owner and the design companies. The remaining 24% considered the benefits of reducing rework. On the other hand, Del Savio et al. (2022) pointed out the importance of training professionals within the Architecture, Engineering, and Construction (AEC)
DOI 10.1201/9781003354222-20
field considering the changes brought by Industry 4.0, which includes managing new technologies, workflows, and methodologies, like VDC, to properly answer to the requirements of the industry. The Virtual Design and Construction (VDC) methodology is proposed to align the benefits with the client and project objectives. The VDC methodology has demonstrated the benefits of visualizing, integrating, and automating tasks, particularly for predicting project outcomes and managing the project towards desired performance (Hassan 2018). To illustrate the VDC application, it was adopted as a case study of constructing a reinforced concrete reaction slab in a civil engineering laboratory in Lima, Peru (Figure 1). The reaction slab is a reinforced concrete element used for full-scale structural testing of a building’s columns, beams, and slabs, among others. This slab has a massive amount of concrete and reinforcement steel. As a result, it sustains extremely high loads, which brings complexities in the construction related to construction restrictions and the high-quality standards required. The main goal of this research is to optimize the construction process time of a reaction slab, conducting the following research question: is it feasible to optimize the construction process of a reaction concrete slab through VDC framework implementation? To answer this, we present the methodology used to develop this work in the Material and Method. Then, in the Results, we represent the traditional and the optimized construction process. Next, in the Discussion, we compare the construction workflows. Finally, the Conclusions summarize the main results and recommendations for future research studies.
157
Figure 1. Civil engineering laboratory.
2 MATERIAL AND METHOD We started this study searching in Scopus and Web of Science basis for papers related to the process optimization with VDC, BIM, and Lean Construction, using the following Boolean equations: •
("OPTIMIZATION") AND ("PROCESS") AND ("VDC" OR "VIRTUAL DESIGN AND CONSTRUCTION") • ("OPTIMIZATION") AND ("PROCESS") AND ("BIM" OR "BUILDING INFORMATION MODELLING") • ("OPTIMIZATION") AND ("PROCESS") AND ("LEAN CONSTRUCTION")
As a result of this search, we found articles divided into 95% BIM, 5% Lean, and 0% VDC, which allowed us to build the research introduction. Next, the VDC framework and the table of interrelationships between its main components were developed. Rischmoller et al. (2018) mentioned that the main VDC components are the client and project objective, BIM (Building Information Modeling), ICE (Integrated Concurrent Engineering), PPM (Project Production Management), and metrics. Then, the traditional workflow, without the VDC, was mapped. Based on this, an optimized workflow was developed under the VDC management approach, including the components of the VDC framework. The production metrics proposed in the VDC framework were monitored throughout the project duration. As indicated by Belsvik (2019), metrics in VDC projects should be used to measure the project result and throughout the project for the continuous improvement of the project processes. Finally, the traditional workflow was compared with the one optimized with VDC to demonstrate the benefits of VDC application in the optimization of construction processes of a reinforced concrete reaction slab. 3 RESULTS The developed VDC framework is shown in Figure 2 with its three components: ICE, BIM, and PPM. The production metrics and controllable factors proposed for the VDC implementation of the construction of a concrete reaction slab are shown in Table 1.
Figure 2. VDC Framework.
158
These were selected, established, and implemented based on the client’s and the project’s needs. Production metrics are yardsticks that allow project teams to make timely course corrections to keep their projects on schedule (Majumdar et al. 2022a, 2022b). Table 1. Production metrics and controllable factors (ICE, BIM, and PPM). Objective
Metrics
Goal
PM_ICEl: Guess attendance
% Attendance) attendees/total guests=*100 % Issues=(topics treated/total topcis=*100 Session duration
> 80%
Application of the + & - % of simulated activities= (simulated/total)*100 Minimum LOD required
yes
%SCR=(constructive restrictions resolved/constructive restrictions encountered)*100 % Reduction time=(optimized flow time/traditional flow time)*100 # of revisions
>=90%
PM_ICE2: % of topics discussed on the agenda PM_ICE3: Session duration CF_ICEl: Include +&- in sessions PM_BIM1: % of simulated activities CF_BIM1: Generate a BIM model of the reaction slab with a LOD 400 PM_PPMl: Percentage of solve of constructive restrictions PM_PPM2: Reduce the construction time by 20%
Figure 3. Visualization of the work and 4D simulation of the placement of the steel.
> 80%
80%
400
>20%
Figure 5. Reaction slab 3D model in LOD 400.
CF_PPM1: Analyze >4 the feasibility of the days proposed for the construction of the slab of the constractor’s schedule PM = Production Metrics | CF = Controllable Factors
For the ICE component, the results considering the 11 ICE sessions were as follows: – An attendance percentage equal to or greater than 80% (goal) was obtained. – All the topics planned on the agenda were discussed. – All meetings lasted less than 100 minutes (goal). – Finally, the Plus & Delta (+ & -) was held at the sessions’ end. This helped measure the ICE sessions’ satisfaction and identify improvement opportunities. For the BIM component, the results were as follows: – It simulated the activities of placing the steel rebars, placing the anchor boxes, and pouring the concrete with the help of the NAVISWORKS program. These simulations enabled constructive restrictions to be detected and lifted (Figures 3 and 4). Kunz &
Fischer (2004) mentioned that models connect the components of the CAD plan with the activities of the design, acquisition, and construction plans. The resulting project model allows decision makers to view the planned construction of a building as an animated 3D model. – Finally, it was developed a model of the reaction slab with LOD 400 (Figure 5). For the PPM component, the results were as follows: – Five constraints were detected with the help of the construction simulations and solved as described in Table 2. In addition, the simulation helped to find restrictions in the slab construction and make decisions to solve them. Kunz & Fischer (2004) stated that models allow more decisions to be made than exist since it is now possible to get involved in the project from an earlier stage so that their knowledge of commercial and engineering can be incorporated into the design of structures, project planning, and organization and to improve coordination in all phases of the life cycle. – With the optimized process, the construction time of the reaction slab was reduced by 44 (51.76%) days. From 85 days to 41 days.
159
Table 2.
Item 1
2 3
4
5
Constructive restrictions and solutions.
Constructive constraint In the plans, the coating on the upper mesh is not considered. Anchor box suspension. Reaction slab concrete temperature control. Lack of an activity for leveling and aligning the anchor boxes in the schedule. 30 m steel bars that will have difficulties when entering the work.
Reaction slab component involved
Corrective action
Steel
Fixed in plans.
Anchor boxes
Anchor boxes were welded. Thermocouples were used.
Concrete
Anchor boxes
This activity was added to the schedule.
Steel
The size of the bar was reduced and then it was welded.
– Finally, the proposed flowcharts were reviewed and validated by the main stakeholders.
Figure 6. Table of interrelationships of VDC objectives.
Figure 6 shows the relationship between the objectives of each of the components. This demonstrates how each goal feeds into the other. Next, in Figure 7, the traditional workflow is shown, and in Figure 8, the optimized workflow. Each of these workflows was carried out and validated in the ICE sessions with the specialists in charge of the project. Figure 7 shows the traditional overall workflow of the reinforced concrete reaction slab construction process. The main stakeholders’ participation in the slab construction process activities is observed. Also,
the processes used for this research are painted in light blue. In addition, in Figure 7, the main activities that can generate rework, such as lost man-hours or unproductiveness and material losses, in the slab construction are appreciated in gray. Finally, a total work execution time of 85 days was estimated. The traditional workflow begins with the need for a laboratory required by the civil engineering department. Within this laboratory, structural tests are planned; a reinforced concrete reaction slab is needed. Next, the bidding process for the work begins. After this, the plans are drawn up, and the contractor reviews them. If the plans are fine, they are approved. Otherwise, they are drawn up again. After this, the work begins with the excavation and, immediately after, the flooring for the reaction slab. Parallel to these activities, the supplier selection process for the acquisition of the slab components (steel, anchor boxes, and concrete) begins. Then, 115 tons of steel are assembled and placed on site. This activity is an estimated duration of 63 days. Then follow the placement, alignment, and leveling of 105 anchor boxes. For this activity, it is an estimated duration of 12 days. After this, the pouring and curing of 220 m3 of concrete follow. For this activity, it is an estimated duration of 1 day. Finally, a work review request is sent to the client to revise and approve. In case of rejection, the contractor performs a survey of observations. Next, the traditional reinforcement workflow and the placement of the steel in place are described. This begins with the contractor’s steel supplier selection process. Then, the contractor makes the steel requirement to the previously selected supplier, and the supplier receives the steel requirement and sends it. Subsequently, the contractor receives the steel and performs quality control. If the product meets the requirement, storage begins. If the product does not meet the requirement, the contractor returns the steel, and the supplier sends a new product. After the steel is stored, the contractor begins setting up (including cutting and reinforcing steel on site) and placing the steel on-site in the following order: bottom mesh steel, wall uprights, middle mesh, and top mesh. A duration of 58 days is estimated for this activity, and it was calculated based on the performance provided by the contractor. After the placement of the upper mesh, the anchor boxes shown in Figure 8 began to be placed with an estimated duration of 12 days. Finally, the hooks and additional hooks began to be placed for an estimated 7 days. Then the traditional workflow of anchor boxes is described. This workflow begins with selecting the supplier of the anchor boxes by the contractor. After selecting the supplier, the contractor submits the request for the anchor boxes, and the supplier produces a prototype of the anchor boxes as a sample for the client and the contractor. Subsequently, the supplier sends the prototype of the anchor boxes, and the supplier and the client receive them. If the prototype is compliant, the supplier makes all the anchor boxes; if it does not comply, the provider makes another
160
Figure 7. The traditional macro flow of the construction process of the reinforced concrete reaction slab.
Figure 8. Anchor boxes.
prototype, and the cycle is repeated. After preparing the anchor boxes, the supplier ships them, and the contractor receives them and performs quality control. If it complies, the contractor receives the anchor boxes, and the cycle is repeated if it does not comply. After this, the anchor boxes are placed, aligned, and leveled. This activity has an estimated duration of 12 days. Finally, the pouring and curing of the concrete continue. Finally, the traditional workflow of concrete pouring is described. This workflow begins with the contractor’s selection process of the concrete supplier. After selecting the concrete provider, the concrete request is sent to the provider, and the provider receives the request. Subsequently, the supplier sends the concrete, and the contractor receives it and performs quality control. Finally, the contractor begins to pour the concrete and cure it if the material is up to standard. If it does not comply, the supplier must resend the concrete, and the cycle repeats until the material complies. Figure 9 shows the optimized overall workflow of the reinforced concrete reaction slab construction process, which superseded the traditional workflow presented in Figure 7. This workflow was validated by the construction team and the laboratory users. The
main stakeholders’ participation in the slab construction process activities is observed. Also, the processes used for this research are painted in light blue and have a separate flowchart with their respective activities. Finally, the total work execution time was 41 days. It all started with the need for a laboratory, which the civil engineering department required. Within this laboratory, structural tests are planned; a reinforced concrete reaction slab is needed. Then, the BIM of the laboratory began with a LOD 300 (level of detail), using Revit and BIM 360 tools to manage and edit the model collaboratively. Immediately after, the bidding process for the work began. After this, the plans were drawn up and reviewed in an ICE session. If the plans were fine, they were approved. Otherwise, they were drawn up again. After this, the BIM of the laboratory advanced to a LOD 400, including the steel reinforcement. The modeling of the Laboratory in LOD 400 lasted 7 days. According to BIM Forum (2021), in LOD 400, the model element is graphically represented within the Model as a specific system, object, or assembly in terms of size, shape, location, quantity, and orientation with detailing, fabrication, assembly, and installation information. Finally, the construction process (4D) simulation continues if the modeling is correct. Otherwise, the model is revised. Parallelly, the process of selecting suppliers for steel, anchor boxes, and concrete began. Then, the work began with the excavation and the flooring for the reaction slab immediately after. Later, assembling and placing 115 tons of steel reinforcement began with an estimated duration of 12 days. Then follow the placement, alignment, and leveling of 105 anchor boxes within approximately 12 days. After this, the pouring and curing of 220 m3 of concrete followed in one day. Also, parallel to the laboratory construction, the simulation and monitoring of said construction was
Figure 9. Overall optimized workflow of the reinforced concrete reaction slab construction process.
161
Figure 10. Steel reinforcement BIM.
developed and validated in the ICE sessions. Finally, a request for review of the work is sent, and the client will decide if he is satisfied or not. If not satisfied, the contractor performs a survey of observations. The optimized workflow of the steel component is described. This flow began with modeling the laboratory’s reinforcement steel, which lasted 7 days. Then, with the model already designed, as shown in Figure 10, the modeler sent the necessary information, as shown in Figure 11. The information sent was the amount of steel, the model, and some sections of the same, to the contractor to begin selecting reinforcing steel. After selecting the steel supplier, he prepared a package proposal for the arrival and placement of the enabled steel. Next, an ICE session was conducted to decide if the package proposal was optimal. With this information, 2 proposals for construction processes were prepared, later shown in an ICE session with all those involved. The main difference between these 2 proposals was the construction sequence of the slab. First, the placement of the reinforcing steel of the adjoining wall with the slab had not been considered. On the other hand, the reinforcing steel of the elements adjacent to the slab was considered in the second proposal. In the first one, the anchor boxes were placed. It was decided which construction process was the most efficient. The contractor then prepared a three-week Lookahead schedule for the arrival of shipped materials with a request for the steel received by the supplier and proceeded to ship the steel. According to the previously sent Lookahead (Figure 12), this steel was received by the contractor and passed through quality control. If the steel received is compliant, it is stored. If not, it is returned to the supplier to send another steel that will pass the quality filter again. After storing the enabled steel, the enabled steel of the lower mesh began to be placed; this lasted 2 days. Then, cut and bent steel was placed on the verticals of the wall in 5 days, and the enabled steel of the intermediate mesh began to be placed, which lasted 1 day. Then the qualified steel of the upper mesh was placed in 2 days, and the anchor boxes were placed. Finally, the hooks and additional hooks were set in two days. The entire placement of the steel lasted 12 days. Next, the workflow of placing the anchor boxes is described. First, the design of the anchor boxes (Figure 13) was done and implemented in BIM. Next, the quantity, dimensions, and drawings of the anchors’
Figure 11. Steel information.
Figure 12. Lookahead.
Figure 13. Anchor box design. Dimensions in millimeters.
boxes were sent to the contractor. After this, the supplier selection process began for the anchor boxes. The supplier then produced and shipped a prototype anchor box for the reaction slab. Then this prototype was reviewed by the client. If it weren’t right, it would have to do it again. Then, the anchor boxes were included in the construction process simulations. Simulations were carried out to choose the most efficient construction process. Subsequently, the supplier began the process of making the 105 anchor boxes.
162
The Lookahead of the arrival of the anchor boxes at the site was then developed, as shown in Figure 12. These were shipped and placed according to Lookahead. After placing all the boxes, the alignment and leveling of the anchor boxes began. If they are not aligned correctly, they can negatively impact the performance of structural tests. These activities lasted an estimated 12 days. Finally, the process of pouring and curing the concrete continues. Finally, the optimized workflow of concrete placement is described. This began with elaborating a BIM considering the concrete placement. After this, the modeler sent the specific information (quantity, plans, model) to the contractor to start selecting a supplier. After selecting a supplier, the supplier developed a concrete casting proposal and technical specifications. Finally, this proposal was discussed in an ICE session, where the following agreements were reached: – Cement with lower heat of hydration was used. – Self-compacting concrete was used. – Ice was placed in the mixer to avoid increased temperature during its transfer. – Temperature control was carried out with thermocouples. – Final curing was done when the concrete slab had already hardened for 2 days. Then, simulations of the concrete mix placement process were elaborated and discussed in an ICE session, and the most efficient process was chosen. Subsequently, the supplier proceeded to send the concrete and passed the contractor’s quality control. Then, if the concrete conditions were met, it was poured. Otherwise, the mixture was returned, and the supplier had to send a new one. Finally, after casting, curing was carried out. These two activities lasted 1 day.
4 DISCUSSION The main difference between the workflows presented in Figures 7 and 9 is the duration of the construction process. With the traditional process, the construction process of the slab would last 85 days. On the other hand, with the optimized process, the duration was 41 days. Therefore, there is a reduction of 44 days in the construction process of the slab. In addition, we can highlight the following differences: – In the traditional workflow (Figure 7), the same actors do not always participate in the processes. This varies according to the process being touched. However, in the optimized process (Figure 9), the same actors always participate in all the construction processes of the reaction slab, either directly (in the field) or indirectly (making decisions through ICE sessions). – In the optimized workflow (Figure 9), the ICE sessions are used for decision-making and discussing how each process would be executed with all stakeholders.
– The traditional workflows (Figure 7) show the potential rework generated by the lack of coordination and collaboration between the stakeholders. However, these reworks steps disappeared in the optimized workflow (Figure 9). – In the traditional workflow (Figure 7), errors or incompatibilities are not detected in advance. On the other hand, in the optimized workflow (Figure 9), with the help of BIM construction simulations, these errors and incompatibilities could be detected and solved in the ICE sessions before construction started. Therefore, according to Fischer and Kunz (2012), it is recommended to use 4D animations to optimize the construction plan or schedule, to engage all stakeholders to look for and understand constraints on the construction process due to spacetime interferences (when one construction activity will interfere with another), find interferences of the construction with ongoing facility operations and user activities, and find interferences between work of different subcontractors.
5 CONCLUSIONS The VDC methodology implementation was crucial for optimizing the construction time of the reinforced concrete reaction slab. This was due to: – Workflow mapping, where it was possible to determine that with the optimized workflow (Figure 9), there was a 44 days (51.76%) reduction in the slab construction process. – It improved collaboration and communication between the stakeholders. As a result, 40% more people were involved with the optimized workflows (Figure 9) than with the traditional workflow (Figure 7). Furthermore, with the involvement of the stakeholders, it was possible to eliminate the constructive restrictions through the ICE sessions before the construction started. – Due to BIM LOD 400 and construction simulations (BIM 4D), errors or incompatibilities are identified and resolved early. – BIM 4D helped identify the most optimal construction process throughout ICE sessions. This was followed through the metric “100% of simulated activities.” – Feedback and continuous improvement through each of the components of the VDC. Each goal has its metrics aligned, which feed each other, as shown in the interrelation graph (Figure 6). Each of these elements contributed to achieving the objective of the client and the project. In addition, the following general conclusions were also reached: – 4D simulations play a very important role as they help visualize and precisely understand the entire construction process on-site. As a result, possible problems (Table 2) that could have occurred on the
163
site were identified, and a preventive solution was proposed before construction began. – The implemented VDC framework helped the contractor understand the complexity of building the reinforced concrete reaction slab and plan his activities properly and in advance to avoid construction delays. Finally, it is worth mentioning that this work was carried out during the Peruvian state of emergency due to COVID-19. Cameras were installed on the construction site to monitor the construction daily and remotely. In addition, all the ICE sessions were carried out virtually.
REFERENCES Bedrick, J., Ikerd, W., & Reinhardt, J. 2021. Level of Development (LOD) Specification For Building Information Models - Part I, Guide, & Commentary. BIM Forum. Belsvik, M. R., Lædre, O., & Hjelseth, E. 2019. Metrics in VDC Projects. 27th Annual Conference of the International Group for Lean Construction, IGLC 2019, 1129–1140. doi:10.24928/2019/0167. Botero Botero, L. F., & Álvarez Villa, M. E. 2004. Guía de mejoramiento continuo para la productividad en la construcción de proyectos de vivienda (Lean construction como estrategia de mejoramiento). Revista universidad EAFIT, 40(136), 50–64. Del Savio, A. A., Galantini, K., Díaz-Garay, B., & Valcárcel, E. 2022. A Methodology for Embedding Building Information Modelling (BIM) in an Undergraduate Civil Engineering Program. Heliyon (in press). Fischer, M., & Kunz, J. 2004, February.The Scope and Role of Information Technology in Construction. In ProceedingsJapan Society of Civil Engineers (pp. 1–32). DOTOKU GAKKAI.
Kunz, J., & Fischer, M. 2012. Virtual Design and Construction: Themes, Case Studies and Implementation Suggestions. Center for Integrated Facility Engineering, Stanford University, 1–2. Franz, B., & Messner, J. 2019. Evaluating The Impact of Building Information Modeling On Project Performance. Journal of Computing in Civil Engineering, 33(3), 04019015. Hassan, H., Taib, N., & Rahman, Z. A. 2018. Virtual Design and Construction: A New Communication in Construction Industry. Paper presented at the ACM International Conference Proceeding Series, 110–113. Majumdar, T., Rasmussen, S. G., Del Savio, A. A., Johannesdottir, K., Hjelseth, E., & Fischer, M. 2022a. A Comparative Analysis of Production Metrics across VDC Implementations. Construction Research Congress 2022 (pp. 1024–1033). doi: 10.1061/9780784483 978.104 Majumdar, T., Rasmussen, S.G., Del Savio, A.A., Johannesdottír, K., Hjelseth, E., & Fischer, M.A. 2022b. VDC in Practice: A Preliminary Categorization of Production Metrics reported in Scandinavia and Latin America. Proceedings of the 30th Annual Conference of the International Group for Lean Construction (IGLC30), 1177–1185. doi: 10.24928/2022/0230 Rischmoller, L., Reed, D., Khanzode, A., & Fischer, M. 2018, July. Integration Enabled By Virtual Design and Construction as A Lean Implementation Strategy. In Proc. 26th Annual Conference of the International. Group for Lean Construction (IGLC), Chennai, India. http://iglc. net/Papers/Details/1547. Silva, J. P., Otoya, J. H., & Alvarado, V. R. E. 2017. Análisis macroeconómico del sector construcción en el Perú. Quipukamayoc, 25(47), 95–101. Yagual-Velástegui, A. M., Lopez-Franco, M. L., SánchezLeón, L., & Narváez-Cumbicos, J. G. 2018. La contribución del sector de la construcción sobre el producto interno bruto PIB en Ecuador. Revista LASALLISTA de investigacion, 15(2), 286–299.
164
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
VDC framework proposal for curtain wall construction process optimization V. Bustamante, J.P. Cedrón & A.A. Del Savio Civil Engineering Department, Universidad de Lima, Lima, Perú
ABSTRACT: A building is made up of structural and non-structural elements. Among these are the curtain wall systems. The curtain walls are glazed elements that can cover the building façade entirely and provide thermal and enclosure properties. Despite its use worldwide, some challenges have been identified in the construction process related to transportation, communication between the stakeholders, and the installation itself. A VDC framework is proposed to overcome these challenges, including an implementation workflow and an interrelationships map between the VDC components. This study specialized on a multiuse 12,000 m2 building under construction in Lima, Peru, resulting in an optimized proposal for curtain wall construction, compared to the traditional construction flow.
1 INTRODUCTION 1.1 Problem statement The most common options for covering the exterior surface of a building are window and curtain wall systems. The latter is more popular due to its aesthetic and constructive homogeneity (Marquis et al. 2017). The construction of curtain wall systems involves installing various parts, such as the glass panel and the vertical and horizontal profiles. Skilled workers install the profiles, anchor the glass panel, and apply silicone at the edges. However, problems have been identified during the installation process, such as the generation of waste material, like unused parts of the curtain wall, poor design of parts and materials for assembly, and incorrect calculations and counts of the curtain wall elements, among others (Chen & Lu 2018; Salimzadeh et al. 2020). According to Kim et al. (2020), the curtain wall construction stage is in the project’s critical path and represents approximately 10-15% of the total cost of the work. Therefore, the possibility of a deviation in this process would directly affect the project deadlines (Muñoz & Pardavé 2018). Hence, the need to analyze productivity, a parameter linked to scheduling, leads to results that facilitate decisionmaking and the ongoing project (Alaghbari et al. 2017; Muñoz & Pardavé 2018). It is possible to analyze this parameter using existing tools and methodologies that facilitate such a process. The VDC methodology offers an analysis through its methodological framework, prioritizing the client’s objectives, project objectives, the importance of ICE sessions, BIM, and Project Production Management (PPM) (Aslam et al. 2021). DOI 10.1201/9781003354222-21
1.2 State-of-art Among the various types of systems of curtain walls used, the stick system comprises lacquered aluminum profiles, horizontal and vertical, and tempered glass panels. The installation starts with the aluminum profiles, and the glass panels are installed from the bottom up (Muñoz & Pardavé 2018). On the other hand, installing curtain wall systems is crucial in controlling and monitoring cost, time, productivity, and waste. Its construction process consists of three main groups of activities, among which are the material unloading process, the panel elevation process, and the installation process (Han et al. 2017). A conventional curtain wall installation process can be visualized in Figure 1, based on the information provided by the contractor in the study case. In 2018, Muñoz & Pardavé demonstrated the contribution of BIM, through a methodological management proposal, in optimizing productivity in “Stick” curtain wall execution processes by reducing the operating cost by 15.04%, reducing the project execution time by 22.1%, and optimizing productivity by 40%. According to Chen & Lu (2018), assembling a multidisciplinary team for curtain wall design and construction phases can also significantly contribute to project quality, cost reduction, time reduction, and waste material reduction. The VDC methodology has been used for its great benefits and contributions to collaborative project management (Lledó & Pérez 2020). Internationally, it is known to have a great synergy with other current methodologies, such as Lean Construction and Lean Project Delivery System (LPDS), in terms of processes, organization, and product (Aslam et al., 2021).
165
Figure 1. Curtain wall standard construction process.
Another valued feature of the methodology is its ability to coordinate and build a reliable 3D model before the construction stage supported by BIM. Coordination is vital for resolving incompatibilities and last-minute design changes in the shortest possible time. Corrales & Saravia (2020) have observed that VDC increases its effectiveness when applied in the early stages of the project and can reduce the variability of a project’s schedule by up to 21%. Productivity is affected by the curtain wall installation process and can be optimized using various tools, such as BIM. However, no information has been found that relates the VDC methodology to the productivity analysis in installing the mentioned curtain wall system. 1.3 Research question Finally, the question arises: how could the productivity of the installation processes of stick curtain wall systems be optimized from a methodological proposal based on the Virtual Design and Construction (VDC) framework in constructing a multipurpose building? The answer to this question is discussed in this research. Therefore, the main objective will be to generate a VDC methodological proposal to optimize the productivity of the execution processes of the stick curtain wall system in the construction of a multiuse building.
2) Literature review, 3) Proposal design: VDC framework and implementation workflow proposal, 4) Validation/Results: Qualitative interviews with key actors, 5) Interrelationships map between the VDC components, 6) Optimized curtain wall construction process. This research has a qualitative approach, and the data collection was based on observational methods. The methods used for data collection include site vis its and interviews with curtain wall systems specialists. The literature review allowed us to identify the VDC implementation gap in curtain wall system processes. Therefore, a VDC framework and implementation workflow is proposed based on VDC concepts and curtain wall installation processes. They were validated through structured qualitative interviews with the client and contractor of the project under study. The feedback about the production objectives and controllable factors sought to validate the methodological proposal.
2 MATERIALS AND METHODS 2.1 Research method The research methodology of this paper can be divided into the following steps: 1) Data collection: Case and curtain wall systems description,
Figure 2.
166
Project’s render. Courtesy: Universidad de Lima.
Figure 3. Proposed workflow for the implementation of the VDC framework.
the project’s objectives. The client’s objectives seek to ensure that the building is sustainable, constructible, functional, and operable. The project objectives seek to reduce time, reduce cost, and control the quality of the work (Kunz & Fischer 2020). Second, regarding the components of the VDC framework, production metrics were defined based on Majumdar et al. (2022a , 2022b). the ICE component is designed for problem-solving through the interaction between the agents involved in the project. In that sense, the production metrics should measure the degree to which stakeholders have timely and significant participation in task review and approval through the number of problems solved in these sessions and control the number of attendees about the number of people related to the project. The BIM component includes the development of a collaborative 3D model of the project in which the number of clashes regarding curtain walls is counted after reaching each maturity level. These levels are associated with a model maturity index (MMI), similarly to the level of development (LOD) concept. Also, the correct labeling of the mentioned elements is measured to ensure future management. Finally, to define the PPM component, the production metrics must measure the weekly productivity through the Percent Plan Complete (PPC), which means planned tasks divided by completed tasks in the installation of the stick curtain walls and continuous monitoring of the actual cost and its variability.
Figure 4. Proposed VDC map of interrelationships.
2.2 Collected data The study case comprises a built area of 12,000 m2, 2 basements and 5 upper levels for classrooms and laboratories, and stick curtain walls indoors and outdoors. The project is in Lima, Peru. The execution period is 11.5 months, including design and construction. Figure 2 shows an image of the project.
2.3 Proposal design The following steps were followed to develop the proposed methodology. First, it is defined the client’s and
3 RESULTS 3.1 Proposal development The client’s main objectives are obtaining the LEED Gold certification and completing the project by the second half of 2022, having the building’s facilities ready to use. In essence, to have a sustainable and functional building. On the other hand, the project’s
167
objectives are achieving a constructible project, an execution time reduction from 11.50 to 10.75 months, saving 15% of the costs at the design stage of the curtain wall, and minimizing 100’% of observations through continuous quality control. Then, within the components of the VDC framework, the ICE component included measuring the degree to which stakeholders have timely and significant participation in task review and approval through monitoring curtain wall issues resolution and assisting key personnel. The BIM component included counting the number of clashes, regarding curtain walls, after reaching each maturity level and measuring the correct labeling of these elements. Finally, the PPM component included tracking the productivity using PPC in the weekly installation of the curtain walls and the variation in the project cost. The proposed methodology ends with generating a VDC framework and a map of interrelationships. Figure 3 shows the implementation workflow proposal from the interviews conducted with the client’s and contractor’s representatives and the information gathered. The framework’s objective is to track productivity, production metrics, and controllable factors to observe their variation over time and perform the respective analysis in the future. Finally, the VDC interrelationships map, shown in Figure 4, helps us to identify the interdependencies between the components of the VDC framework (Sujan & Hjelseth 2021). For example, there is a relationship between the controllable factor of ICE and the BIM and PPM metrics. This is because key professionals can influence the reduced clashes detected and proper productivity management in curtain wall installation. On the other hand, there is a relationship between the ICE metric and the BIM objective because the effective participation of the stakeholders will influence the reduction of the number of clashes detected. Also, there is a relationship between the BIM metric and the ICE objective; this is because the number of clashes detected can be discussed later during ICE sessions, with the active participation of the stakeholders. Finally, there is a relationship between the PPM metric and the ICE objective since the problems encountered in productivity should be discussed during the ICE sessions.
Figure 5. ICE meeting room of the study case.
Figure 6. BIM 3D model of the project.
3.2 Proposal validation To validate the methodological proposal, we interviewed the key project stakeholders to collect their feedback, especially those involved in installing the curtain walls, to validate the proposed metrics and controllable factors. Regarding the interviews, we contacted the specialists representing the client and the contractor. Diagnosis phase: – The interviewees reassured the importance of including the designer in the phase, as he is a crucial actor with the client since he is the one who conceives the idea of the project according to the client’s indications. In addition, they assured the importance of having a common, updated, and reliable data environment accessible to all involved. Objectives formulation phase: – The interviewees indicated that the designer and the client should develop the client’s objectives. But on the other hand, they indicated that the client should be the one to validate the client and project objectives component formulation phase. – ICE component. The interviewees indicated that ICE meetings are an important part of the project. The importance of time management and the accordion model to manage meetings were highlighted. This model involves using sub-groups from the general meeting group to attend specific topics and ensure effective participation (Reed et al. 2017). Figure 5 shows a part of the ICE environment used in the study case. – BIM component. The interviewees highlighted the importance of using the BIM model for clash detection. In addition, they indicated the need to label the curtain walls and have the aluminum profiles and the glass panel modeled at least. Particularly for the project, it was helpful to identify the pre-stressed and post-stressed slabs to reduce the structural errors generated by anchoring the curtain walls to these elements. Figure 6 shows the 3D BIM model used in the project. – PPM component. The interviewees mentioned that to optimize productivity in the installation, it’s essential to have proper management of the transportation of pieces and wall sizing. Finally, they highlighted the importance of running activities
168
Table 1. VDC framework proposed. VDC framework Client objectives
Obtain LEED Gold certification and the project built, functional and operable so that the end users involved can make use of the facilities from the second half of 2022.
Project objectives
Reduce the execution time to 10.75 months, save costs at the design stage and minimize observations through continuous quality control.
Components
Item
Description
Metrics
Goal
ICE
Production
Degree to which intended stakeholders have
# of problems solved × 100 # of identified problems
100%
metrics
timely and significant participation in task review and approval.
Controllable
Assemble the necessary personnel for the
# of attendees × 100 # of people related to the project
90%
factors Production metrics
resolution of problems on site. Number of clashes in BIM after reaching each maturity level.
Controllable
Identify curtain wall modules with a
factors
special label inside the model for tracking purposes.
Production
Percent plan complete (PPC)
PPC =
Controllable
Weekly monitoring of the cost of curtain
Actual cost ($) × 100 Planned value ($)
factors
wall installation.
BIM
PPM
# of clashes in BIM after reaching each maturity level # of curtain walls labeled × 100 # of curtain walls modeled
Planned tasks × 100 Completed tasks
100%
100%
metrics
parallel to the concrete construction process, such as design, storage management, topography, and curtain wall installation. On the other hand, they commented that rescheduling activities are key to controlling productivity, and the planning of transportation and storage should be done from the beginning by the contractor and the supplier. This includes establishing the number of curtain wall elements carried per truck. Control and evaluation phase: – The interviewees mentioned that it would be ideal to have definition checks of the VDC framework at the beginning and end of the workflow to ensure its validity. In addition, they mentioned the importance of having an initial document indicating the responsibilities and commitments of the parties involved in developing the VDC framework. This ensures the continuous elaboration of the proposed VDC framework. Table 1 shows the VDC framework proposed for the multiuse building, validated with the comments and information received.
4 RESULTS DISCUSSION The interviews showed that the approach of a common data environment nurtured by the parties involved is a great contribution to the proposal, thus promoting the participation of those involved and ensuring the quality of the information. Charaja (2018) mentioned that collaboration in actions and information is crucial for the optimal development of the project.
= 3, can be obtained in two steps: (a) by applying CBIPk function on n-tuplets (S1 , C1 , C2 , ...Cn−1 ) and (C2 , ...Cn , S2 ) (nth order) and (S1 , C1 , C2 , ...Cn−2 ) and (C3 , ...Cn , Cn+1 , S2 ) (nth + 1 order) and (b) by applying the IP function on the returned pair of surfaces by the CBIP2 function.
4.4 External high-order 2LSB calculation The examples presented so far are related to building space pairs and the related obtained internal high-order 2LSB surface pairs. External high-order 2LSB surface pairs are calculated in the same manner as the internal the high-order 2LSB surface pairs, with the only difference being the replacement of one of the building space boundary representations (S1 or S2 ) by the boundary representation of the environment surfaces contained in the surface set E. These surfaces are external building areas exposed to either the outside air or the building ground. The surfaces of set E can be obtained by removing from the surfaces of the boundary representations of all building constructions (set C = ∪i Ci ), the surfaces or surface parts that are common among these representations, and the space surfaces in set S = ∪i Si (common boundaries). If no clashes occur and the boundary representations are water-tight, the common boundaries are always shared between two boundary representations. Based on the previous definitions the calculation of the surface set E, can be expressed in math terms as: E = C {CB (C ∪ S)}
Figure 5. Illustration of third order 2LSB surface extraction: (1) CBIP process on space and 2 construction B-reps, (2) IP process on obtained 2LSB surfaces.
(4)
The introduced operation “ ” and the CB function used in (4.4) use the polygon intersection (∩p ) and subtraction (\p ) which act on two coplanar polygons a, b using (Vatti 1992) and return the common part and their difference, respectively. If polygons a and b are not coplanar then a\p b = ∅ and a ∩p b = ∅. More specifically, CB in equation (4.4), is defined as a common boundary function, which when applied on a polygon surface set (obtained in the general
299
case, from the boundary surfaces of multiple watertight boundary representations with no clashes among them), returns polygons or polygon parts common shared by pairs of input polygons: ⎧ ⎫ ⎪ ⎪ ⎪ ⎪ ⎨ ⎬ CB(A) = ai ∩p a j (5) ⎪ ⎪ ⎪ ⎪ ⎩∀ai ,aj ∈A ⎭ i =j
The subtraction operation “ ” applied on two surface sets A and B (A B), returns a surface set in which contains polygonal surfaces obtained by coplanar pair-wise polygon subtraction operations, ai \p bj , ∀ai ∈ A and ∀bj ∈ B: ⎧ ⎫ ⎨ ⎬ A B= a i \p b j (6) ⎩ ⎭ ∀(ai ,bj )∈A×B
The first examined area, displayed in the magnified part A of the section in Figure 7, focuses on the ceiling of the ground floor and external walls of the building. The ceiling contains two building elements stacked one on top of the other which are, from bottom to top: (a) 22cm thick compound ceiling exported as IfcCovering which has three layers: (a1) a 3cm plasterboard, (a2) 5cm rigid insulation and (a3) 14cm air gap and (b) an 18cm thick ceiling slab which has a single layer made of reinforced concrete which is exported as IfcSlab. As it is also indicated in the magnified part A of the section Figure 7, the external walls of the building contain two building elements stacked one after the other which are exported as IfcWalls. These elements are from inside out: (a) a 23cm thick double brick wall which has three layers: (a1) 9cm bricks, (a2) 5cm air gap, and (a3) 9cm bricks and (b) an external facade wall which has a single layer of limestone of 13cm thickness.
In summary and by looking at equation which is based on the polygon subtraction operations “\p ”, the obtained surfaces in the set E, will contain only the surface parts of the construction boundary representation elements in C, which do not have common parts with other boundary surfaces of constructions or space volumes. 4.5 External shading surfaces
Figure 6. Demonstration building.
Using the rationale of the introduced algorithmic extension, the external shading surfaces of buildings formed by stacked multi-element, multi-layered constructions can be calculated. This can be accomplished by replacing the two boundary representations of the inner space shells with the building’s outer shell attached to the outside air. The building’s surfaces of the outer shell attached to the air can be obtained by removing from the surfaces of the outer building shell contained in surface set E, the surfaces or surface parts attached to the ground. 5 APPLICATION EXAMPLE The proposed algorithmic extension is validated on a real BIM model related to a residential building on the island of Crete in Greece in the Akrotiri region at Korakies village built by (ERGOPRISM 2022). The demonstration building has two floors one at basement level and one at ground level, as pictured in Figure 6. The BIM model of this building was detailed enough to contain multiple wall and slab elements stacked one on top of the other reflecting the different phases of the building construction. The BIM model was developed in Autodesk Revit and exported in IFC4 format. Three building areas were examined highlighted in the section Figure 7 in the magnified parts A, B, and C respectively. In these areas different high order 2LSB surface pairs were obtained by the introduced algorithmic extension, as described next.
Figure 7. Building section highlighting the building’s multi-element constructions.
300
The algorithmic extension generates the secondorder 2LSB surfaces in the ceiling surfaces of the ground floor of the building as displayed in part 1 of Figure 8. These 2LSB surfaces have related elements: one element exported in the BIM file as an IfcCovering element (Compound ceiling) and one element exported in the BIM file as an IfcSlab element (Ceiling slab). Additionally, the external second-order 2LSB surfaces related to two IfcWall elements of the BIM file (Brick wall, External facade wall), are also identified as displayed in part 3 of Figure 8.
elements, which are from ground level up: (a) a 10cm thick building pad made of concrete, (b) a 40cm thick foundation slab made of reinforced concrete and (c) 10cm thick cement mortar. The respective external third order 2LSB surfaces are identified correctly by the introduced algorithmic extension as displayed in part 4 of Figure 8. Second-order 2LSB surfaces related to double-wall openings are identified by the introduced algorithmic extension as displayed in part 1 of Figure 8(external) and in part 2 of Figure 8 (internal). The double-wall openings in these cases are formed by single opening volumes attached at their largest surface area. Finally, external shading surfaces related to double slab and double wall elements are identified as pictured in parts 1 and 3 of Figure 8, respectively. 6 CONCLUSIONS
Figure 8. Identified high-order 2LSB surface pairs: (1) Front view, (2) Front view with surface culling, (3) Back view, (4) Back view with surface culling.
The second examined area, displayed in the magnified part B in the section Figure 7, refers to the slab construction between the basement and the ground floor space volumes. This construction consists of two single-layer elements exported in the BIM file as IfcSlabs, which from bottom to top, are: (a) an 18cm thick floor slab made of reinforced concrete and (b) 10cm thick ceramic tiles of the ground floor. The respective internal second-order 2LSB surface pairs are identified correctly by the introduced algorithmic extension, as displayed in part 4 of Figure 8. The third studied area displayed in the magnified part C in the section Figure 7, is related to the construction of the basement slab of the building. This construction is a multi-element construction consisting of three slabs, exported in the BIM file as IfcSlab
The automatic generation of building energy performance simulation models from open BIMs (IFC) involves a considerable challenge, related to building geometry, which is the correct generation of the building’s second-level space boundary (2LSB) surface topology. The generation of this topology becomes even more cumbersome when multi-element building constructions are present in the BIM. Such cases are encountered when modeling building retrofitting projects (facade insulation) or building construction projects built in multiple phases (different elements in different phases). In these cases extending the concept of the 2LSB to higher orders is required to be able to identify the elements of the 2LSB surfaces correctly. When single-element building constructions are present in the BIM model, traditional algorithms such as the Common Boundary Intersection-Projection (CBIP) algorithm, can be used to evaluate the respective first-order 2LSB surfaces. For the calculation of higher-order (≥ 2), 2LSB surfaces an extension of the CBIP algorithm is introduced, analyzed, and tested on a BIM model of an existing residential building. This algorithmic extension identifies correctly the high(2,3)-order 2LSB surfaces of external and internal, double- and triple-element constructions as well as respective opening 2LSB and shading surfaces related to these constructions. The results displayed the correctly identified second and third-order 2LSB surfaces on the demonstration building are also presented. Although generalization to the calculation of nthorder 2LSB surfaces is possible by induction of this algorithmic extension, cases of order (≥ 4) are rarely encountered. The algorithmic extension is going to be included as updates to the online calculation services of an existing cloud platform (Katsigarakis et al. 2022). ACKNOWLEDGEMENTS The research leading to these results has been funded by the European Commission H2020 project
301
“BIM-based holistic tools for Energy-driven Renovation of existing Residences” under contract #820621 (BIMERR) and project “COnstruction-phase diGItal Twin model” under contract #958310 (COGITO). REFERENCES Bazjanac, V. (2009). Implementation of Semi-automated Energy Performance Simulation: Building Geometry. In CIB W, volume 78, pages 595–602. Bazjanac, V. (2010). Space Boundary Requirements for Modeling of Building Geometry for Energy and Other Performance Simulation. In CIB W78: 27th International Conference. buildingSmart (2018). ISO 16739-1:2018 Industry Foundation Classes (IFC) for Data Sharing in the Construction and Facility Management Industries — Part 1: Data Schema. https://www.iso.org/standard/70303.html. Elagiry, M., Charbel, N., Bourreau, P., Di Angelis, E., and Costa, A. (2020). IFC to Building Energy Performance Simulation: A Systematic Review of the Main Adopted Tools and Approaches. In BauSIM conference of IBPSA. ERGOPRISM (2022). Residential House in Korakies Village in Akrotiri Region of Crete. https://www.ergoprism.com/ en/company/. Hitchcock, R. and Wong, J. (2011). Transforming IFC Architectural View BIMs for Energy Simulation. In Building Simulation Conference of IBPSA, pages 1089–1095. Karlapudi, J. and Menzel, K. (2020). Analysis on Automatic Generation of BEPS Models From BIM Model. In BauSIM conference of IBPSA.
Katsigarakis, K., Lilis, G. N., and Rovas, D. (2022). A Cloud IFC-Based BIM Platform for Building Energy Performance Simulation. In European Conference on Computing in Construction. Ladenhauf, D., Battisti, K., Berndt, R., Eggeling, E., Fellner, D. W., Gratzl-Michlmair, M., and Ullrich, T. (2016). Computational Geometry in the Context of Building Information Modeling. Energy and Buildings, 115: 78–84. Lilis, G., Giannakis, G., Katsigarakis, K., and Rovas, D. (2018). A Tool for IFC Building Energy Performance Simulation Suitability Checking. In European Conference on Product and Process Modelling (ECPPM), pages 57–64. Lilis, G. N., Giannakis, G., Kontes, G., and Rovas, D. (2014). Semi-Automatic Thermal Simulation Model Generation from IFC Data. In European Conference on Product and Process Modelling (ECPPM). Lilis, G. N., Giannakis, G., and Rovas, D. (2017). Automatic Generation of Second-Level Space Boundary Topology from IFC Geometry Inputs. Automation in Construction, 76:108–124. Rose, C. M. and Bazjanac, V. (2015). An Algorithm to Generate Space Boundaries for Building Energy Simulation. Engineering with Computers, 31(2):271–280. Vatti, B. (1992). A Generic Solution to Polygon Clipping. Communications of the ACM, 35(7):56–63. Ying, H. and Lee, S. (2021). Generating Second-Level Space Boundaries from Large-Scale IFC-Compliant Building Information Models Using Multiple Geometry Representations. Automation in Construction, 126:103659.
302
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Aspects of BIM-to-BEM information transfer: A tale of two workflows O. Spielhaupter CES Clean Energy Solutions, Vienna, Austria
A. Mahdavi Department of Building Physics and Building Ecology, TU Wien, Vienna, Austria
ABSTRACT: Achieving seamless transfer of information from BIM (Building Information Modelling) to BEM (Building Energy Modelling) has been the objective of multiple efforts. Nonetheless, to improve the state of art in this area, further efforts are needed. To this end, the examination of existing workflows for automated BIM-to-BEM information transfer can be useful. In this context, the present paper reviews available semiautomated approaches for information transfer from IFC-compliant BIM tools to a state-of-the-art building energy simulation application. The starting point of the case study is a synthetic building, which is modelled in two BIM-authoring tools. These models are subsequently exported to the data exchange format IFC, which are in turn transformed into the Input Data Files (IDF) of the simulation tool. The integrity and validity of the resulting geometry and semantic data are analyzed. Moreover, the simulation results obtained based on the two IDF instances are compared with a manually generated base case model. The findings of the case study point to: i) various challenges and constraints with regard to the generation of BIM models intended for subsequent export to BEM; ii) specific issues concerning the transferability of geometry and semantic data from IFC to IDF via the selected workflows; iii) inconsistencies in the resulting simulation results. Moreover, the case study provides pointers to enhancement possibilities of specific aspects of the examined workflows.
1 INTRODUCTION AND BACKGROUND Building Energy Modelling (BEM) is increasingly deployed in AEC (Architecture-EngineeringConstruction) for multiple purposes, including building certification and evaluation. A considerable fraction of data needed for BEM can be obtained from BIM (Building Information Modeling) tools. In this context, the present contribution considers BIM-to-BEM information transfer. The main motivation thereby is to address the challenges involved in automated BIMto-BEM data transfer processes. Data sharing in BIM typically utilize IFC (Industry Foundation Classes) files (O’Donnell et al. 2019; Ramaji et al. 2016). Automated data sharing can reduce the need for manual work of simulation model generation and hence increase the reliability and usability of performance simulation tools (Nielsen 1994; Picco et al. 2015; Ramaji et al. 2020a). Standardized semiautomated BIM-based workflows could replace the tedious and error-prone manual work and save time and costs. Additionally, BIM-based BEM workflows can increase consistency and accuracy when performing BEM in an iterative design process (Gao et al. 2019; Hitchcock et al. 2011; Kamel et al. 2019). Given this background, the present paper provides a brief overview of the current state of semi-automated IFC-based BIM-to-BEM transfer workflows relevant DOI 10.1201/9781003354222-39
for thermal building performance simulation. Moreover, two selected BIM-to-BEM workflows are investigated in detail in terms of a case study. There are many BEM tools (Gao et al. 2019) available, including EnergyPlus (U.S. Department of Energy 2021), TRNSYS (Thermal Energy System Specialists, LLC 2019), IDA Indoor Climate and Energy (EQUA Simulation AB 2020), and Modelica (Wetter 2009). In the present study, EnergyPlus was selected as the test case simulation engine mainly due to the manual check-up possibility of its input data file (IDF). The required data for dynamic energy simulation can be grouped into different categories: 1) location and the related weather file, 2) geometry, 3) construction and materials, 4) space types and their relation to thermal zones, 5) space loads, and 6) heating, ventilation, air-conditioning (HVAC) systems and components (Maile et al. 2007). The bulk of such data could be theoretically obtained through an ideal BIM-to-BEM workflow. The three most frequently used kinds of information transformation workflows are (Ramaji et al. 2020b): i) export BEM directly from BIM; ii) import and transformation of BIM files in a BEM tool; iii) transformation of BIM into BEM outside of BIM and BEM tools. The present paper focuses on the third option, which provides possibilities to standardize data transfer workflows for different BIM authoring and BEM
303
software solutions. However, some available workflows use common file formats of BEM tools as an intermediate step. Therefore, a strict differentiation between the different kinds of workflows is difficult. To transform information outside of BIM and BEM tools, it is required to choose a data file schema that acts as an intermediate schema. Different data file schemas are currently available for BIM-to-BEM data conversion, including IFC (buildingSMART International 2020), gbXML (Green Building XML Schema n.d., Gourlis et al. 2017; Karlapudi et al. 2020) , which is constrained by the usage of the center-line theory (O’Donnell et al. 2019), Honeybee Schema (Ladybug Tools LLC), SEMERGY Building Model SBM, which uses the structure of the Shared Object Model (SOM) from SEMPER (Mahdavi 1996) as a baseline (Fenz et al. 2016; Ghiassi 2013), Simulation Domain Model (Digital Alchemy, Inc. 2021; Giannakis et al. 2019; Nasyrov et al. 2014; O’Donnell et al. 2011 and OpenStudio Model (Alliance for Sustainable Energy, LLC). IFC is currently the only standardized and ISO certified open data format for information exchange in the BIM and BEM environment, offering promising possibilities as a baseline for the BIM-to-BEM data transfer (Ramaji et al. 2016). It is currently not possible to import BIM data exchange format directly into EnergyPlus. However, several third-party tools were developed to bridge this gap. These tools either convert the IFC directly to an IDF or first convert the IFC to a third-party file format and subsequently transform the third-party file format internally into an IDF for use in EnergyPlus. In a comprehensive literature review by Gao et al. (2019), BIM-based BEM workflows with different BIM authoring tools, different exchange formats, IFC, gbXML, etc., and different simulation engines were considered. Numerous workflows have been proposed and implemented. As such they cannot be all examined here. However, workflows that transform an exchange data format into the EnergyPlus IDF data format display notable similarities, lending perhaps justification to limiting the scope of exploration in this contribution to two workflows. 2 APPROACH
correctness; v) Evaluation of efforts (time investment) for manual pre and postprocessing; vi) Comparison of simulation results with a manually created baseline model. 2.1 Case study building The case study building model is an artificial building designed to include unique geometric and semantic properties (Figures 1 and 2). The building has three levels (basement, ground floor and first floor). Two protruding spaces on the ground floor have a flat roof and a sloped roof respectively.
Figure 1. 3D views of the case study building in ArchiCAD.
Figure 2. Ground floor plan in ArchiCAD.
2.2 BIM-to-BEM transformation
A case study with a BIM model is conducted to evaluate BIM-to-BEM transfer workflows. To this end, the BIM model is exported as an IFC file from the proprietary software tools Revit and ArchiCAD. These IFC files are subsequently transformed with available IFC-based BIM-to-BEM tools and workflows. The result of the workflow must be an EnergyPlus Input File (IDF). As a reference case, a manually created IDF is used to compare the IDF and the simulation results to the different IFC-based IDFs and their results. Specifically, the following steps were taken: i) Creation of a detailed building in BIM authoring tools; ii) Export of the BIM model to IFC; iii) Transformation of the IFC with the selected workflows; iv) Review of transformation results concerning consistency and
Two free and publicly available workflows for BIMbased BEM that use the IFC format were evaluated, namely “Python and IfcOpenShell” and “OsmSerializer” (Figure 3). Multiple steps (see the following) were necessary before an IFC file could be transformed and post-processed. 2.3 BIM Export to IFC data format The building designed in a proprietary software tool in the previous step is exported next as an IFC file, a key step in the workflow. The IFC data format can include much information on the designed building, only some of it necessary for simulation. It is possible to choose, via MVD (Model View Definition), a subset of the
304
available data. The best-suited MVD varies from workflow to workflow and is dependent on the desired IFC format. No state-of-the-art MVD is currently available that covers all the information required for the kind of workflows evaluated here. Later manual addition of data lost during the export process to the IFC file can be a cumbersome process. There are still problems due to different data management systems in proprietary BIM authoring tools and the lack of readily customizable IFC-exporters. MVDs ensure that specific data information is included in the exported IFC file. However, there is no automatic check of data correctness. IfcDoc (buildingSMART International 2021b) can be used to create custom MVDs or extend existing MVDs with additional requirements. 2.3.1 Check IFC for consistency Even if the IFC file is in accordance with the chosen MVD, information can still be missing. Therefore, the IFC should be checked in a viewer to evaluate especially space volumes and closed boundary objects (Patel 2020).
to specific standards. Key properties for thermal simulation such as thickness, conductivity, and density are standard in ArchiCAD and Revit. More specific (e.g., surface-related) properties are not defined by default. Data export involves the interface of the BIM authoring tool to IFC. 2.5
BIM authoring tools
ArchiCAD IFC exporter offers extensive possibilities to use and modify pre-defined MVDs. By customizing the settings for data conversion, it is possible to export material properties that would be in the IFC file. Revit does not export these material properties to IFC and gbXML, which is an issue related to the transformation of the built-in Revit schema to the exchange data formats (Nasyrov et al. 2014). Autodesk developed a customized IFC exporter (Autodesk Inc. 2020) that can replace the built-in exporter and export the model with different pre-defined or customized MVDs. However, the export of further material layer properties, such as thermal conductivity or specific heat, is not possible with this IFC exporter. One way to fix this is to use Dynamo, a visual scripting environment for Revit, which accesses objects through the Revit API. Alternatively, a custom made IFC exporter can be used. However, these solutions require extensive knowledge of Revit API (Ramaji et al. 2020b). Table 1. Availability of material properties in the BIM authoring tools and IFC. ArchiCAD Property name [units] Name [string] Roughness [string] Thickness [m] Conductivity [W.m−2 .K−1 ] Density [kg.m−3 ] Specific heat capacity [J.kg−1 .K−1 ] Thermal absorptance [-] Solar absorptance [-] Visible absorptance [-]
BIM + + + + + -
IFC + + + + + -
Revit BIM + + + + + -
IFC + + -
Figure 3. Flow chart of the two evaluated workflows. Legend: available (+), not available (-).
Currently, many different tools are available to view IFC files. However, not all tools have the same possibility to evaluate IFC files in as much details as required for the BIM-based BEM workflows. In the present study, the FZKViewer (KIT 2020) is used to check the created IFC files for consistency. 2.4 BIM to IFC mapping of Entities In the present study, ArchiCAD and Revit are selected as BIM authoring tools. Table 1 indicates the availability of required material properties in ArchiCAD and Revit. Information is available in BIM authoring tools, or custom properties can be created. But it is cumbersome to work in an OpenBIM environment with custom properties that are not defined according
2.6 IFC to IDF mapping of entities To fully understand the transformation workflows, it is important to check the mapping of relevant objects in detail. The mapping of the required IDF class objects from the IFC file for the two chosen workflows is derived from Patel (2020) and Ramaji et al. (2020b). The OsmSerializer workflow uses the OSM (OpenStudio Model) as an intermediate data file schema. Missing material properties in the created IDF data model can be added before further processing via mapping the existing objects to a construction library (Hitchcock et al. 2011). Missing properties could be also added manually.
305
2.7 Conversion of IDF to UTF-8 IFC is encoded in ANSI language format (Shadrina 2015). However, EnergyPlus is not able to read ANSI encoded files successfully. One encoding that does not include special characters is the 8-Bit Universal Coded Character Set Transformation Format (UTF8). If EnergyPlus classes include objects strings with language-wise specific special characters, these must be replaced with the synonym characters compliant with UTF-8 without additional encoding symbols, a transformation process that can be automated with a programming script.
the IDF and assigned to the surface objects. The classes for the HVAC system are added to the IDF, and the system is assigned to the thermal zones. Finally, the pre-defined simulation classes are added to the IDF. 3 RESULTS AND DISCUSSION The evaluation results of the two workflows are summarized in Table 2. Certain issues caused by geometry elements led to critical failures of the workflow. The affected elements were partly replaced with similar elements or neglected in the final model to evaluate the simulation results successfully.
2.8 Output check As EnergyPlus lacks a built-in graphical user interface, the geometry and corresponding semantic data of the resulting IDFs is validated with SketchUp and the OpenStudio plug-in. The resulting IDFs are subsequently checked with the built-in EnergyPlus IDF editor. Critical issues of the defined classes, such as missing references, are highlighted by the software tool. Obvious mistakes in the IDFs are fixable with reasonable effort. To run EnergyPlus, the geometry information can be retrieved from the IFC file. HVAC system information and further simulation options must be defined manually. The EnergyPlus system IdealLoadsAirSystem is added to the IDF files as a simplified HVAC system to evaluate heating and cooling loads. The simulation control classes are added manually, as well as output classes to evaluate the results. The geometry is evaluated by the resulting DXF file, which is a simple geometric representation generated by EnergyPlus. The simulation results are available as a Hypertext Markup Language (HTML) overview file and as a comma-separated values (CSV) file for detailed analysis. A first comparison of the resulting IDFs is made with Notepad++ (Ho 2021) using the compare addon, which highlights deviating text lines. As a second step, the visual comparison of the different IDFs is made by importing the resulting Drawing Interchange Format for AutoCAD (DXF) files of EnergyPlus into a CAD drawing format for AutoCAD (DWG) file. 2.9 Base energy model As a reference case, the case study building is manually created in SketchUp and exported as an IDF file via the OpenStudio SketchUp plugin. First, the geometry of the building is created. Second, openings are drawn and automatically matched to the corresponding surfaces with the OpenStudio plug-in. Third, semantic properties are assigned to the resulting building surfaces. Fourth, the created spaces are assigned to thermal zones. The resulting model is shown in Figure 4. The IDF is further processed in the built-in EnergyPlus IDF Editor and in plain text editor. The predefined constructions and materials are imported into
3.1 Python and IfcOpenShell It is necessary to adapt the python scripts to transform the IFC files successfully. A function to convert the resulting IDF to UTF-8 was implemented in the available scripts. Additionally, certain adjustments were necessary during the design process of the BIM model in ArchiCAD. The reference lines of the building elements are created according to the workflow description by Patel (2020). However, changes to the reference lines for the slabs to the ground were made for a successful transformation.
Figure 4. Manually created building model in SketchUp; visualization of the three different thermal zones.
The results of the transformation process are as follows: The scripts for transforming constructions worked only for multi-layered constructions. No information concerning the roughness was available, hence it was globally set to “MediumRough”. Regarding air layers, if there were varying thicknesses in one or more constructions, all previously created air layer material types were overwritten during transformation, with the resulting IDF having only one type of air layer. The semantic window and door information was missing after the transformation, necessitating the manual addition of the IDF classes. The boundary conditions of the interior walls are correctly assigned to the objects. However, the faces are not consistently orientated in the correct direction. The same applies also to interior openings and slabs between zones. Manual correction is required. The room over two storeys is not converted successfully.
306
Table 2.
Python and IfcOpenShell
OsmSerializer
wall for space over one storey
+
+
wall for space over multiple storeys slab, ceiling, roof, window, door curtain wall skylight
-
+
+
+
-
-
roof
-
+
wall
-
-
Shading element geometry
generic element
+
-
Boundary condition
wall for space over one storey exterior wall for space over one storey interior wall for space over multiple storeys wall to ground window, door exterior window, door interior roof ceiling/roof partly to outside basement floor slab above ground partly to outside
+
+/-
+/-
+/-
-
+/-
+
+
+/-
-
+ +
+ +
+ -
+ -
Constructions wall, slab, ceiling, roof window, door
+
+
+/-
+
Materials
+
-
+/+/-
-
Use case Rectangular surface geometry
Sloped surface geometry
leads to wrong outside boundary definitions of these objects. Given constraints related to structure, it is not possible to convert the geometry with the available python scripts. As such, curtain walls were neglected in the evaluation of the simulation results. Skylight geometry is converted. However, the skylights and the corresponding roofs are not on the same plane and are therefore not converted correctly. Sloped roofs are converted as surfaces without slope with the base zcoordinate. Sloped walls are not converted because they are not recognized by the scripts as walls but as roofs. Due to the wrong definition of these objects, it is not possible to run the scripts successfully. Sloped walls are therefore neglected in the evaluation of the simulation results. Door objects cause issues in the process of exporting to IFC and viewing the file with the FZKViewer. However, the geometry transformation from IFC to IDF works. The dimensions of the doors are larger than intended because the frame of the doors is included. In the design process, it is necessary to define a unique ID for each object.
Summary of transformation results.
wall, slab, ceiling, roof air gap window, door
3.2 OsmSerializer To export and transfer the model from Revit and transform it, the MVD Coordination View 2.0 with 2nd level space boundaries are used. The results of the transformation process are as follows: A space over two storeys is transformed correctly. Sloped roofs are converted correctly. The boundary conditions of walls adjacent to the ground are not defined correctly. The boundary conditions of the interior walls are correctly assigned to the objects. However, the faces are not consistently orientated in the correct direction. Slabs have partly the wrong orientation. There is no differentiation of outside boundary conditions for slabs partly adjacent to ground implemented. Manual correction of the boundary conditions is required. Assignment of the created spaces to thermal zones is not implemented in this workflow and is therefore done manually in SketchUp with the OpenStudio addon. Curtain walls, skylights, shading elements, and interior openings (doors and windows) are not converted at all. It is not possible to retrieve the material information from Revit via the standard IFC exporter. Constructions are exported with the correct thickness but all information regarding the thermal properties of the building objects is missing. 3.3 Simulation Results
Legend: working (+), partly working (+/-), not working (-).
The interior walls of the plenum in the upper floors are missing in the IDF file and the walls on the adjacent zones display wrong boundary definition. There is no differentiation of outside boundary conditions for slabs partly adjacent to ground and exterior walls to ground implemented in the python scripts, which
To compare the simulation results, the transformed constructions with the correct materials definitions of the Python and IfcOpenShell workflow were used for the manual baseline model and the OsmSerializer model. The results excerpted from the EnergyPlus HTML output files are included in Table 3. The diverging building areas can be explained with the calculation methods of the zone areas. In the
307
baseline workflow, the zone area is calculated based on the geometry boundaries. In the workflows based on the IFC the areas are calculated in the workflow. This has also an impact on the volumes. Especially in the Python and IfcOpenShell workflow, the volume is directly extracted from the ArchiCAD room properties. The efforts (time investment) for the selected workflows are compared in Table 4. Overall, the time required for the manual baseline method is the shortest. With geometry modifications in an iterative design process this can change rapidly. The workflows with the semi-automated transformation have an advantage over the manual creation if different geometry variations are to be evaluated. Furthermore, the automatic transformation of the constructions pre-defined in the BIM-authoring tools leads to reduced mistakes in the process of defining the EnergyPlus classes and can therefore improve the replicability and accuracy of the overall process. The tedious efforts for manual definition of constructions and materials can be obsolete with the automated transformation. Table 3.
Simulation results. Python and Manual IfcOpenOsmBaseline Shell Serializer
Name
Building area [m2 ] 300.00 Building volume [m3 ] 1027.00 Gross wall area [m2 ] 397.80 Heating load [MWh.a−1 ] 8.75 Cooling load [MWh.a−1 ] 0.56
240.00 647.58 387.90 9.11 0.26
222.25 812.32 365.26 7.29 0.07
Table 4. Comparison of the efforts (in hours) of the evaluated workflows.
Name Model creation IDF transformation correction IDF simulation classes Total
Manual Baseline
Python and IfcOpenShell
OsmSerializer
2 1.5
4.5 0.5
3.5 1.5
0.5
0.5
1
4
5.5
6
4 CONCLUSION We conducted a review of BIM-based BEM workflows. The focus was on IFC-based approaches for dynamic thermal building performance simulations with EnergyPlus. Two free publicly available approaches were selected for a detailed evaluation in a case study. The different steps of a successful BIM-to-BEM transformation workflow
were described in detail and difficulties were highlighted. The selected workflows were evaluated using an artificial case study building. The building was created in the BIM authoring tools ArchiCAD 24 and Revit 2020. The first transformation workflow uses an IFC file exported from ArchiCAD and transforms this file via Python scripts and the IfcOpenShell library to an IDF file. Building geometry, construction information and thermal zones were transformed. Window construction and material information was manually added. The second transformation workflow uses an IFC file exported from Revit and transforms this file via the add-on OsmSerializer for BIMserver and OpenStudio to an OSM. Building geometry, constructions without material information and space definitions were transformed. The spaces were manually assigned afterwards in SketchUp in combination with the OpenStudio plugin to thermal zones. Issues, such as the wrong orientation of surfaces, were manually corrected. An IDF file was afterwards exported via the OpenStudio plugin. Subsequently, required simulation parameters were manually added to the resulting IDF files of both workflows before running the simulation. The results of both simulations were finally evaluated and compared to each other and a manually created baseline model. It can be concluded for both evaluated workflows that the transformation of most used rectangular building objects was done correctly. Sophisticated geometry objects, such as curtain walls or sloped objects were converted in part incorrectly. Construction elements were in both workflows translated successfully. For the materials however, problems occurred with the export of the BIM authoring tools. The creation of space boundaries to define the relation between different spaces and thermal zones involved mistakes. All these issues caused the need for manual post-processing of the resulting IDF files. The runtime of the computational transformation processes and simulation was not covered in this study as it is highly influenced by the hardware and software environment. Likewise, the efforts for the setup of the BIM-authoring tools and the required tools for the two chosen workflows were not evaluated. The usability of the evaluated workflows was also not addressed. It can be generally concluded that the incorrect data transformation is mainly a topology-based issue. The main issue for BIM-based BEM is the deviation of models created in the design process in a BIM environment. There are currently no binding standards that mandate how objects need to be created in BIM authoring tools. This leads to major problems and incompatibilities in the (semi-)automated transformation processes for energy performance simulation. Due to several different actors and therefore software and BIM environments in different projects, it is difficult to define standards for the unique representation of building objects that are suitable for all project participants. Therefore, automated information transfer
308
buildingSMART International (2021a) BuildingSMART Data Dictionary [Online]. Available at https://www. buildingsmart.org/users/services/buildingsmart-data-dictionary/ (Accessed 17 February 2021). buildingSMART International (2021b) IfcDoc [Online]. Available at https://www.buildingsmart.org/standards/ groups/ ifcdoc/ (Accessed 11 October 2020). Digital Alchemy, Inc. (2021) Simergy [Computer program]. Available at https://d-alchemy.com/products/ simergy (Accessed 24 April 2021). EQUA Simulation AB (2020) IDA Indoor Climate and Energy [Computer program]. Available at https://www. equa.se/ en/ida-ice. Fenz, S., Heurix, J., Neubauer, T., Tjoa, A. M., Ghiassi, N., Pont, U., Mahdavi, A. (2016) ‘SEMERGY.net: Automatically Identifying and Optimizing Energy-efficient Building Designs’, Computer Science - Research and Development, vol. 31, no. 3, pp. 135–140. Gao, H., Koch, C., Wu, Y. (2019) ‘Building Information Modelling Based Building Energy Modelling: A review’, Applied Energy, vol. 238, pp. 320–343. Ghiassi, N. (2013) Development of a Building Data Model for a Performance-Based Optimization Environment, Diploma Thesis, Supervisor: Mahdavi, A., Vienna, University of Technology Vienna. Giannakis, G., Katsigarakis, K., Lilis, G. N., Rovas, D. (2019) ‘A Workflow for Automated Building Energy Performance Model Generation Using BIM Data’, Proceedings of building simulation 2019: 16th IBPSA International Conference and Exhibition, 2-4 Sept., Rome. Rome, International Building Performance Simulation Association, pp. 167–174. Gourlis, G., Kovacic, I. (2017) ‘Building Information Modelling for Analysis of Energy Efficient Industrial Buildings – A Case Study’, Renewable and Sustainable Energy Reviews, vol. 68, pp. 953–963. Green Building XML Schema (n.d.) About GreenBuildingXML [Online]. Available at https://www.gbxml.org/ About_GreenBuildingXML_gbXML (Accessed 10 October 2020). Hitchcock, R. J., Wong, J. (2011) ‘Transforming IFC architectural view BIMS for energy simulation: 2011’, Proceedings of Building Simulation 2011: 12th Conference of International Building Performance Simulation Association, Sydney, 14–16 November. Sydney, International Building Performance Simulation Association, pp. 1089– 1095. Ho, D. (2021) Notepad++ (7.9.1) [Computer program]. Available at https://notepad-plus-plus.org/ (Accessed 21 March 2021). Hong, S. (2020) Geometric Accuracy of BIM-BEM Transformation Workflows: Bridging the State-of-the-Art and Practice, Master Thesis, Ottawa, Ontario, Carleton University. Kamel, E., Memari, A. M. (2019) ‘Review of BIM’s Application in Energy Simulation: Tools, Issues, and Solutions’, Automation in Construction, vol. 97, pp. 164–180. REFERENCES Karlapudi, J., Menzel, K. (2020) ‘Analysis on Automatic Generation of BEPS Models From BIM Model’, BauSIM 2020 Alliance for Sustainable Energy, LLC OpenStudio§[Computer - 8th Conference of IBPSA Germany and Austria: 23–25 program].Available at https://www.openstudio.net/ (Accessed September 2020, Graz University of Technology, Aus24 April 2021). tria; Proceedings. Graz, Austria, 23–25 September. Graz, Autodesk Inc. (2020) IFC for Revit (revit-ifc) (20.3.1.0) Verlag der Technischen Universität Graz, pp. 535–542. [Computer program]. Available at https://github.com/ KIT (2020) Karlsruhe Institute of Technology, Institute Autodesk/revit-ifc (Accessed 2 December 2020). for Automation and Applied Informatics. FZKViewer buildingSMART International (2020) Industry Foundation (6.0, Build 1816) [Computer program]. Available at Classes (IFC) - An Introduction [Online]. Available at https://www.iai.kit.edu/1648.php (Accessed 14 Novemhttps://technical.buildingsmart.org/standards/ifc (Acceber 2020). ssed 31 October 2020).
between BIM authoring tools and commonly used software tools for BEM remains an elusive goal. Future efforts should target the definition of standards for the design of models in BIM authoring tools for the later usage. Specifically, the lack of clearly defined export properties for energy modelling in BIM authoring tools such as Revit or ArchiCAD needs to be addressed. The exported IFC file should include the semantic information from the BIM authoring tool in a structured manner. This should be independent of the type of algorithm used for geometry processing and included intermediate data formats. If extended, Revit IFC exporter could support material layer property export. This would not only benefit BIM-to-BEM workflows for thermal simulations, but also structural or ecological assessments. The usability of tools and their extensions in the context of transformation workflows, as well as their coupling should be enhanced to make the transformation workflows on a regular basis viable to a broader target audience. To avoid issues caused by the intermediate exchange formats, solutions based on the BIM authoring tools API could be developed in future studies (O’Donnell et al. 2019). However, extensive knowledge and maintenance effort for direct API access would be required for this approach. Language sensitivity is a common issue in all researched workflows. Building Smart Data dictionary (buildingSMART International 2021a) is a first step towards solving this issue because elements have a standardized naming that is understandable in many different languages. However, further development for the associated properties of the elements is necessary. BIM-based BEM workflows should integrate a quality check in each process step to ensure the accuracy and correctness of the workflow. A tool for the comparison of BEM models with the input files was developed recently (Hong 2020). The integration of this kind of tools in a standardized BIM-based BEM workflow is worth investigating. Semantic web and connected data approaches, similar to the SEMERGY approach (Fenz et al. 2016), offer an alternative to the manual definition of correct object properties in BIM for later BEM usage (Karlapudi et al. 2020). With the rising complexity of BIM models and the availability of databases encompassing detailed information on building objects, the significance of this approach could rise in the coming years.
309
Ladybug Tools LLC Ladybug Tools [Online]. Available at https://www.ladybug.tools/ (Accessed 1 January 2021). Mahdavi, A. (1996) ‘Semper: A New Computational Environment for Simulation-based Building Design Assistance’, Proceedings of the 1996 International Symposium of CIB W67: Energy and Mass Flows in the Life Cycle of Buildings. Vienna, Austria, pp. 467–472. Maile, T., Fischer, M., Bazjanac, V. (2007) ‘Building Energy Performance Simulation Tools - a Life-Cycle and Interoperable Perspective’, Facil. Eng. (CIFE) Working Pap., vol. 107. Nasyrov, V., Stratbücker, S., Ritter, F., Borrmann, A., Shan, H., Lindauer, M. (2014) ‘Building Information Models as Input for Building Energy Performance Simulation – The Current State of Industrial Implementations’, in Mahdavi, A., Martens, B., Scherer, R. J. (eds) eWork and eBusiness in architecture, engineering and construction: Proceedings of the 10th European Conference on Product and Process Modelling (ECPPM 2014), Vienna, Austria, 17–19 September 2014, Balkema, 1/1/Boca Raton|London|New York|Leiden, CRC Press, pp. 479–486. Nielsen, J. (1994) Usability Engineering, Morgan Kaufmann Publishers Inc. O’Donnell, J. T., See, R., Rose, C., Maile, T., Bazjanac, V., Haves, P. (2011) ‘SimModel: A Domain Data Model for Whole Building Energy Simulation’, Proceedings of Building Simulation 2011: 12th Conference of International Building Performance Simulation Association, Sydney, 14–16 November. Sydney, International Building Performance Simulation Association. O’Donnell, J. T., van Dessel, M., Maile, T. (2019) ‘BIM to Building Energy Performance Simulation: An Evaluation of Current Industry Transfer Processes’, Proceedings of building simulation 2019: 16th IBPSA International conference and exhibition, 2–4 Sept., Rome. Rome, International Building Performance Simulation Association, pp. 92–99.
310
Patel, K. P. (2020) BIM Model Enrichment for Energy Performance Simulation, Master Thesis, Supervisor: Menzel, K., Dresden, Technische Universität Dresden. Picco, M., Marengo, M. (2015) ‘On the Impact of Simplifications on Building Energy Simulation for Early Stage Building Design’, Journal of Engineering and Architecture, vol. 3, no. 1, pp. 66–78. Ramaji, I. J., Memari, A. M. (2020a) ‘Interpreted Information Exchange: Implementation Point of View’, Journal of Information Technology in Construction, vol. 25, pp. 123–139. Ramaji, I. J., Messner, J. I., Leicht, R. M. (2016) ‘Leveraging Building Information Models in IFC to Perform Energy Analysis in OpenStudio’, ASHRAE and IBPSA-USA SimBuild 2016: Building Performance Modeling Conference. Salt Lake City, Utah, pp. 251–258. Ramaji, I. J., Messner, J. I., Mostavi, E. (2020b) ‘IFC-Based BIM-to-BEM Model Transformation’, Journal of Computing in Civil Engineering, vol. 34, no. 3, 04020005, 1–13. Shadrina, A. (2015) Framework for the Transfer of Building Materials Data Between the BIM and Thermal Simulation Software, Diploma Thesis, Supervisor: Mahdavi, A., Vienna, University of Technology Vienna. Thermal Energy System Specialists, LLC (2019) TRNSYS: Transient System Simulation Tool [Computer program]. Available at http://www.trnsys.com/ (Accessed 5 April 2021). U.S. Department of Energy (2021) EnergyPlus (9.2) [Computer program]. Available at https://energyplus.net/ (Accessed 27 March 2021). Wetter, M. (2009) ‘Modelica-based Modelling and Simulation to Support Research and Development in Building Energy and Control Systems’, Journal of Building Performance Simulation, vol. 2, no. 2, pp. 143–161.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Predicting annual heating energy use using BIM: A simplified method for early design phase M.F. Stendahl & M.C. Dubois Division of Energy and Building Design, Faculty of Engineering LTH, Lund, Sweden
ABSTRACT: Building information modeling (BIM) and building energy modeling (BEM) are two key tools to make the transition to net-zero energy buildings (NZEB). However, a recent literature review indicated that the conversion from BIM to BEM is currently dysfunctional, due to limitations regarding competencies, processes, and technology. In this article, a case study is presented to show how the information in the BIM model can be used to calculate the annual heating energy use (kWh/m2 , year) using the simple degree-day method and extracting building envelope surfaces, heated floor area, and heated volume in the BIM model. The proposed method is for very early design phase (EDP) when architects are starting to determine the general building shape, window sizes, etc. This case study is a residential building called Eskilshem, located in Södermanland, Sweden (latitude 59.4◦ N, longitude 16.5◦ E) designed by White arkitekter in 2020. This article presents the workflow and equations embedded in BIM to obtain annual heat energy demand at EDP.
1 INTRODUCTION 1.1 Energy calculations at early design phase Due to climate change and pressures on resources, the demand for a more sustainable built environment with energy-efficient buildings is increasing globally (World Energy Council 2013). As a world leader in building energy-efficiency, the European Union established a set of directives to generally upgrade the building stock and replace energy-inefficient buildings. In this endeavor, one such key directive is the Energy Performance of Buildings Directive (EPBD) (Official Journal of the European Union 2012), which states that all new buildings shall be nearly zero energy buildings by 2020. The nearly zero energy definition varies as a function of country depending on climate and energy production system. In many European countries, the EPBD has been the backbone for developing national building regulations. Sweden, which is used as case study example in this article, has today one of the most ambitious energy regulations in the world (Boverket 2022), which currently set a limit of 75 kWh/m2 , year for primary energy in multi-family residential buildings. This primary energy threshold includes all end-uses, i.e. heating, cooling, ventilation, domestic hot water (DHW), and collective electricity use. The only part that is excluded is the individual electricity use (plug loads, electric lighting, etc.). The most common method to verify compliance with building regulations at the early design phase (EDP) is building energy modeling (BEM) (Samset & Volden 2016), which normally entails dynamic energy simulations, with well-established programs DOI 10.1201/9781003354222-40
such as e.g. EnergyPlus (DOE 2022a), IDA-ICE (Equa 2022), ClimateStudio (Solemma 2022), etc.These programs typically require information about building geometry, construction, thickness and thermal properties of building materials (thermal conduction i.e. U-values, specific heat capacity, etc.), internal heat loads from occupants and equipment (lighting, appliances, etc.), leakage, mechanical ventilation rates (ach, L/s, m3 /h), etc. This information is used to perform so-called dynamic energy simulations, which are hourby-hour calculations of the building thermal balance based on information about outdoor climatic conditions retrieved from a standardized climate file. The climate files, available through e.g. the EnergyPlus weather data (DOE 2022b) and other sources, typically provide hourly data about outdoor air temperatures, direct and diffuse solar radiation, wind speeds, etc. However, this process is quite time-consuming and expensive, as it requires detailed building, operation, and climate information, which is not always available at EDP. Besides competencies (i.e. hiring a skilled energy simulation expert), this process also requires a translation from BIM to BEM models, which is not flawless at the moment, in terms of automation and interoperability, as outlined in a recent literature review (Stendahl et al. 2022). This article proposes an alternative pathway compatible with lower levels of development (LOD) 100300, see Sacks et al. 2018. This pathway consists of working directly in the BIM model to retrieve relevant geometrical information and use well-established equations from building physics, to perform an energy calculation based on the degree-day method (CIBSE 2006; Hagentoft & Sandin 2017). This pathway is
311
developed in this article using a case study, which is a real building project called ‘Eskilshem’, located in Eskilstuna, 87 km west of Stockholm, Sweden (latitude 59.4◦ N, longitude 16.5◦ E). This building was designed by the Scandinavian firm White arkitekter in 2020. The main aim of this article is to demonstrate a simple method to assess annual heating demand (kWh/m2 , year) directly in the BIM model, by extracting building envelope surfaces, heated floor area (Atemp ), and heated volume needed in the energy calculations, which use the interior surfaces of exterior walls, roof, and floor as bounding surfaces for thermal zones. This calculation is useful at EDP during the concept design phase to make sure that the building will be able to comply with national energy regulations. Detailed dynamic energy simulations were performed later in the design process, to confirm compliance. These results are presented and discussed further down in this article.
Figure 1. Revit screenshot of Eskilshem (top) and isolated bloc for the energy analysis (bottom).
2 METHOD
prepare for the energy calculations. This was the most time-consuming step since the original Revit model was not prepared for this purpose from the start. To simplify the process, a section of the model was isolated, shown in Figure 1 (bottom left). Subsequently, a few basic transformations were applied on the original BIM model. These transformations are necessary to extract the correct surfaces, heated floor area, and volume needed in the energy calculation, which requires surfaces and volumes defined from the model’s interior surfaces. The basic transformations applied to the model are described below: – Each thermal zone was created as a “Room” in the menu “Architecture” under “Modify/Rooms” with the same “Name” in the menu “Properties” under the category “Identify Data”. – The height of each thermal zone (“Limit offset”) was defined from floor to floor in the menu “Properties” under “Constraints”. Note that the last floor of the building is an exception, where the limit offset is from floor to ceiling. Figure 2 shows a section of the isolated bloc with the correct limit offset of 2800 mm from floor to floor and 2500 mm from floor to ceiling for the last floor. – Data for interior partitions were not considered in the energy calculation. These surfaces were assigned the parameter “Room Bounding” as unselected in the menu “Walls Properties” under “Constraints”. – Underground construction for the garage was not considered in this analysis. The air in the garage was considered at the same temperature as air inside the apartments and thus, no heat transfer occurred through the floor surface, which was thus considered adiabatic. – The windows and doors elements of the model should be built with area parameters to facilitate data extraction.
2.1 Building geometry The study object Eskilshem is a residential building containing 200 apartments covering approximately 16,600 m2 , with 70 parking spaces in an underground garage, see Figure 1. Besides national building energy regulations (Boverket 2022), for 2020 (85 kWh/m2 , year), Eskilshem had to comply with a more ambitious energy-efficiency target through the “Miljöbyggnad Silver” environmental certification system, which limits primary energy use at 80% of building code level (thus 68 kWh/m2 , year). “Miljöbyggnad” is a Swedish environmental certification system for buildings, similar to BREEAM and LEED, but developed for the Swedish context. Note that the primary energy use target in Miljöbyggnad includes the same end-uses as the national building energy code. 2.2 Transformation of the BIM model Eskilshem was built in the RevitVersion 2020 Environment. The first step consisted of adapting this model to
Figure 2. Revit screenshot of part of the roof section with the correct limit offset.
In Revit, a schedule displays information extracted from the “Properties” of each element. A schedule can provide every instance of the element type (Wing 2019). Once the model adaptation was completed with the basic transformations listed above, five schedules were defined to generate data required in the energy calculations. These schedules included quantities with: The “Room” schedule providing the total interior area in m2 for each floor, attic, and basement heated
312
Figure 3.1. Step 1, each “Room” should have a “Name” under “Identify Data” (Bloc C) used as a filter in the schedule properties.
Figure 3.2. Step 2, select “Schedules” in the “View” menu to create a new schedule with “Multiple” as “Filter list” and “Architecture” as “Structure” selected, then select “Rooms” as “Category”. After selecting “OK”, the schedule “Properties” will automatically open.
Figure 3.3. Step 3, in schedules “Properties” select from “Available fields”: “Name”, “Number”, “Area” and “Level” then select “OK”.
Figure 3.4. Step 4, in schedule “Properties”, select “Filter” by “Name” (Bloc C).
to more than 10◦ C (threshold according to Swedish building regulations). Note that the “Room” name was used as filter in the schedule “Properties”. Figures 3.1 to 3.4 shows and explains the steps to create a “Room” schedule. The “Roof” schedule provided the total roof area in m2 . “Keynote” was used as filter in the schedule “Properties”. A “Keynote” parameter is available for all model elements as materials, and it is added under “Identify Data” in the “Properties” menu of elements or materials, see Wing (2019). The “Wall” schedule provided the total exterior wall area using the interior surfaces in m2 . Again, “Keynote” was used as filter in the schedule “Properties”. The “Windows” schedule provided the total window area in m2 . Again, “Keynote” was used as filter. The windows elements of the model were not built with area parameters; thus, a “Formula” multiplying the windows’ “Height” and “Width” was applied in the calculated value of the schedule “Properties”. The “Doors” schedule provided the total door area in m2 . “Keynote” was also used as filter in the schedule “Properties”. The doors elements of the model were not built with area parameters; therefore, a “Formula” multiplying “Height” and “Width” was applied for the doors in the calculated value of the schedule “Properties.” 2.3 Energy calculation The transformations on the original Revit model allowed obtaining: – the total surface area (A, in m2 ) of walls, roof, windows, doors loosing heat to the outside. – total heated floor area (Atemp, in m2 ). – the total heated volume (V, in m3 ), which can be used to estimate heat losses due to mechanical ventilation and leakage through the envelope. The surface areas (A) and volume (V) can then be used to calculate annual heating demand using the heating degree-day method. Originating from agriculture (Strachey 1878), degree-days (DD) can be used to predict energy use in buildings. Buildings are complicated thermal environments involving a large number of variables (heat loss, occupancy, solar heat gains, window sizes, glazing type, etc.), which influence energy demand. For this reason, a full dynamic thermal simulation is often the preferred method. However, simulations involve a large amount of inputs, skills, as well as time to obtain reliable results, while the information needed in the input is not necessarily available at very early design phase. Furthermore, simplified estimation processes and tools can reduce the amount of effort leading to rapid results (CIBSE 2006). With this DD method, a rough estimate of the energy demand is obtained, which is valuable when most of the energy demand consists of heating needs, according to Verbai et al. (2014), as in a housing project. DD are essentially ‘a summation of the
313
differences between the outdoor temperature and some reference temperature over a specified time period’ (CIBSE 2006; Hagentoft & Sandin 2017). The reference temperature (Tref , also called ‘base temperature’) corresponds to a balance point temperature, i.e. the outdoor temperature at which the heating (or cooling) systems do not need to be in operation to maintain comfort conditions. When the reference temperature is calculated accurately, the correlation between energy use and DD is reliable (D’Amico et al. 2019). Note that DD can apply to heating or cooling systems, leading to the separate concepts of heating and cooling degree-days (referred to as HDD and CDD in this article). A key aspect in the application of HDD is the definition of Tref , which depends largely on the quality of the building envelope. Better building envelopes (i.e. lower U-value) normally correspond to lower Tref since the building can maintain indoor comfort conditions at lower outdoor temperatures, as it does not loose heat as quickly as poorer building envelopes. In Sweden, Tref can be as low as 12◦ C for a passive house and 17◦ C for a conventional construction. Considering Tref of 17◦ C from the start is recommended as it leads to a larger number of HDD and thus, a more pessimistic (i.e. higher) estimate of annual heating loads at the EDP. Furthermore, note that the use of Tref of 17◦ C allows taking into consideration solar radiation and internal heat gains, which both contribute to increase the indoor temperature from 17◦ C to the desired indoor temperature i.e normally around 21◦ C (Neuman 2015). Thus the Tref of 17◦ C was used in the present case study. Annual HDD are normally the summation (over the year) of the differences between Tref (17◦ C) and the diurnal average outdoor temperature (Tout_daily_ave ). Alternatively, when diurnal averages are not available, one may use the difference between Tref (17◦ C) and the annual average outdoor temperature (Tout ), which provides approximately the same HDD. From these values, HDD can be calculated using Equation 1: HDD = (Tref − Tout ) · 365
(1)
The term 365 in Equation 1 is for 365 days/year. This value must be multiplied by the thermal conductance (U-value in W/m2 , ◦ C) of building parts to obtain the annual heat losses through conduction processes (Q_cond ) through the building envelope using Equation 2: Q_cond = (U · A) · HDD · 24
(2)
Note the term 24 in Equation 2 is for 24 h/day, which provides consistency in the units so that the final value is in Wh and not Wday. The ‘heat loss coefficient’ (q_cond = U · A) can be obtained by weighting the U-value by the surface area (A) of each building part, see Table 1. In the Eskilshem case study, typical U-values were assumed for Swedish residential constructions, see Table 1. These values are relatively low and correspond to a well-insulated building envelope.
Note that Equation 2 can be programmed directly in the BIM model to obtain conduction heat losses without having to perform any calculation or simulation outside the BIM environment. This is extremely practical at lower LOD 100–300, when the design process still involves iterations and changes (Andriamamonjy et al. 2019). Table 1. Thermal properties of building parts in the Eskilshem case study. Building part
U-value (W/m2 , ◦ C)
Roof Walls Windows Doors
0.15 0.15 1.00 1.00
Total ( ) Atemp
Surfaces (A) (m2 )
U·A (W/◦ C)
64.68 424.74 86.90 29.56
9.70 63.71 86.90 29.56
605.88 388.08
189.87
The next step concerns heat losses through air leakage and mechanical ventilation.The heat loss coefficient (q_vent_total ) due to leakage through the building envelope and mechanical ventilation was calculated using standard equations involving the air density (ρ = 1.2 kg/m3 ) and heat capacity of air (cp = 1000 J/kg, ◦ C), leakage flow (ϕ_leak ), ventilation flow (ϕ_vent ), and heat recovery of the mechanical ventilation system (η) as described in Equation 3: q_vent_total = ρ · cp · ϕ_vent (1 − η) + ρ · cp · ϕ_leak
(3)
In this equation ρ and cp can be considered as constant at ordinary ambient air temperatures and pressures. What remains to be determined is the mechanical ventilation flow (ϕ_vent ), first term in Equation 3, which was obtained from the minimum requirement by heated floor area (Atemp , in m2 ) in the Swedish building code, i.e. 0.35 L/s, m2 . In absence of ventilation requirements in the building code, one can use air change per hour (ach) instead to obtain the ventilation airflow using Equation 4: ϕ_vent = ach · V
(4)
In Equation 4, V is the total volume of indoor air that must be ventilated, retrieved from the BIM model, see Section 2.3. Besides the minimum ventilation requirement of 0.35 L/s, m2 , some codes require extra ventilation for persons living in the building. A typical value considered is 7 L/s, person. A simple way to account for this additional ventilation flow is to assume an occupancy rate of about 1 person/30 m2 , thus 0,23 L/s, m2 . Note that the occupancy may differ depending on building type, etc. For offices, one would normally consider about 1 person/10 m2 instead. This extra ventilation was not considered in the present case study, but could be added to the heat balance calculation within the BIM model if needed.
314
For air leakage, the airflow was determined from the total surface area of the building envelope obtained from the Revit model (605.88 m2 , Table 1), multiplied by a leakage factor of 0.000015 m3 /m2 /s calculated from two Swedish and European standards (2000, 2020). Note that this volumetric air leakage is given per m2 of building envelope (and not floor area), thus the interest of retrieving total area of building envelope from the BIM model. The heat loss coefficient through leakage (q_leak ) can then be calculated using Equation 5 (which is the same as the term on the right side of Equation 3): q_leak = ρ · cp · ϕ_leak
(5)
The next step consists of multiplying the result from Equation 3 by HDD and 24, as described below: Q_vent_total = q_vent_total · HDD · 24
(6)
These values are then added to conduction heat losses to obtain the annual heating energy use (Q_heating ): Q_heating = Q_cond+ Q_vent_total
(7)
Note that Equations 1–7 are easy to implement directly in the BIM model, provided that the correct surfaces, heated floor area, and volume are extracted, as explained in Section 2.2. It is thus fully possible to obtain the annual heating demand using the DD method with a BIM model at LOD 100–300 (Andriamamonjy et al. 2018).
3 RESULTS In this section, we show how the method presented in Section 2 was applied in the Eskilshem project. 3.1 Calculation of surfaces and volumes from BIM model A control method was devised to check the reliability of the newly developed model in Revit. This control method was a simple ‘manual’ calculation method, meaning that the size of surfaces and volumes were determined one by one by scrutinizing the model and entering dimensions on a separate Excel sheet. Despite extensive efforts invested in adapting the model and creating five schedules, only two of the five schedules’ results successfully matched the control method. However, for windows and doors, the schedule results provided identical results with respect to the control method. The BIM model was investigated following the preliminary results to understand why the “Room,” “Roof,” and “Wall” schedules results did not match the results of the control method.
3.1.1 Room Although the architecture model had the appropriate “Limit Offset” (2800 mm) in the “Room” “Properties”, Revit does not consider the “Limit Offset” setting if the “Rooms Bounding” is selected in the “Properties” of the floors. After this minor modification, the result from the “Room” schedule (Atemp ) returned 388.1 m2 , while it was 388.08 m2 from the control method, respectively, which is a negligible difference. 3.1.2 Roof The preliminary results from the BIM model did not match the control method because the roof of the architecture model was not drawn correctly. The façade of the last floor needs to have limits under the roof, and the roof area should be located behind the façade. Figure 2 shows part of the roof section. When the roof details were correctly drawn, the area from the roof schedule was 64.7 m2 , while it was 64.68 m2 in the control method, which is a negligible difference. 3.1.3 Wall The energy calculation used for the control method was achieved by calculating the total area of the exterior walls from the inside perimeter, bounding the thermal zone. The wall schedule generated in Revit is calculated by default from the center point of the exterior walls. Figure 4 shows the wall calculation from Revit and the control method. One way to solve the default setting in Revit was to duplicate the “Room” schedule. Then in the calculated value of the schedule “Properties,” a “Formula” multiplying “Perimeter” with “Limit Offset” was applied to obtain the inside area of the exterior walls. Following these settings, the wall schedule area was 541.2 m2 (including windows and doors) and the exact same result was obtained from the control method. After minor adaptations in the architectural model to generate quantities and using the control method to understand how the data were generated in Revit, the room, roof, and wall schedules’ results successfully matched the control method.
Figure 4. Revit screenshot of wall default calculation from Revit and the control method.
315
3.2 Energy calculations 3.2.1 Degree days and hours In the case of Eskilshem, the average exterior air temperature (Tout ) of 6.93◦ C was used as this corresponds to the average outdoor temperature in Eskilstuna, for which an *.epw climate data file from 1981-2010 was available at the time of performing this analysis. The climate data files were created by the Swedish Institute of Meteorology and Hydrology (SMHI), and obtained through SVEBY, which is a crossindustry program developing tools for agreements on energy use. Equation 1 returned 3675.55 ◦ C, day (HDD) for the case study. 3.2.2 Heat losses by conduction The heat loss coefficient (U · A) obtained was 189.87 W/◦ C (see Table 1), which was multiplied by a factor 1.25 (thus U · A = 237.34 W/◦ C) to account for extra heat losses through thermal bridges. The result of Equation 2 was 20937 kWh (237.34 W/◦ C ·3675.55◦ C, day · 24 h/day / 1000) for total heat losses by conduction through the building envelope. Considering total heated surface (Atemp ) of 388 m2 , it corresponds to 54 kWh/m2 , year just for conduction heat losses, which is significant. 3.2.3 Heat losses due to air leakage and ventilation This part was calculated using Equation 3. The heated floor area (388 m2 ) was retrieved from the BIM model. A heat recovery of 75% or 0.75 was assumed as this is a standard value used in energy-efficient heat recovery systems.The mechanical ventilation flow rate obtained was 135.8 L/s, which is the same as 0.136 m3 /s (i.e. 388 m2 · 0.35 L/s, m2 / 1000 L/m3 ), which results in a heat flow coefficient of 163.0 W/◦ C (i.e. 1.2 kg/m3 · 1000 J/kg◦ C · 0.136 m3 /s) for mechanical ventilation without heat recovery, and only 40.7 W/◦ C with heat recovery. Note that without heat recovery, the heat loss coefficient due to ventilation is almost the same as for losses through conduction (163.0 W/◦ C versus 189.87 W/◦ C), but heat recovery takes this value down to 40.7 W/◦ C, which outlines the importance of heat recovery systems for energy-efficient buildings. For air leakage through the building envelope, the result obtained was 0.0091 m3 /s. From Equation 5, we obtain: q_leak = 1.2kg/m3 · 1000J/kg◦ C · 0.0091m3 /s = 10.9W/◦ C The total heat loss coefficient through combined ventilation and leakage is thus 173.9 W/◦ C without heat recovery and 51.6 W/◦ C with heat recovery. When fitting these values in Equation 6, the resulting heating demand due to ventilation (mechanical ventilation plus leakage) is 15 340 kWh without heat recovery and 4551 kWh with heat recovery.
3.2.4 Total annual heating load and primary energy use The total annual heating demand obtained can be calculated with Equation 7: Q_heating = 20937 + 15340 = 36277 kWh (without heat recovery) Q_heating = 20937 + 4551 = 25488 kWh (with heat recovery) These values are then divided by the total heated floor area (Atemp , 388 m2 ) retrieved from the BIM model, which results in: 93.5 and 65.7 kWh/m2 , year without and with heat recovery respectively. Since the total primary energy use is limited to 85 kWh/m2 , year in the building code (see Section 2) and 68 kWh/m2 , year in the Miljöbyggnad Silver certification, we know at this stage that a highly energy-efficient heat recovery system (with η > 75%) is a requirement in this case to reach the targets, still leaving very little room for DHW, electricity, etc. In this case, this analysis allowed recommending lower U-values already at EDP, since results indicate that conduction heat losses are significant. 3.3 Dynamic energy simulation Dynamic energy simulations were performed at a later stage of the project by an environmental specialist at White arkitekter, using the software IDA-ICE (Bakerkter, 2020). The building’s energy use was dynamically simulated during a normal year under given conditions with the same climate file described previously in Section 3.2.1, floor plan, building construction, HVAC (heating, ventilation and cooling) systems and operations. The program calculated the building as a multi-zone model, where each room corresponded to a zone. This means that heat conduction through inner walls and floors as well as heat transfer between zones via air flows were taken into consideration. The given conditions and climate file specifications consisted of: – climate data, Eskilstuna, *.epw climate data file from 1981–2010, see Section 3.2.1; – U-value for roof and external walls, 0.11 W/m2 , ◦ C; – U-value for windows and doors, 1.00 W/m2 , ◦ C; – thermal bridges, assumed as 30% (factor 1.30) of the transmission losses according to practice for Miljöbyggnad 3.1; – leakage, 0.3 L/s, m2 for the enclosing envelope at 50 Pa; – occupancy rate; 1.42 person/1 room apartment, 1.63 person/2 room apartment, 2.18 person/3 room apartment with an occupancy of 14 hours/day and 7 days/week; – internal heat loads from equipment (lighting, appliances, etc.) of 30 kWh/m2 , where 70% of these as sensible heating load; heat of recovery, 85%.
316
The results of the dynamic energy simulations are presented in Table 2. Table 2. Annual and primary energy use.
Use energy (kWh/m2 , y) Heating Hot water Surcharge Cold Fans Pumps Lifts Other property el. Total ( )
27.7 28.2 4.0 0.0 5.6 2.0 3.0 1.5
Primary energy factor 1.0 1.0 1.0 1.0 1.6 1.6 1.6 1.6
Primary energy (kWh/m2 , y) 27.7 28.2 4.0 0.0 9.0 3.2 4.8 2.4 79.3
4 DISCUSSION The aim of this study was to validate a simplified method to predict annual heating demand directly in the BIM model at very early design phase i.e. concept design phase with a LOD of 100, using the DD method. The results precisely matched the manual control method developed to ensure the reliability of the created model in Revit. Furthermore, the results were compared to those of a dynamic energy simulation of the real project at the preliminary design phase with a LOD of 200, using the validated software IDA-ICE. Results from the dynamic energy simulation for the primary heating demand were 27.7 kWh/m2 , year. This result is much lower compared to the results obtained with the DD method (65.7 kWh/m2 , year). However, the difference in the results is partially explained by the modification of the program from the concept to the preliminary design phase. For the dynamic energy simulations, the number of apartments was 136 and a kindergarten with three smaller rooms was also considered in the calculation. For these reasons, the occupancy and ventilation rates did not have similar conditions as the DD method. Moreover, internal heat loads from equipment (lighting, appliances, etc.) and people were considered in the dynamic simulations; the U-value for the roof and walls was lower and the thermal bridge factor was higher. Finally, the DD method used 75% heat recovery versus 85% in the dynamic simulations. Only considering the impact of 75% or 85% heat of recovery in the given conditions creates a difference of 6% for the total annual heating demand with the DD method. Feeding the values for heat recovery used in the dynamic simulations in the DD calculation returns 62 kWh/m2 , year. Additionally, reducing the U-values from 0.15 to 0.11 W/m2 , ◦ C and altering the thermal bridge factor from 1.25 to 1.30 further reduces the annual heating demand to 58 kWh/m2 , year.
After implementing these changes in the DD method, the relative difference for annual primary heating energy demand between the two methods was nearly 50% (27.7 kWh/m2 , year versus 58 kWh/m2 , year). Nevertheless, considering the difference in input conditions for the dynamic simulation regarding occupancy, ventilation rates, and internal heat loads, the results are logical i.e. the dynamic method should return lower heating demand as it includes internal heat loads. It is also worth noting that the use of the DD method very early in the design process helped establish an early communication with the architects and builders regarding the necessity for low U-values and high heat recovery, which ensured that the building complied to the code at later stages. We thus conclude that the method presented in this article is relevant for the industry and applicable in other countries, as long as the climate data file corresponds to the context of the project. However, more case studies are needed to confirm these first results and improve the proposed method and reinforce this conclusion. 5 CONCLUSIONS A case study was presented about annual heat load calculations in a BIM model based on the DD method. The case study involved adaptation of the model of a large residential building called Eskilshem located in Eskilstuna, west of Stockholm, Sweden. This case study demonstrates that: – The BIM model must be adapted or prepared according to some specific standard to be able to extract the right surfaces, heated floor area, and volume for the energy calculation. The current way to build the model is not performed according to energy calculations. This method for building the model could be adopted as a standard by the building industry. – Using the DD method, it is possible to predict annual heating energy use at the EDP directly in the BIM model, by implementing a few simple equations as schedules. This method provides information useful in the iterative design process. Providing these schedules directly in BIM would greatly facilitate energy design of buildings even at the concept design phase with LOD 100, which would contribute to reach national energy efficiency targets, also avoiding costly adaptations of the model at more detailed design stages. Finally, commercial tools for energy simulations inserted in BIM, such as e.g. @Revit Energy Optimization, allow performing energy optimization, and these tools should definitely be monitored closely in the coming years as one alternative for performing energy predictions when more information about the project (occupancy, internal loads, etc.) is available. The next step in this research is to implement the method presented in this article as BIM schedules and test them in building practice.
317
ACKNOWLEDGMENTS This study is part of a doctoral research project funded by the Swedish Energy Agency and the Scandinavian firm White arkitekter. The authors also thank Daniel Forgues (ETS, Montreal, Canada) for contributions in the discussions and for initiating this research collaboration. REFERENCES Andriamamonjy, A., Saelens, D., & Klein, R. (2018). An Automated IFC-based Workflow for Building Energy Performance Simulation with Modelica. Automation in Construction, 91, 166–181. https://doi.org/10.1016/j.autcon. 2018.03.019 Andriamamonjy, A., Saelens, D., & Klein, R. (2019). A Combined Scientometric and Conventional Literature Review to Grasp the Entire BIM Knowledge and its Integration with Energy Simulation. Journal of Building Engineering, 22, 513–527. https://doi.org/10.1016/j.jobe.2018.12. 021 Baker, N (2020). Eskilshem, Eskilstuna Energiberäkning, White Arkitekter. Retrieved Through Direct Email to the Author on 2022-05-15. Boverket. 2022. Boverkets Föreskrifter om Ändring i Boverkets Byggregler (2011:6) - Föreskrifter Och Allmänna råd. Accessed 2022-03-09 via https://rinfo.boverket.se/ BBR/PDF/BFS2020-4-BBR-29.pdf CIBSE, 2006. Degree-days: Theory and Application. TM41: 2006. The Chartered Institution of Building Services Engineers. London. D’Amico, A., Ciulla, G., Panno, D., & Ferrari, S. (2019). Building Energy Demand Assessment Through Heating Degree Days: The Importance of a Climatic Dataset. Applied Energy, 242, 1285–1306. https://doi.org/10.1016/ j.apenergy.2019.03.167 Department of Energy (DOE). 2022a. EnergyPlus. Accessed 2022-03-07 via https://energyplus.net/ Department of Energy (DOE). 2022b. Weather Data. Accessed 2022-03-08 via https://energyplus.net/weather Equa. 2022. IDA Indoor Climate and Energy.Accessed 202203-07 via https://www.equa.se/en/ida-ice. Hagentoft C-E, Sandin K (2017). Byggnadsfysik – så fungerar hus. Studentlitteratur, Lund, Sweden. 279 p. Neuman L. (2015). Handbok I energieffektivisering, Del 11, Uppvärmning, Förplanering av gårdsvärmeanläggning.
318
LRF, Europeiska jordbruksfonden för landsbygdsutveckling: Europa investerar i landbygdsområden. Official Journal of the European Union. (2012). Directive 2012/27/EU of the european parliament and of the council of 25 October 2012 on energy efficiency, amending directives 2009/125/EC and 2010/30/EU and repealing directives 2004/8/EC and 2006/32/EC. Sacks, R., Eastman, C., Lee, G., & Teicholz, P. (2018). BIM Handbook: A Guide to Building Information Modeling for Owners, Designers, Engineers, Contractors, and Facility Managers (3rd ed.). Wiley. Samset, K., & Volden, G. H. (2016). Front-end Definition of Projects: Ten Paradoxes and Some Reflections Regarding Project Management and Project Governance. International Journal of Project Management, 34(2), 297–313. https://doi.org/10.1016/j.ijproman.2015.01.014 Solemma. 2022. ClimateStudio. Accessed 2022-03-07 via https://www.solemma.com/climatestudio Stendahl M.-F., Dubois M.-C., Forgues D., Hjelseth E. 2022. Building Information Modeling for Environmental Impact Assessment in Early Design Phases: A Literature Review. Open Journal of Applied Sciences, 2022, 12, 59–81. https://doi.org/10.4236/ojapps.2022. 121006. Strachey Lt-Gen. Sir Richard (1878) ‘Paper on the Computation and Quantity of Heat in Excess of A Fixed Base Temperature Received at Any Place During the Course of the Year To Supply A Standard Comparison with the Progress of Vegetation’Quarterly Weather Report Appendix II Swedish and European Standard. 2000. Thermal Performance of Buildings – Determination of Air Permeability of Buildings – Fan Pressurization Method (ISO 9972:1996, modified). SS-EN 13829. https://www.sis.se/ api/document/preview/29541/. Swedish and European Standard. 2020. and SS 24300-1. Byggnaders energiprestanda — Del 1: Klassning av värmeeffektbehov. Accessed 2022-03-09 via https://www.sis.se/ produkter/byggnadsmaterial-och-byggnader/bygginstallationer/ovrigt/ss-24300-12020/. Verbai, Z., Lakatos, K., & Kalmár, F. (2014). Prediction of Energy Demand for Heating of Residential Buildings Using Variable Degree Day. Energy, 76, 780–787. https:// doi.org/10.1016/j.energy.2014.08.075 Wing, E. (2019). Revit 2020 for Architecture: No Experience Required (2nd ed.). Sybex. pp. 735-778 https://doi.org/ 10.1002/9781119560067. World Energy Council. (2013). World energy scenarios. Composing energy futures to 2050.
Machine learning
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Predicting semantic building information (BIM) with Recurrent Neural Networks B. Mete, J. Bielski, C. Langenhan & F. Petzold Technical University of Munich, Germany
V. Eisenstadt & K.D. Althoff German Research Center for Artificial Intelligence (DFKI)/University of Hildesheim, Hildesheim, Germany
ABSTRACT: Recent advances in technology established artificial intelligence (AI) as a crucial domain of computer science for industry, research and everyday life. Even though computer-aided architectural design (CAAD) and digital semantic building models (BIM) are essential aspects of the contemporary architectural design process, the acquisition of proper data proves challenging and AI methods are absent in established design software. An option to acquire rich data are design protocol studies sequenced through meaningful relations. However, this data requires a framework for pre-processing and training artificial neural networks (ANN). In this paper, we present our research on BIM and AI for autocompletion through suggesting further design steps to improve the design process of the early design stages, based on the methods of the ‘metis’ projects. We propose a recurrent neural network (RNN) model to predict future design phases through sequential learning of cognitive sequences, utilising enriched sketch protocol data.
1 INTRODUCTION Within the last decades, the scientific field of artificial intelligence has been a rapidly growing field of research. Its methods have been further applied in various capacities and varieties in other areas, both for professional operators, e.g., recognition of abnormalities in X-Ray images, and casual everyday users, e.g., text auto-completion of digital keyboards on smartphones. AI methods have also been applied within the architecture, engineering and construction (AEC) industry, mostly focusing on optimization during the later stages of a construction process, e.g., financial, temporal and performance (Abioye et al. 2021). The architectural design process, especially the early stages, are rather untouched because of its complexity. The architectural design process contains a lot of meaningful information of geometrical and semantic nature. Semantic building models (BIM) predominantly formalize the result of a design process. In contrast, we are tracking the steps of the design process that lead to the final design result. In this paper, we propose a novel Deep Learning (DL) approach as a design auto-completion method, to assist the architect during the design decision making process. The overall goal of the methods developed within the ‘metis’ projects is to predict the current and the future design phase, based on design process segmentation using design phases by Laseau (2000), Lawson (2005) and Barelkowski (2013), which is an
DOI 10.1201/9781003354222-41
extended version of the Analysis, Synthesis and Evaluation (ASE) model. Accessing the most recent changes in the design phase and the current status of the design phase, made tangible through the art of hand-drawn sketching (Lawson 2004; Suwa & Tversky 1997), the model is able to suggest possible continuations of the design direction. Furthermore, we aim to build an auto-completion pipeline that operates in real-time and suggest design steps to the architect, in which the user has an option to either reject or accept the proposition.
2 BACKGROUND AND RELATED WORK AI assists humans in various domains of both professional and daily life. Having an intelligent assistant that will aid the architect during designing, is first propounded by Negroponte (1973), which can both predict and suggest new ideas based on architectural design knowledge. First immediate obstacle for such approaches is to collect reproducible data. In order to overcome this drawback, sketch protocol studies have been used to trace the architect’s way of thinking during a design process (Suwa & Tversky 1997). As previous studies only produce qualitative data, a prototype tool was implemented and employed. It enables the development of quantitative results from retrospective sketch protocol studies, including custom categories for manual assigning (Bielski et al. 2022). Thus, both
321
custom and common parameters, e.g. timing and pen pressure on a digital drawing board, can be included for analysing the design process through the art of sketching. Furthermore, in order to process such information in the context of Machine Learning (ML), the data should be quantifiable, and categorizable. Lawson (2004) proposes both temporal and relational segments for classifying sketch protocols, while introducing the subclasses Analysis, Synthesis, Evaluation and Communication for defining the design phases. In addition to that, Laseau (2000) further partitions the Synthesis into two different subclasses as the Exploration and Discovery, while Barelkowski (2013) focuses on dividing the Analysis into Knowing and Understanding for a more distinguished look on the involved knowledge management. This also results in the separation of the Evaluation as a final decision, as well as a tool for creating more information as Evaluation - (informing) Knowing. Hence, the temporal categorization of the design decision making process can be enabled through the relational sequences of design phases. The state-of-art solution for such a categorization problem, is to make use of Artificial Neural Networks (ANNs). Recurrent Neural Networks (RNNs) are a subset of ANNs that include loops, hence it considers several previous input values, while calculating the output. Therefore, the knowledge can persist in the network, allowing the network to come up with predictions within sequential data, such as time sequences. One example of such a learning problem in real life can be predicting stock prices by looking at the previous and the current stock values, or predicting the weather in the following days by having access to the recent weather forecast. In that regard, they distinguish themselves from the ANNs or basic feed-forward neural networks through integrating loop connections in order to include data from the past. However, since neural networks rely on back-propagation that utilizes partial derivatives, having a looped architecture with a long chain, can cause the gradients for the learning weights to either drastically increase, or shrink to 0. This phenomenon is called Vanishing Gradient Problem. Moreover, being vulnerable to the Vanishing Gradient Problem, RNNs are prone to failures while capturing the long-range correlations of sequential data (Hochreiter 1998). There are several neural network architectures, which are subsets of the RNNs that is able to overcome the Vanishing Gradient Problem, such as Gated Recurrent Units (GRU) (Cho et al. 2014) and LongShort Term Memory (Hochreiter 1997). However, LSTMs are much more widely used in the state-of-art networks, which facilitates the development and the maintenance of the project for further improvements. LSTMs include specific gated cells, illustrated in the Figure 1, that allow to store and/or remove parts of the previous information, which enables the model to improve the handling of the long-term dependencies
within the data. Hence, LSTM architecture provides a more robust learning scheme for sequential data.
Figure 1. A vanilla LSTM cell that includes three different kinds of gates (Van Houdt et al. 2020).
3 APPROACH In this section, we present the approach for our supervised learning pipeline. It includes the dataset, the pre-processing and augmentation of this data, the details of the proposed RNN architecture based on the cascaded sequential learning method and finally a learning criterion, resulting in an implemented custom loss function. 3.1 Dataset The dataset consists of five different design processes made by architects, quantified through our open-source sketch protocol analyser tool (Bielski et al. 2022). Each design process data consists of a feature vector, a design phase, and a specific timestamp. All data instances have varying numbers of timestamps, ranging between 4.000 and 18.000 which spans across 15 minutes. Each data instance that corresponds to a specific timestamp includes a feature vector, and the design phase attached to it, which we will refer to as the label. The labels are unique, and have a value among our seven design phases (i.e. Analysis-Knowing, AnalysisUnderstanding, Synthesis-Exploration, SynthesisDiscovery, Evaluation, Evaluation- Knowing, Communication) that represent a more distinguished version of the phases (see Figure 2) of the common design model ASE (Analysis, Synthesis, Evaluation (Lawson 2005). The feature vector consists of distinct information related to the design process for each timestamp. Namely, these information parameters are the pen pressure and geometric coordinates of the pen, gathered from a WACOM tablet used as a digital drawing board during the sketch protocol study, and the sketched elements (e.g. ‘symbol’, ‘line’) and objects (e.g. ‘door’, ‘wall’) that are present in the sketch that are present in the sketch at the respective timestamp. An important step within our approach is the data pre-processing. Even though the data is quantified through the sketch protocol analyzer, there are still categorical values both in the feature vector and the
322
Figure 3. An exemplary representation of the feature mapping. The values on the left shows the unprocessed features with both numeric and non-numeric values. The figure on the right, is the processed and the final version of the feature vector. Figure 2. The design process as an extended version of the ASE model (Lawson 2005).
labels that have to be encoded as numeric values in order for the Artificial Neural Network (ANN) to work with the data. As mentioned before, there are seven design phases (see above) for the labels, while each timestamp is being labeled by only one of them. In order to map that textual information, we are using a categorical data embedding technique called ‘Dummy Variables’(Draper & Smith 1998). In our case, dummy variables create a vector with a length of 7 units, where each unique design phase category is attached to a unique vector element for every timestamp. The resulting vector is a unit vector, where only the value that corresponds to the current design phase is 1, whilst all the remaining vector elements stay as 0. Therefore, after the embedding, instead of having categorical values, the design phase is represented with seven unique unit vectors. For encoding the feature vector, a similar approach is being used, called ‘Multi-Hot-Encoding’, which is a generalized version of dummy variables. This approach is being used, since each categorical feature can appear more than once in the data, therefore it requires an integer variable rather than a binary variable. The different features have varying ranges for their values, which can cause the network to be influenced more by the numerically large values. To be specific, while the encoded features have small integer values, continuous variables like the pen pressure can take values up to a million, which might prioritize the pen pressure value during the learning process. Thus, a final normalization routine must be performed. For this purpose, the last part of the pre-processing step is the L1 normalization, which maps the values of every feature between 0 and 1. This results in changing the feature dataset into a common scale, without deforming the numeric relation between the values since it is a linear map. An exemplary conversion between raw data and the processed features is shown in Figure 3.
3.2 Model As explained in the section 3.1 the training data includes features that are extracted from the design process data, and each timestamp is labelled with a unique design phase. Our model proposes a sequential prediction scheme, using a fixed number of timestamps as an input, and aims to predict the current and the next design phases as the output. In order to overcome the shortcomings of common RNNs, our cascaded model consists of a chain of “processing blocks”, which consists of an LSTM cell, followed by a fully connected layer and a Sigmoid activation function. Each processing block coincides with the features in one timestamp, and they produce an output for their corresponding timestamps. The overall model comprises a chain of processing blocks, and the length of this chain is defined with the parameter “processing window size”. The processing window size parameter is crucial, since we require a long enough chain that the model can capture the correlation between the features well, to predict the next design phase, but short enough that the training is tractable and feasible. The proposed processing window size value of this work is 50. The LSTM cells accept a total number of 41 features and produce 21 output values. The fully connected layers accept 21 features and produce 7 output values. Those 7 output values are indicated with the parameter output size, and it refers to the probabilities for the predictions of all 7design phase labels. Simply, the largest probability value is selected to be the model’s prediction. The fully connected layer is added on top of an LSTM layer is to even capture extra correlations that the LSTM model itself fails to detect. The overall scheme of the learning model, consisting of the chain of ‘processing blocks’ (see Figure 4). 3.3 Learning criteria Each individual data set per sketch data contains more than 10000 timestamps, but only a handful of design phase changes. The biggest learning obstacle is to not overfit the model through the exceeding timestamps with rare to no changes. Overfitting is a common
323
The custom loss function we propose, can be seen in the Equation 2. In the equation P denotes the set that includes the intervals, in which there is a design phase change. Therefore, our loss function, calculates the loss just like the BCE, if the processed batch is not in an interval where there is a design phase change, but it penalizes the term with another parameter when the batch is in an interval with a design phase change.
4 EVALUATION
Figure 4. The model as a chain of processing blocks: Blue shapes represent the feature vector for each timestamp fed into the LSTM cells, yellow blocks represent an individual LSTM cell, green blocks the fully connected layers, white circle the Sigmoid activation function, and finally the purple rectangles the probabilities attached for each design phase for each timestamp, while numerically the largest value eventually becomes the prediction of the model, and w represents the processing window length.
problem in ML (Ying 2019), which arises when the network specifically learns the training data instead of the solution to a general and much wider problem. Along with achieving low accuracy during the evaluation, an overfitted network can be thought of as a memorizing model, instead of a learning model, therefore it should be avoided for a general ML problem. Instead, the most important functionality of the model is to be able to capture the internal dynamics of the periods where there is a design phase with better accuracy. Thus, in order to mitigate and optimize the learning process, we implemented a loss function as a learning constraint. The reason for this is, since a large portion of the design process data continues to stay at the same design phase for a long period of time, and only at several instances, a design phase change can be observed. This fact makes the intervals that have a design phase change more crucial in the learning stage, since the architect’s thought process is most likely to stay the same, when the architect is still in the same design phase. Therefore, instead of using the binary cross-entropy (BCE) which can be seen in Equation 1, we are proposing a custom loss function that augments the BCE loss function. Our loss function penalizes the input sequences that include a phase change, with a large penalty term that can be hyper-tuned. BCE = −
N yi log yˆ
(1)
i
⎧ N ⎪ ⎪ yi log yˆ , if w ∈ /P ⎨− w i CLF = N ⎪ ⎪ ⎩λ ∗ − yi log yˆ , if w ∈ P w
i
(2)
The model has been implemented using the TensorFlow (Abadi et al. 2016) framework and trained the final model with 10 epochs, and 3.500 steps per epoch. The optimizer being used is the Adam (Kingma & Ba 2014) with default TensorFlow learning rate of 0.001. The convergence of the model can be seen from the loss graph in Figure 5.
Figure 5. The loss function with 10 epochs.
Having access to a limited number of design process data requires a pertinent evaluation method. We have selected the k-fold cross validation for our training and evaluation subroutine, where the data is split into training, validation and test data. Furthermore, there is no established or best practices evaluation method for temporal data, since the whole sequence is needed for the network to learn the pattern across the data. Hence, splitting the data instance into two sub-arrays as train and test sets can cause a significant loss of information. In order to remedy the shortcoming of splitting the data into train and test sets for temporal data, there are several techniques proposed, for example successively enlarging both the train and the test data, across different epochs (Cerqueira et al. 2020) However, due to the limited amount of data, we applied another evaluation method, in which the design processes from different architects are used for both training and evaluating the data without separating them into training and test sketches. Several time intervals, that include a design phase change, are selected from these sketches as evaluation intervals. Hence, in the evaluation, we examine if the model can capture the pattern, since preferably the crucial time intervals are used.
324
Our evaluation method consists of creating various numbers of intervals in different sketches, which will then be separated arbitrarily as training and test sets. The important difference is that the intervals are selected among the time sequences which include a design phase change. That way it can be examined whether the model can capture the required pattern or not, since preferably the crucial time intervals are used while calculating the accuracy, as it was explained in Section 3.3 Using the described approach, the accuracy is calculated additionally to the prediction results of the design phases for the entire sketch data.
to Synthesis-Discovery, and Synthesis-Exploration to Evaluation, the results show a high accuracy for the prediction of both the current and following design phase. By implementing a custom loss function, which emphasizes the use of temporal intervals for both training and evaluating, we improved the accuracy for both types of predictions by 6-8 percent.
Figure 7. Comparison of the evaluation results, using the BCE loss function, and our proposed custom loss function. Figure 6. Exemplary evaluation graph: orange lines indicate the prediction values for the whole sketch data, while blue lines represent the ground truth values for the same sketch.
Figure 6 illustrates the prediction values through two lines for all the data within one sketch. While the blue line represents the ground truth values for the respective sketch that indicate the actual design phases of the architect’s design process, the orange line shows the design phase predictions of our model. In this categorical line graph, the y-axis depicts possible design phases, while the x-axis shows the temporal progress with timestamps. As it can be seen in Figure 6, the model predicts the design phases with relatively high accuracy, i.e., 94%. Only a few errors occurred due to the similar patterns in transitioning between design phases. For instance, one repeating error was the model’s inability to capture the design phase SynthesisDiscovery and instead it mistook said phase as the Evaluation. This shows that the transition between the Synthesis-Exploration and the Synthesis-Discovery has similar dynamics to the transition between the Synthesis-Exploration and the Evaluation. Finally, Figure 7 highlights the effect of our custom loss function on evaluation and training success. For both testing and training, our custom function has improved the accuracy of the predictions significantly, compared to the models that have been trained with the same characteristics, but without the BCE loss function. Therefore, the deep learning method, along with the custom loss function, achieves successful results for predicting the current and next design phase of a given design process. To sum up, even though, the model continues to occasionally mistake design phase changes due to similar patterns, specifically Synthesis-Exploration
5 CONCLUSION AND FUTURE WORK The results of our model for both prediction and evaluation are visibly accurate with a percentage of 94 at the end. The implementation of the custom loss function improved the accuracy by 6–8 percent, compared to accuracy of the models trained without the BCE loss function. Thus, we contribute a successful approach for predicting the current and next design phase, based on the categorisation by Laseau (2000), Lawson (2004; 2005) and Barelkowski (2013), using an RNN trained with quantified design process data, to the research field. This novel approach includes the workflow for pre-processing of the design process data, quantified with the sketch analyser tool, the LSTM model architecture and finally, interventions to improve accuracy. Consequently, this novel approach is transferable for predicting custom temporal parameters of various nature of design process data, e.g., design intentions (Lawson 2004), assigned as custom labels within the protocol analyser tool. However, the low number of design process data remains a major limitation for training the model, but far and foremost for evaluating the model’s behaviour for projecting more general results and outlook. Since recruiting a large number of participants and preparing the dataset with manual labelling proves to be too resource-inefficient and cumbersome, Generative Adversarial Networks (GANs) (Goodfellow et al. 2014) can be employed in the future. GANs are generative models and, trained with enough information, can be used to create novel and virtual data. That way, the
325
data retrieval can be automated and fasten the process of dataset preparation. Finally, the ‘metis’ projects aim to ultimately suggest the next design step to the user (e.g., ‘outlining parcel’) during the sketching process to support the architectural design decision making. Thus, we plan to extend the current approach for predicting and suggesting new design phases for further values, such as design intentions and design steps. The individual RNN models for each value type will be connected in a cascading series from the largest segmentation, the design phases, to the smallest, the design step. ACKNOWLEDGMENTS We want to thank the DFG for funding the ‘metis’ projects, as well as the study participants for offering their insight, feedback and design knowledge, as well as their valuable time and sketches. REFERENCES Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean, J., ... & Zheng, X. 2016. {TensorFlow}:A System for {LargeScale} Machine Learning. In 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16) (pp. 265–283). Abioye, S. O., Oyedele, L. O., Akanbi, L., Ajayi, A., Delgado, J. M. D., Bilal, M., & Ahmed, A. 2021. Artificial Intelligence in The Construction Industry: A Review of Present Status, Opportunities and Future Challenges. Journal of Building Engineering, 44, 103299. Barelkowski, R. 2013. Designing More By Knowing Less, Verbeke, J., Pak, B., In Proceedings of the Conference ’Knowing (by) Designing’ at LUCA, Sint-Lucas School of Architecture Brussels, 22–23 May 2013 (pp. 522–531). Ghent, Brussels. Bielski, J., Langenhan, C., Ziegler, C., Eisenstadt, V., Dengel, A., and Althoff, K.D. 2022. Quantifying The Intangible – A Tool For Retrospective Protocol Studies of Sketching During The Early Conceptual De-sign of Architecture. In International Conference of the Association for
Computer-Aided Architectural Design Research in Asia (pp. 403–411). Association for Computer-Aided Architectural Design Research in Asia. Cerqueira, V., Torgo, L., & Mozetiˇc, I. 2020. Evaluating Time Series Forecasting Models: An Empirical Study on Performance Estimation Methods. Machine Learning, 109(11), 1997–2028. Cho, K., Van Merriënboer, B., Bahdanau, D., & Bengio, Y. 2014. On the Properties of Neural Machine Translation: Encoder-decoder Approaches. arXiv preprint arXiv:1409.1259. Draper, N. R., & Smith, H. 1998. “Dummy” Variables. Applied Regression Analysis, 299–325. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., WardeFarley, D., Ozair, S., & Bengio,Y. 2014. GenerativeAdversarial Nets. Advances in Neural Information Processing Systems, 27. Hochreiter, S. 1998. The Vanishing Gradient Problem During Learning Recurrent Neural Nets and Problem Solutions. International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, 6(02), 107–116. Hochreiter, S., & Schmidhuber, J. 1997. Long Short-term Memory. Neural Computation, 9(8), 1735–1780. Kingma, D. P., & Ba, J. 2014. Adam: A Method For Stochastic Optimization. arXiv preprint arXiv:1412.6980. Laseau, P. 2000. GraphicThinking ForArchitects and Designers. John Wiley & Sons. Lawson, B. 2004. What Designers Know. Boston, MA: Elsevier/Architectural Press. Lawson, B. 2005. How Designers Think. 4th edition, Routledge. ISBN 9780080454979. Nicholas Negroponte. The Architecture Machine: Toward a More Human Environment. The MIT Press, Jan. 1973. isbn: 9780262368063. doi: 10.7551/mitpress/8269. 001.0001. url: https://doi.org/10.7551/mitpress/8269. 001.0001. Suwa, M and Tversky, B 1997, ’What Do Architects and Students Perceive in Their Design Sketches? A Protocol Analysis’, Design Studies, 18(4), pp. 385–403. Van Houdt, G., Mosquera, C., & Nápoles, G. 2020. A review on The Long Short-term Memory Model. Artificial Intelligence Review, 53(8), 5929-5955. Ying, X. 2019, February. An Overview of Overfitting and Its Solutions. In Journal of Physics: Conference Series (Vol. 1168, No. 2, p. 022022). IOP Publishing.
326
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
International assessment of artificial intelligence applications in the AEC sector D. Cisterna Karlsruhe Institute of Technology, Karlsruhe, Germany
C. Lagos Pontificia Universidad Católica de Chile, Santiago, Chile
S. Haghsheno Karlsruhe Institute of Technology, Karlsruhe, Germany
ABSTRACT: The construction industry suffers from productivity stagnation and slow digitalization. Its digital landscape is characterized by substantial data heterogeneity and project-specific data silos. To address these challenges, a growing number of tech-companies are developing applications based on Artificial Intelligence (AI). AI can analyze complex data, automate tasks and support decision making. This study aims to identify gaps and opportunities in the industry’s adoption of this technology. A set of 16 variables was used to characterize 236 tech-companies offering AI solutions for Architecture, Engineering and Construction (AEC). Their potential impact was measured through their number of social networks’ followers and latest annual revenue. Several statistical analyses were used to determine company and software characteristics that influenced adoption. The results showed dependency between technology penetration and aspects such as companies’ location, lifetime, size, business model and type of technology. Similarly, the data show undiscovered subjects, which may represent new potential for innovation.
1 INTRODUCTION
2 BACKGROUND
Construction represents approximately 15% of global GDP (Mazhar & Arain, 2015); prior to the COVID-19 pandemic, it had grown to a spending value of about USD$12 trillion and is predicted to continue growing at a 3% annual rate in the coming years (de-Best, 2021). Given the productivity stagnation observed during the last decades in the AEC industry (Momade et al. 2021), transitioning to construction 4.0 is believed to be critical to break the stigma of low productivity (Chui & Mischke, 2019). Artificial Intelligence (AI) is a particularly relevant disruptive technology for AEC, especially given the wide range of opportunities it presents such as automation of processes, improved decision-making and pattern recognition in data, among others (Amann & Stachowicz-Stanusch, 2020). Although AEC adoption of AI has been slow, an increasing number of consolidated companies and start-ups have started developing solutions based on this technology (Darko et al. 2020). Hence, this research focuses on determining the potential of AI adoption in AEC through an in-depth assessment of the types of solutions offered, the characteristics of the companies offering them and their current penetration, to identify gaps and opportunities for new market players.
In recent decades, the development of AI-based applications has accelerated noticeably. AI is becoming ingrained in our daily lives and is widely regarded as the primary force driving digitization (Sousa et al. 2021). This technology uses algorithms that can learn and improve automatically (Kreutzer & Sirrenberg, 2019). To do this, machines or computers must be trained using large amounts of high-quality data which can also be combined with experience to develop expert systems, decision making aid and automate processes (Buxmann & Schmidt, 2019). AI is divided into subfields, including natural language processing, robotics, computer vision, optimization, and automated planning and scheduling (Rao et al. 2021). Other industries such as manufacturing, healthcare, entertainment, and financing have exhibited a rapid adoption curve, while AEC adoption remains limited mostly to study cases. Multiple studies have covered potential benefits and barriers for AI adoption in AEC (Cisterna et al. 2022), mostly signaling that the potential impact of advanced data processing, the development of predictive and prospective algorithms, in addition to the automation of repetitive tasks has been constrained due to the fragmented nature of the industry (Regona et al. 2022).
DOI 10.1201/9781003354222-42
327
Particularly, training useful AI models requires generalization obtained from multiple samples with large number of data, to avoid overfitting solutions for a specific type of project, use, context, or company (Fischer et al. 2021). This has been a complex challenge for construction, where projects, context and challenges vary significantly depending on the scope, country, company characteristics and moment of execution (Ramli et al. 2018). Also, AEC companies have only recently embraced digital systems to support key processes such as managing planning, control, quality, productivity and communication, elements which significantly improve the capability to capture and use high volumes of quality data (Bosch-Sijtsema et al. 2021). However, the last decade has seen a significant increase in the number of companies providing AI solutions, as well as in the range of solutions and uses available. (Abioye et al. 2021). Since most related studies have focused on potential uses, developments, barriers and case studies (Darko et al. 2020), the authors chose to approach the subject from the suppliers’ perspective, by understanding who is investing in providing AI solutions, the characteristics of the main players in the field, differences between these players and newcomers as well as gaps and opportunities for the penetration and adoption of new AI uses in AEC. Therefore, the general aim of understanding AI penetration in AEC is divided into (i) Understanding the characteristics of AI company providers and solutions and (ii) Identifying gaps and industry tendencies where newcomers can take advantage of its growth potential.
3 METHODOLOGICAL APPROACH This research is based on an in-depth review of public domain information about existing and emerging AI solutions offered by companies across the world, that already have websites, landing pages or social network accounts accessible by Google’s search engine. A set of company and solution characteristics, as well as penetration data, such as reported revenue and number of social network followers from the companies were used to gain such understanding. Each company was characterized by five nominal variables and their available AI solutions were characterized by two nominal and nine binary variables. The research techniques used to address the objectives combine discrimination analyses and descriptive statistics to identify large well stablished suppliers as well as emerging medium or small sized players, determining if factors such as the region, lifetime and size of a company influence their starting position and potential penetration of their AI solutions as well as determining if well stablished and emerging companies exhibit different business models, types of solutions or focus on different uses. These comparisons allowed to determine key trends, gaps, and opportunities for AI providers.
3.1 Research sample A thorough review of public domain information was carried out to capture data from existing or emerging companies offering AI solutions for the AEC industry using Google’s search engine. The following search query was used to find over 300 potential companies: {“architecture” or “construction” or “engineering”} AND {“artificial intelligence” or “data science” or “machine learning”} AND {“technology” or “product” or “system” or “solution” or “software”}. Only companies and solutions with existing websites were included in the selected list. Out of the solutions found, 255 had established websites, 19 were at concept or initial pilot studies stages, hence, 236 companies exhibited commercially active solutions.
3.2 Assessment of company penetration The penetration of each company was characterized using two variables, the maximum number of followers exhibited within Facebook, Instagram, LinkedIn, Twitter, and YouTube (Max Followers), and their latest overall annual revenue (Yearly revenue) reported in the business information platform Crunchbase, within 2020 and 2021. 227 companies had at least one active social network account to allow to assess penetration using the maximum number of followers, and 164 companies had public information regarding their annual revenue. Revenues are a direct indicator of a company success, as they are directly tied to the activity ratio or turnover ratios that reflect asset management success (Kaniški & Vincek, 2018). Nevertheless, since not all companies are obliged to publish information of yearly revenue, a complimentary indirect measure of penetration is needed. An indirect way to measure a company’s market penetration is through its social media metrics. The global social media penetration rate reached 49% in 2020, with the highest rates in East Asia and North America. Today, brands are expected to have a social media presence. Social activities and information exchanged inside a company’s target market help comprehend the industry’s evolution. Real-time business pattern disclosure is a potential goldmine for business intelligence. (Murshed, 2020) Several studies have been undertaken to determine the association between social media indicators and customer volume and market success. Results indicate that social media followers perceive higher levels of relationship with the company, report higher service quality, satisfaction and increased perception of value which, in return, increases loyalty (Clark & Melancon, 2013). Furthermore, empirical studies show that after surpassing a critical mass of followers, a positive correlation between share-value and number of social media followers is exhibited in a wide range of companies and value propositions (Paniagua & Sapena, 2014).
328
3.3 Company characterization The companies’ categorical attributes were (1) region, (2) lifetime, (3) number of employees, (4) size classification and (5) business model. First, given that countries such as the United States presented a significantly higher number of companies and solutions compared to emerging regions, the companies were grouped by region into: North America, which includedThe United States and Canada; Europe, which included countries part of the European Union; and Other Regions, which mainly contained solutions from emerging markets such as Africa, Asia-Pacific and Latin America. Second, the company lifetime was classified into five groups, according to the foundation year of the company, obtained from public records. The first group accounted for newcomers that started operations during the pandemic, hence, with three or less years of existence. On the opposite end, companies that existed prior to the 2000’s, hence with at least22 years of existence, were classified as well-stablished companies not native to the 4.0 digital era. The middle groups were separated into companies with 3 to 6, 7 to 10 or 10 to 22 years of existence. Third, company size and number of employees were classified according to the European Commission’s classification for micro, small, medium, and largesized enterprises (SME) (see Table 1), based on the yearly turnover and number of employees. Therefore, four employee groups and four company size groups were obtained. Table 1. Company size classification according to the European Commission (2003). IfM Bonn.
and the core logic of the firms were classified through the business models described in the St. Gallen Business Model Navigator (Hoffmann et al., 2016) and then grouped into categories according to their nature as cloud service (Sowmya et al. 2014) or on-premise software, as shown in Table 2. Table 2.
Business model categorization.
Business Model (BM)
Gallen Business Model St. Navigator definition
Software as a service
Software applications are offered via Internet (e.g., project management software) E-commerce Services are offered via online channels (e.g., business-to-business platforms) Digitalization Existing products or services are offered in a digital variant (e.g., smart home solutions). Sensor as a Services enabled by using Service sensors (e.g., sensor-based data collection) Virtualization Imitation of a traditionally physical process in a virtual environment (e.g., VR/AR for construction support) Solution Total solution of provider integrated product and service offerings.
BM Group Software as a service (SaaS) 116] Platform as a service (PaaS) [34] Hardware as a Service (HaaS) [56]
On-Premise Software (OPS) [30]
[#] Number of companies.
3.4 AI solution characterization
Micro enterprises Small enterprises (excluding micro enterprises) Medium-sized enterprises (excluding micro and small enterprises) No SMEs
Several scholars concur that the business model definition is not standardized (Jensen, 2014) and has often been misused (DaSilva & Trkman, 2012). In this study, Shafer et al. (2005)’s definition will be used: “A business model is a representation of a firm’s underlying core logic and strategic choices for creating and capturing value within a value network”. The strategies
AI solutions were characterized using the information provided by the companies’ web pages and social networks. This allowed to categorize them based on descriptive technology and use-case information, as well as to determine whether specific attributes are present, if it was supported by available information, absent, if it was rejected by available information or unconclusive, if available information did not validate or reject its presence. The solutions were categorized nominally by (1) the construction phase supported by the technology and (2) the type of AI applied to it. First, the construction phases were categorized into: Planning & Design; Construction execution; Operations & Maintenance; or Multiple phases. The latest category was reserved exclusively for general AI solutions that were not targeted to a specific AEC phase. Second, as Table 3 presents, AI uses were categorized into four groups: (1) Unstructured data interpretation containing AI solutions such as natural language processing and
329
Table 3. AI solution categorization. Example of applications
Object detection
Quality issues detection Labor productivity tracking Physical document digitalization
Processing unstructured data from text, images and video to detect, classify and interpret objects and context
Processing of scalar and nominal variables in a structured form such as tables to generate metrics, classification, regression, prediction, or interpretation
Benchmarking Statistical differentiation of samples Regression Schedule and budget models forecasting Classification models
Predictive equipment maintenance
Use of previously trained models such as neural networks to facilitate processes through prescription
Heuristic optimization
Inventory and procurement optimization
Probabilistic calculations
Risk/ hazards estimation and warnings
Expert systems
Use of AI to facilitate decision making
Generative design
Design optimization and 3D printing
Process automation [61]
Aided decision making [56]
Unstructured data interpretation [61]
Examples of AI uses
Structured data augmentation [50]
AI Group Description
Automation of repetitive tasks either physical, administrative or at the decision-making level
Work & trade recognition Natural language processing
3.5 Statistical analyses The statistical tests were divided into two stages. First, a set of discrimination analyses was used to determine which company and solution characteristics served to explain differences in their penetration, to determine which were already in a well-stablished position to offer value and which were emerging players. This stage also allowed to determine differences in revenue gained by different types of solutions, business models and AI uses. The relationship between any attribute and the penetration was measured through a set of statistical analyses (see Table 4), selected based on the type of dependent and independent variables assessed and the distribution of the samples (Kanji, 2006). Nevertheless, in all the cases, a p-value was calculated with 90%, 95% and 99% confidence level to evaluate the null hypothesis h0 “subsamples obtained from categorical separation using the independent variable exhibit the same distribution”, hence, if the p-value was low enough to reject h0, the use of the independent variable allows to explain differences in the dependent variable, thus, allowing to gain information of key factors affecting penetration.
Digital process Inventory assessment automation and order placement Sensorized machinery automation
related to data science, which employs machine learning to obtain new information based on structured and unstructured data; (3) Aided decision making that includes AI solutions based on trained expert systems and models that can deduce, predict and discern from Big Data to trigger alerts, propose recipes or recommend decisions; and (4) Process automation, which includes AI solutions like robotics and generative design that integrate parts of the preceding groupings to deliver solutions that perform physical, conceptual or managerial operations automatically Finally, the binary technological characteristics captured were: (1) Incorporating web or computer apps, (2) mobile apps (tablet or smartphone optimized), (3) availability of Business Intelligence dashboards, and the requirement of special hardware such as (4) cameras, (5) drones, (6) Wi-Fi or other IoT communication networks, (7) sensors and scanners, (8) robots and (9) virtual or augmented reality equipment.
Table 4. sample.
Automated machinery operation
Statistical tests used to determine key factors in
Dependent Independent Samples’ variable variable distribution Test
[#] Number of companies.
computer vision, characterized by the use of unstructured information such as text, images and video to generate structured data via interpretation; (2) Structured data augmentation comprising AI solutions
330
Scale Scale
Binary Binary
Scale Scale
Nominal Nominal
Nominal
Nominal
Normal Nonparametric Normal Nonparametric /
t-test (T) Mann Whitney’s U (MW) ANOVA (AV) Kruskal-Wallis (KW) Chi-squared (CS)
The second stage consisted in a series of descriptive statistical analyses and interpretation to determine industry trends and gain understanding of the relative effects of company position, characteristics and AI solution characteristics in revenue, followers, and competition. Therefore, it allowed the detection of gaps and opportunities identified through the comparison of technologies, business models, AI use cases and well-stablished or emerging players.
Table 6.
Statistical significance of company characteristics.
Nominal variable
N◦ of
Followers (p-value)
Revenue (p-value)
Region N◦ of employees Lifetime Company size Business model Supported phase AI use
3 4 5 4 4 4 4
0 0 0 0 0.045* 0.008 0.673
0.063** 0 0 0 0.007 0.003 0.822
4 RESULTS First, regarding the assessment of new players and well stablished companies, the sample results showed that 57.6% of the companies had six or less years of existence and 42% of the sample belonged to the three to six years of existence lifespan category. Also, 67.8% of the sample represented small (42.4%) and micro (25.4%) companies, while 69.5% of the sample had less than 50 employees. These results support the inference that AI solutions for AEC have significantly grown over the past years, with a significant number of new players, nevertheless, the comparison by region shows that these new players are not equally distributed, as presented in Table 5.
*Statistically significant at 95% confidence level ** Statistically significant at 90% confidence level
Table 7. AI Solution comparison by followers and revenue. AI Use Category
Comp.
Followers
Revenue
Unstr. data interpretation Str. data augmentation Aided decision making Process automation
61 50 56 61
10285 26999 135607 667050
$497M $170M $816M $7464M
Table 8. Table 5.
Statistical significance of AI solutions’ attributes.
Results comparison by region.
Attribute
North Other America Europe regions
% of companies % of followers % of revenue Average followers Average revenue % lifetime < 6 years % big companies % small & micro companies
42.8% 90.7% 85.7% 456015 $3755M 50.5% 21.8% 57.3%
36.0% 8.7% 4.9% 52073 $269M 55.3% 17.7% 71.8%
AI solution characteristic
Followers (p-value)
Revenue (p-value)
Computer or web apps Mobile apps Need for cameras Need for drones Need for sensors or scanners Need for IoT networks
0.000 0.029* 0.013* 0.020* 0.009 0.029*
0.016* 0.016* 0.092** 0.034* 0.060** 0.016*
21.2% 0.6% 9.4% 6494 $1994M 76.0% 4.0% 82.0%
*Statistically significant at 95% confidence level **Statistically significant at 90% confidence level
5 DISCUSSION Second, regarding the assessment of the influence of company and AI solution characteristics on penetration, as Table 6 shows, all descriptive variables, except the AI use, explained differences in at least one of the penetration metrics (number of followers and revenue) with a 99% confidence level.Also, the Business Model explained followers’difference at 95% confidence and Region explained revenue at a 90% confidence level. With regard to AI use, the differences in revenue and followers’ distributions were not statistically significant between categories. Table 7 shows in detail how followers and revenue are distributed according to AI use. Finally, as Table 8 presents, out of the 9 binary AI solution attributes, 6 attributes explained differences in the number of followers and revenue with at least a 90% confidence level. Hence, the comparison of companies and solutions using these variables can allow to determine relevant trends and opportunities.
First, as observed in Table 5, despite only 42.8% of companies belonging to North America, this region concentrates the majority of the well stablished AI providers, as it represented 90.7% of the total followers captured and 85.7% of the reported revenue. Also, despite approximately two thirds of the sample corresponding to new players, only 50.5% of North American companies had six or less years of existence and 57.3% were small or micro companies. In comparison, 76% and 82% of companies from other regions had less than six years of existence or belonged to small and micro company classifications, respectively. These results show that despite North America concentrating most of the well stablished players with the biggest penetration, Europe and Other Regions have recently started to catch up with a relevant number of small-sized companies and young start-ups offering new AI solutions.
331
Second, as Table 9 presents, new and well stablished players seem to be investing in different types of solutions, as represented by their business model. Despite SaaS being the most common business model across all lifespans, it represented 40% in companies with over 22 years of existence against 60% in companies with less than three years. In contrast, On-premises software (OPS) and Platform as a Service (PaaS) solutions are significantly more relevant in companies with over 10 years and specially over 22 years. These results allowed to stablish two inferences: First, new players are significantly opting for new business models as SaaS and HaaS over traditional OPS solutions; second, older well stablished players seem to be transforming traditional OPS solutions into PaaS business models, which require bigger investments probably not achievable yet by new players. Table 9.
Business model composition by lifespan.
Lifespan
OPS
HaaS
SaaS
PaaS
Less than 3 years 3 to 6 years 7 to 10 years 10 to 22 years Over 22 years
16,2% 7,1% 10,2% 19,2% 28,0%
13,5% 34,3% 24,5% 15,4% 4,0%
59,5% 46,5% 55,1% 42,3% 40,0%
10,8% 12,1% 10,2% 23,1% 28,0%
Table 10, which presents the distribution of preferred business models by company size, shows that one-third (33%) of the 236 solutions assessed were SaaS provided by micro to small companies, twice the number of offerings from medium to big companies and equivalent to all the solutions provided by medium to big companies (33%). In the case of HaaS, the number of solutions offered by small and micro companies is four times higher than HaaS alternatives in medium to large alternatives, while the number of OPS and PaaS solutions is approximately equivalent across company size classifications. Table 10. Distribution between business model and company size (revenue). Business model
Micro
Small
Medium
Big
Total
OPS HaaS SaaS PaaS
3% 6% 14% 3%
5% 14% 19% 5%
1% 4% 8% 3%
4% 1% 8% 4%
13% 24% 49% 14%
companies had special hardware requirements, while approximately 65% of medium and big companies invested in AI solutions that did not require additional equipment. Table 11.
Hardware requirements across company sizes.
Company size
Does not require Hardware
Special hardware required
Micro Small Medium Big Total
38,33% 42,00% 62,16% 69,23% 48,73%
61,67% 58,00% 37,84% 30,77% 51,27%
Regarding the different AI uses, Table 6 shows that the category distributions did not present statistically significant differences regarding revenue and followers. This, combined with the similarities in the number of companies belonging to each category allowed to infer that there is no clear trend in the direction of AI applications within the AEC industry and all use cases from data augmentation to process automation are being equally offered. Nevertheless, as Table 7 presents, the average followers and revenue among AI use categories show that well stablished players appear to be investing more on aided decision making and process automation, while unstructured data processing and structured data augmentation seems occupied mainly by small new players. Given that these results are also consistent with the trend discussed before between SaaS, HaaS and PaaS solutions, the study allows to infer that while aided decision making and process automation solutions require significantly large training samples and investment to ensure confident prescriptions, the continuously decreasing costs of computational power and expanding available free source solutions and libraries allow smaller new players to develop AI structured and unstructured data processing solutions that can significantly increase the value proposition of their SaaS and HaaS offerings.
6 CONCLUSIONS
Table 11 presents in the comparison of the requirement of special hardware such as cameras, drones, scanners, sensors and IoT networks, among others. Despite these requirements being approximately equally distributed among the complete sample, it was observed that approximately 60% of micro and small
The use of AI to aid construction processes through data augmentation, interpretation, aided decision making, and automation has an enormous potential to improve AEC productivity. This research analyzed online publicly available information from 236 companies that offered AI powered solutions for needs such as hazard and risk estimation, project performance predictions, computer vision labor tracking, and machinery automation, among others. A set of seven nominal and nine binary variables allowed to characterize the companies and their solutions. Also,
332
each company’s social media followers and yearly revenue were used to discriminate well stablished and new players in the market. Statistical differentiation and descriptive analyses were carried out to determine trends, gaps and opportunities for AEC, allowing for four main conclusions. First, despite well stablished players being mainly concentrated in North America, Europe and other regions produced a significant number of new offerings in recent years, contributing with approximately 57% of the sample. Second, despite SaaS being the preferred business model among the sample, it is significantly more relevant in new players, while PaaS and OPS models seem concentrated in well stablished companies. Third, HaaS solutions have significantly increased in recent years and despite special hardware requirements representing half the sample, 60% of micro and small companies used special hardware for HaaS solutions, compared to less than 36% of medium and big providers. Finally, AI uses are equally distributed among companies, but new players prefer data augmentation through structured and unstructured processing models, while aided decision making and process automation solutions seem to attract more of the well-stablished companies. The study is strongly linked to innovation. This is an extremely dynamic field, as many new applications are created, and many others fail and disappear from the market, on a daily basis. So, the database can never be entirely updated, which is the fundamental limitation of this research. Future study could identify new data sources, update old ones, and create new statistical tests to link them. Secondly, language barriers may have biased and diminished the collection of data from certain regions, such as Asia. Therefore, these findings can serve as a complement to other similar research conducted in other contexts.
ACKNOWLEDGEMENTS This work was supported by the Federal Ministry for Economic Affairs and Climate Action within the framework of the research project SDaC: Smart Design and Construction (SDaC, 2020). REFERENCES Abioye, S. O., Oyedele, L. O., Akanbi, L., Ajayi, A., Davila Delgado, J. M., Bilal, M., Akinade, O. O., & Ahmed, A. 2021. Artificial Intelligence in the Construction Industry: A Review of Present Status, Opportunities and Future Challenges. Journal of Building Engineering, 44, 103299. https://doi.org/10.1016/j.jobe.2021.103299 Amann, W., & Stachowicz-Stanusch, A. (Eds.). 2020. Contemporary Perspectives in Corporate Social Performance and Policy. Artificial Intelligence and Its Impact on Business. Information Age Publishing Inc. Bosch-Sijtsema, P., Claeson-Jonsson, C., Johansson, M., & Roupe, M. 2021. The Hype Factor of Digital Technologies in AEC. Construction Innovation, 21(4), 899–916. https://doi.org/10.1108/CI-01-2020-0002
333
Buxmann, P., & Schmidt, H. 2019. Künstliche Intelligenz. Springer Berlin Heidelberg. https://doi.org/10.1007/9783-662-57568-0 Chui, M., & Mischke, J. 2019. The Impact and Opportunities of Automation in Construction: Voices on Infrastructure. Cisterna, D., Seibel, S., Oprach, S., & Haghsheno, S. 2022. Artificial Intelligence for the Construction Industry – A Statistical Descriptive Analysis of Drivers and Barriers. In J. F. de Paz Santana, D. H. de La Iglesia, & A. J. López Rivero (Eds.), Advances in Intelligent Systems and Computing. New Trends in Disruptive Technologies, Tech Ethics and Artificial Intelligence (Vol. 1410, pp. 283–295). Springer International Publishing. https://doi.org/10.1007/978-3-030-87687-6_27 Clark, M., & Melancon, J. 2013. The Influence of Social Media Investment on Relational Outcomes: A Relationship Marketing Perspective. International Journal of Marketing Studies, 5(4). https://doi.org/10.5539/ijms. v5n4p132 Darko, A., Chan, A. P., Adabre, M. A., Edwards, D. J., Hosseini, M. R., & Ameyaw, E. E. 2020. Artificial intelligence in the AEC industry: Scientometric Analysis and Visualization of Research Activities. Automation in Construction, 112, 103081. https://doi.org/10.1016/j.autcon.2020. 103081 DaSilva, C. M., & Trkman, P. 2012. Business Model: What it is and What it is Not. SSRN Electronic Journal. Advance Online Publication. https://doi.org/10.2139/ssrn.2181113 de-Best, R. 2021. Construction Industry Spending Worldwide From 2014 to 2019, With Forecasts From 2020 to 2035. https://www.statista.com/statistics/788128/ construction-spending-worldwide/?sub1=20220502-085428c4-b80e-d7ee6c064f5d European Commission. 2003. European Commission Recommendation of 6 May 2003 Concerning The Definition of Micro, Small and Medium-sized Enterprises. European Commission. https://www.ifm-bonn.org/en/index/Business definitions/uebersetzen-nach-english-kmu-definition-dereu-kommission Fischer, L., Ehrlinger, L., Geist, V., Ramler, R., Sobiezky, F., Zellinger, W., Brunner, D., Kumar, M., & Moser, B. 2021. AI System Engineering—Key Challenges and Lessons Learned. Machine Learning and Knowledge Extraction, 3(1), 56–83. https://doi.org/10.3390/make3010004 Hoffmann, C. P., Lennerts, S., Schmitz, C., Stölzle, W., & Uebernickel, F. 2016. Business Innovation: Das St. Galler Modell. Springer Fachmedien Wiesbaden. https://doi.org/ 10.1007/978-3-658-07167-7 Jensen, A. B. 2014. Do We Need One Business Model Definition?Advance Online Publication. https://doi.org/10.5278/ OJS.JBM.V1I1.705 (Journal of Business Models, Vol 1 No 1 (2013): Inaugural issue). Kaniški, I., & Vincek, I. 2018. Success Measurement of Business Processes. Tehnièki Glasnik, 12(2), 113–119. https://doi.org/10.31803/tg-20170809204554 Kanji, G. K. 2006. 100 statistical tests (3rd ed.). SAGE. Kreutzer, R. T., & Sirrenberg, M. 2019. Künstliche Intelligenz verstehen. Springer Fachmedien Wiesbaden. https://doi.org/10.1007/978-3-658-25561-9 Mazhar, N., & Arain, F. 2015. Leveraging on Work Integrated Learning to Enhance Sustainable Design Practices in the Construction Industry. Procedia Engineering, 118, 434– 441. https://doi.org/10.1016/j.proeng.2015.08.444 Momade, M. H., Shahid, S., Falah, G., Syamsunur, D., & Estrella, D. 2021. Review of Construction Labor productivity Factors From a Geographical Standpoint. International Journal of Construction Management, 1–19. https://doi.org/10.1080/15623599.2021.1917285
Murshed, N. A. 2020. How Social Media Changes Today’s Marketing Strategies. International Journal of Innovative Science and Research Technology, 5(9), 725–736. Paniagua, J., & Sapena, J. 2014. Business Performance and Social Media: Love or hate? Business Horizons, 57(6), 719–728. Ramli, M. Z., Malek, M. A., Hamid, B., Roslin, N. T., Roslan, M. E. M., Norhisham, S., & Mohd, M. F. 2018. Influence of Project Type, Location and Area towards Construction Delay: a Review on Significance Level of Delay Factors. International Journal of Engineering & Technology, 7(4.35), 392. Rao, T. V. N., Gaddam, A., Kurni, M., & Saritha, K. 2021. Reliance on Artificial Intelligence, Machine Learning and Deep Learning in the Era of Industry 4.0 (Smart Healthcare System Design). Wiley, pp. 281–299.
334
Regona, M., Yigitcanlar, T., Xia, B., & Li, R. Y. M. 2022. Opportunities and Adoption Challenges of AI in the Construction Industry: A PRISMA Review. Journal of Open Innovation: Technology, Market, and Complexity, 8(1), 45. https://doi.org/10.3390/joitmc8010045 SDaC. 2020. Smart Design and Construction Through Artifificial Intelligence. https://sdac.tech/en/ Shafer, S. M., Smith, H. J., & Linder, J. C. 2005. The power of Business Models. Business Horizons, 48(3), 199–207. Sousa, M. J., Osório de Barros, G., & Tavares, N. 2021. Artificial Intelligence a Driver for Digital Transformation (Digital Transformation and Challenges to Data Security and Privacy). IGI Global, pp. 234–250. Sowmya, S. K., Deepika, P., & Naren, J. 2014. Layers of Cloud - IaaS, PaaS, and SaaS: A Survey, Vol.5 (3), 4477– 4480.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Predicting occupant evacuation times to improve building design J. Clever, J. Abualdenien, R.K. Dubey & A. Borrmann Chair of Computational Modeling and Simulation, Technical University of Munich, Munich, Germany
ABSTRACT: Building design requires considering multiple requirements and must fulfill diverse regulations. Therefore, model analysis and simulations are fundamental parts of the design process to find the optimal solution for a given problem. Important decisions are based on a building’s assessed final performance in the early design phases. In particular, the analysis of pedestrian flow dynamics is paramount for public facilities like train stations concerning occupants’ comfort and evacuation behavior. Currently, it requires multiple steps, from preparing the BIM model to performing pedestrian flow analysis, including semi-automated, often manual work that demands high computation times. Therefore, to improve the building design efficiency in terms of time and pedestrian circulation, this paper proposes a framework applying Deep Learning methods. We propose a real-time pedestrian evacuation prediction to replace time-consuming pedestrian dynamics simulations. More precisely, a modular neural network architecture is designed, including a Convolutional Neural Network and a Multilayer Perceptron, that takes floorplan images and building and simulation parameters as input and predicts the crowd evacuation time for a given building model. As a result, a mean prediction accuracy of 15% could be achieved.
1 INTRODUCTION Experts from various interconnected domains form a multidisciplinary design team in construction projects. The resultant design of a building and its performance is strongly influenced by multiple design decisions made by each discipline during the design process. In recent years, the Building Information Modeling (BIM) methodology has become an established and common tool that improves the collaborative work between the different disciplines in the project and provides information throughout the project beginning in the early phases (Borrmann et al. 2018). The design process of a building consists of various stages, where the building is developed from a rough conceptual design to a complex model, including detailed information about all individual components. Especially in the early design phases, fundamental decisions are taken that have a significant impact on the final performance of the building (Knotten et al. 2015). Nevertheless, the required costs and efforts for changes in the design are relatively low (Abualdenien & Borrmann 2019). By comparing the results of numerous simulations and analyses, architects and engineers explore several models and evaluate multiple design options regarding performance. Commonly, analysis and simulations include the structural system, embodied and operational energy during a building’s life-cycle (Abualdenien et al. 2020), and pedestrians’ evacuation behavior and comfort inside a building. BIM offers enormous information about different objects in the model (i.e., DOI 10.1201/9781003354222-43
walls, stairs, zones). For each instance, a geometric representation and a set of properties are accessible (Abualdenien & Borrmann 2019). Moreover, individual simulation information can be added to the model, and, hence, a smooth workflow between BIM-authoring tools and simulation software can be provided. To allow vendor-neutral data exchange, the open standard Industry Foundation Classes (IFC) (BuildingSMART 2020) is widely supported by various existing authoring tools and simulation software and allows an easy exchange of model data. So far, IFC BIM models show a promising possibility to work as a basis for simulation software as many researchers have confirmed (Mirahadi et al. 2019). As the decision-making process highly influences the project outcomes, the application of simulations in that stage help estimate the building’s performance (Abualdenien & Borrmann 2019). Especially, pedestrians’ walking routes are essential when designing a building concerning emergency situations, for the pedestrians’ behavior is strongly dependent on their environment (Low 2000). Specifically, the building’s shape significantly influences efficient crowd routing considering safety and comfort (Hanisch et al. 2003). Thus, this paper aims to improve the integration of pedestrian dynamics simulations into the design phase, notably considering public buildings such as train stations, where emergency evacuation plays a vital role (Løvås 1994). Typical results of pedestrian simulations can be comfort evaluation, walking routes visualization and
335
insights about emergency situations. Nevertheless, integrating simulations into the workflow still consists of multiple steps, beginning with the building models’ export from the BIM-authoring tool, converting and importing them into the simulation software, running the simulation, and postprocessing the final simulation results. Moreover, the usage of agent-based simulation models is computationally expensive, leading to long computation times, and is error-prone (Andriamamonjy et al. 2018). This time-consuming process may obstruct a full investigation of the design space. To overcome these restrictions, the framework proposed in this paper employs Deep Learning (DL) methods to allow real-time predictions of pedestrians’ behavior and walking times. To avoid expensive pedestrian dynamics simulations, especially Machine Learning (ML) approaches can serve as supportive tools or be used as complete replacements (Kim et al. 2019). Since BIM models include a massive set of information, we use them directly as input for the ML model and enable an immediate evaluation of pedestrians’ behavior considering the interaction of multiple design options. The proposed method providing realtime evaluation allows interactive exploration of the solution space, thus enabling designers to find wellperforming solutions in a shorter time. Since public buildings such as transport hubs must fulfill various requirements concerning evacuation time, this paper focuses on train stations. The structure of this paper is as follows: Section 2 provides background information and related research. Section 3 introduces the concept of our approach step by step. In Section 4, details about the implemented neural network are given, whereas Section 5 presents the results. Finally, the last Section 6 summarizes the outcome and discusses future steps.
2 BACKGROUND AND RELATED WORK 2.1 Performance-based building design The designing process of a building consists of many different steps, which result in various decisions and dependencies. Performance-based building design becomes a promising method to maximize the overall building’s performance and reduce critical changes in the final project phases (Mehrbod et al. 2020). The accessibility of sufficient data and information is crucial, as decisions from early design phases can significantly impact the building’s later performance and cost (Østergård et al. 2016). Especially BIM-based approaches allow the usage of comprehensive digital models within the design process, which helps improve the decision-making. For the structural design of a building,A BIM-based optimization evaluation approach was developed by Hamidavi et al. (2020). With this, especially in the design phase, the coordination between architects and structural engineers is improved. Moreover, the authors of Röck et al. (2018) propose considering the
building’s materials for the Life Cycle Assessment and integrating parts into BIM. Hence, the potential effects of the building’s materials become more comprehensible about their embodied energy. 2.2 Pedestrian dynamics analysis and simulation models When designing public buildings such as shopping centers or train stations, especially emergency evacuation is essential (Løvås 1994). For efficient crowd routing inside a building, pedestrian dynamics analysis plays a vital role in safety and comfort while highly dependent on the building’s shape (Hanisch et al. 2003). Research has shown that single pedestrians incline toward polygon-shaped walking routes, where visibility stimulates pedestrians to walk on straight paths for as long as possible. Moreover, while certain areas may appear crowded, pedestrians accept unknown detours and longer traveling times with or without intention (Helbing et al. 2001). Furthermore, neither direct communication nor explicit concepts but intuitive awareness rule a crowd’s self-organizational behavior, notably for crowds with unidirectional pedestrian flows (Helbing et al. 2005). When single pedestrians encounter stationary groups, they interpret them as obstacles and are prone to change their walking paths. Moreover, individual persons tend to adapt to the walking speed of other moving crowds within an overall crowded area (Yi et al. 2015). The choice of the simulation model commonly depends on the number of virtual pedestrians (agents), where three main approaches were developed to model pedestrian behavior. On the one hand, individual agents and their reactions are modeled by microscopic approaches. On the other hand, macroscopic models reflect aggregated person streams. In addition, mesoscopic approaches can handle following individual agents and understanding group behavior (Ijaz et al. 2015). Regarding the findings that only rule-based methods may not necessarily lead to satisfactory results (Yang et al. 2020), Helbing et al. (2000) developed the more general (microscopic) social force model. In this approach, individual agents move with a certain velocity while their repulsive interaction forces consider obstacles and other agents. When it comes to pedestrian crowds, the modeling instead follows a flow mechanism not considering the crowd’s environment and individual agent’s interplay, unlike modeling individual pedestrians’behavior. In particular, the authors of Hughes (2002) present the principle of continuum theory as the basis for crowd representation. In addition, using navigationor guidance fields, the potential field model simulates multiple intentions of pedestrian crowds, introduced in Yang et al. (2020). Moreover, from fluid dynamics the aggregate dynamics, model is derived. Although a common technique, strict cellular automata structuring leads to restrictions in representing reality, where obstacles or densities of pedestrian crowds may not be wholly cell-filling and, hence,
336
lack accuracy (Biedermann et al. 2016). Hybrid models work as an alternative, where specific regions and areas can be assigned to particular modeling approaches representing individual behavior (Biedermann et al. 2021). Furthermore, the optimal steps model (OSM) is not focusing on a rigid spatial grid or dense crowds only. The OSM frees agents from a strict cell representation using continuous space, whereas a discrete stepwise movement is kept (Seitz and Köster 2012).
2.3 Train stations and crowd dynamics In this research, we use train stations as an example of a facility whose design has a significant impact on the pedestrian flows, which has a major impact on the performant and save operation of that facility. It will thus serve as the subject of our investigations and as the basis for the proof-of-concept. Train station designs often vary concerning individual requirements. Usually, train stations provide waiting areas for pedestrians, where the uniform distribution of people over the respective spaces can be observed (Helbing et al. 2001). Besides, studies of crowd dynamics in train stations highlight a notable influence of waiting pedestrians. More precisely, inconveniently placed points of attraction and waiting pedestrians lead to an increase of up to 20% of walking time for arriving passengers leaving the platform area (Davidich et al. 2013). As for the impact of different building elements on crowded areas in train stations, Ma et al. (2013) examined separation modules such as fences and pillars. Using pillars instead of other or no separation modules for non-unidirectional movements, the authors identified an increase in pedestrians’ flow rate. Similarly, improvements in evacuation time could be observed for exit areas when placing pillars close to them (Frank and Dorso 2011).
2.4 Deep learning Until now, the introduction of pedestrian behavior and various simulation models implied their complexity. As a result, performing pedestrian simulations for incredibly complex building designs can quickly yield high computation time. To overcome this issue, the research community more and more involves methods of Artificial Intelligence (AI). Naming a specific category of AI methods, predictive tools can replace time-consuming computations with the support of ML approaches. In doing so, a surrogate function is found and applied to the problem. DL approaches became favored support, especially for dealing with distinct data types and various problems. More specifically, multiple architectures of Artificial Neural Networks (ANNs or NNs) exist to deal with different tasks and purposes. For instance, object detection and segmentation in images and as well as
natural language processing attain different success rates. A well-known feedforward NN is the Multilayer Perceptron (MLP), famous for solving various problems (Nielsen 2015). The MLP is arranged in multiple (hidden) layers that contain any number of connected computational nodes. These nodes store single values that are processed in one direction. A suitable number of nodes and layers for solving a given task sufficiently is essential for the resulting individual NN. The goal of the NN is mapping a given input to the desired output, also known as a classified label, training the network to customize the network to a particular problem. Commonly, applying a backpropagation algorithm to the network makes its parameters optimized and the accuracy improved (Nielsen 2015). When it comes to matrix-like data such as images, Convolutional Neural Networks (CNNs) have become an efficient way of achieving results. Again, the CNN follows a feedforward architecture with multiple layers. Typically, each layer carries out a set of computations. In the first step, a kernel performs the convolution operation for a given input matrix and results in a feature map. Moreover, multiple feature maps can be computed by different kernels in parallel within the same layer returning a feature set, where a kernel can be described as a filter (Goodfellow et al. 2016). Following the convolution, each element of the feature map is processed by a nonlinear activation function, for instance, the rectified linear unit function. Finally, down-sampling reduces the matrix dimensions, commonly done by a pooling operation such as maximum pooling. For all following layers, downsampling helps reduce the computational effort. Additionally, for a given dataset, CNNs can detect and filter out patterns (features) (Goodfellow et al. 2016). The training of a neural network is a process with various parameters and options. Although several optimization techniques help improve the training process, a sufficient amount of data is essential. Moreover, underfitting can occur by providing too little data. The same training data can often lead to overfitting since the network may adapt to the specifics of the examples to a too large extent. An intentional increase of uncertainty within the model can be applied to reduce overfitting by using regularization methods like the dropout. Thereby, the activated nodes are varied almost randomly, leading to the prevention of co-adaptions and, hence, to improved computations (Srivastava et al. 2014). Furthermore, a network’s training process can be enhanced by batch normalization (Santurkar et al. 2018). Each layer’s inputs are normalized before the actual activation of the following computational nodes. Then again, deep dependencies between multiple layers may be partially relieved, also known as decreasing the covariate shift. On the other hand, the integration of batch normalization helps reduce the necessity of regularization methods like dropout (Ioffe and Szegedy 2015).
337
After training an NN architecture for a given task, the gained knowledge can also be used for other purposes. Moreover, a pre-trained network can be retrained for a different dataset or partially reused for new tasks. This process is described as transfer learning and a standard procedure for improving network results, especially when using CNNs (Ribani and Marengoni 2019). Popular network architectures can be found online, such as the VGG16 pre-trained on the ImageNet dataset for classification problems (Simonyan and Zisserman 2014). Giving some examples, the authors of Nishida and Hotta (2018) used CNNs to detect and distinguish cell particles from non-cell particles based on image data. In building design, Geyer and Singaravel (2018) developed a component-based model to estimate heatingand cooling energy within a building. In another approach, flow control, performance, and optimization of fluid dynamics calculations were improved using ML methods (Brunton et al. 2020). Finally, the understanding of pedestrians’ walking routes and densities in a given environment could be predicted in (Clever et al. 2021).
Figure 1. Example of parametric model (Revit/Dynamo). Table 1.
Parameter values for generic train station variation.
Abbreviation
Meaning
Variations
F T W L H E P
No. of floors Distance of tracks No. of tracks Station length Floor height No. of escalators Agents per coach
2 15, 25 2, 3, 4, 5 150, 200, 250, 300 15, 25 1, 2, 3 5, 20, 50
3.2 Parametric models 3 METHODOLOGY 3.1 Hypothesis and aim In this paper, the hypothesis is that real-time predictions can replace time-consuming pedestrian dynamics simulations with DL methods that relate the design information of a building model with individual simulation results. From here, two questions arise: (1) what is a suitable representation of the geometric and semantic design information? (2) Which aspects of the simulation results shall be predicted? These questions are essential for a suitable NN architecture, where layers and parameters must carefully be configured. Contrary to Clever et al., 2021, in this paper we focus on the evacuation times. Thus, we propose a framework that automatically generates a training dataset and predicts results directly from the BIM model, considering specific simulation parameters. Considering the work of Clever et al. (2021), the generic train station models created by a parametric model are reused for this paper. Instead of predicting heatmaps or tracing maps, the present paper focuses on predicting the respective evacuation times of a given train station design. Moreover, the corresponding IFC exports and pedestrian simulations are carried out. In the scope of this paper, we use the pedestrian dynamics simulator crowd:it (Accu:rate 2022), which is based on the optimal steps model (OSM) (Seitz and Köster 2012), for generating the training and validation data. With the available simulation results, postprocessing is applied, and the agents’ walking times are extracted for each model variation. Together with the train station models, the generated dataset is used for training a neural network.
The available train station models by Clever et al. (2021) were generated by a parametric model using Autodesk Revit (Autodesk 2022) and Dynamo (Autodesk 2021). By varying parameters of the design’s geometry according to Table 1, 450 different generic train station models are considered for the dataset. The parametric platform presented in Figure 1 has three escalators at each end, four track lines, a row of two columns, and an elevator in between.
3.3 Floorplan representation As mentioned earlier, specific zones are marked in the BIM models, necessary for the simulation. By assigning different colors to the different zones, a floorplan representation of the models is used as input for the neural network, alongside a vector containing the models’ metadata, as presented in Table 1. Concerning the labeling, Figure 2 shows an example where the pink color represents spawning zones, and white spaces are walkable areas. The output of the simulations is plain numbers serving as the models’ overall evacuation times. In the simulation results (used for training and validation), each time step of each agent is given,
Figure 2. Colored floorplan example.
338
whereas the very last time step would be the time needed to evacuate the building entirely.
Table 2.
Training case
Meaning for CNN
1 2 3
VGG16 pre-trained, no weight updates VGG16 pre-trained, weight updates VGG16 not pre-trained, weight updates
4 NEURAL NETWORK ARCHITECTURE As discussed in Section 2.4, multiple approaches exist to design a neural network architecture depending on the input and output data. This paper uses floorplan images and metadata of the underlying parametric building models as input information, while we predict a single value, i.e., the evacuation time, as output. Hence, we assemble different sub-networks according to the respective data structure to receive the final network architecture. The modular architecture is presented in Figure 3. In the simulation results, all individual time steps are saved for each agent, where the very last step counts as the building’s evacuation time. To avoid including incorrect results, e.g., by agents got stuck, we omitted the last 5% of these time steps. The metadata is structured as a vector of numbers containing various information about the building model. Thus, the metadata input will be processed by an MLP (MLP-in), with three hidden layers. For the floorplan images, we use the basic CNN design of the VGG16 according to (Simonyan and Zisserman 2014) where the input layer is created considering an image resolution of 256 * 256 pixels. Moreover, the dense layers and the final classification layer of the VGG16 are excluded since no image classification is desired. Finally, a concatenation layer (concat) combines the MLP-in and the CNN outputs to create a joined input for the predictive part of the network. Again, an MLP (MLP-out) is used to process the combined input information of the metadata and the floorplan images. During network hyperparameter tuning, especially the CNN base model offers three different modifications we considered for each training setup: (1) the pre-trained weights of the VGG16 network are used as given. In contrast, an update of the weights during training is restricted. (2) The pre-trained weights of the VGG16 are used, but an update of the weights during training is possible. (3) The initial VGG16 weights are not explicitly defined and updated at each training epoch if necessary. The setup cases are summarized in Table 2. According to transfer learning, pre-trained weights shall help improve the overall network training and, thus, its efficiency.
Neural network training setup cases.
5 NEURAL NETWORK RESULTS AND EVALUATION In a first step, we split the dataset into two subsets of training and testing data. We created 450 different train station models, including their respective simulation results. We reserved 25% of the models (∼120) for testing. Due to the different parameters of the models, different resolutions for the respective floorplan images occur, which we adjusted to one equal resolution by resizing. Additionally, we used data augmentation for the training data, doubling the number of projects to 660 while applying random rotation and mirroring to the floorplan images. Furthermore, 20% of the training data was used as validation data in every epoch. Various training setups lead to an optimal batch size of 16 and a total of 150 epochs to avoid overfitting, where the number of hidden layers for the MLP-in and MLP-out are three and four, respectively (see Figure 3). Due to a single value prediction as to the network’s result, we used the mean squared error (MSE) as the loss function to compare the prediction with the ground truth during training and validation. Moreover, we used the Adam optimizer and a learning rate of 0.001. Beginning with training case 1 (see Table 2), the MSE of both training and validation data is shown in Figure 4. The training loss went below an MSE of 1000 after 7 epochs and circle approx. 100 after approx. 20 epochs, while the validation loss needed approx. 20 epochs to go significantly below an MSE of 1000 and circle a loss of approx. 600.
Figure 4. Case 1 – training and validation loss.
Figure 3. Structure of the neural network architecture.
A comparable difference between training and validation loss, as in Figure 4, occurred during all training
339
iterations. Moreover, one must consider that the MSE implies squared error values.The evaluation of the testing dataset was performed on the difference between prediction and ground truth, normalized with respect to the ground truth. The mean of all differences yielded in 20%, while the ground truth’s minimum and maximum evacuation times are approx. 90s and 480s, respectively. Training time was approx. 6.5min, while the simulation time’s minimum and maximum computation times are approx. 0.5min and 262.5min, respectively. For training case 2, the training and validation loss needed approx. 30 epochs for the MSE to go below 1000, as shown in Figure 5. For one thing, the validation loss is slightly lower than in case 1. Then again, the training loss is higher than in case 1 while staying lower than the validation loss. Compared with case 1, the losses of case 2 need around 20 epochs more to decrease to similar loss values. One possible reason can be the update of the CNN’s weights in case 2, where also the update of the MLPin and MLP-out are considered. Concerning the testing dataset, a mean of 19% was achieved. In this regard, a slight difference to case 1 exists. Moreover, training time is almost double with approx. 12.5min.
Figure 6. Case 3 – training and validation loss.
Figure 7. Case 3 – difference between prediction and ground truth.
Figure 5. Case 2 – training and validation loss.
Lastly, Figure 6 shows the MSE for training and validation data of training case 3. Similar to case 2, for both losses, it took approx. 30 epochs to decrease to an MSE below 1000. Both graphs are reasonably comparable to case 2 concerning the loss values. However, the evaluation of the testing data yields a mean of 15%, which is significantly better than for cases 1 and 2. In addition, for case 3, Figure 7 shows a histogram of the testing data evaluation. Altogether, the accuracy of the prediction was 15% with a standard deviation of 8%. We noticed that the majority (i.e. 80% out of 120 test samples) produce less than 20% deviation in the evacuation time (in secs) from the ground truth. As in case 2, training time is approx. 12.5min, thus, double the time as in case 1. Overall, the comparison of the results shows differences between the three training setup cases. Although for case 1, the loss decreases faster than in cases 2
and 3, the accuracy measured by the mean values is best for case 3. Nevertheless, the final evaluation of a mean of 15% for case 3 depends on the particular use case, meaning the necessary individual accuracy for a given problem. Otherwise, for cases 2 and 3, the computation time of the network training is highest. Again, the importance of the necessary time training the network is to be assessed concerning a given problem. A summary of the results can be found in Table 3. Table 3.
Summary of results.
Training case
Mean [%]
Training time [min]
1 2 3
20 19 15
6.5 12.5 12.5
6 CONCLUSION AND FUTURE WORK Conventional pedestrian simulations can easily lead to high computation times and effort concerning computational resources. Hence, a real-time prediction
340
of the evacuation time for a given building model would improve the workflow significantly. In this paper, we designed a modular NN, combining different NN architectures for multiple input- and output data types, providing a real-time prediction of the evacuation time for a given train station geometry. The presented approach considered different neural network architectures, such as the MLP for metadata of the BIM model and the CNN for the corresponding floorplan images. Moreover, we created a new primary network architecture with an individual combination of subnetworks, including the possible option of using pretrained weights for the CNN at training time. The results of the approach allow the conclusion that pre-trained weights of the CNN while prohibiting their update during the entire network’s training may be helpful when the training time of the network shall be reduced. Otherwise, depending on the particular problem, the accuracy of the entire network based on pre-trained weights for the CNN may not be sufficient, whereas an overall trained network without pre-trained weights shows better predictions. Unlike the prediction of pedestrian trajectories in public buildings (Lui et al. 2021), and evacuation routes (Zhang et al. 2021), this paper presents the implementation of an NN-based real-time prediction of evacuation times. In the future, we aim to train a model that can predict both the overall evacuation time and the pedestrian trajectory. An immediate improvement of the presented work can be made by training on more complex data (e.g., multiple floors). In our case, a generic train station model may be relatively similar, leading to misinterpretations of changes in the building design. We see the enormous advantage of including predictive tools in building design. Especially for exploring different building layouts in the early design stages, real-time predictions can significantly help explore the solution space while considering various performance criteria such as evacuation times. Consequently, project time and computational effort will be reduced while developing the optimal design solution that considers multiple factors and dependencies can be fulfilled.
ACKNOWLEDGMENTS We gratefully acknowledge the support of mFUND – Bundesministerium für Digitales und Verkehr in Germany for funding the research project BEYOND. REFERENCES Abualdenien, J. & A. Borrmann (2019, apr). A Metamodel Approach for Formal Specification and Consistent Management of Multi-LOD Building Models. Advanced Engineering Informatics 40, 135–153. Abualdenien, J., P. Schneider-Marin, A. Zahedi, H. Harter, H. Exner, D. Steiner, M. Mahan Singh, A. Borrmann, W. Lang, F. Petzold, M. König, P. Geyer, &
M. Schnellenbach-Held (2020). Consistent Management and Evaluation of Building Models in the Early Design Stages. Journal of Information Technology in Construction 25, 212–232. Accu:rate (2022). crowd:it – the software for state-of-the-art planners. Andriamamonjy, A., D. Saelens, & R. Klein (2018, jul). An Automated IFC-based Workflow for Building Energy Performance Simulation with Modelica. Automation in Construction 91, 166–181. Autodesk (2021). Dynamo BIM. Autodesk (2022). Revit Software. Biedermann, D. H., J. Clever, & A. Borrmann (2021, feb). A Generic and Density-sensitive Method for Multi-scale Pedestrian Dynamics. Automation in Construction 122, 103489. Biedermann, D. H., C. Torchiani, P. M. Kielar, D. Willems, O. Handel, S. Ruzika, & A. Borrmann (2016). A Hybrid and Multiscale Approach to Model and Simulate Mobility in the Context of Public Events. Transportation Research Procedia 19, 350–363. Borrmann, A., M. König, C. Koch, & J. Beetz (Eds.) (2018). Building Information Modeling. Cham: Springer International Publishing. Brunton, S. L., B. R. Noack, & P. Koumoutsakos (2020, jan). Machine Learning for Fluid Mechanics. Annual Review of Fluid Mechanics 52(1), 477–508. BuildingSMART (2020). IFC4 Documentation. Clever, J., J. Abualdenien, & A. Borrmann (2021). Deep Learning Approach for Predicting Pedestrian Dynamics for Transportation Hubs in Early Design Phases. In 28th International Workshop on Intelligent Computing in Engineering, Berlin, Germany. Davidich, M., F. Geiss, H. G. Mayer,A. Pfaffinger, & C. Royer (2013, dec). Waiting Zones for Realistic Modelling of Pedestrian Dynamics: A Case Study Using Two Major German Railway Stations as Examples. Transportation Research Part C: Emerging Technologies 37, 210–222. Frank, G. A. & C. O. Dorso (2011). Room Evacuation in the Presence of an Obstacle. Physica A: Statistical Mechanics and its Applications 390(11), 2135–2145. Geyer, P. & S. Singaravel (2018). Component-based Machine Learning for Performance Prediction in Building Design. Applied Energy 228(July), 1439–1453. Goodfellow, I., Y. Bengio, & A. Courville (2016). Deep Learning. MIT Press. Hamidavi, T., S. Abrishami, & M. Hosseini (2020, nov). Towards Intelligent Structural Design of Buildings: A BIM-based Solution. Journal of Building Engineering 32, 101685. Hanisch, A., J. Tolujew, K. Richter, & T. Schulze (2003). Online Simulation of Pedestrian Flow in Public Buildings. In Proceedings of the 2003 International Conference on Machine Learning and Cybernetics (IEEE Cat. No.03EX693), pp. 1635–1641. IEEE. Helbing, D., L. Buzna, A. Johansson, & T. Werner (2005). Self-organized Pedestrian Crowd Dynamics: Experiments, Simulations, and Design Solutions. Transportation Science 39(1), 1–24. Helbing, D., I. Farkas, & T. Vicsek (2000). Simulating Dynamical Features of Escape Panic. Nature 407(6803), 487–490. Helbing, D., P. Molnár, I. J. Farkas, & K. Bolay (2001). Self-organizing Pedestrian Movement. Environment and Planning B: Planning and Design 28(3), 361–383. Hughes, R. L. (2002). A Continuum Theory for the Flow of Pedestrians. Transportation Research Part B: Methodological 36(6), 507–535.
341
Ijaz, K., S. Sohail, & S. Hashish (2015). A Survey of Latest Approaches for Crowd Simulation and Modeling using Hybrid Techniques. 17th UKSIM-AMSS International Conference on Modelling and Simulation, 111–116. Ioffe, S. & C. Szegedy (2015, feb). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Kim, J., J. Moon, E. Hwang, & P. Kang (2019). Recurrent Inception Convolution Neural Network for Multi Shortterm Load Forecasting. Energy and Buildings 194(2019), 328–341. Knotten, V., F. Svalestuen, G. K. Hansen, & O. Lædre (2015). Design Management in the Building Process -A Review of Current Literature. Procedia Economics and Finance 21, 120–127. Løvås, G. G. (1994, dec). Modeling and Simulation of Pedestrian Traffic Flow. Transportation Research Part B: Methodological 28(6), 429–443. Low, D. J. (2000, sep). Following the Crowd. Nature 407(6803), 465–466. Lui, A. K.-F., Y.-H. Chan, & M.-F. Leung (2021, dec). Modelling of Destinations for Data-driven Pedestrian Trajectory Prediction in Public Buildings. In 2021 IEEE International Conference on Big Data (Big Data), pp. 1709–1717. IEEE. Ma, J., S. M. Lo, W. G. Song, W. L. Wang, J. Zhang, & G. X. Liao (2013). Modeling Pedestrian Space in Complex Building for Efficient Pedestrian Traffic Simulation. Automation in Construction 30, 25–36. Mehrbod, S., S. Staub-French, & M. Tory (2020, jan). BIMbased Building Design Coordination: Processes, Bottlenecks, and Considerations. Canadian Journal of Civil Engineering 47(1), 25–36. Mirahadi, F., B. McCabe, & A. Shahi (2019, may). IFC-Centric Performance-based Evaluation of Building Evacuations Using Fire Dynamics Simulation and Agent-based Modeling. Automation in Construction 101, 1–16. Nielsen, M. A. (2015). Neural networks and deep learning, Volume 2018. Determination press San Francisco, CA.
Nishida, K. & K. Hotta (2018, oct). Robust Cell Particle Detection to Dense Regions and Subjective Training Samples Based on Prediction of Particle Center Using Convolutional Neural Network. PLOS ONE 13(10), e0203646. Østergård, T., R. L. Jensen, & S. E. Maagaard (2016, aug). Building Simulations Supporting Decision Making in Early Design – A Review. Renewable and Sustainable Energy Reviews 61, 187–201. Ribani, R. & M. Marengoni (2019, oct). A Survey of Transfer Learning for Convolutional Neural Networks. In 2019 32nd SIBGRAPI Conference on Graphics, Patterns and Images Tutorials (SIBGRAPI-T), pp. 47–57. IEEE. Röck, M., A. Hollberg, G. Habert, & A. Passer (2018, aug). LCA and BIM: Visualization of Environmental Potentials in Building Construction at Early Design Stages. Building and Environment 140, 153–161. Santurkar, S., D. Tsipras, A. Ilyas, & A. Madry (2018, may). How Does Batch Normalization Help Optimization? Seitz, M. J. & G. Köster (2012, oct). Natural Discretization of Pedestrian Movement in Continuous Space. Physical Review E 86(4), 046108. Simonyan, K. & A. Zisserman (2014, sep). Very Deep Convolutional Networks for Large-Scale Image Recognition. Srivastava, N., G. Hinton, A. Krizhevsky, I. Sutskever, & R. Salakhutdinov (2014). Dropout: A Simple Way to Prevent Neural Networks from Overfitting. Journal of Machine Learning Research 15(56), 1929–1958. Yang, S., T. Li, X. Gong, B. Peng, & J. Hu (2020). A review on Crowd Simulation and Modeling. Graphical Models 111(October 2019), 101081. Yi, S., H. Li, & X. Wang (2015, jun). Understanding Pedestrian Behaviors from Stationary Crowd Groups. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 3488–3496. IEEE. Zhang, Y., Z. Chai, & G. Lykotrafitis (2021, jun). Deep Reinforcement Learning with a Particle Dynamics Environment Applied to Emergency Evacuation of a Room with Obstacles. Physica A: Statistical Mechanics and its Applications 571, 125845.
342
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Inferring interconnections of construction drawings for bridges using deep learning-based methods B. Faltin, P. Schönfelder & M. König Chair of Computing in Engineering, Department of Civil and Environmental Engineering, Ruhr-Universität Bochum, Bochum, Germany
ABSTRACT: The availability of digital models of existing structures plays a vital role in leveraging the full potential of digital planning methods like building information modeling (BIM) in the building’s operation and maintenance phase. Since BIM models are, however, widely unavailable, the reconstruction process must be carried out in advance. It is particularly tedious for engineers to reconstruct 3D models from drawings, since many partial views are included, e.g., sections. Consequently, the interconnections between the views must be established before reconstructing the geometry. This research proposes a deep learning-based method to localize section symbols on construction drawings and recognize their associated denotations. Also, the title of the view is recognized to derive the respective connections between the section symbol and the corresponding view. The proposed pipeline is tested on actual bridge drawings and shows promising results. This paves the way for future works addressing the automatic reconstruction of the bridge geometry.
1 INTRODUCTION Building information modeling (BIM) connects semantic information with geometric representation. This allows the consolidation of all project knowledge into one source, i.e., the digital building model, and thus supports the decision-making process. While BIM is widely utilized in the building design and planning, it is rarely employed in the operation and maintenance phase of a building’s life cycle. This is despite the fact that using a digital model for inspection, condition assessment, and retrofit planning can enhance the efficiency of maintenance management (Sacks et al. 2018). One reason for the limited utilization of BIM in operation is that models of existing buildings are widely unavailable. Therefore the retrodigitization of these assets is essential and must be carried out in advance. Data sources like point clouds, images, or construction drawings can be candidates for extracting the necessary geometric building information. However, the acquisition of point clouds and images must be conducted on-site and might require specialized equipment. Hence, it is quite costly. Contrarily, construction drawings are readily available, either in paper form or as rendered computer-aided drawing (CAD) files. Furthermore, technical drawings are a crucial data source for producing high-quality models, as they contain information about obscured components, which is inherently absent in point clouds. Therefore, when it comes to reconstructing digital building models, drawings are often the data source of choice.
DOI 10.1201/9781003354222-44
One shortcoming of drawings is that they represent 2D projections of a 3D building. Views from various angles must be consulted in order to depict a building in its entirety. In addition, to represent the interior of the structure, virtual sections are conducted, and the resulting views are projected onto the drawings as well. Thus, a technical drawing could hold many individual views of the same structure, e.g., layouts, vertical sections, or detail views. Consequently, the engineer must reconstruct the interconnections between these views and how they come together in the 3D space in order to obtain the digital model. This is a timeconsuming process, which requires a high degree of technical expertise. However, the step is unavoidable before the creation of the digital model can commence. The presented approach aims to automate parts of this process by combining state-of-the-art computer vision techniques with optical character recognition to infer the interconnections between the views and thus reduce the manual effort. The general layout of a technical drawing is illustrated in Figure 1. Each drawing may contain multiple views, which, in turn, consist of a title and the respective geometry. Within each view, section symbols are detected. Each section symbol contains a reference which points to another view in the drawing collection. Also, the view titles are extracted and related to the individual symbols, and thus, the connections between views are identified. It is noted that all experiments are performed using bridge construction drawings, however, the method works independently of the depicted structure type and might also be applied to high-rise buildings. Presumably, it
343
might even be used for mechanical engineering drawings, as long as the used section symbol style is similar enough. The contribution of this paper is the proposal and the implementation of a concept to detect section symbols, recognize view titles and section reference characters, and to organize the extracted information in a semantic graph. In this regard, two distinct deep learning models are trained for their respective tasks purely with synthetic data, making manual annotation unnecessary. The paper is organized as follows: Section 2 provides an overview of the current state of research in automatic engineering drawing analysis. Next, Section 3 describes the implemented pipeline and presents the developed methods. The experimental results of the implementation are shown in Section 4. Finally, in Section 5, the results and limitations of this study are discussed, and an outlook for further research is formulated.
2 RELATED WORKS Limited research has been published towards drawing digitization (Moreno-García et al. 2019), and a minority of publications cover the automatic analysis of infrastructure construction drawings. Most of the studies in the realm of drawing analysis deal with floor plans or other drawing types of high-rise buildings. In recent years, most of these approaches include deep learning aspects, such as convolutional neural networks (CNNs): Lu et al. (2020) present a digital twin reconstruction method including the creation of a column grid. The grid’s nodes are detected using the Hough transform.Also, by means of a neuro-fuzzy network, building elements such as columns, beams and walls are localized in the drawing. Some additional semantic information is collected from text elements via optical character recognition (OCR). Leveraging all the gathered information, an industry foundation classes (IFC) file of the building story at hand is created. Similarly, Zhao et al. (2020) employ a YOLO (You only look once) object detector to localize structural elements in column layout plans and framing plan images. The authors expand their research in (Zhao et al. 2021) by incorporating the better performing Faster R-CNN (region-based CNN) model and by creating an IFC file from the extracted information. Also, drawing annotations are recognized by OCR and are used to attribute the created IFC elements. A broader body of literature is available on the analysis of floor plans, of which only a few research items are mentioned here. For instance, Smith (2007b) propose a method to detect and classify rooms, structural elements, symbols and text annotations in floor plans. They use the YOLO model for detecting regions of interest and perform a pixel-based segmentation of rooms and openings. The extracted layout is then cast into a vector graphics format. Kim et al. (2020) used a conditional generative adversarial network (CGAN) to
transform floor plans into a simplified style version. In this representation, it is better suited for the following vectorization process based on integer programming, which is adopted from (Liu et al. 2017). Kim et al. (2021) trained a modified version of the ResNet50 to recognize floor plan layouts in pixel images to vectorize them afterwards. Although not dealing with construction drawings, but electrical engineering drawings, Elyan et al. (2018) train multiple machine learning models, namely a support vector machine, a random forest and a CNN to recognize domain-specific symbols (e.g. valves, flanges and sensors) in piping and instrumentation diagrams (P&IDs). Prior to training, the unsupervised k-means clustering is run to find actual sub-classes in the data in order to improve the classification accuracy. In (Elyan et al. 2020), they train a CNN to detect and classify symbols. To account for the class-imbalance of symbol occurrences, a generative adversarial network is used to generate additional, synthetic training examples of the minority classes. Yun et al. (2020) develop a R-CNN-based model to localize and classify valves and instruments in P&IDs. To reduce the false positive rate, they also include negative examples in the training set. After applying an elaborate preprocessing pipeline, Yu et al. (2019) use a symbol detector based on AlexNet, the connectionist text proposal network (CTPN), for text recognition and apply traditional computer vision methods for line detection. Similarly, Mani et al. (2020) trained a CNN for symbol detection and used the efficient and accurate scene text (EAST) detector to localize text. As a novelty in this domain, they applied a graph-based approach to infer the line connections between the detected symbols. For automatic detection of symbols in as-built railway plans, Vilgertshofer et al. (2020) train several CNNs simultaneously. The first CNN detects the number of symbols that are present in the given drawing. Secondly, a network specialized on the number inferred by the first CNN is selected and used to predict the position and extent of the symbols. To increase the amount of training data without the need of manually labeling it, the authors develop a process to generate training data synthetically by cropping and assembling cutouts and symbols together. To the best of the authors’knowledge, there exists no publication focusing on inferring the interconnections between the respective views. This research aims to close the identified gap.
3 METHODOLOGY The end-to-end concept for the creation of semantic graph (see Figure 2) for the interconnections of drawing views is summarized in Figure 3. The input to the pipeline is a complete set of construction drawings for one structure, such that there are no dead-end references to views which are not in the drawing collection. The proposed method is based on the assumption that the drawings are already divided into the views, i.e.,
344
Figure 1. Exemplary layout of a technical drawing. It contains multiple views, consisting of a title and its geometry. Section symbols may occur throughout the drawing.
step (1) in the process is still up to the engineer. The next steps are to detect section symbols in all the views, and to detect the title of each view. With both pieces of information, the views can be logically linked together in pairs. The Sections 3.1 to 3.3 address the training of a section symbol detection model capable of step (2) in the process. The recognition of section references is conducted step (3) and briefly explained in Section 3.4. Section 3.5 deals with step (4), the recognition of the view titles. The merging of the collected information in step (5) is briefly described in Section 3.6. 3.1 Data set As described above, the subdivision of the drawings into its views is a prerequisite for the proposed approach. Thus, each view is processed individually, which allows the unambiguous association of views and section symbols. Since the division of a drawing into the views is not part of the automated process, it must be performed manually in advance. Within a view, the title and section symbol are detected. While pre-trained networks can be used for
title recognition, an object detection model has to be trained to localize the section symbols. A section symbol can be further split into two parts, namely section marker, and section reference. The marker describes the direction of the section axis and its viewing direction. The section reference is a unique identifier delivering the interconnection between the views. Typically, the reference is represented as a capitalized character. Two examples of section symbols with marker and reference are depicted in Figure 4. To train a deep learning model to correctly detect the three categories, namely section symbol, section marker, and section reference, a large amount of training data is necessary. This data must be gathered and annotated manually in a laborious process. It is noted that section symbols are rather small compared to the size of a drawing and sparsely occur. Ultimately, this leads to a significant increase of the amount of training data required. To overcome this issue, the training data set may comprise or be enriched by synthetic data. Inspired by the augmentation method proposed by Ghiasi et al. (2021), training images are synthesized based on copypaste operations. In the context of object detection, an
345
Figure 2. Exemplary part of the semantic graph to be generated from the drawing collection in Figure 1.
image can be divided into two types of regions: the foreground, which contains the objects to be recognized, and the background, which represents all the remaining regions. Adopting this differentiation to detection of section symbols, synthetic training images are composed of randomly cropped drawing regions, which serve as the background, and section symbols placed on top of it, which represent the foreground. Similar approaches can be found in (Vilgertshofer et al. 2020, Gupta et al. 2022). For generation of each synthetic image, the following process is carried out: 1. Select section marker template from a predefined set 2. Position section reference relative to marker 3. Augment marker and reference 4. Randomly crop region from drawing 5. Paste section symbol at random position on top of background 6. Calculate bounding box for marker, reference and symbol respectively 7. Save training sample In step (1) the section marker is selected from a set of templates. The set contains slightly different appearances of a section marker, for example, the triangle may not be filled or the line may be solid. This variety is necessary to various existing standards and drawing styles. It is a common phenomenon in actual drawings that the position of the reference relative to the section marker differs between occurrences, e.g., due to already occupied drawing areas nearby. To mimic this irregular characteristic, the reference is rotated and positioned arbitrarily around the marker in step (2). In step (3), both the marker and the reference are altered by resizing, stretching, mirroring and rotation, to further increase the diversity of the data. Next, a region with a defined size is cropped randomly from a drawing in step (4). This region serves as the background
Figure 3. Sequence of processing steps to generate semantic graph for view interconnection.
where the section symbol is placed in step (5) at a randomly chosen position. During the above described process, the position and extent of the marker and reference are continuously tracked, and are used to calculate the respective bounding boxes of the marker, reference, and section symbol in step (6). Finally, in step (7), the training image and annotation is saved. This process is repeated until the desired amount of training data is generated.Admittedly, the chosen number of generated training images is quite arbitrary, but for the proposed approach, 10.000 images are found to be sufficient. In Figure 4, examples of a synthetic and a real image sample are compared. Naturally, by this kind of data synthesis, the generated samples can be slightly unrealistic. For example, sections are typically oriented in perpendicular to other geometries in the image. This is not the case in the synthetic images. For the section detector training, however, this should not be an issue, as the general drawing style is still similar.
Figure 4. Comparison of a real (left) and a synthetic training image (right).
3.2 Data augmentation To ensure the generalizability of the detection network, the training data is augmented once again. In
346
the process, different techniques of data augmentation, described in Table 1, are applied sequentially to each image, i.e., the images are augmented independently from each other. The preprocessed images are batched and then fed to the object detection model to perform the training. Table 1. Used augmentation techniques for model training. The right column shows the probability of an image being altered by the respective transformation during online augmentation. Augmentation technique
Probability
Horizontal flipping Vertical flipping 90◦ , 180◦ or 270◦ rotation Translation Shearing along x-axis Shrinking
50% 50% 50% 25% 25% 25%
The online augmentation procedure is advantageous for two reasons: first, it reduces the amount of required disk space, since each instance of the augmented images is discarded after it is fed to the detection model. Second, it increases the degree of data diversity during training. This can also reduce the risk of overfitting, since no two training epochs contain the same training images.
3.3 Section symbol detection Given the augmented synthetic data set, an object detection model is trained to localize the section symbols, markers and references. Due to its proven performance, availability, and in-depth documentation, the Faster R-CNN architecture is chosen for the task (Ren et al. 2017). In this regard, a pre-configured implementation within the PyTorch framework (Paszke et al. 2019) is used for this study. As described in Section 3.1, three distinct classes for each section symbol a are detected in the input images: a section marker Ma , a section reference Ra and a symbol symbol Sa . The section references’ positions in relation to the section markers vary strongly. Therefore, it is necessary to distinguish between the three classes. For the robust detection of section markers and their respective section references, the raw bounding box detections are postprocessed. Algorithm 1 depicts the logic rules to handle imperfect detections as well as to relate the predictions to each other. In simple terms, the section symbol bounding boxes serve as grouping agents for the actual objects of interest within them. The detection model is trained with a batch size of 32 and a stochastic gradient descent optimizer. The learning rate is scheduled according to the 1cyclepolicy (Smith and Topin 2019) with a maximum learning rate of 1 × 10−3 .
Algorithm 1 Prediction coupling 1: inputs: Sets of detected objects 2: - Section markers M = {m0 , m1 , . . . , mj } 3: - Section references R = {r0 , r1 , . . . , rk } 4: - Section symbols S = {s0 , s1 , . . . , si } 5: for s ∈ S do 6: for m ∈ M do 7: if s encloses m then 8: Couple(s, m) 9: continue with next s 10: end if 11: end for 12: if s has no coupled m and s not enlarged then 13: Enlarge s 14: goto line 6 with enlarged s 15: end if 16: repeat line 6 with r ∈ R analogously 17: end for 18: return All coupled (s, m) and (s, r)
3.4 Section reference recognition To establish the connection between views, the unique identifier of each section must be recognized. Since the reference position is already predicted by the model presented in Section 3.3, the task at hand is to recognize the character of the symbol reference. In a first attempt, the OCR engine Tesseract (Smith 2007a) is used to identify the reference character of the detected section symbol. Despite various trials with different preprocessing techniques, including thresholding, rotation, or erosion and dilation, the engine showed poor results. Presumably, neighboring lines are falsely recognized as characters as well. However, since the images are quite cluttered, there is no obvious remedy to this problem. As an alternative approach, a custom neural network is developed and trained for the described task. The architecture consists of four blocks, containing convolutional layers for feature extraction combined with max-pooling layers to reduce the spatial dimension. The resulting feature vectors are flattened and passed to dense layers that finally predict one of the 26 characters in the English alphabet. The training data is generated based on the process in Section 3.1. The difference here is that only the section reference is pasted onto the background. Also, rather than the bounding box of the character, its class is important in this case, i.e., a character ID. In addition, the training images are significantly smaller, as it is assumed that the references are already well localized by the symbol detection model. Since this is a problem of image classification, as a loss function for the final layer, the categorical cross-entropy is chosen, and the model is trained with an Adam optimizer. Initially, the training is conducted with a learning rate of 1e-3, which is gradually reduced by a factor of 10 if
347
the validation loss stagnates for a few epochs. The network is trained over 70 epochs and shows promising results.
3.5 View title recognition The recognition of the view titles is mainly based on OCR and string matching. First, the upper 30% of each image is separated from the rest and fed into the OCR engine Tesseract (Smith 2007a). This implies the assumption that the title occurs at the top of the view, which holds true for all the data at hand. Tesseract outputs a list of found strings, which is then scanned for certain key strings such as view, section, or detail. Based on a comprehensive key string dictionary, all title candidates are found, and the first to appear on the view (the most upper-left) is chosen as the respective title. In the case that a view is a section view, the section identifier, e.g., A-A, is extracted from the title as a separate piece of information.
3.6 Pipeline output Without further postprocessing, the predictions obtained with the methods described above, are individual pieces of information that need to be linked together. Therefore, the extracted information representing the section marker positions, the section references, view titles and section view identifiers are all stored in an output JSON file. This file is structured according to the knowledge graph structure illustrated in Figure 2.
4 VALIDATION AND TEST RESULTS To give an impression of the trained models’ performances, they are tested with regard to selected object detection metrics. It is noted that the detection of section symbols in this specific case is subject to the following considerations: – The evaluation should focus on the recall score R since the occurrence of many false positives is preferred over many false negatives. False positives may require specialized treatment, however, due to the joint occurrence of section markers and section references, they can be handled quite easily. – The intersection over union (IOU) threshold to consider a detection a true positive can be assumed quite low. This is because, after the raw detection of an object, the predicted bounding box can always be enlarged to make sure that the object of interest is included. In this study, the authors select a value of tiou = 0.50 for the experiments. – For better comparability with similar studies, the class-wise average precision AP scores are included as well. Since a high AP depends on both a high recall and a high precision score, it is quite representative for the overall model performance.
For more detailed definitions of the used performance metrics, the reader is referred to (Padilla et al. 2021). From the training process it is known that the model performs well on a validation data set created as described in Section 3.1. However, due to the synthetic character of the data set, this result is not representative for the actual use of the model. Therefore, a set of 56 real bridge construction drawings is considered for testing the model. They are provided by SD Engineering – A Socotec Company and depict a ramp leading to a highway bridge. In total, 274 views are included, depicting 425 section markers and 403 section references. The difference in the numbers is due to 22 occurrences of section markers without any reference. Since, during the training process, the model is confronted with tiles of fixed size, and scaling similar to real drawings, this should also be the case during testing. Each input drawing is therefore sliced, prior to inference, with a tile side length of 1300 pixels. This yields a total of 951 image files, of which 251 contain at least one of the relevant classes. The report of the results consists of two parts: the validation of the models with synthetic data, which is generated using the same techniques as for the training data, and the testing of the models on actual bridge construction drawings. The results are summarized in Table 2. As far as the validation goes, the models for symbol detection and character recognition are tested by applying them to separately generated data, which has not been used for training. Both show reasonable performances on the validation sets. For the test on real images, it should be noted that the character recognition depends on the correct, prior detection of the section reference bounding box. Therefore, the character recognition accuracy is evaluated based on only the correctly detected references. Examples of positive and negative results can be seen in Figure 5. Given the results of reference detection and recognition, the proposed pipeline achieves an overall reference recognition accuracy of 47.64%. Concluding from the precision-recall curves depicted in Figure 6, section references are detected with a much better precision at high recall values than section markers.
Figure 5. Exemplary presentation of results, positive results on left side, poor detection results on the right.
348
in drawing style or symbol shape is easily handled by automatically creating a new training data set from the drawing collection at hand. The common obstacle of acquiring enough data to train a machine learning model is, therefore, quite elegantly handled. However, the developed program still has some shortcomings: While the section detection, the title recognition and the linking of views is fully automated, the method assumes the input drawings are already split up into views, i.e., they have to be manually divided beforehand. Also, the title recognition algorithm is specifically tailored toward a certain vocabulary and may not generalize well for all unseen data. On a further note, views which are not considered sections (e.g., detail views) are excluded from this study. The authors plan future works which include, first, a comparative study of different object detection model architectures and augmentation techniques to improve the section symbol detection accuracy. Second, the proposed pipeline should be extended such that the views are not only logically connected, but also correctly arranged in 3D space. Finally, machine learning methods will be used to automate the segmentation of entire 2D drawings into views.
Table 2. Object detection performance scores achieved by the proposed model on the synthetic images (validation) and real bridge construction drawings (test).
Validation data – markers – ref. boxes – ref. characters – symbols Test data – markers – ref. boxes – ref. characters – symbols
Object detection R AP
Char. recognition Accuracy
99.20% 94.70% – 99.20%
99.13% 96.39% – 99.42%
– – 82.61%
48.81% 58.40% – 51.00%
47.85% 54.44% – 55.27%
– – 81.59%
ACKNOWLEDGMENTS
Figure 6. Precision-recall curves for the detection of section symbols, section markers and section references for the validation set and the test set.
5 CONCLUSIONS While literature offers various information extraction approaches regarding high-rise drawings or symbol detection methods in other engineering drawings, limited research has been published towards the analysis of bridge construction drawings. This study’s contribution is a novel approach to technical drawing analysis. The proposed method has the potential to aid engineers in the tedious task of finding the interconnections between the numerous views belonging to a collection of drawings. It is shown that by using object detection and OCR methods a process can be set up to automatically infer view connections. Testing the proposed methodology on actual bridge drawings yields promising performance. Moreover, it should be noted that method does not rely on manual data annotation at all. The OCR engine is pre-trained, and the section symbol detector is trained on synthetic data with automatically generated labels. Presumably, this means that even a change
This research is conducted as part of the BIMKIT project, funded by the German Federal Ministry for Economic Affairs and Climate Action. The authors would like to express their gratitude towards Markus Scheffer from SD Engineering – A Socotec Company, who generously provided the drawing data set. REFERENCES Elyan, E., Garcia, C. M., & Jayne, C. (2018). Symbols Classification in Engineering Drawings. In International Joint Conference on Neural Networks, pp. 1–8. Elyan, E., Jamieson, L., & Ali-Gombe, A. (2020). Deep Learning for Symbols Detection and Classification in Engineering Drawigns. Neural Networks 129, 91–102. Ghiasi, G., Cui, Y., Srinivas, A., Qian, R., Lin, T.-Y., Cubuk, E. D., Le, Q. V., & Zoph, B. (2021). Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2917–2927. Gupta, M., Wei, C., & Czerniawski, T. (2022). Automated Valve Detection in Piping and Instrumentation (P&ID) Diagrams. In Proceedings of the 39th ISARC, pp. 630– 637. Kim, H., Kim, S., & Yu, K. (2021). Automatic Extraction of Indoor Spatial Information from Floor Plan image: A Patch-Based Deep Learning Methodology Application on Large-scale Complex Buildings. ISPRS International Journal of Geo-Information 10(12), 828. Kim, S., Park, S., Kim, H., & Yu, K. (2020). Deep floor plan analysis for complicated drawings based on style transfer. Journal of Computing in Civil Engineering 35(2), 04020066.
349
Liu, C., Wu, J., Kohli, P., & Furukawa, Y. (2017). Rasterto-Vector: Revisiting Floorplan Transformation. In IEEE International Conference on Computer Vision, pp. 2214– 2222. Lu, Q., Chen, L., Li, S., & Pitt, M. (2020). Semi-automatic Geometric Digital Twinning for Existing Buildings Based on Images and CAD Drawings. Automation in Construction 115, 103183. Mani, S., Haddad, M. A., Constantini, D., Douhard, W., Li, Q., & Poirier, L. (2020). Automatic Digitization of Engineering Diagrams Using Deep Learning and Graph Search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 176–177. Moreno-García, C. F., Elyan, E., & Jayne, C. (2019). New Trends on Digitisation of Complex Engineering Drawings. Neural Computing and Applications 31(6), 1695–1712. Padilla, R., Passos, W. L., Dias, T. L. B., Netto, S. L., & da Silva, E. A. B. (2021). A Comparative Analysis of Object Detection Metrics with a Companion Open-source Toolkit. Electronics 10(3), 279. Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., & Chintala, S. (2019). PyTorch: An Imperative Style, High-performance Deep Learning Library. In Proceedings of the 33rd International Conference on Neural Information Processing Systems, pp. 8026–8037. Ren, S., He, K., Girshick, R., & Sun, J. (2017). Faster R-CNN: Towards Real-time Object Detection with Region Proposal Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence 39(6), 1137–1149. Sacks, R., Kedar, A., Borrmann, A., Ma, L., Brilakis, I., Hüthwohl, P., Daum, S., Kattel, U., Yosef, R., Liebich, T.,
Barutcu, B. E., & Muhic, S. (2018). SeeBridge as Next Generation Bridge Inspection: Overview, Information Delivery Manual and Model View Definition. Automation in Construction, 134–145. Smith, L. N. & Topin, N. (2019). Super-convergence: Very Fast Training of Neural Networks Using Large Learning Rates. In Artificial Intelligence and Machine Learning for Multi-Domain Operations Applications, Volume 11006, pp. 369–386. Smith, R. (2007a). An Overview of the Tesseract OCR Engine. In 9th International Conference on Document Analysis and Recognition, pp. 629–633. Smith, R. (2007b). Residential Floor Plan Recognition and Reconstruction. In 9th International Conference on Document Analysis and Recognition, pp. 629–633. Vilgertshofer, S., Stoitchkov, D., Borrmann, A., Menter, A., & Genc, C. (2020). Recognising Railway Infrastructure Elements in Videos and Drawings Using Neural Networks. In Proceedings of the Institution of Civil Engineers – Smart Infrastructure and Construction, pp. 19–33. Yu, E.-S., Cha, J.-M., Lee, T., Kim, J., & Mun, D. (2019). Features Recognition From Piping and Instrumentation Diagrams in Image Format Using a Deep Learning Network. Energies 12(23), 4425. Yun, D.-Y., Seo, S.-K., Zahid, U., & Lee, C.-J. (2020). Deep Neural Network for Automatic Image Recognition of Engineering Diagrams. Applied Sciences 10(11), 4005. Zhao,Y., Deng, X., & Lai, H. (2020). A Deep Learning-based Method to Detect Components From Scanned Structural Drawings for Reconstructing 3D Models. Applied Sciences 10(6), 2066. Zhao, Y., Deng, X., & Lai, H. (2021). Reconstructing BIM From 2D Structural Drawings for Existing Buildings. Automation in Construction 128, 103750.
350
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
An ontology-supported case-based reasoning approach for damage assessment A.H. Hamdan & R.J. Scherer Institute of Construction Informatics, Technische Universität Dresden, Germany
ABSTRACT: Despite new technologies in machine vision allow for an automated damage detection, current practices in damage assessment rely mainly on manual evaluations by human experts. Although some new approaches propose a damage assessment via machine learning methods, essential contextual information about the damaged construction is not considered. Contrary to this, knowledge-based approaches have been researched. However, knowledge bases for damage assessment usually contain certain knowledge gaps that result in uncertainties, which still need to be solved manually by experts. Therefore, in this paper a new theoretical approach that utilizes case-based reasoning (CBR) is discussed as additional method for automated damage assessment, which could be utilized in conjunction with knowledge-based approaches. Thereby, the case base of the CBR system would be developed as ontology utilizing the Web Ontology Language (OWL) to be compatible with current knowledge-based approaches, especially the Damage Topology Ontology (DOT).
1 INTRODUCTION Within the last decade, various approaches have been developed for the automated assessment of detected structural damage in constructions. Among existing artificial intelligence technologies for damage assessment, a distinction can be made between analogy-based approaches of machine learning and causal-based approaches of knowledge processing. Current machine learning approaches are primarily used for detection and classification since methodologies like deep learning or convolutional networks are well suited for the recognition of damage patterns (Huang et al. 2018; Hüthwohl et al. 2019; Li et al. 2018). However, these approaches usually consider only the direct features of a damage and ignore contextual information e.g., about the damaged construction or building material, since it is often difficult to cover these data by an appropriate amount of training data. In contrast, knowledge-based approaches consider the buildings context without any training data but rather through digitizing existing expert knowledge in an ontology via rules, which could then be processed by a reasoning engine (Hamdan & Scherer 2019; Hu et al. 2019; Kreyenschmidt 2021; Ren et al. 2019). However, contrary to machine learning approaches, numerical data, such as geometry or scan results, is difficult to process by knowledge-based reasoning. Furthermore, knowledge gaps in the ontology could lead to uncertainties or no results in the damage assessment. To solve the issues of current knowledge-based
DOI 10.1201/9781003354222-45
approaches, a new method of damage assessment is proposed in this paper, which utilizes case-based reasoning (CBR). In CBR previously solved cases to a specific problem are structured in a case base utilizing expert knowledge. The solution to a new problem is then solved based on comparing similarities between the cases in the case base, which in consequence is an analogy-based approach, similar to machine learning methods. However, contrary to most machine learning concepts, CBR requires a significantly smaller amount of data. Therefore, in the CBR-based damage assessment approach previously evaluated damage scenarios are stored in a case base. In this regard, each damage scenario contains information about the damage properties as well as additional contextual information. The case base is modelled as an ontology formalized in the Web Ontology Language (OWL), which allows for knowledge-based consistency checks and reasoning. When a newly detected damage should be assessed, a query, which is formalized in SPARQL, is applied on the case base, containing selected parameters about the damage and construction. By evaluating the similarity between the newly detected damage and the retrieved cases by the SPARQL query, various fitting damage scenarios could be determined, depending on the size and variance of the case base. The proposed approach is discussed, and an exemplary use case is presented in which bending cracks are assessed. Table 1 contains an overview of the used prefixes throughout this research paper.
351
Table 1.
Namespaces and prefixes used in this paper.
Prefix
Namespace
rdf owl bot dot cdo aoi beo props inst
http://www.w3.org/2000/01/rdf-schema# http://www.w3.org/2002/07/owl# https://w3id.org/bot# https://w3id.org/dot# https://wisib.de/ontologie/damage/cdo/ https://w3id.org/aoi# https://pi.pauwel.be/voc/buildingelement# https://w3id.org/product/props/ Prefix for objects of the example case
2 RELATED WORK Already in the early days of the Semantic Web, possible approaches for linking CBRs with ontologies were discussed e.g., by Bergmann & Schaaf (2003), which concluded that strong synergies exist between ontology and CBR on a technological as well as methodological level, but probably require a unified knowledge representation of the similarity model. Currently, approaches to creating case bases in OWL, which are used in the context of CBR, have already been implemented in other disciplines, such as medicine (Oyelade & Kana 2019) or mechanical engineering (Mabkhot et al. 2019). In the architecture, engineering, and construction (AEC) domain, similar approaches have been developed that utilize an ontology consisting of previously modelled cases that are queried against various criteria to determine the best fitting case for a certain problem. For instance, Benevolenskiy et al. (2012) developed a case base in which predefined process patterns were queried and adapted for certain construction tasks via rules, e.g. for the concreting of a column. However, the retrieval of cases was not processed based on similarity of previously solved cases but rather through querying manually configured reference processes. These reference processes were then adapted to the specific project context, similarly to CBR. In the field of damage and defect assessment, Lee et al. (2016) proposed a method for storing cases of construction defects in a RDF triplestore and querying via SPARQL for similar cases when identifying a new construction defect. Furthermore, Xu et al. (2018) developed a CBR system for evaluating defects and identifying possible maintenance solutions, although in this approach the cases were not structured in an ontology.
3 PROCESS MODEL CBR is characterized through multiple subsequent phases that are defined in a process model. Based on the proposed CBR cycle of Aamodt & Plaza (1996) a process model has been developed for CBR-based damage assessment, which is shown in Figure 1.
Figure 1. Process model of the CBR damage assessment (based on the CBR cycle of Aamodt & plaza (1996).
3.1 Damage case retrieval In CBR each case is defined as a pair of a specific problem and a corresponding solution to this problem. Therefore, when a new problem is identified, possible solutions from the case-base of a compatible CBR system could be determined, by retrieving cases with a similar problem and inferring that their corresponding solution is a similar solution to the new problem. Usually, the pair of problem and solution are static and predefined for each case. However, the utilization of ontologies enables an approach, in which the problem and solution can be dynamically defined through the query that retrieves the cases from a given case-base. In the approach proposed in this paper each case is defined as knowledge graph that consists of at least one damage. Thereby, each damage is connected to related properties and contextual objects e.g., affected building elements or adjacent damages, which are also related to other properties and objects. In this regard, the knowledge graph alone does not specify any problem or solution. Therefore, the problem and required solution are specified in a SPARQL query, which is used for retrieving similar cases based on the properties of the damage case, which needs to be assessed. Outsourcing the problem and solution specifications to a machine-readable query statement allows for automatically configuring assessment criteria based on the recorded damage properties and contextual information. For instance, the properties of a crack, such as its height, length, and depth as well as the type of the affected building element could be determined through an initial SPARQL query. In a subsequent step, these properties could be utilized in an algorithm that automatically generates a SPARQL query for case retrieval. Thereby, for each numerical data property e.g., the crack height, a tolerance range is
352
added to define the range of similarity for the cases to be retrieved. 3.1.1 Case similarity When processing CBR, an important issue of the case retrieval is how the similarity between the new case and the retrieved cases is determined. The similarity between two comparable objects of the sets A and B that belong to different cases is usually described through a value between 0 and 1 (see eq. 1) sim : A × B → [0, 1]
(1)
Through the predicate in eq. 1, a case retrieval according one specific criterion could be processed. However, usually multiple criteria need to be considered when comparing the similarity between cases. Therefore, the local similarity measures between cases need to be summarized to a global similarity measure via a constructor function C (see eq. 2). sim = C(simi |i ∈ I )
(2)
Various functions are valid for application in a CBR-based damage assessment. For example, a well applicable function would be the hamming distance method. Thereby, the similarities of two property sets ai and bi , which belong to different cases, are compared with each other (see eq. 4). In addition, a weighting vector g is applied on the comparison method, for weighting the various properties (see eq. 3). g = (g1 , . . . , gn ) , 0 ≤ gi ≤ 1
(gi × simi (ai , bi )|1 ≤ i ≤ n)
(3) (4)
The similarity measures are difficult to consider, when retrieving cases through a single SPARQL query. Therefore, an approach utilizing multiple SPARQL queries in subsequent process steps is recommended. When filtering for damage cases with similar attributes, it is possible to consider similarities through variation of the value range constraints for each case property. Furthermore, constructor functions could be processed in subsequent SPARQL queries e.g., the hamming distance method, to calculate the similarity measurement for each case. The cases could then be sorted based on the similarity to the new damage case that needs to be assessed. 3.2 Damage case reuse After a fitting case has been retrieved, the solution values that are sought can be determined from it. This process could be handled in the same SPARQL query that is also used for retrieving the damage case. For instance, the cause of a newly detected damage is unknown, and a similar damage case has been queried from the case-base. By determining the cause of the queried damage, it could be assumed that the new damage has the same cause. However, it is not guaranteed
that the solution of one case is identical to another, just because they have a high similarity. For this reason, the solution of the retrieved damage case that fits best to the new damage case will be adapted by utilizing predefined adaptation rules. Thereby, the adaptation is a knowledge-based process. The rules could be defined in a specific rule language, such as the Semantic Web Rules Language (SWRL) or the Shapes Constraint Language (SHACL). 3.3 Damage case revise The revision phase is necessary to check and validate the correctness of the determined solution from the case retrieval and case adaptation. This phase could be processed either manually by an expert or, at least to some extent, automatically trough validation rules or algorithms. One potential method for automatic revision would be the application of SHACL shapes that check the resulting case solution based on predefined constraints. The output result could then be exported as validation report. It should be mentioned that the revision process is not limited to rule-based methods. Furthermore, other validation techniques, such as structural analysis or damage prediction simulations could be utilized e.g., for predicting the development of reinforcement corrosion due to chloride ingress (Hájková et al. 2018). 3.4 Damage case retain If the new damage case provides a valid result, it could be integrated into the existing case-base. Through this step, the CBR system will be enriched and optimized. For this reason, the CBR method is often classified as machine learning concept, since the CBR optimization cycle could in theory be processed automatically and fully machine based.
4 KNOWLEDGE MODEL Similarly, to knowledge-based systems, CBR utilizes knowledge models, such as ontologies to solve given problems. In CBR it is established to utilize four types of knowledge models, also called knowledge container. These types are (1) vocabulary, (2) similarity measures, (3) adaptation knowledge and (4) case base (Roth-Berghofer 2004). Figure 2 shows how the knowledge containers are structured and interacting in an ontology based CBR damage assessment. The vocabulary of a CBR system consists of general knowledge about the objects and their properties that define the case problems and solutions. In the knowledge base terms, the resulting vocabulary can also be called terminological box or Tbox. For the ontology-based approach proposed in this paper web ontologies, which are formalized in OWL, are used. Benefits of using OWL ontologies as vocabulary are for example:
353
Figure 2. Knowledge containers used in the ontology based CBR damage assessment.
– Unambiguous identification of terms and properties on a global scale via Uniform Resource Identifiers (URI) – Easily reusable knowledge of various domain specific ontologies by utilizing Linked Open Data principles (Berners-Lee 2009) – Usage of established Semantic Web technologies e.g., SPARQL for querying OWL ontologies or SHACL for ontology validation – OWL statements (also called OWL axioms) are based on description logic. Thereby decidability is guaranteed in the language level OWL DL or lower. For modelling damage cases, existing web ontologies for semantically describing constructions and damages should be used as vocabulary. Thereby, relevant information for damage assessment is the construction topology as well as the localization of damage on a component. For defining the topology, the Building Topology Ontology (BOT) (Rasmussen et al. 2020) could be used, which provides vocabulary for describing relations between building zones, contained building elements as well as connecting interfaces. For describing the topology of bridges, the Bridge Topology Ontology (BROT) (Hamdan & Scherer 2020) could be used in a similar way. Alternatively, a more IFC focused method could be utilized, in which the ifcOWL ontology (Pauwels & Terkaj 2016) is applied as semantic building representation. In this regard, the topology is described through objectified relationships that connect building elements and zones with each other, such as IfcRelAggregates or IfcRelContainedInSpatialStructure. Furthermore, IfcOWL provides additional classes and properties for characterizing the building and its components, such as
building types or geometrical properties. When working with BOT or BROT these additional vocabulary needs to be integrated via other ontologies e.g., the Building Element Ontology (BEO) (Pauwels 2018) or the Building Product Ontology (BPO) (Wagner et al. 2022). For semantically specifying the spatial locations for components in which damages occurs, the Area of Interest Ontology (AOI) (Hamdan & Scherer 2021) could be used. For generically describing damages in an existing construction, the Damage Topology Ontology (DOT) (Hamdan et al. 2019) could be used. In this regard, DOT can be extended with domain specific vocabulary, such as web ontologies for describing damages in reinforced concrete or natural stone facades (Seeaed & Hamdan 2019) or ontologies that add classes and properties for additional functionality, such as an ontology for describing damage mechanics. The vocabulary for describing constructions and the damages that affect them is then used for defining specific damage cases. Furthermore, the vocabulary is referenced in the rules that are used for the adaptation of a case in the reuse phase as well as in the queries that define the problem and solution criteria for the retrieval process. The similarity measures are not explicitly part of the vocabulary but instead defined in the SPARQL queries that are used in this ontological CBR approach. Although specific terminology for defining and grouping the cases of the case base could be added, it is not necessary for the proposed SPARQL-based CBR approach. This is since the problem and the required solution, which form the case, are defined in the SPARQL query. By adjusting the query parameters, new cases can be flexible defined from the same knowledge base.
5 EXEMPLARY APPLICATION The procedure described above in section 3 and 4 is illustrated below using a potential application example. Figure 3 shows how a potential damage case could be modelled utilizing the terminology of existing building and damage ontologies. It should be noted that the example has been simplified to some extend to improve the readability and comprehensibility of the proposed principle. For instance, further information about the damaged building material or adjacent building components are not presented, although they have a significant impact on the damage assessment. The example case describes a concrete beam (inst:ExampleBeam01) that is damaged by a vertical crack (inst:ExampleCrack01). Thereby, the location of the crack in relation to the beam is semantically described through an AOI object (inst:DamagedArea01). The crack is classified as bending crack and transverse crack through respective classes of the Concrete Damage Ontology (CDO). Through reasoning
354
the highest similarity to the assessed case on top of the result set. A potential goal of this query would be the retrieval of similar cracks from the case base to determine possible crack classifications for a crack that is not fully classified. Listing 1. SPARQL query for retrieving all relevant damage cases from the case base and ranking them based on their similarity to the new damage case.
Figure 3. Exemplary damage case describing a crack that affects a beam.
it could be inferred that these classes are a subclass of cdo:Crack, which is a subclass of the DOT class dot:ClassifiedDamage and therefore connected to the DOT ontology. Furthermore, three exemplary attributes were assigned to the crack that define its length, width, and the angle of its crack path. In a similar way the beam is classified via the beo:beam class from BEO. The width, height and span of the beam could be defined through respective data properties e.g., from the PROPS ontology1 , which is still in a conceptual phase. The AOI object that connects the damage representation with the beam representation via the object properties aoi:locatesDamage and aoi:hasAreaOfInterest is also classified as horizontal central area in the lower half of the affected beam. Thereby, AOI provides the functionality of inferring the direct relation between damage and construction component, which is defined via the DOT property dot:hasDamage through semantically reasoning an OWL property chain axiom. It is not necessary to use specific CBR terminology such as classes or properties that need to be defined in an auxiliary ontology. Instead, each damage and associated knowledge graph could be defined as one case that could be retrieved and reused in the CBR cycle. Listing 1 shows a possible SPARQL query for retrieving all damages that fulfill fundamental criteria, such as the affiliation to certain classes or properties and determining the similarity value to a given damage case that needs to be assessed against the CBR system. The retrieved cases are then listed in ascending order based on their similarity, thus ranking the case with 1
https://github.com/maximelefrancois86/props
The first part of the query defines the obligatory criteria that need to be fulfilled for every damage case of the result set. In this regard, each damage must affect at least an instance of beo:Beam via dot:hasDamage and must be classified as crack through the respective class cdo:Crack. In the second part of the example query, the equation for determining the similarity of each retrieved case to the given case is defined. Thereby, the overall similarity considering the three values crack length, crack width and crack angle is determined using the hamming distance method (see eq. 3 & 4 in section 3.1.1). In this regard, for each considered value the difference between the value from a damage case of the case base and the value of the given case is determined and weighted according to a specific weight factor. Finally, the sum of all weighted factors is determined, which represents the overall similarity value that is used for ranking the retrieved cases in ascending order. Through applying this query for a crack that is located on a beam, the exemplary case from Figure 3 could be retrieved thus potentially classifying the new case as bending crack. Based on the characteristics of the damage case to assess, further adaptation rules can be applied that are based on existing expert knowledge. For instance, an indicator for the impact on the structural capacity could be adapted based on specific damage properties, such as the crack width or crack length. It would be subject of further research, to which extend adaptation rules would be necessary in conjunction with the CBR case base to provide accurate and practical damage assessment results. Similarly, rules or shapes could be used for consistency checking and validating the correctness of the damage assessment result. For instance, it must be ensured that a bending crack of relevant size
355
is classified as damage that is significant for the structural capacity of the construction.
6 CONCLUSION In this paper a theoretical approach is proposed for assessing construction damages automatically through CBR, whereby the case base of the CBR is structured as ontology. To utilize existing Semantic Web technologies, OWL is recommended as format to develop the ontological case base as well as SHACL as language for rules and validation shapes to support the reuse and revise cycles of the CBR process. Nonetheless, the proposed approach is not limited to certain formats thus other ontology formats or rule languages could be utilized. One key aspect of utilizing an ontology or knowledge-graph in general as case base for CBR is the overall flexibility in structuring problems and solutions for each case. Instead of clearly distinguishing between problem and solution inside each case, a knowledge graph is defined that just represents information about a damage. The specific problem and corresponding solution are then defined externally in the query used for retrieving the cases and thus could changed flexible depending on the problem and required solution. This also includes defining a query function for determining the similarity between the cases compared to the damage case that needs to be assessed. A potential structure for the case base as well as a sample query for retrieving damage cases for further assessment has been presented in this paper. However, it is still subject of future research to test and verify this approach in a pilot project utilizing a fully developed case base. It must be also researched to what extend adaptation rules should support the CBR case base, especially whether the case base should be structured with a small number of cases but a high utilization of adaptation rules or vice versa. Furthermore, it must be researched, how a revise cycle for verifying the damage assessment results should be managed. Although, a manual revision by a human expert would be feasible, computer-based methods e.g., the utilization of SHACL validation shapes or damage simulations, could lead to an automatization of the revise cycle. CBR could enhance current approaches that mainly rely on knowledge-based assessment of construction damages (Hamdan & Scherer 2019, 2021), especially since the case base could in theory be compatible with current ontologies, such as DOT. It is subject of future research, to develop an appropriate case base and verify the approach in one or multiple test scenarios. REFERENCES Aamodt, A., & Plaza, E. 1996. Case-based Reasoning: Foundational Issues, Method Ological Variations, and System Approaches. Artificial Intelligence Communications, 7(1), 39–59.
356
Benevolenskiy,A., Roos, K., Katranuschkov, P., & Scherer, R. J. 2012. Construction Processes Configuration Using Process Patterns. Advanced Engineering Informatics, 26(4), 727–736. https://doi.org/10.1016/j.aei.2012.04.003 Bergmann, R., & Schaaf, M. 2003. Structural Case-Based Reasoning and Ontology-Based Knowledge Management: A Perfect Match? In Journal of Universal Computer Science (Vol. 9, Issue 7). Berners-Lee, T. 2009. Linked Data. https://www.w3.org/ DesignIssues/LinkedData.html. Accessed: 03.06.2022 ˇ Hájková, K., Šmilauer, V., Jendele, L., & Cervenka, J. 2018. Prediction of Reinforcement Corrosion Due to Chloride Ingress and its Effects on Serviceability. Engineering Structures, 174(August), 768–777. https://doi.org/10.1016/j.engstruct.2018.08.006 Hamdan, A., Bonduel, M., & Scherer, R. J. 2019. An Ontological Model for the Representation of Damage to Constructions. 7th Linked Data in Architecture and Construction Workshop. Hamdan, A., & Scherer, R. J. 2019. A Knowledge-based Approach for the Assessment of Damages to Constructions. 36th CIB W78 2019 Conference. Hamdan, A., & Scherer, R. J. 2020. Integration of BIMrelated Bridge Information in an Ontological Knowledgebase. Linked Data in Architecture and Construction Workshop. Hamdan, A., & Scherer, R. J. 2021. Assumption of Undetected Construction Damages by Utilizing Description Logic and Fuzzy Set Theory in a Semantic Web Environment. European Conference on Product and Process Modelling 2020-20212, 245–252. Hamdan, A., & Scherer, R. J. 2021. Areas of Interest – Semantic Description of Component Locations for Damage Assessment. EG-ICE 2021 Workshop on Intelligent Computing in Engineering. Hu, M., Liu, Y., Sugumaran, V., Liu, B., & Du, J. 2019. Automated Structural Defects Diagnosis in Underground Transportation Tunnels Using Semantic Technologies. Automation in Construction, 107(August), 102929. https://doi.org/10.1016/j.autcon.2019.102929 Huang, Z., Fu, H., Chen, W., Zhang, J., & Huang, H. 2018. Damage Detection and Quantitative Analysis of Shield Tunnel Structure. Automation in Construction, 94(February), 303–316. https://doi.org/10.1016/j.autcon.2018. 07.006 Hüthwohl, P., Lu, R., & Brilakis, I. 2019. Multi-classifier for Reinforced Concrete Bridge Defects. Automation in Construction, 105(December 2018), 102824. https://doi.org/10.1016/j.autcon.2019.04.019 Kreyenschmidt, C. 2021. Domain Ontology for the Classification of Wood Destroying Insects in Existing Timber Constructions. Forum Bauinformatik 2021. Lee, D. Y., Chi, H. lin, Wang, J., Wang, X., & Park, C. S. 2016. A Linked Data System Framework for Sharing Construction Defect Information Using Ontologies and BIM Environments. Automation in Construction, 68(May), 102–113. https://doi.org/10.1016/j.autcon.2016.05.003 Li, D., Cong, A., & Guo, S. 2018. Sewer Damage Detection From Imbalanced CCTV Inspection Data Using Deep Convolutional Neural Network with Hierarchical Softmax. Submited to: Automation in Construction, 101(June 2018), 199–208. https://doi.org/10.1016/j.autcon.2019. 01.017 Mabkhot, M. M., Al-Samhan, A. M., & Hidri, L. 2019. An Ontology-enabled Case-based Reasoning Decision Support System for Manufacturing Process Selection. Advances in Materials Science and Engineering, 2019. https://doi.org/10.1155/2019/2505183
Oyelade, O. N., & Kana, A. F. D. 2019. ’OWL Formalization of Cases: An Improved Case-based Reasoning in Diagnosing and Treatment of Breast Cancer. Int. J. Inf. Secur., Privacy Digit. Forensics, September 2020. Pauwels, P. 2018. Building Element Ontology. https://pi. pauwel.be/voc/buildingelement/index-en.html.Accessed: 03.06.2022 Pauwels, P., & Terkaj, W. 2016. EXPRESS to OWL for Construction Industry:Towards a Recommendable and Usable ifcOWL Ontology. Automation in Construction, 63, 100– 133. https://doi.org/10.1016/J.AUTCON.2015.12.003 Rasmussen, M. H., Lefrançois, M., Schneider, G. F., & Pauwels, P. 2020. BOT: The Building Topology Ontology of the W3C Linked Building Data Group. Semantic Web, 12(1), 143–161. https://doi.org/10.3233/SW-200385 Ren, G., Ding, R., & Li, H. 2019. Building an Ontological Knowledgebase for Bridge Maintenance. Advances in Engineering Software, 130(July 2018), 24– 40. https://doi.org/10.1016/j.advengsoft.2019.02.001
357
Roth-Berghofer, T. R. 2004. Explanations and Case-based Reasoning: Foundational Issues. Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 3155(August 2004), 389–403. https://doi.org/10.1007/9783-540-28631-8_29 Seeaed, M. K., & Hamdan, A. 2019. BIMification of Stone Walls for Maintenance Management by Utilizing Linked Data. 31st Forum Bauinformatik, November. Wagner, A., Sprenger, W., Maurer, C., Kuhn, T. E., & Rüppel, U. 2022. Building Product Ontology: Core Ontology for Linked Building Product Data. Automation in Construction, 133(October 2021), 103927. https://doi.org/10.1016/j.autcon.2021.103927 Xu, Z., Li, S., Li, H., & Li, Q. 2018. Modeling and Problem Solving of Building Defects Using Point Clouds and Enhanced Case-based Reasoning. Automation in Construction, 96(February), 40–54. https://doi.org/10.1016/j.autcon.2018.09.003
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Towards a semantic enriching framework for construction site images C. Zeng & T. Hartmann Technische Universität Berlin, Berlin, Germany
ABSTRACT: Recent applications of deep learning (DL)-based methods in construction have achieved notable momentum in making the construction management process smarter. The deployment of DL-based visual analytics systems needs to semantically enrich the original visual resources to improve situation awareness and scene understanding of the jobsite. This paper proposes a semantic enriching framework that integrates consolidated DL technologies with prior domain knowledge to prompt a traceable, explainable high-level information inferring process to explore semantic information behind visual resources collected from construction sites. The introduced framework starts from feature extraction from multiple aspects; then transforms the acquired facts into ontological assertions to iteratively reach higher-level interpretations achieved by semantic reasoning and querying. A case study, which manually grounds the primitive facts of a sample image and simulates the reasoning process based on the collected facts, is carried out to demonstrate the feasibility of the proposed framework.
1 INTRODUCTION With the ever-increasing use of camera-equipped devices on construction sites and advances in computer vision (CV) algorithms, research into applications based on visual analytics in the construction industry is now seen to be immensely growing. These applications have already shown significant benefits in addressing the deficiencies resulting from traditional construction management methods. The goal of a visual collection and analysis system during construction works is not only to assist perceptual work pertaining to reducing error-prone and labor-intensive site monitoring efforts but also to guide cognitive tasks, such as situation awareness and scene understanding, for construction experts to allow them to focus more on personal experience or knowledge valued decision-making processes. Nowadays, the CV community has provided the construction industry with various off-the-shelf DL visual analytic models (Paneru & Jeelani 2021). These data-driven models are often integrated with perceptual feature patterns, often retrieved in a global feature learning manner that lacks context reasoning beyond convolutions with large receptive fields (Chen et al. 2018). They are able to facilitate the information gathering and processing process at a feature level but fail to exploit the high-level knowledge behind the explicit feature pattern, which can be regarded as “semantic gap”. Visual understanding is a knowledge-intensive process which is decisively shaped by the way commonsense knowledge and experiences are brought to
358
bear (Neumann & Möller 2008). It is even essential for simple perceptual tasks to integrate with cognitive/semantic factors in those cases where visual cues alone fail to express the differences among classes and when the limited number of images hinders the capabilities of deep learning methods (Palazzo et al. 2021). In awareness of such issues, domain knowledge is utilized to assist image classification and scene understanding in (Forestier et al. 2012; Huang et al. 2017). However, employing domain-specific knowledge in a visual analytic system concerning the visual understanding of construction site images is not trivial and imposes several questions: 1. What is the scope of the visual understanding task for the construction site? What is the related domain knowledge to be integrated? 2. How to extract problem-specific knowledge through visual-data analysis? 3. How to fuse domain knowledge and extracted knowledge for high-level information reasoning and retrieve? This paper aims at addressing the above questions by introducing a semantic enriching framework that combines CV algorithms and an ontology model, where the former provides primitive feature representation about the images, and the latter further serves as a computational vehicle for active knowledge representation which permits incremental inference about semantic information based on feature representation together with other sources of contextual information. In this paper, the rationale behind integrating an ontological representation into the visual analytic system can be summarized as the following: DOI 10.1201/9781003354222-46
1. The ontology serves as a conceptual knowledge base to which factual knowledge captured in feature representation can map. A good image understanding is usually a compromise between background knowledge learned a priori and imagespecific observations (Chen et al. 2018). It is therefore essential for the construction site visual understanding system to combine prior knowledge about object attributes, object co-occurrence, hierarchy, pairwise object relationships to infer more advanced knowledge about status, action, activity, etc., otherwise, those ambiguous patterns that share the most of common features will not be easily distinguished. 2. The ontology provides a formalized structure that is capable of integrating heterogeneity information during construction phases. It is even hard for civil engineering specialist who visits the site and reports the current work without looking into any supportive context. Additional contextual information retrieved from the look-ahead plan, BIM models, site layout, and other documents (construction plan, construction progress report, etc.) can be assembled as instances of the central ontology accordingly to define a semantic contextual environment for a specific site. 3. The ontology is capable of its generalization ability. It can be extended at will and is flexible to be adjusted to other sources of informatic evidence (e.g. from various kinds of physical sensors) to conduct multimodal analysis for different granularities of visual understanding.
2 PAPER REVIEW As previously mentioned, DL-based technologies have prompted significant advances in construction management towards different divisions including construction productivity, construction safety, quality management, construction waste management, facilities management, progress monitoring, workspace planning, and so on (Pal & Hsieh 2021). Most of these applications are a result of utilizing feed-forward end-to-end learned ConvNet models. The stacks of convolutions in ConvNet models efficiently aggregate low-level pixel features from images to form high-level feature patterns. The ever-increasing hardware performance and availability of construction field-related datasets provide the necessary conditions to train the vast number of parameters of ConvNet models. Additional breakthroughs come from the optimization and adjustment of the network architecture. Different neural network structures have made diverse intelligient applications in construction industry avaiable. There already exist image classification-based applications towards progress stage recognition (Byvshev et al. 2020), equipment classification (Soltani et al. 2017); object detection applications locating workers (Luo et al. 2018), equipment (Kim et al.
2019), personal protective equipment (PPE) (Chen & Demachi 2021); segmentation methods offering pixel-region-based object localization solutions for equipment (Fang et al. 2020), construction scene (Pour Rahimian et al. 2020), and so on. Rather than identifying and locating concerned objects on the construction site, recent DL-based attempts have managed to explore more semantic information embedded in visual resources. Posture estimation of workers and equipment is realized (Chen & Demachi 2021) to identify the 2D or 3D joint positions of the object skeleton. Such skeleton data can be further processed to output recognized actions of objects (Li & Li 2022). Despite the advance of the above-mentioned methods in extracting feature and low-semantic-level information from visual resources, the nature of convolutional learning, however, prevents the deployed model to gain a global reasoning power that allows pixel regions farther away to directly communicate semantic and spatial information, despite the importance of such information in visual understanding (Chen et al. 2018). Given such deficiency, Bang & Kim (2020) introduce combined CNN-LSTM double-layered models for construction site image caption where the CNN network serves as the encoder transferring scene elements into feature representation, while the LSTM module function as the decoder transferring feature representation into a descriptive sentence. This approach assumes enough examples of semantic descriptions in the training data, which may not adapt to the case when relations between classes grow exponentially with an increasing number of classes. A relevance networks reinforced DL-based method is introduced in (Luo et al. 2018) to recognize diverse construction activity in site images. It is partially based on intuitive notions which do not necessarily provide a consistent knowledge basis to allow the model to be generalizable and adjustable. In a nutshell, the application of conventional DL algorithms in construction management has matured, higher-level semantic information of construction site images, however, is still less explored. Considering the effective learning ability of connectionism (resulting in neural systems) and the reasoning ability of symbolism (resulting in symbolic systems), one might combine these two to gain a deeper visual understanding of construction site images.
3 RESEARCH SCOPE AND METHODOLOGY This paper proposes a visual understanding framework to support the CV approaches and methods for visual understanding of visual resources originating from structural engineering construction sites. Generally, the output of CV methods is encoded into site-related cognitive facts to reason about newly enhanced knowledge pertaining to different granularity of activity information appearing in the scene. The primary focus is on introducing the framework and on verifying it
359
that contains one or more lower-level conceptualizations. Information from previous steps is aggregated to match with activity descriptors to determine the most likely activity and group activity appearing in the image.
4 THE ONTOLOGY MODEL
Figure 1.
Incremental knowledge inference schema.
based on the example from a real construction site. It should also be emphasized that we restrict the type of visual resources to images. Images are often more ambiguous than videos due to comparatively less visual content and cues, at the same time they are more efficient to be processed. These two features can better demonstrate the effectiveness of the proposed framework. A construction project is defined as a set of temporally related activities with specific goals to be achieved. It is therefore essential to take activity as the highest semantic level of view of the construction site images in visual understanding tasks. According to (Aggarwal & Ryoo 2011), human activities can be categorized into four different levels depending on their complexity: gestures, actions, interactions, and group activities. For construction activity, we take the involvement of human as an indispensable factor of it and introduce a similar categorization with five levels: feature, status, action, activity, and group activity. Figure 1 shows an incremental knowledge inference schema following the proposed categorization. Highlevel semantic information can be inferred based on built-in rules by combining low-level features with contextual information from other sources. CV algorithms are responsible for processing the raw data, which refers to original images captured by cameraequipped devices, to generate feature representations. These representations are encoded into ontological assertions that are organized as a named graph at the feature level to infer further semantic information. For the status level, it not only collects evidence from the feature-based named graph but also combines context originating from other information sources (e.g. site layout, look ahead plan, etc.) to gain a non-vague understanding of the status information. At the action level, the same process occurs, but no contextual information is involved. In the proposed framework, activity is described by a predefined activity descriptor
The general schema of the developed ontology is depicted in Figure 2. We reused the InformationEntity concept introduced in (Zheng et al. 2021), which is about the various types of information that are produced or consumed in a construction process. ConstructionSiteImage provides a visual way for describing the construction process, while other CMinformationEntity supports different perspectives for the corresponding process. ConstructionSiteImage has two kinds of representations: FeatureRepresentation and SemanticRepresentation. FeatureRpresentation is generated from different CV methods designed to provide easy-accessed, comprehensive feature representations, while SemanticRepresentation remains to be explored with the support of FeatureRepresentaion and CMinformationRepresentation. In order to promote the integration of heterogeneity information into a consistent symbolic calculation environment, all these representations are transformed into ontological assertions organized as named graphs. In this research, we mainly combine results from object detection, semantic segmentation, and pose estimation mentioned in Section 2 to constitute the main content of FeatureRepresentation, while CMinformationRepresentation encompass context from site plan, look ahead plan, and corresponding acquisition information of visual resources. 4.1 Feature representation Feature serves as the bridge between factual knowledge embedded in the visual resources and conceptual knowledge modeled in the ontological model. While the previous visual understanding attempt (Forestier et al. 2012) try to interpret images based on pixellevel feature representation, the development of DL technologies has equipped the visual analytic system with the ability to directly assign semantic labels or regions to entities based on object detection, semantic segmentation, pose estimation, etc. Basically, an image contains multiple aspects of information. A single type of CV algorithm, however, is usually designed to extract part of such information due to efficiency and performance limitations. In this work, we adopt various DL techniques aiming to achieve a more comprehensive feature representation for construction site images. As shown in Figure 3, Object refers to common objects on construction sites. Depending on whether they can be used as subjects capable of exerting action, we classify them as ActiveObject and PassiveObject, where ActiveObject includes Labor,
360
Figure 2.
Incremental knowledge inference schema.
Machinery, and PassiveObject includes ConstructionComponent, MaterialBatch, and AuxiliaryEquipment. Object detection algorithms can be utilized to associate specific identities to bounding boxes to identify the existence of Object. ObjectFeature represents the intrinsic properties of Object and what Object is exhibited in the image. ObjectFeature consists of QualitativeFeature and QuantitativeFeature. QualitativeFeature corresponds to those features that can be directly assigned with operatable semantic symbols through CV algorithms. For example, 2D or 3D joint positions of Labor and Machinery Skeleton can be learned to infer the Pose of Object (Roberts Dominic et al. 2020); MaterialType of ConstructionComponent and MaterialBatch can be classified with the deep CNN model (Lin et al. 2019). QuantitativeFeature requires further processing in subsequent inference steps to be transferred into the semantic symbol. Subclasses of QuantitativeFeature are Segmentation and Skeleton, the former describes the accurate pixel region that the Object occupies in the image (Chen et al. 2017), while the latter presents possible work patterns of the Object. By analyzing the connectivity and proximity between different pixel areas of Segmentation and Skeleton, one can determine whether there is a potential interaction between Object and Object. Other than ObjectFeature, the spatial relative relationship is also one kind of key primitive features in visual understanding tasks (Cavaliere et al. 2019). ObjectFeature is linked with Object through hasFeature property. The spatial and relative relationship between objects is represented by hasSpatialRelation and hasRelativeRelation. By modeling in this way, the multiple aspects of the feature of the image are transformed into triplets to support further semantic inferences.
4.2 Context Context refers to an identified realm of data, representing the circumstances in which the data can be considered true (Zheng et al. 2021), it provides diverse perspectives of interpretation about the physical world of which the camera takes screenshots. Context participates in the semantic enriching process in terms of instantiated concepts which constrain other possible scene interpretations. We currently introduce three kinds of contexts in this paper. More contexts may be fused in future works given the different complexity of visual understanding tasks. SiteLayout describes construction spaces needed for construction work elements and the spaces occupied by completed work units (Riley & Sanvido 1995). We follow the pattern from (Riley & Sanvido 1995) and defined subclasses of Workspace including WorkArea, PrefabricationArea, StorageArea, RestArea, MaterialPath, PersonalPath. AcquisionInfo provides accurate acquisition contextual information which conditions the resulting acquired images. It comprises Location, Time, Device, ViewAngle. LookAheadPlan pertaining to descriptions of a detailed breakdown of the phase schedule looking forward several days or weeks. By extracting the content from LookAheadPlan, Object can be linked with Activity through hasPlannedActivity property. 4.3 Semantic representation In order to achieve a more comprehensive semantic understanding of construction site images, and to support a traceable and interpretable reasoning process, this paper divides the semantic information into three
361
Figure 3.
Feature representation schema.
aspects, as shown in Figure 4. Construction is considered as “a group of Actors uses a set of Resources to produce a set of Products following certain Processes within a work environment and according to certain conditions.” (El-Gohary & El-Diraby 2010). For structural engineering construction, actors here refer not only to professional workers but also to construction machinery that to some extent replaces workers in large-scale construction works.All aspects of semantic information established are therefore primarily related to corresponding aspects of Labor and Machinery.
4.3.1 Status In this layer, we exploit two types of facts to eliminate some of the uncertainty and ambiguity of the semantic information from construction site images. One is that designated behaviors of professional objects are often carried out in a particular space. For example, the detection of a concrete pump track on a material path indicates the most probable fact that the track is transporting concrete for concrete pouring, while the detection of the same concrete pump track in the work area probably implies it is suspended for concrete pouring. Another fact is that the construction of a structural component often undergoes several work breakdown structure (WBS) steps with distinguishing visual characteristics. This is evident in the case of the construction of cast-in-place concrete structures where formwork, reinforcement, and concrete pouring are necessary working steps. We combine contextual information to take full advantage of such facts. By aligning Location with Workspace, we get to know what kind of Workspace an Object is on. By comparing detected visual features with built rules corresponding to the specific visual pattern, information about which WBS step a constructing structural component is in can be obtained.
Three types of status are defined in the model, namely, ConstructionComponentStatus, LaborStatus, and MachineryStatus. Constructional status of construction components including Preparing, Constructing, Finished. Preparing status indicates that no planned activity-related objects have been detected. Finished status represents the detection of some construction components along with the consistency between the displayed material type of such components and the final material type. Working, nonworking, and transitional states between two consecutive working or nonworking states are assigned to Labor and Machinery as LaborStatus and MachineryStatus respectively. 4.3.2 Action Actions are interactions between active objects or between active objects and passive objects. Crane transports reinforcement and worker drives track are two typical action cases where crane and worker are active objects as the subjects of the action, transports and drive are the actions that active objects perform. Identification of action is achieved by a combination of the status information of the involving objects and the spatial relationship between them. Active objects with working or operating status and passive objects in close proximity to such active objects in the image are taken into consideration for further spatial relationship analysis. The spatial relationship analysis is conducted by comparing connectivity and proximity between pixel regions of segmentation and skeleton. If the pixel regions of interested objects are connected, then they have a mutual isConnectedTo spatial relationship. A threshold is set to distinguish isNearTo and isFarFrom by examining the closest pixel proximity of the candidate regions. Action is related to the typical behavior of professional workers and machinery. For example,
362
Figure 4. Semantic representation schema.
Transporting is the action of a crane; Digging, Hauling, dumping are typical actions of excavators; Nailing, Leveling, etc. are common actions of construction workers. 4.3.3 Activity Construction activity involves a set of flows including labor, equipment, workspace, materials, precedence, information, and external flows (Garcia-Lopez 2017). The presence of labor and/or equipment are regarded as the necessary conditions for defining an activity in this paper. Activity can be further divided into PreparingActivity and WorkingActivity. WorkingActivity is composed of several SubActivity to indicate the detailed WBS steps. The main factor for an activity to be attributed to PreparingActivity or WorkingActivity is whether ConstructionComponent is detected. For example, “crane transports reinforcement” is simply regarded as a PreparingActivity of “ReinforcementTransporting” in StorageArea while as a SubActivity of “ReinforcementPlacement” belongs to WorkingActivity of “CastInPlaceConcreteColumn” if a ConcreteColumn with Constructing status is detected. In order to signify the presence of complex activity, one or more specific activity descriptors are linked with activity through isDescribedBy property. Each activity descriptor independently relies on the DependencySet used to characterize the activity. SubActivity is allowed to reuse the precedent and abstract pattern defined by WorkingActivity. Furthermore, several patterns are aggregated in describing the same activity to
capture the ambiguity of the construction site scene. The matching procedure is carried out by semantically comparing the RDF dataset formed from previous steps against background knowledge about activity dependencies using SPARQL query language. Objects that participate in the same type of activity are grouped together to form a group activity. The concept of group activity is introduced to take into account the situation when there is a large viewing field of construction site images where more than one ongoing activity may show up. 5 CASE STUDY Ontology, which contributes to providing a high-level information abstraction of the scene in a semantic symbol environment, serves as the core of the proposed visual understanding framework. A harmonious conceptual alignment procedure is therefore vital for the performance of the framework. This section presents a case study showing such a procedure to validate the applicability of the proposed ontology modeling for the construction site visual understanding task. It is worth noting that this paper only aims to present a theoretical framework and validate it with a real construction site image and does not involve integration with any DL methods, so the results of the subsequent DL algorithms involved are just manually fabricated output. The test image is shown in Figure 5. At first, CV algorithms are responsible for outputting multiple
363
aspects of information of construction site image. For the example image, object detection, semantic segmentation, pose estimation are deployed in recognizing Object and ObjectFeature. The part of RDF graph base created contains the following content:
Insert: ?m :transporting ?n Where: { ?m a :Crane . ?m :hasStatus ?x . ?x a :Operating . ?m :hasWorkingPart ?y . ?n :isConnectedTo ?y . } All those information is gathered together to match with the built pattern to finally output the ReinforcementTransporting activity result. The corresponding meta matching pattern is as follows. Since only one worker appears in the image, there is no need to further identify group activity. Worker in WorkArea || PrefabricationArea || StorageArea Crane in WorkArea || PrefabricationArea || StorageArea Crane hasStatus Working Crane transporting Reinforcement (OPTIONAL)Worker manupulating Controller 6 CONCLUSION
Figure 5.
Example image from a construction site.
At the second status level, SPARQL-based rules utilizing the facts from the context and the feature level are exploited to deduce the status of Labor, Machinery, and ConstructionComponent. The following rule is about inferring the status of worker_1. A similar rule can also be applied to crane_1 to infer its operating status. Insert: ?x :hasStatus :Working Where: { ?x a :Worker . ?x :in ?y . ?y a :WorkArea || :PrefabricationArea || :StorageArea . ?x :hasPose ?z . ?z a :LaborWorkingPose . } Since worker_1 and crane_1 both have working status, possible actions may be carried out by them. By analyzing the spatial relationship between operatingpart_1 of crane_1 and other objects that are in close proximity to operatingpart_1 in the image, the fact about reinforcementcage_1 isConnectedTo operatingpart_1 is obtained. Based on such fact and the following SPARQL-based rule, the action of crane_1 is further deduced. A similar rule can also be applied to worker_1 to infer the Manipulating action.
To address the deficiency of the current DL-based visual analytic system in semantic reasoning and to improve visual understanding for construction management, this paper proposes a semantic enriching framework combing CV algorithms and ontology to enhance the semantic representations of construction site images. Rather than purely identifying and locating construction-related objects, the use of ontology adds support for enriching the scene description from lower level to higher level. The proposed ontology serves as a formal knowledge representation of the key concepts related to what civil engineering experts perceive from the construction site image. The scope of this paper is limited to the introduction of the framework and the critical concepts constituting the ontology model. An intuitive case study is conducted to demonstrate the effectiveness of the proposed method. With further systematic integration of the ontological model and DL technologies the framework is expected to allow for a deeper visual understanding of visual resources of construction site and pave the way for intelligent cognition-based applications in the construction industry. REFERENCES
364
Aggarwal, J.K. & Ryoo, M.S., 2011, ‘Human Activity Analysis: A review’, ACM Computing Surveys, 43(3), 1–43. Bang, S. & Kim, H., 2020, ‘Context-based Information Generation for Managing UAV-acquired Data Using Image Captioning’, Automation in Construction, 112, 103116. Byvshev, P., Truong, P.-A. & Xiao, Y., 2020, Image-based Renovation Progress Inspection with Deep Siamese Networks, Proceedings of the 2020 12th International Conference on Machine Learning and Computing, 96–104, ACM, Shenzhen China.
Cavaliere, D., Loia, V., Saggese, A., Senatore, S. & Vento, M., 2019, ‘A Human-like Description of Scene Events for a Proper UAV-based Video Content Analysis’, KnowledgeBased Systems, 178, 163–175. Chen, L.-C., Papandreou, G., Schroff, F. & Adam, H., 2017, ‘Rethinking Atrous Convolution for Semantic Image Segmentation’, arXiv:1706.05587 [cs]. Chen, S. & Demachi, K., 2021, ‘Towards On-site Hazards Identification of Improper Use of Personal Protective Equipment Using Deep Learning-based Geometric Relationships and Hierarchical Scene Graph’, Automation in Construction, 125, 103619. Chen, X., Li, L.-J., Fei-Fei, L. & Gupta, A., 2018, Iterative Visual Reasoning Beyond Convolutions, 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 7239–7248, IEEE, Salt Lake City, UT, USA. El-Gohary, N.M. & El-Diraby, T.E., 2010, ‘Domain Ontology for Processes in Infrastructure and Construction’, Journal of Construction Engineering and Management, 136(7), 730–744. Fang, W., Ma, L., Love, P.E.D., Luo, H., Ding, L. & Zhou, A., 2020, ‘Knowledge Graph for Identifying Hazards on Construction Sites: Integrating Computer Vision With Ontology’, Automation in Construction, 119, 103310. Forestier, G., Puissant, A., Wemmert, C. & Gançarski, P., 2012, ‘Knowledge-based Region Labeling for Remote Sensing Image Interpretation’, Computers, Environment and Urban Systems, 36(5), 470–480. Garcia-Lopez, N.P., 2017, An Activity and Flow-based Construction Model for Managing On-site Work – PhD thesis, Stanford University . Huang, H., Chen, J., Li, Z., Gong, F. & Chen, N., 2017, ‘Ontology-Guided Image Interpretation for GEOBIA of High Spatial Resolution Remote Sense Imagery: A Coastal Area Case Study’, ISPRS International Journal of Geo-Information, 6(4), 105. Kim, D., Liu, M., Lee, S. & Kamat, V.R., 2019, ‘Remote Proximity Monitoring Between Mobile Construction Resources Using Camera-mounted UAVs’, Automation in Construction, 99, 168–182. Li, Z. & Li, D., 2022, ‘Action Recognition of Construction Workers Under Occlusion’, Journal of Building Engineering, 45, 103352. Lin, J.J., Lee, J.Y. & Golparvar-Fard, M., 2019, Exploring the Potential of Image-Based 3D Geometry and Appearance
365
Reasoning for Automated Construction Progress Monitoring, Computing in Civil Engineering 2019, 162–170, American Society of Civil Engineers, Atlanta, Georgia. Luo, X., Li, H., Cao, D., Dai, F., Seo, J. & Lee, S., 2018, ‘Recognizing Diverse Construction Activities in Site Images via Relevance Networks of Construction-Related Objects Detected by Convolutional Neural Networks’, Journal of Computing in Civil Engineering, 32(3), 04018012. Neumann, B. & Möller, R., 2008, ‘On Scene Interpretation with Description Logics’, Image and Vision Computing, 26(1), 82–101. Pal, A. & Hsieh, S.-H., 2021, ‘Deep-learning-based Visual Data Analytics for Smart Construction Management’, Automation in Construction, 131, 103892. Palazzo, S., Murabito, F., Pino, C., Rundo, F., Giordano, D., Shah, M. & Spampinato, C., 2021, ‘Exploiting Structured High-level Knowledge for Domain-specific Visual Classification’, Pattern Recognition, 112, 107806. Paneru, S. & Jeelani, I., 2021, ‘Computer Vision Applications in Construction: Current State, Opportunities & Challenges’, Automation in Construction, 132, 103940. Pour Rahimian, F., Seyedzadeh, S., Oliver, S., Rodriguez, S. & Dawood, N., 2020, ‘On-demand Monitoring of Construction ProjectsThrough a Game-like HybridApplication of BIM and Machine Learning’, Automation in Construction, 110, 103012. Riley, D.R. & Sanvido, V.E., 1995, ‘Patterns of ConstructionSpace Use in Multistory Buildings’, Journal of Construction Engineering and Management, 121(4), 464–473. Roberts Dominic, Torres Calderon Wilfredo, Tang Shuai, & Golparvar-Fard Mani, 2020, ‘Vision-Based Construction Worker Activity Analysis Informed by Body Posture’, Journal of Computing in Civil Engineering, 34(4), 04020017. Soltani, M.M., Karandish, S.-F., Ahmed, W., Zhu, Z. & Hammad, A., 2017, Evaluating the Performance of Convolutional Neural Network for Classifying Equipment on Construction Sites, ISARC. Proceedings of the International Symposium on Automation and Robotics in Construction, vol. 34, IAARC Publications. Zheng,Y., Törmä, S. & Seppänen, O., 2021, ‘A Shared Ontology Suite for Digital Construction Workflow’,Automation in Construction, 132, 103930.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Machine learning and genetic algorithms for conformal geometries in design support systems S.J.F. van Hassel, H. Hofmeyer, T. Ezendam & P. Pauwels Eindhoven University of Technology, Eindhoven, The Netherlands
ABSTRACT: To optimise both building designs and their underlying design processes, design support systems exist. For domain specific analyses, these systems benefit from a conformal (CF) representation for the Building Spatial Design (BSD). In a conformal representation, for all entities: the vertices of an entity are, if intersecting another entity, only allowed to coincide with this other entity’s vertices. This paper presents research on whether Machine Learning (ML) and Genetic Algorithms (GA) can be used to obtain a conformal geometry for BSDs. For ML, a neural network is trained to learn the complex relation between BSDs and their conformal representations. GAs are first used to find all quad-hexahedrons in the search space, then to find sets of quad-hexahedrons that form the conformal design. A trained ML model does provide outcomes, but not very useful, even with encoding the configuration type of the design. Differently, the GA finds conformal designs for many instances, even for non-orthogonal designs.
1 INTRODUCTION In the built environment, a building design process is understood as the creation of a plan to realise a building (Hofmeyer & Davila Delgado 2015). Building design processes are unique and iterative, where multiple disciplines are involved (Haymaker et al. 2004). This makes these processes complex, to be supported in the early stages to improve the final design (Kalay 2004), and to reduce costs of design changes (MacLeamy 2004). Ideally, decisions are made where all requirements of the different disciplines are considered. However, a design team is never able to analyse all possible alternatives to make the related well-founded decisions. Design support systems exist that (a) suggest building designs, which can be used by the design teams; (b) simulate design processes (via simulations with partial models), to support the teams process-wise. In all such systems, a representation is used for the Building Spatial Design (BSD). And for domain specific analyses, it is very useful to have a so-called conformal (CF) representation for the BSD. In a conformal representation, for all entities it holds: the vertices of an entity are, if intersecting another entity, only allowed to coincide with this other entity’s vertices. Figure 1 shows BSD non conformal surfaces (grey), which are transformed into a conformal geometry (blue), here via partitioning of surfaces. A conformal geometry is convenient for disciplinespecific models. For example, for structural design, proper finite element meshing, and loading is very
366
Figure 1. Non-conformal BSD surfaces are transformed into a conformal geometry (l); domain related advantages (r).
difficult without a conformal geometry.And if the conformal geometry is used as a blueprint for structural elements, also load paths become more efficient (black arrow). Thermal modelling of the building becomes easier too, as surfaces are either fully internal or external, and so fluxes (the red and blue arrows in Figure 1) for each surface can be correctly described. Current procedural methods to transform a BSD into a conformal design work for orthogonal, and sometimes for specific non-orthogonal designs (Ezendam 2021), whereas realistic building designs may be less specific non-orthogonal. Therefore, in this paper research is presented on whether alternatively Machine Learning and Genetic Algorithms can be used to transform orthogonal and non-orthogonal BSDs into conformal designs.
1.1 Related work For this paper, an existing design support system is used (Boonstra et al. 2018). In this system, the BSD DOI 10.1201/9781003354222-47
Figure 2. BSD described by (a) origin & dimensions for each space, or (b) corner-vertices. On the right, (c) the conformal representation.
is defined as an arrangement of spaces. For orthogonal designs, the spaces are limited to cuboid shapes. For non-orthogonal designs, the spaces are allowed to be different forms of quad-hexahedrons, as long as the ‘walls’ are vertical and the ‘floors’ are horizontal. Spaces of orthogonal designs can be defined either by an origin vector (s) and dimension vector (d) (Figure 2a) or by eight corner vertices (p) (Figure 2b). See also Equations 1, 2 and 3. Non-orthogonal designs are (only) described by eight corner-vertices. s = [xyz]T ;
d = [wdh]T ;
p = [xyz]T
(1, 2, 3)
machine can learn in three different ways (Goodfellow et al. 2017), namely Supervised, Unsupervised, and by Reinforcement learning. Here, Supervised learning will be used, where the dataset is fixed, and the input and associated output are known. The machine learns to predict the output (target) from an input (feature) and the algorithm is trained on the relation between features and targets. Machine learning (ML) will be used to make orthogonal spatial designs conformal. The machine needs to learn the relation between initial spatial designs and their conformal representations. The ML concept is visualised in Figure 3.
For orthogonal designs, the conformal geometry is described by cuboids. See Figure 2c for an example with three cuboids, where each cuboid is described by eight corner-vertices. For non-orthogonal designs, the conformal geometry uses quad-hexahedrons, which are also described by eight corner-vertices. 1.2 Methods This research investigates Artificial Intelligence (AI) techniques to make spatial designs conformal. First, Machine Learning (ML) is used to make conformal geometry predictions. A dataset is generated via the design support system (Boonstra et al. 2018) with building spatial designs as input (features) and corresponding conformal geometries as output (targets). The dataset is used to train a neural network to make spatial designs conformal. Secondly, two Genetic Algorithms (GAs) are used to make spatial designs conformal. A first algorithm uses space coordinates to generate possible quad-hexahedrons. Then a second algorithm is used to find groups of quadhexahedrons that form a conformal design. Initially, the research is focused on orthogonal BSDs with two spaces. If one of the methods functions acceptably, more spaces and non-orthogonal geometries are tried. 2 MACHINE LEARNING (ML) A machine has a level of intelligence if it can learn from external information to improve its own knowledge (Mitchell 1997). The access to quantitative and qualitative data enables a machine to learn certain rules, which can be applied to solve tasks automatically. A
Figure 3. The existing design support system generates a dataset of features and targets, which is used to train a machine that predicts the conformal geometries of new test designs.
Using the existing support system (Boonstra et al. 2018), a transformation is carried out N times, each time for a different spatial design. As such, data is generated at the input and output of the design process. As shown in Figure 3, the input data (features) and the output data (targets) are extracted from the support system, placed in a dataset, and used for the machine learning process. As data is accessible and the input features and output targets are known, supervised ML is used. Finally, a trained model is created that predicts the output by only the input variables. In this way, the algorithm that generates a conformal design is replaced
367
by a trained ML model. The trained ML model is evaluated by several studies, but first the dataset and machine learning model need to be defined, as presented in the next sections.
2.1 Dataset Within the first part of research, using ML, only orthogonal BSDs are considered. Therefore, the conformal representation is defined as an arrangement of cuboids. To start with a basic design, the dataset is based on a building with two spaces, where space 1 is fixed and space 2 is positioned to the right of space 1. Later, space 2 can also be positioned around (front, behind, left) or on top of space 1. For the spatial design, the dimensions vectors (d) of the two spaces are randomly defined. The width h and depth d vary between 4000 and 8000 mm, and the height h varies between 2000 and 6000 mm. The origin vector (s) of space 1 is always defined as [10000 10000 0], and the origin vector of space 2 is based on its position related to space 1. However, the coordinates in the origin vector need to be defined, such that the two spaces are adjacent. The resulting BSDs are classified in five different variants for space 2 relative to space 1: ‘right’, ‘left’, ‘front’, ‘behind’, and ‘top’. Additionally, each variant has several building types, which are given by the size and orientation of space 2 in relation with space 1, which in turn result in different numbers and configurations of the cuboids. As a result, 27 different building types can be generated for each of the variants: ‘right’, ‘left’, ‘front’, ‘behind’, and 81 different building types can be generated for the ‘top’ variant. The number of building types is important in the analysis, which is explained later. The study starts with 1000 data-points where space 2 is positioned to the right of space 1. For this situation, the input data consists of 12 features and the output data of 168 targets (maximal 7 cuboids). One data-point represents the features and targets of one simulated conformal transformation. Features (s and d of BSD) are described by: f = [w1 , d1 , h1 , x1 , y1 , z1 , . . . , zs ]
(4)
where s = number of the space. Targets (p of CF model) are described by: t = x11 , y11 , z11 , . . . , xcp , ycp , zcp
(5)
where c = number of the cuboid; and p = index of the corner-vertex of cuboid c. Note that for the quality of the model, relatively few input data is available for many output options. The data is scaled with mean normalization to reduce the importance of individual features and to improve the training process (Ioffe & Szegedy 2015). Additionally, the data is split into a training set (80%) and testing set (20%).
2.2 Machine learning model A Feed forward Neural Network (FNN) is used to solve the regression task. The architecture of the FNN is shown in Figure 4. The dimensions and origins of the spatial design are allocated in the input layer. The input is connected with one hidden layer to the output. The corner-vertices of the conformal cuboids are allocated in the output layer. Each layer consists of a number of neurons (or nodes). The neurons of one layer are connected to neurons of another layer and have certain weight values. These weights are initialised by the He function (He et al. 2015) and are adjusted in the training process by using the Adam optimisation function (Kingma & Ba 2014). The goal is to minimize the loss between predicted output and the actual output. The loss is indicated by the Mean Squared Error (MSE), and calculated as follows (Kolodiazhnyi 2020): ⎛ ⎞ n m 2 1⎝ 1 MSE = (6) ti [j] − tˆi [j] ⎠ n i=1 m j=1 where n = number of data-points; m = number of output values per data-point; t = actual output; and tˆ = predicted output.
Figure 4.
Neural network architecture.
Four hyper-parameters need to be defined in the analysis: number of hidden layers, number of neurons, learning rate and batch size. These parameters are dependent on the dataset and can be changed. In the analysis, one hidden layer will be used and the number of neurons is equal or slightly larger than the number of nodes in the adjacent input or output layers. Additionally, the initial learning rate is set to 0.01 and the batch size is 1/10 of the number of data-points. 2.3 Analysis The performance of the trained ML model is measured by analysing the loss function (MSE) and the prediction for three test buildings. The loss is calculated for the training data during the training process of 12 iterations. First, test building ‘A’ (Figure 6) is used and needs to be made conformal by a trained ML model. Since the test building has two spaces and space 2 is to the right of space 1, the ML model is trained on a
368
dataset with 1000 data-points (800 training, 200 testing), where space 2 is located to the right of space 1. This dataset is called ‘Single configuration’. Figure 5 shows the loss during training. The training MSE (Not encoded) reduces only slightly during the 12 iterations. The actual conformal model and prediction of test building ‘A’ are visualised in Figures 6b and 6c. The predicted CF model (Not encoded dataset) consists of a set of hexahedrons but bears hardly a resemblance with the actual output.
Figure 5. Training loss of ML model. Trained on Not & One-hot encoded dataset with single configuration, where space 2 is to the right of space 1.
To improve the model, it is realised that typically the number of input values is larger than the number of output values. But here, the input consists of only 12 features and needs to predict 168 output values. Therefore, extra information is applied to the dataset by adding the type of the design as a feature. The 27 different building types of the variant ‘right’ (space 2 at the right of space 1) are included in the input features by using One-hot encoding. One-hot encoding
is used to convert categorical data into numerical data and prevents ranking of the different building types. As shown in Figure 5 (One-hot encoded), now the loss is reduced significantly. In addition, the prediction using the One-hot encoded dataset has many similarities with the actual conformal model, see Figure 6d. Thus it can be concluded that One-hot encoding of the different output types improves the model significantly. And as such the improved model can predict a conformal model, based on a BSD with two spaces, where space 2 is at the right of space 1. However, it has to be tested to what extent the trained model is able to predict alternative designs. Therefore, test building ‘B’, the 180 degrees rotated version of test building ‘A’ is used as an alternative design, see Figure 7a. Using still the ‘Single configuration’dataset, the position of space 2, now to the left of space 1, is not included in the training dataset. Therefore, the trained machine learning model has never seen this configuration and is not able to make proper predictions, as can be seen in Figure 7c. Clearly, a trained ML model can only make predictions of building configurations that are used in the training process. Hence the need for elaborate diversity in the training set. A new dataset is defined, ‘All configurations’, and is created with 1000 data-points for each variant (left, right, etc.). A new ML model is trained and the loss during 12 iterations is shown in Figure 8. Consequently, the conformal prediction of test building ‘B’ using the One-hot encoded dataset on all configurations has more similarities with the actual CF model, see Figure 7d. Lastly, test building ‘C’ is considered, which is the two-space building with the most cuboids (10) in the conformal model, see Figure 9. As shown in Figure 9c, the shape of the conformal model and the number cuboids are predicted by the
Figure 6. Test building ‘A’: (a) spatial design, (b) actual CF model & (c,d) predictions.
Figure 7.
Test building ‘B’: (a) spatial design, (b) actual CF model & (c,d) predictions.
369
Figure 10. GA workflow.
The GA approach here consists of three steps, namely pre-processing of BSD data, a first GA1 to generate quad-hexahedrons, and a second GA2 to generate the conformal model. 3.1 Pre-processing of BSD data Figure 8. Training loss of ML model. Trained on One-hot encoded dataset with all configurations.
In this section, the corner-vertices (p) are used to describe the BSD spaces.Thereafter, a point cloud with vertices (v) is generated by all combinations of existing x, y and z values of the BSD corner-vertices, see Figure 11. Also new non-existing vertices will arise, but vertices outside the spatial design are neglected to reduce computational costs.
Figure 9. Test building ‘C’: (a) spatial design, (b) actual CF model & (c) prediction. Based on ‘All configurations’ One-hot encoded data-set.
trained ML model. However, there is still an error, and the predictions are numerically not perfect. The inaccurate predicted corner-vertices lead to non-cubic results.
Figure 11. Pre-processing data: (a) BSD corner-vertices, (b) existing x, y, z coordinates, and (c) the final point cloud.
3.2 GA1: Quad-Hexahedrons 3 GENETIC ALGORITHMS (GA) A Genetic Algorithm (GA) is based on the evolution theory of Darwin and applies stochastic search to find one or more optimal solutions in a problem (Holland 1975). Unlike most ML algorithms, GAs are not based on gradient-descent optimisation functions, but based on heuristic methods, and use the process of natural selection, reproduction and mutation (Dudek 2013). According to Stanovov et al. (2017), GAs is beneficial in high-dimensional search spaces. A GA workflow is illustrated in Figure 10. GAs start with a population of possible solutions (individuals), which evolve to better solutions, and finally one or more optimal solutions are found. After the population is initialised, the process consists of four main operators: fitness, selection, crossover, and mutation, which are repeatedly executed in every generation. The individual solutions are marked with a fitness value that represents the quality of the solution. Thereafter, the fittest solutions are selected and called parents. A crossover is applied on the parents, where information is exchanged from parents to new solutions, called offspring. To maintain diversity and mimic natural selection, mutations are applied on the individual solutions.
For the first GA, the initial population consists of N individuals, each created by a combination of 8 vertices. These vertices are chosen randomly from the point cloud and should form a quad-hexahedron. When an individual gains the maximum fitness score, it is a quad-hexahedron and saved for the second GA (GA2). The individuals are represented by an 8×3 matrix, see Figure 12, which is efficient (Wallet et al. 1996).
Figure 12. Representation of quad-hexahedron and fitness constraints.
The quad-hexahedrons are labelled with a fitness score based on their shape. For orthogonal shapes, a perfect rectangular cuboid gets the maximum fitness score. For non-orthogonal shapes, a perfect quadhexahedron with vertical walls and horizontal floors
370
obtains the maximum fitness score, see Figure 12. Additionally, non-convex and triangular solutions are eliminated from the population, as well as quadhexahedrons that are located in more than one space of the BSD. All individuals in the population are used as parents and undergo a crossover to create four new individuals, i.e., offspring. Preselection is used and preserves diversity in the population, according to Wong (2015) and Cavicchio (1970). The parents compete with their offspring, and the best 2 individuals are chosen for the next generation. Crossover is alternately applied in three ways: X, Y, and Z. Crossover X uses one-point crossover with crossover point between the 4th and 5th row of the chromosome matrix. Crossover Y uses a three-point crossover with a crossover point per two rows. Crossover Z uses a seven-point crossover with a crossover point per row. Mutation is applied at chromosome (quadhexahedron) and gene (vertex) level. Duplicate quadhexahedrons in the population are replaced by new unique quad-hexahedrons. Furthermore, duplicate vertices in a chromosome are replaced by new vertices, randomly chosen from the point cloud. It is difficult to determine whether all necessary quad-hexahedrons are found. Therefore, GA1 is running n generations, until no new unique quadhexahedron is found. The parameter n is designspecific and can be adjusted. 3.3 GA2: Conformal model In GA2, first the (number q) quad-hexahedrons found in GA1 are used to create a population of candidate conformal models. The initial population consists of M individuals based on a cluster of quad-hexahedrons. In GA2, an individual is called quad-cluster, and represented by an array of zeros and ones. The length of this array is equal to the number q of found quadhexahedrons. The ones in the array form the conformal model, see Figure 13 for an example.
of the BSD (VBSD ), indicated by a fitness volume (fV ), see Equation 7. Secondly, a conformal model should not have intersections between individual quadhexahedrons, indicated by a fitness intersect (fI ), see Equation 8. Intersections are represented by lineline and line-polygon intersections. The fitness values range from zero (best) to one (worst) and are calculated by equations 7 and 8. abs(VCF − VBSD ) ; f V = q i=1 VQi − VBSD
fI =
nInter (7) (nInter + nLinks )
Where VQ indicates the volume of each quadhexahedron. Furthermore, nInter indicates the number of quad-hexahedron pairs that intersect, and nLinks indicates the number of quad-hexahedron pairs that perfectly link. Perfectly linked means that two adjacent quad-hexahedrons are connected by four cornervertices without overlapping faces or edges. Due to the multi-objective fitness function, the design solutions cannot be simply separated by a single fitness score.Therefore, the non-dominated sorting principle (Kalyanmoy 2002) is used to select the best individuals, at the same time maintaining diversity in the population. Finally, the best half of the individuals that are not or less dominated by others will be selected for a next generation. The selected parents undergo a uniform crossover and create new offspring. Uniform crossover outperforms the one-point and twopoint crossover approaches, Syswerda (1989). When two parents have the same binary code, the offspring will also be the same. Therefore, random offspring generation (Rocha 1999) is applied, where the parents are randomly shuffled to create new offspring, but with the same number of quad-hexahedrons. The intermediate population with parents and offspring is mutated, where one or more genes are flipped. The mutation probability of the population is set to 50%. The evolution process stops when a perfect conformal model is found. This is achieved when the volume of the CF model is equal to the volume of the BSD (fV = 0), and there are no intersections present (fI = 0). 3.4 Analysis
Figure 13. Quad-cluster as an array of bits.
The exact number of quad-hexahedrons needed for the final conformal model is unknown. Therefore, the number of quad-hexahedrons in the initial population ranges from 1% to 100% of the total quadhexahedrons. The algorithm will find the best number by the evolving generations and the selection of the fittest. The quad-clusters are assessed by a multi-objective fitness function. First, the volume of the conformal model (VCF ) needs to be equal to the volume
3.4.1 Orthogonal (O) designs The test buildings ‘A’, ‘B’ and ‘C’ from Section 2.3 are also used for the orthogonal GA analysis and are now defined as ‘A-O’, ‘B-O’, and ‘C-O’ (‘O’ for Orthogonal). The results are shown in Table 1 and Figure 14. With v = number of vertices in point cloud; N or M = population size; g = number of generations; q = number of quad-hexahedrons found in GA1; and r = number of quad-hexahedrons in final CF model. Figure 14 shows the BSD with associated point cloud and the correctly predicted CF model for test buildings ‘A-O’, ‘B-O’ and ‘C-O’. There are no complications by using the same GA for all these three test buildings. Therefore, different from a trained ML
371
Table 1.
Results orthogonal GA analysis.
Test building
v
N GA1
g GA1
q GA1
M GA2
g GA2
r GA2
‘A-O’ ‘B-O’ ‘C-O’
26 26 36
512 512 2048
200 200 20,000
12 12 37
8 8 8
5 19 67
6 6 10
for test building ‘A-NO’. Naturally, conformal nonorthogonal designs have more possible combinations, and therefore, show larger computational costs. Figure 15 shows test building ‘A-NO’ and its correctly GA predicted CF model. The results are listed in Table 2. Table 2. Test building
v
N GA1
g GA1
q GA1
M GA2
g GA2
r GA2
‘A-NO’ ‘B-NO’ ‘C-NO’
26 29 36
512 512 2048
200 200 2000
36 60 120
8 – 32
360 – 4500
6 – 6
model, a GA is more generally applicable. Test building ‘C-O’needs the most generations in GA1 and GA2. This has two reasons. Firstly, the position of space 2 leads to only one quad-hexahedron in space 2 and is difficult to be found in GA1. Secondly, the larger number of quad-hexahedrons found in GA1 (q) and needed in the final conformal model (r) result in more combinations, which leads to more computations and generations.
Figure 14. Test buildings ‘A-O’, ‘B-O’ & ‘C-O’. Top row: BSD with point cloud. Bottom row: conformal models generated by GA.
It is shown that the GA is able to make two-space BSDs conformal. However, a GA is only beneficial if it uses less computational time than a process considering all possible solutions (a brute force technique). All possible combinations C are given by Equation 9. C(q, r) =
q! (r!(q − r)!)
(8)
For test building ‘C-O’, deterministic search in GA2 would lead to: C(37, 10)/M = 3.48 · 108 /8 = 4.35 · 107 generations, if the number of quad-hexahedrons in the CF model is known. Now only 67 generations are needed, which indicates that the GA is very helpful. 3.4.2 Non-orthogonal (NO) designs As mentioned earlier, realistic buildings may see nonorthogonal spaces, for which it is challenging to find conformal solutions with existing approaches (Ezendam 2021). Therefore, here a GA will be tested for some test buildings with non-orthogonal shapes. First, test building ‘A-NO’ is evaluated and has the same number of vertices in the point cloud as test building ‘A-O’ from Section 3.4.1. Nevertheless, the number of quad-hexahedrons in GA1 is increased to 36
Results non-orthogonal GA analysis.
Figure 15. Test buildings ‘A-NO’, ‘B-NO’ & ‘C-NO’. Top row: BSD with point cloud and floorplan of the building. Bottom row: Conformal models if generated by GA.
Test building ‘A-NO’ is solved by the presented technique of point cloud generation. However, it is expected that normally the point cloud needs to be expanded to create all necessary quad-hexahedrons. For example, when considering test building ‘B-NO’, with the standard cloud, the GA is not able to find a CF model. If an extra y coordinate is added, relabelling the design as ‘C-NO’, see Figure 15, the total number of quad-hexahedrons q increases to 120 and a CF model is found after 4500 generations. Note that deterministic search would lead to: C (120, 6) /M = 3.65 · 109 /32 = 1.14 · 108 generations. Nevertheless, generating the conformal representation of ‘C-NO’ on current hardware still takes hours.
4 CONCLUSIONS This research investigated the use of ML techniques and Genetic Algorithms for the creation of conformal models in the design and engineering of buildings. This technique is investigated as an alternative to procedural coding approaches. A ML model shows to be able to generate conformal designs, but only when the specific topology (e.g., two spaces side by side) is used in the training process. As such, it is not able
372
to develop a reasoning scheme to find conformal solutions for e.g., 3 spaces in a new configuration, unless it is trained for each specific configuration and in combination with One-hot encoding. For the basic example dataset, there are already 189 different configurations, which implies 189 One-hot encodings. When the number of variables increases for more realistic designs, the number of encodings increases exponentially too. As a result, ML cannot be used. Additionally, ML predictions are compromised numerically. There is a slight error in the results, which makes the predicted conformal model non-cubic, as so unusable. A combination of two GAs is able to create conformal representations of BSDs, at least for the examples tested here. It shows an advantage above deterministic search, for less computations are needed to converge to the optimal solution. In addition, the algorithm is adaptable to solve multiple configurations, and the found conformal geometries are perfect quadhexahedrons without numerical errors. For orthogonal designs, the number of possible cuboids remains relatively low and the GA is able to transform BSDs with two spaces into conformal representations. The GA is also able to transform non-orthogonal designs, but the number of possible solutions increases exponentially. This leads to unfavourably large computation times. In addition, a strategy should be developed to generate a point cloud that for all possible situations ensures a solution. After some preliminary research, a combination of two GAs have been selected to not diffuse the second GA with the relatively simple task of the first GA. Nevertheless, future research could reinvestigate the use of a single GA.
REFERENCES Boonstra, S., Van der Blom, K., Hofmeyer, H., Emmerich, M.T.M., Van Schijndel, A.W.M. & De Wilde, P. (2018). Toolbox for Super-structured and Super-structure Free Multi-disciplinary Building Spatial Design Optimisation. Advanced Engineering Informatics, Volume 36: 86–100. Cavicchio, D. J. (1970). Adaptive Search Using Simulated Evolution. (PhD thesis). Ann Arbor, MI: University of Michigan. Deb, K.(2002). A Fast and Elitist Multiobjective Genetic Algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation. Volume 6(2): 182–197. Dudek, G.(2013). Genetic Algorithm with Binary Representation of Generating Unit Start-up and Shut-down Times for the Unit Commitment Problem. Expert Systems with Applications, Volume 40: 6080–6086. Ezendam, T. (2021). Two Geometry Conformal Methods for the Use in A Multi-disciplinary Non-orthogonal Building Spatial Design Optimisation Framework. (MSc thesis). Eindhoven, The Netherlands: Eindhoven University of Technology.
373
Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. Cambridge, MA: MIT Press. Haymaker, J., Fischer, M., Kunz, J. & Suter, B. (2004). Engineering Test Cases to Motivate the Formalization of an AEC Project Model as a Directed Acyclic Graph of Views and Dependencies. Electronic Journal of Information Technology in Construction, Volume 9: 419–441. He, K., Zhang, X., Ren, S., & Sun, J. (2015). Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification. IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, December 2015. Hofmeyer, H. & Davila Delgado, J.M. (2015). Coevolutionary and Genetic Algorithm Based Building Spatial and Structural Design. Artificial Intelligence for Engineering Design, Analysis and Manufacturing, Volume 29: 351–370. Holland, J.H. (1975). Adaptation in Natural and Artificial Systems. Ann Arbor: University of Michigan Press. Ioffe, S., & Szegedy, C.(2015). Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. Proceedings of the 32nd International Conference on Machine Learning, Lille, France, (July 2015). Vol. 37: 448–456. Kalay,Y. (2004).Architecture’s New Media: Principles,Theories, and Methods of Computer-Aided Design. MIT Press. ISBN 978026211284-0. Kingma, D.P. & Ba, J. (2015). Adam: A Method for Stochastic Optimization. International Conference on Learning Representations (ICLR), San Diego, CA, USA, (May 2015). Kolodiazhnyi, K. (2020). Hands-On Machine Learning with C++: Build, Train, and Deploy End-to-end Machine Learning and Deep Learning Pipelines. Birmingham: Packt Publishing. MacLeamy, P. (2004). Collaboration, Integrated Information and the Project Lifecycle in Building Design, Construction, and Operation. Construction Users Roundtable, WP-2012. Mitchell, T. M. (1997). Machine Learning. New York: McGraw-Hill. ISBN 0070428077. Rocha, M. & Neves, J. (1999). Preventing Premature Convergence to Local Optima in Genetic Algorithms via Random Offspring Generation. Proc. of the 12th International Conference on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems, Cairo, Egypt, (May 1999). Stanovov, V., Brester, C., Kolehmainen, M., & Semenkina, O. (2017). Why Don’t You Use Evolutionary Algorithms in Big Data? IOP Conference Series: Materials Science and Engineering, Volume 173: 12–20. Syswerda, G. (1989). Uniform Crossover in Genetic Algorithms. Proceedings of the 3rd International Conference on Genetic Algorithms, George Mason University, Virginia, June 1989. Wallet, B., Marchette, D. & Solka, J. (1996). A Matrix Representation for Genetic Algorithms. Proceedings of Automatic Object Recognition VI of SPIE Aerosense, ort, May 1996. Volume 2756: 206–214. SPIE. Wong, K. (2015). Evolutionary Multimodal Optimization: A Short Survey. arXiv Online Curated Research-sharing Platform (Cornell University), ref: arXiv:1508.0045.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Predicting the construction duration in the predesign phase with decision trees S. Lauble & S. Haghsheno Institute of Technology and Management in Construction, Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany
M. Franz ModuGen GmbH, Karlsruhe, Germany
ABSTRACT: This paper analyses how well decision trees (DTs) predict the construction duration in the predesign phase. To answer this question, the authors first evaluate the expected prediction accuracy with a survey in the German AEC industry. Second, they compare the prediction accuracy of five DTs (Random Forest, GBR, XGBoost, LightGBM and CatBoost) with artificial neural networks (ANN) and linear regression models (LRMs) in two exemplary data sets from residential projects. The study uses performance indicators mean absolute error (MAE) and mean absolute percentage error (MAPE) as metrics. The results reveal that DTs perform better, with the underlying data sets, than ANNs and LRMs. The expected prediction accuracy of 26% is fulfilled in data set 1 with a MAPE of 13.48% and is nearly reached in data set 2 with a 26.45% MAPE. This shows the potential of using DTs in practice as more and more data in construction is generated. From a practical perspective, the explainabilty of DTs should be further tested in predicting the construction duration.
1 INTRODUCTION Many construction projects show deviations from their original planned and communicated duration. Exemplary delays in major construction projects are the Eurotunnel or the Sydney Opera House (Hall 1982). As the whole project is based on the original predicted key dates, delays can cause additional costs, conflicts, obstruction notices (Braimah 2008) and a loss of reputation. In contrast, project phases with time buffers tie up resources unnecessarily, creating a negative impact on the costs. As deviations are often expected by the project’s stakeholders, deviations should be reduced to a minimal level of acceptance to guarantee the project’s success. Predicting the construction duration within an acceptable level for its stakeholders, is foundational for: – Project feasibility reviews – Investment decisions – Resource planning – Contracts and coordination with partners – Team atmosphere – Definition of the projects’ success. According to Kahneman and Tversky (1979), the reason for inaccurate duration predictions are systematical planning fallacies. There are two major schools of thought explaining planning fallacies, the ‘psycho strategists’ (e.g., Flyvbjerg 2006; Kahneman & Tversky 1979; Siemiatycki 2009), and the ‘evolutionary
374
theorists’(e.g., Love et al. 2011; Odeyinka et al. 2012). According to the psycho strategists, planners often take an “internal view” and give more weight to active processes, than to compare the current project with the results of historical projects. They tend to underestimate risks and overestimate the political as well as economic benefits. The evolutionary theorists in contrast have the opinion that deviations are the result of changes during the project. Both schools of thought show the need of an objective prediction showing the relationship between project content and its duration. For this reason, analytical data-driven models are a highly investigated research topic. These intend to achieve a higher prediction accuracy and hereby reduce deviations. Machine Learning (ML) methods, especially, show a higher prediction accuracy compared to simple statistical methods as part of conventional programming (Chen et al. 2006; Dissanayaka et al. 1999; Lam et al. 2016; Petruseva et al. 2012; Wang et al. 2010). Additionally, when planning in the predesign phase, a high uncertainty exists for planners. Uncertainty is defined as the difference between the amount of information needed for a particular task and the amount of information an organization already possesses (Galbraith 1973). The ability to obtain more information requires transparency, and an explanation of the time influencing factors, as well as their dependencies on the construction project. DOI 10.1201/9781003354222-48
The field of explainable artificial intelligence (XAI) considers both requirements, prediction accuracy and explainabilty. In general, there is a distinction between inherently explainable models and methods for explaining complex black-box models, such as artificial neural networks (ANNs) (Molnar 2019). Arrieta (2020) defines inherently explainable models where a human can stimulate the model with its decomposability and algorithmic transparency. Machine learning decision trees are the only method which meet the requirements to be an inherently explainable model without the support of additional tools (Arrieta 2020). This work therefore tests, if the prediction accuracy of DTs is comparatively sufficient for predicting the construction duration in the predesign phase. On the one hand, a survey determines the acceptance level of deviations from the predesign phase. On the other hand, an exemplary data analysis compares the prediction accuracy of common analytical methods for predicting durations with DTs. The mean absolute error (MAE) and the mean absolute percentage error (MAPE) are used as performance indicators. The survey results are then compared to the results of the data analysis.
2 LITERATURE REVIEW 2.1 Analytical methods to predict the construction duration Numerous analytical methods exist for predicting durations in construction projects. A literature review was conducted with the following keywords: (‘construction’) and (‘project’ or ‘time’) and (‘forecast’ or ‘performance’ or ‘duration’ or ‘causes of delay’ or ‘engineering’ or ‘planning’ or ‘formation’ or ‘influence factor’). 186 publications were selected. With a forward and backward search, a further 189 references were added. To categorize the list of 375 references, the systematical approach for analytics of Davenport and Harris (2007) is used. They define four successive analytical levels: Statistical analysis to investigate dependencies, extrapolations to analyze continuous trends, predictive models to forecast next steps, and optimization to analyze the best next steps. Table 1 summarizes the identified methods with exemplary references. Linear regression models (LRM) are especially frequently used methods. One of the first models for predicting the construction duration was developed in 1969 by Australian researcher Bromilow with the equation: T = K ∗ CB
(1)
Where T is time (in working days), K is a constant describing the general level of time performance for a $1 million project, C is the cost of the contract (in millions), and B is a constant that reflects the sensitivity of time to cost. Based on this equation, other time-cost
models were developed. These models include limited variables and are therefore easy to simulate and interpret. Table 1. Analytical methods for predicting construction durations in the literature. Level
Method
References
Statistical analysis
Relative Importance Index (RII) Correlations Linear regressions Fuzzy systems Box Jenkins Monte Carlo Simulation Artificial Neural Networks (ANN) AI ensemble Genetic algorithm Ant colony Tabu search Simulated annealing
Meng (2012)
Extrapolations
Predictive models Optimizations
Walker (1995) Bromilow (1969) Wu (1994) Agapiou (1998) Albogamy (2014), Moret (2016) Lam (2016), Erdis (2013) Zheng (2004) Kalhor (2011) Jung (2016) Kumar (2011)
In contrast, publications about ANNs show better results, compared to linear regressions, by including more variables (Chen et al. 2006; Dissanayaka et al. 1999; Lam et al. 2016; Petruseva et al. 2012; Wang et al. 2010). ANNs, oriented on the structure and functionalities of the human neural system, transfer what has been learned to new situations. The results of ANNs are generally not comprehensible and based on a “black box” approach. LRMs and ANNs show that as the analytical level of the prediction accuracy increases, the explainabilty decreases. Therefore, this work uses LRMs and ANNs as reference models for explainabilty and prediction accuracy. 2.2 Machine learning decision trees McCarthy et al. (1964) were the first to define the term Artificial Intelligence (AI) with ML as subset of AI. In 1955 they wrote that the “goal of AI is to develop machines that behave as if they had intelligence”. Since natural intelligence includes consciousness, emotions, and intuition, in addition to complex cognitive abilities, the following definition is more appropriate: AI enables computers to learn independently without being explicitly programmed (Samuel 1959). AI models, unlike statistical models of conventional programming, can therefore handle a growing amount of data and can independently recognize (even unknown) patterns in data sets. To reach this goal, four development steps can be seen in the history (Jaakkola et al. 2019): the programming language (1950s), expert systems for specific purposes (1970s), the AI architecture (1980s) and up from 2010s frequently mentioned branches of AI are machine learning (ML) and deep learning (DL) as self-learning
375
applications (Russel et al. 2012). However, from here on, this paper will only focus on ML as a branch of AI. As development in the field of ML progresses, mathematical models also become more extensive and complex. In addition to the LRMs and ANNs previously applied in construction scheduling, this work uses socalled decision trees (DTs) with the goal of regression.
3 MATERIALS AND METHODS 3.1 Method overview This work uses three main methodical parts (Figure 2):
Figure 2.
Figure 1. Exemplary decision tree out of the Random Forests (for data set 2 – San Francisco, USA).
Figure 1 shows a representation of a machine learning decision tree (Random Forest), with a maximal depth of five levels, based on historical project data. Starting from the root of the decision tree, the tree divides at the branches, or edges, to branch nodes or leaf nodes (Myles 2004). The leaf node shows the final decision on the predicted duration. The mean squared deviation serves here as a performance indicator. DTs recursively divide the data set along attributes with the aim that the resulting structure produces the most accurate classification and thus prediction possible. The basic principle during the training of a decision tree is as follows: Within the model, a split of the data is proposed at each branch of the tree according to one or multiple attributes. On this basis, the advantage of each split is calculated, and the one with the highest gain is chosen. The highest gain refers to the split that separates the data into the most coherent subsets. This is called an optimization of the “information gain” (Rockach 2005). The final node shows the prediction (duration), the number of projects in this reference class, and a performance measure. Therefore, DTs have the advantage, that they can represent the human decision-making process visually. Compared to other AI methods, DTs can handle missing attributes well (Sheh 2017), include categories in the prediction, and show good results even with limited data. Advanced extensions of decision trees include Random Forest (Liaw 2002), Gradient Boosting Regression, XGBoost Tree (Chen at al. 2016), LightGBM (Ke 2017), and CatBoost (Prokhorenkova 2018). These DTs are based on combining multiple decision trees.
Overview of methods used.
The core elements to this research are the literature review, the survey (Section 3.2) and the data analysis (Section 3.3) with its performance evaluation (Section 3.4). The survey defines the average maximum deviation where participants are still satisfied. This can vary, depending on organizational role of the participant. Therefore, the survey also contains demographic questions. The performance of the data analysis can vary by the size and structure of the used data set.
3.2 Survey methodology The survey includes four questions about the demographics of the participants. These comprise of the company’s role in projects (client or contractor role, consulting role), project specialization (highbuildings, industrial construction, turnkey construction, civil engineering, transportation construction, prefabricated houses, and others), the size of the company (measured by employees and sales), and the professional experience of the participant. Furthermore, a matrix question was asked in each case about over- and underruns: “I am still satisfied with the result of the construction project if the construction duration is delayed by a maximum of ... % slower (or faster) than the planned duration in the predesign phase”. A pre-test was conducted with two individuals. The first person was from the research community and provided feedback on technical terms as well as the structure of the survey. The second person, a practitioner, reviewed the survey again for comprehensibility. The survey was distributed to stakeholders in the German AEC industry. The survey was available in November 2021. The link to the survey was distributed via LinkedIn groups, as well as to personal contacts to industry experts and scheduling experts. The contacts were asked to distribute the link to one other person each time to get a wider reach of survey participants.
376
A total of 87 participants took part, of which 66 participants completed the survey in full. This results in a dropout rate of 24.14%. According to Hill (1998), Roscoe (p. 163, 1975), and Abranovic (p.307-308, 1997), 30 participants are sufficient in exploratory studies. Chassan (1979), however, calls for a minimum sample of 50 to 100. According to these definitions, the sample size of the survey is sufficient. The participation took a median of three minutes. Notably, the survey was often terminated after the demographic questions.
Table 2.
Key figures of the used data sets.
Characteristics Number of Projects Number of Features Average duration Standard deviation
Data set 1 Data set 2 Tehran, Iran
San Francisco, USA
372
184
32
9
6.27 quarters (equals 564.3 days) 2.09 quarters (equals 180.2 days)
1,179.4 days 808.7 days
3.3 Data analysis The data analysis uses two data sets containing information on new residential construction. Different models for DTs were applied to the data sets and compared to the results of existing models. This not only allows for a comparison of different prediction models, but also for a comparison of the characteristics’ influence of the data set on the prediction accuracy. This is because the two data sets are of different size, each containing the duration of construction at a different granularity with different features. Data set 1 describes construction data of 372 residential buildings in Tehran, Iran. The data set documents the construction execution duration in quarters of year, however, does not document the design duration. The descriptive features in the data set are the costs, project features and economic boundary data. The economic variables include values for each of the five quarters prior to the start of construction (Hossein et al. 2016). The data set is available online at the University of California, Irvine’s “Machine Learning Repository” (2020). The average duration of a residential building is 6.27 quarters with a standard deviation of 2.09 quarters (equals approximately to 564.3 days average duration and 180.2 days standard deviation, with 90 days in a quarter). Data set 2 includes 184 completed apartment construction projects in San Francisco, USA of the San Francisco Government (2020). Only the projects with a permit type of “new construction” and a current status of “complete” were considered to benchmark with data set 1. To clean the data set, lines containing the same content were deleted. The data set includes the available construction time in days, after building permits were issued. Similarly, to data set 1, data set 2 does not document the time required for design. The projects are described by floors, residential units, approved costs as well as the submitted planning documents. Additionally, the data documents the street name and neighborhood area. The average duration in data set 2 is 1,179.4 days.The data set has a higher standard deviation than data set 1, with a standard deviation of 808.7 days. Table 2 describes the main properties of both data sets. The granularity of documentation, the influence of documented projects, as well as features and the standard deviation of the target variable can be compared.
Public-open data: This study assumes that the economic and political situation influences the prediction and therefore will be included in the analysis. The two data sets were further enriched with 65 external data points from OECD data (2020) as well as global economy factors (2020). The goal of the data points selected, is to describe the local economy and politics of the data sets’ regions. Among others, these include key figures on inflation, the corruption and innovation index, and the number of building permits in the respective country. This information was integrated as an example for the first, second and last year of the respective construction project. This results in 195 further features per project (65 features per year).
3.4 Performance indicators The programming was done using Google Colab and the following libraries: TensorFlow (1.14.0), kerasapplications (1.0.6), CatBoost (0.24.3), and shap (0.37.0). The data sets were split into a training and test data set with a ratio of 80/20. The variable k=3 was chosen for a cross-validation. To create a comparative analysis, each tree has a maximal depth of six levels. The study also uses a python-code with defined hyperparameters. The mean absolute percentage error (MAPE) or the mean absolute error (MAE) can measure the performance prediction. The smaller both indicators are, the higher the prediction performance or accuracy is. The MAPE is the ratio of the difference between the actual output value y and the predicted value yˆ to the actual output value y over all data points (Hyndmann et al. 2006). MAPE =
n 100% yi − yˆ i n i=1 yi
(2)
The MAE calculates the difference between the predicted value yˆ and output values y (Hyndmann et al. 2006). The MAE measures the absolute quality. MAPE =
377
n 1 yi − yˆ i n i=1
(3)
Due to the different granularity of the duration in both data sets (quarters and days), the MAPE is more suitable as a unified metric. When comparing to linear regression models, the MAE is better suited for an absolute comparison. With this result, a comparison can be made between existing models, such as ANN, linear regressions, and XAI methods. This comparison provides insights into the applicability of XAI methods for predicting the duration of project phases. Furthermore, the visualization of the decision tree can show the user the reference classes as end nodes of the decision tree. A metric for the deviation within the end nodes serves as a measure of dispersion.
4 RESULTS AND DISCUSSION 4.1 Survey results The full responses can be found on Mendeley (2021). A total of 66 people participated in the survey with 24 (36%) clients, 26 (40%) contractors, and 16 (24%) consultants representing the client or contracting side. It was possible to select multiple options regarding the specialization of the company. The companies specialize mainly in high-rise buildings (44 responses), turnkey construction (31 responses), industrial construction (29 responses) and civil engineering (21 responses). Furthermore, most of the participants were from large or medium-sized companies.
up to 16% and accept a much larger acceleration with up to 35% compared to clients and contractors. The results of the survey are to be understood as a snapshot. Ideally, those involved in the construction project reduce plan deviations, and therefore change the expectation of deviations. Improving scheduling strategies and methods reduces the level of acceptance towards deviations. 4.2 Model-centric analysis To evaluate the prediction performance, the models are applied to both data sets with and without the external data. This way, the added value of the external data can be evaluated separately. Table 4 shows the mean percentage deviation (MAPE) as a comparative result for boosting DTs. In the Iranian data set, the lowest MAPE is 13.5% (equals MAE 0.85 quarters), while in San Francisco data set, their lowest MAPE is 26.5% (equals MAE 423 days). In general, CatBoost shows the best prediction indicators, both with and without the integration of the external data. However, adding the external data further increases the accuracy of the forecast in some cases. Table 4. Comparative evaluation of boosting DTs with the performance indicator MAPE. Data set 1
Table 3. Average percentage of maximum deviations in the planned construction duration in the predesign phase, where participants were still satisfied. Company role
Delay
Acceleration
Average
Client (n = 24) Contractor (n = 26) Consultancy (n = 16) Sum (n = 66)
16% 27% 16% 20%
25% 28% 35% 31%
20% 28% 25% 26%
Data set 2
ML decision tree
*
**
*
**
Random Forest GBR XGBoost LightGBM CatBoost
19.6 17.3 18.1 18.9 16.8
17.8 15.5 15.3 16.8 13.5
36.8 38.7 53.6 48.4 39.7
26.8 27.5 19.1 34.9 26.4
*Without external data; ** with external data; in italic: best result per data set.
Table 3 summarizes the average percentage of maximum deviations for the planned construction duration in the predesign phase, per company role, where participants were still satisfied. On average, participants are still satisfied if construction is delayed by up to 20% compared to the planned duration. If the construction project is faster than planned, an average of up to 31% deviation is accepted. Taking the average value per participant from overruns and underruns of the planned duration, an acceptance level of 26% can be assumed on average. When comparing the roles and their expectations, contractors have the highest acceptance level for delays, with 27%. Whereas clients are on average only satisfied with up to 16% schedule delays. If the construction work is completed faster than planned, clients and contractors have approximately the same expectations. For clients this is 25%, and for contractors 28%. Consultants are still satisfied with delays of
When comparing the two datasets, the San Francisco data set shows worse results. This may be due, first, to the higher standard deviation, second, to the smaller size of the data set, and third, to its features. In data set 1 (Tehran, Iran), more features were documented per project. To reduce the standard deviation, a pre-grouping of the data, for example in short, medium, and long-lasting projects, could improve the prediction accuracy. Next, a comparison is drawn between the result of the CatBoost in data set 2, and the LRMs and ANNs methods. The regression formulas are not applicable to data set 1 (Tehran, Iran) because the documentation is in quarters and the corresponding linear regression formulas require a documentation in days. A conversion of quarters into days is only possible as an approximation. This method is therefore excluded to ensure precise comparability of the results. The ANNs represent the state-of-the-art research on predicting construction durations with ML models.
378
Table 5 shows the results of the comparison using the MAE as the output value for the linear regressions. CatBoost performs better than existing statistical projections (Chan 1999, 2001; Choudhoury 2003; Ireland 1983; Le-Hoai 2009; Ogunsemi 2006), even in comparison to ANNs. In data set 2 the MAE of the ANN is 573 days (36.2% MAPE) and the CatBoost is 423 days (26.5% MAPE). In general, the prediction accuracy shows high deviations to the realized construction duration in data set 2. In comparison, data set 1 has a MAE of the ANN of 3.73 quarters (70.5% MAPE), and with CatBoost a MAE of 0.85 quarters (13.5%).
4.3 Alignment of model prediction accuracy and deviations expectations With the results, it is possible to compare the deviation of the construction duration planned in the predesign phase (delays and acceleration). With the allowed 26% of deviations, data set 1 achieves satisfying results with 13.5% MAPE, and data set 2 almost meets this with a 26.5% MAPE. In data set 1, delay as well as acceleration values stay within the accepting ranges for all roles (clients, contractors, and consultancies). In comparison, in data set 2, the accepted acceleration of up to 31% is met, but the accepted delay, of a maximum of 20%, is not achieved. Furthermore, the expectation of the clients in under- and overruns of the planned duration is not reached. For consultants, the maximum of 16% in delays is below the value of 26.5% achieved with the CatBoost. Accordingly, the satisfying prediction accuracy requirement can only be partially met for data set 2.
Table 5. Comparative evaluation to LRMs andANN for data set 2 with the performance indicator MAE (San Francisco, USA). Data set 2 (MAE in days) Ireland (1983) Chan (1999) Chan (2001) Ogunsemi (2006) Choudhoury (2003) Le-Hoai (2009) Own regression ANN** Random Forest* CatBoost**
relevance of cost variables as included in Bromilow’s forecasting model (Bromilow 1969). The higher prediction accuracy in data set 1 highlights the relevance of including product information in the prediction (e.g. lot area). The research question posed at the beginning can now be answered with respect to the used data sets: Compared to existing methods, DTs show more accurate and explainable results.
509 708 488 508 1.108 585 439 573 587 423
*Without external data; ** with external data; in italic: best result of data set 2.
Even data set 1 shows a higher difference in the prediction accuracy of ANN and CatBoost. The result of 0.85 quarters, which equals approximately to 77 days (with 90 days in a quarter), is by far the best result reached. As a result, the application of DTs can not only improve explainabilty, but also enhance the prediction accuracy compared to linear regressions and ANNs. The lower prediction accuracy of ANNs may be due to the relatively small amount of training data that is often present in construction projects. The addition of the external data further increases the prediction accuracy. Here, only the best external factors are included within the DTs, whereas in ANNs, the inclusion of all features can lead to an overtrained model. When performing a feature optimization, ANNs become even worse. Due to the high difference of the prediction accuracy in both data sets, the documentation structure is relevant when using analytical methods. When analyzing the feature importance with a Random Forest Regressor in the Iranian data set (with external data), the three most important features are: Actual construction costs output, the lot area, and the equivalent preliminary estimated construction cost based on the prices at the beginning of the project in a selected base year. While in the San Francisco data set, the most important features were grouped costs and the housing price in the first year. This confirms the
5 CONCLUSION This publication analyses the prediction accuracy of DTs, regarding their ability to predict the construction duration in the predesign phase. The indicators MAE and MAPE evaluate their performance. First, a survey identifies the average percentage of maximum deviations in the planned construction duration in the predesign phase, where participants were still satisfied. Second, a data analysis with two exemplary data sets of residential buildings (Tehran, Iran and San Francisco, USA) compares the performance of DT to LRMs and ANNs as part of ML. Additional publicopen data enriches both data sets on economic and political variables of the respective country. Finally, the results are compared to the accepted deviations by clients, contractors, and consultants, determined in a survey in the German region. DTs not only show advantages in the prediction accuracy, compared to linear regression models and ANNs, but the explainabilty of the models is also of benefit. For both data sets, the CatBoost has shown particularly promising results. For data set 1 the MAPE is 13.5% (Tehran, Iran) and is 26.5% for data set 2 (San Francisco, USA). The comparison of both data sets shows the number of documented projects, the standard deviations of the duration, and that the documented features have an
379
influence on the prediction accuracy. The data-centric analysis shows the relevance of product information (e.g., lot area, total floor area), price information, and public-open data. Following, documenting especially these features with a growing amount of construction data will get necessary to use DTs in predicting construction duration at an acceptable level. To define an acceptance level, a survey was conducted. Here, 66 people participated. The average accepted deviation is up to 26%, specifically 20% in delays and 31% in accelerations. Results show that, on average, the performance of DTs in data set 1 are met, while for data set 2 the target is nearly met when including external data. Here, especially the expectations of clients and consultancies cannot be met. According to the survey, the client has the highest expectations as he is directly affected by any deviation. Without the external data, the acceptance level is not met for any stakeholder group. In contrast, data set 1 shows its strengths due to the large number of documented features, the size of the data set and the lower standard deviation of the duration. When analyzing if DTs fulfill the acceptance level of deviations, the conclusion is that role and the characteristics of the data set is of importance. It should be noted that the survey is exploratory. A confirmatory analysis of the evaluations should be conducted with a larger sample. Additionally, smaller companies should be added to the research pool. Medium and large companies are overrepresented for the region under consideration in this survey. The data analysis shows that a day-by-day prediction is not yet possible, even with ML methods. Still with ML methods the acceptance level can be met. The question rises whether humans predict more accurately than machines and in which cases a humanor machine-based prediction is preferred. Depending on the results, a weighted combination of human interaction and machines is recommended. To further improve the prediction accuracy data quantity and quality, especially for the cost features and available product information can be enhanced as well a pregrouping of the data can support the results of the DTs. In conclusion, it is necessary to further explore the collaboration between humans and machine learning models in the construction industry, especially when regarding the potential of interpreting DTs.
ACKNOWLEDGEMENT This work was supported by the Federal Ministry for Economic Affairs and Energy of the German Government as part of the research project Smart Design and Construction (SDaC). REFERENCES Abranovic, W.A. 1997. StatisticalThinking and DataAnalysis for Managers. MA: Addison-Wesley, Reading.Agapiou, A.
380
& Notman, D. & Flanagan, R. & Norman, G. 1998.A Time Series Analysis of UK Annual and Quarterly Construction Output Data (1955–1995). Construction Management and Economics, 16(4):409–416. Albogamy, A. & Dawood, N. & Scott, D. 2014. A Risk Management Approach to Address Construction Delays From The Client Aspect. Computing in Civil and Building Engineering, 06:1497–1505. Arrieta, A.B. & Díaz-Rodríguez, N. & Del Ser, Bennetot, J.A. & Tabik, S. & Barbado, A. & García, S. & Gil-López, S. & Molina, D. & Benjamins, R. et al. 2020. Explainable artificial intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges Toward Responsible AI. Information Fusion, 58:82–115. Braimah, N. & Ndekugri, I. 2008. Factors influencing the selection of delayanalysis methodologies. International Journal of Project Management, 26(8):789–799. Bromilow, F.J. 1969. Contract Time Performance Expectations and the Reality. Building forum. 1.3: 70–80. Chan, A.P.C. 1999. Modelling Building Durations in Hong Kong. Construction Management and Economics, 17.2: 189–196. Chan, A.P.C. 2001. Time–cost Relationship of Public Sector Projects in Malaysia. International Journal of Project Management, 19.4: 223–229. Chassan, J.B. 1979. Research Design in Clinical Psychology and Psychiatry. Irvington Pub. Chen, W.T. & Huang, Y.-H. 2006. Approximately Predicting the Cost and Duration of School Reconstruction Projects in Taiwan. Construction Management and Economics 24.12: 1231–1239. Chen, T. & Guestrin, C. 2016. Xgboost: A Scalable Tree Boosting System”. Proceedings of the 22nd acm Sigkdd International Conference on Knowledge Discovery and Data Mining, 785–794. Choudhury, I. & Rajan, S. 2003. Time-cost Relationship for Residential Construction in Texas. Davenport, T.H. & Harris, J.G. 2007. Competing on Analytics – The NewScience of Winning. Harvard Business Press, Boston, Massachusetts. Dissanayaka, S. & Kumaraswamy, M. 1999. Evaluation of Factors Affecting Time and Cost Performance in Hong Kong Building Projects. Construction and Architectural Management 6.3:287–298. Erdis, E. 2013. The Effect of Current Public Procurement Law on Duration and Cost of Construction Projects in Turkey. Journal of Civil Engineering and Management, 19:121–135. Flyvbjerg. B. 2006. From Nobel Prize to Project Management: Getting Risks Right. Project management journal, 37(2):5–15. Galbraith, J. 1973. Designing Complex Organizations. Reading, Mass. Global Economy (2022) The Global Economy. Available at: https://www.theglobaleconomy.com (Accessed: 6 May 2020) Hall, P. 1982. Great Planning Disasters: With a New Introduction. Univof California Press , vol. 1. Hill, R. 1998. What Sample Size is “Enough” in Internet Survey Research. Interpersonal Computing and Technology: An electronic journal for the 21st century, 6(3-4):1–12. Hyndman, R.J. & Koehler, A.B. 2006. Another Look at Measures of Forecast Accuracy. International Journal of Forecasting, 22.4: 679–688. Ireland, V. 1983. The Role of Managerial Actions in the Cost, Time and Quality Performance of High-rise Commercial Building Projects. Construction Management and Economics, 3: 59–87.
Jaakkola, H., Henno, J., Mäkelä, J., and Thalheim, B. 2019. Artificial Intelligence Yesterday, Today and Tomorrow, 2019 42nd International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO): 860–867. Jung, M. & Park, M. & Lee, H.-S. & Kim, H. 2016. Weather Delay Simulation Model Based on Vertical Weather Profile for High-rise Building Construction. Journal of Construction Engineering and Management, 142: 04016007, 01. Kahneman, D. & Tversky, A. 1979. Intuitive Prediction: Biases and Corrective Procedures. Tech. rep. Decisions and Designs Inc Mclean Va: 2. Kalhor, E. & Khanzadi, M. & Eshtehardian, E. & Afshar, A. 2011. Stochastic Time–cost Optimization Using Nondominated Archiving Ant Colony Approach. Automation in Construction, 20(8):1193–1203. Ke, G. 2017. Lightgbm: A Highly Efficient Gradient Boosting Decision Tree. Advances in neural information processing systems: 3146–3154. Kumar, V. & Abdulal, W. 2011. Optimization of Uncertain Construction Time Cost Trade Off Problem Using Simulated Annealing Algorithm. 2011 World Congress on Information and Communication Technologies, 12: 489–494. Lam, K. & Olalekan, O. 2016. Forecasting Construction Output: A Comparison of Artificial Neural Network and Box-Jenkins Model. Engineering, Construction & Architectural Management 23, 2016:302–322. Le-Hoai, L. & Lee, Y. Time-cost Relationships of Building Construction Project in Korea. Facilities, 27:549–559. Liaw, A. & Wiener, M. 2002. Classification and Regression by RandomForest”. R News 2.3: 18–22. Love, P. ED, Edwards, D. J and Irani, Z. 2011. Moving Beyond Optimism Bias and Strategic Misrepresentation: An Explanation for Social Infrastructure Project Cost Overruns. IEEE Transactions on Engineering Management, 59(4):560–571. Lundberg, S.M. & Erion, G.G. & Su-In Lee, S.-I. 2018. Consistent Individualized Feature Attribution for Tree Ensembles. arXiv. Machine Learning Repository (2019) Residental Building Data Set Data Set.Available at: https://archive.ics.uci.edu/ ml/datasets/Residential+Building+Data+Set (Accessed: 6 May 2020). McCarthy, E.J. & Perreault Jr., W.D. 1964. Basic Marketing, Richard D. Irwin, Homewood, IL 181, 1964. Mendely (2021) Results of A Survey on Expectations of Adherence to Project Programs in Construction Projects. Available at: https://data.mendeley.com/datasets/2275r hvkt7/1 (Accessed: 20 December 2021). Meng, X. 2012. The Effect of Relationship Management on Project Performance in Construction. International Journal of Project Management 30, 02. Moret, Y & Einstein, H. 2016. Construction Cost and Duration Uncertainty Model: Application to High-speed Rail Line Project. Journal of Construction Engineering and Management, 142:05016010, 04. Molnar, C. 2019. Interpretable Machine Learning. A Guide for Making Black Box Models Explainable.
381
Myles, A.J. 2004. An Introduction to Decision Tree Modeling”. Journal of Chemometrics: A Journal of the Chemometrics Society, 18.6: 275–285. Odeyinka, H., Larkin, K., Weatherup, R., Cunningham, G., McKane, M. and Bogle, G. 2012. Modelling Risk Impacts on the Variability Between Contract Sum and Final Account. Royal Institution of Chartered Surveyors, London: 1–19. OECD data. 05/06/2020. OECD data. https://data.oecd.org/ Ogunsemi, D.R. & Jagboro, G.O. 2006. Time-cost Model for Building Projects in Nigeria. Petruseva, S. & Zujo, V. & and Pancovska, V.Z. 2012. Neural Network Prediction Model for Construction Project Duration. International Journal of Engineering Research and Technology 1. Prokhorenkova, L. 2018. CatBoost: Unbiased Boosting with Categorical Features. Advances in neural information processing systems: 6638–6648. Rafiei, M.H. & Adeli, H. 2016. A Novel Machine Learning Model for Estimation of Sale Prices of Real Estate Units”. Journal of Construction Engineering and Management 142.2. Rokach, L. & Maimon, O. 2005. Decision Trees. Ed. by Oded Maimon and Lior Rokach. Boston, MA: Springer US, 165–192. Roscoe, J.T. 1975. Fundamental Research Statistics for the Behavioral Sciences [by] John T. Roscoe. Holt Rinehart & Winston, New York. Russell, S.J. & Norvig, P. 2012. Artificial Intelligence. A Modern Approach. Pearson Studium, 3. Samuel,A.L. 1959. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development, 3.3: 210–229. San Francisco Government (2022) Housing and Building Permits. Available at: https://data.sfgov.org/Housing-andBuildings/Building-Permits/i98e-djp9 (Accessed: 6 May 2020). Shapely, LSA. 1953 Value for n—pemon game. Sheh, R. 2017. Why DidYou Do That? Explainable Intelligent Robots, 631. Siemiatycki, M. 2009. Academics and Auditors: Comparing Perspectives on Transportation Project Cost Overruns. Journal of Planning Education and Research, 29 (1):142–156. Walker, D.H.T. 1995. An Investigation Into Construction Time Performan ce. Construction Management and Economics, 13(2):263–274. Wang, U.-R. & Gibson Jr, G.E. 2010. A Study of Preproject Planning and Project Success Using ANNs and Regression Models. Automation in Construction 19.3:341–346. Wu, R.W. & Hadipriono, F.C. 1994. Fuzzy modus ponens deduction technique for construction scheduling. Journal of Construction Engineering and Management, 120(1):162–179. Zheng, D. & Ng, S. & Kumaraswamy, M. 2004. Applying a Genetic Algorithm Based Multiobjective Approach for Time-cost Optimization. Journal of Construction Engineering and Management, 130, 04.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Tower crane layout planning through Generative Adversarial Network R. Li, H.L. Chi, Z. Peng & J. Chen The Hong Kong Polytechnic University, Hong Kong, China
ABSTRACT: Tower cranes are globally utilized in construction projects to transport components vertically and horizontally, which governs the construction schedule and requires proper locations. However, in practice, the layout of the tower crane is mainly decided by the experience of construction contractors or managers, lacking quality assurance. Generative adversarial network (GAN) is an emerging deep learning technology to generate synthetic images with predictive nature, applied in many research areas, especially in automatic design. Given that, this paper proposed a TC-GAN to identify an appropriate tower crane layout. Information on construction projects was gathered and calculated to obtain a high-quality training dataset considering efficiency and safety. Then, a framework derived from cGAN was applied for the TC-GAN generator and discriminator, training on the massive dataset. The learning rate selection was conducted based on evaluating the quality and rationality of the generated image, which validated the TC-GAN performance in tower crane layout planning.
1 INTRODUCTION With the accelerated urbanization in China, the contradiction between the growing urban population and the limited urban land resources has become increasingly apparent (Yu et al. 2019). In 2021, the urbanization rate of China’s resident population reaches 64.72%, which is still below that of 80% in developed countries. Moreover, the urban built-up area reaches 61,000 square kilometers in 2020. Modern buildings are extended upward and downward by increasing the building height and underground depth to expand the living space of the population. High-rise buildings can accommodate more urban population and improve land utilization, which is essential to solve the above contradiction (Wu et al. 2020). Due to the improvement of construction technology and building materials, the height of high-rise buildings in cities has gradually increased, which leads to a heavier vertical transportation load (Ding et al. 2011). Tower cranes are widely utilized on-site transportation machinery to conveying various building materials, including heavy steel and massive panels formwork (Huang et al. 2011). Therefore, due to the building height promotion for the increased lifting height and weight, tower cranes have become the most critical vertical transportation machinery governing high-rise buildings’ construction schedule. Proper tower crane lifting planning can accelerate the construction process by declining transportation time and avoiding secondary handling and installation to reduce construction costs (Lin et al. 2020). In the actual construction practices, contractors mainly rely on their experience to decide the tower crane’s
382
types and location, which sometimes leads to significant deviations in decision qualities. For example, the allowable bearing capacity of soil is often ignored in the selection of crane locations. Since cranes are susceptible to ground conditions, inadequate site investigation prior to the crane operation is a nonnegligible safety issue on construction sites, which has been considered a primary cause leading to serious crane accidents such as crane topples (Chen et al. 2022). Therefore, it is vital to provide suitable Tower Crane Layout Planning (TCLP) to scientifically solve unsatisfactory efficiency and safety concerns. Adopting deep learning methods in executing construction activities is becoming more prevalent in the construction industry to produce safe and efficient decision suggestions. Deep learning uses the backpropagation technique to show how a machine should adjust its internal parameters, revealing detailed structure and features in massive data sets (Huang et al. 2011). Based on training on an adequate high-quality dataset, the internal parameters of networks can be obtained, and a decision will be generated after uploading the corresponding parameters and receiving input data. Recently, numerous deep learning methods have been successfully adopted in various design and transportation decision-making, such as automated architectural home design (Huang &Zhang 2018; Zhou et al. 2021). Therefore, this study proposes TC-GAN, one of the emerging deep learning techniques, to automatically provide a rational tower crane layout in high-rise buildings considering construction efficiency and safety, which can assist managers and constructors in decision making. Section 2 reviews the existing research DOI 10.1201/9781003354222-49
on the layout of tower cranes and the related deep learning algorithms of image-to-image translation. Section 3 elaborates on the central tower crane layout problem, including the employed database and its collection method, evaluation metrics, and research problem assumptions. Section 4 proposes TC-GAN to generate the proper tower crane layout results, describing the neural network architecture of the generator and discriminator. Section 5 displays the implementation procedure and results of TC-GAN, discussing the learning rate selection of TC-GAN. Finally, conclusions are drawn in Section 6. 2 LITERATURE REVIEW 2.1 TCLP problem The TCLP problems are used to be regarded as operations research problems with optimization objectives under various constraints, which aim to determine the layout of tower cranes to enable vertical and horizontal transport of components on the construction site efficiently (Lecun et al. 2015). Scholars proposed several mathematical models to address this problem scientifically, considering complicated construction situations. For the mathematical models, research before 2000 was mainly focused on simplifying the hook movements as mathematical models in stationary construction situations. In1983, Walter and Richard proposed a mathematical perspective model to establish the optimal location of a single crane within a construction site (Rodriguez-Ramos & Francis 1983). Later, the dynamic programming model, multiple linear regression model, mixed-integer- linear programming (MILP) model, and binary mixed-integer linear program (BMILP) model were gradually applied to solving TCLP problems to minimize the crane transportation time or rental cost (Abdelmegid et al. 2015; Furusaka & Gray 1984; Yeoh & Chua 2017; Zhang et al. 1999). Meanwhile, various constraints were gradually considered, such as the lifting radius of the tower crane, rental, labor and setup cost, maximum and minimum load weight, tower crane type, and area of tower crane base (Ali Kaveh & Yasin 2020; Marzouk & Abubakr 2016; Zhang et al. 1999). Various optimization methodologies were applied to efficiently address the TCLP mathematical model, obtaining the exact and near-optimal tower crane layout planning. Branch-and-bound technology was applied to obtain the exact optimal planning with the consideration of priority for urgent material, taking plenty of computing time for traversing all possible locations (Yeoh & Chua 2017). Metaheuristics and hybrid-metaheuristics algorithms were introduced to find a nearly optimal tower crane location within a tolerable calculation time range. Genetic algorithm (GA) was a regularly used algorithm in obtaining the optimal layout of the tower crane (Tam et al. 2001). In addition to GA, other metaheuristics algorithms, such as the firefly algorithm (FA), colliding bodies optimization (CBO), enhanced colliding
bodies optimization (ECBO), vibrating particles system (VPS), and enhanced vibrating particles system (EVPS), upgraded sine cosine algorithm (USCA), whale optimization algorithm (WOA), slap swarm optimization (SSA), sine cosine algorithm (SCA), also have been applied (Kaveh & Vazirinia 2018, 2020; Wang et al. 2015). Several hybrid-metaheuristics algorithms were also applied, including artificial neural network (ANN) - GA and particle bee algorithm (PBA), which combined honey bees (bee algorithm, BA) and birds (particle swarm optimization, PSO) (Lien & Cheng 2014; Tam & Tong 2003). However, using these optimization algorithms is based on extracted data from the building plan, which is labor-intensive to obtain and tough visualization of results. As reviewed, most existing research has focused on proposing various mathematical models and optimization algorithms. Although these methods perform exceptionally well under specific optimization objectives and nearly comprehensive constraints, some drawbacks exist, such as unstable accuracy and success rate. In addition, the calculation time exponentially increases when facing large-scaled and highcomplexity cases, making them unable to generate suitable solutions within a tolerable time. In summary, these methods require much manual input and are time-consuming to calculate the result. Moreover, the output of these numerical models is abstract and complex for a site manager to interpret fully. Hence, it is necessary to propose a more intelligible method. 2.2 Generative Adversarial Networks (GANs) Generative adversarial networks (GANs) are generative model frameworks and excel at image generation. The GAN includes two parts: the generator and discriminator, as shown in Figure 1. The generator’s objective in GAN is to generate a synthetic image or video into a desired style or domain. The discriminator evaluates the authenticity of the input image and distinguishes whether it is a real picture or a synthetic image, training on an adequate high-quality paired dataset. Moreover, the loss as feedback from the discriminator will send back to the generator. The generator will update the parameter in the corresponding neural networks. Repeating the process further increases the quality of the generated synthetic image. Current popular networks for the image production direction are deep belief network encoder (DBN), Variational auto-encoder (VAE), and GAN (O’Connor et al. 2013; Razavi et al. 2019). Concerning the quality of synthetic images, GAN far exceeds other networks. Several frameworks have been conducted to improve the performance of GAN networks. Pix2pix is an extraordinary framework of cGANs that replaces noise input with paired datasets, showing remarkable results for an image translation task (Isola et al. 2017). Based on the applied condition, the pix2pix generator converts the input image into the target
383
3.1 Dataset
Figure 1. Process of GANs training.
image, and the discriminator ensures that the condition is met by considering pixel-wise loss. Perceptual Adversarial Networks (PAN) introduced a new approach for image-to-image translation that focuses on minimizing perceptual loss rather than pixel loss between paired pictures (Wang et al. 2018). However, Pix2pix and PAN are all frameworks for the paired dataset, which are difficult to collect and obtain in some domains. Thus, some researchers focus on unpaired datasets to propose several frameworks such as Discover cross-domain relations with GAN (DiscoGAN), Cycle-GAN, Dual GAN (Dual-GAN), and Star GAN (Star-GAN) with unpaired datasets. They were proposed for image-to-image translation (Choi et al. 2018; Kim et al. 2017; Yi et al. 2017; Zhu et al. 2017). Combing VAE and GAN architectures, an unsupervised Image to Image Translation (UNIT) model was proposed for image style transfer (Liu et al. 2017). Then, Multi-modal UNIT (MUNIT) was proposed to expand UNIT, generating different results based on a given domain image (Huang et al. 2018). GAN is far superior to other methods in image generation problems, and the improved GAN is sufficient to generate realistic images. The results of generated images show excellent quality and detailed diversity and the more robust applicability of the proposed framework. In the architectural-constructionengineering industry, GAN has exhibited outstanding performance in architectural layout design and hospital operating department layouts (Rahbar et al. 2022; Zhao et al. 2021). Nevertheless, there is a lack of research on using GAN or deep learning algorithms and developing new approaches to learning the nonlinear relationship of the location of a tower crane from images. Therefore, this study is expected to develop an image-based tower crane layout approach based on GAN to close the research gap.
3 TOWER CRANE LAYOUT GENERATION PROBLEM This study is supposed to generate a diverse set of proper tower crane layouts based on design construction layout images with efficiency and safety objectives. The following sections explain dataset, evaluation metrics, and the assumptions.
A paired dataset is selected to train the generator and discriminator to make the generator better learn the tower crane layout features in high-rise buildings. Training on a high-quality dataset is crucial for generating high-quality synthetic images. Therefore, this study obtains the dataset based on the construction experience of TCLP for high-rise buildings. As for the high-rise building, attached tower cranes are the most commonly used vertical transportation machinery. It can be elevated by adding the standard section under the control cabin. When exceeding the independent height of the designated building structure, the crane must be attached to the building for lifting. The dataset is generated for such attached tower cranes that need to be located on the attached route outside the building within a certain safety distance. The optimal tower crane layout is obtained by calculating the optimal layout planning considering construction efficiency and safety concerns. The safety concern includes the soil conditions where cannot locate tower crane. The construction efficiency is reflected by the transportation time affected by the distance between the supply area, demand area, and the location of the tower crane. A weighting function is used to generate a Pareto optimal tower crane layout to better obtain the appropriate location of the tower crane, as shown in Eq (1). The value of α is decided based on workers’ experience, and actual construction needs to reflect the weight relationship between efficiency and safety concerns. The transportation time is calculated based on the hook movement, as Eq. (1)-Eq. (10) in Li et al. (2018). Eq (2) illustrates the influencing factor for the safety condition of the tower crane location, which is the distance between the location and the inaccessible area with weak soil conditions. For the safety consideration, the distance is better to be larger than a certain distance meter β. Ci = Ti + ∂ • Si Si =
Di , D i ≺ β 1, Di ≥ β
(1) (2)
where Ci refers to the total cost of the tower crane layout at location i; Ti refers to the transportation time; Si refers to the cost of the safety condition, and the value of parameter α reflects the weight relationship between transportation efficiency and safety concerns. In addition, the layout must be generated for a single tower crane, and the tower crane is forbidden to be located in unsuitable ground conditions. Based on the restrains of the safe condition and efficiency and the multiple objectives proposed, the optimization model ofTCLP can be formulated as the discrete optimization problem as below: Objective function:
384
min Ci • ai • bi
Subject to: I
ai = 1
(3)
i=1 I
– Only one tower crane is required for each construction project. – The tower crane can complete all lifting tasks along the attachment route at any location. Further research to address the limitation mentioned above will be carried out in the future.
(bi − 1) = 1
(4) 4 TC-GAN
i=1
ai , bi ∈ 0, 1
where ai and bi are all binary variables. ai equals 1 if the tower crane is located at i, and 0 otherwise. bi equals 1 if the soil condition of location i is suitable for the tower crane. Then, this paired dataset is cropped and resized into the desired size of the generator of TC-GAN to enable the training process of the generator and discriminator. 3.2 Evaluation metrics This section used proficient perceptual studies and computer vision-based evaluation to evaluate the performance. Five construction design images were selected for the test, and the corresponding tower crane layouts were generated using TC-GAN. The evaluation includes the quality and rationality of the synthetic images. The quality of the generated image is measured in a computer vision-based evaluation method called the intersection-over-union (IoU) score. The IoU score is commonly used to reflect the accuracy degree of the predicted image for the location information, referring to the ratio of the intersection with the real image in the generated image to the concatenation of the two areas. The IoU score can reflect the quality and accuracy of the generated image compared with the calculated optimal tower crane layout planning. In this study, the generation image includes the construction boundary, soil condition, the potential tower crane layouts, the building plan, supply areas, and the optimal tower crane layout planning. Proficient perceptual studies were applied to measure the rationality of synthetic images. The generated and calculated optimal construction tower crane layout were presented to subjects at the same time for them to assign one of three ratings: real (+1), hard to tell (+1), and fake (−1), respectively.
Instead of solving the TCLP optimization problems based on a mathematical model and corresponding optimization solvers, the proposed approach, TCGAN, is trained on a paired dataset consisting of original images and layouts suggested through solving the optimization problems. Pix2pix is a typical framework of cGAN based on the paired dataset which has been successfully applied in various domains such as wheel hub manufacturing and structural design of shear walls (Liao et al. 2021; Shen et al. 2019). Thus, TC-GAN applied the framework of pix2pix to realize high performance. The following sections detail the network structure of the TC-GAN. 4.1 TC-GAN generator The generator’s objective is to automatically generate a proper tower crane layout based on the input image. The input image of TC-GAN is the label images with building data, construction boundary, attachment route, supply area, and soil condition. The TC-GAN generator generates a proper tower crane layout based on the label images. The generated tower crane layout needs to be within the attachment route to reveal the design rationality. Moreover, keep a safe distance from a weak soil condition area that cannot locate the tower crane (i.e., their distance is better, not less than 10m). Compared with the dataset in Isola et al. (2017), the dataset of crane layout planning is simpler with less color and irregular shape. Therefore, the U-Net was applied in the neural network architecture of the generator. During the down-sampling phase, some simple characters will be captured in the shallow layers of the network, including color and shape. As the data flows into the deeper neural network layers, some capture obtained in the shallow layers can be lost, leading to a bad performance of the generator. To avoid information loss, skip connections are applied, linking the shallow layers with their corresponding deeper layers.
3.3 Assumptions
4.2 TC-GAN discriminator
In contrast to the actual process of tower crane layout planning, a few restrictive assumptions to simplify the problem are set:
The function of the discriminator is to determine whether the input is real TCLP images or synthetic ones (the generator’s output). The training procedure aims to adjust the generator’s parameters to simulate data distribution and realistic TCLP images while the discriminator evolves. In the discriminator of TCGAN, PatchGAN has been applied to segment the input image in an N × N patch, which is smaller than the full size of the input image. In TC-GAN, the N was selected as 70.
– The supply area, demand area, and tower crane layout are represented as the center points of the projection in the horizontal direction. – The value of the location of weak soil conditions, construction buildings, and construction boundaries are known in the design phase.
385
5 IMPLEMENTATION AND RESULTS The proposed TC-GAN was implemented in PyTorch and utilized a workstation with NVIDIA GeForce RTX 3090 and Inter(R) UHD Graphics 750. Our model adopts an ADAM optimizer (b1 = 0.5, b2 = 0.99) and is trained for 200 epochs. 5.1 Dataset generation Figure 2 is one pair of the dataset. The central part of the image is the building plan with the location of columns, walls, stairs, and elevators. The blue rectangle represents the construction boundary. The black square with a cross refers to the potential location of the tower crane on the attachment route. The blue circle with a cross refers to the center location of the supply area. The solid gray rectangle represents the areas with weak soil conditions. As shown in Figure 2(b), the solid red rectangle refers to the optimal layout of the tower crane considering construction efficiency and safety. The construction images in the dataset were cropped into 256×256 to avoid tensor mismatch problems during training TC-GAN.
Figure 2.
Figure 3. Synthetic images through TC-GAN under different learning rates (The resolution of the images is 256×256.).
One of the paired datasets for TC-GAN.
5.2 Results Image quality and rationality metrics were identified to evaluate the performance of the proposed TC-GAN. To obtain optimal parameters, TC-GAN was trained by a set of learning rates, including 0.0001, 0.0002, 0.0005, 0.001, 0.002, 0.005 and 0.01. The implementation results ofTC-GAN under different learning rates were shown in Figure 3, based on a selected case. As shown in Figure 3, the performance of the TC-GAN generator is not tolerable for the low equality on the generated image when the learning rate equals 0.0005 and 0.01. Then, the performance under the rest of the learning rates was evaluated through the proposed image quality and rationality metrics. 5.3 Results evaluation Five typical construction layout cases were selected to evaluate the performance of the generated image from TC-GAN. For the image quality, the results of IoU for TC-GAN synthetic images are shown in Figure 4. The average IoUs, in descending order, were 90.56%, 89.74%, 82.04%,76.49% and 66.42% under the learning rate 0.0001, 0.0002, 0.001, 0.002 and 0.005, respectively. When the learning rate equaled 0.0001,
the average IoU reached the highest value, reflecting the high accuracy of the construction information generation on the images.
Figure 4. IoU of the synthetic images under different learning rates.
As for rationality, five practitioners with construction engineering experience were recruited to judge the reasonableness of the synthetic images. The results were collected and analyzed. The results are shown in Table 1. In contrast with other results, when the learning rate equaled 0.0001, the average and standard
386
deviation of the user rating reached the best outcomes. In addition, this also revealed that the TC-GAN could generate an appropriate tower crane layout. Table 1. Rating for the rationality of the generated tower crane layout. Learning rate
Average
Standard deviation
0.0001 0.0002 0.001 0.002 0.005
3.20 2.80 2.80 −1.80 −0.60
1.17 1.60 1.94 1.72 1.62
6 CONCLUSION Tower cranes are vital construction machinery that governs the construction schedule for high-rise buildings. This study proposes an image-based TC-GAN to automatically generate a proper Tower Crane Layout Planning (TCLP) considering the transportation time and soil condition to improve the design efficiency and avoid safety concerns. The results were evaluated by IoU and average user rating, respectively. The TC-GAN performance under different learning rates was evaluated regarding image quality and rationality, and the one that equaled 0.0001 was selected. It revealed that the proposed TC-GAN was workable and appropriable for assisting the managers withTCLP problems. A limitation of this study is that the TC-GAN is only applied for attached cranes which can be extended to other types of cranes by expanding the dataset. Furthermore, the approach for finding suitable hyperparameter configurations, such as learning rate, can only be acquired manually, which is time-consuming with low adaptability. REFERENCES Abdelmegid, M.A., Shawki, K.M., & Abdel-Khalek, H. 2015. ‘GA Optimization Model for Solving Tower Crane Location Problem in Construction Sites’, Alexandria Engineering Journal, 54(3), 519-526. Chen, J., Chi, H.L., Du, Q., & Wu, P. 2022. ‘Investigation of Operational Concerns of Construction Crane Operators: An Approach Integrating Factor Clustering and Prioritization’, Journal of Management in Engineering, 38(4), 04022020. Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., & Choo, J. 2018. ‘Stargan: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation’, In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 8789–8797). Ding, C. 2013. ‘Building Height Restrictions, Land Development and Economic Costs’, Land Use Policy, 30(1), 485–495. Furusaka, S., & Gray, C. 1984. ‘A Model for the Selection of the Optimum Crane for Construction sites’, Construction Management and Economics, 2(2), 157–176.
387
Huang, C., Wong, C.K., & Tam, C.M. 2011. ‘Optimization of Tower Crane and Material Supply Locations in a Highrise Building Site by Mixed-integer Linear Programming’, Automation in Construction, 20(5), 571–580. Huang, W., & Zheng, H. 2018. ‘Architectural Drawings Recognition and Generation Through Machine Learning’, Proceedings of the 38th Annual Conference of the Association for Computer Aided Design in Architecture (ACADIA). Huang, X., Liu, M.Y., Belongie, S., & Kautz, J. 2018. ‘Multimodal Unsupervised Image-to-image Translation’, In Proceedings of the European conference on computer vision (ECCV) (pp. 172–189). Isola, P., Zhu, J.Y., Zhou, T., & Efros, A.A. 2017. ‘Imageto-image Translation With Conditional Adversarial Networks’, In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1125–1134). Kaveh, A., & Vazirinia, Y. 2018. ‘Optimization of Tower Crane Location and Material Quantity Between Supply and Demand Points: A comparative study’, Periodica Polytechnica Civil Engineering, 62(3), 732–745. Kaveh, A., & Vazirinia, Y. 2020. ‘An Upgraded Sine Cosine Algorithm for Tower Crane Selection And Layout Problem’, Periodica Polytechnica Civil Engineering, 64(2), 325–343. Kim, T., Cha, M., Kim, H., Lee, J.K., & Kim, J. 2017. ‘Learning to Discover Cross-domain Relations with Generative Adversarial Networks’, In International conference on machine learning (pp. 1857–1865). PMLR. LeCun, Y., Bengio, Y., & Hinton, G. 2015. ‘Deep Learning’ Nature, 521(7553), 436–444. Liao, W., Lu, X., Huang, Y., Zheng, Z., & Lin, Y. 2021. ‘Automated Structural Design of Shear Wall Residential Buildings Using Generative Adversarial Networks’, Automation in Construction, 132, 103931. Li, R., Fu, Y., Liu, G., Mao, C., & Xu, P. 2018 An Algorithm for Optimizing the Location of Attached Tower Crane and Material Supply Point With BIM. In ISARC. Proceedings of the International Symposium on Automation and Robotics in Construction (Vol. 35, pp. 1–8). IAARC Publications. Lien, L.C., & Cheng, M.Y. 2014 ‘Particle Bee Algorithm for Tower Crane Layout with Material Quantity Supply and Demand Optimization’, Automation in Construction, 45, 25–32. Lin, J., Fu, Y., Li, R., & Lai, W. 2020 ‘An Algorithm for Optimizing the Location and Type Selection of Attached Tower Cranes Based on Value Engineering’ In ICCREM 2020: Intelligent Construction and Sustainable Buildings (pp. 106–117). Reston, VA: American Society of Civil Engineers. Liu, M.Y., Breuel, T., & Kautz, J. 2017 ‘Unsupervised Image-to-image Translation Networks’ Advances in neural information processing systems, 30. Lu, Y., & Zhu, Y. 2021 ‘Integrating Hoisting Efficiency Into Construction Site Layout Plan Model for Prefabricated Construction’, Journal of Construction Engineering and Management, 147(10), 04021130. Marzouk, M., Abubakr, A. 2016 ‘Decision Support for Tower Crane Selection with Building Information Models and Genetic Algorithms’, Automation in Construction, 61, pp. 1–15, O’Connor, P., Neil, D., Liu, S.C., Delbruck, T., & Pfeiffer, M. 2013 ‘Real-time Classification and Sensor Fusion with a Spiking Deep Belief Network’, Frontiers in neuroscience, 7, 178. Rahbar, M., Mahdavinejad, M., Markazi, A.H., & Bemanian, M. 2022 ‘Architectural Layout Design Through
Deep Learning and Agent-based Modeling: A Hybrid Approach’, Journal of Building Engineering, 47, 103822. Razavi, A., van den Oord, A., & Vinyals, O. 2019 ‘Generating Diverse High-resolution Images with VQ-VAE’, In 2019 The International Conference on Learning Representations (ICLR) Rodriguez-Ramos, W.E., & Francis, R.L. 1983 ‘Single Crane Location Optimization’, Journal of Construction Engineering and Management, 109(4), 387–397. Shen, Z., Dang, H., Sun, M., & Zhou, X. 2019 ‘August. Application of Generating Adversarial Networks in Highlight Removal of Wheel Hub Surface’, In 2019 International Conference on Advanced Mechatronic Systems (ICAMechS) (pp. 12–15). IEEE. Tam, C.M., & Tong, T.K. 2003‘GA-ANN Model for Optimizing the Locations of Tower Crane and Supply Points for High-rise Public Housing Construction’, Construction Management and Economics, 21(3), 257–266. Tam, C.M., Tong, T.K., & Chan, W.K. 2001‘Genetic Algorithm for Optimizing Supply Locations Around Tower Crane’, Journal of Construction Engineering and Management, 127(4), 315–321. Wang, C., Xu, C., Wang, C., & Tao, D. 2018 ‘Perceptual Adversarial Networks for Image-to-image Transformation’ IEEE Transactions on Image Processing, 27(8), 4066–4079. Wang, J., Zhang, X., Shou, W., Wang, X., Xu, B., Kim, M.J., & Wu, P. 2015 ‘A BIM-based Approach for Automated Tower Crane Layout Planning’, Automation in Construction, 59, 168–178. Wu, K., de Soto, B.G., Adey, B.T., & Zhang, F. 2020 ‘BIMbased Estimation of Vertical Transportation Demands
During the Construction of High-rise Buildings’, Automation in Construction, 110, 102985. Yeoh, J.K., & Chua, D.K. 2017 ‘Optimizing Crane Selection and Location for Multistage Construction Using a Fourdimensional Set Cover Approach’, Journal of Construction Engineering and Management, 143(8), 04017029. Yi, Z., Zhang, H., Tan, P., & Gong, M. 2017 ‘Dualgan: Unsupervised Dual Learning for Image-to-image Translation’, In Proceedings of the IEEE international conference on computer vision (pp. 2849–2857). Yu, J., Zhou, K., & Yang, S. 2019 ‘Land Use Efficiency and Influencing Factors of Urban Agglomerations in China’, Land Use Policy, 88, 104143. Zhang, P., Harris, F.C., Olomolaiye, P.O., & Holt, G.D. 1999 ‘Location Optimization for a Group of Tower Cranes’, Journal of Construction Engineering and Management, 125(2), 115–122. Zhao, C., Yang, J., Xiong, W., & Li, J. 2021 ‘Two Generative Design Methods of Hospital Operating Department Layouts Based on Healthcare Systematic Layout Planning and Generative Adversarial Network’, Journal of Shanghai Jiaotong University (Science), 26(1), 103–115. Zhou, C., Huang, B., & Fränti, P. 2021 ‘A Review of Motion Planning Algorithms for Intelligent Robots’, Journal of Intelligent Manufacturing, 1–38. Zhu, J.Y., Park, T., Isola, P., & Efros, A.A. 2017 ‘Unpaired Image-to-image Translation Using Cycleconsistent Adversarial Networks’, In Proceedings of the IEEE international conference on computer vision (pp. 2223–2232).
388
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Consideration of detailing in the graph-based retrieval of design variants D. Napps & M. König Chair of Computing in Engineering, Department of Civil and Environmental Engineering, Ruhr-Universität Bochum, Bochum, Germany
ABSTRACT: During early design processes, interdisciplinary experts frequently exchange building information at different Levels of Development to generate designs fulfilling multiple requirements, by ideally using building information models. This paper presents an opportunity for architects to gather inspirations for upcoming projects based on previous designs and knowledge in relation to the detailing of building elements. For this purpose, two approaches from the research group FOR 2363 are combined. Initially, this refers to the variant management, which enables to find and compare multiple similar design options to an existing building structure, through a subsequent retrieval process. The second research focuses on the detailing of components and potential uncertainties associated with geometric and semantic information. Regulations for unifying these approaches are formulated and the similarity calculation is specified. This leads to an optimization of the retrieval process including the consideration of the detailing of components. The usability is afterwards demonstrated.
1 INTRODUCTION 1.1 Research gap and problem statement Early design stages of buildings are of great importance for the successful realization of a project, because at this stage the decisions are made that combine the complex ambitions and requirements of a building with a compatible design. Although the impact of early design decisions on buildings and their life cycles are significant, the early phases of the project and decision-making applications still receive limited attention. (Østergård et al. 2016) In this context, the exchange of information on the building’s respective requirements and development stages, between the stakeholders at different detailing levels, becomes particularly relevant and challenging. (Abualdenien & Borrmann 2018) The specification of building elements with a different detailing in the early design phases has the potential to capture multiple requirements and, in conjunction with multiple decision-making processes and design variants of a building, can have a major impact on the final building performance and total cost. A supporting decisionmaking tool for architects has already been introduced, but the variant management, where it is based on, is actually not linked to a specific Level of Development (LOD). (Napps et al. 2021) Building Information Modeling (BIM) continues to increase its importance for interdisciplinary communication, cooperation and collaboration between all project partners. For this reason, BIM has been DOI 10.1201/9781003354222-50
increasingly adopted by the Architecture, Engineering and Construction (AEC) industry in recent years. (Young et al. 2009) The European Union proposes to combine BIM with other digitization technologies to achieve the full potential of digital transformation. (Locatelli et al. 2021) Knowledge graphs are increasingly utilized in the AEC sector to capture and analyse complex relationships. In the context of BIM, these are usable for the analysis of structural relationships between entities. (Ji et al. 2022) By capturing all the different stakeholder requirements for a new building, digital buildings are constantly being adapted and modified to evaluate the advantages and disadvantages of different designs of a building and to comply with legal requirements. (MITRE 2014) The Case- Based Reasoning (CBR) approach has already been used to assist architects in finding similar 2d floor plans in the early design phases. (Sabi et al. 2017) However, a direct adaptation of existing buildings or building sections into a new BIM model is difficult and can only result from an advanced similarity comparison of existing projects for example with the help of a database. (Napps et al. 2021) At the beginning of the design process, the uncertainty in how the design may evolve is high due to incomplete or unknown information (Knotten et al. 2015), so that the upcoming project and the database can have different LODs, because there is no specified standard available for the early design phases. This negatively affects the whole retrieval process and thereby complicates the design decisions for the architects. Resulting in a challenge
389
for the variant management, since available options for a variant with a modified development status can lead to errors. 1.2 Previous research and proposed solution The paper is developed in the context of the German Research Foundation (DFG) project FOR 2363 and combines two fundamental research areas in the early BIM domain within this research group. The first one is the variant management in combination with a similarity based retrieval process of building information models. (Napps et al. 2021) This enables individual entities in a BIM model to be classified into three categories and to identify similar objects or layouts stored in a database. Therefore a graphical representation of the different variants was proposed, based on the internal structure of the Industry Foundation Classes (IFC) according to the version IFC4. (Mattern & König 2018) Using a developed plug-in, Napps et al. have currently established a possibility to assign building elements directly in a BIM environment to one of the three variants.In case of a customization of these components, the possible option is thus automatically stored in the database. (Napps et al. 2022) The second research topic is the LOD approach. Thereby the sequential refinement of the geometric and semantic information is described. (Abualdenien & Borrmann 2018) Parallel to the first approach, Abualdenien et al. used a graph structure for the detailing of building elements and developed a multi-LOD meta-model, which explicitly specifies the LOD requirements of each building element type, taking into account the possible uncertainties. (Abualdenien & Borrmann 2019) In various scientific disciplines, the use of graph frameworks is beneficial for analysing and extracting information. They are used in the construction industry in the context of BIM because of their ability to describe complex digital models and internal relationships. (Abualdenien & Borrmann 2021) This research aims to identify a graph-based approach for architects to search for different components in different LODs, which can be transferred from similar designs to the current project. This leads into flexible adjustments and detailed inspiration, especially regarding the use of multiple variants at different stages of detailing. For a more specific variant management, which can easily adapt throughout all development stages of a new project, it is necessary to provide the architects with different input parameters, as well as the required detailing information. Concerning to these conditions architects are able to retrieve a similar existing project with the option to exchange multiple variants in relation with its detailing. A change of the LOD as a filtering option allows multiple modifications and adjustments and is important to compare analytical parameters, for example for life cycle analysis. Consequently, it is possible to use the variant management independently of a projectrelated Level of Development, as it can be adapted to the requirements. Moreover, several detailing levels of
variants are stored for the same digital model, resulting in a greater variety of possible options being included in the database for retrieval for future projects. Additionally, the research provides an opportunity to use graph-based methods in the field of civil engineering in context with the IFC data structure, exported from a building information model. A more detailed outline of the paper is as follows: Section two discusses the background and the related work, including the combination of the variant management, the retrieval process and the LOD, as well as the established graphical concept. Section three provides the graph-based framework as well as an enhanced more precise similarity calculation for the retrieval process. Section four demonstrates the applicability of the developed approach through a selected case study. Finally, section five summarizes the results, highlights the contributions for the industry and provides an outlook for future research.
2 BACKGROUND Interdisciplinary networking of spatial planners, architects, civil engineers and other project participants requires effective communication and exchange methods through the planning process of new buildings. Building Information Modeling and the Industry Foundation Classes offer an opportunity to fulfil the requirements in an international and standardized format. (Schapke et al. 2018) The international IFC standard (ISO 16739) provides the fundamentals for this data exchange. Developed and continuously updated by buildingSMART, IFC is an object-based file format that organizes and interconnects all the entities of a building in an object-based inheritance hierarchy. (ISO 2018) This internal architecture is important in terms of the graph-based framework of the variant management and the retrieval process with the detailing options of variants. 2.1 Variant management The variant management is a synonym for the option management and was introduced to classify different elements in a BIM model. This categorization is based on three types, namely structure, function and product variants. Using this assignments, it is subsequently possible to create possible design options for the assigned features. (Mattern & König 2017) Furthermore, it is used to identify similar element options for a building, which results from a graph-based matching comparison of similar previous projects. (Napps et al. 2021) Any associated impact of a relevant change to the floor plan is covered by the structure variant. It is hardly possible to divide structure and function variants, because they are hardly related. Additionally, structural options are influencing the subsequent planning process. Function variants are related to different building objects that belong to systems that
390
have the same function. One of the characteristics is the material, e.g. insulation, as well as the loadbearing capacity. Replacing an object by a similar object or changing the attributes of a single object represents the third variant type, the product variant. They have no influence on other building elements. Mostly, numerous objects of the same kind are affected by the assignment of a product option. (Mattern & König 2018) This enables the architect to select a particular option from a pool of possible components, allowing him to obtain and explore different inspirations for the design of a building concept. Based on this, a design can be customized by alternative available options. 2.2 The retrieval process CBR processes have already been used to leverage solutions that have already been identified to solve an existing problem. In this context a retrieval process has been introduced recently which is based on generated graphs concerning the IFC structure of BIM models. (Napps et al. 2021) The process represents the first part of the CBR cycle, which in the end enables to use an already solved problem to find a solution for an existing problem. This process is used to retrieve building graphs according to the query parameters defined by the architect. Therefore it is divided into four main sections: (1) During the retrieval process, the architects specify parametric requirements and their respective weightings for an upcoming project using a developed interface (e.g. net area, number of floors and rooms utilization, construction method, etc.). (2) These requirements are queried using an inexact pattern matching and (3) similarity calculation in a graph database to obtain the optimal result. (4) Finally, further adjustments can be made by subsequently querying the database for available alternative options for a component that is assigned to a variant type. The inexact pattern matching is carried out using the subgraph matching, where the described subgraph (G’) is searched in a different graph (G) obtained from a database. The higher the weighting of elements in the first part of the process, the greater their importance for the design solution. Computational similarity calculation of different building designs and the graph pattern matching originate from previous research experience in this subject. (Napps et al. 2021) The best solution to an emerging problem described by an architect is the most similar design draft of an already existing project. The nearest neighbour algorithm has gained acceptance for the case-based reasoning and determines the similarity of two cases for any number of attribute values. A weighting and normalization of the attributes is integrated, so that the importance of the features of a digital model (e.g. storey size, nuber of spaces) can be determined. Both qualitative and quantitative similarity determination is possible. The similarity results in a value
between 0 and 1, where 0 indicates no similarity and 1 indicates total similarity. (Watson 1995) SH (p, v) =
n
i=1 ωi Si (pi , vi ) n
i=1 ωi
(1)
All attributes that do not exist in a matching case are assigned the value 0 for the similarity calculation. 2.3 Level of development With the Level of Development specification, the development of a building model is managed during the design and construction process. This improves the quality of decision-making through the design process. (BIMForum 2021) LOD describes the BIM elements at a specific stadium, including specifications and completeness that represent their information quality. Beginning at a coarse Level of Development, the grading scale refines in the higher development levels. This increases the quality and information content of the features in the model components in increasing the scale. However, it does not determine which development level is to be reached at which stage in a project, but rather allows the user to determine the course of the model (see Figure 1).
Figure 1. Levels of development.
The American Institute for Architects (AIA) introduced a LOD schema which is commonly used in the construction industry. These specification addresses LOD 100, LOD 200, LOD 300, LOD 350 and LOD 400. While LOD 500 is included in the AIA’s LOD definition, it is no longer necessary to define it, as it is not an indicator for the process and requires a target-performance comparison. The different LODs have minimum requirements and in continuing levels, the previous level’s information is always included as well. (Abualdenien & Borrmann 2018) Two important requirements are combined by the abbreviation LOD, the geometry level (LOG) and the information level (LOI). LOG provides information about quantifiable components, whereas LOI includes non-geometric qualitative information. (Karlapudi et al. 2021) The terminology of the Level of Development concept varies in a number of countries, so it is often used as the Level of Detail (LoD) with a different numbering system. In this paper, the LoD is considered the input of an element, while the LOD is the reliable output. This is equal to the approaches of the BIMForum. (BIMForum 2021)
391
2.4 Graph representation For the graphical representation of the IFC data, the (labelled) property graph (LPG) approach is utilized. This enables object-oriented programming and overcomes the limitations of the triple centred approach (RDF), because of their reduced structure. Instances of relationships of the same type can also be identified so that they can be qualified or attributed. (Alocci et al. 2015) The LPG approach provides efficient storage and fast graph traversal, which is essential for many graph analytic applications. (Purohit et al. 2020) Consequently, this concept is capable of capturing the diversity of building information and relationships in order to be able to query them on a subsequent phase. The BIM models are exported as an IFC file and can be stored in a graph database. For this operation, the java implemented open source database of Neo4j is used.
Figure 2. BIM-based graph representation.
3.2 Integration of the detailing 3 METHODOLOGY In the next sections, the methodological framework for linking the two approaches is discussed. It focuses on the graph-based enhanced determination of the similarity calculation for the retrieval process as well as the integration of the detailing into the whole process. 3.1 Definition of property graphs The labelled property graph is a set of vertices (V) that are connected by edges (E), which define an ordered structure of the graph by start and end points. Specific properties are contained in each of the vertices as well as their edges (see Figure 2). Labels (L) provide a grouping of the vertices and assign their role within the dataset. (Robinson et al. 2015) The previous approach of finding similar structures in a graph using subgraph matching is still valid. Mathematically, the description of generated subgraphs (attributed or labelled) can be thus defined as G = (V , E, L) with Lv = (lv1 , lv2 , ...) and Le = (le1 , le2 , ...). Both, vertices and edges can be determined by weightings. (Kriege & Mutzel 2012) As the details of the entities of a variant type affect the qualitative information of the vertices, they initially only effect the stored information, i.e. the properties. According to this definition, the graph structure does not differ for varying the LOI of an entity. However, changing the LOG will affect the topology of the graph. This is because altering the LOG of a building component may affect the graph structure if it changes the geometric information and thus the connections of the vertices. These facts will be illustrated and discussed later on in a supporting visual representation. This graph approach applies to the entire graph-based retrieval and can also be applied to the intended combination of the variant management and the detailing of building components, since a variety of information can be stored in property graphs.
In order to capture the impact of the changes regarding the LODs among the three variant types, it is first necessary to define the different possibilities of LODs for the variant types. Variation of variants in different LODs leads to an improved efficiency of the retrieval process, as an increased choice of design options at different early planning stages can be provided to inspire the architects (see Figure 3). This combination of the variant management with LOD is restricted by certain rules that are a consequence of the individual definitions of the approaches.
Figure 3. Detailing of an external wall.
According to the initial definition by Mattern and König, similar options to a variant type are an alternative design at the same Level of Development. (Mattern & König 2018) Consequently, a new option results for each change in the LOD. Within the LOD definition, the placement and or the geometric representation of an element may vary if the level is changed. Since these components are mostly important for design decisions, they are the priority for the option management. Individual solutions for the three types are provided for these changes, including the chosen limitation of the LOD (see Table 1). Structure variants can be recognisable as an object even in the earliest development stages. Accordingly, this variant type can be considered between LOD 100 to LOD 400. Differentiated elements are included in the function variants. Columns are only geometrically approximated starting at LOD 200, likewise the partitioning of rooms. The load-bearing capacity of interior walls is specified precisely from LOD 300 and can
392
Table 1.
Possible LODs of the variant types.
Variant
Exemplary entities
LOD range
Structure Variant Function Variant Produt Variant
IfcCurtainWall IfcColumn, IfcWall IfcDoor, IfcWindow
LOD 100 - 400 LOD 200 - 400 LOD 300 - 400
only be approximated earlier. The limitation of the function variants is consequently that the acquisition is possible from LOD 200 and is increasable up to LOD 400. Specific regulations for the product variants apply, as the geometric dimensions of the assigned components can not be changed according to its definition. Thus, the product variant is limited by LOD 300 to LOD 400. However, this allows a change of properties and materiality. To minimize the influence of a modified detailing on the graph structure in order to avoid influencing the results of the subgraph matching, further rules for the integration of the detailing are defined. This is most important for the walls of a building. Although a wall may vary in its geometric position and thickness, its relationship with the rooms should remain the same, namely as a complete space boundary. This ensures that walls are connected to each other and that there is no unwanted gap in a room, independently of the LOD. 3.3 Enhanced retrieval process and similarity measurement The elaborated retrieval process is subsequently adapted in terms of accuracy. This establishes more opportunities for the architect’s input and output, which can be adapted to the intended Level of Development, and a more precise retrieval of similar designs. In the first step, the user is thus supported with more precise results on the similarity matching of his parameters with a floor plan or building to be found. Second the user can specify an explicit LOD for a variant type and search for an option in the same LOD. This is realized by implementing a filter option (see section 4.3). In terms of improving the precision, for the whole process an additional similarity calculation is integrated. Therefore, the calculation of the footprint similarity by Tversky is adopted to provide more accurate results. The extension of similarity enables a more precise capturing of attributes that are not available in both cases. In this process, structural similarities are identified and calculated. Based on this similarity concept, existing elements in the hierarchical structures of two feature graphs can be compared in various ways. (Tversky 1977) Tversky’s footprint similarity calculation is used complementary to obtain better results in combination with the generalized similarity. ST (p, v) =
α |f (p) ∩ f (v)| α |f (p) ∩ f (v)| + β |f (p) ∪ f (v)| + γ |f (v) ∪ f (p)|
(2)
All relevant features (f) of the cases p and v can be captured and α, β, γ ∈ R are constants which allow
for different treatment of the various components. For common features of two cases α = 0 is minimal and α = 1 is maximal importance. For the calculation of the Jaccard or Tanimoto similarity the following applies: α = β = γ = 1. The Dice similarity coefficient differs and sets α = 1 and β = γ = 0.5. (Rahnama & Hüllermeier 2020) By adding α to the formula, the importance of the same attributes can be determined by the architect. Since both the Hemming algorithm and the Tversky similarity calculation are normalized and weighted, they can be summed with intended significances. The respective weighting of the similarity values is to be determined to ensure that ωi + ωj = 1. ωi SH + ωj ST = Stotal
(3)
The importances of the approaches is initially set to equal values. However, the operators can individually determine which approach they want to include with higher weightings in the total similarity calculation.
4 DEMONSTRATION In the following, the graph-based framework for the retrieval and detailing of design variants is demonstrated using an exemplary search query. Thereby, the effects on the variant types as well as on the similarity calculation regarding the process are presented and visualized in depending on the detailing. 4.1 User operation Architects benefit in a first instance from the qualitative improvement of the retrieval process of a building, as the results of the extended similarity calculation capture both the weighted sum of the individual features as well as the structure and thus the deviations of the IfcElements. Regarding to the new requirements for the retrieval process depending on the detailing, users are supported with alternatives for identifying similar features. There is currently the possibility for the user to weight the requirements that are important for the emerging project. This implementation is based on the similarity calculation according to Hemming. The extension by Tversky’s similarity calculation uses a separate weighting, which means that no modifications to the previous input interface of the architect are required. In case the user retrieves a similar graph in the database representing the solution of the problem, there are two possibilities to modify a variant in a LOD according to the architect’s specific or less specific requirements.There are two possibilities to find an option in a different LOD: (1) Manual adjustment of the Level of Development regarding the elements in the BIM model or IFC file by the architect. (2) Subsequent search for alternative variant options with a correspondingly different LOD from the database.
393
Following on from the second aspect the detailing requires another interface wherein the operator can define the level for the respective variant type and the associated entities in the BIM model or IFC file. Once the architect has retrieved a building design that matches the problem definition, but is not satisfied with the development level of individual variants, an LOD can be specified for all defined entities that the user would like to receive as a possible option (see Figure 4). All variants in an existing BIM model can thus be set to the same Level of Development, if this is desired by the architect, in order to specify the planning and to be able to carry out subsequent evaluations.
option from LOD 200 onwards so that the design of the building is not significantly changed and generic placeholders are assigned to the elements. These levels of the structure variants operate in the same way as the function variants, which are explained in the following retrieved floor plan, which was most similar to the architects defined parameters.
Figure 4. User interface.
Already in the first determination process, the detailing of elements is indirectly applied, since the detailing can be influenced both by the weighting of the entities and by the prior limitation of the variant types. Reducing the importance of windows in the retrieval process of an overall similar floor plan leads to more flexibility in the quantity, shape, size and position for finding a similar alternative. Once a floor plan layout is found, the customized option management allows the windows marked as product variants to be exchanged, but only within the respective restrictions of the LOD (see Section 3.2). Since the positioning is not changeable, that otherwise would affect the structure, thus only semantics of the variant can be changed, e.g. glass thickness, material, opening properties. The proposed operation allows the architect to filter this alternative option to a variant in a modified detail. 4.2 Effects on variant types in changing the LOD Varying the detailing of a particular variant can affect the floor plan as well as the exported graph structure. The following section identifies the effects that a change in the detailing of variant types has for the retrieval process. Four different options of an interior wall are created for the demonstration. Accordingly, this demonstration serves only as a minor sample and control of the procedure for the functional options. In the retrieval process, the input parameters of the architect are compared with four digital models stored in the database. Structure variants grant the maximum flexibility in detailing for the retrieval process. A typical structure variant has already been shown in figure three. By the fact that this variant type is capable by starting at LOD 100, options to a variant in a different level have a great impact on the overall design, storey size, building shape and many other aspects. Accordingly, it should be only possible to search for an alternative
Figure 5. Detailing and uncertainty of a function variant.
The change in the function variant in its detailing is illustrated (see Figure 5). Exemplified is a loadbearing interior wall that depending on the Level of Development may be non-specific in its position, thickness, length, and material, as well as other properties. However, due to the chosen integration of the LOD, the wall is always a closing and separating boundary of one or more rooms (office, hallway, breakroom). Thus, the changes in the geometric specifications of this function variant have an impact on the related spaces. As a result, the extent of these effects depends on the weightings of the rooms, e.g. the requested volume and the variant itself, i.e. the width of the wall. A significant weighting of the office volume limits the width of the functional variant or specifies the positioning and thus limits the uncertainty of the element (green marked range). Accordingly, the detailing of a possible option for the function in the process needs to be restricted to higher levels. Low weighting of the parameters allow an increased choice of variants, as this enables different Levels of Development. Thus, the weights of the input parameters affect the variability concerning the detailing of a similar functional option. Product variants are only identifiable from LOD 300 on, because of defined positions and dimensions. The detailing is mostly limited to the specification of
394
properties. These are not recognisable in the 2d view, but can be displayed in a 3d view or in the IFC file. Hence, the architect is limited in the ability to search for alternative solutions. The strongest effects on a change in detailing of variant options in a floor plan are most evident in the first LODs, where both geometric and semantic information are uncertain. Thus, the structure variants have the greatest impact in the case of a transformation. Even the effects on the geometric positioning of the illustrated function variant can have significant effects on the design by varying the level. The detailing of a product variant, on the other hand, has negligible impact on the design of a building, but does have an impact on analytical, non-geometric parameters specified by the properties of the components. If the architect searches for a similar option to a variant, the variant is at least in LOD 200, because the received floor plan already contains the element to be changed, therefore it corresponds at least to this level.Changing the detailing affects the floor plan, but are limited by the fact that a matching solution needs to have the same relationships and connections to other elements as the version to be replaced. 4.3 Effects on similarity and the graph in changing the LOD Replacing a variant with an option of the same type containing modified detailing has an impact on the similarity of these components and affects the overall similarity of the floor plan. In the following, these effects on the similarity and the graph structure are summarized. Calculating the similarity between an option to a variant in a different LOD reveals a correlation that affects all variant types. When searching for a similar option, the same similarity calculation can be made as for retrieving a similar design. However, this calculation is not appropriate for the change in detail. If a comparable entity in a higher stage of development is requested, the similarity with a possible option can be calculated using the proposed calculation (Stotal ). In this case, a possible option will have more properties stored in the vertices. On the other hand, if an optional building element is searched for by an architect based on a vague design idea in preferred a lower LOD than it is in the BIM model, the problem description has more information stored in the vertices and the option searched for is missing this attributes.The higher the difference between the searched LOD and the origin level, the more dissimilar they are, therefore at this point a filtering is suggested regardless of the similarity. The selected option does have an influence on the overall similarity value if an option to a feature is to be replaced in a higher, lower or same level. Since this process is a filtering process and an extension in terms of an adaptive adjustment of the design (section four of the retrieval process) the overall similarity is changed, but it is of no importance for the process, since the similarity has already resulted in the selection
of a floor plan (first section of the retrieval process). It is therefore possible to search for a similar option in an alternative LOD if it is specified in terms of its requirements (the desired LOD). This ensures that changes in detailing do not affect the retrieval and calculation of the overall similarity and that the processes can be considered separately. Adjusting the detailing of IfcEntities only affects the graph structure if the geometric position of an element is imprecise. Using the definition (see Table 1) and the visual representation (see Figure 5) a variant can be changed in its geometry, but the dependencies stay the same. This exemplary floor plan can be transformed as a graph consisting of 119 vertices and 196 edges. The relationships of the wall with the connected IfcElements and the included IfcDoor are highlighted. Since these relationships are regulated and can not be changed, the graph structure remains the same even if the detailing of the wall is modified.Thus, the geometric change is meaningfully limited and the graphical representation remains. These dependencies are formulated by the user at the beginning of the process. Many dependencies get clearer by considering the IFC-graph structure resulting from the floor plan. Effects of an alternative option on the graph only result if the architect specifies vague structure variants (LOD 100) and changes them to higher level. Since variants are already formulated as placeholders they can thus be set in relation to other entities. Another change in the graph occurs by replacing an entity with a different entity that performs the same function and belongs to the same variant type, e.g. replacing a load-bearing wall by columns.
5 CONCLUSION AND FUTURE WORK The opportunities of the variant management are very flexible, contain a high potential of extension possibilities and it simplifies the early planning phases of buildings. This study introduced a possibility for architects to use the modified retrieval process for a current project, depending on the detailing of the individual assigned building elements to provide decision assistance. This results in new possibilities for retrieving alternative similar variants and adopt the respective LOD for the upcoming building, which enables a more accurate process and multiple impressions for architects. Furthermore, the research approach contributes to the industry, because an optimal retrieval process results in time savings for architects and civil engineers can use the implementation of LODs to consider analytical parameters and select optional variants in relation to the client’s requirements. This is becoming increasingly important in terms of sustainability, living conditions, profitability, amortization and energy efficiency. Although the demonstrated approach provides many advantages, there are still limitations in the real-world application of complex cases. Considering complex buildings covering several storeys with a
395
high number of variant options and retrieving solutions from a large building graph database can be detrimental to speed and memory capacity in the retrieval process. There is also the problem of the availability of data from digital models for the case base. Continuing the process, the scale will be extended so that, in addition to the detailing of the option management, the entire building design can be queried in different LODs. Since this is dependent on the weighting of different features of the project, a possibility needs to be identified to set the detailing in dependence on the weighted input parameters of the architect.As an orientation, it can be assumed that with a low weighting of an entity the detailing may be allowed to vary more than in the case of a high weighting, because of an important particular design idea of the architect.In addition, the performance of the approach in terms of complexity needs to be investigated further to ensure that the processes operate efficiently in terms of speed and storage. REFERENCES Abualdenien, J. & Borrmann, A. (2018). Multi-lod Model for Describing Uncertainty and Checking Requirements in Different Design Stages. In J. Karlshoj & R. Scherer (Eds.), eWork and eBusiness in Architecture, Engineering and Construction, pp. 187–195. Boca Raton, FL: CRC Press. Abualdenien, J. & Borrmann, A. (2019). A Multi-lod Model for Visualizing Building Information Models’ vagueness. Computing in Civil Engineering, 248–255. Abualdenien, J. & Borrmann, A. (2021). Pbg: A Parametric Building Graph Capturing and Transferring Detailing Patterns of Building Models. International Conference of CIB W78 (38), 1–10. Alocci, D., Mariethoz, J., Horlacher, O., Bolleman, J. T., Campbell, M. P., & Lisacek, F. (2015). Property Graph vs rdf Triple Store: A Comparison on Glycan Substructure Search. PLOS ONE 10(12), e0144578. BIMForum (2021). Level of Development (lod) Specification: Part i, Guide & Commentary. ISO (2018). Industry Foundation Classes (ifc) for Data Sharing in the Construction and Facility Management Industries. Ji, S., Pan, S., Cambria, E., Marttinen, P., & Yu, P. S. (2022). A Survey on Knowledge Graphs: Representation, Acquisition, and Applications. IEEE Transactions on Neural Networks and Learning Systems 33(2), 494–514. Karlapudi, J., Valluru, P., & Menzel, K. (2021). Ontological Approach for Lod-sensitive Bim-data Management. Linked Data in Architecture and Construction Workshop (9), 103–114. Knotten, V., Svalestuen, F., Hansen, G. K., & Lædre, O. (2015). Design Management in the Building Process – A Review of Current Literature. Procedia Economics and Finance 21, 120–127.
Kriege, N. & Mutzel, P. (2012). Subgraph Matching Kernels for Attributed Graphs. International Conference on Machine Learning (12), 291–298. Locatelli, M., Seghezzi, E., Pellegrini, L., Tagliabue, L. C., & Di Giuda, G. M. (2021). Exploring Natural Language Processing in Construction and Integration with Building Information Modeling: A Scientometric Analysis. Buildings 11(12), 583. Mattern, H. & König, M. (2017). Concepts for Formal Modeling and Management of Building Design Options. Computing in Civil Engineering, 59–66. Mattern, H. & König, M. (2018). Bim-based Modeling and Management of Design Options at Early Planning Phases. Advanced Engineering Informatics 38, 316–329. MITRE (2014). Systems Engineering Guide. Bedfort, MA and McLean, VA. Napps, D., Pawlowski, D., & König, M. (2021). Bim-based Variant Retrieval of Building Designs Using Case-based Reasoning and Pattern Matching. International Association for Automation and Robotics in Construction (38), 435–442. Napps, D., Zahedi, A., König, M., & Petzold, F. (2022). Visualisation and Graph-based Storage of Customised Changes in Early Design Phases. International Association for Automation and Robotics in Construction (39), 191–198. Østergård, T., Jensen, R. L., & Maagaard, S. E. (2016). Building Simulations Supporting Decision Making in Early Design – A Review. Renewable and Sustainable Energy Reviews 61, 187–201. Purohit, S., van Nhuy, & Chin, G. (2020). Semantic Property Graph for Scalable Knowledge Graph Analytics. Big Data. Rahnama, J. & Hüllermeier, E. (2020). Learning tversky similarity. pp. 269–280. Springer, Cham. Robinson, I., Webber, J., & Eifrem, E. (2015). Graph Databases: New Opportunities for Connected Data (Second edition ed.). Beijing: O’Reilly. Sabi, Q. u., Bayer, J., Eisenstadt, V., Bukhari, S. S., & Dengel, A. (2017). Semantic Pattern-based Retrieval of Architectural Floor Plans with Case-based and Graph-based Searching Techniques and Their Evaluation and Visualization. International Conference on Pattern Recognition Applications and Methods (9), 50–60. Schapke, S.-E., Beetz, J., König, M., Koch, C., & Borrmann, A. (2018). Collaborative Data Management. In Building Information Modeling, pp. 251–277. Springer, Cham. Tversky, A. (1977). Features of Similarity. Psychological Review 84(4), 327–352. Watson, I. (1995). An Introduction to Case-based Reasoning. pp. 1–16. Springer, Berlin, Heidelberg. Young, N. W., Jones, S. A., Bernstein, H. M., & Gudgel, J. E. (2009). The Business Value of Bim: Getting Building Information Modeling to the Bottom Line. Smart Market Report.
396
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Retrieve information from construction documents with BERT and unsupervised learning M. Shi, T. Heinz & U. Rüppel Technical University of Darmstadt, Darmstadt, Germany
ABSTRACT: The exploitation of using text documents from precedent projects for decision-making in the construction industry is still at a low level. One reason is that the in unstructured natural language formulated information cannot be processed directly by computer programs and the search is conducted by keywordsmatch, which is inefficient and imprecise. To make the information of unstructured text document accessible in digital processes without introducing additional manual work, we propose using natural language processing and unsupervised learning methods to automatedly extract information from unstructured textual documents. This paper describes an NLP-based pipeline that includes methods for data acquisition and preprocessing, different transformer-based embedding methods, and subsequent downstream tasks. Our proof-of-concept is trained on documents from different waterways construction projects in the German language. Because of the unsupervised learning and available language models, this pipeline can be generalized to other languages and construction types.
1 INTRODUCTION The digitization of the construction industry is constantly making document management and archiving more effortless and more efficient. Storing documents is possible in just a few steps. Numerous documents are created in every project process, from conception to demolition. In those textual documents stored information is essential for decision-making in the following phases and projects. However, finding the needed information can be very time-consuming with the growing number of documents. One reason is that information is usually formulated in natural language that computer programs cannot directly process. Many document management systems support metadatabased “exact-match” or wildcard search, which only displays results when the searched keywords match the manually given documents’ metadata, such as title or tags. This search method requires the metadata to cover the document’s content correctly and users to have precise knowledge about keywords that lead to a high hit rate. Otherwise, the search engine returns many irrelevant documents, which causes additional selection efforts. Another disadvantage is that synonyms and typing errors can not be handled, which leads to the exclusion of relevant documents. In recent years, various natural language processing (NLP) and machine learning methods have been used to automatedly process information retrieval tasks, inter alia, automated extraction of risk cases from
DOI 10.1201/9781003354222-51
accidents reports, extraction of poisonous contract clauses, and compliance checking. Those research focused only on one type of document and mostly applied rule-based methods or supervised learning. Both cases rely heavily on manual work for rule formulation or hand labeling. There is a research gap in using unsupervised and deep learning methods to retrieve information from unstructured project documents for a more intelligent and efficient document search. Thus, the authors propose a pipeline that automatedly extracts data and information from unstructured textual documents. Those extracted data and information can be used to assist the semantic search process later on, which is out of the scope of this paper. The pipeline uses the language model BERT to embed words. Three unsupervised learning-based downstream tasks, topic modeling, named entity recognition, and keywords extraction, are taken to extract the content of every document. Besides the downstream tasks, we suggest using cosine similarity to process user queries and return corresponding files with match probabilities based on sentence embeddings. This paper is organized as follows. Section 2 gives a literature review of the NLP application in the AEC industry. In section 3, the methodologies used in this paper are explained. Afterward, we describe the proposed pipeline and the proof-of-concept implementation in section 4. The following section 5 evaluates and discusses results from the implementation. Finally, the conclusion and future work are given in section 6.
397
2 RELATED WORKS NLP techniques have been increasingly drawing attention in the construction domain during the last decades. This is especially true for automated information extraction and retrieval, which save engineers from heavy manual work in various disciplines. One of the research fields is using NLP to support safety engineering and risk management. Tixier et al. proposed a rule-based framework to empower construction companies to extract structured datasets from unstructured accident reports (Tixier et al. 2016). Zou et al. presented combining the TF-IDF-based vector space model and semantic query expansion upon WordNet to retrieve risk cases, such as risks in designs and safety risks on construction sites, from previous construction projects (Zou et al. 2017). Later on, Zhang et al. applied different statistical machine learning algorithms and ensemble models to analyze construction accident reports (Zhang et al. 2019). NLP also supports project engineers in analyzing bid documents based on project data from the past and current risks for Engineering-ProcurementConstruction (Kim et al. 2020). Lee et al. applied rule-based NLP in international construction contracts for automated extraction of poisonous contract clauses to support contractors (Lee et al. 2019). Another application domain is automated building code interpretation and compliance checking (Fuchs & Amor 2021). The most explored strategies for information extraction from building codes are rulebased, ontology-based, or deep learning-based (Fuchs & Amor 2021). The deep learning-based approach is gaining more attention because of its higher scalability compared to the other two methods (Fuchs & Amor 2021) (Moon et al. 2021) (R. Zhang & ElGohary 2019) (R. Zhang & El-Gohary 2020). Leng et al. applied the Bidirectional Long Short Term Memory network with Conditional Random Field to automatedly extract entities and relations from unstructured textual data about Mechanical, Electrical, and Plumbing systems (Automatic MEP Knowledge Acquisition Based on Documents and Natural Language Processing 2019). Zheng et al. compared the performances of BERT-based language models with that of traditional deep learningbased (not Transformer-based) word embeddings on text classification and NER tasks in their preprint paper (Zheng et al. 2022). Their results showed that the BERT-based models outperformed the traditional deep learning-based embeddings. They contributed the better performance of BERT-based models to their masked language structure, which can create contextualized word representations (Ethayarajh 2019). The previous research mainly focused on one domain (compliance checking, contract management, etc.). Information extraction and retrieval through multidisciplinary and different project phases are understudied. In addition, most of the research base on machine learning and deep learning are supervisedlearning methods that require much manual labeling
work. The potential of unsupervised deep learning methods is not researched enough. Thus, this paper proposes a pipeline to bridge the gap in applying unsupervised deep learning methods to multidisciplinary textual data.
3 METHODOLOGIES 3.1 Natural language processing & Information retrieval NLP is an inter-discipline of computer science and linguistics. With NLP techniques, computers can process and analyze large amounts of textual data for various tasks. Information retrieval (IR) refers to finding the material that satisfies an information need within large unstructured data collections (Manning et al. 2009). IR is mainly used in search engines such as Google or enterprise information management systems, where the system prepares documents according to user needs. Various models have been explored since the beginning of research in IR. These models can be roughly clustered into Boolean retrieval, vector space model, probabilistic retrieval, and the language model (Manning et al. 2009). With the development of deep learning algorithms and high-performance hardware, deep neural networks have been thriving in IR research in recent years. One of the application fields is training language models that can be used for a sequence of downstream tasks such as NER and question-answering. 3.2 Embedding and language model For deep learning models to process human language, words must be transformed into numeric representations, specifically to vectors in a continuous vector space. This activity is called word embedding. Besides word embedding, sentence embedding and document embedding are also commonly used in different NLP tasks. Multiple word embeddings and language models were developed, and various algorithms (Mikolov et al. 2013) have been explored to preserve the semantic information in the vector representations. A language model is a function that builds probabilities over strings drawn from some vocabulary based on the surrounding word sequences (Manning et al. 2009). One of the representative language models is Transformer. The Transformer-based language model is a network architecture that does not require sequential processing of tokens, which makes the training process parallelizable and faster (Vaswani et al. 2017). As a transformer-based model, BERT, which stands for Bidirectional Encoder Representations from Transformers, is pre-trained on two tasks with huge datasets (Devlin et al. 2018). The first task is masked language modeling, where 15% of the tokens are masked and BERT should predict them from the context. Then BERT is trained with the following sentence prediction task so that the model can capture the relationship
398
between two sentences (Devlin et al. 2018). After the computationally expansive pre-training, BERT is capable of learning the contextual relations both at the token and sentence level and has a high generalization ability to unseen data. As a sequence, the user can fine-tune parameters in the BERT model to various downstream tasks, e.g., NER, with a relatively small dataset (Devlin et al. 2018). In the fine-tuning process, task-specific inputs and outputs are plugged into BERT, and all parameters will be calculated further based on the pre-trained parameter values. Compared to the pre-training process, fine-tuning is computationally inexpensive (Devlin et al. 2018).
3.3 Named entity recognition and keyword extraction Named entity recognition (NER) refers to recognizing information-unites like names, including person, organization, location, and numeric expressions from unstructured data (Nadeau & Sekine 2007). Those recognized entities are then assigned corresponding entity labels. NER plays an essential role in NLP systems for IR, relation extraction, etc. (Hedderich et al. 2021). Pre-trained word embeddings have been proven to outperform traditional feature-engineered supervised methods in NER tasks (Hedderich et al. 2021). Automatic keyword extraction is another method that is commonly applied to IR systems. Keywords provide a compact representation of a document’s content that supports IR (Rose et al. 2010) and can be used to search over large textual materials efficiently.
3.4 Topic modeling Topic modeling, especially LDA-based document models, is a statistical tool to extract documents’ topics and has been used for information retrieval for a long time (Wei & Croft 2006). Topic modeling discovers topics with corresponding probabilities from a document based on the statistical distribution of words. Furthermore, this method allows assigning keywords to each topic that the human reader can interpret. In addition to classical models like LDA, several embedding-based topic modeling techniques have shown their advantages (Grootendorst 2022).
4 CONCEPT AND IMPLEMENTATION Based on the described methodologies, we proposed a pipeline for retrieving information from multidisciplinary project documents to assist information retrieval. This pipeline comprises five steps: data acquisition, preprocessing, embedding and downstream tasks training (NER, keyword extraction, topic modeling), similarity analysis, and semantic search. The results from all three downstream tasks are intended to serve users of an archiving system with additional information in the form of metadata without having to look explicitly into the document. The three tasks are designed to complement each other and provide information that users can view at a glance. This paper focuses on the first four steps. In each subsection, we first describe the concept and then give information on used libraries for implementation. The implementation of the proposed pipeline is achieved by utilizing several libraries in python. The described pipeline steps are implemented in a Google Colab Notebook to provide an easy way to recreate the development environment and a structured overview of the single pipeline steps. 4.1 Data acquisition For data acquisition, 346 PDF project documents, ranging from 1 to 645 pages, were provided by the German Federal Waterway Engineering and Research Institute. The project documents are available both in coded PDF files and scanned PDF files, requiring different reading methods. Unlike coded or vector-based PDFs, in which characters can be read as initially created, scanned PDFs consist of pixels that need to be “transformed” to characters by Optical Character Recognition (OCR). Afterward, we stored the acquired textual data in a relational database for further processes. To extract the text from the 346 PDF project documents, we used Apaches Tika for non-OCR documents and Googles Tesseract to detect text in OCR documents whose content is not accessible with Tika. The two data acquisition methods could extract textual data from 343 PDFs, which equals 5.35 Megabytes in CSV format. The captured texts are saved in a simple SQLite-DB to make them easily accessible for further steps.
3.5 Cosine similarity
4.2 Preprocessing
Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between two vectors and determines whether two vectors are pointing in roughly the same direction (Han et al. 2012). In NLP, cosine similarity is a helpful measure of how similar two words, sequences, or documents are. Though the cosine value mathematically ranges from -1 to 1, it is usually used in [0,1] to measure text similarity. The closer the cosine value to 1, the greater the match between two vectors (Han et al. 2012).
The preprocessing has a decisive influence on the performance of the following learning models in NLP. Essential preprocessing components include tokenization, stopwords elimination, stemming (Vijayarani et al. 2015), and lemmatization. The selection of preprocessing depends on the downstream tasks (topic modeling, NER, and keywords extraction). To optimize the performance of each downstream task, we iteratively undertook different preprocessing components and derived two strategies: weak and strong
399
preprocessing. The weak preprocessing eliminates meaningless characters, such as ASCII and special characters, and is applied to NER. While the strong preprocessing has additional components, eliminating general and domain-specific stop words (words that are not important for the content of a document), text lowering, and lemmatization. The strong preprocessing is applied to topic modeling and keywords extraction. The data preprocessing step utilizes the pythonbased NLP libraries spacy and NLTK, which are mainly used for removing stop-words and lemmatization. In addition, we expanded the preprocessing step with multiple small cleaning functionalities, which aim to remove additional content like URLs or special characters. Also, a corpus-specific list of special stop words is removed from the text, which was successively retrieved by evaluating the results of the downstream tasks. 4.3 Embedding & training downstream tasks The preprocessed data are then embedded with BERT language models. As mentioned in section 3.2, the pre-training process requires large training data and high computational capacity, which small institutions or individuals do not possess. Thus, several organizations published their pre-trained BERT models for free use in communities. This paper applied multiple pre-trained BERT variants depending on downstream tasks. 4.3.1 Topic modeling Since the available documents are raw data and their available classes from the document management system are of errors, we propose to use the unsupervised method, topic modeling, to cluster the documents into multiple categories. We implement the topic modeling task based on the BERTopic model (Grootendorst 2022). BERTopic generates document embedding with pre-trained Sentence-Bert (Reimers & Gurevych 2019). Sentence-BERT is a modification of the pre-trained BERT model that converts sentences and paragraphs instead of words to fixed-sized vectors. With direct sentence embedding, sentence similarity can be performed highly efficiently on modern hardware (Reimers & Gurevych 2019). Afterward, BERTopic clusters these embeddings with HDBSCAN (McInnes et al. 2017) and finally generate coherent topic representations with the class-based TF-IDF method. The topic modeling is implemented by using the python library BERTopic (Grootendorst 2020/2022). To utilize recognized topics with human-understandable information, BERTopic assigns probabilities to possible topics and returns a word list that represents each topic. 4.3.2 Named entity recognition Unlike Topic modeling, where semantically meaningful sentence embeddings are more suitable, NER
concentrates more on wordwise embeddings. In addition, domain-specific terms and abbreviations are important for information retrieval and need to be recognized by the NER model. Thus we employed the RoBERTa for word embedding, which stands for Robustly optimized BERT approach (Liu et al. 2019). RoBERTa employed multiple training decisions in BERT and significantly improved the performance. After the embedding, we applied FLAIR, one of the state-of-the-art models for sequence labeling tasks like NER (Akbik et al. 2019). FLAIR allows fast and simple integration of pre-trained transformer models. The NER task needs to extract entities of type location (LOC), organization (ORG), and miscellaneous (MISC) from documents. In the implementation, we used the ner-germanlarge model which is based on the embedding model XLM-RoBERTa and FASTTEXT (Schweter & Akbik 2021) embeddings that are fine-tuned on NER-Tasks. 4.3.3 Keywords extraction For keyword extraction, embeddings of individual keyword candidates are compared with the embeddings of sentences or a document based on cosine similarity. Thus, the embedding is, like in topic modeling, handled by Sentence-BERT. In our case, we used the pre-trained paraphrase-multilingual-MiniLM-L12-v2 Sentence-Transformer. The keyword extraction leverages KeyBERT, a python library that provides functionality that uses BERT or Sentence-BERT embeddings and cosine similarity to recognize important keywords in documents by comparing the embedding of the whole document with embeddings of single words from those documents. 4.4 Similarity analysis In the similarity analysis, we focus on the similarity of user queries with the documents. We again utilized Sentence-BERT Models to create embeddings of the whole document corpus, and user queries are embedded with the same model in an ad-hoc manner. We then use cosine similarity to compare the embedded documents with the embedded query and return the five most similar documents. Cosine similarity can solve the exact match problem mentioned in the instruction section. The used library is the same as in keyword extraction, the paraphrase-multilingual-MiniLM-L12v2 Sentence-Transformer.
5 EVALUATION AND DISCUSSION To evaluate the results produced by the proposed pipeline, we took a subset of 14 documents from the provided document corpus (see Table 1). We first manually evaluated the extracted information from a domain-specific point of view for all three downstream tasks. An overview of the validation is listed in Table 2. Afterward, statistical metrics are applied in the
400
Table 1.
Evaluation dataset.
Nr.
Pages
Data acquisition
Content
1 2 3 4 5 6 7 8 9 10 11 12 13 14
645 34 13 1 5 59 48 9 4 238 45 63 2 4
standard standard standard standard standard standard standard standard standard OCR OCR OCR OCR OCR
Operating handbook weir rolling mill Component overview and acceptance test Barrage structural survey Supplement to measuring program for structural measurement at Hofen lock Datasheet -Position switch Audit report of the structural auxiliary building structures Dimension of individual components in weir system Bored pile log Absolute rotary encoder bus interfaces Installation assembly and operating instruction of clutches, brakes, lifting devices, etc. Overview of component orders (from 1923) Structural data and load case calculations of a weir system Test report dam panels of the underwater inspection lock Datasheet and product description of lightening with motion detector
topic modeling task. To maximize the heterogeneity, we composed the evaluation subset concerning the following criteria: document length (between 1 and 645 pages), the diversity of topics covered, and the document quality (both OCR documents and non-OCR documents). Table 2.
Results’ overview of downstream tasks.
Downstream tasks
Dataset
Recognized cases
Recognition rate (%)
Keyword extraction NER* Topic modeling
14 14 14
10 11/12/9 11
71.42 76.18 78.57
*The recognition rate of NER is an average rate based on LOC: 78.54%, ORG: 85.71%, MISC: 64.29%.
Due to the various purpose of each downstream task, we set different success criteria for each task. The requirement for the keyword extraction is to give a quick overview of the most important terms from a document. Therefore, the list of keywords should comply with the content and have an adequate length. 10 of 14 evaluation documents fit those criteria. After analyzing the datasets and respective outcomes, we found that two reasons were accountable for the failed keyword extraction of those four files. One reason that all downstream tasks have in common is that some documents could not be read correctly because of the bad scan quality of PDFs, which makes delivered results unusable. Another reason is that the number of extracted keywords correlates with the length of the document, which results in many irrelevant keywords for very long document with multiple subjects. The NER should assign recognized terms with LOC, ORG, or MISC entities. For all three cases, the named entity extraction is considered as successful when >80% of the extracted entities fit the corresponding category. This task achieved good results in
9 of 14 evaluation cases. For LOC entities, information like countries, cities, or street names is extracted correctly in 11 documents (recognition rate 78.54%). Entities of type ORG were recognized with a similar rate of 85.71% (12 of 14 documents). MISC entities delivered information for technical standards or construction components and tools. MISC entities are correctly found in 9 out of 14 documents (64.29%). The NER task failed in some cases due to the same issues the keyword extraction failed: low qualitative data acquisition and long documents with multiple subjects. The topic modeling is considered successful when the identified topics represent the documents’ content. Therefore, the learned model gives every topic name derived from the topic-defining keywords. If at least two of the top 3 assigned topics fit the scope of the document, an assignment was considered to be correct, which was the case in 11 of 14 documents. An additional challenge when evaluating topic modeling results is finding the optimal parameter combination of the number of topics a model should find in the corpus and the applied preprocessing methods in the previous step. Therefore we introduce a coherence score based on FASTTEXT-embeddings and cosine similarity to evaluate the topic modeling results. In order to determine the effects of the preprocessing as well as the optimal number of topics, we gather the coherence scores from different topic models, each combined with three levels of preprocessing, as shown in Figure 1. It is clear to see that strong preprocessing outperforms variations with weak or no preprocessing. The strong preprocessing utilizes additional stop-word removal and lemmatization, while the weak preprocessing only applies less advanced methods like URL or punctuation removal. A topic’s coherence score is measured by calculating the mean of cosine similarities from all pairwise combinations of words describing a topic, as shown in Figure 2. The coherence score of a topic model is then calculated by taking the mean
401
The results also showed some shortcomings. For scanned PDF documents, data acquisition is essential for downstream processes and a more sophisticated method for OCR analysis is needed. In addition, the low performance of the downstream tasks on the document “Operating handbook weir rolling mill” showed that unsupervised learning methods that are based on statistical models can not handle large documents with a high diversity of subjects. 6 CONCLUSION Figure 1.
Preprocessing and number of topic optimization.
of the coherence scores from all topics belonging to a topic model. In opposition to the keyword extraction and the named entity recognition, the topic modeling is not affected by the length of a document.
Figure 2. Similarity matrix of a Topic.
For the evaluation of the similarity analysis, three example queries were held against the embedded document corpus. For each query, the top five most similar text parts from the whole corpus were returned and evaluated from a domain-specific point of view. On average, three of the five returned most similar text parts were considered meaningful and fitting for the query case. One of the main problems we faced during the evaluation of the similarity analysis was again related to the varying length of the documents. The longer the document is, the more likely the document can cover queries. For one query, for example, 4 of 5 of the most similar text parts were found in the same document, which is also the longest one. Though evaluation results of the pipeline are far from the level for industrial implementation, the results can be taken as satisfied considering the fact that no hand-labeled data were used. The results shows the research potential in using available general pre-trained, BERT-based language models to extract information from domain-unspecific project documents. Another conclusion from the results is that appropriate preprocessing and hyperparameters play essential roles in model performances even with large pre-trained language models.
In this paper, we proposed a pipeline for extracting information from multidisciplinary project documents to aid the information retrieval and described the necessary steps from acquiring data from raw documents through preprocessing and embedding to extracting representative information with downstream tasks. We execute the proof of concept by implementing the concept on 343 real-world project documents with open source python libraries. The results show that transformer-based text embeddings can offer a strong foundation for exploring new methodologies to extract information from textual documents. Also, the results point out that applying the proper text extraction procedure and preprocessing methods has a considerably significant impact on the quality of the results. By analyzing the documents in which the downstream tasks failed, we found that long documents with multiple subjects will harm the performance of statistical unsupervised learning models. A solution could be splitting large documents into multiple subdocument with single subject to execute the downstream tasks separately. This work’s main contribution is to combine deep learning-based language models and unsupervised training processes for information retrieval in AEC project documents and give an overview of potential NLP models. Even though the documents are mainly in German, the pipeline itself is language-independent and can be transferred to use cases in other languages where the language models are available. We also conclude that even though the results from unsupervised models are promising, they are not enough for extracting information from project documents for industrial implementation. Future work could be fine-tune BERT-based models with labeled data and supervised learning methods to extract information from the document, and compare the performance with the results in this work. The needed labeled data for the supervised learning methods should be generated automatically in first place, based on domain-specific knowledge. ACKNOWLEDGEMENTS We thank the German Federal Waterway Engineering and Research Institute and the German Federal Waterways and Shipping Administration for providing the data.
402
REFERENCES Akbik, A., Bergmann, T., Blythe, D., Rasul, K., Schweter, S., & Vollgraf, R. (2019). FLAIR: An Easy-to-Use Framework for State-of-the-Art NLP. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), 54–59. https://doi.org/10.18653/v1/N19-4010 Automatic MEP KnowledgeAcquisition Based on Documents and Natural Language Processing. (2019, September 18). Jia-Rui Lin’s Page. https://linjiarui.net/en/publications/ 2019-09-18-automatic-mep-knowledge-acquisition-base d-on-nlp Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. https://doi.org/ 10.48550/arXiv.1810.04805 Ethayarajh, K. (2019). How Contextual are Contextualized Word Representations? Comparing the Geometry of BERT, ELMo, and GPT-2 Embeddings. ArXiv:1909. 00512 [Cs]. http://arxiv.org/abs/1909.00512 Fuchs, S., & Amor, R. (2021). Natural Language Processing for Building Code Interpretation: A Systematic Literature Review. 10. Grootendorst, M. (2022). BERTopic: Neural topic modeling with a class-based TF-IDF procedure. ArXiv:2203.05794 [Cs]. http://arxiv.org/abs/2203.05794 Grootendorst, M. (2022). BERTopic [Python]. https://github. com/MaartenGr/BERTopic (Original work published 2020) Han, J., Kamber, M., & Pei, J. (2012). 2—Getting to Know Your Data. In J. Han, M. Kamber, & J. Pei (Eds.), Data Mining (Third Edition) (pp. 39–82). Morgan Kaufmann. https://doi.org/10.1016/B978-0-12-381479-1.00002-2 Hedderich, M.A., Lange, L.,Adel, H., Strötgen, J., & Klakow, D. (2021). A Survey on Recent Approaches for Natural Language Processing in Low-Resource Scenarios. ArXiv:2010.12309 [Cs]. http://arxiv.org/abs/2010.12309 Kim, Y., Lee, J., Lee, E.-B., & Lee, J.-H. (2020). Application of Natural Language Processing (NLP) and Text-Mining of Big-Data to Engineering-Procurement-Construction (EPC) Bid and Contract Documents. 2020 6th Conference on Data Science and Machine Learning Applications (CDMA), 123–128. https://doi.org/10.1109/CDMA473 97.2020.00027 Lee, J.,Yi, J.-S., & Son, J. (2019). Development ofAutomaticExtraction Model of Poisonous Clauses in International Construction Contracts Using Rule-Based NLP. Journal of Computing in Civil Engineering, 33(3), 04019003. https://doi.org/10.1061/(ASCE)CP.1943-5487.0000807 Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019). RoBERTa: A Robustly Optimized BERT Pretraining Approach. ArXiv:1907.11692 [Cs]. http://arxiv.org/abs/1907.11692 Manning, C. D., Raghavan, P., & Schütze, H. (2009). Introduction to Information Retrieval. 569. McInnes, L., Healy, J., & Astels, S. (2017). hdbscan: Hierarchical Density Based Clustering. Journal of Open Source Software, 2(11), 205. https://doi.org/10.21105/joss. 00205 Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013). Distributed Representations of Words and Phrases andTheir Compositionality. Proceedings of the 26th International Conference on Neural Information Processing Systems - Volume 2, 3111–3119.
403
Moon, S., Lee, G., Chi, S., & Oh, H. (2021). Automated Construction Specification Review with Named Entity Recognition Using Natural Language Processing. Journal of Construction Engineering and Management, 147(1), 04020147. https://doi.org/10.1061/(ASCE)CO.1943-786 2.0001953 Nadeau, D., & Sekine, S. (2007). A Survey of Named Entity Recognition and Classification. Lingvisticae Investigationes, 30(1), 3–26. https://doi.org/10.1075/li.30.1.03nad Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. ArXiv:1908.10084 [Cs]. http://arxiv.org/abs/1908.10084 Rose, S., Engel, D., Cramer, N., & Cowley, W. (2010). Automatic Keyword Extraction from Individual Documents. In M. W. Berry & J. Kogan (Eds.), Text Mining (pp. 1–20). John Wiley & Sons, Ltd. https://doi.org/10.1002/9780470 689646.ch1 Schweter, S., & Akbik, A. (2021). FLERT: Document-Level Features for Named Entity Recognition.ArXiv:2011.06993 [Cs]. http://arxiv.org/abs/2011.06993 Tixier, A. J.-P., Hallowell, M. R., Rajagopalan, B., & Bowman, D. (2016). Automated Content Analysis For Construction Safety: A Natural Language Processing System to Extract Precursors and Outcomes From Unstructured injury reports. Automation in Construction, 62, 45–56. https://doi.org/10.1016/j.autcon.2015.11.001 Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need. ArXiv:1706.03762 [Cs]. http://arxiv.org/abs/1706.03762 Vijayarani, S., Ilamathi, J., & Nithya. (2015). Preprocessing Techniques for Text Mining—An Overview. International Journal of Computer Science & Communication Networks, 5(1), 7–16. Wei, X., & Croft, W. B. (2006). LDA-based document models for ad-hoc retrieval. Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval - SIGIR ’06, 178. https://doi.org/10.1145/1148170.1148204 Zhang, F., Fleyeh, H., Wang, X., & Lu, M. (2019). Construction SiteAccidentAnalysis UsingText Mining and Natural Language Processing Techniques. Automation in Construction, 99, 238–248. https://doi.org/10.1016/j.autcon. 2018.12.016 Zhang, R., & El-Gohary, N. (2019). A machine learningbased approach for building code requirement hierarchy extraction. 2019 CSCE Annual Conference. https:// par.nsf.gov/biblio/10110925-machine-learning-based-app roach-building-code-requirement-hierarchy-extraction Zhang, R., & El-Gohary, N. (2020). A Deep-Learning Method for Evaluating Semantically-Rich Building Code Annotations. In L.-C. Ungureanu & T. Hartmann (Eds.), EG-ICE 2020Workshop on Intelligent Computing in Engineering, Proceedings (pp. 285–293). Universitatsverlag derTU Berlin. http://www.scopus.com/inward/record.url? scp=85091017805&partnerID=8YFLogxK Zheng, Z., Lu, X.-Z., Chen, K.-Y., Zhou, Y.-C., & Lin, J.-R. (2022). Pretrained Domain-Specific Language Model for General Information Retrieval Tasks in the AEC Domain. Computers in Industry, 142, 103733. https://doi.org/10.1016/j.compind.2022.103733 Zou, Y., Kiviniemi, A., & Jones, S. W. (2017). Retrieving Similar Cases For Construction Project Risk Management Using Natural Language Processing Techniques. Automation in Construction, 80, 66–76. https://doi.org/10.1016/ j.autcon.2017.04.003
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Automated floorplan generation using mathematical optimization Y. Zhong & A. Geiger Karlsruhe Institute of Technology, Karlsruhe, Baden-Württemberg, Germany
ABSTRACT: This paper presents a tool for automated floorplan generation by using mathematical optimization. The floorplan generation is transferred into an optimization problem, the building outlines and the rooms are parametrized as axis-aligned polygons and rectangles, and the requirements for the size, position and adjacency of each room are formulated as constraints in the optimization. The model is built as a mix-integer nonlinear programming and can be solved by a mathematical programing solver. In the presented paper, a tool is introduced for defining the input requirements, performing the floorplan layouts and exporting the results as an IFC model which can be further processed in openBIM compatible tools. The result shows that the developed optimization model fulfills the requirements of a room book and provides appropriate floorplan proposals.
1 INTRODUCTIONS
2 STATE OF THE ART
In the architectural design, one of the challenges is to optimize the relationship between different spaces, when their own sizes and positions are considered as well. Combining these factors a layout design can be created at an early stage, that can be used for a further refinement and simulation, taking external elements into account, such as the urban surroundings or the natural environment. This paper introduces a tool for generating floorplan proposals automatically based on a set of requirements for rooms or spaces. The size and positional requirements of each room and their adjacencies are parametrized, as well as the outline of each floor in the building. The generation of floorplans is modelled as a mixed integer nonlinear programming (MINLP) problem, which can be solved by various methods and toolkits. The goal is to provide different floorplan proposals that can be used in the early stage of architectural design and they can be exported as IFC data format for further adjustments, validation or simulation. The Layout Designer presented in this paper is a tool to generate early design of floorplan layouts, which meet the spatial and relational requirements. It is part of the research project Smart Design and Construction (SDaC). The goal of this project is to discover the possibilities of artificial intelligence in the architecture and construction industry. The Layout Designer serves in this project for the use case automated floorplan generation. The results from Layout Designer can be exported as IFC data format and be used by the partners in this project for other use cases.
There have been several researches on how to generate layouts automatically. One of the challenges of automated layout generation is to fulfill the requirements of the users, which can variate in different manners. How to formulate these requirements, so that the computer is able to “understand” them and solve the problem, has been a central task of automated layout generation. Various methods have been used for floorplan generation. (Lopes et al. 2010) uses procedural modeling techniques to create suitable layouts based on geometric grids. (Merrell et al. 2010) uses data-driven methods to fulfill high-level requirements in the automated floorplan generation. (Schneider et al. 2011) describes a method using evolutionary algorithms for the generative process, in which the spaces are limited as rectangular grids. (Hempel et al. 2015) formalizes the requirements for the design of a hospital, and prioritizes them based either on practical knowledge or according to user customization. With the developed tool Early Design Configurator different layout proposals can be provided in the early stage, and these proposals can be used in other tools for further validation or simulation. (Egor et al. 2020) proposes an evolutionary strategy in the Magnetizing Floorplan Generator, in which the accessibility between rooms are emphasized. Artificial Intelligence has been a popular topic in recent years. In the construction industry, its potential is also being discussed and relevant applications are being developed. (Chaillou 2019) uses Generative Adversarial Networks (GAN) to generate different apartment building design.The networks are trained by
404
DOI 10.1201/9781003354222-52
a large number of labelled floorplans to generate footprints, split rooms and place furniture. (Hu et al. 2020) introduces Graph2Plan, in which they present a deep learning framework to generate floorplans by learning the rooms’ relationship through a graph. (Nauata et al. 2021) combines GAN and relationship graph to generate floorplans. The layout generation is based on a set of networks learning the relation-graphs of the rooms, and the generated layouts can accurately reproduce these relationships. (Wu et al. 2018) develops a MIQP-based methods for layout design, in which the layout generation is formalized as a mixed-integer quadratic programming. This method provides the users enough freedom to customize their requirements on the layout design. (Shekhawat et al. 2020) presents GPLAN for floorplan generations based on graph theory and linear optimization. 3 METHODOLOGY In this chapter the method of floorplan generation is introduced. The generation is transformed to an optimization problem, and the model is considered as mix-integer nonlinear programming (MINLP). The formulation of the problem starts by creating the outline of the building with one or multiple floors. In this developed tool, the outline of each floor can be different but all are restricted as rectilinear polygons. The second step is to create a list of requirements for each room. It includes the size, position of them and their adjacencies. The requirement list can be created manually by the user in the program, or imported from external sources and presented as a table. These requirements are regarded as constraints in the optimization problem. In the following it will be introduced how they are parametrized for calculation. In order to simplify the modelling of optimization problem, every room is modelled as a rectangle, which is sufficient for early stage design. It concludes four variables: (xi , yi ) is the coordinates of the bottom left corner of the rectangle, and (li , wi ) describes its length and width. Furthermore, based on the amount of floors in the building, a set of auxiliary variables Fi = (f0i , f1i , . . . fni ), is used to describe on which floor the room is located. These auxiliary variables are binary and their sum is restricted to be one, which is: n fki = 1
(1)
k=0
Last but not least, another set of auxiliary variables is used to indicate whether or not two rooms are on the same floor, and they are defined as follows: fij =
n fki fkj (i < j)
(2)
Figure 1. Parameters of a room. i indicates the index of each room.
be regarded as a bounding box subtracted by several rectangles. The inside constraint can be expressed as: ⎧ ⎪ ⎨
−M (1 − fki ) ≤ xi ≤ Lk + M (1 − fki ) −M (1 − fki ) ≤ yi ≤ Wk + M (1 − fki ) , ⎪ ⎩ −M (1 − fki ) ≤ xi + li ≤ Lk + M (1 − fki ) −M (1 − fki ) ≤ yi + wi ≤ Wk + M (1 − fki )
(3)
where Lk and Wk are the length and width of the bounding area of each floor. A large constant M is used here to ensure the inequality when fki equals 0 (which means the room is not on the floor k). To express the constraint regarding to the subtracted rectangles, these rectangles are also formalized as “rooms” that should not be covered by any actual room. No rooms should overlap with each other and this constraint is described as follows: ⎧ x ≥ xj + lj − M 1 − θijR fij ⎪ ⎪ i ⎪ L ⎪ ⎨ xj ≥ xi + li − M 1 − θij fij yi ≥ yj + wj − M 1 − θijT fij , (4) ⎪ B ⎪ ≥ y + w − M 1 − θ f y ⎪ ij ij ⎪ ⎩ jR i L i T θij + θij + θij + θijB − fij ≥ 0
Figure 2. Outline of a floor in the building. The upper right part is considered as a room that cannot be covered by any other rooms, and the whole polygon could be considered as its bounding box subtracted by this small rectangle. d where θijk (d = R, L, T , B) describe the four directions of the relative position of two rooms: right, left, top and bottom. The last inequation ensures that when two rooms are on the same floor, they should not overlap at least in one of the four directions, otherwise they shouldn’t be on the same floor. Again, the constant M is used here to ensure the inequality in general cases. The size of the rooms can be restricted either by setting the range for the length and width or by setting target values for both dimensions and put the size error in the objective function. Setting a range for the values are regarded as constraint and is expressed as follows:
k=0
One of the basic constraints is that a room must be inside of the contour of each floor, which can
405
lmin ≤ li ≤ lmax wmin ≤ wi ≤ wmax
(5)
Another option for size control is the aspect ratio. It is a constraint that limits the difference between the length and width of the room. Giving a range (rmin , rmax ) for the aspect ratio and this is defined as follows: li ≥ rmin wi li ≤ rmax wi
(6)
The position of a room can also be restricted by setting specific point of the bottom left corner. By using a point (xs , ys ) and this constraint is defined as follows: x i = xs y i = ys
(7)
Figure 3. Boundary constraint. The room should be on one of the three south edges of the building, two of which are on the same floor and the third one is on the other floor.
An alternative of restricting a room’s position is to put it on at least one of the boundaries of the building. This is also necessary, e.g., when designing an apartment in which the bedroom should be on the south side for more sunshine. In this case assuming there are n south edges on the floor k, a set of auxiliary, binary variables βkl (l = 1, 2, . . . , n), and the vertices of each edge are ((xkl1 , ykl ), (xkl2 , ykl )). This constraint is described as follows: ⎧ xi ≥ xkl1 − M (1 − βkl fki ) ⎪ ⎪ ⎨ xi + li ≤ xkl2 + M (1 − βkl fki ) , (8) ≤ ykl + M (1 − βkl fki ) yi ⎪ ⎪ βkl fki ≥ 1 ⎩ k=0,l=1
where the first three inequations ensure that the bottom edge of the room rectangle and one of the south boundaries in the building have overlap, and the last inequation ensures that at least one of such south boundaries should exist (see Figure 3). The adjacency of two rooms is also important in layout design and can be done by different types according to practical need, such as a door connection, an open connection or a wall connection. Assuming that a bathroom should be located next to the main bedroom with a door connection of length d, such constraint can be expressed as follows: ⎧ xi ≤ xj + lj − dcij ⎪ ⎪ ⎪ ⎪ ⎨ xj ≤ xi + li − dcij yi ≤ yj + wj − d 1 − cij , ⎪ yj ≤ yi + wi − d 1 − cij ⎪ ⎪ ⎪ ⎩ fki fkj = 1
Figure 4. Adjacency constraint. Top: two rooms are connected horizontally with a minimum contact length d. Bottom: two rooms are connected vertically.
building, and to minimum the area errors of all rooms. Let Ak (k = 0, 1, . . . , n) be the area of each floor, the objective function is expressed as follows: Ocovered =
Ak − li w i k=0
(10)
i
(9)
k=0
where the auxiliary variable cij describes whether it is a vertical or horizontal connection. When cij = 1, it means the two rooms are connected and their bottom/top edges have overlap. In this case, the first two inequations ensure a minimum length for the overlap so that a door can be placed, and the third and fourth inequations denote a forced overlap of their edges (see Figure 4). The last equation ensures these two rooms have to be on the same floor. The objective function of the optimization consists of two parts, to maximum the covered area of the whole
Another objective function is the size error when target sizes for the rooms are specified, and it is defined as follows: 2 2 li − lit + wi − wit (11) Osize = i
in which (lit , wit ) is the target size of the room. Combining both objective functions above with weight parameters λcovered and λsize yields the final objective function: min λcovered Ocovered + λsize Osize
R,θ ,ρ,c
406
(12)
4 APPLICATIONS In this chapter, the tool SDaC Layout Generator is presented to demonstrate how a layout is generated automatically. The interface of the program is implemented in C++ with wxWidget toolkit and SCIPoptSuite (Bestuzheva 2021) is used as the mathematical programming solver. It should be noted that, this program is still a prototype and needs improvement on user interaction and performance optimization.
Figure 6. Example of topology view to show the relationships of the rooms.
Figure 5.
within a reasonable range. The two rooms with indices 5 and 6 have the type “stair” so their size and position is restricted to be exactly the same while they have to on different floor. Moreover, room 0 has the type “entrance” so that it has to be on one of the edges on the ground floor. The results are exported as IFC format and its 3D-Model can be viewed in other software (see Figure 8).
Requirement set dialog to create or edit a room.
Starting by drafting the outline of each floor of the building, the users can also set the building parameters such as wall strength, door width and floor height, etc. Some of these parameters are applied in the optimization process while the others are used for IFC exportation. The next step is to create a requirement set, in which the users define every room that is needed. Figure 5 shows a dialog in which users can set every details of a room. A topological graph is also shown to present the specified adjacency of the rooms (see Figure 6). It is also supported that users can import a requirement set into the program and make further adjustment.
5 RESULTS AND DISCUSSIONS In this chapter, an example is shown in which a twostory house’s interior layout is generated. First of all the house’s outline is designed with a rectangle and an L-shape polygon. The requirement set includes 13 rooms, describing their size expectation, ideal positions and the required adjacencies. A topological graph is drawn to help users to read the connectivity more clearly. In Figure 7 the layout of each floor is shown. In order to show the adjacencies that are defined in the requirements, doors between rooms are displayed by a white rectangle. Walls are added to present the layouts more realistically. As is shown in the results, all the position and adjacency constraints in the requirement set are fulfilled, while the size errors are controlled
Figure 7. Floorplan example of a two-story house. On the top is the ground floor and on the bottom is the first floor.
407
Another example is to regenerate an apartment design from the architect Alvar Aalto (see Figure 9). All the size, position and adjacency constraints from the original layout are extracted and imported in the optimization model; the results are shown in Figures 10 and 11. The generated layout is similar to the original one when most of the constraints are also fulfilled.
Figure 8. 3D model of the two-story house example. The data is in IFC format.
Figure 10. Regenerated floorplan with constraints extracted from Figure 9.
Figure 9.
Apartment Design from Alvar Aalto.
Comparing to the results in another study which also uses mathematical optimization, one of the main differences between this method and Wu’s method (Wu et al. 2018) is that, the marginal area in the results is not automatically filled by a second-level optimization. This can still be avoided by setting additional constraint:Ocovered = 0. Another difference is that the level variables are added in the optimization model, which enable the user to have more flexibility on the design for a building with multiple floors. This inevitably increases the complexity of the optimization model and influences the performance. However, the average time to obtain an optimal result is still acceptable. 6 CONCLUSIONS AND LIMITATIONS In this paper, the tool for generating floorplan layouts automatically by using mathematical optimization is
Figure 11. 3D model of regeneration floorplan from Alvar Aalto’s apartment.
presented. A mixed integer nonlinear programming is modified to formulate the problem, in which the layout generation is fully parametrized and the requirements for rooms are accordingly modelled as constraints in the optimization. The results show that this method can be used to generate layouts on multiple floors, whether the level constraint is specified or not. Regenerating a layout by extracting constraints from other floorplans is also proved to be possible. At the end, the result is exported as IFC format, which can be visualized as a 3D model and could also be used for further adjustments, validation or simulation. The results in this study show that there is still much to improve. A multithreading framework could be implemented for the SCIPoptSuite toolkit to improve the performance on more complex or large-scale modelling. It is also possible to extract other feasible but not optimal solutions during the optimization, which
408
could represent different layouts, because in architectural design variation is also important. Furthermore, artificial neural networks have the potential to help completing the requirement set, such as building a more complete topological graph or controlling the size error. ACKNOWLEDGEMENT The work presented in this paper belongs to the SDaC project and is supported by Federal Ministry for Economic Affairs and Climate Action (BMWi). REFERENCES Bestuzheva, K., et al. 2021. The SCIP Optimization Suite 8.0. Chaillou, S. 2020. Archigan: Artificial Intelligence x Architecture. Architectural Intelligence: 117-127. Springer, Singapore. Egor, G., et al. 2020. Computer-aided Approach to Public Buildings Floorplan Generation. Magnetizing Floorplan Generator. Procedia Manufacturing 44: 132–139. Hempel, S., et al. 2015. Generating Early Design Alternatives Based on Formalized Requirements and Geospatial Data. Proceedings of the CIB W (Vol. 78).
409
Hu. R., et al. 2020. Graph2plan: Learning Floorplan Generation From Layout Graphs. ACM Transactions on Graphics (TOG), 39(4): 118–1. Lopes, R. et al. 2010. A Constrained Growth Method for Procedural Floorplan Generation. Proc. 11th Int. Conf. Intell. Games Simul. Merrell, P. et al. 2010. Computer-generated Residential Building Layouts. ACM SIGGRAPH Asia Papers: 1–12. Nauata, N., et al. 2021. House-GAN++: Generative Adversarial Layout Refinement Network Towards Intelligent Computational Agent for Professional Architects. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern recognition: 13632–13641. Schneider, S., et al. 2011. Rethinking Automated Layout Design: Developing A Creative Evolutionary Design Method For The Layout Problem in The Architecture and Urban Design. Design Computing and Cognition’10: 367–386. Springer, Dordrecht. Shekhawat, K., et al. 2020. GPLAN: Computer-generated Dimensioned Floorplans for Given Adjacencies. arXiv preprint arXiv: 2008.01803. Stamm-Teske, W., et al. 2010. Raumpilot Wohnen: 173. Krämer. Wu, W., et al. 2018. MIQP-based Layout Design for Building Interiors. Computer Graphic Forum. Vol. 37. No. 2.
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
Model checking
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Information extraction and NLP for the interpretation of building permits: An Italian case study S. Comai Department of Civil, Architectural, Environmental Engineering and Mathematics, University of Brescia, Brescia, Italy
E. Agrawal Department of Information Engineering and Computer Science (DISI), University of Trento, Trento, Italy
G. Malacarne Fraunhofer Italia Research, Bolzano-Bozen, Italy
M. Sadak Ontopic, Bolzano-Bozen, Italy
S.M. Ventura & A.L.C. Ciribini Department of Civil, Architectural, Environmental Engineering and Mathematics, University of Brescia, Brescia, Italy, Ontopic, Bolzano-Bozen, Italy
ABSTRACT: Assessing project conformity involves extracting information from building regulations and creating rules to verify it. Nowadays, automated systems are already available that assess project conformity but require manual intervention on rule creation, making the process time-consuming and prone to errors. To solve this limitation, this research proposes a new system for extracting information from regulatory codes, which combines the OpenIE6 model with rule-based NLP methods. The articles considered for the search are those containing quantitative values. They were used to train the proposed models to ‘learn’ the context of words in a sentence or document. From the normative articles, triplets (subject-relation-object) were automatically extrapolated and used for the creation of conformity rules. A case study is proposed in which the new data mining technology is applied and the conformity analysis is performed. The research is part of a larger project that aims to make the entire compliance process automatic.
1 INTRODUCTION 1.1 Digital building permit The Architecture, Engineering, and Construction industry (AEC) is going through an era of digital transformation, and numerous efforts by governments and academia have been made in developing new solutions to automate the building process. A key step of the building procedure is verifying design compliance with the regulatory requirements set by municipalities for new constructions and renovations. In this sense, a building permit is the authorisation to start the construction phase of a building project and it is part of a multidisciplinary process of spatial planning for quality and usability together with a sustainable and controlled development of the built environment (Noardo et al. 2022). Building permitting is traditionally a manual, subjective, error-prone and time-consuming process, with a high risk of interpretability, inconsistency in assessments and delays DOI 10.1201/9781003354222-53
in the construction process (Malsane et al. 2015). Moreover, its digitalisation is often mistaken for a dematerialisation of documents as opposed to effective data management. Compliance to building permit requirements within a digital ecosystem could be assured by means of rule-based model checking (Eastman et al. 2009). Compliance automation requires, as first step, the analysis and extraction of the constraints expressed by the municipal codes and the subsequent creation of a set of rules to verify the prerequisites in an information model. Current systems for extracting data from regulatory documents is partially automatic (Niemeijer et al. 2014), as it requires several manual interventions on various fronts, making the extrapolation process long and tedious. In some cases, the Information Extraction (IE) processes turn out to be totally manual (Malsane et al. 2015). Hjelseth and Nisbet (2010) added manual annotations to the text of the building code by adopting a semantic mark-up methodology (i.e, RASE - Requirement, Applicability,
413
Selection, and Exception) and converted them into a form that can be processed by the computer using predefined rules. Zhang and El-Gohary (2013b) and Zhou and El-Gohary (2016) used IE rules developed by experts based on the syntactic information of building code sentences (part-of-speech) and used construction domain ontologies to extract the semantic information elements. This methodology has high performance but, relying on annotations, they are not flexible and scalable enough. Machine learning-based IE methods rely on machine learning models to capture the syntactic and semantic patterns of a text. The application of this method identifies of all the semantic entities of the language such as subject, predicate, requirement, etc., defining the overall meaning of a sentence. Song (2018) proposes the use of Natural Language Process (NLP) to pre-process data and analyse them semantically, but there turn out to be several errors in the analysis of morphemes. In light of the problems that have emerged, the research proposed in this paper presents a system for extracting requirements from a regulatory code that combines the OpenIE6 model with rule-based Natural Language Process (NLP) methods. These neural network-based models are trained using millions of textual inputs from various sources, allowing them to “learn” the context of words in a sentence or document. 1.2 Literature review For several years, the AEC industry has been investigating a methodology that would allow automatic translation of building code articles for compliance verification and building permit issuance. Hjelseth and Nisbet (2010) was the first to define an objective methodology to analyse and extract information from the regulatory regulation. Despite the novelty of the research, the methodology proposed involves the use of manual tagging technology making the process time consuming and still subjective. In order to address these limitations, research has begun to investigate new technologies to analyse regulations automatically. In 2013, the first studies were conducted to extract information from codes using pattern matching based rules and conflict resolution rules (Dimyadi & Amor 2013; Salama & El-Gohary 2013; Zhang & El-Gohary 2013a, b). Through an ontology, the semantic features of the text (concepts and relations) are recognized and, thanks to phrasal tags, the sentence structure, separation and sequencing of semantic information elements are analysed. This process, even though it was well structured, still requires a high level of manual intervention at each step of the methodology. The following year, Niemeijer et al. (2014), investigated the interpretation of building codes by NLP. The text sample, called corpus, is subjected to the part-of-speech tagging process. An approach based on the combination of deterministic grammar with
supervised learning is used and this involves considerable manual work. Dimyadi et al. (2016) introduced the Visual Code Checking Language (VCCL) which is based on the use of a graphical notation to represent the rules defined in the building code in a language that can be read simultaneously by machine and human. Experts in the field are unlikely to have computer skills so using a visual approach will make it easier to set up and understand the rule. Also in this case setting up the rule is still a manual, time-consuming and risky process. The following year, Zhang and El-Gohary (2017) proposed a new system that semantically processes natural language with data-driven EXPRESS techniques to automatically extract and transform regulatory and design information. NLP aims to capture the meaning of sentences to facilitate their full understanding by computers. The normative information extraction and transformation module consists of the normative information algorithm and the normative information transformation algorithm (Zhang & El-Gohary 2012). The information extraction algorithm aims to extract the normative requirements from a normative document into a semantic representation of the information triple. Each triple contains information instances (e.g., “subject”, “compliance control attribute”) and can be defined as Semantic Information Elements (SIEs). However, this process proves to be critical during the information extraction and interpretation activity. In recent years, as reported by Noardo et al. (2022), research investigating the applicability of NLP has grown significantly. Xiaorui et al. (2020) propose the use of the NLP approach based on syntactic and semantic rules combined with a structure inspired by transfer learning trying to reduce POS labelling errors by 82%. The results obtained with these rule-based approaches turn out to be limited, and have led to the exploration of ontology-based methods that provide access to domain-specific semantic information and semantic features. Zhang and El-Gohary (2017a) has contributed by proposing several ontology-based architectures to build text classification, information extraction, fully automatic code checking (Granados Buey et al. 2016). Thus, an additional step being taken in research is the analysis of Deep Learning (DL) technologies since network-based models better understand the context of words. Song et al. (2018) applies the word2vec model to learn the semantics of building code articles and the context in which the word is placed. In such a method, the model initially ‘learn’the word sentence semantics by word embedding techniques, then applied this semantic knowledge in classifying the articles in distinct topics. These models are not ideal for the AEC industry because regulations have complex semantic relationships, so Li et al. (2020) came up with a joint extraction model using a hybrid deep learning algorithm for identifying all subject entities in a sentence, their associated object entities and predicate relations.
414
1.3 Structure of the paper In the following paragraphs, the methodology developed to automate data extraction from building regulations is described. Subsequently, information requirements for the preparation of building models to be submitted for compliance checking are described. This is followed by the case study to which the methodology was applied. Limitations of the research and further considerations for future works are also introduced.
2 RESEARCH METHODOLOGY For this research, the authors have proposed an information extraction approach based on the Iterative Grid Labeling Open Information Extraction (OIE) model that requires minimum manual efforts in order to extract key information from building code articles using semantic understanding. This is combined with the traditional rule-based NLP methods for key information entity extraction and the creation of a structured format for data that can be used to validate the incoming design proposals in the Industry Foundation Classes (IFC) format. Opted for the Iterative Grid Labelling (IGL) based model proposed by Kolluru et al. (2020) in OpenIE6 for generating de-conjugated article sentences and then extracting triplets from these simpler sentences. Finally, the key entities from these triplets are extracted using rule-based NLP and a uniform database is created that can be utilized in downstream tasks for building permit proposal validation. Each step as shown in Figure 1, is discussed in detail in the following sections.
Data pre-processing is crucial for the performance of the models on the texts extracted from the documents. It is important to consider all the constraints around the task of information extraction and to limit the scope of the research around the articles which define building regulations empirically with a quantitative value and a unit of measurement, the articles which come under this category were selected. To facilitate the filtering of non-empirical articles, the first-level selection is done using a simple regular expression. As the text is in the Italian languate, data were analysed using spaCy’s sentence parser before any further processing. spaCy’s is a software that allows the extraction of information on a large scale and extrapolates words or phrases from complex sentences. Normative articles usually consist of several lines of text, so by using spaCy’s it_core_news_ms Italian language model, it was possible to automatically separate articles into sentences. In order to ensure that the selected articles contain information about the regulatory constraints that can be evaluated against the design proposal submitted in the neutral IFC data format, the article data were filtered using ‘quantulum3’, a Python package used for quantity and unit of measurement identification extraction. Furthermore, the articles with only quantitative numeric entities with proper units of measurements have been selected for further processing and building plan validation. Another important step in data processing is the translation of Italian text to English. The reasoning behind the translation to English is: – the OpenIE6 model used in the implementation is pre-trained using Bi-directional Encoder Representations from Transformers (BERT) models which are trained in the English language texts; – to make the proposed approach language agnostic and allow building codes from other languages also utilise it for information extraction. The BERT model captures each word having its own representation, more appropriately called embedding. Each layer performs an attention calculation on the embedding of the words of the previous layer, and then creates a new intermediate embedding with the same size as the embeddings produced by the previous layers. 2.2 Conjunctive extraction
Figure 1. Phases and sub-phases of the research.
2.1 Data pre-processing The primary and most important step in any NLPbased task is the pre-processing of the textual data. Data collected from the source mostly come in a raw and unstructured format which needs to be prepared in a format that can be processed by the algorithms, for example, information extraction models in this case.
There are multiple parameters ranging from the complexity level of building code articles to the length of the sentences that have a lot of correlated, conjunctive arguments that make the extraction of all the key information elements from these article sentences difficult. Therefore, a dedicated step for conjunctive relation extraction has been included in the model pipeline. In this step, the conjunctive coordination analyser presented in OpenIE6 was used. The coordination analyser is built on the Iteractive Grid Labeling (IGL) system as proposed by Kolluru et al. (2020). IGL works
415
on a 2-D grid structure with M × N rows and columns. They have defined the coordination analysis as a hierarchical labeling task where all the conjunctive words are predicted at the same level of the hierarchy. To explain this in an intuitive way, the M × N grid can be described as M being the maximum depth of the hierarchical levels and N will be the number of words in each sentence. At each iteration level, the model assigns a label to each word which could be either a CC (coordinated conjunction), CONJ (belonging to a conjunct span), or SEP (Separator), or N (None). 2.3 Triple extraction In this step, the simplified sentences are processed through an Iterative Grid Labeling based Open Information Extraction model (OpenIE6) to obtain the ‘subject;relation;object’triplets. In the context of OpenIE6, information extraction can be defined with the help of the below-given equations. For a given sentence S, which is constituted of multiple words wi , S = {w1 , w2 , w3 . . . , wN }
(1)
the objective of OpenIE is to extract a set of extraction triplets in the form of subject, object and relation. E = {E1 , E2 , E3 , . . . , EM }
(2)
where, Ei = {subject; relation; object}
(3)
and each subject, relation and object is constituted of wi s. The Iterative Grid Labeling technique developed by Kolluru et al. (2020) assigns a label to each word in the sentence which is later used to extract the triples from the sentence. The model is devised on a BERT and Self-attention transformer layers based architecture which starts with generating the word embeddings of the words in the input sentence through a BERT encoder. It captures the context of usage of the word and assigns each word a context-based embedding. While feeding the data to BERT, some additional tokens like ‘is’, ‘of ’, ‘from’ are added to aid the token prediction by OpenIE. These contextualised embeddings are then passed on to self-attention transformer layers where it iteratively generates label embeddings for each word embedding. At each iteration level, these label embeddings are fed to a fully-connected label classifier, which predicts a tag for each token. There are four labels (S)ubject, (R)elation, (O)bject and (N)one which are predicted for each word iteratively by the classifier. Thus the predicted label is passed through a label embedding module which generates embeddings for the predicted labels. To capture the dependencies among the predicted labels and contextual information on the input these label embeddings are added back to the previous output from the self-attention transformer layer. Therefore, when the
next level of extraction happens the information label prediction is added to the embeddings, making the dependencies among extractions captured. Finally, a softmax layer assigns probabilities to each predicted label at every iteration and the label with maximum probability is extracted by taking the argmax of these probabilities. 2.4 Entity database creation For the final task, the triplets extracted in the form of a tuple of ‘subject’, ‘relation’ and ‘object’ are further segregated in key entities that can be stored in a structured format to create a database for IFC proposal validation. The first part of the triple, which refers to subject entities, is usually composed of the main entity of the sentence. Whereas, the relation part of the triple contains information in the form of an action which is related to the subject. Similarly, the last part of the triple has information on the object towards which the action is performed. Therefore, as a first level approach a basic text POStagging is conducted on these triples. In this step, a Natural Language ToolKit (NLTK), python library that enables text analysis, based POS-tagger was utilized which assigns each element of the triples a tag based on the grammatical structure. These tags are either a noun, verb, adverb or adjective etc. As the second step, the unit of measurement conversion was performed to create uniform units in the extracted entities. It is observed that there are instances where a unit of measurement is expressed in a textual form in a given sentence which gets separated into a triplet as it is. For example, ‘meters’instead of ‘m’or ‘square meters’ instead of ‘m2’. Therefore, all the units are converted into standard SI units of measurement system by using the ‘quantulum3’ package. For addressing the quantitative relations, in the second and third part of the triple, a dictionary-based approach was used to identify the relative quantifiers. The negative conditional arguments such as ‘not’, “never”, “doesn’t”, “isn’t”, “must not”, “shouldn’t” etc. are put together in a dictionary to identify them in the triples and further match it with the quantitative comparator words like ‘smaller’, ‘lower’, ‘less’, ‘more’, ’exceed’, ‘protrude’, ‘pass’, ‘least’, ‘minimum’. These dictionaries are defined in order to assign a proper mathematical comparator notation to the rule which is required to validate the information in the building proposal. Hence, the words were assigned to a combination of six attributes that seem most important for expressing the building code rule properly so that it can be further used to validate the data from the IFC file. These six attributes are “Subject”, “Compliance checking attribute”, “Comparative element”, “Quantity”, “Unit”, “Relative restrictive element”. 2.5 Building model preparation The drafting of an information specification is a fundamental phase for data management and to define
416
the information requirements necessary for the building model preparation (Eastman et al. 2009). The definition of information requirement allows to optimising the control of the constraints expressed in the regulatory codes. In the language to indicate a building element or an object there are more synonyms, and this confuses the univocal identification of an object. This might lead to difficulties in the generation of univocal rules to verify specific requirements. In order to solve this problem, it has been necessary to establish a well-defined nomenclature of the elements to be verified. Figure 2, shows an extract of the classification defined for the case study described in this paper. The basic data that must be present in the information specification table are: – the object considered; – the IFC class to which the object belongs; – the typology of classified object; – nomenclature/classification adopted.
– to support public administrations in facing the ongoing digital revolution by providing innovative software applications; – to demonstrate the positive impact on the daily work of designers and companies working with public administrations. One of the GEOBIMM expected results is the GEOBIMM4BP, a prototype application that will demonstrate the potential of the integration between BIM and GIS in making the building permit process faster and more efficient. The GEOBIMM4BP has two main perspectives: (1) a process-based perspective by means of a digital process supported by BIM models and (2) a rule-checking perspective by means of BIM-GIS automatic checks of building permit requirements. The case study described within this paper refers to the approach selected to perform research on the second perspective.
3.2 The GEOBIMM4BP prototype application
Figure 2. Example of object nomenclature in the information model.
In addition, one column has been added for possible specifications/notes that can be added to the object name. Therefore, the name of a building element is composed by two parts: the nomenclature established a priori and a customised additional specification. The need to add the specifications (i.e., fifth column in the Figure 2) arose because some premises may have different rules to respect depending on the generic object or specific type. These specifications will have to be shared with the assigned professional who will have to create the information model of the building under construction or renovation by correctly naming the constructive elements.
The GEOBIMM4BP prototype application works is illustrated in Figure 3 and is described as follows: the final user (i.e. a Municipality Officer) will upload the BIM model of a building, georeference it, and display it in context. When launching the checking system, the BIM model will be queried to automatically extract the data required to perform the checking against the requirements defined within the building permit process and, in particular, with those listed into the building code of the Municipality. Subsequently, the data extracted from the BIM model will be compared with those required by the building code. In the meantime, the requirements listed in the building regulations of the Municipality will be extracted using
3 CASE OF STUDY A case of study has been developed as a way to test the proposed methodology. The case study is part of the Geographic Building Information Model Management (GEOBIMM) research project, which is following introduced.
3.1 The GEOBIMM project The GEOBIMM project has three main objectives: – to make the process of managing a public work more efficient by reducing the execution time of technical processes for a faster, more transparent and more aware public administration;
Figure 3.
417
The GEOBIMM4BP process map.
the Open Information extraction pipeline shown in Figure 1. The building regulation of the Municipality of Merano was selected as case study. The Municipally of Merano, a small town located in the North of Italy and associated partner of the GEOBIMM project, supported the research team providing the building code and explaining the building permit process, as described in Fauth et al. (2022). 3.3 Information extraction of the building regulation The OpenIE model implementation was divided into three major steps. After (1) the data were collected and pre-processed, (2) they were used by the downstream OpenIE models for Conjunctive extraction and (3) triplet extraction. From the Merano building regulations, 49 articles were selected, which created a dataset of 107 single sentences after the data preprocessing phase. The sections of the building regulations used for this purpose are: ‘Technical Standards Requirements for Hygienic-Sanitary Content’, ‘Architecture; Landscape Green cover’, ‘Rules for Building Construction activities’, ‘Obligatory Building Permit requirements’, ‘Protection of public safety, the appearance and decor of the built-up areas and public streets’, ‘Environmental and landscape protection’ and ‘Maintenance Standards’. In the subsequent step, the data were tokenised using two separate tokenisers. First, the tokenisation was performed using the NLTK library and the output tokens were stored separately in the model. Next, BERT based tokeniser was used for tokenising the input text. Since BERT is known to create multi-token embeddings for a word, the NLTK tokenisation was used to retrieve the original token count of a sentence. These BERT embeddings were then fed to Iteractive Grid Labeling based on conjunctive coordination extraction model. The OpenEI6 model iteratively labels each word with one of the predefined conjunctive labels. The conjunctive labels were used in the implementation are ‘CP_START’, ‘CC’, ‘CP’, ‘SEP’ and ‘None’, where CP represents a Conjunctive Phrase, CC represents a Coordinating Conjunction and ‘SEP’ represents separators like a comma. Each of these labels were given a numeric representation and the IGL-based Conjunctive extraction model labels each word in the sentence with one of these given labels. The OpenIE6 model was used to extract for conjunctive coordination analysis using “BERT-largecased” pre-trained weights. One of the important hyper parameters used for conjunctive models is the maximum number of iterations in the IGL process, which is set to M=3 for conjunctive extraction. The simple sentences extracted from the conjunctive extractor contain the information on the subject entity relations. To extract these values, the Triplet extraction model, which extracts tuples from these simple sentences was used.
This OpenIE6 Interactive Grid Labeling extraction model uses the pre-trained weights of “BERT-basecased”. This BERT model also utilises the masked language modelling method combined with Next Sentence Prediction, it contains 12 encoder layers and generates 768 hidden dimensions for each input word. The hyper parameter for maximum number of iterations in the IGL part was set to M=5. This is done to ensure the maximum possible labels for a word are predicted for each input sentence and finally the label with the highest probability score is assigned to the word four main labels used in this process are: namely subject(S), object(O), relation(R) and none(N). Finally, from the extracted triplets, a general pattern of information representation was developed. From the ‘subject’ part of the triplet, the object and the compliance control attributes were extracted using the POS tagging combined with dependency parsing techniques. The ‘relation’ part of the triple generally represents the obligatory definition of the implied rule, such as ‘shall be’, ‘may not be’, ‘must have’ etc. 3.4 Entity database creation Following the extraction of the properties and the values associated with them, a manual verification of the conformity of the IFC model of the building in question was carried out. As a first step, the objects in the model were classified and then the rules for verifying the presence of the property or value (e.g., height, width, distance, etc.) were set up control rules were set up by carrying out an initial classification of the objects (e.g., rooms, sanitary facilities) present in the model and then setting the rule. The previously set information specification allowed the creation of structured rulesets specific to each element. 4 DISCUSSIONS In this research, 1514 items were used to train data mining models. After teaching the algorithm sentence composition and identifying the grammatical parts of the sentence, it was possible to proceed with the case study. 49 articles were selected and were preprocessed. From this phase, 107 useful sentences were generated. From these sentences, 305 simple sentences were extracted using conjunctive extraction. Some of these were not correctly separated, so the total of correct conjunctive extraction is 291 sentences. The 291 sentences were then analysed through the triplet extractor which generated 291 triplets. After a check, 279 triplets were found to be correctly extrapolated. Thus, it is possible to evaluate the extrapolations performed against three metrics: recall, precision and F-measure (Li et al. 2020). Recall, means the ratio of the correct extrapolations to the total number of starting extrapolations (4). R=
418
number of correct extractions total number of extractions existing in the data (4)
Precision refers to the ratio of triplet extraction to entity extraction (5). P=
number of correct extractions total number of extractions
(5)
F-measure means the combination of precision and recall with the addition of a parameter β to assign value weights (6). Where β = 1. F − measure =
(β2 + 1)PR (β2 P) + R
(6)
Metrics
R
P
F-measure
Conjunctive extraction Triplet extraction
0.935 0.897
0.954 0.958
0.944 0.926
5 CONCLUSIONS 5.1 Results This research aimed to investigate the application of Natural Language Processing (NLP) and Open Information Extraction (OpenIE) methods to extract information from building codes and represent it in a structured form. as a first step, it was necessary to pre-process the data contained in the building regulations. Using a regular expression, the building codes were divided into empirical (quantitative) and nonempirical articles. For the research, the quantitative articles were taken into consideration. Subsequently, each article was automatically divided into separate rows. Each row was filtered using quantulum3 to extrapolate the quantity to be verified and the unit of measurement. In addition, each word making up the sentence was grammatically labelled. This labelling enables the extrapolation of triplets in the form of subject, object and relation using the Open IE6 model. the triplets extracted in the form of a tuple of ‘subject’, ‘relation’ and ‘object’ are further segregated in key entities that can be stored in a structured format to create a database for IFC proposal validation. To verify the conformity of an IT model, it was necessary to define the information content and nomenclature of the elements to be verified.
of all, quantitative articles from the building regulations of several Italian municipalities were selected manually. Moreover, the scope of the implemented models has been limited to the articles containing quantitative information about the entities. The extraction of properties and values for building compliance was done automatically while the compliance of the IFC model, taken as a case study, was done manually. 5.3 Future works In future research, it would be useful to expand the pool of regulatory articles taken into consideration, both including further building code section and qualitative requirements. By addressing the dataset size issue, the model performance can be evaluated in a better way and the required refinements can be introduced into the proposed architecture. Also, due to the unavailability of labelled dataset, the model has been evaluated by manually verifying the extracted data. Although the performed well given the circumstances, it is required to perform evaluations using labelled data to confirm the usability of the models. In addition, the validation of the presence of an object within the information model and its associated property was done manually. In the future, it is planned to implement the proposed model for the validation of design proposals within an automated platform.
ACKNOWLEDGEMENTS The proposed case study is part of the GEOBIMM project. The GEOBIMM project has been financed by the European Regional Development Fund of the Autonomous Province of Bolzano – Investments in growth and employment ERDF 2014–2020. REFERENCES
5.2 Limits of the research The scope of a research like thus us extremely vast, and so is the complex nature of building code regulations which makes the research conducted to validate the applicability limited to keep things under control. However, this opens up opportunities for further explorations, addressing the complex issues and possibilities to apply futuristic techniques to these problems. There have been several assumptions made during the implementation and research of this project. First
419
Dimyadi, J., Solihin, W., Preidel, C. & Borrmann, A. 2016. Towards Code Compliance Checking on The Basis of A Visual Programming Language, Journal of Information Technology in Construction, 21: 402–421. Dimyadi, J. & Amor, R., 2013. Computer-assisted Regulatory Compliance Checking. Proceedings of the 12th New Zealand Computer Science Research Student Conference (NZCSRSC). Eastman, C., Lee, J.M., Jeong, Y.S., & Lee, J.K. 2009. Automatic Rule-based Checking of Building Designs. Automation in Construction, 18(8): 1011–1033. Fauth, J., Malacarne, G. & Marcher, C. 2022. Digitalization of the Building Permit Process - a Case Study. Proceedings of the CIB World Building Congress. Granados Buey, M., Garrido, A.L., Bobed, C. & Ilarri, S. 2016. The ais project: Boosting Information Extraction From Legal Documents By Using Ontologies. Proceedings of the 8th International Conference on Agents and Artificial Intelligence (ICAART), 2: 438–445. Hjelseth, E. & Nisbet, N. 2010. Exploring Semantic Based Model Checking. Proceedings of the 2010 27th CIB W78 International Conference.
Kolluru, K., Adlakha, V., Aggarwal, S., Mausam, & Chakrabarti, S. 2020. Openie6: Iterative Grid Labeling and Coordination Analysis for Open Information Extraction. Computation and Language. Li, F., Song, Y. & Shan, Y. (2020). Joint Extraction of Multiple Relations and Entities From Building Code Clauses. Applied Sciences, 10(20):7103. Malsane, S., Matthews, J., Lockley, S., Love, P.E., & Greenwood, D. 2015. Development of An Object Model For Automated Compliance Checking. Automation in Construction, 49: 51–58. Niemeijer, R.A., De Vries, B. & Beetz, J. 2014. Freedom Through Constraints: User-oriented Architectural Design. Advanced Engineering Informatics 28(1): 28–36. Noardo, F., Guler, D., Fauth, J., Malacarne, G., Mastrolembo Ventura, S., Azenha, M., Olsson P. & Senger, L. 2022. Unveiling the Actual Progress of Digital Building Permit: Getting Awareness Through a Critical State of the Art Review. Building and Environment, 213. OpenIE6: https://github.com/dair-iitd/openie6 Salama, D.M. and El-Gohary, N.M. 2013. Semantic Modelling For Automated Compliance Checking. Computing in Civil Engineering, pp. 641–648. Song, Y., Kim, J. Lee, J.K. 2018. NLP and Deep Learningbased Analysis of Building Regulations to Support Automated Rule Checking System. Proceedings of the 35th
420
International Symposium on Automation and Robotics in Construction (ISARC), pp. 586–592. Xiaorui, X. & Zhang, J. 2020. Building Codes Part-ofspeech Tagging Performance Improvement By Errordriven Transformational Rules. Journal of Computing in Civil Engineering, 34(5): 04020035. Zhang, J. and El-Gohary, N.M. 2012. Extraction of Construction Regulatory Requirements From Textual Documents Using Natural Language Processing Techniques. Computing in Civil Engineering, 453–460. Zhang, J. & El-Gohary, N.M. 2013. Semantic NLP-based Information Extraction From Construction Regulatory Documents For Automated Compliance Checking. Journal of Computing in Civil Engineering, 30(2): 04015014. Zhang, J. & El-Gohary, N.M. 2013. Information Transformation and Automated Reasoning For Automated Compliance Checking in Construction. ASCE International Workshop on Computing in Civil Engineering. Zhang, J. & El-Gohary, N.M. 2017. Integrating Semantic Nlp and Logic Reasoning Into a Unified System For Fullyautomated Code Checking. Automation in Construction, 73: 45–57. Zhou, P. & El-Gohary, N.M. 2016. Ontology-based Automated Information Extraction From Building Energy Conservation Codes. Automation Construction, 74: 103–117.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Process-based building permit review – A knowledge engineering approach J. Fauth & W. Müller RIB Information Technologies AG, Stuttgart, Germany
S. Seiß Bauhaus-Universität Weimar, Weimar, Germany
ABSTRACT: Digitalization and applications of intelligent systems in theAEC sector increasingly demand data. Required data is often not accessible but necessarily needed to interpret information and finally to transform into knowledge. For this reason, suitable data must be collected, analyzed, and formalized, before data can be used to feed systems in regard to automation. In scientific practice, the problem of how data is to be conducted for knowledge representation appears frequently. Using the example of building permit reviews, this paper presents qualitative expert interviews as a convenient method in terms of knowledge engineering. Subsequently, obtained data is used to develop an ontology representing building permit authorities. The approach is validated by automating the subprocess of participation of agencies of public interest during the building permit review. The approach represents a solid method to generate data as well as demonstrates transferring knowledge in an ontology through rules and queries.
1 INTRODUCTION In the course of digitization and data use, research repeatedly encounters the problem that no suitable data is available or is not provided by practice. However, in order to investigate expert knowledge for scientific approaches, it is necessary to evaluate data. If data is not available, it needs to be generated. Possibilities of data acquisition offer empirical studies, including qualitative data collection. Although qualitative data collection is underrepresented in the AEC sector (Simon 2019), it is an important source of data to collect implicit expert knowledge. However, missing data cannot be used to formalize explicit knowledge and thus cannot be implemented for expert systems. The problem of missing data also exists exemplarily for building permit process issues. However, approaches are not comprehensive as they mainly focus on content reviews in terms of automated code compliance checking (e.g., Eastman 2009; Nawari 2018; Zhang & El-Gohary 2019). Ontologies are widely used for knowledge representation in different disciplines. Hinkelmann et al. 2010 developed an ontology in the context of e-government using the example of building permitting and considering formal aspects in the process. In the AEC industry, many ontologies have been developed over the past years. For example, Pauwels & Terkaj (2016) developed the ifcOWL based on the express schema of the ifc to describe building information models by semantic web standards. The DiCon ontology is published by Törmä (2021). DiCon provides an ontology DOI 10.1201/9781003354222-54
to describe construction systems. In addition, specialized ontologies for building code compliance checking are created. Li et al. 2021 developed an ontology to check railways for compliance. Bouzidi et al. (2012) developed a code checking ontology for French building codes based on owl and semantic query language (SPARQL). Pauwels et al. 2015 investigated different methods for code compliance checking in the AEC industry by semantic web standards. By now, developed ontologies neglect the overall view of building permit process and focus exclusively on a regulatory check of building applications. Therefore, it is necessary to develop an ontology which is able to handle building permit processes of a building permit authority and to align existing building code compliance checking ontologies. This paper aims to describe a knowledge engineering approach how explicit knowledge can be generated from qualitative data. Therefore, the example of the participation process as a subprocess of the building permit process is used. The participation subprocess describes the involvement of agencies of public interest (AoPI) involved in the building permit review. For this purpose, qualitatively collected and analyzed data material containing implicit knowledge is transformed into explicit knowledge. Transcriptions of qualitative expert interviews and their analyses are used as main data sets. The interviews were conducted in several international studies in previous research to investigate the building permit process. This research (1) introduces qualitative expert interviews as an appropriate method to generate data for AEC sector, using
421
building permit processes as an example.This will provide incentives to use this method for other cases as well. Furthermore, this approach (2) results in a generalized semantic model of the participation subprocess during a building permit review. The proposed system can be used as a basis for further developments and comparisons between other countries. The remainder of this paper is structured as follows: Section 2 clarifies the understanding of knowledge engineering and developments of ontologies. Section 3 gives insights on the research methodology, while Section 4 introduces the implementation and validation part of the approach. The paper closes with a summary and conclusion in Section 5.
2.2 Ontologies and semantic web Ontologies describe a conceptual framework of knowledge in a specified domain. Therefore, ontologies represent a part of the real world in a semantic model. (Synak et al. 2009). The semantic model of an ontology is described by semantic web languages. The basic representation language of the semantic web is the Resource Description Framework (RDF). RDF graphs organize data in triples of subject, predicate and object. Figure 2 illustrates such a triple, in which subjects and objects are drawn as nodes, and the predicates are drawn as edges between the nodes. Therefore, RDF graphs are called directed and labelled triple stores (Hitzler et al. 2007, W3 2014).
2 BACKGROUND 2.1 Implicit and explicit knowledge In order to create a knowledge-based system, structured information about the subject is required. As Figure 1 illustrates models serve this purpose by providing a structured representation of a system, which itself abstracts reality. Therefore, models are consequently a system themselves and an abstraction of reality. Models can also represent systems in each granularity, like a reduced system of a more complex system (Schneeweiß 1991). Knowledge can be separated into explicit and implicit knowledge. Explicit knowledge can be easily formalized and transmitted between individuals. Implicit knowledge can be seen as personal experience knowledge and is therefore hard to formalize. (Polany 1983) explained implicit knowledge as “we can know more than we can tell”. Explicit and implicit knowledge as well as users and organizations interact together and create highly valuable organizational knowledge. (Nonaka & Takeuchi 1995). Explicit knowledge is based on implicit knowledge (Sveiby 1997). In order to implement implicit knowledge into an expert system, it must be acquired. Activities for acquiring knowledge is named knowledge acquisition. Besides knowledge integration and maintenance of the knowledge base, knowledge acquisition is part of knowledge engineering. According to Karbach & Linster (1990) and Bimazubute (2005), knowledge engineering refers to the development of expert systems. It becomes clear that acquiring and formalizing implicit knowledge into explicit knowledge is crucial for (rule-based) expert systems. Empirical methods like interviews can be used to acquire and formalize implicit knowledge to integrate this knowledge into a rule-based system (Horvath 2000).
Figure 2. RDF triples.
RDF is extended by more expressive languages in form of Resource Description Framework Schema (RDF-S) and Ontology Web Language (OWL). RDFS for example, extends the schema of RDF by the definition of classes and properties. OWL enables the definition of constraints between classes, properties, and entities (Allemang et al. 2020). OWL uses RDFS to describe ontologies. Generally, knowledge of ontologies is separated into an assertion box (A-Box) and terminology box (T-Box). Semantic web applications relying on a dedicated rule language can be extended by a rule box (R-Box). The T-Box contains intentional knowledge (terminology) and describes a given domain by vocabulary like a database schema in the form of classes, properties, and relations. The described concept of the T-Box will not change over time change over time. The rules are stored in a separated R-Box and are executed by an inference engine in combination with A- and T-Box (Pauwels et al 2015, Baader et al. 2017). An overview is illustrated in Figure 3. The A-Box represents the facts associated with the vocabulary terms of the T-Box. Therefore, the A-Box instances defined classes with individuals of the real world. The A-Box contains extensional knowledge, the knowledge about a specific situation, which will. 3 RESEARCH METHODOLOGY This approach pursues the objective of describing the path from data collection to knowledge representation
Figure 1. Abstraction and representation in system theoretical context (Fauth 2021, Schneeweiß 1991).
422
and the integrated rules describe and structure the participation subprocess and related information like the building application or authorities.
Figure 3. Overview implementation path for any rule-checking application relying on a dedicated rule language. (Pauwels et al. 2015).
designated knowledge acquisition as part of a knowledge engineering approach. Figure 4 summarizes the procedure of explicit knowledge generation using qualitative expert interviews as a source of implicit knowledge. Since no suitable data is available at the initial stage of this study, it was generated in the form of qualitative expert interviews. With the help of a semi-structured interview guideline, the interviews are recorded and transcribed. Scientific transcriptions of the data material are suitable for further analyses (e.g., with special software) and serve traceability of the analyses. In this status, the data material contains implicit knowledge. Qualitative content analysis transforms the data into information. The analysis results are interpretations, which can be described as explicit knowledge at this stage. Subsequent formalizations of the results provide knowledge representation and knowledge base. Depending on the purpose of the investigation, different modeling and notation languages are suitable for the formalization. In this approach, Business Process and Notation (BPMN) and Decision Modeling Notation (DMN) languages are used. Based on formalized knowledge, the proposal for the solution of the problem follows - here the development of an ontology together with rules. Finally, the developed system is able to represent and infer knowledge as well as to visualize the information accordingly. The knowledge acquisition process presented here is based on the example of the participation subprocess of building permit authorities involved in the building permit process and is described in detail in Section 4. Specifically, it is about supporting the building officials in their decision, based on cases, which other AoPI have to be involved for a certain project. Possible AoPI and specific motivations to involve them are extremely diverse. Therefore, this paper examines three exemplary AoPI in more detail. The ontology
Figure 4. Overview of knowledge acquisition as research methodology.
4 IMPLEMENTATION AND VALIDATION In preliminary studies (e.g. Fauth 2021), building permit authorities were interviewed regarding the participation subprocess. On the basis of this data material and generated knowledge, the participation process in building permit authorities is analyzed and formalized first. During a building permit review, various agencies of public interest (AoPI) need to be involved. Involvement means that all AoPI are requested to submit a statement regarding their area of expertise. This process is named participation and is proceeded by the building permit authority (in this case). The term “AoPI” refers to all authorities and agencies that must be or can be legally involved in any way in a planned construction project. In addition to AoPI, other institutions and experts are also commissioned to provide opinions or evidences. For simplification in this study, other institutions and experts to involved in the building permit review process are grouped with AoPI together since the participation subprocess is similar. Interview results show that participation is individual processed by building authorities. For example, it is common practice for some authorities to always commission certain AoPI, regardless of the proposed project. It was also found, that inexperienced or nonspecialist staff in particular involve as many AoPI as possible to avoid any mistakes in the participation process. This unnecessarily inflates the process and leads to possible extensions of the review process. Other authorities ask AoPI whether they see the need to be involved. This also requires additional preliminary review of the documents and can lengthen the process.
423
Due to the individuality in each building permit authority, the interview results have to be abstracted into a system and represented in a BPMN (ISO/IEC 19510 2013). Accordingly, the participation subprocess is divided into mandatory Selection (P1) and optional selection (P4). Subsequently, the individual and specific AoPI are selected and assigned (P2 & P5). A decision is to be made if an AoPI needs to be involved (refers to P1) in the review, and if yes, which specific AoPI is to be involved and assigned (e.g., specialist authority in a specific municipality) (refers to P2). Even if it is not necessary to selectAoPI, participation can still be optional and additional (P4 & P5). The process steps, including the process steps after selections (P3, P6 & P7), are illustrated in Figure 5. Due to the large number of AoPI, there is a problem in practice considering every mandatory AoPI for participation, represented by P1. The proposed approach aims to develop a system which supports the decision to be made (in P1) automatically based on requirements and rules. Therefore, it is presumed that the selection of the mandatory AoPI depends on the respective construction project proposed in a building application. Additionally, the developed system considers the selection of optional AoPI (P4) as well. Since decisions and their influencing parameters cannot be represented within a BPMN model, DMN is used to establish the link between business modeling and decision logic and can be used for rule implementation (OMG 2021). Table 1 shows a simplified decision table in the sense of DMN for the participation of (a) nature conservation authority, (b) historic preservation authority, and (c) proofing structural engineering office. The specifications in the table are assumptions. In addition, the requirements represent only a small part of the possible requirements and parameters for involving AoPI. Based on the defined process requirements in BPMN and DMN an ontology is created. The ontology provides a schema (T-BOX) and rules (R-Box) to describe the necessary data for the building permit process. The illustration of the ontology focus on the relevant elements of the building authorities participation subprocess and neglected further ontology content. The ontology is implemented by the owl schema to provide basic logic like the definition of subclasses and possible extensionality by import functionalities. Figure 6 illustrates the major classes and
Figure 5.
relations of the ontology for building permit authorities (OBPA). The classes are illustrated as rectangles, and arrows illustrate the directed relations. As illustrated in Figure 6, the classes are related to a specific domain to provide a better overview and understanding of the ontology. The domain of the building application describes the building application as well as the construction project. The building application is processed by examiners. The class examiner is a subclass of person. The building official, the public officer and the proofing structural engineer represent different kinds of examiners and are subclasses of the class examiner. All examiners are employed by a specific authority. In OBPA, building officials work for a building permit authority, and public officers work for nature conservation and historic preservation authorities. Proofing structural engineers are employed in proofing structural engineering offices. OBPA contains further classes, which cannot be illustrated due to the limited space. Further extension of the ontology is possible by aligning ontologies like the Building Ontology Topology (BOT) ontology by owl: import to import complete ontologies or direct references like foaf:person to use specific classes. BOT, for example, is a lightweight ontology to describe buildings and can be used to describe the referenced construction project in more detail (https://w3c-lbd-cg.github.io/bot/). Table 1.
DMN logic to select AoPI.
Input
Output
Historical building
Distance to nature conservation
Gross floor area (GFA)
a) true
-
-
a) false b) -
>=50m
-
b) c) -
1000m2
c) -
-
fun( greater than 5 Baseline: Adjust tokenisation of schema terms to decrease the token length: expr -> expression, fun -> function ... 6 Higher standard deviations for BLEU scores due to selecting the best run according to F1-Scores.
4.2 LRML normalisation experiments We investigate the LRML normalisation in the lowest possible granularity. The clause alignment (Step 1) can be programmatically separated from the added implicit information (Step 2). Since Steps 3-5 were conducted in sequence, there is no clear separation of these steps leading to a combined investigation. Fixing all LRML converter warnings increased the number of LRML rules. We excluded these additional rules from the initial comparisons and added them in the last experiment to allow fair comparison. Table 2 shows the high impact of adjusting and re-generating the LRML rules, which led to 8.5% Table 2.
4 EVALUATING THE LRML CORPUS
Establishing the T5-AMR baseline.
Improving results with data normalisation.
Setting
Size+
BLEU*
F1-Score*
Baseline + Step 1 + Step 2 + Step 3-5 + References + Entities + Add. Data
599 (542/57) 599 (542/57) 599 (542/57) 5881 (534/54) 588 (534/54) 588 (534/54) 760 (704/56)
36.8% (2.5) 37.6% (2.3) 38.2% (1.9) 58.5% (4.2) 55.0% (5.8) 59.7% (3.2) 60.7% (3.4)
35.0% (0.4) 37.6% (0.4) 37.5% (0.3) 46.0% (0.9) 46.4% (0.7) 46.5% (0.8) 48.0% (0.8)
4.1 General setup and baseline enhancements The following experiments use the setup described in Section 2.2. T5-AMR is trained using the Huggingface library, the T5-base tokeniser, and the following hyperparameters: Batch size: 4, Learning rate: 2e-4, Weight decay: 1e-4, Warmup steps: 100, Epochs: 30, Early stopping patience: 5, Beam size: 5 and Others: Default. We run each experiment with three random seeds and report the average validation F1-Score and BLEUScore with associated standard deviations from the epochs with highest F1-Scores.
432
+ Number of samples: Dataset (Training/Validation set) * Standard deviation in brackets 1 Some rules were merged to avoid duplication.
rise in F1-Score. In contrast, adjusting the alignment, normalising references and entities, and adding rules caused less progress, and auxiliary input of implicit knowledge degraded the results. 4.3 Value conditioning According to Section 4.2, the semantic changes to the LRML rules had the highest impact on performance. Since these changes require high manual effort, we investigate value conditioning, a method of inputting additional information about an LRML rule, as an alternative to fixing the issues. This technique is closely related to prompt engineering and prefix-tuning, two popular techniques to fine-tune and improve language models (Li & Liang 2021). We appended eight different tags (see Table 3) about complex LRML constructs and encoding issues to the input text and tested the impact. As a control group, we deleted the tagged rules from the dataset. We expected the parsing performance to rise or stay similar if the removed rules are too complex or introduce too much noise. Otherwise, the scores should drop because of the decreased dataset size. The results in Table 3 suggest that the NSP has no benefit in knowing details about the structure of the rules. To the contrary, concatenating additional information consistently worsened the model’s F1-Scores. The additional information or grouping of certain LRML rules seems to confuse the NSP. Also, there is no clear correlation between removing rules and adding the tag. So, a re-evaluation after improving the NSP performance will be considered to test if these
results are caused by existing problems in the NSP or the ineffectiveness of the method. Nevertheless, removing selected samples indicated that translating clauses using abstractions and implicit knowledge and leaving out irrelevant information is especially problematic for automation. Also, removing the clauses containing the key-construct brought improvements. For an explanation, we investigated the files mostly containing this encoding style. 33 out of the 114 samples are from the documents B1/AS1 and C/AS1, which have the worst parsing performance, according to Table 4. Table 4.
Removed*
Setting
Size+
BLEU F1-Score BLEU F1-Score
Baseline Ignore1 Implicit2 LOD3 Medium3 Coarse3 Loop4 Define5 Key6 Document7
704/56 126/22 138/15 704/56 205/22 112/6 24/2 126/1 114/4 704/56
60.7% 49.0% 60.7% 56.6% 60.0% 52.3% 59.8% 59.1% 54.9% 59.6%
+
48.0% 43.5% 46.6% 45.0% 45.2% 45.9% 46.1% 46.0% 45.9% 46.5%
62.2% 48.8% 61.1% 49.5% 57.5% 58.0% 55.4% 58.3% 58.5%
48.7% 47.7% 47.1% 46.4% 48.8%
Affected rules in training/validation set * The standard deviation was removed due to space limitation 1 Some parts of the regulatory statement were not encoded. 2 The LRML rule contains unknown information. 3 Level of detail – Refers mostly to the number of concepts in entities. Fine-grained, medium-grained, and coarse-grained. 4 The LRML rule contains a loop or similar construct. 5 The LRML rule contains variable definitions. 6 Reference encoded with a document and number expression. 7 Input the AS specifier (e.g. b1as1) to learn about implicit references or document specific translation styles.
Document
Size+
BLEU*
F1-Score*
Baseline1 B1/AS1 C/AS2 Others2 G14/VM13 D1/AS13 B2/AS13 E2/AS14 G15/AS1 G12/AS2 E1/AS1 B1/AS3 G13/AS1 G13/AS2 G12/AS1 Random4/5 Random5 full
574 (518/56) 580 (518/62) 573 (518/55) 529 (518/11) 556 (518/38) 528 (518/10) 526 (518/8) 760 (518/242) 539 (518/21) 573 (518/55) 579 (518/61) 578 (518/60) 550 (518/32) 571 (518/53) 570 (518/52) 600 (518/82) 760 (684/76)
58.8% (6.4) 52.9% (1.0) 59.4% (1.4) 48.2% (9.5) 65.3% (0.6) 66.7% (1.0) 55.9% (8.5) 58.6% (3.7) 64.8% (0.4) 59.5% (4.4) 61.1% (3.3) 61.4% (3.1) 64.1% (2.0) 67.5% (3.2) 68.4% (2.5) 73.9% (2.5) 74.2% (3.7)
46.0% (0.5) 40.7% (1.3) 42.8% (0.6) 47.2% (1.4) 48.2% (1.6) 49.2% (1.9) 50.2% (1.9) 50.8% (0.2) 52.4% (1.8) 53.4% (1.6) 53.8% (0.9) 54.9% (1.1) 57.4% (0.6) 57.6% (1.2) 59.6% (1.8) 67.3% (1.2) 71.6% (4.0)
+ Number of samples: Dataset (Training/validation set) * Standard deviation in brackets 1 Comparable baseline with 518 training samples 2 E3/AS1(4), G14/AS1(4), G1/AS1(2), G4/AS1(1), C/VM(1) 3 Baseline validation documents 4 Up to 82 rules of E2/AS1 were randomly selected to train the other documents. 5 Average of three splits with different random seeds.
Table 3. Addressing complex LRML expressions. Conditioning*
Document specific comparison.
Finally, LRML clauses using variable definitions and loops seem to cause no harm considering the current parsing performance. In particular, variable definitions might be more predictable as expected. 4.4 Document specific evaluation Finally, we use our experimental setup to compare the quality and complexity of the LRML rules per regulatory document. To allow fair comparisons, we train all documents with the same number of samples (i.e. 518 for E2/AS1). Additionally, we report the results for random splits of the reduced and full datasets as upper boundaries on removing the need to generalise across topics. We draw three conclusions from the results in Table 4: 1. Different AS for the same building code have high conceptual or encoding similarities. So, results are closer to the upper limit. See G12 and G13.
433
2. The baseline has worse results than the individual documents in the baseline, indicating the benefit of having a more versatile training set. 3. The suggested method can be used to identify problematic documents. For example, B1/AS1 and C/AS2 need further investigation. Table 5 compares B1/AS1 and C/AS2 against similar-sized documents with medium performance based on the complexity classes introduced in Table 3. Both documents have a high number of define- and key-constructs. C/AS2 has the topic “Protection from Fire” and was encoded by experts using loops, calculations, and complex constructs to achieve high executability. B1/AS1 contains clauses that describe textual changes to referenced standards. Since the LRML expressions used to encode these changes are unique to this AS, the NSP has not learned to generate expressions similar to the human translations. Table 5.
Complexity comparison.
Setting
B1AS1
CAS2
B1AS3
E1AS1
All Ignore1 Implicit2 Medium3 Coarse3 Loop4 Define5 Key6
62 8 5 17 2 0 31 17
55 14 5 11 2 8 20 16
60 4 16 9 13 6 8 7
61 15 1 25 7 2 10 1
1−6
See Table 3 for descriptions.
5 DISCUSSION The experiments in Section 4.2 confirmed the efficacy of the treatments suggested in Section 3. Furthermore, Section 4.3 indicates that the untreated problems of Section 2.3 lead to genuine concern and should be treated in future work. We can conclude that manual information extraction from building regulations is not necessarily the best solution. In addition to time, cost and effort, this process can cause inconsistencies, under- and over-specification, unsound logic, and complex interpretations. The freeform data extraction is seen as an insufficient solution if the necessary tool support and input validation are missing (Biemann et al. 2017). Three enhancements are suggested: First, the LKMD should be fully integrated into the extraction process to validate all entities and relations, suggest relevant terms, and allow safe and quality-assured dictionary management. Second, retrieving LRML expressions for related phrases could prevent inconsistent translations (Song et al. 2018). Finally, the logical soundness of the LRML representation should be evaluated to identify missing relations (Wang et al. 2018). While the NSP performance is not yet sufficient for fully automated translation, we see a payoff for this method in validating manual translations and for the integration into a semi-automatic translation process. Having established an upper limit of 71.6% F1-Score
for the random validation split is a good sign for the feasibility of seq2seq models for LRML parsing. Nevertheless, the high disparity compared to the generalisation split might not be limited to the generalisation capability but also be connected to the training data variety and repetitiveness of weaker models’ generated output. In particular, the average length ratio between the translations and references was 0.93 for random splits (Table 4), 1.11 for the best model in Table 2, and 1.45 for the baseline model in Table 2. 6 CONCLUSIONS AND FUTURE WORK This work brings us closer to the ambitious task of automatically translating building regulation end-toend into semantic representations useable for ACC. We improved the parsing performance on a generalisation validation split from 33.7% to 48.0% F1-Score by improving the LRML tokenisation, the alignment between regulatory statements and LRML rules and the consistency and soundness of the LRML rules. For random validation splits the NSP has achieved 71.6% F1-Score. Furthermore, we have identified LRML constructs, encoding styles and entire documents that remain problematic. These problems can be resolved iteratively using an NSP to indicate problems and provide alternative translations. The next step in this research will be to systematically translate topically and structurally unrelated international building regulations to provide an unbiased test set to verify the scalability and correctness of the improved LRML corpus. Additionally, we will test more sophisticated model architectures, training procedures and decoding strategies to resolve problems unrelated to the dataset quality. Finally, we plan to ground the NSP with the bSDD and LKMD and teach it about the construction domain. ACKNOWLEDGEMENTS This research was funded by the University of Canterbury’s Quake Centre’s Building Innovation Partnership (BIP) programme, which is jointly funded by industry and the Ministry of Business, Innovation and Employment (MBIE). REFERENCES
434
Amor, R. & Dimyadi, J. 2021. The Promise of Automated Compliance Checking. Developments in The Built Environment 5. Artstein, R. 2017. Inter-annotator Agreement. In Handbook of Linguistic Annotation (pp. 297-313). Springer, Dordrecht. Banarescu, L., Bonial, C., Cai, S., Georgescu, M., Griffitt, K., Hermjakob, U., Knight, K., Koehn, P., Palmer, M. & Schneider, N. 2013. Abstract meaning representation for sembanking. In Proceedings of the 7th Linguistic Annotation Workshop and Interoperability With Discourse. Beach, T.H., Hippolyte, J.L. & Rezgui, Y. 2020. Towards The Adoption of Automated Regulatory Compliance
Checking in The Built Environment. Automation in Construction, 118. Biemann, C., Bontcheva, K., Eckart de Castilho, R., Gurevych, I. & Yimam, S.M. 2017. Collaborative Webbased Tools for Multi-layer Text Annotation. In Handbook of Linguistic Annotation (pp. 229–256). Springer, Dordrecht. Bhuiyan, H., Olivieri, F., Governatori, G., Islam, M.B., Bond, A. & Rakotonirainy, A. 2019, December. A Methodology for Encoding Regulatory Rules. In MIREL@ JURIX. buildingSMART. 2022. buildingSMART Data Dictionary. https://github.com/buildingSMART/bSDD. Eurostat. 2017. Which Sector is the Main Employer in the EU Member States?Available at: https://ec.europa.eu/eurostat/ web/products-eurostat-news/-/DDN-20171024-1. Accessed 3 Jun. 2022. Devlin, J., Chang, M.W., Lee, K. & Toutanova, K. 2018. Bert: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv preprint arXiv:1810.04805. Dimyadi, J., Fernando, S., Davies, K. & Amor, R. 2020. Computerising the New Zealand Building Code for Automated ComplianceAudit. New Zealand Built Environment Research Symposium (NZBERS). Eastman, C., Lee, J.M., Jeong, Y.S. & Lee, J.K. 2009. Automatic Rule-based Checking of Building Designs. Automation in Construction, 18(8), pp.1011-1033. Fuchs, S. & Amor, R. 2021. Natural Language Processing for Building Code Interpretation: A Systematic Literature Review. In Proc. of the Conference CIB W78, Vol. 2021. Fuchs, S., Witbrock M., Dimyadi, J. & Amor, R. 2022. Neural Semantic Parsing of Building Regulations for Compliance Checking. In Proc. of the Conference CIB W78, Vol. 2022. Hjelseth, E. & Nisbet, N. 2011, October. Capturing Normative Constraints By Use of the Semantic Mark-up RASE Methodology. In Proceedings of CIB W78-W102 Conference. Hurren, K. 2019. Building Consents Over Two Decades | BERL. Available at: https://berl.co.nz/economicinsights/building-consents-over-two-decades. Accessed 31 May 2022. Jascob, B. 2022. Amrlib. Github. https://github.com/bjascob/ amrlib Li, B., Schultz, C., Dimyadi, J. & Amor, R. 2021. Defeasible Reasoning for Automated Building Code Compliance Checking. In ECPPM 2021–eWork and eBusiness in Architecture, Engineering and Construction (pp. 229–236). CRC Press. Li, F., Song, Y. and Shan, Y., 2020. Joint Extraction of Multiple Relations and Entities from Building Code Clauses. Applied Sciences, 10(20), p. 7103. Li, X.L. & Liang, P. 2021. Prefix-tuning: Optimizing Continuous Prompts for Generation. arXiv preprint arXiv:2101.00190. Ministry of Business, Innovation and Employment. 2017. Acceptable Solutions and Verification Methods For New Zealand Building Code Clause D1 Access Routes (Vols. 2nd edition, amendment 6). Ministry of Business, Innovation and Employment. 2021. Building and Construction Sector Trends. Annual Report 2021. Available at: https://www.mbie.govt.nz/ dmsdocument/16973-building-and-construction-sector-tr ends-annual-report-2021-pdf. Nazarenko, A., Levy, F. & Wyner, A.Z. 2016. Towards a Methodology for Formalizing Legal Texts in LegalRuleML. In JURIX (pp. 149–154). Papineni, K., Roukos, S., Ward, T. & Zhu, W.J. 2002, July. Bleu: A Method for Automatic Evaluation of Machine
435
Translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (pp. 311–318). Poesio, M., Chamberlain, J. & Kruschwitz, U. 2017. Crowdsourcing. In Handbook of Linguistic Annotation (pp. 277–295). Springer, Dordrecht. Preidel, C. & Borrmann, A., 2018. BIM-based Code Compliance Checking. In Building Information Modeling (pp. 367–381). Springer, Cham. Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W. & Liu, P.J. 2019. Exploring the Limits of Transfer Learning With a Unified Text-totext Transformer. arXiv preprint arXiv:1910.10683. Song, J., Kim, J. & Lee, J.K. 2018. NLP and Deep Learningbased Analysis of Building Regulations to Support Automated Rule Checking System. In ISARC. Proceedings of the International Symposium on Automation and Robotics in Construction (Vol. 35, pp. 1–7). IAARC Publications. Statistics New Zealand. 2022. Building Consents Issued: April 2022 | Stats NZ. Available at: https://www.stats.govt. nz/information-releases/building-consents-issued-april-2 022. Accessed 31 May 2022. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł. & Polosukhin, I. 2017. Attention is all you need. Advances in Neural Information Processing Systems, 30. Wang, C., Tatwawadi, K., Brockschmidt, M., Huang, P.S., Mao, Y., Polozov, O. & Singh, R. 2018. Robust Text-tosql Generation With Execution-guided Decoding. arXiv preprint arXiv:1807.03100. Wu, J., Sadraddin, H.L., Ren, R., Zhang, J. & Shao, X. 2021. Invariant Signatures of Architecture, Engineering, and Construction Objects to Support BIM Interoperability Between Architectural Design and Structural Analysis. Journal of Construction Engineering and Management, 147(1). Wu, J. & Zhang, J. 2022. Model Validation Using Invariant Signatures and Logic-Based Inference for Automated Building Code Compliance Checking. Journal of Computing in Civil Engineering, 36(3), p.04022002. Wyner, A.Z. & Governatori, G. 2013. A Study on Translating Regulatory Rules from Natural Language to Defeasible Logics. In RuleML (2). Zhang, J. & El-Gohary, N. M. 2015. Automated Information Transformation for Automated Regulatory Compliance Checking in Construction. Journal of Computing in Civil Engineering. 29(4). B4015001. Zhang, J. & El-Gohary, N. 2016. Semantic NLP-based Information Extraction From Construction Regulatory Documents for Automated Compliance Checking. Journal of Computing in Civil Engineering, 30(2). Zhang, R. & El-Gohary, N. 2019. A Machine Learning-based Approach for Building Code Requirement Hierarchy Extraction. In 2019 CSCE Annual Conference. Zhang, R. & El-Gohary, N. 2020, November. A MachineLearning Approach for Semantically-Enriched BuildingCode Sentence Generation for Automatic Semantic Analysis. In Construction Research Congress 2020: Computer Applications (pp. 1261–1270). Zhang, R. & El-Gohary, N. 2021. Semantic Representation Learning and Information Integration of BIM and Regulations. In Computing in Civil Engineering 2021. Zhou, P. & El-Gohary, N. 2017. Ontology-based Automated Information Extraction From Building Energy Conservation codes. Automation in Construction. 74. pp. 103–117.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Using RASE semantic mark-up for normative, definitive and descriptive knowledge N. Nisbet & L. Ma The Bartlett School of Sustainable Construction, UCL, London, UK
ABSTRACT: The RASE methodology has gained some attention as a means to expose the logical objectives and individual metrics found in regulatory and contractual documents. This paper will explore the role of RASE in exposing not only normative documents but also both definitive resources such as dictionaries, thesauri and classification tables, and descriptive resources such as BIM models, contract diaries, product data or technical journalism. A single common execution framework allows any of these resources to be cross-compared, independent of domain or language. This support for mixed modalities opens the potential for RASE to be used as a core concept across multiple domains and information types.
1 INTRODUCTION
Table 1. Type of document content.
The Architectural, Engineering and Construction (AEC) industry has been relatively slow to adopt digitalization compared to other sectors. This means that the construction sector retains a heavy dependance on documents not only as evidence such as certificates and photographs but also for holding the primary sources of information, even when there are alternatives available such as Building Information Modelling BIM and its evolution into asset information management. RASE (AEC3 2021) makes explicit the logical structure within knowledge content of documents and other stored formats. Previous work has examined how RASE semantic mark-up of documents can act as a precursor to a wide range of existing applications and knowledge representations, by treating them as presentations of RASE knowledge (Nisbet 2022). In contrast, this paper examines how the RASE semantic mark-up can used to render the content of documents directly operable, without the need for any intermediate representations. This should offer advantages in terms of accuracy and efficiency. It can also offer advantages in terms of privacy and security for knowledge content owned by governmental and commercial entities. Table 1 gives a breakdown of types of content found in documents. Individual documents may contain several types of content. The content may be text or tables. Documents may also contain multi-dimensional and complex information. The types of content in italic are not considered further. The current scope of interest has been set around normative, definitive and descriptive knowledge. This
Breakdown
Document content
1 1.1 1.2 1.3 2 2.1 2.2 2.2 3 3.1 3.2 3.3 4 4.1 4.2 5 5.1 5.2 6 6.1 6.2
Descriptive (picturing) Description Illustration Narrative Definition Synonyms and translations Classification Equations and algorithms Normative (expectations) Regulations Requirements Recommendations Argumentation (convincing) Argument Comparison and contrast Exposition (explaining) Analysis Cause and effect Evidence (verification) Audio-visual content Certificates and affidavits
436
is to exclude argumentation and exposition (types 4 and 5) where the knowledge content is dynamic and potentially inconsistent, though any conclusion derived from an argument or exposition may be considered as knowledge and marked-up. It also excludes the evidential content of documents (type 6) which may be supportive of knowledge content but is not subject to reasoning. DOI 10.1201/9781003354222-56
2 METHOD
Table 3.
This paper adopts a design-science paradigm to allow the exploration of the domain. It offers working descriptions of the three kinds of explicit knowledge in order to test their utility. It presents the relevant algorithms developed in a series of experimental applications for exploiting that knowledge both in isolation (section 3) and in combination with other knowledge resources (section 4). Each experiment is reported with source material, its RASE mark-up and an outline of relevant algorithms. Limitations are noted.
Knowledge type
High levels
Lowest level
Normative Definitive Descriptive
section concept entity
metric term property
3 THREE KINDS OF OPERABLE KNOWLEDGE At least for the three kinds of knowledge in scope, RASE asserts that there are four roles performed by the metrics found in phrases and the sections found in the structure of written documentation. The metrics and sections may or may not exactly match the presentational structure of the document, so, for example, an exception section may be a separate paragraph following the main requirement. Table 2.
RASE Types used in mark-up.
Sections
Metrics
Requirement Section
Requirement (normative) Reference (definitive) Report (descriptive) Application Selection Exception
Application Section Selection Section Exception Section
The addition of RASE mark-up to a document or its association to tabular and multi-dimensional data, creates a simple hierarchy. Nisbet et al (2022) showed that this hierarchy can be mapped to other representations using a simple tree traversal algorithm which ensures that each objective section and every metric test is visited methodically. When executing (as opposed to reporting) over a RASE document it may not be necessary to visit every section or metric and so further heuristics can be safely used to accelerate the processing. RASE as originally described (Nisbet 2008) was focused on normative knowledge (type 3). In applying that knowledge it has been necessary to also consider definitive knowledge (type 2) and descriptive knowledge (type 1). 3.1 Normative knowledge In response to issues around accuracy and efficiency as reported in automated regulation code compliance checking, Nisbet (2008) proposed the RASE methodology as a means to capture and render operable the normative content of Building Codes and Regulations, thereby eliminating the requirement for domain
RASE knowledge constituents.
expertise, code expertise and model expertise to come together to re-interpret the regulations. Examples of normative content include much of the legal, regulatory and contractual documents. It also includes other requirements from clients or third parties. They are characterized by ‘Requirement’ knowledge content (Table 2). A feature of the formal style of some legal and regulatory content is the complexity of the chapter, paragraph and sentence structures. Normative knowledge content acts to create expectations. These expectations may be met by descriptive knowledge content such as a BIM model, or the knowledge of a user, or by measurements taken from the real world. An application (AEC3 2022) can traverse a normative document to convert the text and tables into an interactive checklist. The checklist can include the original text with added input boxes to allow a user to answer the questions implied by the metric phrases. This is particularly useful if the normative knowledge applies to a single entity, such as a proposal, site or building overall. Initially the overall result is ‘unknown’. As each metric is answered, the document can update the overall result or hide sections that have been satisfied as ‘as required’, ‘excepted’ ‘not applicable’ or ‘not selected’, or ‘false’ If any result is ‘unknown’then it and the relevant sections and metrics below remain visible. More work is required to decide how the outcome can be preserved and documented.
Figure 1.
Table 4.
Example regulation as a form (NSW 1995).
RASE mark-up used in Figure 1.
RASE type
Metrics
Selection Selection Application Requirement
development is ‘dwelling house’ development is ‘attached’ type is ‘development’ height above ground 0) return f if(n(u,R) > 0) return u return t if c is A if(n(t,A) > 0) return t if(n(u,A) > 0) return u return f if c is S if(n(f,S) > 0) return t if(n(u,S) > 0) return u return f if c is E if(n(t,E) > 0) return f if(n(u,E) > 0) return u return t return u
Metric evaluation.
tfu evaluateMetric(m1) v = m1.value g = m1.target c = m1.comparator if c is ">=" then return (v >= g) if c is ">" then return (v > g) if c is "= 0 for all x ∈ ss , D(x) → min
3 3.5 4 4.5
(1)
However, utilizing raw index values can lead to deviation in searching, especially for sub-attributes of D on different scales. Moreover, if the search for feasible solutions results in an empty set, the users should consider including input parameters with a more global impact and improve the adaptation methods, which might also involve the analysis of C. Thus, transformation approaches are anticipated to conduce to metrics (C, D) on comparable scales. Finally, the search evolves toward the global maximum point symbolizing the closest valid variant during the adaptation, providing a compliant BIM model that deviates the least from the input model.
4 PROOF OF CONCEPT 4.1 Scenario description In order to demonstrate the applicability of the Model Healing framework, a prototype of adaptive building design for the German standard DIN 18232-2 (DIN, 2007) is realized. This rule addresses security on room level in terms of smoke and fire protection, which ensures smoke is vented from escape routes. An academic one-space BIM model is created, which
Required Ventilation Area [m2 ]
Height of the Smoke Layer [m]
1
2
3
4
5
0.5 1 0.5 1.5 1 2 1.5 1
4.8 3.4 3.0 2.5 3.0 2.5 3.0 3.5
6.2 4.4 8.7 3.6 6.2 3.1 5 8.4
8.2 5.8 11.3 4.7 8 4.1 6.5 10.7
11 7.8 15 6.4 10.6 5.5 8.7 13.9
15.4 10.9 20.4 8.9 14.4 7.7 11.8 18.6
Fire Classification
4.2 Proposal application 4.2.1 Compliance checking The compliance checking examines the minimal ventilation area and solely considers this evaluation criterion. Due to the simplicity of this one-space model, we developed a checking algorithm to evaluate the compliance instead of taking existing model checkers for model evaluation. After the initialization of the design, compliance checking is executed following the overall workflow (Figure 3). 4.2.2 Solution space formation Compared with complex BIM models from practice, this one-space model contains evident variational features related to the checking rule: objects (number and location of smoke vent) and their characteristics (type and dimensions of smoke vent). In the proposed framework, the end-users determine the adaptable parameters from relevant variational features associated with the checking failures based on their understanding of the overall design process and corresponding constraints. In the investigated scenario, the room dimensions and the fire classification
454
transformation function flog is employed to moderate the distribution irregularity by keeping their original signs. The min-max normalization coefficients (α,β) in fnorm are captured from the non-negative values to convert positive values into [0, 1], facilitating investigating the feasible region (Ci >= 0) where the designs comply with the checking rules. flog and fnorm for D are expressed in Equation 4 and 5, and same transformations are applied to C.
Figure 3. Fragment of Dynamo script for Model Healing: a,b) primary input, c) adaptable input regarding the checking rule, d) initial modelization, and e) rule checking on the initial design.
to resist are considered to be fixed. Therefore, the input parameters of the solution space can be easily elected among all corresponding variational features in Table 2. We assume the smoke vent in square shapes and utilize identical dimensions for every smoke vent according to practical experience. This simplification filters the solution space into a multiple-dimensional space with two adaptable input parameters: the number of the smoke vent n and the length of each square smoke vent l. Table 2.
log (Di + 1) , − log (−Di + 1) ,
fnorm (Di ) =
Di − β α
(if Di >= 0) (if Di < 0)
(4) (5)
where α = max (D1 ,. . .,Di ), β = 0, i ∈ N ∗ .
Relevant features of the initial design.
Parameter
h
z
fire strength
n
l
Value
3.0 m
0.5 m
Class 5
3
1.75 m
After being evaluated by the checking algorithm, the initial design shows that it did not meet the requirement on minimum smoke ventilation area Amin . Based on practical experience, we vary the input parameters with comparable values following a two-sided truncated normal distribution. We fix the smoke vent family with limited length l in [1 m, 3 m] and the number of smoke vent n in [1, 9]. Afterward, the built-in checking algorithm is reapplied to all adapted designs. 4.2.3 Multicriterial searching toward Model Healing The two indices (C, D) of the healing metric for the ith adapted design are expressed in Equation 2 and 3. Ci = ni ∗ li2 − Amin , i ∈ N ∗ (2) ! Di =
flog (Di ) =
2
ni −1 ninit
+
li linit
2 −1
(3)
where ni , ninit represent the number of the smoke vent of the ith adapted design and the initial design, and li , linit are the length of each square smoke vent of the ith adapted design and the initial design. Since the original “distance” on input parameters (n, l) are on different scales, a preliminary transformation is adopted to calculate the weighted quasidistance D (Equation 3). Moreover, a logarithmic
Figure 4. Illustration of the raw indices and other transformed indices
The raw indices and transformed indices are illustrated in Figure 4. In this prototype, the normalizedlogarithmic indices (Ci∗ , Di∗ ) are adopted. flog (Ci ) − βlog,C Ci∗ = fnorm flog (Ci ) = αlog,C
(6)
flog (Di ) − βlog,D Di∗ = fnorm flog (Di ) = αlog,D
(7)
where αlog,C = max flog (C1 ), . . . , flog (Ci ) , αlog,D = max flog (D1 ), . . . , flog (Di ) , βlog,C = βlog,D = 0, i ∈ N ∗. The optimal design is selected according to the performance threshold and performance function in Equation 1, leading to compliant designs that deviate the least from the initial design (Figure 5). The solution space concerning D is described in Figure 6, depicting the feasible and infeasible regions within the overall adaptation boundaries. The results achieved Model Healing by selecting the closest feasible option within the solution space.Those adapted options far from the initial design (in Figure 5) have a significantly different length or
455
5 CONCLUSION AND FUTURE WORKS
Figure 5. The results of searching ized-logarithmic indices (C ∗ , D∗ ).
via
normal-
This paper proposed a novel framework for automatic building design adaptation for non-compliant building designs. It addressed the gap that code compliance checking and automated design improvement have not been conjointly investigated to automate the design process. The proposed framework investigates the rule checking results and selects adaptable parameters from design features associated to non-conformance. Within the framework, the compliance violation is considered the improvement target, which eases the design improvement from iterative processes. Based on the variational design, comparable variants are created by varying input parameters within the solution space. Metric indices representing the checking conformance and dissimilarity to the initial design are calculated to support model evaluation. The searching among adapted designs finally leads to a compliant BIM model that deviates the least from the input model. An experiment on German smoke and fire protection regulation shows that the proposed framework is applicable. This paper addresses the existing gap in interpreting rule checking results to support automatic building design adjustment. We acknowledge the following limitations and corresponding future works: •
Figure 6. Solution space malized-logarithmic indices (n, l).
Industrial BIM models comprising enormous components will result in high-dimensional spaces, reducing the model representation effectiveness and hampering the broader adoption of this framework. The proposed variational features and hierarchical levels need further investigation to help select feasible input parameters of the solution space for practice building designs concerning specific checking rules. • The scope of the experiment is limited. Employing an academic model simplifies the failure checks and solution space formation. Built-in algorithms temporarily undertook model checking. Nevertheless, challenges might be revealed when the checking demands expand to undeveloped checking rulesets in model checkers. Further experiments on industrial-practice BIM models and practical model checkers will expand the framework’s application.
concerning the norof quasi-distance D:
number of smoke vents. The initial design and optimal design selected from the adaptation considered the final solution for this experiment are illustrated in Figure 7. In this way, the initial model was “healed” to a compliant design by slightly enlarging the dimensions of smoke vents and adding one more smoke vent in the ceiling. The requirement of building codes is satisfied, and the initial design trend has also been kept.
ACKNOWLEDGEMENTS This research was funded by the International Graduate School of Science and Engineering (IGSSE) / TUM Graduate School. REFERENCES Figure 7. 3D views: (a) the initial design, (b) the adapted design.
456
Abualdenien, J. & A. Borrmann (2019). A Meta-model Approach for Formal Specification and Consistent Management of Multi-lod Building Models. Advanced Engineering Informatics 40, 135–153.
Amor, R. & J. Dimyadi (2021). The Promise of Automated Compliance Checking. Developments in the Built Environment 5(October 2020), 100039. Anandan, S., S. Teegavarapu, & J. D. Summers (2006). Issues of Similarity in Engineering Design. Proceedings of the ASME Design Engineering Technical Conference 2006, 1–10. Borrmann, A., M. König, C. Koch, & J. Beetz (2018). Building Information Modeling: Technology Foundations and Industry Practice. Brown, N. C., V. Jusiega, & C. T. Mueller (2020). Implementing Data-driven Parametric Building Design with a Flexible Toolbox. Automation in Construction 118. Brown, N. C. & C. T. Mueller (2019). Quantifying Diversity in Parametric Design: A Comparison of Possible Metrics. Artificial Intelligence for Engineering Design, Analysis and Manufacturing: AIEDAM 33(1), 40–53. Cornelius Preidel; & André Borrmann; (2017). Refinement of the Visual Code Checking Language for an Automated Checking of Building Information Models Regarding Applicable Regulations. Congress on Computing in Civil Engineering, Proceedings (2017), 157–165. Dimyadi, J. & R. Amor (2013). Automated Building Code Compliance Checking – Where is it at? Proceedings of the CIB World Building Congress 2013 and Architectural Management & Integrated Design and Delivery Solutions (AMIDDS) (380), 172–185. Dimyadi, J., R. Amor, & M. Spearpoint (2016). Using Bim to Support Simulation of Compliant Building Evacuation. eWork and eBusiness in Architecture, Engineering and Construction - Proceedings of the 11th European Conference on Product and Process Modelling, ECPPM 2016 (September), 511–520. Dynamo. https://dynamobim.org/. Accessed: 2022-03-20. Eastman, C., J. m. Lee, Y. s. Jeong, & J. k. Lee (2009). Automatic Rule-based Checking of Building Designs. Automation in Construction 18(8), 1011–1033. Eastman, C., P. Teicholz, R. Sack, & K. Liston (2011). BIM Handbook, A Guide to Building Information Modeling for Owners, Managers, Designers, Engineers, and Contractors, Second Edition. Etalle, S. (1996). An Introduction to Logic Programming Through Prolog Michael Spivey, Prentice Hall International Series in Computer Science, Englewood Cliffs, NJ, 1996. ISBN 0-13-536047-1. Science of Computer Programming. Gassmann, O. & M. Zeschky (2008). Opening up the Solution Space: The Role of Analogical Thinking for Breakthrough Product Innovation. Creativity and Innovation Management 17(2), 97–106. Gerber, D. J., S. H. Lin, B. Pan, & A. S. Solmaz (2012). Design Optioneering: Multi-disciplinary Design Optimization through Parameterization, Domain Integration and Automation of a Genetic Algorithm. Simulation Series 44(8 BOOK), 33–40. Graff, L., H. Harbrecht, & M. Zimmermann (2016). On the Computation of Solution Spaces in High Dimensions. Structural and Multidisciplinary Optimization 54(4), 811–829. Granadeiro, V., J. P. Duarte, J. R. Correia, & V. M. Leal (2013). Building Envelope Shape Design in Early Stages of the Design Process: IntegratingArchitectural Design Systems and Energy Simulation. Automation in Construction 32, 196–209. Hjelseth, E. & N. Nisbet (2010). Exploring Semantic Based Model Checking. Proceedings of CIBW78 Conference 27, 16–18. Ismail, A. S., K. N. Ali, & N. A. Iahad (2017). A Review on Bim-based Automated Code Compliance Checking
457
System. International Conference on Research and Innovation in Information Systems, ICRIIS. Lee, G., R. Sacks, & C. M. Eastman (2006). Specifying Parametric Building Object Behavior (bob) for a Building Information Modeling System. Automation in Construction 15(6), 758–776. Lee, J. k. (2011). Building Environment Rule and Analysis (bera) Language and its Application for Evaluating Building Circulation and Spatial Program. Lin, V. C., D. C. Gossard, & R. A. Light (1981). Variational Geometry in Computer-aided Design. Proceedings of the 8th Annual Conference on Computer Graphics and InteractiveTechniques, SIGGRAPH 1981 15(3), 171–177. Ma, W., X. Wang, J. Wang, X. Xiang, & J. Sun (2021). Generative Design in Building Information Modelling (bim): Approaches and Requirements. Sensors 21(16). Markus, Z. & E. v. H. Johannes (2012). Computing Solution Spaces for Robust Design. International (February), 1102–1119. Nawari, N. O. (2012). Automating Codes Conformance. Journal of Architectural Engineering 18(4), 315–323. Nawari, N. O. (2019). A Generalized Adaptive Framework (gaf) for Automating Code Compliance Checking. Buildings 9(4). Peffers, A., E. Charles, J. The, M. Rossi, & W. Hui (2006). The Design Science Research Process. 17028071, 83–106. Preidel, C. & A. Borrmann (2015). Automated Code Compliance Checking Based on a Visual Language and Building Information Modeling. 32nd International Symposium on Automation and Robotics in Construction and Mining: Connected to the Future, Proceedings. Preidel, C. & A. Borrmann (2016). Towards Code Compliance Checking on the Basis of a Visual Programming Language. Journal of Information Technology in Construction 21(July), 402–421. Robert, C. P. (1995). Simulation of Truncated Normal Variables. Statistics and Computing 5(2), 121–125. Shahin, T. M. (2008). Feature-based Design – An Overview. Computer-Aided Design and Applications 5(5), 639–653. Solibri Office. https://www.solibri.com/solibri-office. Accessed: 2022-03-20. Solihin, W., J. Dimyadi, & Y. Lee (2019). In Search of Open and Practical Language-driven Bim-based Automated Rule Checking Systems. Advances in Informatics and Computing in Civil and Construction Engineering (January). Solihin, W., J. Dimyadi, Y. C. Lee, C. Eastman, & R. Amor (2020). Simplified Schema Queries for Supporting Bimbased Rule-checking Applications. Automation in Construction 117(March), 103248. Wu, J. & J. Zhang (2022). Model Validation Using Invariant Signatures and Logic-Based Inference for Automated Building Code Compliance Checking. Journal of Computing in Civil Engineering 36(3). Zhang, J. & N. M. El-Gohary (2016). Semantic NLP-based Information Extraction from Construction Regulatory Documents for Automated Compliance Checking. Journal of Computing in Civil Engineering 30(2), 04015014. Zhang, J. & N. M. El-Gohary (2017). Integrating Semantic NLP and Logic Reasoning into a Unified System for Fully-automated Code Checking. Automation in Construction 73, 45–57. Zhang, R. & N. El-Gohary (2021). A Deep Neural Network-based Method for Deep Information Extraction Using Transfer Learning Strategies to Support Automated Compliance Checking. Automation in Construction 132(October 2020), 103834.
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
A multi-representation method of building rules for automatic code compliance checking Z. Zhang, N. Nisbet, L. Ma & T. Broyd The Bartlett School of Sustainable Construction, UCL, London, UK
ABSTRACT: In the Architecture, Engineering and Construction (AEC) industry, design review is an important step that often leads to project delays, as the typical manual compliance checking process is error-prone and timeconsuming. As an approach to accelerate this process and achieve a better quality of design, automatic compliance checking (ACC) has been researched for several decades. Rule interpretation and representation is a bottleneck of ACC. It focuses on the interpretation of regulations and the representation of them in a suitable computerreadable form. Despite extensive research efforts, a rule representation method that is suitable to represent all types of rules has yet to be proposed. To address this issue, this research proposed a multi-representation method that provides a “mix and match” for different representations and different types of rules, thereby representing all types of rules with suitable representations. This research is valuable to both academia and industry as it enables the representation of rules with less knowledge loss and more accuracy.
1 INTRODUCTION In the Architectural, Engineering and Construction (AEC) industry, the design needs to be checked against regulations before the construction permit can be obtained. The compliance checking is traditionally conducted manually by rule experts, suffering from low efficiency, low accuracy and high cost. The errors may result in project delays or poor operational performance. To address these issues, automatic compliance checking (ACC) has been researched for more than 50 years. There have been two perspectives to approach ACC, one looks at the easy retrieval and query of building model data, and the other focusses on the interpretation and representation of regulations. The latter is claimed to be the bottleneck of the ACC process, which can take up to 30 % of the total time for implementing a rule (Solihin & Eastman 2016). Recognizing the important role of rule interpretation and representation in understanding and retaining construction design knowledge, various methods were developed to represent building codes. While these methods have been useful to handle various aspects of the knowledge domain, none of them has all the capabilities required to represent regulations (Macit ˙Ilal & Günaydın 2017). There also lacks a thorough understanding of rules, representation methods and the relationships between them; and the methodological backdrop of the current representation development is weak (Zhang et al. 2022). This paper thus aims to: 1) compare the methods of rule representation and identify their capabilities;
458
2) select suitable representation methods for different types of rules; 3) propose a multi-representation method that is capable of representing all types of rules. The relationships between rule types, capabilities, suitable representation for each rule class and a multi-representation method are shown in Figure 1.
Figure 1. Relationships among class of rules, capabilities, and representation methods.
The upper boxes and arrows show the overall relationships of class of rules, capabilities and the multi-representation method to be developed in this paper. For each column, it has several subsets as shown in the lower boxes (i.e. rule classes, capability sets and representations). The multi-representation method can only be developed when different types DOI 10.1201/9781003354222-59
of rules and capabilities required to represent them are understood. Once the suitable representation methods for different classes of rules are selected, a “mix and match” approach can be used to establish the multi-representation method.
Table 1. Required and desired capabilities for rule representation
Aspects
Required capabilities
Rule features
1) Requirement 2) Applicability 3) Selection 4) Exception 5) Definition 6) Outcome 7) Logical relationship 1) No calculation or simple calculation 2) Function and algorithm 3) Simulation 1) Hierarchy 2) Cross-reference 1) Easy to use and understand 2) Dictionary 3) Independent of the rule engine 4) Independent of the data model
2 METHOD In this paper, design science research (DSR) is adopted. It is suitable for this research as it aims to provide practical solutions for industry problems (i.e. a representation method for building rules). DSR has five steps, including awareness of problem, suggestion, development, evaluation and conclusion (Vaishnavi 2007). This paper presents the awareness of problem and suggestion as a solid foundation for developing a rule representation method. The suggestion of a new representation method is based on abductive reasoning. It involves the researchers selecting the “best” based on a set of pragmatic criteria. In this paper, the authors adopted the pragmatic criteria (Table 1) proposed in our previous work (Zhang et al., unpubl.). We seek to select the most capable representations based on those criteria. The evaluation will also help understand which representation is suitable for each type of rule.
3 EVALUATION OF EXISTING REPRESENTATION METHODS 3.1 Evaluation criteria Zhang et al. (unpubl.) proposed 16 required and two desired capabilities (Table 1) for representing building regulations. It is so far the most comprehensive set of criteria. Thus, we have chosen these criteria to evaluate the representations. The evaluation of representation methods was twofold. Firstly, the current capabilities of representations were evaluated against the above-mentioned criteria. Secondly, they were also assessed regarding their potential of developing certain capabilities given their characteristics. 3.2 Evaluation and results Some early ACC developments use procedural code to represent and execute rules. An example of this is the CORENET e-PlanCheck project in Singapore (Solihin 2004). It is a hard-coded approach, where the system is a “black box” with a built-in rule engine and procedural codes for executing rules, thus the representation is not independent of data model and rule engine. The system mainly focused on individual rule provisions and neglected cross-references. Because of the binary nature of the programming language, the checking outcomes only include “true” or “false”, whilst other actions or side-effects cannot be represented. The system also requires extensive programming knowledge
Rule intensity
Rule organization Implementation
Desired capabilities NA
NA
NA 1) Conciseness 2) Translatability
to develop, which domain experts typically do not have. However, its advantage lies in dealing with rules of high intensity using algorithms and functions and considering hierarchies of rules and logical relationships within a rule provision to get correct checking outcomes. Production rule has been widely used for representing rules. It follows the pattern of “if then ”. An example in the AEC industry is Tan et al. (2010). They proposed decision tables to present regulations in a concise and compact way, where each row is a production rule. Decision tables can represent applicability, selection and requirements in its cells. The actions of the checking include “pass”, “fail” and “exceptional”. Cross-references among rule provisions were also considered. However, the deficiency of this method is: 1) it cannot represent unknown and other side effects; 2) no explicit logical relationships are shown within each row; 3) it lacks consideration of the rule context (e.g. definitions of terms); 4) it failed to reveal the implicit knowledge and assumption of rules, thus making it difficult to check rules with higher intensity and/or performance-based rules. Solibri (2022) is the most widely-adopted commercial ACC software. It used parametric tables to represent rules, which is easy to use after some simple training. Currently, there are 55 rule templates to check model and data quality and consistency, objects and properties, and geometries. The rule templates include object and space groups and some other pre-defined applicability such as “air well”. APIs are also available to allow experienced users to build their own rule templates. However, the rule templates only cover a limited
459
number of objects, relationships and algorithms. Thus, Solibri has limited capability to deal with rules with higher intensity (e.g. rules requiring functions, algorithms and simulations). In addition, it must depend on the built-in rule engine and internal data model to finish the check. Like many other methods, it also only includes “pass” and “fail” as outcomes. The logical relationships within one rule provision are not clear. Any cross-references among rules and the hierarchy of regulations are not represented. Although Solibri allows the user to implement any definitions (e.g. fire compartment) using a mixture of rules and individual picks of entities, the broader context for applying a rule is limited to using a second rule as a “gate-keeper”. Predicate logic is a well-defined function in logic that can generate results of true/false (or undefined). It has been used to represent building rules by many researchers. It has quantifiers to distinguish “ALL” instances from the case that there “EXISTS” an instance, making it possible to represent applicability, selection and exception. It also uses logical connectives to denote conjunction, disjunction and negation. For example, Rasdorf & Lakmazaheri (1990) proposed a logic-based SASE (Standards Analysis, Synthesis and Evaluation) model that used predicate logic statements to represent rules. This model showed rule categories and relationships by linking rules using a tree-like organizational sub-model. The expressions of predicate logic are also independent of rule engine, with the potential of being independent of the building data model depending on the specific software. However, the drawbacks of predicate logic are: 1) as a two-value logic, it only has true or false as outcome; 2) it can become very lengthy and complex when the rules are complex; 3) it was not used to present definitions and other context; 4) it does not reflect the hierarchy of the regulations. Having its root in predicate logic, conceptual graph (CG) was initially proposed by Sowa (1976) and later extended by Solihin & Eastman (2016) for representing building regulations, where different shapes were used to represent concept nodes, conceptual relations and represent functions intuitively. Further extensions include representing cross-references and exceptions, and denoting either a derived concept or a concept that requires functions/algorithms/simulations. CG also uses “or” and “not” to represent logical disjunction or negation, and the lines mean logical conjunction by default. Its graphical representation makes it easy to use and understand. Despite being a fairly comprehensive representation method, it has several drawbacks, including its current inability to represent applicability and selection and to show outcomes other than “pass” and “fail”; it cannot show hierarchies of rules; it does not represent definitions; and rules are represented using a model-dependent IFC format. It also does not have a dictionary to link the data model ontology to the rule ontology. However, an independent representation of rules might be possible in its future developments. Visual Code Checking Language (VCCL) is a language-driven ACC solution proposed by Preidel &
Borrmann (2016). It is a user-friendly visual programming language aiming to provide transparent “white boxes” (as opposed to black boxes) for rule checking. It has tailored methods for building rules, including logical, mathematical, geometric-topological, relational, building model related and utility methods. They are executed by the built-in rule engine. Each rule provision can be represented by linking multiple input ports, output ports and method nodes. Nevertheless, VCCL still failed to consider a broader context. It only has binary results including “true” and “false”. It is also restricted by the IFC format and thus does not have a dictionary to map ontologies. In addition, it failed to consider the hierarchy of regulations and cross-references among rules, but the authors argue it might be possible for VCCL to incorporate these in a later version. RASE (Hjelseth & Nisbet 2011) is a semanticbased approach. RASE has been updated several times, here we evaluate the latest version of RASE (Beach et al. 2015) published in academic literature. It represents the semantic constructs of rules, including requirements, applicability, selection and exceptions. Its striking feature is the consideration of broader context by incorporating definitions and/or titles in the rule documents. It keeps the rule text as is and uses mark-ups to mark different semantic constructs. The mark-ups are easy to use once the user has some knowledge of semantic constructs. In addition, keeping the rules as is has several advantages: 1) the structure of regulations is retained, including cross-references, hierarchy, and logical relationships (although logical relationships may not be explicit); 2) the representation is independent of the rule engine and model data. To link the design and the regulation ontologies, a dictionary was developed. As for checking outcomes, RASE uses an open world assumption and its checking outcomes include “pass”, “fail”, “unknown” and other side-effects such as “add credits”. A limitation is that it cannot explicitly present the knowledge needed to address the rule intensity. More recently, some scholars developed representations using Semantic Web. Rules are represented using SPARQL that has similarities to SQL. The queries can then be applied to RDF/OWL representations of a model, often with extensions to simplify the model and underlying schema. The syntax in the semantic web requires specialized training. There has been little progress in exploiting geometry or computationally intensive queries, thus, it has a low capability of dealing with rule intensity. In Pauwels et al. (2011), they used N3Logic to represent building rules, which is essentially a form of “if then ”, with logical connectives to denote logical relationships. This production-rule-like representation limits expressiveness. Consequently, it does not have the capability of representing exceptions, outcomes other than “pass” or “fail”. It also did not consider a broader context such as definitions. The semantic web representations have potential to be independent of the data model, although most of them currently use IFC.
460
Table 2.
Evaluation result of representation methods.
```
Capability ``` ``` A
Method
Procedural Code Decision Table Solibri Predicate Logic Conceptual Graph VCCL RASE Semantic Web NLP
√ √ √ √ √ √ √ √ √
B
C
D
√ √ √ √
√ √
√
P P √
√ √
E
P √
× × × √
× P × P P P √
× ×
× ×
× ×
× √ √
× √
F √ √ × √ √ P × √ ×
G
I √
3 2 2 3 3 2 P 1 1
× × × × × √
J
K
N
O
Q
References
× √
L H H M H H H L M
× × × P × × √
× √
× √
× √ √ × √
× P P × √
× √
× √
Solihin (2004) Tan et al. (2010) Solibri (2022) Rasdorf & Lakmazaheri (1990) Solihin & Eastman (2016) Preidel & Borrmann (2016) Beach et al. (2015) Pauwels et al. (2011) Zhang & El-Gohary Nora (2017)
× √ √
√
P √ √
×
×
√ ×
Note: A = Requirements; B = Applicability/Selection; C = Exception; D = Definition; E = Outcome (Other actions/Side effects); F = Logical Relationships; G = Rule Intensity (level 1-3); I = Hierarchy; J = Cross-reference; K = Easy to Use and Understand; N = Dictionary; O = Independent of Rule Engine; Q = Independent of Data Model; H/M/L = High, Moderate, Low; P = Potential of developing the capability.
However, it will rely on the rule engine regardless of the representation form, and a dictionary can be used to link ontologies. Regarding rule organization, it can represent hierarchies of rules and cross-references. Aiming at a fully-automated ACC system, semantic natural language processing (NLP) methods were proposed by researchers. Most research has narrowed the scope to the simplest subject-predicate sentences, ignoring subsidiary clauses, titles, lists and other matters. Zhang & El-Gohary (2017) have tried various approaches to build on the named-object identification, such as assuming all terms are requirements and matching sentences to pre-defined templates. They did not pay attention to the representation of exception, applicability and selection. The checking outcome includes “pass”, “fail” and “unknown”, but no other actions were considered. The method can only deal with rules with relatively low intensity. In addition, the focus is mainly on rule provisions, without considering the cross-references and hierarchies of rules. Despite its defects, this method recognized the importance that the representation needs to be independent of both the rule engine and the data model, and it used logical rules and facts to achieve that. In general, it requires some knowledge of logic to use but some simple training would suffice. As a result of the analysis above, a summary of the shortlisted representation methods and their capabilities are shown in Table 2. The meaning of numbers is explained in the notes under Table 2. From Table 2, it is evident that no existing representation method has all of the capabilities required to represent building rules, although methods such as CG and RASE check most of the boxes. To minimize the knowledge loss during the rule interpretation and representation processes, a more well-rounded representation method is needed.
representation method needs to be independent of the rule engine and data model, it is envisaged that a multirepresentation method could collaboratively address this issue by mixing and matching representations with different capabilities. To achieve this, it is important to first understand what are the different types of rules and what can be the criteria to distinguish them. Hence, in this section, the authors propose a new classification by analyzing example regulations: Health Building Notes (Department of Health and Social Care 2017) andApproved Documents (Ministry of Housing 2010).
4.1 Manual or automatic checking A frequently mentioned topic in ACC is what type of rules can be checked automatically. Researchers held different opinions regarding this issue, which are ultimately reflected in their classifications. For example, Soliman-Junior et al. (2020) used quantitative/qualitative/ambiguous and ability to be translated into logical rules as criteria for classifying rules, where they claimed rules that cannot be translated into logical rules can only be checked manually. However, as suggested by Zhang et al. (unpubl.), some rules are automatically checkable when they are more thoroughly analyzed. In this paper, the authors argue that only rules that include words or phrases that are subjective (i.e., based on or influenced by personal feelings, tastes, or opinions) cannot be checked automatically. It is because this type of rule is typically open to interpretation. It is thus difficult to have a specific object, or attribute value or value range to check, thereby making it difficult to be automated. For example, Rule 1 shows requirements regarding quality and aesthetics. Different people may have different ideas of “pleasant and welcoming”.
4 A CLASSIFICATION OF BUILDING RULES In Section 3, it was recognized that there is no single representation method that is suitable to represent all types of rules based on the capabilities. As the suitable
461
“The main entrances and reception areas should be pleasant and welcoming.” Rule 1. A subjective requirement, HBN 00-01, Appendix 1, (Department of Health and Social Care 2017)
Other rules are automatically checkable once sufficient information has been provided to define or clarify any insufficiently defined words or phrases (e.g., close to). Notably, although it has been a common belief among many scholars that only rules related to the design stage can be checked, the authors argue that operational rules can also be checked if they are appropriately interpreted. Rule 3 is an operational requirement for the shower seat. This requirement can inform the design and be interpreted as a requirement for adjustable shower seats. “…The position of the shower seat should be adjusted between uses as required.” Rule 2. An operational rule, HBN 00-02, Rule 4.48 (Department of Health and Social Care 2017) 4.2 Rule classifications Based on the analysis of rules that have the potential to be automatically checked, a new classification method is proposed in this section. Incorporating the classification criteria proposed by scholars such as Solihin & Eastman (2015), Macit ˙Ilal & Günaydın (2017) and Hjelseth & Nisbet (2011), the new classification method recognized that a single criterion might neglect other criteria that distinguish rules within the class from each other. Hence, in this paper, we proposed a new classification using four criteria, namely semantic constructs, self-contained or linked explanatory, rule intensity and prescriptive or performance-based. 4.2.1 Semantic constructs Semantic constructs concern the components of rule provisions with specific semantic meanings. Hjelseth & Nisbet (2011) proposed the RASE method, recognizing four semantic constructs of building rules: Requirement (R), Applicability (A), Selection (S) and Exception (E). Developed from their method, the authors included three more semantic constructs here: definitions (optional), outcomes (compulsory) and logical relationships (optional). Definition is closely related to requirements, in the sense that it provides context for requirements and clarification for terms. Studies such as Zhang et al. (2022) stressed the importance of considering a broader context (e.g. definitions and titles) when analyzing rules. Rule 3 shows the definition of the term TER. It is important for accurate calculation of TER to meet energy efficiency requirements. “The Target CO2 Emission Rate (TER) is the minimum energy performance requirement for a new building based on the methodology approved by the Secretary of State in accordance with regulation 25. It is expressed in terms of the mass of CO2 emitted per year per square metre of the total useful floor area of the building.” Rule 3. Definition of TER, Approved Document L2A, Rule 2.2 (Ministry of Housing 2010)
Outcomes are sometimes implicit in rule texts. They can take many forms. They can be as simple as pass or fail. Taking an open-world assumption (Hustadt 1994), they can also be unknown when the information is not sufficient to make a judgement. They may include other actions and side-effects such as “adding 5 credits (points)” in BREEAM (Building Research Establishment 2018), where the final credits and rating are based on the accumulation of credits awarded in different evaluation aspects. Note that all rules have compulsory semantic constructs, but each rule can have different types of optional semantic constructs. As a result, one rule can have more or less semantic constructs than another rule. For example, the rule constructs of Rule 4 are marked below. The underlined part is the requirement. The part in bold is the selection, which is slightly different from “applicability” as a selection offers alternative subjects. It thus applies to both staff use and patient use. “Help call should be provided if the room is for staff or patient use.” Rule 4. HBN 00-02,Rule 3.24, Semantic constructs in a rule (Department of Health and Social Care 2017) Logical relationships concern the logical connectives between words or phrases in rules. The same set of words and phrases, if linked using different logical connectives, can have very different meanings. Logical relationships mainly include “and”, “or” and “not”, meaning logical conjunction, disjunction and negation, respectively. 4.2.2 Self-contained or linked explanatory Another criterion is self-contained and linked explanatory. Initially proposed by Macit ˙Ilal & Günaydın (2017), this criterion has a different meaning here. Self-contained means that a rule provision is selfexplanatory (i.e. does not rely on other rule provisions). Linked explanatory rules refer to rules that do not have complete meanings themselves. They have to be looked at together with other rules. A typical example of linked explanatory rules is cross-references. Rule 5 shows a rule provision with cross-reference. It means that paragraph 5.27 needs to be consulted to make sure the rule is fulfilled. “A landing should be provided at the top and bottom of each flight of stairs. The minimum clear landing depth is 1200 mm but must equal the clear stair width between handrails (see paragraph 5.27)” Rule 5. A rule provision with cross-reference, HBN 00-04, Rule 5.4 (Department of Health and Social Care 2017) 4.2.3 Prescriptive or performance-based rules Prescriptive rules often present requirements explicitly, meaning that the path to compliance is clearly stated (e.g. the required properties, objects, and relationships). However, for performance-based rules,
462
only the expected performance was asked, rather than how it can be achieved. The designers need to use their expertise to make sure the design meets the performance requirements. Performance-based rules are also not straightforward to check, as they either require domain experts to make the implicit knowledge explicit by rule interpretation or computationally intensive simulations. Rule 6 provides the required air change rate. It requires ventilation knowledge to interpret this rule into a machine-readable rule. “The room will be considered fit for purpose if, with the ventilation system operating and all doors closed, the following parameters are achieved: the patient’s room has an air change rate of at least 10 per hour;” Rule 6. A performance-based rule, HBN 04-01 Supplement 1, Rule A2.12 (Department of Health and Social Care 2017) 4.2.4 Rule intensity Many studies have mentioned the difficulty of automatically checking building rules (Eastman et al. 2009), including: 1) rules written in human language can be ambiguous, subjective and thus hard to interpret and represent using machine-readable representations, and 2) some rules require intensive computations. Since the term “rule complexity” has been adopted to describe rules with one or more of the abovementioned characteristics without a consensus by academics and practitioners, the authors propose the term “rule intensity” in this paper. Compared with the multiple facets of “rule complexity”, rule intensity only concerns the computational power required to check the rule automatically. Based on this, rules can be classified into three categories: 1) No calculations or simple calculations 2) Functions or algorithms 3) Simulations No calculations or simple calculations means the requirement is straightforward and does not require complex calculation. For example, rules that only require checking the existence of objects or the property value of an object fall into this category (Rule 7). Rules that include only simple arithmetic calculations and no spatial or physical calculations also fall into this category. “Bidets should be fitted with a sensor-operated over-rim supply.” Rule 7. A rule checking the existence of simple objects, HBN 00-02, Rule 2.3 (Department of Health and Social Care 2017) Rules with slightly higher intensity are included in the functions or algorithms category. This type of rule may involve combinatorial issues that deal with multiple objects and possibilities to compliance (Solihin & Eastman 2015). For example, Rule 8 requires clear space around the toilet on the bath side to allow the accessibility of mobile hoist.
“The room layout utilises the minimum clear space requirement to the side of the toilet for mobile hoist transfer (that is, 1150 mm from the centreline of the toilet to the nearest obstruction), on the bath side of the toilet only.” Rule 8. A rule that requires function/ algorithm to check, HBN 00-02, Rule 2.22 (Department of Health and Social Care 2017) Rules with the highest intensity are rules that require simulations. Many rules in this category are performance-based rules and they are typically found in fire codes, energy requirements, etc. They typically ask for a proof-of-solution or a feasible design to achieve performance requirements. Rule 9 asks for a design to satisfy the energy performance requirement as set by regulation 24. The approved software suggested will run simulations to help generate energy performance calculations and predictions. “The TER must be calculated using one of the calculation tools included in the methodology approved by the Secretary of State for calculating the energy performance of buildings pursuant to regulation 24.…” Rule 9. A rule that requires simulation, Approved Document L2A, Rule 2.3 (Ministry of Housing 2010)
5 PROPOSED RULE REPRESENTATION METHOD 5.1 Representation selection for different types of rules Given the evaluation in Section 3.2 and the analysis of rules in Section 4, suitable representations can be selected by comparing the rule characteristics and the capabilities of representations (Figure 1). As the combination of different criteria for classifying rules can generate too many rule types, the authors only provide several examples here to demonstrate the process. For example, we can regard that a rule has requirements, definitions, and a checking outcome that includes other side-effects as Rule Class A. For Rule Class A, RASE would be the most suitable representation, as only RASE have both the capabilities of representing definitions and side-effects. There could also be a Rule Class B that has a requirement of level 3 rule intensity. In this case, conceptual graph could be the most suitable representation, as CG is easier to use when the rule intensity is high. 5.2 A multi-representation method for all types of building rules A suitable rule representation method should be capable of representing all types of rules. To achieve this, three representation methods, namely RASE, predicate logic and conceptual graph, were selected to cover all capabilities, which forms a multi-representation
463
method. Figure 2 shows when to select which representation. RASE can deal with most of the required capabilities, except it does not have explicit logical relationships and does not make the implicit knowledge and assumptions explicit. It is ideal for larger scale documents because the time-saving advantage of markups is especially evident when dealing with many documents. Predicate logic deals with specific problems and individual rule provisions. It has a sound theoretical base, and it is very suitable for representing building rules. Conceptual graphs come into play when the rule intensity is high and the predicate logic statements get lengthy and hard to read.
6 DISCUSSION It came with no surprise that from the evaluation of representations, no single representation meets all the 16 requirements. If our analysis is right, the proposed multi-representation method would address this issue. Nevertheless, the evaluation result shows several issues that need to be aware of when developing a new representation method.
demonstration and easier understanding, only rules with relatively low intensity were selected. This may result in the underestimation of rule intensity (Solihin & Eastman 2015) or sacrifice the completeness of ACC system. Which rules require manual checking are often decided arbitrarily without sufficient evidence. The authors’experience shows that rules that are seemingly “uncheckable” byACC systems can be checkable when being carefully interpreted. 4) The lack of attention to rule organization Rule organization includes hierarchy and crossreferences. While many methods have the capability to present cross-references, they rarely consider hierarchy of rules. Hierarchies may affect checking outcomes as they denote different constraint levels of regulation documents and the superiority and inferiority of regulations. 5) The dependency on model data and rule engine It seems that because BIM model is the most commonly used data model, most representations use objects, attributes and relationships in IFC directly to avoid the need to map among different ontologies. As a result, although ACC can be done, the representation is dependent on rule engine and building model data. This limits the expressiveness of representations and makes the representations difficult to be updated when regulations are revised.
7 CONCLUSION
Figure 2. method.
Selection diagram for the multi-representation
1) The neglection of the broader context of rules As shown in Table 2, only two methods can deal with definitions. Most methods lack the consideration of including the broader context such as regulation document titles, section descriptive sentences and definitions. However, these texts provide valuable context such as the applicability of rule provisions. 2) Only considering binary outcomes Existing representations showed a general inability to represent outcomes other than “pass” and “fail” (except RASE). As the regulations become more complex, it is vital that representations are equipped with “unknown”, other actions and side-effects to represent as many types of rules as possible. 3) Only dealing with rules with low intensity In the current research and practices, scholars and practitioners seem to rush to develop an ACC system for proof-of-concept or implementation. For clearer
This paper evaluated existing representation methods for building rules. It proposed a new classification method of rules and developed a multi-representation method to represent different types of rules. To the best of the authors’ knowledge, this is the first research that mapped different types of rules to their suitable representation and proposed a representation method that meets all of the required capabilities. The method can help check more rules automatically, thereby improving the quality and efficiency of the design review process. This paper inevitably has limitations. Firstly, more representation methods can be evaluated, and more suitable representations may be found. Secondly, the evaluation results might not be 100% accurate, as 1) representations were evaluated based on the descriptions in corresponding papers without implementing the method; and 2) many representations are still being developed. Thirdly, the method is proposed but not tested. The testing of this proposed representation will be included in our future research. REFERENCES
464
Beach, T. H., Rezgui, Y., Li, H. & Kasim, T. 2015. A Rulebased Semantic Approach for Automated Regulatory Compliance in the Construction Sector. Expert Systems with Applications 42: 5219–5231. Building Research Establishment 2018. BREEAM New Construction 2018 (UK).
Department of Health and Social Care 2017. DH Health Building Notes. In: CARE, D. O. H. A. S. (ed.). Eastman, C., Lee, J.-m., Jeong, Y.-s. & Lee, J.-k. 2009. Automatic Rule-based Checking of Building Designs. Automation in Construction 18: 1011–1033. Hjelseth, E. & Nisbet, N. Capturing Normative Constraints By Use of the Semantic Mark-up RASE Methodology. 2011 2011. 1–10. Hustadt, U. Do We Need the Closed World Assumption in Knowledge Representation? Knowledge Representation Meets Databases, 1994 1994. Macit ˙Ilal, S. & Günaydın, H. M. 2017. Computer Representation of Building Codes for Automated Compliance Checking. Automation in Construction 82: 43–58. Ministry of Housing, C. L. G. 2010. Approved Documents. Pauwels, P., Van Deursen, D., Verstraeten, R., De Roo, J., De Meyer, R., Van de Walle, R. & Van Campenhout, J. 2011. A Semantic Rule Checking Environment for Building Performance Checking. Automation in Construction 20: 506–518. Preidel, C. & Borrmann, A. 2016. Towards Code Compliance Checking on the Basis of A Visual Programming Language. Journal of InformationTechnology in Construction 21: 402–421. Rasdorf, W. J. & Lakmazaheri, S. 1990. Logic-based Approach for Modeling Organization of Design Standards. Journal of Computing in Civil Engineering 4: 102–123. Solibri. 2022. Solibri Model Checker [Online]. Available: https://www.solibri.com/solibri-office [Accessed 31/08/2021 2021]. Solihin, W. Lessons Learned From Experience of Codechecking Implementation in Singapore. BuildingSMART Conference, 2004 Singapore. BuildingSMART.
465
Solihin, W. & Eastman, C. 2015. Classification of Rules for Automated BIM Rule Checking Development. Automation in Construction 53: 69–82. Solihin, W. & Eastman, C. 2016. A Knowledge Representation Approach in BIM rule Requirement Analysis Using the Conceptual Graph. Journal of Information Technology in Construction 21: 370–402. Soliman-Junior, J., Formoso, C. T. & Tzortzopoulos, P. 2020. A Semantic-based Framework for Automated Rule Checking in Healthcare Construction Projects. Canadian Journal of Civil Engineering 47: 202–214. Sowa, J. F. 1976. Conceptual Graphs for a Data Base Interface. IBM Journal of Research and Development 20: 336–357. Tan, X., Hammad,A. & Fazio, P. 2010.Automated Code Compliance Checking for Building Envelope Design. Journal of Computing in Civil Engineering 24: 203–211. Vaishnavi, V. K. 2007. Design Science Research Methods and Patterns: Innovating Information and Communication Technology. Auerbach Publications. Zhang, J. & El-Gohary Nora, M. 2017. Semantic-Based Logic Representation and Reasoning for Automated Regulatory Compliance Checking. Journal of Computing in Civil Engineering 31: 04016037. Zhang, Z., Ma, L. & Broyd,T. 2022.Towards Fully-automated Code Compliance Checking Of Building Regulations: Challenges For Rule Interpretation And Representation. 2022 European Conference on Computing in Construction. Ixia, Rhodes, Greece. Zhang, Z., Nisbet, N., Ma, L. & Broyd, T. [No date]. Capabilities of Rule Representations for Automatic Compliance Checking. Unpublished.
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
Processes
Taylor & Francis Taylor & Francis Group http://taylorandfrancis.com
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
Development of augmented BIM models for built environment management L. Binni, B. Naticchia, A. Corneli & M. Prifti Polytechnic University of Marche, Ancona, Italy
ABSTRACT: Traditional surveys in the built environment are time-consuming and usually result in enormous amounts of data that are difficult to manage and contain biases. Therefore, BIM modeling of both geometries and related information leads to inconsistent and incomplete models. Alternatively, preliminary BIM models result overly detailed. In both scenarios, the models must be reworked later slowing down the design process. The proposed methodology combines point cloud surveying technique, photogrammetry, and BIM within a game engine platform to define a workflow for an incremental model semantic enrichment that leads to an augmented BIM environment. The case study prototype allows stepwise accurate integration of detailed BIM objects by easing positioning them in the scene in accordance with the overlapped aligned images. This approach provides a way to enrich the BIM model only when required avoiding reworks, reducing working time and costs.
1 INTRODUCTION Creating a Building Information Model (BIM) of an existing building, commonly referred to as an as-is BIM, requires the acquisition of the building’s actual state. This has been challenging since the emergence of BIM in the architecture, engineering, construction, and operation (AECO) industry. Despite the boost to the use of BIM provided also by the Italian legislation, most existing buildings are still not maintained, renovated or deconstructed through BIM (Macher et al. 2017). Major building stocks owners and asset managers (e.g., banks, governments, investment funds) often find themselves in need of digitizing their assets without compromising daily functions. In many cases even the mere access to the premises is a source of disruption to functionalities, and it is therefore desirable that this lasts as short a time as possible. Concurrently, another aspect related to building digitization that must be taken into account concerns the information that needs to be modeled. Modeling information is indeed a time-consuming and therefore cost-producing process (Wang & Kim 2019; Tang et al. 2022). Thus, extensive information modeling of a building has proven to be a losing approach. For this reason, uses of BIM that enable focused modeling have been established. However, this must coordinate with the sometimes limited ability to access buildings, which dictates the need for massive data collection so as not to interfere repeatedly with asset functions. Thus, keeping in mind the twin instances of creating as little disruption to building functions as possible DOI 10.1201/9781003354222-60
and at the same time collecting as much information as possible, it is necessary to take advantage of efficient and extensive data collection techniques. Laser scanning is a widely used technique to achieve this goal. Nevertheless, laser scanners allow the collection of information about the geometry of objects in the form of Point Clouds (PCs) that are devoid of a whole range of information. In fact, the scan-to-BIM process remains largely a manual process on the one hand because of the huge amount of data and on the other hand because of the absence of semantic information within the PCs (Wang & Kim 2019). One datum still little exploited for building modeling are spherical images. These lack metric information but allow semantic information about assets to be derived even a posteriori. This research aims to define a process for creating as-is augmented BIM models through the combination of PCs and spherical images which leads to an incremental modeling approach.
2 BACKGROUND 2.1 Building surveying Building surveying is considered as the process of acquiring information about the built environment. Retrieval of all required information in its correct representation is one of the challenges in all information flow management processes (Eastman et al. 2009; Preidel & Borrmann 2015). Construction typology, materials, geometry, and condition are examples of information that could be retrieved.
469
Depending on tools and procedures two different surveying techniques are commonly used nowadays: direct and indirect building surveying. The direct approach consists in the use of simple measurement tools, such as metric distance measurers, plumb bob, and so forth, that allows in-site geometric measurements.This method leads to a time-consuming and error-prone drawing process. Furthermore, manually measured data cannot be automatically linked to the BIM platform, which will cause additional data conversion steps and take up another data storage space (Xu et al. 2020). Indirect approaches differ from direct approaches since they enable off-site measurement. Consequently, they are faster, more economical, and accurate then direct methods. Most of them are currently based on photogrammetry and 3D scanning. Both photogrammetry and 3D scanning lead to the creation of PCs. The taxonomy of BIM information distinguishes geometric, semantic, and topological types (PrattHartmann 2004; Schlueter A. & Thesseling F 2009). Geometric information directly relates to the shapes and forms of facilities, whereas semantic information captures their intrinsic properties (e.g., functionality), and topological information gives the relationships among these objects (Xue et al. 2021). The scan-to-BIM process involves three tasks: modeling component geometry, assigning an object category and material properties to a component, and defining the relationships between components (Macher et al. 2017). Among these tasks, however, only geometric modeling is best supported by point cloud processing, for which there are several algorithms for identifying planes, surfaces, and geometric shapes. Conversely, the semantic component of the information is often more difficult to recover relying only on the PC, which, especially if not combined with colors, makes the identification of building components complicated, above all if this is done retrospectively. 2.2 Spherical images use in construction industry (Pereira & Gheisari 2019) stated that overall, three main areas employing 360-degree panoramic technology in the construction industry have been identified: interactive learning, reality background to augmented information, and visualization of safe and unsafe situations. Most applications of spherical photos involve visualizing the current state of buildings and overlaying them on existing BIM models (Dave et al. 2018). Research efforts in the construction industry have focused on using 360-degree panoramas as a backdrop of reality to augment the information they contain (e.g., 3D models, 2D images, audio, text). Cote et al. (2013) used 360-degree panoramas to visualize the surface of a road, where augmentations (e.g., 3D models, images, annotations) show a virtual excavation and illustrate underground services on it.
Eiris et al. (2017) described the process of using 360-degree panoramas to create a virtual tour within complex construction projects for asset management and documentation. In contrast, Barazzetti et al (2020). describe the workflow implemented for deterioration mapping based on 360 images, which highlights pathologies on surfaces and quantitatively measures their extent. However, the opportunity to capture the entire scene around a 360-degree camera has not yet been fully exploited. Three-hundred-and-sixty-degree cameras allow non-expert to quickly record large sites with multiple spaces because the entire scene around the photographer is captured. This in fact reduces the number of images needed in a condition mapping project, making it an effective solution for quickly documenting environments (Barazzetti et al. 2020). Spherical images can be exploited as a great support of modeling since all components of the environment can be recognized in the photos, and therefore even at a distance of time it is not complicated to interpret the data. On the other hand, 360-degree images lack any measurement data and therefore it is difficult to use them alone to support modeling. This research project aims to develop a process for the combined exploitation of point clouds and spherical images to support BIM building modeling.
3 METHODOLOGY 3.1 Workflow The followed approach establishes a professional workflow (Figure 1). Since images implicitly contain semantic, they can be used to enrich low-level information given by a basic geometric model. The first matter to be addressed in correctly superimposing spherical images to a 3D model within a Game Engine (GE) concerns the spatial position and orientation of the camera at the time the image was captured, which is referred to as Camera Pose Estimation (CPE). In this paper, CPE is addressed through the development of a resection model followed by an alignment model. Then, the model enables semantic enrichment through the addition of BIM objects and information, the historicization of multiple surveying phases, and the measurement intended as the possibility of digital surveying. Paragraph 3.2 introduces the CPE issue, while 3.2.1 and 3.2.2 respectively focus on the methodologies for the resection and alignment models. Paragraphs 3.3 presents the methodology for the semantic enrichment of the model. 3.2 Camera pose estimation Different methods have been introduced in photogrammetry and computer vision to estimate the pose of the camera starting from equirectangular or perspective images. Among those, the pose is usually solved as the Perspective n-Point (PnP) problem. Solutions
470
Figure 1. Workflow for an augmented BIM model exploiting spherical images and 3D scanning surveying techniques. The implementation of the workflow in a real case study is presented in section 4. Accordingly, the processes required to define the Initial Scene and to use the Resection & Alignment algorithm are discussed in section 4.1.1, while the Model enrichment in the Augmented BIM model is discussed in section 4.1.2.
of the PnP problem are mainly developed with three reference points (P3P) as in (Yu et al. 2021) or four reference points (P4P) as in Hu & Wu (2002). The study of the PnP problem mainly consists of two aspects: designing fast and stable algorithms to find solutions, and classify the results finding conditions under which the problem has either one or more solutions. While the first aspect has many results, the second is still open. The smallest subset of the PnP problems that yields a discrete number of solutions is the P3P (Gao et al. 2003), that is the most basic case and represents the special case in all other PnP problems (n > 3). However, direct methods for PnP lead to more than one solution and do not consider the redundancy in observations which should strengthen the solution from a statistical point of view as explained by Alsadik (2016). Besides, more recent approaches are based on the use of angular conditions and are referred as Resection models. By considering redundancy in observation and being robust to improper initial values, those introduce a stable and reliable solution that converges to a global minimum even with wrong preliminary values. An approach based on oblique angles and vectors geometry that defines a robust Resection model was developed to estimate the camera pose by Alsadik (2020). Once the camera has been shifted in its actual position by the translation vector t, equirectangular images can be aligned to the three-dimensional model by finding the optimal rotation. To preserve shape and size, a Euclidean transformation must be applied. In fact, Euclidean isometry allows geometric transformation of the space through rotations and translations that preserve the distances between each pair of points included in the considered sets. A rigid Euclidean transformation of a vector v is generally defined as:
where T (v) is the transformed vector; R is an orthogonal transformation matrix; and t is the vector which gives the translation of the origin. In this paper, a Resection model based on the one developed by Alsadik (2020) is built to find the optimal camera pose within the GE platform i.e. Unity. Then, equirectangular images are aligned with the model by means of the method demonstrated by Ho (2011).
T(v) = R × v + t
where col i,j are the x-coordinates in pixels for the control points i and j; rowi,j are the y-coordinates in pixels
(1)
3.2.1 Resection model Since equirectangular images are exploited to enrich the model within the GE, the Resection model required to find the camera pose is built following the approach defined in (Alsadik 2016, 2020). This model is based on angular conditions represented by oblique angles instead of horizontal and vertical angles. Spherical trigonometry laws are used to derive oblique angles while vector geometry is applied to relate the derived angles, the camera unknown orientation parameters, and the coordinates of the object points. Three points are necessary to define the inclined angle in the orientation problem and a minimum of three oblique angles is necessary to define the object space coordinates of the camera. In this paper four reference points are used. Firstly, equirectangular images of the case study are taken, and vertical (β i , β j ) and horizontal (θ) angles relative to the first two reference points i and j within a spherical triangle are computed as follow: θ = pixel ◦ × colj − coli βi,j = pixel ◦ ×
471
"
image height 2
(2) #
− row i,j
(3)
for the control points i and j; and pixel˚ is the pixel size in degree. Then, the oblique angle γ between the two points can be computed from the cosine rule: cos γ = cos θ × cos βi × cos βi + sin βi × sin βi (4) Since four reference points are considered, the process above is repeated for six times to find six oblique angles. Those can now be derived using the approximate camera 3D coordinates applying the vector dot product: cos γ =
dxi × dxj + dyi × dyj + dzi × dzj Li × L j
(5)
requires a minimum of three reference points to be solved. In this work, four reference points are used to strengthen the stability of the solution and the convergence to an optimal minimum as shown in the experimental tests made in (Alsadik 2016, 2020). Equation 1 can be expressed in a matrix form as follow: B=R × A + t
where A and B are the matrixes of the 3D points with known correspondences. The first step is finding the centroids through Equations 11a, 11b.Accordingly, the covariance matrix ( ) can also be computed as shown by Equation 12:
where dxi,j , dyi,j , and dz i,j are the differences in coordinates between the camera and the control points i and j; and Li and Lj are the spatial distances between the camera and the observed control points i and j. Equation 5 is reformulated as the Equation 6 to further compute the partial derivates relative to unknown coordinates of the camera (∂F / ∂X P ), (∂F / ∂Y P ), and (∂F / ∂Z P ) through the system of equations 7. F = Li × Lj × cos γ − dxi × dxj + dyi × dyj + dzi × dzj (6) ⎧ cos γ × XP − Xj × Li ∂F ⎪ ⎪ = Xi − 2XP + Xj + ⎪ ⎪ ∂X Lj ⎪ ⎪ ⎪ P ⎪ ⎨ ∂F cos γ × Y − Yj × Li = Yi − 2YP + Yj + ⎪ ∂YP Lj ⎪ ⎪ ⎪ ⎪ ⎪ cos γ × Z ⎪ ∂F P − Zj × Li ⎪ ⎩ = Zi − 2ZP + Zj + ∂ZP Lj (7)
The solution is applied by least square adjustment as shown in Equation 8 in a matrix form to compute the vector of corrections to the camera coordinates
= (δ Xp , δ Yp , δ Zp ): −1
= B × Bt × Bt × F
(8)
Where B =[(∂F / ∂X P ), (∂F / ∂Y P ), (∂F / ∂Z P )] is the matrix of partial derivates to the camera coordinates XP , YP ,ZP . Furthermore, the covariance matrix ( ) of the camera coordinates is computed as: −1
= σ02 × B × Bt
(10)
centroid A =
N 1 i A n i=1
(11a)
centroid B =
N 1 i B nr i=1
(11b)
= (A − centroid A ) × (B − centroid B )T
(12)
where Ai and Bi are vectors [3×1] relative to the coordinates of the points [x, y, z]. Then, the optimal rotation is computed by means of the Kabsch algorithm, Kabsch (1978), that minimizes the Root-Mean-Square Deviation (RMSD) between two paired sets of points. The RMSD is the measure of the differences between values predicted and observed. The Kabsch algorithm is based on the Singular Value Decomposition (SVD) that is a factorization of a matrix into three other matrices. SVD (M) = [U, S, V]
(13)
M =U × S × VT
(14)
where M is the factorization m × n matrix; U is the m × m complex unitary matrix; S is the m × n complex unitary matrix; and V is the n × n complex unitary matrix. The image rotation problem can be now solved by decomposing the covariance matrix as follow: SVD ( ) = [U , S, V ]
(15)
R = V × UT
(16)
The result is a [3×3] matrix. In our case study, the solution is expressed in terms of quaternions.
(9)
3.2.2 Model-image alignment model The alignment between two sets of corresponding 3D point data can be achieved by finding optimal rotation and translation (Arun et al. 1987). Since the translation vector of the camera has already been computed, the rotation matrix must be calculated. This problem
3.3 Incremental model semantic enrichment Superimposing and aligning equirectangular images to the three-dimensional model of the environment means building a virtual environment inside which the model itself can be upgraded to progressively define a more detailed BIM model. Consequently,
472
this procedure allows avoiding over detailed preliminary modeling that always leads to reworks. Practically, the semantic enrichment of the model can be performed exploiting Unity prefabs in the scene. Prefabs are reusable instances that can be either created directly into the GE or imported elements such as IFC files. Prefabs allow creating, configuring, and storing objects with all their components and properties. Prefab instances can be gradually added in the scene and positioned in their actual location. In this work, the incremental model semantic augmentation is performed by importing and positioning in the scene five BIM objects (saved as IFC files) as prefab instances. To define a stepwise workflow, the prototype saves the entire model as an IFC file, that can be subsequently opened to apply further modifications, additions, or validations always within the GE (Figure 1).
4 SYSTEM IMPLEMENTATION 4.1 Application development The application is developed by means of a case study that concerns the Construction section of the Department of Civil and Building Engineering, and Architecture (DICEA) of the Polytechnic University of Marche.The development follows the workflow shown in the previous section. 4.1.1 Virtual environment Following the approach shown in Figure 1, PC surveying technique is firstly performed using the ZEB Horizon GeoSLAM to derive a basic geometric 3D model composed of simple elements such as surfaces representing walls and floors (Figure 2). The basic three-dimensional model of the DICEA department is then imported into Unity.
Figure 2. The PC of the DICEA Department and the basic 3D model built upon the PC.
Then, a photographic surveying is performed using the Ricoh Theta 360◦ panoramic camera, capturing 23 equirectangular images of the department: 1 image for smaller rooms, 2 images for larger rooms, and 7 images along the hallway. Each equirectangular image is imported in the Unity scene as a material (Figure 3).
Figure 3. Three equirectangular images imported into the GE as new materials.
The Unlit Shader is assigned to the material to directly exploit the natural lighting in the images that do not require additional light sources. These materials are then applied to the interior surface of spheres with arbitrary dimension, such that the camera can be subsequently fitted into them and approximately placed around the actual image capture position. Since the default sphere mesh known as UV sphere is basically a rectangular grid wrapped around a sphere shape and compressed into a single point at the poles, equirectangular mapping would cause distortions due to the high concentration of triangles at the poles. For this reason, a polyhedric octahedron sphere is used. This kind of mesh is composed of congruent and regular polygonal faces and does not lead to distortions (Flick 2022). At this point, the initial scene (Figure 1) has been defined. Then, the CPE algorithm based on both the resection and model-image alignment models is built to accurately overlap the geometric and photographic environments. The algorithm must be fed with 4 reference points (4 points in the image and the corresponding 4 points in the 3D model). The process consists in manually selecting a total of eight points in the current scene as the example shown in Figure 4. The algorithm primarily corrects each sphere’s position by means of the vector of corrections to the camera coordinates ( ). Since the camera is positioned at the origin of a sphere, correcting the sphere’s position means translating the camera to the actual image capture position. Once the translation is executed, the algorithm aligns the equirectangular image to the 3D model in the background depending on the rotation matrix R. The location of the spheres and the orientation of the images are then saved. A radar map of the DICEA department is included in the Graphical User Interface (GUI) to navigate the model i.e. to shift the camera among the spheres to visualize the different rooms (Figure 5). The camera can rotate around its origin and zoom the field of view for a more detailed visual investigation of the environments. 4.1.2 Enrichment Once the camera can be moved among the different environments of the model, and equirectangular images are aligned with the three-dimensional geometry, the case study model is enriched with five BIM objects imported as prefabs into Unity (Figure 5). The BIM objects considered in this work are a venting unit,
473
Figure 4. Alignment procedure: manual selection of the first out of four reference points (in this example the visible upper vertex of the column is selected both on the image and on the 3D model).
Figure 5. Graphical User Interface. The view can be switched between the images and the model (upper images). Transparency is given to the lower image to show the accuracy of the alignment. (A) map for navigating the model; (B) BIM objects instancing menu; (C) two of the five IFC elements imported into the GE as prefabs: Radiator and LED light prefabs; (D) two of the four reference points used for the resection and alignment processes of the current scene; (E) distance measurement result tab.
a radiator, a light switch, an electrical outlet, and a LED light. An algorithm based on the Unity RaycastHit structure, snaps, and vector differences is developed. The RaycastHit structure is used to retrieve information back from a ray projected towards an object in the scene. Snaps can be used to locate the object in the environment by selecting a reference point (a vertex of the prefab). Objects can also be rotated in the scene to find the optimal position. The enriched scene can be saved as an IFC file to preserve modifications. The updated file can be
imported again in the GE if further modifications or validations are required. Figure 6 shows the enrichment process on the GUI. Besides, measurements can be taken by clicking two points in the image and the results are shown in the GUI (Figure 5). 5 CONCLUSIONS Based on the combination of PC surveying technique, photogrammetry, and BIM within a GE, this paper
474
Figure 6. Model semantic enrichment process: adding a radiator into the scene. (1) and (2) comparison between the spherical image and the model; (3) enriched scene.
proposes a professional workflow that allows the creation of a virtual environment that consists in an augmented as-is BIM model. This workflow has been implemented through the development of a professional application within the GE Unity, which exploits spherical images and PCs. The prototype, by means of the Resection and Alignment algorithm implemented in this work, firstly allows the alignment of the 360◦ images, the PC, and a basic 3D model built upon the PC. Then, the aligned virtual environment can be enriched within the same application by adding additional BIM objects that can be imported in the scene as Unity prefabs. The BIM enrichment is both accurate and made extremely easy as it involves placing those prefabs in the scene only by selecting on the image one reference vertex of the object you wish to add. The augmented model can be saved as an IFC file and imported again for incremental enrichment. Provided that the geometry of the environment does not change dramatically during the time, the proposed virtual environment establishes a repository of photographic data that can be used should the need arise either for upgrading the BIM model progressively or as an aligned photographic collection of stages of processing in construction sites. Furthermore, the prototype enables digitized measurements of the model exploiting the collected spherical images. This could be convenient in the measurement of voids, which is a complex and error-prone procedure to perform using a PC only. The application implemented following the proposed methodology achieves the following results:
in the scene (i.e. BIM objects) only when required. This stepwise process allows avoiding overdetailed preliminary modeling. – The aligned environments allow a direct detection of modeling errors in the scenes. – Surveying historicization, by superimposing and aligning equirectangular images taken in consecutive processing steps to the BIM model. – Digitized measurements within the augmented scene. These results highlight the ability of the proposed model to accelerate and facilitate the surveying and modeling phases in the management of the built environment, reducing working errors, time, and costs. Thus, the tool is directly applicable in the AECO industry. Field tests will be conducted to certify the efficiency of the prototype and that time, costs, and errors of the preliminary building surveying and modeling stages will be significantly reduced compared to traditional methods. Future development should consider the following improvements: – Implementation of the proposed prototype exploiting Mixed Reality (MR) technologies. – Implementation of the application within Common Data Environments for modeling step historicization. REFERENCES
– BIM model progressive semantic enrichment. The model can be gradually upgraded adding prefabs
475
Alsadik, B. 2016. A Modified Method for Image Triangulation Using Inclined Angles. ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLI-B3: 453–458.
Alsadik, B. 2020. Robust Resection Model for Aligning the Mobile Mapping Systems Trajectories at Degraded and Denied Urban Environments. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 1: 215–222. https://doi.org/10.5194/isprsannals-V-1-2020-215-2020 Arun, K. S., Huang, T. S. & Blostein, S. D. 1987. LeastSquares Fitting of Two 3-D Point Sets. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-9(5): 698–700. Barazzetti, L., Previtali, M. & Scaioni, M. 2020. Procedures for Condition Mapping Using 360˚ Images. ISPRS Int. J. Geo Inf., 9, 34. Côté, S., Trudel, P., Desbiens, M., Giguère, M. & Snyder, R. 2013. Live Mobile Panoramic High Accuracy Augmented Reality for Engineering and Construction. Proceedings of the Construction Applications of Virtual Reality (CONVR), London, England. Dave, B., Buda, A., Nurminen, A. & Främling, K. 2018. A Framework for Integrating BIM and IoT Through Open Standards.Automation in Construction, 95, 35–45. Eastman, C., Lee, J., Jeong, Y. & Lee, J. 2009. “Automatic Rule-based Checking of Building Designs.” Automation in construction, 18(8), 1011–1033. Eiris Pereira R., Moud H.I. & Gheisari M. 2017. Using 360-Degree Interactive Panoramas to Develop Virtual Representation of Construction Sites. In: Proc. Lean & Computing in Construction Congress (LC3), Vol. 3 (CONVR), Heraklion, Greece, 775–782. Flick, J. 2022. Retrieved from Creating an Octahedron Sphere in Unity. binPress. https://www.binpress.com/creatingoctahedron-sphere-unity/ Gao, X.-S., Hou, X.-R., Tang, J. & Cheng, H.-F. 2003. Complete Solution Classification for the PerspectiveThree-Point Problem. Pattern Analysis and Machine Intelligence, IEEE Transactions On, 25: 930–943. https://doi.org/10.1109/TPAMI.2003.1217599 Ho, N. 2011. Retrieved from Finding Optimal Rotation and Translation Between Corresponding 3D points. Nghia Ho. http://nghiaho.com/?page_id=671#comment-469 Hu, Z. Y. & Wu, F. C. 2002. A Note on the Number of Solutions of the Noncoplanar P4P problem. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(4): 550–555. https://doi.org/10.1109/34.993561 Kabsch, W. 1978. A Discussion of the Solution for the best Rotation to Relate Two Sets of Vectors. Acta Crystallographica Section A, 34(5): 827–828.
Luhman, T., et al. 2014. Close-Range Photogrammetry and 3D Imaging. Berlin: Walter de Gruyter GmbH. Macher, H., Landes, T. & Grussenmeyer, P. 2017. From Point Clouds to Building Information Models: 3D SemiAutomatic Reconstruction of Indoors of Existing Buildings. Applied Sciences, 7, 1030. Pereira, R.E. & Gheisari, M. 2019. 360-Degree Panoramas as a Reality Capturing Technique in Construction Domain: Applications and Limitations. Pratt-Hartmann I. 2004. Fragments of Language, J. Logic Lang. Inform., 13 (1), 207–223, https://doi.org/10.1023/ B:JLLI.0000024735.97006.5a Preidel, C. & Borrmann, A. 2015. Automated Code Compliance Checking Based on a Visual Language and Building Information Modeling. ISARC. Proceedings of the International Symposium on Automation and Robotics in Construction. 10.13140/RG.2.1.1542.2805. Schlueter A. & Thesseling F. 2009. Building Information Model Based Energy/exergy Performance Assessment in Early Design Stages, Automation in Construction, 18, (2), 153–163, https://doi.org/10.1016/j.autcon.2008.07.003. Tang, P. Huber, D. Akinci, B. Lipman, R. Lytle, A. 2010. Automatic Reconstruction of As-built Building Information Models From Laser-scanned Point Clouds: A Review of Related Techniques. Automation in Construction, 19, 829–843. Tang, S., Li, X., Zheng, X., Wu, B., Wang, W., & Zhang, Y. 2022. BIM Generation from 3D Point Clouds by Combining 3D Deep Learning and Improved Morphological Approach. Automation in Construction, 141, 104422. https://doi.org/10.1016/J.AUTCON.2022.104422 Wang, Q., & Kim, M.-K. 2019. Applications of 3D Point Cloud Data in the Construction Industry: A Fifteen-year Review From 2004 to 2018. Advanced Engineering Informatics, 39, 306–319.https://doi.org/https://doi.org/10. 1016/j.aei.2019.02.007 Xu, Z., Kang, R. & Lu, R. 2020. 3D Reconstruction and Measurement of Surface Defects in Prefabricated Elements Using Point Clouds. Journal of Computing in Civil Engineering, 34(5), 04020033. Xue F., Wu L. & Lu W. 2021. Semantic Enrichment of Building and City Information Models: A Ten-year Review. Advanced Engineering Informatics, 47, 101245. Yu, Q., Xu, G., Dong, W. & Wang, Z. 2021. Solving the Perspective-three-point Problem With Linear Combinations: An Accurate and Efficient Method. Optik, 228, 165740. https://doi.org/https://doi.org/10.1016/j.ijleo. 2020.165740
476
ECPPM 2022 – eWork and eBusiness in Architecture, Engineering and Construction – Hjelseth, Sujan & Scherer (Eds) © 2023 the Author(s), ISBN 978-1-032-40673-2
BIM-based disaster response: Facilitating indoor path planning for various agents A. Dugstad, R.K. Dubey, J. Abualdenien & A. Borrmann School of Engineering and Design, Technical University of Munich, Munich, Germany
ABSTRACT: During path planning, rescue operations in buildings are often hindered due to the lack of critical building information. This paper proposes a real-time navigational platform combining various information sources to mitigate this problem. Various operating agents such as first-responders and UGVs receive specialized assistance by integrating semantic information from digital building models with incident-related real-time data. The proposed methodology includes three steps: (1) Pre-configuration of conditions influencing path choices according to agents’ capabilities and safety preferences. (2) Representation of the building geometry using a navigation graph. Inclusion of semantic information by weighting the graph with incident-related information, compared with the conditions per agent type and semantic information from the digital model. (3) Development of a knowledge database containing conditions for various types of agents. Evaluation in a real-world scenario revealed that paths calculated by the planning module were faster, shorter, and safer than those determined by first responders. 1 INTRODUCTION An indoor disaster scene consists of areas dynamically changing in accessibility and safety during incidents, which poses a challenge for operating first responders. The lack of accurate real-time information (e.g., disaster location and its spread, crowd density and room occupancy, structural damages, etc.) often reduces the efficacy of disaster response. The current challenges are magnified by the growing size and complexity of modern built environments. Currently, there are no platforms that can enable a remote administrator to assimilate the real-time status of the building and acquire important evacuation details to identify the safest and speediest path to the destination for various agents and disaster types. Recent developments in the architecture, engineering, and construction (AEC) industry regarding building information modeling (BIM) have unfurled new opportunities to meet disaster situations with more detailed support for first responders using digital tools and methods. The availability of new technologies, such as unmanned vehicles, and the increased access to digital information due to the digitalization of the construction industry, have opened up new opportunities to address those challenges (Directionsmag 2019). The utility of path planning in outdoor areas has already been proved with mapping solutions, such as Google Maps or Open Street Maps, throughout the last decades. Indoor areas can also be of interest with regard to path planning, as demonstrated by Google Maps’ recent indoor mapping and path planning solutions (Google, 2022). Though, the generation of indoor mapping and path planning is not new (Liu et al., 2021). Geometry-based solutions range from simple ones that only consider the connectivity DOI 10.1201/9781003354222-61
between different areas to more complex ones that also consider properties such as the agent’s size. In this work, an agent is regarded as an object or person that can move through space, depending on its locomotion type, physical abilities, and constraints. In the case study of this paper, first responders, unmanned aerial vehicles (UAVs), and unmanned ground vehicles (UGVs) are used as agents. Solutions that rely not only on geometric but also semantic sources can take into account differing abilities of agents (Khan 2015). Semantically rich building models form the basis for such advanced mapping approaches. BIM is a methodology for generating and using those comprehensive, informative representations of built facilities consisting of three-dimensional components and a vast amount of additional non-geometrical information (Borrmann, et al., 2015). Disaster management is one area of application requiring indoor mapping and path planning. Various researchers (Beata et al., 2018) have already investigated what information is to be collected, how to collect it, and how to present it efficiently to operating first responders. While significant efforts have been made to present as much information as possible to first responders, how this information is to be processed, to provide first responders with easily manageable information, remains under-researched. Existing solutions focus on a limited amount of obstacle avoidance, such as avoiding fire ignition points, and providing solutions only for first responders (Chou et al., 2019; Wang et al., 2019). State-of-theart solutions also commonly overlook how disasterrelated support must be provided for UAVs and UGVs as well, since they are expected to operate with first responders in future incidents (Directionsmag, 2019).
477
This paper addresses the above-mentioned research gap and fuses the rich disaster-related information with semantically rich building models and various operating agents’ abilities and priorities. A novel method for planning paths through disaster scenes is proposed, combining wide-ranging information from digital building models to address the challenge of large, complex buildings for various types of agents while also considering varied disaster types. In preparation for this study, expert interviews and surveys were conducted with robot manufacturers and first responders. A qualitative list of relevant incident-related obstacles and their effects on operating agents was generated (Dugstad et al., 2021). It was determined that both direct and indirect effects of incidents on agents must be considered. A direct effect is the deadly threat that fire has to humans. An indirect effect of fire is its impact on the functionality of building elements, decreasing the safety of areas connected to these building elements. BIM can be used to generate a geometric representation of the environment and provide semantic information to determine the effects of incidents on the built environments and thus indirectly on first responders. A knowledge database developed in this study allows the collection of agent-related properties.These properties can be used to compare real-time information from the incident with the geometric and semantic information of the built asset and the agent’s priorities regarding accessibility, safety, and distance of a path through the incident. The environment-abstraction methodology, knowledge database, and weighting function were implemented. Finally, the methodology was tested in a case study where a selected number of parameters was implemented. The time for first responders to orient themselves in a disaster scene and plan a path, which represents the conventional procedure in a disaster scenario, was compared with the time of the tool to plan a path through the scene while taking into account the determined parameters. The path planning module was found to determine faster and safer paths more quickly than the first responders. 2 STATE OF THE ART This section provides a comprehensive literature review about indoor mapping, their applicability to planning paths for various agents, and possibilities to integrate BIM-based information into disaster technologies and, more precisely, path planning methods. 2.1 Indoor map generation Typical approaches for indoor map generation based on BIM are the grid-based map-generation (Lin et al., 2013), and the graph-based map generation (Cheng et al., 2014; Hamieh et al. 2020; Teo & Cho 2016), as well as combinations of those (Liu et al., 2021; Zhou et al., 2020). Grid-based maps allow precisely setting start and end points for fine-granularity path planning. Their efficiency with algorithms like A* (Hart et al.,
1968) is, however, low due to their comparably high grid-resolution (Yao et al., 2010). By contrast, graphbased maps are generated based on some logic, such as topology. Topological indoor mapping approaches overcome the inefficiency of grid-based maps (Zhou et al., 2020). They are typically generated by skipping the height on each floor and applying Poincaré Duality (Liu et al., 2021). Nodes may represent spaces that are connected depending on their connectivity through transition elements (Hamieh et al., 2020). Topological maps allow no precise selection of start and end points and optimal shortest path planning due to their usually comparably lower resolution (Zhou et al., 2020). A graph-based example for overcoming in-precise path generation with a logic-based approach is presented by (Kneidl et al., 2012). 2.2 Path planning for various agents Maps are generated based on the idea of representing the connectivity of an environment’s accessible areas. It is important to note that the accessibility differs for different types of agents. (Khan 2015) investigated how the accessibility of an indoor environment is different for flying, rolling, and walking locomotion types. Not only is this accessibility constrained by the moving object’s size but also by its moving ability, involving parameters such as whether it can climb stairs and ramps or use windows. In their study, they define a navigable space for each locomotion type. 2.3 BIM-based disaster technologies BIM-based disaster technologies can be subcategorized into three stage. In the first stage, BIM is used to visualize the environment in which incidentrelated parameters are detected (Beata et al., 2018). First responders can use those visualization environments to orient themselves, where specific gas concentrations of, for example, carbon monoxide are high or low and where fire is located. Some researchers wanted to visualize the status quo of disaster-related parameters and how the situation may develop (Chen et al., 2018; Wang et al., 2021). Their approaches include using BIM with a fire simulation software to visualize how incidents may develop during a disaster. 2.4 Path planning tools A common approach to planning paths through a disaster-related environment is to either only visualize hazardous situations and plan the shortest path independently (Rueppel & Stuebbe 2008), or generate a graph-based map as explained in Section 2.1 and weight this graph with information that goes beyond distance information. A path planning algorithm searches for a path with the optimal weight. Thus, hazardous situations can be bypassed by adjusting the weights of nodes and edges corresponding to a hazardous area. (Chou et al., 2019) introduce such a weighting-based approach to determine safe paths
478
for first responders in fire scenarios. They generate a graph where the nodes are set based on the location of Bluetooth sensors that can detect fire ignition points in an indoor environment. The weight of the edges is initially set according to the distance of the nodes to each other. When a fire is located by a Bluetooth sensor, the weight of the edges connected to the corresponding node increases, forcing the algorithm to search for an alternative path around fire ignition points.
3 METHODOLOGY This study proposes an end-to-end real-time path planning approach that considers both the direct and indirect effects of incident-related obstacles on first responders and various agents operating in disaster scenarios. The proposed method consists of five parts, as demonstrated in Figure 1. The data source and exchange depicted in Section 3.1 sums up the information existing before an incident. It consists of a semantically rich building model, in Industry Foundation Class (IFC) format in the frame of this study, and information gathered from first responders stored in a knowledge database, which was developed within this study. The graph data model described in Section 3.2 was designed to hold every information necessary to represent the indoor environment within a weighted graph. The weighting framework Section 3.3 uses the information from the data exchange and real-time information from sensors from the field to weight the graph with the graph structure according to the graph data model. The path planning module’s dynamics are described in Section 3.4. It consists of the functionality that the graph is constantly re-weighted according to any newly fetched information from the sensors from the field, and paths can be dynamically planned
accordingly. The usability of the approach is validated in Section 3.5. 3.1 Data exchange The BIM methodology is suitable for generating building models providing this project’s necessary information. It allows for creating semantically rich building models that provide detailed information such as fire resistance. BIM is a standard methodology used in the construction industry to generate building models. Most newly developed construction processes base on the BIM methodology, and BIM models are thus available for disaster response for many buildings. They will be extensively available in the future. BIM models can be produced with software such as Autodesk Revit and exported in various data formats. IFC is a commonly used export data format that makes building information easily accessible via libraries such as Ifcopenshell in Python. The knowledge database, which was developed within this study, is based on a line of expert interviews that were conducted to gain an overview of the needs of first responders (Dugstad et al., 2021). The interview results contain qualitative situations that the first responders would consider positive or negative in different disaster scenarios. The specified situations can directly or indirectly impact the first responder’s path choices. Fire is, for instance, considered both, impacting first responders’ path choice directly and indirectly. The direct component may be the inability of first responders to cross the fire. The indirect component is that a fire impacts an agent’s path when the functionality of the building construction is threatened under fire. Furthermore, the importance of avoiding or preferring different situations differs. Four different cases are to be distinguished: A path with a specified situation is preferred, can be used, should be avoided
Figure 1. Framework Overview. Workflow presenting all developed modules and their interactions.
479
or must be avoided. For instance, a preferred path is a path via stairways or emergency routes in a fire incidence. A path with no special occurrences is a path that can be used. A path with a specified temperature may be defined as a path that should be avoided, and a path with fire is, for instance, a path that must be avoided. Direct conditions can be considered by gathering the information from the field on where a specified parameter such as fire is occurring and comparing it with the information on how this parameter should be handled for a specified agent. Here the applicability of the methodology for various agents comes into the picture. A database that defines how different occurrences shall be handled for a first responder, UAV, or UGV can adjust paths for the specified agent. Indirect conditions can be considered by gathering the field information on where a specified parameter such as fire is occurring and comparing it with how this impacts the building. For instance, the impact of fire on a building can be analyzed by using the semantic information about the fire resistance of building elements such as walls from the semantically rich BIM Model. The fire resistance holds the information on how long a standard fire can penetrate a building element until it loses its functionality. The database stores how the resulting effect should be handled. Users can, for instance, choose that they would like to avoid areas where the duration of fire exceeds the guaranteed fire resistance of the connected building elements. 3.2 Graph data model The study considers empty multiple-floor indoor environments with straight floors. Operating agents are considered intelligent, meaning they are able to navigate around small obstacles on their own. The interviews conducted in (Dugstad et al., 2021) revealed three significant requirements for paths in disaster environments. Paths must be short, accessible, and safe (Dugstad et al., 2021). Thus, the indoor map generated for path planning has to be of a type that allows the calculation of short, accessible, and safe paths for various agents. To do so, the map must hold information about (1) the connectivity of different locations to each other, (2) the distances between different connected locations, (3) the accessibility of the different areas for specific agents, and (4) the safety in different connected areas for various agents. A weighted graphbased approach was found to be useful in aligned studies (Chou et al., 2019) and will also be applied in this study. The logic described in the following section explains how a graph must be built to allow it to be weighted according to the abovementioned criteria. In Section 3.3 the weighting of the graph according to those criteria is described. (1) and (2) The shortest path within a room is always the direct one from the entering point of the room to the exit point. If the two transition points are visible to each other, the shortest path is a straight line between the two points. For cases where the transition points are not visible, a vertex-based graph-generation approach
Figure 2. The floor plan, given to the participants, contains all information about the scene. Fire resistance is indicated as F30 – F180, standing for 30 and 180 minutes fire resistance, respec-tively. Temperature is given per room. The existence and dura-tion of a fire in a room are given. Blocked doors and windows are marked in orange. All other doors and windows are consid-ered open. Stairs have stair height, and ramps have slope prop-erties. Start and goal position are indicated.
by (Kneidl, et al., 2012) can be used. Together with a path search algorithm such as A* or Dijkstra, the shortest path within a room can be determined. The proposed methodology to get the shortest path between two points in a building is thus, to represent all transition points in the building as nodes, connect them when they are associated with the same room, and store the waypoints retrieved as the room-based path search in a list of waypoints within the edge. The correct distance can then be calculated by retrieving the distance between these waypoints. This graph data model represents the geometric connectivity of the building. (3) The type of transition elements determines whether an agent can access an area that is delimited by this transition element or not. A rolling agent may, for instance, not be able to access stairs. Storing the information about a transition type in the corresponding node allows weighting the node according to the accessibility for a specified agent afterward. Further information, which is relevant to determine an agent’s ability to use this transition point may be the transition elements’ size and geometry. (4) Multiple edges always correspond to only one space in a building due to the nodes representing transition elements, that are located on geometrical space boundaries. The definition of a safety level per room allows weighting every edge according to the corresponding safety occurrences in the room. Every node is thus fitted with a function to calculate the accessibility weight, and every edge is thus fitted with a function to calculate the safety and distance weight. here listing facts use either the style tag List signs or the style tag List numbers. 3.3 Graph weighting The graph weight consists of three parts: distance, accessibility, and safety. Each of these three categories is conducted per agent on each node and edge of the graph. The distance weight is applied to the edges, which connect the nodes (transition elements). While there is no path request, this weight is the distance. The accessibility weight is applied to the nodes, which represent transition elements. The weight is
480
Figure 3. Graphs representing the connectivity of the area, depicted in Figure 2, reduced by the edges and nodes which must be avoided for a first responder (a) and (d), a UAV (b) and (e), and a UGV (c) and (f). Purple nodes represent doors. Blue nodes represent stairs, dark green nodes represent ramps. The start position of a requested path is depicted as a light green node and goal position as a yellow node. The view of the graph in the first row corresponds to the view of the floor plan in Figure 2.
either 1 if this transition element represents a preferred one for the agent, 2 if the agent can use the transition element, 3 if the agent prefers to avoid the transition element, or 4 if the agent cannot use this transition element at all. The weighting algorithm considers not only varying types of transition elements but also their properties, such as the blockage of a door or window, the slope of a ramp, and stair height. The weighting is conducted with the information collected for each agent’s abilities in the knowledge data bank. The safety weight is applied to the edges. Each room holds one safety weight and corresponds to multiple edges. All edges located in one room hold the same safety weight. The safety weights are similar to the ones of the accessibility weight, of values between 1 and 4, which describe whether the respective area is to be avoided. For this first development stage, a building fitted with sensors that can detect fire and temperature in every room is assumed, and the information measured by them is the following: whether or not there is a fire in a room, how high the room temperature is, and the duration of a fire. The building model holds the information on the fire resistance of walls (i.e., how long a wall can withstand a fire until it loses its functionality). The knowledge database holds the information about the preferences per agent per incident. The safety weight is determined by comparing building model information with agent preferences and the real-time information from the scene. Both the direct impact (i.e., fire affecting the safety level of a room for the agent) and indirect impacts (i.e., fire affecting the stability of the walls in a room and thus
the safety level of a room for the agent) are taken into account in this weighting step.
Figure 4. Paths, which were found by the path planning module (green) and the participants (yellow and purple) for UAVs.
When a path is requested for a specified agent, the weights of the graph are dynamically updated based on the agent and real-time information from the scene. The graph is reduced by all edges and nodes, which hold a weight of 4. The path planning tool searches through the reduced graph, which is weighted according to the distance for the x shortest paths. The x shortest paths are re-weighted according to their safety- and accessibility weight, which at this point only consists of values between 1 and 3, due to the exclusion of all inaccessible nodes and edges. The re-weighting is conducted by checking the properties that store the safety and accessibility information per agent. Each node and edge can contain multiple weights due to numerous incidents (fire and temperature, for instance) influencing the weight. The highest weight per edge and
481
Figure 5. Paths, which were found by the path planning module (green) and the participants (yellow) for first responders.
node is selected. The shortest x paths get assigned the averaged safety and accessibility weight over all edges and nodes. To incorporate the distances of the paths, their weight is interpolated to a value between 1 and y depending on whether their value is equivalent to the shortest distance between start and goal node or y = t times the shortest distance. Each path consists of a distance, accessibility, and safety weight between 1 and 3. The first responders can conduct the prioritization, for instance, accessibility over safety in the step of editing agent-based information in the knowledge database. Based on this prioritization, the paths are weighted, and the most optimal one is proposed. 3.4 Dynamic graph Technologies are considered deployed to report information on windows’ and doors’ opened or locked status. Similar to the data from the sensors that determine fire and temperatures, this information is a “new detection”. Every agent-related graph is constantly reweighted per new detection, activating the dynamic path planner per the current situation in the building. The dynamics of the path planning consist of the realtime update of graph weights with real-world information via the sensors and intelligent technologies. A client can request a path from the path planning module specifying the agent for which the path should be planned as well as the start and the goal position.The path planning module reduces the graph according to the methodology described in Section 3.3. The closest node to the start and goal position is searched and connected to the graph. The shortest x paths from the start to the goal position are generated using the Dijkstra Algorithm (Dijkstra & others, 1959). We used Dijkstra path-planning algorithm mainly due to its easy implementation and linear time complexity. Paths are weighted according to the safety, accessibility, and distance with respect to the corrected distance, described in Section 3.3. The path with the lowest weight is suggested. 3.5 Validation and simulations A prototype of the developed concept was implemented and validated in a study involving a complex, extensive building at the inner-city campus of the
Technical University of Munich (TUM), which was prepared for a possible fire event. The prototype allows for a generation of graphs based on built environments automatically. Correct paths for transition points per room that are not visible to each other were incorporated manually in this version. The building model was prepared with Revit 2022 and exported into IFC4 format. Properties such as the slope of ramps, heights of stairs, and the fire resistance of walls were adjusted in the model. The current implementation of the module checks which rooms share the same wall to weigh the impact of fire on the safety of an adjacent room depending on the fire resistance of a wall separating these rooms. Thus, one wall is required to be shared by only two spaces in this implemented version. Outdoor areas are one space. The case study aims to test the performance of firefighters to plan paths to different goal positions in a building that is affected by a fire incident to validate the possible assistance which can be achieved with an approach proposed within this study. To avoid confusion between the use of the terminology for firefighters representing the participants and the firefighter representing agents for which the paths were planned, the firefighters within this study are from now on referred to as the participants. A mock-up of fire incidents was simulated using a pictorial representation on a 2D drawing of a floor plan. The participants were given floor plans of the scene that were enriched with information about the scene and the abilities of three different agents for which paths should be planned. The three different types of agents considered are a UAV, which can transition to any transition element despite closed ones; a UGV, which can use doors, stairs, and ramps but is constrained by ramp slopes and stair heights and a first responder, which is also able to use stairs, ramps, and doors, but is not constrained regarding ramp height. While the building is five stories high, the scenario was prepared for one storey to compare the developed methodology with the participants’ ability in an easy scenario.
Figure 6. Paths, which were found by the path planning module (green) and the participants (yellow) for UGVs.
The participants’ speed was counted by measuring their time from receiving the floor plan to planning a path. The result was split up into the gross time, which includes the time participants used to orient themselves and check the properties of the agents,
482
and the net time, which was the time participants only used to orient themselves within the scene. The study was recorded to measure this time correctly. The determined paths by the participants and the path planning module were compared by their weights. To keep it comparable, the parameters measured in the scene were temperature, and fire as well as blocked windows and doors. Fire is processed regarding its direct impact on the safety of a path and its impact on the functionality of walls using the difference between the duration of a fire and the minutes that the wall should be able to persist the fire. The accessibility was weighted according to the type of the transition elements and, for ramps, according to their slope, and stairs, according to the stair height that an agent can climb. A further parameter used in the weighting schema was whether a door was opened or closed. 4 SIMULATION AND RESULTS To make the graph clear to the reader, the BIM model was reduced to one storey. The graphs which were produced for the three types of agents based on accessibility and safety (both weighted < 4) are presented in Figure 3. The different colors of the nodes represent transition elements. It can be seen that the graphs (a) and (d), representing the connectivity graph for the implemented first responder, and the graphs (c) and (f), representing the connectivity for the implemented UGV, are very similar due to the similar abilities of the implemented first responder, and UGV. The graph for the UGV is accessibility-wise reduced by one node, representing a slope that is too steep for the UGV, and further nodes due to its lower temperature resistance.
Due to the UAV’s ability to use windows, graphs (b) and (e) consist of more nodes and edges than the other graphs. The reduction of the UAV’s graph with regards to accessibility due to two blocked windows in the upper right of the floor plan in Figure 2 can be detected in Figure 3 upper right corner. The path planning module was implemented with x=10,100,1000, representing the number of the shortest paths during step 1 of the planning. Factor x was introduced in Section 3.3. Factor t, introduced in Section 3.3, was chosen as 3, such that a path three times as long as the shortest distance between start and goal node is weighted with factor 3. The shortest distance between start and goal was determined as 53 m. Any path, which is 53m long, is thus weighted with 1. Any path, equal or more than 53m times 3=159m long, is weighted with 3. The values in between were interpolated. The optimal path for all three types of agents is depicted in green in Figures 5, 4 and 6. The calculation time to determine the paths was under 1 second when using x=10 and x=100 shortest paths in the algorithm, and 7 seconds when using x=1000. The participants in the case study needed, on average, 03:49min gross and 02:14min net to plan a path. The net time is more accurate since it excludes the time used to check for agents’ parameters such as a maximum slope. Operating first responders are expected to know those parameters and thus won’t need to look them up. With 03:03min, participants took the most time to search for the path for the first responders, which may, however, also correlate to the fact that this was the first one they searched for. The shortest time (01:35min) was used for finding a path for the UAV, which is able to use windows and can thus be considered easier than the other cases. On average,
Table 1. Comparison of the paths found by the path planning module and the participants during the case study. Agent indi-cates who the path was planned for. Path planner represents the planning person or computer. Calculation time refers to the time which was used to plan the path. Gross time stands for the time in total, including the time which was used to look up parameters by participants. Net time stands for the time that was solely used to search for a path in the scene. The path length of the defined paths is given in meters. The weight consists of the accessibility weight of the nodes, the safety weight of the edges and the dis-tance weight of the edges of each path. The total weight averages all three weights. Calculation time [min] Agent
Path planner
Total
Net
Weight [–]
Path length [m]
A
S
D
Total
FR
Participant 1 Participant 2 Participant 3 PPM
05:37 03:13 08:20
02:45 01:49 04:36