20. Internationales Stuttgarter Symposium : Automobil- und Motorentechnik [1. Aufl.] 9783658309947, 9783658309954

Citation preview

Proceedings

Michael Bargende · Hans-Christian Reuss Andreas Wagner Hrsg.

20. Internationales Stuttgarter Symposium Automobil- und Motorentechnik Band 2

Proceedings

Ein stetig steigender Fundus an Informationen ist heute notwendig, um die immer komplexer werdende Technik heutiger Kraftfahrzeuge zu verstehen. Funktio­ nen, Arbeitsweise, Komponenten und Systeme entwickeln sich rasant. In immer schnelleren Zyklen verbreitet sich aktuelles Wissen gerade aus Konferenzen, Tagungen und Symposien in die Fachwelt. Den raschen Zugriff auf diese Infor­ mationen bietet diese Reihe Proceedings, die sich zur Aufgabe gestellt hat, das zum Verständnis topaktueller Technik rund um das Automobil erforderliche spe­ zielle Wissen in der Systematik aus Konferenzen und Tagungen zusammen zu stellen und als Buch in Springer.com wie auch elektronisch in Springer Link und Springer Professional bereit zu stellen. Die Reihe wendet sich an Fahrzeug- und Motoreningenieure sowie Studierende, die aktuelles Fachwissen im Zusammen­ hang mit Fragestellungen ihres Arbeitsfeldes suchen. Professoren und Dozenten an Universitäten und Hochschulen mit Schwerpunkt Kraftfahrzeug- und Moto­ rentechnik finden hier die Zusammenstellung von Veranstaltungen, die sie selber nicht besuchen konnten. Gutachtern, Forschern und Entwicklungsingenieuren in der Automobil- und Zulieferindustrie sowie Dienstleistern können die Procee­ dings wertvolle Antworten auf topaktuelle Fragen geben. Today, a steadily growing store of information is called for in order to understand the increasingly complex technologies used in modern automobiles. Functions, modes of operation, components and systems are rapidly evolving, while at the same time the latest expertise is disseminated directly from conferences, congres­ ses and symposia to the professional world in ever-faster cycles. This series of proceedings offers rapid access to this information, gathering the specific know­ ledge needed to keep up with cutting-edge advances in automotive technologies, employing the same systematic approach used at conferences and congresses and presenting it in print (available at Springer.com) and electronic (at Springer Link and Springer Professional) formats. The series addresses the needs of automotive engineers, motor design engineers and students looking for the latest expertise in connection with key questions in their field, while professors and instructors working in the areas of automotive and motor design engineering will also find summaries of industry events they weren’t able to attend. The proceedings also offer valuable answers to the topical questions that concern assessors, researchers and developmental engineers in the automotive and supplier industry, as well as service providers.

Weitere Bände in der Reihe http://www.springer.com/series/13360

Michael Bargende · Hans-Christian Reuss · Andreas Wagner (Hrsg.)

20. Internationales Stuttgarter Symposium Automobil- und Motorentechnik Band 2

Hrsg. Michael Bargende FKFS/IVK Universität Stuttgart Stuttgart, Deutschland

Hans-Christian Reuss FKFS/IVK Universität Stuttgart Stuttgart, Deutschland

Andreas Wagner FKFS/IVK Universität Stuttgart Stuttgart, Deutschland

ISSN 2198-7432 ISSN 2198-7440  (electronic) Proceedings ISBN 978-3-658-30994-7 ISBN 978-3-658-30995-4  (eBook) https://doi.org/10.1007/978­3­658­30995­4 Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der Deutschen National­ bibliografie; detaillierte bibliografische Daten sind im Internet über http://dnb.d­nb.de abrufbar. © Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung, die nicht ausdrücklich vom Urheberrechtsgesetz zugelassen ist, bedarf der vorherigen Zustimmung des Verlags. Das gilt insbesondere für Vervielfältigungen, Bearbeitungen, Übersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen. Die Wiedergabe von allgemein beschreibenden Bezeichnungen, Marken, Unternehmensnamen etc. in diesem Werk bedeutet nicht, dass diese frei durch jedermann benutzt werden dürfen. Die Berechtigung zur Benutzung unterliegt, auch ohne gesonderten Hinweis hierzu, den Regeln des Markenrechts. Die Rechte des jeweiligen Zeicheninhabers sind zu beachten. Der Verlag, die Autoren und die Herausgeber gehen davon aus, dass die Angaben und Informa­ tionen in diesem Werk zum Zeitpunkt der Veröffentlichung vollständig und korrekt sind. Weder der Verlag, noch die Autoren oder die Herausgeber übernehmen, ausdrücklich oder implizit, Gewähr für den Inhalt des Werkes, etwaige Fehler oder Äußerungen. Der Verlag bleibt im Hinblick auf geografische Zuordnungen und Gebietsbezeichnungen in veröffentlichten Karten und Institutionsadressen neutral. Verantwortlich im Verlag: Markus Braun Springer Vieweg ist ein Imprint der eingetragenen Gesellschaft Springer Fachmedien Wiesbaden GmbH und ist ein Teil von Springer Nature. Die Anschrift der Gesellschaft ist: Abraham-Lincoln-Str. 46, 65189 Wiesbaden, Germany

WELCOME IN STUTTGART The car is currently undergoing a complete reinvention. And for us there is a lot at stake: our role as technological pioneers, our economic strength, many jobs and conservation of our natural resources. The Stuttgart International Symposium for Automotive and Engine Technology, now in its 20th year, is intended to enable a fruitful exchange of expertise and creative ideas in order to meet this great challenge. I am delighted to be the patron for this event, and I welcome the participants from all around the world to our state capital! The automotive industry is an especially important component of our economic power, particularly in Baden-Württemberg. Going forward, these key industries must adapt to ecological and political climate requirements. But we can only shape the future of our automotive economy by working together with everyone involved in this process. My state government is also supporting this development with the cross-sector “Baden-Württemberg Automotive Industry Strategic Dialog”, because the transition to sustainable and digital mobility requires expertise from diverse areas – from business, science, trade unions, associations, civil society and politics. With more than 100 lectures and 800 participants, the Stuttgart International Symposium is one of the largest congresses for vehicle and engine development in Europe. So I would like to take the opportunity to thank the organizers of this congress, and wish all the guests inspiring conversations and exciting new knowledge! Winfried Kretschmann Prime Minister of the State of Baden-Württemberg

V

A WARM WELCOME The zero carbon car – technical challenges of the future The mobility transition and intermodal mobility; car sharing versus shared space, digitalization, automation and networking, e-fuels, batteries or fuel cells after all? The current discussions concerning the future of personal transport are generally characterized by uncertainty and open questions, rather than clear answers and confidence. This is perturbing the industry: unsettling it on the one hand and on the other hand making it more important than ever to get to grips with the issues involved. With its focus on “The challenges of future technology” and its diverse program of lectures, the 20th Stuttgart International Symposium grapples with these questions. It hopes to help the industry approach these future-related topics and provide a platform for – potentially controversial – discussions. The program structure has also changed. In response to requests from our participants we are now offering more opportunities for collaboration and networking: » In the World Café, you can work on specific solutions for current problems in our industry. The results of this creative unit will be presented directly to a wide audience directly before the podium discussion. » We'll be celebrating the Stuttgart International Symposium's 25th birthday with a stand party in the exhibition. » A poster session and a tour of the exhibition are further new features. You can also expect some optical changes on location, along with some familiar elements! We are curious to hear your feedback and look forward to seeing you in Stuttgart on March 17 and 18, 2020! Prof. Dr. Michael Bargende Prof. Dr. Hans-Christian Reuss Prof. Dr. Andreas Wagner

VI

INDEX – Volume 2 SECTION 1

STRATEGIES FOR COMMERCIAL VEHICLE DRIVETRAINS 2030 Chairperson: Prof. Dr. Thomas Koch The future of combustion engines from the perspective of rail applications Andreas Wegmann, M. Gayer, J. M. Voith SE & Co. KG, VTA; B. Gay, M. Pister, Liebherr Machines Bulle SA

1

Strategies for commercial vehicle drivetrains 2030 Philip Scarth, FPT Motorenforschung AG

3

The commercial vehicle in 10 years and beyond – Still with a classic ICE? Christian Weiskirch, A. Kammel, TRATON GROUP

5

CONNECTED CAR Chairperson: Prof. Dr. Andreas Wagner 5G network infrastructure & autonomous driving – The next generation C-V2X technology Walter Haas, HUAWEI TECHNOLOGIES Deutschland GmbH

7

Welcome to the photonic age in automotive industry Martin Enenkel, Jenoptik Optical Systems GmbH

9

SEAT’s digital transformation in two dimensions Sebastian Grams, SEAT S.A., VW Group

23

VII

INDEX – Volume 2 FUELS Chairperson: Prof. Dr. Stefan Pischinger Methanol as an electricity-based fuel for plug-in hybrid electric vehicles Tobias Bieniek, K. Rößler, F. Otto, Mercedes-Benz AG

31

E-fuels as chemical energy carrier under the aspects of costs and efficiency Martin Rothbart, J. Rechberger, AVL List GmbH

47

Potential analysis and virtual development of SI Engines operated with DMC+ Michael Grill, C. Wagner, M.-T. Keskin, M. Bargende, FKFS; L.Cai, H. Pitsch, ITV, RMTH Aachen University; S. Blochum, LVK, TU München

49

ELECTRIC DRIVETRAINS Chairperson: Prof. Dr. Michael Auerbach Supporting the synthesis of electric drive systems with scenario management Adrian Braumandl, F. Marthaler, K. Bause, F. Ranly, IPEK, Karlsruher Institut für Technologie (KIT)

75

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems Marius Panzer, C. Wagner, A.-L. Winkler, A. Wöll, R. Bernewitz, MANN+HUMMEL GmbH

89

Breaktor™ – advanced protection for high voltage circuits in electric vehicles Till Wagner, K. Calzada, Eaton Electrical Products Ltd

VIII

101

INDEX – Volume 2 E-MOBILITY Chairperson: Jürgen Schenk Range prediction and e-mobility – The complexity behind a length scale Rafael Abel, F. Beutenmüller, S. Stein, TWT GmbH

111

Calculation of driving consumption of an electric city bus Joscha Reber, L. Braun, EvoBus GmbH

113

Entrance to an electrified last mile ecosystem project “bring your own battery” Markus Geiger, B. Budaker, csi entwicklungstechnik GmbH; G. Lange, C. Schmidt, AUDI AG

129

How accurate can a range calculation of an electric vehicle be? Lisa Braun, EvoBus GmbH

139

IX

INDEX – Volume 2 SECTION 2

TESTING Chairperson: Prof. Dr. Karl-Ludwig Haken Testing in times of big data and machine learning Hendrik Bohlen, P. Assendorp, Werum Software & Systems AG

155

Testing ADAS end-of-line – Avoid the hazardous effects Frank Heidemann, SET GmbH

157

Consumption-relevant load simulation during cornering at the vehicle test bench VEL Martin Gießler, P. Rautenberg, F. Gauterin, Karlsruhe Institute of Technology (KIT)

159

BIG DATA Chairperson: Prof. Dr. Eric Sax AI and big data management for autonomous driving (AD) Frank Kraemer, IBM

173

Semantic segmentation of solid state LiDAR measurements for automotive application Sören Erichsen, M. Schmidt, Ibeo Automotive Systems; J. Nitsch, Ibeo Automotive Systems, ETH Zurich; A. Schlaefer, Hamburg University of Technology

179

Method for a scenario-based and weighted assessment of map-based advanced driving functions Carl Esselborn, M. Eckert, M. Holzäpfel, E. Wahl, Dr. Ing. h.c. F. Porsche AG; E. Sax, Karlsruhe Institute of Technology (KIT)

193

X

INDEX – Volume 2 FUEL CELL Chairperson: Prof. Dr. Helmut Eichlseder Validation of fuel cell control units with powerful simulation platforms for fuel cells Abduelkerim Dagli, M. Seeger, MicroNova AG

209

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level Johann C. Wurzenberger, T. Glatz, AVL List GmbH; D. Rašić, G. Tavčar, AVL-AST d.o.o.; I. Mele, A. Kregar, T. Katrašnik, University of Ljubljana

219

Holistic design of innovative cathode air supply for automotive PEM fuel cells Michael Harenbrock, A. Korn, A. Weber, MANN+HUMMEL GmbH; E. Hallbauer, MANN+HUMMEL Innenraumfilter GmbH & Co. KG

235

ECMS based on system-specific control parameter adaption of a fuel cell hybrid electric vehicle Sergei Hahn, J. Braun, H. Kemmer, Robert Bosch GmbH; H.-C. Reuss, IVK, Universität Stuttgart

251

SENSORS & ACTUATORS Chairperson: Prof. Dr. Karl-Ludwig Krieger Using machine learning methods to develop virtual NOx sensors for vehicle applications Robert Fechert, B. Bäker, S. Gereke, F. Atzler, IAD, TU Dresden

265

Image-based condition monitoring of a multi-LED-headlamp Pascal Janke, J. Cai, HELLA GmbH & Co. KGaA; M. Niedling, L-LAB; T. Bertram, RST, TU Dortmund

281

Measurement and testing of lidarsensors Andy Günther, B. Bäker, IAD, TU Dresden

297

XI

INDEX – Volume 2 SOFTWARE Chairperson: Prof. Dr. Lennart Löfdahl Credibility of software-in-the-loop environments for integrated vehicle function validation Indrasen Raghupatruni, S. Burton, M. Boumans, T. Huber, A. Reiter, Robert Bosch GmbH

299

Big data driven vehicle development – Technology and potential Tobias Abthoff, P. Fank, D. Boja, NorCom Information Technology GmbH & Co. KGaA

315

Continuous development environment for the validation of autonomous driving functions Sebastian Lutz, M. Behrendt, A. Albers, IPEK, Karlsruher Institut für Technologie (KIT)

327

HYBRID POWERTRAINS Chairperson: Prof. Dr. Christian Beidl Impact of future 48 V-systems on powertrain operation under real-driving conditions Daniel Förster, M. Timmann, R. Inderka, J. Strenkert, Mercedes-Benz AG; F. Gauterin, Karlsruhe Institute of Technology (KIT)

345

Model based development of optimum control strategies for hybrid electric vehicles Christoph Pötsch, A. Cvikl, J. C. Wurzenberger, AVL List GmbH

361

Virtual powertrain – Vehicle simulation on the engine test bench with an implemented P2 topology Sebastian Lachenmaier, L. Cross, C. Ferrara, A. Greis, M. Wüst, D. Naber, Robert Bosch GmbH

377

MAHLE modular hybrid powertrain equipped with passive MAHLE jet ignition Neil Fraser, M. Berger, MAHLE International GmbH; S. Reader, M. Bassett, A. Cooper, MAHLE Powertrain Limited

393

XII

INDEX – Volume 2 SECTION 3

EMISSIONS Chairperson: Prof. Dr. Georg Wachtmeister Phase change cooled manifold for RDE compliant powertrains Thomas Arnold, M. Krause, IAV GmbH

409

Analysis of scavenging air post-oxidation by means of 3D-CFD simulation including reaction mechanism Rodolfo Tromellini, J. Przewlocki, F. Cupo, M. Bargende, IVK, Universität Stuttgart; M. Chiodi, FKFS

421

Use of ogygenate blends as inflammation aid in diluted mixtures Moritz Grüninger, O. Toedter, T. Koch, IFKM, Karlsruher Institut für Technologie (KIT)

439

EE-ARCHITECTURE Chairperson: Prof. Dr. Gerhard Hettich Fault tolerant electric energy supply system design for automated electric shuttle bus Marcus Goth, D. Keilhoff, H.-C. Reuss, IVK, Universität Stuttgart

441

Concepts of functional safety in E/E-architectures of highly automated and autonomous vehicles Dennis Niedballa, H.-C. Reuss, IVK, Universität Stuttgart

457

Toolchain for architecture development, modeling and simulation of battery electric vehicles Carl Friedrich Hettig, P. Orth, FEV Europe GmbH; M. Deppe, T. Pajenkamp, dSpace GmbH; C. Granrath, J. Andert, RWTH Aachen University

471

XIII

INDEX – Volume 2 HEATING & COOLING Chairperson: Prof. Dr. Stefan Böttinger Model-based approach for on-demand temperature control Alexander Herzog, IAV GmbH; M. Fuchs, IfT, Leibniz Universität Hannover; A. Vagapov, IAV GmbH und IfT, Leibniz Universität Hannover

485

Preheating components with metal hydrides or lime – Small, high power, no additional energy Mila Kölbig, I. Bürger, M. Schmidt, M. Linder, ITT, Deutsches Zentrum für Luft- und Raumfahrt e.V. (DLR)

501

Fast running detailed battery thermal management models based on 1D-3D synergetic approach Dig Vijay, N. Framke, P. Stopp, Gamma Technologies GmbH

511

Artificial Intelligence in predictive thermal management for passenger cars Felix Korthals, M. Stöcker, Daimler AG; S. Rinderknecht, TU Darmstadt

529

AGILE DEVELOPMENT Chairperson: Prof. Dr. Tobias Flämig Continuous integration in powertrain software – Today and tomorrow Daniel Heß, D. Volquard, R. Siedel, F. Feyerherd, IAV GmbH

545

Managing software evolution in embedded automotive systems Lukas Block, IAT, Universität Stuttgart

557

Agile systems engineering for critical systems Christof Ebert, Vector Consulting Services; F. Kirschke-Biller, Ford

573

XIV

INDEX – Volume 2 POWERTRAIN SIMULATION Chairperson: Prof. em. Dr. Günter Hohenberg Discretization and heat transfer calculation of engine water jackets in 1D simulation Florian Mandl, M. Grill, M. Bargende, FKFS/IVK, Universität Stuttgart

583

Optical investigations for the optimization and calibration of 3D-CFD injection models Simon Hummel, A. Vacca, M. Reichenbacher, K. Müller, A. Kächele, M. Koch, M. Bargende, FKFS/IVK, Universität Stuttgart

605

Accelerated assessment of optimal fuel economy benchmarks for developing the next generation HEVs Pier Giuseppe Anselma, G. Belingardi, Politecnico di Torino

623

BATTERY Chairperson: Prof. Dr. Andreas Friedrich Front loading approach in battery development for generation update Paul Schiffbänker, N. Dejanovic, AVL List GmbH

641

CO2-neutral battery production in Europe – How to make it happen? Robert Stanek, M. Hackmann, P3 automotive GmbH

655

Greater sustainability with a second life of used electric vehicle batteries Jürgen Kölch, EVA Fahrzeugtechnik GmbH

663

Increased safety for battery electric vehicles by using heat-resistance stainless steels Stefan Lindner, Outokumpu Nirosta GmbH

673

XV

SPEAKERS, CHAIRPERSONS Dr. Rafael Abel TWT GmbH

Lukas Block IAT, Universität Stuttgart

Dr. Tobias Abthoff NorCom Information Technology GmbH & Co. KGaA

Florian Bock AUDI AG

Martin Angerbauer IVK, Universität Stuttgart

Dr. Ulrich Bodenhausen Vector Consulting Services GmbH und Ulrich Bodenhausen AI Coaching

Pier Giuseppe Anselma Politecnico di Torino

Hendrik Bohlen WERUM Software & Systems AG

Thomas Arnold IAV GmbH

Prof. Dr. Stefan Böttinger Universität Hohenheim

Prof. Dr. Frank Atzler TU Dresden

Adrian Braumandl Karlsruher Institut für Technologie (KIT)

Prof. Dr. Michael Auerbach Hochschule Esslingen

Dr. Lisa Braun EvoBus GmbH

Prof. Dr. Bernard Bäker TU Dresden

Chhaya Chavan hofer-pdc GmbH

Federico Coren Thomas Bareiß MdB Bundesministerium für Wirtschaft und Energie TU Graz Prof. Dr. Michael Bargende FKFS/IVK, Universität Stuttgart

Abdülkerim Dagli MicroNova AG

Prof. Dr. Christian Beidl TU Darmstadt

Prof. Dr. Klaus Dietmayer Universität Ulm

Youssef Beltaifa Hochschule Karlsruhe

Prof. Dr. Christof Ebert Vector Consulting Services

Dr. Peter Berlet IAVF Antriebstechnik GmbH

Sven Eckelmann Hochschule für Technik und Wirtschaft Dresden

Lutz Berners Berners Consulting GmbH

Prof. Dr. Lutz Eckstein RWTH Aachen University

Tobias Bieniek Daimler AG

Dr. Torsten Eder Mercedes-Benz AG

Dr. Ulrich Eichhorn Arne Bischofberger IPEK, Karlsruher Institut für Technologie (KIT) IAV GmbH

XVI

SPEAKERS, CHAIRPERSONS

Prof. Dr. Helmut Eichlseder TU Graz

Prof. Dr. Bernhard Geringer TU Wien

Dr. Gerald Eifler ElringKlinger Motortechnik GmbH

Dr. Martin Gießler Karlsruher Institut für Technologie (KIT)

Martin Enenkel Jenoptik Optical Systems GmbH

Dietmar Goericke Forschungsvereinigung Verbrennungskraftmaschinen e. V.

Jürgen Erhardt Erhardt GmbH Fahrzeug + Teile Sören Erichsen Ibeo Automotive Systems GmbH Carl Esselborn Dr. Ing. h.c. F. Porsche AG Sebastian Ewert MAHLE GmbH Robert Fechert TU Dresden Prof. Dr. Tobias Flämig DHBW Stuttgart Fabian Fontana IVK, Universität Stuttgart Daniel Förster Mercedes-Benz AG Dr. Günter Fraidl AVL List GmbH Neil Fraser MAHLE International GmbH

Marcus Goth IVK, Universität Stuttgart Andreas Graf itemis AG Dr. Sebastian Grams SEAT S.A., VW Group Dr. Michael Grill FKFS Moritz Grüninger Karlsruher Institut für Technologie (KIT) Andy Günther TU Dresden Dr. Andreas Haag Robert Bosch Automotive Steering GmbH Walter Haas HUAWEI TECHNOLOGIES Deutschland GmbH Sergei Hahn Robert Bosch GmbH Prof. Dr. Karl-Ludwig Haken Hochschule Esslingen

Prof. Dr. Andreas Friedrich Deutsches Zentrum für Luft-und Raumfahrt e.V. Dr. Michael Harenbrock (DLR) MANN+HUMMEL GmbH Prof. Dr. Frank Gauterin Karlsruher Institut für Technologie (KIT)

Dr. Martin Härtl TU München

Markus Geiger csi entwicklungstechnik GmbH

Frank Heidemann SET GmbH

XVII

SPEAKERS, CHAIRPERSONS

Martin Heiderich Honda R&D Europe (Deutschland) GmbH

Pascal Janke HELLA GmbH & Co. KGaA

Alex Heron-Himmel Dr. Heribert Kammerstetter Deutsches Zentrum für Luft-und Raumfahrt e.V. AVL List GmbH (DLR) Satheesh Kandasamy Dr. Alexander Herzog SIMULIA Corporation, A DASSAULT IAV GmbH SYSTEMES Company Daniel Heß IAV GmbH

Roland Kemmler Mercedes-Benz AG

Prof. Dr. Dr. Gerhard Hettich EAST Consulting

Moritz Kilper Daimler AG

Carl Friedrich Hettig FEV Europe GmbH

Dr. Felix Kistler TWT GmbH

Johannes Hipp TU Darmstadt

Manuel Klauß NTT DATA Deutschland GmbH

Christian Hochfeld Agora Verkehrswende

Gunnar-Marcel Klein MANN+HUMMEL International GmbH & Co. KG

Dr. Andreas Höfer Valeo Siemens eAutomotive Germany GmbH Prof. em. Dr. Günter Hohenberg Simon Hummel IVK, Universität Stuttgart Nicolas Ide Robert Bosch GmbH Marcus Irmer CLM, TH Köln Naser Jafarzadehpour Mercedes-AMG GmbH Felix Jakob AKKA Technologies Taleb Janbein Robert Bosch GmbH

XVIII

Prof. Dr. Thomas Koch Karlsruher Institut für Technologie (KIT) Bernhard Kockoth ViGEM GmbH Dr. Mila Kölbig Deutsches Zentrum für Luft-und Raumfahrt e.V. (DLR) Dr. Jürgen Kölch EVA Fahrzeugtechnik GmbH Dr. Jens König Deutsches Zentrum für Luft- und Raumfahrt e.V. (DLR) Felix Korthals Daimler AG Frank Kraemer IBM

SPEAKERS, CHAIRPERSONS

Prof. Dr. Karl-Ludwig Krieger Universität Bremen

Marius Panzer MANN+HUMMEL International GmbH & Co. KG

Prof. Dr. Ferit Küçükay TU Braunschweig

Dr. Anna-Antonia Pape TWT GmbH

Michael Kühn E@motion GmbH

Prof. Dr. Nejila Parspour Universität Stuttgart

Sebastian Lachenmaier Robert Bosch GmbH

Daniel Paula CARISSMA, TH Ingolstadt

Eugen Pfeifer Dr. Andreas Leich Deutsches Zentrum für Luft-und Raumfahrt e.V. AUTOMOTEAM GmbH (DLR) Prof. Dr. Stefan Pischinger FEV Group GmbH Stefan Lindner Outokumpu Nirosta GmbH Dr. Wilfried Plum OWI gGmbH Prof. Lennart Löfdahl Chalmers University of Technology Dr. Christoph Pötsch AVL List GmbH Prof. Giovanni Lombardi University of Pisa Indrasen Raghupatruni Robert Bosch GmbH Sebastian Lutz Karlsruher Institut für Technologie (KIT) Joscha Reber EvoBus GmbH Florian Mandl IVK, Universität Stuttgart Prof. Dr. Dr. Wolfram Ressel Universität Stuttgart Nicolas Marmann SET GmbH Prof. Dr. Hans-Christian Reuss FKFS/IVK, Universität Stuttgart Wolfgang Müller-Pietralla Volkswagen AG Konstantin Riedl TU München Eiji Nakai Mazda Motor Corporation Rainer Röck Ingenieurbüro Röck Dennis Niedballa IVK, Universität Stuttgart Martin Rothbart AVL List GmbH Denis Notheis Karlsruher Institut für Technologie (KIT) Prof. Dr. Hermann Rottengruber OvGU Magdeburg Martin Novák Icon Technology & Process Consulting Ltd.

XIX

SPEAKERS, CHAIRPERSONS

Prof. Dr. Günter Sabow Wirtschafts- und Industrievereinigung Stuttgart e.V. Mustafa Saraoğlu IfA,TU Dresden Prof. Dr. Eric Sax Karlsruher Institut für Technologie (KIT) Philip Scarth FPT Motorenforschung AG Jürgen Schenk P3 automotive GmbH Paul Schiffbänker AVL List GmbH Jürgen Schlaht Siemens Mobility GmbH Prof. Dr. Siegfried Schmauder IMWF, Universität Stuttgart Christian Stach Robert Bosch GmbH Robert Stanek P3 automotive GmbH Ulrich Steinbach Ministerium für Wissenschaft, Forschung und Kunst Baden-Württemberg Dr. Antonio Torluccio Automobili Lamborghini S.p.a. Rodolfo Tromellini IVK, Universität Stuttgart Daniel Trost FKFS Antonino Vacca IVK, Universität Stuttgart

XX

Dr. Dig Vijay Gamma Technologies GmbH Prof. Dr. Georg Wachtmeister TU München Prof. Dr. Andreas Wagner FKFS/IVK, Universität Stuttgart Till Wagner Eaton Dennis Wedler IK, TU Braunschweig Dr. Andreas Wegmann J.M. Voith SE & Co. KG | VTA Dr. Christian Weiskirch TRATON GROUP Prof. Dr. Dr. Michael Weyrich Universität Stuttgart Wilhelm Wiebe DHBW Mannheim Ursel Willrett IAV GmbH Johannes Winterhagen Redaktionsbüro delta eta Dr. Teddy Woll Daimler AG Frank Wolter FEV Europe GmbH Kai Wolter Karlsruher Institut für Technologie (KIT) Dr. Johann Wurzenberger AVL List GmbH David Wüterich SEG Autmotive GmbH

The future of combustion engines from the perspective of rail applications Andreas Wegmann, M. Gayer J. M. Voith SE & Co. KG | VTA B. Gay, M. Pister Liebherr Machines Bulle SA

This manuscript is not available according to publishing restriction. Thank you for your understanding.

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_1

1

Strategies for commercial vehicle drivetrains 2030 Philip Scarth, FPT Motorenforschung AG

This manuscript is not available according to publishing restriction. Thank you for your understanding.

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_2

3

The commercial vehicle in 10 years and beyond – Still with a classic ICE? Dr.-Ing. Christian Weiskirch, Dr. rer. nat. Andreas Kammel TRATON GROUP

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_3

5

The commercial vehicle in 10 years and beyond – Still with a classic ICE?

Abstract The Commercial Vehicle Business is at the beginning of more disruptive change driven by electrification, automation and digitization. New technologies will be introduced with growing frequency. In the same time the demands from legislation but even more important from society to step out of fossil energy sources towards CO2-neutral transport solutions forces the manufacturers of Commercial Vehicles to continue to improve on the classic technologies but as well manage the transition into the new areas. Long term the future for the Commercial Vehicles will be electric without any doubts, but the point in time when this will happen is discussed controversially between governments, politicians, manufactures but even within departments in the companies. And this discussion is not based on TCO-calculations, the purely TCO based tipping points will be reached in only a couple of years. Necessary infrastructures, customer acceptance, global scale and various uncoordinated incentives and initiatives are discussion topics that lead to a diffuse and fuzzy view on the total picture – which crystal ball is really the right one? Same goes for the transition period, where the manufacturers will earn the majority of their income with the classic technologies also to finance the development of the next generations. Various opportunities are already available, from more evolutionary green drop-in fuels to more revolutionary hydrogen in internal combustion engines or fuel cells. Also in this more or less well known technologies it’s unclear on which horses to put the bets on, but it is clear that it is impossible to develop solutions for all options. Looking back in time the engineers have already demonstrated that whatever challenge a new fuel brings – the necessary technology will be made available. But can that be afforded for shrinking volumes or even niche-applications? The future for in city applications like busses, distribution trucks and typical special vehicle for garbage collection, street cleaning and such is quite clear electric. The ecosystem is known and aside some infrastructure challenges and also necessary adoptions in operations the transition will happen quite fast. Also regional distribution and regional haulage can be handled with electrified vehicles quite soon. For busses operating interurban but for sure long distance maybe different solutions need to applied. Long haulage operation on the other hand can be handled with electric vehicles if you change the typical operation pattern or with the expected improvement in battery and charging technology. The situation becomes different, if the vehicle is used for work instead of transport. It makes heavy machinery mobile and with the internal combustion engine it delivers the power for the operation in almost every accessible spot in the world. And a refill of the high density energy storage is in this case easy to handle. Can we change this within 10 years or will the modern ICE survive much longer?

6

5G network infrastructure & autonomous driving – The next generation C-V2X technology Walter Haas, HUAWEI TECHNOLOGIES Deutschland GmbH

This manuscript is not available according to publishing restriction. Thank you for your understanding.

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_4

7

Welcome to the photonic age in automotive industry Martin Enenkel, Jenoptik Optical Systems GmbH

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_5

9

Welcome to the photonic age in automotive industry

1 The Photonic Age in Automotive Industry at start of page The Photonic Age within the Automotive Industry has started already long time ago for more than 100 years. Coming from candles to illuminate drive ways to today’s PIXEL Lighting and Laser beams in car headlamps or Organic-LED (OLED) in Backlighting. Long Range LiDAR-Systems (Figure 1) supports Autonomous Driving and adequate driver information systems inform pilots from multicolor head up displays (Figure 2). Car to Car commutation is used as a bi-directional communication to ensure safety handshakes. Femto-Second Laser Equipment enables manufacturers to do micro drilling to define weakness sweat points at structure elements (Figure 3 +4) with batch size one.

Figure (1)

Figure (2)

Photonics is enabler for: – Autonomous driving and connected cars – Digitalization

Figure (3)

10

Figure (4)

Welcome to the photonic age in automotive industry Beam shaping optics allows to adjust laser tools to the customer needs in terms of material to handle, process timing and fit, from and function. Compared with multi axis robot cells, photonics makes the car production fast, more stable and more flexible (Figure 5+6).

Figure (5)

Figure (6)

Today’s automotive industry is not only about car manufacturing or car parts and equipment, it is also about the environment. Traffic control, law enforcement (Figure 7) and safety and security (Figure 8) are major investment areas for photonics in the automotive industry.

Figure (7)

Figure (8)

Photonics makes roads and communities safer

11

Welcome to the photonic age in automotive industry

2 Jenoptik Divisions are addressing areas in the Automotive Industry Jenoptik addresses the automotive industry with his 3 divisions ● Light and Production (L&P) ● Light and Optic (L&O) ● Light and Safety (L&S) Light and Production Division (L&P) provides engineering with focus on smart manufacturing and process automation for automotive industrial customers (Figure 9). Therefore Photonics makes the car production faster, more stable and more flexible.

Figure (9) Division Light & Production for car manufacturing

Light and Optics Division (L&O) provides engineering with focus on state of the art opto-electronics. From component level, over modules (Figure 10) to full system solution for full integration as Tier#2 supplier. Therefore Jenoptik brings Photonics at the heart of our OEM customers.

Figure (10) Division Light & Optics for car on board systems

12

Welcome to the photonic age in automotive industry Light and Safety Division (L&S) provides Imaging based solutions for Public Safety in combination with intelligent data management, camera vision and machine learning software to address road safety (Figure 11).

Figure (11) Division Light & Safety for road safety

Therefore through Photonics Jenoptik makes Road and communities more safe.

3 Light & Production 3.1 Photonics Makes the Car Production Faster, More Stable and More Flexible Laser processing with precise and safe 3-D-laser machines for perforation, cutting and welding of plastics, metals and sensitive materials enables faster, more stable and more flexible manufacturing (Figure 12).

Figure (12) Laser Processing

13

Welcome to the photonic age in automotive industry The Metrology Business Unit within L&P supports through Photonic Metrology Methods for (Figure 13): ● High-precision Metrology ● Roughness & Contour Metrology ● Form Metrology ● Optical Shaft Metrology ● In-Process Metrology ● Visual Surface Inspection

Figure (13) Metrology

Automation group within L&P Division keeps deep engineering competence for automated assembly equipment (Figure 14).

Figure (14) Integration & Automation

14

Welcome to the photonic age in automotive industry

3.2 Future Business Trends & Challenges in Automotive Production Vacuum lamination manufacturing processes require micro holes & or slits inside interior or exterior components to process, but invisible to detect by human eye (Figure 15).

Figure (15) Pitching

Weakening of the steering wheel covers to create a tear line for invisible airbags (Figure 16).

Figure (16) Scoring

Visible and invisible light takes over essential functions for more safety and comfort. Therefore surface structuring by laser processing of interior/ exterior components for light integration enables new functionalities (Figure 17).

15

Welcome to the photonic age in automotive industry

Figure (17) back-lighting

Car body manufacturing with laser by melting of the joining partners renew common processes (Figure 18).

Figure (18) Remote Welding Metal

Laser carving the tread geometries into slicks generates individual customizes tire surface geometries for specified application conditions (Figure 19).

Figure (19) Tire Carving

16

Welcome to the photonic age in automotive industry

4 Light & Optics 4.1 Photonics Makes the Car Autonomous, Smarter and Bathed in Light Optoelectronics solution for car sensing like Short, Mid and Long-Range LiDAR Sensorik (Figure 20), Driver information System like Head Up Displays ( Figure 21) or Dust Particle Sensors (Figure 22) are only possible through customized high end optics. The Optic design could be athermal for high temperature range, complex and/or freeform shape. LiDAR Systems require high end components like Multi-Lens-Arrays (MLA), Defractive-Optical-Elements (DOE) made of Glass or Polymer materials. Also for customized SAC & FAC’s with Laser Edge Emitters are from the industry required. Active alignment System enables highest accuracy of optical components in complex systems architectures for reliability and reproducible data collection.

Figure (20) LiDAR

Figure (21) HUD

17

Welcome to the photonic age in automotive industry

Figure (22) Particle Sensor

4.2 Future Business Trends & Challenges in Automotive Sensing and Interior Light ● Homologated standardized LiDAR testing are not established ● use cases are growing as accidents occur ● there is not one LiDAR position that fits all use cases but the use case will define the position of the LiDAR(s) ● integration into headlamps preferred place of LiDAR for car manufacturers (at corners best Field of View) ● Lidar has to be able to understand severity of contamination to trigger cleaning ●  Cleaning of LiDAR glass or headlamp glass has to get smarter/ automated ● Fusion of LiDAR with camera and radar offers the required redundancy ● LiDAR systems may require different light sources ● Automotive certified MEMS mirrors for micro scanning of in & outbound signal

5 Light & Safety 5.1 Photonics Makes Roads and Communities Safer Photonics at road sides makes roads and communities safer. Camera vision systems compare with machine learning algorithm enables car classification solutions for car for road charge by toll enforcement (Figure 23).

18

Welcome to the photonic age in automotive industry

Figure (23)

Jenoptik Traffic Tower (Figure 24) at Roadside ensures safe traffic through law enforcement (e, g, speeding, redlight…).

Figure (24)

Automated number plate recognition (ANPR) by camera vision systems handle more than 100 million plate captures per day (Figure 25).

19

Welcome to the photonic age in automotive industry

Figure (25)

5.2 Future Business Trends & Challenges in Road Safety Public Authorities are focusing on ● Enforcing speed limits on spots or between two points of road ● Automatic monitoring of “Emissions Zone” to ensure only vehicles with low CO2 can enter zone (e.g. Ghent: city center) ● ANPR capture data stored for minimum of 2 years, enabling lifestyle pattern analysis of criminals or terrorists ● Extending Toll Collect from highways to secondary main roads

6 Summary 6.1 Your Take Away by Jenoptik ● Modern photonics traffic safety solutions are around us every day which supports safety, security and environment protection ● Deep engineering competence in automated laser processing enables new designs for safety and comfort ● Efficient, precise and safe 3D laser machines are available for perforation, cutting and welding of plastics, metals and sensitive materials ● Different LiDAR systems may require different light sources and car positions, because there is not one positions in the car that fits all

20

Welcome to the photonic age in automotive industry ● Lidar Systems have to fulfill many several use cases, but there is not one position and system that fits all ● Homologated standardized LiDAR testing are not established yet and this needs to be pushed by the industry

Bibliography 1. Power Point Presentations of the DVN Workshop LiDAR Conference December 2019 2. Figure (1) https://www.alamy.de/stockfoto-las-vegas-nv-usa-8-januar-2016-velodyne-lidar-puck-bild-fur-selbstfahrer-autos-auf-dem-display-wahrend-der-ces-2016fridayphoto-gen-blevinsla-daily-newszumapress-kredit-zeigen-gen-blevinszumadrahtalamy-live-news-92895875.html, 12.2019 3. Figure (2) https://automotive-technology.de/bmw-head-up-display/, 12.2019 4. Figure (3) unknown 5. Figure (4) unknown 6. Figure (5) Jenoptik 7. Figure (6) Jenoptik 8. Figure (7) Jenoptik 9. Figure (8) Jenoptik 10. Figure (9) Jenoptik 11. Figure (10) Jenoptik 12. Figure (11) Jenoptik 13. Figure (12) Jenoptik 14. Figure (13) Jenoptik 15. Figure (14) Jenoptik 16. Figure (15) unknown 17. Figure (16) unknown 18. Figure (17) unknown 19. Figure (18) unknown

21

Welcome to the photonic age in automotive industry 20. Figure (19) unknown 21. Figure (20) https://www.alamy.de/stockfoto-las-vegas-nv-usa-8-januar-2016-velodyne-lidar-puck-bild-fur-selbstfahrer-autos-auf-dem-display-wahrend-der-ces-2016fridayphoto-gen-blevinsla-daily-newszumapress-kredit-zeigen-gen-blevinszumadrahtalamy-live-news-92895875.html, 12.2019 22. Figure (21) https://automotive-technology.de/bmw-head-up-display/, 12.2019 23. Figure (22) 24. Figure (23) Jenoptik 25. Figure (24) Jenoptik 26. Figure (25) Jenoptik

22

SEAT’s digital transformation in two dimensions Dr. Sebastian Grams, SEAT S.A., VW Group

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_6

23

SEAT’s digital transformation in two dimensions

Abstract In this paper we describe a two-dimension compass used in SEAT as a digitalization framework to define and prioritize IT projects to generate as much benefit possible for the company. Key works: digitalization, digital transformation, digital journey, digital strategy

1 Introduction During the last 5 years there has been a ‘new wave’ of digital transformation programs has gained popularity in large manufacturing organizations where the core business are products or services not directly associated with information technologies. This digital transformation does not only include the roll-out of specific technologies inside the company to support existing processes, but companywide culture changes as well to embrace the digitalization initiatives that explore the possibilities to generate new types of digital services beyond the core business. When describing a long-term evolution through digitalization we may refer as the “digital journey” of an organization and we start hearing about ‘hype’ concepts in every office of promises that will transform the company like: agile, scrum, kanban, apps, UX, MVPs, digital workplace, VR, big data, Machine learning, Robotic process automation, etc. Car manufacturers are no exception to the rule as old brands have been challenged by new commers to the mobility industry.

1.1 Challenge for the company For traditional companies (as Car manufacturers), the technical enablement responsibility for the digitalization journey usually resides in the IT department. Together, IT departments and board members or top management of the organization must decide where and when to invest to: ● Optimize productivity to defend their position in the market ● Search for new opportunities in the digital age where currently the new challengers born in the digital age, act as 100% data driven companies This situation generates a sense of urgency for the ‘old fashion’ companies to invest in digitalization, yet for most of the management and IT departments the question remains the same: Out of all the technologies and solutions out in the market, how to know what is the right criteria for a company to invest on?

24

SEAT’s digital transformation in two dimensions For these organizations, there are natural impediments to execute a successful digital transformation including: ● Silos thinking from different business units with different objectives, often contradictory ● Tight rules for budget management and low flexibility for execution changes, opposite to current agile development practices ● Low percentage of edge technologies knowledge, a weak position in the digital age ● Fear of failure, leaving no space for a culture of innovation or pivot to reconduct strategy

2 Defining a digital vision at SEAT We have defined for SEAT different fields of action for digitalization focus to evolve the company from car manufacturer into data-driven top player in the mobility as a service business. This we call SEAT’s digitalization vision (See figure 1). In a first field of action, digital services associated to the connected car can bring on top revenue from car sales. However, on a second level, as customer behaviour has changed and for many people owning a car is no priority, we have started providing car as a service and mobility as a service through digital platforms where the monetization is not anymore directly coming from the car buyer but form the €/km provided to the customer. On the top of this offerings, data exchange and digital transactions associated to mobility beyond the vehicle itself, will enable data monetization for B2B scenarios. Accompanying the previous concepts explained, there should be a continuous effort to achieve processes efficiencies (reduce cost, increase productivity and maintain quality levels) through digitalization. Examples of this could be paperless processes, Workflow automation, Deployment of collaboration tools & Data management tools for strong governance practices.

25

SEAT’s digital transformation in two dimensions

Figure 1: Digital vision to transform SEAT into a top data-driven service provider

2.1 After defining a Digital Vision Through the definition of a digital vision we set a common final objective. However, as stated before, siloed organizations can find themselves in internal conflicts and harsh situations where every investment decision starts to be questioned. Even worst, Business Units with large budgets can start acquiring digital solutions without a long-term view in benefit for the whole company. How to create a common language for all the company? How to generate trust between different department’s managers to understand the impact of different digitalization initiatives? How to agree on something that might be/sound too abstract like ‘robot process automation’ or ‘network renovation’? how to reduce stress when prioritizing GDPR compliance changes?

3 Setting a same language: Digitalization framework A graphical representation of the sequenced end-to-end supply chain without business unit names can allow every person in the organization to easily understand how the core business is composed. The following figure represents SEAT’s digitalization compass, also known as axis model. Though this compass, SEAT IT has been able to define a common digitalization language for the company.

26

SEAT’s digital transformation in two dimensions

Figure 2: SEAT’s axis model as compass though digital journey

The axis model is composed of 2 transversal axes representing two dimensions. On one dimension (the vertical axis) our core business is represented. Over the whole axis we represent the complete end-to-end value chain, starting from R&D, purchasing, Order management, Inbound logistics, Manufacturing control, Back office functions like HR/Financial Controlling/Communication, Outbound logistics, Omnichannel and final customer. Right in the middle of the axis, the factory as the core of the supply chain. The first half of the axis represents the value chain ‘into the factory’, the second half of the axis, once the product is manufactured, represents the value chain ‘out of the factory’. As a transversal dimension, a new horizontal axis represents the supply chain of new business models in the digital age, in our example we assume that the car/vehicle manufacturer will become a Mobility as a Service provider. Therefore, the value chain includes fleet management back office, data management, payment systems, digital platforms for commercialization towards final users. Right in the middle of the axis we represent the connected vehicle as the core of this value chain. The first half of the horizontal axis represents the value chain as ‘enablers for future offerings’, the second half of the horizontal axis, once the product is fully digitalized and generating data (e.g. connected car), represents the value chain for the ‘future digital offerings.

27

SEAT’s digital transformation in two dimensions Finally, in the bottom of the compass, a common base represents all the IT ‘compulsory’ projects needed to run the daily business, for instance software upgrades or legal compliance modifications such as GDPR requirements.

4 How to use the axis model? The compass becomes a framework for all the company where all IT and digitalization initiatives are represented as a ‘dot’ located close to the value chain concept that describes the best, where the benefit will be in case of executing that project. Every person that refers to this ‘map’ should relatively easy understand not only a ROI perspective, but to identify easily where the impact of the investment should take effect, no matter if it is a non-technical profile. To make it as a systematic procedure, as part of the demand management activities of the IT department, all new projects that the Business units request to be considered for the following investment round are mapped (in common agreement between IT and the Product/process Owners) to the axis model right where the supply chain will be affected, or of applicable, where the new business supply chain is enabled. Also, as part of the limitations for investment, if a project is required to keep the business running or a legal obligation in IT systems requires investment, these projects are mapped to the base of the axis model.

Figure 3: Projects mapped in the canvas axis model

28

SEAT’s digital transformation in two dimensions

4.1 Benefits of the axis model The consequence of having this visual representation provides an easy and transparent way to document, classify and discuss the projects contained in the digitalization portfolio from a broader perspective for the whole company in such a way that we avoid: ● Silos planning, as no business unit represents positions in the compass ● Cannibalizing IT capacity, as it is clear the size of the portfolio ● Find synergies in a common strategic planning, to see where same technologies apply for different problems As explained, projects that pay-off into our canvas axis cross in either of the 2 dimensions: X, Y and these are the ones that the company should consider the most strategic programs. Projects without a clear contribution to the axis cross are discarded. For consistency with the digital vision, digitalization programs require as well to identify how data from digitalization is monetized, therefore a translation of payback in terms of €/service & €/km or optimize the €/car and direct manufacturing savings in the supply chain is required. This way of translating in figures the IT projects, helps the IT department to show in transparent way how digitalization budget is distributed in the company.

Figure 4: Projects clustered according to their contribution and the distribution of the budget assigned

29

SEAT’s digital transformation in two dimensions

5 Transformation journey at SEAT With a Digital vision defined for the company and the 2-dimension digitalization framework, IT department in SEAT had as well a huge challenge ahead: Deliver the best and modern solution in the fastest way to really become a strategic player for the company. Here we summarise some of the people transformation measures: ● Transforming common old-style project manager profiles into more technical lead approach. ● Creation of the Agile Center of Excellence with scrum masters and agile coaches to facilitate the project delivery. ● Creation of a Software Development Center to generate internal knowledge and find economical efficiencies. Though the creation of this Software Developer Center, called SEAT:CODE, in a Startup ecosystem we have guaranteed to reduce the time to market of the solutions, increase our capacity, generated cost reductions and transform the organization from different business units through learning by doing process as Product owners. The company has started to leave behind its own limitations like fear to fail and lack of software development technical know-how. The following figure shows 3 examples of products delivered by SEAT:CODE as main technical enabler for the Digital transformation in SEAT. First, we can identify on which axis the project is classified, either serving to the new business models enabled by digitalization, or a supply chain optimization to the core business. On the right side of the compass, a description of each project is done and a brief description about the delivery lead time and project complexity/scope.

Figure 5

30

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles T. Bieniek, K. Rößler, F. Otto Mercedes-Benz AG

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_7

31

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles

1 Introduction Electricity-based fuels allow a climate neutral operation of powertrains, including internal combustion engines (ICE). However, today such energy sources are characterized by higher costs compared to fossil fuels, which is not acceptable to the customer. A plug-in hybrid electric vehicle appears to be a beneficial solution. With a conveniently designed battery capacity, most of the average driving distance can be realized electrically, while the ICE has advantages for long-distance trips. In this case, the fuel costs play a minor role and the drawbacks of the battery electric vehicles, such as the limited range, long charging times, insufficient charging infrastructure, high weight and costs of the battery, can be minimized. There is a controversial discussion about the most suitable synthetic fuel. Methanol is a well-known raw material in the chemical industry and a promising candidate as it has favorable properties when used in internal combustion engines. Furthermore, it is liquid at ambient conditions, which is interesting from an infrastructural point of view. As a spark-ignition (SI) fuel, methanol additionally has advantages over other liquid electricity-based fuels in terms of cost [1]. The joint research project MEEMO, which is supported by the German Federal Ministry for Economic Affairs and Energy, aims to contribute to the discussion of such electricity-based fuels.

2 Properties of Methanol To highlight one of the main differences, methanol (CH3OH) is a pure substance while gasoline consists of a variety of hydrocarbons. With regard to the molecular structure, the hydroxyl (OH) group causes a distinct polarity of the methanol molecule. Thus, intermolecular forces are characterized by hydrogen bonds leading to different material properties compared to conventional gasoline [2,3]. This chapter discusses these characteristics and their impact on the operation of internal combustion engines in relation to gasoline. Please note that in order to simplify the comparison, isooctane (C8H18) will be considered as a reference fuel for gasoline in the discussion below.

2.1 Oxygen Content In contrast to gasoline, alcohol fuels incorporate oxygen in their molecular structure, which influences the stoichiometric air-to-fuel ratio (AFR) and the fuel’s mass-based energy content. Among the different alcohols, the mass-specific oxygen content increases with declining length of the molecule’s carbon chain leading to methanol having

32

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles Table 1: Properties of methanol and isooctane [2, 4–6] Chemical formula Molecular weight [kg/kmol] AFR [kg/kg] Lower heating value (LHV) [MJ/kg] LHV per unit mass of air (LHV/A) [MJ/kg] LHV per unit mass of mixture (LHV/M) [MJ/kg] Heat of vaporization (HoV) [kJ/kg] HoV per unit mass of air (HoV/A) [kJ/kg] Specific CO2 emissions [g/MJ] Adiabatic flame temperature [K]

Isooctane

Methanol

C8H18 114.23 15.04 44.30 2.95 2.76 302.1 20.09 69.58 2276

CH3OH 32.04 6.43 19.93 3.10 2.68 1102.0 171.33 68.92 2143

the largest oxygen content (around 50% of mass) [3]. Hence, the use of methanol as an alternative fuel has the largest impact on the aforementioned properties, which are explained in depth in the following subsections.

2.1.1 Air-to-Fuel Ratio Regarding the air-to-fuel ratio, the high oxygen content of methanol results in a lower additional air mass requirement for stoichiometric combustion compared to gasoline (see Table 1). In the context of direct injection (DI) internal combustion engines, the necessary fuel mass is more than doubled compared to gasoline (AFR ≈ 14.7), if constant volumetric efficiency is assumed. Especially with regard to high engine loads, this significantly affects the design of the fuel system in order to keep injection times and consequently mixture formation on an acceptable level. As a result, even higher pressures in case of DI or an additional port fuel injection (PFI) might be required.

2.1.2 Lower Heating Value (LHV) As mentioned above, the oxygen fraction also plays an important role in conjunction with the contained energy per unit mass of fuel. In this case, the parameter of interest is the lower heating value (LHV), which consequently is the lowest for methanol compared to other alcohols and 55% lower than that of isooctane. In the DI engine concept, however, the amount of energy per unit mass of inflowing air is of greater interest. This parameter incorporates the air-to-fuel ratio, increasing the heat release and consequently the brake mean effective pressure (BMEP) by 5% during combustion of methanol relative to isooctane (see Table 1), assuming equal volumetric efficiency and a start of injection (SOI) after the inlet valves are closed. In case of an

33

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles external carburetion (PFI), the energy content per unit mass of the mixture is considered. Here, Table 1 shows the opposite relationship, where the energy content for a methanol-air mixture is 3% lower. Nevertheless, the enhanced charge cooling effect of methanol due to its higher heat of vaporization might compensate this disadvantage [2].

2.2 H/C Ratio The hydrogen-to-carbon (H/C) ratio of a fuel has an impact on the formation of the two main combustion products, carbon dioxide (CO2) and water (H2O). Methanol is characterized by an increased H/C-ratio compared to the reference fuel isooctane [2]. Referring to the energy-specific carbon dioxide emissions, methanol reduces the output of CO2 by one percent compared to isooctane (see Table 1). Verhelst et al. [2] state that, this reduction potential in specific CO2 emissions rises to seven percent when conventional gasoline is considered instead of isooctane.

2.3 Heat of Vaporization As already mentioned at the beginning of this chapter, molecular interaction is governed by hydrogen bonds in the case of methanol. This applies to all alcohol molecules, in which these forces decline with increasing length of the molecule’s carbon chain [3]. Since the strength of intermolecular bonds influences the energy needed to change the state of aggregation from liquid to gaseous, the foregoing relation among the alcohols also applies to the heat of vaporization. Consequently, this parameter is the highest for methanol and 3.65 times greater in comparison with isooctane. This gap rises to a more than eightfold increase, when the cooling effect per unit mass of air is considered (see Table 1), which is beneficial for engine operation in many respects. At first, cylinder gas temperatures are reduced, resulting in lower heat loss throughout the engine cycle and thus in a greater efficiency [2]. This gets even more interesting in the area of the engine’s rated maximum power where, in case of conventional fuel, full load enrichment is commonly applied in order to decrease exhaust gas temperatures and hence protect the outlet valves and the turbine entry from thermal stress at high power outputs. As Chapter 4 will show, the strengths of methanol are particularly visible in this region of the engine map. Moreover, since temperature is a main factor in the formation of NOx, decreased in-cylinder temperatures may reduce emissions of this pollutant. This is further supported by an increased heat capacity of the exhaust gas and a reduced adiabatic flame temperature (see Table 1) [2]. In terms of engine operation with conventional gasoline, knocking phenomena are a common issue at higher loads, leading to retarded ignition angles in order to reduce the

34

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles pressure and temperature level of the unburned gas and hence inhibit self-ignition processes. The negative consequence is a loss in engine efficiency since fuel energy tends to heat up the exhaust gas instead of producing mechanical energy at the crankshaft under these circumstances. Besides the increased octane rating (methanol: 109 [2], gasoline: 95), the high cooling potential of methanol makes it possible to circumvent this problem to some extent, especially in DI engines. As chemical kinetics in the unburned gas follow an Arrhenius approach, a reduced temperature in the combustion chamber leads to decelerated reaction rates 𝑟 (Eq. 1) [7]. 𝑟 = 𝐾∙𝑒

(1)

This enables the flame front to capture the unburned gas zone before it self-ignites, thereby suppressing knocking. Thus, engine operation at optimum ignition angles is possible in a wider map range, resulting in increased thermal efficiency.

2.4 Boiling Characteristics and Vapor Pressure With regard to the combustion engine’s cold-start capability, the fuel’s boiling characteristics play an important role. Since methanol is a one-component liquid, it has a single boiling point at 64.7 °C at ambient pressure [5]. In contrast, gasoline involves several species with different individual boiling points. This results in a boiling curve describing the amount of evaporated fuel mass as a function of its temperature. Considering temperatures relevant for engine cold-start, the vapor pressure of gasoline is consequently higher compared to methanol due to gasoline’s low-boiling components [5]. Hence, together with the aforementioned large heat of vaporization, methanol fuel negatively affects cold-start performance necessitating special measures in order to ensure acceptable starting times. The latter are explained in detail in Chapter 4.

3 Experimental Setup Engine tests run at Daimler’s Research & Development division in Untertürkheim, Germany. Investigations are based on a Mercedes-Benz M282 4-cylinder gasoline engine and focus on stationary engine operation on the one hand and cold-start tests on the other. The latter are performed via the integrated starter motor within a climate test-bed on which the intake air as well as the engine coolant can be conditioned down to - 20 °C. The following table shows the basic specifications of the M282 engine. Due to methanol’s higher knock resistance, different engine configurations are utilized, incorporating increased compression ratios. Furthermore, a larger turbocharger is installed (configuration 4) in order to benefit from the alcohol’s advantageous properties at the engine’s rated maximum power, which is detailed in Chapter 4. Fuel injection is realized using

35

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles centrally positioned six-hole direct injection valves with the associated high-pressure pump (HPP) delivering a maximum rail pressure of 250 bar. With regard to the applied fuel types, tests are performed using gasoline, which consists of a mixture of 90% unleaded gasoline and 10 % ethanol, as well as IMPCA-specified methanol, which is referred to as M100 in the following sections. Table 2: Engine data and configurations [8] Mercedes-Benz M282 Cylinder Displacement [cm3] Bore [mm] Stroke [mm] pmax [bar] Configuration Compression ratio Turbocharger

4 1332 72.2 81.3 120 1 10.5 GT12

2 12.0 GT12

3 14.0 GT12

4 14.0 GT15

4 Experimental Results In this chapter the results of the stationary engine tests as well as the cold-start investigations are presented. The former are carried out at stoichiometric combustion conditions across the entire engine map. With regard to the engine’s mechanical stress, a peak cylinder pressure pmax of 120 bar is accepted, representing the average over all four cylinders and 500 cycles.

4.1 Wide-Open Throttle (WOT) 4.1.1 Fuel Impact The following lines focus on the influence of the applied fuel on the engine’s behavior at wide-open throttle conditions based on the standard engine configuration (see Table 2, configuration 1). Maximum engine load is achieved through adaptation of the ignition angle on the one hand and the waste-gate position on the other.

36

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles Gasoline

M100

Knock

pmax T3

HPP BMEP [bar]

BMEP [bar]

IMEP 30 25 20

30 25 20

115 kW

15 10

TC RPM

1000

2000

3000 4000 RPM

130 kW

15 10

5000 6000

1000 2000

3000 4000 RPM

5000

6000

Figure 4.1: BMEP limiting parameters in case of gasoline (left) and M100 (right)

Figure 4.1 shows the obtained brake mean effective pressures for both fuels and their limiting parameters, which differ significantly. Starting with the gasoline-fueled engine test, the knock limitation over the entire engine speed range is apparent. Hence, retarded ignition angles are necessary in order to allow higher boost pressures (waste-gate closed further) and thus increased BMEP. This procedure is restricted by the engine smoothness σIMEP at lower engine speeds (1500 - 3000 rpm) as well as the exhaust gas temperature T3 at elevated revolutions. As a result, the engine’s maximum power output when 40

1100 M100 Gasoline

1000

30

900

25

800

T3 [°C]

MFB50 [°CA]

35

20

700

15

600

10

500

5

1000

2000

3000 4000 RPM

5000

6000

Exceeded exhaust gas temperature

400

M100 Gasoline 1000

2000 3000 4000 RPM

5000

6000

Figure 4.2: 50 % mass fraction burned MFB50 (left) and exhaust gas temperature T3 (right)

running on gasoline at stoichiometric conditions is 115 kW at 5000 rpm. In contrast, M100 does not show knocking combustion phenomena, which is why pmax becomes the limiting factor. Hence, earlier ignition angles are realized, which is represented on the left side of Figure 4.2 by means of the 50% mass fraction burned value MFB50. This also positively affects the engine smoothness. However, further BMEP improvement is inhibited by the maximum delivery rate of the high-pressure pump due to methanol’s

37

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles more than halved energy content. As already introduced in Chapter 2, the large heat of vaporization of methanol creates a strong charge cooling effect which significantly lowers T3 (approx. 85 °C, see Figure 4.2) and hence enables an improved maximum power output (130 kW at 5000 rpm). Even higher engine performance is prevented by the maximum turbocharger speed at engine revolutions above 4500 rpm.

4.1.2 Skipped Intake Air Cooling Chapter 2 introduced the beneficial properties of methanol regarding knock resistance, such as an elevated octane rating or the significantly increased cooling potential of the combustible mixture. This opens up the ability to reduce the performance of the intake air cooling system, which could be a beneficial measure from a cost and packaging perspective. Hence, this section focuses on corresponding WOT tests based on engine configuration 2. 140

p2 TCRPM TCRPM

100

30 25 20

40

127 kW 1000

2000

3000 4000 RPM

80 60

130 kW

15 10

cooled non-cooled

120

T2 [°C]

BMEP [bar]

cooled non-cooled

5000

6000

20

1000

2000

3000 4000 RPM

5000

6000

Figure 4.3: BMEP (left) and charge air temperature T2 (right) for cooled and non-cooled intake air

As the dotted line on the right hand side of Figure 4.3 indicates, intake air temperatures T2 are significantly elevated without charge air cooling. The aim of this investigation is to approximate the brake mean effective pressures of the WOT results with cooled inflowing air. This is achieved by a waste-gate that is closed further in order to compensate the density-loss due to the hotter intake air. In accordance with the left-hand side of Figure 4.3, comparable power outputs can be generated without running into knock issues even in the case of non-cooled charge air. Slight disadvantages in terms of BMEP only need to be accepted at revolutions above 3500 rpm as the maximum allowed intake air pressure p2 and again the turbocharger speed become limiting parameters. Besides this, the pmax restriction is maintained in the case of turned-off charge air cooling.

38

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles

4.1.3 Effect of Turbocharger and Injector Adjustment The two foregoing sections already revealed that the methanol-fueled engine would benefit from an enlarged turbocharger. For this reason, further wide-open throttle measurements are conducted using a 10% increased turbine as well as a 20% scaled-up compressor diameter, the results of which are represented by Figure 4.4. The latter are based on engine configurations 3 and 4 including a compression ratio of 14.0. As indicated on the left, the scaled-up turbocharger allows notably elevated loads at high revolutions without running into critical shaft speeds. As a result, the engine’s rated maximum power output rises by 20% up to 155 kW at 5500 rpm. 300

T2

30 25

155 kW

20 15 10

129 kW 1000

2000

3000 4000 RPM

250 tinj [CAD]

BMEP [bar]

standard TC scaled-up TC

5000 6000

standard TC scaled-up TC

200 150 100 Reduced injection duration due to increased injector flow

50 0

1000 2000

3000 4000 RPM

5000

6000

Figure 4.4: Effect of turbocharger adjustment on BMEP (left) and injection duration (right)

However, this also leads to raised air temperatures downstream from the compressor T2 in this operating range, forming the restricting factor of this engine set-up. Due to this limitation, analyses without charge air cooling were not conducted at this engine configuration. As expected, the poorer flow-mechanical response of the bigger turbocharger results in a delayed low-end torque. The right plot in Figure 4.4 unfolds a further methanol specific issue regarding the injection duration. Even in the case of the standard turbocharger and moderate power outputs (129 kW at 5500 rpm), injection lasts nearly 270 CAD. Originating from this injection system, a power output of 155 kW would consequently require injection durations of around 325 CAD, which might lead to inappropriate mixture formation. Thus, engine configuration 4 is additionally equipped with injectors that are characterized by a 40% increased flow rate in order to avoid excessive injection durations.

39

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles

4.2 Part Load Operation 4.2.1 Fuel Consumption of Gasoline and M100 Based on engine configuration 2, analyses focus on the effect of the applied fuel on the brake specific fuel consumption (BSFC) at medium load ranges. For this reason, Figure 4.5 at first illustrates the absolute advance of MFB50 of M100 compared to gasoline. Here, the strong knock limitation of the non-alcohol fuel becomes obvious at

MFB50 [CAD]

BSFC [%] 20

-10 15 -8

-10

-6

IMEP [bar]

IMEP [bar]

20

-2 -4

10

-8

15

-6

-4

10 1500

2500 RPM

3500

1500

2500 RPM

3500

Figure 4.5: Changes in MFB50 (left) and BSFC (right) of M100 relative to gasoline

low revolutions and high loads. An explanation for this is that low engine speeds lead to more time being available for chemical reactions in the unburned gas zone, which causes pre-ignition and hence knock phenomena. As known from the investigations at wide-open throttle conditions, M100 does not tend to show knock occurrences. Thus, M100 allows optimum ignition angles (MFB50 at 8 CAD) over nearly the entire engine map shown here, which is one reason for the distinct improvement in fuel consumption described by the right-hand side of Figure 4.5. Further positive aspects are the reduced heat loss, due to lowered gas temperatures (high heat of vaporization) and an increased laminar burning velocity [2]. It should be noted that no comparative data exists for an engine load of 20 bar (IMEP) at 1500 rpm since this operation point could not be reached with gasoline.

4.2.2 Fuel Consumption of M100 at High Compression Ratios Concentrating on the operation with M100, Figure 4.6 gives evidence of the absolute values for MFB50 and BSFC of the M282 engine equipped with a compression ratio of 14.0 (configuration 3). The fuel consumption is normalized to a lower heating value of 42500 kJ/kg in order to enable better comparability to typical gasoline. Obviously,

40

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles M100 allows for optimum MFB50 angles until cylinder peak pressures pmax are reached at higher loads. This is beneficial regarding fuel consumption with a minimum of 200 g/kWh (at 2500 rpm and 17 bar IMEP), which corresponds to an engine efficiency of 42%. Further, BSFC tends to decrease with rising engine loads since gas exchange losses also decline.

MFB50 [CAD] 20 15

20

10

8

IMEP [bar]

IMEP [bar]

9

BSFCnorm [g/kWh]

10

200

15

205 210

10

8

230

5

5 1500

2500 RPM

3500

1500

2500 RPM

3500

Figure 4.6: 50% mass fraction burned (left) and normalized BSFC (right) at a compression ratio of 14.0

4.2.3 Emissions Chapter 4.2 concludes by taking a closer look at the part load emissions of M100 compared to gasoline. Engine set-up 2 is considered for this purpose as in section 4.2.1. Referring to the left diagram in Figure 4.7, M100 exhibits advantages in terms of nitrogen oxide (NOx) emissions. Since the formation of such pollutants is favored at increased temperatures [9], an explanation for the reduced emissions in case of methanol is its lower adiabatic flame temperature (see Table 1) as well as its large heat of vaporization. Since peak temperatures rise with earlier MFB50 angles, this benefit of M100 declines towards lower engine speeds and elevated loads (see Figure 4.5 and 4.7). In terms of the particle emissions, denoted on the right-hand side of Figure 4.7, a massive reduction of over 85 % is achievable when the engine runs on methanol. As described by Svensson et al. [10], this is due to methanol containing only one carbon atom while precursors of soot require at least C2 molecules. They emphasize this, by means of computational studies of methanol combustion showing only a slight tendency to form soot at high temperatures under fuel rich conditions [10].

41

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles  NO [%]

PN [%]

x

-20

-8

-99

-99

20

20 IMEP [bar]

IMEP [bar]

-10 -7 15 -20

-99 15 -88

-30 10

10

1500

2500 RPM

3500

-92

-96

-84

1500

2500 RPM

3500

Figure 4.7: Relative changes in nitrogen oxide (NOx) (left) and particle number (PN) emissions (right) of M100 compared to gasoline

4.3 Cold-Start Tests Engine start under cold conditions is a common issue when it comes to alcohol fuels. With regard to methanol, reasons for this are among others its distinct charge cooling potential, the relatively low vapor pressure and its elevated lower flammability limit compared to gasoline [2]. One further aspect is that methanol does not contain low-boiling fractions that support the engine’s cold-start capability [2,5].

1

Signal

Injection Ignition

0 30.0

25.0

20.0

15.0

10.0

5.0

0.0

-5.0

CAD BTDC Figure 4.8: Schematic diagram of the applied cold-start injection strategy at -20 °C according to [11]

However, according to [2] and [3], charge stratification appears to be an applicable solution in case of direct injection as temperatures are increased and hence evaporation is assisted at the end of the compression stroke. Hence, the cold-start investigations in this work also utilize late injection as Figure 4.8 reveals. The strategy applied at temperatures of -20 °C is illustrated here. Injection is split into two pulses in order to elevate the turbulence in the combustion chamber and thus support mixture formation. The ratio

42

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles of these injections is 70 to 30%, with the first one starting at 27 CAD BTDC while the second one begins at 10 CAD BTDC. Ignition is located between these pulses at 19 CAD BTDC, overlapping approximately 3 CAD with the first injection. Applying this strategy, acceptable engine cold-starts can be realized, as Figure 4.9 illustrates. However, there is a certain time gap for pressure build-up (PBU) ΔtPBU of around 0.5 seconds between the releases of the first injection in both cases. While fuel injection starts at around 130 bar in the case of gasoline, M100 necessitates a rail pressure of 240 bar in order to enable proper engine starts, which underlines its aggravated mixture formation. Gasoline M100

RPM

1800

 tPBU

1400 1000 600

Temperature: -20 °C

200 0.0

0.5

1.0

1.5

2.0

2.5

3.0

Time [s] Figure 4.9: Time of engine start-up for gasoline and M100 at -20 °C according to [11]

Moreover, engine ramp-up times are extended compared to gasoline, indicating that combustion does not run properly in some cycles. Hence, further measures are required in order to overcome this disadvantage.

5 Challenges Regarding the Application of M100 Besides the advantageous thermodynamic properties, the application of methanol as a fuel for internal combustion engines also involves challenges regarding mechanical wear, which was most obvious in the fuel system and on the cylinder liners. Regarding the fuel system, the use of M100 leads to cavitation erosion at the injectors and the high-pressure pump, which affects the spray quality and the system’s pressure retention. Moreover, the cylinder liners exhibited areas with significant material loss. One possible reason for the increased liner wear is suboptimal spray targeting, which, in conjunction with the extended injection times, washes away the lubricant film and leads to elevated abrasion. Another theory is that intermediates of methanol combustion, especially formic acid, react with the iron of the cylinder liner and result in significant material loss. This is supported by investigations of Ryan et al. [12].

43

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles

6 Summary, Conclusion and Outlook This paper investigated the impact of pure methanol fuel (M100) on the operation of turbocharged, direct injection combustion engines. For this purpose, analyses focused on stationary engine tests as well as cold-start experiments. The former revealed significant benefits from a thermodynamic perspective, such as a reduced tendency to knock even at elevated inlet air temperatures, allowing for improved engine efficiency as well as an increased maximum rated power output at stoichiometric engine operation due to lowered exhaust gas temperatures. Moreover, methanol showed slight advantages regarding the output of nitrogen oxides especially at lower engine loads, as well as significantly reduced soot emissions in part load operation. Since alcohol fuels negatively affect cold-start capability, a particular strategy was developed involving increased fuel pressures and late injection consisting of two injections, with the ignition timing lying between those pulses. With respect to the impact of M100 from a mechanical point of view, the fuel system was obviously a limiting factor. On the one hand, the flow rate of the high-pressure pump did not allow for increased low-end torque, while on the other hand the fuel injectors reached their injection duration limits when it came to the engine’s maximum power output. Hence, future engine tests will utilize an engine set-up that operates at a pressure level of 350 bar instead of 250 bar with an increased flow rate. Additionally, direct injection is complemented by a port fuel injection in order to further reduce injection times of the DI system and thus enable improved mixture formation. Besides this, spray targeting is aligned more centrally, which helps to avoid cylinder liner wetting. With regard to the engine start, further investigations will focus on decreased injection pressures as well as improved ignition timing during engine ramp-up in order to further reduce the engine’s cold-start time.

Bibliography 1. U. Kramer, F. Ortloff, S. Stollenwerk, R. Thee, Defossilisierung des Transportsektors - Optionen und Voraussetzungen in Deutschland, 2018. 2. S. Verhelst, J.W.G. Turner, L. Sileghem, J. Vancoillie, Methanol as a Fuel for Internal Combustion Engines, Progress in Energy and Combustion Science 70 (2019) 43–88. https://doi.org/10.1016/j.pecs.2018.10.001.

44

Methanol as an electricity-based fuel for plug-in hybrid electric vehicles 3. R.J. Pearson, J.W.G. Turner, Renewable Fuels, in: Comprehensive Renewable Energy, Elsevier, 2012, pp. 305–342. 4. Verein Deutscher Ingenieure, VDI-Gesellschaft Verfahrenstechnik und Chemieingenieurwesen, VDI-Wärmeatlas: Mit 320 Tabellen, 11th ed., Springer Vieweg, Berlin, 2013. 5. H. Menrad, A. König, Alkoholkraftstoffe, Springer, Vienna, 1982. 6. Methanol Institute, Methanol Gasoline Blends: Alternative Fuel for Today’s Automobiles and Cleaner Burning Octane for Today’s Oil Refinery. 7. R. van Basshuysen, F. Schäfer, Handbuch Verbrennungsmotor, Springer Fachmedien Wiesbaden, Wiesbaden, 2017. 8. H. Schnüpke, T. Maass, S. Zimmer, A. Rehberger, M. May, J.-C. Schmitt, T. Langer, M. Proust, O. Mohsen, P. Trochet, A. Arandyelovitch, Modern, Compact and Efficient: M 282 - The New 1.4-Liter Gasoline Engine from Mercedes-Benz, in: 26th Aachen Colloquium Automobile and Engine Technology 2017. 9. J. Warnatz, R.W. Dibble, U. Maas, Combustion: Physical and Chemical Fundamentals, Modeling and Simulation, Experiments, Pollutant Formation, 4th ed., SpringerVerlag Berlin Heidelberg, Berlin, Heidelberg, 2006. 10. E. Svensson, C. Li, S. Shamun, B. Johansson, M. Tuner, C. Perlman, H. Lehtiniemi, F. Mauss, Potential Levels of Soot, NO x HC and CO for Methanol Combustion, in: SAE Technical Paper Series, SAE International400 Commonwealth Drive, Warrendale, PA, United States, 2016. 11. S. Kaltenecker, Experimentelle Untersuchung des Kaltstartverhaltens von Methanol und Methanol-Mischkraftstoff an einem aufgeladenen Ottomotor mit Direkteinspritzung, 2019. 12. T.W. Ryan, T.J. Bond, R.D. Schieman, Understanding the Mechanism of Cylinder Bore and Ring Wear in Methanol Fueled SI Engines, in: SAE Technical Paper Series, SAE International400 Commonwealth Drive, Warrendale, PA, United States, 1986.

45

E-fuels as chemical energy carrier under the aspects of costs and efficiency Martin Rothbart, J. Rechberger AVL List GmbH

This manuscript is not available according to publishing restriction. Thank you for your understanding.

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_8

47

Potential analysis and virtual development of SI Engines operated with DMC+ Cornelius Wagner M. Sc., Dr.-Ing. Mahir-Tim Keskin, Dr.-Ing. Michael Grill, Prof. Dr.-Ing. Michael Bargende FKFS Dr.-Ing Liming Cai, Prof. Dr.-Ing. Heinz Pitsch ITV RWTH Aachen Sebastian Blochum M. Sc. LVK TU Munich

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_9

49

Potential analysis and virtual development of SI Engines operated with DMC+

1 Introduction In order to take measures for climate protection, CO2 emissions are to be reduced by 80-95% by 2050 in the European Union [1]. In order to keep the internal combustion engine competitive in terms of CO2 and pollutant emissions to Battery Electric Vehicles (BEV) or Fuel Cell Electric Vehicles (FCEV), alternatives to fossil fuels must be found. The choice of fuel plays a decisive role here. The main idea is to produce fuels synthetically with the help of renewable energies. This can lead to a CO2-neutral fuel. In [2] a fuel for gasoline engines was presented, which is characterized by special features. The following is an analysis of the potential of this synthetic fuel compared to gasoline (RON95 E10). As the analysis of the synthetic fuel on the test bench is expensive and very time-consuming, the first step is to assess the potential by means of simulation. In order to be able to make reliable statements about the behavior of this fuel via 0D/1D-Simulation, models must be adapted to the synthetic fuel. The simulations were performed with the FKFS UserCylinder® [3], [4], [5] and [6].

2 Fluid Properties of DMC+ DMC+ consists of 60 % Dimethyl carbonate (DMC), 35 % Methyl formate (MeFo) and 5 % ethanol (mass percent). DMC and MeFo are produced synthetically by carbonylation of methanol. Detailed properties of the single components are presented in [7]. According to [2] the addition of ethanol has a positive effect on the stability of combustion in engine operation. With regard to the blend of the three components this fuel with excellent properties in terms of particle emission and combustion stability is presented in [2]. It and is characterized by: – High knock resistance – Good evaporation properties – Resistant to cold – Resistant to hydrolysis – Mostly soot-free combustion – Sulphur free – Very low NOx raw emission – No catalyst poisons Figure 1 shows the structural formula of the individual components.

50

Potential analysis and virtual development of SI Engines operated with DMC+

Figure 1. Structural formula of the individual components of DMC+.

Since the properties of fuels have a strong effect on the process control in the combustion engine, the different adjustments are broken down as follows.

2.1 Characteristic Values of DMC+ The following table shows key values of DMC+ compared with RON95 E10. Table 1. Fuel key properties RON95 E10 vs. DMC+ [2]. Lower Heat Value (mass-based) Lower Heat Value (vol.-based) Density RON/MON Enthalpy of vaporization A/F ratio stoich.

RON95 E10 42.0 MJ/kg 31.4 MJ/l 748 kg/m³ >= 95/85 420 kJ/kg 14.3 kg/kg

DMC+ 15.8 MJ/kg 16.2 MJ/l 1024 kg/m³ > 110/100 459 kJ/kg 4.9 kg/kg

Due to LHV of DMC+, a larger amount of fuel mass must be supplied to the system to obtain a similar performance. As a result, a large quantity of fuel is available for evaporation. This has a direct effect on the intake air, which cools down considerably. However, the volumetric consumption is relatively lower due to the density which is 37% higher. Taking the improved efficiency into account (as demonstrated later), this means on balance that the fuel tank has to be roughly 1.5 times bigger than in a gasolinepowered vehicle to achieve the same cruising range. As can be shown the enthalpy of vaporization is ca. 9 % higher for DMC+ compared to RON95 E10, which further enhances the effect described above. In addition, the stoichiometric air consumption is lower with the same fuel energy, which means that less intake air has to be cooled. The significantly greater knock resistance is expressed by the RON, which is increased by more than 15.

51

Potential analysis and virtual development of SI Engines operated with DMC+

2.2 Ignition Delay DMC+ Accurate prediction of engine knocking is essential for reliable simulation results. Therefore, the ignition delay in the quasidimensional models must be calculated correctly. In order to facilitate the numerical simulation of ignition delay times and laminar flame speeds of DMC+, a chemical mechanism was developed as part of this study. The wellvalidated chemical mechanism [8] for a set of C0-C4 hydrocarbon species including ethanol served as the starting point of the model development. The specific-mechanisms of DMC and MeFo were extracted from the mechanisms of Dooley et al. [9] and Sun et al. [10], respectively, and integrated into the mechanism of Blanquart et al. [8]. This mechanism has been developed further and was validated successfully against the literature data of ignition delay times and laminar flame speeds of dimethyl carbonate, methyl formate and ethanol. The mechanism is called Cai Mechanism (see also [11]). Figure 2 shows the ignition delay times for RON 95 E10 and the investigated fuel DMC+ (preliminary results from reaction-kinetics calculations) at a stoichiometric air ratio and without EGR. DMC+, unlike RON95 E10, does not show a two-stage ignition (TSI)/ negative temperature coefficient (NTC) behavior. This has an extreme effect on the area of temperatures lower than 1.2 1000/K. The ignition delay times increase by over an order of magnitude below this value. Since the relevant part of the engine operation is between approx. 0.8 and 1.2 1000/K at pressures higher than 20 bar, the significantly lower knocking tendency can be seen directly.

Figure 2. Ignition delay times simulated with Cai Mechanism of DMC+ and RON95 E10.

52

Potential analysis and virtual development of SI Engines operated with DMC+ As the DMC+ results were only preliminary and a quick estimation of the knock behavior of DMC+ was to be achieved, an adaption of existing knock models was used at first (the implementation of a proper DMC+ knock model is intended for the upcoming months). It was decided to use the older Schmid model for DMC+ [12] rather than the newer and generally much better Fandakov model [13], [14] (which is used for the simulation of RON95 E10) due to the lack of TSI/NTC behavior, allowing to use an Arrhenius equation for the ignition delay times. Figure 3 shows the ignition delay times already shown above with the corresponding curve of an Arrhenius equations. While with RON95 E10 the representation of the ignition delay times via an Arrhenius approach is insufficiently accurate, this is much better suited for DMC+ due to the missing two-stage ignition behavior. In order not to underestimate the knocking tendency, the coefficients for the Arrhenius equation are assumed conservatively for DMC+. The 50 and 100 bar curves are always below the simulated values. This makes the modelled fuel more prone to knocking. The Schmid model is adjusted as follows: First, load points where knock occurs are simulated with the RON95 E10 engine model. Then the Fandakov model is adjusted against validation data of test-bench measurement. Afterwards the knock model is switched to Schmid model with a model adjustment until the knock prediction matches with the Fandakov model. Finally, Arrhenius coefficients and fuel properties are replaced with data of DMC+.

Figure 3. Arrhenius equation curves for the fuels RON95 E10 and DMC+.

53

Potential analysis and virtual development of SI Engines operated with DMC+

2.3 Laminar Flame Speed DMC+ The laminar flame speed S_L in dependence of temperature T and pressure p is determined using the Heywood equation [15] with parameters adapted for DMC+ in the 0D/1D-Simulation model. 𝑆 =𝑆

,

(1)

with: 𝛼 = 2.1 − 0.8(𝛷 − 1)

(2)

𝛽 = −0.05 + 0.22(𝛷 − 1)

(3)

𝑆

(4)

,

= 0.108 + 0.23(Φ − 1.06)

Figure 4 shows the agreement of the laminar flame speed for the Heywood equation with results from a reaction kinetic simulation for DMC and MeFo. For the Heywood equation one set of parameters is used for both species. These parameters depend on pressure and temperature. Only the two main components of DMC+ are considered. Ethanol generally has a higher laminar flame speed. This exclusion results in a conservative modelling as a slow flame speed has a bad effect on the efficiency of the engine.

Figure 4. Laminar flame speed represented with Heywood equation vs. reaction kinetic simulation.

54

Potential analysis and virtual development of SI Engines operated with DMC+ The simulation was performed for DMC with the model according to Sun et al. [10], while MeFo was performed with the model according to Dooley et al. [9]. In the upcoming months, the reaction-kinetics based approach via a Heywood correlation will be substituted with a spline-based approach similar to [16], [17] and [18], which will ensure even more accurate results. In the range of lean mixtures until rich mixtures to values of Φ ≈ 1.2 deviation of laminar flame speed is less than 15% for boundary conditions relevant for engine operation (50 bar 800 K, 100 bar 1000 K). In this paper only Φ = 1 is considered, with a deviation of less than 4 %. However, it is interesting to note that laminar flame speed for lean mixtures does not drop as steep as it happens for RON95 E10, making DMC+ also suitable for lean combustion concepts (see Figure 5), which will be investigated in ongoing work on the matter.

Figure 5. Laminar flame speed of DMC+ compared to RON95 E10 [16].

3 Engine models In the first step, the potential of DMC+ is determined with the help of an efficient state of the art gasoline engine, which is adapted for operating with the synthetic fuel. Then a concept is discussed that is optimized for the special properties of DMC+ to further highlight the potential of the fuel. The engine model presented below serves as the basic model.

55

Potential analysis and virtual development of SI Engines operated with DMC+

3.1 Basic Model VW EA211 TSI evo is a direct-injection gasoline engine characterized by a high efficiency over a wide range of the characteristic map (see Figure 6). Decisive factors are the variable nozzle turbocharger, Miller process, a high compression ratio and cylinder deactivation in part load [19]. This engine can be arguably seen as state-of-the-art for current gasoline engines in terms of efficiency. The specifications are shown in Table 2. Table 2. Specification of VW EA211 TSI evo [19]. Max. Power

96 kW@5500 RPM

Max. Torque Displacement volume Compression Ratio Stoichiometric Ratio EGR

200 Nm 1500 cc 12.5:1 1 internal

Results of RON95 E10 are simulated with higher quality models for knock behavior and laminar flame speed. Knock behavior is simulated with the aforementioned Fandakov model while the laminar flame speed is modeled with the spline approach of [16], [17] and [18]. Figure 6 shows the brake efficiency as a function of torque and engine speed for the engine operated with RON95 E10. [19] provides measurement data regarding the specific fuel consumption of the engine. In order to obtain accurate results, the simulation must be adapted with regard to models of the turbulence and combustion based on [5], autoignition based on [14], pressure loss, and friction based on [20]. The deviation of the simulation from the measurements in the relevant map area is less than 2 %. The engine achieves its peak efficiency of 38.4% at 3500 RPM and 120 Nm.

56

Potential analysis and virtual development of SI Engines operated with DMC+

Figure 6. Simulation of the characteristic map of VW EA211 TSI evo RON95 E10.

3.2 Basic Model with DMC+ In order to be able to generate results with the synthetic fuel, modifications described in section fluid properties must be done. The calorific values of the gaseous fuel is calculated from the ratio of C-, H- and O-Atoms with [21]. Also the values of fuel gas are calculated with methods described in [21]. Using the so-called equilibrium condition and the component approach, the fluid properties for different fuel gases can be precisely determinate. Due to the low heat value of DMC+ the parameters of the injector model must be adjusted. Injection delivery rate is increased by factor of three. This is the rate of fuel injection when the injector is held open. In practice, adaptions of the fuel injection system will be unavoidable, for instance by combining direct injection (DI) with port fuel injection (PFI) or by adapting the injector design. The effects of the combination with PFI and DI will be investigated at another research site with regard to an efficiency study is important here, as this can have negative influence on the volumetric efficiency. Figure 7 shows the change in brake efficiency between the RON95 E10 and DMC+ simulation calculated with Equation 5. 𝐺𝑟𝑜𝑤𝑡ℎ𝐵𝑟𝑎𝑘𝑒𝐸𝑓𝑓. =

, ,

∙ 100 − 100

(5)

57

Potential analysis and virtual development of SI Engines operated with DMC+

Figure 7. Growth in brake efficiency of VW EA211 TSI evo with DMC+.

The increased efficiency of DMC+ is clearly visible in the area of high load. Reasons for this are the better knock resistance, especially in the low-end torque area. At high engine speed and low loads efficiency decrease and the simulation with RON95 E10 becomes more efficient. Three load points marked with X in Figure 7 will be used to analyze the differences between an engine powered by RON95 E10 or DMC+. The X at top-right is a full load point at rated power (FL), bottom-left is a part load point (PL1) and bottom-right is a second part load point (PL2).

3.2.1 Analysis of specific load points Figure 8 shows aforementioned load points with their fractionation of fuel energy in the cylinder at a stationary engine operating point. All values are related to the supplied fuel energy for one cycle (100% energy).

58

Potential analysis and virtual development of SI Engines operated with DMC+

Figure 8. Analysis of three engine load points RON95 E10 vs. DMC+.

A comparison of the full load point (FL) shows that the energy content of the exhaust gas is significantly lower for DMC+. This is achieved through a better process control due to the higher knock resistance. The knock model predicts no knocking at this point for DMC+. Therefore, the engine can be operated with an optimum combustion center at 8 ° crank angle (CA) after FTDC compared to ca. 21 ° CA for RON95 E10. The heat loss is slightly increased by ca 1 % of the fuel energy because of the earlier combustion center of DMC+. The other energy losses are similar. This means that a large part of fuel energy can be converted into mechanical work for DMC+ which is equivalent to a higher brake efficiency. The combustion centers of the second load point (PL1) are similar for both fuels. Nevertheless, a slight efficiency advantage can be seen for DMC+. Due to the significantly higher fuel mass that has to be supplied to the system, a cooling effect of the combustion chamber through fuel evaporation can be seen. For comparison, the maximum cylinder temperature for DMC+ is ca. 2200 K whereas for RON95 E10 it is ca. 2400 K. As a result, the heat loss is reduced which leads to a slightly higher efficiency. At the third load point (PL2), DMC+ suffers a combustion loss. This can be attributed to the lower flame speed compared to RON95 E10. The burn duration for the conversion of fuel from 10 to 90 % lasts 35 ° CA for DMC+ while the duration for RON95 E10 is 24 ° CA. Especially at high engine speed and low loads, this effect is visible. The flame does not reach all combustion chamber walls until the end of cycle. This

59

Potential analysis and virtual development of SI Engines operated with DMC+ results in a 2% lower efficiency compared to RON95 E10 for the load point under consideration. This disadvantage is not significant at higher loads, as the burn duration of DMC+ becomes shorter with increasing load due to the optimum combustion center while the late ignition timing of RON95 E10 in the knock area makes the burn duration longer.

3.2.2 Investigation on the low LHV of DMC+ The effects of the lower stoichiometric air to fuel ratio for DMC+ is to be analyzed in this section. Therefore, an operation condition is discussed which has approximately the same air mass flow. From this consideration, important conclusions can be drawn regarding the amount of energy that can be supplied (Table 3) and the cooling properties of the fuel (Figure 9 and Figure 10). Table 3. Integral values of the different load points RON95 E10 vs. DMC+. RON95 E10

DMC+

Engine Speed Air Mass Torque Fuel Mass Fuel Energy Cylinder Mass

2500 RPM 392 mg 140 Nm 28 mg 1180 J 432 mg

389 mg 160 Nm 81 mg 1318 J 483 mg

MFB50

17

8

DMC+ achieves a higher torque with the same air volume. Due to the fluid properties, ca. 11 % more fuel energy can be supplied to the system. This is due to the oxygen content in the fuel mentioned above. The fuel carries a part of its oxidizing agent. Furthermore, the engine process can be performed with an earlier center of combustion for DMC+. Figure 9 shows the mean temperature (mass-weighted unburned and burned zone), pressure and burned fuel fraction curve over crank angle.

60

Potential analysis and virtual development of SI Engines operated with DMC+

Figure 9. Load Point with equal charge air RON95 E10 vs. DMC+.

Already 120 ° before FTDC the mixture is 50 K cooler in the combustion chamber for DMC+. During compression, the temperature difference increases – mainly as a consequence of the ideal gas law – to more than 100 K. The isentropic exponent is ca. 1.32 for the mixture with RON95 E10 and ca. 1.30 with DMC+ in the compression stroke at 50 ° CA before FTDC. The temperature then rises steeper for DMC+ due to the earlier MFB50. In the expansion stroke, the mixture cools down quicker again, which can be attributed to the earlier combustion center. At EVO, the mixture is ca. 150 K colder. Also, the peak temperature is lower for DMC+, reducing NOx raw emissions. The difference in temperature and pressure is illustrated in Figure 10. The initial cooling effect (-120 – -15° CA) and the more efficient process management (-15 – 150° CA) is very clearly visible here.

Figure 10. Difference of temperature (red) and pressure (blue) between DMC+ and RON95 E10.

61

Potential analysis and virtual development of SI Engines operated with DMC+ This leads to the fact that in the example shown, the low LHV is advantageous for the engine process.

3.2.3 Full load curve Due to the fact that the Schmid Model does not predict knocking for DMC+ in the above presented results, the theoretically possible full load curve shall now be investigated. For this purpose, the pressure limitation in the cylinder is omitted. However, the knock control remains active. This consideration makes little sense for RON95 E10 under the aspect of having an efficient engine. Due to the knock limitation, the ignition timing would have to be shifted so far back that reaching higher loads would not be possible much less an economic operation.

Figure 11. Full load curve RON95 E10 vs. DMC+.

Ignoring a possible peak pressure limit (simulated values can be up to 260 bar), rated power can be increased by a factor of 1.8 and the max. torque can be increased by a factor of 2.6 even. Basically, DMC+ allows the design of an almost knock-free SI engine with a liquid fuel, a very attractive combination. While the combustion center of RON95 E10 at 1500 RPM is already at 29 ° CA, for DMC+ it is only at 15 ° CA despite significantly higher torque. At the same time, these results obviously point to the conclusion that a DMC+ engine can be downsized heavily if rated power is to be held constant, which is investigated in the following section.

62

Potential analysis and virtual development of SI Engines operated with DMC+

3.2.4 Challenges to overcome Since the results were generated simulatively and the first experiments on the test bench have just started, the challenges that an engine fired by DMC+ must face will be discussed here. The low AFR holds challenges regarding spray preparation and homogenization. Wall impingement and fuel evaporation properties are phenomena that are difficult to predict in 1D flow simulation because they are highly dependent on 3D effects. A simple evaporation model is used as described in [22]. Wall impingement is ignored. During the project, these phenomena are examined in more detail with the help of 3D CFD simulation and then taken into account in subsequent engine simulations. As the results cannot yet be referred to in this paper, this is ignored. Downsizing increases the ratio of surface area to volume with the number of cylinders remaining constant. This increases the wall heat transfer. Also downsizing has an effect on the valve size which become smaller and thus the charge exchange is more difficult. Higher charge pressures are necessary. Due to the fuel properties of DMC+, less exhaust enthalpy is available for the turbocharger. Transient effects such as cold start behavior or load steps were also not considered and must be included in later investigations.

3.3 DMC+ adapted Engine In order to take advantage of the knock resistance of DMC+, the engine displacement and the compression ratio are adjusted. Since the available exhaust gas enthalpy changes with the process management and fluid properties, the turbocharger must be adapted, too.

3.3.1 Parameter selection To find the optimum relation between displacement volume and compression ratio, a variation of these parameters in a DOE study is carried out. For this, a partial load point and the operating point at rated power were examined. Only the operating point at rated power is shown below, since the results are similar for both load points. For comparison, the brake efficiency before optimization at rated power is 39.02 % (1500 cc and CR 12.5). The displacement volume was varied with a constant stroke/bore ratio. Due to the changing frictional behavior of the engine, approaches from [20] are selected which consider the change in compression ratio and displacement volume.

63

Potential analysis and virtual development of SI Engines operated with DMC+

Figure 12. Brake efficiency at rated power as a function of displacement volume and compression rate for DMC+.

It is noticeable that there are no values in the upper left area. This is due to geometry limitations when the compression volume gets too small and basically leaves no space for the spark plug if a flat piston is maintained. As expected for the Otto cycle, the efficiency increases with increasing compression rate. As the displacement decreases, the efficiency improves due to reduced friction losses until surface/volume ratio of the burn chamber becomes too disadvantageous. The combination of 900 cm3 displacement volume and a compression ratio of 19:1 was finally chosen for the downsizing engine with a brake efficiency of 42.6%. This high compression ratio can be achieved because of three reasons: ● High knock resistance of the fuel ● Enormous cooling effect due to the high amount of fuel mass ● Miller-Cycle is realized at high loads This extreme downsizing has effects in reality that are difficult to represent in the simulation. Since the spark plug and the combustion chamber geometry are not modeled discretely influences of downsizing on the efficiency of combustion are difficult to predict. Also the ignition process, which becomes more complicated as the displacement volume decreases, is not simulated here. However, these issues will be taken into account in later studies. It is assumed, that this does not greatly affect the validity of the calculations.

64

Potential analysis and virtual development of SI Engines operated with DMC+

3.3.2 Characteristic map Figure 13 shows the characteristic map of the downsized engine with DMC+. It is noticeable that the only area where the efficiency is worse is in the area of low end torque (grey area), which can be explained by a number of design choices that are not compulsory. ● DMC+ reduced knock considerably, but using efficiency optimal combustion centers deprives the turbine of a great amount of exhaust enthalpy urgently needed at low engine speeds. This is further enhanced by the lower exhaust temperatures of DMC+. ● It was thus decided to voluntarily use later MFB50 timings, sacrificing efficiency to gain torque. Alternatively, a different turbocharger sizing or an adaption of the exhaust path with the aim of reduced wall heat losses could be used. The problem would also be decisively mitigated when using an "eTurbo", i.e. installing an electric machine on the turbocharger shaft. As the valve timing has not been optimized, a miller cycle is implemented in this area. The valve timing has to be optimized. For the remaining operating range, a considerable improvement in brake efficiency up to 20% can be observed.

Figure 13. Growth in effective efficiency of downsized VW Evo EA211 with DMC+.

The potential of DMC+ is to be clarified again in Figure 14 by means of three load points. A moderate partial load point in blue, the point with highest brake efficiency of RON95 E10 in green and the full load point in yellow.

65

Potential analysis and virtual development of SI Engines operated with DMC+ Even for the best point of RON95 E10 there has been a slight improvement with the switch to DMC+. With the downsizing engine, the efficiency increases by more than 10 %. The highest potential for increasing brake efficiency is offered by the full load point due to the fuels knock resistance.

Figure 14. Comparison of DMC+ engine concepts based on three load points.

4 RDE-Simulation The performance of the engine concepts is now being investigated using a virtual Real Drive Emission (RDE) Profile. The results are based on a longitudinal model of a middle class vehicle (E-Segment). The engine is modelled based on the characteristic maps from previous calculations. The given virtual route complies with the regulations in [23]. It is ca. 62 km long. Figure 15 shows the performance of the basic engine concepts with RON95 E10 and DMC+ and the DMC+ adapted Engine on an RDE-Profile. In the lower area the velocity of the vehicle can be seen, which is given as default value for all engine models in the simulation. In the middle area the required fuel energy for the length of the RDE is shown. While above, the total NOx-raw emissions for the RDE are presented.

66

Potential analysis and virtual development of SI Engines operated with DMC+ 210 Cumulative NOx [g]

180 RON95 E10 DMC+ Basic Engine DMC+ Downsizing Concept

150 120 90 60 30 0

108 90 72 54 36 18

Vehicle Velocity [km/h]

140

Cumulative Fuel Energy [MJ]

126

0

120 100 80 60 40 20 0 0

1000

2000

3000 Time [s]

4000

5000

6000

Figure 15. Comparison of Engine Concepts based on an RDE-Profile.

The DMC+ Basic Engine reduces NOx-raw emissions by 40 %. With the adapted engine, emissions can be reduced even by 60 %. Due to the strong combustion chamber cooling of DMC+ less NOx-raw is produced. A similar behavior can also be seen for cumulative fuel energy. While the energy consumption is comparable for the basic engine with DMC+ and RON95 E10, the amount of energy required for the adapted engine is 10 % lower. Table 4. Average brake efficiency over RDE-Profile. RON95 E10 Brake efficiency

32.4 %

DMC+ Basic Eng. 32.9 %

DMC+ Adapted 36.0 %

67

Potential analysis and virtual development of SI Engines operated with DMC+ As can be seen in Table 4, the increase in efficiency is significant for the DMC+ adapted engine in an RDE-Profile.

5 Validation For the following validation of the presented models, one-cylinder test bench data is used. A load variation between 3 and 11 bar indicated mean pressure (IMEP) was measured at a distance of 2 bar. In parallel the engine speed between 1000 and 3000 RPM in steps of 500 RPM was measured in each case for the two fuels. The fuel RON95 E5 was used for the measurement. Therefore, in the following, DMC+ is compared with RON95 E5.

5.1 Investigation of the measurement In order to first assess the effects of the fuels on the internal engine operation the gross indicated efficiency of the measurement is investigated. Since friction is not relevant for the forthcoming pressure trace analysis (PTA), brake efficiency is not considered. Eng. Speed: 2000 RPM

38.0

IMEP: 7 bar

RON95 E5 DMC+

37.5

ig [%]

37.0 36.5 36.0 35.5 35.0 34.5 1.0

boost pressure [bar]

0.9 0.8 0.7 0.6 0.5 0.4 0.3 3

4

5 6 7 8 9 10 Indicated Mean Pressure [bar]

11

12

1000

1500 2000 2500 Engine Speed [RPM]

3000

Figure 16. Test bench measurement of gross indicated efficiency and boost pressure of RON95 E5 and DMC+.

68

Potential analysis and virtual development of SI Engines operated with DMC+ The measurement in Figure 16 shows that DMC+ has a higher gross indicated efficiency ηig than RON95 E5. The only exception is load point 7 bar IMEP and 2000 RPM. It is assumed that this is an exception, as MFB50 is constant. It can also be seen that the pressure in the inlet channel is lower than ambient pressure due to the throttle valve for both fuels. Only measurements with throttle valve angle < 90 ° were available for the work. DMC+ must be throttled more strongly due to the lower air demand which is why the inlet channel pressure is lower. Since the charge exchange is not taken into account for the gross indicated efficiency (only high pressure) this may result in a lower indicated efficiency (high pressure + low pressure).

5.2 Measurement vs. Simulation of the test bench engine A PTA is used to determine the combustion process of the measurement. For this purpose, a model is created that corresponds to the combustion chamber of the one-cylinder test bench engine. The pressure curve of the measurement is used to calculate the heat release rate. In the next step the combustion process is simulated with the models from the previous chapters. The main focus is on the combustion model with the laminar flame speed and the turbulence model. No knocking occurs at these operating points. Therefore, the knock model cannot be validated. Figure 17 shows the combustion process of the load point 2000 RPM and 7 bar IMEP for both fuels RON95 E5 and DMC+. Furthermore, the comparison between PTA and simulation is shown. It can be seen that the heat release rate with DMC+ is, as suspected, lower. This is due to the lower laminar flame speed. Therefore, the burning duration is longer. This change between RON95 E5 and DMC+ is correctly predicted by the combustion model (without adjustments), which means that the change in laminar flame speed is calculated accurately. Since the aim here is only to assess whether the trends and tendencies of the models correspond, this is sufficient. Therefore, the model is well suited to represent the different fuels.

69

Potential analysis and virtual development of SI Engines operated with DMC+ 45 RON95 E5 Sim. RON95 E5 Meas. DMC+ Sim. DMC+ Meas.

40 35

dQ_B [J/°KW]

30 25 20 15 10 5 0 -40

-20

FTDC 20 40 Crank Angle [°]

60

80

Figure 17. Heat release rate for RON95 E5 and DMC+ at 2000 RPM and 7 bar IMEP.

With the aforementioned findings, the predicted quality of the gross indicated efficiency between simulation and measurement is assessed. Figure 18 shows the difference of the simulated gross indicated efficiency and the gross indicated efficiency of the PTA. Efficiency is overestimated for both fuels in the simulation. In average the overestimation of the gross indicated efficiency is about 1 % for both fuels. On this basis, the variation is minor. This means that load or speed variation can be represented very well. At low loads the difference increases for RON95 E5. With this analysis it can be assumed that the quality of the simulation results in the presented paper is reliable.

70

Potential analysis and virtual development of SI Engines operated with DMC+ RON95 E5 12 Indicated Mean Pressure [bar]

11 10 9

 ig [%]

1.5

8 7

0

6

1.8

0.25

5 0.5

4 0.75

3 DMC+

1

Indicated Mean Pressure [bar]

12 11

1.25

10

1.5

9

1.75 1.0

8

0.8

0.5 2

7 6 5 4 3 1000

1500 2000 2500 Engine Speed [RPM]

3000

Figure 18. Difference of gross indicated efficiency Simulation vs. Measurement for RON95 E5 and DMC+.

6 Summary/Conclusions The engine of the future must be highly efficient and emit as few pollutants as possible. Furthermore, the fuel should be CO2 neutral. DMC+ is predestined as a synthetic fuel to meet these demands assuming it is produced with renewable energy sources. The most important properties of DMC+ are the high knock resistance and the large cooling potential. The brake efficiency already improves when simply exchanging the used fuel in the simulation model. In order to exploit the full potential of DMC+, the engine design has to be adapted. In this way, as has been shown, downsizing and higher compression ratios can be used to improve brake efficiency even further, leading to more than 20% higher efficiencies compared to a state-of-the-art gasoline engine. This results to an improvement of brake efficiency in an RDE-Profile of ca. 11 %.

71

Potential analysis and virtual development of SI Engines operated with DMC+ In order to generate even more reliable results, the models must be adapted to accurately predict the behavior of DMC+ in comparison with measurements on the test bench. The main focus is on the validation of the knock model as there are no measurements for knocking behavior yet. This work will be done in the upcoming months and will presumably provide even more impressive results, as so far all model adjustments were done conservatively. As it is an engine concept of the future, in the next step, a similar procedure will be carried out with a high performance engine, featuring a high turbulence concept and an efficiency-improved turbocharger, among others. Additionally, the potential of lean combustion will be investigated. All in all, peak gross indicated efficiency values higher than 50% should be easily feasible in such a concept.

7 Bibliography [1]

European Union, “A Clean Planet for all; A European strategic long-term vision for a prosperous, modern, competitive and climate neutral economy,” COM(2018) 773 final, 2018.

[2]

Härtl, M., Stadler, A., Blochum, S., Pélerin, D, “DMC+ as particulate free and potentially sustainable fuel for DI SI engines,” Zukünftige Kraftstoffe, 2019, doi:10.1007/978-3-662-58006-6.

[3]

Grill, M. and Bargende, M., “The cylinder module,” MTZ Worldw 70(10):60–66, 2009, doi:10.1007/BF03227984.

[4]

Grill, M., Schmid, A., Chiodi, M., Berner, H.-J. et al., “Calculating the Properties of User-Defined Working Fluids for Real Working-Process Simulations,” SAE Technical Paper 2007-01-0936, 2007, doi:10.4271/2007-01-0936.

[5]

Grill, M., Billinger, T., and Bargende, M., “Quasi-Dimensional Modeling of Spark Ignition Engine Combustion with Variable Valve Train,” SAE Technical Paper 2006-01-1107, 2006, doi:10.4271/2006-01-1107.

[6]

Grill, M. and Bargende, M., “The Development of an Highly Modular Designed Zero-Dimensional Engine Process Calculation Code,” SAE Int. J. Engines 3(1):1–11, 2010, doi:10.4271/2010-01-0149.

[7]

Maier, T., Härtl, M., Jacob, E., and Wachtmeister, G., “Dimethyl carbonate (DMC) and Methyl Formate (MeFo): Emission characteristics of novel, clean and potentially CO2 -neutral fuels including PMP and sub-23 nm nanoparticle-emission characteristics on a spark-ignition DI-engine,” Fuel 256:115925, 2019, doi:10.1016/j.fuel.2019.115925.

72

Potential analysis and virtual development of SI Engines operated with DMC+ [8]

Blanquart, G., Pepiot-Desjardins, P., and Pitsch, H., “Chemical mechanism for high temperature combustion of engine relevant fuels with emphasis on soot precursors,” Combustion and Flame 156(3):588–607, 2009, doi:10.1016/j.combustflame.2008.12.007.

[9]

Dooley, S., Burke, M.P., Chaos, M., Stein, Y. et al., “Methyl formate oxidation: Speciation data, laminar burning velocities, ignition delay times, and a validated chemical kinetic model,” Int. J. Chem. Kinet. 42(9):527–549, 2010, doi: 10.1002/kin.20512.

[10] Sun, H., Yang, S.I., Jomaas, G., and Law, C.K., “High-pressure laminar flame speeds and kinetic modeling of carbon monoxide/hydrogen combustion,” Proceedings of the Combustion Institute 31(1):439–446, 2007, doi:10.1016/ j.proci.2006.07.193. [11] Cai, L., Ramalingam, A., Minwegen, H., Alexander Heufer, K. et al., “Impact of exhaust gas recirculation on ignition delay times of gasoline fuel: An experimental and modeling study,” Proceedings of the Combustion Institute 37(1):639–647, 2019, doi:10.1016/j.proci.2018.05.032. [12] Schmid, A., Grill, M., Berner, H., Bargende, M, “Ein neuer Ansatz zur Vorhersage des ottomotorischen Klopfens,” 3. IAV Tagung: Ottomotorisches Klopfen, Tagungsband, Berlin, 2010. [13] Fandakov, A., Grill, M., Bargende, M., and Kulzer, A.C., “Two-Stage Ignition Occurrence in the End Gas and Modeling Its Influence on Engine Knock,” SAE Int. J. Engines 10(4):2109–2128, 2017, doi:10.4271/2017-24-0001. [14] Fandakov, A., Grill, M., Bargende, M., and Kulzer, A.C., “A Two-Stage Knock Model for the Development of Future SI Engine Concepts,” SAE Technical Paper 2018-01-0855, 2018, doi:10.4271/2018-01-0855. [15] Heywood, J.B., “Internal combustion engine fundamentals,” Mechanical engineering, ISBN 978-1260116106, 2018. [16] Hann, S., Grill, M., and Bargende, M., “Reaction Kinetics Calculations and Modeling of the Laminar Flame Speeds of Gasoline Fuels,” SAE Technical Paper 2018-01-0857, 2018, doi:10.4271/2018-01-0857. [17] Hann, S., Urban, L., Grill, M., and Bargende, M., “Prediction of burn rate, knocking and cycle-to-cycle variations of binary compressed natural gas substitutes in consideration of reaction kinetics influences,” International Journal of Engine Research 19(1):21–32, 2018, doi:10.1177/1468087417732883.

73

Potential analysis and virtual development of SI Engines operated with DMC+ [18] Hann, S., Grill, M., Bargende, M., “A Quasi-Dimensional SI Combustion Model Predicting the Effects of Changing Fuel, Air-Fuel-Ratio, EGR and Water Injection,” in: SAE Technical Paper 2020-01-0574, 2020. [19] Eichler, F., Demmelbauer-Ebner, W., Theobald, J., Stiebels, B., et al., “Der neue EA211 TSI®evo von Volkswagen,” (37. International Vienna Motor Symposium), 2016. [20] Huß, M., “Übertragung von Motoreigenschaften mit Hilfe charakteristischer Skalierfunktionen zur Simulation verschiedener Varianten von Ottomotoren,” München, Technische Universität München, Diss., 2013, Universitätsbibliothek der TU München, München, 2013. [21] Grill, M., Chiodi, M., Berner, H.-J., and Bargende, M., “Calculating the thermodynamic properties of burnt gas and vapor fuel for user-defined fuels,” MTZ Worldw 68(5):30–35, 2007, doi:10.1007/BF03226830. [22] Grill, M., “Objektorientierte Prozessrechnung von Verbrennungsmotoren,” 2006, doi:10.18419/opus-4076. [23] THE EUROPEAN COMMISSION, “COMMISSION REGULATION (EU) 2016/427 of 10 March 2016 amending Regulation (EC) No 692/2008 as regards emissions from light passenger and commercial vehicles (Euro 6),” 2016.

Acknowledgment The research project is financed by BMBF (Federal Ministry of Education and Research), due to a decision of the German Bundestag. The authors would like to thank the BMBF for providing financing. Supported by:

On the basis of a decision by the German Bundestag Also the authors would like to thank Martin Härtl from LVK at TUM for exchange of project progress.

74

Supporting the synthesis of electric drive systems with scenario management Adrian Braumandl, Florian Marthaler, Katharina Bause, Felix Ranly IPEK – Institute of Product Engineering at Karlsruhe Institute of Technology (KIT) Kaiserstraße 10, 76131 Karlsruhe, Phone: +49 721 608 46731 Mail: {adrian.braumandl, florian.marthaler, katharina.bause}@kit.edu

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_10

75

Supporting the synthesis of electric drive systems with scenario management

1 Motivation The vehicle market in Germany is currently influenced by many different factors. Stricter emission regulations, exhaust gas scandals, high traffic densities and a large number of available models with different drive system topologies are just some of the characteristics. In order to successfully establish vehicles on the market, customer needs and conditions of use must be taken into account during their development. Aspects such as operational range and sustainability of the vehicles as well as legal questions must also be considered. In this work, these aspects will be summarized and put into context.

2 State of the art To be competitive, companies need to identify customer demands as early as possible to be able to satisfy them with corresponding products. [1] Since the potential to influence the development process regarding costs and efficiency is the highest during the early stages of product development [2,3], it is necessary to also identify the relevant product attributes as early as possible. Thus, product developers have to be aware not only of demands for products currently in development but also of demands for future products. [4] Because of this companies require information regarding these future demands, future customer needs and future market boundaries. Various tools are available in the field of future management: Forecasts provide short-term and quantified statements about the future, from which operational recommendations for action can be derived, but reliability reduces with increasing time horizon. [5] Trends provide mediumterm statements about the future and mainly consider individual developments in isolation. In case of complex, highly connected and highly interdependent economic systems the research of trends is not sufficient since the development of markets is rather dynamic. [6,7] As it is not possible to foresee the future, scenarios, which describe different possible futures, are developed. [8] Scenarios provide long-term, internally consistent fture worlds with the aim of producing as complete an overall picture as possible of the relevant section of the future. [9] In the area of product engineering, mainly the procedure according to Gausemeier and Plass is used: During scenario preparation, the scenario field is defined and narrowed down for the respective application. In the subsequent scenario field analysis, influencing factors are identified and their interactions are checked in a network analysis. The results are the relevant key factors for modeling the future product environment. In scenario forecasting, possible developments of the individual key factors are determined, thus forming the projection portfolios. Scenario building is used to form consistent projection bundles and thus to build the scenarios. Finally, the scenarios formed are interpreted during scenario transfer. Scenario management, the actual usage of scenarios in strategic leadership, is more than the actual

76

Supporting the synthesis of electric drive systems with scenario management scenario building process. The essential goal is recognizing potentials and risks to accordingly support strategic decision making. Use cases are leadership, technology management and product planning, especially when support is necessary in complex decision-making processes.[10] Besides identifying strategic potentials in product development, scenario analysis is also particularly useful in synthesizing the strategic orientation of research. [27,28]

3 Procedure The procedure used in this work corresponds to the classic scenario building process. It is divided into five steps which are explained below.

3.1 Scenario preparation In this part of scenario creation, the topic area is defined. Clusters of influence are defined and explained. Concrete clusters of influence that are considered in this work are, among others, politics, energy supply, raw materials, manufacturers, customers/users and the charging infrastructure. Politics set decisive framework conditions through its legislation, such as the setting of emission limits [11,12] and environmental zones [13] as well as various support programs for industry and private. [14] Due to the demand of electric vehicles for various raw materials such as lithium, cobalt and platinum, influencing factors such as the required amount of raw materials must be considered. The manufacturers' sphere of influence is manifold. New components such as electric motors and batteries reduce the influence of manufacturers in the upstream area of vehicle production. [15,16,17] In addition, electric vehicles contain fewer components. This reduces the number of potential maintenance operations. [16] As a result, the share of after-sales in the downstream area might decrease. In addition to these changes, manufacturers must, among other things, rethink the design of vehicles depending on the functions they are to perform. [15,18] The cluster customers and users is representing their interests and priorities during the vehicle purchasing process as well as the vehicle usage. The cluster charging infrastructure includes influencing factors such as the choice of charging locations, the number of charging devices and the standardisation of charging plugs.

3.2 Scenario field analysis In this part, the clusters of influence described above are examined. Influence factors from the clusters of influence are determined (see Table 1).

77

Supporting the synthesis of electric drive systems with scenario management Table 1: Influence factors (excerpt) Cluster Raw materials

Influence factor Quantity of raw materials Availability of raw materials Environmental damage caused by raw material extraction

In the next step, the defined influence factors are evaluated by a networking and relevance analysis. The sequence of the analyses is briefly explained below. ● Networking analysis The network analysis consists of two parts, the direct impact analysis (DEA) and the indirect impact analysis (IEA). It can be determined, how strongly an influence factor influences other influence factors („active factors“) and how strongly influence factors are influenced by other influence factors („passive factors“). ● Relevance analysis The influence factors are relatively compared to each other regarding their relevance resulting in a ranking considering all comparisons made. Key factors are now being determined by identifying active factors with high relevance in their respective cluster of influence.

3.3 Development of projections Projections are used to define changeable properties of key factors. The projections are considered separately for each key factor. When creating the projections, no attention is paid to whether a projection contradicts the projection of another key factor. The creation of a projection will be illustrated in the following example. The considered key factor is the „Quantity of raw materials“. Possible projections are as follows: ● New occurrences ● New deposits that can be mined are discovered and the extraction processes become more efficient. Thus, larger quantities can be extracted from the deposits. [19]

78

Supporting the synthesis of electric drive systems with scenario management ● Drying up of resources ● Environmental incidents (earthquakes, extreme heat, pollution, political decisions) lead to a drying up of raw material sources. As a result, smaller quantities of a resource are available. [20] ● Increase in consumption ● Raw materials are becoming scarcer due to an increased consumption. As a result, they and the associated products are becoming more expensive. This ultimately affects the user, who has to deal with higher prices. [21] ● New raw materials ● Technological progress is bringing new raw materials into focus, which can be obtained in a more environmentally friendly way (e.g. NMA batteries). A broader range of technology and products is emerging, which can be selected according to the application. [22]

3.4 Scenario building There are multiple ways to conduct the scenario building phase. While the consistency matrix-based approach has been established [10] it is rather time consuming and usually needs software support. [23] MARTHALER, ALBERS ET AL. analysed the feasibility of a catalogue-based approach towards the scenario building phase by comparing it to the consistency matrix-based method. This catalogue-based approach transfers the projections of the key factors into a morphological box. Then, one initiator factor per cluster is chosen. Consistent projections of the initiator factors are linked to each other to create initial scenarios. Consistent projections of the leftover key factors are then assigned to these initial scenarios. Only one projection per key factor may be used in a scenario. [24] Depending on the size and relevance of multiple initiator factors can be selected per influencing field. Besides removing the software requirement and reducing time demand, the traceability of the scenario building process is improved by using the catalogue absed approach. An excerpt of the scenario table is shown below (Table 2). “Availability of raw materials” was selected as initiator factor in the cluster “Raw materials” to show the general procedure with a concrete yet simple example.

79

Supporting the synthesis of electric drive systems with scenario management Table 2: Scenario catalogue (excerpt) Initiator factor

X

Cluster

Raw materials

Politics

Influence factor Availability of raw materials Support for charging infrastructure

Projection 1A 1B 1C 2A 2B

Description Politic stability in producing countries (Civil) Warfare in producing countries Diversification of raw materials Additional funds

Scenario A B X

X X

Cancellation of funds

X

3.4.1 Scenario A – “Electromobility: The panacea” The subsidies for research and development will remain constant in the future as electric vehicle technology reached a competitive level. In order to further spread electric mobility and remove obstacles, additional funds will be approved for the expansion of charging infrastructure. In order to facilitate the expansion, the residential property law will be relaxed. Technological progress makes it possible to set up a smart grid. The costs of electricity generation and infrastructure expansion are largely borne by government subsidies. Raw material extraction processes become more efficient, so that larger quantities can be obtained and sold. Thanks to constant economic growth in raw material producing countries their politic situation is stable and raw materials are always available. Due to the increased amount of available raw materials, manufacturers are able to bring inexpensive vehicles onto the market. These are affordable for a large part of the population. With the growing experience of the manufacturers, an increasing share of electric vehicle models is designed to suit electric drive systems instead of converting conventional platforms. These vehicles convince the users with their performance, which further increases the use of the vehicles. As the number of vehicles increases, so does the charging infrastructure demand. Legal requirements thus specify a necessary amount of charging stations on public and company parking lots.

3.4.2 Scenario B – “Mobility diversity: The challenge of individual mobility” The subsidies for research and development will remain constant in the future as electric vehicle technology reached a competitive level. An approach that is open to different

80

Supporting the synthesis of electric drive systems with scenario management technologies is followed by legislation, leaving the customers the freedom of choice. Due to the technical capabilities of electric vehicles, additional funding for infrastructure development is not necessary. Since the expansion of the infrastructure is not being additionally promoted, electricity tariffs are at a higher price level. Demand for raw materials used in batteries increases since an increasing amount of products are equipped with batteries. This leads to a faster shortage of raw materials. As a result, prices for the associated technologies increase. To counteract, manufactures researching technological solutions using different raw materials. The number of competing technological solutions on the automotive market is thus growing. There are more alternatives for the users to chose regarding their mobility and an option suiting their needs can be chosen. Besides battery electric vehicles other alternatives are replacing conventional vehicles.

3.5 Scenario transfer To analyse the impact of the changes described in the scenarios compared to today’s world, product projections are created in addition of the scenarios.

3.5.1 Vehicle projections To give an articulate example how scenario management can benefit the developer, simplified projections of electric vehicles based on currently available vehicle models with different characteristics are considered in this chapter. Lithium-Ion battery electric vehicles (BEVs), fuel cell electric vehicles (FCEVs) and plug-in hybrid electric vehicles (PHEVs) as well as electric sports cars are analysed regarding their suitability to the created scenarios.

3.5.1.1 Large BEV A large passenger car equipped with a large lithium-ion battery to deliver high operational ranges. The large BEV is comparatively heavy and uses a lot of raw materials to build. Furthermore, it takes rather long to recharge. It is not tailored to a specific use case but the attempt to create a one-fits-all battery electric car capable of replacing cars with internal combustion engines. Requires high power charging infrastructure.

3.5.1.2 Large FCEV A large passenger car equipped with a fuel cell and a high-pressure hydrogen storage system that is capable of delivering very high operational ranges and being refueled as fast as conventional vehicles. While it is not tailored to a specific use case it excels at long distance journeys, only requiring a basic network of hydrogen stations.

81

Supporting the synthesis of electric drive systems with scenario management

3.5.1.3 Large PHEV A large passenger car equipped with an internal combustion engine, an electric machine besides a fuel tank and a battery capable of providing energy to complete shorter trips in electric mode. PHEVs are expensive due to the integration of a powerful electric machine and a corresponding battery in the conventional drivetrain. PHEVs are not tailored to a specific use case, however they come with the advantages of both conventional as well as electric vehicle regarding their driving performance and capability of locally emission-free mobility.

3.5.1.4 Small BEV A small passenger car equipped with a small lithium-ion battery suited for urban use, where a particularly high range is not necessary. Raw material requirement is low, reducing the costs and increasing customer acceptance. Requires charging infrastructure in urban areas.

3.5.1.5 Small FCEV A small passenger car equipped with a fuel cell and a high-pressure hydrogen storage system. The small FCEV is rather expensive and usable space is limited. Furthermore, while the vehicle is capable of providing decent operational range, driving comfort is subpar on long trips. Either a sophisticated hydrogen network or possibilities for decentralized hydrogen production are required to suit the frequent need of refueling.

3.5.1.6 Small PHEV A small passenger car equipped with a small internal combustion engine and an electric machine as well as a fuel tank and a lithium-ion battery capable of providing energy for short trips in electric mode. Intergration of electric machine and battery into a conventional drivetrain lead to high costs and limited space. The need for raw materials for the batteries and pollutants emitted on longer trips lead to a low sustainability rating.

3.5.1.7 Electric sports car Electric sports cars are focused around driving performance. Due to the torque characteristics of electric machines they can outmatch conventional sports cars when it comes to acceleration from lower speeds or standstill. Despite being fitted with large batteries capable of providing high power the operational range is not exceptional. An overview over some of the attributes of the projected vehicles is given in Table 3.

82

Supporting the synthesis of electric drive systems with scenario management Table 3: Characteristics of vehicle projections (excerpt). Legend: ++ very good, + good, O average, - bad, -- very bad

Operational Range Costs Sustainability Charging / refueling duration Efficiency

Large BEV O -

Large FCEV + + ++

Large PHEV ++ O -++

Small BEV ++ + O

Electric sports car --

+

O

-

++

O

4 Results The vehicle projections described above will now be compared with the individual scenarios. For this purpose, the scenarios are examined with regard to their technological characteristics and statements about the automobile market. Table 4 provides an overview. Table 4: Evaluation of scenarios with regard to technical characteristics of the environment or the required technical characteristics of the vehicles (excerpt). Legend: ++ very positive, + positive, O neutral, - negative, -- very negative Costs Sustainability Application-specific

Scenario A ++ O --

Scenario B + ++

The following is an example of how the evaluation of the properties is carried out. First of all, "costs" are considered. In scenario A, electric mobility receives ongoing subsidies. In addition, new raw material deposits are tapped. This combination makes it possible to keep manufacturers' costs for the production of vehicles low. For these reasons, the “costs” item in scenario A was rated "very positive". In scenario B, manufacturers have to struggle with many competing products. Electric mobility is not particularly promoted. In addition, the demand for raw materials is very high, so prices rise. At the same time, customers are less interested in owning a vehicle, because they are no longer dependent on it due to other mobility offers. This combination increases production costs for manufacturers. Therefore, the “costs” item in scenario B was rated “negative".

83

Supporting the synthesis of electric drive systems with scenario management Because of the competition between different mobility offers in scenario B, the application level of electric vehicles is important in order to prevail on the market against other products, hence the rating "very positive". In scenario A, on the other hand, there are no significant competing products for e-mobility. Therefore, it is not particularly important that the vehicles serve a specific purpose. They are capable of asserting themselves on the market independently of special niches. In the next step, the vehicle projections are now to be crossed with the characteristics of the scenarios and compared. The result is shown in Table 5. Table 5: Consistency check of vehicle projections and scenarios (excerpt). Legend: ++ very positive, + positive, O neutral, - negative, -- very negative Large BEV Large FCEV Large PHEV Small BEV Small FCEV Small PHEV

Scenario A + O O ++ --

Scenario B O ++ O ++ O

● Scenario A The highest consistency with this scenario is shown by the small BEV, followed by large BEV concepts. There is a high level of consistency in terms of costs and price. The small BEV prevails over large concepts in terms of design, resulting in the higher consistency of this vehicle projection. The small PHEV shows the strongest inconsistency. Especially because of its low sustainability, efficiency and high costs and prices, this vehicle doesn’t fit into this scenario. ● Scenario B Since a multitude of mobility offers is available, the application-specific design of the vehicles is of great importance. The small BEV is the best fit to this scenario. With its sustainability, efficiency and clear focus on urban applications, it is highly consistent with scenario B. Large FCEVs also fit into the scenario. They cover long-range applications. The installation of gearboxes in purely electric vehicles can be considered to increase efficiency in a wider range of application. [25] This year the German government adopted the so-called climate package. [26] It contains measures further promoting electric mobility in Germany. Finally, the scenarios are now to be compared with measures of the climate package. One measure of the

84

Supporting the synthesis of electric drive systems with scenario management climate package envisages having one million charging points in Germany by 2030. In addition, a charging point obligation is to be introduced for gas stations. This measure corresponds to scenario A and contradicts scenario B. In addition, investments are to be made in the German power grid. The grid operators are to be supported in investing in the intelligence and controllability of the power supply system. This measure also supports scenarios A and B, as they include the development of a Smart Grid. In addition, existing measures such as BAFA subsidies and tax advantages for electric vehicle owners are to be expanded, as described in scenario A. In addition to measures promoting electromobility, measures are also planned to promote products competing with private transport. Several financial measures are planned to modernise rail transport. Bus fleets are to be converted to electric, hydrogen-based or alternative drive concepts. This is rather consistent with scenario B, which follows an approach that is open to different technologies. Summing up, the probability of both presented scenarios is on a comparable level as p.e. most legislative measures taken in the meantime are consistent.

5 Conclusion In this paper, clusters of influence regarding electromobility have been defined. Subsequently, influence factors from the clusters of influence were determined and evaluated with regard to various variables. With the help of this evaluation it was possible to reduce the number of influence factors to essential key factors. These were used to create future projections, which were then used to create a catalogue of scenarios in the next step. As a result, two scenarios were presented to show different possibilities of how electric mobility could develop in Germany. In addition to the scenarios, different vehicle projections were created. By comparing the vehicle projections with the scenarios, the vehicle projections could be analysed with regards to their consistency for different scenarios. Finally, the image of the vehicle projections was refined by comparing different current political measures and highlighting specific characteristics. The results of this work confirm the feasibility of using the catalogue-based approach analysed by MARTHALER, ALBERS ET AL. to generate scenarios and using the gained knowledge to synthesise consistent technological systems that fulfill future needs.

6 Acknowledgements This paper presents exerpts of work in the project „Methoden zur arbeitsteiligen, räumlich verteilten Entwicklung von H2-Brennstoffzellen-Fahrzeugen in Kooperation mit China – MorEH2“. The authors are grateful to the German Federal Ministry of Education and Research funding this project.

85

Supporting the synthesis of electric drive systems with scenario management

Bibliography 1. Cooper, Robert G. (2013): New Products. In Kenneth B. Kahn (Ed.): The PDMA handbook of new product development, vol. 46. 3rd ed. Hoboken, N.J.: Wiley, pp. 1–34. 2. Fischer, Jan O. (2008): Kostenbewusstes Konstruieren. Praxisbewährte Methoden und Informationssysteme für den Konstruktionsprozess. Berlin: Springer. Available online at http://dx.doi.org/10.1007/978-3-540-78313-8. 3. Ehrlenspiel, Klaus; Kiewert, Alfons; Lindemann, Udo; Mörtl, Markus (2014): Kostengünstig Entwickeln und Konstruieren. Berlin, Heidelberg: Springer Berlin Heidelberg. 4. Albers, Albert; Gausemeier, Juergen (2010): Von der fachdisziplinorientierten zur vorrausschauenden und systemorientierten Produktentstehung. In BMBF (Ed.): 10. Karlsruher Arbeitsgespräche Produktionsforschung 2010. Karlsruhe, März 9-10, 2010, pp. 248–256. 5. Fink, Alexander; Siebe, Andreas (2011): Handbuch Zukunftsmanagement. Werkzeuge der strategischen Planung und Früherkennung. 2., aktualisierte und erw. Aufl. Frankfurt am Main: Campus-Verl. 6. Gausemeier, Juergen; Fink, Alexander; Schlake, Oliver (1998): Scenario Management. In Technological Forecasting and Social Change 59 (2), pp. 111–130. DOI: 10.1016/S0040-1625(97)00166-2. 7. Hajkowicz, Stefan; Reeson, Andrew; Rudd, Lachlan; Bratanova, Alexandra; Hodges, Leonie; Mason, Claire; Boughen, Naomi (2016): Tomorrow's digitally enabled workforce: Megatrends and scenarios for jobs and employment in Australia over the coming twenty years. CISRO. Brisbane, Queensland. 8. Gausemeier, Jürgen; Pfänder, Thomas; Lehner, Anne-Christin (2017): Strategische Unternehmensführung mit Szenario-Management. In Dieter Spath, Engelbert Westkämper, Hans-Jörg Bullinger, Hans-Jürgen Warnecke (Eds.): Neue Entwicklungen in der Unternehmensorganisation. 1. Aufl. 2017. Berlin, Heidelberg: Springer Vieweg (VDI-Buch), pp. 97–109. Available online at https://doi.org/10.1007/978-3662-55426-5_15. 9. Fink, A., Schlake, O., Siebe, A., 2001. Erfolg durch Szenario-Management : Prinzip und Werkzeuge der strategischen Vorausschau. Campus-Verl. 10. Gausemeier, Jürgen; Plass, Christoph (2014): Zukunftsorientierte Unternehmensgestaltung. Strategien, Geschäftsprozesse und IT-Systeme für die Produktion von morgen. 2., überarb. Aufl. München: Hanser. Available online at http://www.hanser-elibrary.com/action/showBook?doi=10.3139/9783446438422.

86

Supporting the synthesis of electric drive systems with scenario management 11. Lewicki, P., 2013. Pkw und leichte Nutzfahrzeuge [WWW Document]. Umweltbundesamt. URL https://www.umweltbundesamt.de/themen/verkehr-laerm/emissionsstandards/pkw-leichte-nutzfahrzeuge (accessed 7.18.19). 12. Frondel, M., 2018. Stellungnahme von Prof. Dr. Manuel Frondel (RWI Essen) zum Vorschlag der Europäischen Kommission für eine Verordnung zur Festlegung von Emissionsnormen für neue Personenkraftwagen und für neue leichte Nutzfahrzeuge bis 2030. Deutscher Bundestag, Ausschuss für Umwelt, Naturschutz und nukleare Sicherheit, Ausschussdrucksache 19(16)68-E 13. Umweltbundesamt, 2012. Umweltzonen in Deutschland [WWW Document]. Umweltbundesamt. URL https://www.umweltbundesamt.de/themen/luft/luftschadstoffe/ feinstaub/umweltzonen-in-deutschland (accessed 7.18.19). 14. BMWi, n.d. Rahmenbedingungen und Anreize für Elektrofahrzeuge und Ladeinfrastruktur [WWW Document]. URL https://www.bmwi.de/Redaktion/DE/Artikel/Industrie/rahmenbedingungen-und-anreize-fuer-elektrofahrzeuge.html (accessed 6.17.19). 15. Kampker, A., Vallée, D., Schnettler, A. (Eds.), 2018. Elektromobilität: Grundlagen einer Zukunftstechnologie, 2. Auflage. ed. Springer Vieweg, Berlin. 16. Seeberger, M., 2016. Der Wandel in der Automobilindustrie hin zur Elektromobilität – Veränderungen und neue Wertschöpfungspotenziale für Automobilhersteller. St. Gallen. 17. Strathmann, T., 2019. Elektromobilität als disruptive Innovation: Herausforderungen und Implikationen für etablierte Automobilhersteller, BestMasters. Springer Gabler, Wiesbaden, Germany. 18. Keichel, M., Schöller-Schwedes, O. (Eds.), 2013. Das Elektroauto: Mobilität im Umbruch, ATZ/MTZ-Fachbuch. Springer Vieweg, Wiesbaden. 19. Inman, M., 2014. Natural gas: The fracking fallacy. Nature 516, 28–30. https://doi.org/10.1038/516028a 20. Samadi, S., Fischedick, M., Lechtenböhmer, S., Thomas, S., 2011. Kurzstudie zu möglichen Strompreiseffekten eines beschleunigten Ausstiegs aus der Nutzung der Kernenergie. 21. Elsner, H., Homberg-Heumann, D., Huy, D., Lutz, R., Moldenhauer, K., Plein, M., Schauer, M., Schmidt, S., Schmitz, M., Sievers, H., Szurlies, M., Wilken, H., 2019. Deutschland - Rohstoffsituation 2018. 22. Kurzweil, P., Dietlmeier, O., 2018. Elektrochemische Speicher: Superkondensatoren, Batterien, Elektrolyse-Wasserstoff, rechtliche Rahmenbedingungen, 2., aktualisierte und erweiterte Auflage. ed. Springer Vieweg, Wiesbaden.

87

Supporting the synthesis of electric drive systems with scenario management 23. ScMI AG, 2019. Scenario Manager TM. URL https://www.scmi.de/de/software/scenario-manager (accessed 7.18.19). 24. Marthaler, F., Albers, A., Gesk, J., Siebe, A., 2020. A Comparison of the Consistency Based and the Catalogue-based Approach of Defining Future Scenarios. XXXI ISPIM Innovation Conference (submitted) 25. Klimakabinett Deutschland, 2019. Eckpunkte für das Klimaschutzprogramm 2030. 26. Albers, A., Reichert, U., Ott, S., Radimersky, A., 2016. Software supported development of a Battery Electric Vehicle powertrain considering the efficiency. 16th International VDI Congress Drivetrain for Vehicles 2016 27. Marthaler, F., Albers, A., et al., 2019. Strategic Potential Identification for enabling an overarching generational Product Development: A qualitative study in the LiveLab IP – Integrated Product Engineering. Stuttgarter Symposium der Produktentwicklung 2019, pp- 233–252. 28. Albers, A., Marthaler, F., et al., 2018. Development methods for 2030: An interpretation of scenarios in the application of methods. NordDesign 2018 August 14 – 17, 2018 Linköping, Sweden

88

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems Marius Panzer, Claudia Wagner, Anna-Lena Winkler, Alexander Wöll, Dr. Richard Bernewitz MANN+HUMMEL GmbH, Ludwigsburg

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_11

89

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems

Abstract Changing drivetrain architectures and designs such as highly integrated e-Axles and hybrid transmissions require new filtration solutions. Especially when it comes to the joint integration of all drivetrain components such as the direct cooling of the engine and power electronics and even the direct cooling of the battery cells. For these tasks the requirements for differential pressure and oil cleanliness are less and the dielectric oil properties and the insulating function, for instance, come more and more into focus. Latest used filter analyses have shown that the abrasion and particle generation during the driving cannot be neglected when it comes to the direct integration of electric motor and power electronics into the oil loop. Therefore filtration solutions are needed which combine immediate highest efficiencies within compact installation space and low differential pressure. To protect all system components starting from the oil pump MANN+HUMMEL developed a new generation of filters using their depth filter media MULTIGRADE eM-CO. The new concept increases the degrees of freedom for the development of the lubrication circuit significantly. Offering an ultra-compact and flexible installation space, lowest differential pressure or highest filtration efficiencies for system reliability. To bear the mentioned electric and dielectric properties in mind which are especially important for the direct cooling of electric components such as the engine, power electronics and the battery cells a new coolant drying cartridge has been developed. The specially chosen drying agent reduces the water induced conductivity change of the oil and keeps the oil in the original condition. The oil humidity can be monitored by a proven sensor system which also indicates the service of the coolant drying cartridge. With this knowledge and expertise, MANN+HUMMEL developed a sensor-supported oil management system with intelligent cooling function and integrated oil drying for highest oil quality over the entire life of the system.

90

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems

1 Introduction The development of a climate and resource friendly transportation system is one of the major trends in the last years. To reach the target of maximum 1.5°C temperature rise compared to pre-industrial levels set by the COP21 Conference in Paris [1], emissions from transportation, accounting for 24-25% of the total CO2 emissions, have to be reduced significantly. With a rising number of congested cities, high levels of air pollution and stricter statutory requirements, companies are working towards even more sustainable solutions. Driving pleasure without local emissions is therefore considered as one of the biggest acceptance criteria for the mobility change towards electrified cars. Therefore mobility related companies are focused on the electrification of the powertrain. In this respect the highly integrated e-Axle is a particularly promising concept. The easy scalable e-Axles are a system which combines the electric motor, gearbox and power electronics. This saves installation space, components and wiring due to the fact that the e-Axle is integrated directly on the drive axle. According to the system suppliers, this enables up to 20 percent less weight and between five to ten percent greater efficiency. [2] After the launch of the first fully electric cars or trucks from all major automakers it’s the time now for the advanced development of the next generation of vehicles. Not only the lessons learned and the experience of the first car generation is considered but also the everyday suitability such as range extension and charging speed comes into focus. Together with the new developments of thermal management fluids and the advanced development of the other system components the unification of the thermal management loops are possible. [3] Oil as a non-conductive cooling and lubrication liquid is therefore suitable to play a significant role in this lubrication ant thermal management unification. The following chapters will highlight challenges and solutions for oil cleaning within highly integrated drivetrains.

2 Particles in the cooling and lubrication loop As all the drive components are integrated in an e-Axle, a common oil circuit is sufficient for cooling and lubrication. This advantage, however, has a drawback. Mechanical as well as electronic components are exposed to the same oil and metal particles and chips out of the transmission, which can penetrate the power electronics and electric motor and cause serious damage. Not only the chance of an electric short is given but also the abrasive particles can harm the Rotor/Stator system e.g. by the destruction of the winding coating. After the coating

91

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems is partly permeable by oil the copper-corrosivity of the oil can destroy the windings and therefore lead to a system failure. Additionally the strong magnetic forces within the electric motor can lead to a particle deposition and therefore a blocking or a malfunction of the motor cooling is given. Even an oil change cannot remove the particles after they have been deposited therefore an instant removal of the particles out of the system is needed. Older publications for automatic transmission suggest that the initial dirt is the main source of contamination [4] within the transmission system. Even though modern transmission production is much more sensitive towards a high initial component cleanliness this stays an important factor. Especially because initial dirt can be rather large (> 2000 µm) and consists of hard and abrasive particles (e.g. Splinters, Sand etc.). Nevertheless, the generated dirt during driving should not be neglected. Figure 1 shows the increase of the filtered dirt over the mileage of a BEV with a high integrated e-axle. As can be seen in Figure 1, the dirt amount within the first 80.000 miles is nearly linearly increasing. Depending on the initial dirt amount and the axle configuration an axle lifetime of 300.000 km this can finally add up to 1.5 - 4.5 g of dirt.

Figure 1: Collected amount of dirt over the first 100.000 km of a high integrated e-axle

The composition of the dirt is thereby strongly depending on the system architecture. Whereas in systems without direct electric motor cooling larger amounts of magnetic particles can be found in the filters these particles are nearly absent when the oil is in direct contact with the rotor/stator. It is assumed that the magnetic forces within the electric motor leads to a deposition of these magnetic particles and therefore they are

92

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems not transported further down the oil gallery. Additionally mostly aluminum based or in case of a starting motor failure copper based particles can be found in the system. Apart from the inorganic and metallic particles it needs to be mentioned that also larger amounts of organic contamination can be found in the system. Here the particles are mostly plastic debris, sealing abrasion, soft fibers or epoxy contamination. Unlike the metallic particles are these particles softer and less abrasive, therefore the system risk is smaller. If a closer look is taken on the collected particle sizes it can be observed that the collected abrasive metallic particles can be differentiated in two major groups. The first group are the larger splinter-like particles. Mainly from initial debris but also larger two dimensional chips or even broken-off metallic parts. These particles can be up to 5000µm in length. On the other hand there is a huge amount of much smaller particles generated during by the driving. Over 90% of the collected particles are between 1 and 100µm in size with an average at about 20µm.

Figure 2: Particle size distribution of the collected amount of dirt of a high integrated e-axle

To ensure a long axle-lifetime, all these particles need to be efficiently, quickly and permanently removed from a reliable working system. Therefore filtration solutions are needed which combine highest efficiencies within compact installation space and low differential pressure.

93

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems

3 Fully synthetic filter media MULTIGRADE eM-CO designed for e-mobility needs To take the particle analysis and the high integration of transmission, electric motor and power electronics into consideration it becomes evident that an instant oil cleanliness is very important. This prevents the deposition of the particles within the engine and ensures the lifetime of the system components such as fast running bearings or sensitive gaskets and valves. The filtration experts at MANN+HUMMEL have taken on the challenges that come with alternative drivetrains and developed suitable transmission oil filters for e-Axle applications. The filters are equipped with the new MULTIGRADE eM-CO filter media. The use of a depth filter media enables the secure entrapment of the separated particles in the 3D fiber matrix. This prevents the reintroduction of particles or even long chips into the system, a danger which is common with conventional strainer systems. The fully synthetic filter media grades MULTIGRADE eM-CO are free of hard glass fibers and have a very high chemical resistance at considerably lower differential pressure in comparison to conventional transmission oil filters fitted on the suction-side (Figure 3). MANN+HUMMEL utilizes consistent filter media layers without uncontrolled artificial media bypass, something which is required for conventional products due to their high loss of differential pressure. [5]

Figure 3: Enlargement of MANN+HUMMEL MULTIGRADE eM-CO filter media 3D structure without artificial bypass holes for a secure entrapment of the separated particles

To ensure the system reliability all filter media are tested according to ISO 16889 with the restriction that for the more open filter media larger sensors and ISO Coarse (A4) or test dust with particles up to 300 µm is used. The ISO 16889 filter testing is under the given circumstances the only standardized measurement which gives reliable results

94

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems for the immediate effect of a filter on the particle reduction. It also takes the constant particle generation by the axle system into consideration and allows at every point of the test a statistically proofed result by the presence of a given base upstream gravimetric level (BUGL). As can be seen in Figure 4 the different MANN+HUMMEL media portfolios provide an excellent filtration base for each customer system and filtration task. The different media ranges are between 99.5 % efficiency at 5 µm(c) for highest oil cleanliness up to 99.5 % at 180 µm for an excellent ration of differential pressure and filtration efficiency.

Figure 4: ISO 16889 results at the clogging point (CP) for different MANN+HUMMEL filter media grades used within transmission filters either on suction side or on pressure side

The MULTIGRADE eM-CO filter media require less space as the engineers at MANN+HUMMEL were able to exploit the company’s proven pleating technology. This gives the customer a larger degree of freedom with regard to higher filter fineness or dimensioning of the pump, as the energy dissipation is minimized by the pleating technique. The filter media always retains its shape and performance, even when exposed to coldness or high differential pressure. Stability is ensured by a drainage grid (Figure 3) or adhesive lines which maintain the ideal gap between the pleats. All this enables more long-term use of the complete filter surface area and is especially useful within suction side filter development.

95

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems The degrees of freedom for the customer to choose from by the pleating technology are (Figure 5) [6]: ● Increasing the system reliability by the use of up to 100 times more efficient filter media at the same installation space and differential pressure ● Reducing the needed installation space by 83% keeping the differential pressure and the efficiency un a comparable level ● Increasing the energy efficiency of the system by reducing the differential pressure by 75% keeping the installation space and the efficiency comparable

Figure 5: The MANN+HUMMEL pleating technology provides different degrees of freedom graphical represented by the comparison of two suction sider filters

96

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems

4 Additional filtration tasks Not only the abrasive risks of the oil contamination through particles are in the focus when it comes to e-Axles, also another important oil condition needs to be considered. Especially with the unification of cooling circuits with the same oil the dielectric and electric properties of the oil becomes more and more important. In particular at highly integrated e-Axles, where the oil is used to cool the motor and the power electronics it is important to have low conductivity and a high breakdown voltage. Due to oil ageing, temperature and humidity, the oil can lose its strong insulating properties, leading to severe risks such as electric shortages and destruction of the e-Axle. The oil ageing and the temperature are correlated and can be influenced by the additives and the thermal management system. It is very important to remove the humidity as good as possible out of the system. By using a MANN+HUMMEL coolant dryer cartridge it is possible to keep the induced humidity increase in the oil on a very low level, as the dissolved water is constantly removed from the oil. [7] The coolant dryer cartridge can be easily and sensor-based serviced to ensure an always low humidity level without oil change. (Figure 6).

Figure 6: Progress of relative humidity [%] over the time during the treatment with the oil dryer cartridge (right) after the intentional adding of 2000ppm water to DEXRON VI ATF fluid at laboratory scale recorded with the humidity sensor available for series application

97

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems

5 Modular thermal management system For highest integration, the newly developed filters can be arranged in a customer specific and flexible housing design. However, either as standalone filters within the customer given installation space or integrated in a compact modular thermal management system, all filters ensure maximum system protection against abrasive particles and water, at smallest installation spaces (Figure 7). Many years of plastic expertise at MANN+HUMMEL and the use of modern simulation methods such as the finite element method (FEM) and mold filling simulations in the development process ensure that all components are able to withstand the highest mechanical requirements. The engineers also exploited the possibilities of simulations with computational fluid dynamics (CFD) to enable efficient flow characteristics for the filter housing in particularly tight installation spaces, to minimize energy losses by a minimal differential pressure with maximum system protection. Additionally the long MANN+HUMMEL expertise in acoustic design optimization can be exploited. Together with an intelligent oil cooling, the MANN+HUMMEL oil management modules protect e-Axles and cooling oil circuits at a very high level, by including filters for the protection of bearings, gears and channels, for instance, as well as controlled and optimized oil flow and humidity sensor-supported oil drying. As a result, cleanest oil at constant high quality from the first second is ensured.

Figure 7: Pressure, temperature and humidity sensor supported oil management system with intelligent cooling function and integrated oil drying for highest oil quality over the whole system lifetime

98

Protection of e-axles and innovative drivetrains with multipurpose oil filter systems

Bibliography 1. UNFCCC, Paris Agreement, 2015, http://unfccc.int/files/essential_background/convention/application/pdf/english_paris_agreement.pdf 2. Robert Bosch GmbH: https://www.bosch-presse.de/pressportal/de/en/the-start-uppowertrain-for-electric-cars-the-bosch-e-axle-offers-greater-range-121216.html, accessed on 2018-28-09. 3. Automobil-Industrie, Petronas Lubricants International: Hydraulische Systeme bei E-Autos: Flüssig bleiben, https://www.automobil-industrie.vogel.de/hydraulischesysteme-bei-e-autos-fluessig-bleiben-a-884390/ accessed on 202-01-22. 4. John. G. Eleftherakis, Abrahim Khalil, Doug Parnell, Bill Pizzo and Dan Haggard Advances in Automatic Transmission Cleanliness; SAE Paper 2001-01-0372 5. MANN+HUMMEL GmbH: https://www.mann-hummel.com/en/the-company/magazines/automotive-news/automotive-news-ausgabe-022017/filtration-competence-fore-mobility/ accessed on 2019-18-10. 6. MANN+HUMMEL GmbH: https://www.mann-hummel.com/en/areas-of-expertise/ solutions-for-alternative-drives/maximum-filter-performance-for-the-e-axle/ 7. MANN+HUMMEL GmbH: https://www.mann-hummel.com/en/areas-of-expertise/ solutions-for-alternative-drives/clean-drives-further/ accessed on 2019-18-10.

99

Breaktor™ – Advanced protection for high voltage circuits in electric vehicles Till Wagner, Kevin Calzada Eaton Electrical Products Ltd

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_12

101

Breaktor™ – Advanced protection for high voltage circuits in electric vehicles

1 Electric vehicle trends The vehicle electrification era is here, bringing with it a wave of innovative technologies and exponential advancements. However, there are important safety implications to be considered as these technologies and market trends are adopted. What are these trends and how do they impact circuit protection of electric vehicles?

Figure (1) EV market requirements for mass adoption.

The market today sees demand and development towards decrease charging times, increased range requirements, and a wider variety of vehicle models. All of which are impacting the electrical system performance and design. Decreasing charging times result and a demand for higher voltage charging and systems. Due to the higher overall power levels that can be achieved as the system voltages are increased, platform and charging voltage fall into two segments around 400VDC and 800VDC. Today’s electric vehicle can be mainly found at the 400VDC with some exception and plans to develop 800VDC system voltage platforms.

102

Breaktor™ – Advanced protection for high voltage circuits in electric vehicles

Figure (2) xEV voltage falls into two segments [IHS, 2019].

While a trend analysis and market requirements are pointing towards an increased system voltage level for future platforms, planning data is indicating a steady deployment of 400VDC system voltage throughout future years. Technological challenges are slowing down the adoption of 800VDC system architectures, though it does provide clear benefits in the long run.

Figure (3) Technology limitations slowing 800-volt adoption [IHS, 2019].

Increased range requirements influence the size of batteries and the total power provided by the battery. As a result, battery development will see increased battery capacities, which result in the directly related available fault current that can be delivered to a vehicle electrical system in the event of a short circuit on the battery. The manufacturers must consider the increased fault current in the protection component requirements.

103

Breaktor™ – Advanced protection for high voltage circuits in electric vehicles

Figure (4) xEV fault current directly related to battery capacity [IHS, 2019].

On top of that, the motor power and current is ramping up. By that, circuit switching and protection devices will face higher requirements. They must withstand the higher operating currents and cyclic requirements.

Figure (5) xEV protection devices need to withstand higher operating currents [IHS, 2019].

Summing up, all three trends are challenging developments for circuit protection and need careful consideration. The decreased charging time requirement promotes higher system voltages. Increased range results in higher fault currents. And the variety of models planned indicates a higher operating current requirement.

2 Critical protection considerations Conventional system architecture design for main protection involves a coordinated fuse and contactor solution within the power distribution box. The two components must be coordinated to ensure coverage of the full range of possible fault scenarios, which can derive from different battery states of charge and route causes.

104

Breaktor™ – Advanced protection for high voltage circuits in electric vehicles

Figure (6) Conventional xEV circuit protection.

The conventional circuit protection solution of fuses and contactors has contradicting design requirements. Current levels in a vehicle can vary widely from 1

(9)

Finally, the drive consumption over a certain time is given as: 𝐸

=

𝑃

(𝑡)𝑑𝑡 

(10)

3 Low resolution problem The basic problem of drive consumption prediction is the general and temporal inaccuracy of future velocity data due to unpredictable environmental conditions such as weather, traffic density or weight influenced by the number of passengers [5], [6]. However, the physical equations commonly used to calculate the drive consumption require a certain resolution. Otherwise, the result only reflects the actual demand to a limited extent. When predicting a future velocity profile for a city bus, it is likely that we use

116

Calculation of driving consumption of an electric city bus average velocities for a specific distance. Especially if we do not know which characteristics such as traffic signs appear along the route. Such speed intervals can last from a few seconds to, in the worst case, the time between the bus stops on the timetable. Consequently, a route with ten stops would consist at least of nine average velocity values. As shown in Figure 1, a poorer resolution affects the calculation accuracy of forces and powers and leads to a lower drive consumption, as some forces can no longer be represented. In the following, several influences are examined in more detail based on 174 real city bus trips with a total distance of about 8200 km and a measured velocity profile resolution of 0.2 s (5 Hz). In order to generate trips with a lower resolution, we used an algorithm to split the measured data into specific constant speed intervals, see Figure 2. We then applied the energy approach to these intervals in order to determine the general accuracy. Two main problems have emerged, which influence the calculation of the drive consumption: ● Reduction of the velocity peaks due to average values ● Continuity loss of velocity profile due to jumps between average values

Figure 1: Decreasing drive energy consumption at different velocity profile resolutions while considering a constant efficiency value

Figure 2: Artificially generated low-resolution velocity profile based on measured velocity data

117

Calculation of driving consumption of an electric city bus

3.1 Problem of velocity reduction The reduction of the velocity peaks influence the forecasted air and roll resistance in Figure 1. The air resistance varies squared with the velocity, which in turn means that the power required is proportional to the velocity cubed. Higher velocities in particular therefore lead to a sharp increase in energy consumption. By having longer time intervals with average velocities, the air resistance is likely to decrease. Since the velocity of a city bus in general is rather moderate, the average speed is mostly below 30 km/h [17]. Thus, the change of the air resistance due to longer intervals is in relation to the other resistance not decisive, see Figure 3. The roll resistance depends on the roll coefficient cr, which in general varies with velocity and ground conditions. In the case of a city bus with low average velocities and different ground conditions, it is difficult to assume the right value at every point, especially for larger time intervals. Therefore, we assume a constant roll coefficient. Small resolution changes in the range of few seconds have an influence around 10 percent, which is due to the efficiency calculation. Thus, in Figure 4 we can see a jump between the high resolution of 0.2 s and the lower resolution of 30 s. However, with longer time intervals (> 30 s) the influence becomes less and a change in roll resistance is hardly noticeable. Therefore, we did not investigate this effect further in this publication.

Figure 3: Change in drive consumption due to air resistance at different resolutions considering a constant efficiency value

118

Figure 4: Change in drive consumption due to roll resistance at different resolutions considering a constant efficiency value

Calculation of driving consumption of an electric city bus

3.2 Continuity loss of velocity profile If the resolution of the velocity profile decreases below 1 Hz, time sequences are higher than 1 s, the velocity curve changes from a continuous curve to a discontinuous function having jumps at every sequence end. This leads to problems calculating the slope and acceleration power. The necessary slope power results from the multiplication of equation (4) with the velocity v. If only a discrete velocity function is used for the drive resistance calculation, the following applies to the time-average in general: (11)

sın(𝛼) 𝑣 ≠ sın(𝛼) 𝑣̅  

Table 1: Example velocity vector and possible errors multiplying time-average values v(t) [m/s] 5 3 1 0

α(t) [°] 0 2 5 5

sin(𝛼) 0 0.03 0.09 0.09

sin(𝛼)𝑣 0 0.10 0.09 0

sın(𝛼) 𝑣

sın(𝛼)

𝑣̅

sın(𝛼) 𝑣̅

0.05

0.05

2.25

0.11

Table 1 shows an example velocity vector with corresponding slope angle values. The example calculation of the correct time-dependent result as well as the different possibilities of calculating the corresponding average validate equation (11). The false separated multiplication of the average speed and slope angle results in a false slope resistance. Therefore, we cannot simply multiply the time average of the velocity with the time average of the slope values. However, in the case of the city bus the topography of the route is mostly known. Thus, we can conduct precise location-depend slope values for the calculation. Hence, we do not examine the dependency of the slope resistance on larger time intervals of the vehicle speed further. What is more, we already explained in section 2 that the calculation of the overall drive consumption based on the total drive resistances in equation (6) is no longer valid if the vehicle speed function is discontinuous. Instead, we proposed the energy approach, which calculated the acceleration power separately derived from the energy conservation approach, see equation (8). Figure 5 and figure 6 show the difference between the acceleration power share in the total consumption for the force approach of equation (6) and the energy approach of equation (8) at different velocity resolutions. Both approaches tend to underestimate the acceleration consumption at larger time intervals, see figure 6, since not all acceleration phases are represented by the average velocities. Apart from this, the force approach results in a larger underestimation of the acceleration energy, even at a resolution of 0.2 s (5 Hz). The reason for this is the neglect of the second velocity v2, to which acceleration is applied.

119

Calculation of driving consumption of an electric city bus

Figure 5: Comparison of the acceleration power share in consumption for energy and force approach, resolution 0.2 s

Figure 6: Comparison of the acceleration power share in consumption for energy and force approach, resolution 60 s

Focusing on the acceleration power, the discontinuity of the generated velocity data leads to unrealistic performances, which can no longer be carried out by the electric drive. For example a large difference in neighboring average velocities divided by a short time, for example 10 km/h in 1 s. Note that without the correct calculation of the acceleration power of equation (7), the drive consumption becomes more and more error-prone. Various publications attempt to solve this problem by inserting stops or acceleration phases over several seconds between two intervals at average velocities [11], [12]. In addition, they also adjust the average velocity values so that the total distance travelled does not change. However, these modifications are intended to make the velocity profile more realistic, but manipulate driving consumption in a certain way. Furthermore, the acceleration consumption decreases with longer time intervals due to the lower frequency of accelerations and thus, a lower efficiency loss, see Figure 7. In addition to unrealistic high powers, which can exceed the maximum powers of the electric drive, the motor efficiencies are no longer representative, because the drives are designed for real drive situations. By still using the original efficiency map while calculating with lower resolutions using average speeds, the total drive consumption is getting significantly wrong, compare Figure 11, adjustment 1. Accordingly, we must define a drive efficiency corresponding to the "normal" velocity profile, for example a constant value.

120

Calculation of driving consumption of an electric city bus

Figure 7: Change in drive consumption due to acceleration resistance at different resolutions considering a constant efficiency value

4 Proposals for correction of data and physical model In the previous section, we showed that the velocity resolution influences every calculated drive resistance and thus the resulting drive consumption. Especially the loss of continuity leads to several significant problems while calculating the drive consumption for velocity profiles with low resolutions. We now propose two solutions reducing the inadequateness’s of calculating the acceleration force and the usage of the drive efficiency map in such cases. All investigations are based on a set of 174 test drives with a total distance of about 8,200 km.

4.1 Fitting acceleration consumption The fitted acceleration resistance curves of the test drives in Figure 8 show a similar behavior over all drives. The acceleration resistance share significantly reduces with decreasing vehicle velocity resolution. Now we must find a solution to derive the right acceleration consumption based on such an incorrect reduced value due to a lower velocity profile resolution. One possibility to correct the decrease would be a “best fit” line with which one can deduce the original value from every point. The basis for this approach is that the decrease always behaves in the same way. We used a constant efficiency value for the physical model, in order to minimize influences of the efficiency map problems described in Section 3.2. The “best fit” method consists of a fifth degree polynomial and shows the course of the acceleration energy over the different interval lengths, see Figure 8 and Figure 9. Although, we can identify trends and a region about where the original value should be, these are subject to inaccuracies of up to 50%. Furthermore, a data-rich look up table is required, since up to now we were not able to find a generally valid function.

121

Calculation of driving consumption of an electric city bus

Figure 8: Polynomial fits on drive consumption due to acceleration resistance for all 174 trips

Figure 9: Spread of polynomial fits on drive consumption due to acceleration resistance

4.2 Modification of velocity data and efficiency map The efficiency map of electric drives is usually given in a speed-torque matrix. With the vehicle velocity and the required drive force, it is possible to calculate the required speed and thus torque of the electric drive. The corresponding efficiency for the calculation can then be conducted from the matrix. One solution for lesser velocity resolutions is not to use the torque-dependent efficiency map, but to transfer a purely velocity-dependent variant for average velocities. In this way, power peaks outside the map could be taken into account, which are generated by the instant transition between two velocity intervals. For this purpose, we examined all high resolution driving data and their original engine efficiencies in relation to speed and we derived a new efficiency map. It is important to weight the efficiencies according to the sought-after performances, as these have a greater influence on overall driving consumption [18]. 𝜇̅

=

∑

𝜇 ∑

(𝑡) (𝑡) 𝑃 (𝑡) 𝑃

(12)

Equation 7 applied to different velocities gives the new velocity-dependent efficiency map. Another possibility, as in various publications, e.g. [14], [13], and [19], is to use a single motor efficiency for all different powers, regardless of whether mechanically possible or not. Again, equation (12) is used, but this time independent of the velocity. A further possibility to solve the discontinuity of the lower resolution velocity profile is the insert of stops or acceleration phases over several seconds between two intervals

122

Calculation of driving consumption of an electric city bus at average velocities [5], [6]. For this purpose, we scoured the original trip data for real stops and these are entered into the longer intervals as shown in Figure 10. Furthermore, we have provided the transition between two constant speeds with a constant realistic acceleration, if there is no stop close to it. However, the adjustment of the velocity profile results in a shortening of the total distance, which leads to a reduced driving consumption. Therefore, we must increase the average speed to regain the original distance. We did not take this into account in this study, but should do so in the future, in order to further improve the calculated drive consumption.

Figure 10: Scheme for inserting acceleration phases and stop times

We applied all these described efficiency adjustments to the total energy demand in order to assess the overall impact. Thus, our goal is to become an idea about the influences and find promising solutions for a further general approach. Therefore, we evaluated the spread over our total measured driving data. If the spread is small, the found results are valid for all considered measured data and thus are robust against environment conditions and influences from different trip characteristics. Further, the closer the found values are to one, the better we are able to reproduce the measured drive consumption. Efficiency map and velocity profile adjustments in the calculation of the drive consumption: 1. Torque-dependent efficiency map 2. Torque-dependent efficiency map with insertion of stops and acceleration 3. Constant efficiency coefficient 4. Constant efficiency coefficient with insertion of stops and acceleration 5. Speed dependent efficiency map 6. Speed dependent efficiency map with insertion of stops and accelerations

123

Calculation of driving consumption of an electric city bus

Figure 11: Calculation adjustments vs. measured consumption for a 30 s resolution

Figure 12: Calculation adjustments vs. measured consumption for a 60 s resolution

Figure 13: Calculation adjustments vs. measured consumption, resolution 120 s

Figure 14: Calculation adjustments vs. measured consumption, resolution 180 s

Figure 11 to figure 14 show the influence of these six adjustments for different velocity resolutions. We compare the ratio of the calculated drive consumption with the measured one for four different velocity resolutions. Especially at higher resolutions of 30 s, 50 percent of the data of version 1 and 5 are inside a spread of 17 percent (quartile range) over all test drives, which makes them difficult for finding a representative value. The introduction of additional stops and accelerations in adjustment 2, 4, and 6 seem to

124

Calculation of driving consumption of an electric city bus reduce the quartile range to a minimum of 4 percent at version 6. Even at lower resolutions, version 6 is generally the best option, with a maximum quartile range of 9 percent at 180 s. The overall results are given in Table 2. Table 2: Upper quartile, lower quartile and range of the adjustment of the drive consumption calculation with different velocity profile resolutions 1 1.20 1.03 60s 0.96 0.86 120s 0.88 0.76 180s 0.85 0.74 30s

0.17 0.10 0.12 0.11

2 0.90 0.81 0.87 0.79 0.84 0.76 0.84 0.74

0.09 0.10 0.08 0.10

3 0.96 0.86 0.86 0.75 0.82 0.69 0.80 0.67

0.10 0.11 0.13 0.10

4 0.95 0.85 0.86 0.76 0.83 0.70 0.80 0.65

5 0.10 1.18 1.01 0.10 0.94 0.85 0.13 0.84 0.76 0.15 0.83 0.71

0.17 0.09 0.08 0.12

6 0.93 0.89 0.87 0.81 0.83 0.75 0.81 0.72

0.04 0.06 0.08 0.09

5 Conclusion In order to calculate the drive energy consumption for future trips, a backward approach for longitudinal dynamics is usually used. However, if the necessary velocity data is not available in a high resolution, changes have to be made, both to the data and to the physical model to obtain a representative energy consumption. We attributed the occurring problems to two main aspects: the reduction of the velocity peaks due to average values and the continuity loss of the velocity profile due to jumps between average values. The peak reduction mainly affects the air resistance, which depends on the square of velocity. Since the velocity of a city bus in general is rather moderate, the change of the air resistance due to longer intervals is in relation to the other resistances not decisive. A first consequence of the continuity loss is that we must correct the classic backward approach based on the driving force, since it is only approximately correct. The approximation is invalid in cases where the velocity profile does not have a minimum resolution of at least 1 s. Instead, the principle of energy conservation must be taken into account when calculating the acceleration power. As a result, the second velocity v2 to which the acceleration is applied is not neglected. Furthermore, the continuity loss mainly affects the acceleration resistance and the engine efficiency map. The jumps between average velocities generate forces, which can no longer be mapped. In general,

125

Calculation of driving consumption of an electric city bus the predicted drive consumption decreases with decreasing velocity resolution compared to the measured drive consumption. In the evaluation, no mathematical function could be found that would indicate the original consumption with the usual application of the physical equations. Accordingly, various modifications of the velocity profile and the energy prediction model were determined to improve the acceleration resistance and the used efficiency for calculating the overall consumption. The best modification is the introduction of constant acceleration and stopping times combined with the use of a speed-dependent drive efficiency map. Over all test drives, it resulted in a quartile range of 4% for a resolution of 30 s and 9% for a resolution of 180 s. In future investigations we will further adjust the previously discussed velocity profile with added acceleration by increasing the velocity to correct the average value. The discovered roll consumption change at small resolutions (1-10 s) will also be examined more accurately.

Bibliography 1. Bundesministerium für Verkehr und digitale Infrastruktur, “Sofortprogramm Saubere Luft 2017-2020”, 2017. [Online]. Available: https://www.bmvi.de/Shared Docs/DE/Artikel/G/sofortprogramm-saubere-luft-2017-2020.html. [Accessed: 02.01.2020] 2. Bundesministerium für Wirtschaft und Energie, “Energieeffizienzstrategie 2050”, 2019. [Online]. Available: https://www.bmwi.de/Redaktion/DE/Publikationen/Energie/energieeffiezienzstrategie-2050.html [Accessed: 02.01.2020] 3. Global Covenant of Mayors for Climate & Energy, "Our Commitment to Green and Healthy Streets Fossil-Fuel-Free Streets Declaration," 2017. [Online]. Available: https://www.c40.org/other/green-and-healthy-streets. [Accessed: 02.01.2020] 4. H. Hielscher, "Vier Städte wollen bis 2030 auf Elektro-Busse umstellen", WirtschaftsWoche, 01.11.2018. 5. D. Pevec et. Al., "Electric Vehicle Range Anxiety: An Obstacle for the Personal Transportation (R)evolution," 4th International Conference on Smart and Sustainable Technologies (SpliTech), 2019, Croatia 6. H. Fechtner and B. Schmülling, "Survey of Current Vehicle Mass Estimators in the Context of Future Mobility," 2nd IEEE International Conference on Intelligent Transportation Engineering (ICITE), 2017, Singapore 7. J. Asamer et. Al., "Sensitivity analysis for energy demand estimation of electric vehicles”, Transportation Research Part D: Transport and Environment, Vol. 46 (2016), 182–199

126

Calculation of driving consumption of an electric city bus 8. M. Delavaux et. Al., "Comparison between Forward and Backward approaches for the simulation of an Electric Vehicle”, Vehicle Power and Propulsion Conference (VPPC), 2010, France 9. C.J.J. Beckers et. Al., "Energy Consumption Prediction for Electric City Buses”, 13th ITS European Congress, 2019, Netherlands 10. T. Halmeaho et. Al., "Experimental validation of electric bus powertrain model under city driving cycles”, IET Electrical Systems in Transportation, Vol. 7, Iss. 1 (2017), 74–83 11. J. Wang et. Al., "Battery electric vehicle energy consumption prediction for a trip based on route information”, Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, Vol. 232, Iss. 11 (2018), 1528–1542 12. L. Thibault et. Al., "A Unified Approach for Electric Vehicles Range Maximization via Eco-Routing, Eco-Driving, and Energy Consumption Prediction”, IEEE Transactions on Intelligent Vehicles, Vol. 3, Iss. 4 (2018), 463–475 13. C. Bredberg and J. Stjernrup, "Driver Modeling, Velocity and Energy Consumption Prediction of Electric Vehicles”, Lund University, Lund, June 2017. 14. M. Gallet et. Al., "Estimation of the energy demand of electric buses based on realworld data for large-scale public transport networks”, Applied Energy, Vol. 230 (2018), 344–356 15. D. Zhou et. Al., "Online energy management strategy of fuel cell hybrid electric vehicles based on time series prediction”, IEEE Transportation Electrification Conference and Expo (ITEC) , 2017, USA 16. M. T. Sebastiani et. Al., "Evaluating Electric Bus Operation for a Real-World BRT Public Transportation Using Simulation Optimization”, IEEE Transactions on Intelligent Transportation Systems, Vol. 17, Iss. 10 (2016), 2777–2786 17. L. Braun, "How accurate can a range calculation of an electric vehicle be?”, 20. internationales Stuttgarter Symposium, 2020, Germany 18. L. Braun, "Fahrer- und fahrsituationsabhängige Bewertung unterschiedlicher Elektromotorkonzepte”, Karlsruhe Institute of Technology, Karlsruhe, 31.12.2012. 19. I. K. Reksowardojo et. Al., "Development of Engine Power Capacity Calculation Method for Range Extender and Case Study in Medium-Size Electric Bus”, 5th International Conference on Electric Vehicular Technology (ICEVT), 2018, Indonesia

127

Entrance to an electrified last mile ecosystem project “bring your own battery” Markus Geiger, Dr. Bernhard Budaker csi entwicklungstechnik GmbH Gunnar Lange, Christian Schmidt Audi AG

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_15

129

Entrance to an electrified last mile ecosystem project “bring your own battery”

1 Introduction In a world of urbanization, the lockout risk for vehicles in cities is continually increasing. Emissions, air pollution and the increasing volume of traffic as well as the resulting traffic jams in cities are the main drivers for this development. A sustainable interface between interurban and urban mobility will become necessary. Due to its high efficiency and emission-free drive, electric mobility in combination with innovative last mile concepts will cover these needs.

1.1 Challenges of urban mobility The volume of traffic is increasing rapidly, especially in inner-city areas and in industrial metropolitan areas. The consequences are increased fuel consumption and thus CO² emissions, more noise, delays and stress for drivers. Innovative mobility and infrastructure concepts (e.g. figure 1) are becoming much more important in these metropolitan areas.

Figure 1: Urban city of the future [1]

130

Entrance to an electrified last mile ecosystem project “bring your own battery”

1.2 Last Mile Concepts as intersection between interurban and urban mobility In order to relieve the metropolitan areas and reduce emissions, there are already various solutions available using various mobility and transport concepts. One of them is a car-free city, which is no longer just a utopia. Many cities have already completely banned cars from the city center or at least significantly reduced their presence [2]. Besides significantly better air quality and reduced noise, traffic safety is also a very positive aspect of car-free city centers. In Houten, Madrid, Paris and Copenhagen, for example, different mobility concepts for car-free inner urban regions are already being developed and implemented. Last but not least due to the demand of society to reach the workplace or the inner city in a time-efficient and possibly stress-free way, the number of micromobility solutions in cities and metropolitan areas has been rising very strongly in the recent years. Escooters and E-bikes are no longer a rarity, but can be found in large numbers in almost every bigger city. Car manufacturers have also developed some products in the field of micromobility. One example is the Audi etron scooter of the Audi AG shown in figure 2. With a maximum range of 20 km and a top speed of 20 km/h, it is an attractive alternative to the automobile in metropolitan areas.

Figure 2: Audi etron e-Scooter [3]

131

Entrance to an electrified last mile ecosystem project “bring your own battery”

1.3 Challenges of micro mobility The main challenge of micro mobility is to keep up the safety of all participants in road traffic. In many cases, accidents occur due to unexpected behavior by users of LEVs (light electric vehicles). Car drivers, for example, often expect a pedelec to behave in a similar way to an ordinary bicycle. The speed, energy and power of a pedelec are often being underestimated and in combination with the constantly growing traffic volume, dangerous situations and even accidents can occur very quickly. In addition, the rider of a pedelec up to a maximum speed of 25 km/h does not need a special driving license. In order to reduce this safety risk, there are already many training courses and other measures for learning how to handle a LEV in an appropriate, safe and environment-oriented manner. The manufacturers and users of a pedelec must be fully conscious of the fact that by integrating an electric drive they are transforming the originally muscle-powered bicycle into a machine. This means that higher safety requirements must be fulfilled in the development, manufacture and use of pedelecs. Furthermore, the riding behavior of the driver must be adjusted to the increased risk of danger.

2 Innovative Sharing Concept Together with csi entwicklungstechnik, Audi AG has developed the innovative sharing concept "bob - bring your own battery" consisting of an e-bike and battery system. These systems represent the interface between the world of vehicles and the world of micro-mobility. The Audi location in Ingolstadt was chosen as an application scenario. The main idea of the innovative sharing concept is based on the unlocking/locking of the NFC lock on a pedelec fleet with the bob battery system. One battery system can be unlocked to several pedelecs, which also allows a limited fleet solution for different groups of people or departments in a company.

2.1 Motivation Audi employees with their working places in the satellite buildings outside the company grounds often do not have a time-efficient and comfortable way of keeping appointments on the company site. This is mainly due to the very high volume of traffic around the company ground and the parking situation at the Audi site in Ingolstadt. The aim of the user-oriented micro mobility concept “bob” is to enable an easy access to an electrified sharing community.

132

Entrance to an electrified last mile ecosystem project “bring your own battery”

2.2 The user-oriented sharing concept The innovative Last Mile concept is based on the existing Audi AG company bike concept with around 1500 bikes. An electrified version of these company bikes allows Audi employees to be offered an expanded portfolio of company mobility. With the business model "bob - bring your own battery" the user receives the practical, personalized and portable battery system and can thus upgrade the existing company bikes with integrated e-drive into pedelecs. The option of continued use of the modified company bikes as muscle-powered bikes ensures dual mode and therefore an expansion of the existing company mobility. In a first demonstration project, a sharing concept for PLEVs (Personal Light Electric Vehicle) was developed and implemented in first vehicle prototypes. In this project phase of the sharing concept, the technical feasibility was proven by a first functional prototype. In a further project phase, the individual components of the sharing concept and their design were adapted and modified. After the technical implementation of the last mile concept, a pilot phase within Audi AG at its Ingolstadt location will follow. This will be used for validation of the business model.

3 Last Mile Concept – bob: “bring your own battery” In the pilot phase of the bob project, a very small series of e-bikes and battery systems will be built and then tested by trained Audi employees in a longer test phase. The overall system has an appropriate degree of robustness and functional stability so that the test phase can be carried out as smoothly as possible and without system failures. Nevertheless, these systems are currently still functional prototypes and not series products. Due to the strict regulations on the company site, the existing and already tested company bikes will be used for the implementation of the innovative sharing concept. Therefore, only selected safety-relevant components must be tested after finishing the modification. The required type of testing/certification was identified in the scope of a comprehensive certification analysis with recommendations for action. The technical implementation of the e-bike and the battery system will be followed by a more extensive rollout within Audi. An expansion of the concept for external customers is also planned.

133

Entrance to an electrified last mile ecosystem project “bring your own battery”

3.1 E-Bike System Figure 3 shows the extended company bike with electric wheel hub drive on the front wheel, power electronics, holder for the bob battery system and NFC lock. The electrical cables are routed through the bike frame, which keeps the design as simple as possible even after the modification to a pedelec and minimizes the risk of failure or damage.

Figure 3: Company bike of Audi AG converted to a pedelec [4]

The 36 V electric hub drive is spoked into the rim of the front wheel, has an output of 250 W and supports the user of the pedelec up to a maximum speed of 25 km/h. The company bikes are converted by replacing the front wheel. In addition, the modified power electronics and sensors are mounted on the frame between the pedal bearing and the rear wheel. To keep the complexity of the bike as low as possible, the function control unit is integrated into the battery system. Only by inserting the battery system into the battery holder on the bike side will be the bike awakened to an e-bike.

134

Entrance to an electrified last mile ecosystem project “bring your own battery”

3.2 Battery System The battery system includes a 36 V Lithium-Ion battery pack with a capacity of 2500 mAh, whereby the maximum range of the pedelec is about 20 kilometers and therefore enough for the validation scenario considered in the project. Figure 4 shows the battery system consisting of a control unit PCB, LED elements, USB connectors and an active NFC module. The bob battery system can be controlled with a rotary/push button of a current infotainment module from a car model of Audi AG.

Figure 4: bob battery system [4]

The control unit PCB controls all functions and is the core of the battery system. For example, it detects the signals of the rotary/push button and converts them depending on the current operating mode for the corresponding function control. In addition, the control unit board also handles all the lighting scenarios. Depending on the operating mode, these are differentiated between a visualization of the current support level, the current battery level, the charging animation during the charging process and the assisted opening/closing of the NFC lock. The Audi logo and the lock symbol on the battery system are the illuminated surfaces for the visualization. They are made of a partially transparent material, produced in silicon print and bonded to the 3D printed and black colored housing of the battery system. Furthermore, the control unit board inside the battery system controls the function of the charging and discharging process of the battery pack as well as the 5V charging of various end devices via the USB interfaces type A and type C on the bottom of the battery system. During the charging process of the battery pack, a current sensor on the control board analyzes the current charge state. The discharge process of the battery pack only takes place when the bob battery system is inserted into the holder on the

135

Entrance to an electrified last mile ecosystem project “bring your own battery” frame of the pedelec and if communication with the power section of the pedelec can be established. If communication is successfully established, this feedback is given by a separate communication pin, which is analyzed by the control unit board. Figure 5 shows the bottom of the battery system containing the two USB interfaces and the connector system for data and voltage transmission to the power section of the pedelec.

Figure 5: bottom of the bob battery system [4]

The active NFC module in the battery system is also activated by the control unit board for a defined period after the rotary/push button on the battery system has been pressed. After successful activation, the NFC lock on the pedelec can be unlocked by the bob battery system. Therefor the bob battery system, with the lock symbol flashing blue in this state, must be positioned to the NFC symbol on the NFC lock of the pedelec. A click is an acoustic feedback of the unlocked NFC lock for the user.

4 Results of the project The technical feasibility of the sharing concept has already been tested and successfully proven within the scope of a Prove of Concept (POC). Possibilities for changes to improve the design as well as the functions of the POC were identified and implemented in the second project phase - the pilot phase. Some examples are the materials used, the adapted control unit board of the bob battery system and the change from rear-wheel drive to front-wheel drive.

136

Entrance to an electrified last mile ecosystem project “bring your own battery”

Figure 6: bob system including the pedelec and battery system [4]

At the end of the second project phase, a very small fleet of pedelecs and bob battery systems with an innovative sharing concept will be available. This fleet will be used in a final project phase for a longer pilot phase at Audi AG in Ingolstadt to validate the concept.

5 Conclusion Emissions have serious effects on the climate and the environment. The earth is becoming ever warmer as more and more CO2 is released into the atmosphere. Electromobility counters this trend and, in contrast to combustion engines, does not emit CO2 while driving - provided that the production and energy of the batteries is based on renewable energies. Increasing urbanization is intensifying the situation in urban areas. Last mile concepts based on electric mobility can relieve the roads in these areas and thus reduce air pollution and emissions. In addition, the quality of life of the people can be increased. Stress, noise and delays caused by traffic bottlenecks can be significantly reduced. The innovative and user-oriented Last-Mile Sharing concept "bring your own battery" takes these aspects into account and thus represents an innovative and efficient solution approach. The pilot phase at the Audi location in Ingolstadt will show how userfriendly, efficient and robust the current bob functional models already are, and where there is potential for improvement for transferring them to a series product.

137

Entrance to an electrified last mile ecosystem project “bring your own battery”

6 Outlook There are already many last-mile concepts in the field of electromobility and micro mobility. The sharing concept based on the NFC module is one of the innovations in the bob concept with Audi AG. In addition to the longer pilot phase, a certification based on the existing certification analysis with recommendations for action must also be carried out. The last mile concept "bring your own battery" is initially intended for internal Audi use. This concept will expand company mobility by providing a time-efficient and innovative LEV system.

References 1. https://www.infineon.com/cms/de/discoveries/elektromobilitaet/ 2. https://www.t-online.de/auto/recht-und-verkehr/id_83858914/in-diesenstaedten-fahren-bereits-fast-keine-autos-mehr.html 3. https://www.autobild.de/artikel/audi-e-tron-scooter-15627631.html 4. Audi AG

138

How accurate can a range calculation of an electric vehicle be? Lisa Braun ([email protected]) EvoBus GmbH, Mannheim

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_16

139

How accurate can a range calculation of an electric vehicle be?

Abstract The accurate remaining range calculation for electric vehicles is crucial in reducing range anxiety. Thus, a reliable method to assess the quality of the used remaining driving range (RDR) calculation is required, which allows us to take into consideration multiple trips, ensuring RDR accuracy on all given environment conditions. Both the evaluation method and the evaluation criteria play a key role in finding an accurate RDR calculation. We extend a previous method, combining the Mean Absolute Percentage Error with a quantile to take into consideration not only the best result but also several good results for each test trip. Instead of the previously used inverse driven distance as evaluation criteria we use a defined optimal RDR, which results from the present on board energy and the specific average consumption of the complete test trip. We can show that this new approach improves our assessment resulting to RDR calculations with a higher precision over a wide range of test data.

1 Motivation Replacing internal combustion engines by electric motor driven vehicles is crucial for improving air quality in cities, alongside with reducing greenhouse gas emissions, slowing down climate change [1]. Especially on public transportation, in particular the use of electrified city buses will play a key role [2]. Therefore, twelve major cities declared in 2017 that they would switch their entire bus fleet to zero-emission buses by 2025 [3]. Some German cities will follow until 2030 [4]. One of the main challenges for public transport operators to implement electric buses is the limited range compared to diesel buses. For this reason, an accurate range prediction is crucial for reliable operation. The operator must trust the estimated remaining driving range (RDR) to plan the operation accordingly avoiding unnecessary SOC buffers, preventing breakdowns. The correct RDR calculation could therefore be seen equivalent to an increase of battery capacity [5]. Thus, a correct RDR specification is a decisive competitive advantage, as a higher battery capacity increases both costs and weight. However, what is the correct RDR? While there are many approaches to calculate the RDR, the accuracy of calculated RDR is not always determined. In addition, methods of assessing accuracy are rarely discussed. Most commonly used methods for validation are absolute error (AE) or relative error (RE) of the start and end values , e.g. [5], [6], [7], followed by root means square error (RMSE) [8], mean absolute percentage error (MAPE) [9] and R-squared in data-based models [10]. Further, the available RDR depends highly on the present environment conditions, such as outside temperature, number of passengers, traffic density or route topology, as they affect the consumption and the usable battery capacity. Using an appropriate validation criterion may still not result

140

How accurate can a range calculation of an electric vehicle be? in a good RDR calculation, because the calculation may only match certain environmental conditions. Therefore, several test-drives with different operation scenarios should be used for validation to ensure that the identified advantages or disadvantages of an RDR calculation are not limited to certain operating conditions. In a previous work, we evaluated some of these validation criteria for an exemplary RDR calculation based on 25 test-drives of an electric city bus with a total distance of 3631 km [11]. This paper is an extension of this work. We extended the database to 186 test-drives with a total distance of 11,025 km while we replaced the formerly used self-created test-drives by actual operation data with realistic load and auxiliary use. Section 2 explains the challenges of RDR calculation on electric vehicles. Section 3 describes briefly the evaluation method described in [11] and used test data for the evaluation with that deriving to a proposed optimal range calculation based on challenges of Section 2. Section 4 uses the evaluation method, validates it with an exemplary RDR calculation and compares it to the discussed expansion of section 3. The final section concludes.

2 Challenges of the RDR calculation The range anxiety in electric vehicles results from the combination of long charging times, short drivable distances and range prediction errors [12], which due to the shorter driving distance have a higher impact compared to vehicles with internal combustion engine [5]. Therefore, a high accuracy is essential to reduce or avoid range anxiety. Further, there are two use cases for an electric bus, where we need the RDR. The operator first needs a precise RDR to plan the distribution of the bus fleet. Before the operation starts, it must be clear if the bus can complete the planned trip to avoid delays or cancellations due to breakdowns. Second, the driver needs a reliable information if the vehicle is capable of completing the planned trip. Therefore, both range prediction before starting the trip and range updating during the trip must be accurate. To avoid breakdowns, accuracy at the end of the available range is particularly important. The RDR depends on two factors: the usable on-board energy WBatt stored in the battery and the specific consumption Bd,aux per driven kilometer of the drive train and auxiliary consumers. 𝑅𝐷𝑅 =

,

(1)

In a vehicle with internal combustion engine, consumption depends on the operating conditions. In an electric vehicle, also the battery behavior and therefore the usable onboard energy changes with the conditions. Thus, to calculate the RDR reliably before the operation, both the energy demand and the battery behavior must be accurately predicted, e.g. [13], [14] or [15]. During the operation, the battery management system can

141

How accurate can a range calculation of an electric vehicle be? provide the present usable on-board energy and hence, only the consumption of the trip must be forecast. This nevertheless is challenging, as the energy consumption changes significantly under different environment conditions. For example, during winter season lack of waste heat leads to a considerably higher consumption for heating the passenger compartment when compared to vehicles with internal combustion engines [16], [17]. For proper operation, the lithium ion battery also requires heating or cooling. These energy consumers on a city bus are significantly higher than on a passenger car, as the door opening at every bus stop leads to a significant loss of thermal energy due to air exchange and cold passengers in winter or hot passengers in summer enter the bus and must be heated or cooled as well. Therefore, on hot summer or cold winter days the consumption of the auxiliaries can be significantly higher than the consumption of the drivetrain, while at mid temperatures it is much smaller. Another challenge in the on-board RDR calculation of an electric vehicle is the recuperation. During recuperation the present consumption of an electric vehicle reduces, but can be inverted. Depending on the calculation method, this can notably reduce the specific consumption while the stored energy is increased. Thus, the RDR can increase significantly, see equation (1). Especially in mountainous areas and depending on the used algorithm, these effects can significantly influence the dynamic of the RDR, see also [11]. The question is, would a driver rely on an ever-changing RDR value or would she / he rather expect a reliable information about whether, the wished destination could be reach or not. The question here is how to assess, if an RDR calculation leads to reliable results?

2.1 Criteria to evaluate the accuracy of the RDR Evaluations often use the measured inverse driven distance to evaluate the calculated RDR, e.g. [5], [11], [18]. With the exception of reference [11], all found references combine this approach using AE or RE evaluating the change in RDR prediction compared with total driven distance of a test vehicle. Having the RDR estimation before the trip begins; this seems to be an appropriate evaluation method, as it is irrelevant for the operator, if the RDR becomes more accurate during a trip. If the vehicle is already planned on a specific route and is not capable of completing, this derives into changes in the operation and higher effort for the operator. However, in [11] we already explained why it is not sufficient to use the AE or RE for RDR calculations during a trip, as it does not evaluate the dynamic behavior of the calculation during the trip. Further, the evaluation with the driven distance during the operation assumes that the RDR is accurate, if it decreases by one kilometer for each driven kilometer. In this case, the RDR cannot be accurate during recuperation, when it increases due to additionally charged energy although the vehicle continues to increase its driven distance. What is more, if the test drive does not completely discharge the

142

How accurate can a range calculation of an electric vehicle be? battery, it is not clear what the actual driving range would have been [11]. Further corrections are required, if the inverse driven distance is used to assess the RDR. Therefore, to evaluate on board RDR calculations the inverse driven distance might not be the most fitting evaluation criteria. Thus, some references evaluate directly the forecast of the energy consumption instead of the RDR; see for example [7], [9] and [19]. With this evaluation approach, energy recuperation is not an issue, as both the forecasted, and the measured consumption could increase or decrease during operation. However, even if consumption forecast is precise that does not necessarily reflects a reliable RDR calculation. We already explained that the present energy consumption could change significantly due to changing environment conditions. The evaluation of the energy consumption forecast therefore does not replace the RDR calculation, only adds to it. Another approach is to compare SOC behavior during the drive for evaluation; see for example [20] and [21]. This approach is also valid during recuperation. Further, the information about the SOC also combines the information of remaining energy stored in the battery and the energy consumption like the RDR. However, navigation normally relay on distances. Therefore, it is easier for the operator and the driver to understand a RDR value instead of a remaining SOC value. The SOC is sufficient for estimating, if it is possible to complete a trip, however is not very useful for planning another one as the information about how many kilometers the bus could cover with the remaining SOC is not available.

3 Methodology to evaluate the accuracy of the RDR As described in the previous chapter, there are different evaluation methods for the RDR, each with advantages and disadvantages. In summary, we need an evaluation criterion, which on the one hand is taking into consideration the specific characteristic of an electric drive such as the SOC behavior. On the other hand, it shall reflect what the operator and the driver would see as reliable RDR information. Something the inverse driven distance would deliver. In addition, we should be able to include all of our test drives in the evaluation with little effort, not just those with full battery discharge. Figure 1 to Figure 4 show four different example test-drives, T1, T2, T3 and T4, out of our 186 samples with different operating conditions and depth of discharge (DOD). Figure 2 and Figure 4 show charging sequences within respectively at the end of the trip. If we use the driven distance as evaluation criteria, we would need to delete the charging sequences for the evaluation, as they would falsify our results. Further, we can see that the relation between the reduction of SOC and the inverse driven distance changes from T1 to T4 due to different environment conditions such as the stated average speed vavg and average temperature TAvg. The low average temperature during T4

143

How accurate can a range calculation of an electric vehicle be? reduced the drivable range, whereas on T2 the middle average temperature and middle average speed lead to a particularly higher one. As a result, pinpointing where, on the y-axis, the inverse driven distance should be placed to be able to assess the calculated RDR, becomes very hard. Consequently, the question arises of the correct RDR of an electric vehicle, on the basis of which we can evaluate the accuracy of our RDR algorithm. From now on, we refer to this correct RDR value as the “optimal RDR”.

Figure 1: optimal RDR vs. 96 km trip distance for vavg: 12 km/h, TAvg: 20°C, DOD: 55%

Figure 2: optimal RDR vs. 226 km trip distance vavg: 16 km/h, TAvg: 21°C, DOD: 116%

Figure 3: optimal RDR vs. 74 km trip distance vavg: 18 km/h, TAvg: 32 °C, DOD: 43%

Figure 4: optimal RDR vs. 44 km trip distance vavg: 8 km/h, TAvg: 4 °C, DOD: 64%

Considering the charged on-board energy, the optimal RDR would be the distance a vehicle is able to drive until it completely discharged its battery. As a result, the optimal RDR at the start of a trip corresponds to the driven distance at the end, if the stored energy is completely drained. In retrospective, the RDR shown to an operator before a trip would be correct, if we use the correct specific average consumption and the real available battery energy of the trip for its calculation, see equation (2). 𝑜𝑝𝑡𝑖𝑚𝑎𝑙 𝑅𝐷𝑅 =

, ,

(2)

Therefore, approaches predicting the RDR often include both a vehicle model and a battery model to be able to forecast these two values, see e.g. [13] and [21]. The advantage of using the optimal RDR of equation (2) for evaluation instead of the driven distance is that we can evaluate the calculated RDR for every trip length. In contrast to

144

How accurate can a range calculation of an electric vehicle be? the driven distance, the y-axis values are clear due to the considered usable battery energy. What is more, the stored on board energy changes due to recuperation and thus the optimal RDR increases alongside, see Figure 1 to Figure 4, solving the previously discussed problem of the inverse driven distance as evaluation criteria during recuperation phases of the vehicle. The same applies for the charging process in Figure 2. Thus, for the evaluation of the RDR we can also use trips with charging sequences as both the calculation and the evaluation value change with the on-board energy. In summary, the optimal RDR combines the advantages of the SOC evaluation and standard unit of kilometers used in navigation. Thus, the most accurate calculation that is possible is the RDR calculation, which is closest to the described optimal RDR.

3.1 Proposed evaluation method for RDR algorithms In [11] we proposed an evaluation method to assess the quality of the RDR calculation. As we do not only focus on the RDR calculation for an operator before a trip, but also on the RDR calculation for driver’s display during the trip, we defined a method to evaluate not just one RDR value, but a forecast curve. Further, the evaluation method should enable the assessment of a large number of test data. We found that both MAPE and RMSE showed reliable results when evaluating the accuracy of such a RDR calculation. However, since not just one but several test-drives were used for the evaluation, the x-quantile of the best results of each test drive was used for the evaluation instead of only taking the minimum value per trip into account. This results in the RDR algorithm with the highest occurrence of optimal values as the most accurate calculation method. Therefore, this evaluation method enables us to find the RDR calculation with the highest robustness against the influence of different operating conditions, when using enough operating condition variations in the evaluation data. We now propose to exchange the used inverse driven distance as evaluation criteria through the defined optimal RDR of equation (2). Further, as the RMSE and MAPE have shown similar results we reduce the evaluation to the MAPE only as it puts a heavier penalty on positive errors [22], which we see positive because too low RDR values do not cause any breakdowns.

3.2 Test data used for the evaluation The full electric bus eCitaro was used on all test-drives [23]. We extended the test data used in [11] from 25 test-drives with 3631 km to 186 test-drives with 11,025 km. Further, all test-drives without load or auxiliary use were replaced, as they do not represent realistic operation scenarios. In contrast to the test-drives described in [11], the highest measured average speed during a trip is now reduced to 34 km/h and the lowest average

145

How accurate can a range calculation of an electric vehicle be? temperature is 4 °C. In total, the trip distance varies from 15 km to 227 km, the temperature ranges from -0.7 °C to 40.5 °C and the highest measured speed is 66.4 km/h. Thus, our test data is lacking very low temperatures and express bus lines with higher speeds e.g. over motorways. Table 1 to Table 3 cluster the considered trips due to their specific characteristics. Table 1 shows the number of trips within a defined distance cluster. Notice that most of the trips recorded are well below 100 km. The average values considered for the average temperature cluster TAvg in Table 2 and average speed vAvg cluster in Table 3 refer each to one trip. Both clusters sum up the distances of the trips within these clustered value ranges. The data shows that most of the trips recorded are in a typical temperature range for Central Europe and have the average speed footprint of inner-city operation that is typical for a city bus. Table 1: Number of test-drives within different distance clusters Cluster

1

2

3

4

5

s [km] no [-]

15-50 86

50-100 81

100-150 14

150-200 1

>200 4

Table 2: Distance driven during trips within different average temperature clusters Cluster

1

2

3

4

5

6

7

TAvg [°C] s [km]

0-5 206

5-10 1,234

10-15 2,296

15-20 2,976

20-25 2,293

25-30 976

>30 1044

Table 3: Distance driven during trips within different average speeds Cluster

1

2

3

4

5

6

7

vAvg [km/h] s [km]

0-5 43

5-10 1,044

10-15 2,117

15-20 6,665

20-25 1,046

25-30 77

>30 33

146

How accurate can a range calculation of an electric vehicle be?

4 Validation of the evaluation method For the validation of the evaluation method, we created with an example algorithm 36 different RDR calculations for each recorded trip. Therefore, we changed four different parameters of this example RDR calculation: parameter A and C have three variations while parameter B and D have two variations. Figure 5 and Figure 6 show the best resulting RDR calculation for T3 respectively T2 considering the proposed MAPE to the optimal RDR (MAPE-opt), the previously used MAPE to the inverse driven distance (MAPE-dist) and the AE to the optimal RDR. For T3 in Figure 5 both RDR calculations with minimum MAPE are the same and very close to the optimal RDR while the best RDR calculation due to its AE has especially at the beginning of the trip a higher dynamic than the optimal RDR. This shows the first disadvantage when considering the AE: the dynamics of the calculation throughout a drive are not evaluated. Nevertheless, this is important if the RDR found shall be as precise as possible. The second disadvantage is that the placement in the y-axis and thus the shown value of the RDR is not taken into consideration but only its change. While the change could be correct, the values must not. Thus, we recommend that also for use cases where only the RDR at the beginning of a trip would be of interest, the complete RDR curves are evaluated to find the best calculation approach as the AE is lacking significant information for deciding if a calculation is sufficiently accurate.

Figure 5: RDR curves of calculations with minimum MAPE vs optimal RDR (MAPE-opt), minimum MAPE vs. inverse driven distance (MAPE-dist) and minimum absolute error (AE) of T3

Figure 6: RDR curves of calculations with minimum MAPE vs optimal RDR (MAPE-opt), minimum MAPE vs. inverse driven distance (MAPE-dist) and minimum absolute error (AE) of T2

147

How accurate can a range calculation of an electric vehicle be? T2 in Figure 6 includes two charging sequences and while the RDR calculation with minimal MAPE-opt is still very close to the optimal RDR as well as the minimum AE curve, the RDR calculation with a minimal MAPE-dist is not. It shows not only much higher RDR values than the optimal RDR but also a much higher dynamic throughout the test drive. The problem of the MAPE-dist is the considered charging sequences. The RDR calculation has a minimum MAPE-dist when it is closest to the inverse driven distance, which decreases over the complete evaluation. However, the considered charging processes mean that the RDR curve is alternately above and below the distance. Therefore, the curve with high values is optimal, since it is close to the continuously decreasing inverse driven distance for a long time. Figure 7 and figure 8 show changes in the MAPE over the 36 conducted RDR calculations for the previously discussed test-drives T1, T2, T3 and T4. The MAPE-opt in figure 7 shows a distinct pattern relating high and low MAPE values to different parameter combinations for T2, T3 and T4. In contrast, the evaluation with MAPE-dist only shows for T3 a similar pattern. Both T2 and T4 include charging sequences and thus the discussed problems occur while assessing with the inverse driven distance.

Figure 7: MAPE-opt of RDR calculations

Figure 8: MAPE-dist of RDR calculations

Figure 9: MAPE-opt occurrence of the best three calculations of each test-drive (Q=0.83)

Figure 10: MAPE-dist occurrence of the best three calculations of each test-drive (Q=0.83)

148

How accurate can a range calculation of an electric vehicle be? Figure 9 and Figure 10 show the corresponding occurrence of the three best MAPE values of all test-drives, so the quantile is 0.083. As verified in our previously conducted study, the MAPE-dist shows a significant high occurrence of optimal values for parameter B1, especially in combination with A3. In contrast, the highest occurrences of optimal values for the MAPE-opt is for different parameter A1 combinations. The best MAPE-dist combination of A3 & B1 does not show a high occurrence for the MAPEopt. Therefore, the new evaluation with the optimal RDR leads us to different RDR calculations as optimal results. In our case, instead of the parameter combination A3 & B1 & C3 & D1/2, the best RDR calculations are obtained with the parameter combination A1 & B1 & C1/2/3 & D1. Thus, not only the final combination changes but also the parameters, which have a higher influence on the found accuracy of the RDR calculation. We can explain this effect with the differences in the two evaluation criteria. The continuously decreasing inverse driven distance leads to parameter combinations, which reduce the dynamic behavior of the RDR curve to follow this reduction as close as possible. As the optimal RDR follows the dynamic of the SOC and thus changes with higher dynamic, it leads to RDR algorithms with parameter combinations allowing a faster change of the RDR curve. We already showed in figure 5 and figure 6 that the found RDR calculations are either similar to the ones found using the previous method or the accuracy is even higher. Consequently, exchanging the inverse driven distance by optimal RDR as evaluation criteria shows the expected results. Figure 11 to figure 16 show the MAPE-opt occurrence for selected cluster defined in section 3.2. Here is possible to notice that in all six figures the MAPE-opt occurrence is high for parameter A1 especially combined with parameter B1. For parameter C and D there is no obvious distinction, of which combination would be the best choice. Therefore, there is no difference in the found most accurate RDR calculation dependent on the defined environment conditions: average temperature range, average speed range or trip distance.

Figure 11: MAPE-opt occurrence (Q=0.83) for all test-drives with vavg below 10km/h

Figure 12: MAPE-opt occurrence (Q=0.83) for all test-drives within vavg of cluster 6

149

How accurate can a range calculation of an electric vehicle be?

Figure 13: MAPE-opt occurrence (Q=0.83) for all test-drives with Tavg below 10°C

Figure 14: MAPE-opt occurrence (Q=0.83) for all test-drives within Tavg of cluster 6

Figure 15: MAPE-opt occurrence (Q=0.83) for all test-drives with distances below 50km

Figure 16: MAPE-opt occurrence (Q=0.83) for all test-drives with distances within cluster 2

5 Conclusion It is expect that Electrified city buses will play a key role on improving the air quality in cities and on the reduction of greenhouse gas emissions. One of the main challenges for their deployment is their limited range compared to diesel buses. Thus, an accurate range prediction is crucial for both the operation planning as well as the reliable operation without breakdowns. In a former work, we proposed an evaluation method for assessing the quality of the RDR calculation. The MAPE between calculated RDR and inverse driven distance of the bus during different test-drives proofed to be a good criterion for the evaluation. As we wanted to take into consideration a high number of test-drives, we extended the method considering not only the algorithms with minimum MAPE, but also the x-quantile of the smallest MAPE values for each test drive. Thus, we are able to find a RDR algorithm with a high robustness against the influence of different operating conditions. However, the inverse driven distance lacks the information about what the real driving

150

How accurate can a range calculation of an electric vehicle be? distance would have been, if the battery was not completely discharged during the test drive. Further, it does not increase during recuperation or charging sequences but decreases or stays constant so during these operations there is for sure an error between the RDR and the evaluation criteria, which can lead to high MAPE values. In retrospective, the accurate RDR results from the charged on-board energy and the total consumption a bus had during its test drive. Thus, it is possible to calculate an optimal RDR curve with the correct specific average consumption and available battery energy. The highest achievable accuracy of a RDR algorithm is achieved, if it corresponds to this optimal RDR. Thus, this optimal RDR curve now replaces the inverse driven distance in the evaluation method. We evaluated this new approach with 186 test-drives with 11,025 km driven distance. The results show that our method allows on the one hand, to quickly identifying parameter dependencies of a RDR algorithm and, on the other hand, to find a reliable, robust RDR calculations.

Bibliography 1. Bundesministerium für Verkehr und digitale Infrastruktur, “Sofortprogramm Saubere Luft 2017-2020”, 2017. [Online]. Available: https://www.bmvi.de/Shared Docs/DE/Artikel/G/sofortprogramm-saubere-luft-2017-2020.html. [Accessed: 02.01.2020] 2. Bundesministerium für Wirtschaft und Energie, “Energieeffizienzstrategie 2050”, 2019. [Online]. Available: https://www.bmwi.de/Redaktion/DE/Publikationen/ Energie/energieeffiezienzstrategie-2050.html [Accessed: 02.01.2020] 3. Global Covenant of Mayors for Climate & Energy, "Our Commitment to Green and Healthy Streets Fossil-Fuel-Free Streets Declaration," 2017. [Online]. Available: https://www.c40.org/other/green-and-healthy-streets. [Accessed: 02.01.2020] 4. H. Hielscher, "Vier Städte wollen bis 2030 auf Elektro-Busse umstellen", WirtschaftsWoche, 01.11.2018 5. J. Hong et. Al., "Accurate Remaining Range Estimation for Electric Vehicles", 21st Asia and South Pacific Design Automation Conference (ASP-DAC), 2016, China 6. S. A. Birrell et. Al., "Defining the accuracy of real-world range estimations of an electric vehicle", 17th International Conference on Intelligent Transportation Systems (ITCS), 2014, China 7. K. Gebhardt et. Al., "Applying stochastic methods for range prediction in E-mobility", 15th International Conference on Innovations for Community Services (I4CS), 2015, Germany

151

How accurate can a range calculation of an electric vehicle be? 8. M. Kubicka et. Al., "About prediction of vehicle energy consumption for eco-routing", 19th International Conference on Intelligent Transportation Systems (ITSC), 2016, Brazil 9. S. Grubwinkler et. Al., "Range Prediction for EVs via Crowd-Sourcing", Vehicle Power and Propulsion Conference (VPPC), 2014, Portugal 10. F. Foiadelli et. Al., "Energy Consumption Prediction of Electric Vehicles Based on Big Data Approach", International Conference on Environment and Electrical Engineering and Industrial Commercial Power Systems Europe (EEEIC / I&CPS Europe), 2018, Italy 11. L. Braun and T. Nagel, "Methods to evaluate the quality of a remaining range algorithm," 32nd Electric Vehicle Symposium (EVS32, 2019), France 12. S. Heath et. Al., "Do you feel lucky? Why current range estimation methods are holding back EV adoption", IET Hybrid and Electric Vehicles Conference 2013 (HEVC 2013), 2013, UK 13. J. A. Oliva et. Al., "A Model-Based Approach for Predicting the Remaining Driving Range in Electric Vehicles”, Annual Conference of the Prognostics and Health Management Society, 2013, USA 14. H. Rahimi-Eichi and M.-Y. Chow, "Big-Data Framework for Electric Vehicle Range Estimation”, 40th Annual Conference of the IEEE Industrial Electronics Society, 2014, USA 15. S. Sautermeister et. Al., "Influence of Measurement and Prediction Uncertainties on Range Estimation for Electric Vehicles”, Transactions on Intelligent Transportation Systems, Vol. 19 Iss. 8 (2018), 2615-2626 16. L. Horrein et. Al., "Impact of Heating System on the Range of an Electric Vehicle”, Transactions on Vehicular Technology, Vol. 66 Iss. 6 (2017), 4668-4677 17. P. Iora and L. Tribioli, "Effect of Ambient Temperature on Electric Vehicles’ Energy Consumption and Range: Model Definition and Sensitivity Analysis Based on Nissan Leaf Data”, World Electric Vehicle Journal, Vol. 10, Iss. 1 (2019) 18. C. Pan et. Al., "Driving range estimation for electric vehicles based on driving condition identification and forecast”, AIP Advances, Vol. 7, Iss. 10 (2017), 105206 19. S. S. Sonalikar and S. D. Shelke, "Estimation of Remaining Range of Electric Vehicle Using Kalman Filter”, International Conference on Inventive Research in Computing Applications (ICIRCA 2018), 2018, India

152

How accurate can a range calculation of an electric vehicle be? 20. K. Sarrafan et. Al., "Accurate range estimation for an electric vehicle including changing environmental conditions and traction system efficiency”, IET Electrical Systems in Transportation, ISSN 2042-9738, Vol.7 Iss. 2 (2017), 117-124 21. S. Pai and M. R. Sindhu, "Intelligent driving range predictor for green transport”, IOP Conference Series: Materials Science and Engineering, Vol. 561 (2019) 22. J. Armstrong and F. Collopy, "Error measures for generalizing about forecasting methods: Empirical comparisons”, International Journal of Forecasting, Vol. 8, Iss.1 (1992), 69-80 23. „Der neue eCitaro”, 2018. [Online]. Available: https://www.mercedes-benzbus.com/de_DE/models/ecitaro.html. [Accessed: 02.01.2020].

153

Testing in times of big data and machine learning Hendrik Bohlen, Paul Assendorp Werum Software & Systems AG

This manuscript is not available according to publishing restriction. Thank you for your understanding.

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_17

155

Testing ADAS end-of-line – Avoid the hazardous effects Frank Heidemann, SET GmbH

This manuscript is not available according to publishing restriction. Thank you for your understanding.

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_18

157

Consumption-relevant load simulation during cornering at the vehicle test bench VEL Dr.-Ing. Martin Gießler, M. Sc. Philip Rautenberg, Prof. Dr. rer. nat. Frank Gauterin Karlsruhe Institute for Technology (KIT)

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_19

159

Consumption-relevant load simulation during cornering at the vehicle test bench VEL

1 Introduction Standardized driving cycles (e.g. WLTC) define speed points for a straight driving road load calculation in simulation and on roller chassis dynamometers. Within real rides on public roads, straight driving road loads are increased by curve resistances during curve driving. Furthermore, for curve driving the wheels has to be steered, so that the steering system needs to be actuated to overcome the necessary steering torque at the steered wheels. This extra loads increase the energy or fuel consumption and the resulting emissions especially during rides on curvy country or urban roads. Therefore, curve loads have to be considered when developing new inner city vehicles for delivery, public transport or individuals. Trends like boom of online trade with increased parcel delivery, urbanization, expansion and modernization (electrification) of local road bound public transport and last but not least emission standards increase the need to consider curve loads within the vehicle development process. Within this publication the authors present the theory on consumption relevant curve loads. This loads have been simulated on the full vehicle test bench VEL (Vehicle in the Loop) of the Institute of Vehicle System Technology at the Karlsruhe Institute of Technology (KIT). Selected test results on the influence of curve loads performed on the VEL are finally presented.

2 Consumption relevant loads during curve driving The curve resistance A vehicle’s drivetrain delivers the drive torque needed to overcome the driving resistances caused by operation-depending rolling resistance of the tires [1], the air drag, the inclination resistance, losses in the drivetrain and the force to pull a trailer and accelerate the vehicle. Additionally slip losses when transmitting torque from the tire to the road [2] and the side force and slip angle depending curve resistance [3] have to be taken into account for more precise calculation of road load. Since the wheel motion during loading and unloading depend on the installed chassis and its kinematic, the slip and the slip angle will also change with wheel motion. This complex interactions and the tire force characteristic are mapped within the vehicle simulation environment (CM) used for the later presented results. During curve driving a lateral acceleration caused by centrifugal force will act at the vehicles center of gravity. The wheels are running under a slip angle  towards the

160

Consumption-relevant load simulation during cornering at the vehicle test bench VEL tangential speed v in Figure 1 to develop side forces FY at the tires, which cosine component carry these centrifugal force. The sine component Eq. (1) describes the curve resistance.

Figure 1: Curve driving resistance FR, (without caster and pneumatic trail) caused by side force Fy of a wheel rotating with slip angle  to the speed v.

𝐹

,

= 𝐹 ∙ 𝑠𝑖𝑛 𝛼

Eq. (1)

The curve resistance force divided by the wheel load FZ give us the possibility to describe a wheel load independent increase of rolling resistance caused by the curve resistance as ∆𝑓𝑅𝛼 : ,

=

∙ 𝑠𝑖𝑛 𝛼



∆𝑓

= 𝜇 ∙ 𝑠𝑖𝑛 𝛼

Eq. (2)

The average lateral friction force coefficient µY to be transmitted in average by each wheel (tire-track-contact) can be calculated balancing the acting forces: 𝑚 ∙ 𝑣 ⁄𝑅 = 𝑚 ∙ 𝜇 ∙ 𝑔



𝜇 =𝑣 ⁄ 𝑅∙𝑔

Eq. (3)

The lateral slip angle can be determined with a wheel load independent cornering stiffness of the tire within the linear progression up to the maximum lateral transmittable force coefficient, Eq. (4). The increase of rolling resistance by curve resistance can then described with Eq. (5).

161

Consumption-relevant load simulation during cornering at the vehicle test bench VEL

Figure 2: Typical progression of side force coefficient µY over slip angle  and the slop of the linear increase defined as wheel load normalized cornering stiffness Cµy.

𝛼 = μ ⁄𝐶 ∆𝑓

Eq. (4)

= 𝑣 ⁄ 𝑅 ∙ 𝑔 ∙ sin 𝑣

𝑅∙𝑔∙𝐶

Eq. (5)

100% 50% 25% 13% 6% 3% 2% 1%

single freq. cummaltive

100

Frequency

Within a master thesis at KIT in Cooperation with Fraunhofer ICT-NAS [4] a ride around Karlsruhe was defined and driven that meets the requirements of an RDE (Real Driving Emissions) track according to the regulation EU 2016/ 427. The GPS data of this trip allows us to determine the curve radii, [5]. The analysis of curve radii clarifies that 13% of the driven curve radii had been smaller than 60 m, Figure 3. Based on this determined curve radii, for this RDE ride a significant portion of curve resistance coefficient of 30% had been evaluated, Figure 4.

Curve Radius [m] Figure 3: Left: Frequency of distribution of curve radii. Right: RDE Ride starting from Karlsruhe via Wörth am Rhein to Speyer and back.

162

Consumption-relevant load simulation during cornering at the vehicle test bench VEL

𝑎

,

(2)

3.6 Weighting, aggregation and decision making Applying the evaluation metrics in section 3.5 leads to aggregated results for each route. In order to reduce these route-based results to a more manageable consolidated variable, the results are aggregated separately for each metric. This aggregation can be performed locally in smaller regions or on a bigger scale, e.g. a whole country or continent. Generally, the routes generated in section 3.2 try to cover as much of the road network as possible while avoiding simulating the same roads several times. This approach represents the importance of the respective road only indirectly by searching for the shortest route in time and therefore favoring bigger roads like arterial roads and freeways. In order to include a measure for the probability of a car equipped with an automated driving function driving on a specific road, the weighting step is included in the concept.

200

Method for a scenario-based and weighted assessment of map-based advanced … The final step of the concept is using the derived and aggregated information about the usability of a driving function with a specific feature set and a defined operating range to decide about which features genuinely add value to the driving function. Combined with the other aspects mentioned in section 2, valid decisions can be made based on objective measures.

4 Evaluation with an exemplary implementation In order to proof the applicability of the concept, we present an exemplary implementation using a PCC with the objective to assess a specific feature set. As test region we choose Western Europe. It is vast enough to cover different characteristics of road networks, while still being sufficiently small enough to perform the simulations in a reasonable timescale. However, because of the prevalent left-hand traffic, the United Kingdom and Ireland are not included in the model to avoid additional complexity of the underlying traffic rules.

4.1 Route and environment generation To show the advantages of the route generator in terms of parametrization, two different routing scenarios are compared, for which the parameters and characteristics can be found in table 1. Scenario 1 (S1) is supposed to represent shorter commutes of around 30 km length, while S2 aims at longer travels of around 270 km route length. The histograms in figure 3 show the resulting route lengths for both scenarios. Depending on the road network characteristic between the origin and destination points, the routes can be significantly longer than the theoretical minimum, given by the circle’s radius. If the map database cannot provide the entire road attribute for the generated route, only the longest route section with available data is taken. That explains the existing routes with a length smaller than the original radius. Table 1: Parameters and characteristics for the routing generator Scenario Radius S1 S2

𝛿

20 km 𝜋/4 200 km 𝜋/8

Overlap 0.25 0.25

Distance share of freeway links 11% 86%

Link uniqueness 78% 64%

Link coverage 35% 6%

With increasing radius, the area of the circle rises by the square of the radius. In order to maintain a high link coverage as defined in section 3.2, the angular resolution for the second scenario is raised. Since the second scenario aims at modelling longer travels, it

201

Method for a scenario-based and weighted assessment of map-based advanced … comes with a high share of freeway roads in terms of travelling distance. The link coverage is relatively small for S2, since the freeway roads only represent a small portion of the overall road network. For the generated routes, the link- and node-related road attributes are queried from the map data stored in a PostgreSQL database. From the link sequence, traces for the attribute variables are generated. Since the link shape points have varying spatial sampling rates, the traces are interpolated to a constant sampling rate of five meters.

Figure 3: Route length distributions for both scenarios

4.2 Simulation and Evaluation The equidistant traces are input for the simulation in the next step. In order to demonstrate the feasibility of our method, we model a simple predictive control with focus on longitudinal features. The function’s features address the use cases speed limits, curves, turnings, roundabouts, stop signs and yield signs, which means that the vehicle’s speed is adjusted to drive comfortably and abiding by the law through these use cases. All of these use cases present position-dependent upper speed constraints. From the use case speed limit, a very simple constraint 𝑣 , (𝑠) is derived using the identity function. Another constraint is calculated from the road curvature, when assuming a maximum in curves, given by [12]. target lateral acceleration 𝑎 , 𝑣

,

(𝑠) =

𝑎

,

⁄𝜅(𝑠)

(3)

For turnings, a maximum 𝑣 , (𝑠) speed is calculated based on a simple characteristic line depending on the turning angle. Having defined a speed constraint for curves and turnings, a roundabout can be described as a combination of those two situations. For stop and yield signs, a simple target velocity is defined. The resulting velocity constraint is calculated as the minimum of all constraints at each position 𝑣 (𝑠) = 𝑚𝑖 𝑛 𝑣

202

,

(𝑠), 𝑣

,

(𝑠), 𝑣

,

(𝑠), 𝑣

,

(𝑠), 𝑣

,

(𝑠) .

(4)

Method for a scenario-based and weighted assessment of map-based advanced … Based on this velocity constraint, a maximum speed is calculated, provided that the and 𝑎 , . longitudinal acceleration 𝑎 does not exceed the defined thresholds 𝑎 , 𝑎

,

≤𝑎 ≤𝑎

,

(5)

An example is shown in figure 4, in which the velocity constraints are depicted with black solid lines. For this application of a predictive longitudinal control, two main evaluation metrics are introduced that are supposed to represent the usability of the automated driving function. The first metric is a measure for the takeover-rate of the function, which we define as the number of potentially necessary driver takeovers per unit time driven. Takeover-Rate =

#

(6)

For the given feature set, the driver has to take over control in situations where the vehicle has to yield to other traffic, including traffic lights, stop signs, yield signs and other intersection situations, for example a left turning with oncoming traffic. Since it is difficult to determine in how many cases traffic is present indeed, a worst-case scenario is assumed where the driver has to interfere at each of these situations to give way to other vehicles. Even though there is a feature for slowing down at a stop sign, the driver is still responsible for checking the right-of-way situation. This conservative approach will result in a higher takeover-rate than expected in real-world testing. However, it makes it possible to compare different regions on a relative scale.

Figure 4: Example for the longitudinal control based on velocity constraints

Experience constitutes the second metric, measuring in how many situations the driver actively perceives the function controlling the speed beyond keeping a constant pace like a conventional ACC. This metric can be thought of as measuring the value added by the predictive features compared to the ACC features.

203

Method for a scenario-based and weighted assessment of map-based advanced … Experience =

(7)

The road sections with a velocity change are shaded in figure 4. A curvy road can elicit different experiences depending on the speed limit. For a high speed limit, the driving function’s curve feature has to slow down before each curve and accelerate afterwards, which is quite perceptible for the driver, hence the function has a high experience on this road. However, if the speed limit is restrictive, curves can be taken while driving at the speed limit. In this case, a normal cruise control without any of those additional features would be sufficient. These two metrics are calculated for each route and saved in the database together with the geographical coordinates of the route center. The geographical reference can be used in the aggregation step to spatially merge the route results in a certain area.

4.3 Weighting and aggregation Since we have no access to a source of meaningful data that could help to weigh the routes separately depending on the number of vehicles that operate on it, we do not include the weighting in our exemplary implementation. However, one way to determine a link-dependent probability is to look at the overall traffic distribution in the road network. An estimate can be calculated using floating car data [13] that contain speed information combined with a position reference. Under the assumption that the share of vehicles which provide floating car data is approximately the same on the whole road network, the number of floating car observations per road link related to the overall number of observations could be used as a weighting factor. In order to make decisions based on the evaluation in the previous section, an aggregation of the route-individual results is helpful and can be carried out on different levels. The routes’ characteristics can be merged geographically, for example using map tiles as depicted in figure 5. The evaluation for the experience 𝑒 for all m routes that fall into a map tile k are aggregated using a weighted average 𝑒 =∑

𝑒 𝑙 ⁄∑

𝑙 ,

(8)

with the individual route length 𝑙 . Depending on the resolution of these map tiles, it is possible to observe local anomalies like shown in figure 5. For example, in regions with more curved roads like in the Alps, the experience is higher. On the other side, the takeover-rate appears to be higher in dense urban areas, as more intersections lead to a higher number of potential right-of-way-situations. However, in order to generate compact numbers, it is also possible to further aggregate the results to a country level or even to a single value for the whole investigated region. Table 2 shows the aggregated results for the regarded region in Western Europe.

204

Method for a scenario-based and weighted assessment of map-based advanced … Experience

65

60

55

50

45

40

-10

-5

0

5

10

15

Longitude

Figure 5: Map tile-based aggregation and visualization of the experience (left) and necessary takeover-rate (right)

Table 2: Aggregated results of the two investigated scenarios Scenario Experience S1 14.6% S2 4.60%

Takeover-Rate 18.9 1/h 6.02 1/h

Simulated road length 1,103,500 km 116,409 km

4.4 Decision-making The last step of the proposed method for evaluating automated driving functions is to derive decisions from the results provided in the steps before. For this exemplary implementation, the assessment of the usability is paramount. With different feature and parameter sets, the influence on the defined metrics can be used in order to make statements about the usability. Combined with other dimensions like costs, feasibility, strategic orientation and legal regulations, a decision on the function’s features can be made. From table 2 it can be seen that the driving function has a considerable higher experience in S1, but also a high takeover-rate. Focusing on freeway roads (S2), the experience and takeover-rate decrease, as there are a smaller number of use cases and fewer intersections with right-of-way-laws on these roads.

205

Method for a scenario-based and weighted assessment of map-based advanced … Looking at figure 5, the results can also be of help in other applications. The knowledge about the distribution of specific situations in a larger region supports the planning of real-world tests in order to ensure that the selected test routes are a good representation of the overall road network, since real-world tests often can only cover a small fraction of the basic population in a statistical sense. On the other side, it is possible to use the information about the scenario locations to determine specific testing routes with a high number of occurring situations. This approach comes with the advantage of a reduction of test kilometers while maintaining a high stimulation of the driving function.

5 Conclusion and Outlook In this work, we showed a new method for evaluating automated driving function concepts by means of a simulation using the link-layer of map data as input. An approach for route generation in order to create realistic trips while covering large parts of the road network was presented. The concept was exemplarily implemented with a predictive longitudinal control. For the evaluation of the usability of a driving function with a specific feature set, the metrics of takeover-rate and experience were derived. Although the approach is very scalable due to the availability of map information for the whole road network, the concept has a reduced significance at the same time. The map data only model parts of the static parameters. Many more inputs are considered by a driving function deployed in a vehicle. Additionally, map data can be partly incomplete and outdated, so it is important to use up-to-date maps. However, the presented approach provides objective indications about the regiondepending use case distribution, which is very useful in several stages of the development process. Future research can be carried out on optimized route generators, more detailed simulations and additional metrics for evaluation.

Bibliography [1] SAE International, „SAE J3016: Taxonomy and Definitions for Terms Related to Automated Driving Systems for On-Road Motor Vehicles“, in Surface Vehicle Recommended Practice, 2018. [2] G. Bagschik, T. Menzel, A. Reschka und M. Maurer, „Szenarien für Entwicklung, Absicherung und Test von automatisierten Fahrzeugen“ in 11. Workshop Fahrerassistenzsysteme. Hrsg. von Uni-DAS e. V, 2017. [3] S. Geyer, M. Baltzer, B. Franz, S. Hakuli, M. Kauer, M. Kienle, S. Meier, T. Weißgerber, K. Bengler, R. Bruder und others, „Concept and development of a

206

Method for a scenario-based and weighted assessment of map-based advanced … unified ontology for generating test and use-case catalogues for assisted and automated vehicle guidance,“ IET Intelligent Transport Systems, Bd. 8, pp. 183-189, 2013. [4] C. King, J. Bach, S. Otten und E. Sax, „Identifikation von Fahrszenarien während einer virtuellen Testfahrt,“ INFORMATIK 2017, Lecture Notes in Informatics, Gesellschaft für Informatik, Bonn, 2017. [5] J. Bach, S. Otten und E. Sax, „Model based scenario specification for development and test of automated driving functions,“ in 2016 IEEE Intelligent Vehicles Symposium (IV), 2016. [6] J. Bach, M. Holzäpfel, S. Otten und E. Sax, „Reactive-replay approach for verification and validation of closed-loop control systems in early development,“ SAE Technical Paper, 2017. [7] J. Bach, J. Langner, S. Otten, E. Sax und M. Holzäpfel, „Test scenario selection for system-level verification and validation of geolocation-dependent automotive control systems,“ in Engineering, Technology and Innovation (ICE/ITMC), 2017 International Conference on, 2017. [8] P. H. L. Bovy und S. Fiorenzo-Catalano, „Stochastic route choice set generation: behavioral and probabilistic foundations,“ Transportmetrica, Bd. 3, pp. 173-189, 2007. [9] P. H. L. Bovy, S. Bekhor und C. G. Prato, „Route sampling correction for stochastic route choice set generation,“ Transportation Research Board 88th Annual Meeting, 2009. [10] O. Damberg, J. T. Lundgren und M. Patriksson, „An algorithm for the stochastic user equilibrium problem,“ Transportation Research Part B: Methodological, Bd. 30, pp. 115-131, 1996. [11] F. Xie und D. Levinson, „Measuring the structure of road networks,“ Geographical analysis, Bd. 39, pp. 336-356, 2007. [12] M. Mitschke und H. Wallentowitz, Dynamik der Kraftfahrzeuge (VDI-Buch, 5., überarb. u. erg, Wiesbaden: Springer Fachmedien, 2014. [13] C. De Fabritiis, R. Ragona und G. Valenti, „Traffic estimation and prediction based on real time floating car data,“ in Intelligent Transportation Systems, 2008. ITSC 2008. 11th International IEEE Conference on, 2008.

207

Validation of fuel cell control units with powerful simulation platforms for fuel cells Michael Seeger ([email protected]), Abduelkerim Dagli ([email protected]) MicroNova AG, Unterfeldring 6, 85256 Vierkirchen

Validation of fuel cell control units with powerful simulation platforms for fuel cells

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_23

209

Validation of fuel cell control units with powerful simulation platforms for fuel cells

1 Initial situation Reducing the CO2 standard values through the EU is showing effect: Car manufacturers, cities and municipalities are looking for alternatives to vehicle propulsion systems in face of imminent fines. In this context, a study by the Roland Berger management consultancy notes a high demand for hydrogen technology. Hydrogen fuel cells are considered a key technology for reducing greenhouse gases that are harmful to the climate and human health. Such vehicles can boast of short refueling times and longer ranges, particularly in fleet operation and heavy load applications. The 89 European municipalities recognized this as well: over the next five years, it’s planned to invest up to 1.8 billion euros in city buses, cars and delivery vans with fuel cells, as well as in electrolysers for hydrogen production and combined heat and power plants. In the medium term, it’s planned to invest in appropriately equipped garbage trucks, trains and ships as well. However, this new drive technology requires sustainable and reliable protection1.

2 HiL Technology Test methods using hardware-in-the-loop (HiL) technology are one method of securing such drives. Hardware-in-the-loop refers to a procedure in which an embedded system (e.g. real electronic control unit or real mechatronic component, i.e. hardware) is connected to an adapted counterpart via its inputs and outputs. The latter is generally called HiL simulator and simulates the real environment of the system in the vehicle. From the perspective of the test, hardware-in-the-loop is a method for securing embedded systems, for support during development and for early commissioning of machines and systems.

1

See Claudia Russo, Head of Marketing & Communications Germany, Austria and Switzerland, Roland Berger GmbH, 30.11.2018

210

Validation of fuel cell control units with powerful simulation platforms for fuel cells A typical HiL test system usually includes the following components: ● A computing unit for calculating large and complex simulation models and for connecting the real-time models with the corresponding I / O. ● I / O interfaces for connecting the control unit to the test system. Different boards are used for different test scenarios, with dedicated or freely configurable I / O interfaces. There are also components such as load boards and the possibility of specific electrical fault simulation for various application scenarios. ● A simulation of bus and network communication to test communication between two or more ECUs. Not-yet existing network participants are simulated. ● One or more control units that can also be tested with new functions or control unit software. They are connected to the test system and are typically referred to as devices under test (DUT). ● Software for configuring and automating HiL testing. It runs on an operating PC, as does the software for parameterizing the simulation model and the visualization of the simulation.

3 Testing of fuel cell stacks using HiL MicroNova developed the “NovaCarts Fuel Cell” HiL simulator for testing fuel cell control units. NovaCarts Fuel Cell simulates the fuel cell stack and the environment of the associated control unit in the vehicle and can be expanded for future technologies of fuel cell control units by means of a firmware update. The versatile, scalable HiL system is suitable for the complete protection of new functions in control units for fuel cell stacks (Fuel Cell Control Units, FCCU). Its modular structure and extensive expansion options enable flexible adaptation to different test requirements (e.g. power emulation / powerless simulation). The parameters and controllers used for the respective simulation can be changed directly in the software, which eliminates the need for time-consuming hardware replacement. In addition, test engineers can quickly and easily adapt NovaCarts Fuel Cell to future requirements, such as new communication interfaces or updated HV architectures, through a firmware update. An open model platform with cycle times of a few microseconds and high I / O dynamics was used for the HiL simulator. This enables the development of new FCCU algorithms as well as the use of real parts, spare loads and residual bus simulation. In addition to the SAE standard J2799 for communication between the vehicle and the hydrogen filling station, resistance simulations can also be carried out for simulating temperature sensors with negative or positive temperature coefficients (NTC or PTC).

211

Validation of fuel cell control units with powerful simulation platforms for fuel cells The system behavior is supported by simulation models for fuel cells and thus enables a closed-loop test setup. It meets the high requirements of the complex control units: Using model simulation, the correct functioning of the FCCUS in development can be tested. These simulations also shorten development times and reduce costs at the same time. The simulation of recurring processes in particular has the advantage that new development versions can be tested under the exact same criteria as the previous versions.

Figure 1 NovaCarts FuelCell HiL

212

Validation of fuel cell control units with powerful simulation platforms for fuel cells In combination with the HiL system, "NovaCarts Battery" is also capable of simulating functions of the connected battery, such as state-of-charge (SoC) and state-of-health (SoH) controls and cell balancing mechanisms.

4 Combining HiL with established CVM-Systems In addition to the classic basic-HiL-system, the actual cells can and will be checked for testing the control unit. In addition to the use of real cells, a so-called cell simulation can also be used. When using real fuel cells, common cell voltage monitoring systems (CVM) monitor the fuel cell stack as a complete system. However, this has the disadvantage that, in the event of a malfunction, no statement can be made as to where exactly this occurs. Single cell monitoring, on the other hand, enables deep insights into the interior of the stack. Critical operating states are not only recognized, they can also be localized precisely. The requirements for a CVM system are complex. For stationary use on test benches, for example, high level of measurement accuracy and a high sampling rate are required. A modular measuring system with a measuring frequency of up to 1 kHz per measuring channel is required for this. Such a system usually consists of 1 to 42 ten-channel voltage detection modules, a link module for connecting the supply and communication lines and a bus termination module. Similar to the Lego principle, all modules can be plugged together without additional wiring, so that up to 420 channels and thus 420 individual cells of a fuel cell stack can be scanned fully synchronously. In order to investigate in detail how a fuel cell system reacts to changes in state such as a load change, the measurement accuracy must be sufficiently high. For applications with a low number of channels or lower sampling rates, data transmission via CAN bus is sufficient. For more complex tasks, there should also be an option to implement additional analysis functions without the need to integrate an additional PC. While the existing FPGA ensures that the serial data from the measuring module is made available for evaluation in a parallel data structure, microprocessors enable PCindependent data processing. An implemented realtime patch ensures that reactions are possible in real time.

213

Validation of fuel cell control units with powerful simulation platforms for fuel cells The NovaCarts Fuel Cell HiL for testing the control unit can be easily combined with common CVM systems due to used communication interfaces being consistently compatible.

5 Combining HiL with the cell simulation for one Battery Another possible combination is the use of a so-called cell simulation of batteries, which is a much simpler, more reproducible and process-reliable method. The modular HiL system NovaCarts Battery offers one of the most powerful and accurate cell simulations on the market. To complete the set-up, a so-called CMCe (Cell Management Controller) with cell simulation cards is installed in a second cabinet, which in addition to simulating individual cells can also represent the balancing of the cell network. The fully digital cell simulation is possible thanks to powerful field programmable gate arrays (FPGAs) on the NovaCarts Cell Simulation Board.

Figure 2 NovaCarts Cell Simulation Board

214

Validation of fuel cell control units with powerful simulation platforms for fuel cells In order to simulate even small deviations and differences in the individual cells of the stack, the highly precise mapping of current jumps and voltage dips is necessary. Since the software-controlled internal resistance of the cells is installed in this cabinet and can be changed quickly and in real time, the simulation of the cells is possible without any problems. In addition, the open and powerful model platform with cycle times of a few microseconds and high I / O dynamics enables the development of new FCCU algorithms such as B. State-of-Health (SoH), State-of-Charge (SoC), electrochemical impedance spectroscopy. Using chemical and electrochemical simulation models, the current state of the stack can also be calculated and displayed very precisely. The simulation of capacitive and inductive balancing mechanisms enables validation of FCCU with passive and active cell balancing. The high signal quality arises above all from short and stable connections to the control unit and from the fault simulation directly attached to the output.

Figure 3 NovaCarts Failure Insertion Board

215

Validation of fuel cell control units with powerful simulation platforms for fuel cells The error simulation card enables the following errors to be fed into the signal lines: ● interruption ● cross-circuit between signals ● short circuits to FailRail1 and FailRail2 The card also offers an additional connection for real loads and replacement loads. It is supplemented by a resistance simulation for simulating temperature sensors with a negative or positive temperature coefficient (NTC or PTC).

Figure 4 NovaCarts Resistor Simulation Board

Thanks to the high channel density of the resistor card, even HiL systems with many I / O’s can be implemented compactly and cost-effectively. Four channels are combined in a group with galvanic isolation of up to 1,000 V. This makes the card ideal for simulating temperature sensors, which are required for testing FCCU, among other things.

216

Validation of fuel cell control units with powerful simulation platforms for fuel cells The external shunt simulation module was specially developed for testing battery management systems and can also be used when testing fuel cell control units.

Figure 5 NovaCarts Shunt Simulation Module

It replicates the voltage of the entire battery very precisely. It can be installed in the immediate vicinity of the corresponding BMS measuring output to ensure the high accuracy required for shunt measuring resistors. Thanks to the high dynamics of more than ten kilohertz, the module is even suitable for demanding battery applications, such as the simulation of starter batteries or future BMS functions. In addition, automobile manufacturers and suppliers can also use the module to simulate special performance jumps that typically occur in the event of errors. Hence, the two test systems NovaCarts Fuel Cell and NovaCarts Battery can be used to test the entire string of fuel cell stack and drive battery.

217

Validation of fuel cell control units with powerful simulation platforms for fuel cells

6 Summary The HiL simulator "NovaCarts Fuel Cell" from MicroNova offers one of the most powerful platforms in this field and thus supports the development of alternative and electrified drives. The modular and scalable HiL system is ideal for securing new functions in control units for fuel cells. In combination with the HiL system "NovaCarts Battery", the functions of the connected battery can also be simulated, thus fulfilling the requirements for protecting the new drive technology.

218

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level J. C. Wurzenberger, T. Glatz AVL List GmbH, Austria D. Rašić, G. Tavčar AVL-AST d.o.o, Slovenia I. Mele, A. Kregar, T. Katrašnik University of Ljubljana

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_24

219

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level

1 Introduction Hydrogen-powered vehicles comprising low temperature proton exchange membrane fuel cells (LT-PEMFC) appear to be a viable, emission free, solution characterized by short refueling times and long driving ranges. However, additional optimizations are required on a component and system level to achieve further cost reduction while ensuring adequate service life and sufficiently high energy conversion efficiencies to boost the market shares of fuel cell electric vehicles (FCEVs). Frontloading is an effective measure to shorten product development cycles and to reduce development costs. Frontloading is particularly important in the development process of FCEVs, which are inherently fuel cell hybrid vehicles, comprising two electrochemical devices. The fuel cell (FC) covers lower frequency power demands while higher frequency power demands are covered by a battery. Since the FCEV powertrain features a complex component topology comprising several physical domains (i.e. mechanical, electrical, fuel cell system, battery system, thermoregulation and controls), system layout and system level optimization are key development steps. The development of an efficient, right-sized, durable and cost optimized FCEV represents a significant system level challenge, which calls for the application of advanced system simulation tools to efficiently explore the vast design space of material, geometry and operation parameters. Electrochemical devices, fuel cell and battery, are prone to degradation during their lifetime. Aging phenomena need to be considered already during initial system layout studies when targeting to approach engineering limits. Traditionally, system level vehicle models rely on empirical models for fuel cell and battery, intended to replicate component performance characteristics. Empirical models feature several limitations. They do not have any prediction capability and need to be trained on measured data, which are rarely available in early development stages. Empirical models are trained on limited datasets, while optimization studies aim to explore the system outside of the trained regions. This is specifically pronounced in transient operating regimes of electrochemical devices. Real electrochemical systems feature multiple interactions on interand intra-component level. This influences performance and degradation, which, in general, cannot be captured by empirical models in a way to preserve full causality of the underlying phenomena. This is crucial for the assessment of degradation phenomena, which are triggered by local values of potentials, concentrations and temperatures, referred to as degradation stimuli. Degradation phenomena feature complex interrelations of various degradation mechanisms occurring at different rates and at different values of degradation stimuli.

220

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level This paper attempts to answer the above listed challenges by presenting a paradigm shift in system level modeling of FCEVs. This is achieved with a multi-domain modeling framework combining models of different physical domains featuring different scales, frequencies and numerical requirements. The core of the innovative approach are spatially and temporarily resolved, mechanistically based, models for fuel cell and battery. The models are fully integrated into the overall FCEV model running in transient operating conditions. The models predict the electrochemical behavior based on their inherent design and material properties being one of the key merits for early stage development. Spatially and temporarily resolved degradation stimuli are crucial for a proper prediction of local degradation rates. The system level modelling approach is demonstrated by replicating the behavior of a commercial FCEV, the Toyota Mirai. Validation simulations confirm the reliability of the proposed modelling framework on the system level. The interaction of components and domains is shown by a credible replication of essential causalities. Spatial and temporal resolved results of fuel cell and battery reflect the impact of transient duty cycles. Local growth rates of the solid electrolyte interphase on the anode side of the battery are modeled. In the fuel cell, the rate of carbon corrosion and the growth rate of platinum particles at the cathode side are assessed and demonstrate the merits of the proposed multi-domain approach.

2 Model The following sections discuss a system level simulation framework, followed by a base description of the battery and fuel cell models including a discussion on the applied degradation models.

2.1 Multi-Physical Modeling Framework The simulation of FCEVs on system level requires modeling of several different physical domains. Figure 1 shows a schematic of the main domains and how they interact. A network of rotational mechanical degrees of freedom (1) describes the transport of mechanical power from the e-motor to the wheels. An electrical network (2) models the transport of electrical power, generated in the fuel cell, to the e-motor, the battery and other consumers. A compressible gas-path network (3) describes the transport air and hydrogen to the fuel cell. A thermal network (4) of lumped masses describes the conductive transport of energy from the-motor, battery, fuel cell to heat sinks represented by the gas or liquid cooling systems. An incompressible liquid cooling network (5) transports heat via radiators to the environment. An information flow network (6) controls the plant model in an open and closed loop configuration.

221

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level A multi-domain simulation environment, implemented by AVL CRUISETM M (see Tatschl and Wurzenberger [1]), is applied in this study. Several characteristics empower the simulation of multi-physical systems. Each domain, featuring individual types of model equations, uses a tailored numerical solution technique and an individual time stepping. A generic co-simulation master coordinates the co-simulation of the different domains and conservative flux balancing over the domain borders is applied.

Figure 1: Multi-physical domains in multi-rate setup.

2.2 Electrochemical Battery Model The electrochemical battery model follows the efforts from Doyle et al. [2]. Figure 2 shows a unit cell comprising a porous electrode of metal oxide, a separator layer permeable only for Li-ions and a porous electrode of graphite. A 1D model is applied assuming that the major gradients occur between the two collectors. The model describes the transport of Li-ions (potential e and mass ce) in the electrolyte, the transport of electrons ( potential s) in the solid phase and the transport of heat represented by the fluxes 𝐼 , 𝑁 , 𝐼 , 𝐻 , respectively. The inter- and extracalation of Li-ions (𝑁 ,) into the solid structures is described by a transient, spherical, 1D diffusion model. The driving force for the intercalation of Li-ions is the potential difference between the electrolyte and the solid structure. An essential aspect of the applied modelling depth is the spatial resolution of gradients. Degradation processes are, in general, triggered by local nonuniformities and/or by local values of potentials, concentrations and temperatures. A major concern of Li-ion batteries is the capacity and power fading during their lifetime. Li-ion batteries are a complex system and phenomena leading to performance degradation are comprehensive and not fully understood. In an extensive study Vetter et. al. [3] discussed degradation mechanisms at the carbon anode and the lithium metal

222

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level oxide cathode. One dominant aging effect at the carbon anode takes place at the electrolyte/electrode interface. At this interface, the electrolyte tends to decompose and to form of a solid electrolyte interphase (SEI). The SEI formation is a continuous reaction taking place in parallel to the intercalation of Li-ions. The rate of SEI growth tends to decrease with increasing layer thickness and it is enhanced by high operating temperatures and charging to high state of charge (SOC) values. The SEI layer stores lithium that no longer can be cycled, thus it reduces the reversible capacity of the battery. The SEI film also increases the impedance of the battery. Li-ions need to overcome an additional diffusive resistance to intercalate into the solid structure. The additional volume required by the film changes the porosity of the electrode that can cause contact losses of active particles and therefore reduce the electric conductivity of the solid, which all decrease battery voltage and increase the rate of heat generation.

Figure 2: 1D+1D battery model. Transport of Li-ions, electrons and heat.

In a recent work, Yang et al. [4] propose a mass balance for the SEI layer considering an irreversible deposition rate on spherical particles including a diffusive transport term that limits the SEI formation rate with increasing film thickness. Figure 3 comprises these correlations. The total lithium reaction rate 𝑟 is the sum of the interaction rate 𝑟 and SEI formation rate 𝑟 and thus expresses the competing character of the two phenomena. The irreversible SEI formation applies a Tafel type equation and the lithium intercalation rate follows the Butler-Volmer equation. The variables, kSEI, SEI, ka, kc represent model parameters, Ts is the temperature and ce, cEC, cs,s and cEC,s denote the lithium and ethylene carbonate (EC) concentration in the electrolyte and at the particle surface, respectively. The reaction current density is given by i0, F, R are the Faraday and ideal gas constant and SEI and a denote reaction overpotentials. The overpotentials apply an additional resistance depending on the SEI film thickness SEI. The EC mass balance combines in a steady correlation a diffusive transport and a reaction term. Here, De represents the diffusion coefficient that scales the EC concentration differences between the bulk and the particle surface. The SEI mass balance is determined

223

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level by the SEI formation rate, where SEI is the density of the film, MGSEI is its molar mass and ap is the volume specific particle surface.

Figure 3: Anode reaction mechanisms. Lithium intercalation and parasitic SEI formation.

Further well-known battery degradation mechanisms, like lithium plating and dendrite growth, collector corrosion, binder decomposition and metal dissolution are left out in this modeling study.

2.3 Fuel Cell Model The fuel cell model considers a physically based spatial and temporal resolution relying on a computationally efficient method combining 1D numerical and 2D analytic solutions, called Hybrid 3D Analytic-Numerical (HAN), Tavčar and Katrašnik [5]. Figure 4 shows the concept of the HAN approach. The fuel cell stack of parallel channels is reduced to one representative unit. This unit is numerically discretized in axial flow direction and each axial slice is further split into seven cuboids representing different computational domains. The different domains influence each other by providing boundary conditions for each other. The species balance equations in the channels and GDLs (gas diffusion layer) in the x-y plane (see Figure 4) describe an advection and diffusion and a pure diffusion problem, respectively. This suitable for analytical solution methods when introducing a velocity potential being the spatial derivative of the velocity. The species concentration vector holds values of reactant and product species and the corresponding diffusion coefficients are noted as matrix of ternary diffusion coefficients appropriately considering multispecies diffusion phenomena. The species balances of the individual planes are connected in axial direction by source and sink terms representing advection in axial direction. The flow velocity in axial direction is calculated with the help of a steady pressure drop correlation and the assumption of laminar (Hagen Poiseuille) velocity

224

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level profiles in radial direction. The transient enthalpy balance of membrane, GDL and bipolar plate is solved in a lumped manner exchanging heat with the gas flows going through the channels.

Figure 4: Modeling concept of HAN.

At the cathode side of the membrane, the reaction of protons H+ with oxygen O2 and electrons e- is modeled with a Butler-Volmer type of equation. The exchange current density depends on temperature and the partial pressures of reactants and products. The proton transport through the membrane is modeled by means of local potential gradients and an ohmic resistance. This resistance depends on temperature and the membrane water content at the anode and the cathode side. The flux of liquid water through the membrane is driven by diffusion and electro-osmotic drag. Here the diffusivity of water and the thickness of the membrane are considered. The solution of the fuel cell balance equations leads to a temporal and spatial resolution of species concentrations, electric potentials and temperatures. These local states are essential to assess the impact of operating conditions on fuel degradation phenomena in a physical causal manner. Figure 5 schematically represents the modeling framework of intertwined catalyst degradation mechanisms presented in Kregar et. al [6]. High temperatures, high local electric potentials between the ionomer phase and the catalyst material and high humidity represent degradation stimuli (black block in Figure 5) causing an increased rate of formations of hydroxide and oxide groups on carbon support and Pt (green block in Figure 5). The presence of hydroxide and oxide groups affects the catalyst degradation in two ways. Firstly, the oxide layer shields the Pt surface from electrochemical dissolution, however, despite this effect, the rate of platinum dissolution increases at high local electric potential differences leading to “Ostwald” ripening of Pt particles. This process describes the net growth of catalyst particles, driven by the dissolution of small particles due to their higher surface tension, and redeposition of material on larger particles, resulting in a net loss of catalyst surface area (blue blocks and arrows in Figure 5). Some dissolved Pt is also lost due to diffusion into the membrane (purple block and

225

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level arrow in Figure 5). Secondly, the formation and interaction between surface hydroxide groups on carbon and Pt leads to electrochemical corrosion of carbon support (orange block and arrows in Figure 5) being most pronounced at the intersection between particles and their support. This leads to detachment of Pt and its agglomeration into larger particles (red blocks and arrows in Figure 5). The catalyst degradation model simulates the effects of both particle growth processes. It tracks the rate of formation of hydroxide and oxide groups and the rates of electrochemical Pt dissolution and carbon corrosion. The model calculates the resulting changes in particle size distribution, from which the particle size redistribution and consequently the catalyst surface area can be determined.

Figure 5: Causal chain and schematic presentation of intertwined catalyst degradation mechanisms.

3 Results This section comprises results showing different aspects of FCEV modeling on vehicle and on component level. Validation results are presented and simulation data on the spatial and time resolved behavior of the battery and the fuel cell are shown. The impact of driving conditions on battery and fuel cell degradation is discussed.

3.1 Validation on Component Level The detailed electrochemical models of battery and fuel cell were validated in preceding efforts from the authors of this study. Wurzenberger et al. [7] simulated polarization curves of an LMO (lithium manganese oxide) battery for different discharge rates. The results showed a good match with given reference data taken from 3D-CFD simulations and measurements. Tatschl et. al. [8] simulated fuel cell polarization curves and compared the results with experimental data and 3D-CFD simulations. The model was assessed at different temperatures, pressures and excess air ratios.

226

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level

3.2 Validation on Vehicle Level The simulation framework presented in this study is used to model the Toyota Mirai vehicle. Figure 6 shows the basic model layout and key parameters taken from LohseBusch et al. [9].

Figure 6: System layout and model parameter of Toyota Mirai 2017 modeled in CRUISETM M.

The model focuses on the electric network featuring the detailed electrochemical component models for battery and fuel cell. These components are placed in parallel and feed, depending on the operating strategy, power to the electrical machine. DC/DC converters are used to operate the fuel cell, battery and e-motor on different voltage levels. The power from the e-motor is transferred to the wheels via a final drive and differential. The fuel supply system for air and hydrogen and the cooling system are modeled in a simplifying manner. A Li-ion based battery system is chosen to demonstrate the capability of the proposed electrochemical model. Hence, a virtual battery pack is applied featuring the same energy and nominal voltage as the original NMh (Nickel metal hydride) battery of the Toyota Mirai. A driver model drives the vehicle following a given target velocity. The torque demand from the driver is translated via a Power Control Unit (PCU) into corresponding air and fuel supply flows for the fuel cell. The PCU also takes care of the power balancing between fuel cell and battery, including the recovery of kinetic energy during braking. Here it needs to be mentioned that the PCU model does not claim to be close to the actual PCU applied in the Toyota Mirai vehicle. The control model is as complex as needed to run the virtual vehicle in transient conditions and to provide plausible responses of key performance indicators during transient drive cycle simulations.

227

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level

Figure 7: FCEV in WLTP. Simulation and measurement data of e-motor (EM), fuel cell (FC) and battery (B).

Figure 7 shows results from a WLTP (worldwide harmonized light vehicle test procedure). Several key performance indicators of the electrical machine, the fuel cell and the battery are compared to measurements taken from Lohse-Busch et al. [9]. The comparison reveals a very good match for almost the entire driving task except for the phase between 1050 s 1300 s. During this phase, the measured power of the e-motor deviates from the simulated one. Since, aside from this deviation, the simulated power nearly fully coincides with the measured one, this isolated difference cannot be attributed to an additional power consumption that is not sufficiently covered by the model. Figure 8 presents the fuel consumption and battery SOC, simulated for the WLTP cycle. The flat segments of the fuel consumption go along with either stand-still or deceleration phases of the WLTP. The fuel consumption deviates by 4% from the measurements

228

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level underlining that the model was reasonably well parameterized. The curve of the battery SOC reveals all phases of regenerative braking (SOC increases) and that the torque demand during highly transient acceleration phases (1050 s to 1300 s, 1450 s to 1700 s) is rather covered by energy from the battery (SOC decreases) than from the fuel cell. This power split results from the vehicle control strategy. The SOC is neutral over the entire drive cycle, which is essential to correctly access the fuel consumption.

Figure 8: FCEV in WLTP. Fuel consumption and battery state of charge.

3.3 Intra-Component Assessment Figure 9 presents spatial and time resolved data of the battery. The battery SOC (Figure 9a) shows that the battery undergoes distinct charging phases with increasing stoichiometries at the anode. The highest intercalation values can be observed close to the separator layer. Similar, the Li-ion concentration in the electrolyte (Figure 9b) shows the highest concentration near the current collectors in the electrode that is currently de-intercalating lithium. Li-ions diffuse to the opposite electrode and shows lower electrolyte concentrations when being intercalated. The spatial and time resolution of the stoichiometry and electrolyte concentrate reflect the power demand and thus their relation to the vehicle parameters, the control strategy and the driving profile. The spatial non-uniformities revealed by the model are influenced by the electrode thickness and type of active materials. These parameters influence the porosity, ionic and electric resistances and hence the power and capacity of the battery. This closes the causal chain between power demand originating from a driving task to the power feedback influenced by intra-component phenomena, in a physically consistent manner.

229

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level

Figure 9: Battery in WLTP: Solid stoichiometry in metal oxide and graphite (a) and electrolyte concentration (b) over the cell thickness and time.

Figure 10 shows spatial and time resolved data of the fuel cell. The spatial gradients of the current density (Figure 10a) are influenced by the local excess air ratios, humidity and temperature. With the continuous depletion of oxygen along the cathode channel the reactant availability becomes smaller and reduces the efficiency of the fuel cell. The local water content (Figure 10b) is the result of transport phenomena in the membrane and the power demand. In turn, the water content also influences ohmic resistance of the membrane and thus the provided power. It enhances the proton transport through the membrane up to a point where too much water floods the catalyst hindering the access of the reactants to the reaction sites.

Figure 10: Fuel cell in WLTP: Current density (a) and membrane water content (b) over the length of the fuel cell channel and time.

The performed analysis reveals a strong dependency of intra-component gradients on operating conditions such as excess air ratio, humidity and power demand. Vice-versa,

230

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level local maldistributions influence the power output of the fuel cell and thus feedback to the system level of the vehicle.

3.4 Degradation on Vehicle Level Figure 11 shows the growth of the SEI layer during the WLTP. The layer thickness is taken from two distinct points in the graphite electrode. Near the separator, the SEI layer thickness shows a faster growth which can be explained by the higher overpotentials at this location. The partially flat phases of the SEI thickness coincide with the phases where the vehicle accelerates, and the battery is discharged. In opposite direction (see last 100 s) the most pronounced SEI growth takes places during the period of fastest charging. The thickness of the SEI layer (a few nanometers) leads to a total capacity loss of the battery of around 20% when upscaled to 6000 h of operation.

Figure 11: FCEV in WLTP. SEI layer growth at the positions in the graphite electrode.

The fuel cell catalyst degradation model was calibrated based on experimental data, obtained during 3000 h of fuel cell operation (see Borup et al. [10]). As seen in Figure 12a, the degradation of the catalyst layer is not very pronounced during the vehicle operation. It can also be seen that the proposed degradation model replicates well the particle size growth in the entire size range. Figure 12b shows the growth of the mean particle size during the WLTP. The particle growth strongly depends on the local electric potential in the catalyst layer, which controls the rates of deteriorating electrochemical processes. Although the overall change in particle size is small (~10-5), the differences in degradation between inlet and outlet of the cathode channel are clearly visible. This can be explained by the differences in humidity, which increases along the cathode channels, being one of the main precursors of oxidation processes in the catalyst, leading to its degradation.

231

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level

Figure 12: Comparison of catalyst degradation model to experimental data (a) and growth of catalyst particles during WLTP (b) caused by the difference between electrode potential and potential of immediately adjacent ionomer (c).

4 Conclusion This study presents a FCEV model applying electrochemical, spatially resolved, battery and fuel cell models on system level. The base models and models for degradation are discussed. Validation results on vehicle level and insights into temporal and spatial resolved battery and fuel cell characteristics are shown. The impact of component aging is simulated for a drive cycle. From this, the following conclusions can be drawn. ● The results gained from the multi-physical simulation framework emphasize the capability to assess multiple interacting domains of different scales at the same time. The validation on vehicle level demonstrates the applicability of comprehensive component models on vehicle level. The validation shows that key performance indicators can be meet with reasonable effort of model parameterization and control modeling. ● The application of spatial resolved electrochemical battery and fuel cell models on system level demonstrates the successful closure of the causal chain between intra-component phenomena and vehicle driving performance. Intra-component phenomena influence the vehicle performance and the driving conditions influence the component behavior. ● Spatial resolved, electrochemical models are mandatory to investigate component degradation. The application of these models on vehicle level interactively reflects the impact of driving conditions, vehicle parameters, component characteristics and control strategies. This allows for unprecedented optimization capabilities very early in the development process.

232

FCEV simulation – Electrochemical battery and fuel cell models on vehicle level

Bibliography 1. Tatschl, R. and Wurzenberger, J. C., “Antriebsstrangsimulation mit dem kommerziellen Code CRUISETM M,” In Grundlagen Verbrennungsmotoren 9th Ed. edited by G. P. Merker and R. Teichmann, 1403-1439. Springer, 2019, ISBN-10: 365823556X. 2. Doyle, M., Fuller, T. F. and Newman, J., “Modeling of Galvanostatic Charge and Discharge of the Lithium/Polymer/Insertion Cell,” J. Electrochem. Soc. 140(6):1526-1533, 1993, doi:10.1149/1.2221597. 3. Vetter, J., Novák, P., Wagner, M. R., Veit C. et al., “Ageing mechanisms in lithium-ion batteries” J. of. Power Sources 147(1-2):269-281, 2005, doi:10.1016/j.jpowsour.2005.01.006. 4. Yang, X.-G., Leng, Y., , Zhang, G., Shanhai Ge., S. et al., “Modeling of lithium plating induced aging of lithium-ion batteries: Transition from linear to nonlinear aging” J. of Power Sources 360:28-40, 2017, doi:/10.1016/j.jpowsour.2017.05.110. 5. Tavčar, G. and Katrašnik, T., “An Innovative Hybrid 3D Analytic-Numerical Approach for System Level Modelling of PEM Fuel Cells,” Energies 6(10):5426-5485, 2013, doi:10.3390/en6105426. 6. Kregar A., Tavčar G., Kravos A., and Katrašnik T., “Predictive virtual modelling framework for performance and platinum degradation modelling of high temperature PEM fuel cells,” Energy Procedia 158:1817-1822, 2019, http//doi.org/ doi:10.1016/j.egypro.2019.01.426. 7. Wurzenberger, J. C., Rašić, D., Tavčar, G., Glatz, T., Mele, I. and Katrašnik, T., “FCEV Performance Assessment - Electrochemical Fuel Cell and Battery Modelling on Vehicle Level” SAE Technical Paper, 2020 (accepted). 8. Tatschl, R., Fink, C., Tavcar, G., Urthaler, P. et al., “A Scalable PEM Fuel Cell Modelling Approach to Support FCEV Component and System Development,” Europ. Battery, Hybrid and Fuel Cell Electric Vehicle Congress, Geneva, 2017. 9. Lohse-Busch, H., Duoba, M., Stutenberg, K., Iliev, S. et al., “Technology Assessment of a Fuel Cell Vehicle: 2017 Toyota Mirai,” Argonne National Laboratory, Engery Systems Division, Report # ANL/ESD-18/12, 2018. 10. Borup, R. L., Mukundan R. More K. L., Neyerlin K. C et al., “PEM Fuel Cell Catalyst Layer (MEA) Architectures”, on 233th Meeting of the Electrochemical Society, 2018-05-13, Seattle US, https://permalink.lanl.gov/object/tr?what= info:lanl-repo/lareport/LA-UR-18-24447.

233

Holistic design of innovative cathode air supply for automotive PEM fuel cells Dr. Michael Harenbrock, Alexander Korn, Andreas Weber MANN+HUMMEL GmbH Eva Hallbauer MANN+HUMMEL Innenraumfilter GmbH & Co. KG

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_25

235

Holistic design of innovative cathode air supply for automotive PEM fuel cells

1 Introduction 1.1 Motivation Climate change is one of the major threats to mankind. To reach the target of maximum 1.5°C temperature rise compared to pre-industrial levels set by the COP21 Conference in Paris, CO2 emissions from transport must be reduced significantly. Fuel cell technology can play a major role in reducing these emissions. Compared to storing energy in batteries, fuel cells decouple vehicle weight from electric driving range as the energy is generated on-board and does not need to be stored beforehand. This opens up applications which cannot be easily covered by Battery Electric Vehicles (BEV) powered by Li ion batteries like heavy-duty, long-haul transport, with short fueling time and higher driving range requirements [1].

Figure 1: Favorable fields of application for alternative propulsion technology

For automotive applications, Proton Exchange Membrane (PEM) fuel cells are the technology of choice as they allow start-stop operation required in transportation. As in Internal Combustion Engines (ICE), oxygen is taken from ambient air, while hydrogen is stored in pressure tanks. Figure 2 shows a generic fuel cell system [2].

236

Holistic design of innovative cathode air supply for automotive PEM fuel cells

Figure 2: Simplified layout of a typical automotive system, adapted from [2]

To enable market penetration of fuel cell technology, the fuel cells must reach the required system lifetime. In addition, system cost has to be reduced for widespread technology adoption. The paper showcases a holistic system approach for designing complete cathode air paths, highlighting both aspects mentioned above.

1.2 Methodology Analysis of the functional requirements of the cathode air intake path reveals similarities with the air path for Internal Combustion Engine Vehicles (ICEV) with regard to typical air flow and temperature requirements [3].

Figure 3: Typical pressure and temperature levels in cathode air paths

This is plausible as ICE and fuel cell use oxygen as one reactant to generate power – ICE by combustion with hydrocarbon-based fuels, while fuel cell make use of an con-

237

Holistic design of innovative cathode air supply for automotive PEM fuel cells trolled electrochemical, catalyzed reaction with hydrogen. The overarching methodology is to adapt existing solutions established for ICE air management to fuel cell applications with similar power rating and oxygen demand, applying state-of-the-art technology. Through this approach, component cost is reduced while energy efficiency and acoustic performance are improved through this holistic approach recent work. One major difference between fuel cell and ICE is the exhaust temperature and composition. As the cathode exhaust mainly consists of oxygen-depleted air with process water, there is no need to use exhaust pipes made out of metal. The use of plastic instead of stainless steel offers potentials for cost and weight reduction.

2 Cathode air paths The cathode air path – inlet side´s main components are cathode air cleaner, water separators, compressor, air ducts, air-cooler, resonators, flaps, humidifier, and a media distribution plate. The cathode air path – exhaust side contains a water separator, a resonator and air ducts.

Figure 4: Balance-of-plant components for cathode air and coolant loops

2.1 Cathode air filter To achieve the required system lifetime, clean air supply is essential. Gases like NOx, SO2 and NH3 poison the catalyst, leading to higher stack degradation rates [4, 5]. Efficient gas removal is realized with functionalized activated carbons, binding the target gases through polar interactions [6]. Figure 5 shows the positive effect of adsorption [7].

238

Holistic design of innovative cathode air supply for automotive PEM fuel cells

Figure 5: Positive effect of activated carbon cathode air filter on short stack degradation [7]

It compares the degradation of two stationary short stacks, operated in parallel under real-life conditions with ambient air, with one of the short stacks having been protected by a filter equipped with NOx-specific activated carbon. The binding capacity of an adsorber for a given target gas is specified per gram of adsorber material. To ensure protection against these gases for an intended service interval, the total mass of gas to be separated in that period must be known. As concentration levels consist of background and local contributions, peak concentrations can occur locally which can have a specifically high negative impact and are not considered in the typical averaged air quality data [8]. To overcome this problem, collecting concentration data under real driving conditions with high spatiotemporal resolution is required. As part of the ALASKA project [9-12], measurement campaigns were carried out in Germany, e.g. on a representative route around Stuttgart / Germany, often used to estimate real-life fuel consumption in passenger cars by FZ Jülich. The route length is 93 km, the drive duratin is about 2 hours, and it includes an uphill climb up to 550 m altitude with the following road type distribution: ● City traffic: 13% ● Country roads: 16% ● Federal highways: 52% ● Highways: 19%

239

Holistic design of innovative cathode air supply for automotive PEM fuel cells Based on the measurements, the NO, NO2, NH3 and SO2 loads per minute of fuel cell operation were calculated per road type [10]. Using the assumption of an average daily driving time and distribution between driving time on highway, country and city traffic, annual loads were estimated which serve as a basis for filter element design, especially the required amount of adsorber material, see table 1 [10]. Table 1: Annual gas loads for NO, NO2, NH3 and SO2 Gas NO NO2 NH3 SO2

Load –highway 0.32 mg/min 0.22 mg/min 0.08 mg/min 0.01 mg/min

Load - country 0.17 mg/min 0.14 mg/min 0.01 mg/min 0.01 mg/min

Load – city 0.12 mg/min 0.07 mg/min -

Average load 2.9 g/year 2.0 g/year 0.1 g/year 0.1 g/year

As specific interactions between activated carbons and gases are utilized to capture contaminants, gases with different molecular properties require different activated carbons for effective removal. To ensure sufficient capacity and selectivity towards all relevant gases special activated carbons can be used that are modified to target specific gases either by means of physical or chemical interactions. Figure 6 shows breakthrough curves of four modified activated carbons with two different target gases [12].

Figure 6: Breakthrough tests according to ISO 11155-2 with two test gases and four adsorbers

A filter element was designed based on this data. Of the different design options, a concept was chosen which combined a pleated filter bellow with adsorber stacks which met the pressure drop requirement (figure 7).

Figure 7: ALASKA filter element

240

Holistic design of innovative cathode air supply for automotive PEM fuel cells The required balanced protection against NH3, SO2 and NOx was proven by adsorption capacity tests at 50% relative humidity and an air flow of 100 m³/h. The tests were carried out at concentration levels significantly exceeding atmospheric concentrations to achieve full loading / breakthrough with acceptable test duration [12]. Table 2: Adsorption test results for NO, NO2, NH3 and SO2, ISO 11155-2 Gas NH3 Toluene SO2 NOx n-Butane

Concentration 30 ppm 80 ppm 30 ppm 30 ppm 80 ppm

Average Capacity 0.2 g/filter 77.0 g/filter 15.5 g/filter 13.7 g/filter 8.7 g/filter

In addition to total binding capacity, the capability of a filter to prevent breakthrough caused by concentration peaks is essential for efficient stack protection. As an example, even a high NH3 concentration peak, probably caused by NH3 slippage from a vehicle equipped with an SCR exhaust after-treatment device, could be managed during the test campaign, as shown by the low breakthrough concentration on the filter media´s downstream side (figure 8).

Figure 8: Breakthrough performance of filter media against NH3 at 10 l/min. airflow, highway

The findings show that fuel cells can effectively be protected against airborne contamination. Capturing different target gases in a balanced way is essential to achieve the required system lifetime.

241

Holistic design of innovative cathode air supply for automotive PEM fuel cells

2.2 Water separator In addition to gaseous and particulate airborne contaminants, the separation of liquid water from cathode air plays an important role [13]. Depending on system design and functional requirements, water separators can be integrated at different positions as shown in figure 9 at positions 1, 6, and 7.

Figure 9: Positions for water separators in cathode air paths

The task of water separator (figure 9, position 1) is to prevent ingress of liquid water from ambient air into the cathode air path. To keep the filter element fully functional, water loading, e.g. through splash water or heavy rainfall, has to be prevented to avoid blocking of the adsorber pores by water. While supplying the fuel cell stack with air which contains a sufficient amount of humidity to prevent membrane dehydration and thus performance degradation caused by ohmic losses, the entry of liquid water has to be prevented. Water entry into the stack on the cathode side would lead to blocking of pores in the gas diffusion layer, resulting in reduced transport of oxygen to the catalyst and thus power loss. The water separator (figure 9, position 6) separates droplets and liquid water which might result from oversaturation of the cathode inlet air caused by the humidifier. On the cathode exhaust side, the remaining exhaust pressure can be used to drive a turbine which supports the compressor on the inlet side, thus reducing the overall power consumption of the compressor. To protect the turbine blades from impact of large water droplets or even ice crystals, an additional water separator can be applied upstream of the turbine (figure 9, position 7) [2]. Water separation on the exhaust side also prevents water splashing out of the tailpipe, e.g. at low temperatures.

242

Holistic design of innovative cathode air supply for automotive PEM fuel cells Passive water separators for cathode air systems are applied as actively driven systems like disc separators require additional electric power which would decrease the overall system energy efficiency. Diffusion separators (coalescers) are efficiently removing smaller droplets, but they create a relatively high pressure drop which is often not acceptable because of the additionally required compressor power which influences noise and energy efficiency. Consequently, axial cyclones are applied for pressure dropoptimized water separation (figure 10).

Figure 10: Functional principle of an axial cyclone water separator

Axial cyclones can be mounted in vertical or horizontal position. If the component is mounted in horizontal position, efficient separation requires slightly higher water loads, whereas positioning the separator in a vertical position also enables water droplet removal at lower water mass flows as the separation is supported by the force of gravity. In addition to the optimization of the gravimetric water separation efficiency, special care has to be taken on the pressure drop resulting from the specific design. Table 3: Typical performance data for axial cyclone water separators Gravimetric separation efficiency (droplet size 10 – 100 µm) Mass air flow range Pressure loss

: 70 – 90% Horizontal: 20–100% of mmax Vertical: 0 – 100% of mmax p: < 30 hPa

Adjustment of the swirl generator design offers opportunities for pressure drop optimization.

2.3 Humidifier To supply power at high energy efficiency, proton transport from anode to cathode through the membrane has to be fast. In case the air supplied to the stack at temperatures

243

Holistic design of innovative cathode air supply for automotive PEM fuel cells above 50 – 60°C has a low humidity, the amount of water produced inside the cell is not sufficient to maintain the required humidity level of the membrane. As a consequence, the membrane dries out which results in ohmic losses as ionic conductivity. Thus, humidification of the cathode inlet air is required. Often, an external gas-to-gas humidifier is used, although internal humidification is also possible [2]. As the pressure level in the wet exhaust is lower than on the inlet side, humidity has to be transferred against a macroscopic pressure gradient. Membranes are used as the active material inside the humidifier. The membrane has to be gas-tight, but still capable of water transport by a solution-diffusion mechanism. Either hollow fiber or flat sheet membranes are used with the latter offering benefits in terms of pressure drop. Humidifiers can be designed out of plastic which offers weight reduction potential and more design options. They can be operated in counter-, cross- or parallel flow, see figure 11 [14].

Figure 11: Flat-sheet humidifiers: counter-flow (left), cross—flow (middle), parallel flow (right)

The product design shown in figure 12 consists of a box-shaped humidifier element with flat-sheet membranes, operated in cross-flow. The element is fixed in a basic frame structure with two ribbed walls. The four plates containing the in- and outlet connections are designed as separate parts.

Figure 12: Plastic flat-sheet humidifier

Ribbing helps to prevent deformation of the humidifier housing which might be caused by the overpressure inside the system. Through this design, the central components – humidifier element, frame, and two walls – can be standardized for a given fuel cell

244

Holistic design of innovative cathode air supply for automotive PEM fuel cells power rating, whereas adaption to concrete packaging requirements and connections can be realized by modification of the four end plates.

2.4 Connecting the components 2.4.1 Air ducts Connecting the individual components yields the full system. To design air ducts for cathode air inlet systems, the different temperature and pressure levels must be considered as shown in figure 3. Standard air ducts can be used at the low pressure inlet side, upstream of the compressor (figure 3, condition 2). The highest material requirements are for the charge air ducts directly located downstream of the compressor (figure 3, condition 3) as the air temperature rises to levels up to 200°C. As in ICEV, an air cooler can be integrated into the charge air ducts. Standard air ducts can also be used at the exhaust side as pressure and temperature levels are moderate (figure 3, conditions 5-6). When selecting the material, hydrolysis-resistant grades have to be employed to avoid material degradation and leaching of contamination from the parts´ surfaces into the air stream. As during operation the complete system will be encountering vibration, elements for movement compensation must be integrated into the ducting. Typical elements are bellows and agile joints.

Figure 13: Movement compensation by bellows (left) and agile joints (mid, detail left)

Bellows as shown in figure 13 are capable of compensating larger movements, while agile joints are effective for movement compensations up to approximately 3 mm. In addition to temperature and pressure stability, material selection for ducts must ensure that the ducts´ surfaces do not introduce contamination into the system themselves, e.g. through reaction with hot water vapor, liquid water, or additive leaching effects.

2.4.2 Flaps & sensors To enable intelligent air flow management, further components like sensors and flaps, can be added.

245

Holistic design of innovative cathode air supply for automotive PEM fuel cells

Figure 14: Position of sensors and flaps

Flaps in positions 11a (figure 14) manage the air flow to and from the fuel cell stack. Integration into a manifold-like plastic part yields a media distribution module. An additional last-chance particle filter can be added to ensure that no particles released from cathode air path components downstream the cathode air filter enter the stack. Figure 15 shows a media distribution module, containing cathode air inlet and cathode air exhaust outlet routing to stack, as well as connections for the coolant in- and outlet.

Figure 15: Media distribution module for the cathode side

A flap on position 11b (figure 14) enables bypassing of the humidifier in case that the humidity level of the cathode air is already in the target range. Thus, energy savings can be realized that would be consumed if the air passed through the humidifier. A cathode bypass (figure 14, position 11c) is also integrated. The function is similar to waste-gates applied for turbo-charged combustion engines. If the fuel cell stack runs at full power, and e.g. braking rapidly reduces the air flow to zero, the stack´s inlet flap will close. As the compressor still runs at high rpm, pressure pulses result. To avoid this, the pressurized air can be directly supplied to the cathode exhaust flow, bypassing humidifier and stack. Efficient flow management requires sensor integration inside the system (figure 14, positions 8), supplying necessary data on system pressure, temperature, humidity and air mass flow, e.g. high-precision, automotive hot–wire mass flow sensors [2].

246

Holistic design of innovative cathode air supply for automotive PEM fuel cells

2.4.3 Silencers Improvement of acoustics is an important challenge for electrified powertrains as noises generated by radiation, flow or the compressor are not masked by an engine. Different silencers can be applied on different positions of the systems (figure 16, positions 9a, 9b, 9c), transferring functional principles from ICEV air paths [3].

Figure 16: Position of resonators

At the low pressure side (figure 16, position 9a), noises from the inlet orifice and noise radiation from the ducts must be removed. Main source of the noises is the compressor. On the high pressure side (figure 16, position 9b), noise radiation from the ducts must be removed. Noise sources are mainly the compressor and flow noises of the stack and the humidifier. On the exhaust side (figure 16, position 9c), noises from the exhaust orifice as well as noise radiation from the ducts must be removed. Noise sources are the turbine and flow noises, namely stack, humidity, and water drainage with throttle. Tuning is typically designed for 20 dB noise reduction in the requested frequency range. It can be achieved by modifying the porous surface area and the resonator volumes. All four chambers function in reflective mode and work for a wide frequency spectrum.

Figure 17: Reflective 4-chamber broad band silencer

In the silencer shown in figure 18, reflective chambers 1 and 2 are combined with absorptive chambers 3 and 4 which achieve smooth dampening at frequencies above 500 Hz.

247

Holistic design of innovative cathode air supply for automotive PEM fuel cells

Figure 18: 4-Chamber broad band silencer

Multi-chamber resonators combine acoustic tuners fitting to low and medium frequency range. In this example, chambers 1 to 4 act as Helmholtz resonators, chamber 5´s function is that of a 1/4-tube.

Figure 19: Multi-chamber broad band silencer

3 Summary and conclusion The analysis shows that many of the established products from air management in ICEV can be applied for cathode air paths, e.g. through adaption of the duct diameters to reflect the different air mass flow. Compressor noises can also significantly be reduced, as well as noises from other components. If all components are regarded as one holistic system, overall acoustic performance can be optimized. It also allows to improve the pressure loss of the cathode air inlet path which helps to reduce the parasitic losses caused by the compressor. The components shown will need to be adapted for individual fuel cell systems designs, and not all components, e.g. water separators and silencers, might be required on all positions, yet the solutions described can act as a toolbox for improved system design. This approach helps to overcome cost challenges related to comparatively low annual production volumes which prevent realization of economies of scale.

248

Holistic design of innovative cathode air supply for automotive PEM fuel cells

Acknowledgement Part of this work was supported by the German Federal Ministry for Economic Affairs and Energy by partially funding the project ALASKA [03ET6036].

Bibliography 1. Hydrogen Council, “How hydrogen empowers the energy transition,” https://hydrogencouncil.com/wp-content/uploads/2017/06/Hydrogen-Council-Vision-Document.pdf, January 2017 2. James, B. D., Huya-Kouadio, J. M. et al., “Mass Production Cost Estimation of Direct H2 PEM Fuel Cell Systems for Transportation Applications: 2018 Update,” http://www.sainc.com/what-we-do/energy-consulting, 2019 3. Korn, A., Weber, A. et al., “Ansaugsysteme“ from Handbuch Verbrennungsmotor, 7. Auflage, Wiesbaden, Springer Vieweg, 2015 4. Zamel, N., and Li, X., “Effect of contaminants on polymer electrolyte membrane fuel cells,” Progress in Energy and Combustion Science, 37(3):292-329, 2011 5. Li, H., Shi, Z. et al., “Impurities in Fuels and Air,” from Encyclopedia of Electrochemical Power Sources, Amsterdam, Elsevier, 2009 6. Diersch, S., Harenbrock, M., “Contamination Control for LT PEM Fuel Cell Systems,” from 30th International Electric Vehicle Symposium (EVS30) Vol. 2, Red Hook, Curran Associates, 2018, 731-739, ISBN 978-1-5108-6370-5 7. Misz, U., “Evaluierung der kathodenseitigen Schädigungsmechanismen durch partikuläre und gasförmige Luftschadstoffe mit Hilfe von elektrochemischen Messmethoden zur Standzeiterhöhung von PEM-Brennstoffzellen (Kathodenluft II), https://www.iuta.de/igf-docs/ab_-_kathodenluft_ii_16325n_2012-06-29.pdf, 2015 8. Ehlers, C., “Mobile Messungen – Messung und Bewertung von Verkehrsemissionen,“ Schriften des Forschungszentrums Jülich, Reihe Energie & Umwelt / Energy & Environment 229, Jülich, Zentralbibliothek Verlag, 2014, ISBN 978-3-89336989-8 9. Misz, U., Talke, A. et. al., “Effects, Damage Characteristics and Recovery Potential of Traffic-induced Nitric Oxide Emissions in PEM Fuel Cells under Variable Operating Conditions,” Fuel Cells 18(5):594-601, 2018 10. Talke, A., “Der Einfluss von ausgewählten Luftschadstoffen auf die Brennstoffzelle unter fahrzeugnahen Betriebsbedingungen,“ Ph.D. thesis, Fakultät für Ingenieurwissenschaften, Abteilung Maschinenbau, Universität Duisburg-Essen, 2017

249

Holistic design of innovative cathode air supply for automotive PEM fuel cells 11. Talke, A., Misz, U. et al., “Influence of Nitrogen Compounds on PEMFC: A Comparative Study,” Journal of The Electrochemical Society, 165(6):F3111-3117, 2018 12. Wurth, S., Abschlussbericht Verbundvorhaben ALASKA: Teilprojekt: Ermittlung der Filterkapazität, 2017 13. Li, H., Tang, Y. et al., ”A review of water flooding issues in the proton exchange membrane fuel cell,” Journal of Power Sources 178(1):103–117, 2008 14. Brandau, N., “Analyse zur Zellinternen Befeuchtung eines Polymerelektrolytmembran-Brennstoffzellenstapels,“ Ph.D. thesis, Fakultät für Maschinenbau, Technische Universität Carolo-Wilhelmina, Braunschweig, 2013

250

ECMS based on system-specific control parameter adaption of a fuel cell hybrid electric vehicle Sergei Hahn, Jochen Braun, Dr. Helerson Kemmer Corporate Sector Research and Advance Engineering, Robert Bosch GmbH Prof. Dr. Hans-Christian Reuss Institute for Internal Combustion Engines and Automotive Engineering (IVK), University of Stuttgart

ECMS based on system-specific control parameter adaption of a fuel cell hybrid …

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_26

251

ECMS based on system-specific control parameter adaption of a fuel cell hybrid …

1 Introduction The powertrain of a fuel cell hybrid electric vehicle (FCHEV) can be realized by a combination of a proton-exchange membrane (PEM) fuel cell system (FCS) and a battery system (BS). The use of two power sources in one vehicle increases the complexity but also combines the advantages of both systems, such as using an energy source in FCS for realizing high ranges in combination with the high dynamics of BS [1]. Furthermore, the hybrid powertrain provides an additional degree of freedom in the control of power demand. Therefore, a well-tuned energy management strategy (EMS) can ensure a high-efficient operation of the FCHEV. A well-known and efficient method, which is applicable to every hybrid energy system and satisfies the real-time requirements in a vehicle, is the Equivalent Consumption Minimization Strategy (ECMS) [1]. This method uses an equivalence factor for weighting the power provided by the battery, so that the sum of fuel cell and battery power can be treated as an equivalent fuel consumption [2]. The authors in [3]–[5] have demonstrated that a well-tuned factor can result in almost optimal fuel consumptions. A drawback of such an offline tuning is the need of a priori knowledge of the driving cycle. Hence, this procedure is not sufficient in terms of robustness and optimality for different driving cycles [5]. To realize an online adaption of the equivalence factor, it is possible to use an adaption method based on a feedbackcontroller, see [1] and [6]–[8]. An additional improvement of the robustness of the EMS is possible by optimizing the parameters over different driving cycles [8]. Furthermore, there are several aging and degradation mechanisms in battery systems as well as in fuel cell systems [9]–[13], which can be influenced by the operation conditions. The authors in [14] and [15] investigated an optimal power split based on Dynamic Programming (DP) between hybrid power sources to avoid too high degradation rates. Such a performance loss of these systems can be expressed in the form of a change of the system behavior, e.g. a reduction of the battery capacity [11] or a change of the shape of the polarization curve [13]. Consequently, this fact requires also a re-optimization of the ECMS parameters in order to adapt the operation strategy to ensure a high efficiency and improve the robustness of the whole system. Based on this hypothesis, this paper introduces a new approach to adapt the control parameters using information of the actual system parameters.

2 Hybrid power system model The combination of a FCS and a BS results in a hybrid power supply system. The possibility to charge or discharge the battery offers one degree of freedom in choosing the operation point in order to minimize the hydrogen consumption over a driving cycle.

252

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … Fig. 1 depicts the relationship of the power flows within this hybrid system. A timedependent power demand Pd(t) has to be equal to the sum of the power Pfcs(t) provided by the FCS and the power Pbs(t) provided by the BS: 𝑃 𝑡 =𝑃

𝑡 +𝑃

𝑡

(1)

The resulting hydrogen mass flow ṁh is directly coupled to the power provided by the FCS. The amount of energy, which can be provided or absorbed by the battery, is described by the current state of charge ξ. Hence, it is necessary to describe the relationship between the power split and the state dynamics. These dynamics can be determined using the battery current Ib, which is assumed to be positive in the discharge direction, the initial state of charge ξ0, the nominal capacity Q0 of the battery and the following equations [4]: 𝜉 𝑡 =𝜉 − ( )

=−

𝐼 𝜏 d𝜏

(2)

( )

(3)

ṁh(t) BS Battery FCS

= = Pbs(t)

Pfcs(t)

Pd(t) Fig. 1: Power flow of the hybrid power system.

253

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … Ib Ri(,sgn(Ib))

=

Uoc0()

Ub

Fig. 2: Equivalent circuit model of the battery.

To calculate the battery current a simple and steady-state equivalent circuit model is used, which is depicted in Fig. 2. Following this model, there is a relation between the open circuit voltage Uoc, the internal resistance Ri of the battery, the battery current Ib and the terminal voltage Ub of the battery: 𝑈 (𝜉, 𝐼 ) = 𝑈 (𝜉) − 𝑅 𝜉, sgn(𝐼 ) ∙ 𝐼

(4)

The battery specific parameters can be modelled with a dependence on the state of charge and the direction of current flow [4]. With (4) the output power of the battery Pb can be expressed as a function of Ib and ξ: 𝑃 =𝑈 ∙𝐼 =𝑈 ∙𝐼 −𝑅 ∙𝐼

(5)

If a DC-DC converter within the BS is considered, a constant efficiency factor ηbsc can be used to depict the reduction of the effective power by the additional losses. With this assumption, the BS power can be stated as: 𝑃 (𝑡) = 𝜂

( )

∙ 𝑃 (𝑡)

(6)

This equation leads in combination with (1) and (5) to a relation between the current power demand, the FCS power and the battery current: 𝑃 −𝑃

=𝜂

(

)

∙ 𝑈 ∙𝐼 −𝑅 ∙𝐼

(7)

3 Operation strategy The minimization of the hydrogen consumption over a driving cycle with the duration tf is an optimal control problem. To find a solution for this problem the cost function,

254

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … the optimization objective and the constraints of the system have to be defined. Finally, the optimization problem can be expressed using a mathematical formulation: (𝑡) = min

min 𝐽 𝑃 ( )

𝑚 𝑃

( )

( )

⎧ ⎪ ⎪ 𝜉(0) ⎪ s.t. 𝜉(𝑡 ) ⎨ ⎪ 𝜉(𝑡) ⎪ ⎪ ⎩𝑃 (𝑡)

(𝑡) d𝑡

(8)

( )

=− =𝜉

(9)

=𝜉 ∈ 𝜉 ∈ 𝑃

,𝜉 ,𝑃

,

,

ξmax and ξmin are the maximal and minimal values for the battery state of charge. The power limits of the FCS are denoted as Pfcs,min and Pfcs,max. In order to ensure a charge sustaining operation the final usable energy within the battery must be equal to the initial amount of energy.

3.1 ECMS According to the Minimum Principle, see [1], the optimal control input has to minimize the Hamiltonian H, which is defined as followed for the given hybrid system: 𝐻 = 𝑚 (𝑃

) − 𝜆(𝑡) ∙

( )

(10)

The coefficient of the dynamic part of the equation is the co-state λ(t). This parameter can be considered as an equivalence factor and transforms the power provided by the battery into an equivalent fuel consumption. The co-state is dependent on the system and the driving cycle, so in most cases this parameter is not known a priori [1]. A simple approach to overcome this drawback and to improve the robustness of the operation strategy is to implement a proportional controller, which realizes the estimation of the co-state [1]: 𝜆(𝑡) = 𝜆 − 𝑘 ∙ 𝜉

(𝑡) − 𝜉(𝑡)

(11)

This approach forces an adaption of the co-state if the deviation with respect to a reference state of charge ξref increases. This time-dependent reference allows further information based improvements, e.g. a depletion of the battery during high velocities for recuperating the kinetic energy afterwards or a prediction-based trajectory planning in the future [7]. The respective initial co-state λ0 and the proportional gain kP have to be optimized to fulfill the requirements in terms of efficiency and robustness. Due to aging

255

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … effects, the system parameters might change and the parameters should be re-optimized to maintain the performance. Unfortunately, such an optimization is very expensive and hard to realize online.

3.2 ECMS with adaptive parameters Due to this disadvantage, it is necessary to describe the ECMS parameters as functions of the system parameters. This is the proposal of this paper. Such parameters would lead to an adaption of the controller. If the Hamiltonian (10) is a convex function, the local minimum is also the global minimum [1]. Let us assume that this is the case and the local minimum is within the interval [Pfcs,min, Pfcs,max], then the minimum of (10) can be found using the partial derivative: (

)

=

(

)

−

( )

) !

(

∙

=0

(12)

In combination with (7) an interrelation between the equivalence factor and the system parameters is found: 𝜆(𝑡) = −

∙

∙𝑄 ∙

∙

(13)

Now this equation can be used to limit the search space. For a constant load and a constant battery reference voltage Uoc,0, the power demand has to be covered by the FCS [8], whereby the previous equation can be simplified: 𝜆 =−

∙𝑄 ∙𝑈

(14)

,

,

An additional consideration of the absolute limits of the hybrid system can further reduce the search space. The maximum and minimum reasonable values of the equivalence factors can be found using the current appropriate power limits of the hybrid system: 𝜆

(𝜉) = −

∙𝑄 ∙

∙

(𝜉) = −

∙𝑄 ∙ ,

∙

(15)

,

,

𝜆

,

∙

, ,

∙

(16)

The control parameters of the ECMS controller can now be expressed with (14)–(16) to realize a system dependent controller parameter set. The initial equivalence factor λ0 is a function of a reference FCS power Pfcs,0, which is equal to the average power of a cycle with constant power demand [8]. The proportional gain of the controller can be

256

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … linked with the boundaries of the equivalence factor and the boundaries of the battery state of charge. This ensures an appropriate adaption of the proportional gain after a possible change of the system parameters and prevents the system to run into the state of charge limits and remain there. The following controller laws produced good simulation results in terms of efficiency and robustness: 𝜆(𝑡) = 𝜆 − 𝑘 (𝑡) ∙ 𝜉 (𝑡) − 𝜉(𝑡)

(17)

with ( )

𝑘 (𝑡) =

(

( )

) ( )

(

( )

)

𝜆 =−

, 𝜉(𝑡) > 𝜉

(𝑡) ∧ (𝜉

(𝑡) ≠ 𝜉

)

, 𝜉(𝑡) ≤ 𝜉

(𝑡) ∧ (𝜉

(𝑡) ≠ 𝜉

)

∙𝑄 ∙𝑈

(18)

,

(19)

,

4 Simulation results The aim of this section is to evaluate and compare the performance of the introduced methodology with a conventional proportional controller. For this purpose, a quasi-stationary vehicle model is used, which executes a backward calculation of an exemplary FCHEV model in order to calculate the demanded power. For the battery model, the dataset of a generic and fictive lithium-ion battery model with a capacity of 1.6 kWh is used. The state of charge-dependent parameters for the discharging resistance, the charging resistance and the open circuit voltage are shown in Fig. 3. Open Circuit Voltage Discharging Resistance

Charging Resistance

Open Circuit Voltage, Uoc [V]

300

0.5

250

Operation Window

0.4

200

0.3

150

0.2

100

0.1

50

Internal Resistance, Ri [Ω]

0.6

350

0

0 0

0.2

0.4 0.6 State of Charge,  [-]

0.8

1

Fig. 3: Battery model data as a function of the state of charge.

257

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … The FCS is described by a current-voltage-characteristic of a PEM fuel cell stack, depicted in Fig. 4, which is based on heuristic data. Here a total number of 452 fuel cells provide a maximum power of about 114 kW and a maximum current flow of 420 A. This available power is reduced due to the energy consumption of the necessary auxiliary equipment. The consumption of these units can be approximated as a function of the fuel cell stack current, as described in [1]. An exemplary correlation, which is based on empirical data, is shown in Fig. 4 as well. Afterwards the derivation of the hydrogen consumption and the FCS efficiency can be performed. Fig. 5 illustrates the resulting system-specific graphs with the lower heating value of hydrogen as a reference. Current-Voltage-Characteristic Auxiliary Power Consumption

Stack Power 140

500 Current Limit

400

120

350

100

300

80

250 200

60

150

40

100

Power, P [kW]

Stack Voltage, Ust [V]

450

20

50

0

0 0

100

200 300 Stack Current, Ist [A]

400

500

Fig. 4: Power-characteristic of the stack and the FCS auxiliary units.

FCS Efficiency

Hydrogen Consumption 2.5

60

2

50 40

1.5

30

1

20 0.5

10 0

0 0

20

40 60 FCS Power, Pfcs [kW]

80

100

Fig. 5: Efficiency and hydrogen consumption of the FCS.

258

Hydrogen Consumption, ṁh [g/s]

FCS Efficiency, fcs [%]

70

ECMS based on system-specific control parameter adaption of a fuel cell hybrid …

4.1 Operation strategy optimization In order to optimize the introduced operation strategy in terms of efficiency and robustness two performance indices are introduced. The first performance index Jopt is the relative additional hydrogen consumption for the WLTP Class 3b driving cycle, which is used for the certification of new vehicles in the European Union. The second performance index Jrob is the relative additional hydrogen consumption averaged over the three driving cycles CADC Urban, CADC Rural and CADC Motorway 150, which is an indication for the robustness of the operation strategy with focus on hydrogen consumption. The absolute hydrogen consumption of each cycle is referred to a benchmark solution. In this paper, the minimum possible hydrogen consumption is determined using dynamic programming (DP). For this purpose, a DP script for MATLAB®, which can be downloaded from the ETH Zürich website, was used [16]. The aim of such a distinction between these two performance parameters is to find a parameter set with good performance for the certification cycle as well as high robustness for all the “extreme” use cases urban, rural and highway. To avoid any falsification of the hydrogen consumption due to a difference of the battery state of charge at the beginning and at the end of the driving cycle the cycle is iterated until a steady-state condition of the battery state of charge is reached. The performance of different parameter sets, which were defined based on a Latin Hypercube Sampling, are depicted in Fig. 6 for both methods. What can be observed is that, even with a reduced number of parameters, the resulting Pareto front of the adaptive ECMS is comparable to the Pareto front of the conventional ECMS approach. Conventional ECMS

Adaptive ECMS

1

Jrob [%]

0.8

0.6

0.4

0.2

0 0

0.2

0.4

0.6

0.8

1

Jopt [%]

Fig. 6: Performance of the different parameter sets for the two operation strategies with reference to DP.

259

ECMS based on system-specific control parameter adaption of a fuel cell hybrid …

4.2 Robustness over lifetime These results provide Pareto-optimal parameter sets, which allow a comparison of the robustness in terms of efficiency throughout the lifetime. Additionally, the relevant parameters for the begin of life (BOL) and the end of life (EOL) have to be defined. In this paper, a capacity loss of 20 % and an increase of the internal cell resistance of 50 % is assumed for the battery at EOL. Fig. 7 exemplifies the effect of fuel cell degradation on the current-voltage-characteristic. An EOL degradation of 10 % at 0.6 V fuel cell voltage was assumed as a reference. The adaptive ECMS approach allows an immediate online update of the control parameters λ0 and kP without any further optimization if information about the current state of health of the hybrid system is available. Next, the results for BOL and EOL conditions can be compared for both ECMS strategies. Table 1 shows the absolute hydrogen consumption for the WLTC and the absolute hydrogen consumption averaged over all CADC use cases. It can be observed that the relative consumption with reference to both optimization criteria can be reduced throughout the lifetime. Fig. 8 highlights the high performance of the adaptive ECMS approach by a comparison with the respective benchmark solutions for BOL and EOL. Whereas the stability of the adaptive approach is maintained, the conventional approach gets unstable over time. It is possible to improve the stability of the conventional approach by introducing an integrator part, see [1] and [2]. However, this would lead to a further increase of the optimization effort [2]. End of Life

Begin of Life

Fuel Cell Voltage, Ufc [V]

1.2 1 0.8 0.6 0.4 0.2 0 0

0.5

1 1.5 Current Density, ifc [A/cm²]

2

Fig. 7: Effect of fuel cell degradation on the current-voltage-characteristic.

260

2.5

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … Table I. Simulation results for BOL and EOL Conventional ECMS kg/100 km

Cycle

Adaptive ECMS kg/100 km

Rel. difference

0.927 1.008

+0.0 % –0.1 %

0.946 1.049

–11.7 % –15.4 %

BOL WLTC Ø CADC

0.927 1.009

WLTC Ø CADC

1.071 1.24

EOL

Conventional BOL

Conventional EOL

Adaptive BOL

Adaptive EOL

log(Jrob) [%]

10

1

0,1 0,1

1

10 log(Jopt) [%]

Fig. 8: Effect of system degradation on the robustness throughout the lifetime for the two operation strategies with reference to DP.

5 Conclusion Within this paper, a new approach for the realization of an ECMS controller parameter adaption throughout the lifetime is proposed. This methodology enables an online adaption of the operation strategy based on current information about the battery capacity, the internal battery resistance and the FCS characteristic. A simple and quasistatic simulation model of the hybrid system was used to compare the conventional ECMS

261

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … approach with the novel adaptive strategy. In a first step, an optimization of the parameters of both strategies was implemented in order to optimize the efficiency and the robustness. For this optimization, the results were referred to a benchmark, which was obtained by DP. For this simulation model, the resulting Pareto front of the new methodology has a comparable quality in terms of efficiency and robustness, even with a low optimization effort. Subsequently, one Pareto-optimal parameter set was used for each strategy to demonstrate the effect of a change of the system parameters. The comparison between the ECMS approach has shown that the adaptive approach has a higher stability than the conventional ECMS approach throughout the lifetime. This article neglected any temperature effects on the FCS and battery power. Therefore, in future studies a further examination of the adaptive strategy with a temperaturedependent behavior is necessary. Furthermore, an experimental validation with a FCS and a BS could enable an analysis of the performance for realistic aging scenarios. Finally, this methodology in general can be used for the control of other hybrid systems as well, e.g. internal combustion engines in combination with a battery. The studies using the introduced approach revealed good performance for FCHEV with small batteries in which a charge sustaining operation strategy is recommendable. With this adaptive strategy, it is possible to realize stability throughout the lifetime without an integrator part.

Bibliography 1. L. Guzzella, and A. Sciarretta, Vehicle Propulsion Systems: Introduction to Modeling and Optimization. 3rd ed., Heidelberg, Germany: Springer-Verlag Berlin Heidelberg, 2013. 2. S. Onori, L. Serrao, and G. Rizzoni, Hybrid Electric Vehicles: Energy Management Strategies. London, UK: Springer-Verlag London, 2016. 3. A. Sciarretta, M. Back, and L. Guzzella, “Optimal control of parallel hybrid electric vehicles,” IEEE Trans. Control Syst. Technol., vol. 12, pp. 352–363, May 2004. 4. D. Ambühl, “Energy management strategies for hybrid electric vehicles,” Ph. D. dissertation, Measurement and Control Laboratory, ETH Zürich, Zürich, Switzerland. 2009. 5. C. Musardo, G. Rizzoni, Y. Guezennec, and B. Staccia, “A-ECMS: An adaptive algorithm for hybrid electric vehicle management,” European Journal of Control, vol. 11, pp. 509–524, 2005.

262

ECMS based on system-specific control parameter adaption of a fuel cell hybrid … 6. M. Koot, J. T. B. A. Kessels, B. de Jager, W. P. M. H. Heemels, P. P. J. van den Bosch, and M. Steinbruch, “Energy management strategies for vehicular electric power systems,” IEEE Trans. Veh. Technol., vol. 54, pp. 771–782, May 2005. 7. T. Nüesch, A. Cerofolini, G. Mancini, N. Cavina, C. Onder, and L. Guzzella, “Equivalent consumption minimization strategy for the control of real driving NOx emissions of a diesel hybrid electric vehicle,” Energies, vol. 7, pp. 3148–3178, 2014. doi:10.3390/en7053148 8. F. Odeim, “Optimization of fuel cell hybrid vehicles,” Ph. D. dissertation, Department Energy Technology, University of Duisburg-Essen, Duisburg, Germany. 2018. 9. M. Broussely et al., “Main aging mechanisms in Li ion batteries,” Journal of Power Sources, vol. 146, pp. 90–96, August 2005. 10. J. Vetter et al., “Ageing mechanisms in lithium-ion batteries,” Journal of Power Sources, vol. 147, pp. 269–281, September 2005. 11. R. Wegmann, V. Döge, and D. U. Sauer, “Assessing the potential of a hybrid battery system to reduce battery aging in an electric vehicle by studying the cycle life of a graphite|NCA high energy and a LTO|metal oxide high power battery cell considering realistic test profiles,“ Applied Energy, vol. 226, pp. 197–212, June 2018. 12. J. Wu et al., “A review of PEM fuel cell durability: Degradation mechanisms and mitigation strategies,” Journal of Power Sources, vol. 184, pp. 104–119, September 2008. 13. D. Bezmalinovic, B. Simic, and F. Barbir, “Characterization of PEM fuel cell degradation by polarization change curves,” Journal of Power Sources, vol. 294, pp. 82–87, October 2015. 14. T. Fletcher, R. Thring, and M. Watkinson, “An energy management strategy to concurrently optimise fuel consumption & PEM fuel cell lifetime in a hybrid vehicle,” International Journal of Hydrogen Energy, vol. 41, pp. 21503–21515, December 2016. 15. F. Martel, Y. Dubé, S. Kelouwani, J. Jaguemont, and K. Agbossou, “Long-term assessment of economic plug-in hybrid electric vehicle battery lifetime degradation management through near optimal fuel cell load sharing,” Journal of Power Sources, vol. 318, pp. 270–282, June 2016. 16. O. Sundstrom, and L. Guzzella, “A generic dynamic programming Matlab function,” IEEE Int. Conf. on Control Appl., pp. 1625–1630, July 2009.

263

Using machine learning methods to develop virtual NOx sensors for vehicle applications Robert Fechert, Prof. Bernard Bäker, Stephan Gereke, Prof. Frank Atzler Technische Universität Dresden, Dresden Institute of Automobile Engineering

Using machine learning methods to develop virtual NOx sensors for vehicle …

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_27

265

Using machine learning methods to develop virtual NOx sensors for vehicle …

1 Abstract The use of combustion engines in the powertrain of vehicles will continue to play an important role for some years to come. In order to be able to comply with the legal framework conditions in the future, a further reduction in emissions must be achieved. An optimally designed exhaust aftertreatment system plays a decisive role in achieving this goal. In this paper, the development of a virtual model-based NOx sensor for diesel engine drive systems is examined. A NOx raw emission value that is as accurate as possible forms the basis for optimized operation of the SCR system that is widely used in exhaust aftertreatment today. The sensor concept to be developed should be executable in real time on existing vehicle hardware. The development of the virtual NOx sensor based on machine learning procedures. The necessary database is created by measuring real driving cycles on a dynamic combustion engine test bench. After an extensive pre-processing of the measured data, in which different combinations of model input data and time delay strategies are examined, a comparison and evaluation of different model approaches is carried out. The presented concept of a real-time capable virtual NOx sensor, which is based on the data-based modelling of raw NOx emissions, can represent the dynamic operating behaviour of combustion engines very well and can therefore be used in the vehicle at low cost.

2 Introduction In the future world of powertrains there will be a mixture of different technologies. It is the government's goal to have 25% of the vehicle fleet replaced with electrified cars by 2030. However, this also means that 75% of vehicles will be equipped with a "conventional" combustion engine. These statistics show that the combustion engine remains a very popular and suitable drive source for mobile applications. The focus will therefore continue to be on low-emission diesel engine concepts, especially for NOx reduction, combined with the reduction of CO2 emissions. To decrease the emissions of nitrogen oxide, it is helpful to be able to quantify and predict them during engine operation. Thus, both the raw emissions can be optimally operated by adjusting the operating point as well as the SCR system, which is now available in almost all diesel powertrains. Up to now, NOx sensors have mostly been used as a hardware solution for the determination. The present contribution investigates the use of a low-cost and real-time capable virtual NOx sensor for the exhaust aftertreatment of diesel engines, which is based on the data-based modelling of raw NOx

266

Using machine learning methods to develop virtual NOx sensors for vehicle … emissions by means of machine learning methods. Due to the rapid further development of such algorithms in the last years, they can be used for a multitude of technical problems today and are far superior to classical modelling approaches in the area of such complex problems, both in parameterisation and especially in calculation effort. The aim is to model the dynamic operating behaviour of combustion engines, which is particularly relevant to emissions, in the best possible way. This is achieved by adding time-delayed measurement signals as model input variables. The training of the models is initially carried out using a relatively small number of measuring points, which were recorded on a highly dynamic combustion engine test bench with a sampling rate of 100 Hz.

3 Use of virtual NOx sensors As a use case for the virtual NOx sensor, the determination of raw emissions for the optimal operation of the SCR system in vehicle diesel engines is assumed in this paper. In this context, SCR technology is currently regarded as the most efficient means of reducing NOx emissions. The formation of nitrogen dioxide begins at a combustion temperature of over 1250°C, where the formation of thermal NO is to be expected. The nitrogen and oxygen required for the formation is available in the cylinder charge after the gas exchange. The exact representation of the formation process can be described using the Zeldovich mechanism. The formation of prompt NO is closely connected with the reaction of the fuel radical CHn and occurs primarily in the flame front at temperatures from about 1000K [1]. Due to its instability, nitrogen dioxide (NO2) is formed during oxidation with oxygen. This gas is colourless, irritates the mucous membranes and has a strong pungent smell. The effects can include an increase in airway resistance, possible changes in lung function, impairment of infection defence and permanent damage to the lungs. Due to the latter effects on humans, a binding limit value for NO2 emissions for road traffic areas was set by the EU Parliament in 2010. In currently implemented powertrain configurations, the limit values for the valid Euro6D-Temp standard are safely undercut, even in the required RDE boundary conditions. The high NOx reduction rates of over ≥ 98% can be demonstrated to be achievable over a wide engine range with constant fuel consumption. In order to achieve this, it is necessary to optimize SCR catalytic converter technology and provide the exact amount of reducing agent carrier (AdBlue). In addition to quantity adjustment according to requirements, current investigations are focusing on the evaporation and deposit-free decomposition of ammonia. The dosage of AdBlue is largely dependent on the engine operating point, the exhaust gas composition and the exhaust gas temperature.

267

Using machine learning methods to develop virtual NOx sensors for vehicle …

Figure 1: Comparable system diagram of the test vehicle [2]

For the design of an optimal SCR operation strategy, the exact knowledge of the parameters mentioned is a basic requirement. Currently, the exhaust system contains a large number of sensors that are solely responsible for analysing and monitoring the exhaust gas cleaning function. Figure 1 gives an overview of the similar schematic structure of an exhaust gas treatment system. The number of sensors, combined with their respective individual costs, shows the enormous costs that are incurred. Using the virtual NOx sensor, the operating point-dependent emissions should be derived by a trained algorithm and thus become predictable. The advantage of this approach will be the predictive control of the entire exhaust aftertreatment system. Expected NOx emissions can therefore be reduced much better by early and appropriate dosing. The currently sporadic overdosage and the associated ammonia slip can thus be avoided. Conversely, the methodology can also be applied to an NOx-minimum engine operation strategy. For example, map ranges with high expected NOx emissions can be

268

Using machine learning methods to develop virtual NOx sensors for vehicle … specifically blocked for operation during warm-up (light-off of the exhaust aftertreatment system not yet reached). When the critical catalyst temperature is exceeded and NOx reduction is therefore possible, these operating ranges are released again and made available to the driver.

4 Creation of the database Before going into more detail in the following chapter about the steps that are necessary for a model developement, a brief description of how to obtain the training and test data required for the model is given.

4.1 Measurement setup As a target value for the data-driven modelling of the NOx sensor, the concentration of NOx raw emissions, which are produced during combustion in the real combustion engine, is recorded. The highly dynamic test bench available at the Chair of Combustion Engineering and Drive Technology is used to provide the learning data.

Figure 2: Test bench

269

Using machine learning methods to develop virtual NOx sensors for vehicle … A vehicle model developed by the institute is implemented in the test bench setup, where parameters can be adapted individually with regard to vehicle, driver and operating and shifting strategy. The results of the engine test bench were compared with the vehicle on the road and roller test bench measurements carried out in advance. The structure of the test bench is shown in Figure 2. A four-cylinder diesel engine with an application close to Euro 6 is used as the test vehicle. The engine has a turbocharger with variable turbine geometry and a common rail injection system. The exhaust gas can be returned to the combustion chamber via an uncooled high-pressure EGR or a cooled low-pressure EGR. In the present experiment, the defined velocity characteristics are given to the model as a boundary condition. In the course of the measurement, the various thermodynamic operating variables and the continuous exhaust gas analysis are stored in the data memory.

4.2 Measuring program To obtain the measured data, speed profiles of about 8.5 hours driving time were recorded on the engine test bench. In addition to the WLTC, four RDE-like driving cycles were recorded. The maximum speed in the cycles is about 100 km/h. Two of the driving cycles show a higher dynamic in the sense of a higher vapos value. The focus of the measurement data was initially placed on the driving behaviour in urban and suburban areas, as the training of the models in the dynamic operation of the combustion engine (acceleration and shifting operations) is particularly demanding. The challenge of mapping these in the model was met by using the highest possible number of these operating situations in the model training data. The driving cycles were generated on the basis of real driving speed curves using Markov models [3]. Table 1: Characteristics of the measured driving cycles Parameter Driving duration Route length Stops/km Maximal speed Average speed Max. accelaration Max. decelaration

270

Unit [s] [km] [1/km] [km/h] [km/h] [m/s2] [m/s2]

Cycle 1 7558 46,2 2,78 99,9 24,8 2,4 -3,0

Cycle 2 7417 50,4 2,17 100,0 26,6 2,2 -2,8

Cycle 3 7314 81,4 0,82 96,9 42,9 3,3 -3,6

Cycle 4 7272 77,2 0,79 96,9 40,9 3,3 -3,8

WLTC 1800 23,1 0,34 131,0 53,3 1,7 -1,5

Using machine learning methods to develop virtual NOx sensors for vehicle … Table 1 shows some static characteristics of the measured speed profiles. In comparison to the WLTC, it can be seen that the other four driving cycles have a lower average speed but higher maximum values for acceleration and deceleration. Cycle 1 and 2 will also clearly have a lower dynamics than Cycle 3 and 4. Emission models often show the greatest model errors, especially in transient engine operation. After an initial analysis of the model results on the basis of the cycles, it was decided in a further measurement campaign to improve the model quality in dynamic operating behavior to record individual load jumps (LJ) with a jump in engine torque at constant engine speed on the test bench.

4.3 Preparation of the measured data The aim in selecting training data must be to cover all operating conditions relevant to the later model. If later model excitations lie between the training data, they can be interpolated, but extrapolation outside the trained operating range usually fails. The load jumps were recorded at a sampling rate of 100 Hz and then sampled down to 10 Hz for further processing. The cycles were recorded at 10 Hz. In order to limit the complexity of the models, it was decided not to consider the cold start phase at first. In these first seconds after the start of the combustion engine, increased emissions occur because both the components and the lubrication system have not yet reached the operating temperature. Real measured data are subject to outliers which, from a statistical point of view, do not meet the expectations with regard to the other series of measurements. Depending on the model approach, such outliers make modelling more difficult or may even lead to the fact that no usable model can be found. Smoothing the measured values with a filter approach is therefore essential. In the presented study, the data were smoothed with a moving average filter after evaluating several filters. Due to the high temporal resolution of 10 Hz, no information is lost even in the area of the particularly relevant NOx peaks, but only noise influences are suppressed. Then the data is scaled to the interval between 0 and 1, as this makes them easier to process for machine learning procedures. In the context of data-based modelling, the available data is usually divided into training, test and validation data. In the presented case, the four driving cycles and the WLTC were divided into four data blocks in order to be able to cross-validate the model approaches applied to the data. This procedure allows a better picture of the quality of the models for different input data. Since the data are time-dependent, it is always necessary to use coherent data sets for the individual blocks. In addition, the retrospective time must be taken into account when forming the data blocks.

271

Using machine learning methods to develop virtual NOx sensors for vehicle …

5 Presentation of model approaches Models have the goal of replacing real processes with mathematical descriptions. Depending on the planned application of the models, certain phenomena are purposefully neglected in order not to increase the complexity of the model unnecessarily. This paper uses data-based model approaches from the field of machine learning. Machine learning itself is a subfield of artificial intelligence. While classical programming techniques apply human-defined rules to input data, machine learning algorithms are able to recognize patterns and laws in data. Thus the statistical or mathematical models created can learn rules (training), which can then be applied to unknown input data. The database (training data set) required for the model training must be correspondingly large and completely map the input combinations to be learned. The modelling of the nitrogen oxide concentration is to be carried out as a regression problem. In contrast to fuel consumption and CO2 emissions, exhaust gas components such as nitrogen dioxide cannot be modelled statically and thus on the basis of a map. This is due to the particular relevance of dynamic operation for these emissions, which is expressed by the fact that the highest concentrations occur in these situations. Herzig suggests a maximum look-back time of 3 s for a measuring point of the nitrogen oxide concentration between turbocharger and oxidation catalyst to represent the dynamics [4]. For dynamic models, this means that the output variable depends not only on the current input variables but also on past input variables, thus giving the system a memory. The realization is done by the approach of external dynamics in a kind of memory structure.

5.1 Selected model approaches The selection of the model approaches is on the one hand based on approaches used in the literature for similar problems [5] and on the other hand on the possibility of being able to depict nonlinear dynamic regressions with increasing model complexity. For the model approaches with external dynamics the Support Vector Machine for Regression (SVR) and the LOcal LInear Model Tree (LOLIMOT) were selected. Basically, almost all static model approaches can be used to map external dynamics by realizing the dynamics through memory and feedback. The SVR was chosen as the limit for the lower model accuracy due to its low complexity and the associated short calculation time. For the training of many nonlinear models, nonlinear iterative optimization methods are necessary, which are often accompanied by high computation times. For this reason, the local LOLIMOT algorithm is used as a second approach with external dynamics, which approximates local linear models and can thus be trained with linear optimization methods. The Long-Short-Term-Memory-Network (LSTM) represents a global

272

Using machine learning methods to develop virtual NOx sensors for vehicle … approach with internal dynamics and has been chosen due to the great success in the recent past with similar problems. Details of the model approaches can be found in the corresponding literature [6, 7, 8]. The Python library scikitlearn was used to calculate the SVR approach, while the LOLIMOT models were created using the Python implementation https://github.com/ 4204015/Trajektorienplanung/blob/master/lmntools.py. For the LSTM network the Python Deep Learning library Keras was used.

5.2 Selection of model input parameters Data-based modelling approaches draw their knowledge purely from the data they are trained with and derive their model parameters from this data. The selection of the model inputs has a direct effect on the training duration and the quality of the model. Accordingly, not an unlimited number of input variables should be used when creating a model. A pre-selection of the input variables is therefore always recommended. If the selection process happens automatically, the term Feature Subset Selection is generally used in the context of machine learning. Since in this case models with external dynamics (SVR, LOLIMOT) and internal dynamics (LSTM) are considered, the time delays of the input variables must be selected in addition to the actual selection of the input variables for the models with external dynamics. Table 2: Model inputs with minimal time shift and selected time shift Model input

min. time shift 𝚫𝝉𝒎𝒊𝒏 Pedal position 6 Engine speed 14 Intake air mass flow 8 Pressure suction pipe 9 Temp. near intake manifold >30 Temp. near exh. manifold >30 Cylinder head temperature 26 Setting high pressure EGR 3 Low pressure EGR setting 4 Lambda value exhaust gas 9 Injection times (pre-; main-; 4, 5, 3 post-)

SVR

LOLIMOT

LSTM

1, 7 10 -; 1,9; 3

1, 7 1, 15 5, 21 1, 15 3, 7 13 1, 10 -

1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20 1, 2, …, 20

273

Using machine learning methods to develop virtual NOx sensors for vehicle … Willing to limit the large number of measurement signals, a selection of signals according to expert knowledge was made in a first step. This is based on technical documentations and publications on dynamic modelling of nitrogen oxide emissions. In order to ensure an onboard functionality of the model to be developed, the focus of the investigations will be on the use of sensor variables of the engine control unit, whereas external signals of the test bench will not be used. For this reason, other emission signals (HC, CO2) will also not be used as model input for the NOx model. The exhaust gas composition is only monitored via a lambda sensor behind the turbocharger. The measured variables shown in table 2 were selected as possible model input variables. Engine speed and pedal position are the basic parameters that determine the operating point of the combustion engine. The formation of nitrogen oxide emissions is largely determined by the intake products air and fuel and the boundary conditions prevailing during combustion. These are represented by the temperature sensors on the intake and exhaust manifold and in the cylinder head. Exhaust gas recirculation systems are used to reduce nitrogen oxide within the engine. These systems return exhaust gas to the combustion chamber, thus lowering the temperatures in the combustion chamber, which in turn inhibits NOx formation. In a further step it was examined whether there are stationary signals. In addition, quantities which show a strong correlation (𝑅 > 0,9) with others were removed by means of a correlation matrix based on forumla 1, as these do not indicate any gain in information. 𝑅 𝑥, 𝑦 =

∑ ∑

( (

̅ )( ̅) ∑

) (

)

(1)

Correlating courses were found for the pre-injection and main injection times in the different cylinders, so that only cylinder one was considered. With the current selection of variables according to table 2, the focus was on 13 signals from which 390 possible input variables can be derived to represent the dynamics at a sampling rate of 10 Hz and a review time of 3 s. Since the model approaches with external dynamics can hardly reasonably process such a large number of input variables, the final input data set should contain significantly fewer signals. An applied Feature Subset Selection ensemble of filter, wrapper and embedded approaches [9] did not show the desired results in the present case. Based on these findings, the following alternative procedure was developed in Herzig between [4]. In a first step, it is useful to determine a minimum time difference Δτ the undelayed 𝑥(𝑡) and the derived signal 𝑥(𝑡 − 𝜏) of the possible input variables by means of an autocorrelation function. To limit the complexity of the resulting models and still cover a wide time horizon, the autocorrelation threshold should be set to 0.8.

274

Using machine learning methods to develop virtual NOx sensors for vehicle … The time shift can take integer values between 1 ≥ 𝜏 ≥ 30, which results from the maximum look-back time of 3 s at a recording frequency of 10 Hz. Table 2 shows the resulting minimum time differences for the selected input variables. Due to the inertia and a resolution of 1°C of the temperature sensors, most temperature signals exceed the maximum look-back time with respect to their minimum time differences. It is therefore sufficient to consider one time difference per temperature signal. To find the ideal time delays, three time delay strategies were examined. In variant 1, the one with the smallest mean model error (MES, see chapter 6) is selected as the first time delay and then the minimum time difference of the signal concerned is added for the other relevant time delays. For the first time shift, variant 2 proceeds as variant 1 and selects the others on the basis of the minimum model error, while ensuring that the minimum time difference is observed. Variant 3 uses exactly one time step for the first time shift and then adds the minimum time difference for the others. The final selection of the input variables for the SVR and LOLIMOT is done automatically using the wrapper technology Feed Forward Selection. In several iterations, the inputs that improve the model are selected on the basis of the previously selected input variables and the possible candidate data. The evaluation criterion is the same as the MSE using the four cross-validated models per approach. If there is no improvement over three iterations, the calculation is aborted. Since the optimal model structure can vary depending on the input variables in the LSTM approach, the Feed Forward Selection would be very computationally intensive and the model is trained with the complete set of input variables. Table 2 shows the selected inputs depending on the model approach. It can be seen that after only five iterations no model improvement could be achieved for the SVR. LOLIMOT, on the other hand, can process significantly more input variables in a meaningful way.

6 Evaluation of the model results The chapter evaluates the model approaches with the input variables determined according to Table 2, both on the basis of statistical parameters and the time course of the model predictions. The mean square error (MSE) is used to evaluate the model quality. A model with a smaller MSE can be assumed to be the better one. It is calculated from the arithmetic mean of the quadratic deviation between the estimated values 𝑦 and the measured values 𝑦 (cf. formula 2). The MSE weights larger deviations more heavily than smaller ones, but due to the squaring it is not very suitable for interpreting the actual model error.

275

Using machine learning methods to develop virtual NOx sensors for vehicle … 𝑀𝑆𝐸 = ∑

(𝑦 − 𝑦 )

(2)

Another evaluation criterion for regression problems is the coefficient of determination 𝑅 , which corresponds to the square of the correlation coefficient 𝑅 (cf. formula 1) and takes on values in the interval [0, 1]. If all errors are zero, 𝑅 takes the value 1. Small prediction errors mean a small dispersion between estimated and measured values and thus a high coefficient of determination. The accuracy of predictions can be expressed with the mean absolute error MAE. It calculates the arithmetic mean over the amounts of the deviations between the estimated and measured values (cf. formula 3). 𝑀𝐴𝐸 = ∑

|𝑦 − 𝑦 |

(3)

Since the training and validation data were scaled to the interval [0, 1], the normalized MAE (NMAE) is used in this paper. A model can be described the better, the closer it comes to a NMAE of zero and a coefficient of determination of one. In addition, the error integral over all measuring points was evaluated for model evaluation. For this purpose, the magnitude of the error between measured value and prediction is added up.

6.1 Validation results The worst model quality shown in table 3 results for the SVR approach, which was used as the basic model in this paper. In all three statistical parameters it shows worse values than the LOLIMOT and LSTM approaches. However, the training time of the SVR is also significantly lower than with the other variants. It should also be noted that the Feed Forward Selection of the input variables (see table 2) selected the lowest number of signals for the SVR. The special suitability of the LOLIMOT models for internal combustion engine processes as already shown in [10] is again apparent here. LOLIMOT achieves the best model quality, but the algorithm also calculates a very complex model, which accordingly takes a long time. Despite the fact that the LSTM network has much more knowledge about the dynamic processes due to the inclusion of all time-delayed signals of the input variables, no better model quality is shown. However, it is to be expected that a larger amount of input data, as it is usual for neural networks, can still deliver significantly better results. An effect of the load jumps on the model quality cannot be proven on the basis of the statistical key figures. The models with a reduced parameter set, which were only trained with the basic input variables pedal position and motor speed, show a worse model quality except for the LSTM approach.

276

Using machine learning methods to develop virtual NOx sensors for vehicle … Table 3: Validation results Model

SVR SVR without LJ SVR red LOLIMOT LOLIMOT without LJ LOLIMOT red LSTM LSTM without LJ LSTM red

Inputs/ Estimators 4/5 4/5 2/4 7 / 13 7 / 13 2/4 13 / 130 13 / 130 2 / 20

𝑹𝟐

NMAE MSE

Training time [s]

63,93 63,36 50,93 84,05 84,2

0,0386 0,0427 0,047 0,0232 0,0259

0,00267 0,00331 0,00364 0,00118 0,00142

173 173 180 5113 4431

Error integral [ppm] 2202707 2207174 2683282 1327455 1337178

69,73 68,19 77,00 81,16

0,0334 0,0323 0,0292 0,0242

0,00224 0,00236 0,00208 0,0014

694 1452 1353 1164

1904843 1848431 1510433 1383548

6.2 Dynamic prediction The LOLIMOT model was selected for the observation of the dynamic subsequent behaviour. Figure 3 shows the predicted courses for the model trained with the full set of parameters, the model trained without load jumps and the model trained with the basic input values in comparison with the measured values. Additionally the plot shows the vehicle speed, the accelerator pedal position and the rpm. For a better overview, only a section of approx. 250 s from the validation run is shown. It can be seen that the course of the predicted NOx concentration follows the measured values well overall. As expected, there is almost no deviation of NOx concentrations in stationary areas, whereas speed and accelerator pedal gradients show increased differences. The NOx peaks are well hit in terms of time, although the model underestimates the height of the peaks depending on the type of training. For comparison, a map-based prediction signal is plotted in the upper plot, which shows already at first glance that such an approach is unsuitable for the present problem. Compared to the other model approaches, LOLIMOT is particularly good in these dynamic areas with high NOx concentrations, which was also shown by a study that excluded these areas, since it was hardly possible to achieve better model quality in these areas. A model training which includes even more of these highly dynamic operating point changes of the combustion engine could probably improve the models even further.

277

Using machine learning methods to develop virtual NOx sensors for vehicle … The model with a reduced parameter set shows especially the dynamic ranges in a worse way, whereas the model trained without load jumps predicts these peaks even a bit better than the model trained with load jumps. This is an area where further investigations will have to provide clarity in the future.

Figure 3: Time sequences of the validation

7 Summary and outlook In this paper, the development of a virtual NOx sensor using data-based modelling of raw nitrogen oxide emissions for diesel engine propulsion was investigated. A possible application for such a sensor is, for example, the operation of the SCR system in exhaust gas aftertreatment. The Support Vector Machine for Regression (SVR), the LOcal LInear Model Tree (LOLIMOT) and the Long Short Term Memory (LSTM) were used as model approaches. The database used was generated by measuring a small set of real driving profiles on a highly dynamic combustion engine test bench. Within the scope of an extensive preprocessing of the data, the database was prepared for a use in dynamic models. In the

278

Using machine learning methods to develop virtual NOx sensors for vehicle … present case, dynamic means that past points in time of the input variables are also included in the model. In order to make the number of model inputs manageable with regard to the calculation effort, individual time shifts of the input variables must be selected for the SVR and LOLIMOT models with external dynamics. As a result, the LOLIMOT and LSTM approaches in particular show a good model quality with a coefficient of determination of 0.8 or more. This should also be evaluated with regard to a training data set that is small for machine learning problems and initially created without much effort. More extensive data sets for model training and a further optimization of the model parameters lead to the expectation of even better model quality. It is also possible to predict the emission instead of the concentration via the exhaust gas mass flow rate. This seems to be particularly interesting in the sense that by integrating the exhaust gas aftertreatment into the model section, a prediction of the endof-pipe emissions seems to be within the realm of feasibility. It can be assumed that the approach can be quickly transferred to different types of internal combustion engines or other emission components. The virtual sensor is therefore interesting in terms of cost-effective use and low-maintenance operation compared to classic hardware sensors.

Bibliography 1. Günther Peter Merker, Christian Schwarz, Gunnar Stiesch, Frank Otto, „Simulating Combustion“, Springer, 2005 2. VW, „ Die neue Dieselmotoren-Baureihe EA288“, VW Selbststudienprogramm 514, 2013 3. Roman Ließner, Robert Fechert, Bernard Bäker, „Derivation of Real Driving Emission Cycles Based on Real-World Driving Data“, 3rd International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS), Porto, 2017 4. Paul Herzig, „Methoden maschineller Lernverfahren zur Abschätzung des Emissionsverhaltens von Fahrzeugen“, Master thesis Technische Universität Dresden, 2019 5. Christoph Hametner, Christian Mayr, Stefan Jakubek, „ Dynamic NOx emission modelling using local model networks“, In: International Journal of Engine Research, Vol. 15(8), 2014 6. Alex J. Smola, Bernhard Schölköpf, „ A tutorial on support vector regression“, In: Statistics and computing, 14.3, 2004

279

Using machine learning methods to develop virtual NOx sensors for vehicle … 7. Oliver Nelles, „ LOLIMOT-Lokale, lineare Modelle zur Identifikation nichtlinearer, dynamischer Systeme“, In: at-Automatisierungstechnik, 45.4, 1997 8. Sepp Hochreiter, Jürgen Schmidhuber, „LSTM can solve hard long time lag problems“, Advances in Neural Information Processing Systems 9 (NIPS), Denver, CO, USA, 1996 9. Verónica Bolón-Canedo, Amparo Alonso-Betanzos, „Recent Advances in Ensembles for Feature Selection”, Springer, 2018 10. Rolf Isermann, „Elektronisches Management motorischer Fahrzeugantriebe“, Vieweg+Teubner, 1. Auflage, 2010

280

Image-based condition monitoring of a multiLED-headlamp Pascal Janke, Jiayu Cai HELLA GmbH & Co. KGaA, 59552 Lippstadt, Germany Dr. Mathias Niedling L-LAB, Research institute for automotive lighting and mechatronics, 59552 Lippstadt, Germany Univ.-Prof. Dr.-Ing. Prof. h.c. Dr. h.c. Torsten Bertram Technical University Dortmund, Institute of Control Theory and Systems Engineering, 44227 Dortmund, Germany

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_28

281

Image-based condition monitoring of a multi-LED-headlamp

1 Introduction LED-Matrix headlamps use single controllable LEDs to realize AFS lighting functions (advanced frontlighting system) such as glare-free highbeam. In those modules, each LED is responsible for a specific illumination angle on the street. By European law, in case of a failure of one LED, the driver must be notified [1]. Usually, LED monitoring is achieved by measuring the voltage and current [2]. In the common case of a series connection of the LEDs, it is not possible to detect malfunctioning of a specific LED exactly because the measurement of the operating parameters can only detect a cumulated discrepancy [3]. Moreover, it is not possible to measure the luminous flux of every single LED to get information about the current status of each emitter. Knowing not just the state (on / off), but also the exact luminous flux has a lot of advantages. For example, it is possible to adjust the LED's brightness in dependence on the current temperatures. In cold winter, you could have a brighter street illumination. In hot summer evenings, the LEDs could be protected from high temperatures and, by this extend the lifetime and prevent possible mortality. Especially the last is an important step towards sustainability. The objective of this work is to achieve accurate monitoring of the state and exact condition of each LED of the Matrix-headlamp system. Therefore, an image-based monitoring-procedure using a CMOS sensor is investigated. Such a procedure requires image processing techniques considering the high effort in computer vision. In the following, the approach using a camera for monitoring single LEDs of a Matrix headlamp is presented. Also, a first approach is made for detecting the exact condition of the LEDs and compare it to the desired target value. In the beginning, an overview of the current state of the art on LED-monitoring and image processing techniques is given. After-wards the experimental setup and the results are presented.

2 State of the art 2.1 LED Monitoring The LEDs used in a headlamp are subject to degradation, as all LEDs. It is distinguished as long-term and temporary degradation. The long-term degradation depends on the lifetime of the LEDs and their operating parameters. High power LEDs last 12,000 to 30,000 hours. This means, after that time, the LEDs have a luminous flux of 70% of a new LED. The lifetime is highly dependent on the current and the temperature. High temperatures are also the reason for the temporary degradation. It is possible that an increasing temperature from 25°C up to 120°C, which is the highest temperature permitted in a headlamp because of the soldering, will result in a decreasing luminous flux

282

Image-based condition monitoring of a multi-LED-headlamp up to 80% of its origin. High temperatures also increase the possibility of a total failure of the LEDs [4]. As mentioned in the introduction, total failure has to be detected, and the driver must be notified. But the detection of the exact condition of the LEDs, caused by long-term and temporary degradation is the goal. By detecting a degradation, it is further possible to react to this and, for example, control the LEDs by its actual luminous flux. This way a perfect road illumination for the driver can be guaranteed, as well as avoiding further damage. Suitable for this are the use of various photodiodes with appropriate evaluation [5]. Photodiodes are an accurate and cheap way for monitoring light, but they cannot simply distinguish between the light sources whereas image sensors like a CMOS or CCD sensor have directional information because of their array of photodiodes. Of course, an optic is needed in front of the sensing chip, which often can be ordered in a fixed combination. For the purpose of monitoring the LEDs of a headlamp, a CMOS sensor has the most advantages and will be used.

2.2 Image processing techniques For the purpose of monitoring all, in this case, 84 LEDs and distinguish them, several segmentation techniques are needed. Morphological operators must be applied for creating a region of interest. In this region, the greyscale value of the picture is the base for a comparison to the not degraded, healthy state.

2.2.1 Grayscale Most camera sensors are taking pictures with color information. Those RGB pictures are measured by color filters inside the sensor. For this work, information about light intensity is needed. Therefore, the RGB picture needs to be converted into a grayscale picture. This is achieved by using formula (1), which was introduced as the NTSCstandard in 1953. Y is called luma and is calculated from the pre-corrected non-linear chromaticity signals. Luma represents the brightness of a picture [6]. 𝑌 = 0.299 × 𝑅 + 0.587 × 𝐺 + 0.114 × 𝐵

(1)

2.2.2 Otsu thresholding method The simplest segmentation methods are based on global or local threshold values. By specifying or calculating a threshold 𝑇, a grayscale image 𝑓 𝑥, 𝑦 transfers into a binary image 𝑏 𝑥, 𝑦 . The affiliation of an image point to a segment (0 or 1) is decided by comparing the grayscale value with a threshold value.

283

Image-based condition monitoring of a multi-LED-headlamp 𝑓 𝑥 =

1, for 𝑓 𝑥, 𝑦 0, for 𝑓 𝑥, 𝑦

T T

(2)

If the binarization is based on the entire image used to calculate the threshold, we talk about global thresholding. Otherwise, we talk about local thresholding. But it is difficult to calculate the threshold optimum for the respective application. Most methods calculate the threshold according to a target criterion. The target criteria are often based on the grayscale histogram, see figure 1. The most common method to determine the optimal threshold value is Otsu thresholding. The entries of the histogram are expressed as probabilities of a discrete random variable is taken, and the grey value is determined for which, after an analysis of variance, the histogram is best separated into two connected parts [7].

Figure 1: Calculation of the threshold based on the grayscale histogram: (left) origin grayscale picture; (middle) histogram of the origin picture; (right) picture after segmentation. [8]

2.2.3 Erosion and Dilatation The basic procedures of morphology are erosion and dilation. Morphological methods are suitable for various tasks of image evaluation: pre-processing, noise suppression, edge extraction and detection. The methods require already segmented images. The erosion of a set 𝐺 by a structuring element 𝑆 completely contained in 𝐺 is defined as the pixelwise displacement of the element over the whole image. A reference point is defined for the element. If the structuring element fits completely into 𝐺, the pixel of the image at the position where the reference point of the element is located belongs to the eroded set 𝐺 ⊖ 𝑆. Similarly, the dilation of a set 𝐺 by a structuring element 𝑆 completely preserved in 𝐺 is defined as the displacement of the element 𝑆, but unlike erosion, the pixel of the image at the location where the reference point of the element is located belongs to the dilated set if there is a pixel of the structuring element that fits in 𝐺. I.e. erosion reduces the set, while dilation increases the set. Figure 2 illustrates how erosion and dilation of the set 𝐺 by 𝑆 works.

284

Image-based condition monitoring of a multi-LED-headlamp

Figure 2: Illustration of erosion and dilation: (top) Original set G, structuring element S and the reference point for S; (middle) Result of the erosion of G by S; (bottom) Result of dilation of G by S. [9]

The operation opening is defined as the concatenation of an erosion and a subsequent dilation. The operation with reverse order is defined as closure.

2.2.4 Contour tracking procedure The classical contour-following method is the best known and most important elementary segmentation method [10]. This is later used to check how many regions are segmented and verify if 84 regions are found.

3 Experimental setup In the presented investigation the HD84 Matrix headlamp, which is the current state of the art, is used as a test object. This headlamp consists of the light source unit with the LEDs and a heatsink, a primary (total internal reflection) optic in front of the LEDs, a lens holder and a secondary (collimator) optic. The LEDs are aligned in 3 rows. As seen in figure 3 (right), the first row and half of the second row have the purpose of the low beam. All three rows are used for a high beam light distribution. This module needs an additional light source for the prefield illumination, which will not be further considered in this work. In figure 3 (left), the basic structure of the HD84 headlamp is shown. For development, a Raspberry Pi 3B+ is used in combination with the OpenCV database.

285

Image-based condition monitoring of a multi-LED-headlamp

Figure 3: HD84 Matrix-LED system: (left) Main module with the 3 LED rows; (right) illumination areas of the LEDs. [11]

3.1 Camera and positioning The camera sensor chosen is the Raspberry Pi Camera Module V2.1. It uses the Sony IMX2019 sensor with 8 megapixel resolution. It has been chosen because its size is fitting in the corresponding position in the headlamp. Also, it has the desired field of view (FOV). In addition its settings, like exposure time and ISO-settings, can be controlled very well and without further implementation effort. The camera must not interfere with the light beam. But it has to be able to detect every LED. Thus, the camera is placed as shown in figure 4. It is between the primary and secondary optics. There, according to light simulations, only stray light from the primary optics should be detected from the camera. To confirm the simulation, the prototype has been tested with a luminance camera and compared to its origin.

Figure 4: Positioning of the camera on the main module. The circuit board is placed outside of the lens holder and the camera inside is facing towards the primary optics under a specific angle.

286

Image-based condition monitoring of a multi-LED-headlamp

3.2 Computation process For automated condition monitoring during an operating headlamp, several functional procedures are required. At first, the correct exposure times for the images need to be found. Then the picture has to be divided into sections according to the corresponding LEDs. Also, the image noise must be examined to avoid wrong intensity information. At last, the grayscale values for all LEDs have to be detected and saved as the reference values. In the following, only the highbeam light distribution will be evaluated, since in this state all LEDs are operating. But the procedure for every other light distribution is the same.

3.2.1 Exposure time Finding the correct exposure time for the images is essential. To do this, the exposure time must be higher than 5 ms to avoid flickering of the image. The upper limit must be determined to avoid overexposure. Finding the correct time is implemented in an automated process: while operating with the desired light distribution, in this case highbeam, the exposure time is extended until the highest grayscale values of 255 is reached. The exposure time before will be saved. But with this exposure time, not all LEDs can be detected. The LEDs on the far sides of the module are operating at a much lower PWM duty cycle than the ones in the middle. The next step is to reduce the exposure time until all LEDs can be detected. The exposure time beneath that is the time which will be saved. In this setup two different exposure times are needed to detect all LEDs. Figure 5 shows the separation of the LEDs to the corresponding times. Figure 6 (left) shows the picture taken by the camera at 37 ms exposure time.

Figure 5: Separation of the LEDs to their corresponding exposure time for detection. The numbers are the label of each LED.

287

Image-based condition monitoring of a multi-LED-headlamp

3.2.2 Separation of LEDs and region of interest After setting up the camera, the healthy state information of all LEDs is available at previously found exposure times. Then for each LED a pixel is positioned whose grayscale value changes with the change of state of this LED. This value is the criterion for condition monitoring. According to the Lambertian model, the relationship between the illuminance on a pixel and the luminous flux of an LED can be modeled [12].

Figure 6: Finding the region of interest (ROI) of LED No. 15: (left) picture of all LEDs operating; (right) binary picture after applying the Otsu thresholding method and closure for LED 15.

To apply the Lambertian model, it is necessary to detect the region of interest for each LED on the pictures. Therefore 84 pictures are taken with just one LED operating. This picture is transferred into grayscale and segmented with the Otsu thresholding method. Afterward, the morphological operation closure is applied to get a binary picture as shown in figure 6 (right) for the example of LED No. 15. Closure is needed to minimize the overlapping of adjacent LEDs. The pixels are saved as the region of interest (ROI) for each LED.

3.2.3 Image noise and dark current In general, image noise is different between random and local noise (so-called pattern noise) like the Fixed Pattern Noise and Photo Response Non-Uniformity (PRNU) and Dark Signal Non-Uniformity. The random noise is represented by 100 images in the same LED state and also under the same LED temperature of 25°C. Figure 7 illustrates the distribution of 100 grayscale values from the pixels for LED No. 15. With the data, a normal distribution is approximated in each case. Since determining the exposure time and positioning of the pixel, the random noise is already taken into account, the standard deviation of the noise compared to the grayscale value difference of the pixel between on and off state of the LED is very small. This means that the random noise when monitoring the on/off state of an LED is not important. Fixed Pattern Noise and Photo Response Non-Uniformity are the image noise, which always occurs only on certain pixels. When the pixels are positioned, the pixels with disturbing Fixed Pattern Noise and the Photo Response Non-Uniformity are already eliminated.

288

Image-based condition monitoring of a multi-LED-headlamp

Figure 7: Analysis of the random noise for LED No. 15. The sum of grayscale values for the according ROI is shown for 100 pictures taken.

Figure 8: Grayscale offset for LED No. 15 for its ROI in connection to the LED temperature and the resulting dark current noise.

289

Image-based condition monitoring of a multi-LED-headlamp However, the noise caused by dark current, which is closely related to the temperature is disturbing, because the temperature changes dramatically in highbeam mode. Figure 8 shows the grayscale offset values of the ROI for LED No. 15 with rising LED temperature when all LEDs are on. Those values are the mean values subtracted by the value at 25°C. This will later be taken into account as an offset.

4 LED condition monitoring With this experimental setup, it is now possible to get grayscale values that are linked to a corresponding LED. The next step is to determine the threshold corresponding to the LEDs condition.

4.1 LED state For detecting the on and off state of each LED, it is necessary to determine the threshold for each state. Because there are only two possible states, the Lambertian model can be modified. Therefore, the model is simplified by assuming only one LED is mainly responsible for the illumination of a specific pixel on the camera sensor. The other LEDs have just a small influence because they are further away. For getting the threshold, two pictures are made. One with just one LED operating and one with all LEDs operating but not the one before. Looking at the maximum value in the corresponding ROI of the first picture pictures and the minimum value of the second picture, a high difference can be monitored. The average of these two values will be stored as the threshold. Because of just the two possible states, it is sufficient to use only the maximum values in one region. This saves computation time.

4.2 LED luminous flux To detect the luminous flux multiple, but finite states are defined. This is necessary because the LEDs in a headlamp are controlled by the controller area network (CAN). When the headlamp control unit gets a CAN value for an LED, it is controlling and dimming the LED with PWM. In this case, the CAN values for the LED have a maximum of 64 discrete states with a step size of 1.6. The CAN to PWM ratio has an exponential factor to counteract the non-linearity of the human perception of brightness. This can be seen in the grayscale values for LED No. 15 in figure 9. The figure shows the mean values with an interval of 3 times the standard deviation. By this, the measured grayscale value has a probability of 99.7% to be in this state. This interval is needed because of the random image noise and therefore the mean value is not enough as the threshold. The CAN value maximum in figure 9 is 86.4 due to its standard value for a highbeam light distribution.

290

Image-based condition monitoring of a multi-LED-headlamp For example, fi a measured grayscale value (G2 in the figure) lies within an interval, the monitoring program outputs the corresponding CAN signal. If the measured value lies in several intervals (G3) or a gap (G1), the program outputs a corresponding CAN signal interval. This is done by the program for every LED. The total computation time for a highbeam light distribution, including taking the two pictures, is approx. 8 s. with this hardware.

Figure 9: Grayscale interval of every CAN value for LED No. 15, with three example grayscale measurements G1 - G3.

5 Validation The validation of the image-based condition monitoring of the LEDs is tested by comparing the input values of the LEDs with the output values of the algorithm. The program itself just gives output values, if they are differing from the standard highbeam light distribution. Otherwise, it gives the information that every LED is in a normal state.

5.1 LED failures Most important for the headlamps is to monitor the state on or off for each LED, because an unwanted off-state means a malfunction and has to be reported to the driver. Thus, the validation is done by switching off several LEDs and monitor the output of the program.

291

Image-based condition monitoring of a multi-LED-headlamp Table 1 is showing the comparison of this evaluation. It shows which LEDs were shut off, in which temperature range the validation took place, how often this scenario has been tried, and at last how often the program detected the correct state. Mostly the scenarios have been evaluated five times in a short time. But for some scenarios there have been many more tests; those are highlighted. The reason was to test the program under higher temperatures, so these have been evaluated up to 90°C. As seen in table 1, every single failure was reported correctly by the program, even at high temperatures. Table 1: Validation of the LED state for different LEDs shut off Number of the LED(s) shut off1

Temperature interval

Amount of evaluations

Amount of correct detections

3 56 73 4 + 82 33 + 59 42 + 43 33 + 73 28 + 56 29 + 58 + 83 32 + 34 + 60 14 + 15 + 16 28 + 32 + 71 17 + 58 + 61 1 + 43 + 83 3 + 55 + 84 60 + 70 + 72

[38°C 90°C] [34°C 36°C] [29°C 32°C] [37°C 39°C] [30°C 90°C] [29°C 90°C] [37°C 39°C] [36°C 37°C] [31°C 90°C] [37°C 38°C] 38°C 38°C [38°C 39°C] [34°C 36°C] [37°C 38°C] [36°C 37°C]

91 5 5 8 92 103 8 5 92 5 5 5 5 5 5 5

91 5 5 8 92 103 8 5 92 5 5 5 5 5 5 5

5.2 LED degradation For evaluation, a LED from the corner (LED No. 4) and a LED from the middle (LED No. 15) are used and presented. The specified CAN signals and the program outputs those LEDs are shown in tables 3 and 4. Here the program outputs correspondingly detected CAN values at different specified CAN input values. The tests are also performed at different temperatures. The temperatures at which the program is validated are listed in temperature ranges. During validation, it can be seen that the result of the 1

According to figure 5

292

Image-based condition monitoring of a multi-LED-headlamp program often fluctuates. The reason is that the result for monitoring the same given CAN value is different at different temperatures due to the image noise. In addition, the program gives only one interval as a result of low CAN signal levels according to figure 10. Therefore, the minimum and maximum detected CAN signal from the results of the program are listed as an interval in the tables. The deviation between specified and detected CAN values are shown in the last column. The smaller the deviation, the more precise is the detected CAN signal. Table 2: Validation of the exact condition of LED No. 4 Target CAN value

Detected CAN value

Temperature interval

Deviation

41.6 41.6 33.6 22.4 19.2

[38.4 40.0] [36.8 38.4] [25.6 30.4] [16.0 17.6] < 16.0

[37°C 39°C] [39°C 61°C] 39°C [38°C 39°C] [34°C 38°C]

[1.6 3.2] [3.2 4.8] [3.2 8.0] [4.8 6.4] [3.2 19.2]

Table 3: Validation of the exact condition of LED No. 15 Target CAN value

Detected CAN value

Temperature interval

Deviation

88.0 83.2 72.0 65.6 65.6 57.6 46.4 46.4 38.4 22.4

88.0 [83.2 84.8] 72.0 64.0 [60.8 64.0] [49.6 52.8] [32.0 38.4] [22.4 36.8] [16.0 25.6] < 16.0

[36°C 39°C] [33°C 39°C] 39°C [35°C 39°C] [39°C 60°C] [31°C 38°C] [31°C 39°C] [39°C 60°C] [38°C 39°C] [34°C 38°C]

0 [0 1.6] 0 1.6 [1.6 4.8] [4.8 8.0] [8.0 14.4] [9.6 24.0] [12.8 22.4] [6.4 22.4]

Both tables are showing, that a (virtual) degradation in the range of 5 – 30% of the LEDs is well detected, with just a small deviation. Dimming the LED further down, which represents a higher (virtual) degradation of the LEDs, result in higher deviations. The supposed reason is signal noise. The grayscale offset set against the dark current noise and in addition the grayscale interval, in order to compensate random noise, are not precise enough for a proper image noise compensation. But still, for a first approach, small degradations from (virtual) long-term degradation and as well as temporary degradation can be monitored with just a small deviation. In the case of LED No. 4, the

293

Image-based condition monitoring of a multi-LED-headlamp maximum difference for small degradation is a CAN value of 8. In this case that results in a PWM difference of 1.25%, which is a luminous flux of 1.88 lm. In the case of LED No. 15, until a (virtual) degradation of the CAN value of 65.6, there is no higher deviation than 1.6. Due to the exponential CAN to PWM ratio, small steps in high CAN values result in larger steps in PWM. Saying this, the deviation at CAN value 83.2 results in a PWM difference of 3.8% (5.7 lm). The deviation of 4.8 at CAN value 65.6 and high-temperature results in a PWM difference of 4.05% (6.1 lm). Out of a total of 150 lm, which each LED can generate, a maximum difference of 6.1 lm seem not much. But for a possible closed-loop control the deviation would cause too much fluctuation. So, further improvement needs to be done to monitor the exact condition of the LEDs.

6 Conclusion and outlook In this work, two types of image-based condition monitoring have been realized, namely the monitoring of the LED state (on/off) and the monitoring of the LED luminous flux. The monitoring of the LED state works exactly in all cases. The LED luminous flux monitoring runs with a small deviation when the LED is dimmed (up to 30%). But with lower dimming values therefore also lower luminous flux, the deviation increases. The advantage of this type of monitoring is that the LEDs can be monitored during runtime in the field. This makes it possible to implement the monitoring program in the ECU of the headlamp. This enables the headlamp itself to compensate for the degradation of the LEDs by increasing the CAN signal at the output of the program. Furthermore, the cost of a CMOS sensor is much lower compared to several photodiodes or other measuring devices. But there are also limitations when using this type of monitoring. First, various methods have been used to suppress image noise. However, the image noise, especially dark current noise, can falsify the result. Secondly, LED luminous flux monitoring is based on the assumption that only one LED degrades. Therefore, if several LEDs are degraded, the result is most likely distorted. To overcome the disadvantages of surveillance, a better camera with lower image noise can be used. It could also be attempted to use several cameras to monitor the LED luminous flux. In addition, the monitoring should be extended in case of degradation of several LEDs.

294

Image-based condition monitoring of a multi-LED-headlamp

Bibliography 1. UN Economic Commision for Europe: Regelung Nr. 48 der Wirtschaftskommission der Vereinten Nationen für Europa (UN/ECE). Einheitliche Bedingungen für die Genehmigung der Fahrzeuge hinsichtlich des Anbaus der Beleuchtungs- und Lichtsignaleinrichtungen, 2011 2. A. Herzog, T. C. Khanh. Die elektrische, thermische und optische Charakterisierung der Degradationsmechanismen von Leuchtdioden im UV-A Bereich. 13. Internationales Forum für den lichttechnischen Nachwuchs, Lux Junior 2017 3. T. Sutharssan, S. Stoyanov, C. Bailey, and Y. Rosunally. Prognostics and healthmonitoring of high power LED. Micromachines, 3(1):78–100, 2012. 4. T. Vos, P. den Breeijan und W. D. van Driel. “Quality and Reliability in Solid-StateLighting: Qua Vadis?”, In: Solid State Lighting Reliability Part 2. Springer International Publishing, 978-3319581743, 2018. 5. P. Janke, D. Kliebisch, B. Giepen, J. Holtmann, F. Ernst, T. Bertram, A. Trächtler. Condition Monitoring eines intelligenten Frontscheinwerfers. VDI Mechatroniktagung 2019, Paderborn, Deutschland. 6. T. Strutz. Bilddatenkompression, 5. Auflage, Springer Vieweg, 2017 7. N. Otsu. A threshold selection method from gray-level histograms. IEEE transactionson systems, man and cybernetics, 9(1):62–66, 1979. 8. R. Fisher, S. Perkins, A. Walker, and E. Wolfart. Hypermedia image processing reference. England: John Wiley & Sons Ltd, 1996. 9. J. Beyerer, F. P. León, and C. Frese. Automatische Sichtprüfung: Grundlagen, Methoden und Praxis der Bildgewinnung und Bildauswertung. Springer-Verlag, 2016. 10. H. Süße and E. Rodner. Bildverarbeitung und Objekterkennung. Springer Fachmedien, Wiesbaden, 2014. 11. F.-J. Kalze, U.-C. Knaack, and B. Böke. Mehrzeiliges Scheinwerfersystem in der Mercedes-Benz E-Klasse. ATZ-Automobiltechnische Zeitschrift, 118(2):16–23, 2016. 12. H. Yang, J. M. Bergmans, T. C. Schenk, J.-P. M. Linnartz, and R. Rietman. An analytical model for the illuminance distribution of a power LED. Optics express, 16(26): 21641–21646, 2008.

295

Measurement and testing of lidarsensors Andy Günther, B. Bäker IAD, TU Dresden

This manuscript is not available according to publishing restriction. Thank you for your understanding.

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_29

297

Credibility of software-in-the-loop environments for integrated vehicle function validation Indrasen Raghupatruni, Simon Burton, Marcus Boumans, Thomas Huber, Antonia Reiter Robert Bosch GmbH

Credibility of software-in-the-loop environments for integrated vehicle …

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_30

299

Credibility of software-in-the-loop environments for integrated vehicle …

Abstract Highly Automated Driving (HAD) technology is an enabler for innovation in the automotive industry through the integration of complex vehicle functions for comfort, safety and economy. However, the safety implications of these functions greatly increase the validation requirements as a pre-requisite for their release into the intended environment. Software-in-the-Loop (SiL) environments are emerging as an alternative to traditional testing approaches such as Hardware-in-the-Loop (HiL) systems. This is due to the strong evolution of vehicle simulation technologies as well as the virtualization of electronic control units, vehicle networks, vehicle simulation technologies coupled with the availability of increased computing power. In this paper we provide insights into a novel approach for arguing the credibility of SiL environments for validating the complex functional and non-functional requirements of HAD. These approaches include methodologies to extract test scenarios from real field-data and automatically evaluate their relevance, selection of test cases that validate the credibility of the given SiL environment and a unique metric to indicate the accuracy of the SiL environment in comparison to the reference system. With the help of a HAD use case we then demonstrate the advantages of this novel approach compared to methods currently employed in the automotive industry. Keywords: Software-in-the-Loop, Credibility of SiL, Software Verification & Validation, Highly Automated Driving, Real Field Data.

300

Credibility of software-in-the-loop environments for integrated vehicle …

1 Introduction Highly Automated Driving (HAD) functions are typical examples of a current trend towards the integration of complex vehicle functions for comfort, safety and economy. However, the safety implications of these functions greatly increase the requirements on verification and validation as a pre-requisite for their release into the intended environment. The lack of precise specification of the operational design domain as well as the lack of specific guidance in existing safety standards make the challenge of demonstrating the safety of such systems even more demanding. One approach to the verification and validation of such systems is to uncover critical scenarios through an analysis of field data collected during millions or billions of kilometers of vehicle road-based testing. This is an extremely time consuming and expensive process with few guarantees that an adequate coverage of critical situations has been achieved. For the residual failures rates required by HAD functions, measuring the probability of hazardous events through vehicle testing therefore does not scale, even with large fleets of continuously running vehicles. Software-in-the-Loop (SiL) based testing is emerging as an essential extension to traditional testing approaches such as Hardware-in-the-Loop (HiL) systems or vehicle prototypes. This trend is enabled by the evolution of vehicle simulation technologies, virtual electronic control units, vehicle networks and vehicle simulation technologies coupled with the availability of increased computing power.

Figure 1: Role of simulation for validation of HAD functions

The use of simulation when testing one or more components of a HAD system addresses the issues of controllability, observability and repeatability by stimulating the software

301

Credibility of software-in-the-loop environments for integrated vehicle … interfaces of the system within a controlled, synthetic environment [1]. By using synthetic test data, a greater coverage of the application domain and scenario classes can be achieved, including those conditions that are difficult or dangerous to reproduce in the real world, allowing a huge number of critical situations to be examined within a safe context [2], see Figure 1. If simulated tests can be demonstrated to be representative they can be used to form an argument to reduce the amount of driving hours required to form a statistical argument for freedom of unacceptable risk in the final system, thereby making the release argument for HAD more economically feasible. In the context of HAD, software updates for safety and security in the product operation phase have be provided for many years and this requires the test environments to be available for long periods, this is only possible with a fully virtualized approach, such as SiL. The use of SiL environments to validate HAD functions has been described in [3][4]. [3] in particular, provides insights into the use of a separation of concerns test methodology to create an optimal SiL environment for addressing functional and nonfunctional requirements. However, in order for simulation-based evidence to be used as part of the overall system safety argument and release procedure, the credibility of the virtual test environments in comparison with the intended operational domain context must be ensured. This paper is organized as follows. Section 2 describes the overall context of V&V evidence within an overall safety assurance strategy for HAD functions. Section 3 describes the need to argue the representativeness of the test evidence provided via simulation and provides the main contribution of this paper, which is a framework for creating such a confidence argument. Section 4 describes how such a confidence argument can be presented and under which conditions it would be valid. The framework is then illustrated using the example of a HAD function and reference platform in Section 5. The paper concludes with an evaluation of the approach, conclusions and further work.

2 The contribution of V&V Evidence to the Assurance Case The release of HAD for use on public roads must be supported by a rigorous and structured argument for its safety. Figure 2 illustrates a framework for a systematic approach to assuring the safety of HAD function and the key questions that should be answered: A systematic Domain Analysis forms the basis of an understanding of the environment in which the system should operate. Classes of scenarios in which the system should operate are defined and analyzed to identify dominant properties of the environment relevant to the safe operation of the system.

302

Credibility of software-in-the-loop environments for integrated vehicle … Domain analysis Which safety properties must be maintained within an open context environment?

Design for Safety

Assurance case

Which system design (e.g. sensor configuration) leads to the required performance?

How to argue a level of residual risk commensurate to societal and legal expectations?

V&V evidence Which V&V evidence can provide confidence in the argument for safe behaviour and tolerable residual failure rate?

Figure 2: Framework for safety assurance of Highly Automated Driving

A set of high-level safety goals for the system are then defined based on this domain analysis. This includes a systematic hazard and risk analysis involving not only a consideration of a failure in the function of the ego-vehicle, but also a systemic view of intrinsically hazardous conditions within the interaction between the ego-vehicle and its environment that must be avoided. Since the safety goals are defined based on this “determined context” for the vehicle operation, the validity of the safety assurance argument that is produced is therefore also restricted to this set of scenario classes and identified properties. The resulting scope is known as the operational design domain. Typical classes of scenarios included in the analysis could include for example, motorway driving at speed, in heavy traffic and relevant environmental properties may include, weather and lighting conditions, road surface as well as the behavior and type of other traffic participants. Ensuring that all “relevant” classes of scenarios and their dominant safety properties have been identified is a key task of the validation of such systems. The top-level safety goals of the system are also defined within the context of societal and legal expectations on safe driving behavior. The rigorous analysis of such contextual requirements therefore also forms an important component on delimiting the scope of the safety assurance argument. Design for Safety involves applying an iterative approach to refining the safety goals in coordination with system design decisions. At each level of refinement, assumptions made in the design are explicitly stated and analyzed, so that they may also be later validated. A functional and technical system design should be derived that is capable

303

Credibility of software-in-the-loop environments for integrated vehicle … of achieving the safety goals, even in the presence of inevitable insufficiencies and faults within individual components. Component insufficiencies may be, for example, limitations of particular sensor components in poor light conditions. Meeting the safety goals requires a means of determining and specifying which insufficiencies of individual components are acceptable in all parts of the Sense, Understand, Decide, Act chain. In addition, it must be determined how such insufficiencies either propagate through, or are minimized by, the chosen system design (e.g. through a combination of diverse sensor channels with non-overlapping insufficiencies). The system architecture must be robust both to functional insufficiencies and faults of the individual components themselves, as well as to the uncertainties inherent in the domain or sensing channels. The requirements on V&V evidence will dramatically increase for HAD systems in comparison to previous vehicle functions. In addition, evidence must be provided to validate that the scenario classes and associated properties provide adequate coverage of the real scenarios encountered in the target domain. Such validation is required at the system level as well as to confirm the assumptions made for each component. Due to the complexity of the domain and the impractical number of driving kilometers that would be required to provide statistically relevant test results [5], sources of validation other than standard vehicle-level system tests will be required. These may include simulation as well as the statistical analysis of large amounts of data captured in the field (e.g. during the operation of previous generations of a driving function). The simulation should be based on the same semantic data model as was used to describe the scenario classes and properties during the domain analysis. In this way, the validation activities (both virtual and in the real-world) can be actively used to iteratively refine the domain model and system design until a convincing argument for safety within the particular operational design domain can be made. By providing a structured argument that the HAD function achieves safe behavior for all conditions that meet the set of assumptions describing the target application domain, the Assurance case justifies that the residual level of risk associated with this function is acceptable. This will rely not only on technical, but also societal, ethical and legal considerations. This argument shall be supported by an evidence from the V&V, both for functional sufficiency of the HAD function for its determined context as well as the validity of the set of assumptions used to define the target application domain. This paper focusses on the contribution of V&V evidence to argument the assurance case of the HAD functionality.

3 Demonstrating the credibility of SiL Testing Results In order to be argue that the billions of kilometers of real-world driving required to create a release argument for HAD can be significantly reduced to a tractable level by

304

Credibility of software-in-the-loop environments for integrated vehicle … employing SiL approaches, the credibility of test and analysis results created in this manner must be carefully considered. This requires that the SiL environment must have a set of characteristics which directly relate to the V&V targets that have to achieved through simulation. Examples of such credibility characteristics have been listed in Table 1. Table 1: Credibility characteristics relating to the V&V target V&V Target Evaluation of the performance of the HAD decision making function.

Credibility Characteristics – Does the simulated behavior of the environment correspond to real-world scenarios, including edge cases? – What coverage of the scenarios found in the application domain can be achieved in the simulation.

Evaluation of the performance of the HAD perception function.

– How true is the simulated environment to that perceived by the system sensors in the realworld? – What coverage of environmental conditions that can have a detrimental effect on the perception performance can be simulated?

Evaluation of the performance of the HAD function within the target E/E Architecture

– Which properties of the target HW and SW environment are simulated and how accurately, e.g.: – Distribution of functions to tasks with an Operating System distributed across microcontroller or -processor cores. – Timing characteristics and data integrity across a network used to for communication between different components of the HAD function.

Evaluation of the robustness of – Which set of potential hardware and software the HAD function against failures failures can be simulated within the SiL? in the target E/E Architecture

305

Credibility of software-in-the-loop environments for integrated vehicle … The consequences of the credibility characteristics mentioned in Table 1 is to ensure that evidence presented to the assurance case created using the SiL fulfill the criteria [6] described in Section 4.

Figure 3: Methodology for providing V&V evidence with credible SiL environment

This paper focuses on the V&V target to evaluate the performance of the HAD decision making function. A methodology for providing V&V evidence from a credible SiL environment which contributes sufficiently to the assurance case of HAD functionality is illustrated in Figure 3. The components of this framework comprises of the following: – SiL environment to satisfy the V&V target of the HAD decision making function comprises of virtual ECUs for perception, control and actuation, environment models and vehicle models, this is discussed in detail in the case study of Section 4. – Credibility assessment method comprising the reference system and the SiL environment. State-of-the-art standards [7] and guides [8] and the corresponding credibility assessment procedures for the modelling and simulation (M&S) such as the Credibility Assessment Scale (CAS) [7] and Predictive Capability Maturity Model (PCMM) [9] require a reference system to assess the simulation maturity for decision making based on simulation results (≥ Level-1 for CAS; ≥ Maturity Level 1 for PCMM). Here the reference system is field data obtained from test vehicles with HAD functionality. – Field data is used to provide test cases generated for simulation for validating the HAD functionality. A scaling method for validation of HAD functionality has to solve three main problems:

306

Credibility of software-in-the-loop environments for integrated vehicle … – How to transform recorded data from open-loop replay to closed-loop SiL test cases? – How can you derive additional trusted test cases from real situations which you cannot reproduce (e.g. changes in speed, wind, lighting condition etc.)? – What are your arguments for why you trust the derived SiL test cases? Figure 3 gives a graphical overview about the scaling method of continuously deriving SiL test cases for HAD validation, which provide new information to existing information like complex, still unknown or rare scenarios – the so-called circuit of relevant scenarios [10]. As shown in the Figure 4, one core part is a so-called “Reconstructor” of real-world data to a simulation scenario, which can be executed as closed-loop scenario, but still representing the original real-world scenario without modifications. This Reconstructor component also must be contained in the credibility assessment methodology.

Figure 4: Reconstructor to convert recorded field data to test cases

4 Confidence Arguments for the Credibility of the SiL environment Figure 5 shows an excerpt from a system-level assurance case, where the SiL environment is used to present evidence that a particular performance claim of the system of interest is met according to the Goal Structuring Notation (GSN) [11]. In order for the assurance case to be valid, the contribution of the evidence provided by the SiL environment to the performance claim must be credible. This is indicated by the assurance claim point (ACP) on the relationship between evidence and the performance claim [6], the ACP for SiL based test method is established in Figure 6.

307

Credibility of software-in-the-loop environments for integrated vehicle … Operating environment {defined operating environment} Benchmark {benchmark for satisfactory performance}

System of Interest {Scope of system and its interfaces}

Performance Claim {System under Test} satisfies {property} to {required level of performance} in defined operating environment

ACP

SiL-based Performance Evidence {evidence demonstrating that claim is satisfied}

ACP

Controlled vehicle tests based Performance Evidence {evidence demonstrating that claim is satisfied}

SiL Test environment {Used to provide evidence} ACP

Real-world field tests based Performance Evidence {evidence demonstrating that claim is satisfied}

Figure 5: Excerpt from a system-level assurance case

Considering that the functional, non-functional and credibility requirements for the SiL environment are provided with a desired level of detail as a part of the system requirements for the intended maturity level of system release and for the desired application domain, the sub-goals [7] contributing to the confidence argument G1, for the performance claim of SiL environment to validate HAD functionality are elaborated below. Note that the triangles are added under the sub-goals to show that they are not addressed in the case study of Section 5. – GC1: Validation domain has to be derived based on the domain analysis and is defined by the validation test cases. The validation domain [13] could be a superset of the application domain, intersect or disjoint from the application domain. To provide argumentation for the safety of the HAD functionality tested with SiL, the validation domain has to be a superset of the application domain and methods such as Search-based testing [14] can be used to contribute to the test coverage. Also field data can be used to contribute to the generation of edge-cases. – GC2: Configuration of the SiL environment and the level of detail of plant models virtual ECUs etc. shall be done considering the V&V targets to be addressed as elaborated in Section 3. – GC3: Calibration data used for configuring the SiL environment has to be from reliable sources and the data used to perform the tests within the SiL are representative of the target environment.

308

Credibility of software-in-the-loop environments for integrated vehicle … Requirements functional, non-functional safety & credibility requirements GC1: Validation domain Test cases for validating the SiL environment cover the application domain and intra-/extrapolation constraints are considered [13]

G1 Confidence Argument: SiL environment satisfies credibility criteria to required level of performance for the defined ODD

Application Domain test requirements that shall be fulfilled by the SiL environment GC4: Reference System Reference system used for the required credibility level is made available for validation of SiL

GC5: Robustness GC2: Configuration SiL environment is adapted or configured considering the separation of concerns approach [3]

GC7 SiL environment is validated for the defined application domain

GC6: Subject matter experts

GC3: Calibration Data Data used to calibrate or configure the SiL environment is obtained from reliable sources and the same data is used to calibrate the reference system

Robustness of the SiL environment is adequate i.e. the SiL environment produces consistent results with the desired accuracy every time

Sn1 Validation results of the SiL environment are documented and satisfy the credibility requirements

Relevant subject matter experts are involved in the development, credibility assessment and the release argument based on SiL.

Figure 6: Evidence for the performance claim in the context of SiL based testing

– GC4: A reliable reference system is required to validate the SiL environment for contributing to the overall credibility assessment of the SiL environment according to the state-of-the-art standards discussed in Section 3. – GC5: Robustness measures have to demonstrate that there is a demonstrable correlation between the performance of the system under testing within the SiL environment and its performance within the target environment. – GC6: Subject matter experts have to be involved in the development as well as the credibility assessment of the SiL environment especially to provide required validation metrics [15]. – GC7: The SiL environment has to be validated for the defined validation domain with a specific metric based on the accuracy requirements of the quantities of interest of the use-case. The solution of this sub-goal is demonstrated in case study of Section 5.

5 Case Study Environment models such as road profile from navigation, traffic and the sensor models such as the Camera, Radar and other surround sensors provide input to the virtual ECU in the SiL environment for decision making and motion planning functions. In this usecase the environment models are reconstructed from field data recorded from a test vehicle and are the test scenarios for the decision making functions in the SiL as shown in Figure 4.

309

Credibility of software-in-the-loop environments for integrated vehicle … The motion planning algorithm of the HAD function is responsible for controlling the lateral and longitudinal vehicle actuators. The control of longitudinal motion of the ego vehicle requires the vehicle velocity to be predicted, ranging from a few meters up to few kilometers. The prediction is based on the current vehicle velocity, sensor fusion data as well as the navigation data and needs to be accurate to control the accelerator and brake actuators. Figure 7 illustrates the vehicle velocity predicted for each second starting from the vehicle start, for e.g. after 3s from the vehicle start, the vehicle velocity would be 10kmph and the vehicle would have travelled 50m. For providing the validation argument for the SiL environment, the predicted vehicle velocity at a distance of 50m (i.e.) is chosen as the quantity of interest (QoI).

Figure 7: Predicted vehicle velocity

Figure 8 shows the flow diagram to assess the accuracy of the predicted velocity with respect to the corresponding signal from the reference system. As mentioned earlier the test cases extracted from field data are executed in the SiL and reference system. Execute test cases in SiL Environment

Measurements

Validation test cases extracted from field data

Determine validation metric

Compare signals and compute the overall rating of QoI Execute test cases in Reference system

Measurements

Credibility assessment of SiL

Determine Quantities of Interest (QoI)

Figure 8: Flow diagram for validation of SiL

The predicted vehicle velocity output signal measurements from these test executions are then compared for correlation with a ISO/TS 18571 [12] based validation metric which is a combination of corridor method and EEARTH method. Figure 8 shows the validation plots for computing the corridor, phase, magnitude and slope metrics. To compute the overall score of the signals the ISO validation metric requires weighing factors to be provided and as discussed in Figure 6 of the GSN the sub-goal GC6 requires involvement of the subject matter experts to be responsible, for determining

310

Credibility of software-in-the-loop environments for integrated vehicle … the level of accuracy of the signals based on the credibility requirements of the SiL for releasing the HAD system.

Figure 9: ISO 18517 based phase, magnitude, slope and corridor score of prediction function

Table 2 below provides sample weighting factors as determined by the experts and the computed metric scores. The overall score for predicted vehicle velocity is then calculated with this metric and results in the overall confidence rating of 0.811. As per the expert decision this rating is of sufficient accuracy in our case study. Table 2: Metric scores and weighing factors to compute overall ISO rating Metric

Score

Weighting factor

Corridor

0.872

0.4

Phase

0.998

0.2

Magnitude

0.506

0.2

Slope

0.811

0.2

311

Credibility of software-in-the-loop environments for integrated vehicle … This validation argument combined with the evidence from all the other arguments (GC1-GC6) provides the performance claim that the SiL environment for validating the HAD functions satisfies credibility criteria to required level for performance and for the defined application domain.

6 Conclusion and remarks In this paper we established the importance of SiL for evaluating the performance of the HAD decision making function, also providing insights to systematically develop and argue the credibility of the SiL environment for the intended application domain of a specific HAD function. Using the approach mentioned in this paper, functional properties of the SiL can be validated given the availability of measurements from the reference system for the given scenario, but the complete coverage of application domain cannot be achieved due to open context of the system being validated. This requires the framework mentioned Section 2 to be implemented rigorously for various use-cases, by constantly improving the credibility of SiL environments using the field data and is the scope of our future work.

Bibliography 1. Boumans M., Johannaber M., Schulmeister U: Consistent application of systems engineering and simulation for cross-domain function integration. 19. Internationales Stuttgarter Symposium. Proceedings. Springer Vieweg, Wiesbaden (2019) 2. MacGregor, J. and Burton, S.: Challenges in Assuring Highly Complex, High Volume Safety-Critical Software. In International Conference on Computer Safety, Reliability, and Security (pp. 252-264). Springer, Cham. (2018). 3. Raghupatruni, I., Goeppel, T., Atak, M., Bou, J., Huber, T. : Empirical Testing of Automotive Cyber-Physical Systems with Credible Software-in-the-Loop Environments. In The 8th IEEE International Conference on Connected Vehicles and Expo (ICCVE), Graz (2019). 4. Son, T. D., Bhave, A., & Van der Auweraer, H. (2019, March). Simulation-Based Testing Framework for Autonomous Driving Development. In 2019 IEEE International Conference on Mechatronics (ICM) (Vol. 1, pp. 576-583). IEEE. 5. Bokc, T., Maurer, M., & Farber, G. (2007, June). Validation of the vehicle in the loop (vil); a milestone for the simulation of driver assistance systems. In 2007 IEEE Intelligent Vehicles Symposium (pp. 612-617). IEEE.

312

Credibility of software-in-the-loop environments for integrated vehicle … 6. Burton, S., Gauerhof, L., Sethy, B. B., Habli, I., & Hawkins, R. (2019, September). Confidence Arguments for Evidence of Performance in Machine Learning for Highly Automated Driving Functions. In International Conference on Computer Safety, Reliability, and Security (pp. 365-377). Springer, Cham. 7. Standard for Models and Simulations, NASA-STD-7009A. NASA, 2016. 8. Schwer, L. E. (2007). An overview of the PTC 60/V&V 10: guide for verification and validation in computational solid mechanics. Engineering with Computers, 23(4), 245-252. 9. Oberkampf, W. L., Pilch, M., & Trucano, T. G. (2007). Predictive capability maturity model for computational modeling and simulation (No. SAND2007-5948). Albuquerque, NM: Sandia National Laboratories. 10. Pütz, A., Zlocki, A., Küfen, J., Bock, J., & Eckstein, L. (2017, June). Database approach for the sign-off process of highly automated vehicles. In 25th International Technical Conference on the Enhanced Safety of Vehicles (ESV) National Highway Traffic Safety Administration. 11. Goal structuring notation community standard version 2. Technical report, Assurance Case Working Group (ACWG) (2018). https://scsc.uk/r141B:1?t=1. Accessed 04 June 2019. 12. ISO/TS 18571:2014. https://www.iso.org/standard/62937.html. Version: August 2019. Accessed 09 August 2019. 13. Roy, C. J., & Oberkampf, W. L. (2011). A comprehensive framework for verification, validation, and uncertainty quantification in scientific computing. Computer methods in applied mechanics and engineering, 200(25-28), 2131-2144. 14. Gladisch, C., Heinz, T., Heinzemann, C., Oehlerking, J., von Vietinghoff, A., & Pfitzer, T. (2019, November). Experience Paper: Search-Based Testing in Automated Driving Control Applications. In 2019 34th IEEE/ACM International Conference on Automated Software Engineering (ASE) (pp. 26-37). IEEE. 15. Zhang, Q., Xing, L., Sun, Y., He, Y., & Yang, J. (2019, April). Credibility Assessment of Modeling and Simulation Requirement. In 2019 International Conference on Modeling, Simulation, Optimization and Numerical Techniques (SMONT 2019). Atlantis Press.

313

Big data driven vehicle development – Technology and potential Dipl.-Ing. Philippe Fank, Dipl.-Ing. Dankmar Boja, Dr. Tobias Abthoff NorCom Information Technology GmbH & Co. KGaA

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_31

315

Big data driven vehicle development – Technology and potential

Abstract The success of automotive development in the field of autonomous driving and electrification increasingly relies on the ability to gather, manage and analyze large sets of data. Thereby, the development processes and utilized tools must adapt to provide the right environment to face emerging challenges. What are those challenges and how can an integrated Big Data platform help engineers to succeed? In this article, we discuss the four main challenges and requirements we identified which are linked to vehicle development in the era of Big Data. They emerge from the integration-, collaboration-, automation- and innovation-aspects of such a platform. With a use case, we furthermore demonstrate the capabilities of a Big Data platform in practice.

Introduction Future vehicle development will intensify existing challenges and pose new ones to development processes, used tools and responsible engineers: Data and its variety will continue to increase dramatically with new development goals such as autonomous driving and electrification. Information from unstructured documents and heterogeneous data from different domains must be reliably gathered, merged and analyzed. The scalability, efficiency and reliability of these data-driven processes and the used tools are crucial for the achievement of these goals and the commercial market success of the developed vehicle. As a result, analysis tools based on Big Data technologies are increasingly used, as the underlying technical paradigm is best suited to meet the aforementioned challenges. In addition to implementation as "on-premise" clusters, cloud solutions are used more and more by OEMs and their suppliers. Here, individual cloud developments and standard cloud adapted and extended for OEM-specific requirements are both used. We outline how the main requirements - clustered in four aspect - can be met with the technological advantages of a user-friendly, multi-infrastructure and multi-cloud big data platform. Such a platform helps engineers to successfully master data access, management and analysis. We further shed light on the criteria to be considered when deciding about the infrastructure set-up of the platform. Finally, we illustrate with a realworld use case how Big Data platforms can be utilized for developing scalable data analytics algorithms. The realization of this use case started with a proof-of-concept phase and was then transformed to a widely applicable and comprehensive analysis workflow for industrial R & D departments. Here, scalable refers to equally to the data size and the number of engineers involved in using these tools in practice.

316

Big data driven vehicle development – Technology and potential

Four aspects of sensor data analysis with a Big Data Platform From the specific challenges listed in the introduction, requirements for a Big Data platform can be derived. These requirements can be allocated to four different aspect which are briefly outlined.

Figure 1 – Overview of requirements derived from aspects of a Big Data platform

317

Big data driven vehicle development – Technology and potential Although the four aspects shown in Figure 1 do not follow a hierarchical order, they are strongly interdependent. In the following, we discuss each section in more detail: ● Integrate ● Measurement Data derived from vehicle fleets and test benches may be carried out all over the world. Therefore, the data could be distributed across many different cluster environments (on premise and cloud). However, engineers need to be able to analyze all the data directly after recording it. From this it follows that generic analytics approaches need to be designed in a way such that analyses can be performed simultaneously on all cluster environments – as if the data was stored in one place. ● Data must not only be available but also utilizable through a suitable Big Data format. Ideally, the transformation of raw data can easily be adjusted to evolving industry standards. ● And finally, management of data, code, documents and results must be implemented adjustable to ensure worldwide seamless accessibility unnoticed by the end user as enabled by a single-sign-on security mechanism. ● Collaborate ● Based on metadata, users must be enabled to quickly select the relevant data for their specific analysis. ● Any solution must be able to overcome the gap between procedural structures of the development process and the organizational structures of the enterprise to allow easy and transparent file sharing. ● Ideally, the data management also includes features to support a compliant data lifecycle management and self-acting data retention policies. As a result, the storage and usage of data is closely connected, and development processes can be accelerated. ● Innovate ● Especially in an early stage of data analysis, short iteration periods are necessary to allow a flexible and expeditious approach. It is best enabled by an interactive analysis environment which fulfills the mentioned requirement of working with highly distributed data as described above. ● Specialized automotive libraries can reduce the initial effort for data analysis by providing a standard way (API) for carrying out operations and by domain specific convenience functions e.g. for time series analysis.

318

Big data driven vehicle development – Technology and potential ● Additionally, cutting edge Open-source libraries have to be usable in order to keep up with trends and innovations. Especially the domain of Machine Learning provides a vast range of frameworks which continuously incorporate the latest research. ● Automate Once an analysis algorithm is well tested and finalized, its capabilities have to be usable by a larger group of users. This step poses an obstacle in many ways: ● Some users possibly lack specific knowledge of the analysis and just want to create results based on a variation of a small set of input parameters. ● Third parties might only be granted rights to execute the analysis and use its results without revealing the intellectual property embedded in the code. ● Multi-step analysis algorithms are ideally packed in modules so they can easily be reused and rearranged in other analysis workflows. These workflows in practice have to be trigged event-based (as new data is made available) or time-based (creating reports regularly). We firmly believe that any Big Data platform solution has to meet these requirements. In the next section, we show how the capabilities of DaSense as an example of an integrated Big Data platform solution satisfies these requirements.

How the DaSense Big Data platform meets the technical requirements Following the structure of four different aspects of a Big Data platform, DaSenses features to satisfy the listed requirements from the previous section can be summarized.

Integrate – Big Data technology for enterprise requirements Enterprises need to decide on the underlying infrastructure of a Big Data platform. Since every IT-department has to find ways to manage tight budgets with broad and increasing requirements from engineering departments, costs and flexibility emerge as the key criteria for this decision. Since there is no single truth, the best solution can vary for each company and each use case. Ideally, a Big Data platform is able to adapt to both infrastructure set-ups using its Big Data API– a fact which was demonstrated with DaSense for various customers. The analysis of data stored on various sites is supported by the Distributed Query Engine (DQE) of DaSense.

319

Big data driven vehicle development – Technology and potential

Figure 1 – Distributed Query Engine (DQE) allows to analyze and run complex queries with DaSense on various data sets which are stored decentralized

In this case, different instances of DaSense are connected to a large network and data analysis is managed by a job engine which schedules and distributes single computations. Using the job engine, DaSense is capable to dynamically distribute computing jobs between on-premise instance and cloud instance of DaSense for optimal usage of the available and with respect to legal requirements and data security. As a premise of this approach, there exist two types of jobs, which are either only allowed to be executed in the on-premise instance of DaSense or jobs that can be executed on-premise and on the cloud instance. After jobs are registered and categorized to either one of the described types, they are queued and assigned accordingly to the two instances. With the arrival of new jobs, a currently executed job can be re-organized by stopping it, transferring a snapshot of the job and resuming it on a difference instance to realize a cost-optimal usage of the available resources. The underlying hardware again is flexible – the job engine of DaSense is capable of running jobs on CPUs and GPUs following the need for computational expensive calculations in the fields of Machine Learning, Deep Learning and Artificial Intelligence.

320

Big data driven vehicle development – Technology and potential

Collaborate – Facilitate collaboration along development processes DaSense uses a fast-access index to manage uploaded documents. The aforementioned ingestion features the extraction and additional computation of metadata for each file which are then included in the index allowing for a fast definition of relevant data scope from the available data. DaSense furthermore is based a flexible and evolving storage structure contrary to a fixed folder structure prone to duplication and silo structures representation organizational structures instead of a process-oriented data usage.

Figure 2 – Example of data management with DaSense – searching for measurement files with a customary filter including metadata and facets which are assigned to each file independent from a hierarchical folder structure

Company-specific rules for data lifecycle management are enforced by DaSense. Thereby legal requirements for storing documents can be reliably separated from regular housekeeping of outdated information.

321

Big data driven vehicle development – Technology and potential

Innovate – Flexible environment for each use case through variable libraries DaSense supports the automated ingest and conversion of many common measurement data formats, e.g. MDF, ADTF, ROSbag, Parquet, HDF5, etc. to be utilized in an integrated Managed Analytics Environment. Additionally, DaSense can provide a containerized multi-user environment with customizable libraries tailored to each use case: Besides easy use of state-of-the-Art Open-source Software for Statistics, Machine Learning and Deep Learning, several industry specific libraries were developed by NorCom to provide convenient analysis features for special application: Available library extensions and possible applications in DaSense Extension Time Series Trace data Object lists Documents

Application PowerTrain/EE/ADAS EE ADAS Natural Language Processing

Figure 3 – Use DaSense to benefit from provided library extensions and Open-source software tailor to a specific use case

322

Big data driven vehicle development – Technology and potential

Automate – Establish productive data analysis workflows DaSense features the enclosure of code in DaSense Apps. Thereby, the code is accessible to the user through a graphical user interface. The apps are organized in an app store for convenient management and access control of apps by administrators. The results of an app can be of various formats (pdf-reports, html-plots, etc.), downloaded by the user or persisted in a database. With persisted results, different apps can be linked in such a way, that they access generated results and perform further computation. Using this mechanism, a comprehensive workflow can be mounted from independently developed apps. The code of the underlying script itself cannot be accessed by the user which allows to share the app to other parties without the risk of losing intellectual property. Using DaSense’s REST API, individual apps and workflows can be triggered automatically either event-based (e.g. the availability of new data) or time-based (e.g. to create regular reports). Each job and its parameters are logged to ensure traceability and repeated execution - for example on an updated data set. In the next section we illustrate how DaSense was used in practice to implement a scalable data analysis workflow for an enterprise environment.

Data analysis use case implemented with DaSense In the following we describe the evolution of a data analysis use case from the proofof-concept phase to a comprehensive and extensible analysis workflow. The evolution follows the iteration of the implemented analysis, a fluctuating and growing group of users and their level of technical knowledge, required obfuscation of the generated code and connection to auxiliary convenience functions for preparing data and persisting analysis results: Fluctuating torque loads in a vehicle’s powertrain during shifting operation of automatic gearboxes can lead to oscillations. They can impair passenger comfort and interfere with controller performance. Usually, these events are detected and logged manually by test drivers during driving routines. In this use case, we describe how DaSense was used to objectify the detection of these bucking events and to continuously gain statistical knowledge of possible root causes, guiding engineering efforts and supporting design decisions in the development process. Thereby we can illustrate how the describes requirements present themselves in practice and how the Big Data platform solution was used to overcome them.

323

Big data driven vehicle development – Technology and potential

Outline of analysis as proof-of-concept Since the described events occur only rarely, the data set was extended beyond the tests which already included the manually labeled events. In this maximized data set, potential events were isolate based on the condition, that the timeseries value for the current gear differs the timeseries value for the target gear. For this step, we used a pythonbased programming language (DSL) specifically developed by NorCom for the efficient and convenient analysis for time series of senor data. The manually identified events have then been used as labeled input data for a machine learning based classification approach. In this step we could rely on the latest version of Open-source Machine Learning frameworks and for test purposes replace the framework with various other available Open-source libraries. Short iterations in the interactive development environment allowed for a quick exchange and progress between data scientist and automotive engineers and parallel execution capabilities of DaSense accelerated the hyperparameter tuning. Once all events from the extended data set were classified, possible root causes were derived with correlation analysis. Therefore, the time intervals for a detected event were used mask various other signals. With input from the engineering department, combined with an agnostic data science approach on all given signals we were able to gain insight to statistical correlations of different signal values and the occurred events. In summation, the capabilities of DaSense during the proof-of-concept phase emerged clearly when: ● creating a homogeneous data set for the analysis ● working with time series data to extracting potential events ● integrating Open-source machine learning frameworks ● transforming proof-of-concept solution to comprehensive analysis workflow as described in the following section

Transform analysis in a comprehensive workflow The analysis included multiple steps, e.g. defining the relevant data scope, isolate possible events, classify possible events, perform correlation analysis for identified events and persist results.

324

Big data driven vehicle development – Technology and potential

Figure 4 – Implemented data analytics algorithm can be encapsulated in a custom DaSense App and combined with standard DaSense Apps into a comprehensive workflow

Since this workflow includes features that are reoccurring, we were able to reuse proven solutions from existing workflows. In practice, this was implemented by encapsulating the code in a DaSense-App, which was then connected to the other steps, which again were (already) transformed to apps. To link DaSense Apps, the user can benefit from a graphical interface by which all available apps can be managed. Those apps are easyto-use independently to prior knowledge of the data analysis. User-friendly interfaces allow individual control of a predefined set of relevant parameters in each step. Figure 5 shows the GUI for DaSense SearchApp to identify time intervals, which satisfy a set of defined criteria. The results can be displayed in plots and persisted in a structured database, which can be accessed with standard Business Intelligence (BI) tools. This way, the execution of the complete workflow becomes traceable and reproduceable. Results can easily be compared and interpreted in the context of the set parameters in each step. The app workflow allowed the interactive configuration and execution of each step. ● Searching for potential events in a continuously increasing data set ● Executing of classification algorithm to detect events in new data

325

Big data driven vehicle development – Technology and potential ● Performing correlation analysis with a flexible set of signals ● Persisting events in databases or extraction of raw measurement data

Customer benefit With the use of DaSense, the customer was able to make quick progress developing the described algorithm with input from various domain experts, put to productive use company-wide and share the results with stakeholders along the automotive development process – saving valuable time and resources in the race for innovations.

Summary In this article, we have elaborated the challenges linked to data-driven automotive development processes and derived requirements for an integrated Big Data platform. We described how the features of the DaSense platform meet those requirements. With the description of a real-world use case, we illustrated how this platform can facilitate data analysis from the stage of proof-of-concept to a comprehensive and productive analysis workflow in an enterprise environment.

326

Continuous development environment for the validation of autonomous driving functions Sebastian Lutz, Dr. Matthias Behrendt, Univ.-Prof. Dr.-Ing. Dr. h. c. Albert Albers Institute of Product Engineering – IPEK at the Karlsruhe Institute of Technology (KIT)

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_32

327

Continuous development environment for the validation of autonomous driving functions

Abstract The development of driver assistance systems and automated driving functions is associated with new challenges for the development process. More and more comprehensive functions have to be realized in series production at low cost. This problem pushes established development environments to their limits. This contribution shows new approaches to solve this problem by showing how to achieve a high degree of reproducibility and stability with flexibly configured validation environments, making development more efficient. For this purpose, individual aspects of the environment can be modularly adapted with regard to the desired goals. For example, such a module consists of a virtual environmental simulation or a sensor stimulator. This combination can simulate selected environmental impressions to vehicle sensors, which can be combined with a roller test bench or a test track in the further course of development. This avoids breaks between currently used environments and combines the advantages of virtual function testing with the validity of real test drives. Such a validation environment is presented for investigations on a camera-based driver assistance system. For this purpose, a self-developed camera stimulator is used for early functional verification as well as for applications on the complete vehicle on the chassis dynamometer and the test track.

1 Introduction Today’s development of vehicles is an increasing challenge due to more and more complex system connections. The demand for electrification, connection with the driver and automated driving in vehicles is increasing the pressure on manufacturers and developers. To meet these challenges, new test procedures must be generated and run as soon as possible in order to be able to reproduce individual test scenario parameters (such as weather or traffic conditions). Modern test benches offer reproducible development environments [1] in which sensor error functions can be tested more safely than on the road. Additionally it is possible to carry out external sensor manipulation tests. In addition to the interaction between sensors and the vehicle control systems, advanced driver assistance systems (ADAS) and automated driving (AD) functions must also be checked for interaction with other domains of vehicle development. The longitudinal vehicle control system, for example, also has an impact on the powertrain and the associated energy efficiency or NVH and driving comfort aspects [2]. In order to be able to make decisions about the System of Objectives for ADAS or AD as early as possible, preliminary tests on test benches help to check target values for

328

Continuous development environment for the validation of autonomous driving functions energy consumption or to carry out fast plausibility checks. These test benches have to be developed, so virtually configurable test scenarios help to get fast ready for action. This is supported by fewer dependencies in the test parameters, which enables only to influence individual test parameters. In addition, boundary limits can be determined in a secure test environment and thus development goals can be ensured. In this context current validation environments for ADAS and AD have two established environments. On the one hand, there are as many tests and investigations as possible to be carried out in a completely virtual environment [3]. This avoids the need for expensive prototypes and the tests can be carried out much faster than in real time. On the other hand, only the real test track or road offers a high representativeness, because of the same environment conditions like in customer's hands [4]. One of the disadvantages is an expensive validation over many thousands test kilometers. The presented approach will show ways to use these two established development environments fluently with each other and combine their advantages, taking into account the respective limitations.

2 Requirements for a suitable validation environment To be able to develop and adjust future ADAS or AD systems, two general developments must be taken into account. On the one hand, the depth of detail of individual components is increasing, which requires more expertise, for example, in sensors and their physical principles. On the other hand, more and more information is exchanged within the vehicle, so that networking and dependencies between individual systems are increasing [5]. These trends create challenges at different points in the development process. For example, validation must be able to take more details from the environment into consideration [6]. In addition, a validation should be able to be carried out quickly on known test benches, so knowledge about system behavior must be more and more available in the development process. This means that tests can be set up much faster and then repeated quickly, while only individual influences can be varied and their influence on the vehicle's behavior analyzed. If individual environment influences and effects can be understood in greater depth, the demands on the validation environment increase, so that it must also be able to refine its output quality in terms of realism. As a result, the requirements on generating more details of the simulated environment are constantly changing. One way of solving this problem is using more complex software solutions for the environmental simulation. This easy exchange of only the virtual environment simulation is then enabled by standardized interfaces and the modular structure. This allows only individual influences to be exchanged and therefore only individual parts of the environment and not several connected effects, like the environment perception of other vehicle sensors, to be exchanged.

329

Continuous development environment for the validation of autonomous driving functions The consistent use of elements of the validation environment, like the use of the same stimulator over many validation environments, helps to better understand the influences on the test results. Additionally the knowledge of their strengths and weaknesses increases over several development steps, so the validation environment can be used more effectively. At present, road testing offers the great advantage of being representative for customer use. On the other hand, these test drives often have the problem, that they can not be repeated exactly the same way. Test bench environments do not have this difficulty in reproducibility. For this reason, efforts are being made to shift from test drives on the road to the use of test benches. In this transition it is then necessary to provide the required customer representation. One approach is the transfer of recordings of real test drives, allowing a fast repetition of a situation from the test drive in a test bench environment as often as required. Individual parameters can then be varied during the maneuver, increasing the understanding of vehicle behavior.

3 Principle design of validation environments for automated driving To comply with the previously presented requirements for a suitable validation environment as comprehensively as possible, the IPEK-X-in-the-Loop-Approach offers a good basis. This approach is shown in Figure 1 and consists of the three system models for vehicle development: driver, vehicle and environment as well as the maneuvers and test cases. On the one hand, this allows the modeling of the interactions between the three system models. On the other hand, it should be noted that the systems involved can be considered on different levels of abstraction. This approach can be used to represent investigations on vehicle subsystems such as the engine or the transmission as well as on the overall "vehicle" system.

330

Continuous development environment for the validation of autonomous driving functions

Figure 1: The IPEK-X-in-the-Loop-approach [7]

When setting up validation environments, the realization of a model, like the “Vehicle” or the “Environment”, is extended by the possibility to represent individual components or entire systems, physical or virtual. The interaction between virtual and physical system components needs interfaces, which transform parameters and conditions bidirectional. The approach presented provides a fundamental understanding of the development of automated driving functions, from software development e.g. with Software-in-theLoop (SiL) at the beginning of a development process to integration in the overall vehicle. This provides a seamless transition between the development steps and enables environment interaction with reality to be integrated in small steps from a lot of virtualization to more and more physical testing. Because of the modular design, the properties of the validation environment can be refined in a targeted manner and as many surrounding components as possible can be reused. Figure 2 shows the IPEK-X-in-the-Loop-approach applied to the validation of a camerabased adaptive cruise control (ACC) or autonomous emergency brake (AEB) system. In this case, an ECU and therefore a virtual driver is the controlling instance when driving the vehicle in longitudinal direction instead of a human driver. The functionally relevant sensors, such as the wheel speed sensor or the mono camera, are shown in the vehicle and are extended by a rest vehicle model. This model represents the necessary rest systems of the vehicle, which, depending on the development objective, should not be examined in detail, but nevertheless provide the necessary input variables for the operation of the vehicle system. The maneuvers and test cases are based on the environmental conditions and speed profiles relevant for such systems. For the ACC system, for example, the driving behind a target vehicle

331

Continuous development environment for the validation of autonomous driving functions with moderate acceleration and deceleration provides a common scenario. Driving onto a stationary target vehicle, as it is the case at traffic lights, for example, represents the transition to scenarios for an emergency braking system. The maneuvers and test cases as well as the vehicle system in development derive demands on the "environment" system. This must provide the vehicle with realistic impressions and be able to implement the test cases and maneuvers. As with the "vehicle" model, the "environment" model can also be implemented by a mixed physical-virtual solution. A virtual environmental simulation is used to generate the scenario and the resulting environmental impressions for the camera. The scenarios generated in this way are then fed to the vehicle via physical interfaces, depending on the stage of development.

Figure 2: The IPEK-X-in-the-Loop-approach applied on the validation of an automated driving system

332

Continuous development environment for the validation of autonomous driving functions A possible implementation of such a validation environment for the development on a complete vehicle is shown in Figure 3. Information from the virtual environmental simulation is transferred to all relevant interfaces on the vehicle. The vehicle system in development is placed on a roller test bench, which can simulate the road with corresponding driving resistances from wind and road gradient to the vehicle drive train. A camera stimulator is mounted in front of the vehicle camera. This stimulator displays the image from the virtual environment representing real world, which is generated from the perspective of the interior mirror. A simulation solution such as IPG CarMaker (CM) or Vires Virtual Test Drive can be used as virtual environment simulation. These solutions are real-time capable, i.e. they can generate simulation data in a test bench environment with constant cycle times. Additionally they offer universal interfaces for the integration of stimulators, which means environmental data can be transferred directly to the vehicle via CAN or Ethernet with UDP to stimulators.

Figure 3: Setup of a validation environment based on the chassis dynamometer

An environment set up like this will be presented in the next chapter, with individual components such as the camera stimulator and the virtual environment simulation being used both in a HiL environment and for validation on the test track.

333

Continuous development environment for the validation of autonomous driving functions

4 Continuous validation of a camera-based vehicle control system The following chapter shows the application of the IPEK-X-in-the-Loop approach for continuous validation of a camera-based automated driving function. Starting with an early validation on the ECU with connected sensor, the basic features of the development environment, like the virtual simulation environment or the camera stimulator, are built up and then further developed for validation with the vehicle. This is then continued in combination with the test track. During validation, a large part of the nesecceary activities handling the creating of a suitable validation environment. Therefore, in the validation of a camera-based vehicle automation system presented here, the design of the system model "environment" is discussed more in detail. Figure 4 shows an overview of the different modules that have been integrated into the validation environment. In general, both physical and virtual solutions are used. On the top level are the elements that interact with the vehicle at the overall vehicle level. On the left side is the virtual environment simulation, which can consist of a simple model in MATLAB/Simulink or a complex simulation environment like IPG CarMaker. On the right, there are the physical features, which can be the test bench hall or a real environment for road tests. The virtual as well as physical realizations for the vehicle sensors are on subsystem level. The environment impressions for the vehicle camera can be generated by images from a 3D simulation or a previously generated image catalogue, which provides appropriate impressions. The virtual impressions are then fed directly into the vehicle through physical components such as a camera stimulator, suitable connections for a data injection or test dummies in front of the car. Since the camera-based vehicle automation system also evaluates data from the wheel speed sensors, the validation environment also supplies the vehicle with information regarding the driving velocity from simulation. For this purpose, in addition to the direct data injection, either a roller test bench can set the wheels in motion or simulate resistances, or the vehicle can be moved on the road or on a test track.

334

Continuous development environment for the validation of autonomous driving functions

Figure 4: Elements for the realization of the validation environment

In the following, three different versions of the validation environment are discussed. We start with a setup for the vehicle camera with connected control unit. The system is then operated in combination with a roller test bench with the complete vehicle. Finally, the complete vehicle is tested on a test track in combination with the camera stimulator.

4.1 Validation of the camera-sensor with connected control unit The validation environment in this setup includes the virtual environment simulation and the camera stimulator. In addition to the 3D simulation images for the camera stimulator, speed values are generated from the environment simulation and transmitted directly to the ECU via CAN. It suits investigations on the influences added by the sensor to the controller behavior and image processing algorithms fine. Figure 5 on the right shows results from a project with AVL Germany, which was used for a validation process. A dismounted vehicle camera stands on a tripod in front of the stimulator. This displays a virtual target vehicle in a driving scenario. With the help of this setup, the correct recording of environmental impressions by the camera can be tested. In this case the evaluation of the distance between the ego car and the target vehicle can be checked.

335

Continuous development environment for the validation of autonomous driving functions

Figure 5: Setup for the validation of the Camera-Sensor with connected control unit

Measurements will then help to investigate on the distance recognition performance of the camera system by comparing the distances in the virtual simulation to the measurements coming from the camera control unit. For example, the behavior can be categorized by the timespan in which the system detects the target and an accurate distance determination is guaranteed. Because of the good reproducibility the system behavior in this environment remains very stable over multiple test runs and does not change its characteristics. The challenge by using a camera-stimulator is the correct and recoverable alignment in front of the stimulator. Because of the direct access to the control unit in this validation step, the camera raw-data are available. With this information, it is possible to match the impressions coming from the stimulator with the camera perception.

4.2 Validation with the complete vehicle on the roller test bench The validation on a roller test bench transfers the virtual road from the environmental simulation via the rollers to the wheels without any required modifications on the vehicle electronics. The development environment created in this way is shown in Figure 6 and enables adaptation to all other sensors and subsystems in the vehicle. In this way only speed information has changed to real wheel speed, which enables an evaluation in the vehicle, instead of digital injected values. In addition, maneuvers can be repeated very quickly, since no obstacles need to be positioned and the vehicle must not be driven back to the starting point on a test track. It is only necessary to set the vehicle to the initial speed and restart the simulation.

336

Continuous development environment for the validation of autonomous driving functions The six degrees of freedom in positioning have different influences on the distance recognition. A decent solution for minimizing these influences is an image calibration program, which moves and rotates the image on the stimulator screen. This procedure is then continued until the behavior known from the street can be observed. Therefore, this program is also able to quantify differences in positioning of the stimulator.

Figure 6: Setup for the validation with the complete vehicle on the roller test bench

Figure 7 shows the speed profile of the vehicle during the last few meters in a drive-up maneuver. The scenario is based on a target vehicle standing at a distance of 100 m, which is driven onto with activated ACC system set at 30 km/h. Because of legal reasons the setting speed of 30 km/h means that, the vehicle actually approaches the target vehicle in the scenario at 28 km/h. The deceleration phase begins at a distance of approximately 40 meters. The speed decreases constantly up to a distance of about eight meters and deviates only very slightly from this trend if the test is repeated.

337

Continuous development environment for the validation of autonomous driving functions

Figure 7: Measurement evaluation for an approach maneuver until stop on the roller test bench

The ability to control relevant environmental conditions, such as ambient light and humidity, minimizes unintended influences on the vehicle sensors and reduces deviations in the repetition of the test. This high reproducibility of the environment enables the developer to observe the effects of only slight changes in system application parameters. Also, the fast repetition of tests enables a wider spectrum of tests scenarios be carry out, which reduces development risks.

4.3 Validation with the complete vehicle on the test track Validation on the roller test bench has its advantages in terms of rapid test repetition and reproducibility. However, there are also limitations, such as different tire contact with the road surface, or inability to drive maneuvers with steering intervention on the wheels. In order to be able to investigate such scenarios during development, the investigations can be continued on a test track, which is shown in figure 8. In this way as few parameters as possible are changed in the same step, which means that both the environmental simulation and the camera stimulation are used without any changes. This enables a quick shift for the investigations from the roller test bench to the test track. In our setup, it was possible to carry out measurements on the test track directly after test bench tests, because both were close together. In this case, the installation on the vehicle, especially the positioning of the camera stimulator, has not changed. When

338

Continuous development environment for the validation of autonomous driving functions repeating the test, only the vehicle must have sufficient space on the test track and no test targets must be repositioned.

Figure 8: Exemplary setup for the validation with the complete vehicle on the test track

The diagram in Figure 9, shows the speed profile during a drive-on maneuver to a target vehicle located 100 m away in stillstand like in the environment before. Because of the same setting speed of 30 km/h, the test vehicle approaches the target vehicle at 28 km/h until a remaining distance of about 40 meters. In the test evaluation, it becomes apparent that the results vary more because of uncontrollable environment influences, like the contact between tire and street or different light conditions on the camera, so that the reproducibility in this environment decreases. For example the moving vehicle could induce vibration on the camera stimulator, which influences the image recognition. The biggest advantage of the test track is the condition of the moving vehicle with real street tire contact as well as the ability to drive maneuvers with steering. The best solution is a combination of test drives at a test bench completed with a few tests on the test track.

339

Continuous development environment for the validation of autonomous driving functions

Figure 9: Measurement evaluation for an approach maneuver until still stand on the test track

4.4 Comparison of results between test bench and test track Figure 10 shows measured values from the test track compared to those from the roller test bench. The remaining distance at the time of the start of the deceleration is at a comparable level with 42 m. Up to a distance of 5 meters, the curves are similar, with an almost constant deceleration being maintained up to a remaining distance of 6 meters. Underneath a speed of 2 km/h the speed curves differ. On the test track the vehicle comes to halt at a distance of 4 m. On the test bench, the deceleration decreases for a short moment and the vehicle then comes to halt at a distance of 2.5 m. Although the measurements on the test track with three test runs are less, the value range extends over a similar area.

340

Continuous development environment for the validation of autonomous driving functions

Figure 10: Comparison of the measurement between roller test bench and test track

In general, the results of both environments are comparable and show that it is not just possible to transfer validation activities from the test track to the roller test bench. With this transfer it is even possible to get a better reproducebility, which enables more detailed investigations on small influences in the control algorithms. The differences at a remaining speed of 2 km/h can be explained by the roller test bench control, which differs slightly at this speed to the road resistance behaviour. In drive-on tests with real target vehicles, similar deceleration responses are obtained as with stimulation on the test track. With this in mind, validation with the help of sensor stimulators on the test track as well as on the test bench is quite comparable. Additionally it can be shown that a decreasing reproducibility is resulting with increasing realistic environmental influences by using sensor stimulators to replace real world impressions with virtual generated scenarios.

341

Continuous development environment for the validation of autonomous driving functions

5 Summary and Conclusion This contribution shows an approach for setting up different validation environments for driver assistance systems or automated driving. The target is always to adjust as few elements as possible on the development environment during the further development in order to remain comparable with the previously used environment. In this way, investigations are quickly made possible and controlled as well as reproducible conditions are available when necessary. In addition, the development risk can be minimized by continuous validation, maximizing the knowledge about the product to be developed. Central elements of this approach are sensor stimulators, which supply the vehicle with information from a virtual environmental simulation. The focus here is not to develop more accurate stimulators. Instead investigations are presented to demonstrate the benefits of such a stimulator based development environment. The procedure is based on the IPEK-X-in-the-Loop approach, which enables reproducible investigations on vehicle phenomena that normally can only be observed on a driving car. In this context, a roller test bench serves as a link between the vehicle and the virtual test environment. Besides the roller test bench a camera-stimulator connects the vehicle-camera to the 3D-Simulation of the environment simulation, which enables validation and testing of a ACC or other automated driving function based on similar sensors. According to this understanding, the camera stimulator and the roller test bench are always sensor stimulators to create a link between the virtual test environment and the real world environment. When using these devices, it must be ensured that the sensors are reacting in the same way like in the real environment. An initial calibration is a necessary part of setting up the validation environment. Basically, this step ensures that exactly the information from the virtual environment is perceived by the vehicle in the way it needs to fulfil its function. To demonstrate the performance and transformability of the approach, part of a development process is presented, beginning with a validation on the camera-sensor and the connected ECU. As part of the next step, an environment will be proposed in which the entire vehicle is operated on a roller test bench with virtual test maneuvers. As an intermediate step to testing on the road with artificial traffic obstacles, the entire vehicle with connected stimulators is operated on a test track, performing the same maneuvers as in the previously presented validation environment. This allows the results to be compared. One result of this comparison is the knowledge, that validation with the help of sensor stimulators on the test track as well as on the test bench is quite comparable. Finally, the method presented in this paper has shown a flexible and reproducible validation environment for ADAS or automatic driving functions. This modular way of

342

Continuous development environment for the validation of autonomous driving functions setting up new validation environments enables a situational and therefore more efficient way of developing such functionality and makes more and more complex vehicle systems developable.

Bibliography [1] S.-O. Müller, M. Brand, S. Wachendorf, H. Schröder, T.Szot, S. Schwab, B. Kremer, Integration vernetzter Fahrerassistenz-Funktionen, ATZextra, Heidelberg: Springer-Verlag, 06.2009 [2] A. Fen, Simulative Ermittlung von Key-Performance-Indikatoren komfortorientierter Fahrerassistenzsysteme, ATZ, Heidelberg: Springer-Verlag, 01.2018 [3] G. Sievers, S. Graf, M. Peperhowe, C. Seiger, Simulation von Sensoren im SiLund HiL-Umfeld, ATZextra, Heidelberg: Springer-Verlag, 07.2018 [4] K. Golowko, D. Szolnoki, S. Schreiber, Fahrerassistenzsysteme reproduzierbar testen, ATZ, Heidelberg: Springer-Verlag, 04.2013 [5] J. Schrepfer, J. Mathes, V. Picron, H. Barth, Automatisiertes Fahren und seine Sensorik im Test, ATZ, Heidelberg: Springer-Verlag, 01.2018 [6] M. Herrmann, Sensormodelle für die Entwicklung und Absicherung automatisierter Fahrfunktionen, ATZextra, Heidelberg: Springer-Verlag, 03.2019 [7] U. Lindemann, Handbuch Produktentwicklung Kapitel 6: Verifikation und Validierung im Produktentstehungsprozess, München: Carl Hanser Verlag, 2016

343

Impact of future 48 V-systems on powertrain operation under real-driving conditions Daniel Förster, M. Timmann, R. Inderka, J. Strenkert Mercedes-Benz AG F. Gauterin Karlsruhe Institute of Technology (KIT)

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_33

345

Impact of future 48 V-systems on powertrain operation under real-driving conditions

Abstract 48 V-hybrid drives represent a cost-efficient technology that can significantly reduce the fuel consumption of passenger cars, while requiring limited integration effort. To achieve the maximum benefit with a 48 V-system while maintaining low overall system costs, a detailed understanding of the relations between conflicting optimization goals such as efficiency, comfort and cost, especially under real-driving conditions, is essential. This paper analyzes this conflict in a 48 V-system with P2 topology, where the optimization goals are represented by CO2 emissions (efficiency), engine start frequency (comfort) and effective battery load (component cost). Characteristic real-driving scenarios are investigated using a longitudinal-dynamics simulation model. An ECMS operating strategy (OS) is implemented and control parameters are defined in order to manipulate the conflicting goals. The results reveal the sensitivity of OS parameters to conflicting optimization goals. The relationship between optimization goals is identified through Pareto fronts. Pareto optimal solutions are selected for each scenario using weighting factors. Finally, the effects of different Pareto solutions on the powertrain operation in terms of the distribution of ICE-off duration are identified.

1 Motivation Due to decreasing legal CO2 limits for passenger cars, automotive manufacturers have to introduce new technologies to increase powertrain efficiency. There exists a goal conflict between fulfilling customer requirements, in terms of functionality, comfort and price, as well as providing significant fuel consumption reduction. 48 V-hybrid drives provide multiple new hybrid functions such as recuperation or electric sailing, which reduce vehicle fuel consumption by up to 15 %, while requiring low integration effort [1]. Therefore, 48 V-hybrid drives represent a technology which is suitable for short-term integration in many vehicle segments in order to reduce the average fuel consumption in manufacturers’ passenger car fleets. In the system design process, component requirements are identified based on system requirements such as efficiency, comfort and total system cost. To also make the technology available for cost-sensitive vehicle segments, the right-sizing of the system components is particularly crucial for 48 V-hybrids. Consequently, a detailed understanding of conflicting goals from system requirements is necessary. This paper discusses this conflict for a 48 V-hybrid drive with a P2 topology.

346

Impact of future 48 V-systems on powertrain operation under real-driving conditions

2 48 V-System Design 2.1 Approach

Figure (1) Structure of 48 V-system, system environment and system interfaces

The product development process in the automotive industry can be represented by the V-Cycle [2]. The stakeholder requirements are identified based on customer needs, legal regulations and company strategy. On the left side of the V-Cycle, a functional product architecture, followed by the product design, is derived based on stakeholder requirements. The architecture defines different levels of abstraction, where each level can be considered as a system consisting of several sub-systems, each of which interact with its environment through a defined interface. The system design is derived from the functional architecture and relevant system requirements. The 48 V-system is a subsystem of the vehicle system and is defined by its topology, components and operating strategy (OS), which is outlined by the structure in Figure (1). The characteristics in each category are considered as system parameters that determine the resulting component requirements and can be interpreted as output of the 48 V-system design stage in the V-Cycle. In this approach, the sensitivity between system parameters and the fulfillment of system requirements is analyzed to propose a sensible compromise between

347

Impact of future 48 V-systems on powertrain operation under real-driving conditions

conflicting goals. The main responsibilities of the 48 V-system are to provide efficiency- or comfort-increasing hybrid functions to the powertrain and to supply the vehicle powernet with electric energy. Therefore, the influences on the 48 V-system are defined as mechanical loads, representing loads exchanged with the powertrain system, and electrical loads, considering electric consumers’ energy consumption.

2.2 System Parameters 2.2.1 Topology The system topology provides possible power-flow paths in the powertrain and determines the resulting hybrid functions. For 48 V-applications, parallel topologies are considered which can be integrated into the vehicle powertrain with few changes in conventional components. Parallel hybrid topologies are denoted as P0–P4. P0 and P1 hybrids are realized as belt-driven starter generators (BSG) or integrated starter generators (ISG) and have the potential to replace conventional auxiliaries such as the generator or the starter. The main benefit of these systems results from recuperation (RCP) of braking energy, which can be reused later to supply the electric powernet or to assist the vehicle propulsion by boosting (BST). Additionally, the load point of the internal combustion engine (ICE) can be varied through load-point-shift (LPS). P2 hybrids require an additional clutch (K0), located between the ISG and the ICE, which allows electric driving (ED) with a disconnected ICE at moderate powertrain loads [1]. The realization of ED in 48 V-hybrids, with a reduced electrical performance compared to full hybrids, is a promising as well as challenging objective. While disconnecting the ICE at moderate powertrain loads can increase its efficiency substantially, the implementation of transition control of the ICE operating state has to be carefully considered, since it is potentially compromising system requirements in terms of comfort and performance. To achieve the full potential of P2 hybrids in terms of ED availability, an additional start system for the ICE might be necessary to guarantee a fast and comfortable ICE start in all possible scenarios. Furthermore, it might be necessary to reserve electric power in order to provide the desired vehicle dynamics response before the ICE can contribute torque after starting.

2.2.2 Components Electrical components with relevant impacts on hybrid system functions are the EM and the battery (BAT). The EM interface can be defined by a maximum torque characteristic and its component efficiency. The BAT has to be designed such that it can cope with the occurring electric load profiles in terms of energy capacity and maximum power. Therefore, the main parameters defining the component impacts on system performance of a 48 V-system can be considered maximum electrical system power and

348

Impact of future 48 V-systems on powertrain operation under real-driving conditions

the corresponding battery energy capacity. Further system influences, such as thermal derating or alterations in voltage level induced by battery state of charge (SOC), strongly depend on the specific component design, the design of which and the corresponding influential parameters are unknown from a top-down perspective in system design. Thus, it is considered reasonable to apply an abstract level of physical detail in the component models for system design investigations, where only primary system influences, such as maximum system power and associated component losses, are considered. The results can be used to identify relevant component requirements, while further effects of specific component designs have to be considered iteratively in later development stages.

2.2.3 Operating Strategy The OS in a 48 V-hybrid is responsible for the energy management and torque split coordination between the EM and the ICE. The goal in the so-called energy management problem is to minimize fuel consumption by the chosen torque split, while taking multiple constraints into account. Constraints can occur from component limitations such as maximum voltage or torque, whereas further limitations result from system requirements. For example, the frequency of ICE starts might negatively influence the NVH experience and the amount of reserved power for acceleration events influences the instant torque response of the vehicle propulsion. Multiple concepts of operating strategies exist while differing in terms of complexity, optimality and required input data [4]. Rule-based strategies benefit from easy vehicle behavior interpretation resulting from logical rules, whereas comprehensive system analysis has to performed beforehand to guarantee that the implemented rules result in the expected performance. Optimization-based strategies are founded on formal mathematical optimization concepts. They can be categorized into global optimization concepts, such as dynamic programming (DP), which identifies the global optimum but requires high computation effort and knowledge of the future driving profile, and instantaneous optimization, such as equivalent consumption minimization strategies (ECMS), that seek to minimize a local cost function at each time step based on a fuel equivalence factor. Once the equivalence factor is estimated iteratively, ECMS can lead to a near-global fuel consumption minimum while the constraints can easily be incorporated in the cost function [3]. Component limitations, such as SOC thresholds, as well as time-variant parameters, for example instantaneous SOC, which effects efficiency, cannot be considered in the local optimization process. These factors may compromise the solution’s optimality. However, as proposed in Section 0, those limitations should be neglected in the system design stage due to the uncertainty about the final component design. This makes ECMS a suitable tool for simulation studies in the early development stage.

349

Impact of future 48 V-systems on powertrain operation under real-driving conditions

2.3 System Influences The 48 V-system influences in Figure (1) are categorized as mechanical loads from the powertrain system and electrical loads from the powernet system. In the system design stage, the detailed analysis of such load influences on the system behavior is necessary to guarantee an optimal, as well as robust, system performance in relevant real-driving scenarios. The mechanical loads on the 48 V-system result from longitudinal vehicle dynamics and vehicle parameters such as mass or drag coefficient, which define the driving resistance. The OS controls the torque split between the EM and the ICE, which itself is influenced by component losses in the powertrain system, to overcome the vehicle driving resistance. Therefore, mechanical loads from the system context are defined by vehicle parameters, powertrain components and the vehicle longitudinal dynamics. While the effects of vehicle parameters and component losses can be modeled by physical equations or numeric loss maps, longitudinal dynamics are more complex due to multiple influences from the driving environment and driver. Driving cycles (DCs) can be applied to model those dynamics. A DC either represents a combination of multiple driving environments (e.g. urban and highway), such as the world harmonized light-duty test cycle (WLTC), or it models one characteristic scenario representative for one specific driving environment and a specific driver type. In vehicle certification tests, only few short DCs are considered to complete the test procedure within reasonable time. However, one typical DC cannot cover the full range of possible driving scenarios and there is a risk that some scenarios are underrepresented. Consequently, in system design analysis it is considered sensible to additionally investigate different characteristic driving cycles (CDCs), each representing a specific driving scenario type to guarantee robust system performance. The total electric consumer energy consumption at the 48 V and 12 V voltage levels is defined as electrical loads. This system influence is especially important for the design of 48 V-hybrids, since its increased voltage level enables new comfort consumers, such as electric air conditioning or electric suspension among others, all of which can increase the powernet energy consumption significantly in real-driving scenarios.

2.4 System Requirements The system requirements of a 48 V-system can vary based on specific stakeholder requirements. However, the reduction of CO2 emissions is considered the main motivation for the development of 48 V-systems in this research. In addition, comfort and vehicle performance requirements, influenced by the 48 V-system, have to be taken into account. In order to make the 48 V-technology available for the integration into many vehicle segments, the limitation of overall system costs represents a further important system requirement. The system costs consist of a fixed amount for the integration of a new system architecture and a variable contribution, which is influenced by specific

350

Impact of future 48 V-systems on powertrain operation under real-driving conditions

component requirements. For instance, the battery load profile determines the necessary component dimensioning in terms of cell characteristics and or thermal management.

3 Simulation Study In this paper, the sensitivity of 48 V-system design parameters on exemplary system requirements is analyzed in real-driving scenarios using the simulation framework in Figure (2). The selected system parameters and influences defined in Section 2 are discussed in the following sections.

Figure (2) Overview of powertrain simulation framework for 48 V-system analysis

3.1 System Parameters A P2 topology is considered in the simulation study, since it provides a reasonable compromise between integration effort and available hybrid functions. Compared to less difficult P0 or P1 topologies, it benefits from additional ED functionality with disconnected ICE at low powertrain loads, which is realized by an additional clutch K0. To reduce complexity, the impacts of an additional start system or EM power preserved for ICE start or vehicle agility are neglected. The control of ED and the corresponding transitions in the ICE operating state 𝑏 are focused on in this analysis. For 48 V-

351

Impact of future 48 V-systems on powertrain operation under real-driving conditions

hybrids with a P2 topology, a sensible maximum electric power can be estimated approximately at 20–25 kW based on literature [1, 5–7]. This value exceeds the current products on the market utilizing a P0 or P1 topology, which typically have a maximum power of 10–15 kW [8]. The increased maximum power can be explained as a tradeoff between improved availability of hybrid functions, such as ED or RCP, and increasing system costs. The 48 V-components in the simulation model in Figure (2) are represented by power-loss maps of a 25 kW PSM design for ISG applications and a simplified inner-resistance model of a Li-Ion battery. The inner-resistance 𝑅 and the are considered as constant reference values, since the final open-circuit voltage 𝑈 battery design and the corresponding parameter dynamics are unknowns at the system design stage. The battery energy capacity is considered by an integral of energy deviation from an initial state in (1), which defines the criterion for a neutral energy balance in the simulation. The EM losses are obtained from power-loss maps based on the corresponding component speed, torque and battery voltage 𝑈 . Δ𝐸

𝑡 =

𝑈

𝜏 𝐼

𝜏 𝑑𝜏

(1)

A cost-based ECMS approach is implemented to control the torque split between the ICE and the EM as well as transitions between ED and ICE operation. The performance and transitions in ICE function is adapted based on battery energy deviation ∆𝐸 operating state 𝑏 . The applied control laws are defined in (2)–(6). 𝑃: min 𝐽 𝑥 𝑡 , 𝑢 𝑡 𝐽 𝑥 𝑡 ,𝑢 𝑡

= 𝜆 (𝑥(𝑡 , 𝑢(𝑡 ) 𝑚 𝜆 𝜆

𝜆

(𝑥(𝑡)) = 𝜆 𝑥(𝑡) = 𝑘

∀ 𝑡 ∈ 𝑡 ,…,𝑡 𝑢(𝑡) + 𝜆

(2) (𝑥(𝑡))

𝑃

𝑥(𝑡) + 𝜆 Δ𝐸

(𝑢(𝑡)) 𝑄

(3) (4)

(𝑥(𝑡))

(5)

𝐸

𝑥(𝑡), 𝑢(𝑡) = 1+𝑘

,

,

𝑓𝑜𝑟 𝑏

𝑢(𝑡) = 0 ∧ 𝑏

𝑥(𝑡) = 1

1+𝑘

,

,

𝑓𝑜𝑟 𝑏 0,

𝑢(𝑡) = 1 ∧ 𝑏 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

𝑥(𝑡) = 0

(6)

To calculate the optimal trajectory of torque split based on (1), the performance function 𝐽(𝑥(𝑡), 𝑢(𝑡)) in (2) is minimized at each time step, where 𝑥(𝑡) is the state vector, representing current battery energy deviation and ICE operating state and 𝑢(𝑡) is the torque split control parameter. The performance function weighs the fuel consumption

352

Impact of future 48 V-systems on powertrain operation under real-driving conditions

𝑚 and the battery energy consumption 𝑃 based on the fuel’s lower heating value 𝑄 and an equivalence factor 𝜆 . The equivalence factor is influenced by an iteratively estimated constant value 𝜆 , which guarantees a neutral battery energy balance. can be reduced by Furthermore, the required total battery energy capacity Δ𝐸 , the parameter 𝑘∆ , which increases electric energy cost at negative energy deviations Δ𝐸 and vice-versa. The parameter is selected based on a parameter variation, accepting 1 % increase in CO2 while reducing the required battery energy capacity. A hysteresis for the ICE operating state transition is implemented using the parameter 𝜆 in (5), where the ICE operating state 𝑏 is to be interpreted as one for the state K0 closed and zero for the state K0 open. The hysteresis effect on exemplary system requirements is analyzed in this study, while the transition for ICE start and stop events are represented by separate hysteresis parameters 𝑘 , and 𝑘 , respectively.

3.2 System Influences The goal of the simulation study is to analyze the sensitivity of system parameters on system requirements in customer-oriented real-driving scenarios. Thus, the inputs of the simulation framework in Figure (2) have to be chosen properly. The mechanical loads on a 48 V-system are primarily influenced by the vehicle parameters and the longitudinal vehicle dynamics. The combination of both factors define the vehicle driving resistance, which has to be overcome by the powertrain system. Consequently, the vehicle parameters of an exemplary C-segment vehicle in Table (1) are applied to investigate influences of different characteristic real-driving cycles (CDCs), shown in Figure (3), on the 48 V-system. The CDCs are identified based on a data-driven analysis on extensive real-driving measurements [9]. Furthermore, the mechanical system loads are influenced by the powertrain system, since the efficiency of components, such as the ICE, is incorporated into the torque split control of the OS. Therefore, the efficiency of the primary conventional powertrain components ICE, GB and differential (DT) in Figure (2) are modeled based on power-loss maps. The losses are obtained based on the corresponding component speed, torque and, in case of GB, the selected gear. Table (1) Vehicle parameters of an exemplary C-segment vehicle 𝑐 [-] 0,26

𝐴 [m2] 2,2

𝑐 [-] 0,001

𝑚 [kg] 1500

353

Impact of future 48 V-systems on powertrain operation under real-driving conditions

Figure (3) Initial 600 s of CDCs with average speed 𝑣̅ and standstill share 𝑥 (b) Urban High, (c) Extra-Urban, (d) Highway

. (a) Urban Low,

A statistical evaluation of 48 V-hybrid vehicle measurements is carried out to investigate the total consumer load under different ambient conditions. The considered vehicles are equipped with state-of-the-art 48 V-consumers such as electric air conditioning, electrical water pump or electrically assisted turbocharger. For this study, the electrical loads are simply considered as the average consumer load at 20 °C ambient temperature based on the statistical distribution in the field measurement data.

3.3 System Requirements and Experiment Design In this study, the relationship between 48 V-system parameters and system requirements is analyzed for a P2 topology in different real-driving scenarios. The focus is on the implementation of a transition control between ICE operating states, under consideration of its influences on exemplary system requirements, such as efficiency (Δ𝐶𝑂 ), comfort (start frequency 𝐹 , ) and system cost (represented by effective battery load 𝑃 , ). The transition control is defined by hysteresis parameters 𝑘 , and 𝑘 , for ICE start and stop from (6). Accordingly, the design space is defined by two OS parameters, while the explored function space is represented by three exemplary system requirements. Requirements regarding vehicle performance and agility are neglected in this research to reduce complexity. Consequently, the full EM power is available for hybrid functions in the simulation.

354

Impact of future 48 V-systems on powertrain operation under real-driving conditions

4 Results

Figure (4) Impact of OS parameters on 48 V-system requirements in four CDCs

The hysteresis parameter effects are analyzed in four CDCs based on simulation results from the framework in Figure (2). The influence of 𝑘 , and 𝑘 , on exemplary system requirements is displayed in Figure (4) for an interval of 0–0.4 and 0–0.8 for the variables respectively. The sensitivity of ICE start hysteresis is significantly higher compared to the ICE stop hysteresis, which is also reflected by the difference in variation intervals. This fact can be explained by the distribution of powertrain loads in a DC. The share of high loads requiring ICE operation is typically less than the share of moderate loads, where the OS favors ED in order to increase efficiency. Therefore, a hysteresis on ICE start will be effective during a larger timeshare, at moderate loads, as in the comparably rare phases of ICE operation. The surfaces in Figure (4) reveal two conflicting optimization goals for the OS parameters. Start frequency 𝐹 , decreases with a higher hysteresis value for both ICE start and stop, while additional emissions increase. The investigated OS parameters have Δ𝐶𝑂 and effective battery load 𝑃 , the highest impact on the Urban Low and Urban High scenarios resulting from a high initial start frequency, combined with a high share of ED.

355

Impact of future 48 V-systems on powertrain operation under real-driving conditions

Figure (5) Pareto fronts of OS parameters in investigated CDCs

Pareto fronts for the fulfillment of the system requirements are extracted from the parameter variation in each CDC and are displayed in Figure (5). The increase in CO2 corresponding to a reduction of start frequency is significantly larger in urban scenarios. In extra-urban scenarios, the start frequency can be reduced by more than 50 % with an increase of less than 1 % in CO2. However, simultaneously the maximum effective battery load of all CDCs in Figure (5)(b) increases by up to 1 kW, which potentially increases the system costs. It can be concluded that the reduction of start frequency is CO2-sensitive in urban scenarios, while it is potentially cost-sensitive in extra-urban scenarios. An optimal compromise can be selected from Pareto fronts in (7) by introducing weighting factors to each system requirement and consequently minimizing the cost function in (8). In the cost function, each solution from the Pareto fronts is normalized based on the maximum and minimum value occurring in the corresponding dimension. The global maximum and minimum values from all CDCs are applied for the individual normalization, since a constant ratio between system performance and cost is desirable to provide a robust system behavior, while utilizing the maximum component potential in all CDCs equally. In this study, the weighting factors 𝑤 , , 𝑤 and 𝑤 , in (8) are set to 2, 1 and 1 respectively. 𝑃: min 𝑓 𝑥

356

,𝑥

,

,𝑥

,

(7)

Impact of future 48 V-systems on powertrain operation under real-driving conditions

𝑓 𝑥 𝑤

𝑥 ΔCO

,

− ΔCO , − ΔCO +𝑤

,𝑥

,𝑥

,

,

𝑥

+𝑤

,

𝑥 ,

𝑃

,

,

,

=

,

F , −𝑃 −𝑃

−F

,

−F

, ,

, ,

(8)

, ,

. ,

,

Figure (6) (a) Optimal OS parameters from Pareto fronts based on objective function (7), (b) average propulsion power for ICE start and stop

Figure (6)(a) shows the resulting hysteresis values for 𝑘 , and 𝑘 , from the identified Pareto solutions. The stop hysteresis is higher in all CDCs due to the distribution of powertrain loads discussed earlier. A start hysteresis is selected only for urban scenarios, which can be explained by the comparably high impact on effective battery load in extra-urban scenarios. Based on the simulation results, the resulting average propulsion power for ICE start and stop is evaluated and displayed in Figure (6)(b). The values vary based on the DC-specific equivalence factor in (4), the instantaneous battery energy deviation in (5) and the instantaneous component speed, which is influenced by the GB shifting-strategy. Errorbars exemplify the distribution interval of one standard deviation 𝜎 in each direction. With the selected Pareto solutions, the ICE is started at 10–15 kW and stopped at -5–0 kW average propulsion power based on the specific CDC. From the results in Figure (6) it can be concluded that a scenario-based adaption is sensible for hysteresis parameters in an ECMS approach. Finally, the cumulated probability density functions (CDFs) for ICE-off durations in different CDCs are evaluated in Figure (7). Four different scenarios are compared. The baseline scenario represents a conventional vehicle with automatic start-stop function, where the combustion is

357

Impact of future 48 V-systems on powertrain operation under real-driving conditions

stopped at vehicle standstill. Additionally, three scenarios from the previously discussed Pareto fronts are compared. The optimum scenario is identified by the cost function in (8) and represents the results achieved with the parameters from Figure (6)(a). The Min. CO2 and Min. Fice,st scenarios define the individual minimum of the conflicting optimization goals for CO2 reduction and start frequency. Dots on the CDFs identify the corresponding 50 % percentile.

Figure (7) Cumulated density functions of ICE-off duration. Dots represent 50 % percentiles. (a) Urban Low, (b) Urban High, (c) Extra-Urban, (d) Highway

The total ICE-off percentage in the investigated P2 system varies between 80 % in Urban Low and 30 % in the Highway scenario. This represents an increase of 300 % compared to baseline in case of Urban Low. It is concluded that ICE operation time decreases drastically when introducing 48 V P2 systems, which has to be taken into account for the development of further vehicle systems such as cabin air conditioning or exhaust gas aftertreatment. Furthermore, the duration of ICE-off phases are shorter in a P2 system, especially in the Min. CO2 scenario, where frequent ICE starts and stops enable an efficiency-optimal operation at each time step. Based on the 50 % percentiles, the average ICE-off duration in the Min. CO2 scenario can be estimated between 15 and 40 s in the investigated CDCs. The optimum scenario identified in this study can increase this duration by 30–40 % with a maximum increase of 𝑘

𝜉

.

(16)

Therefore, we obtain the desired value for the heat capacity flow rate from the power series that displays the faster convergence in the considered operation point. This combination allows us to utilize the method in regions, where a mere power series expansion, as the one outlined in Sec. 2.1, would fail. Yet a further aspect of the combined moment-cumulant approach is of computational nature. It is well-known from mathematical statistics, that moments and cumulants are not independent but related to each other by algebraic means, such that 𝑘 =∑

(−1)

(𝑖 − 1)! 𝐵 , (𝑚 , 𝑚 , … , 𝑚

).

(17)

Here 𝐵 , represent the incomplete Bell polynomials. Due to the mathematical analogy to the generating functions known from mathematical statistics, this very relationship also persists in the given context. From this, we may draw the conclusion, that in order the derivatives of the natural logarithm do not have to be calculated to determine 𝜉 explicitly but can be obtained from algebraic equations contained in Eq. (17). Therefore, the gained performance of the controller is achieved at the price of little additional computational effort.

491

Model-based approach for on-demand temperature control

3 Application to the temperature control problem Now we seek to exemplify the method explained in the preceding chapter for the system outlined in Fig. 1.

3.1 Model for the charge air cooler In order to apply the control strategy discussed in Chapter 2, we first formulate the functional expression for the moment generating function, viz. Eq. (6). To this end, the DTC of the CAC is modeled by a heat exchanger, which exhibits one shell-side with two tube-side passes. From this, we have the following characteristics [14, 16, 17] ,

𝑃= (

(

(18)

)

)

where differences between the two passes, due to distinctions in the effective heat transfer areas and heat transfer coefficients are taken account of by 𝜖. To determine the effective thermal conductibility 𝛾, we follow along the lines of Ref. [12], where we have to explicitly take the temperature dependence into account. A validation of model (17) based on real-driving situations reveals an overall mean absolute error of below 2K.

3.2 Model for the coolant mass flow The actuating variable under consideration in our context is given by the rotational of the pump. Therefore, in order to facilitate our control method, we have speed 𝑛 to relate the coolant mass flow 𝑚 , to this quantity. To do so, we utilize the model outlined in [12], where, for the system under consideration, we find a linear depend. In Fig. 2, a comparison between the measured (full line) ence, e.g. 𝑚 , ∼ 𝑛 normalized mass flow 𝑟: =

, , ,

,

(19)

and the respective modeled value (dotted line) is displayed. Here, we have as the maximum coolant mass flow measured within the system. The used 𝑚 , , depicted time series have been obtained by concatenating various real-driving measurements under different ambient conditions. Overall, we find acceptable agreement between the two signals with a mean absolute deviation of 7.35%.

492

Model-based approach for on-demand temperature control

4 Results and Discussion In order to apply the control strategy explained in Chapter 2, using Eq. (10) we insert the DTC of the CAC, Eq. (18), into Eq. (11). Taking terms up to second order into account, we calculate the moments 𝑚 , 𝑚 as well as 𝑚 offline, to implement them as

Fig. 2 – Comparison between the measured normalized coolant mass flow (full line) and the modeled value (dotted line) for various driving situations.

characteristic maps such that the derivatives do not have to be calculated explicitly by the automotive ECU. From this, utilizing Eq. (17), we obtain the corresponding cumulants. Thus, the overall solution can be calculated by the ECU under real-time conditions. The code was implemented and provided to the automotive ECU with the help of a software bypass. To evaluate the applicability of the method, test measurements were carried out for various real-driving situations and different ambient conditions. Representative results from this measurement campaign are shown in Figs. 3 and 4 respectively. Fig. 3 displays the performance of the controller at an ambient temperature which varied between 20°C and 28°C for a constant desired value for the charge-air temperature (full black line).2 We compare the actual charge-air temperature shown as the dashed black line with this target. The dotted black lines mark a boundary of a five-Kelvin distance to the desired temperature value. We observe that the charge-air temperature hardly ever exits this zone. The grey dashed line represents the gas temperature upstream to the CAC. It 2

Here and in the following, we utilize the axis notation that 𝑇 , , , corresponds to the maximum desired value for the CAC within the respectively shown measurement.

493

Model-based approach for on-demand temperature control is clearly displayed, that the driving situations are of a highly dynamical nature, as can be seen both, by the absolute difference between 𝑇 , , and 𝑇 , , as well as by the steep temperature peaks occurring for the temperature upstream to the cooler. This circumstance is also underlined by the normalized heat capacity flow rate for the charge air, as displayed in the inset of Fig. 3 by the grey full line. For this graph, we also observe steep peaks corresponding to distinct load changes. As can be seen from the same inset, the corresponding quantity for the coolant displays a comparatively smooth variation with time. Overall, we find a mean absolute control error for all measurements with contant desired value of 1.46K. In order to estimate, whether an on-demand temperature requirement is feasible within the presented method, measurements were conducted where the desired value for the charge-air temperature is changed abruptly. This scenario can be viewed as a stepwise excitation of the system and as such is very challenging for the controller. Fig. 4 exhibits the results for an ambient temperature, which varied between 18°C and 27°C. Again, the controller has been tested under dynamical driving situations, as displayed by the

Fig. 3 – Illustration of the temperature control performance: The charge-air temperature downstream to the cooler (black dashed line) is compared with the corresponding desired value (black full line). The dotted lines respectively illustrate a five-Kelvin gap to the desired value. The highly dynamical driving situations are displayed by the working fluid temperature upstream to the heat exchanger (grey dashed line). Inset: The dynamical driving situations are also reflected by the normalized heat capacity flow rates for the charge air (full grey line). In contrast to this behavior, the corresponding value for the coolant (dashed black line) exhibits a comparatively smooth variation in time.

494

Model-based approach for on-demand temperature control gas temperature upstream to the CAC (grey dashed line). Overall, we find good control performance for the charge-air temperature (dashed black line) in comparison with the corresponding desired value (full black line). Again, the dotted lines mark a five-Kelvin distance to 𝑇 , , , . In spite of the overall good control quality, unlike in Fig. 3, here we find distinct situations, where 𝑇 , , exceeds the five-Kelvin band. They appear between 1500s and 2200s in the shown measurement. A closer inspection reveals, that the strong deviations are due to a combination of two effects: On the one hand, the comparatively high desired temperature values in these segments lead to low values of . This causes long reaction times within the cooling fluid. Prior to the strong 𝑛 charge-air temperature peaks, highly dynamical driving situations can be observed, which lead to an extremely fast reaction within the working fluid side of the CAC. The comparatively inert cooling fluid is unable to follow these abrupt changes.

Fig. 4 – Illustration of the temperature control performance for a non-constant desired value (full black line): The charge-air temperature downstream to the cooler is shown as a black dashed line. The dotted lines respectively illustrate a five-Kelvin gap to the desired value. The highly dynamical driving situations are displayed by the working fluid temperature upstream to the heat exchanger (grey dashed line). Inset: The dynamical driving situations are also reflected by the normalized heat capacity flow rates for the charge air (full grey line). In contrast to this behavior, the corresponding value for the coolant (dashed black line) exhibits a comparatively smooth variation in time.

495

Model-based approach for on-demand temperature control The best-suited reaction in such a situation is to run the coolant pump at full rotational speed. This is actually observed for the measuring points under consideration as shown (dashed grey line), for which the right-hand side in Fig. 5. Here we witness, that 𝑛 ordinate applies, is at the maximum value for the high control deviations. In contrary to these situations, Fig. 6 gives a closer inspection on the time-series for changes in the desired value, where the aforementioned combination of long reaction times within the cooling circuit and the highly dynamical nature in the intake air-system is absent. Under these circumstances, 𝑇 , , (black dashed line) does follow 𝑇 , , , (full black line) far better and we find 𝑇 , , , − 𝑇 , , ≤ 5K except for short periods after abrupt changes in the desired value. This delay is due to the intrinsic thermal masses within the system. Again, also the gas temperature upstream to the CAC is depicted by the dashed grey line to demonstrate the dynamics within the measurement. Moreover, the inset contains the heat capacity flow rates for the charge air (full grey line) and the coolant (dashed black line). We observe, that the dynamical driving situations are again

Fig. 5 – Closer inspection of the temperature control performance for a non-constant desired value (full black line) in situations with high control deviations: The charge-air temperature downstream to the cooler is shown as a black dashed line. The dotted lines illustrate a five-Kelvin gap to the desired value. The rotational speed of the pump (dashed grey line), for which the right-hand side ordinate applies, displays, that the controller runs the pump at maximum load for measuring points with high control deviations. However, due to the long reaction times of the cooling system for the high desired values, the control deviations cannot be avoided.

496

Model-based approach for on-demand temperature control displayed in the time series of Γ . Finally, the separated excitations in the system, i.e. the changes in 𝑇 , , , , as well as the load variations appear in Γ , which is obtained, from the control strategy. We conclude this chapter with some final remarks on the measurements containing the varying desired value use-case. First, it has to be stated, that the step-like changes utilized randomly within our investigations represent a worst-case scenario rather than a realistic behavior in a vehicle, where a suitable coordinator usually shapes a smooth behavior of the desired temperature value. Additionally 𝑇 , , , in a highly dynamical driving situation is unlikely to remain comparatively high persistently. Instead, for these situations a lowering of the target value is more appropriate, as this offers benefits concerning driving comfort and power density. Therefore, we actually expect the control

Fig. 6 – Illustration of the temperature control performance for a non-constant desired value (full black line): The charge-air temperature downstream to the cooler is shown as a black dashed line. The dotted lines illustrate a five-Kelvin gap to the desired value. The temperature upstream to the charge-air cooler (dashed grey line) underlines the dynamical nature of the driving situations. Inset: The heat capacity flow rates for the charge air (full grey line) and the cooling fluid (dashed black line) display that the working fluid exhibits rapid changes while the time series for the coolant varies smoothly. Both, the step-like changes as well as the sudden load variations are distinguishable in this signal. However, due to their separate occurrence, the controller is able to fulfil the on-demand temperature requirement overall.

497

Model-based approach for on-demand temperature control quality in an industrial application to be even higher than within the discussed worstcase scenario. However, in spite of the demanding operation profile, overall we find the absolute control error for all investigated measurements with a step-like varying desired value to be at 2.18K. In particular, due to the model-based nature of the control, this performance turns out to be widely independent of the ambient temperature.

5 Summary and outlook In this publication, we have reported on a recently suggested control algorithm. The method is based on a power series expansion of the underlying physical model with respect to the correction which has to be taken by the controller. The formal analogy to a moment generating function as known from mathematical statistics inspired us to define a cumulant generating function, which again may be expanded into a power series. Thus, we established two independent methods, which determine the coolant mass flow necessary to ensure good control quality. From a comparison of the respective remainder terms, we decide which method is more appropriate in the given working point. This circumstance extends the applicability as compared to a mere power series expansion of the physical model. Due to the algebraic relations between the cumulants and the moments, the additionally gained control performance is achieved at the price of little additional computational effort. The control algorithm has been applied to an indirect charge-air cooling system. To this end a physical model for the heat exchanger and the coolant pump has been utilized. The implementation of the method on an automotive ECU allowed to estimate the control performance under demanding conditions, including dynamical driving situations and step-like variations in the desired value. Altogether we find a mean absolute temperature deviation of 1.46K for the scenario of a constant desired value. The far more challenging case of step-like on-demand temperature requirements, exhibits a slightly reduced control quality, mainly induced by an intrinsic lack of response within the cooling fluid to the highly dynamical driving situations. However, averaging over all measurements containing rapid changes in the desired value, the mean absolute temperature deviation of 2.18K constitutes a very promising result. Ongoing research at IAV comprises the application of the outlined control approach to other applications in the context of on-demand control problems. Furthermore, the generalization of the method to multivariate control with a focus on thermal management applications is investigated currently.

498

Model-based approach for on-demand temperature control

Bibliography 1. Roberts, A., Brooks, R., Shipway, P., Energ Convers Management 82, 327, 2014. 2. Mollenhauer, K., Tschoeke, H. (eds.): "Handbook of Diesel Engines", SpringerVerlag Berlin-Heidelberg (2010). 3. Käppner, C., Garrido Gonzales N., Drückhammer, J., Lange, H., Fritzsche, J., Henn, M. in: Bargende, M., Reuss, H. –C., Wiedemann, J. (eds.): "17. Internationales Stuttgarter Symposium Proceedings", Springer Fachmedien Wiesbaden GmbH (2017). 4. Kadunic, S., Baar, R., Scherere, F., Zegenhagen, T., Ziegler, F., 22. Aachen Colloquium, (2013). 5. Wawzyniak, M., Art, L., Jung, M., Ben Ahmed, F., ATZ worldwide, 09, 47 (2017). 6. Graf, F., Lauer, S., Hofstetter, J., Perugini, M., MTZ worldwide, 10, 39 (2018). 7. Nöst M., Doppler, C., Klell, M., Trattner, A., In: Watzenig, D., Brandstätter, B. (eds.): "Comprehensive Energy Management – Safe Adaption, Predictive Control and Thermal Management, Automotive Engineering: Simulation and Validation Methods”, Springer International Publishing AG, Cham (2018). 8. Saygili, Y., Eroglu, I., Kincal, S., Int. J. Hydrogen. Energy, 40, 615 (2015). 9. O’Keefe, D., El-Sharkh, M. Y., Telotte, J. C., Palanki, S., J. Power Sources, 256, 470 (2014). 10. Castiglione, T., Pizzonia, F., Bova, S., SAE Int. J. Mater: Manf. 9(2), 294, (2016). 11. Herzog, A., Skorupa, F., Meinecke, R., Frase, R., MTZ worldwide 5, 24 (2014). 12. Herzog, A., Pelka, C., Skorupa F., in: "Energy and Thermal Management, Air Conditioning, Waste Heat Recovery". 1st ETA Conference, December 1-2, 2016, Berlin, Germany (2017). 13. Herzog A., "Verfahren zur Lösung instantaner Regelungsprobleme”, Patent application (submitted 2019). 14. VDI Heat Atlas, 2nd edn. p. 34 ff., Springer-Verlag, Berlin-Heidelberg (2010). 15. Abramowitz M., Stegun, I. (eds.): "Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables”, National Bureau of Standards, Applied Mathematics Series 55. 16. Roetzel, W., Spang, B., Fortschr.-Ber. VDI, vol. 19(18). VDI-Verlag, Düsseldorf (1987). 17. Roetzel, W., Spang, B., Chem. Eng. Res. Des. 67, 115 (1989).

499

Preheating components with metal hydrides or lime – Small, high power, no additional energy Mila Kölbig (0711/6862-214, [email protected])a Inga Bürger (0711/6862-492, [email protected]) a Matthias Schmidt (02203/601-4091, [email protected]) b Marc Linder (0711/6862-8034, [email protected]) a Deutsches Zentrum für Luft- und Raumfahrt (DLR), Institut für Technische Thermodynamik a Pfaffenwaldring 38-40, 70569 Stuttgart b Linder Höhe, 51147 Köln

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_44

501

Preheating components with metal hydrides or lime – Small, high power, no additional energy

1 Motivation Operation temperature of vehicle components in internal combustion engines is far above ambient temperature, in particular compared to winter ambient temperatures of down to -20 °C. At the cold start phase at the beginning of a driving cycle, the systems’ performance is highly inefficient, which leads to increased pollutants (NOx, CO, HC) and degradation of the components. 60-80% of all pollutants of the whole driving cycle are generated during this cold start phase. It takes several minutes until the system reaches operation temperature. The cold-start inefficiencies can be attributed to either the catalytic converter, lubricant or combustion wall being too cold. Later on in the driving cycle, waste heat is available. [1-4] The cause of the effect of the lubricant temperature on the vehicles’ emission is the strong temperature dependency of its viscosity. At low temperatures, the high viscosity leads to increased pump work at increased pressure as well as to increased frictional losses in the engine. These effects lead to decreased engine efficiency. Due to the strong temperature dependency of lubricant viscosity, small temperature increases at low temperatures result in large decreases of the viscosity. Hence the viscosity can be reduced by 65 % in the first 20 °C of heating. For a preheater application this means a small temperature increases can reduce friction and its effects on pollutants considerably. Later on, the lubricant reaches temperature of 110 to 130 °C. [4, 5] For the catalyst in the exhaust gas treatment of vehicles, the chemical reactions require at least 250 °C. The light of time describes the duration until the catalyst reaches temperatures where 50 % conversion efficiency is reached. This takes at least 150 to 210 s. Later on, the system operates at temperatures of up to 450 °C. [4, 6]  In order to reduce pollutants and degradation during cold start, this work investigates systems which can preheat vehicle components very quickly and be regenerated only by the available waste heat.

2 Approach Reversible gas/solid reactions are considered for the described preheating challenge. The chemical reaction of a solid A with a gas B to a solid AB releases heat ΔH, as depicted in Figure 1, left to right. This can be used to generate heat for preheating. If (waste) heat is available, the back reaction can be triggered and the gas is released, Figure 1, right to left. The gas is stored separately during the storage phase of the preheater to prevent the back reaction. The reaction has a defined correlation between temperature and pressure which is schematically depicted in a van’t Hoff plot at the bottom of Figure 1. The logarithmic pressure is shown over the reciprocal temperature and the equilibrium approaches a line. For every pressure, the reaction has a certain

502

Preheating components with metal hydrides or lime – Small, high power, no additional energy equilibrium temperature and vice versa. This diagram is a useful tool to predict the reaction temperature level at a certain pressure.

Figure 1. Gas/solid reaction and equilibrium line in van’t Hoff plot

There are several advantages of gas/solid reactions which make them attractive to be used as a preheater in vehicles. ● The heat is stored as chemical potential. Therefore, no thermal losses occur over time. This allows long parking periods without losing functionality. ● The heat can be generated on demand right when it is needed by adding the reaction partner. ● The temperature level is determined by the pressure level. Different materials work at different temperature and pressure ranges. This allows a wide area of application. ● The back reaction can be triggered by waste heat. Hence, no additional energy is required for preheating This paper presents two specific materials optimized for two preheating use cases: metal hydride for preheating lubricants and lime for preheating catalysts.

2.1 Preheating lubricant with metal hydrides 2.1.1 Metal hydrides Metal hydrides are metal alloys which absorb and desorb hydrogen in an equilibrium reaction as described above. They can react very quickly even at low ambient temperatures. Since hydrogen is not available in conventional vehicles, the hydrogen has to

503

Preheating components with metal hydrides or lime – Small, high power, no additional energy be stored in another metal hydride. This lead to coupled, strongly interdepended reactions for the operation of the preheater. Such a system is schematically given in Figure 2. The hydrogen never crosses the systems boundary and is not consumed but only reversibly exchanged between the two metal hydrides. The materials chosen in this work operate at pressure levels below 30 bar.

Figure 2. Coupled metal hydride reactions in a closed system

For heat generation at temperatures of -20 °C, no publications with coupled metal hydride reactions could be found other than the authors own work [7, 8]. At higher temperatures, some work was published [9-12]. A fairly comparable work is the investigation by Qin et. al [13]. They showed thermal peak power of 0.48 kW/kgMH at 20 °C.

2.1.2 Experimental and reactor design The described use case of preheating lubricant is used to define the boundary conditions for the metal hydride system. The preheating reaction is investigated at temperatures between -20 and 20 °C. Regeneration temperature level from the lubricant is considered to be between 90 and 130 °C. Experiments were performed in laboratory scale investigating the thermal power transferred to a heat transfer fluid. This shows the potential of the systems for vehicle applications. The metal hydrides Hydralloy C5 (Ti0.95Zr0.05Mn1.46V0.45Fe0.09) from GfE Metalle und Materialien GmbH (Germany) and LaNi4.85Al0.15 from WholeWin (China) were chosen for the use case of preheating lubricants. Metal hydrides are powders with small particle size in the range of several micrometers. They have a very low thermal conductivity of around 1 W/mK. For high thermal power, both gas transport of the reaction partner through the powder and the thermal conduction of the generated heat have to be optimized. This is a great challenge for the small particle powders. Great care was taken for the reactor design to enable high thermal power output. The container for the metal hydrides is based on a tube bundle heat exchanger. The metal hydride is positioned on the shell side and heat transfer fluid flows through the tubes. The maximal travel distance for the hydrogen through the powder is 38 mm. The maximal distance of the powder to the tubes is around 1 mm, allowing fast heat transfer. The reactors contain 615 g of Hydralloy C5 and 960 g of LaNi4.85Al0.15. More details on the material selection, reactor optimization and experimental method are given in [7, 8].

504

Preheating components with metal hydrides or lime – Small, high power, no additional energy

2.1.3 Results

Thermal power in kW/kg

The main results found for the metal hydride preheater are given in the following. Different influence factors were investigated, such as regeneration temperature, mass flow of the heat transfer fluid and ambient temperature. The grates effect on the thermal power by far was posed by the ambient temperature. Results for the specific thermal power over time for different ambient temperatures are given in Figure 3.

1.5

20°C

1.0 0°C -20°C

0.5 0.0 0

100

200

300

Time in s Figure 3. Specific thermal power over time for different ambient temperatures [7, 8]

When the reaction is triggered at t = 0 s, the thermal power increases steeply to values of 0.6 kW/kgMH at -20 °C, 1.1 kW/kgMH at 0 °C and 1.6 kW/kgMH at 20 °C. After a peak is reached, the thermal power decreases again. The faster reaction at higher temperature leads to a shorter duration for full conversion and the thermal power decreases faster. Only at -20 °C, the reactions take longer than 300 s for completion. Compared to literature (see above), the reached value at -20 °C is in the same range, however at a temperature 40 K lower. With the given performance of the system, the dimensions for a possible preheater should be estimated. Heating 2 l of lubricant by 20 K in order to minimize the negative effects due to viscosity requires 0.2 MJ, according to Kunze et. al [14]. With the shown performance of around 0.5 kW/kgMH, 1 kg of metal hydride would take 400 s for heating the lubricant. If 2 kg of metal hydride are be used, the heating time decreases to 200 s. This shows the great potential of the metal hydride system run without additional energy.

505

Preheating components with metal hydrides or lime – Small, high power, no additional energy A much easier system could be applied in fuel cell vehicles. Fuel cells also have a cold start problem. The produced water freezes and expands, which causes mechanical degradation [15]. The available hydrogen infrastructure in fuel cell vehicles offers different pressure levels (at least 30 bar in tank, around 2 bar in fuel cell), which can be used as free driving force without any hydrogen consumption. Hydrogen form the tank is absorbed at the beginning of driving and heat is generated to preheat the fuel cell. The lower pressure of the fuel cell is than used during driving to trigger the desorption reaction. Therefore, only on metal hydride is required for the preheater, because no hydrogen has to be stored in a second one. The higher pressure level allows even higher thermal power. In previous work of the author [7, 8] values of over 5 KW/kgMH at -20 °C and hydrogen pressure of 10 bar could be shown.

2.1.4 Conclusion ● Very fast reactions of coupled metal hydrides at low temperature were proven ● The thermal power of the coupled reactions depends strongly on ambient temperature ● The thermally driven, coupled reactions reach thermal power values of over 0.5 kW/kgMH at -20 °C ● With 2 kg of metal hydride, 2 l of lubricant could be heated by 20 K within 200 s

2.2 Preheating catalysts with lime 2.2.1

Lime and water

The reaction of calcium oxide with water to calcium hydroxide offers a high reaction enthalpy of around ΔH = 100 kJ/mol, as given in equation (1). Hence, the material offers a high energy density. 𝐶𝑎𝑂( ) + 𝐻 𝑂(

)

↔ 𝐶𝑎(𝑂𝐻)

( )

+ ∆𝐻

(1)

Additionally, calcium oxide/ hydroxide is nontoxic, widely available and cheap. The described reaction is usually considered with water vapor. The equilibrium reacts at around 500 °C and 1 bar water vapor [16]. This temperature level would fit the use case of preheating the catalyst in vehicles. As described above, preheating the catalyst to 250 °C very quickly would enable the cleaning reactions earlier and hence could save a lot of pollutants. The preheater can be regenerated while driving at the operation temperature of the catalyst of up to 450 °C. In order to save the water for the next preheating cycle, it has to be condensed at ambient temperature. The preheater then would consist of on reactor with calcium oxide generating heat for preheating and a connected heat exchanger condensing and storing water.

506

Preheating components with metal hydrides or lime – Small, high power, no additional energy

2.2.2 New operation mode for preheating However, for preheating in vehicles, water vapor at ambient temperatures of -20 °C is very low and no thermal energy is available for evaporation. Therefore, a new operation mode was developed. If liquid water is added to the reactant, the initial reaction is extremely fast. Compared to water vapor, the reaction with liquid water reaches lower temperature levels. In simple laboratory test, temperatures of up to 300 °C could be shown. This could already be high enough to allow the catalyst to work. The temperature level can be increased further, however, by using the reaction heat with liquid water to evaporate water and allowing the pressure to rise. This leads to reactions with water vapor at high higher pressure levels and hence at higher reaction temperature. Consequently, the thermal power and the end temperature can be increased further. The concept of the liquid-vapor lime preheater is depicted in Figure 4. The equilibrium of the lime-water reaction is shown schematically in the van’t Hoff diagram. At , the reaction partner is added liquidly. For use cases below ambient temperature 𝑇 0 °C, an antifreezing compound can be added. The reaction with liquid water increases the temperature and enables evaporation, shown by the lower arrow in Figure 4. In a closed reactor, the pressure rises and the reaction temperature increases further along its equilibrium, see upper arrow in Figure 4. This allows heat for preheating at ) depending on the reached pressure level. During high temperature levels (𝑇 driving, the system can be regenerated by waste heat from the catalyst at its operation temperature 𝑇 and the released water is condensed and stored in a condenser.

Figure 4. Fast reaction with liquid water for evaporation allowing high reaction temperature

507

Preheating components with metal hydrides or lime – Small, high power, no additional energy This newly developed operation mode allows high reaction temperature, high energy density and high thermal power at low temperatures. With the high reaction enthalpy, 1 kg of calcium oxide can store 0.5 kWh and requires 320 g of water for complete reaction. The system is under investigation at the moment to show its potential.

2.2.3 Conclusion ● The newly developed liquid-vapor lime preheater has the potential for fast heat-up to high temperatures, e.g. for catalysts ● The system reaches reaction temperatures of 500 °C at 1 bar of water vapor ● 1 kg of calcium oxide can store 0.5 kWh of thermal energy

3 Conclusion 60-80 % of all pollutants of the whole drive are generated within the first minutes of driving when the system is not at operation temperature yet [2, 3]. Shortening this start-up time by preheating specific components such as the lubricant or the catalyst can save many pollutants and reduce degradation. The present paper proposes reversible gas/solid reactions which generate heat on demand and can be regenerated by onboard surplus energy. Due to the storage of the thermal energy as chemical potential by separating the reaction partners, the heat is stored free of loss as long as required. The proposed metal hydride preheater can heat lubricant to temperatures where the most negative effects due to viscosity can be reduced to a minimum. With the shown high thermal power of up to 0.5 kW/kgMH at -20 °C, 2 l of lubricant can be preheated within 200 s by 2 kg of material. The system is regenerated by waste heat and hence doesn’t require additional energy. The newly developed liquid-vapor lime preheater has the potential for fast heat-up with high thermal power to high temperatures, e.g. for catalysts. 1 kg of material can store 0.5 kWh of thermal energy. The system is regenerated by waste heat during driving and no additional energy is required.

508

Preheating components with metal hydrides or lime – Small, high power, no additional energy

Bibliography 1. J. D. Trapy and P. Damiral, "An Investigation of Lubricating System Warm-up for the Improvement of Cold Start Efficiency and Emissions of S.I. Automotive Engines," (in English), SAE technical paper, 1990. 2. M. S. Reiter and K. M. Kockelman, "The problem of cold starts: A closer look at mobile source emissions levels," Transportation Research Part D: Transport and Environment, vol. 43, pp. 123-132, 2016. 3. R. Cipollone, D. Di Battista, and M. Mauriello, "Effects of oil warm up acceleration on the fuel consumption of reciprocating internal combustion engines," Energy Procedia, vol. 82, pp. 1-8, 2015. 4. A. Roberts, R. Brooks, and P. Shipway, "Internal combustion engine cold-start efficiency: A review of the problem, causes and potential solutions," Energy Conversion and Management, vol. 82, pp. 327-350, 2014. 5. G. Andrews, A. Ounzain, H. Li, M. Bell, J. Tate, and K. Ropkins, "The use of a water/lube oil heat exchanger and enhanced cooling water heating to increase water and lube oil heating rates in passenger cars for reduced fuel consumption and CO2 emissions during cold start.," SAE technical paper, vol. 2007-01-2067, 2007. 6. H. Li et al., "Study of Thermal Characteristics and Emissions during Cold Start using an on-board Measuring Method for Modern SI Car Real World Urban Driving," (in English), SAE Int J Engines, 2008. 7. M. Dieterich, I. Bürger, and M. Linder, "Open and closed metal hydride system for high thermal power applications: Preheating vehicle components," International Journal of Hydrogen Energy, vol. 42, no. 16, pp. 11469-11481, 2018 2017. 8. M. Kölbig, "Coupled metal hydride reactions for preheating vehicle components at low temperatures," ed, 2018. 9. P. M. Golben, D. DaCosta, and G. Sandrock, "Hydride-based cold-start heater for automotive catalyst," Journal of Alloys and Compounds, vol. 253-254, pp. 686-688, 1997. 10. D. H. Dacosta, M. Golben, and D. C. Tragna, "Metal hydride systems for the hydrogen planet," in Proceedings of the 14th World Hydrogen Energy Conference, R. D. Ventor, T. F. Bose, and N. Veziroglu, Eds., ed. Montreal: IAHE, 2002. 11. I.-S. Park, J.-K. Kim, K. J. Kim, J. Zhang, C. Park, and K. Gawlik, "Investigation of coupled AB5 type high-power metal hydride reactors," International Journal of Hydrogen Energy, vol. 34, pp. 5770-5777, 2009.

509

Preheating components with metal hydrides or lime – Small, high power, no additional energy 12. Z. Z. Fang et al., "Metal hydrides based high energy density thermal battery," Journal of Alloys and Compounds, vol. 645, pp. S184-S189, 2015. 13. F. Qin, J. Chen, M. Lu, Z. Chen, Y. Zhou, and K. Yang, "Development of a metal hydride refrigeration system as an exhaust gas-driven automobile air conditioner," Renewable Energy, vol. 32, pp. 2034-2052, 2007. 14. K. Kunze, S. Wolff, I. Lade, and J. Tonhauser, "A systematic analysis of CO2reduction by an optimized heat supply during vehicle warm-up," SAE technical paper, 2006. 15. J. Mishler, Y. Wang, P. P. Mukherjee, R. Mukundan, and R. L. Borup, "Subfreezing operation of polymer electrolyte fuel cells: Ice formation and cell performance loss," Electrochimica Acta, vol. 65, pp. 127-133, 2012. 16. M. Schmidt, A. Gutierrez, and M. Linder, "Thermochemical energy storage with CaO/Ca(OH)2 – Experimental investigation of the thermal capability at low vapor pressures in a lab scale reactor," Applied Energy, vol. 188, pp. 672-681, 2017.

510

Fast running detailed battery thermal management models based on 1D-3D synergetic approach Dig Vijay, Nils Framke, Peter Stopp Gamma Technologies GmbH

Fast running detailed battery thermal management models based on 1D-3D…

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_45

511

Fast running detailed battery thermal management models based on 1D-3D…

Abstract Efficient energy management of electric vehicles, EVs, in terms of range optimization and cabin climate control, undoubtedly depends on the efficient battery thermal management. Due to inherently transient system behavior of EVs, the dynamic behavior of battery packs is almost always examined in conjunction with the entire vehicle. As a result, it is desired to have battery models that are not only fast running but also accurate. In this work a 1D-3D synergetic approach is presented for battery thermal management that is both fast running and accurate. In a 1D-3D approach flow through the cooling channels is based on 1D Navier-Stokes equation and thermal conduction through the battery pack is based on 3D heat equation. Moreover, the battery thermal model is integrated with a lumped electric model, whereas the electric cell model is the result of detailed electrochemical battery modelling. The battery modules designed for this work are based on indirect liquid cooling. Accordingly, pouch cells are vertically assembled with the cooling plates to achieve the maximum area contact. The final model is implemented to test several design criteria such as cell temperature, uniformity of temperature distribution, SOC level and thermal runaway. Furthermore, the numerical outcomes of the 1D-3D synergetic approach are also validated with a detailed 3D CFD conjugate heat transfer model of a single cell and plate assembly. In conclusion, 1D-3D synergetic approach is accurate enough for component testing and fast enough for entire vehicle testing.

1 Introduction and Objective With the advancements in the battery technology, over the last decade the world has seen a political desire to go for a cleaner energy. Almost every automotive manufacture today is working on electric vehicles (EVs), but there is still a long way ahead to produce EVs that are competitive in terms of price and driving range. Charging infrastructure is also a challenge to overcome. Besides all the desired improvements in the cell chemistry (e.g. energy density, power density, cycle life, fast charging and weight), battery thermal management is one common challenge for all the EVs. To name some of the challenges; cooing and heating of the battery packs, cold start and fast charging, uniformity of temperature at the cell as well as at the pack level. Additionally, in comparison to conventional vehicles, in case of EVs the flow circuits are more complex and interactive through active valving. The control strategies are ever more challenging in terms of defining the input signals for the components that simultaneously control the temperatures of cabin, battery, e-motor and other power electronics. Moreover, these controls are also responsible for preventing the uncontrolled operating conditions that otherwise can lead to battery failures and thermal runaway.

512

Fast running detailed battery thermal management models based on 1D-3D… With the above challenges in mind, due to inherently transient system behavior of EVs, the dynamic behavior of battery packs is almost always examined in conjunction with the entire vehicle. Therefore, during the predevelopment stages, EVs stand in need of numerical methodologies that are not only accurate but also computationally less expensive when it comes to integrated system analysis. Moreover, these numerical methodologies should be able to deal with zero flow conditions through robust flow models, as zero flow is a frequently occurring phenomena in case of EVs. Accordingly, the objective of this paper is to present a synergetic 1D-3D fluid flow and thermal solution that is both fast running at system-level and accurate at component-level. The models developed in this work are used to test both component and system behaviors such as maximum cell temperatures, uniformity of temperature distribution, SOC level under different charge-discharge cycles and thermal runaway. The numerical outcomes of the 1D-3D synergetic approach presented in this work are also validated through detailed 3D CFD numerical models.

2 Material and Methods 2.1 Problem description The battery module investigated in this work is based on indirect liquid cooling. It consists of 20 pouch cells assembled together, whereas 21 vertical cooling plates are placed between the cells, as shown in Figure 1-A. The module is designed to have two inlets on the left-hand side (both front and back) and two outlets on the right-hand side (both front and back), as shown in Figure 1-B & Figure 1-C. The modular assembly of cell and cooling plate is extendable to achieve different module sizes, see Figure 1-D. In case of all the odd numbered cooling plates (count from the front side) the vertical direction of flow is from bottom to top. In Figure 1-E the direction of flow through the first cooling plate (disassembled) can be seen, whereas in Figure 1-F the direct mapping of flow over the cell surface can be seen by completely removing the first cooling plate. Similarly, in case of all the even numbered cooling plates the vertical direction of flow is from top to bottom, In Figure 1-G and Figure 1-H the direction and mapping of flow through second cooling plate can be seen. In case of all the cooling plates (21 plates) the horizontal direction of flow is always from left to right. Because of the current packing layout, the coolant cross flow is achieved in subsequent cooling plates, which in turn helps in achieving uniform cell temperature, SOC and cell aging.

2.2 Cell characterization Battery cells investigated in this work were designed through an electrochemical model using GT-AutoLion. In GT-AutoLion, for a given cell chemistry, shape and size, it is possible to predict the cell performance well before the first test data is available [1-2].

513

Fast running detailed battery thermal management models based on 1D-3D… The principal cell data required to characterize the pouch cells used in this work can be seen in Table 1.

Figure 1. Design of the battery module based on pouch cell

Considering the focus of this work is to address the problems related to battery thermal management, 1D-3D synergetic approach presented in this work is indirectly complemented with the electrochemical cell models developed in GT-AutoLion rather than having a direct coupling, which otherwise is also possible [3-4]. The indirect coupling is achieved by characterizing RC battery equivalent models from detailed electrochemical cell models. Such a workflow is a common practice, where the cell experts deliver the RC equivalent models capable of supplying the heat rejection boundary conditions as a function of cell temperature and SOC etc. Hence, with a goal to extract the circuit parameters for an electric-equivalent battery model, Hybrid Pulse Power Characterization (HPPC) simulation results were produced from the electrochemical model created in GT-AutoLion. The battery characterization tool of GT-SUITE can take the current and voltage traces of HPPC tests or simulations and automatically extracts the circuit parameters required for the electrical-equivalent circuit models consisting of several RC branches. The outcome of battery characterization in terms of open circuit voltage and internal resistance can be seen in Figure 2. Table 1. Principal cell data used in GT-AutoLion Shape Capacity Dimensions (W × H × T) Cathode Active Material Anode Active Material Tabs (W × H × T) Cathode Tab Material Anode Tab Material

514

Pouch Cell 36.29 Ah 140 mm × 190 mm × 9 mm LFPO Graphite 30 mm × 20 mm × 3 mm Aluminum Copper

Fast running detailed battery thermal management models based on 1D-3D…

Figure 2. Circuit parameters for electric-equivalent battery model

2.3 1D-3D component analyses At Gamma Technologies (GT), for last several years a 1D-3D synergetic approach has been successfully implemented in GT-SUITE [5--7] for the component analyses of IC engines, e-motors, turbochargers, fuel cells, cabin, power electronics and of course battery packs. In a 1D-3D synergetic approach the fluid flow is solved one-dimensionally, where the flow model involves the solution of 1D Navier-Stokes equations, namely the conservation of continuity, momentum and energy equations. Moreover, the 1D flow domain is discretized through a staggered grid, where the scalar variables such as pressure, temperature, density, internal energy, enthalpy and species concentrations are solved at the centroid of a finite volume and the vector variables such as mass flux, velocity and mass fraction fluxes are solved at the boundaries of a finite volume. On the other hand, the thermal analyses in a 1D-3D synergetic approach is completely three-dimensional, where the heat equation is solved over a 3D finite element (FE) mesh. Accordingly, for the work presented in this work the 3D FE meshes were created in GEM3D, which a preprocessing tool of GT-SUITE. The key aspects of a 1D-3D synergetic approach are fast simulation time owing to 1D flow circuits and accurate component temperatures owing to 3D FE thermal domain. Once the 1D flow and 3D thermal domains are defined, the next step in a 1D-3D synergetic approach is to define the solid-fluid contact interface. In GT-SUITE, the convection boundary condition, also known as Newton boundary condition, see Eq. (1), is implemented to exchange the information of wall temperature and heat flux at a fixed time interval, whereas the time steps used to solve the flow and the thermal domain can be different. −𝑘

,

=ℎ 𝑇

,

− 𝑇 0, 𝑡

(1)

515

Fast running detailed battery thermal management models based on 1D-3D…

2.3.1 Modular battery modeling Owing to the modular nature of battery packs, at GT a methodology is implemented in which the focus is only on the repetitive building blocks of the system. For example, in case of the battery module presented in this work, if we consider the assembly of a cooling plate and battery cell as a single unit cell then we only need an array of two such unit cells to create the entire module. The reason that two units are required and not one has to do with the fact that there are two different types of cooling plates, as shown in Figure 3-A and Figure 3-B. However, in order to complete the model an additional assembly of end connector (both front and back side) and a single unit of cooling plate without a battery cell is also required as in the module under question there are 21 cooling plate surrounding 20 battery cells.

Figure 3. Two repetitive assemblies of cell and plate

With a focus on each unit cell, as shown in Figure 3-A and Figure 3-B, two assemblies containing 1D flow and 3D thermal domains were created in GEM3D (a preprocessing tool of GT-SUITE). These assemblies contain several input parameters such as flow and thermal properties, initial and boundary conditions, ambient conditions, thermal contact resistances etc. Furthermore, all the input parameters can be defined as variables changing from case to case. Moreover, such assemblies can be connected to each other through flow, thermal and electrical connections. In view of this, following steps are performed in GEM3D to prepare two compound templates: ● Thermal domain – Conversion of cell and cooling plate into 3D finite element mesh and assigning isotropic/anisotropic thermal properties. ● Flow domain – Conversion of cooling channels into 1D flow network with suitable discretization length and assigning flow properties. ● Electrical domain – Characterizing an electrical-equivalent lumped battery model with open circuit voltage and internal resistance.

516

Fast running detailed battery thermal management models based on 1D-3D… ● Internal and external connections – Assigning conductive connection between cell and plate, convective connection between fluid and cooling plate, boundary selection for natural convection to ambient, boundary selection to define external conduction connections to neighboring cells and plates that are residing inside neighboring compound templates. ● Internal and external signals – Defining external signals to electrically connect the lumped electric battery models residing inside neighboring compound templates, defining internal signals to connect thermal and electric cell domains to each other. Implementing the above steps in GT-SUITE allows for the modular building and scaling of the battery modules or packs. In Figure 4 on the left side, the model buildup is demonstrated in the form of two highlighted base assemblies. Moreover, Figure 4 also shows the 3D contours of cell temperatures for a steady state simulation, where the batteries are discharged at a C-rate of 2C and the coolant inlet temperature and single plate flow rate were 25 °C and 2 g/s, respectively. The low temperature gradients in Figure 4 demonstrate the effective cooling of the battery module in terms of uniform cell temperature distribution.

Figure 4. 3D contours of cell temperatures in GT-POST (post processing tool of GT-SUITE) for a steady state module discharge at a C-rate of 2C

517

Fast running detailed battery thermal management models based on 1D-3D…

3 Model Validation The 1D-3D synergetic approach presented in this work assumes the fluid flow to be one dimensional, which is a realistic assumption for regular flow geometries such as the one implemented in this work. However, to support this argument, additional 3D CFD simulations were performed to validate the assumption of simplified 1D fluid flow. Accordingly, two separate validation models were built, one that focused on the pressure drop by only considering the fluid domain in an isothermal problem. Another one focused on the heat transfer by considering the solid and fluid domain in a conjugate heat transfer problem. For each 3D CFD simulation (fluid or solid-fluid) an equivalent 1D or 1D-3D GT-SUITE model is prepared. Figure 5 shows the comparison of two isothermal flow simulations conducted in 3D CFD and in GT-SUITE, respectively. When converting the semi-circular distributor volumes of the 3D flow geometry to a 1D flow volume, it is reduced to its main characteristics such as volume, expansion area and flow losses due to directional changes. It is therefore represented as a generic circular shape in Figure 5.

Figure 5. Pressure drop comparison at a mass flow rate of 3.6 g/s between 1D GT solution and 3D CFD solution.

Furthermore, the current battery module was sized to determine the range of pump flow rate needed for a battery electric vehicle (BEV) and it was estimated that 15 such battery modules (3 parallel rows of 5 battery module in series) will be needed for a medium sized vehicle. At a single cooling plate level, it would mean a mass flow rate of 3.6-14.4 g/s with a channel Reynolds number ranging between 175-705. Accordingly, Figure 6 shows the comparison of two different solution for an entire range of pump flow rate, where the non-calibrated GT model predicts the pressure drop with an error of 3-9 % in comparison to 3D CFD results. This is acceptable when considering that

518

Fast running detailed battery thermal management models based on 1D-3D… the 1D-3D approach allows for very fast iterations on the geometry with minimal time spend for model buildup and modifications as well as its capabilities to run in integrated system models of the full cooling circuit of an EV. In the absence of test date this error can be eliminated by scaling the pressure drop with the help of GT recommended friction enhancing multipliers that are known from the extensive calibration studies on similar topics. On the other hand, if the test data or 3D CFD results are available for a set of steady state operating points then it is recommended to calibrate the GT model to benefit from the model predictivity in transient simulations. For all the further analysis, the pressure drop was calibrated to match the 3D CFD results, as also shown in Figure 6.

Figure 6. Pressure drop comparison between 1D GT solution and 3D CFD solution

For the second validation a conjugate heat transfer model of coolant and cooling plate was built both in 3D CFD and compared to the 1D-3D approach. For conjugate heat transfer problem only a single cooling plate and coolant flow was considered. Battery cells were not included in this analysis and a constant heat flux was directly imposed on the front and the back side of the cooling plate. Heat transfer simulations were also conducted for the same range of mass flow rate as for the isothermal simulations. Moreover, no further calibration was done on the GT models. Accordingly, Figure 7 shows the comparison between the two solutions. Based on the conjugate heat transfer validation the cooling plate temperatures difference between the two models was found to be less than 0.25 K, which is also true for the bulk coolant temperature. Hence, the above results validate the capability of the 1D-3D approach to be accuracy at “component level”, allowing for high quality fast running models at “system level”.

519

Fast running detailed battery thermal management models based on 1D-3D…

Figure 7. Comparison of conjugate heat transfer between 1D-3D GT solution and 3D CFD solution at a plate mass flow rate of 3.6 g/s, coolant inlet temperature of 300 K and heat rejection of 10 W imposed on both sides of the cooling plate

4 Results and Discussion 4.1 Steady state module design analyses In this section the results of the steady state design analyses are presented. The results presented in Figure 9 show the implementation of 1D-3D synergetic approach to determine the range of coolant flow rate needed to keep the maximum cell temperature under the safe limit. Unlike in a purely 1D system model, where only lumped cell temperature can be considered, in the current approach it is possible to determine the local peak temperatures, which is very important for module design analyses. Accordingly, Figure 9 shows the outcomes of steady state simulation in terms of module cell temperature averaged over the 3D mesh under different discharge C-rates as a function of coolant flow rate through a single cooling plate. For all the four cases in Figure 9 the coolant inlet temperature and ambient temperature are set to 25 °C.

520

Fast running detailed battery thermal management models based on 1D-3D…

Figure 8. Module-averaged cell temperatures under different discharge C-rates as a function of single plate coolant flow rate

Another typical design analysis is to investigate the effect of thermal contact resistance between the cell and the cooling plate, which again has a 2D spatial effect on the cell temperature. For the results presented in Figure 9 thermal contact conductance of 1e5 W/m2K was assumed between the cell and the cooling plate. Before making this assumption a separate analysis was conducted, where the thermal contact conductance was varied at a constant plate coolant flow rate 1.8 g/s and C-rate of 2C. Correspondingly, it can be seen in Figure 8 that beyond a thermal conductance of 2500 W/m2K the change in cell temperature is very low. Therefore, it is assumed that a thermal contact conductance of 1e5 W/m2K is high enough for efficient battery cooling.

Figure 9. Effect of thermal contact resistance on the cell temperature

521

Fast running detailed battery thermal management models based on 1D-3D…

4.2 Transient thermal behavior of charging-discharging To test the performance of the 1D-3D approach, several standard transient analyses were performed. The first investigation was to charge the battery module through a CCCV (constant current, constant voltage) charging cycle, as shown in Figure 10. During the entire charging cycle each cooling plate was supplied with a coolant flow rate of 1.8 g/s at a coolant inlet temperature of 25 °C. The entire charging cycle lasted for 1.2 hours (simulated time) to reach a SOC level of 99.9 %, whereas for the entire cycle the increase in module temperature was less than 1 K. The key argument here for 1D-3D approach is the simulation time required on a standard PC, which was 1.1 h computational time for a charging cycle of 1.2 h leading to 0.91 as the factor of real time. The dynamic changes in the system thermodynamics during a CCCV charging cycle can be significantly different when compared to some other charge or discharge cycle i.e. different type of simulation can lead to longer or shorter simulation times in order to reach convergence at every time step. Therefore, it is required to investigate different transient processes to test the model performance.

Figure 10. Module-averaged cell temperatures during a typical CCCV (constant current, constant voltage) charging cycle

522

Fast running detailed battery thermal management models based on 1D-3D… Accordingly, the battery module under question was investigated under two different discharging scenarios. In the first scenario the module was completely discharged at different constant C-rates, as shown in Figure 11, where each cooling plate was supplied with a coolant flow rate of 1.8 g/s at a coolant inlet temperature of 25 °C. In Figure 11 the thermodynamic states of the system (temperature, heat rejection etc.) is different for each case. However, from the numerical side all the 5 cases are identical, as observed during the simulation that each case reported a similar and relatively lower computational time as a factor of real time.

Figure 11. Module-averaged cell temperatures over complete discharge cycles with different C-rates at a constant single plate coolant flow rate of 1.8 g/s

In the second scenario the module was completely discharges over multiple WLTP cycles, as shown in Figure 12, where each cooling plate was supplied with a coolant flow rate of 1.44 g/s at a coolant inlet temperature of 25 °C. The discharge current profile of a WLTC cycle on a hot day (30 °C) is taken from a BEV system model published elsewhere [8]. In the BEV system model presented by Mezher et al. [8] 15 identical battery modules (3 parallel rows of 5 battery module in series) are used. Accordingly, to use the discharge current profile in this work the current profile from [8] was reduced by a factor of 3. For the chosen cell chemistry, the range of a BEV equipped with the given battery modules is estimated to be around 200 km under multiple WLTP cycles. In terms of computational time, even in this case the 1D-3D approach showed very promising results.

523

Fast running detailed battery thermal management models based on 1D-3D…

Figure 12. Range estimation of a BEV with multiple WLTP drive cycles

4.3 Thermal runaway One of the important tasks under battery thermal management is to avoid such situation in which a battery failure can lead to thermal runaway. Therefore, the current battery module was used to test the functionality of a 1D-3D approach in the context of thermal runaway simulations, as shown in Figure 13. With two inlets on the front and back side of the battery module, the mass flow distribution through 21 cooling plates forms a U-shape with maximum flow rate through the end cooling plates (plate no. 1 and 21) and minimum flow rate through the middle cooling plate (i.e. plate no. 11). Consequently, cell no. 10 and 11 tend to have higher temperatures compared to neighboring cells. For that reason, it is assumed that cell 10 and 11 would age faster in comparison to other cells. Therefore, cell no. 10 was chosen for thermal runaway analysis. A control logic was defined to initiate thermal runaway once critical temperature is reached, releasing high amount of heat over a short period of time. Based on the available literature [9] it is possible to estimate the amount of heat rejected during such cell explosions. However, such study was not done as part of this work and with prime focus on 1D-3D synergetic approach, the amount of heat released was appropriately chosen to make sure that the exploded cell reaches a maximum skin temperature above 400 °C.

524

Fast running detailed battery thermal management models based on 1D-3D…

Figure 13. Thermal runaway propagation under different cooling scenarios

Accordingly, for all the 5 cases shown in Figure 13 the module was discharged at a C-rate of 3C and the entire system was initialized at 25 °C. In Case-A simulating a cooling pump failure, see Figure 13-A, cell no. 10 took 13 minutes to reach 50 °C (point of explosion). Subsequently, the neighboring cells, for which the auto-ignition temperature was set at 125 °C, also enter a critical state to bring the entire module into thermal runaway in 18 minutes. Changing the auto-ignition temperature to 100 °C accelerated

525

Fast running detailed battery thermal management models based on 1D-3D… the entire thermal runaway process to 13 minutes, see Case-B in Figure 13-B. In comparison to Case-A in Case-C the battery module was continuously cooled at a plate flow rate of 2 g/s. Consequently, no thermal runway took place in this situation for autoignition temperature of 125 °C and beyond, whereas, changing the auto-ignition temperature to 100 °C caused the thermal runaway, see Case-D in Figure 13-D. Hence, the given flow rate of 2 g/s in Case-D could only decelerate the entire thermal runaway from 13 minutes (Case-B) to 19 minutes but could not avoid it. To avoid the thermal runway in Case-D, the plate flow rate was further increased to 4 g/s, see Case-E in Figure 13-E.

5 Conclusion A synergetic 1D-3D solution for battery thermal management is presented in this work to manifest the combined features of fast running 1D system models and highly detailed 3D component models. Accordingly, it was demonstrated through several steady state and transient simulations that a 1D-3D approach can be both accurate and fast running at the same time. Moreover, the functionality of 1D-3D models is also proved for complex heat transfer processes such as thermal runaway. The most important benefit of this approach is the easy integration of battery management system (BMS) with the vehicle system models. Additionally, it facilitates the faster interaction of different flow systems and interdependent components with BMS, which is needed to device control strategies that are quite complex in case of electric vehicles. The 1D-3D approach implemented within the framework of GT-SUITE, which is a product of Gamma Technologies, exhibit a broader scope not only for BMS but also for other key components of xEVs.

Bibliography 1. Gamma Technologies, Modeling a Commercial Li-Ion Cell with Basic Cell Design Inputs Available from Supplier, www.gtisoft.com 2. Gamma Technologies, Select the Right Li-ion Cell Chemistry and Cell Design for any application with AutoLion, www.gtisoft.com 3. Jim Kalupson, Gang Luo, Christian E. Shaffer, AutoLion: A Thermally Coupled Simulation Tool for Automotive Li-Ion Batteries, SAE 2013 World Congress & Exhibition 4. Gang Luo, Christian Shaffer, and Chao-Yang Wang, Electrochemical-thermal coupled modeling for battery pack design, The Electrochemical Society, PRiME 2012

526

Fast running detailed battery thermal management models based on 1D-3D…

5. Thomas Payet-Burin (PSA Groupe), Engine heat distribution model for coolant warm-up and regulation analysis, European GT Conference 2019, www.gtisoft.com 6. G. Fossaert, R. Corbishley, N. Gunter, T. Strauss (Mahle Powertrain), Thermal Management Optimization of Hybrid Vehicles in GT-SUITE, European GT Conference 2019, www.gtisoft.com 7. Remy Fontaine (Opel), Simulating thermal management systems and engine heat distribution, European GT Conference 2016, www.gtisoft.com 8. Haitham Mezher, Dig Vijay, Nils Framke, Peter Stopp, Integrating Electrochemical Battery Models into Component & System Simulation: Focus on Thermal-Management, SIA Simulation Numérique 2019 9. Andrey W. Golubkov, Sebastian Scheikl, Rene Planteu, Gernot Voitic, Helmar Wiltsche, Christoph Stangl, Gisela Fauler, Alexander Thalera and Viktor Hacker, Thermal runaway of commercial 18650 Li-ion batteries with LFP and NCA cathodes – impact of state of charge and overcharge, RSC Adv., 2015, 5, 57171

527

Artificial Intelligence in predictive thermal management for passenger cars Felix Korthals, Dr.-Ing. M. Stöcker Daimler AG Prof. Dr. S. Rinderknecht TU Darmstadt

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_46

529

Artificial Intelligence in predictive thermal management for passenger cars

Abstract Operating artificial neural networks (ANN) and especially the training of ANN requires an available database. Using the cloud connection of vehicles, measurement data from single vehicles and from whole fleets can be collected in big scale. This paper investigates and evaluates the potential of enhancing predictive thermal management (TM) with the help of artificial intelligence (AI). The model-based, predictive control of cooling circuits is combined with ANN. For this purpose parts of the vehicle model, which preprocess predictive data for cooling circuit control, will be replaced by ANN. The vehicle model needs to be calibrated initially during the development process and has to be adapted manually to changes in hard- and software constantly. For this purpose, the utilization of ANN provides a big advantage. The automated calibration in the form of training and retraining the ANN is done in the cloud. Training data used for this purpose are control unit signals. In practice, the ANN is using predictive data as inputs to calculate the states of the vehicle model for the route lying ahead. Using this system, continuous adaptation and refinement of calibration within development process and later for customers can be realized. Overall, the effort of manually calibrating parameters can be reduced and the process can be automated.

1 Introduction Thermal management in passenger cars is guaranteeing optimal engine temperature for all driving states. Within the last years, it has evolved from solely mechanical connected components to electrical devices that are all controlled individually. Components like electric cooling pumps, fans, air panels and valves need to be controlled to adjust coolant temperatures and volume flows. To ensure best performance of the vehicle, the combustion engine, the battery or both, in case of hybrid powertrains, need to be kept within an optimal temperature range. At the same time, all parts have to be protected against overheating. The system holds a big complexity and nowadays is controlled by characteristic curves and maps. This results in a large number of more than 5000 parameters (labels), which have to be calibrated. The calibration process is done manually and requires a high degree of system understanding and experience. In conventional drivetrains, the thermal management system has a significant impact on the performance and working life of an engine. Even more important, energy consumption of the engine and all controlled parts in the cooling circuit is directly influenced by the control strategy. This affects the emissions of CO directly. In order to reduce these emissions

530

Artificial Intelligence in predictive thermal management for passenger cars and enable an energy efficient control system, predictive thermal management has been introduced in [1].

1.1 Predictive thermal management Predictive thermal management prepares the cooling circuit for upcoming events and obstacles on the route lying ahead of the vehicle. Coolant pumps and valves can be activated in advance to cool down the system in front of a hill or other obstacles requiring higher engine performance. This approach is more efficient than just reacting to the driver inputs. It is shown in [1], that a predictive cooling strategy reduces the activation times of the actuators in the cooling system. Thus, the system requires less energy compared to a conventional cooling strategy. Predictive control of thermal management only works with a model-based approach (cf. Figure 1). Physical models are fed with predictions of the route lying ahead. The predictive information like speed limits, slope, curvature and road class is available from the map data of the navigational system. It follows the specifications of the Advanced

Figure 1: Physical Vehicle Model with the predicted vehicle horizon as input, is calculating the performance horizon as output. The output is fed into the subsequent Cooling Circuit Model (CC-Model) to calculate heat inputs and coolant temperatures. In this figure only relevant models are illustrated.

Driver Assistant Systems Interface Specification (ADASIS) and provides predictive route information (ADAS data) for the most probable path [2]. The most probable path is available independently from an activated route guidance. The predictive data are prepared in a Horizon Reconstructor (HRC) and are provided as predicted vehicle horizon (PVH) as described in [1, 3, 4]. The PVH is fed into physical models, describing and vehicle dynamics. The models compute the corresponding engine speed 𝑁 engine torque 𝑀 for any given route (cf. Figure 1). These outputs produce a performance horizon. This is fed into a detailed model of the high temperature cooling

531

Artificial Intelligence in predictive thermal management for passenger cars circuit (CC-Model), representing the thermal behavior of engine, batteries, pumps, valves, pipes and other components of the circuit. The CC-Model computes the expected heat input of the engine into the coolant and subsequently, all coolant temperatures between the components of the cooling circuit. With this information, the necessary actuation of the actuators in the cooling circuit can be calculated in advance.

1.2 Replacing physical models with ANN This work aims to reduce the manual calibration effort for the physical models. Thereand 𝑀 , will be replaced by ANN. The fore, phys. vehicle models, calculating 𝑁 ANN models can be trained and evaluated in an automated workflow (cf. Figure 3). This workflow can serve to automatically calibrate parameters for cooling circuits of various vehicle model series and variants. Instead of manually calibrating hundreds of labels, only the hyperparameters of the workflow (cf. Section 2.3) have to be defined in advance. The phys. vehicle model was chosen to prove the concept of combining ANN and predictive TM. The in- and outputs all are vehicle signals (cf. Figure 2), which easily can be measured on a test car. Thus, these models offer a good opportunity to

Figure 2: Training of the ANN models. Training inputs and targets are composed out of preprocessed measurement data. The output targets are compared to the ANN outputs to assess the performance of the ANN.

validate the concept with real life measurements. It has to be clarified, that the training of the ANN is done with measurement data consisting of vehicle speed (𝑣 ), slope (𝑚 ) and gear state (𝐺 ) as inputs, targets are 𝑁 and 𝑀 (cf. Figure 2). Whereas, for the application of the ANN models combined with the phys. models, the inputs will be ADAS data provided by the PVH (cf. Figure 4). The PVH provides speed limits and slope, whereas the gear state is not available and has to be calculated separately. The calculation of the gear state is not part of this work and will be investigated individually. As the ANN models are entirely based on measurement data, they are considered to be black box models [5]. In an automotive application replacing physical models by

532

Artificial Intelligence in predictive thermal management for passenger cars black box models is not commonly done. This is due to the non-deterministic output domain of a black box model. To avoid any physically or technically illegitimate behavior of the control system, the outputs of the ANN need to be monitored. If the output signals do not comply with a predefined output range, the predictive control will fall back into a non-predictive mode. This procedure is implemented in [1]. It guarantees functionality of thermal control, if there are no ADAS data available for a certain part of the route.

2 Methods To replace physical models with ANN (cf. Figure 4), their required field of application needs to be clarified. ANN exist in various architectures, each specialized for a certain task. After a matching architecture is found, the internal ANN setup (layers and neurons) needs to be adjusted to the defined task. This is done by an automated workflow (cf. Figure 3), which is collecting and preprocessing the training data, training different

Figure 3: Automated Workflow. Collecting measurement data from vehicle, preprocessing, training of different ANN variants & selection of best performing ANN.

models and ranking them to select the best performing ANN. A selection criterion is defined in Section 2.2, which helps to check all appropriate models for the best setup. Figure 3 shows how the five relevant raw measurement signals (cf. Section 1.2) are collected and preprocessed, to serve as training inputs and targets for the training of and different ANN models. To create two ANN, which serve best in calculating 𝑁 𝑀 , the workflow iterates through the different combinations of hyperparameters (cf. Section 2.3), creates corresponding ANN, trains them and evaluates their performance. An ANN calculating both signals at once is not part of this work. The described workflow could collect and process measurement data in a cloud environment, to train and adapt models to hard- and software changes of a vehicle or a whole fleet. Combining the trained ANN model with the given physical vehicle models is visualized in Figure 4. Once the ANN are integrated in the existing process described in Section 1.1, they have to process the inputs from the PVH instead of measurement data. These

533

Artificial Intelligence in predictive thermal management for passenger cars inputs are given in coarse grained intervals. Whereas, the training of the ANN aims to generate good performing ANN with fine grained inputs (measurement data).

Figure 4: The phys. vehicle model is combined with ANN. Models for engine speed and engine torque prediction were replaced by ANN.

2.1 Artificial neural networks for time series regression ANN can serve a variety of different tasks. From binary classification over image processing to sequence prediction there are various common architectures. Since the inand outputs of the phys. vehicle model are sequences of vehicle signals (time series), the ANN needs to be able to process this kind of data. Thus, a regression model for time series prediction needs to be trained. Recurrent neural networks (RNN) are specialized in this task and can process multiple sequences as in- and outputs [6]. The most basic RNN architecture is called Vanillan RNN [7]. The Vanillan RNN is built out of an input layer followed by one recurrent hidden layer and an output layer (cf. Figure 5). Each layer consists of a certain number of neurons. The construction and function is roughly based on neurons in a human brain. They are activated, if an activation function reaches a corresponding threshold. Artificial neurons hold a mathematical activation function, which generates an output signal as input for the subsequent layer. Like visualized in Figure 5 the inputs (𝑥) of the neurons are modified by weight matrices (𝑤 ) and bias vectors (𝑏 ), varied during training to create an ANN with optimal outputs [6]. The subscript 𝑛 indicates the input (i), hidden (h) and output (o) layer. The recurrent hidden layer enables the ANN to learn time dependencies. This is done by feeding back the

534

Artificial Intelligence in predictive thermal management for passenger cars network state as input for the following time step with 𝑧 being the lag operator [7]. The polygons in Figure 5 represents the activation functions.

Figure 5: Simple Vanillan RNN. Hidden Layer with recurrent feed back.

Simple RNN architectures struggle to learn long-term dependencies due to the vanishing gradient problem as described in [7]. The vanishing gradient describes a phenomenon where the weights and biases are not updated anymore, while training. Hence, no improvement of the ANN output can be reached. To overcome this problem long shortterm memory (LSTM) cells are one of the most successful and broadly used options [6]. Introduced in 1997 by Hochreiter and Schmidhuber [8], LSTM cells offer a structure which takes long- and short-term dependencies into account. Characteristic features of LSTM cells are the cell state and different gate layers. The cell state propagates information through time, gate layers process the cell input and the cell state to define the cell output and the new cell state. The cell state is updated every time step. Different layers (gates) process the input signal and the cell state. The forget gate defines, which information of the cell state can be forgotten and which part should be kept depending on the actual cell state and input value. The new input gate defines what information of the input signal needs to be combined with the actual cell state. The output gate defines, which part of the cell state should be the output of the LSTM cell for the given input value at this time step. Further explanations can be found in [7, 8]. RNN constructed with LSTM cells are commonly called LSTM networks. In this case, the hidden layers hold LSTM cells. Due to the clear advantage of LSTM networks mentioned in [8], the ANN in this work are LSTM networks.

2.2 Selection criterion for choosing the ANN architecture To find an ANN model, that replaces the corresponding physical model best, many different LSTM network setups are investigated in an automated workflow. For this, a selection criterion is defined. Since the task of the ANN models is to compute the performance horizon out of the given PVH, the error of the calculated values, compared to the training targets (measurement signals), is an important criterion. A small error between these two time series indicates an ANN, which has adapted the qualities of the physical model. As selection criterion the Root Mean Squared Error (RMSE) is used. It

535

Artificial Intelligence in predictive thermal management for passenger cars is defined as square root of the quadratic sum of the residuals, which is minimized through regression [9]. Small RMSE values indicate a good accordance of the calculated signal and target signal [7]. RMSE =

∑

𝑦 −𝑦

,

(1)

The target signal is defined as 𝑦 , calculated values as 𝑦𝑖 . The number of time steps of the measurement is 𝑛.

2.3 Hyperparameters for automated network design The automated workflow for the investigation of different RNN architectures varies different hyperparameters to create and test ANN models. To find the best working LSTM network structure, the number of hidden recurrent layers is varied in a range between one and eight layers. Furthermore, the hyperparameters learning rate (𝐿𝑅), optimizer function and window size of the input sequence are varied. In this work, classic gradient descent and the Adam optimizer [10] are tested as optimizer function. The hyperparameter 𝐿𝑅 defines how strong the weights and biases are updated during training [11]. 𝐿𝑅 is varied between 0.005 and 0.05 with discrete steps of size 0.005. The window size determines, how many time steps are given in the input sequence to learn from during training and is varied as 𝑥 in the interval 𝑥 ∈ ℝ: 5 ≤ 𝑥 ≤ 250 . ANN are trained in epochs. In one epoch all available training data is shown to the ANN once and weights and biases are adapted. In this work 250, 500 and 1000 training epochs were compared. Activation functions used are the ReLU (Rectified Linear Unit) function for the input neurons and the sigmoid function for the output neurons, to scale the output in a range between zero and one. Training is done by minimizing a loss function with the help of an optimizer. This function depends on the error between the calculated and the target output signal. The loss function adopted in this work is the Mean Squared Error (MSE) [7] and is defined as MSE =

∑

(𝑦 − 𝑦 ) ,

(2)

where the target signal is defined as 𝑦 , calculated values as 𝑦𝑖 . The number of time steps of the measurement is 𝑛. Optimization is done by adjusting the weights and biases using the standard learning algorithm called back propagation trough time [7]. In back propagation calculated errors of the ANN output layer are propagated back layer by layer, while adjusting the weights in order to minimize the error [12].

536

Artificial Intelligence in predictive thermal management for passenger cars

2.4 Data selection and preprocessing To obtain good model quality, preprocessing the measurement data is obligatory. For this process, a preprocessing workflow is set up. At first, all measurements with missing values are discarded. To enable the ANN to learn short- und long-term dependencies in time series, it is crucial to provide sequences with linked information (time steps). In a next step, sequences with stationary driving states are shortened to 5 s length. Longer stationary sequences do not hold additional information, concerning the signals used for training. Stationary driving states are defined as sequences of constant velocity and slope, that is constant engine torque and subsequently constant heat induction. Sequences of dynamic driving states are kept without truncation. This is done to focus on dynamic driving states while training. Therefore, it is necessary to make sure the training data set is not dominated by stationary sequences. To reduce the noise of the training data a Savitzky-Golay filter [13] is used. This low pass filter is specially designed for smoothing noisy data. As defined in Equation (3), filtering is done by fitting each data value 𝑓 (time step), of a measurement signal, to a polynomial function over a window between 𝑛 and 𝑛 centered at 𝑓 . The filtered data value is represented by 𝑔 . The corresponding filter coefficients 𝑐 can be defined by Eq. (4) with the help of design matrix 𝐴. Fitting is done with the least-squares method. 𝑔 =

𝑐 = 𝐴 =𝑖 ;

𝑐 𝑓

(𝐴 ∙ 𝐴)

(3)

𝑛

𝑖 = −𝑛 , … , 𝑛 ; 𝑗 = 0, … , 𝑀

(4) (5)

The very noisy slope signal is filtered with a polynomial function of order 𝑀 = 4 and a window including 401 time steps (𝑛 = 𝑛 = 200). With this setup, heavy outliers are smoothed notably. Filtering slope is crucial because the position of the sensor in the car creates a very noisy signal. Whereas, engine speed and engine torque are provided as filtered signals for the control units of the vehicle. During preprocessing all training data are normalized, which scales the data in a range between zero and one. In this way, all signals are equally weighted for training. This is necessary to prevent the training from being dominated by signals with higher absolute values. For all signals, the autocorrelation function (ACF) is computed. The ACF compares the signal with itself shifted in time [5]. The more the signal is shifted the less the two

537

Artificial Intelligence in predictive thermal management for passenger cars sequences correlate, if no seasonality does exist. In general, it can be visualized how much a time step in a sequence depends on previous ones. Especially for the training target signals, the ACF helps to estimate the dependencies on earlier time steps [5]. See Section 3.3 for application of the ACF. After preprocessing is done the whole data set is divided into a training, validation and test set. Only the training set is used to optimize weights and biases. The ANN is trained for several epochs. After a certain amount of epochs, the ANN is tested on the validation set to make sure no overfitting to the training data is occurring. At the end of training the test data set is used to check the ANN performance on unseen data sets. In this work, the ANN models are trained with a data set from Mercedes Benz W 213 model series with a four cylinder gasoline engine, 48V on-board system and electric cooling pump. The data was collected during test runs. The lengths of the measurements are between 1500 s and 4500 s.

3 Results In this section, the ANN models generated by the automated workflow (cf. Figure 3) and their calculated outputs are presented and discussed. Best performing ANN were found by the lowest RMSE values. The best performing ANN for engine speed calculation is equipped with two hidden LSTM layers and uses the Adam Optimizer. Training lasted for 250 epochs with a fixed learning rate, 𝐿𝑅 = 0.02. The window size of the input sequence was 100 time steps. For engine torque the best model setup is also defined by two LSTM layers and the Adam optimizer. The window size is 50 time steps, 𝐿𝑅 = 0.05 with 250 epochs of training. Thus, both ANN have the Adam Optimizer and a two LSTM layer setup in common. In the next paragraphs, the calculations of engine speed and engine torque by the two ANN models are discussed. For better visualization of the ANN model outputs, characteristic sequences from two test measurements are plotted.

3.1 Engine speed ANN model performance The best performing ANN for engine speed calculation was fed with two test measurements to evaluate its performance for processing unknown data. The test set produced an average RMSE = 0.03. In Figure 6 two characteristic sequences from two different test measurements are shown. Each column shows one sequence. Vehicle speed and slope as ANN inputs are plotted in the first row. The target and the calculated signal are visualized in the second row of Figure 6.

538

Artificial Intelligence in predictive thermal management for passenger cars The left sequence displays a higher vehicle speed and varying slope, whereas in the right sequence, vehicle speed is lower but more dynamically varying and slope is changing in a smaller range. Thus, sequences are referred to as “high speed” and “dynamic” sequence. The ANN produces an overall RMSE = 0.02 for the whole measurement of the “high speed” sequence. This results in a good curve fitting for the visualized sequence. The measurement corresponding to the “dynamic sequence” is reproduced by the ANN with an overall RMSE = 0.04. It is shown, that the trained ANN is able to calculate engine speed for different driving states and performance levels.

Figure 6: ANN output for engine speed calculation. Sequences of two test measurements. The first row shows vehicle speed and slope for the corresponding output signal in the second row. Left measurement with overall RMSE = 0.02. Right measurement with overall RMSE = 0.04.

The good performance of the engine speed (𝑁 ) calculation can be explained by considering the formula used for the physical model. In [1] this formula is used as follows. 𝑁

=

𝑣

∙ 30 ∙ 𝑖 ∙ 𝑖 3,6 ∙ 𝜋 ∙ 𝑟

(6)

The computation of 𝑁 with the information of the PVH is mostly depending on the predicted vehicle speed (𝑣 ), on the gear ratio (𝑖 ) and the axle transmission ratio (𝑖 ). The actual gear state is defining 𝑖 , whereas 𝑖 is a constant, vehicle specific value. a constant factor, concerning the vehicle wheels and tires, is used among With 𝑟

539

Artificial Intelligence in predictive thermal management for passenger cars with the other constant values. By providing vehicle speed and gear state as input sigand 𝑖 , nals for training, the ANN is provided with the information concerning 𝑣 which explains the good RMSE value.

3.2 Engine torque ANN model performance For the calculation of engine torque a second model was trained. Testing this model with the test measurements produces an average RMSE = 0.105. In Figure 7, the same two characteristic sequences as described in Section 3.1 are shown. The target and the

Figure 7: ANN output for engine torque calculation. Sequences of two test measurements. The first row shows vehicle speed and slope for the corresponding output signal in the second row. Left measurement with overall RMSE = 0.11. Right measurement with overall RMSE = 0.10.

calculated signal for engine torque is visualized in the second row of Figure 7. The measurement containing the “high speed” sequence reaches an overall RMSE = 0.11, whereas the other test measurement reaches an overall RMSE = 0.10. The ANN outputs are following the trend of the target signal as shown in Figure 7, but differ in some areas between 5 % and 20 % from the normalized target signal. Taking a closer look at the physical model and the formula computing engine torque (𝑀 ) in [1] gives first approaches for explaining the ANN models performance. The used formula in [1] is defined as follows.

540

Artificial Intelligence in predictive thermal management for passenger cars 𝑀

=

𝑟 ∙𝑅 𝑖 ∙𝑖 ∙𝜂

(7)

In Eq. (7) driving resistances (𝑅 ) and constant powertrain efficiency (𝜂 ) are taken into account together with 𝑖 , 𝑖 and 𝑟 . As described in Section 3.1, 𝑖 depends on the actual gear state, which is an input signal of the ANN. The driving resistances are composed of different, vehicle specific resistance components. Considered were drag, climbing resistance, friction and inertia. The inputs of the ANN contribute to this in indirect form with vehicle speed and slope. In further work, it will be examined, if the lack of information about driving resistances, could be compensated by adding additional inputs. Figure 7 also shows constant discrepancies between calculated and target signal, if the normed engine torque reaches a plateau with minimum values, which equals a negative torque in the raw measurement data. It needs to be examined, whether the ANN performance is improving, if sequences of negative torque signals are rejected from the training data.

3.3 Discussion The two ANN models created to calculate 𝑁 and 𝑀 , show many similarities concerning their architecture. As stated in [7], advantages of the Adam Optimizer were confirmed. For both signals a model architecture with relatively little amount of LSTM layers was sufficient. The option of creating deeper RNN was not improving the ANN performance. This indicates the assumption, that the relatively small number of in- and outputs can be modeled with a two layered LSTM sufficiently. For a next step, the input data and loss functions seem to hold more effective options to further improve performance of the ANN (cf. Section 3.1 and 3.2). To evaluate the models differing performances, the ACF is computed for the two predicted signals (cf. Figure 8). Considering the ACF clearly shows, that engine torque is much harder to predict as engine speed. This is indicated by the steep declining curve

Figure 8: Comparing autocorrelation functions of engine speed and engine torque for a measurement in the training data set.

541

Artificial Intelligence in predictive thermal management for passenger cars progression of the ACF within the first lags. Compared to 𝑁 , 𝑀 is much less depended on former time steps. This also provides the explanation for the much smaller input window size for the engine torque model (cf. Section 3). Looking at the automated workflow it can be stated, that a significant reduction of the calibration work has been achieved. Compared to the conventional process, where manual calibration of hundreds of labels for the physical vehicle model had to be done, the workflow only needs to be provided with the hyperparameter ranges described in Section 2.3. The created workflow could easily be implemented on a cloud environment, where vehicle measurement data is collected and preprocessed. The training in a cloud environment offers scalable computational power for much larger amounts of data and more complex ANN.

4 Conclusion The results demonstrate working ANN models, which can replace physical models in the original setup of the model-based control (cf. Figure 4). The utilization of ANN models reduces the extend of manual parameter calibration to a minimum. For that, an automated workflow was set up, which preprocesses training data, iteratively trains multiple ANN with different architectures and hyperparameters and finally selects the best ANN model for the investigated task by a selection criterion. The models were trained to predict the performance horizon as precise as possible. For engine speed, the results were already satisfying, whereas the engine torque models predictions need further improvement. In future work, additional input signals will be given to add information about the driving resistances. It will be also examined, in which way the gearbox has to be considered. If a simple model is not sufficient, a further ANN could serve this task. Furthermore, the optimizer function and the loss function will be investigated closely to improve the ANN outputs. In a subsequent step, the composition of the data space will be investigated. The collected vehicle measurement data need to be joined into a data set for robust training. Therefore, the amount of the different driving states, which need to be represented in the training data set, has to be defined. This is to make sure automatically collected data can be composed to a data space, which serves all training goals. In a next step, the automated workflow will be transferred to the CC-Model to enable this complex system to profit from the same advantages as the vehicle model.

542

Artificial Intelligence in predictive thermal management for passenger cars

References [1]

M. Stöcker, Modellbasiertes Thermomanagement mit Navigationsdaten zur Energieeffizienzsteigerung der Fahrzeugkühlung. Dissertation - TU Darmstadt, 1st ed. Aachen: Shaker, 2018.

[2]

C. Ress, D. Balzer, A. Bracht, S. Durekovic, and J. Löwenaus, ADASIS Protocol for Advanced In-Vehicle Applications: ADASIS Forum. [Online] Available: http://durekovic.com/publications/default.html. Accessed on: Jan. 06 2020.

[3]

M. Steinhart, Entwicklung eines ADAS-basierten Kreuzungsidentifikationsmodells für das vorausschauende Thermomanagement in Personenkraftfahrzeugen: Master Thesis, 2014.

[4]

K. Weiler, Entwicklung eines Fahrer- Fahrzeugmodells zur Vorhersage der Motorbetriebspunkte und der Fahrzeuggeschwindigkeit aus ADAS-Streckendaten: Master Thesis, 2014.

[5]

O. Nelles, Nonlinear System Identification: From Classical Approaches to Neural Networks and Fuzzy Models: Springer, 2001.

[6]

I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning: MIT Press, 2016.

[7]

F. M. Bianchi, E. Maiorino, M. C. Kampffmeyer, A. Rizzi, and R. Jenssen, “An overview and comparative analysis of Recurrent Neural Networks for Short Term Load Forecasting,” CoRR, 2017.

[8]

S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Computation, vol. 9, 1997.

[9]

F. Kirschbaum, “Modellbasierte Applikationsverfahren: Vorlesungsskript,” 2017.

[10] D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” Dec. 2014. [Online] Available: http://arxiv.org/pdf/1412.6980v9. [11] Y. Bengio, “Practical recommendations for gradient-based training of deep architectures,” Jun. 2012. [Online] Available: http://arxiv.org/pdf/1206.5533v2. [12] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning Internal Representations by Error Propagation,” 1985. [13] W. H. Press, Numerical recipes in C: The art of scientific computing, 2nd ed. Cambridge, New York: Cambridge University Press, 1992.

543

Continuous integration in powertrain software – Today and tomorrow Daniel Heß, Daniel Volquard, Ronald Siedel, Fabian Feyerherd IAV GmbH, Berlin

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_47

545

Continuous integration in powertrain software – Today and tomorrow

1 Powertrain Software Development 1.1 Frontloading at the feature development Today's powertrain ECU development uses Agile SW development, continuous integration, testing and deployment. Typically, the aim is to achieve high maturity of the function modules in the integrated system as quickly as possible. Further it is necessary to have high reactivity to changes in requirements (innovations, market changes, etc.). Software functions in new, more modular software architectures are developed in shorter cycles, system requirements are broken down to encapsulable requirements and developed modularly as individual functions. Ideally, therefore the function modules are commissioned, pre-dated and validated in the integrated system by the technical experts in rapid prototyping loops in beforehand. For this purpose, it must be possible to create individual, integrated stands for the system as quickly as possible. Each functional module thus achieves the highest possible level of maturity as early as possible, and has already been validated in the respective current integrated environment for robustness. People responsible for a function module want to order a SW integration themselves and have it available as soon as possible for the validation of their development context (download from a SW repository) The software integration process must be automated and secured (self-testing, logging) in such a way that no detailed integration knowledge is required for the build (Fire and Forget).

1.2 Challenges of Industrialization As opposed to the above feature driven, agile approach with a high frequency of changes, later on in the calibration phase of the project the main goals of a software are high reliability and stability. Throughout the homologation process the software logic must not change anymore. The release pattern during this phase of the project is typically not as-soon-as-possible anymore, but based on fixed milestones and previously agreed functional content. A typical reason for this lies in the fact, that vehicle tests cannot be fully automated, especially when drivability relevant functionalities are changed (by calibration or software logic) a manual test has to be performed to see whether the drivability can be accepted by the customer.

546

Continuous integration in powertrain software – Today and tomorrow Some tests are very cost intensive as they expensive resources as test bench, roller and climate chamber and are bound to certain seasons (summer- / winter trip). In contrast to pure software tests as in the development stage, these cannot be repeated as often as desired in the integrated system. Homologation, approval and legislation are very formal processes that require a certain time period during which the change of the system should be minimized to avoid the risk of having to start all over again.

1.3 Continuous Integration Due to the relatively rigid application process explained above, the procedure for integrating new software modules was in the past regularly determined by a milestonedetermined development process. Functional requirements were first collected and jointly developed. The developed code was then handed over to the calibrators at various handover times. However, the functionality could then already be outdated and overtaken by new findings. With continuous integration, changes can be implemented all the time, which allows for simplified troubleshooting and faster testing, especially in complex systems, where the interaction between the functions regularly outweighs the influence of the individual function. In general, each software change can be sooner and more frequently on the test bench. As continuous integration seems to be the perfect match for the challenges of rapid feature development, it creates challenges for calibration, validation and industrialization.

2 Current CI Processes and toolchain

Overview of the continuous integration (CI) process and toolchain

547

Continuous integration in powertrain software – Today and tomorrow

2.1 Verification of sources To avoid time and resource consuming errors during the integration process it makes sense to find or estimate errors as early as possible. Therefore, we do the incoming inspection of the source code with the following inspection mechanisms: – File-based checks: – Check whether all files required for integration are available – Check if the versions of the files are the ones that were planned for this software – Check if there are duplicate files in complex file structures (same file in different locations? With different content?) – Function-based checks: – If all functions are defined – Are all defined functions used – Signal based checks – If all signals are defined – Are all defined signals declared – Multiple declared The check set is continuously extended to further increase the probability of a successful build. Additionally, checks differ in their results with respect to pure software quality or integration abort. If all checks are positive, the probability of a positive integration result is very high!

2.2 Triggering the integration The start of a software integration can be triggered in different ways. This depends on whether an integration is requested due to functional changes in the software or whether the integration only serves to validate a change in the tool chain. – Software development builds – Manual triggered (e.g. test integration) – Time triggered (e.g. daily, weekly etc.) – Check-in triggered (e.g. new functionality implemented) – Toolchain development builds – Same trigger types can be used to validate toolchain changes – Source code which was already successfully build is rebuild and result is compared to validate toolchain after software change

548

Continuous integration in powertrain software – Today and tomorrow

2.3 Validation of integration result After a successful build, the resulting binary is automatically checked. – Automated checks – All code parts are in dedicated memory sections – All signals defined in configuration management system are integrated and accessible – Error level of all log files arose during the process – Notification – Final integration report includes the summarized integration results as a delivery – Formal review process is defined and established – An automated review procedure is currently under development

2.4 Possible degree of automation In short, evaluation and implementation of functional requirements can never be fully automated or only to a very limited extent. This applies for the implementation of new functionalities into the control software as well as for the tool chain. Also occurring errors that lead to nonfunctional builds have to be tracked regularly by software developers. Therefore it is important to have a very high degree auf automation for the remaining part within the system boundaries of integration itself. In principle, there is of course still potential in the input analysis or evaluation of the integration results and the integration with regard to virtualization, resource usage and the stability of the tool chain itself. This potential is being exploited step by step as part of the further development of the tool chain. To complete the entire continuous process, also automation of testing, delivery, deployment is an important goal. Due to current limitations such as the lack of remote access to flash interfaces, but also for safety reasons, a number of manual activities are still necessary today in order to actually test the integrated software with a meaningful basic data in the system, a vehicle. Some of these limitations are described in more detail below.

549

Continuous integration in powertrain software – Today and tomorrow

3 Difficulties in consumption of frequent software builds 3.1 Complex Powertrain

Complex Powertrain

In a vehicle, the powertrain comprises the main components that generate power and deliver that power to the road. This includes the engine, transmission, drive shafts, differentials and four or more control units with a special software: 1.

Engine control unit

2.

Transmission control unit

3.

Central gateway unit

4. High voltage control unit

550

Continuous integration in powertrain software – Today and tomorrow These controllers are connected with different kinds of bus architectures (CAN, Flexray, LIN…). Each control unit needs a verified dataset for current project phase (e.g. test bench, prototype vehicles, road approval). Each controller has to be flashed individually by cable and could have different supplier, architecture or need an own flash tool. A Powertrain software system reflects this complex structure and needs huge effort in testing new or change parts of the software or the calibration of it.

3.2 Tests at System Level In order to be able to test at system level, a suitable data set is required for integration into the software. The algorithmic structures are therefore calibrated specifically with data parameters. Typically for a powertrain ECU there are about 50,000 parameters (characteristics, maps). Initial parameter values guarantee code security – not logical function, new functions are initially given "neutral" data, which naturally makes their testing more difficult. With a high expenditure of time, all parameters must be "populated" with a multitude of measurement and analysis tasks according to the current test requirements. These vary according to powertrain, emissions, project phase. Especially in early project phases these calibration data are not yet mature enough to be relevant for a meaningful test. Compared to software integration, the integration of calibration data into the software means a considerable effort. Furthermore, the test system (test bench, HiL system or prototype vehicle) has to be physically available and has to be set up and maintained. The system complexity leads to "days" instead of "hours" of validation, e.g. 10,000 signals to be tested are to be checked for consistency and correct function between Engine Controller and Gateway for Flexray. System test results must be checked manually to ensure that the results are valid. A calibration engineer who has to deliver a data set cannot use a daily software delivery - which therefore cannot be tested and checked quickly.

4 Short term patches and Workarounds 4.1 Focus of change With need of dataset and system test in current process, the problem can be reduced if software changes are classified in different change types, classes or intensions. For example: – When change has no impact on module interfaces, no need of testing the whole system.

551

Continuous integration in powertrain software – Today and tomorrow – When no combustion or emission structure affected, no need to deal with certification for road release. – When torque path is not affected, no need to deal with regulations of functional safety. – When evaluation purpose is intended, reduced quality and safety precautions can be allowed. If 100% test case specification coverage for change is integral requirement to accept change, safety of code can be approved by automated system test after continuous integration.

4.2 Dataset Data parameters of unchanged modules or maintained structures can be overtaken for testing. With overtaken data parameters test can run right after integration and test coverage is granted. IAV recommends complete specification of new calibration parameters as integral requirement to accept the change in the algorithmic structure. This calibration should be testable and not just “switch off” the new functionality. IAV uses automated process to overtake parameters along with possibility to integrate parameter specification from change management system. This reduces the need of manual work by a calibration Engineer for previously known functions.

4.3 Continuous Testing For each software change, specified test types are defined depending on the change class / type / intensity and the corresponding measurements are performed automatically. Similarities of characteristic system behavior during the system test (e.g. on the basis of typical traces of emission cycle test, full load test, communication tests, ...) are automatically compared with reference results via cross correlation. If the similarity of the system behavior exceeds a certain threshold value, the software can be automatically forwarded to the function requester for data processing and evaluation of the test.

552

Continuous integration in powertrain software – Today and tomorrow

4.4 Continuous Evaluation Usually vast amount of measurement data is created on multiple test systems (SIL to production vehicle) during development. There are various tools that can be used to check this data and use it for the further development. Data science based methods can also be used to evaluate results of automated system tests furthermore, e.g. IAV Mara is designed to incorporate various methods to acquire, store, filter and recognize abnormalities in measurement streams.

4.5 Summary: Optimized Process

Optimized process

1. Developer can request build via configuration management system after supplying new software specification 2. After module has implemented, continuous build is triggered, software is integrated 3. Dataset is prepared from previous release and specified values if supplied 4. Software and dataset is loaded to SIL or ECU on HiL system with proper configuration

553

Continuous integration in powertrain software – Today and tomorrow

5 Long term solution scenario (2030+) 5.1 Vision Recent developments in the automotive industry suggest that the complexity of the powertrain will decrease significantly. Today's plug-in hybrid with combustion engine, its exhaust gas treatment, (automatic) manual transmission as well as electric motor, battery and power electronics is historically the most complex drive system. To operate an electric motor, less than 50% of the ECU resources are required, which are consumed by today's exhaust gas after-treatment and its diagnostics (25,000 parameters, 2MB Flash, 50M Lines of Code). An electric wheel hub motor does not require a gearbox at all and therefore no transmission control. Furthermore, the electric motor is much easier to control or model for testing due to its simpler function and almost no interaction with external operating conditions compared to the combustion engine. The power source (battery, hydrogen cell, flux-capacitor) is also easier to regulate and has only two major points of interaction with the drive: voltage and current.

5.2 Software of the future In the upcoming years, Software will become one of the strategic development areas for the automotive sector. This is because the continuous expansion of vehicle functionality – becoming more connected, managing the (electric, hydrogen cell) powertrain, AD – is enabled by increasingly powerful Software just as much as by Hardware. This trend need a new “lightweight-system-structure”. This system structure can be handled with different and more recent software solutions. Leaner and exactly predefined interfaces open options to use modern software concepts as micro services or software as service. Which can form their own business models (With less or no impact on complete system no need of full system test when changing a single module - appbased software concept). System parts, which can be fully represented by formula or described suitable by numerical approaches, will not need excess number of calibration parameters, which simplifies the integration of these parts.

5.3 Architecture of the future The overall trend is a transition from a decentralized architecture (components connected by a central gateway in the E/E architectures), in which functions are running on dedicated ECUs with high SW-to-HW integration, towards more centralized systems with dedicated domain controllers. Finally, the architecture is expected to evolve into

554

Continuous integration in powertrain software – Today and tomorrow virtual domains, in which one control unit runs functions or (micro-)services of different domains (e.g., infotainment and so on). The centralization will go along with a separation of HW and SW, leading to vehicle systems be built as a layered architecture with clear abstraction points at operating system (OS) and middleware layers. When using central intermediate code storage, changing modules in vehicle becomes as easy as loading an app from store. While this evolution will occur across domains over time, infotainment and driving assistance are expected to be the forerunners, as areas of high performance and low safety or latency criticality are easier and more beneficial to transform. With smarter and/or smaller system interfaces, software can be designed to run on any capable controller in the system (distributed computing) and only when computing results are needed (software as service). Car2X Systems can easily be implemented when telemetric data can be collected by nearby vehicles using a remote service.

6 Conclusion and outlook The separate handling of calibration data and the calibration process hinders a continuous process for the creation of a complete powertrain software. The current monolithic system architecture is unfavorable for continuous processes. The need for complex test setups for verification and validation is an obstacle to continuous development. In the lecture, aids and workarounds could be shown to deal with these difficulties today. Future software architectures and development approaches are expected to fully support continuous software integration. Thereby the fully electric drive is much less complex in terms of modeling and simulation. This will make it possible in future to provide and use continuous processes for the entire development cycle.

555

Managing software evolution in embedded automotive systems Lukas Block University of Stuttgart, Graduate School of Excellence advanced Manufacturing Engineering / Institute of Human Factors and Technology Management IAT

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium,, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_48

557

Managing software evolution in embedded automotive systems

1 Introduction Customer requirements evolve frequently in the automotive domain. A the same time, rapid integration of innovative solutions is important to maintain and extend the customer base (Vitale et al. 2020). Thus, automotive software components are subject to constant change and variations, even when the vehicle is already on the road (Guissouma et al. 2018). Agile development methods support these alternations on an organizational level. However, they also fuel the frequency of these changes. Software updates in general ease the change of functionality, while over the air software (OTA) updates allow to alter vehicles already on the road (Khurram et al. 2016). Consequently, new software packages are released in short, irregular timeframes, instead of model years, and they are directly available to the customer. The previously long life cycle of the entire software system is divided into many smaller, preferably decoupled life cycles of software functions and hardware components (Block et al. 2019). Instead of reconciled software releases, the software grows piecemeal: Smaller junks of software with partial system updates prevail. The software evolves. Accordingly, additional challenges in software architecture arise: Partial software releases must maintain the compatibility and thus their interfaces to the remaining system. In distributed environments, the dispersal of software makes it difficult to control the change and its propagation to whoever is implementing an evolving interface (Senivongse 1999). Domino effects might occur, with changes throughout the whole system and a loss of development speed. Thus, alternations should be internal and a change of the interfaces must be prevented or the effect must be limited. However, frequent compatibility maintenance of partial updates might lead to suboptimal design decisions (Weyns et al. 2015). The architecture erodes. Ultimately, this accumulates a vast amount of architectural technical debt in the long run (Martini and Bosch 2016). Consequently, evolution strategies to maintain interface compatibility and prevent technical debt are necessary, to support continuous and frequent software releases. Service oriented vehicle architectures like the AUTOSAR Adaptive Platform decouple software components through well-designed service interfaces and ease the maintenance of compatibility. However, they fail to provide solutions for proactive change management and encapsulation. Thus, the research question is as follows: How should the evolution of software components in embedded automotive systems be managed, to maintain compatibility and mitigate the risk of architecture erosion? This publication transfers software design principles from the web service domain to the challenge at hand. A set of patterns for web service evolution is introduced and their system of objectives is adjusted to suit the automotive domain. As a result, a set of guidelines for automotive software evolution emerges, that focuses on dependability requirements and compatibility between modified parties.

558

Managing software evolution in embedded automotive systems

2 Approach The following approach focuses on so-called “breaking changes” in software evolution: A software component evolves in a way, which breaks the contract previously established with its communication partner (Xavier et al. 2017). Such a contract is usually defined through the software’s interface. Most publications (e.g. Xavier et al. (2017)) however limit this interface to its syntax definition in code. Within this paper, we follow the idea of Design by Contract (Meyer 1992): An interface is a multitude of syntactical and semantical agreements. onnelou et al. (2008; 2011) for example cluster problematic changes into structural changes (syntax), behavioral changes and policy-induced changes. Applying the Liskov Substitution Principle (Liskov 1998), compatibility and breaking as well as non-breaking changes can be identified through the relation of the interface definitions in these three classes. Consequently, an interface change in the context of this work is everything that requires a communication partner, either a provided or required port, to update the respective interface it implements. Changes that only affect “inner” code of a software component are not considered within this paper. Managing breaking changes in distributed systems is an idea that has existed in the web service domain for more than 20 years (see e.g. Senivongse 1999). There, service evolution “is the disciplined approach for managing service changes” (Papazoglou et al. 2011). Services are distributed, loosely coupled software components (Wang et al. 2014; Fehling 2015): They try to make little assumptions, when communicating with another party. The entities are independent in terms of the platform they run on, their physical and virtual address, their timing and their format (Fehling 2015, pp. 65–66). As such, service-oriented software architectures already provide certain flexibilities, when it comes to altering an interface provisioned at a certain port. However, diverse change requirements can lead to situations, in which alternations of an interface also requires adaptation of the corresponding communication partners. Existing research on service evolution encapsulates and thus enables change via more general evolution strategies (see e.g. Wang et al. 2014) and specific design patterns (see e.g. Kaminski et al. 2006). Evolution strategies manage the transition towards new versions of a service. Design patterns are implementation templates to maintain the compatibility with ports, implementing the previous interface version. The approach followed here describes embedded automotive software from the viewpoint of long-lasting software systems, which underlay evolution. This temporal perspective on interface design and versioning allows transferring the findings from web services to the automotive domain. As both are distributed systems, a transfer is reasonable. Furthermore, service-oriented software architectures for automotive systems

559

Managing software evolution in embedded automotive systems have gained increasing attention in recent years (see e.g. Maul et al. (2018) and AUTOSAR Document 706). However, there are points that distinguish web services from embedded automotive services: – Criticality and incompatibility: Incompatibility in web services might only lead to the loss of business value. Failure to communicate in the automotive domain can be safety critical. As such, the strategies and design patterns, which are presented in the remainder of this work, must prevent incompatibility of the provided and required ports. – Controllability of communication partners: In the web service domain, the communicating parties are almost always not under control of the entity, triggering the change (i.e. the evolver (Senivongse 1999)). This is not necessarily true for the automotive domain. However, alternations of the communication partner might be restricted due to resource limitations, accessibility (e.g. no update capability) or criticality of the component. Furthermore, it might make sense to roll out a functionality as quick as possible and adapt the depending software components later on. – Knowledge about communication partners: Automotive software systems are specified through their respective requirements. In contrast to web services, the respective communication partners are thus well-known and documented. Change impact analysis is applicable (see e.g. Käßmeyer et al. (2015) and Block et al. (2019)). – Dependability and non-functional requirements: Automotive software is limited by a multitude of non-functional requirements (Liebel et al. 2018). It is restrained by hardware resources and must provide a certain level of service quality, while fulfilling dependability requirements. Consequently, resource consumption, performance and real-time requirements as well as implementation and testing efforts influence possible evolution pattern decisions. Web service evolution approaches have to be adapted to consider these points. Section 3 describes service evolution strategies. It addresses the advantages and disadvantages of co-existing software interface versions as well as their transition. Section 4 introduces different design patterns. They allow maintaining compatibility and can be used to realize the evolution strategies from section 3.

3 Software evolution strategies A software evolution strategy is a top-level procedure model about how an expected or unexpected change to a software interface should be handled. It does not concern the implementation details or how compatibility is realized. It manages the coexistence and transition of different interface versions. However, implications for the implementation level can be derived (see section 4). Consequently, service evolution strategies focus

560

Managing software evolution in embedded automotive systems on the question “What to do?” rather than “How to do it?”, whenever an interface evolves. Five types of elementary software evolution strategies exist in literature. In the following, these five patterns are presented and further developed to match the previously mentioned criteria for the automotive domain. Subsequently, they are analyzed regarding their advantages and disadvantages to derive guidelines for their usage. As such, they serve as generic and atomic building blocks for more comprehensive evolution strategies.

3.1 Instant transition strategy Instant transition means, that only one version of a software interface is supported and maintained at once (Kaminski et al. 2006). The previous interface version is not used by any input/output port in the system. Therefore, all communication partners must be changed through the same release, in which the altered software interface is.The instant transition strategy describes the most trivial approach to software evolution: It builds up on the existing change and software release management approaches on an organizational level. Additionally, it eases implementation, because there is no need to consider backwards compatibility. Only one interface version has to be maintained and compatibility issues due to different software versions cannot arise. Moreover, integration testing has to be conducted only once. However, an important precondition exists, which must hold for this strategy to be applicable: All software components using the altered interface have to be under control of the change-triggering entity. Furthermore, altering these components must be possible. Even if these preconditions are fulfilled, adapting all affected software components might require a lot of effort. Consequently, instant transitions may delay releases and hinder the continuous integration of new functionality. The creation of new business value, performance advantages and bug fixing cannot be realized instantaneously. Finally yet importantly, the entirety of changes might be too large for a single software update to be delivered collectively; especially over the air.

3.2 Compatibility strategy The compatibility strategy tries to guarantee, that an updated version of a software component can still communicate with a software component, implementing an old version of the interface (Wang et al. 2014). If for example the new and old version are incompatible, both interfaces are still maintained. Consequently, changing an interface definition and applying the compatibility pattern ensures, that there is no direct impact on the dependent software components (Wang et al. 2014). Partial and incremental alternation of interfaces is possible. Development speed is held up. This strategy should be

561

Managing software evolution in embedded automotive systems applied, if parts of the system require a new interface, while others cannot be changed at all. Following the compatibility pattern means, that a multitude of compatibility requirements must be fulfilled. However, compatibility cannot be maintained easily and architectural erosion might take place (see section 1 and section 4). The system can be stuck in architectural technical debt (Martini and Bosch 2016). Furthermore, compatibility requirements for each communication partner must be examined (Wang et al. 2014). In general, the advantages and disadvantages of section 3.1 can be inversed.

3.3 Transition strategy The transition strategy focuses on gradually fading out the old version of an interface and introducing the new one (Wang et al. 2014). The number of ports implementing the old interface decreases with every software release and update, while the number of communications via the new interface version increases. Thus, the transition strategy starts with the integration of the new interface version while the previous is still supported. The strategy is therefore quite similar to the compatibility strategy. They are often applied together. However, the previous interface is faded out afterwards. This solves the problems of the compatibility pattern in terms of architecture erosion. Therefore, the strategy should be employed to enable a smooth transition of software applications towards the new interface version (Wang et al. 2014). It is most suitable in scenarios, where the interfacing parties are changeable in general, but changing them might need some time (e.g. additional major effort; negotiations with and information of the partners, who control them). Development and release speed can be maintained and deprecated interfaces are only kept for a certain timeframe.

3.4 Split strategy Until now, the integrity or wholeness of a software component and its interfaces has not been touched. Wang et al. (2014) describes the split-map strategy for web services as an approach, which splits the interface into two new interfaces: One in which all stable operations are accumulated and another interface that includes operations with frequent and expected changes. They then require the communicating parties to map their old interfaces to the new ones. Within this paper, the second step of remapping is seen as part of other strategic building blocks like the instant transition or the compatibility strategy. Thus, the split strategy, defined in this paper, only addresses the strategic division of an interface in a stable and a change-prone part. Such interface divisions reduce future impacts on communication partners, which only use the stable part (Wang et al. 2014). Consequently, the split strategy makes sense, if only some software components require the updated interface. A new major version

562

Managing software evolution in embedded automotive systems must be released, if an interface is split, which is already existing and in use. The previously presented transition strategies can be employed to cope with this issue. During architectural development, it reduces the future impact on the connected software components without any side effects. Finally, the split strategy can allow for better service maintenance if operations in both interfaces are not overlapping (Wang et al. 2014).

3.5 Merge strategy Joining two software components as an evolution strategy may make sense, when exceptional circumstances prevail (Gamma et al. 2011, p. 187): There might be two interfaces, which are only indirectly connected through a third software component – let’s call it client. Still, the two interfaces are closely coupled: A change in one interface triggers a change in the client. The client must adapt to make the state of the altered interface consistent with the state it receives through the other interface. As such, a merger of the two interfaces may release the client from the burden of keeping the two interfaces and their states consistent. The joint interface evolves as one and is responsible for its inherent and consistent states. Consequently, the merge strategy hinders knock-on effects towards the client, because future changes are concentrated on a single interface and arise simultaneously (Gamma et al. 2011, p. 185; Wang et al. 2014). To summarize, evolution strategy patterns serve as elementary, reusable building blocks to build more comprehensive evolution strategies. For example, the compatibility pattern might be combined with the transition pattern, to support the fast deployment of a new functionality and adapt the affected components in the subsequent releases. However, these top-level strategies must be put into practice on the implementation layer. The next section discusses possible design patterns for the previously presented evolution strategies.

4 Design patterns for software evolution The evolution’s context determines the particular implementation of an evolution strategy in practice. As such, the non-functional requirements (see end of section 2) set boundaries to possible implementations and therefore limit the applicability of certain patterns. In most contexts, implementation of the changing interface and the underlying code base is subject to every-day software development. It is not discussed further in this section. However, it is challenging to find a suitable implementation strategy to maintain compatibility with the previous interface version (Kaminski et al. 2006). Thus, the implementation of the presented strategic patterns is discussed, if they are combined with the compatibility pattern.

563

Managing software evolution in embedded automotive systems The structural design patterns of Gamma et al. (2011, pp. 137–218) serve as a starting point. Further research from the web service domain is incorporated. Consequently, three different design patterns arise: Duplication, Adapter and Façade. Additionally, the usage of runtime mediation techniques is shortly discussed in section 4.4. Each pattern is examined regarding its resource consumption and performance or real-time requirements.

4.1 Duplication Duplication looks like one of the simplest implementation techniques for compatibility at first: Each interface version and its implementation are kept as separate software components. Interface versions and implementations can be arbitrarily changed. This is especially helpful, if quality of service attributes are loosened (e.g. timing constraints) and no other design pattern is applicable. Performance requirements are always fulfilled, because each version is run separately. Fading out an old version is easy, because the previous version of the software component is removed, without triggering any side effects. Consequently, the duplication pattern supports all strategic evolution patterns. However, it suffers from other defects: Firstly, resource consumption doubles, because two instances of nearly the same software are deployed. Secondly, the code of two disjointed but similar components needs to be maintained. Finally, state inconsistencies might arise, if two completely separated software components, are responsible. Thus, at some level of detail, both software versions must overlap to maintain a consistent behavior regarding the state. This can introduce its own problems: It might become necessary to maintain an “inner” interface for state management, that remains compatible with all versions of the software (Kaminski et al. 2006). Consequently, the duplication design pattern is of special interest for state-less services, if resource consumption is not a restricting factor.

4.2 Adapter An adapter is an operator, which converts the new interface version to the previous version, the communication partner expects. It is also known as the wrapper (Gamma et al. 2011, p. 139) or mapping pattern (Senivongse 1999). The adapter accumulates the differences between the current and the previous version of the software’s interface (Senivongse 1999; Kaminski et al. 2006) and allows the communication of otherwise incompatible software components (Gamma et al. 2011, pp. 139–140). Consequently, adapters avoid code duplication and encapsulate different software versions (Kaminski et al. 2006). Each piece of code is clearly assigned to one or more versions of the software: Either to the currently available software interface and its implementation or to an adapter. Thus, version transition automatically cleans up dead code. Removing an

564

Managing software evolution in embedded automotive systems adapter also removes its associated code. This is especially valuable in the case of “chain of adapters” (see Kaminski et al. 2006): The code specific to the behavior of each version is encapsulated within a separate adapter, which then calls the previous interface version. The latest interface version contains the implementation of the functionality it provides (see Figure 1). As new versions evolve, the only code that needs to be edited is the software’s current codebase and the adapter for the previous version. Consequently, adapters support the compatibility pattern joint with the transition, split and merge pattern.

Interface v1

Interface v2

Interface vn

Current Interface

Implementation Adapter vn  vn+1

Current code base

… Implementation Adapter v1  v2

Implementation Adapter v2  v3

(Primary implementation of the SwC)

Figure 1: Structure of the design pattern „chain of adapters“ from Kaminski et al. (2006)

However, the adapter might be responsible for functionality the latest version does not provide (Gamma et al. 2011, p. 140). It must incorporate additional information and operations. A major question is how this information can be gathered. As such, the adapter pattern is constraint by past and subsequent interface versions in certain circumstances (Kaminski et al. 2006). Especially performance relaxations of the adapted software cannot be corrected. Furthermore, passing a chain of adapters during runtime imposes additional performance losses. The loss is particularly dependent on the implementation pattern of the adapter: It can be implemented as a separate software component, which calls the most previous interface version, or as an additional interface of the same software component. The first implementation pattern offers better flexibility in terms of localization: If an adapter is implemented as a separate software component, it can be assigned to any suited processing unit within the distributed system. However, this imposes higher performance losses. Another advantage and challenge at the same time is the fixture of detected software defects (bugs), which is discussed in detail in Kaminski et al. (2006).

4.3 Façade A façade provides a joint interface to a set of interfaces. This makes it easier to split and merge the subordinate interfaces (Gamma et al. 2011, pp. 185–188). In comparison

565

Managing software evolution in embedded automotive systems to adapters, façades only distribute calls to the inner interfaces and orchestrate their dependencies. The implementation of the subordinate interfaces perform the actual work (Gamma et al. 2011, p. 185). Consequently, it removes the communication partners from the burden of keeping the state of the two interfaces consistent. As such, it allows implementing the merge strategy without altering the already existing interfaces. Splitting the interfaces and providing the previously joint interface through a façade enables the split strategy. Because façade is often a more lightweight implementation for compatibility, the performance losses are expected to be lower than for adapters. However, the same question in terms of implementation patterns arises.

4.4 Mediation Runtime mediation of different software versions is an approach often discussed in the web service domain (see e.g. Ponnekanti and Fox (2004) and Leitner et al. (2008)): Ondemand creation of adapters by a middleware or self-adapting web services can free the development team from manually implementing the above patterns. Ideally, compatibility is always maintained automatically. However, mediation during runtime violates performance requirements in most cases and prevents proper application. However, it can be used during development of the new interface version: Source code generation and semi-automated implementation of adapters and façades is possible (see e.g. Senivongse (1999) and Hummel and Atkinson (2009)).

5 Application In this section, the above options and guidelines for usage are applied to choose the evolution pattern for two AUTOSAR systems: A central locking system, realized on the AUTOSAR Classic Platform and a sensor fusion and trajectory-planning algorithm, based on the AUTOSAR Adaptive Platform. They display two distinct, real-world evolution scenarios and are implemented on platforms with differing degrees of flexibility.

5.1 Application 1: Evolution of a central locking system The central locking system is an experimental setup for a car with wireless key access, two doors, a rear lid and a tank flap. The model conforms to the system description in AUTOSAR Document 268 and is based on the system of a major automotive manufacturer. A more fine-grained description can be found in Block et al. (2019). The evolution scenario is as follows: The method to unlock the car needs to be extended. Besides the wireless key, image based authentication through biometrical or temporary valid image codes should be possible. This mainly affects the central software component for authentication. Two options prevail:

566

Managing software evolution in embedded automotive systems – The already existing software component can be extended, resulting in a significant change of the interface. However, the software (and hardware) components for the wireless key receiver and the related processing algorithms are not under control of the evolver. Consequently, the compatibility strategy with the adapter design pattern must be applied for the existing authentication method. In this case, it is not necessary to change the subsequent software components, responsible for opening the locks. – Alternatively, an additional software component can be introduced, which is responsible for the new authentication methods. This equals the compatibility strategy with design pattern duplication. The interface to the wireless key software components remains unchanged. Yet, resource consumption is significantly higher in this case and the subsequent software components need to be changed. However, they are all under control of the evolver. Thus, the instant transition strategy can be applied to adapt them. Performance, especially timing requirements are not critical in both cases. In the end, the second option is selected: Firstly, it is questionable if an alternation to this extend is just a new software and interface version or a completely new functionality. The separation into two software components takes this into account. Secondly, additional sensor hardware is necessary to realize the use case. Thus, additional resource consumption is a minor problem, because new hardware elements must be introduced anyway. Finally, the new feature is optional. The selected solution eases variant management.

5.2 Application 2: Evolution of an autonomous driving system The second application addresses the alternation of an interface, which is implemented in a sensor fusion algorithm, in a trajectory-planning algorithm and in multiple smart sensors (see Figure 2). The interface describes complex object lists, containing temporal-spatial information of real world objects. The smart sensors send these object lists to the sensor fusion algorithm through the respective interface. The fusion algorithm then creates an environment model. The environment model is also represented through an object list. Thus, it is forwarded through the same interface to the trajectory-planning algorithm. Consequently, the evolution of this interface affects all three types of software components. In total, the interface evolved in two steps: The first evolution comprises the alternation of the list structure to a nested hierarchy. Consequently, additional dependencies of values on a semantic level are introduced. Furthermore, an additional variable is added and the real world unit of an integer value is changed. In the second evolution step, the assigned coordinate system of one variable is altered. Syntactically, nothing changed.

567

Managing software evolution in embedded automotive systems The interface alternation between the smart sensors and the fusion algorithm (see Figure 2) is managed following the compatibility strategy, combined with the transition strategy. Compatibility is maintained with the chain of adapters design pattern. The main reason to apply this scheme is that the evolver of the interface does not manage the smart sensors. However, the implementers of the smart sensors signaled willingness to transition to the new interface within a certain time. Thus, the chain of adapters maintains compatibility and supports the slow conversion towards the changed interface versions. Additionally, the mapping between the versions is possible and only consumes little resources. The necessary, additional information is retrievable. Consequently, nearly no performance losses are expected. Coding efforts to maintain compatibility is limited to writing the adapter classes with only a few lines of code. Smart sensor 1 Smart sensor 2 …

…

Sensor fusion

Trajectoryplanning

Smart sensor n

Figure 2: Schematic excerpt of the system, showing the evolving interface and respective ports

For the ports between fusion and trajectory-planning algorithm, an instant transition scheme is chosen. They are directly under control and much closer related. Additionally, it is a one-to-one mapping.

6 Discussion and Conclusion Automotive software components are subject to constant change and variations. Thus, strategies to handle software evolution in embedded automotive systems are of crucial importance: Alternations in code might lead to a domino effect, with changes throughout the whole system. Within this paper, strategies and design patterns for software evolution are introduced. Insights from the web service domain are adapted and adjusted to suit the requirements of automotive applications. A set of guidelines supports the development of different evolution strategies to maintain compatibility while architecture erosion and technical debt is mitigated.

568

Managing software evolution in embedded automotive systems Consequently, the presented approach gives first insights into the topic of software evolution in the automotive domain. To the best of our knowledge, this is the first work to address this idea in the automotive context. However, it is based neither on a comprehensive literature review nor on extensive empirical studies1. Thus, further systematic application of the mentioned evaluation criterions combined with empirical evidence and further literature research might make the results more valuable. The resultant set of relevant design criterions may even serve as an input to discuss the design of longlasting vehicle APIs. Overall, the proposed strategy and design patterns provide reusable templates and guidelines to encourage well-defined service evolution while minimizing architectural erosion.

Bibliography 1. Block, L.; Riedel, O.; Herrmann, F. (2019): A Lifecycle Model to Support Continuous Component Evolution in Embedded Automotive Systems. In M. Bargende, H.-C. Reuss et al. (Eds.): 19. Internationales Stuttgarter Symposium. Automobil- und Motorentechnik. Stuttgart, 03/19/2019-03/20/2019. Wiesbaden: Springer Vieweg (Proceedings). 2. AUTOSAR Document 706, 3/29/2018: Explanation of Adaptive Platform Design. 3. AUTOSAR Document 268, 10/31/2018: Explanation of Application Interfaces of the Body and Comfort Domain. 4. Fehling, C. (2015): Cloud Computing Patterns. Identification, Design, and Application. PhD Thesis. University of Stuttgart, Stuttgart. Institute of Architecture of Application Systems. 5. Gamma, E.; Helm, R.; Johnson, R,; Vlissides, J. (2011): Design patterns. Elements of reusable object-oriented software. 23rd ed. Boston: Addison-Wesley (AddisonWesley professional computing series). 6. Guissouma, H.; Klare, H.; Sax, E.; Burger, E. (2018): An Empirical Study on the Current and Future Challenges of Automotive Software Release and Configuration Management. In T. Bures, L. Angelis (Eds.): 44th Euromicro Conference on Software Engineering and Advanced Applications (SSEA 2018). Prague, 08/29/201808/31/2018. Piscataway, NJ: IEEE, pp. 298–305.

1

Empirical experiments are running at the time of writing. However, results are not available yet.

569

Managing software evolution in embedded automotive systems 7. Hummel, O.; Atkinson, C. (2009): The Managed Adapter Pattern. Facilitating Glue Code Generation for Component Reuse. In S. H. Edwards, G. Kulczycki (Eds.): Formal foundations of reuse and domain engineering, vol. 5791. 11th International Conference on Software Reuse (ICSR 2009). Falls Church, VA, 09/27/2009 – 09/30/2009. Berlin: Springer (Lecture Notes in Computer Science, 5791), pp. 211–224. 8. Kaminski, P.; Müller, H.; Litoiu, M. (2006): A design for adaptive web service evolution. In B. H. C. Cheng (Ed.): Proceedings of the 2006 International Workshop on Self-Adaptation and Self-Managing Systems. 2006 International Workshop on SelfAdaptation and Self-Managing Systems. Shanghai, China, 5/21/2006 - 5/22/2006. ACM Special Interest Group on Software Engineering. New York, NY: ACM, pp. 86–92. 9. Käßmeyer, M.; Schulze, M.; Schurius, M. (2015): A process to support a systematic change impact analysis of variability and safety in automotive functions. In D. C. Schmidt (Ed.): Proceedings of the 19th International Conference on Software Product Line. Nashville, Tennessee, 07/20/2015-07/24/2015. New York, NY: ACM, pp. 235–244. 10. Khurram, M.; Kumar, H.; Chandak, A.; Sarwade, V.; Arora, N.; Quach, T. (2016): Enhancing connected car adoption. Security and over the air update framework. In: 2016 IEEE 3rd World Forum on Internet of Things (WF-IoT). With assistance of Geoff Mulligan. 2016 IEEE 3rd World Forum on Internet of Things (WF-IoT). Reston, VA, USA, 12/12/2016 - 12/14/2016. Institute of Electrical and Electronics Engineers. Piscataway, NJ: IEEE, pp. 194–198. 11. Leitner, P.; Michlmayr, A.; Rosenberg, F.; Dustdar, S. (2008): End-to-End Versioning Support for Web Services. In : IEEE International Conference on Services Computing (SCC ‘08). 2008 IEEE International Conference on Services Computing (SCC). Honolulu, HI, USA, 7/7/2008 - 7/11/2008. IEEE Computer Society. Piscataway, NJ: IEEE, pp. 59–66. 12. Liebel, G.; Tichy, M.; Knauss, E.; Ljungkrantz, O.; Stieglbauer, G. (2018): Organisation and communication problems in automotive requirements engineering. In Requirements Engineering 23 (1), pp. 145–167. DOI: 10.1007/s00766-016-0261-7. 13. Liskov, B. (1998): Data Abstraction and Hierarchy. In SIGPLAN Notices 23 (5). 14. Martini, A.; Bosch, J. (2016): Architectural Technical Debt in Embedded Systems. In INCOSE International Symposium 26 (1), pp. 1029–1043. DOI: 10.1002/j.23345837.2016.00209.x. 15. Maul, M.; Becker, G.; Bernhard, U. (2018): Serviceorientierte EE-Zonenarchitektur Schlüsselelement für neue Marktsegmente. In ATZelektronik 13 (1), pp. 36–41.

570

Managing software evolution in embedded automotive systems 16. Meyer, B. (1992): Applying ‘design by contract’. In Computer 25 (10), pp. 40–51. DOI: 10.1109/2.161279. 17. Papazoglou, M. P. (2008): The Challenges of Service Evolution. In Z. Bellahsène, M. Léonard (Eds.): Advanced Information Systems Engineering, vol. 141. CAiSE: International Conference on Advanced Information Systems Engineering. Montpellier, France, 06/16/2008 – 06/20/2008. Berlin: Springer (Lecture Notes in Computer Science, 5074), pp. 1–15. 18. Papazoglou, M. P.; Andrikopoulos, V.; Benbernou, S. (2011): Managing Evolving Services. In IEEE Softw. 28 (3), pp. 49–55. DOI: 10.1109/MS.2011.26. 19. Ponnekanti, S. R.; Fox, A. (2004): Interoperability Among Independently Evolving Web Services. In H.-A. Jacobsen (Ed.): Proceedings of the 5th ACMIFIPUSENIX international conference on Middleware, vol. 3231. New York, NY: SpringerVerlag New York Inc (ACM Digital Library), pp. 331–351. 20. Senivongse, T. (1999): Enabling flexible cross-version interoperability for distributed services. In Z. Tari (Ed.): Distributed Objects and Applications. International Symposium on Distributed Objects and Applications. Edinburgh, UK, 09/06/1999 09/06/1999. Los Alamitos: IEEE Computer Society Press, pp. 201–210. 21. Vitale, J.; Bowman, K.; Singh, R.; Robinson, R.; Giffi, C.; Hecker, M. et al. (2020): 2020 Global Automotive Consumer Study. Is consumer interest in advanced automotive technologies on the move? Deloitte Development LLC. 22. Wang, S.; Higashino, W. A.; Hayes, M.; Capretz, M. A. M. (2014): Service Evolution Patterns. In : Proceedings of the International Conference on Web Services. 2014 IEEE International Conference on Web Services (ICWS). Anchorage, AK, USA, 06/27/2014 – 07/02/2014: IEEE, pp. 201–208. 23. Weyns, D.; Caporuscio, M.; Vogel, B.; Kurti, A. (2015): Design for Sustainability = Runtime Adaptation ∪ Evolution. In M. Galster, I. Crnkovic (Eds.): Proceedings of the European Conference on Software Architecture Workshops. European Conference on Software Architecture (ECSA 2015). Dubrovnik, Cavtat, Croatia, 9/7/2015 - 9/11/2015. New York, New York: Association for Computing Machinery (ICPS), pp. 1–7. 24. Xavier, L.; Brito, A.; Hora, A.; Valente, M. T. (2017): Historical and impact analysis of API breaking changes. A large-scale study. In M. Pinzger, G. Bavota et al. (Eds.): SANER 2017. 2017 IEEE 24th International Conference on Software Analysis, Evolution and Reengineering (SANER). Klagenfurt, Austria, 02/20/2017 02/24/2017. Institute of Electrical and Electronics Engineers. Piscataway, NJ: IEEE, pp. 138–147.

571

Agile systems engineering for critical systems Christof Ebert Vector Consulting Services Frank Kirschke-Biller Ford

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_49

573

Agile systems engineering for critical systems

1 The Magic Triangle of Cost, Quality and Innovation Participants in our 2020 Vector survey observe three significant challenges [1]. Cost and efficiency have emerged as the single most relevant short-term challenge, indicating the need to succeed in a fast-changing world with unclear business drivers. At the same time, quality is further growing compared to our last-year survey as short- and mid-term target. Obviously, the threats of product liability and global visibility of insufficient quality have reached technology companies. Time is gone when software could mature with the early adopters at a bleeding edge. What is delivered must be mature. Innovation obviously matters for all companies Figure 1 captures the results of recent Vector survey. The horizontal axis provides shortterm challenges, while the vertical axis shows more mid-term obstacles. Each reply allowed up to five challenges in both dimensions, adding to the sum of 100%. Thanks to a response rate of nearly 5% covering different domains, the survey well represents different B2B business models and diverse regions in the world. According to the Vector industry survey, companies struggle with three major forces along the magic triangle: ●

Cost and Efficiency: i.e., development cost, price pressure, rework, IT effort

●

Quality: i.e., functional safety, cybersecurity, governance, compliance

●

Innovative products and digital transformation: i.e., service orientation, product management, business processes, convergence of IT and embedded technologies.

Cost first is a recipe for failure. A simple example is canceling trainings of engineers to save on cost. In turn they are demotivated and lack the necessary energy and know-how to master the crisis. A viscous circle will start. Fig. 2 shows this negative pattern which we currently see growing. With sole focus on cost pressure and cost reduction programs, innovation and quality suffers. At the same time good people start looking for alternatives which further reduces innovation and competences. Rework and delays mean more cost, which drives more cost pressure. It is a downward moving cycle that many companies currently feel. Red-ocean scenarios will inevitably grow if companies focus too much on cost. In fact, we have seen several once strong players drowning in cost savings programs starting last year, and now being unable to recuperate along the necessary software innovation. Once big players such as Thyssen-Krupp need to sell valuable assets just to survive. Mastering this magic triangle mans a strong balance of energy with an agile approach, that is “time boxing” and “expense boxing” with a restrictive portfolio management to

574

Agile systems engineering for critical systems avoid that scarce resources are wasted in firefighting. Let us look towards recommendations for these themes which we have identified in our projects and in numerous interviews with industry experts over the past months.

2 Agile Systems Engineering Agile development has arrived in automotive electronics. Globally distributed teams, large projects, continuous updating "over the air", safety-critical systems as well as complex hardware and systems engineering require a targeted adaptation of agile methods. The biggest and most difficult challenge, however, is the change in thinking. And this does not only concern engineers. The automotive industry is at a turning point. Global growth has slowed down, but at the same time successful companies are growing with a shift towards ACES (Autonomy, Connectivity, E-Mobility, Services). These four key drivers represent a huge transformation, for OEMs as well as Tier-1. Traditional mechanics are increasingly being replaced by electronic systems and software functions, and traditional vehicle use by mobility services. The transformation is most evident in system and software development. Softwareintensive systems are under enormous market pressure. While they have to be uncompromisingly innovative in terms of technology and customer function, the markets demand ever shorter cycle times and continuous efficiency increases while maintaining high safety standards. Flexibility and costs must be constantly improved in global competition. Bringing innovations with good quality to the market and increasing efficiency at the same time is the main topic of the current Vector customer survey "Industry Trends 2019" [1]. Traditional development processes that achieve innovation and quality through a complex development and validation process are only partially effective for highly innovative products. Our customer projects and studies show that rework costs can be reduced by 20-50% over the entire product life cycle by improving requirements and systems engineering [1-4]. Agile techniques are the essential lever for flexibility and efficiency. But how is agility implemented? Textbook methods do not fit to industrial practice in automotive electronics. Performance, functional safety, cycle security and a growing global variety of variants require specifically adapted development processes - starting with systems engineering. Governance and traceability must be balanced in the specific context with lean procedures. More recent agile methods such as SoS, LeSS and SAFe are suitable for large projects, but at the same time complex and associated with many implementation risks. They

575

Agile systems engineering for critical systems limit agility by extremely high overhead due to complex processes and requirements. [5]. In addition, they do not address the specific challenges of critical systems such as functional safety and the agile collaboration of several development partners. Our conclusion after almost two decades of agile transformation processes is that each organization needs a dedicated, adapted process model that fits its boundary conditions of market, technology and culture. A standard for everyone does not fit, as the CMMI with its hard-defined maturity levels already showed. This still applies today to SAFe and other complex agile method frameworks. We have therefore adapted the necessary agile system and software development processes together with and for Ford individually - without using a pre-designed framework from the method kit. The initial focus was on agile requirements and systems engineering, because requirements and architecture decisions influence all further steps. Agile systems engineering is the decisive success factor for the efficient development of software-intensive electronics. It supports the continuous and incremental development of requirements up to validation and above all the further development in the life cycle according to SOP. Due to the increased degree of abstraction with a focus on the customer function at the beginning of the requirements development and analysis, problem descriptions are much clearer, easier and less redundant. This not only increases the development speed, but also ensures clearly understandable domain concepts within the project. Models help with the consistency of system requirements over software requirements up to design and validation In the implementation with Ford, we identified four key levers for agile systems engineering (Fig. 1): [2]. 1) Adapted agile methodology ● Hierarchical scrum for system engineering and orchestration of cross design teams, ● Combined feature roadmaps with accelerated design decisions ● Flexible synchronization of early software and hardware deliveries and reduction of synchronization points 2) Modeling and simulation ● Modeling and simulation as well as model-based system technology and hardware simulation ● Architectural analysis, technical debts and their impact on efficiency ● Early experience of functions

576

Agile systems engineering for critical systems 3) Collaboration models ● Agile cooperation of teams distributed worldwide – also from different suppliers – with requirements, design, development, validation and integration. – Efficient and tool-based distributed knowledge management ● Mapping of central roles such as an embedded safety officer in all Scrum teams. 4) Quality and efficiency ● Stringent regression techniques for the agile partial deliveries ● Reuse of hardware and software elements, with traceability for end-to-end change management ● Key figures on efficiency in the development process, e.g. Function points, benchmarks and trend-opportunity indicators The biggest challenge with agile system engineering is that the goal is defined and further developed during the development based on iterations (Fig. 2). Development is therefore not primarily driven by a rigid process, but rather by an evolving product. A cornerstone at Ford is a feature-based process that is based on a feature dictionary with subordinate functions and software components. This company-wide information model forms the basis for global cooperation and reusability. Project-specific feature implementation planning serves as backlog input for ECU development in an orchestrated approach and enables agile and coordinated functional growth. The prioritization of functions and definition of a minimally sustainable product as well as periodic integration phases for the solution are essential elements of the planning. An OEM in particular sees the critical dependencies on suppliers: Agile systems and vehicle technology require full transparency from and with the suppliers. This includes regular software maturity reports from suppliers with a focus on delivery dates, featurebased implementation and test status reports, which include earned value management, full transparency in open questions, tool integration with defined interfaces between OEM and supplier, and the coordination and traceability of distributed features to close the circuit. Agile systems engineering encompasses the entire product lifecycle, not just the activities associated with requirements analysis as was previously the case. Therefore, software and system engineering require a comprehensive PLM solution. The underlying information model must therefore take different perspectives and layers of abstraction in order to move from vehicle functions to component requirements and system specifications.

577

Agile systems engineering for critical systems In order to achieve consistency between these different levels, content is handled transparently and reused in related documents. This facilitates the traceability from the functional level via the system and the components as well as the simple regression activity and the consistency in derived documents. This also takes into account the central requirement of traceability of requirements according to the automotive SPICE.

3 Results Ford has achieved the following results with this agile transformation: ●

Reliability: Self-organization considerably reduces the amount of cooperation in comparison to the classic project management approach.

●

Transparency: The transparency of the project status and thus the project management is significantly improved.

●

Efficiency: speed and quality are improved. Management and teams recognize and appreciate improvements and agile team spirit. In particular, the early definition of requirements and the acceptance of changes during development as part of an agile process are essential for efficiency.

●

Complexity: Product complexity is better managed and controlled

●

Quality: Reuse and embedded quality responsibility create better quality

Of course, these results require professional change management. For agile transformation in such a complex context, we recommend power users on board, sensible piloting, parallel coaching, training, support, continuous improvement. Planning the changes is also important. We recommend the integration of improvements and necessary costs directly into the annual budgets of the product lines. Agile development has arrived in automotive electronics. Globally distributed teams, large-scale projects, safety-critical systems or hardware and system technology have shown that agile technologies can also be adapted for safety-critical systems. It is not a question of compulsorily using a complex procedure model, but rather of selecting individual agile methods as suitably as possible and then continuously optimizing them based on practical experience. This is the only way that the agile culture arrives at the engineers and their management. The biggest and most difficult change is the change of mindset. Changing practices only does not make a company agile if the underlying culture and thinking do not change. The agile transformation is therefore also a central management task. Or in the words of the poet prince and politician J.W. von Goethe: "It is not enough to know, you have to use it; it is not enough to want, you have to do it."

578

Agile systems engineering for critical systems

Fig. 1: The vicious circle driven by cost pressure at the expense of competence and quality

Fig. 2: Four success factors for agile Systems Engineering

579

Agile systems engineering for critical systems

Fig. 3: Agile Systems Engineering creates consistency and reduces overheads (Source: Ford)

Authors Christof Ebert is managing director and co-founder of Vector Consulting Services. As a consultant, he supports companies worldwide in product development, strategy and in agile transformations. In his work on supervisory boards and as a business angel, author, sought-after speaker and professor at the University of Stuttgart and the Sorbonne in Paris, he combines industry and innovation. He can be reached at: [email protected] Frank Kirschke-Biller headed the global core software processes at Ford. Since 2000 he has been with Ford in various management positions in the areas of infotainment, integration and electronics architecture, networks, diagnostics and software development. Previously, he headed the department for sensors and system technology at a startup in the field of mechatronics. Frank Kirschke-Biller studied electrical engineering at the University of Duisburg.

580

Agile systems engineering for critical systems

Bibliography 1. Vector: Industry Trends 2020: Trends for Thriving Twenties: Industry Survey and Stimulus. www.vector.com/trends 2. Frank Kirschke-Biller et al: Agile Systems Engineering at Ford 3. https://vector.com/portal/medien/vector_consulting/publications/AgileSystemsEngineering_Vector_Ford.pdf 4. Ebert, C.: Systematisches Requirements Engineering. ISBN: 978-3-86490-139-3, Dpunkt-Verlag, Heidelberg, Germany, 6. komplett überarbeitete Auflage, 2019 5. Ebert, C. and J.Favarro: Automotive Software. IEEE Software, ISSN: 0740-7459, vol. 34, no. 3, pp. 33-39, May/Jun 2017 6. Ebert, C. and M. Paasivaara: Scaling Agile. IEEE Software, ISSN: 0740-7459, vol. 34, no. 6, pp. 98-103, Nov/Dec 2017

581

Discretization and heat transfer calculation of engine water jackets in 1D simulation Florian Mandl, Michael Bargende, Michael Grill IVK (University of Stuttgart), FKFS

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_50

583

Discretization and heat transfer calculation of engine water jackets in 1D simulation

1 Abstract The industry is working intensively on the precision of thermal management. By using complex thermal management strategies, it is possible to make engine heat distribution more accurate and dynamic, thereby increasing efficiency. Significant efforts are made to improve the cooling efficiency of the engine water jacket by using 3D CFD. As well, 1D simulation plays a significant role in the design and analysis of the cooling system, especially for considering transient behaviour of the engine. In this work, a practice-oriented universal method for creating a 1D water jacket model is presented. The focus is on the discretization strategy of 3D geometry and the calculation of heat transfer using Nusselt correlations. The basis and reference are 3D CFD simulations of the water jacket. Guidelines for the water jacket discretization are proposed. The heat transfer calculation in the 1D-templates is based on Nusseltcorrelations (𝑁𝑢 = 𝑁𝑢(𝑅𝑒, 𝑃𝑟)), which are derived from 3D CFD simulations. Furthermore, additional modifications in the calculation routines and helpful criterions allow to handle the specific characteristics of the complex water jacket geometry userfriendly. The 1D results are compared in detail with the 3D CFD reference and show very good agreement for various operating points regarding inlet mass flow rate and inlet temperature. The methodology is developed using a passenger car engine (Test Engine I) and validated on a large engine (Test Engine II). Whereas the passenger car engine has a water jacket with the cylinders in cross flow, the large engine has single cylinder cooling with flow along the cylinder axis. The analysis of two engines with different water jacket flow characteristic shows the general validity of the developed methodology.

2 Introduction In the design and optimization of the water jacket, component cooling (protection) is the primary objective. In addition, the potential for increasing efficiency on the engine side (reduction of frictional power) and for reducing emissions must be exploited. The requirements are becoming increasingly complex, since modern concepts require not only cooling but also rapid heating of the engine. Rapid engine warm-up has a positive effect on fuel consumption and ensures reliable exhaust gas aftertreatment (lightoff temperature of the catalytic converter). However, some phenomena inside the water jacket can only be measured with great effort or not at all. The ever-present time and cost pressure makes the integrated use of 1D and 3D CFD simulations an even more important component of cooling circuit and especially water jacket development. The 1D simulation enables fundamental concept studies at an early stage of development. It impresses with short computing times, which also enable the analysis

584

Discretization and heat transfer calculation of engine water jackets in 1D simulation of transient processes. The 3D CFD analysis provides information on local flow velocities and heat transfer coefficients [1]. In the 1D simulation, real geometries are modelled with the help of elementary flow objects such as pipe sections, orifices and branches. These geometrical simplifications should reflect the real physical processes as well as possible and have to be made accordingly. The heat transfer calculation in the 1D simulation is based on dimensionless quantities (Nusselt correlations), which are derived from considerations of similarity theory. Software for 1D flow simulation provides the common correlations (Colburn, Gnielinski, Dittus-Boelter, Sieder-Tate) [2] [3]. These approaches and their area of applicability are described in detail in the standard literature [4] [5] [6]. They are applicable for simple flow problems under certain conditions, but do not provide satisfactory results for the complex flow geometry of a cooling water jacket [7] For concrete applications the heat transfer in the 1D simulation has to be calibrated with the help of 3D CFD and/or measurements [8]. A basis and overview of the thermal modelling of the engine in the 0D/1D simulation can be found in the literature [9] [10]. In the following, a practice-oriented general methodology for discretization and heat transfer calculation (solid to coolant) of a cooling water jacket is presented.

3 1D Discretization of the Cooling Water Jacket In this context, the term discretization refers to the spatial subdivision of the real three-dimensional flow region into sub-volumes. In the 1D simulation, these subvolumes are modelled by suitable objects, which should reproduce the physical threedimensional behavior as accurately as possible. In general, discretization has an influence on the quality of the results of a 1D simulation, so that the question of the ideal discretization strategy arises. In the following, the strategy is defined on the basis of guiding points. In general, a finer resolution (discretization) of the flow problem in the 1D simulation leads to a higher model complexity regarding model creation, calibration and application. Therefore, the discretization should not only depend on the geometric and physical properties of the problem, but should also be appropriate to the application scenario of the simulation model.

585

Discretization and heat transfer calculation of engine water jackets in 1D simulation

3.1 Application Scenarios, observed Variables and Boundary Conditions of a 1D Cooling Circuit Model The advantage of 1D simulation lies in the short computing times, even on standard office hardware. This enables a relatively inexpensive consideration of many variants of a problem as well as the investigation of transient processes. Modern combustion engines, which can fulfil today's technical and legal requirements, are equipped with variable technologies, e.g. with regard to fuel injection or valve train. This development towards higher variability takes place in engine’s thermal management, too [1]. For example, variable pumps, various fluid circuits and intelligently controlled (thermostatic) valves optimally regulate heat supply and heat dissipation. Due to the dynamic behaviour of a vehicle engine and its thermal boundary conditions, as well as the important warm-up behaviour after a cold start, a transient consideration of the cooling system is necessary. When creating a 1D model of a cooling water jacket, the later application purpose and the desired observation variables should be defined and taken into account. If, for example, the combustion engine is considered as a simple heat source in the overall cooling system, the real cooling water jacket geometry can be abstracted up to a single 1D flow volume (see Figure 1a). With such a strong simplification the geometry of the real water jacket remains nearly unconsidered, apart from total volume and (heattransferring) surface. Consequently, the flow properties, and thus also the heat transfer, can only be modelled globally for the cooling water jacket. The usable and realistic observable variables of such a 1D model are the total heat input of the engine into the coolant circuit or the coolant outlet temperature.

Figure 1: a) Engine coolant jacket modelled with a 1D flow element; b) Engine coolant jacket modelled with subdivision into engine block and cylinder head; c) Water jacket discretization schematic illustration: Engine block and cylinder head are subdivided along the cylinders and into exhaust and intake sides.

In this case, the effort involved in creating and calibrating models is low, also the need for information about the problem and the boundary conditions for modelling. Such a simple 1D water jacket model provides only a few global result variables

586

Discretization and heat transfer calculation of engine water jackets in 1D simulation (e.g. total heat input, coolant outlet temperature) in the calibrated operating ranges. This means that no local result variables (e.g. component temperatures) are available and the prediction capability is very limited. Starting from the simplest modelling shown in Figure 1a, the obvious and reasonable model refinement is the subdivision of the cooling water jacket into engine block and cylinder head (see Figure 2b). The subdivision of the cooling water jacket into block and head makes sense for several reasons: ● The thermal boundary conditions differ in the two regions. ● The flow characteristics in the block are different from those in the head, so it is recommended to describe the two regions separately with regard to heat transfer. ● Usually the cooling water jacket has an outlet in the engine block and in the cylinder head. The distribution of the input mass flow to the two outlets including outlet temperatures can be modelled. For a finer discretization of the water jacket, as presented in the following, information from the inside of the water jacket is required to calibrate the 1D simulation model. For this purpose, a 3D CFD simulation of the water jacket is recommended. To determine the ideal degree of discretization, the following criteria can already be derived from the above simple modelling approaches: ● The discretization must be adapted to the later application purpose of the 1D water jacket model. ● A finer discretization means higher calibration effort. ● A finer discretization also increases the number of boundary conditions, which should be present according to the selected discretization. The modelling of the cooling water jacket with one or two 1D flow objects cannot reflect the real flow characteristics, i.e. the heat input is also treated abstractly in such a model. With a finer discretization of the water jacket geometry, the flow can be reproduced more realistically in the 1D simulation and thus the heat transfer can be treated in more detail. In the following section, the discretization of the water jacket Test Engine I is described. Figure 1c illustrates the basic principle. For reasons already described, the division into engine block and cylinder head is applied, and in addition each cylinder is divided into an intake and an exhaust side in both regions.

587

Discretization and heat transfer calculation of engine water jackets in 1D simulation

3.2 1D-Discretisation of the Cooling Water Jacket of the Test Engine I Test Engine I is a turbocharged 4-cylinder passenger car diesel engine with 2.0l displacement, the cooling water jacket geometry is shown in Figure 2 (left). On the intake side (i.e. on the engine side where the intake valves are located) on cylinder 1 there is the cooling water pump or the inlet to the engine water jacket. Also on the intake side in the engine block at cylinder 4 is an outlet of the water jacket. The second outlet is located in the cylinder head on the intake side of cylinder 4.

Figure 2: 3D geometry of the cooling water jacket Test Engine I with an inlet and two outlets (left), stream lines of the cooling water jacket Test Engine I (right).

The discretizing/division of the real 3D water jacket geometry into sub-regions is based on the geometric basic elements, but the model complexity should also be kept in mind, i.e: ● Summarize regions according to the flow characteristics. ● Realistic boundary conditions for 1D flow elements must be given. ● For a discretization according to Figure 1c, a 3D CFD simulation of the water jacket should be used to calibrate the 1D model. ● The heat transfer, as a further result variable and reference from the 3D CFD, should be considered in the discretization. Figure 3 shows the water jacket geometry and the discretization selected here (right) from above: The engine block region is shown in blue, the cylinder head region in red and the cylinder head gasket in magenta.

588

Discretization and heat transfer calculation of engine water jackets in 1D simulation The water jacket around the cylinder liner is divided into approximately 90° segments; in addition, a small area in the constriction between two cylinders is discretized as a separate flow volume. The reasons for this are: ● A flow direction reversal of approx. 270° takes place here, which tends to lead to pressure losses and, under certain circumstances, to an increased heat transfer due to interaction with the wall. ● In order to come close to the basic idea of the 1D flow simulation, it is recommended to set the interfaces orthogonal to the flow direction. ● In the area of the constriction between two cylinders there is also a flow interface to the cylinder head. ● The thermal load at the constrictions is higher.

Figure 3: Discretization of the water jacket of Test Engine I, top view.

The complex water jacket geometry in the cylinder head area is divided as desired along the cylinders and into inlet and outlet sides. This makes sense with regard to flow field and thermal boundary conditions. Flow interfaces are combined according to the flow characteristics.

4 Modelling of the Heat Transfer in the Cooling Water Jacket The calculation of the heat transfer between wall (W) and fluid (F) is done here in the 1D simulation according to Newton's approach (Eq. 1) [2] [6]. 𝑄 = 𝛼 ⋅ 𝐴 ⋅ (𝑇

𝑇 )

(1)

589

Discretization and heat transfer calculation of engine water jackets in 1D simulation The heat transfer coefficient 𝛼 depends on a characteristic length, the thermal conductivity 𝜆 of the fluid, as well as its characteristic flow velocity 𝑣, density 𝜌, dynamic viscosity 𝜂 and the isobaric specific heat capacity 𝑐 .

4.1 Nusselt Correlation as a Heat Transfer Approach In order to formulate the heat transfer coefficient with its dependencies, an approach with dimensionless numbers is chosen from the similarity theory. Eq. 2 shows the relationship between 𝛼 and the Nusselt number 𝑁𝑢, which describes the forced convective heat transfer between solid surface and flowing fluid. 𝛼 = ⋅ 𝑁𝑢

mit

𝑁𝑢 = 𝑁𝑢(𝑅𝑒, 𝑃𝑟) = 𝑎 ⋅ 𝑅𝑒 ⋅ 𝑃𝑟

(2)

The Nusselt number can be formulated by the Reynolds number 𝑅𝑒 and Prandtl number 𝑃𝑟. Eq. 2 shows the general form of a so-called Nusselt correlation with three constants 𝑎, 𝑏, 𝑐. In the literature it can be found suitable Nusselt correlations with certain constants 𝑎, 𝑏, 𝑐 [4] for geometrically simple flow problems. For the heat transfer in the complex water jacket geometry, these correlations do not provide results with satisfactory accuracy [7]. Therefore, it is necessary to determine the constants 𝑎, 𝑏, 𝑐 of the Nusselt correlation for the cooling water jacket using 3D CFD simulations.

4.2 Nusselt Correlation in 1D Simulation For the development of the methodology regarding 1D simulation of the heat transfer in the engine water jacket, the 3D CFD serves as reference in this work. Figure 4 shows a section of the water jacket in the region of the cylinder liner. In this figure, the coolant flows from left (index 1) to right (2) through the volume along the cylinder, while a partial mass flow leaves the volume upwards (3) towards the cylinder head. The figure is intended to illustrate the problem: The 3D CFD reference on the left in the figure is to be modelled with a 1D approach, which is schematically shown on the right. The cooling water jacket is modelled here in the 1D simulation with general 1D flow volumes. These represent the respective discretized area as an elementary volume in which all quantities (pressure, fluid temperature, material properties, etc.) are constant with regard to the spatial distribution. Boundary conditions such as a wall temperature are also set as constant values, and averaged values are also available at the flow interfaces (index "FIF"; inlets and outlets of the volume). Thus, it is necessary to calculate the heat transfer as accurately as possible with these mean values and boundary conditions of the 1D flow simulation using a Nusselt correlation (Eq. 2).

590

Discretization and heat transfer calculation of engine water jackets in 1D simulation

Figure 4: Flow field and heat transfer coefficient from 3D CFD (left) in compared to the 1D model approach (right).

Figure 4 right schematically illustrates a general 1D flow volume. The illustration shows that information on the fluid movement is only available at the flow interfaces. can be used to formulate a Nusselt correlation 𝑅𝑒 and the Prandtl number 𝑃𝑟 (Eq. 3) for each according to Eq. 2 and to determine a heat transfer coefficient 𝛼 flow interface of the 1D volume. 𝛼

=

,

⋅ 𝑁𝑢

=

,

⋅ (𝑎 ⋅ 𝑅𝑒

⋅ 𝑃𝑟

)

(3)

It should be noted that in this calculation the flow boundary conditions are used directly from the interfaces (index "FIF"), since they are undefined within the 1D volume. Temperature-dependent material quantities, such as those found in the Prandtl number, are used from within the 1D volume (index "1DVol"). With Eq. 3, a heat transfer coefficient 𝛼 is now defined for each flow interface of a for 1D volume. In the next step, to obtain a common heat transfer coefficient 𝛼 the entire 1D volume, a mass flow-weighted mean value is calculated according to Eq. 4 [2]. 𝛼

=

∑

𝛼

⋅𝑚

where

𝑚

=∑

𝑚

(4)

The weighting of 𝛼 with the mass flow 𝑚 (sign-less) at the respective flow interface represents a reasonable, fundamental approach.

4.3 Calibration and Modifications of the Heat Transfer Calculation The Nusselt correlation (Eq. 2) contains three parameters 𝑎, 𝑏, 𝑐, which are adjusted using 3D CFD simulations of the water jacket. In the 3D CFD evaluation, global parameters (e.g. total heat input, outlet temperatures) as well as subregions of the

591

Discretization and heat transfer calculation of engine water jackets in 1D simulation water jacket are considered analogous to the 1D discretization and evaluated with regard to heat input, temperature, mass flows, etc.

4.3.1 Operating Points for Calibration A cooling water jacket in the combustion engine is essentially characterized by two operating parameters: The fluid mass flow through the water jacket and the fluid temperature. Both variables are included in the Nusselt correlation and thus in the heat transfer calculation: ● The fluid mass flow and the water jacket geometry result in flow velocities, of which the Reynolds number 𝑅𝑒 depends. ● The fluid temperature directly determines the heat flow between fluid and wall according to Eq. 1 and has an indirect influence on the heat transfer via temperature-dependent material quantities, e.g. in the form of the Prandtl number 𝑃𝑟. Which operating points are necessary for the calibration of the 1D heat transfer, and thus must also be simulated in the 3D CFD, depends, among other things, on the later application purpose of the 1D model: ● The temperature dependence of the material values is low for an engine at operating temperature (80-90°C). If the 1D coolant jacket model is to cover only these operating ranges later, a coolant-temperature variation can be omitted. For the Prandtl number exponent 𝑐 in the Nusselt correlation, the default value can be 𝑐 = 0.4 [6] and thus reduce the number of calibration parameters to two. ● With respect to fluid temperature, the operating points for model calibration should cover the temperature range that the 1D model should reflect. During a cold start (20°C) or even under low temperature conditions (< 0°C), the viscosity increases exponentially. ● In general, at higher engine speeds, more heat must be dissipated from the combustion engine, which can be achieved by increasing the coolant mass flow. In classic mechanically driven coolant pumps, the mass flow depends directly on the engine speed. The Reynolds number 𝑅𝑒 depends directly on the fluid velocities, so that its exponent 𝑏 must be calibrated on the basis of the mass flow or engine speed range to be considered.

4.3.2 Heat Transfer Calculation for low Reynolds Numbers A Nusselt correlation according to Eq. 2 describes the heat transfer of a turbulent flow. The Reynolds number 𝑅𝑒 characterizes the present flow state with regard to turbulent or laminar flow, with a transition region in between. Thus, for low Reynolds

592

Discretization and heat transfer calculation of engine water jackets in 1D simulation numbers, due to low flow velocities and/or low fluid temperatures (i.e. high viscosity), the heat transfer calculation must be modified.

4.3.3 Weighting Factor for Flow Interfaces The common heat transfer coefficient of a 1D flow volume 𝛼 corresponds to the mass flow-weighted mean value of the individual flow interface-related 𝛼 (Eq. 4). (Eq. 5) should be evaluated for For the 1D model calibration, the parameter 𝑥 , plausibility. 𝑥

,

=∑

⋅

where

⋅

∑

𝑥

,

=1

(5)

𝑥 , corresponds to the fraction to which a flow interface "FIF" contributes to the common 𝛼 . In combination with the analysis of 3D CFD results of the corresponding region, flow interfaces can be identified with an implausibly high contribution. An example of this is the cooling water outlet in the block on Test Engine I. In the discretization applied here, this coolant outlet represents one flow interface (of a according to Eq. 4, total of five in this 1D flow volume). In the calculation of 𝛼 this provides a high contribution due to the relatively high outgoing mass flow, which cannot be supported by the analysis of the 3D CFD results. In this case, the flow interface of the cooling water outlet is weighted too much and the actual heat input of the 1D flow volume is not modelled correctly. In order to maintain the previous discretization in such cases, a additional factor 𝐹 is introduced for the weighting of 𝛼 , see Eq. 6. With this extension in the calculation approach, the influence of individual flow interfaces can be reduced. However, this intervention should be justified by further information (e.g. 3D-CFD). =

𝛼 where

𝑚

1 𝑚

=

𝛼 𝑚

⋅𝐹

⋅𝑚

⋅𝐹 and

(6) 𝐹

∈ [0, 1]

This method can also be justified for the flow interfaces at the cylinder head gasket.

5 Results of Test Engine I The aim of this research work is to achieve an accuracy of ±5% for the 1D cooling circuit model in terms of total heat transfer. Figure 5 shows the percentage deviation of the heat transfer between 1D model and 3D CFD reference. The heat input total (green), in the engine block (blue) and in the cylinder head (magenta) is evaluated for the calibration and validation points.

593

Discretization and heat transfer calculation of engine water jackets in 1D simulation The coolant mass flow rises (at constant inlet temperature) for the calibration points 1-7, i.e. the Reynolds number also shows an increase. The Prandtl number remains almost constant, there is only a minimal increase due to the increasing viscosity (at higher mass flow rates the fluid temperature at the outlet decreases). Points 8-12 increase the coolant inlet temperature (at constant mass flow), resulting in a drop in the Prandtl number. Due to the decreasing viscosity, the Reynolds number rises slightly over these operating points. Thus the calibration points cover the mass flow and fluid temperature influence with regard to heat transfer and the factors or exponents of the Nusselt correlation can be determined in a comprehensible and tangible way.

Figure 5: Percentage deviation (1D - 3D CFD) of the heat input into the coolant. Green: Deviation in the entire water jacket, blue: Deviation in the engine block, magenta: Deviation in the cylinder head.

For the calibration points 1-12 in Figure 5, the percentage deviation of the heat input between 1D and 3D CFD is in the range of ±2% for the engine block and cylinder head. The total heat input is even within ±1% accuracy, as the slight deviations in the block and head are partly compensated. The result of this calibration can be evaluated as very good: The influence of 𝑚 and 𝑇 (or 𝑅𝑒 and 𝑃𝑟) on the heat transfer is reproduced very well in the separately calibrated block and head regions. The respective parameters 𝑎, 𝑏, 𝑐 of the Nusselt correlation are constant over the operating points.

594

Discretization and heat transfer calculation of engine water jackets in 1D simulation The validation points 13-18 are further variations concerning 𝑚 and 𝑇 . Also for these operating points the model calibration fulfills the desired accuracy. The deviations of the total water jacket, the block and the head behave in the tendency the same and supply an accuracy of ±4%. Overall, the results are very satisfactory and should always be considered in relation to the accuracy of measurements or boundary conditions.

6 Validation of the Methodology using Test Engine II The discretization methodology was explained and the approaches for heat transfer calculation in the 1D simulation were formulated on the basis of Test Engine I. The results of the first phase of the project were presented. This has to be validated by means of Test Engine II. In Test Engine I, a passenger car diesel engine, the cylinder liners are typically in cross flow of coolant. Test Engine II, on the other hand, is a large engine with single cylinder cooling in which the liner flow is typically along the cylinder axis. The validation using a water jacket with different flow characteristics is intended to ensure the general validity of the approaches and methodology.

6.1 Specification Test Engine II Test Engine II is an MTU 20V 4000, i.e. a diesel-powered 20-cylinder V-engine of the MTU Series 4000 [11]. For the cooling circuit simulation, however, only one cylinder bank is considered, since the water jacket of the two banks is not directly connected. The 3D geometry of the water jacket of a cylinder bank is shown in Figure 6 top. The cooling water jacket can be geometrically divided into the following regions: Inlet rail, outlet rail, liner, combustion chamber roof, cylinder head. The liner, combustion chamber roof and cylinder head of a cylinder unit and are geometrically identical for all cylinders.

595

Discretization and heat transfer calculation of engine water jackets in 1D simulation

Figure 6: Top: 3D geometry of the water jacket of Test Engine II (one cylinder bank). Bottom: 1D flow model (here in GT-POWER) of the cooling water jacket.

6.2 3D CFD Simulation Water Jacket Test Engine II The 3D CFD simulation also provides a detailed flow field of the cooling water jacket for Test Engine II. This is considered in the discretization and in general the 3D-CFD is the reference for the 1D cooling water jacket simulation. The 3D CFD model is created and configured in the same way as Test Engine I. Figure 7 shows the stream lines of a cylinder and wall-near temperatures as a result of the 3D CFD.

Figure 7: 3D CFD simulation of Test Engine II: Wall-near temperatures and stream lines.

596

Discretization and heat transfer calculation of engine water jackets in 1D simulation The figure (see also Figure 6) illustrates the geometry and flow characteristics of the cooling water jacket, which can be described in the following points: ● The incoming cooling water is distributed to the individual cylinders via an inlet rail (in the image background). ● The liner of each cylinder is surrounded by a water jacket. ● In the lower region of the liner, the individual cylinders are connected to each other (below in the picture). ● Cooling water enters the individual cylinders (water jackets) from behind to the liner and there is a characteristic flow along the cylinder axis. ● Via branched pipes in the combustion chamber roof region (reddish in the picture), the coolant reaches the cylinder head. ● There is no connection between the individual cylinder heads. ● From the cylinder heads the cooling water flows into an outlet rail (top of picture) and leaves the cooling water jacket. Figure 7 also shows the temperatures close to the wall, based on the following thermal boundary conditions: ● Inlet and outlet rail are considered adiabatic ● The cylinder liner inside (greenish) with wall temperature 𝑇 and outside adiabatic ● Combustion chamber roof (branched pipes) with wall temperature 𝑇 ● Cylinder head inside at the inlet ducts with wall temperature 𝑇 batic (blue)

, outside adia-

● Cylinder head inside at the outlet ducts with wall temperature 𝑇 batic (blue)

, outside adia-

The coolant is "EGL-5050" with the temperature-dependent fluid properties according to [12]. To determine the Nusselt correlation, a variation is performed simulatively (3D-CFD) with respect to coolant inlet temperature 𝑇 and inlet mass flow 𝑚 (see Figure 9 top). Here, between minimum and maximum mass flows rate there is a factor of about 2, while the covered temperature range is about 50 K. At the operating points #13 and #14 the flow characteristic of the water jacket was changed in the following way:

597

Discretization and heat transfer calculation of engine water jackets in 1D simulation ● For OP#13 the flow connections between the liners were closed. ● Also for OP#14, in addition the input mass flow rate 𝑚 evenly to all cylinders.

is set distributed exactly

6.3 Discretization of Test Engine II The discretization of the large engine is presented below. The inlet and outlet rails are modelled according to the common method in the 1D simulation with pipes, pipe bends and branches (T-piece) and are regarded here as adiabatic. The focus is on the discretization of the heat transfer areas in the cylinder. Figure 8 shows the result of the discretization of a cylinder unit. Due to equality, only one cylinder needs to be discretized and the resulting 1D model is used for all cylinders.

Figure 8: Discretization of the cooling water jacket of a cylinder unit of the Test Engine II (abstract representation of the geometry).

As shown in Figure 8 on the right, the cylinder unit is essentially divided into three regions (from bottom to top): 1. Liner (light blue and blue): The liner region is divided into four 90° segments. The flow interfaces are taken into account so that each segment has an interface to the combustion chamber roof region. In the rear dark blue

598

Discretization and heat transfer calculation of engine water jackets in 1D simulation segment "L2" is the inlet to the cylinder unit, the two segments "L3" and "L4" in the foreground each have a flow interface to the adjacent liners. 2. Combustion chamber roof (the branched pipes in yellow, brown and grey): The branched pipes in the combustion chamber roof area are divided into three flow volumes with different flow interfaces. A more detailed discretization of the pipes, bends and branches is dispensed with in terms of model complexity. 3. Cylinder head (divided into inlet (blue) and outlet (red and magenta)): The water jacket in the cylinder head is divided into three volumes. "CHI" represents the region of the intake ports, "CHE1" and "CHE2" the exhaust ports, flowing from "CHE2" into the outlet rail. The coolant enters the cylinder head volumes from the combustion chamber roof. A main flow (see Figure 7) "CHI" → "CHE1" → "CHE2" occurs. In accordance with this discretization, the relevant parameters such as fluid temperature, heat input or mass flows are also evaluated in the 3D CFD.

6.4 1D Model and Calibration of Test Engine II Figure 6 shows the 3D geometry of the water jacket and the resulting 1D flow model. In the 1D model, the different regions of the cooling circuit are color-coded: Inlet rail (cyan), liner (blue), combustion chamber roof (yellow), cylinder head (orange) and outlet rail (red). The ten cylinder units are geometrically identical, whereby one unit consists of the following 1D elements: 10 general 1D volumes (4 liner, 3 combustion chamber roof, 3 cylinder head) and 20 flow interfaces (1 from inlet rail, 1 to outlet rail, 1 to adjacent liner, 17 internally between the 1D volumes). In general, the number of volumes should be kept as low as possible in terms of model complexity. For a conventional cooling water jacket of a passenger car engine, as in Test Engine I, six 1D volumes per cylinder are an acceptable upper limit: four volumes for the region around the liner, two for the cylinder head (1 intake, 1 exhaust). With Test Engine II (large engine), three volumes are added for the branched pipes in the combustion chamber roof region, as well as another volume to model the cylinder head geometry. This also results in a higher number of flow interfaces. To calibrate the heat input, the first step is to decide for which regions of the cooling water jacket a Nusselt correlation is to be determined. For Test Engine I, a correlation is derived for the regions engine block and cylinder head, since the thermal boundary conditions and flow characteristics differ in block and head.

599

Discretization and heat transfer calculation of engine water jackets in 1D simulation Since the thermal boundary conditions and the flow characteristics in the liner, combustion chamber roof and cylinder head regions also differ significantly on Test Engine II, three separate correlations are recommended.

6.5 Results of Test Engine II In the following, the results of the Test Engine II are presented. Analogous to Test Engine I, the relative deviation of the heat transfer between 1D model and the 3D CFD reference for the operating points (see Figure 9 top) is considered for model calibration and validation. Figure 9 shows the percentage deviation of the heat input for the entire water jacket (black) and all liners (blue), combustion chamber roof areas (yellow) and cylinder heads (red). For the calibration and validation points (with the exception of OP#1), the deviation between 1D and 3D simulation with respect to total heat input is in the range of ±2.5%. This model quality is also achieved for the combustion chamber roof and cylinder head regions.

Figure 9: Percentage deviation (1D - 3D CFD) of the heat input into the coolant. Deviation in the entire water jacket (black) and in the liner (blue), combustion chamber roof (yellow) and cylinder head (red) areas.

In the water jacket region of the liners, larger deviations of more than ±5% occur for some operating points and thus also represent the dominant part with regard to the

600

Discretization and heat transfer calculation of engine water jackets in 1D simulation deviation of the total heat input. For example, the calculated heat transfer in the liner region is 10% too low for validation points #8 and #9 in the 1D model. Especially at these operating points it should be noted that this is an extrapolation towards low inlet temperatures 𝑇 . The individual liners differ in their flow behavior. Particularly in the front cylinder units, the coolant flow enters the liner in varying proportions via the inlet rail and adjacent cylinder unit. On the other hand, the flow in the combustion chamber roof and cylinder head regions shows a more uniform picture. Figure 10 shows the relative deviation of the heat input for the liner (top), combustion chamber roof (center) and cylinder head (top) for the individual cylinder units resolved.

Figure 10: Percentage deviation (1D - 3D CFD) of the heat input into the coolant. Deviation of the individual cylinder units, divided into the regions liner (top), combustion chamber roof (center) and cylinder head (bottom). The thicker line represents the mean deviation.

In addition, the mean deviation of the corresponding area is shown as a thick line (corresponds to the result in Figure 9). The spreading width of the cylinder units is smallest for the combustion chamber roof region and is on average ∼5% and maximum ∼10%. For the liner and cylinder head regions, the average spreading width is ∼10% with some outliers up to 20%. The cylinder units #3 (in green) and #5 (in cyan) stand out especially in Figure 10. These two units also show different behaviour with regard to their inflow. Liner #2 has almost exclusively inflow from the inlet rail. Liner #3, on the other hand, is supplied in approximately equal proportions by the inlet rail and by unit #2. Liner #4

601

Discretization and heat transfer calculation of engine water jackets in 1D simulation behaves similarly to #2 (main part of inlet rail) and liner #5 similarly to #3, so the flow is approximately the same. In the rear half of the cylinder bank this changeable behaviour subsides and the units are supplied mainly via the inlet rail. This correlation between conspicuous deviations with regard to heat transfer and the different flow into the cylinder unit provides an approach for improving the model calibration. A simple approach would be to adjust the weighting of the respective flow interfaces in the front cylinder units. Thus, the model quality could be improved in the corresponding units and also with regard to total heat input without having to introduce or determine cylinder-dependent Nusselt correlations. The small range of deviation for the operating points #13 and #14 (flow interface between adjacent liners closed) also indicates that there is potential for improvement here. The accuracy of the 1D model with regard to total heat input is ±2.5% for the simulated operating points, as well as for the combustion chamber roof and cylinder head regions. In the region of the liner, a deviation of 10% may occur. The spreading width of the deviations for the individual cylinder units is acceptable for the cylinderindependent calibration used here. A starting point for improving the model calibration was presented. Overall, the methodology for simulating the heat transfer in the 1D cooling water jacket, which was developed on the basis of Test Engine I, can be transferred to a large engine.

7 Summary A practice-oriented universal method for the creation of a 1D water jacket model has been developed. The focus was on the discretization strategy of 3D geometry and the calculation of heat transfer using Nusselt correlations. Basis and reference are 3D CFD simulations of the water jacket. A discretization strategy has been presented on the basis of a passenger car engine with cylinders in cross flow of the coolant. This is based not only on the geometry of the cooling water jacket, but also on physical properties such as flow field and heat transfer. In addition, the later application scenario should be taken into account when creating the model. A too fine discretization has to be avoided, because this requires a higher number of boundary conditions and the model creation and calibration becomes too complex. The heat transfer calculation in the 1D simulation is based on Nusselt correlations (𝑁𝑢 = 𝑁𝑢(𝑅𝑒, 𝑃𝑟) = 𝑎 ⋅ 𝑅𝑒 ⋅ 𝑃𝑟 ) derived from 3D CFD simulations. By using a mass flow and temperature variation of the coolant, calibration over the entire engine operating map is possible. Furthermore, additional modifications in the calculation

602

Discretization and heat transfer calculation of engine water jackets in 1D simulation routines and helpful criteria make it possible to handle the specific properties of the complex water jacket geometry in a user-friendly way. The 1D results have been compared in detail with the 3D CFD reference and showed very good agreement. For different operating points with regard to fluid temperature and coolant mass flow, a model accuracy of ±5% with regard to heat input has been achieved. The methodology has been successfully validated on a large engine with single cylinder cooling and coolant flow along the cylinder axis. The investigation of two engines with different size ratios and flow characteristics of the water jacket has confirmed the general validity of the methodology.

Bibliography 1. J. Dohmen, R. Barthel and S. Klopstein, "Virtuelle Kühlsystementwicklung", MTZ, 12 2006. 2. Gamma Technologies, GT-SUITE Flow Theory Manual, 2018. 3. LMS Imagine, LMS AMESim Userguide, 2013. 4. VDI-Wärmeatlas, 11. Auflage, Berlin, Heidelberg: Springer Vieweg, 2013. 5. W. Kay, Convective Heat and Mass Transfer, New York: McGraw-Hill, 1993. 6. Incropera, DeWitt, Bergmann and Lavine, Fundamentals of Heat and Mass Transfer, John Wiley & Sons, Inc., 2007. 7. K. Robinson, J. G. Hawley, G. P. Hammond and N. J. Owen, "Convective coolant heat transfer in internal combustion engines," Automobile Engineering, 2002. 8. D. Ghebru, Modellierung und Analyse des instationären thermischen Verhaltens von Verbrennungsmotor und Gesamtfahrzeug, Dissertation: Karlsruher Institut für Technologie, 2013. 9. A. J. Torregrosa, P. Olmeda, J. Martín and C. Romero, "A Tool for Predicting the Thermal Performance of a Diesel Engine," Heat Transfer Engineering, 2011. 10. L. Fonseca, P. Olmeda, R. Novella and R. Molina Valle, "Internal Combustion Engine Heat Transfer and Wall Temperature Modeling: An Overview," Archives of Computational Methods in Engineering, 2019. 11. M. Kurreck, W. Remmels, M. Eckstein, O. Bücheler and V. Wachter, "Die neue Generation der MTU-Motorbaureihe 4000," MTZ, pp. 356-362, 05 2007. 12. Brines, ASHRAE Fundamentals Handbook, Chapter 20: Physical Properties of Secondary Coolants, 2001.

603

Discretization and heat transfer calculation of engine water jackets in 1D simulation

Acknowledgments The research project (FVV 1266) was performed by Institute for Internal Combustion Engines and Automotive Engineering (IVK) at University of Stuttgart under the direction of Prof. Dr.-Ing. Michael Bargende and by Institute of Mobile Systems (IMS) at Otto von Guericke University Magdeburg under the direction of Prof. Dr.-Ing. Hermann Rottengruber. Based on a decision taken by the German Bundestag, it was supported by the Federal Ministry for Economic Affairs and Energy (BMWi) and the AIF (German Federation of Industrial Research Associations eV) within the framework of the industrial collective research (IGF) programme (IGF No. 19378 BG). The project was conducted by an expert group led by Dipl.-Ing. Yann Drouvin, (Toyota Motorsport GmbH). The authors greatfully acknowledge the support received from the funding organisations, from the FVV (Research Association for Combustion Engines eV) and from all those involved in the project. Thanks also go to MTU Friedrichshafen GmbH and Volkswagen AG for providing data and Siemens Industry Software GmbH for supporting with software.

604

Optical investigations for the optimization and calibration of 3D-CFD injection models Simon Hummel, Antonino Vacca, Marc Reichenbacher, Karsten Müller, Andreas Kächele, Markus Koch, Michael Bargende IVK (University of Stuttgart), FKFS

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_51

605

Optical investigations for the optimization and calibration of 3D-CFD injection models

Abstract In this paper, a methodology is presented, which enables a calibration of 3D-CFD injection models by the use of optical investigations obtained from a constant volume chamber. Engine simulation is still progressing and as emission legislations and efficiency goals get more and more important, the level of detail in simulation must also rise. At such a detailed level, 3D-CFD is time-consuming and demands high computational resources. An approach to overcome this problem is the use of empirical models, which calculate specific phenomena based on fundamental research. Fuel injection is a phenomenon that is up to date not fully understood as a high number of very complex processes has to be considered to predict the atomization behavior of the spray in detail. Therefore, a direct numerical simulation of spray formation would not be feasible in the development of internal combustion engines. The 3D-CFD-tool QuickSim, which has been developed at FKFS, embeds empirical models in its conventional CFD code, which makes it very efficient in terms of computational time. The work focuses on a methodology for a direct comparison of measured and simulated properties to describe the spray. Those properties include the droplet diameters, velocities, the overall spray shape (targeting) and liquid penetration. All of those are vital parameters which determine the fuel distribution inside the combustion chamber, which is the basis for a well-engineered combustion strategy. By the comparison of simulation and measurement, the models can be calibrated for higher accuracy and hence better predictability of the results.

1 Introduction Injection modelling is one of the most critical points when simulating internal combustion engines due to the high number of influencing factors and the complex statistic nature of the injection event itself. The goal of injection modelling in the 3D-CFDtool QuickSim bonds the need of a good description of the spray together with a time effective integration for a simulation of several engine cycles [1]. A successful integration and calibration of the injection models call for a new methodology which considers input parameters coming from optical measurements. As the spray development itself is not directly simulated in all physical details, but instead represented by a model, one is forced to find parameters that describe several properties in a sufficient way. Those parameters can be of a macroscopic or of a microscopic nature; both of them are vital to predict mixture formation and combustion precisely.

606

Optical investigations for the optimization and calibration of 3D-CFD injection models Start of Procedure

Experiment: Spray Diagnostics

Measurement Data

Default Initialization Parameters Simulation: Constant Volume Chamber

Modified Initialization Parameters

Comparison

Sufficient Quality?

n

y Legend: Information

Simulation: Engine

Process

Final Results

Fig. 1: Model optimization procedure

Fig. 1 depicts a basic methodology for the entire optimization process. A base calibration of the injection model is performed by modeling the same constant volume chamber that is also used in the optical experiments. The first simulation is performed by choosing a default set of initialization parameters by experience. The results can be compared to the experiments and a modified set of initialization parameters is fed into the next simulation. After a few of those iterations, the model quality can be significantly improved and the final simulations can be conducted using the model inside the engine simulation with the respective geometries, combustion and flow field. For chapters 2-4, a solenoid 5-hole injector was investigated.

2 Experimental Setup Main component of the equipment for optical spray diagnostics is a constant volume chamber with optical access, which can be seen in the Fig. 2. The chamber is machined from massive stainless steel to avoid any corrosion with all possible test fluids. The optical accesses provide a transparent diameter of 100 mm. The chamber can be cleaned over a third access that can be unscrewed. The injector is mounted at the top of the chamber and protrudes slightly out of a conical-shaped roof. In the circumference of the injector tip, there is a circle of small holes providing a constant flow of nitrogen into the chamber with the aim to purge remaining droplets and fuel vapor out of the volume. Nitrogen is supplied from a gas bottle with a pressure regulator and is used due to its inert properties, which are preventing the chance of inflaming the injected fuel. The drain is seated at the bottom and leads the evacuated gases into the building’s suction system. The temperature of the chamber cannot be conditioned, hence the experiments were carried out at ambient temperature (≈25°C). As a surrogate

607

Optical investigations for the optimization and calibration of 3D-CFD injection models fuel for gasoline, n-Heptane is used as it shows similar results for the spray properties [2]. The fuel is pressurized by the use of a static nitrogen system. The injectors can be electronically actuated by standalone amplifiers, triggered by a synchronizing TTLsignal.

2.1 Imaging A very important tool of spray investigation is high speed photography. By recording high resolution video data, the spatial development of the spray clouds can be determined and evaluated. In engine application, the spatial distribution of the fuel within the combustion chamber is highly influencing engine performance, efficiency and emissions. Piston wetting is an example where the liquid penetration of the spray is too deep and hence droplets hit the metallic surface of the piston during operation (e.g. at very low in-cylinder densities). The used camera was adjusted to a setting of 256x192 pixels at 23000 fps, which appeared to be a good compromise between spatial and time resolution when performing evaluations on liquid phase penetration. Axial liquid phase penetration is the value of interest, followed by others such as spray angle, radial penetration or the tip velocity profile. Also, the hydraulic delay between electronic start of injection (rising TTLedge) and actual start of the liquid fuel jet can be determined by imaging. The spray is illuminated from the same side as the camera is located, hence the droplets scatter light back into the camera. Fig. 2 shows the application of the camera on the chamber with its optical access:

Fig. 2: Application of camera and illumination

608

Optical investigations for the optimization and calibration of 3D-CFD injection models To avoid reflections from walls, which would disturb image processing, a lightabsorbing curtain was placed behind the spray region.

Fig. 3: Raw spray images

Fig. 3 shows three raw images with different time stamps for one injection of a multihole injector. The scaling (mm/pixel) is determined by recording a metal scale piece of a known length. The processing of the images will be described later on.

2.2 Phase Doppler Anemometry Tests For the analysis of local phenomena in the spray, a PDA system is used. In particular, the velocity of droplets and the droplet diameter can be measured within a small measurement volume on different positions within the spray. The same optical chamber is used to perform the PDA experiments. The 2D-system used is made by TSI Inc. and uses a Coherent ArIon-Laser to supply the wavelengths of 488 nm and 514.5 nm. The signal processor (FSA4000) is able to measure two velocity components and perform a phase measurement for diameter determination by using three separate detectors for Channel 1. A photo detector module (PDM1000) converts the collected scattered light into voltage signals. The time base for each injection is also provided by a synchronizing TTL-signal. The connected software performs the whole system control, parametrization and data management. For a three axes system is used to move the chamber. The PDA transmitter and receiver optics remain on a fixed location and only the chamber is moved. The traverse system is connected to the PDA software via a serial port.

Fig. 4: Schematic of PDA-setup

609

Optical investigations for the optimization and calibration of 3D-CFD injection models The proper choice of spatial measurement locations is vital for several reasons. First aim is to achieve a good quality of data at all, as there are regions in the spray where measurement quality is lower or where measurements are not possible at all. When measuring too close to the nozzle exit, the spray is not fully broken up and contains ligaments, whereas the PDA-technique is focused on spherical droplets. Also, the droplet number per volume can be high enough, that several droplets pass the measurement volume at a time. This leads to signal patterns that cannot be evaluated by the processing algorithm.

Fig. 5: Measurement location within spray plume

Fig. 5 shows a fully developed spray plume in the coordinate system. A distance to the nozzle exit of 50-80% of the maximum liquid penetration appeared to be a good compromise, where the spray is fully broken up, not too dense and hence data rates are sufficient. When measuring at a distance downstream of 80% of the maximum liquid penetration, data rates decrease as too few droplets reach this region at all. For the measurements, the spray was traversed orthogonally to the plume’s axis at a z-coordinate of 30mm downstream the nozzle. Band pass filters were chosen to suit the feasible frequency spectra for this injector application. Low negative velocities should be detectable when measuring in vortex regions and higher positive velocities appear in the center of the jet. The setup shown above allows a velocity range of roughly -10…130 m/s. Parameters such as PMT (Photo multiplier tube) voltage, threshold voltage etc. were optimized for good data rates along with a good evaluation efficiency for the burst signals. In general, high data rates are not beneficial if most of the signals cannot be validated. Also, detector saturation must be avoided, as this leads to problems in burst center detection and hence coincidence criteria. Software coincidence was enabled to have both velocity components on each measured droplet. For one measurement position within the spray 7500 droplets were collected. This was used as a compromise between measurement time and a good statistical validity of the results. More measured droplets

610

Optical investigations for the optimization and calibration of 3D-CFD injection models would lead to a better statistical validity, but measurement time would increase proportionally or spatial resolution would have to be decreased to use the same time schedule for the measurements. As a rough estimate, the collection of data at one spatial point takes round about 90-180 seconds.

3 Numerical Description The numerical investigations within the presented work were accomplished with the 3D-CFD tool QuickSim, which has been developed and constantly improved at FKFS/IVK Stuttgart during the last years. It considers coarser meshes compared to traditional approaches, using ICE-adapted computational models, e.g. for the fuel, the working fluid description or the wall heat transfer process. The goal is the reduction of simulation time despite considering a fully integrated approach for the virtual engine development. This allows the extension of the simulation domain to a full engine (including intake and exhaust manifold), but also to the calculation of several successive operating cycles in a convenient time. Consequently a more stable flow field can be reproduced minimizing the influence of initial and boundary conditions [1], [3]. This in turn endorses to test many engine configurations considering the reciprocal influences of the different engine parameters. Consequently QuickSim uses a multiphase RANS approach where turbulence is reproduced employing the k-ԑ models. The combustion process and the flame propagation are based on the two-zone Wellermodel [4] adapted to GDI-combustion for spark ignition flame propagation. The jet atomization is a highly complex process due to the numerous influencing factors such as the nozzle geometry, the fuel specification, discharge velocity and turbulence within the jet along with pressure and temperature conditions [5]. The primary droplet breakup consists of the formation of ligaments and droplets from a coherent liquid phase, close to the injector’s orifice. It mainly depends on jet velocity and nozzle flow turbulence. The primary droplet breakup and the injector’s internal flow are not directly included in the injection model of QuickSim. To overcome this simplification literature recommends to thoroughly calibrate the secondary droplet breakup parameters as well as the initial droplet size, by means of optical spray measurements [6]. The employment of this approach has some limits such as the simulation of supercritical injection conditions, which result in a supersonic gas jet. Such a simulation would require a highly resolved computational mesh, having a cell discretization length that is even smaller than the cross-sectional width of the injector’s orifice [7]. With the aim of still capturing the fuel propagation within the combustion chamber, appropriate initialization parameters must be found. For the purpose of virtual engine development the injection simulation is integrated into the engine cycle simulation, meaning that it has to be combined with a RANS calculation method to be able to ensure a certain accuracy in reproducing the real phenomena within the flow field.

611

Optical investigations for the optimization and calibration of 3D-CFD injection models

Fig. 6: Mesh of constant volume chamber

In Fig. 6, the computational mesh of the constant volume chamber is shown. As said before, by combining 3D-CFD models with a dedicated formulation for engine-only processes. QuickSim does not require a structured grid refinement, but it can employ more coarse meshes. To reproduce the injection chamber only 250000 cells are used for the region where the fluid is injected, with a discretization length of 1.5 mm (total chamber volume 2.33 liters). A coordinate system is defined to describe the injector mounting position in QuickSim and the injector is not physically integrated into the mesh, except for its housing. The injector spray geometry is modeled on the basis of four characteristics angles, which define the direction of the spray jets for multi-hole injectors or the jet width for hollow-cone injectors. The combination of these parameters, together with SMD, the geometrical definition of the initialization region and the initial droplets location allow to initialize the desired sprays geometry in accordance to the spray targeting. These angles are shown in Fig. 7. Among these settings, the direction of the jets is defined by means of spherical coordinates, setting the position of the jet center axis. The zenith angle θ identifies the radial distance from the injector center axis, and the azimuth angle φ describes the jet orientation starting from the reference direction, given by x-axis. The angle α adjusts the width of the jets. For hollow-cone injectors the jet center axis coincides with the injector center axis z, thus zenith and azimuth angles are equal zero. One additional angle γ is used to reproduce hollow-cone sprays, while the angle α corresponds to half of the spray cone angle.

612

Optical investigations for the optimization and calibration of 3D-CFD injection models

Fig. 7: Geometric spray definition in QuickSim [8]

Since, the injector’s internal flow and the primary breakup are neglected, the droplets’ physical properties are initialized in a conical region which is defined through the parameters Lmin, Lmax and a characteristic angle, as shown in Fig. 8:

Fig. 8: Initialization domain of the spray in QuickSim [8]

A well-defined number of non-interacting droplets, with identical physical properties, is grouped into samples, called parcels. 1000 droplets per parcel are considered in the calculation, which appeared to be a good compromise in previous investigations for a sufficient model quality vs. calculation time. Moving on, the calculation framework in QuickSim consists of a blend formulation between Lagrangian and Eulerian approach, where the conservation equations of mass, momentum and energy for the dispersed phase are written for each individual parcel. A Rosin-Rammler function is used to initialize the diameter distribution, representing the SMD. This simplified injection model needs then as input position, size, velocity, direction of movement and temperature of each droplet cluster. The variation of these parameters, together with the

613

Optical investigations for the optimization and calibration of 3D-CFD injection models experimental data, are used to calibrate the penetration curve and SMD for the simulations. Once the droplets are generated into the computational mesh the interaction with the surrounding media is of fundamental importance to reproduce the jets progression, the air entrainment and thus the mixture formation. The break-up model adopted is that of Reitz and Diwakar. It considers two possible breakup regimes, the “Bag breakup”, where non-uniform pressure field around the droplet causes it to expand in the lowpressure wake region and eventually disintegrate when surface tension forces are overcome, and “Stripping breakup” in which liquid is sheared or stripped from the droplet surface [9]. One limit of the breakup model settings of the software is to neglect inter-droplet collisions and coalescence.

4 Analysis of Experiment and Simulation 4.1 Macroscopic Spray Parameters Alongside 3D-CFD simulations an imaging software was built and validated through an experimental test campaign at IVK/FKFS Stuttgart, where Mie-scattering and Shadow imaging can be processed in a similar way. The purpose was to create a repeatable methodology to analyze spray formation and propagation using the same tool to evaluate measurements and 3D-CFD simulations. Two different scripts run parallel over the measured and the simulated injection respectively. Both evaluate the level of luminance for the injection frames. The image processing via thresholding itself is known from literature, e.g. [10]. The injection electric signal is the camera trigger to start recording. The early frames without injection are detected and the background is stored to be subtracted from the frame where the injection happens. This mechanism is created in order to delete any source of light except for the reflecting fluid jet and to evaluate the injector’s hydraulic delay with respect to the electric signal. A scaling factor is automatically applied by recording a characteristic length in the injection chamber. The videos are transformed into gray scale images and then converted to binary using a variable thresholding algorithm (based on Otsu). The binary frames are collected into an array and for each matrix the boundary pixels are stored tracing the exterior jet contour. Each row of these matrices contains the row and column coordinates of a boundary pixel. The tool then automatically characterizes the jet for each injection timing computing axial penetration, tip speed and angles corresponding to maximum upper and lower vertical penetrations, as represented in Fig. 9.

614

Optical investigations for the optimization and calibration of 3D-CFD injection models

Fig. 9: Imaging tool applied to Mie-scattering with time flow analysis

Among the boundary regions detected, through numerical methods, the algorithm can individuate the main spray contour or the different jets composing the spray, depending on the injector used.

Penetration [mm] Angle [Deg]

Penetration [mm]

Exit Speed [m/s]

Once the experimental data have been collected and characterized, the calibration of the 3D-CFD simulation is realized by applying the same tool with a customized script to the simulation outputs also evaluating the main geometric parameters of the spray.

Fig. 10: Comparison of macroscopic spray parameters

615

Optical investigations for the optimization and calibration of 3D-CFD injection models With reference to Fig. 10, experimental results and simulations are compared considering the above mentioned geometric parameter and recursively the numerical analysis can be improved. In the right upper figure, the spray exceeds the boundary of the optical access, which is why the vertical penetration stops at a certain value. Finally a graphic analysis between numerical results (lower row) and measurements (upper row) is run to verify the reliability of the simulated jet shape as shown in Fig. 11. The first frame (0.1ms) is zoomed in for better visibility.

Fig. 11: Direct comparison of spray shape

Using this approach the respective injector models were calibrated in the 3D-CFD simulation environment, varying injection pressure (from 100 to 200 bar) and chamber back pressure (1 to 5 bar). When using this automated procedures the effort for the simulation calibration was reduced and the model predictability could be enhanced.

4.2 Microscopic Spray Parameters For a representative comparison of PDA and simulation data and a subsequent optimization of the 3D-CFD tool QuickSim, it is necessary to use suitable and identical measurement points. For this reason, a MATLAB tool was developed which can generate injector-specific measurement grids by entering the injector data. The generated measurement grids, with defined measuring point positions, can be used both for the evaluation of the simulation results and for the control of the traversing unit of the PDA system. With the help of another MATLAB tool developed at FKFS/IVK, the obtained PDA and simulation data can be analyzed separately and compared with each other. As a result, the simulation settings can be optimized based on the actual results measured with PDA technology. A linear measuring grid with a defined injector

616

Optical investigations for the optimization and calibration of 3D-CFD injection models distance orthogonal to the jet center axis is arranged to represent an entire jet crosssection. The center of the measurement grid was located on a z-coordinate of 30mm. After Calibration

Before Calibration

15

15

10 5 0

10 min_lab max_lab mean_lab QuickSim -5 0 5 Vertical distance to spray axis [mm]

min_lab max_lab mean_lab QuickSim

5

0

-6

-4 -2 0 2 4 Vertical distance to spray axis [mm]

6

Fig. 12: Comparison of droplet velocities and SMD in the cross-section line

Fig. 12 demonstrates the optimization of the simulation parameters regarding the initialized droplet velocity and the spray describing angles. The left column presents the comparison between experiment and simulation before adjusting the initialization values. The first initialized droplet velocity in the simulation is too slow and the spray cone is larger than in reality, which is shown by the red line in the upper left image. Furthermore, the jet center in the simulation is slightly offset inwards compared to the PDA results, despite the targeting given by the manufacturer being entered. After running iterative simulations with modified initialization parameters, the right column represents the improvements obtained. Obviously, the droplet velocity has now a good agreement with the measurements of PDA. At the same time, the jet centers of the simulation and PDA results are now very close and the spray cone angle of the simulation is much more stretched, similar to the experimental values. As a consequence of the increase of the droplet initialization velocity in the simulation parameters, a slight decrease of the SMD is caused, because of the stronger interaction of the droplets with the surrounding media, as it can be visible comparing the two lower diagrams in Fig. 12. The reduction of the SMD could normally be compensated by increasing the initialization diameter of the droplets. Nevertheless in QuickSim, there was no significant improvement in the resulting droplet diameter at the location of interest when increasing the initialized SMD, which probably makes a revision of the droplet breakup model necessary.

617

Optical investigations for the optimization and calibration of 3D-CFD injection models

5 Real Engine Simulation The purpose of the measurements carried out with the injection chamber was to characterize different injectors and to provide data for the calibration of the injection model into the 3D-CFD software QuickSim. On the other side, as said, this injection model aims to be integrated into a full engine simulation. Specifically the robustness of the model has to be tested in a real engine flow field, where the gas exchange processes through the valves, the piston stroke as well as the combustion event are considered. For this reason an already calibrated single-cylinder engine model is selected to test three different injector types against each other. The end of injection was kept constant as the injectors have slightly different flow rates. An overview is given in Tab. 1: Tab. 1: Injection conditions for simulated load point

Type Inj. pressure DOI EOI

Injector 1

Injector 2

Injector 3

Solenoid/5 holes 150 bar 62 deg 300 deg

Piezo/hollow cone 150 bar 55 deg 300 deg

Solenoid/5 holes 150 bar 72 deg 300 deg

The selected operating point was 2000rpm, 20 bar IMEP (max. torque) with retarded spark timing close to FTDC. The engine has a displacement of 333cm³. Fig. 13 shows the injection and evaporation behavior in comparison. It can be seen, that injector 3 has an advantage as the fuel evaporates earlier and more gradual during the compression stroke: Inl.Valve Exh.Valve. Ign & Comb.

55 50 45

Mass [mg]

40 35 30 25

Injected Evaporated

20

Injector 1

15

Injector 2

10

Injector 3

5 0 -360

-270

-180 Crank Angle [deg]

-90

Fig. 13: Fuel evaporation behavior

618

FTDC

Optical investigations for the optimization and calibration of 3D-CFD injection models Besides this, numerous important topics can be addressed of which only one will be shown in this work. The spatial λ-distribution within the cylinder is highly influencing the combustion event. At spark timing, it is important to have a lambda value close to 1.0 in the spark plug region. Fig. 14 depicts the comparison of the three injector types right before the spark timing. Obviously, Injector 3 is the optimum case as the inhomogeneity is lower compared to injector 1 and 2. Besides that, the region in the vicinity of the spark plug is less rich and closer to the optimum value of 1.0. The lean region is smaller, which leads to a faster overall combustion event. Despite of this benefit, the wider spray of injector 3 tends to cause liner wetting, which can be evaluated by extracting the fuel mass in the cells adjacent to the liner.

Fig. 14: Lambda distribution in the combustion chamber, 4°CA b. FTDC for injector 1 (a), 2 (b) and 3 (c)

The results show the opportunities that a well calibrated 3D-CFD-simulation offers when it comes to specific questions in engine development.

619

Optical investigations for the optimization and calibration of 3D-CFD injection models

6 Conclusion and Outlook In this work, an approach was presented to calibrate injection models for a fast 3DCFD-tool by the use of experimental spray investigations. The models were calibrated using a direct comparison between the experiments, performed in a constant volume chamber, vs. a simulation of exactly the same chamber geometry. After a recursive calibration procedure, the tuned models can be integrated into an engine model and are able to simulate the characteristics of different injectors for one engine. By doing so, an effective development process can be carried out, using all sorts of further variations such as combustion chamber shape, piston shape, valve timings or turbulence intensity. In further work, the models have to be improved to move onto an even better and more detailed representation of the injection event itself. This can be realized by the use of additional, specialized submodels for certain phenomena.

Bibliography 1. M. Chiodi: An Innovative 3D-CFD-Approach towards Virtual Development of Internal Combustion Engines, University of Stuttgart, Ph.D. Thesis, Springer, 2010. 2. G. Wigley, M.Mehdi, M. Williams, G. Pitcher, J. Helie: The effect of fuel properties on liquid breakup and atomisation in GDI sprays. In ICLASS, Kyoto, Japan, Paper ID ICLASS06-075, 2006. 3. M. Wentsch, A. Perrone, M. Chiodi, M. Bargende et al.: Enhanced Investigations of High-Performance SI-Engines by Means of 3D-CFD Simulations. In SAE Technical Paper Series, number 2015-24-2469, 2015, doi: 10.4271/2015-24-2469. 4. H.-G. Weller, S. Uslu, A.-D. Gosman, R.-R. Maly, R. Herweg, B. Heel: Prediction of Combustion in Homogeneous-Charge Spark-Ignition Engines. In International Symposium COMODIA 94. 5. R. D. Reitz: Atomization and other breakup regimes in a liquid jet. PhD thesis, Princeton University, 1978. 6. G. P. Merker, R. Teichmann: Grundlagen Verbrennungsmotoren. Springer Fachmedien Wiesbaden, 2014. 7. M. Baratta, A. E. Catania, E. Spessa, L. Herrmann, K. Roessler: MultiDimensional Modeling of Direct Natural-Gas Injection and Mixture Formation in a Stratified-Charge SI Engine with Centrally Mounted Injector. In SAE Technical Paper Series, number 2008-01-0975, 2008, doi: 10.4271/2008-01-0975.

620

Optical investigations for the optimization and calibration of 3D-CFD injection models 8. M. Wentsch: Analysis of Injection Processes in an Innovative 3D-CFD Tool for the Simulation of Internal Combustion Engines. University of Stuttgart, PhD thesis, Springer, 2019. 9. CD-adapco: Star-CD - Methodology, Commands, User guide v4.18, 2012. 10. D. Martin, P. Pischke, R. Kneer: Investigation of the influence of multiple gasoline direct injections on macroscopic spray quantities at different boundary conditions by means of visualization techniques. In Int. J. Engine Res. Vol. 11, 2010

621

Accelerated assessment of optimal fuel economy benchmarks for developing the next generation HEVs Pier Giuseppe Anselma, Giovanni Belingardi Department of Mechanical and Aerospace Engineering (DIMEAS) Center for Automotive Research and Sustainable Mobility (CARS) Politecnico di Torino, Torino, Italy

Accelerated assessment of optimal fuel economy benchmarks for developing …

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_52

623

Accelerated assessment of optimal fuel economy benchmarks for developing …

1 Introduction Hybrid electric vehicles (HEVs) are forecasted remarkably spreading in the global vehicle market over the next years due to their capability of reducing fuel consumption and tailpipe emissions while simultaneously tackling current limitations for on-board storable electrical energy [1]. Nevertheless, design procedures and sizing methodologies for HEV powertrains are consistently complicated compared to both battery electric vehicles (BEVs)that are powered by electric motors solely, and conventional vehicles that are powered by internal combustion engines (ICEs) solely [2]. Difficulties in HEV powertrain design and sizing relate both to the amount of power components embedded (i.e. one ICE, one or multiple electric motor/generators (MGs), a high-voltage battery, dedicated power electronics) and the necessity of a related proper energy management strategy (EMS) [3]. EMSs for HEVs can be classified into off-line and on-line ones. Off-line EMSs profit from the overall knowledge of preselected driving missions to optimize the HEV powertrain operation. They can be used to assess the fuel economy capability of HEV powertrain architectures and component sizes [4][5] or to calibrate and benchmark online EMSs [6][7][8]. On the other hand, on-line EMSs do not require the knowledge of future driving conditions and they can therefore be implemented in the on-board ECU of HEVs. This paper focuses on off-line EMSs for HEVs. In this framework, examples of popular optimization strategies for HEVs include Dynamic Programming (DP), the Pontryagin’s Minimum Principle (PMP) and the Power-weighted Efficiency-based Analysis for Rapid Sizing (PEARS) [9]. DP represents the most popular HEV optimization approach as it can return a global optimal solution, nevertheless it suffers from curse of dimensionality [10]. The PMP can approximate the global optimal solution provided by DP under few assumptions, however it does not always guarantee the global optimum and requires the calibration of the tuning factor for the electrical energy consumption [11]. The PEARS algorithm is computationally rapid and it can satisfy the charge-sustaining criterion without recurring to iterative calculation, however it exhibits non-uniform proximity with the global optimum and it can be applied to a limited set of HEV architectures represented by the multimode power split powertrains [12]. Therefore, there remains a need for a validated universal off-line control strategy suitable for rapid sizing of all types of HEV powertrains [13]. To answer the illustrated drawbacks, this paper aims at presenting a new off-line EMS for HEVs suitable for rapid sizing of hybrid powertrain components named Slopeweighted Energy-based Rapid Control Analysis (SERCA). This EMS has been developed aiming at achieving a good approximation of the optimal HEV fuel economy benchmark provided by DP while simultaneously reducing the computational cost by

624

Accelerated assessment of optimal fuel economy benchmarks for developing … around two orders of magnitude. Moreover, the SERCA algorithm can be easily implemented for different kind of HEV powertrain architectures including parallel, seriesparallel and power split. The rest of this paper is organized as follows: the HEV powertrain architectures under study are firstly presented and modeled. The off-line optimal control problem is then introduced for all the retained HEV architectures. DP is subsequently recalled as global optimal, yet computationally expensive, benchmark EMS. SERCA is then illustrated and implemented for all the HEV layouts. Simulation results and conclusions are finally given.

2 HEV powertrain architectures and modeling This section firstly aims at presenting the considered HEV powertrain architectures including parallel, series-parallel and power split layouts. Then, the adopted HEV modeling approach will be discussed.

2.1 HEV powertrain layouts 2.1.1 Parallel P2 Among the different hybrid architectures, parallel HEVs have been selected by many car manufacturers as their first step into vehicle electrification [14]. In a parallel HEV, the tractive power is combined: both the ICE and the MGs can contribute to the vehicle propulsion, i.e. their corresponding torques are additive. When the MG is large enough, it can drive the HEV by itself or simultaneously with the ICE. Particularly, the MG can be used to shift the ICE operating points to a higher-efficiency area by acting as a generator or a motor depending on the power demand being higher or lower. Particularly in the P2 architecture, one MG is placed between the ICE and the gearbox input and a clutch connection allows for eventually disengaging it from the ICE crankshaft. A possible scheme of the parallel P2 HEV layout is illustrated in Figure 1.

2.1.2 Series-Parallel P1P2 Series-parallel HEV powertrain architectures embed two MGs. Particularly for the P1P2 layout, other than the MG located in the P2 position, an additional MG is mounted directly on the ICE crankshaft as shown in Figure 2. When the clutch is engaged, ICE, MG1 and MG2 exhibits the same angular speed and the propelling torque can be arbitrarily distributed among these three power components. On the other hand, when the clutch is disengaged, the HEV powertrain can either operate in pure electric mode (i.e. ICE and MG1 are not activated) or in series mode (i.e. the ICE is turned on and MG1 serves as generator in providing electrical energy to the battery).

625

Gearbox

Differential

Clutch

Rear

Rear

MG

Front

Differential

Clutch

ICE

Figure 1: Parallel P2 HEV layout

ICE

MG1

MG2

Gearbox

Front

Accelerated assessment of optimal fuel economy benchmarks for developing …

Figure 2: Series-Parallel P1P2 HEV layout

2.1.3 Power split Power split HEV architectures are much successful and represent a large portion of the current population of commercially available full HEV powertrains. They consist of one or multiple planetary gear (PG) sets, which are very compact and can realize a continuously variable transmission. PG sets embed a ring gear, a sun gear and a carrier and they constitute the power split device (PSD), which is responsible for directing the power fluxes between the components of the hybrid powertrain [15]. Thanks to the PG kinematics, in a power split HEV the rotating speed of the ICE can be decoupled from the speed of the vehicle, thus enhancing fuel economy potential and flexibility in the operation. The power split HEV layout retained in this paper is illustrated in Figure 3 and refers to the well-known Toyota Hybrid System® [16]. In this HEV configuration, ICE, MG1 and output shaft are respectively linked to carrier, sun gear and ring gear of a PG set, while MG2 is directly linked to the output shaft through a reduction gearset.

Rear

Differential

ICE

MG1

Front

PG

MG2

Figure 3: Power split HEV layout

2.2 HEV powertrain modeling Overall, the HEV powertrain architectures considered in this paper are modelled adopting a backward quasi-static approach [17]. In this modeling procedure, required speed, torque and power values for the hybrid powertrain components are directly derived from the vehicle speed profile defined in the retained driving mission. In other words, the actual vehicle speed always matches the targeted profile of the driving mission. The power components (i.e. ICE and MGs) are taken into account through their empirical

626

Accelerated assessment of optimal fuel economy benchmarks for developing … operational maps with torque and speed as independent variables, while an internal resistance model is adopted for the battery. Here, values for voltage and internal resistance are assumed to be constant and independent from the battery state-of-charge (SoC), as it has been demonstrated that it is still possible to achieve a globally optimal solution with this hypothesis for charge-sustaining (CS) problems [18].With regards to the vehicle, road resistance forces are evaluated using experimental road load coefficients. More details about the powertrain and vehicle model can be found in [4].

3 Off-line HEV control In this section, the HEV off-line control problem is presented that is typically considered in HEV powertrain design and sizing processes and in development and calibration procedures of on-line EMSs. Subsequently, DP is recalled as universally accepted numerical tool to return the global optimal solution for the HEV off-line control problem.

3.1 Control problem The optimal off-line control problem for an HEV aims here at minimizing a multi-target cost function 𝐽 that considers, among other terms, estimated fuel consumption (EFC) and number of ICE activations over a certain period as example. The resulting mathematical formulation is stated in (1): min 𝐽 =

𝐿 (𝑡)𝑑𝑡

subject to: 𝑆𝑜𝐶(𝑡 ) = 𝑆𝑜𝐶(𝑡 𝑆𝑂𝐶 = 𝑓(𝑆𝑂𝐶, 𝜔 𝜔

,𝑇

) ,𝜔

≤𝜔

≤𝜔

𝜔

≤𝜔

≤𝜔

𝜔

≤𝜔

≤𝜔

≤𝑇

≤𝑇

𝑇

≤𝑇

≤𝑇

𝑇

≤𝑇

≤𝑇

𝑃

≤𝑃

≤𝑃

𝑇

𝑆𝑜𝐶

,𝑇

)

(1)

≤ 𝑆𝑜𝐶 ≤ 𝑆𝑜𝐶

627

Accelerated assessment of optimal fuel economy benchmarks for developing … Where 𝐿 (𝑡) represents the instantaneous cost function which needs specific definition depending on the retained HEV powertrain architecture Pi as it will be detailed later in this section. Charge-sustaining (CS) criteria is defined by imposing equivalent battery SoC values at the beginning and the end of the considered time period. Finally, speed, torque and power of power components (battery, ICE, MG1 and MG2 as well where applicable) are restricted within the correspondent actual operating regions. In the follow-up of this section, DP will be illustrated as numerical approach to solve this control problem.

3.2 Dynamic Programming DP is by far the most commonly adopted approach to solve the HEV optimal control problem. It involves generating a globally optimal solution backward along a time horizon by searching through all feasible discrete control actions for all the state grid points [19]. While DP is demonstrated achieving global optimality under a wide range of operating conditions, its major drawback refers to the computational power and computational time needed for exhaustively searching through all the possible solutions [20]. Specifically considered control variables, state variables and cost functions for the retained HEV powertrain layouts will be illustrated below.

3.2.1 Parallel P2 When controlling a parallel P2 HEV, three levels of decision need accomplishment at each time instant [7]: 1

Which gear is to be engaged in the gearbox;

2

Whether to propel the vehicle in pure electric mode (MG only) or in hybrid operation (ICE+MG);

3

In case the hybrid mode is selected, how to split the required torque between ICE and MG.

This leads to embed the two terms illustrated in (2) for the control variable 𝑈 and the value of ICE torque 𝑇 . ing the gear number # 𝑈

=

# 𝑇

includ-

(2)

In a backward HEV modelling approach, pure electric or hybrid operation are particularly distinguished by the sign of the ICE torque being null or positive, respectively. As regards the state variable 𝑋 , its formulation considered here is reported in (3).

628

Accelerated assessment of optimal fuel economy benchmarks for developing … 𝑆𝑜𝐶 = 𝐼𝐶𝐸 / #

𝑋

(3)

The battery SoC is retained in order to achieve CS operation, while a binary term defining the ICE state (i.e. on/off) and the engaged gear number are considered in order to account for comfort and smooth HEV operation. Particularly, the optimal control solution identified by DP throughout the driving mission should avoid an excessive number of ICE de/activation and gear shifting events. This is performed by formulating the instantaneous cost function (whose integrated value over the driving mission needs minimization) 𝐿 as follows: 𝐿

= 𝑚

+ 𝛼 ∙ 𝐼𝐶𝐸

+ 𝛼 ∙ 𝑔𝑒𝑎𝑟

(4)

Where 𝑚 represents the instantaneous rate of fuel consumption as given by the ICE and 𝑔𝑒𝑎𝑟 denote ICE activation and gear shifting events, fuel table, while 𝐼𝐶𝐸 respectively, which can be detected by means of the corresponding state terms. 𝛼 and 𝛼 are constant weighting factors.

3.2.2 Series-Parallel P1P2 The required control decisions for a series-parallel P1P2 layout are similar to the ones related to the parallel P2 HEV architecture, yet few additions need to be made. Obtained levels of decisions are reported as follows: 1

Which gear is to be engaged in the gearbox;

2

Whether to keep the clutch engaged or disengaged;

3

In case the clutch is engaged, how to split the required torque between ICE, MG1 and MG2;

4

In case the clutch is disengaged, whether to operate in pure electric mode (i.e. only MG2 is activated) or in series mode (ICE and MG1 are activated as well);

5

In case the series hybrid operation is selected, which values of speed and torque assign to ICE and MG1.

requires additional terms to handle As a result, the corresponding control variable 𝑈 the increased control complexity for this HEV layout, namely the MG1 torque 𝑇 , the ICE speed 𝜔 and the binary clutch status 𝐶𝑙 / (i.e. engaged or disengaged):

629

Accelerated assessment of optimal fuel economy benchmarks for developing …

𝑈

# 𝑇 𝑇 = ⎨ 𝜔 ⎪ ⎩𝐶𝑙 ⎧ ⎪

⎫ ⎪

/

⎬ ⎪ ⎭

(5)

As concerns state variable 𝑋 and cost function 𝐿 , their formulations for the series-parallel P1P2 HEV architecture equal their counterparts for the parallel P2 layout.

3.2.3 Power split Power split HEV powertrain layouts exhibit different terms in their control variables compared to parallel and series-parallel architectures. The control variable 𝑈 considered in this paper for the power split HEV layout contains speed and torque values for the ICE and it is formulated in (6). 𝜔 (6) 𝑈 = 𝑇 In a backward modeling approach, values of speed and torques for both the MGs can indeed be evaluated starting from the control variable terms following the planetary gear kinematics and dynamics [15]. In the considered power split HEV layout, the embedment of a gearbox is not strictly necessary in the hybrid transmission since the capability of PG sets of operating as an electrically variable transmission (eVT). As consequence, the state variable 𝑋 and the cost function 𝐿 for the power splti HEV architecture can be simplified as in (7) and (8), respectively. 𝑋

𝑆𝑜𝐶 = 𝐼𝐶𝐸 /

(7)

𝐿

= 𝑚

(8)

+ 𝛼 ∙ 𝐼𝐶𝐸

4 Slope-weighted Energy-based Rapid Control Analysis (SERCA) In this section, the Slope-weighted Energy-based Rapid Control Analysis (SERCA) is described as a novel approach for the HEV off-line optimal control problem. This methodology can be divided in three phases, as illustrated in Fig. 4: the division into subproblems, the definition of the generalized optimal operating points and the energy balance realization process [21][22].

630

Accelerated assessment of optimal fuel economy benchmarks for developing … Step A: Sub-problems exploration Step B: Definition of optimal operating hulls Step C. Energy balance realization process Outcome: - Estimated fuel consumption - Time series of HEV control and state variables over the driving mission

Figure 4: Workflow of SERCA

4.1 Sub-problems exploration The first step of SERCA aims at exploring the possible solutions of each sub-problem, particularly represented by the single time point of the retained driving mission. The sub-problems are characterized with the specific values of current vehicle speed and desired acceleration, respectively. Similar to DP, discretized arrays for the control variable terms are firstly created. Each possible control sub-solution for the specific subproblem is thus represented by a certain combination of control term values. Following the backward HEV modeling approach, values for EFC and variation in the battery SoC can then be assessed for each feasible sub-solution.

4.2 Definition of optimal operating hulls Once all the possible sub-solutions are identified for a specific subproblem (i.e., a time point of a target driving mission), they can be assessed based on EFC and battery SoC depletion. Examples of sub-solution comparisons for the same sub-problem corresponding to a current vehicle speed value of 35 km/h and a requested vehicle acceleration from the driver of 0.1 m/s2 are illustrated in Fig. 5, Fig. 6 and Fig. 7 for the parallel, the series-parallel and the power split HEV layouts, respectively. In all the three figures, a positive value of SoC depletion means that the battery is providing energy to power the vehicle. This corresponds both to pure electric operation and to hybrid operation in case the ICE is not providing enough power to propel the vehicle by itself. On the other hand, a negative value of battery depletion means that the ICE is providing more power than the amount needed to satisfy the power demand coming from the algebraic sum of vehicle resistance forces and inertia load related to the requested vehicle acceleration. In this case, the excess ICE power can be used to charge the high-voltage battery.

631

Depleted SoC [%]

Accelerated assessment of optimal fuel economy benchmarks for developing …

Depleted SoC [%]

Figure 5: Sub-solutions comparison example for a parallel P2 HEV layout (vehicle speed = 35 km/h, vehicle acceleration = 0.1 m/s2)

Depleted SoC [%]

Figure 6: Sub-solutions comparison example for a series-parallel P1P2 HEV layout (vehicle speed = 35 km/h, vehicle acceleration = 0.1 m/s2)

Figure 7: Sub-solutions comparison example for a power split HEV layout (vehicle speed = 35 km/h, vehicle acceleration = 0.1 m/s2)

632

Accelerated assessment of optimal fuel economy benchmarks for developing … The general descending trend of the point cloud reminds how battery recharging can be achieved through the gradual increase of fuel consumption. The shape of the cloud of points for the hybrid electric sub-solutions differs according to the specifically retained HEV powertrain layout as shown in Fig. 5, Fig. 6 and Fig. 7, respectively. As example, for the P2 HEV layout in Fig. 5, a single curve can be observed for each feasible gear number that can be traced by varying the value set to the ICE torque. On the other hand, for series parallel P1P2 and power split architectures in Fig. 6 and Fig. 7 respectively, a cloud of points can be recognized rather than single curves. This is due to the additional degree of freedom related to the capability of varying the speed of the ICE. This representation can be interpreted as a sort of Pareto frontier for all the feasible subsolutions of the HEV powertrain in the considered sub-problem. The sub-solutions at the lower edge of the point cloud thus correspond to the optimal ones, as they exhibit the highest ratio between charged battery energy and correspondently consumed fuel. As consequence, these sub-solutions should be considered for eventual hybrid operation in an attempt of reaching the global optimal solution in a considered driving mission. A discrete operating hull is therefore stored for each sub-problem that is represented by the optimal sub-solutions of the Pareto frontier. Then, the slope between two adjacent points (k-1) and k of the optimal hull is defined as θ in (9). 𝜃(𝑘 − 1, 𝑘) =

=

( ) ( )

(

) (

)

(9)

After the optimal operating hull is identified and stored and the slope for each optimal sub-solution is computed for all the sub-problems of the considered driving mission, the energy balance realization process can be performed.

4.3 Energy balance realization process The last stage of SERCA aims at efficiently solving the optimal HEV off-line control problem for the overall considered driving mission. First it is assumed that, when feasible, the HEV powertrain operates all the time points in pure electric mode. Particularly, in the Pareto frontiers of Fig. 5, Fig. 6 or Fig. 7 the pure electric point with the lowest depleted SoC value is considered and the hybrid powertrain is set to operate according to the corresponding control variables in the considered sub-problem. The total required electrical energy EEV is then obtained by summing the depleted (or charged) battery energy in each point where pure electric mode is selected. Consequently, the time point i exhibiting the highest value of slope ( |θi| =|θMAX| ) is selected for hybrid operation. The corresponding control variables are set to operate in

633

Accelerated assessment of optimal fuel economy benchmarks for developing … the identified time point. Then, the variables related to the overall driving mission operation are updated in (10). E

=E

m

_

+ 𝑆𝑜𝐶 =m

_

(10)

+m

Particularly, the value of required electrical energy needed is reduced by the charged battery energy in correspondence with the selected point i. Meanwhile, the global fuel consumption mfuel_TOT is increased with the increment provided by the selected hybrid operating point. The electric-to-hybrid operation replacement is carried out iteratively until the value of overall electrical energy consumed in the retained driving mission EEV becomes null or negative. Finally, the corresponding EFC and the hybrid powertrain operation for the considered driving mission can be extrapolated in this way.

5 Simulation results The SERCA algorithm aims at achieving optimality for the HEV off-line control problem solution and simultaneously reducing the associated computational cost. In this section, several driving missions are considered to evaluate the performance of SERCA when applied to various HEV powertrain architectures while benchmarking it with the well-known DP approach. Table 1 illustrates the vehicle and powertrain data considered in this paper. In general, vehicle and battery data for a full HEV model have been retained from Amesim® software, while lookup tables for power components and battery have been derived from [23] and scaled appropriately in order to get an hybridization factor of around 0.45 for all the three HEV considered architectures [24]. Table 1. Vehicle and powertrain data Component

Vehicle

ICE Transmission (P1 and P1P2 layouts)

634

Parameter Mass Wheel dynamic radius Road Load coefficient A Road Load coefficient B Road Load coefficient C Capacity Maximum power Maximum torque Gear ratios Final drive ratio

Value 1000 Kg 0.317 m 100 N 5 N/(m/s) 0.5 N/(m/s)^2 1.2 l 89 kW @ 4000 rpm 230 Nm @ 2000 rpm [3.7; 2; 1.5; 1; 0.8] 3

Accelerated assessment of optimal fuel economy benchmarks for developing … Transmission (power split layout) MG (P2 layout) MG1 (P1P2 and power split layouts) MG2 (P1P2 and power split layouts) Battery

PG ratio (ring/sun) MG2 to output shaft ratio Final drive ratio Maximum power Maximum torque Maximum power Maximum torque Maximum power Maximum torque Capacity Open-circuit voltage Internal resistance Temperature

3.27 1.26 3.27 72 kW 240 Nm 22 kW 74 Nm 50 kW 167 Nm 19 Ah 123.62 V 54.54 mΩ 25 °

Driving missions simulated here particularly refer to the Urban Dynamometer Driving Schedule (UDDS), the Highway Fuel Economy Test (HWFET), the Worldwide Harmonized Light Vehicles Test Procedure (WLTP) and the New European Driving Cycle (NEDC). All the reported computational times (CTs) refer to a desktop computer with Intel Core i7-8700 (3.2 GHz) and 32 GB of RAM. In all the simulations, a CS operation has been simulated by imposing equal battery SoC values at the beginning and the end of the driving missions. Table 2 and Table 3 report obtained simulation results for all the retained HEV powertrain architectures focusing on the EFC and the CT, respectively. Concerning EFC, the SERCA algorithm is demonstrated capable of predicting fuel economy results close to the DP global optimal benchmark over different driving missions. Considering the UDDS as example, the increase in the EFC value obtained by SERCA is limited within 0.76 %, 0.13 % and 0.85 % for the P2, P1P2 and power split hybrid powertrain architectures, respectively. On the other hand, looking at CTs, the SERCA algorithm is proven achieving remarkable savings compared to the DP benchmark. In the UDDS case as example, CTs required to perform a simulation using SERCA only represent the 1.46 %, 0.09 % and the 14.05 % of the DP counterpart for the P2, P1P2 and power split hybrid powertrain layouts, respectively, while giving comparable values for the EFC. In this framework, the SERCA reveals more efficient compared to the current state-ofart. The objective realization of the CS HEV operation particularly allows avoiding recursive calculation, thus suggesting the successful implementation of SERCA for effective rapid sizing of hybrid powertrain architectures.

635

Accelerated assessment of optimal fuel economy benchmarks for developing … Table 2. Simulation results for EFC P2 Driving mission UDDS HWFET WLTP NEDC

P1P2

Power split

DP

SERCA

DP

SERCA

DP

SERCA

293.25 g 581.43 g 723.47 g 308.66 g

295.49 g 587.21 g 723.95 g 307.69 g

234.99 g 546.95 g 757.74 g 287.55 g

235.29 g 551.58 g 766.68 g 287.55 g

243.22 g 554.62 g 771.54 g 290.12 g

245.28 g 557.17 g 779.26 g 291.62 g

Table 3. Simulation results for CT P2 Driving mission UDDS HWFET WLTP NEDC

P1P2

Power split

DP

SERCA

DP

SERCA

DP

SERCA

274 s 115 s 184 s 188 s

4s 2s 7s 3s

22260 s 13179 s 30506 s 24100 s

20 s 17 s 29 s 14 s

747 s 450 s 4584 s 3123 s

105 s 69 s 289 s 199 s

6 Conclusions This paper presents the application of a novel rapid near-optimal EMS named slopeweighted energy-based rapid control analysis (SERCA) to various HEV powertrains including parallel, series-parallel and power split layouts. The operating steps of SERCA have been detailed, particularly the division into sub-problems, the construction of the generalized optimal operating hulls and the energy balance realization process. The SERCA addresses the problem of effective rapid component sizing for HEV powertrains. The illustrated energy management strategy is validated based on a comparison of the resulting SERCA EFC values with the globally optimal solution provided by DP over several driving missions. Results for several different driving missions particularly reveal a narrow difference contained within 0.99 %, 1.18 % and 1.00 % at maximum for the parallel P2, the series parallel P1P2 and the power split HEV powertrain architecture, respectively. Moreover, the SERCA algorithm is demonstrated achieving remarkable computational rapidness compared to DP. Future work may consider the implementation of the SERCA in a design methodology for rapid component sizing of various HEV powertrains. Finally, an on-line energy management strategy may be developed based on the SERCA and implemented in an on-board control logic. For instance, offline SERCA optimization may be considered

636

Accelerated assessment of optimal fuel economy benchmarks for developing … to derive optimal control policies [25]. Alternatively, SERCA may rapidly provide near-optimal benchmarks for recently developed artificial intelligence-based on-line HEV EMSs [26].

Bibliography 1. B. Bilgin et al., "Making the Case for Electrified Transportation," in IEEE Transactions on Transportation Electrification, vol. 1, no. 1, pp. 4-17, 2015. 2. P.G. Anselma, G. Belingardi, “Comparing battery electric vehicle powertrains through rapid component sizing”, Int. J. Electric and Hybrid Vehicles, vol.11, no.1, pp. 36-58, 2019. 3. A. Biswas, A. Emadi, "Energy Management Systems for Electrified Powertrains: State-of-the-Art Review and Future Trends," in IEEE Transactions on Vehicular Technology, vol. 68, no. 7, pp. 6453-6467, July 2019. 4. P.G. Anselma, Y. Huo, J. Roeleveld, A. Emadi, G. Belingardi, “Rapid Optimal Design of a Multimode Power Split Hybrid Electric Vehicle Transmission”, Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, vol. 233, no. 3, pp. 740-762, 2019. 5. P.G. Anselma, G. Belingardi, A. Falai, C. Maino, F. Miretti, D. Misul, E. Spessa, “Comparing Parallel Hybrid Electric Vehicle Powertrains for Real-world Driving”, 2019 AEIT International Conference of Electrical and Electronic Technologies for Automotive, Torino, Italy, 2019, pp. 1-6. 6. P.G. Anselma, Y. Huo, J. Roeleveld, G. Belingardi, A. Emadi, “Real-time rule-based near-optimal control strategy for a single-motor multimode hybrid electric vehicle powertrain,” 2018 FISITA world congress, Chennai, India, Oct. 2018, pp. 1-14. 7. P. G. Anselma, A. Biswas, J. Roeleveld, G. Belingardi and A. Emadi, "Multi-Fidelity Near-Optimal on-Line Control of a Parallel Hybrid Electric Vehicle Powertrain," 2019 IEEE Transportation Electrification Conference and Expo (ITEC), Detroit, MI, USA, 2019, pp. 1-6. 8. P.G. Anselma, Y. Huo, J. Roeleveld, G. Belingardi, A. Emadi, “From Off-Line to On-Line Control of a Multimode Power Split Hybrid Electric Vehicle Powertrain”, IFAC-PapersOnLine, vol. 52, no. 5, pp. 141-146., 2019. 9. G. Belingardi, P.G. Anselma, M. Demic, “Optimization-based controllers for hybrid electric vehicles”, in Mobility & Vehicle Mechanics, vol. 44, no. 3, pp. 53-67, 2018.

637

Accelerated assessment of optimal fuel economy benchmarks for developing … 10. J. Lempert, B. Vadala, K. Arshad-Aliy, J. Roeleveld and A. Emadi, "Practical Considerations for the Implementation of Dynamic Programming for HEV Powertrains," 2018 IEEE Transportation Electrification Conference and Expo (ITEC), Long Beach, CA, 2018, pp. 755-760. 11. Kim, N., Rousseau, A.: “Sufficient conditions of optimal control based on Pontryagin’s minimum principle for use in hybrid electric vehicles”, Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Automobile Engineering, Vol 226, Issue 9, 2012, pp 1160 – 1170. 12. P.G. Anselma, Y. Huo, E. Amin, J. Roeleveld, A. Emadi, G. Belingardi, “Modeshifting Minimization in a Power Management Strategy for Rapid Component Sizing of Multimode Power Split Hybrid Vehicles,” SAE Technical Paper 2018-011018, 2018. 13. P.G. Anselma, G. Belingardi, “Next generation HEV powertrain design tools: roadmap and challenges,” SAE Technical Paper 2019-01-2602, 2019. 14. Yang, Y., Ali, K. A., Roeleveld, J., Emadi, A., “State-of-the-art electrified powertrains - hybrid, plug-in, and electric vehicles”, in International Journal of Powertrains, vol. 5, no.1, pp.1-29, 2016. 15. Schulz M., “Circulating mechanical power in a power split hybrid electric vehicle transmission”, Proc IMeche, Part D: J Automobile Engineering, 2004; 218: 1419– 1425. 16. Matsumura, M., Shiozaki, K., and Mori, N., "Development of New Hybrid Transaxle for Mid - Size Vehicle," SAE Technical Paper 2018-01-0429, 2018. 17. L. Guzzella, A. Amstutz, "CAE tools for quasi-static modeling and optimization of hybrid powertrains", IEEE Transactions on Vehicular Technology 1999; 48(6): 1762-69. 18. N. Kim, S. Cha, and H. Peng, “Optimal Control of Hybrid Electric Vehicles Based on Pontryagin's Minimum Principle,” IEEE Transactions on Control Systems Technology, vol. 19, no. 5, pp. 1279-1287, Sept. 2011. 19. O. Sundstrom and L. Guzzella, "A generic dynamic programming Matlab function," 2009 IEEE Control Applications, (CCA) & Intelligent Control, (ISIC), St. Petersburg, 2009, pp. 1625-1630. 20. J. Lempert, B. Vadala, K. Arshad-Aliy, J. Roeleveld and A. Emadi, "Practical Considerations for the Implementation of Dynamic Programming for HEV Powertrains," 2018 IEEE Transportation Electrification Conference and Expo (ITEC), Long Beach, CA, 2018, pp. 755-760.

638

Accelerated assessment of optimal fuel economy benchmarks for developing … 21. P. G. Anselma, Y. Huo, J. Roeleveld, G. Belingardi and A. Emadi, "Slope-Weighted Energy-Based Rapid Control Analysis for Hybrid Electric Vehicles," in IEEE Transactions on Vehicular Technology, vol. 68, no. 5, pp. 4458-4466, May 2019. 22. G. Buccoliero, P. G. Anselma, S. Amirfarhangi Bonab, G. Belingardi and A. Emadi, "A New Energy Management Strategy for Multimode Power Split Hybrid Electric Vehicles," in IEEE Transactions on Vehicular Technology, in press, 2020. 23. Dabadie, J., Sciarretta, A., Font, G., and Le Berr, F., "Automatic Generation of Online Optimal Energy Management Strategies for Hybrid Powertrain Simulation," SAE Technical Paper 2017-24-0173, 2017. 24. J. M. Tyrus, R. M. Long, M. Kramskaya, Y. Fertman and A. Emadi, "Hybrid electric sport utility vehicles", IEEE Transactions on Vehicular Technology 2004; 53(5): 1607-22. 25. P. G. Anselma, Y. Huo, J. Roeleveld, G. Belingardi and A. Emadi, "Integration of On-Line Control in Optimal Design of Multimode Power-Split Hybrid Electric Vehicle Powertrains," in IEEE Transactions on Vehicular Technology, vol. 68, no. 4, pp. 3436-3445, April 2019. 26. A. Biswas, P. G. Anselma and A. Emadi, "Real-Time Optimal Energy Management of Electrified Powertrains with Reinforcement Learning," 2019 IEEE Transportation Electrification Conference and Expo (ITEC), Detroit, MI, USA, 2019, pp. 1-6.

639

Front loading approach in battery development for generation update Nenad Dejanovic, M.Sc.Eng., Product Manager Batteryvvv DI Paul Schiffbänker, Product Manager Electrification AVL List GmbH, Hans-List-Platz 1, 8020 Graz, Austria

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_53

641

Front loading approach in battery development for generation update

Abstract Automotive OEMs are facing many challenges today. One of them is to rapidly increase within the next couple of years, the portfolio of their electric vehicles, and to do this, in the shortest possible development time and with the minimum occurred costs. Considering the battery as one of the critical components within the electric powertrain, the solution is to approach battery development for new electric vehicles based on an already existing battery pack project. The battery development process at AVL (AVL GBDP – Global Battery Development Process) is implemented within a project organization structure, having defined quality-gates and time triggered milestones, with defined activities and deliverables that will be synchronized, checked and reported at each quality gate and milestone review. Having the GBDP, as the basis of a lead battery development, this paper will provide an analysis of how front-loading can support the development of derivate battery packs during a generation update. Keywords: Battery Derivate Development, Validation, Simulation, V-process

1 Generation update: What is a battery derivate? One of the many challenges the automotive OEMs are facing today is to rapidly increase within the next couple of years, the portfolio of their electric vehicles and to do this in the shortest possible development time and with the minimum occurred costs. If we look at the battery as one of the critical components within the electric powertrain, from the point of cost, performances, safety, producibility and sustainability, the solution is to approach the battery development based on an already existing battery pack project. Such a battery pack is then considered a lead battery, while the new battery developed is a battery derivate. Different scenarios could be the basis for a derivate battery development. For instance, change of sales regions that would lead to different homologation requirements; utilization of new battery cell formats or cell chemistries that could impose new safety, design or performance considerations, or possibly integration of the existing battery pack into a new to be developed powertrain/or vehicle, requiring new mechanical, electrical and thermal interfacing requirements.

642

Front loading approach in battery development for generation update

2 AVL Battery Derivate Development process 2.1 Steps in a derivate battery project The following diagram, (see Figure 1), shows the steps taken at AVL during a battery derivate development project having the lead battery project as the basis.

Figure 1: Diagram of steps taken during a battery derivate development project

Review of all base project documentation is the first step in the derivate development project. The scope of documentation that is reviewed includes: requirements and technical specifications, 2D/3D datasets, safety documentation, quality documents, validation reports, supplier documentation and assembly documentation. Delta assessment will then follow to identify: ● Deltas on “top-level” requirements and boundaries - assessing changes in targets, changes in localization (homologation requirements) and changes by use of new components. ● Deltas on performance requirements, such as: battery pack power/current in charge and discharge and the energy capacity. ● Deltas on design and integration – assessing changes in DMU (geometry and weight), thermal design, structure design, electric design, tightness and pressure release and their impact on part/component interfacing and integration.

643

Front loading approach in battery development for generation update Integration impact assessment is performed by analysing the battery pack considering its Subsystems (SS) (see Figure 2), as well as from the point of its main Functions Groups (FG) (see Figure 3)

Figure 2: Battery pack Subsystem Energy capacity & Power delivery Store energy Charge power Discharge power

LV functions

Active conditioning

EMC

El. Isolations Current carrying performance

Cooling Heating

EMC emissions EMC immunity

BMS control functions

Ambient robustness

Lifetime endurance

Safety functions

Cell monitoring / pack monitoring SoX estimation and energy management Power delivery Contactor control Isolation monitoring Thermal conditioning control Identification and Communication Operating mode Usage logging Fault management and diagnostics

Ambient temperature Humidity Pressure / Altitude

Operation time Mileage Mission profile

Crash safety HV safety Chemical safety Fire resistance, Flammability Thermal safety / Overheating / Thermal shock Functional safety

Resistance to media

Ingress protection

Corrosion Resistance to chemical media

Ingress of fluids Ingress of solids

Packaging Size Weight Fixation Interfaces

Abuse /Misuse Electrical abuse Mechanical abuse Thermal abuse Handling abuse

Mechanical robustness

Legal

HV E/E functions

Documentation

Mechanical vibrations Mechanical shocks Additional mechanical loads

Homologation Transport

El. Isolations Current carrying performance Switch-on Switch-off

Labels, Marketings and Manuals

Figure 3: Main Function Groups (FG)

Integration risk assessment during a derivate battery pack development project is realized by evaluating the technical risks related to: design, performance, producibility, durability and safety. To assess these risks, three different scenarios in a battery derivate project are presented. The Risk Diagram (see Figure 4), provides a general view for

644

Front loading approach in battery development for generation update each of the presented scenarios, assessing the risk level related to the changes that are made to the lead battery and implemented in the derivate battery development project.

2.2 Scenario 1 – Changes on cell level Following two cases are considered:

2.2.1 Increase of cell energy density The approach in this case would be to use the existing battery module from the lead battery project but utilize a new battery cell with higher volumetric & gravimetric energy density, with similar geometry and cell connection points. Such a change would result in a battery pack that has higher installed energy content (kWh) and higher power (kW) capabilities. Increasing the cell energy density is always a matter of additional safety that needs to be evaluated, as the cell with higher energy content will also have a different venting and swelling behaviour. Additionally, the increase of cell energy content, achieved by an increase of specific columbic Ah capacity, leads to also increased cell charge/discharge current capabilities. In terms of above presented main Functions Groups (FG), during the derivate battery development project especially the Safety Functions, Energy capacity & Power delivery, Abuse/Misuse and Current carrying performance in the HV E/E Functions Groups need to be considered. Higher risk level marked on the Risk Diagram shows that the increase of cell energy density can potentially have a high impact in reducing pack safety and performances/functionality. This would mean that considerable additional efforts and attention would need to be implemented during the derivate battery development to mitigate these risks.

2.2.2 Decrease of cell energy density In this case, the existing battery module from the lead battery pack is kept but a new cell that has lower energy density with similar geometry and cell connection points is used. Such change reduces the energy capacity and power capabilities for the derivate battery pack. Considering the Functions Groups, reducing the cell energy density will not require additional safety considerations to the already conducted in the lead battery project. Assessments are needed primary from point of Energy capacity & Power delivery Functions Group, to confirm that the reduced energy and power capabilities, will meet the new targets. Also, considerations from point of HV E/E Functions group, could lead to possible optimizations on E/E component level.

645

Front loading approach in battery development for generation update Low risk level marked on the Risk Diagram shows that the reduction of cell energy density will have low impact on pack safety. Such low risk levels could support stakeholders and project managers in decisions to optimize the cost and duration of the derivate battery concept development phase by applying simulations on virtual prototypes to reach the Targets Agreement quality gate, as per AVL GBDP, without having to build additional hardware.

2.3 Scenario 2 – Changes on module level This scenario describes the case of derivate battery pack that requires less energy content than the lead battery. This can be achieved by making changes on the module level - reducing the number of modules in the lead pack. As this would lead to a battery pack with lower energy content and reduced power capabilities compared to the lead battery pack, the safety impact would be like described in case 1.b)., but, in this case, additional considerations would have to consider that reducing the number of modules, will reduce the battery pack nominal voltage. Assessment is required primary focusing on HV E/E Functions group to evaluate possible optimizations on E/E component level and Energy capacity & Power delivery Functions Group to confirm that the reduced energy and power capabilities, will meet the new targets. Interfacing requirements towards other components within the powertrain, especially towards the inverter needs to be assessed considering the reduced pack nominal voltage. Medium risk level marked on the Risk Diagram shows that reducing the number of modules within the lead battery pack, may lead to some reduction of functionalities/performances in interfacing the pack to other system components. This would require special attention to be taken by the stakeholder and project management team to mitigate the risk.

2.4 Scenario 3 – Changes to interfaces on pack level In this scenario the utilization of a battery pack from the base project and integration into a newly developed vehicle with a longer wheelbase is considered. Such a change primarily needs to assess new interfacing requirements (e.g. connectors, plugs, brackets for additional stiffness, etc.) into the vehicle sub-structure and additional mechanical loads and shocks the battery will be subjected to. It needs also analysis of how the interfacing changes have impacted the pack vibrational behaviour (e.g. eigenfrequencies). Taking this into consideration, assessment for derivate development is required specially from the point of Packaging, Mechanical Robustness and Design for Production & Service Functions Groups. Medium risk level marked on the Risk Diagram shows that using an existing battery pack for a new vehicle may lead to some additional design activities and producibility issues, that could impact duration and costs of the project.

646

Front loading approach in battery development for generation update

Figure 4: Risk Diagram

● All previously described steps are taken during the Feasibility phase in Derivative Battery development project, ending with a Technical Feasibility Report for the derivate development project. ● Completing the feasibility phase, derivate concept design phase will follow, with base battery project reworks. Further concept development steps will depend on the prior risk assessment. For derivates with low risk, completing the concept development phase, with the Targets Agreement (TA) deliverables reached according to AVL GBDP quality gate, could be realized by simulations on the virtual prototype. Serial derivate development with B- sample design and C- sample design phases, will then directly follow, representing steps towards technically optimized designs, with production intent, having hardware built, capable for full verification testing. In this way, virtual development approach for low risk derivate projects reduces the total development time. For derivatives with higher risk, a decision might be to go into an optional loop, with an A/B sample design, simulations, prototype build and testing and a simulation model correlation. Serial derivate development with a B- sample design and C- sample design phases, would then follow, after the A/B design has passed the quality gate. AVL derivative battery development process is shown on picture below (see Figure 5).

647

Front loading approach in battery development for generation update

Figure 5: AVL derivate battery development process

3 Virtual development approach in battery derivative development Frontloading and a virtual development approach are critical in reducing development time efforts and costs. As stated above, changes only on the cell level can have a very high risk for the battery derivate development project. Cell properties influence global battery behaviour. For example: ● Venting and therefore hazard behaviour can be very different if cell chemistry is changed (see Figure 6) ● Different swelling behaviour causes different global stiffness and modal behaviour as well as different hazard levels in crush (see Figure 7)

Figure 6: Cell venting behavior as function of cell chemistry

Figure 7: Cell swelling

It is thus necessary to perform a gap analysis between base and new cell, as basis for planning of derivate development efforts.

648

Front loading approach in battery development for generation update

3.1 Virtual development includes performing the following activities regarding the new selected cell: ● Electro-thermal model calibration & update of thermal contact from new cell chemistry to coolant ● Stiffness & mechanical failure criteria out of new cell test program ● Macro-level approach for modelling new cell out-of-plane mechanical behaviour ● Measuring and determination of modal properties for the new cell (see Figure 8) ● Consideration of stiffness change of cell stack (module) over lifetime due to swelling behaviour of new cell and nonlinear behaviour of compression pads ● Evaluating changes to the swelling as a cause to a different crush behaviour over lifetime ● New cell venting measurements and behaviour ● Re-assessment (from base battery project) of delay time from 1st cell in thermal runaway until neighbouring cells are triggered to fire (see Figure 9)

Figure 8: Modal properties and eigenfrequencies

Figure 9: Thermal Propagation

649

Front loading approach in battery development for generation update As can be seen from many of the activities, deep cell understanding is essential for cell properties characterization and modelling, as well as integration and simulation of cell behaviour within a battery pack. If the lead battery development process is already is serial development phase (B or C sample design), application of virtual development approach can help in optimize the process by using the correlated B/C sample simulation models from the lead battery and updating them with the new cell properties, resulting in an updated derivate battery model (see Figure 10)

Figure 10: Optimized derivate development by virtual approach

Figure 11: AVL Generic Validation Plan as basis for gap analysis

650

Front loading approach in battery development for generation update

4 DVP for derivate development project Once the delta assessment and the risk assessment have been completed, the necessary changes that need to be made in the design verification and validation plan (DVP) and possible additional testing that would need to be included in the program can be defined. These changes might include additional homologation, durability-oriented tests, performance and functional tests. AVLs structured approach in delta assessment considering all stakeholder requirements, and assessment of all risks for the derivate battery development project, is an enabler to define an updated and optimized DVP, from point of costs and testing time. Example of an AVL generic validation plan for a B-sample battery pack is shown below (see Figure 11).

5 Derivate Battery Assembly and Production When starting a derivate battery pack development project, “design for assembly” is crucial, as it must be balanced between the level of adaptations on the existing lead battery design, necessary - to satisfy the new boundaries and targets, and the minimization of the additional investments and changes needed on the existing, lead battery pack production line. To reduce new investments and timing during generation updates, the derivate battery pack should be built on the same initially as flexible planned production line. Unfortunately, the possibility to use the production line established for the lead battery pack for a future derivate battery is today typically not considered when initial investments are made, and production line planning for the lead battery pack is done. This results in having a production line that does not have the flexibility and required additional capacity for the derivate battery production and any increased production quantity, achieved by increasing the variable (labor) cost, could result in increased marginal cost of production or possibly even average total cost of production, having diminishing return to scale. Considering the different scenarios for derivate battery pack development presented in Article 2, from the production perspective, it can be assumed that for scenarios 1 and 2, there are no significant changes required in derivate battery assembly and production line, compared to the lead battery. It is needed to check whether the existing production line can satisfy the increased production output, if both packs shall still continuously be produced. It also needs to be

651

Front loading approach in battery development for generation update checked if any additional measures are required by the personnel and tools during the assembly process due to higher currents and voltages (Scenario 1) or higher weights (scenario 3), and if additional requirements in the scope of EOL testing are given. For scenario 3 more efforts are needed in production of the derivate battery pack. The changes to interfaces on pack level can be considerable, requiring a re-evaluation of all single process steps, making it possible to decide if any additional capital investments to produce the derivate battery pack are necessary. The additional investments activities might only be for some changes on the existing production line or these could be much substantial for establishing completely new production lines. To evaluate the additional needed investments, a “gap” analysis must be performed, covering the following topics: ● Production Sequences ● Production Tact Time ● Single Production Steps: Parts/BOM (handling) & Tools and Fixtures ● Logistic & Supplier ● Quality Steps ● Safety and ● EOL-Test

6 Conclusion and final discussion As mentioned in this paper, automotive OEMs are today facing many challenges on the path to increase the portfolio of their electric vehicle offerings in the next couple of years. The electrification path demands significant, costs, efforts and time. Derivate battery development is an option to reduce them, but to be successful in this endeavour, a systematic derivate battery development process needs to be established to assess the level of all required changes for the lead battery. AVLs battery derivate development process is such a process. Additionally, when starting a derivate battery pack development project, we also need to balance between the level of adaptations on the lead battery design – necessary to satisfy the new boundaries and targets, whilst at the same time to minimize the additional investments and changes needed on the existing, lead battery pack production line.

652

Front loading approach in battery development for generation update Derivate battery development requires deep knowledge on the battery technology, from cell to pack level, covering all Subsystems and Functions groups and understanding how critical the changes on a subsystem or component level might be on the performances, functionality or safety on the system/vehicle level.

Bibliography 1. “AVL Battery development process”, AVL internal process guideline 2. Dr. Bernhard Brunnsteiner: “Front Loading Ansatz bei Zellchemiewechsel in der Batterie Entwicklung“, AVL presentation, Emotive 2019

653

CO2-neutral battery production in Europe – How to make it happen? Robert Stanek, Markus Hackmann P3 automotive GmbH)

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_54

655

CO2-neutral battery production in Europe – How to make it happen?

1 Introduction eMobility is a clear trend in the automotive industry and will soon dominate new vehicle registrations. The providence of a sufficient electric range is one of the major challenges on the way there. An enabler that is associated with this issue is battery technology. With higher energy densities, achieved through advancing developments in battery technologies, costs are shrinking and electric vehicles can be equipped with more battery capacity. In 2030, about 45% of new vehicle registrations in the EU will have an electric powertrain. Driven by CO2 Emission regulations on the major automotive markets in China, the U.S. and Europe automotive OEMs are obliged to introduce CO2 emission free and emission reduced vehicles into their product portfolios. On the European market this tension is even more preset due to the strictness of CO2 targets for 2020 and 2025 compared to the other markets and recent industry developments. The decreasing market share of diesel vehicles and the introduction of the WLTP test cycle have put OEMs in Europe under even more pressure to change their propulsion system portfolios. CO2 compliant fleets must be introduced to prevent financial penalties and image losses.

1: Annual sales shares of vehicles propulsion systems in Europe driven by CO2 regulations [P3 CO2 Compliance Simulation]

The solution seems to be the introduction of electric vehicle platforms starting after 2019 as announced by many European and international OEMs. One of the targets of these platform strategies is to benefit from economies of scale for the component costs of the electric powertrain and to in consequence achieve competitive prices for the new electric vehicles.

656

CO2-neutral battery production in Europe – How to make it happen? The battery systems and the battery cells constitute the major cost share of the electric powertrain (up to 40% for medium range vehicles) and thus play a particular role within the sourcing and vehicle price definition strategy.

2: Key driving factors for carbon emission free electric vehicles

A sustainable change of mind in regard to climate change increased the pressure on OEMs to produce holistically carbon free vehicles in the future and is driven by ecological, political, economical and social factors. To meet the target of ZERO carbon emissions in 2050 worldwide, car makers must not only sell BEV vehicles but also produce these without emitting CO2and offer e.g. also renewable electricity for charging of these. This will include especially a carbon emission free active battery material manufacturing and battery manufacturing. Especially the collaboration between the key elements of legislation, certification and the availability of technologies needs to go in hand to build up a sustainable value chain with lower emissions for measurable effects including direct and indirect emissions (vehicle, used energy).

3: Key elements of the e-mobility value chain with leading elements towards CO2 effects and potentials

657

CO2-neutral battery production in Europe – How to make it happen?

2 State of the art and current industry performance Currently the supply market of lithium-ion battery cells is dominated by South Korean and Japanese players as these introduced this technology to the market. These players used to supply earlier low volume xEV (electric vehicle) projects with lithium-ion cells from their domestic production locations. With the upcoming volume scale up for electric vehicles driven by CO2 regulations these players have built dedicated production locations in (Eastern) Europe to follow the necessary supply chain requirements of OEMs. Especially the initial stages of the battery value chain offer high value add potential due to necessary investments and technology differentiation potentials based on elaborated know-how and experience of established players.

4: Key value chain elements with identified value added factors and ratios as well as necessary key capabilities (Note – Value Add based on addressable cost share consisting of CAPEX depreciation, overheads and profit pools)

Most of the required sub-components and raw materials (e.g. nickel, cobalt, lithium and graphite) for these new production locations will initially be sourced from domestic and established supply chains in Asia. This is as well caused due to the lithium-ion technology is very sensitive towards changes of the applied materials and components for the manufacturing process.

658

CO2-neutral battery production in Europe – How to make it happen?

5: Several major OEMs already announced a carbon free vehicle production.

The topic of this talk and elaboration will consequently be to discuss the chances for European players and industry segments to still enter these supply chains due to e.g. missing infrastructure and local production facilities of the established competitors as well as open technology fields. Moreover, a focus point of the presentation will be to show P3’s view on success factors and roadblocks for new entrants in the battery cell manufacturing business to participate and compete with the dominant market players. In this context the depth of vertical integration into the value chain of battery cell production will be most of interest as it defines the amount of potential chances and risks. These aspects will also be discussed with regards to location-based advantages of the European industry. Among other points the availability of CO2 friendly energy grids, resulting in beneficial CO2 friendly component and raw material production will be considered. In view of the initial motivation of setting CO2 targets for the automotive industry this specific characteristic of the European location is a focus of this presentation to raise the argument of well to wheel CO2 implications.

3 Preliminary results and findings Along the full value chain of batteries there are several factors driving the emissions footprint with battery materials as e.g. nickel, cobalt, manganese, lithium and graphite being a part of these. Especially the mining, processing and shipping is causing most of the carbon footprints and is highly distributed across the globe. The mining, processing and transportation of the electrode raw materials is causing about 2/3 of the total carbon footprint of today’s battery cells and is especially relevant to localized markets as Europe due to lacking mining.

659

CO2-neutral battery production in Europe – How to make it happen?

6: Example case for europena based cell manufacturer with according supply chain and emission footprint

By today all running battery manufacturing facilities are build up on either small footprints for specific industries or copies of existing installations in Asia with no focus on energy consumption or green technologies. Initial projects like Northvolt are trying to change this. To evaluate the full emission footprint starting from the refining of electrode material up to a fully assembled batter pack many factors and simulations of machine consumables and value chains need to be considered. For a current battery pack this includes details of approximately 10 relevant steps and integration levels. CO2 value chain of a conventional 60-75 kWh EV battery pack with a standard global material value chain for European OEMs with domestic module and system assembly.

7: Reference evaluation of carbon emission footprint of traction battery system with 60-75kWh installed capacity (P3 simulation)

660

CO2-neutral battery production in Europe – How to make it happen? Given the european environmental factors and e.g. the option to use in the future emission free electricity sources as e.g. hydro, solar or wind power, these could be used for material refinement as well as cell and battery manufacturing and leverage up to 75% of the emissions for such batteries, especially with colocated material mines and refinements. The overall potential for European „New Entrants“ to reduce emissions is significant and the relevant players need to identify local industry advantages (mining and material refinement options) and (re-)evaluate their targeted value chain. The P3 simulation shows that for an localized manufactured battery system including the cells in Scandinavia the CO2 emissions can be cut by 75%, if the active material manufacturing (cathode & anode) as well as the cell manufacturing is executed with renewable energies.

8: P3 Simulation of full impact on battery emission footprint with currently all potential levers used

The key challenge still exists as of todays and future necessary battery material volumes almost none of the key materials is actively sourced and refined in Europe. This is caused on the one side due to missing availability of the materials as well as negative cost positions due to highly cost efficient setups for material refinement in Asia and strictly industrialized value chains. Europe therefore will still be highly dependent on Asia cell material supply and also integrated the carbon footprint of these materials within the battery products manufactured. This open end needs to closed to a continous loop including battery material recycling to estabilsh a sustainable, competitive and reliable material supply within Europe.

661

Greater sustainability with a second life of used electric vehicle batteries Dr. Jürgen Kölch, EVA Fahrzeugtechnik GmbH

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_55

663

Greater sustainability with a second life of used electric vehicle batteries

1 Introduction Worldwide more and more fluctuating renewable energy, such as solar and wind power, is being strongly increased (so called “Energiewende”). The expansion of energy storage to compensate these energy fluctuations will become an important issue in the future. There is the possibility to store e.g. photovoltaic (PV) electricity during sunny days and give it back to the home grid as needed during night or cloudy days. The batteries aged during the life of an electric vehicle can be reused for a variety of stationary applications. In combination with photovoltaics, higher private consumption rates can be reached at a module level. At the level of complete vehicle high-voltage batteries, DC quick charge stations with low power connections or load management (peak shaving) represent an example for industrial clients. In the largest configurations in the MW area, used electric vehicle batteries can provided grid support or they can be used to stabilize the energy from wind parks. In order to ensure that the second life of the battery is just as long as it was in the vehicle, the state of charge range of the battery is limited, among other things. In addition, components related to the energy storage system must be checked for their suitability for stationary use and replaced if necessary. The connection of various aged vehicle batteries poses particular challenges. The environmental benefits of an electric vehicle face repeated criticism. Along with the use of CO2-intensive, coal-based electricity for vehicle operation, the ecological rucksack of the battery is the primary focus. When the battery is reused at the end of the vehicle’s service life (Battery 2nd Life, B2L), the environmental benefits far exceed those of just regular recycling.

2 Photovoltaic electricity for electric vehicles 2.1 Falling photovoltaic module prices At the end of 2018 the worldwide installed photovoltaic (PV) power reached 500 gigawatts according to Wirth [1]. The modules account for about half of the total costs of a PV system. A doubling of the installed power leads to 24% price reduction for photovoltaic modules. In the last 15 years (2000 to 2015) the module price for one kilowatt peak (kWp) fell down from 6,000 to 600 EUR [1]. For one kWp six to ten square meters of module area is needed. For Germany 800 to 1,000 kWh per kWp and year can be expected. In case of unfavorable roof orientations (e.g. orientation West or East and an assuming roof pitch of 40°) about 25% yield loss compared to an ideal south-facing roof can be assumed. However, with a west/east-orientation the doubled number of modules can be installed compared to a south-roof. Even a northern-roof with a pitch of 20° can be contemplated with about 30% yield losses.

664

Greater sustainability with a second life of used electric vehicle batteries Figure 1 shows the global installed photovoltaic (PV) capacity in gigawatts (GW; peak capacity) plotted double-logarithmically against the module price.

Figure 1: Average PV price for one watt (peak) depending on the installed PV capacity, Source: Own presentation according to Fraunhofer ISE [1].

2.2 Achieving grid parity for photovoltaic electricity in Germany With the introduction of the Renewable Energies Act (EEG) in Germany in 2000, all the generated solar electricity was fed into the grid because the payment was significantly higher than the price for domestic electricity. Since 2013, we have now reached the time when it is more economical to cover your own energy consumption using solargenerated electricity first and then feed only the surplus solar electricity into the grid (grid parity). The more the domestic electricity price increases in the future (and the more the price of solar electricity falls), the greater the cost advantage will be for covering your own consumption with solar electricity generated independently (see figure 2).

665

Greater sustainability with a second life of used electric vehicle batteries

Figure 2: Comparison between the cost trend in the REA price paid for photovoltaic systems with an installed capacity of less than 10 kW and domestic electricity prices. Source: Own presentation according to Quaschning [2].

2.3 Used batteries for a stationary energy storage The growing market for electric vehicles will increase the electricity demand on the one hand, but it will also increase the return flow of used vehicle batteries. These batteries are no longer good enough for the mobile application but are still good enough to be used for a stationary application. The following applications have been identified from several possible applications for used electric vehicle batteries. They have been divided by size into three areas.

2.3.1 Small battery application When the vehicle high-voltage battery (HVB) is stripped down to the module level, systems can be set up with the modules via parallelization, e.g. for a private photovoltaic (PV) power storage system. They are often designed for a voltage range of smaller than 60 V DC as safety extra low voltage (SELV). Protection against direct contact does not have to be provided within this voltage range. This is currently provided in the single-family home with a private user as system operator. In Germany, the generation of PV power has generally been cheaper than household electricity costs (grid parity) since 2013. From this point on, a market has opened for battery storage in order to

666

Greater sustainability with a second life of used electric vehicle batteries in-crease the share of private consumption and reach a higher economic efficiency of the PV system due to decreasing battery costs.

2.3.2 Medium battery application For an output power in the order of 100 kW, the entire system can comprise several HVS units. To get this power at 400 V AC with typical available inverters a higher DC battery system voltage of 800 V is needed. Two HVS with 400 V are connected serially to 800 V. A realized battery rack with four HVS is shown in Figure 3. Each of the two upper and lower HVS are interconnected to one unit and can be scaled according to the application in power and energy with a suitable design of the inverters. Possible applications for this system are large commercial PV systems or peak shaving (load management) systems to streamline a company's power supply. These battery packs can also be used to support DC fast charging stations for electric vehicles which do not require a strong grid connection. Maybe stationary battery storage systems could help the currently discussed installations of DC charging stations with more than 300 kW charging power.

2.3.3 Large battery application For the primary control reserve a minimum lot size of one megawatt must be available for the German Transmission System Operators [3]. For this size, the rack construction in the medium application would be possible but requires a lot of space. A denser construction is the configuration of individual modules in battery cabinets, which can be placed in containers or suitable halls. The removed modules of two electric vehicles are in one battery cabinet with 800 V DC. In addition to a frequency stabilization application, the buffering of a large wind park is also possible.

667

Greater sustainability with a second life of used electric vehicle batteries

Figure 3: Combination of battery cabinets with individual modules.

3 Design criteria for Battery 2nd Life applications in contrast to the e-vehicle A typical vehicle is only in operation approx. 1-2 hours per day. The state of charge (SOC) is heavily utilized (completely empty to completely charged) and the temperature range of the battery can vary widely (sometimes extremely). On the other hand, this is different for stationary applications: The battery storage system needs to be operational 24/7, 365 days per year. The SOC band is limited and is set up for lower currents and thus lower outputs when compared to the vehicle. The battery temperature has a comfort zone of 20 to 25 °C. All the measures listed should enable the longest possible service life for the second life of the battery.

668

Greater sustainability with a second life of used electric vehicle batteries

3.1 Design of E/E components Due to operation with low currents (C rates) as in vehicles and other design parameters, E/E components such as fuses and contactors must be newly designed and implemented for stationary use. This also affects various cable cross sections and new high-voltage cables that cannot be applied from the vehicle to the stationary application.

3.2 Packaging When the HVB is broken down to the module level, the modules must be properly dimensioned geometrically in battery cabinets. For safety and E/E components, a separate service level is a good idea. It should also be easily accessible for the subsequent replacement of components. The planning of high-voltage and communications lines also plays an important role regarding electromagnetic compatibility (EMC).

3.3 Heat management HVBs are often liquid-cooled in the electric vehicle. The effort to do this for stationary applications is extremely high so that air-cooling is a better alternative. When battery cabinets with individual modules must be cooled, the air is blown through several, individual fans from the front through each individual module level. (see figure 4).

Figure 4: Air cooling flow of individual modules on the bottom from the front.

669

Greater sustainability with a second life of used electric vehicle batteries

3.4 System architectures with 800 V and 400 V voltage level Typical vehicle high-voltage batteries (HVB) have a DC voltage of a max. 400 V DC. This makes it difficult to reach the usual AC voltage level with a standard inverter of 3 x 400 V AC (low-voltage grid). The typical operating range of wide-spread industrial inverters requires double the amount of DC voltage (800 V DC) on the part of the HVB. This results in considerable changes to E/E components, and the maintenance of minimum air gap and creepage distance. In addition, two similar aged HVBs must be connected to each other in series. (see figure 5).

Figure 5: Wiring from the energy storage system to the transmission grid (800 V system architecture).

4 Batteries as stationary energy storage in the smart grid When charging electric vehicles in the Smart Home, two different views collide. From an OEM point of view the electric vehicles always has the sovereignty over the charging process and knows the state of charge and maybe the time of the next vehicle trip. Furthermore, a desired preconditioning (pre-heating of the vehicle in winter or pre-cooling in summer) can be coordinated for greater range and comfort. In contrast the home automation system handles the vehicles like an electrical load which can be switched on and off like e.g. an electric heating element in the hot water storage. The rude onand off-switching can also lead to charge interruptions in some cases. This different

670

Greater sustainability with a second life of used electric vehicle batteries approach led to complex communication of the vehicle with the home energy management system (HEMS). Thus, the HEMS provides the vehicle with a recommendation for a preferred charging time with expected time-resolved PV performance. The vehicle takes this information as an input to calculate its own charging schedule and the result is reflected back to the HEMS. The HEMS plans to set the time and energy level of the electric vehicle into its timetable. What sounds rather good in theory works in practice only if the user didn’t forget to plug the electric vehicle.

5 Outlook The use of used electric vehicle batteries is still a new topic. There are still relatively few electric vehicles on the market and used vehicle batteries for a stationary application will only become available when their range and performance are no longer adequate for the customer. This means that the relevant quantities of used electric vehicle batteries will only become available in several years with an upward trend and this coupled with an increase in the number of electric vehicles. However, new, future battery generations with significantly increased range with an attractive upgrade offer could give customers enough incentive to replace their vehicle battery much sooner, and thus the topic of Battery 2nd Life (B2L) would become relevant much more quickly than is currently the case due to the rapid rise in the number of old vehicle batteries.

Bibliography 1. Wirth, Harry, Fraunhofer ISE: Recent Facts about Photovoltaics in Germany, Last update: January 7, 2020, https://www.ise.fraunhofer.de/content/dam/ise/en/documents/publications/studies/recent-facts-about-photovoltaics-in-germany.pdf, accessed on 2020-01-09. 2. Quaschning, Volker, in: BWK, Volume 64 (2012), No. 7/8, pg. 25-28. 3. German transmission system operator’s common internet platform: Common tendering for primary control reserve, https://www.regelleistung.net/ext/static/prl?lang=en, accessed on 2020-01-09.

671

Increased safety for battery electric vehicles by using heat-resistance stainless steels Stefan Lindner, Outokumpu Nirosta GmbH

© Springer Fachmedien Wiesbaden GmbH, ein Teil von Springer Nature 2020 M. Bargende et al. (Hrsg.), 20. Internationales Stuttgarter Symposium, Proceedings, https://doi.org/10.1007/978-3-658-30995-4_56

673

Increased safety for battery electric vehicles by using heat-resistance stainless steels

1 Introduction Passenger cars with combustion engines dominate the 20th century because of their significant expanded range, the quick refuel process as well as availability and price of the fossil fuels. During the last years, electric vehicles experience a renaissance from their previous developments at the end of the 19th century to one enabler for future mobility [1]. Especially social desirability of our society as well as resulting political and legal framework, but also the increased price and limitedness of fossil fuels, promote a change in the type of drive. This trend is picked up by the automotive industry with significant investments into the technology of battery electric vehicles [2]. The development is further reinforced by simultaneously staged megatrends like autonomous driving, car sharing, driving in urbanized areas often connected with environmental zones, or the topic of “last-kilometre” transportation for goods or people mover. New design aspects for battery electric components but also for the surrounding car body must be nowadays integrated to ensure the same passenger safety as reached with combustion engine vehicles. From these changing requirements also growing demands for application-specific material concepts can be derived and are getting more challenging. Figure 1 illustrates an overview of the different material related requirements for the example of the battery compartment. Beside passenger safety related to material strength and ductility, lightweight, processability for joining and forming, recyclability and the material-specific CO2-footprint can be highlighted as some major requirements for the application field of battery electric vehicles (BEV) [3, 4]. Improved as well as newly developed stainless steels can fulfil the high expectation on materials for future mobility by offering superior energy absorption, highstrength properties connected with specific hardening mechanism and 100 % recyclability. Thereby, the property of heat-resistance, well-known from exhaust systems and heat exchanger, is for the first time in car body engineering directly connected with passenger safety in the case of spontaneous ignition of the battery cells. Also in this case, stainless steels represent the winning formula with experience over decades in high-temperature applications.

674

Increased safety for battery electric vehicles by using heat-resistance stainless steels

Figure 1: Overview of several material requirements for battery compartments

2 Stainless steels Steels with a chromium content of at least 10.5 % are classified as stainless steels according to DIN EN 10088-1 [5]. Moreover, the presence and level of other alloying elements which can be sub-divided into ferrite and austenite formers determine the resulting microstructure, which in turn defines the respective physical and mechanical properties of the specific stainless steel grade. Thereby, stainless steels can be classified depending on their microstructure into ferritic, martensitic, meta-stable or fully austenitic and “Duplex” (ferrite + austenite) stainless steels, view figure 2.

Figure 2: Microstructure and lattice structure of stainless steels

675

Increased safety for battery electric vehicles by using heat-resistance stainless steels In contact with atmospheric oxygen, stainless steels offer a natural and stable chromiumoxide surface layer resulting in a naturally high level of corrosion protection. Since their invention in 1912, stainless steels represent with their 100 % recyclability the most recycled material in the world whereby the quality is preserved in the process. With a proportion > 85 % of recycled content (> 2,000,000 tons per year), stainless steel produced by Outokumpu significantly decreases the carbon footprint of the complete transportation supply chain. Stainless steels represent an established high-performance material in transportation engineering for decades. Their extensive range of properties enables their use in exhaust systems, airbag tubes and trim systems as well as crash-relevant structural and chassis components. In addition to their enormous formability, they are also notable for good weldability, adhesion properties and paintability. Thereby, the properties of stainless steels perform a continuous reinvention to support the ongoing improvement and changing demands in automotive manufacturing, view figure 3.

Figure 3: Timeline of stainless steel properties in the automotive world

A new generation of stainless steels with a stable one-phase, fully austenitic microstructure can be developed because of a special hardening mechanism called TWIPeffect (Twinning-Induced Plasticity) into the material category of ultra-high strength steels, view figure 4, [6]. Responsible for this characteristic profile are the coordinated alloying elements like chromium and manganese reaching specific stacking fault energy. As a result, the cold-formable material enables an intensive work-hardening during cold-rolling of the material but also during cold-forming of the component or during impact situation of the vehicle. At the same time enormous energy absorption can be realized. Furthermore, the component-manufacturer is able to adjust the local desired properties of the material in dependence of the cold-forming degree.

676

Increased safety for battery electric vehicles by using heat-resistance stainless steels

Figure 4: Mechanical-technological values of Forta H-series

Design engineers can create higher strength or higher ductility areas inside one component where the usage conditions of the components require it. The Forta H-series is applicable over the complete automotive process chain because of its good similar and dissimilar weldability. Further, a high formability for complex formed parts which contributes directly to component stiffness is given. Because of the well-balanced alloying system, the conditions defined in [7] are fulfilled to avoid any kind of delayed cracking. Beside chromium, the other main alloying element manganese works as an austenite former to provide a stable fully austenitic microstructure even after forming and welding as well as non-magnetic properties. Chromium influences the mechanical-technological values of the TWIP material in a positive way [8], supports the solubility and homogeneity of the other elements and builds up a chromium-oxide passivation layer on the steel surface. Based on this passivation layer the material works without any kind of coating like galvanizing. In combination with the established cathodic dip coating in the automotive manufacturing process, the passivation layer presents a very good corrosion protection system even after a lattice cut or stone chipping. The natural chromium-oxide passivation layer repassivates again after a mechanical injury and avoids an undercut of the injured dip coating process.

3 Design concepts for battery electric vehicles With range and availability two key demands in point of customer acceptance for batterypowered electric vehicles (BEV) can be mentioned. To fulfil those expectations, a high degree of space utilization in the vehicle floor to host the battery modules is required and results in complex design and engineering demands. Further, the battery modules have to be protected from environmental influences such as corrosion, extreme temperatures and mechanical injuries like stone chips as well as deformation

677

Increased safety for battery electric vehicles by using heat-resistance stainless steels to prevent thermal collapse. Moreover, to sustain crash requirements constitute a further challenge, especially for side and underfloor intrusion where the intrusion space is limited. Additionally, thermal management needs to ensure that the batteries are maintained at their ideal operating temperature offering maximum range. All of this leads to the battery compartment as a key component in the design of battery-powered electric vehicles. The material stainless steel and its associated construction concepts can make a significant contribution to enable the application requirements of BEV. Stainless steels further offer advantages in mass production as well as recycling, carbon footprint and overall operating costs.

3.1 Heat-resistance as key requirement for safety Stainless steels are used since decades as the material of choice for exhaust systems as they are offering heat-resistance as well as high-temperature corrosion resistance properties. Heat resistance means stable mechanical technological values even at higher temperatures. For the field of battery-powered electric vehicles and a potential self-ignition of the battery cells after a crash situation with a critical deformation of the batteries, heat-resistance is directly connected to passenger safety. After a crash, the structure must be functional since a fire department can rescue the occupants out of the vehicle which means in time between 9 and 12 minutes. Figure 5 points out the results of an austenitic stainless steel in relation to aluminum and coated carbon steel. During a self-ignition of the batteries, the expectable temperature stays over 1,200 °C.

Figure 5: Result of structural test for different metal materials and temperatures

678

Increased safety for battery electric vehicles by using heat-resistance stainless steels

3.2 Compartments for battery cells The battery compartment covers the battery modules with the battery cells and should therefore have an installation space as large as possible to increase the driving range. A deep-drawn shell solution represents a battery compartment simple in geometry and scalable up to mass-produced volumes with short manufacturing times. By using austenitic stainless steels and their high tendency to significant work-harden during deepdrawing, such a shell can be manufactured from highly ductile materials resulting in ultra-high strength parts. Therefore, the intrusion resistance of the component can be adjusted to a safe level. Thereby the interior radii as a safety space between the shell wall and the integrated battery modules must be dimensioned to ensure the necessary safety. An interior radius in the range of between seven and 12 mm represents not a limitation during deep-drawing austenitic stainless steels. The advantage of a deepdrawn battery compartment beside the fast cycle times is the saved packaging space. Furthermore, additional production steps like welding in the battery “clean room” can be avoided. Thus, thermal distortions during welding and subsequent cleaning operations which increase usually production costs are excluded with a shell construction. Figure 6 illustrates a complete stainless steel battery compartment design including the shell solution but also offering the integration of necessary surrounding parts like the thermal management system, the underride protection or the crash frame at the sides.

Figure 6: Lightweight stainless steel battery compartment

3.3 Thermal management Having more than 7,000 battery cells inside a BEV and a lost to waste heat of 5% of the efficiency of lithium-ion traction batteries, a critical amount of heat inside the battery compartment must be lead away. At temperatures above 45 °C, battery damage

679

Increased safety for battery electric vehicles by using heat-resistance stainless steels occurs due to chemical reactions inside the cells. At the other extreme with a temperature below −5 °C, performance and charging capacity decrease and thus lowering vehicle range. In a damage case, the mostly liquid coolant must not come into direct contact with the battery cells. Therefore, the ideal solution is passive cooling, which lead to additional advantages in production and assembly. By using a thin ferritic stainless steel in contact with the battery modules and a thicker austenitic stainless steel as a protection to the environment, the advantages of stainless steels in point of their physical properties can be used. The different metal thicknesses and physical properties like heat conductivity, deliver highly efficient cooling of the battery modules. On the contrary the insulating effect of the outer-located austenitic stainless steel helps to reduce the heat transfer with the environment, view figure 7.

Figure 7: Heat conduction influenced by material thickness and physical properties

It is further possible to combine the before mentioned technologies into one solution connecting the different advantages. Figure 8 illustrates a stainless steel shell solution where during deep-drawing cooling channels are directly integrated into the floor. Subsequently a flat sheet can be attached, e.g. with bonding, to realize a closed battery housing with a passive cooling system.

680

Increased safety for battery electric vehicles by using heat-resistance stainless steels

Figure 8: Deep-drawn stainless steel shell with integrated cooling channels

3.4 Crash frame und underride protection The crash frame is positioned between the battery shell and rocker panels and must absorb the impact energy during a crash situation. Alongside geometric design, material selection is a crucial factor. In addition to high residual ductility reaching in an ideal case values of A80 >50%, the component must also display very high strength. Using strain-hardenable austenitic stainless steels, it is possible to produce complex shapes as roll-formed profiles, which produce geometrically high stiffness due to the complexity of their form. A twin-chamber profile is a workable design where the profile faces the rocker panel and absorbs the energy. The inside profile generates high intrusion resistance. The forming possibilities of austenitic stainless steels enable rollformed geometries that could previously only be achieved with extruded aluminum profiles. Resistance spot welding or laser beam welding with ductile seams and highpower transmission are further benefits offering a mass-production assembly. A second manufacturing concept for the crash frame is internal high-pressure forming using an active fluid medium for producing extremely stiff pillow-plate absorbers, view figure 9.

Figure 9: Deep-drawn stainless steel shell with integrated cooling channels

681

Increased safety for battery electric vehicles by using heat-resistance stainless steels The reachable height to thickness ratio of the overlapped, tightly welded stainless steels sheets is on the level of r h/t = 60. Also in this case, the combination of manufacturing process and used material results in significant material hardening, while the geometry delivers high stiffness. Just like the two-chamber principle with roll-formed profiles, an outward-facing energy-absorption area can be defined using a section of thin exterior sheet. On the other hand, the internal section works as stiffener and must have a higher sheet thickness, ideally to a factor between 2.5and 3.5 related to the thinner outer sheet. The space in between can be used for other applications such as conveying the coolant for the thermal management. The pillow-plate frame can be flat formed and the side walls folded upwards, thus enabling flexible and scalable production unimpaired by tooling investment costs. The crash-frame concepts presented here were successfully validated through simulation by using two different test scenarios: the static crushing load case (F = 200 kN) and the dynamic crash load case where a pole is intruded at a speed of 32 km/h. No damage of the internal battery modules occurs. In order to protect the thermal management and, in particular, the batteries against critical intrusion from the underfloor area, the battery housing is supplemented by an underride guard. This skid plate requires a careful balance of energy absorption capacity and high resistance to intrusion. Moreover, stiffness is a vital issue as the underride guard covers the entire surface of the battery modules. The system can be positively influenced by integration of corrugations. This part of the vehicle floor is also subject to high corrosive load what makes stainless steel to the material of choice. The model of a mechanical compression spring embedded in a double floor is a suitable physical operating principle and a cross-industry innovation derived from ship-building industry. Stainless steels can be cold formed to produce high-strength spring elements, fulfilling stiffness classes up to C1900 in accordance with DIN EN 10151 [9], which equates to tensile strengths of 1900 – 2000 MPa. Already since the 1850s the principle of double floor systems is mandatory in shipbuilding for transatlantic passenger transport because of safety reasons. During an underbody impact, this construction method causes the partially sharp-edged underfloor impactor to glance off, resulting in an angle-based reduction of impact forces. Established spring parameters from classic mechanical engineering, such as spring deflection and spring rate, can be used to configure the system. A significant advantage of spring elements is the principle of force distribution to surrounding elements as used in mattresses. It converts point transmission of force into a planar load. This general physical operating principle for spring elements can be transferred to different types of stainless-steel profiles.

682

Increased safety for battery electric vehicles by using heat-resistance stainless steels

4 Summary Innovative material concepts with a wide range of properties make an important contribution towards satisfying the complex issues of future mobility concepts. Thereby stainless steels represent a key enabler. The innovative combination of materials and manufacturing technologies can increase the acceptance and efficiency of future transportation concepts. Using the properties offered by stainless steels, the design demands of safety in combination with lightweight and sustainability represent not a contradiction in structural design. Further, stainless steels as high performance materials are applicable for high-temperature resistance which will be for battery-powered electric vehicles directly connected with passenger safety. It also offers significant improvements in areas of overall cost of mass production, carbon footprint and recycling. Because of consequently using stainless steels also a multi-generation usage of vehicle components can be mentioned as an outlook to future transportation engineering.

Bibliography 1. Karle, A.: Elektromobilität – Grundlagen und Praxis, Carl-Hanser-Verlag, 2015 2. N.N.: Fortschrittsbericht 2018 – Markthochlaufphase, Nationale Plattform Elektromobilität, Berlin, 05/2018 3. Müller, B., Zachäus, C., Meyer, G.: European strategic processes towards competitive, sustainable and user-friendly electrified road transport, 30th International electric vehicle Symposium, Stuttgart, 10.09.2017 4. Zhou, Q.: Electric Mobility in China – Developments, Opportunities & Challenges, 30th International electric vehicle Symposium, Stuttgart, 10.09.2017 5. DIN EN ISO 10088-2. Stainless steels - Part 2: Technical delivery conditions for sheet/plate and strip of corrosion resisting steels for general purposes; German version EN 10088-2:2014 6. Lindner, S.: Forta H-Series – Ultimate lightweight solutions with high strength austenitic materials, 16th International Symposium, Stuttgart, 2016 7. Ratte, E.: Wasserstoffinduzierte verzögerte Rissbildung austenitischer Stähle auf CrNi(Mn)- und Mn-Basis, dissertation IEHK RWTH Aachen, Shaker Verlag, Band 8/07, 2007 8. Bracke, L.: de Cooman, B.; Liebeherr, M.; Akdur, N.: Phase transformations in High Strength Austenitic FeMnCr Steels, Solid-solid phase transformations in inorganic materials, Volume 1, S.905-910, 2005 9. DIN EN ISO 10151. Stainless steel strip for springs - Technical delivery conditions; German version EN 10151:2002

683