New Developments and Applications in Sensing Technology [1 ed.] 3642179428, 9783642179426

Citation preview

Lecture Notes in Electrical Engineering Volume 83

Subhas Chandra Mukhopadhyay, Aimé Lay-Ekuakille, and Anton Fuchs (Eds.)

New Developments and Applications in Sensing Technology

ABC

Subhas Chandra Mukhopadhyay School of Engineering and Advanced Technology (SEAT) Massey University (Manawatu Campus) Palmerston North, New Zealand E-mail: [email protected] Aime Lay-Ekuakille Dipartimento d’Ingegneria dell’innovazione University of Salento Lecce, Italy E-mail: [email protected] Anton Fuchs Institute of Electrical Measurement and Measurement Signal Processing Graz University of Technology Graz, Austria E-mail: [email protected]

ISBN 978-3-642-17942-6

e-ISBN 978-3-642-17943-3

DOI 10.1007/978-3-642-17943-3 c 2011 Springer-Verlag Berlin Heidelberg  This work is subject to copyright. All rights are reserved, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965, in its current version, and permission for use must always be obtained from Springer. Violations are liable to prosecution under the German Copyright Law. The use of general descriptive names, registered names, trademarks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. Typeset & Cover Design: Scientific Publishing Services Pvt. Ltd., Chennai, India. Printed on acid-free paper 987654321 springer.com

Guest Editorial

This special issue titled “New Developments and Applications in Sensing Technology” in the book series of “Lecture Notes in Electrical Engineering” contains invited papers from renowned experts working in the field of sensing technology. A total of 17 chapters describe the advancement in the area of smart sensors and sensor networks design, measurement techniques, signal processing, and efficient algorithms in recent times. The 17 carefully selected chapters are extended versions of the conference papers presented at the 4th International Conference on Sensing Technology (ICST 2010), held at University of Salento, Leece, Italy from June 3 to 5, 2010. This special issue has focussed on different aspects of modern sensing technology, i.e. intelligent measurement, information processing, adaptability, recalibration, data fusion, validation, high reliability and integration of novel and high performance sensors. The aspects and methods are used for applications in material testing and analysis, communication, quality inspection, biomedical and environmental measurements. The selection of the chapters in this special issue reflects the range of requirements and suitable approaches for current challenges in sensing technology. While future interest in this field is ensured by the constant supply of emerging modalities, techniques, and engineering solutions, many of the basic concepts and strategies have already matured and now offer opportunities to build upon. We are very happy to be able to offer the readers of ”Lecture Notes in Electrical Engineering” such a diverse special issue both in terms of its topical coverage and geographic representation. We hope that the readers will find it interesting, thought provoking, and useful in their research and practical engineering work. We would like to extend our wholehearted thanks to all the authors who have contributed their work to this special issue. Subhas Chandra Mukhopadhyay, Guest Editor School of Engineering and Advanced Technology (SEAT), Massey University (Manawatu Campus) Palmerston North, New Zealand [email protected] Aime Lay-Ekuakille, Guest Editor Dipartimento d’Ingegneria dell’innovazione University of Salento Lecce, Italy [email protected] Anton Fuchs, Guest Editor Institute of Electrical Measurement and Measurement Signal Processing, Graz University of Technology Graz, Austria [email protected]

VI

Guest Editorial

Dr. Subhas Chandra Mukhopadhyay graduated from the Department of Electrical Engineering, Jadavpur University, Calcutta, India in 1987 with a Gold medal and received the Master of Electrical Engineering degree from Indian Institute of Science, Bangalore, India in 1989. He obtained the PhD (Eng.) degree from Jadavpur University, India in 1994 and Doctor of Engineering degree from Kanazawa University, Japan in 2000. During 1989-90 he worked almost 2 years in the research and development department of Crompton Greaves Ltd., India. In 1990 he joined as a Lecturer in the Electrical Engineering department, Jadavpur University, India and was promoted to Senior Lecturer of the same department in 1995. Obtaining Monbusho fellowship he went to Japan in 1995. He worked with Kanazawa University, Japan as researcher and Assistant professor till September 2000. In September 2000 he joined as Senior Lecturer in the Institute of Information Sciences and Technology, Massey University, New Zealand. He is working currently as an Associate professor with the School of Engineering and Advanced Technology of Massey University, New Zealand. His fields of interest include Sensors and Sensing Technology, Electromagnetics, control, electrical machines and numerical field calculation etc. He has authored over 200 papers in different international journals and conferences, edited nine conference proceedings. He has also edited seven special issues of international journals as guest editor and seven books with Springer-Verlag. He is a Fellow of IET (UK), a senior member of IEEE (USA), an associate editor of IEEE Transactions on Instrumentation and Measurements. He is in the editorial board of e-Journal on Non-Destructive Testing, Sensors and Transducers, Transactions on Systems, Signals and Devices (TSSD), Journal on the Patents on Electrical Engineering, Journal of Sensors. He is in the technical programme committee of IEEE Sensors conference, IEEE IMTC conference and IEEE DELTA conference. He was the Technical Programme Chair of ICARA 2004, ICARA 2006 and ICARA 2009. He was the General chair and co-chair of ICST 2005, ICST 2007, ICST 2008, IEEE Sensors 2008, ICST 2010. He has organized the IEEE Sensors conference 2009 at Christchurch, New Zealand as General Chair. He is co-editor in chief of the International Journal on Smart Sensing and Intelligent Systems (www.s2is.org). Aimé Lay-Ekuakille has a Master Degree in Electronic Engineering, a Master Degree in Clinical Engineering, a Ph.D in Electronic Engineering from Polytechnic of Bari, Italy. He has been technical manager of different private companies in the field of: Industrial plants, Environment Measurements, Nuclear and Biomedical Measurements; he was director of a Health & Environment municipal Department. He has been a technical advisor of Italian government for high risk plants. From 1993 up to 2001, he was adjunct professor of Measurements and control systems in the University of Calabria, University of Basilicata and Polytechnic of Bari. He joined the Department of Innovation Engineering, University of

Guest Editorial

VII

Salento, in September 2000 in the Measurement & Instrumentation Group. Since 2003, he became the leader of the scientific group; hence, he is the co-ordinator of Measurement and Instrumentation Lab in Lecce. He has been appointed as UE Commission senior expert for FP-VI (2005-2010). He is still: chair of IEEE-sponsored SCI/SSD Conference, member of Transactions on SSD and Sensors & Transducers Journal editorial board. He is Associate Editor of the International Journal on Smart Sensing and Intelligent Systems. He is currently organizing the next ICST2010 in Lecce, Italy. He is a member of the following boards and TCs: Association of the Italian Group of Electrical and Electronic Measurements (GMEE), SPIE, IMEKO TC19 Environmental Measurements, IEEE, IEEE TC-25 Medical and Biological Measurements Subcommittee on Objective Blood Pressure Measurement, IEEE-EMBS TC on Wearable Biomedical Sensors & System and included in different IPCs of conferences. Aimé Lay-Ekuakille is a scientific co-ordinator of different international projects. He has authored and co-authored more than 95 papers published in international journals, books and conference proceedings. His main researches are on Environmental and Biomedical instrumentation & measurements and, measurements for renewable energy. Anton Fuchs was born in Graz, Austria, in 1977. He received the Dipl.Eng. degree in telematics from Graz University of Technology in 2001 and joint the Institute of Electrical Measurement and Measurement Signal Processing, Graz University of Technology in 2002. He worked as a researcher, project manager, and lecturer and received the Doctoral degree in technical science in 2006 from Graz University of Technology. He was Visiting Researcher and Research Fellow at the Centre for Bulk Solids and Particulate Technologies, University of Wollongong, Australia in 2004 and in 2007/2008 respectively. In 2009 he received the venia docendi for “Process Instrumentation and Sensor Technology” from Graz University of Technology and became Associate Professor. Anton Fuchs is now with the Virtual Vehicle Competence Center in Graz, Austria. He is still Associate Professor and Distinguished Lecturer at Graz University of Technology. His main research interests include capacitive sensors and the measurement of transported material in industrial conveying processes. Anton Fuchs is author and co-author of more than 90 scientific papers and patents.

Table of Contents

Detection of Micro-cracks on Metal Surfaces Using Near-Field Microwave Dual-Behaviour Resonators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Julien Kerouedan, Patrick Qu´eff´elec, Philippe Talbot, C´edric Quendo, Alain Le Brun Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. Mason, A. Shaw, A.I. Al-Shamma’a Planar Electromagnetic Sensor for the Detection of Nitrate and Contamination in Natural Water Sources Using Electrochemical Impedance Spectroscopy Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M.A. Md Yunus, S.C. Mukhopadhyay Current Reconstruction Algorithms in Electrical Capacitance Tomography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M. Neumayer, H. Zangl, D. Watzenig, A. Fuchs Non-destructive Control of Metallic Plate with Magnetic Techniques . . . . L. Battaglini, P. Burrascano, A. Canova, F. Ficili, M. Ricci, D. Rossi, F. Sciacca Gas Sensing Characteristics of Pure and ZnO-Modified Fe2 O3 Thick Films . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N.K. Pawar, D.D. Kajale, G.E. Patil, S.D. Shinde, V.B. Gaikwad, Gotan H. Jain Design and Construction of a Configurable Full-Field Range Imaging System for Mobile Robotic Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . D.A. Carnegie, J.R.K. McClymont, A.P.P. Jongenelen, B. Drayton, A.A. Dorrington, A.D. Payne Cr2 O3 -doped BaTiO3 as an Ammonia Gas Sensor . . . . . . . . . . . . . . . . . . . . Gotan H. Jain, S.B. Nahire, D.D. Kajale, G.E. Patil, S.D. Shinde, D.N. Chavan, V.B. Gaikwad

1

15

39

65 107

123

133

157

Physical and Electrical Modeling of Interdigitated Electrode Arrays for Bioimpedance Spectroscopy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . M. Ibrahim, J. Claudel, D. Kourtiche, B. Assouar, M. Nadi

169

Water Quality Assessment through Smart Sensing and Computational Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . O. Postolache, P. Silva Gir˜ ao, J.M. Dias Pereira

191

X

Table of Contents

Multi-spectral Analytical Systems Using LIBS and LII Techniques . . . . . Satoshi Ikezawa, Muneaki Wakamatsu, Yury L’vovich Zimin, Joanna Pawlat, Toshitsugu Ueda

207

Electromechanical Sensors Based on Carbon Nanotube Networks and Their Polymer Composites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . P. Slobodian, P. Riha, R. Olejnik

233

Novel Planar Interdigital Sensors for Detection of Domoic Acid in Seafood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A.R. Mohd Syaifudin, K.P. Jayasundera, S.C. Mukhopadhyay

253

Nano-Biosensor Development for Biomedical and Environmental Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D.M.G. Preethichandra, E.M.I. Mala Ekanayake

279

Nondestructive Evaluations of Iron-Based Materials by Using AC and DC Electromagnetic Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Koji Yamada, Jiaoliang Luo, Masato Enokizono

293

STACK: Sparse Timing of Algorithms Using Computational Knowledge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vasanth Iyer, S. Sitharama Iyengar, Garmiela Rama Murthy, Kannan Srinathan, Mandalika B. Srinivas, Regeti Govindarajulu A New Approach to Estimation of Protein Networks for Cell Cycle Based on Least-Squares Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Takehito Azuma, Masachika Kurata, Noriko Takahashi, Shuichi Adachi Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

305

321

333

Detection of Micro-cracks on Metal Surfaces Using Near-Field Microwave Dual-Behaviour Resonators Julien Kerouedan1,2, Patrick Quéffélec1, Philippe Talbot1, Cédric Quendo1, and Alain Le Brun2 1

Lab-STICC (UMR 3192), Université de Bretagne Occidentale, CS 93836, 6 avenue Le Gorgeu, 29238 Brest Cedex 3, France 2 EDF R&D / STEP, 6 quai Watier, BP 49, 78401 Chatou Cedex, France [email protected]

Abstract. The aim of this paper is to demonstrate that micro-cracks at the surface of metals can be detected and imaged by using near-field micro-wave resonators. It deals with two novel sensors: a first-order dual-behaviour resonator (DBR) filter and a first-order DBR filter with an open-ing in the ground plane. Measurements were mainly carried out on a stainless steel mock-up with several EDM (i.e. manufactured by Electron Discharge Machining) rectangular surface notches presenting widths between 0.1 and 0.3 mm and depths between 0.5 and 3 mm. The results presented here show the high sensitivity of the DBR probes and their ability to differentiate between notches of different depths and notches of different widths. Keywords: electromagnetic sensors, non-destructive testing (NDT), microwaves, near-field resonator.

1 Introduction The fatigue and ageing of metal materials under operation conditions are major concerns in energy production plants. An early and non-destructive diagnostic of surface defects would allow one to carry out relevant preven-tive maintenance operations avoiding unnecessary replacements or early repairing of healthy components. Today, most of the automated non-destructive testing (NDT) solutions available to detect the surface-breaking defects are based on ultrasonic [1] or Eddy current techniques [2]. Despite their high sensitivity and resolution, they are unable to meet all the requirements of every real situation. Eddy current testing sensitivity to different external parameters sometimes makes signal analysis difficult, and ultrasonic techniques are not always suitable for the detection of small depth defects located near the inspection surface. Consequently, it sounded us relevant to evaluate the potential of microwavebased techniques to detect the surface defects with a depth shallower than 3 mm. Over the last years, microwave far- and near-field approaches were investigated to detect surface defects. With a far-field characterization [3], the spatial resolution is of the order of a half wavelength (λ/2). So, to detect micro-cracks it is necessary to work at very high frequencies, which causes high measurement equipment costs. In addition, at very high fre-quencies, signal-to-noise ratio (SNR) issues can appear. As a result, because of their poor spatial resolution, the far-field methods are un-suitable for S.C. Mukhopadhyay et al. (Eds.): New Developments and Appl. in Sen. Tech., LNEE 83, pp. 1–13. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

2

J. Kerouedan et al.

detecting small depth defects. On the other hand, the use of near-field techniques permits to significantly improve the spatial resolution. Indeed, with a near-field characterization, the spatial resolution primarily depends on two parameters: the size of the probe end and the lift-off distance (the lift-off distance is the distance between the probe end and the sample under test). Consequently, many near-field NDT methods have been recently developed to detect surface and subsurface defects in various materials over a wide range of frequencies. Two main categories of near-field techniques are reported in the literature: the methods based on reflection coefficient measurements and the methods relying on the use of resonators. The first type of near-field techniques consists in measuring the reflec-tion coefficient at the end of either an open-ended rectangular waveguide [4-6] or an open-ended coaxial line [7, 8]. The presence of a crack near the aperture of these open-ended structures induces changes in the reflection coefficient’s magnitude and phase. Due to the use of wellknown open structures of propagation, these methods are easily implemented. However, the detection of a micro-crack requires the measurement of magnitude variations of about the hundredth of dB and of phase variations of a few degrees. Thus, the detection of a surface micro-crack with this first near-field technique is achievable but it requires high levels of accuracy and re-liability for the measurement equipment to detect the very small magnitude and phase variations caused by the micro-crack. The second type of near-field techniques relies on the use of resonators to measure the changes in the resonance frequency and quality factor induced by the interaction with the surface flaws. These variations can be detected using a network analyzer. The resonators presented in the literature are terminated with either a tip [9, 10] or an electric or magnetic dipole [11] in order to guide the radiation issued from the resonator toward the sample under characterization. Many studies have highlighted the link between the spatial resolution of near-field resonators, the size of the probe end and the lift-off distance, and have showed that methods of material characterization resting on the use of resonators make it possible to obtain a spatial resolution lower than the micron at microwave frequencies [9]. These resonant techniques have proven to successfully image defects and non-uniformities in various metals [9-11]. The two major advantages of using near-field resonators to detect surface defects are the dependence of the spatial resolution on the probe end size and the lift-off distance, and the direct detection of variations in the resonance frequency and quality factor by using a network analyzer. The aim of our study was the detection of several rectangular EDM (i.e. manufactured by Electron Discharge Machining) notches with widths between 0.1 and 0.3 mm and depths shallower than 3 mm on the surface of a 20 mm thick austenitic stainless steel plate. After the analysis of the existing microwave-based techniques, it appeared to be interesting to investigate how well a near-field resonant probe could detect the rectangular EDM notches. In order to minimize the costs of the probe fabrication and measurement equipment, we focused our research on the realization of microstrip sensors with resonance frequencies of about 10 GHz. In addition, for practical reasons, we imposed a lift-off distance greater than or equal to 50 μm. To detect and image the surface defects, we developed a novel detection technique using the reflection (S11) and transmission coefficients (S21) of a dual-behaviour resonator (DBR) band-pass filter [12-14]. The main feature of this device is the high selectivity of the DBR resonator, which permits an easy measurement of the frequency shift with a network analyzer. In addition, the high sensitivity of the DBR probe makes it possible to detect micro-cracks at operating frequencies near 10 GHz.

Detection of Micro-cracks on Metal Surfaces

3

This paper presents the design and the results obtained with two original near-field resonant probes based on dual-behaviour resonators with and without an opening in the ground plane. Measurements carried out on different notches (with widths between 0.1 and 0.3 mm and depths between 0.5 and 3 mm) are presented, showing the influence of the depth and width of the notch on the behaviour of the DBR sensors. The sensitivity of the two probes is discussed.

2 First-Order DBR Filter Probe 2.1 Description of the Probe A first-order DBR band-pass filter results from the association of two different parallel band-stop structures [12, 13]. Fig. 1 shows (a) the design and (b) the reflection (S11) and transmission parameters (S21) of this first-order DBR band-pass filter. The filter that we studied was realized on an alumina substrate. It included two openended stubs. The low-frequency stub (LF stub for low-frequency stub) of length, l1, and characteristic impedance, Zs1, brings a transmission zero in the lower attenuated band, whereas the high-frequency stub (HF stub for high-frequency stub), of length, l2, and characteristic impedance, Zs2, brings a transmission zero in the upper attenuated band. A band-pass response between the lower and upper rejected bands is created by constructive recombination (Fig. 1.b). (a)

Zs1, l1 Z0

Z0

Zs2, l2

S11 (dB) and S21 (dB)

(b)

0 -10 -20 -30 -40

S21

S11

-50 5 6 7 8 9 10 11 12 13 14 15

Frequency (GHz)

Fig. 1. First-order DBR band-pass filter: (a) design, (b) reflection (S11) and transmission (S21) parameters

4

J. Kerouedan et al.

The characterization of a metallic component is realized by moving the DBR filter probe at a constant lift-off distance d above the surface (the lift-off distance d is defined as the distance between the probe end and the surface). Fig. 2 shows (a) a schematic and (b) a photograph of the first-order DBR probe over a notch. (a) LF stub Metallic strip Port 1

Substrate Port 2 HF stub

Z WHF 0 Y

X

h W

d Notch Metal

(b)

Fig. 2. (a) Schematic and (b) photograph of the first-order DBR band-pass filter probe over a notch; W and h represent the width and depth of the notch, d is the lift-off distance, and the length L of the notch is in the y-direction

The principle of micro-crack detection consists in measuring the changes in the capacitive coupling created between the HF stub (of width WHF) and the metal sample under test. The probe-metal coupling can be described by a coupling capacitance [14, 15] that decreases when a defect is located below the HF stub of the DBR probe. This variation of capacitive coupling induces an increase in the frequency of the transmission zero associated with the high attenuated band, fHF, but also an increase in the central frequency, f0, i.e., the frequency associated with the minimum of the filter reflection parameter S11. On the other hand, the frequency of the transmission zero associated with the low attenuated band, fLF, remains unchanged due to a lack of LF stub-defect interaction. The influence of the probe-metal coupling on the behaviour of the first-order DBR filter sensor was studied in detail in [14]. This study highlighted that for a given coupling capacitance the shifts in fHF are more important than the shifts in f0. The frequency fHF obtained by measuring the transmission parameter S21 of the DBR filter is thus the best indicator of the changes in the probe-metal interaction.

Detection of Micro-cracks on Metal Surfaces

5

2.2 Experimental Results 2.2.1 Measurements Carried Out on a Stainless Steel Mock-Up Measurements were carried out on a 20 mm thick austenitic stainless steel plate (conductivity σ = 1.4 × 106 S.m-1) with several EDM rectangular notches. The experimental setup is shown in Fig. 3. It consists of an Agilent PNA E8364A (45 MHz – 50 GHz) network analyzer and a motorized three-axis displacement device holding the probe.

Computer

Network analyzer

Probe

3-axes displacement device Steel plate with EDM notches Fig. 3. Photograph of the experimental setup

0 Notch Faultless metal

S21 (dB)

-5 Stainless steel -10 -15

Notch

-20 -25

Faultless metal -30 12,5 12,6 12,7 12,8 12,9 13,0 13,1 13,2

Frequency (GHz) Fig. 4. Measured transmission parameter S21 of the 50 μm wide HF stub filter set at d = 50 μm over a faultless metal plate and over a 200 μm wide, 3 mm deep and 10 mm long notch

6

J. Kerouedan et al.

ΔfHF (MHz)

The first experiments were focused on the detection of a 200 μm wide, 3 mm deep and 10 mm long EDM notch. Fig. 4 shows the transmission parameter (S21) responses measured with a 50 μm wide HF stub filter, for a lift-off distance d = 50 μm, when the probe was set over a faultless metal plate and over the middle of the notch. The presence of the notch causes an increase in fHF (ΔfHF = 51 MHz) and f0 (Δf0 = 26 MHz) (data not shown). In order to image the notch, the variations of the HF stub frequency (ΔfHF) were measured during several 1D scans over the notch. Fig. 5 shows four 1D scans in the x-direction with a 10 μm incrementing step measured with the 50 μm wide HF stub DBR filter for four different values of d between 50 and 150 μm. This figure highlights that ΔfHF is the highest in the middle of the notch, and that the DBR filter sensitivity decreases when the distance d increases. Additional measurements were carried out over a 200 μm wide, 1 mm deep and 10 mm long notch and over a 200 μm wide, 0.5 mm deep and 10 mm long notch in order to examine the influence of the notch depth on the response of the first-order DBR probe. Fig. 6 shows the values of ΔfHF measured with the 50 μm wide HF stub filter when 1D scans are performed in the x-direction over the 0.5, 1 and 3 mm deep notches (probe incrementing step = 10 μm, and lift-off distance d = 50 μm). The three plots presented in Fig. 6 differ by their height i.e. by their maximum ΔfHF value. The maximum ΔfHF values associated with the 0.5, 1 and 3 mm deep notches are 42, 45, and 51 MHz, respectively. These experimental results highlight the influence of the notch depth on the ΔfHF variations and show that the first-order DBR probe enables us to differentiate between notches of different depths in the stainless steel mock-up. In order to examine the influence of the notch width on the response of the firstorder DBR probe, other experiments were performed over notches of 100 μm and 300 μm width while maintaining the other dimensions unchanged: 1 mm depth and 10 mm length. 60 55 50 45 40 35 30 25 20 15 10 5 0

Stainless steel

-500 -400 -300 -200 -100 0

d = 50 µ m d = 75 µ m d = 100 µ m d = 150 µ m

100 200 300 400 500

x- position (µm)

Fig. 5. 1D scans (x-direction) over a stainless steel plate with a 200 μm wide, 3 mm deep and 10 mm long notch measured with the 50 μm wide HF stub DBR filter for four different lift-off distances d. Variations of the HF stub frequency (ΔfHF) as a function of the x-position; x = 0 corresponds to the middle of the notch in the stainless steel mock-up.

ΔfHF (MHz)

Detection of Micro-cracks on Metal Surfaces

60 55 50 45 40 35 30 25 20 15 10 5 0

Stainless steel

7

h = 3 mm h = 1 mm h = 0.5 mm

-500 -400 -300 -200 -100 0 100 200 300 400 500

x- position (µm)

Fig. 6. 1D scans (x-direction) over three 200 μm wide and 10 mm long notches differing by their depth h measured with the 50 μm wide HF stub DBR filter for a lift-off distance d = 50 μm. Variations of the HF stub frequency (ΔfHF) as a function of the x-position; x = 0 corresponds to the middle of the notches in the stainless steel mock-up.

ΔfHF (MHz)

The experimental results associated with these two notches were compared to those obtained during the characterization of the 200 μm wide, 1 mm deep and 10 mm long notch. Fig. 7 shows the values of ΔfHF measured with the 50 μm wide HF stub filter when 1D scans are performed in the x-direction over the 0.1, 0.2 and 0.3 mm wide notches (probe incrementing step = 10 μm, and lift-off distance d = 50 μm). 80 75 70 65 60 55 50 45 40 35 30 25 20 15 10 5 0

Stainless steel

w = 0.3 mm w = 0.2 mm w = 0.1 mm

-500 -400 -300 -200 -100 0 100 200 300 400 500

x- position (µm)

Fig. 7. 1D scans (x-direction) over three 1 mm deep and 10 mm long notches differing by their width W measured with the 50 μm wide HF stub DBR filter for a lift-off distance d = 50 μm. Variations of the HF stub frequency (ΔfHF) as a function of the x-position; x = 0 corresponds to the middle of the notches in the stainless steel mock-up.

8

J. Kerouedan et al.

The three bell shape curves presented in Fig. 7 differ by their height and by their width. The maximum ΔfHF values associated with the 0.1, 0.2 and 0.3 mm wide notches are 22, 45, and 71 MHz, respectively, and the widths of the curves associated with the 0.1, 0.2 and 0.3 mm wide notches are about 330, 430 and 540 μm, respectively. These experimental results show the ability of the first-order DBR probe to differentiate between notches of different widths in the stainless steel mock-up. 2.2.2 Measurements Carried Out on an Aluminium Mock-Up Measurements were carried out on a 20 mm thick aluminium plate (conductivity σ = 37.5 × 106 S.m-1) with several EDM rectangular notches, using the same experimental setup as in the study with the stainless steel mock-up (§ 2.2.1, Fig. 3). Fig. 8 shows the values of ΔfHF measured with the 50 μm wide HF stub filter when 1D scans are performed in the x-direction over a 100 μm wide, 1 mm deep and 1 mm long notch and over a 100 μm wide, 0.5 mm deep and 1 mm long notch located on the aluminium mock-up (probe incre- menting step = 10 μm, and lift-off distance d = 50 μm). 30

Aluminium

ΔfHF (MHz)

25

h = 1 mm h = 0.5 mm

20 15 10 5 0 -500 -400 -300 -200 -100 0 100 200 300 400 500

x- position (µm)

Fig. 8. 1D scans (x-direction) over two 100 μm wide and 1 mm long notches differing by their depth h measured with the 50 μm wide HF stub DBR filter for a lift-off distance d = 50 μm. Variations of the HF stub frequency (ΔfHF) as a function of the x-position; x = 0 corresponds to the middle of the notches in the aluminium mock-up.

The two plots presented in Fig. 8 differ by their height i.e. by their maximum ΔfHF value. The maximum ΔfHF values associated with the 0.5 and 1 mm deep notches are 14 and 21 MHz, respectively. These experimental results and those obtained with the austenitic stainless steel mock-up (§ 2.2.1, Fig. 6) highlight the ability of the firstorder DBR probe to differentiate between notches of different depths. Fig. 9 shows the values of ΔfHF measured with the 50 μm wide HF stub filter when 1D scans are performed in the x-direction over a set of three notches of the same

Detection of Micro-cracks on Metal Surfaces

9

length and depth (1 mm), but of different widths (0.1, 0.2 and 0.5 mm). The lift-off distance between the probe and the surface of the aluminium mock-up is d = 50 µm (probe incrementing step = 10 μm). 160

ΔfHF (MHz)

140

Aluminium

120

w = 0.5 mm w = 0.2 mm w = 0.1 mm

100 80 60 40 20 0 -500 -400 -300 -200 -100 0 100 200 300 400 500

x- position (µm)

Fig. 9. 1D scans (x-direction) over three 1 mm deep and 1 mm long notches differing by their width W measured with the 50 μm wide HF stub DBR filter for a lift-off distance d = 50 μm. Variations of the HF stub frequency (ΔfHF) as a function of the x-position; x = 0 corresponds to the middle of the notches in the aluminium mock-up.

The three curves presented in Fig. 9 differ by their height and by their width. The maximum ΔfHF values associated with the 0.1, 0.2 and 0.5 mm wide notches are 21, 37, and 110 MHz, respectively, and the widths of the curves associated with the 0.1, 0.2 and 0.5 mm wide notches are about 370, 460 and 780 μm, respectively. These measurement results and those obtained with the stainless steel mock-up (§ 2.2.1, Fig. 7) highlight the ability of the first-order DBR probe to differentiate between notches of different widths.

3 First-Order DBR Filter Probe with Open Ground Plane 3.1 Description of the Probe This second probe was developed to increase the lift-off distance d. In a previous work [14], we have demonstrated the feasibility of increasing the lift-off distance by increasing the HF stub width of the DBR filter. For example, we have shown that a 100 μm wide HF stub first-order DBR filter allows one to detect a 200 μm wide, 3 mm deep and 10 mm long EDM notch on the surface of a stainless steel plate until d = 150 μm [14]. To allow the notch detection for d > 150 μm, our idea was to create an opening in the ground plane below the HF stub of the first-order DBR filter so as to increase the radiation of the sensor. Simulations were performed with the finite element (FEM)-based commercial software package HFSSTM in order to determine the optimal shape and size of this

10

J. Kerouedan et al.

opening. For a 100 μm wide HF stub filter, the best compromise between sensitivity and resolution was obtained for a rectangular opening with a width Wo = 1 mm. Figs. 10 and 11 show respectively a schematic and two photographs of the DBR probe with open ground plane.

Fig. 10. Schematic of the first-order DBR filter probe with open ground plane

(a) BF stub

HF stub (b) Ground plane

Opening below the HF stub

Fig. 11. (a) Top view photograph and (b) bottom view photograph of the first-order DBR filter probe with open ground plane

3.2 Experimental Results All the measurements were carried out with the same experimental setup (Fig. 3) and the same stainless steel mock-up as in the study with the first-order DBR filter probe. Fig. 12 shows the transmission parameter (S21) responses measured using a 100 μm wide HF stub filter with open ground plane, for a lift-off distance d = 300 μm, when the probe was set over a faultless metal plate and over the middle of a 200 μm wide, 3 mm deep and 10 mm long EDM notch. This figure highlights an increase in fHF (ΔfHF = 33 MHz) induced by the presence of the notch.

Detection of Micro-cracks on Metal Surfaces

11

Fig. 13 gives the variations ΔfHF measured using the 100 μm wide HF stub filter with open ground plane when 1D scans are performed in the x-direction over the 200 μm wide, 3 mm deep and 10 mm long notch with a 50 μm incrementing step and for six different values of d between 100 and 650 μm. This figure shows that the DBR probe with open ground plane can easily detect the notch until d = 400 μm while it is impossible with the classic DBR filter probe (§ 2.2.1, Fig. 5).

0

Stainless steel

Notch Faultless metal

S21 (dB)

-5 -10 -15

Notch

-20 -25 Faultless metal -30 13.6 13.7 13.8 13.9 14.0 14.1 14.2

Frequency (GHz) Fig. 12. Measured transmission parameter S21 of the 100 µm m wide HF stub filter with open ground plane set at d = 300 µm m over a faultless metal plate and over a 200 µm m wide, 3 mm deep and 10 mm long notch

120

d = 100 µm d = 180 µm d = 250 µm d = 300 µm d = 400 µm d = 650 µm

Stainless steel

ΔfHF (MHz)

100 80 60 40 20 0 -2000 -1500 -1000 -500

0

500 1000 1500 2000

x- position (µm)

Fig. 13. 1D scans (x-direction) over a stainless steel plate with a 200 μm wide, 3 mm deep and 10 mm long notch measured using the 100 μm wide HF stub DBR filter with open ground plane for six different lift-off distances d. Variations of the HF stub frequency (ΔfHF) as a function of the x-position; x = 0 corresponds to the middle of the notch in the stainless steel mock-up.

12

J. Kerouedan et al.

At last, Fig. 14 shows the values of ΔfHF measured using the 100 μm wide HF stub filter with open ground plane when 1D scans are performed in the x-direction over a 200 μm wide, 3 mm deep and 10 mm long notch and over a 200 μm wide, 1 mm deep and 10 mm long notch, with a 50 μm incrementing step and d = 300 μm. The difference between the maximum ΔfHF values associated with the 1 mm deep notch (ΔfHF = 22 MHz) and the 3 mm deep notch (ΔfHF = 33 MHz) illustrates the influence of the notch depth on the ΔfHF variations. 40

ΔfHF (MHz)

35

Stainless steel

h = 3 mm h = 1 mm

30 25 20 15 10 5 0 -2000 -1500 -1000 -500

0

500 1000 1500 2000

x- position (µm)

Fig. 14. 1D scans (x-direction) over two 200 μm wide and 10 mm long notches differing by their depth h measured using the 100 μm wide HF stub DBR filter with open ground plane, for a lift-off distance d = 300 μm. Variations of the HF stub frequency (ΔfHF) as a function of the xposition; x = 0 corresponds to the middle of the notches in the stainless steel mock-up.

4 Conclusion A near-field microwave NDT method using dual-behaviour resonator filters was investigated in order to detect surface defects in metals. The detection principle was validated theoretically and experimentally. The experimental results obtained with the stainless steel mock-up and the aluminium mock-up showed the good spatial resolution and the high sensitivity of the DBR filter probes. In addition, the measurements carried out on EDM notches highlighted: 1) the influence of the width and depth of the notch on the HF stub frequency variations (ΔfHF) and 2) the enhancement of the detection sensitivity by using a first-order DBR filter with open ground plane. Further investigations will be aimed at studying the dependence of the ΔfHF variations on the dimensions (depth, width and length) of the notch. We also plan to evaluate the potential of the DBR probes to detect fatigue cracks and stress corrosion flaws.

Detection of Micro-cracks on Metal Surfaces

13

References 1. Krautkrämer, J., Krautkrämer, H.: Ultrasonic Testing of Materials, 4th edn. Springer, Berlin (1990) 2. Moore, P.O.: Nondestructive Testing Handbook (Electromagnetic Testing), 3rd edn., vol. 5. American Society for Nondestructive Testing (ASNT), Colombus (2004); Udpa, S.S. (technical ed.) 3. Shirai, H., Sehiguchi, H.: A simple crack depth estimation method from backscattering response. IEEE Transactions on Instrumentation and Measurement 53, 1249–1253 (2004) 4. Yeh, C.-Y., Zoughi, R.: A novel microwave method for detection of long surface cracks in metals. IEEE Transactions on Instrumentation and Measurement 43, 719–725 (1994) 5. Ghasr, M.T., Carroll, B., Kharkovsky, S., Austin, R., Zoughi, R.: Millimeter-wave differential probe for nondestructive detection of corrosion precursor pitting. IEEE Transactions on Instrumentation and Measurement 55, 1620–1627 (2006) 6. McClanahan, A., Kharkovshy, S., Maxon, A.R., Zoughi, R., Palmer, D.D.: Depth evaluation of shallow surface cracks in metals using rectangular waveguides at millimeter-wave frequencies. IEEE Transactions on Instrumentation and Measurement 59, 1693–1704 (2010) 7. Wang, Y., Zoughi, R.: Interaction of surface cracks in metals with open-ended coaxial probes at microwave frequencies. Materials Evaluation 58, 1228–1234 (2000) 8. Ju, Y., Saka, M., Uchimura, Y.: Evaluation of the shape and size of 3D cracks using microwaves. NDT&E International 38, 726–731 (2005) 9. Tabib-Azar, M., Su, D.-P., Pohar, A., Leclair, S.R., Ponchak, G.: 0.4 μm spatial resolution with 1 GHz (λ = 30 cm) evanescent microwave probe. Review of Scientific Instruments 70, 1725–1729 (1999) 10. Wei, T., Xiang, X.-D., Wallace-Freedman, W.G., Schultz, P.G.: Scanning tip microwave near-field microscope. Applied Physics Letters 68, 3506–3508 (1996) 11. Wang, R., Li, F., Tabib-Azar, M.: Calibration methods of a 2 GHz evanescent microwave magnetic probe for noncontact and nondestructive metal characterization for corrosion, defects, conductivity and thickness nonuniformities. Review of Scientific Instruments 76, 54701 (2005) 12. Quendo, C., Rius, E., Person, C.: Narrow bandpass filters using dual-behavior resonators. IEEE Transactions on Microwave Theory and Techniques 51, 734–743 (2003) 13. Quendo, C., Rius, E., Person, C.: Narrow bandpass filters using dual-behavior resonators based on stepped-impedance stubs and different-length stubs. IEEE Transactions on Microwave Theory and Techniques 52, 1034–1044 (2004) 14. Kerouedan, J., Quéffélec, P., Talbot, P., Quendo, C., De Blasi, S., Le Brun, A.: Detection of micro-cracks on metal surfaces using near-field microwave dual-behavior resonator filters. Measurement Science and Technology 19, 105701 (2008) 15. Kleismit, R.A., Kazimierczuk, M.K., Kozlowski, G.: Sensitivity and resolution of evanescent microwave microscope. IEEE Transactions on Microwave Theory and Techniques 54, 639–647 (2006)

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design A. Mason, A. Shaw, and A.I. Al-Shamma’a Liverpool John Moores University, Liverpool, United Kingdom

Abstract. There is a growing trend in the use of intelligent Wireless Sensor Networks (WSNs) for a wide range of applications. In the early part of the decade the underlying hardware was largely in prototype form and used for small scale demonstration systems, but there is now growing interest in applications which are commercially viable. This work began on the premise that the sensor hardware has gradually become smaller, yet there are still a few peripheral components which are lagging behind; namely the battery and antenna. Here, a novel antenna design is presented; this antenna is of a practical size for use in WSNs, whilst also offering improved energy consumption over commonly used monopole antennas.

1 Introduction Antennas are critical to the operation of wireless communication systems such as those used for radio, television, and mobile phones. They are often taken for granted by an end user of such a product – many consumers are blissfully unaware of how much antenna design can impact on device performance, size and energy consumption. Antenna design is often forgotten in WSNs since the devices are deployed in close proximity to one another (i.e. with a separation of 10m or less). As a result, device communication with simple wire antennas is a simple and affordable solution, but is not necessarily efficient. This leads to data corruption during wireless transmission which can result in three possible scenarios: • Loss; the data is irreparable and is lost forever – in this case the energy put into capturing, processing and transmitting the data is wasted. • Recovery; some protocols may allow data recovery, implying that there is a permanent data overhead which incurs additional energy consumption. • Retransmission; important data may be repeatedly transmitted until the intended message is correctly received – this ensures reliability, but leads to wasted energy. Therefore, it is desirable to have a system which minimises data corruption in order to improve efficiency, particularly when one takes into account the fact that data transmission from a typical sensor node consumes three times more energy than data processing alone [1]. In addition to the energy problem, the physical form of the standard monopole antenna is considered to be unsuitable for many applications. In particular the authors S.C. Mukhopadhyay et al. (Eds.): New Developments and Appl. in Sen. Tech., LNEE 83, pp. 15–37. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

16

A. Mason, A. Shaw, and A.I. Al-Shamma’a

have been involved in the use of WSNs for inventory management [2, 3] and spot weld monitoring [4]. In these situations it is undesirable for the sensor node, or mote, antenna to protrude from the object to which it is attached, since both applications involve high speed movement which could damage or destroy the antenna should it be snagged. The devices used as part of the authors work are the commonly known MicaZ motes [1], one of which is shown in Figure 1. One can clearly see here that the monopole antenna protrudes some way from the mote; the length of the protrusion is linked to the operating frequency of the mote, and therefore its wavelength (λ) [5]. In this case the operating frequency is 2.45GHz, and λ = 122mm. Monopole antennas are typically λ/4 in length [6] and there is no exception here as the MicaZ antenna measures approximately 40mm1. The initial thought in this situation was to simply flatten the antenna and effectively shield it from damage. This scenario is shown clearly in Figure 2.

Fig. 1. Berkeley MicaZ mote with monopole antenna attached

Fig. 2. Mote antenna held in place with a nylon cable tie to prevent snagging 1

Note that λ/4 = 31mm at 2.45GHz, but the MicaZ antenna includes a sheath which is slightly longer than the antenna so as to offer limited protection.

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

17

A simple experiment, using one mote as a transmitter and another as a receiver, demonstrated that flattening the antenna caused significant signal strength degradation. When the transmitter and receiver are separated by 1m of air, the signal strength was found to be 10dBm less if the antenna was flattened, compared to if the antenna was in its normal position (see Figure 1). As a result of these findings it was thought that designing a new antenna for the motes might be a more effective solution, since significant losses in signal strength ultimately lead to data loss.

2 Designing a New Antenna 2.1 Antenna Requirements Since industry disliked the idea of the existing MicaZ monopole antenna alternative types were considered. Wire antennas such as dipoles are typically in the order of λ/2 in length, and loop antennas often have a circumference equal to λ. We can see that dipole and loop antennas would be larger than the standard monopole antenna supplied with the MicaZ and therefore likely to be an even greater concern for industrial use. Smaller sizes are possible for loops and dipoles, but they do not make effective radiators [7, 8]. The best remaining option was a PCB antenna. With size and practicality being major concerns of industry it seemed that a low profile PCB antenna would be ideal. It was thought that such an antenna would be suitable for retrofitting to the current MicaZ motes, and in the future could possibly be integrated with the mote circuitry in a combined PCB design. In order to facilitate this, an aim was set of creating an antenna no greater than the size of MicaZ PCB (i.e. - 57mm × 32mm). Although PCB antennas do have their advantages, it is noted in literature that they tend to suffer from a narrow impedance bandwidth, quite often in the order of just a few percent [5]. The impedance bandwidth [9] refers to the ability of an object to absorb or transmit energy into its surroundings; in the case of antennas, the later is desirable. Impedance bandwidth can be calculated using Equation 1, where fu is the upper operating frequency, fl is the lower operating frequency and f0 is the centre frequency. fu and fl refer to the points where the energy transmitted by an antenna is ≥ 88.9%. In some texts this is also referred to as the point where the voltage standing wave ratio (VSWR) is ≤ 2 [10], and describes the range over which antennas are effective radiators.

⎛ f − fl ⎞ ⎟⎟ × 100 Bandwidth (%) = ⎜⎜ u ⎝ f0 ⎠

(Equation 1)

For the MicaZ mote, fl = 2.485GHz, fu = 2.400GHz and f0 = 2.443GHz [1], since the devices support multiple frequency channels for reduced interference. These figures lead to a minimum impedance bandwidth requirement of 3.48%. In addition to these requirements it was also thought that the antenna should have good directional properties (i.e. seek to radiate equally well in all directions). In literature this is often better defined as directivity, but this parameter is often difficult to quantify accurately, therefore this work takes a qualitative approach. Low directivity is critical for WSNs since it is often impossible to control the orientation of nodes during

18

A. Mason, A. Shaw, and A.I. Al-Shamma’a

deployment – in mobile applications (e.g. inventory management) orientation may also vary significantly with time. Antennas with only a single plane of polarisation cannot communicate well (if at all) with those orientated perpendicular to themselves [11]; monopole and dipole antennas suffer noticeably from this issue. To summarise before continuing, the new antenna was required to: • • • •

operate at the centre frequency (f0) 2.443GHz. have an impedance bandwidth greater than 3.48%. be no larger than 57mm×32mm×1.6mm. have a low directivity.

2.2 Coplanar Waveguide (CPW) Antenna During a review of literature relating to antenna design, it was discovered that Nithisopa et al [12] had designed a broadband co-planar waveguide (CPW) fed slot antenna which, in simulations, had proven suitable for use over the range of approximately 2.35-2.70GHz, resulting in an impedance bandwidth of 14%. The term CPW refers to the way in which the antenna is fed; two parallel slots are cut into a copper surface to act as a transmission line feed to the radiating elements of the antenna itself – this is illustrated in Figure 3. The radiating elements come in many different forms, although it appears that the slot type is popular. The copper surrounding the feed slots acts as a ground plane which promotes more uniform radiation than one would experience with similar structures such as patch antennas [5, 13, 14]. Based upon the work conducted by Nithisopa, an Ansoft HFSS [15] model was created as shown in Figure 3. The model was set up by following strict guidelines [16] provided by the developer of HFSS for the creation of CPW models. Table 1 gives information relating to the dimensions illustrated in Figure 3. Dimensions W1 and W2 are of particular importance in impedance matching the antenna to a typical 50Ω transmission line. Impedance matching is vital in antenna design in order to ensure that as much power as possible from the radio transceiver is transferred to the propagation medium via the antenna [17]. Poor matching leads to power being reflected by the antenna back toward the transceiver, resulting in reduced transmission range, wasted energy and potential damage to the transceiver itself. Figure 4 shows the difference in simulated performance as a result of using FR4 instead of Duroid substrate, as in Nithisopa’s work. The reason for changing substrate was simply a case of using materials to hand at the time for prototype manufacture, but one can see that the increase in dielectric constant (εr) reduces the impedance bandwidth. For Duroid εr ≈ 2, but for FR4 εr ≈ 4. Despite the decrease in impedance bandwidth FR4 still resulted in an impedance bandwidth – calculated to be 10% – far exceeding the requirements for this application. Table 1. CPW dimensions (mm) h

pcbX

pcbY

W1

W2

H1

H2

L1

1.6

90.0

45.0

0.5

2.4

23.0

10.5

39.0

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

19

Fig. 3. CPW antenna structure and dimensions

Fig. 4. Simulated reflected power for FR4 and Duriod substrates

Another interesting property of Nithisopa’s CPW antenna was its low directivity which is a significant advantage of CPW design when one considers other types of PCB antenna. Balanis [18] gives a guide to constructing a simple patch antenna, a structure mentioned in passing earlier. This structure consists of a rectangular conductive patch above a larger conductive ground plane. The two conductive layers are separated by a dielectric substrate, as shown in Figure 5. A comparison of the CPW antenna and a patch antenna created using the Balanis guide shows that the CPW antenna has a favourable radiation pattern for applications requiring low directivity; this is evidenced in Figure 6. The xz and yz planes are of particular interest at the 180° position where the patch antenna experiences attenuation in the order of 20dB – this is a direct result of the ground plane preventing transmission in this direction. The downside for the CPW antenna is increased attenuation in the plane of the PCB when compared with the patch antenna. However this is a reasonable compromise as the loss is much smaller than that caused by the patch antenna ground plane.

20

A. Mason, A. Shaw, and A.I. Al-Shamma’a

Fig. 5. A simulated patch antenna, constructed using a guide by Balanis [18]. Note that the model is transparent so that the electric vector field is visible across the entire structure.

Fig. 6. Simulated radiation patterns of the (a) patch and (b) CPW antennas

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

21

Despite the CPW antenna providing a bandwidth much greater than that required and having low directivity, it was still too large. Therefore further investigation ensued with CPW antennas remaining the focus of attention. 2.3 Folded Coplanar Waveguide (FCPW) Antenna In order to solve the issue of size with the CPW antenna, the largest dimension (pcbX) was considered. This dimension had to accommodate the radiating slots of length L1, which were ultimately responsible for radiating power into the transmission medium (i.e. the surrounding air). Thinking of the slots as being analogous to the two arms of a standard wire dipole antenna, some thought was given to what might happen if the slots were folded, therefore allowing them to be accommodated by a shorter pcbX dimension. Folded wire dipole [19] antennas are created by taking the two radiating elements of a standard λ/2 dipole and folding them to form a closed loop; this transition is shown in Figure 7. By applying a similar train of thought to the CPW antenna a new folded co-planar waveguide (FCPW) antenna was created. The simulation model for this antenna is shown in Figure 8, and Table 2 shows the dimensions used for this structure after applying parametric analysis to the model in order to optimise its radiation characteristics. The simulated radiation pattern is shown in Figure 9, and is not too far removed from that of Nithisopa’s CPW antenna. The simulated impedance bandwidth was calculated to be 61% - this is discussed further later.

Fig. 7. Converting the CPW antenna to a folded CPW antenna, with new dimensions also labelled Table 2. FCPW dimensions (mm) h

pcbX

pcbY

W1

W2

W3

H1

H2

H3

H4

L1

1.6

40.0

27.5

0.5

4.0

1.5

2.0

4.0

0.5

8.5

39.0

22

A. Mason, A. Shaw, and A.I. Al-Shamma’a

Fig. 8. FCPW antenna simulation model, showing the electric vector field surrounding the antenna

Fig. 9. FCPW antenna simulated radiation pattern

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

23

Given the promising impedance bandwidth and directivity results, the largest improvement with the FCPW antenna, in terms of achieving the initial goals, was size reduction. The initial requirements, stated in Section 2.1, were 57mm×32mm×1.6mm. The FCPW antenna is 40mm×27.5mm×1.6mm. This gives a total surface area saving of 39.7%, indicating that the antenna could be suitable for WSN nodes smaller than the Berkeley MicaZ. 2.4 Validating the Design The prototype FCPW antenna is shown in Figure 10. It has a bulkhead type SMA connector attached so as to allow connection to various devices (including the Berkeley MicaZ via an inter-series MMCX to SMA adaptor). Use of an SMA connector was convenient for experimentation purposes, the results of which are presented in the next section. It is imagined that this connector would not be necessary if the antenna were used in practise – instead the antenna could be connected directly to the radio transceiver of a wireless sensor node. The centre conductor of the SMA connector is soldered to the centre copper strip, whilst the outer conductor is connected to the antenna ground plane on either side of the centre strip. The simulated results for reflected power and those measured using the Anritsu VNA show a reasonable agreement (see Figure 11). At 2.45GHz the reflected power is just 2.6% - this means that 97.4% of the power incident to the antenna should be radiated. The impedance bandwidth during simulation was found to be 61%, whilst the measured bandwidth is 53%. This is more than adequate for the successful operation of the MicaZ mote over all of its selectable frequencies.

Fig. 10. Prototype FCPW antenna post construction

24

A. Mason, A. Shaw, and A.I. Al-Shamma’a

Fig. 11. Simulated vs. measured reflected power for the FCPW antenna

3 Radiation Pattern Measurement 3.1 Methodology Simulations gave an indication of the FCPW antenna radiation pattern, however physical confirmation of this was required in order to validate the design. Testing antennas is not a trivial task as one must take a number of steps to ensure that measurements are not subject to interference from surrounding sources; this is particularly relevant here due to the wide range of uses for the ISM 2.4GHz frequency band (e.g. WLAN infrastructure).

Fig. 12. Inside a typical anechoic chamber

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

25

Small antennas are typically tested in almost ideal conditions inside a structure known as an anechoic chamber [20]. Such chambers are usually shielded with a thick conducting metal, thus utilising the skin effect [21] to prevent signals from the outside world entering the chamber. An example of such a chamber is shown in Figure 12. Unfortunately, an anechoic chamber was not available for this work since they are costly to rent, and even more costly to build. Therefore an alternative method for testing was put into operation. It was thought that since WLAN infrastructure is typically localised to urban areas in the UK, tests could be carried out in a rural scenario where interference would be reduced. In addition, it was thought that a rural location, such as a large open field, would allow transmitted signals to simply travel away from the antenna and into space. This avoids issues with electromagnetic phenomena such as multipath fading [22-24].

Fig. 13. Antennas used for experimentation, including orientation indicators; (a) monopole (b) FCPW and (c) horn

26

A. Mason, A. Shaw, and A.I. Al-Shamma’a

In addition to the FCPW antenna two further antennas were required; a reference (for comparison purposes), and a receiver. The reference antenna constructed was a λ/4 monopole with a ground plane, designed to approximate the performance of the default MicaZ monopole antenna. A horn antenna was chosen as the receiver since they are well known to have a high directivity [25] and are therefore unlikely to be affected by signals which originate from anywhere but directly in front of the horn aperture. In addition, horn antennas are highly orientation sensitive and only accept radiation in a single plane when they are used at their fundamental operating frequency. The chosen horn had dimensions appropriate for use at 2.4GHz. All three antennas are shown in Figure 13. Both the receiver (horn) and transmitter (monopole or FCPW) needed to be suspended 1.5m above ground level in order to prevent reflections from the floor causing multipath interference. In order to do this, tripods were employed, as shown in Figure 14. The receiving horn antenna was bolted to the top of a large surveyor’s tripod. Since it was assumed that the ground in a rural setting would not be particularly level, a base was made for this tripod with a large bolt at each corner allowing the base to be levelled – a plumb line hanging from the centre of the tripod was used for levelling. A standard camera tripod with full height adjustment was used to mount the transmitter. Much of the tripod was made out of plastic rather than metal which was considered to be useful in terms of reducing its impact on the experimental results (e.g. due to coupling). The mounting here was moveable through 360°, allowing rotation of the transmitting antenna in order to obtain radiation pattern measurements.

Fig. 14. Experimental setup in an open field

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

27

The two antennas were placed 1.5m apart to ensure that the measurements were in the antenna far field (as opposed to the near field). The far field of the monopole antenna is 0.276m, but Balanis [26] notes that this is not a fixed rule and so it seemed appropriate to allow additional range. An Anritsu MS2024A vector network analyser (VNA) was used – the horn antenna was connected to its input port, and the transmitter to its output. The VNA was then used to measure the return loss2, in dB, at 2.45GHz. All measurements were repeated three times, and the averages used in the following sections; further repetitions were not practical because of changeable weather conditions and life time of the Anritsu VNA’s battery. 3.2 Measured Radiation Patterns The results of these experiments represent the measured return loss arbitrarily normalised to -50dB. The measurements shown are relative to the antenna orientations indicated in Figure 13. For the monopole antenna the electric field is polarised in the z direction, whilst for the FCPW it is taken to be in the y direction. With the pyramidal horn being a flared waveguide, it is possible to assume [27] that the electric field is parallel to the z axis of the waveguide (see Figure 13). This means that turning the horn antenna through 90° allows one to consider how the antennas perform when there is a polarisation mismatch. The measurements shown in Figures 15 and 16 include the situation where the antennas polarisations are matched and mismatched respectively. This is important for this work since we cannot often guarantee the relative orientation of the antennas of sensor nodes in a WSN.

Fig. 15. Monopole antenna radiation patterns with (a) matched and (b) mismatched polarisation

2

The ratio of received power (input) to transmitted power (output).

28

A. Mason, A. Shaw, and A.I. Al-Shamma’a

Fig. 16. FCPW antenna radiation patterns with (a) matched and (b) mismatched polarisation

3.3 Discussion Looking first at the results of the monopole antenna (Figure 15), we see good correlation of the measured radiation patterns with those from known and respected literature [28, 29]. This is an excellent indication of the validity of the measurement method used in this work. When polarisation between the monopole and horn antennas is matched, as in Figure 15(a), we see in the xy plane that the monopole antenna acts isotropically; this is indicated by the reasonably uniform distribution of measured radiated power. However, in the xz and yz planes significant attenuation occurs in the z axis. Unlike infinite ground planes, their finite counterparts do not completely block transmission [13] but there is definite attenuation (in the order of 5-10dB) due to the presence of the monopole ground plane between 90° and 270° (not inclusive). The maximum measured power occurs in the yz plane at 310°, and is 8.07dB. The patterns are also generally quite symmetrical in the z axis, again as one would expect, so a similar peak can be found at 50°. A wire antenna is said to have a polarisation which is parallel to the wire [11]. Therefore it is not surprising that, as shown in Figure 15(b), there is a significant loss experienced with the monopole antenna when polarisation is mismatched. This is particularly evident in the xy plane, where the radiated power reaches -40dB. There is some improvement in the xz and yz planes, although it is suspected that this is a result of reflections incident to the ground plane. It is possible that the ground plane is acting as a poor parabolic dish antenna in these cases. Let us now consider the case of the FCPW antenna beginning with the radiation pattern results when polarisation is matched, as shown in Figure 16(a). Radiation is weakest in the plane of the PCB since electric field lines formed as a result of opposing charges on the FCPW ground plane converge at the PCB edges but cannot join. This is further highlighted in Figure 17. A drop in the radiated power of approximately 10dB is also present at the 90° and 270° degree positions due to the lack radiating elements covering these positions. Whilst a 10dB loss is significant, it is not nearly as significant as the loss that HFSS reported (shown as a 35dB loss in Figure 9). Another noteworthy feature is that the FCPW antenna, whilst not

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

29

outputting peak power near that of the monopole antenna, does not drop much below 10dB, whilst the monopole antenna drops to almost -20dB along the z axis. Looking finally at the most exciting results from the FCPW antenna, shown in Figure 16(b), this antenna is much more tolerant of polarisation mismatch than the monopole antenna. In the xy plane there is typically a 2-3dB loss when compared with the monopole measurements shown in Figure 15(b); the same reductions in radiated power are present at the 90° and 270° positions. These reductions apply to the yz plane also. The most notable feature, however, is the xz plane which displays peaks of up to 4.05dB which is more than experienced by the antenna in the Figure 16(a). So, the question is, why does the FCPW antenna exhibit such results? Looking at Figure 18 one can see that the vector fields radiate perpendicular to the PCB in most cases. A result of this is the drop in radiated power shown in Figure 16(b) in the xz plane at 0°, because directly above the antenna there is a polarisation mismatch, whilst there is an increase shown nearer the edges of the board. At the edges of the PCB the electric field appears to change direction so that it is nearly orthogonal to the field in the centre. This indicates that the polarisation of the electric field is not strictly linear, as is the case with the monopole antenna. Figure 19 serves to reinforce this argument. Here the electric field is shown to be parallel to the PCB; in the centre of the antenna the field is parallel to the y axis, but near the edges this changes and the field becomes almost parallel to the x axis. The reason for this occurring is most likely due to the fold introduced to the antenna. Normally the electric field would form across the narrowest dimension of the radiating slots, but at the fold the electric field can form between the centre strip and the ground plane, opposing the direction of the central electric field. Since the results in Figure 16(a) and 16(b) are not identical (i.e. the two radiation patterns are not of the same magnitude) this indicates that the antenna has an elliptical rather than circular polarisation. This means that the antenna is not truly omni-directional, but does exhibit properties which are far more favourable than a simple monopole antenna. The next section of this chapter looks at the practical implications of these properties.

Fig. 17. FCPW antenna simulated electric field vector plot showing the formation of fields at the edges of the PCB

30

A. Mason, A. Shaw, and A.I. Al-Shamma’a

Fig. 18. HFSS simulated vector field pattern for the FCPW antenna showing the formation of multiple polarisations (side view, xz plane)

Fig. 19. HFSS simulated vector field pattern for the FCPW antenna showing the formation of multiple polarisations (top view, xy plane)

4 Energy Efficiency Measurements 4.1 Experimental Setup Whilst a measurement of the antenna’s radiation patterns was useful to understand how power was distributed around the FCPW antenna, it is important to highlight the

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

31

real world implications of these measurements. Since the objective is to replace the MicaZ antenna, it was thought that this could involve a direct comparison of the existing monopole antenna and the FCPW antenna. Therefore experimental results were obtained which compared the two antennas in operation when attached to a MicaZ mote. A PC application was written which utilised the Received Signal Strength Indicator (RSSI) feature of the motes to calculate a ten second average for signal strength, along with the number of data packets received (out of a maximum of 1000). It was decided to record both of these items of data to ensure not only a good signal strength, but also that data transmission was taking place. Figure 20 shows a screenshot of this application.

Fig. 20. Bespoke application for recording mote RSSI and packet loss information

Tests were conducted outdoors, and the setup as shown in Figure 21. The base station and a moveable MicaZ node were placed 1.5m above ground level, and the distance between the two was then increased in 1m increments, starting at 0.1m and ending at 25m3. The transmission power level of both the base station and the moveable node were set to 0dBm (i.e. 1mW). Various orientations of both the standard monopole and FCPW antenna were investigated, including what happened when a polarisation mismatch occurred.

3

The first interval was 0.9m in order to accommodate the 0.1m initial spacing.

32

A. Mason, A. Shaw, and A.I. Al-Shamma’a

Fig. 21. RSSI and packet loss measurement setup

4.2 Measured Results The results obtained for the standard MicaZ monopole antenna are shown in Figures 22 and 23, whilst those for the FCPW antenna are shown in Figures 24 and 25. Each figure includes six sets of results. The associated orientation of each set of results is given in terms of the plane which is parallel to the base station monopole antenna.

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

33

The orientations are relative to those shown in Figures 13. For a polarisation mismatch the base station antenna was moved from being vertical to being horizontal with respect to the Earth.

Fig. 22. MicaZ monopole antenna RSSI as a function of distance

Fig. 23. MicaZ monopole antenna packet loss as a function of distance

34

A. Mason, A. Shaw, and A.I. Al-Shamma’a

Fig. 24. MicaZ FCPW antenna RSSI as a function of distance

Fig. 25. MicaZ FCPW antenna packet loss as a function of distance

4.3 Discussion Comparing the RSSI results (Figures 22 and 24), the monopole antenna performs best when it is vertical with respect to Earth (i.e. with the xz and yz planes) and there is no polarisation mismatch. For the FCPW antenna, the best performance is achieved when

Improving the Energy Efficiency of Wireless Sensors through Smart Antenna Design

35

the copper face of the antenna is facing the base station (i.e. the xy plane) and there is no polarisation mismatch. It is notable that whilst the monopole antenna gives the highest RSSI result (yz plane), it displays the largest variation in results also. Comparing the results in the yz plane with no polarisation mismatch and those in the xy plane with a polarisation mismatch there is a 20-25dBm difference. Looking at the best and worst results for the FCPW, the difference is limited to a maximum of almost 10dBm. The RSSI results do not tell the whole story however, which is why the packet loss results (Figures 23 and 25) are also included. These results show that the monopole antenna experiences heavy packet loss when it is orientated in the xy plane, even at a distance of just 2m. As the distance increases this packet loss varies greatly, and on numerous occasions peaks at over 50% loss. For the FCPW antenna however, packet loss does not appear to occur at all until a distance of 11m, and even then it is not as pronounced as that experienced by the monopole antenna. It is thought that these results are due to the FCPW antenna having an elliptical polarisation, as discussed in Section 4. This is a significant finding, since it is likely that WSN nodes will be deployed in close proximity to one another in many applications, and therefore often it is short range communications ( 0 then   ′ ′ meas |θ )π(θ ) 5: Evaluate the MH acceptence ratio α(θ, θ′ ) = min 1, π(C π(C meas |θ)π(θ)

6: Draw u ∼ U(0, 1) 7: if u < α(θ, θ′ ) then 8: θ = θ′ 9: π(θ) = π(θ ′ ) 10: π(C meas |θ) = π(C meas |θ ′ ) 11: end if 12: end if 13: end for

necessary. The previous explained reconstruction algorithms aimed on reconstructing the material values of the individual finite elements within the pipe. For the statistical approach presented in this subsection, we want to demonstrate an approach using a shape representation. Hence, the approach is able to reconstruct sharp material transitions. In the further we will give a short introduction about the three missing topics. Shape models: For the first a decision for an appropriate shape model θ has to be made. Figure 7 already depicted an exemplary shape to represent the closed contour of an inclusion. In principles, it would be sufficient to store the coordinates of the corner points to build a shape description. However, a number of methods were developed, which can be used to represent closed contours, e.g. • • • •

Fourier models [44]. Spline representations. Radial basis functions [21]. Level set methods [25].

Level set functions may be a less appropriate shape description in concern with statistical inversion theory, as they were initially invented to describe the evolution of shapes. However, we stated them for completeness. When using the finite element method for solving the forward problem, the contour has to be mapped onto the permittivity values of the finite elements. This mapping causes an additional approximation error, as the obtained permittivity distribution on the finite elements will in general never correspond exactly to the contour. In the case of a boundary element method, the contour can be directly used. Generation of candidates: The generation of a proposal candidate θ′ out of the current candidate θ becomes one major issue when implementing

Current Reconstruction Algorithms in Electrical Capacitance Tomography

99

a sufficient MCMC methods. Concerning with shape reconstructions, the generation of a proposal candidate is refereed as move [22]. Figure 11 depicts the so called corner move. The idea of the corner move is to randomly

Fig. 11. Exemplary move to generate a proposal candidate θ ′ .

pick a corner point of the shape and then to change the position randomly. In essential, this is the only move necessary to obtain a Markov Chain which reaches an equilibriums distribution. This comes by the fact, that by the corner move every possible shape can be generated out of the initial shape. However, to increase the performance additional moves like rotation, transition or scaling [22] are mandatory. Prior distributions: The last point to explain more in detail concern the choice of the prior distribution π(θ). As already mentioned, the most simple prior is given by only rejecting infeasible solutions. We only want to explain here one more specific prior given by    1 c(θ)  π(θ) ∝ exp − 2 −1 I(θ) (113) 2σpr 2 Γ (θ)π

where c(θ) denotes the circumference and Γ (θ) denotes the area of the inclusion. I(θ) forms the prior proving the feasibility of θ. Hence, this prior aims on the ratio of the circumference of the inclusion with respect to the circum2 ference of a circle with the same area as the inclusion. σpr is the variance, by which this ratio can can be controlled.

To conclude this subsection it has again to be mentioned, that MCMC methods form a simple way to solve inverse problems of any kind. The major advantage of the method is the fact, that by the posterior distribution an evaluation of the quality of the result becomes possible. Also the fact that available prior knowledge can be incorporated in a natural way make the method attractive.

100

M. Neumayer et al.

The major drawback of the approach is the increased computation time due to the necessity of using a sampling technique to obtain the posterior distribution.

5

Reconstruction Results

This section provides reconstruction results for some of the presented algorithms. Figure 12 depicts a photography of the real material distribution. Several plastic rods and pipes have been placed inside the sensor. The gray colored objects are PVC rods, with a relative permittivity of εr = 3.5. The white colored rod is of another material, which has a slightly lower permittivity. The pipe is out of teflon which has a permittivity of about 2.2. As the material distribution is of complex nature, the tomographic algorithms using volumetric descriptions are preferable for the reconstruction. An algorithm for shape reconstruction would not be suitable for this material distribution. Hence, to demonstrate the ability of statistical inversion theory, we will demonstrate the reconstruction of a single inclusion.

Fig. 12. Photography of the real distribution.

Figure 13 depicts the reconstruction result obtained by the OIOR algorithm. Figure 14 depicts the result obtained by OFOA and OSOA. As both algorithms are out of the class of back projection methods and thus having the same computational costs, it makes sense to compare this results. The result obtained by the OIOR is a strongly blurred image. One can imagine that inclusions are situated in the lower part of the pipe. However, it is not possible to say something more about the specifics of the material distribution. Also the drawback of the decreased sensitivity in the center of the pipe and the fact, that for the design of the OIOR no prior knowledge was used comes fully to hand. In comparison

Current Reconstruction Algorithms in Electrical Capacitance Tomography

101

Fig. 13. Reconstruction result obtained by the OIOR algorithm.

the four results obtained by the OFOA and the OSOA offer results of increased quality. Although the OFOA is not able to reconstruct the interior of the teflon pipe. However, the air region on the right side of the half circle formed PVC rod is visible. Compared to the OIOR algorithm, the images firstly allow to interpret the distribution. The two results obtained by the OSOA clearly demonstrate the increased quality due to the nonlinear approach. The teflon pipe is now clearly visible. One can also see, that the choice of the prior distribution (rod like or Gaussian). The reconstructed permittivity distributions for the OFOA and the OSOA are higher compared to the real material values. Summarizing one can say, that the use of prior knowledge can help to highly increasing the quality of the reconstruction result. Figure 15 depicts the reconstruction result obtained by the nonlinear method. Compared to the results obtained by the OSOA, the quality is slightly increased. The result contains no artifacts in the upper region of the pipe as they can be seen in the results obtained by the OFOA and the OSOA. However, as the nonlinear method is based on an iterative algorithm, the reconstruction time increase by a factor of about 105 . The resluts obtained by the Kalman Filter are very similar to that of the nonlinear method. Hence, they are not depicted. Figure 16 depicts the result obtained by applying statistical inversion to reconstruct an elliptic object. Figure 16(a) depicts the conditional mean and the MAP-estimate as point estimates of the reconstruction result. One can see good accordance to the real distribution. As mentioned, a least squares estimator would also provide the MAP estimate for Gaussian processes. Hence, with an deterministic approach using a suitable shape model the same result could be obtained. The big advantage of statistical inversion theory is offered in figure 16(b). The so called scatter plot depicts randomly chosen points of the shape model. Hence, out of the variance of the scatter plot one can quantify the uncertainty of the point estimates.

102

M. Neumayer et al.

1

(a) OFOA with rod like data.

(b) OFOA with Gaussian data.

(c) OSOA with rod like data.

(d) OSOA with Gaussian data.

1.5

2

2.5

3

3.5

4

Fig. 14. Reconstruction result obtained by the OFOA and OSOA algorithm.

Current Reconstruction Algorithms in Electrical Capacitance Tomography

1

2

103

3

Fig. 15. Reconstruction result obtained by the nonlinear algorithm.

True Object MAP CM

(a) Conditional mean and MAP estimate.

(b) Scatter plot.

Fig. 16. Reconstruction result obtained by statistical inversion.

6

Conclusion

In this book chapter several reconstruction algorithms for ECT have been presented. The palette reached from linear methods which are suitable for real time applications, to nonlinear methods which offer increased possibilities, to fully statistical methods, which not only provide the single reconstruction result but a statistic about the solution of the problem. The chapter started with an introduction about ECT, containing physical aspects about the electrical effects within the sensor, different measurement principles as well as calibration schemes. In section 2 we tried to define a classification

104

M. Neumayer et al.

scheme for reconstruction algorithm based on the representation of the reconstructed data. By this we wanted to demonstrate that, although the presented algorithms perform a full tomographic reconstructions, some of the presented methods offer the probability for parameter reconstruction tasks. Section 3 presented some numerical tools. In section 4 five algorithms are presented in more detail. Finally some reconstruction results are presented for real measurement data.

References 1. Feynman, R.P.: Feynman Lectures On Physics (3 Volume Set). Addison-Wesley Longman, Amsterdam (1998) 2. Olmos, A.M., Carvajal, M.A., Morales, D.P., Garcia, A., Palma, A.J.: Development of an Electrical Capacitance Tomography system using four rotating electrodes. Sensors and Actuators A: Physical 148(2) (2008) 3. Hadamard, J.: Sur les problmes aux drives partielles et leur signification physique, pp. 49–52. Bull. Univ. of Princeton (1902) 4. Soni, N.K., Paulsen, K.D., Dehghani, H., Harov, A.: Finite Element Implementation of Maxwell’s equations for image reconstruction in electrical impedance tomography. IEEE Transactions on Medical Imaging 25(1), 55–66 (2006) 5. Wegleiter, H., Fuchs, A., Holler, G., Kortschak, B.: Analysis of hardware concepts for electrical capacitance tomography applications. In: Proceedings of the IEEE Sensors Conference, vol. 7(3), pp. 436–519 (2005) 6. Wegleiter, H.: Low-Z Carrier Frequency Front-End for Electrical Capacitance Tomography Applications, Ph.D. thesis, Graz University of Technology, Austria (2006) 7. Zangl, H., Watzenig, D., Steiner, G., Fuchs, A., Wegleiter, H.: Non-Iterative Reconstruction for Electrical Tomography using Optimal First and Second Order Approximations. In: World Congress on Industrial Process Tomography WCIPT5 (2007) 8. Neumayer, M., Steiner, G.: Impact of Wave Propagation Effects in Electrical Tomography. In: Proc. of the Conference on the Computation of Electromagnetic Fields, Compumag (2009) 9. Neumayer, M., Steiner, G.: Industrial Process Tomography for Complex Impedance Recostruction. In: Proc. of the 13th International IGTE Symposium (2008) 10. Zangl, H., Neumayer, M.: A Fast Gain Invariant Reconstruction Method for Electrical Tomography. In: Proc. of the 14th International IGTE Symposium (2010) 11. Neumayer, M., Watzenig, D., Zangl, H.: An H∞ Approach for Robust Estimation of Material Parameters in ECT. In: World Congress on Industrial Process Tomography WCIP6 (2010) 12. Neumayer, M., Of, G., Schwaigkofler, A., Steinbach, O., Steiner, G., Watzenig, D.: Shape Determination in ECT using an Energy Norm Formulation. In: Proc. of the 14th international IGTE symposium (2010) 13. Watzenig, D., Brandner, M., Steiner, G.: A particle filter approach for tomographic imaging based on different state-space representations. Jnl. of Measurement Science and Technology 18(30) (2007) 14. Schwarzl, C., Watzenig, D., Fox, C.: Estimation of contour parameter uncertainties in permittivity imaging using MCMC sampling. In: 5th IEEE Workshop on Sensor Array and Multichannel Signal Processing (2008)

Current Reconstruction Algorithms in Electrical Capacitance Tomography

105

15. Steiner, G., Watzenig, D.: Electrical Capacitance Tomography with physical bound constraints. In: SICE Annual Conference (2008) 16. Simon, D.: Optimal State Estimation: Kalman, H-infinity, and Nonlinear Approaches. John Wiley & Sons, Chichester (2006) 17. Process Tomography Limited url: http://www.tomography.com/ (visited on 30.9.2010) 18. Yang, W.Q., Spink, D.M., York, T.A., McCann, H.: An image-reconstruction algorithm based on Landweber’s iteration method for electrical-capacitance tomography. Jnl. of Measurement Science and Technology 10 (1999) 19. Yan, H., Shao, F.Q., Wang, S.: Fast calculation of sensitivity distributions in capacitance tomography sensors. Electronics Letters 34(20), 1936–1937 (1998) 20. Fletcher, R.: Practical Methods of Optimization, 2nd edn. John Wiley and Sons Ltd., New York (2001) 21. Uhlii˜r, K., Patera, J., Skala, V.: Radial Basis Function method for iso-line extraction. Elect. Comp. and Informatics, 439–444 (2004) 22. Watzenig, D., Fox, C.: A review of statistical modelling and inference for electrical capacitance tomography. Jnl. of Measurement Science and Technology 20(5) (2009) 23. Kaipio, J.P., Somersalo, E.: Statistical and computational inverse problems. Applied Mathematical Sciences, vol. 160. Springer, New York (2004) 24. Brandst¨ atter, B., Holler, G., Watzenig, D.: Reconstruction of inhomogeneities in fluids by means of capacitance tomography. Jnl. for Comp. and Math. Electrical and Electronic Eng. 22, 508–519 (2003) 25. Osher, S., Fedkiw, R.: Level set methods and dynamic implicit surfaces. Applied Mathematical Sciences, vol. 153. Springer, New York (2003) 26. Liu, S., Fu, L., Yang, W.Q., Wang, H.G., Jiang, F.: Prior-online iteration for image reconstruction with electrical capacitance tomography. IEE Proceedings of Science, Measurement and Technology 151(3), 195–200 (2004) 27. Landweber, L.: An iterative formula for Fredholm integral equations of the first kind. American Journal of Mathematics 73(3), 615–624 (1951) 28. Yang, W.: Design of electrical capacitance tomography sensors, Meas. Sci. Technol. 21(4) (2010) 29. Hansen, P.C.: Rank-Deficient and Discrete Ill-Posed Problems. SIAM Monographs on Mathematical Modeling and Computation, vol. 4 (1998) 30. Watzenig, D., Steiner, G., Brandst¨ atter, B.: Managing Noisy Measurement Data by Means of Statistical Parameter Estimation in Electrical Capacitance Tomography. In: Proc. of the 13th International IGTE Symposium (2004) 31. Wegleiter, H.: Low-Z Carrier Frequency Front-End for Electrical Capacitance Tomography Applications, Dissertation, Graz University of Technology (2006) 32. Brandst¨ atter, B., Holler, G., Watzenig, D.: Reconstruction of inhomogeneities in fluids by means of capacitance tomography. Int. Journal for Computation and Mathematics in Electrical and Electronic Engineering (COMPEL) 22(3), 508–519 (2003) 33. Soleimani, M., Lionheart, W.T.: Nonlinear Image Reconstruction for Electrical Capacitance Tomography Using Experimental Data. Meas. Sci. Technol. 16, 1987– 1996 34. Scott, D.M., McCann, H.: Process Imaging for Automatic Control. CRC Press, Taylor & Francis (2005) 35. Tikhonov, A.N., Arsenin, V.Y.: Solutions of Ill-Posed Problems, Washington, DC (1977) 36. Isaksen, O.: A review of reconstruction techniques for capacitance tomography. Measurement Science and Technology 7, 325–337 (1996)

106

M. Neumayer et al.

37. Yang, W.Q., Peng, L.: Image reconstruction algorithms for electrical capacitance tomography. Measurement Science and Technology 14(1) (2003) 38. Zienkiewicz, O.C., Taylor, R.C.: The Finite Element Method, 5th edn. The Basis, vol. 1. Elsevier, Amsterdam (2000) 39. Polycarpou, A.C.: Introduction to the Finite Element Method in Electromagnetics. Synthesis Lectures on Computational Electromagnetics 1(1), 1–126 (2006) 40. Hastings, W.K.: Monte Carlo sampling methods using Markov chains and their applications. Journal Biometrika 57(1), 97–109 (1970) 41. Kortschak, B., Brandst¨ atter, B.: A FEM-BEM approach using level-sets in electrical capacitance tomography. Int. Journal for Computation and Mathematics in Electrical and Electronic Engineering (COMPEL) 24(2), 591–605 (2005) 42. Watzenig, D., Holler, G., Brandst¨ atter, B.: Adaptive regularization parameter adjustment for reconstruction problems. IEEE Transactions on Magnetics 40(2), 1116–1119 (2004) 43. Brebbia, C.A., Telles, J.C.F., Wrobel, L.C.: Boundary Element Techniques: Theory and Applications. Springer, New York (1984) 44. Zahn, C.T., Roskies, R.Z.: Fourier descriptors for plane closed curves. IEEE Transactions on Computers 21(3), 269–281 (1972) 45. Simon, D., Chia, T.L.: Kalman filtering with state equality constraints. IEEE Transactions on Aerospace and Electronic Systems 38(1), 128–136 (2002) 46. Kalman, R.E.: A new approach to linear filtering and prediction problems. Journal Basic Engineering 82, 35–45 (1960) 47. Xie, C.G., Huang, S.M., Hoyle, B.S., Thorn, R., Lenn, C., Snowden, D., Beck, M.S.: Electrical capacitance tomography for flow imaging: system model for development of image reconstruction algorithms and design of primary sensors. In: IEE Proc. Circuits, Devices, Systems, pp. 89–97 (1992) 48. Brandst¨ atter, B.: Optimal design with adjoint state approaches, Ph.D. thesis, Graz University of Technology, Austria (1999) 49. Grewal, M.S., Andrews, A.P.: Kalman Filtering: Theory and Practice Using MATLAB, 2nd edn. Wiley Interscience, Hoboken (2001) 50. Kim, H.C., Kim, K.Y., Park, J.W., Lee, H.J.: Electrical impedance tomography reconstruction algorithm using extended Kalman filter. In: Proceedings of the IEEE International Symposium on Industrial Electronics, ISIE, vol. 3, pp. 1677–1681 (2001)

Non-destructive Control of Metallic Plate with Magnetic Techniques L. Battaglini1, P. Burrascano1, A. Canova2, F. Ficili2, M. Ricci1, D. Rossi3, and F. Sciacca3 1

Università degli Studi di Perugia, Dip. di Ingegneria Industriale 2 Politecnico di Torino, Dip. Di Ingegneria Elettrica 3 AMC Instruments srl, spin-off del Politecnico di Torino

Abstract. Magnetic inspection techniques are nowadays widely used in system for Non-destructive Test (NDT) of various type of application. One of the most important application is the magnetic inspection of metallic object, that use family of sensor (Hall, GMR and coils) sensible to the magnetic flux (Hall and GMR are sensible to intensity of flux, and coils are sensible to the variation of the flux). System based on this techniques, in general, are able to read the sensor and, applying a proper software processing to the data collected, give information about the external an internal state (damage, presence of inclusion) of the object under test. An example of this method is the magneto-inductive inspection of metallic rope, widely used for the inspection of cableways. In this paper we will present a different application of this techniques, were the object that we want to test is a metallic plate (or sheet) and the output of a sensor designed ad-hoc for the application is combined to numerical techniques that can give a visual representation of the state of the object. Keywords: Non-destructive techniques, Magnetic Inspection, Electronic Sensor Design, Software Processing.

1 Introduction The various techniques for Non Destructive Testing of Materials (NDT) (based on electromagnetic sensors, ultrasound, X-ray, thermal, ...) are technologies that still continue their development today, started from the 50s. There is a growing requirement of both manufacturers and end-consumers for an increase in the quality of products, even in the case of large scale productions. The objective is that of obtaining a production which is able to have a complete control of possible defects; this result calls for an exhaustive verification: random inspections, performed by sampling the lot, show to be not adequate for the highest quality levels –even if are adopted sophisticated sampling plans-. The manufacturers move thus to the verification of the entire production by making use of “Online Non Destructive Testing” (Online NDT) techniques. Applying this latter approach to product inspection calls for accuracy, reliability, and the ability to process very large amount of raw data in each time unit (high throughput), in order to be effectively introduced in mass productions. S.C. Mukhopadhyay et al. (Eds.): New Developments and Appl. in Sen. Tech., LNEE 83, pp. 107–122. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

108

L. Battaglini et al.

It is possible to apply this approach of Online NDT to a larger and larger variety of types of industrial productions due to the impressive growth, developed in the recent years, of the signal processing capacity. This increased processing power has improved greatly the potential of some of these techniques, enabling them to respond to requests for a constant improvement of efficiency of production processes. In particular this increase in the processing power has allowed the evolution in the direction of a control of the entire production, directly performed into the production line. To obtain useful information from the raw data detected by the sensors, highresolution imaging systems are needed. Moreover systems often have to operate in real time, with limited acquisition times and adverse conditions of signal to noise ratio. A typical example of this approach can be observed in metal industries where Online NDT techniques begins to be adopted even in the case of automotive products or for household appliances: inspection procedures at the end of the line and non subjective, which are capable of detecting both included and surface defects, are currently introduced; they rely on a set of measurements and computer aided decisions. The great complexity of an entire production line monitoring is necessary to consider carefully the project is the system hardware and software processing, optimizing the overall performance of the system. The report presented here, we propose the results of a collaborative work between 1) the Faculty of Engineering, University of Perugia – Campus of Terni, 2) the Faculty of Engineering at Politecnico of Torino and 3) AMC Instruments SrL, a spin-off of Politecnico of Torino. All the skills were complemented at their best: those needed for designing the system architecture, those needed for processing multidimensional data and the skills needed for the realization of a product already engineered for its possible use in a production system. This cooperative work enabled us to develop an applied research study that appears to have excellent potential application. The paper presented here describes the different aspects of this project, motivates the different choices and shows the results that we get by applying the resulting system to a number of particularly significant benchmarks. The paper is organized as follows: in the first part the concept of the system and the realization of the sensor are described. Then are described the interface and the software developed for the processing of the information. Finally, in the last section, are reported and commented on the results of some laboratory measurements for some cases of particular interest. The last section draws some conclusions and indicates the possible development prospects.

2 Concept The basic idea of this work is to develop an electronic system able to detect (and, if possible, characterize) fault present in an manufactured product (in this case is speaking about metallic sheet), using magnetic techniques. A system like this is made by three fundamental block: • • •

Sensing Head (or Sensor) Acquisition System Software (and processing platform)

Non-destructive Control of Metallic Plate with Magnetic Techniques

109

The sensor has the purpose of get the basic information, in the analog domain. This is an hardware block, that has mechanics and electronics components. The electronic components are sensors and other mixed-signal blocks (filters, amplifiers, muxs…). The acquisition system is made basically by an acquisition board, that has the task of convert the information from analog to digital domain, under some specification, in order to make the results usable. This is a digital block, that use programmable logic and A/D converters. Then we have the software, that use a standard support (desktop PC) in order to connect with the previous block. This logic is widely used in automatic measurement system, and is commonly known as PC-based measurement system. The use of a PC-based system reduce the complexity of the acquisition block (acquisition board and software), that in the other case must be designed ad-hoc. The software block has the purpose of use the digital information obtained form the two previous block, in order to provide the final output. This block make an extensive use of numerical algorithim, in order to correctly process the data and obtain easy readable information. Figure 1 shows a basic diagram of the measurement chain.

Fig. 1. Basic block diagram of the system

3 Sensor Topology The firs problem to solve is, of course, how to get the right initial information, or more specifically, how to get a signal (possibly an electrical signal in voltage) that is, somehow, proportional with the presence of a discontinuity in the metallic sheet. The idea is to create a constant magnetic flux orthogonal to the sheet itself. In this condition the presence of a discontinuity (given from a fault or an inclusion) in the material, generate a variation in the flux, displaying the presence of a fault. In order to make the information readable form a standard acquisition system, the signal must be converted from magnet to electrical domain. In order to do this, some Hall effect sensor can be used (the use of Hall sensor instead GMR or Coils will be clarified later on). So the typical setup of this system, use a permanent magnet, with its north pole orthogonal with the sheet and an array of sensor between the magnet and the sheet (see figure 2). This type of configuration allow to reveal this kind of lack: • Discontinuity (holes, abrasions, scratch…) of various type on magnetic plate • Non-magnetic inclusions on magnetic plate • Magnetic inclusions on non-magnetic plate

110

L. Battaglini et al.

Fig. 2. Setup of the sensor. Hall probe are placed between metallic sheet and PM.

Obviously isn’t possible to reveal discontinuity in non-magnetic plate (ex: aluminium plate). Magnetic inclusion in magnetic material can generate little flux variation. The possibility of reveal such variation depends on factor like resolution of sensor and background noise. It’s important to notice that this kind of system born modular, so it can be adapted to sheet of different dimension. So, in order to monitor a sheet of a given dimension (where the dimension is a multiple of the dimension of the sensing head) more sensor can be mounted in modular way (see figure 3).

Fig. 3. Example of a modular system

This is a fundamental aspect for system that, like this, are intended to be used for monitoring object during production cycle.

4 Technological Choiches and Specifications Starting from the idea discussed before, the following step was a proper design of the electronic board of sensing head.

Non-destructive Control of Metallic Plate with Magnetic Techniques

111

Fig. 4. Position of the Array in the reference frame

The sensing head should have the following specification (see figure 4): 1. 2. 3.

Should have a resolution of 1mm in the X direction (refer to figure 4) Should work with a sliding speed of the sheet up to 2m/s (typical on-field speed 22 seconds, but even at 10 seconds, a precision of 0.6 mm could be obtained. Employing the 8 × 8 binning, the system could generate range images between 15.7 and 68.5 Hz with a best precision of 2.5 mm [15]. The penalty for this however, was significantly reduced spatial resolution. The primary remaining acquisition time limitation was the employment of the digital camera attached to the image intensifier. Affordable cameras at the time possessed a slow frames-per-second acquisition rate, typically 50 Hz to 100 Hz, with higher rates only possible with pixel binning. This system was also severely limited by the utilisation of the image intensifier. This device was costly (~6000 euro) and bulky as can be seen in Figure 6, where the image intensifier is the tube-like device attached to the camera. Furthermore, the intensifier requires three independent voltages, one of the order of -50 to +10 V, another ~700 V, and the third at a level of approximately 6 kV (note the presence of the high voltage power supplies in the foreground of Figure 6). Finally, the range processing was performed off-line by an externally interfaced PC. So whilst successful, this form of the FFRIS was limited to being a bench-top device.

Design and Construction of a Configurable Full-Field Range Imaging System

141

Fig. 6. Original FFRIS configuration featuring image intensifier and high-voltage external power supply

5 Real-Time Capable Design As mentioned, the primary limitations of the previous systems were the bulk and expense of the image intensifier and the limitations in frame acquisition imposed by the requirement for a commercial digital camera. The next iteration in the system development [9] replaced the Image Intensifier and the digital camera with a PMDTechnologies PMD19k-2 image sensor chip. This CMOS based PMD sensor permits the gain modulation of the imaging pixels to be controlled on-chip, effectively replacing the shuttering function previously performed by the image intensifier and the image acquisition function of the camera. This vastly reduced the bulk and power requirements of the FFRIS, rendering the system potentially portable. Furthermore the component costs of the system were reduced by over 80%! One disadvantage however, is that the pixel array sizes of these chips are currently limited – although significant improvements are expected in the near future. At this point, the decision was made to produce a semi-modular system comprising • • • •

an FPGA development board (based on the Stratix III), an Illumination board containing the laser diodes and driver circuitry, an Imaging Sensor board (with the sensor and lens attached as a daughter board), a VGA/Ethernet board for VGA output display and interfacing to a PC.

This is illustrated in Figure 7 [9]. Note the circular arrangement of the illumination diodes on the right. This semi-modular system provided the capability to upgrade or change one of the boards without necessitating a full system redesign, for example the sensor or illumination diodes could be exchanged with little effect on the remaining hardware.

142

D.A. Carnegie et al.

Fig. 7. System architecture of modular bench-top full-field imaging system

In the previous design iteration an FPGA controlled the gain of the laser diodes and produced a shuttering signal. Now that an FPGA was embedded in the system, the capability existed to utilise it to generate the modulation frequencies and hence eliminate the requirement of employing the DDS chips. This significantly simplified the FFRIS board design. Furthermore, by choosing an FPGA with sufficient capacity, the received frames could be stored and processed on-chip to determine the phase and hence the range to objects in those frames. Eliminating the requirement for an external PC is a significant improvement for mobile robotic applications. A Stratix III Development kit hosting an Altera Stratix III EP3SL150 FPGA was employed. This FPGA was chosen primarily for its easily reconfigurable phase-locked-loop resources which provide the ability to reconfigure the phase, frequency and duty cycle of the output channels in real time. This FPGA also has enough on-chip logic and memory resources to buffer the frames and hence calculate the full-field range images in real-time. The FPGA now incorporated [9] (again change with previous xxx) the tasks of: • • • • •

Driving the modulation signals Controlling data retrieval from the sensor Calculating range images Timing and control of signals for VGA display of range image data Handling JTAG interface and Ethernet connections to a PC in order to receive and process user commands or transferring data for long-term storage

The Illumination board comprised independently driven infrared (808 nm) and visible red (658 nm) laser diodes employed in two independent banks of 8. These diodes were driven in a controlled current configuration (by the FPGA) with a continuous

Design and Construction of a Configurable Full-Field Range Imaging System

143

total optical output power of 800 mW. Lasers of different wavelength were employed to investigate how their different reflectivities affected the performance of the FFRIS. Extra care is required in using the IR lasers however, since eye damage can easily result due to the beam being undetectable by the unassisted retina, preventing a natural blink reflex. Parallel work by the authors (for example [15,17]) provided a theoretical framework to select optimum frequencies within the achievable system bandwidth. A problem encountered is that range ambiguity occurs if the phase shift exceeds 2π and is a significant issue as the modulation frequency is increased. The maximum unambiguous range du is inversely related to the modulation frequency fm as:

du =

c 2 fm

Equation 3

Hence for a 10 MHz modulation frequency ranges up to 15 metres can be resolved unambiguously, but at 60 MHz, this reduces to 2.5 metres. In other words, at this modulation frequency the system would not be able to resolve the difference between an object located at 3 metres distance from one located at 5.5 metres. This problem has been resolved [16,18] by employing two modulation frequencies simultaneously, however a discussion of this is outside the scope of this chapter. The PMD 19K-2 3D Video Sensor Array from PhotonICs is employed as the imaging sensor. This sensor features a 160 × 120 array of pixels grouped into four independently modulated blocks of 40 × 120 pixels. Each modulation block presents a capacitive load of 250 pF at the driver interface. To drive these blocks at modulation frequencies of 10 MHz and above, ultra high current pin drivers (EL7158 from Intersil) are employed.

Fig. 8. Assembled FFRIS utilising Stratix III FPGA development board

144

D.A. Carnegie et al.

For this particular sensor, a rise and fall time of 12 ns can be achieved, limiting the maximum modulation frequency (without skew) to 41 MHz. Maximum contrast is achieved with a modulation input of +2.5 V. At a modulation frequency of 41 MHz, the imaging sensor’s modulation inputs have an impedance of approximately 15.5 Ω each and so the current draw per modulation input is 160 mA, with a worst case total current draw of 640 mA. The complete assembled system is illustrated in Figure 8.

6 Portable Version The system described so far demonstrates many of the requirements that we wish for our final system, however its size is constrained by the large physical dimensions (210×180×70 mm) of the Stratix III development board as can be appreciated in Figure 8. For the portable version, suitable for mounting on a mobile robot, the modular construction of the previous design is retained to allow a versatile system since it can be expected that the FPGA will be replaced in the future with a more powerful version, the illumination diodes may be exchanged for different wavelengths, bandwidth or intensity, and certainly advances are expected in sensor capacity. The new system has the arrangement of Figure 9 [19] which illustrates how the four boards sandwich against each other to minimise physical dimensions. The boards are separated and attached to each other using M3 PCB standoffs. Physical sizing is determined in the first instance by the circular cut-out for the Illumination board. This must have a minimum diameter of 35 mm to accommodate the optical lens (diameter 30 mm). The laser diodes of the Illumination board have an 8 mm diameter, and to accommodate 16 requires them to be centred on a circle with a minimum diameter of 43 mm. To accommodate the on-board voltage regulation and laser driving circuitry, and the 6 mm diameter of the standoffs, a board design of 100 mm square is required (considerably smaller than the dimensions of the Stratix III development board). The system is designed to be generic and flexible in terms of the components it will tolerate. The device is designed to meet the processing requirements of a 1 Mpixel image although we expect it to be some time before such a sensor becomes inexpensively available. Similarly, the current ratings for the components are overspecified which has obvious implications for the power source. Again this is done deliberately to accommodate circuit additions in the future if required. The following sections detail the function of each board and a brief description of how this was implemented. Communication between the boards is facilitated by a generic I/O interface that carries 17 differential transmit lines, 17 differential receive lines, a +3.3 V rail spread across 20 pins, a +12 V rail spread across 19 pins, 4 single ended I/O lines, and 4 JTAG lines (TDO, TDI, TMS, TCK). This is implemented with a 172 pin high speed mezzanine male connector. The form of this connector is modelled on the external interfaces provided on the Stratix III development kit.

Design and Construction of a Configurable Full-Field Range Imaging System

145

Fig. 9. System architecture of portable modular full-field imaging system

6.1 FPGA Board This board is the “heart” and “brains” of the entire system. It is configured to internally contain the NIOS processor which performs many of the system control function. The FPGA board connects to the image sensing/capture board via one of the two generic digital I/O ports and via these ports controls the image capture and illumination processes. The FPGA board also connects to the External Interface board to facilitate the FPGA sending captured images to peripheral devices such as a PC or VGA monitor. The FPGA selected has to have sufficient I/O lines to access the generic I/O ports and to interface to external RAM. As discussed, the prototype version utilised the Altera Stratix III family of FPGAs. For this portable implementation, the Altera Cyclone III EP3C40 is preferred due to the substantially lower cost. This version of the Cyclone family contains 39,600 logic elements, 4 PLLs, ~ 1 Mbit RAM, and 535 User I/Os. This is deemed sufficient for our initial purposes, but should a future iteration require additional processing of the raw images, then the EP3C120 contains three times the number of logic elements and RAM and provides for an easy upgrade path (pins are mostly identical). A 50 MHz oscillator is employed as the system clock. The design of the full-field range imaging system utilises four discrete blocks of memory for processing and storage of images. The prototype system implements the RAM required for image processing on the FPGA’s internal 5499 Kbit static random access memory. This is sufficient for storage and processing of image frames from the 160 × 120 (19K2) pixel PMD sensor but not for higher resolution sensors and so provision for external memory must be provided.

Fig. 10. Block diagram of FPGA board

146 D.A. Carnegie et al.

Design and Construction of a Configurable Full-Field Range Imaging System

147

6.2 Memory Requirements Five different types of memory are required for this system: • • • • •

Accumulator – required for the storage of images during image processing Output buffer – stores the images to be accessed by the NIOS processor, the VGA output and/or the Ethernet output NIOS program – stores the firmware of the NIOS processor NIOS Ethernet frame buffer – required to store image frames to output to a peripheral computer Flash FPGA configuration memory – stores the FPGA configuration to be loaded at start up.

In determining the choice of memory to satisfy the above requirements, it is important to note that the Cyclone III family supports several high speed external memory interfaces to allow external memory to be connected to the FPGA with little customisation of hardware or software. These supported memory devices include DDR, DDR2 and SDR SDRAM. DDR2 SDRAM is preferred over the older memory types for this application. The accumulator is used to store the accumulated real and imaginary terms (up to 16 bits each) for each pixel needed to calculate the phase [20]. Hence for a 1 Mpixel image this would require 4 MB. For an image processing routine operating at a clock frequency of 10 MHz, the accumulator must operate at a data rate of 320 MBit/s (32 bits/pixel × 10 MHz). The accumulator must read and write a pixel value each time it performs a mathematical operation. This increases the required data rate of the accumulator to at least 640 MBit/s. Depending upon the application, the output buffer may be required to hold between 1 and 4 frames at 16 bits per frame. An application that only requires the distance information would simply require 1 frame, if amplitude was also required then 2 frames would be necessary and if the raw pixel data is needed then all 4 frames would need to be stored. Hence for a 1 Mpixel frame, in the worst case, 8 Mbytes are required. This output buffer is accessed by the NIOS processor, the VGA output process and the image processing process. All three accessing processes are clocked at 10 MHz and so the memory must be able to transfer data to each process at a speed of 320 Mbit/s. Time multiplexing is utilised to handle the three accessing processes. This increases the required data transfer speed to 960 Mbit/s (320 Mbit/s by each process). For the accumulator and output buffer, the Micron MT47H64M8 DDR2 SDRAM is selected. This device has a memory size of 512 Mbit and operates at a clock frequency of 333 MHz. It has an 8-bit wide data bus allowing data transfer rates of 2.7 Gbit/s. For the NIOS program, the Micron MT47H32M16 DDR2 SDRAM is selected, effectively being a 16 bit variant of the H64M8 above. The NIOS Ethernet buffer RAM stores frames that are to be output to a peripheral computer via the Ethernet connection. We have designed 1 Gbyte of memory, sufficient to buffer 256 1 Mpixel output frames. This is implemented using the M47H512M8 DDR2 SDRAM which has a memory size of 4 Gbit and operates at a frequency of 333 MHz. It has an 8-bit wide data bus allowing data transfer rates of

148

D.A. Carnegie et al.

2.7 Gbit/s. Two of these devices operating in parallel increase the memory density to 8 Gbit (1 Gbyte) and double the data transfer rate to 5.3 Gbit/s. The Flash configuration is 64 Mbyte which is deemed to be easily sufficient for any current or future sized configuration files. 6.3 Illumination Circuit Board As explained previously, 16 laser diodes are incorporated into the Illumination board and are chosen to be the 130 mW CWML101J27 Mitsubishi devices, operating at 660 nm. These are driven by the iC-HK 155 MHz laser switches to provide up to 150 mA continuous current or 700 mA maximum pulsed current. A block diagram representation of this board is provided in Figure 10. It is important at start-up that the diode current be gradually increased to allow the laser to reach a steady operating temperature since at low temperatures the high currents could potentially cause catastrophic optical damage. The diode current in earlier iterations of the board was controlled by an on-board microcontroller. This functionality is now undertaken by the FPGA, and the diodes are modulated from the FPGA via a control signal marked in Figure 11 in the form of a Two-Wire Interface TWI bus. This control signal feeds into a digital to analogue convertor (DAC) (AD5311 from Analog Devices) and then to the laser switch. A protection circuit has also been included to switch off the laser diodes if the control or modulation signals from FPGA board become disconnected from the Laser Illumination board.

Fig. 11. Block diagram for Illuminator Board

Design and Construction of a Configurable Full-Field Range Imaging System

149

A DIP switch allows the user to select between running the laser diodes at maximum power or half power. The half power option operates the laser diodes at a safer optical intensity when output power is not a critical consideration for image ranging a scene. 6.4 Image Sensing and Capture Board This board connects to the FPGA via the generic digital interface. This board handles all the digitising of the captured images from the sensor and all the signal modifications to drive the image sensor. It also busses power and modulation control signals from the FPGA board to the illuminator board. It is illustrated in block diagram form in Figure 12.

Fig. 12. Block diagram for the Image Sensing and Capture Board

The FFRIS requires control over the modulation of the sensor and must be able to access the raw pixel values. These requirements eliminate many of the image sensors currently on the market. We chose not to wait for an improvement over the PMD 19k sensor to become available, and elected instead to design the Image Capture board so that it could immediately host the PMD Daughter board from the previous system incarnation. As this image board is only a sub-system of the complete FFRIS, it will be straight-forward to design a modified board in the future as appropriate sensors become available (this is the motivation behind the modular design – to easily implement hardware changes without incurring a system redesign). The PMD daughter board connects to this board so that the imaging sensor is centred on the board, ensuring that the attached optical lens will be aligned with the circular cut-out of the Laser Illumination board.

150

D.A. Carnegie et al.

The image sensor has two analogue video outputs. A 16-bit Imaging Signal processor ADC (AD9826 from Analog Devices) is provided to convert the two analogue video streams from the image sensor into digital frames to be processed by the FPGA. This ADC can sample the video streams at a maximum rate of 15 million samples per second. Each 16-bit data word is transferred to the FPGA in two sequential bytes on an 8-bit data bus via the generic I/O interface. The generic I/O interface between the FPGA and Image Capture board utilises LVDS signals to improve the quality of the transmitted signals. LVDS receivers and transmitters convert the LVDS signals to the single-ended CMOS signals required by the image sensor, ADC and high current pin drivers. The LVDS receiver/transmitters are the SN65LVDT388/389 8-way high speed drivers, and the repeater is the SNLVDS100. The Image Capture board busses the modulation and control signals to the Illumination board via Serial Advanced Technology Attachment (SATA) connections. 6.5 External Interface Board This board provides all the necessary communication drivers to allow the system to connect with peripheral devices such as an external computer or VGA monitor. It also provides communication interfaces to allow configuration and management of the imaging system by an external control computer. Figure 13 presents a block diagram of the main sub-systems of the External Interface board.

Fig. 13. Block diagram of the External Interface board

Design and Construction of a Configurable Full-Field Range Imaging System

151

Image frames can be displayed on an external standard 640 × 480 resolution VGA monitor over a VGA interface allowing users to examine four frames of raw and processed data simultaneously in real time. The VGA output is driven by a triple 10bit high speed video DAC (ADV7123, from Analog Devices). This DAC has three separate 10-bit input ports that drive three complementary outputs to produce the analogue red, green and blue video streams to display images on a VGA monitor. All the VGA data and control signals are sent over the generic I/O interface from the FPGA board to the External Interface board. The External Interface board uses a standard VGA jack to allow the connection of any standard VGA monitor. An 8-bit microcontroller (ATMega32U4 from Atmel) incorporating a USB controller is employed to provide the communication interface between the control computer and the range imaging system (which contains a Mini–B USB connector). The MCU provides full-speed and low-speed USB connections and allows the range imaging system to appear as a USB device to any USB host running on an external computer. The External Interface board contains a RJ-45 jack with an integrated Ethernet interface to allow long term storage of processed and raw image frames on an external computer. A stand-alone Ethernet controller (DM9000A from Davicom) provides both the MAC and PHY transceiver and is interfaced to the FPGA via a 16-bit data bus over the generic I/O interface. The NIOS processor is responsible for writing image frames to the Ethernet controller chip.

7 Power Supply As can be seen from Figures 10 – 13, the component boards require a plethora of voltages, including +1.2, +1.8, +2.5, +3.3, +5.0, +12 V. These are provided on the component boards by an appropriate regulator, but such designs are straight-forward and will not be further described here. The power consumption of the image ranger sub-systems has been calculated based on the maximum ratings of the components utilised on each board plus a future expansion capacity. The resultant maximum rating of each board is (rounded up to the nearest 5 W): • • • •

FPGA Board – 25 W. External Interface Board – 5 W. Image Sensing and Capture Board – 10 W. Laser Illumination Board – 50 W.

The FPGA board power consumption is primarily set by the configuration of the FPGA, specifically the speed and the number of resources utilised, 20 W being the measured output of the working Stratix as configured in the bench-top system. Based on the power consumption calculation the input power supply must be able to deliver 90 W of power. The power supply will typically be from a 12 V source and hence the supply must be capable of providing a maximum of 7.5 A. The External Interface board incorporates protection and filter circuits to provide a clean power supply to the unregulated power bus that runs through the entire system. A 40 A rated Schottky diode (48CTQ060SPBF, from Vishay) provides reverse polarity protection

152

D.A. Carnegie et al.

on the input voltage and two power-out connectors, fused at 10 A each limit the total current. An inductor-capacitor (LC) filter, rated at 23 A, provides filtering on the input power supply to prevent high frequency noise.

8 Results The completed and assembled unit is illustrated in Figure 14 [19]. With the modular board arrangement, it has dimensions of 120×200×120 mm (excluding protruding connectors) which is approximately 30% of the size of the bench top system of Figure 8. This comparison is illustrated in Figure 15. Different connectors could be employed to further reduce the device’s physical size, but this is not a priority since the current sizing can easily be accommodated on our mobile robots. With the PMD sensor, 19200 simultaneous range measurements at sampling frequencies of up to 24 Hz to centimetre precision have been achieved. Sub-centimetre acquisition requires sampling at rates of 0.1 Hz or lower. Range disambiguation has been implemented [16,17] with little loss of precision.

Fig. 14. Completed portable FFRIS

The field of view of the system is limited only by the optical hardware. Using a 16 mm focal length lens results in a field of view (horizontal × vertical) of 22.2° × 16.5°. Care must be taken with the selection of lens as wider angle lenses in particular will introduce optical distortion. Calibration of the system for different lens types will mitigate this effect. Figure 16 displays a sample (face of a mannequin) output from the system where colour has been used to illustrate the range. This image was taken with a modulation frequency of 36 MHz and a frame integration time of 20 ms. As required, changes in acquisition time, laser intensity, and modulation frequencies can easily be made by a user.

Design and Construction of a Configurable Full-Field Range Imaging System

153

Fig. 15. Comparison of portable and bench top FFRIS

Fig. 16. Colour enhanced range full-field range image

9 Summary Reviewing the original specifications, this completed FFRIS satisfies all of the essential and desirable requirements. Specifically: • •

The system is able to operate at configurable modulation frequencies up to 40 MHz. All modulation and synchronisation frequencies are generated within the FPGA and are precisely frequency locked with each other.

154

D.A. Carnegie et al.

• •

• • •

•

•

•

The illumination intensity of the laser diodes can be easily changed in software. Full-field range measurements are calculated by the on-board embedded FPGA. This range can be visually displayed, or provided in numerical form for another processor. No external PC is required unless the configuration details are to be altered or the frames transferred for long-term storage or analysis. The lasers are protected from over-current during turn-on or in the event of loss of modulation or control signals. The device can be entirely powered by the 12 V batteries that power most of the fleet’s mobile robots. The system has been currently over-engineered allowing for a maximum power consumption of 90 W. In operation 50 – 60 W is a more typical figure, but this can increase particularly if greater intensity is required from the laser diodes. Regardless, for the mobile robots in our fleet, the robot’s locomotive motors easily dominate power requirements. The device is inexpensive. The largest cost is the sensor (currently 1000 euro), with the next most expensive component being the FPGA which is of the order of 100 euro. Total component cost of the system is approximately 2000 euro. It is compact with the most significant room for improvement being a change of the connectors. Whilst commercial production could further reduce the system size, it is comparable to other systems currently on the market. Acquisition time and modulation frequency can be varied in software to provide versatility between real-time operation where precision may not be so critical through to longer acquisition, high precision measurements when greater environmental detail is required.

Whilst other devices on the market may outperform our system in one or more of spatial resolution, frame rate, depth precision or size, none out perform us on all of these criteria. The most significant advantage of our solution is that it is extremely configurable and provides a solution for a greater range of situations than any other systems we have investigated.

References [1] Sharp GP2Y3A003K and GP2Y2A002K at http://sharp-world.com/products/device/catalog/index.html [2] Sick, http://www.sick.com/ [3] Jongenelen, A.P.P.: Development of a Configurable Range Imaging System for Unambiguous Range Determination, PhD Thesis under examination, Victoria University of Wellington (2010) [4] Christie, S., Hill, S.L., Bury, B., Gray, J.O., Booth, K.M.: Design and Development of a Multi-Detecting Two-Dimensional Ranging Sensor. Measurement Science and Technology 6, 1301–1308 (1995)

Design and Construction of a Configurable Full-Field Range Imaging System

155

[5] Blais, F.: Review of 20 years of range sensor development. Journal of Electronic Imagine 13(1), 231–243 (2004) [6] Carnegie, D.A., Cree, M.J., Dorrington, A.A.: A high-resolution full-field range imaging device. Review of Scientific Instruments 76(8) (2005) [7] Dorrington, A.A., Carnegie, D.A., Cree, M.J.: Towards 1 mm depth precision with a solid-state full-field range imaging system. In: Proceedings of SPIE: Sensors, Cameras, and Systems for Scientific/Industrial Applications, San Jose, CA, USA, vol. 6068 (2006) [8] Dorrington, A.A., Cree, M.J., Payne, A.D., Conroy, R.M., Carnegie, D.A.: Achieving sub-millimetre precision with a solid-state- full-field heterodyning range imaging camera. Measurement Science and Technology 18, 2809–2816 (2007) [9] Jongenelen, A.P.P., Carnegie, D.A., Payne, A.D., Dorrington, A.A.: Development and characterisation of an easily configurable range imaging system. In: Proceedings of the 24th International Conference Image and Vision Computing (NZ), Wellington, New Zealand, pp. 79–84 (2009) [10] Payne, A.D., Dorrington, A.A., Cree, M.J., Carnegie, D.A.: Characterization of modulated time-of-flight range image sensors. In: SPIE – 7239 3D Imaging Metrology, San Jose, CA, USA (2009) [11] Büttgen, B., Seitz, P.: Robust optical time-of-flight range imaging based on smart pixel structures. IEEE Transactions on Circuits and Systems I: Regular Papers 55, 1512–1525 (2008) [12] PMDTechnologies, http://www.pmdtec.com/ [13] MESA Imaging, http://www.mesa-imaging.ch/ [14] Canesta Inc., http://canesta.com/ [15] Dorrington, A.A., Cree, M.J., Carnegie, D.A., Payne, A.D., Conroy, R.M., Godbaz, J.P., Jongenelen, A.P.P.: Video-rate or High-Precision: A Flexible Range Imaging Camera. In: Proceedings SPIE Image Processing: Machine Vision Applications, San Jose, CA, USA, vol. 6813 (2008) [16] Jongenelen, A.P.P., Carnegie, D.A., Payne, A.D., Dorrington, A.A.: Maximizing precision over extended unambiguous range for TOF range imaging systems. In: Proceedings of the 27th IEEE International Instrumentation and Measurement Technology Conference, Austin, TX, USA, pp. 1575–1580 (2010) [17] Jongenelen, A.P.P., Bailey, D.G., Payne, A.D., Dorrington, A.A., Carnegie, D.A.: Analysis of Errors in ToF Range Imaging with Dual-Frequency Modulation. IEEE Transactions on Instrumentation & Measurement (publication pending) [18] Payne, A.D., Jongenelen, A.P.P., Dorrington, A.A., Cree, M.J., Carnegie, D.A.: Multiple frequency range imaging to remove measurement ambiguity. In: 9th Conference on Optical 3-D Measurement Techniques, Vienna, Austria, pp. 139–148 (2009) [19] McClymont, J.: The Development of Extrospective Systems for Mobile Robots. ME Thesis, Victoria University of Wellington (2010) [20] Jongenelen, A.P.P., Bailey, D.G., Payne, A.D., Carnegie, D.A., Dorrington, A.A.: Efficient FPGA Implementation on Homodyne-Based Time-of-Flight Range Imaging. Journal of Real-Time Image Processing, Special Issue (2010) (publication pending)

Cr2O3-doped BaTiO3 as an Ammonia Gas Sensor Gotan H. Jain1,*, S.B. Nahire2, D.D. Kajale1, G.E. Patil1, S.D. Shinde3, D.N. Chavan4, and V.B. Gaikwad3 1

Materials Research Lab., Arts, Commerce and Science College, Nandgaon 423 106, India [email protected] 2 G.M.D. Arts, B.W. Commerce and Science College, Sinnar, India 3 Materials Research Lab., K.T.H.M. College, Nashik 422 005, India 4 Department of Chemistry, Arts Commerce and Science College, Lasalgaon 422 306, India

Abstract. The thick films of pure BaTiO3 (BT) were prepared by screenprinting technique. The gas sensing performances of these films were tested to various gases by using static gas sensing system at various operating temperatures. The pure film showed maximum response to H2S gas at 350oC but poor selectivity. Different wt% of Cr (0.56, 5.27 and 6.07) was added in BaTiO3, base material, followed by sintering at 550oC for 30min. The thick films of such powder were prepared by screen-printing technique. The thick films of this Crdoped BT were prepared and tested to various gases. The Cr2O3-doped BT film (5.27wt %) showed maximum response to ammonia gas at 350oC and suppresses the response to H2S gas. The response of 5.27wt% film was observed to be the most amongst the 0.56 and 6.07wt. The selectivity of the Cr2O3-doped BT was found to be more against the other gases. The 90% response and recovery levels were attained within 3 and 20 s, respectively for Cr2O3-doped BT (5.27wt %) film. The very short response and recovery time are the important features of this Cr2O3-doped BT film to NH3 gas. Keywords: BaTiO3, thick films, NH3 gas sensor, sensitivity, selectivity.

1 Introduction In the recent, sensors have attracted a great deal of attention from scientists and engineers. Even in the near future, it is expected to gain importance in view of the construction of more or less intelligent ensembles, which integrate actuating, sensing and computing subsystems. Detection of various gases using solid-state chemistry has generated a great deal of interest, both in academia and in industry. Although much research has been focused on sensors based on SnO2 technology, other inorganic oxides are receiving increased attention. These include binary oxides, such as oxides of titanium, tungsten, and gallium, and more complex ternary oxides. Compounds having pervoskite structures are among one of the most important classes of ternary oxides. Barium titanate (BaTiO3) is one of the most intensively investigated *

Corresponding author.

S.C. Mukhopadhyay et al. (Eds.): New Developments and Appl. in Sen. Tech., LNEE 83, pp. 157–167. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

158

G.H. Jain et al.

ferroelectrics and has been widely used in electronic industry in applications for capacitors, thermocouples, transducers, sensors and actuators, etc. A number of studies have focused on the chemistry and physics of the response of these materials to gases. Depending on the conditions, these compounds can behave as n- or p-type semiconductors. Because of their structural similarity, similar mechanisms of interaction with gases are expected to occur for these compounds, although the relative importance of the mechanisms for any specific operating condition would depend, in each instance, on the specific compound. It is well known that a number of pervoskite oxides (ABO3) have been used as gas sensor materials because of their stability in thermal and chemical atmospheres. So over the last decade, the pervoskite oxide ceramics such as BaTiO3 (BT) have created and promoted interest in chemical sensors. It is capable of detecting a particular gas in the high temperature region, 175-450oC, near and above the temperature giving the maximum resistivity [1]. Modifications in the microstructure, the processing parameters and also the concentration of acceptor/donor dopants can vary the negative temperature coefficient of the resistance (NTCR) and conductivity of BaTiO3. It is also known in literature [2-6] that PTCR (positive temperature coefficient region) disappears completely when donar-doped BaTiO3 was annealed at high temperatures in atmosphere of low oxygen partial pressure. BaTiO3 is well known for the detection of CO [8, 9], CO2 [10-12], humidity [13], etc. Various attempts have been made to improve the selectivity and sensitivity of BaTiO3 by using dopants and additives [14, 15]. There are a few reports dealing with BaTiO3-based gas sensors. Efforts are, therefore, made to develop BaTiO3 - based gas sensors and for the improvement in its sensing performance by doping and modifying the surface of the thick films. Pure and modified BaTiO3 are observed to be most sensitive to H2S gas. Some well-known materials for H2S gas sensing are SnO2-ZnO-CuO [16], SnO2-Pd [17], SnO2-Al2O3 [18], SnO2CuO [19-24], SnO2-CuO-SnO2 [25], and ZnSb2O6 [26]. Researchers have developed various types of sensors by adding different additives [1,27-29] into semiconducting BaTiO3. The sensing materials modified by incorporating different additives, either by doping or dipping technique. The sensing performance of pure and modified BaTiO3 films was studied in terms of the change in conductance in the presence and absence of gases.

2 Experimental 2.1 Preparation of BaTiO3 Powder Powders of Ba(OH)2.8H2O and TiO2 of analytical reagent grade were ball milled to mix thoroughly at the same molar concentrations. The mixture was sintered at 1000oC for 6h to obtain BaTiO3 [30, 31]. The fine-grain powder of BaTiO3 was obtained by milling in a planetary ball mill for 2h. The sub micron size powder was then used to formulate the paste for printing of thick films. 2.2 Preparation of BaTiO3 Thick Films The thixotropic paste was formulated by mixing the fine powder of BaTiO3 with a solution of ethyl cellulose (a temporary binder) in a mixture of organic solvents such as

Cr2O3-doped BaTiO3 as an Ammonia Gas Sensor

159

butyl cellulose, butyl carbitol acetate and terpineol, etc. The ratio of the inorganic to organic part was kept at 75:25 in formulating the paste. This paste was screen printed [32, 33] on a glass substrate in a desired pattern. The films were fired at 550oC for 30 min. Silver contacts were made for electrical measurements. 2.3 Preparation of Cr2O3-doped BaTiO3 Thick Films Different wt% of CrO3 was added in BaTiO3, base material, followed by sintering at 550oC for 30min. CrO3 is not thermally stable above its melting temperature (197oC). At higher temperature, it loses oxygen to give stable Cr2O3. In this way, the Cr2O3doped BaTiO3 powder was obtained. The thick films of such powder were prepared by screen-printing technique. 2.4 Thickness Measurements The thickness of the thick films was measured by using the Taylor-Hobson (Talystep, UK) system. The thicknesses of the films were observed in the range from 65 to 70µm. The reproducibility in thickness of the films was possible by maintaining the proper rheology and thixotropy of the paste.

3 Characterization Results 3.1 Structural Properties Fig. 1 shows the X-ray diffractogram of a Cr2O3-doped BaTiO3 thick film, for 5.27wt% of Cr. The observed peaks are matching well with the reported data [34] of BaTiO3 confirming the single phase of the compound. The presence of separate Cr2O3 peaks indicates the composite nature of the material i.e. independent identity of Cr2O3. The average grain size calculated from Scherrer formula was 278 nm.

Fig. 1. X-ray diffractogram of Cr2O3-doped BaTiO3 thick film

160

G.H. Jain et al.

3.2 Microstructural Analysis Fig. 2(a) depicts a SEM image of an unmodified BaTiO3 thick film fired at 550oC. The film consists of voids and a wide range of particles with particle sizes ranging from 200 to 1330nm distributed non-uniformly. Fig. 2(b-d) depictS SEM images of Cr2O3-doped BaTiO3 thick films fired at 550oC with 0.56, 5.27 and 6.07wt% of Cr, respectively. The agglomeration of particles increases as Cr2O3 wt% increases. The change in doping concentration changes the particle sizes. The particle sizes ranging from 0.3 to 1.0μm (Fig. 2(b)), 0.5 to 1.0μm (Fig. 2(c)), and 0.66 to 2µm (Fig. 2(d)) were observed.

(a)

(c)

(b)

(d)

Fig. 2. SEM images of (a) unmodified BaTiO3 film and Cr2O3-doped BaTiO3 films with (b) 0.56wt%, (c) 5.27wt%, and (d) 6.07wt% of Cr.

3.3 Elemental Analysis The constituent elements such as Ba, Ti, O and Cr associated with various films are represented in Table 1. It reveals from the table that the film with 5.27wt% of Cr was observed to be the most oxygen deficient as compared to other samples. This deficiency could be attributed to the larger oxygen adsorption capability of the sample.

Cr2O3-doped BaTiO3 as an Ammonia Gas Sensor

161

Table 1. Quantitative elemental analysis

Samples Unmodified BaTiO3 Cr2O3-doped BaTiO3 : 0.56wt% Cr : 5.27wt% Cr : 6.07wt% Cr

Ba 80.46 82.08 79.07 76.39

wt% of Ti O 12.63 6.91 13.17 4.19 11.30 4.36 12.28 5.26

Cr 0.56 5.27 6.07

3.4 Electrical Conductivity of Cr2O3-doped BaTiO3 Films Fig. 3 represents the variation of conductivity with temperature for the pure and Cr2O3-doped BaTiO3 (BT) films. The legends suffixed with ‘a’ are the graphs for the conductivities of the films in the air ambient, while legends suffixed ‘g’ are the graphs for conductivities of the films in the NH3 gas ambient. It is clear from the graphs that the conductivity is varying approximately linearly with temperature for all films. The conductivity of Cr2O3-doped BaTiO3 films was observed to be increased.

Log(conductivity)1/Ohm-m

-0.2

BT(a) BT+0.56wt%Cr(a) BT+5.27wt%Cr(a) BT+6.07wt%Cr(a)

-1.2

BT(g) BT+0.56wt%Cr(g) BT+5.27wt%Cr(g) BT+6.07wt%Cr(g)

-2.2 -3.2 -4.2 -5.2 -6.2 1.3

1.5

1.7

1.9

2.1

2.3

2.5

2.7

1000/T(oK-1)

Fig. 3. Variation of electrical conductivity with temperature

3.5 Gas Sensing Performance 3.5.1 Gas Response of Unmodified BaTiO3 Film with Operating Temperature The variation of H2S gas response with operating temperature ranging from 100 to 450oC is shown in Fig. 4. The response goes on increasing with the temperature, attains its maximum (350oC) and then decreases with further increase in temperature. It is clear from graph that the optimum operating temperature is 350oC.

162

G.H. Jain et al.

60

Gas response

50 40 30 20 10 0 100

150

200 250 300 350 Operating Tem p.( oC)

400

450

Fig. 4. H2S gas (100ppm) response with operating temperature

3.5.2 Selectivity Fig. 5 shows the histogram of the selectivity of pure BaTiO3 film to various gases. The table attached to histogram shows the gas response values to various gases. It reveals that H2S gas is most selective against CO2 and poor selective against NH3 gas.

Gas response

50 40 30 20 10 0 Pure BT

CO

LPG

7

11

H2S

NH3

Ethanol

CO2

53.38

21

10

3.2

Gases

Fig. 5. Selectivity of pure BaTiO3 films to various gases

3.5.3 Gas Response of Cr2O3-doped BaTiO3 Films with Operating Temperature Fig. 6 depicts the variation of NH3 gas response with operating temperature. The response to NH3 gas goes on increasing with temperature for pure and doped films. The film with 5.27wt% of Cr was observed to be the most sensitive to NH3 gas and its optimum operating temperature was 350oC.

Cr2O3-doped BaTiO3 as an Ammonia Gas Sensor

120

Pure BT BT+0.56w t%Cr BT+5.27w t%Cr BT+6.07w t%Cr

100 Gas response

163

80 60 40 20 0 100

150

200

250

300

350

400

450

o

Operating temp.( C)

Fig. 6. Variations in NH3 gas (100ppm) response with operating temperature

3.5.4 Selectivity Fig. 7 shows the bar diagram of the selectivity of pure and Cr2O3-doped BaTiO3 films to various gases at optimum operating temperature. The table attached to bar diagram indicates the gas response values to various gases. It is observed that the pure BaTiO3 film showed highest H2S gas response while Cr2O3-doped BaTiO3 films showed highest response to NH3 gas.

Gas response

120 90 60 30 0

CO

LPG

H2S

NH3

Ethanol

CO2

Pure BT

7.0

11.0

53.4

21.0

10.0

3.2

BT+0.56w t%Cr

0.8

0.8

21.0

41.0

2.3

1.0

BT+5.27w t%Cr

1.2

1.3

4.2

111.7

3.6

2.3

BT+6.07w t%Cr

1.8

1.6

10.0

71.0

4.3

2.6

Gases

Fig. 7. Selectivity of Cr2O3-doped BaTiO3 films to various gases

164

G.H. Jain et al.

3.5.5 Response and Recovery Time of Cr2O3-doped BaTiO3 Sensor The transient response of Cr2O3-doped (5.2wt%) BaTiO3 film to NH3 gas is depicted in Fig. 8. The gas response of this film was found to be largest at 350oC. The 90% response and recovery levels were attained within 3 and 20s, respectively for this sample. The very short response and recovery time are the important features of this Cr2O3-doped BaTiO3 film to NH3 gas.

100oC 200oC 300oC 400oC

100

Gas response

80

150oC 250oC 350oC 450oC

60 40 20 0 0

4

8

12 Time(s)

16

20

24

Fig. 8. Transient response of Cr2O3-doped (5.27wt%) BaTiO3 to NH3 gas

4 Discussions It is known that atmospheric oxygen molecules are adsorbed on the surface of Cr2O3doped BaTiO3 semiconductor oxide in the forms of O- and O2- thereby decreasing the electronic conduction. Atmospheric oxygen molecules take electrons from the conduction band of Cr2O3-doped BaTiO3 to be adsorbed as O-BaTiO3. The reaction is as follows: O2(g) + 2e- → 2O-BaTiO3

(1)

The Cr2O3-doped BaTiO3 is more oxygen deficient as compared to pure BaTiO3. The excess Ba ions (due to oxygen vacancies) act as donors [35]. When reducing gas molecules like NH3 react with negatively charged oxygen adsorbates, the trapped electrons are given back to conduction band of Cr2O3-doped BaTiO3. The energy released during decomposition of adsorbed ammonia molecules would be sufficient for electrons to jump up into conduction band of Cr2O3-doped BaTiO3, causing an increase in the conductivity of sensor. The possible reaction is: 2NH3 + 3O-BaTiO3 → 3H2O + N2 + 3e-

(2)

For this reaction to proceed to the right hand side, some amount of activation energy has to be provided thermally. An increase in operating temperature surely increases the thermal energy so as to stimulate the oxidation of NH3 (equation (2)). The reducing gas

Cr2O3-doped BaTiO3 as an Ammonia Gas Sensor

165

(NH3) donates electrons to Cr2O3-doped BaTiO3. Therefore, the resistance decreases, or the conductance increases. This is the reason why the gas response increases with operating temperature. The point at which the gas response reaches maximum is the actual thermal energy needed for the reaction to proceed. However, the response decreases at higher operating temperatures, as the oxygen adsorbates are desorbed from the surface of sensor [36]. Also, at high temperatures the carrier concentration increases due to intrinsic thermal excitation, and the Debye length decreases. This may be one of the reasons for the decreased gas response at high temperatures. When the optimum amount of Cr (5.27wt%) is incorporated into the BaTiO3 material, the Cr2O3 species would be uniformly distributed. Due to this, not only the initial resistance of the film is high but this amount would also be sufficient to promote the catalytic reaction effectively and the overall change in the resistance on exposure of ammonia gas leading to high sensitivity. When the amount of Cr2O3 on the surface of base material, BaTiO3, is less than the optimum, the dispersion may be poor and the sensitivity of the film is observed to be decreased since this amount may not be sufficient to promote the reaction effectively. On the other hand, as the amount of Cr2O3 on BaTiO3 surface is more than the optimum, an additive Cr2O3 would be distributed more densely. As a result, base material BaTiO3 would be masked and the overall change in the resistance on the exposure of gas would be smaller leading to lower response to ammonia gas.

5 Conclusions Following statements can be made from the experimental results. 1) 2) 3) 4)

The thick films of unmodified BaTiO3 were sensitive to H2S gas. The Cr2O3-doped BaTiO3 was observed to be semiconducting in nature. The Cr2O3-doped BaTiO3 was most sensitive and selective to NH3 gas. The resistance of the Cr2O3-doped BaTiO3 films in ambient air was observed to be very high. 5) The resistance of the Cr2O3-doped BaTiO3 films was observed to decrease suddenly upon exposure to NH3 gas at optimum operating temperature. 6) The fast recovery of the sensor could be attributed to the larger oxygen deficiency in BaTiO3. The larger oxygen deficiency would enable BaTiO3 to adsorb more oxygen ions, helping the sensor to recover fastly. 7) Cr2O3-doped BaTiO3 was observed to be more sensitive to NH3 gas than unmodified BaTiO3.

References 1. Zhou, Z.G., Tang, Z.L., Zhang, Z.T.: Studies on grain-boundary chemistry of pervoskite ceramics as CO gas sensors. Sens. Actuators B 93, 356–361 (2003) 2. Haayman, P.W., Van Dam, R.W., Klaasens, H.A.: Method of preparation f semiconducting materials. German Patent 929350 (1995) 3. Jaffe, P., Cook Jr., W.R., Jaffe, H.: Piezoelectric Ceramics, p. 94. Academic Press, New York (1971)

166

G.H. Jain et al.

4. Sahuri, O., Wakino, K.: Processing techniques and applications of positive temperature coefficient thermistors. IEEE Trans. Component 10, 53 (1963) 5. Ravi, V., Kutty, T.R.N.: Current limiting action of mixed phase BaTiO3 ceramic semiconductors J. Appl. Phys. 68, 4891 (1990) 6. Kutty, T.R.N., Ravi, V.: Varistor property of n-BaTiO3 based current limiters. Appl. Phys. Lett. 59, 2691 (1991) 7. Zhou, Z.G., Tang, Z.L., Zhang, Z.T.: Studies on grain-boundary chemistry of pervoskite ceramics as CO gas sensors. Sens. Actuators B 93, 356–361 (2003) 8. Tang, Z.T., Zhou, Z.G., Zhang, Z.T.: Experimental studies on the mechanism of BaTiO3 based PTC CO gas sensor. Sens. Actuators B 93, 391–395 (2003) 9. Ishihara, T., Kometani, K., Nishi, Y., Takita, Y.: Improved sensitivity of CuO-BaTiO3 capacitance type CO2 sensor. Sens. Actuators B 28, 49–54 (1995) 10. Liao, B., Wei, Q., Wang, Q.Y., Liu, Y.X.: Study on CuO-BaTiO3 semiconductor CO2 sensor. Sens. Actuators B 80, 208–214 (2001) 11. Haeusler, A., Meyer, J.U.: A novel thick film conductive type CO2 sensor. Sens. Actuators B 34, 388–395 (1996) 12. Wang, J., Xu, B.K., Liu, B.F., Zhang, J.C., Zhang, T.: Improvement of nanocrystaline BaTiO3 humidity sensing properties. Sens. Actuators B 66, 159–160 (2000) 13. Wagh, M.S., Patil, L.A., Seth, T., Amalnerkar, D.P.: Surface cupricated SnO2-ZnO thick films as a H2S gas sensor. Mater. Chem. Phys. 84, 228–233 (1985) 14. Kanefusa, S., Nitta, M., Haradome, M.: High sensitivity H2S gas sensor. J. Electrochem. Soc. 132, 1770–1773 (1985) 15. Lantto, V., Romppainen, P.: Response of some SnO2 gas sensors to H2S after quick cooling. J. Electrochem. Soc. 135, 2550–2556 (1988) 16. Tamaki, J., Maekawa, T., Miura, N., Yamazoe, N.: CuO-SnO2 element for highly sensitive and selective detection of H2S. Sens. Actuators B 9, 197–203 (1992) 17. Manorama, S., Sarala Devi, G., Rao, V.J.: Hydrogen sulfide sensor based on tin oxide deposited by spray pyrolysis and microwave plasma chemical vapor deposition. Appl. Phys. Lett. 64, 3163–3165 (1994) 18. Sarala Devi, G., Manorama, S., Rao, V.J.: Gas sensitivity of SnO2/CuO heterocontacts. J. Electrochem. Soc. 142, 2754–2756 (1995) 19. Tamaki, J., Shimanoe, K., Yamada, Y., Yamamoto, Y., Miura, N., Yamazoe, N.: Dilute hydrogen sulfide sensing properties of thin film prepared by low pressure evaporation method. Sens. Actuators B 49, 125–186 (1998) 20. Vasiliev, R.B., Rumyantseva, M.N., Yakovlev, N.V., Gaskov, A.M.: CuO/SnO2 thin film heterostructures as chemical sensor for H2S. Sens. Actuators B 50, 186–193 (1998) 21. Mangamma, G., Jayaraman, V., Gnanasekaran, T., Periaswami, G.: Effects of silica addition on H2S sensing properties of CuO-SnO2 sensors. Sens. Actuators B 53, 133–139 (1998) 22. Yuanda, W., Maosong, T., Xiuli, H., Yushu, Z., Guorui, D.: Thin film sensors of SnO2CuO-SnO2 sandwich structure to H2S. Sens. Actuators B 79, 187–191 (2001) 23. Tamaki, J., Yamada, Y., Yamamoto, Y., Matsuoka, M., Ota, I.: Sensing properties of dilute hydrogen sulfide of ZnSb2O6 thick film prepared by dip-coating method. Sens. Actuators B 66, 70–73 (2000) 24. Yamazoe, N., Kurokawa, Y., Seiyama, T.: Effects of additives on semiconductor for gas sensor. Sens. Actuators B 4, 283–289 (1983) 25. Lee, M.S., Meyer, J.U.: A new process for fabricating CO2-sensing layers based on BaTiO3 and additives. Sens. Actuators B 68, 293–299 (2000)

Cr2O3-doped BaTiO3 as an Ammonia Gas Sensor

167

26. Ishihara, T., Kometani, K., Nishi, Y., Takita, Y.: Improved sensitivity of CuOBaTiO3capacitive type CO2 sensor by additives. Sens. Actuators B 28, 49–54 (1995) 27. Lee, J.D.: Concise In-organic Chemistry, 5th edn., p. 698 28. Manku, G.S.: In-organic Chemistry, pp. 465–467 29. Patil, L.A., Wani, P.A., Sainkar, S.R., Mitra, A., Pathak, G.J., Amalnerkar, D.P.: Studies on “fritted” thick films of photoconducting CdS. Mater. Chem. Phys. 55, 79 (1998) 30. Aslam, M., Chaudhary, V.A., Mulla, I.S., Sainkar, S.R., Mandale, A.B., Belhekar, A.A., Vijaymohan, K.: A highly selective ammonia gas sensorusing surface-ruthenated zinc oxide. Sens. Actuators A 75, 162–167 (1999) 31. Chaudhary, V.A., Mulla, I.S., Vijaymohan, K.: Impedance studies of an LPG sensor using surface ruthenated tin oxide. Sens. Actuators B 55, 127–133 (1999) 32. Niranjan, R.S., Chaudhary, V.A., Sainkar, S.R., Patil, K.R., Mulla, I.S., Vijaymohan, K.: Surface ruthenated tin oxide thin-film as a hydrocarbon sensor. Sens. Actuators B 79, 132– 136 (2001) 33. Chaudhary, V.A., Mulla, I.S., Vijaymohan, K.: Comparative studies of doped and surface modified tin oxide towards hydrogen sensing: synergistic effects of Pd and Ru. Sens. Actuators B 50, 45–51 (1998) 34. ASTM Data Manuals, pp. 34–129 35. Ishihara, T., Kometani, K., Hashida, M., Takita, Y.: Application of mixed oxide capacitor to the selective carbon dioxide sensor. J. Electrochem. Soc. 138, 173 (1991) 36. Cotton, F.A., Wilkinson, G.: Advanced Inorganic Chemistry, 2nd edn., p. 828. Interscience Publishers, John Wiley & Sons (1967)

Physical and Electrical Modeling of Interdigitated Electrode Arrays for Bioimpedance Spectroscopy M. Ibrahim1, J. Claudel1, D. Kourtiche1, B. Assouar2, and M. Nadi1 1

Electronic Instrumentation Laboratory of Nancy, Nancy University, France 2 Institute Jean Lamour, Nancy University, France

Abstract. This paper concerns a theoretical and electrical modelling of interdigital sensor in a wide band frequency. A theoretical approach is proposed to optimize the use of the sensor for bioimpedance spectroscopy. CoventorWare® software was used to modelize in three dimensions the interdigital sensor system for measuring electrical impedance of biological medium. Complete system simulation by Finite element method (FEM) was used for sensor sensitivity optimization. The influence of geometric parameters (number of fingers, width of the electrodes, …), on the impedance spectroscopy of biological medium was studied. A high level description of the sensor and the biological medium was also developed under VHDL-AMS with SystemVision® software from mentor graphics. The simulation results are compared with measurements obtained with a true interdigitated sensor illustrating a good correlation. This shows that even the theoretical model is simple, it remains very effective.

1 Introduction Electrical impedance measurement has been demonstrated as a potential useful approach in biomedical applications. This method allows to determine the physiological status of ex vivo or living tissues as well as their electromagnetic characterization [1]. The changes induced by some pathologies could be associated with variations of essential tissue parameters such as the physical structure or the ionic composition that can be reflected as changes in the passive electrical properties. The range of applications derived from this technique is quite wide [2, rigaud et morucci]. Planar interdigitated electrode arrays have become more prominent as a sensor device due to the ongoing miniaturization of electrodes and the low cost of those systems [3]. An important advantage of these sensor devices is the simple and inexpensive mass-fabrication process and the ability to use these devices over a wide range of applications without significant changes in the sensor design [4-5]. Typically these sensors have been used for the detection of capacitance, dielectric constant and bulk conductivity in biological medium [6-7]. Basically, the structure consists on two S.C. Mukhopadhyay et al. (Eds.): New Developments and Appl. in Sen. Tech., LNEE 83, pp. 169–189. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

170

M. Ibrahim et al.

parallel coplanar electrodes whose design (width, gap between electrodes, length) is repeated periodically [8]. This paper, based on previous work [9], presents a new approach of physical and electrical modelling system of a biological sensor. The electrical and physical modelling of the Interdigital sensor and the medium was developped by using COVENTORWARE® and Systemvision (MENTORGRAPHICS®) sofware respectively. Section two describes the correlation between design parameters and frequency behavior in coplanar impedance sensors. By developing total impedance equations and modeling equivalent circuits we propose a theoretical optimization of the geometrical parameters of the sensor. One objective was to get the optimal ratio between the width of the electrodes and the gap. The third section gives a description of the sensor and medium model with the finite element method FEM using CoventorWare® software. We studied the influence of the medium’s physical properties on the frequency sensor response. We simulated the influence of electrodes number and we found the number 16 as optimized for a cross section 1mm*1mm. In the fourth section, we give a description of electrical model for IDT sensor with VHDL-AMS (SystemVision software®). This software provides an electrical approach that can be readily used in current electronic design flow to include distributed physics effects. VHDL-AMS language permits to simulate the sensor and medium. It allows fast simulations to validate a simplified model and to serve as a reference to power conditioning. The sensor manufacturing is described in the fifth section. A test bench based on a measurement system composed by RCL meter connected with computer was built to test the sensor. Preliminary bioimpedance measurements were done on calibrated ionic solution of NaCl. Section six concludes on the validity of models and presents the perspectives.

2 Theoretical Aspect 2.1 Description of Interdigital Sensors Interdigital sensor is equivalent to a parallel plate capacitor (Figure 1) [10-11]. An electric field is created between the positive and negative electrodes (instantaneous polarity) shown on figure 1 (a) and (b) respectively. When a medium is placed on the sensor, the electric field across the medium under test is also shown on figure 1 (c). The dielectric properties of the material as well as the geometry of the material under test affect the capacitance and conductance between the two electrodes. The variation of the electric field can be used to determine the properties of the material depending on the application. To use them in bioimpedance domain, a potential difference is applied between two electrodes and the electrical impedance between the electrodes is measured. The electrodes of the interdigital sensor are coplanar, so the measured impedance will have a very low signal-to-noise ratio.

Physical and Electrical Modeling of Interdigitated Electrode Arrays

171

Fig. 1. Funtion principle of an Interdigital sensor.

The main dimensional characteristics parameters of such a pair of electrodes are (figure 2): 1. 2. 3. 4.

The number of digits N. The length of digit L. The digit width W. The distance between a digits S.

Fig. 2. Interdigitated sensor structure and dimensional parameters

2.2 Equivalent Circuit Model Figure 3.a gives the configuration of the planar structure when switched as an interdigitated impedance cell. When such a cell is immersed in an electrolyte, the simplified equivalent electrical circuit is represented by figure 3.b.

Fig. 3. (a) Configuration of interdigitated impedance cell and (b) its equivalent circuit

172

M. Ibrahim et al.

The electrical elements in the equivalent circuit modelize the physical phenomena that determine the total electrical impedance (Z) detected in the measurement cell (figure 3.a). Thus, the equivalent model elements could be expressed in terms of physical quantities. The resistance RSol of the resistance is the sensing element and is related to the electrolyte conductivity σSol by the cell constant KCell [12]: RSol =

KCell

(1)

σSol

The cell constant KCell is equal to [13] : KCell =

with K (k ) = ∫

1 0

2

(N - 1)L

1

(1 - t² )(1 - k²t² )

()

K k

. K

( 1 - k² )

⎛π W ⎞ . ⎟ ⎝2 S+W⎠

dt and k = cos⎜

(2)

Where N is the number of fingers, S the finger spacing, W the finger width and L the finger length. The function K(k) is the incomplete elliptic integral of the first kind [14]. So, the cell constant depends entirely on the geometry of the sensor. The lead resistance RLead is the result of the series resistances of the connecting wires. Direct capacitive coupling between the two electrodes is represented by the cell capacitance CCell given by: CCell =

ε 0. ε r, Sol

(3)

KCell

with εr,sol ≈ εr,water = 80. One model element which is not drawn in figure 1.b is a capacitor representing the direct capacitive coupling between the connecting wires. This capacitor comes in parallel with CCell and will therefore virtually increase the observed cell capacitance. The impedances that explain the interface phenomena occurring at the electrodeelectrolyte interfaces, are simplified to the double layer capacitances CDL. These are depending on the electrode material and the electrolyte solution but, for horizontal electrode surfaces, they can be approximated by: CDL = 0.5.A. CDL, Surface = 0.5.W.L.N.CDL, Surface

(4)

where A is the electrode surface and CDL,Surface the characteristic of the double layer capacitance of the electrode-electrolyte system. One must notice that the factor 0.5 is the result of CDL determined by only half of the electrode surface A. The characteristic of the double layer capacitance CDL, Surface is supposed to be equal to the characteristic capacitance of the Stern layer for electrolytes having a quite high ionic strength. This characteristic capacitance of the Stern layer is approximated by CStern, Surface = 10-20 μF/cm2 [15, 16].

Physical and Electrical Modeling of Interdigitated Electrode Arrays

173

Based on the equivalent circuit of figure 1.b, the total observed impedance can be expressed as

( )

Z jω = 2RLead +

Z1

(5)

j.ω.CCell. Z1 + 1

Where Z1 = RSol +

2 j.ω.CDL

2.3 Theoretical Optimization of the Sensor

Impedance, Ohm

Figure 4 shows a schematic graph of total impedance of equivalent circuit (Figure 3.b). There are three zones in the impedance spectrum, which correspond to the three kind of elements in equivalent circuit. The frequency dependent property of these zones can be analysed using the equivalent circuit mentioned above. As shown in Figure 3.b, there are two parallel branches (CCell and CDL). When the frequency is not adequatly higher than fHi, the current cannot cross the middle of the dielectric capacitor. That is, the capacitor is inactive, and just acts as an open circuit. Then , the total impedance corresponds to the double layer capacitance and solution resistance in series. Although both CDL and RSol provide to the total impedance below fHi, each of them dominates at different frequencies.

Cdl

Rsol

Ccell

Frequency, Hertz

Fig. 4. Schematic diagram of total impedance–frequency plots

The CDL becomes essentially resistive at the frequency lower than fLow, and it contributes mainly to the total impedance value: Z≈

2 + jω .CDL.RSol j.ω .CDL

(6)

174

M. Ibrahim et al.

and fLo ≈

1 π. RSol CDL

(7)

The impedance increases with the decrease in the frequency (double layer region). However, above fLow, double layer capacitance offers no impedance. This is explained by the fact that only the resistance of the solution contributes to the impedance while below fHi the influence of CCell is not yet indicative, the total impedance is independent of the frequency (resistance of the solution zone). This results into a frequency band, restricted by fLo and fHi, in which the results (e.g. the conductivity) can be deduced from the observed impedance using:

( )

Z jω = 2RLead + RSol

(8)

To optimize the impedance cell leads to maximize the plateau width in the curve of figure 4. When the frequency is higher than fHi, the current cross the middle of the dielectric capacitor instead of crossing the electrolyte solution resistance. That is, the branch (CDL + RSol + CDL) is inactive, and the branch (CCell) is active. In this zone, the dielectric capacitance of the medium governs the total impedance, and the double layer capacitance and medium resistance could be neglected. Thus, the total impedance value is inversely proportional to the frequency: RSol

Z≈

(9)

j.ω .CCell.RSol + 1

and fHi ≈

1 2. π .RSol.CCell

(10)

Or in terms of conductivity parameters:

σSol

fLo ≈

(11)

0.5.π .W. L. N. CDL, Surface . KCell

and fHi ≈

σSol 2.π . ε 0. ε

(12) r, Sol

Note that the higher boundary frequency, fHi, is not dependent on the geometry, according to the theory, when the wiring capacitance is not present. Obviously, maximising the width of the plateau can only be done by decreasing the lower boundary frequency. In order to make the lower boundary frequency (11) as low as possible, the geometrical term

Physical and Electrical Modeling of Interdigitated Electrode Arrays

W. L. N. KCell (N, L, S, W)

175

(13)

should be maximised. When using a square structure of L*L, one variable can be eliminated since L*L : L = N. (W+S) - S With L in mm and S in microns L+S≈L Finaly L = N. (W+S)

(14)

However, it is more illustrative to introduce a factor a = S/W. Using the substitutions: W=

L N

.

1

and S =

(a + 1)

L N

.

a

(a + 1)

(15)

Which are based on equation (15) together with the ratio a, expression (13) becomes: 2.L

.

1

(N - 1) (a + 1)

()

K k

. K

( 1 - k² )

= X (N, L) * Y(a)

(16)

Where

(

)

X N, L =

2.L

(N - 1)

and Y(a ) =

1

(a + 1)

()

K k

. K

( 1 - k² )

The function K has the same meaning as it had in equation (2). This optimisation expression, which has to be maximised in order to minimise flo, can be split into two parts. The first is the factor X (N, L), showing that the cell size L*L must be as large as possible while the number of fingers must be reduced. Since there is no maximum in the desired cell size, with respect to the optimisation of flo, the value L can be chosen arbitrarily. A cell size of about 1*1 mm² will be used in the modeling. The optimal number of fingers N has a minimum for N = 2 since this is the lowest possible number of fingers. On the other side, the sensitivity of the impedance measurement depends on the number of fingers. Then, the modelling allows us to study the influence of the number of fingers on the impedance measurement. The highest the number of fingers, the highest the sensitivity. The factor X (N, L) is related to the sum W+S according to equation (14), the second factor in expression (16) has only the ratio W/S = a as a parameter. In figure 5 this factor is plotted as a function of a, a varying from 0 to 10. For a = 1, the finger width is equal to the gap between them. It can be seen that this is not the optimal ratio, for a = 0.66 the optimisation function has a maximum which means that the finger width should be approximately 1.5 the gap width W = 3S/2. This maximum for the function denoted in the figure as f (a) is equal to 0.66.

176

M. Ibrahim et al.

Therefore, when designing a square structure, the design rule, based on the maximum frequency range criterion becomes according to equation (14): L = (2.5N-1).S. When the minimum number of fingers, N = 2, is also taken into account, the design rule becomes : L = 5S. 0.7

Y(a) 0.6 X: 0.66 Y: 0.51

0.5

Y(a)

0.4

0.3

0.2

0.1

0

0

1

2

3

4

5

6

7

8

9

10

a

Fig. 5. Optimisation of the S/W ratio.

3 COVENTORWARE® Modeling 3.1 Model Description In this section, the model of the sensor loaded by the blood medium is described. This model was developed for simulation with the finite element method FEM using CoventorWare® software. We used the module MEMS electro quasistatic harmonic response proposed by the software. 3.1.1 Sensor Modeling The structure of the impedance sensor used in this simulation is at micron scale. It is composed by layers of glass and platinum and is showed in figure 6. A glass layer of length 1 000 µm and width of 1 000 μm has been defined as a substrate thickness of 1 000 μm to carry the system (gray layer on figure 6). The glass is a good electrical insulator (10-17S.m-1) with a relative permittivity around 5-7. As the glass has a very small permittivity, we do not need to put an insulating layer between the substrate and the electrodes. Next, we define a mask of platinum electrodes (thickness 1 μm) known as a good conductor 9.66*106 S·m-1, using the graphical editor of CoventorWare. This is deposited on the glass (red layer on figure 5). The effective region of electrodes forms a square 1 000 μm*1 000 μm.

Physical and Electrical Modeling of Interdigitated Electrode Arrays

177

Fig. 6. 3D view of a planar interdigitated electrode arrays.

The characteristic parameters of the electrodes, the length of digit L, the number of digits N, the digit width W and the distance between a digits S, were selected according to formulas of optimization given in section 2: L is fixed at 1000 μm, W=

3.S 2

μm, S =

L

(5.N/2 − 1)

μm

and N is a variable. For example, N = 4 electrodes, S = 1000/(5.4/2-1) = 111 μm and W = 3.111/2 = 167 μm. 3.1.2 Modeling the Medium The medium modeled is composed by two layers: The two layers DL, that describe the interface phenomena occurring at the electrode-electrolyte interfaces, are simplified to a single equivalent layer. This is the first layer shown on figure 7 (green layer). The second layer is the blood (blue layer). The double layer can be formed from interface phenomena platinum electrodemiddle blood, with a thickness about 50 A° [17]. The relative permittivity of the layer DL (thickness 50 A°) for medium blood is about 97 [18]. The 50 A° thickness creates problems for the mesh system, then they were shifted to a thickness of 10 µm (equal to 50 A°*2*1000) and a relative permittivity 194 000 (equal to 97 *2*1000) in order to not change the capacity of this layer. One can notice that DL layer is almost insulating. We tested the blood as a biological medium with 0.7 S/m as conductivity and 80 at a frequency of 1 Ghz as relative permittivity [19]. The layer of the blood was equal to 500µm of thickness.

178

M. Ibrahim et al.

Fig. 7. 3D view of a planar interdigitated electrode arrays and the blood medium.

3.2 Results and Discussion A 1 volt sinusoïdal signal between terminals of interdigitated electrodes and a frequency range from 100 Hz to 1000 GHz was applied. The biological impedance was measured for different cases of electrode configuration. For the first three models, we arbitrarily chose N equal 16 electrodes. Figure 8 shows the influence of the double layer DL on biological impedance of the blood medium. Simulation results are obtained with and without interface double layer DL. One can notice that the impedance is constant in the second case (without DL) and does not take into account the cut off frequencies fLo and fHi. The impedance consists of only resistance in parallel with capacitance (negligible). Figure 9 shows the influence of the conductivity of the blood medium. Simulation results are obtained for two different conductivity : 0.7 and 9 m/s. When the conductivity increases, the RSol decreases and the plateau shifts to a lower impedance. In addition, a change in the height of the plateau implies a change in the boundary frequencies fLo and fHi, (equations 11 and 12 of section 2). Figure 10 gives the influence of medium permittivity on biological impedance and the boundary frequencies fLo and fHi. By comparing simulation results for two different permittivities, one can observe that the permittivity does not affect the impedance for small frequencies (less than 107 Hertz). For high frequencies when the permittivity increases, the impedance decreases and the fHi takes a smaller value which is similar to the equations (12) of the second paragraph. Figure 11 gives simulation for N fingers. Six different structures were used with respect to the conditions of optimization explained above in the paragraph sensor modeling. The case N equal 2 is taken as reference since this is the lowest possible number of fingers. In this case we can see that the impedance shows an unpredicted resonance at high frequencies.

Physical and Electrical Modeling of Interdigitated Electrode Arrays

179

For N equals 8 a resonance is not observed but an unstable plateau occurs for high frequencies. Where N equals 12, there is not resonance or instability, but the distinction of fHi is not clear. The two curves of 16 and 20 electrodes are almost together, and we can distinguish three regions of frequencies between fLow and fHi. For N equal 30, the impedance curve shows a resonance at high frequency.

10

10

Impedance, Ohm

10

10

10

10

10

7

16 electrodes, with interface double layer DL 16 electrodes, out interface double layer DL

6

5

4

3

2

1

0

10 2 10

10

4

flow

10

6

10

8

fhigh1010

10

12

Frequency, Hertz

Fig. 8. Simulated impedance of a blood medium deposited on the structure of the sensor number of fingers 16 electrodes, with and without interface double layer DL.

10

10

Impedance, Ohm

10

10

10

10

10

10

8

16 electrodes, conductivity = 0.7 s/m 16 electrodes, conductivity = 9 s/m

7

6

5

4

3

2

1

0

10 2 10

10

4

flow

flow 10

6

10

8

fhigh 10

10

fhigh

10

12

Frequency, Hertz

Fig. 9. Impedance-frequency characteristics for t conductivities of 0.7 S / m and 9 S/m

M. Ibrahim et al.

7

10

16 electrodes, permittivity = 80 16 electrodes, permittivity = 5200

6

10

5

Impedance, Ohm

10

4

10

3

10

2

10

1

10

0

10 2 10

4

10

6

10

10

8

10

12

10

10

Frequency, Hertz

Fig. 10. Influence of the blood permittivity on the impedance. 7

10

2 electrodes 8 electrodes 12 electrodes 16 elctrodes 20 electrodes 30 electrodes

6

10

5

10

Impedance, Ohm

180

4

10

3

10

2

10

1

10

0

10 2 10

4

10

6

10

8

10

10

10

12

10

Frequency, Hertz

Fig. 11. Biological impedance of the medium at various the number of digits N.

Physical and Electrical Modeling of Interdigitated Electrode Arrays

181

4 VHDL-AMS Modeling 4.1 Model Description In this section, we describe the model of the sensor and the medium that we developed and simulated using the high-level behavioral language VHDL-AMS using SystemVision software. The systems libraries was used to describe the model as an electrical circuit. 4.1.1 Electrical Modeling for Sensor and Medium In VHDL-AMS, the sensor and medium are described as an electrical circuit. In this circuit, the impedance of medium depends on the geometry of the sensor; therefore one must model the sensor loaded by the medium. The general model, in figure 12, is obtained by symmetry from the simple model between two classical plane electrodes; that represents the impedance between two fingers.

Fig. 12. General electrical model for an interdigitated sensor with medium

CDL, e represents the double layer capacity per finger, ZMed, e and CCell, e the impedance of medium and the cell capacitance between two fingers. These components are governed by the same equations (1), (3) and (4), but with a different cell constant, which does not contain the term N and (N-1) : the electrodes form factor Ke (equation 17). Ke =

2

.

L K

()

K k

( 1 - k² )

(

)

= KCell. N - 1 and ZMed =

ZMed,

e

(N - 1)

= Zbe(jω ). KCell

(17)

This model can be easily be simulated in VHDL-AMS, but its remains very difficult to write its equations. It is necessary to simplify it in order to allow a simple use. To do this, one supposes that the effect of the double layer can be divided into two equal parts. We divide the interface capacity into two equal parts to obtain a parallel circuit between the electrodes (figure 13). This new model can be simplified in a simple circuit with 4 components : 2 double layer capacitors CDL, the cell capacitor CCell and the medium impedance ZMed. This is the circuit presented in the

182

M. Ibrahim et al.

theoretical part in figure 3.b. This simplified model proves that one can find the conductivity and permittivity of medium by using cell constant.

Fig. 13. Steps to simplify the model.

4.1.2 VHDL-AMS Description The sensor loaded by a blood sample is described, using VHDL-AMS, like a dipole consisting on simple passives components. Their values are calculated, from the geometric characteristics of the sensor and the medium. For example, the electrode constant is calculated with the Euler method in a loop (figure 14). The final circuit is realized by placing basic components with a loop; it is the “Port Map” function (figure 15). constant nu constant ki constant kip

: real : real : real

:= W/(W+G); :=sin(nu*MATH_PI/2.0); :=sqrt(1.0-ki**2);

pure function K(x:real) return real is variable y,nbr,pas,t,result: real; begin result:=0.0; pas:= 1.0/100000.0; for nbr in 1 to 99999 loop t:=real(nbr)/100000.0; y:= result + pas*(1.0/sqrt((1.0-t**2.0)*(1.0-(x**2.0)*(t**2.0)))); result:=y; end loop; return y; end K; constant Fac:real:= L*K(ki)/(2.0*K(kiP)); --------------------------------------------------------------------------------------constant ccel: real:=PHYS_EPS0*(epssub+epsech)*Fac;

Fig. 14. Example of calculation of constants (here: the electrodes form factor and CCell).

Physical and Electrical Modeling of Interdigitated Electrode Arrays

183

begin proc1 :for i in 1 to (N/2) generate c1:entity WORK.capa(ideal) generic map (cap => cdl) port map ( P1 => P1, P2 => NM(2*i-1)); c2:entity WORK.capa(ideal) generic map (cap => cdl) port map ( P1 => P2, P2 => NM(2*i)); end generate proc1; proc2 :for i in 1 to (N-1) generate c1:entity WORK.res(ideal) generic map ( re => rsol) port map ( P1 => NM(i), P2 => NM(i+1)); c2:entity WORK.capa(ideal) generic map ( cap => ccel)

Fig. 15. Placement of components by Port Map for an electrolytic medium.

4.2 Results and Discussion 4.2.1 Ionic Solution Sample An ionic solution is characterized by its conductivity σSol. So, the medium impedance ZMed,e is a resistance. This model has been simulated in VHDL-AMs using SystemVisionTM with the same parameters used in ConventorWare® model. We choose N=16, L=1mm, W=38µm, S=26µm and σSol=0.7 S/m. The frequency analysis simulation is made by connecting an alternative current source to the sensor. An AC current of 1A was applied at a frequency varying from 100Hz to 1GHz. Figure 15 gives the simulation result of the impedance variation. The central plateau is the resistance of the solution. Impedance (Ohm) 10

10

10

4

- simulation for 16 electrodes, conductivity=0,7 S/m

3

2

10

2

10

3

10

4

10

5

10

6

10

7

10

8

10

9

Frequency (Hz)

Fig. 16. Simulated impedance of a sensor with 16 electrodes, for an electrolytic medium with a conductivity of 0.7 S/m.

184

M. Ibrahim et al.

4.2.2 Blood Sample The electric and dielectric behavior of blood sample present more properties than a simple ionic solution. It is constituted by free cells in an electrolyte : the blood plasma. The cells are composed by an electrolyte which is contained in an insulating membrane. The classical modeling is a resistance for the electrolyte, and a résistance in series with a capacitor is given by Fricke’s model (figure 17). We take for ZMed,e the equation of figure 17 with the electrodes form factor. For this simulation, we keep the same geometric parameters than the previous simulation for ionic solution. The values for blood’s parameters are σP=1.5 S/m, σC=1 S/m, Cm=1.75 µF/cm² and Ø=55%. The surface capacity of membrane is high but less than the capacity of double layer. Its effect appears at higher frequency. The results of the simulation are given in figure 18 and figure 19.

Fig. 17. Fricke’s Model for blood and its equivalent impedance in [Ohm.m]. With rP, rC, Cm, a and Ø the resistivity of plasma in [Ohm.m], the intern resistivity of blood cells in [Ohm.m], the membrane surface capacity in [F/m], the radius of blood cells in [m] and the volume in percentage of blood cells. Impedance (Ohm) 4

10

3

- Simulation for 16 electrodes; blood model

10

2

10

2

10

3

10

4

10

5

10

6

10

7

10

8

10

9

10

Frequency (Hz)

Fig. 18. Simulated impedance of a sensor with 16 electrodes, for a blood medium.

Physical and Electrical Modeling of Interdigitated Electrode Arrays y 10

p

185

y

-2

Conductivity (S) 10

10

-4

-6

Capacity (F) 10

10

10

-8

-10

-12

10

2

10

3

10

4

10

5

10

6

10

7

10

8

10

9

Frequency (Hz)

Fig. 19. Conductance and capacity of a sensor with 16 electrodes for blood medium.

The figures 18 represents the impedance; one can see two plateaux which correspond respectively to the plasma resistance and plasma resistance in parallel with blood cell resistance. The value of capacitance is difficult to evaluate in this type of curve, but it is easily find in the figure 19. Figure 19 gives the conductance and the capacity versus the frequency. For the conductivity, each plateau represents respectively the plasma resistance and the plasma resistance in parallel with the blood cell resistance. In capacity, each plateau represents respectively the double layer capacity, the blood cells capacity and CCell.

5 Experimental Validation 5.1 Sensor Manufacturing The sensor was provided by our colleague from the IJL team (Institut Jean Lamour, Henri Poincaré-Nancy 1 University). It was obtained by a deposit of 500 nm platinum on an insulating glass substrate in a 5 steps process: • • • • •

Sensor cleaning with acetone and isopropanol. Deposition of a platinum by ion-beam sputtering. UV lithography: deposit resin, mask application, insolation and development. Ion beam etching. Removal resin with acetone and isopropanol.

Its geometrical parameters are N=100, W=4µm, S=8µm and L=1000µm. This is a first prototype sensor for which we recycled a mask designed for SAW interdigitated sensor previously developped at IJL. A printed circuit board (PCB) was designed and built to connect the sensor with an appropriat measuring instrument. The connections between the sensor and the PCB were realized using a gold wires bonding figure 20.

186

M. Ibrahim et al.

Fig. 20. Sensor and PCB connection and its gold wires bonding and partial microscopic view of the fingers

5.2 Measurements The measurement system is composed of an LCR meter HP4284A controlled by © VEE software with a GPIB interface. It allows a fast and automatic measurements between 20Hz and 1 MHz. A photography of this system is given in the figure 20a. The measurements were performed with a calibrated drop of ionic solution. This solution contains 0.9% of NaCl, and has an approximate conductivity of 0.72 S/m. We placed a drop directly on the sensor, as shown on figure 21.b. The sensor connections (bonding) were not isolated, and can cause some errors of measurements. These first measurements were done just to validate the model . Figure 22 shows the measurement results compared to VHDL-AMS simulation results.

(a) Fig. 21. (a): Measurement system using the HP4284A PRECISION LCR METER. (b): Deposit of a drop of calibrated solution on the sensor.

Physical and Electrical Modeling of Interdigitated Electrode Arrays

187

(b) Fig. 21. (continued) Impedance (Ohm) 5

10

4

Measured impedance for a NaCl calibrated solution: conductivity = 0.72 S/m

10

Simulated impedance for an ionic solution: conductivity = 0,72 S/m 3

10

2

10

2

10

3

10

4

10 Frequency (Hz)

5

10

6

10

7

10

Fig. 22. Comparison between measurements and the two simulations

Simulation of complex system by the finite element method with CoventorWare®, takes lot of memory for computation. For N equal 100 electrodes, our simulation equipment was not able to model and simulate the whole system. The preliminary experimental measurements agree with the simulation. One can see a plateau at higher frequency, at a level close to that of simulation, but the instrument frequency limitation do not allow to check the precise level. The slope of the curve is slightly lower in measurement, because the real system response is not exactly the same those classical passive electronic components.

188

M. Ibrahim et al.

6 Conclusion This paper presents a comparative approach for simulation of biological sensor modeling in physical and electrical domains using two softwares. CoventorWare® software for three dimensional interdigilal sensor simulation techniques to analyse the influence of the physical properties of the medium and the impedance response was used. The simulation results are in agreement with the theoretical equations of optimization. This optimization method used for bioimpedance spectroscopy sensor is obtained from theoretical equations, by developing total impedance equations and modeling equivalent circuits. The equations given relate the cutoff frequencies to the geometric parameters of the sensor and physical properties of the measured medium. A geometric structure of the sensor was proposed. The use of a square cross section permit to eliminate one of the geometric parameters of the sensor, that simplifies the optimization and the analysis of the sensor. Electrical modeling of the interdigital sensor and the medium is carried out with VHDL-AMS software from MENTOR GRAPHICS®. The use of VHDL-AMS language shows the advantage to combine multiphysical domains. The approach can be readily used in current electronic design flow to include distributed physics effects into modelling and simulation process with VHDL-AMS. Simulations results give similar results as physical simulation. However, all the physical properties are not represented, especially at high frequency. The useful properties are correctly simulated. The use of behavioural models in simulation simplify physics and explore interactions between different domains in a reasonable amount of time compared to physics modelling with CoventorWare® software. The simulation results of the impedance obtained with VHDL AMS don’t show any resonance because all the geometric parameters, such as thickness of the medium and the interactions between ions were not include in the model. The experimental results obtained with a sensor, designed by the IJL (Institut Jean Lamour, Nancy University) team, are in agreement with those obtained by simulation. The future goal is to design a specific sensor by optimizing its dimensions for blood measure samples. It will be necessary to design a tank on the active area of the sensor, to avoid measurement errors, and do measurement at higher frequency. The simulation and measured curves present many similarities; the preliminary experiment measures are satisfactory. The next goal is to realise our own sensor by optimizing dimensions to measure blood samples. It will be necessary to design a tank limited to the active area of the sensor, to reduce measurement errors, and allows measurement at higher frequency.

References [1] Faes, T.J., Meij, H.A., de Munck, J.C., Heethaar, R.M.: The electric resistivity of human tissues (100 Hz-10 MHz): a meta-analysis of review studies. Physiol. Meas. 1999 20, R1-10 (1999) [2] Katz, E., Willner, I.: Electroanalysis 15(11), 913–947 (2003) [3] Mukhopadhyay, S.C.: Sensing and Instrumentation for a Low Cost Intelligent Sensing System. In: SICE-ICASE International Joint Conference, pp. 1075–1080 (October 2006)

Physical and Electrical Modeling of Interdigitated Electrode Arrays

189

[4] Mukhopadhyay, S.C., Gooneratne, C.P., Demidenko, S., Sen Gupta, G.: Low cost sensing system for dairy products quality monitoring. In: Proceedings of 2005 International Instrumentation and Measurement Technology Conference, IEEE Catalog Number 05CH37627C, pp. 244–249 (2005) ISBN 0-7803-8880-1 [5] Van Gerwen, P., Laureyn, W., Laureys, W., Huyberechts, G., Op De Beeck, M., Baert, K., Suls, J., Sansen, W., Jacobs, P., Hermans, L., Mertens, R.: Nanoscaled interdigitated electrode arrays for biochemical sensors. Sens. Actuators, B 49, 73–80 (1999) [6] Timms, S., Colquhoun, K.O., Fricker, C.R.: J. Microbiol. Meth. 26, 125 (1996) [7] Geng, P., Zhang, X., Meng, W., Wang, Q., Jin, L., Feng, Z., Wu, Z.: Electrochim. Acta 53, 4663 (2008) [8] Igreja, R., Dias, C.J.: Sensors and Actuators A 112, 291–301 (2004) [9] Borkholder, D.: Based biosensors using microelectrodes. Ph.D. dissertation, Stanford University, Palo Alto (1998) [10] Mamishev, A., Sundara-Rajan, K., Yang, F., Du, Y., Zahn, M.: Interdigital sensors and transducers. Proceedings of the IEEE 92, 808–845 (2004) [11] Sundara-Rajan, K., Byrd II, L., Mamishev, A.V.: Moisture content estimation in paper pulp using fringing field impedance spectroscopy. IEEE Sensors Journal 4, 378–383 (2003) [12] Olthuis, W., Streekstra, W., Bergveld, P.: Sensors and Actuators B 24-25, 252–256 (1995) [13] Jacobs, P., Varlan, A., Sansen, W.: Design optimisation of planar electrolytic conductivity sensors. Medical & Biological Enfineering & Computing, 802–810 (November 1995) [14] Abramowittz, M., Stegun, I.: Handbook of mathematical functions. Dover Publications Inc., New York (1965) [15] Dahmen, E.A.M.F.: Electroanalysis Theory and applications in aqueous and non aqueous media and in automated chemical control. Elsevier, Amesterdam (1986) [16] Bard, A.J., Faulkner, L.R.: Electrochemical methods, fundamentals and applications. John Wiley and Sons, New York (1980) [17] Kovacs, G.T.A.: Introduction to the theory, design, and modeling of thin-film microelectrodes for neural interfaces. In: Stenger, D.A., McKenna, T.M. (eds.) Enabling Technologies for Cultured Neural Networks, pp. 121–165. Academic, London (1994) [18] Bard, A.J., Faulkner, L.R.: Electrochemical Methods. Willey, New York (2001) [19] Jaspard, F., Nadi, M., Rouane, A.: Dielectric properties of blood: an invetigation of haematocrit dependance. Physiological Measurement 24, 134–147 (2003)

Water Quality Assessment through Smart Sensing and Computational Intelligence O. Postolache1,2, P. Silva Girão1, and J.M. Dias Pereira1,2 1

Instituto de Telecomunicações, IST, Av. Rovisco Pais, 1049-001 Lisboa, Portugal Tel.: +351 21 8417974 [email protected] 2 Escola Superior de Tecnologia de Setúbal (LabIM), Instituto Politécnico de Setúbal, 2910-761 Setúbal, Portugal Tel.: 351 265 790000 [email protected] Abstract. Surface water quality monitoring is one of the important activities in the environmental monitoring domain and implies complex measurement activities in order to obtain physical, chemical and biological characteristics of the water. Some of these characteristics are able to be measured in the field but imply the utilization of specific water quality sensors that are used by operators as individually units or, preferably, are part of distributed water quality monitoring networks particularly when monitoring extensive areas. Two concepts are nowadays associated with environment monitoring networks: smart sensing nodes and computational intelligence algorithms. Thus, different smart sensing nodes deliver data that are used by advanced processing units for different purposes, namely: (1) to evaluate the characteristics of water based on measurement channel indirect modeling; (2) to perform the short time and long term forecasting of these characteristics; (3) to detect pollution events and anomalous functioning; (4) to perform data recovering using intelligent algorithms such as neural network and adaptive neuro-fuzzy. The overall operation of the network is optimized if its nodes are provided with functionalities such as auto-identification, networking plug-and-play, auto-calibration, and fault detection. IEEE 1451 family of standards define all aspects necessary not only to transform a sensor into a smart sensor, but also to interface or integrate sensors in networks. In the paragraphs that will follow, we propose the architecture of a smart sensing node suitable for a distributed water quality monitoring network that is IEEE 1451 compatible. The emphasis is placed on the identification of each sensor – which permits individual addressing - and on the algorithms for multivariable characteristics modeling that prove to be very useful for accurate direct digital readout of water quality parameters. Keywords: IEEE 1451, smart sensor, RFID tag, neural network, adaptive neuro-fuzzy, water quality monitoring.

1 Introduction Quality monitoring of surface waters is an important issue to guarantee that they are adequate to the required uses [1][2]. Different measuring solutions have been S.C. Mukhopadhyay et al. (Eds.): New Developments and Appl. in Sen. Tech., LNEE 83, pp. 191–206. springerlink.com © Springer-Verlag Berlin Heidelberg 2011

192

O. Postolache, P.S. Girão, and J.M.D. Pereira

proposed by equipment providers and integrators, and by the R&D community. They include, as a rule, expensive equipments that also include proprietary protocols associated with sensing channel data acquisition, data processing and data logging. YSI, Campbell Scientific [3][4] are examples of manufacturarers that provides equipments in the area of field water quality monitoring which assure the measurement of multiple water parameters (e.g. pH, temperature, conductivity, turbidity, etc.) as so as parameters as depth and water flow. Such equipments present wired communication interfaces such as RS485 or SDI12 [5] that permits to develop water quality monitoring networks. The acquired data is usually sent from individual equipments to a central location using RS232 to wireless modems for relatively short distances such tenth of km (e.g RS232 3G/HSDPA modem from SIMCOM, RS232-RF modems from XStream) or even using RS232 Satellite Modem (e.g SLIN 0011AA - NAL Satellite Modem). In order to assure the auto-identification, auto-calibration and compatibility between different devices of and extended water quality monitoring network the smart sensor network technology is considered. Thus a standard for smart sensors, IEEE1451.X [6] that was developed in the late 1990’s was considered as an interesting solution for the water quality monitoring field. The standard permits an easy development of smart transducer manufacturing and an increasing connectivity of smart sensors to networks. Nowadays, the IEEE1451.X family is a set of protocols for wired interfacing (IEEE1451.0, IEEE1451.1, IEEE1451.2, IEEE1451.3, IEEE1451.4, IEEE1451.6) and wireless interfacing (IEEE1451.5 and IEEE1451.7) of smart sensors suitable for distributed applications including environment quality monitoring. The identification of different smart sensors is performed through the utilization of a memory embedded in the smart sensor and includes information about the transducer included in the so-called Transducer Electronic Data Sheet (TEDS). The direct access to the transducer manufacturing and calibration information is available only for IEEE1451 compatible transducers through TEDS [7], which limits the interest and importance of this standard, since many real systems are characterized by analogue outputs (4-20mA). To overcome this problem, IEEE1451.4 [8] protocol represent an interesting solution. Two different implementations of IEEE1451.4 are considered in the present chapter. The first one uses a 1-wire memory while the second uses the memory of an UHF RFID tag to store the Basic TEDS information [9] providing wired and wireless connection capabilities, respectively. As part of distributed water quality monitoring network, the hardware of a water quality monitoring node (sensing elements, conditioning circuits, acquisition, RFID tags and reader, and communication) must usually be complemented with signal processing blocks to perform different tasks such as data linearization, data compensation, short and long term prediction of pollution events (duration and concentration)[10-12]. Considering the nonlinearity of the single or multivariable characteristics associated with water quality measuring channels (e.g. conductivity measurement channel), intelligent data processing algorithms, such as neural networks and adaptive neuro-fuzzy [13] represents good option to improve the accuracy of the measurements through accurate models of measurement channels.

Water Quality Assessment through Smart Sensing and Computational Intelligence

193

2 Distributed Smart Sensing Network As mentioned before, the smart sensing node as part of water quality dis-tributed monitoring was the result of the work developed by a research team, including the chapter authors, to develop hardware and software for a system designed to perform the measurement of water quality and the sound monitoring on the Sado Estuary region, a well known place as a habitat for bottlenose dolphin families. The nodes (SSNi) constitute a network and are wireless connected to a base station (B_STAT) based on a personal computer (PC) land installed (Fig.1). Additionally a field measurement station (F_STAT) expressed by a laptop PC can be also considered [14].

Fig. 1. The sensing node distribution on the Sado Estuary (SSN1,SSN2,SSN3 – smart sensing nodes, F_STAT – field measurement station embedded on a ship, B_STAT – base station on the land).

Fig. 1 presents a set of three monitoring nodes that are distributed in three important localizations regarding the dolphin groups daily motion trajectories inside the Sado Estuary. The base station that is located on the land receive the data from the smart nodes and performs additional tasks such as intelligent processing, data logging and web based data publishing. 2.1 Smart Sensing Node’s Architecture Considering the Basic TEDS memory support, two architectures for water quality smart sensing nodes were considered and represented in Fig 2. the differences between the presented architectures being related to the existence of 1-wire network (Fig. 2. a) or RFID-system ( Fig.2. b). Important components of the nodes are: the WQ measurement module connected to low-cost 4-20mA transducers, the 1-wire uLAN or the RFID reader, the GPS module, the Wi-Fi interfacing module connected to a high gain Yagi antenna, and a power supply module including a solar panel.

194

O. Postolache, P.S. Girão, and J.M.D. Pereira

PW module

PW module

SW & Wi-Fi module

uLAN/ serial

S/Eth

SW & Wi-Fi module

WQ meas. module

WQ meas. module

RFID Reader

GPS

GPS

ant1

sens sw ant2

MEM1

WQT1

MEM2

tag1

MEMn

WQT2

WQTn

a)

WQT1

tag2

tagn

WQTn

WQT2

b)

Fig. 2. WQ Smart Sensing Node with Wi-Fi communication capabilities a) with TEDS stored in a 1-wire memory, b) with TEDS stored in the RFID tag memory (Tj- WQ sensors; ant1, ant2 – RFID reader antennas; tagj- RFID tags, SW&Wi-Fi module - Ethernet switching and Wi-Fi interfacing, sense SW – RFID correspondence establishment module, PW module – power supply module, GPS- geographic positioning system).

More details regarding the hardware and software components of the distributed water quality monitoring system including smart nodes IEEE1451 compatible are described in the following sub-sections. 2.1.1 WQ Parameters Sensing The water quality measurement module is constituted by a set of water quality transducers (WQT1, WQT2..WQTj) having 4-20 mA analog outputs. The measurement of only four parameters was considered: temperature, pH, conductivity and turbidity, which are measured using WQ101, WQ201, WQ301 and WQ770, from Global Water. A multi-channel current to voltage converter module based on RCV420 from Burr-Brown, assures a 0-5V output voltage for 4-20mA input current. The voltage signals are applied to the analog channels (ACH0-ACH3) of a data acquisition module (Ipsil IPu8930) that includes a 16-bit analog-to-voltage converter (ADC) and an Ethernet interface. The acquisition module is connected to the SW&Wi-Fi module that permits the wireless communication with the base PC. The compatibility of the water quality transducers with IEEE1451 standard is assured adding a 1-wire memory [15] to each transducer or an UHF RFID [16]. Based on the information received from sensors’ memories, the selection of previously

Water Quality Assessment through Smart Sensing and Computational Intelligence

195

calculated internal parameters of intelligent processing blocks are used to perform the voltage-to-water quality parameter (V-wqp) conversion. 2.1.2 IEEE1451.4 Implementations Identification represents an important feature of the smart sensors. Embedding the RFID technology at the level of the WQ smart sensing node, the transducer tracking can be done during calibration, testing in laboratory but also during on-field operation. As was mentioned above two kinds of memories are attached to the water quality transducers: 1-wire memories and UHF RFID tag memory. Both memory supports are able to store specific information denominated Basic TEDS. Table 1 identifies the Basic TEDS fields that are mandatory to comply with IEEE1451.4 and that are stored in the 1-wire or RFID tag’s memory. Table 1. Implemented Basic TEDS fields.

Components Manufacturer ID Model Number Version Letter Version Number Serial Number

Number of bits 14 15 5 6 24

Allowable range 17-16381 0-32767 A-Z 0-63 0-16777215

The Basic TEDS .Dot4 field is filed in with the values of each transducer (smart sensor) according to the specification given by the manufacturer. Two solution of TEDS’ implementation were considered: Basic TEDS based on 1-wire memory solution The WQ transducers are connected to the WQ measurement module that includes and Ethernet DAQ unit expressed by uP8930 from Ipsil that has a 16-bit ADC. The DAQ Ethernet port is connected to the sensing node Switch and Wireless Communication block (SW&Wi-Fi module). The transducer identification uses the Basic TEDS information codified according with IEEE1451.4 standard. Thus, the information for each transducer of the WQTj, as indicated in the Table 1, is stored in a 1-wire memory (DS2433) associated to each transducer and accessed through a 1-wire MicroLAN Coupler (DS2409) (Fig.3). A 1-wire to RS232 protocol converter is used to assure the connection between the 1-wire MicroLAN and the RS232 port of the SB72-EX (high performance Serial-to-Ethernet device) that is followed by an Ethernet/Wi-Fi bridge (D-LINK DWL-G820). The data from the transducers (WQT1, WQR2 …WQTj) memories (M1, M2 …Mj), that store the TEDS information (see Table 1), are individually read by the land unit (expressed by a PC) using a set of adapters including the MicroLAN coupler, RS232-to-Ethernet bridge, Ethernet-WiFi bridge. The land unit software uses the received data to extract extended information that is stored in water quality transducer database that is denominated Virtual TEDS [17].

196

O. Postolache, P.S. Girão, and J.M.D. Pereira

WQT1

WQTj

M1 TEDS

Mj TEDS

1-wire

1-wire

4-20mA

4-20mA

1-wire

1-wire MicroLAN coupler

MUX uC

1wire/ RS232

ADC Ethernet

SB-72EX (PORT1)

Ethernet Switch

Fig. 3. The block diagram of Basic TEDS based 1-wire memory implementation

Basic TEDS based on UHF-RFID tag memory solution The architecture of the WQ measurement module is similar to the architecture presented on the Basic TEDS based on 1-wire memory solution. The difference is related to the existence of UHF RFID reader that replaces the elements such as 1-wire MicroLAN and the existence of RFID tags that replaces 1-wire memories. The RFID tags are attached as labels to the WQTj transducers (Fig. 4). WQ transducer Ti

RFID tag

4-20mA

tagj DAQ ACHi

Fig. 4. IEEE1451.4 standard implementation based on RFID tags and WQ transducers characterized by 4-20mA current output.

Water Quality Assessment through Smart Sensing and Computational Intelligence

197

As it can be observed in Fig. 4, the low cost passive transponders (tagj, j=1…n, where tagj are expressed by ALL-9540-02 World Tag 860-960MHz in the present case) are attached to the water quality transducers WQTj’, as labels attached to the transducer cables the compatibility with IEEE1451.4 standard is assured. The Basic TEDS (Transducer Electronic Data Sheet) is stored in the tag’s memory. The main characteristics of the used tags are: EPC class1, RF communication protocol ISO/IEC18000-6 CEPC Class 1 Gen 2 (generation 2 – read and write many times), EPC memory size: 96 bits, access control: 32 bits, kill code: 32 bits. The typical distance between the reader and tags (read range) is of about 4m, which is enough for the present application. The identification process of the sensor with RFID label begins when the reader (ALR-8800 from Alien) is switched on: it starts emitting a signal in the selected frequency band (860MHz-960MHz). The tags reached by the reader’s field will “wake up” (supplied by the field itself) [18]. In order to discriminate between the received information from the multiple tags, an anti-collision algorithm is implemented at the reader level [19]. Considering the memory size of the used tags (96 bits for ALL-9540), the TEDS information storage is restricted to Basic TEDS. These informations can be written using the capabilities of ALR-8800 RFID reader that is designed to program EPC Class 1 Generation 2 tags [20]. Two external circular polarized antennas, ant1 and ant2 (ALR-8610-AC) operating at 850-875MHz and with 6dBi gain, are used to read the tag fixed on the transducers or to write the memory of the tags, including new elements, during operation. The ant 1 works as the transmitting antenna while ant 2 works as the receiving antenna (Fig.2.b). An important advantage of the UHF RFID tags is that they are easily attached to the transducer cable and they can be read and written wirelessly. However, a drawback is the impossibility to direct perform the identification tag-measuring channel that is done using an additional sense sw module (see Fig. 2). The module works under reader DIO port control, which switches on the measuring channel at the same time the procedure for the identification of the tag-measuring channel is run, thus assuring that the signal acquired on a specified channel (e.g. ACH0) corresponds to the

Fig. 5. System Efficiency, SE, versus antenna – tag distance for different attenuations of the emitted power of transmission RF signal

198

O. Postolache, P.S. Girão, and J.M.D. Pereira

identified transducer (e.g tag0). The transducers are connected one by one to the ADC of the uP8930, and at the same time, the corresponding tag is read. A verification procedure implemented by the software installed in the base station detects the existence of a new transducers connected to the WQ measurement module and simultaneously is obtained the information for the attached tag. The correspondence transducer-tag is stored in a file. Based on TEDS stored in the tag’s memory (e.g. TEDS1=1C02 0042 06A5 9000 is connected to ACH0), the processing parameters from intelligent virtual extended TEDS are extracted. Related to RF identification, a practical approach concerning the evolution of system efficiency, SE, defined as the relation between the number of detected tags and total number of tags, for different distances between ant1 and ant2 and the UHF tags (tagj) was carried out. Some of the obtained results are presented in Fig. 5.

3 Intelligent Modelling of Measuring Channels Direct and inverse modeling of sensors characteristics using intelligent algorithms such as artificial neural network and neuro-fuzzy systems are reported in literature [11-13][21][22] and represents one of the field of interest for the chapter authors . The purpose of the direct modeling is to obtain a neural network or an adaptive neurofuzzy designed model of the measuring channels in such way that the outputs of the considered channel and the model match closely. The direct model corresponds to the calibration curve model obtained for a given measuring channel, while inverse model uses the measuring channel output data (acquired voltages) to extract the information related the measured physical value (e.g. water pH). In the particular case of water quality monitoring application, the data received by the base station from the smart sensor node trough Wi-Fi communication are processed using inverse model coefficient that are stored in a database named Intelligent Virtual TEDS (IV-TEDS). The model selection (coefficient selection) is performed using the Basic TEDS information that is stored in the memory associated to each measuring channel (1-wire memory, passive RFID tag memory). Two intelligent algorithms were considered in order to perform a comparison between the modeling accuracy and model implementation complexity for well known multilayer perceptron artificial neural network (MLP-NN) [23][24] and adaptive neuro-fuzzy (ANFIS) [25][26] algorithms. Accurate models conduct to accurate digital reading of the water’s physical characteristics (e.g. temperature, conductivity, pH, turbidity). Thus using as input values the normalized acquired voltages from the corresponding measuring channel and the model coefficients (MLP-NN or ANFIS coefficients) stored on the IV-TEDS database the normalized values of water quality parameters are calculated. In the particular case of the multilayer perceptron neural network (MLP-NN) represented in Fig. 6, the specific weights and biases (model coefficients) are obtained during the training phase based on a training algorithm such as the Backpropagation or the Levenverg Marquardt [23]. Training uses a training set constituted by known values of the WQ parameters (normalized values) and the common influence factors (e.g. temperature) and the corresponding voltage values obtained at the output of the measuring channels. Fig. 6 depicts the overall processing scheme that includes the neural network processing block associated with WQ measuring channel inverse

Water Quality Assessment through Smart Sensing and Computational Intelligence

199

modeling with external common factor compensation [27]. Fig. 7 shows the diagram including a set of six layers associated to an ANFIS processing structure, WQ measurement channel inverse model, that includes the fuzzification, the implication and (if needed) defuzzification stages.

Vm

Vi N ViN

VmN OL

IL

HL

wqpNNN N-1 wqpNNN

Fig. 6. Water quality parameter neural processing scheme (Vm – main input voltage, Vi- influence factor voltage, IL – neural network input layer, HL- neural network hidden layer, OLneural network output layer, N, N-1 norm. and denormalization blocks).

IL

Vm

mf-IL RL mf-OL wSO

O

weighted sum

Vi wqc

normalization factor

Fig. 7. A Water quality parameter ANFIS inverse model designed to extract the WQ values with external influences (e.g. temperature influence) correction ( IL-input layer, mf-ILmembership function input layer, RL-rule layer, mf-OL – membership function output layer, wSO-weighted sum output, O-output, wqc-compensated value of measured water quality parameter).

200

O. Postolache, P.S. Girão, and J.M.D. Pereira

Regarding the inverse modeling can be mentioned that during the measurement, the Vm and Vi (e.g.Vm=VC, Vi= VT ) acquired voltages are converted into values of water quality parameters (e.g. conductivity C). Referring to the neural network architecture, two single input-single output neural networks, one for VT →T and another for VTU→TU conversion, and two dual inputsingle output neural networks, one for (VpH, VT)→ temperature compensated pH and another (VC, VT)→ temperature compensated C conversion were designed and implemented. The networks have one single hidden layer (HL), the training algorithm in all four cases was Levenberg Marquardt and sum-square errors (SSE) of 1E-4 are the stop condition. The MLP-NNs training and test are performed using a data set of voltage values delivered by the transducers during the calibration phase. The calibration solutions used were formazin {10, 20, 40, 100, 200, 400, 800, 1000}[NTU] for TU transducer, buffer solutions {4, 5, 6, 7, 8, 9} for pH transducer, and {84, 447, 1500, 2764}[uS/cm] for conductivity. Calibrations were performed in a laboratory for different temperatures, TЄ[5; 30]°C. A study concerning the number of neurons in the hidden layer and neural network inverse model accuracy can be done. Good results are normally obtained for tansignoid [28] neurons in the hidden layer expressed by the following numbers: nhidden|MLP-NNT =6, nhidden|MLP-NNpH =7, nhidden|MLP-NNC =12, nhidden|MLP-NNTU =5. After MLP-NNs design, the calculated neuron weights and biases are stored in the intelligent virtual TEDS whose organization is presented in Table 2. On-line neural conversion and compensation blocks that use intelligent virtual TEDS information are based on the following relation:

wqp = W2 ⋅ tanh(W1 ⋅ V + B1) + B2

(1)

where W1 and W2 represent the weights matrices, B1 and B2 the biases matrices of the neural network designed to obtain the WQ parameter value (wqp) using the acquired normalized voltages VTU, VC, VpH and VT included in vector V. Referring to the ANFIS model architecture, the first layer of neurons is denominated input layer (IL) and receives the input from the measuring channels (acquired voltage values). The second layer, membership function input layer (mf-IL), calculates the fuzzy membership degree to which the input voltage values (e.g. Vm=VpH) are mapped from the input voltage intervals to the unit interval through a membership function, mf. This mf can be defined in linguistic terms highlighting the advantage in this sense of the ANFIS models representation. For example, the pH of water under test is not defined in a crispy sense as acid or neutral but rather as 0.5 acid and 0.5 neutral. Each node (neuron in the neural network sense) of the mf layer includes an mf for one of the inputs (e.g Vm=VpH and Vi=VT). Fig. 7 shows the ANFIS architecture with two inputs and four corresponding membership functions. The mf functions used in the present work are of the triangular, trapezoidal and Gaussian type, the last one being defined by:

(

µ Aij x si ,cij , σij

)

⎛ ⎛ x − c ⎞2 ⎞ j ij = exp ⎜ − ⎜ ⎟ ⎟ ⎜⎜ ⎜ 2 ⋅ σ ⎟ ⎟⎟ ij ⎠ ⎝ ⎝ ⎠

(2)

Water Quality Assessment through Smart Sensing and Computational Intelligence

201

where the cij, σij internal parameters are adjusted during the adaptive neuro-fuzzy network training phase. The third layer is the rule layer (RL) and represents associations between the input and the output variables. The number of rules (n - number of rules layer neurons) is normally included in the 4 to 100 interval for the models associated with WQ measuring channels taking into account their non-linearity. The rules syntax is characterized by the following structure: If xS1 is A1 and xS2 is B1, then f1 = p1 ⋅ xS1 + q1 ⋅ xS2 + r1 If xS1 is A2 and xS2 isB2, then f2 = p2 ⋅ xS1 + q2 ⋅ xS2 + r2 … If xS1 is An and xS2 is Bn , then fn = pn ⋅ xS1 + qn ⋅ xS2 + rn

(3)

which corresponds to a Sugeno fuzzy inference system [29]. In (3) the Ai and Bi are linguistic terms. For example, in the case of the ANFIS model for pH measurement channel, A1 represents “high acid”, B1 “low temperature”, xS1 and xS2 are model input values, and p1…pn and q1…qn, r1…rn are consequent parameters. Table 2. TEDS for a two input – one output neural network.

ANN (TEDS) Fields Name Number of ANN inputs Number of ANN output Number of ANN layers

Value

Comments

ninput

The input variables are the sensors’ output voltages

noutput

The output variable is the temperature corrected value Hy

N

The number of ANN layers includes the input, the hidden and the output layers

ANN training stop condition

α

The ANN training stop condition is expressed by a sum square error value (SSE)

ANN hidden layer neurons weights matrix

W1

The hidden layer neurons weights are used to calculate the hidden neurons output values

B1

The hidden layer neurons biases are used to calculate the hidden neurons output values

W1

The output layer weights are used to calculate the ANN output

B2

The output layer biases are used to calculate the ANN output

ANN hidden layer neurons biases vector ANN output layer weights transposed vector ANN output layer biases

202

O. Postolache, P.S. Girão, and J.M.D. Pereira

The fourth layer calculates the degrees to which output membership functions, oSic are matched by input data:

oSic = w i ⋅ fi

(4)

where wi is the firing strength of rule i. Layer five includes summation of rule outputs and firing strength, the former sum being divided by the latter on the sixth layer to yield the overall output of the system. In the particular case of single input – single output ANFIS model for a WQ measuring channel, the output corresponds to a water quality parameter such as temperature or turbidity while for the measuring channels such as pH, electrical conductivity, or dissolved oxygen the output corresponds to the respective temperature corrected values. The ANFIS model design uses a backpropagation hybrid algorithm [30] as training algorithm. We designed and implemented the ANFIS models using the MATLAB fuzzy toolbox. Referring to the models accuracy, the performance criteria based on the maximum absolute error (emax) and root mean square error (rms) were evaluated. Table 3 highlights the numerical values for pH measuring channel. Table 3. Accuracy of FuNNpH versus FuNN’s architecture and training type – FuNN testing phase.

No. mf 2

4

10

mf type Triangular trapezoidal Gauss Triangular trapezoidal Gauss Triangular trapezoidal Gauss

Bkp Hyb No. of epoch 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2

Bkp.

Hyb. emax

0.687 0.681 0.688 0.683 0.686 0.685 0.682 0.681 0.684

0.0031 0.0258 0.0071 8.861E-5 0.0064 0.0016 0.0001 0.0001 0.0001

As shown in Table 3, the number of training epoch is independent on the training algorithm (Backpropagation (Bkp) or Hybrid (Hyb)) while the accuracy depends on the number and type of adaptive neuro-fuzzy membership functions. For the presented case the best results are obtained for four triangular membership functions. Comparing the MLP-NN and ANFIS as intelligent algorithm used to obtain the direct and inverse modeling of water quality both algorithm were proven to be universal approximators with good results on direct digital readout of WQ values starting from acquired voltages associated with WQ measurement channels. Thus in particular measurement conditions and for particular measurement ranges both applied methods are characterized by errors lowers than 1%. Thus, a comparison between the obtained results is presented in Table 4.

Water Quality Assessment through Smart Sensing and Computational Intelligence

203

Table 4. The evolution of relative errors associated with water quality measurement using the designed intelligent algorithms: multilayer perceptron neural network, adaptive neuro-fuzzy inference system

Intell. algorithm\errors

εT(%)

εpH(%)

εC(%)

εTU(%)

MLP-NN

0.22

0.26

0.23

0.18

ANFIS

0.31

0.59

0.42

0.25

The results presented in Table 5 corresponds to the temperature (T), pH, conductivity (C) and turbidity (TU) variation ranges that corresponds to the particular case of Sado River Estuary (ranges (5°C ≡

t max

(2)

0

∑ N

=

H int k V k ( x , y , θ )

∑V

k =0 N

R

k

( x , y ,θ )

(3)

k =0

Here, Hint stand for the internal fields to give rises of the jumps. The physical mechanism was interpreted in the ref. [3]. For an example , the analytical results of the strain anisotropy is demonstrated in an iron-based sample with the tensile stresses of 500MPa in Fig. 6. Here, Fig.6(a) shows the at a heavily strained position, Fig.6(b), a slightly strained position, respectively. Apparently, the heavily strained position shows the larger anisotropy than that at the slightly. These results are reflected by the strength of the average pinning forces at each position.

Fig. 6. The expectations () of magnetic Barkhausen noises observed at differently strained positions in a sample

2.3 Leakage Flux The observations of the leakage flux distribution was performed at the sample surfaces by using a fluxgate sensor which does not emit flux during the flux-gate switching actions [16], and with the spatial resolution of 50µm with a condition of the lift-off distance of 50µm from the sample surface. The leakage flux distribution of the normal components ΦL(x,y) is displayed in Fig. 7 at positions (x, y) and its gradient dΦL(x,y)/dx along with x direction, in which the magnetization was performed before measurements. The leakage flux generates around at positions where the permeability become smaller than those at surrounding positions. This mechanism is quite natural to consider the flux accommodation along the sample. Namely, the gradient of flux (-dΦ L (x,y)/dx )x=x0 becomes larger at x=x0 due to the total flux µ(x,y)S(y) become smaller at the position with the cross section S(y). Therefore, the gradient of leakage flux along the applied strain at the sample surface depicts the internal flux fluctuations or in another word, the permeability distribution along the strained direction. Note here, that even a sample without inhomogeneity of

298

K. Yamada, J. Luo, and M. Enokizono

material constants and with the same cross-section along x direction shows leakage fluxes. However, the leakage flux shows an only monotonic increase or decrease along the sample direction x.

Fig. 7. Sample deformation, the leakage flux distribution and the derivatives in an A533B sample.

Fig. 7 shows a leakage flux distribution and the sample deformations in a sample with the maximum thickness deformation about of -0.8%. The leakage flux occurs not only by the thickness deformation but also by the permeability fluctuation as shown in the top figure. The second top figure shows 2-dimentional distribution of the leakage flux and ΦL(x,0). The bottom figure shows the distribution of -dΦL(x,y)/dx and -dΦL(x,0)/dx along x direction in B&W. These distributions show most plausible to understand the physical mechanism as explained just before. However, we must be careful to examine whether the geometry effects are included or to examine whether the lift-off distance of the Hall element is constant during the measurement over the scanning. 2.4 AC Conductivity and Magnetic Susceptibility We developed a sensor using small transformers with open poles, attached at materials under evaluation [17]. Fig. 8 illustrates the schematic structure of the probes. The

Nondestructive Evaluations of Iron-based Materials by Using AC and DC

299

Fig. 8. The experimental set-up of magnetic ac parameters using two transformer probes (Tr1,Tr2).

output voltages (V0j=1,2) of the two probes are connected in reverse to obtain the difference signal. Here, we performed the basic experiment to characterize a probe output (V0) as a function of frequency f. For this purpose, we define the ac magnetic circuit like that in the ac circuit theory as Φ (ω ) =

NI ( ω ) ∑ R j (ω ) j

,

(4)

where, N stands for the primary coil turns, I( ω), the current and ω(=2πf), angular frequency, respectively. The magnetic resistance Rj(ω) is defined as frequency dependent and complex variable, and the suffix (j) runs over several circuit elements. The current I( ω) is also dependent on frequency as I0 exp(-iωt),( i ≡ − 1 ) . Therefore, we obtain the output voltage V0 as V o (ω ) = i

nN ω I ( ω ) ∑ R j (ω ) j

.

(5)

Here, n stands for the winding turns in the secondary coil. Now, the transfer function C( ω) is defined by C (ω ) ≡ K =

V 0 (ω ) ω I (ω )

1 ∑ R j (ω ) j

.

(6)

Here, K stands for a constant (=-i/Nn) It is easily understood that C(2πf) reflects the flux conductive properties as that in the dc magnetic circuit. Fig. 9 shows the drastic change of the magnetic transfer function C(f), between the probe without air-gap immersed with the original transformer core, and with the airgap. Namely the magnetic transfer function without air-gap increases by 50 times larger than that with air-gap of 0.2 mm. By using the probe in this study, we can obtain the stable value of magnetic susceptibility by some calibration in the frequency range less than 1 kHz as shown in Fig.10. In the lower figure of Fig.10, the eddy

300

K. Yamada, J. Luo, and M. Enokizono

Fig. 9. The experimental characteristics of a transformer type probe with completely closed magnetic circuit, immersed with the original transformer core (open circle), and a probe with open magnetic poles attaching at the same core surface (closed circle).

current losses in Cupper plates are enhanced in a range larger than 1kHz. Note here that C(f) is normalized by that with air-gap and attached no sample. Hence, the value of C(f) larger than unity means that the magnetic field inside the probe increased than that in air. In other words, we can determine the magnetic susceptibity and conductivity precisely in comparison with the difference material with calibrated permeability. In this way, we attained the resolution of 0.2% for conductance changes in conductive materials and the susceptibility changes of 0.1 % for magnetic materials, respectively.

Fig. 10. The permeance of several magnetic materials (upper figure) as a dust core (5mmt), a sheet for shield(10µmt) and a soft iron plate(1mmt) and cupper plates with different thicknesses (lower figure).

Nondestructive Evaluations of Iron-based Materials by Using AC and DC

301

In addition to these results, the phase shifts were also observed by excitation current I and the difference output VD in Fig.8 reflects the preceding or retarding phase caused by the imaginary part of the magnetic resistances.

3 Discussions We presented many NDE tools in this study together with the additional experiments on the ac magnetic diagnosis tools. In this discussion, we examine the validity of ac magnetic circuit and the experimental results. Namely, here, we suppose the complex variables of the magnetic resistance R(ω) as Z( ω) in the conventional circuit theory. Further, 3 components are proposed in each probe of Tr1 and Tr2 as shown in Fig. 11.

Fig. 11. Equivalent magnetic components composed of RLC (the leakage flux), RLG (air gap at the magnetic poles) and RLG (the material).

Namely as shown in Fig. 11, the component RLC originated by the leakage flux around magnetic core positions expanded from the winding coil and RLG , by the airgap between the magnetic pole and sample surface, respectively. The dependence of gap resistance on the frequency could be neglected as pure resistance (real number). Now, in general, the difference signal of Tr2 and Tr1 (=ΔC21(f)) is exactly expressed by the difference outputs of Tr2 and Tr1 (=ΔC21(f)) as Δ C 21 ( f ) ≈

1 RM 2 [ RL G RL G + R M

− 2

RM 1 ] . RL G + R M 1

(7)

Note here that RLG1=RLG2=RLG due to the exactly the same current flows in the primary coils of the two probes with the same structure. Now, ΔC21(f) behaves always positive due to the increasing function of RM in the case of real numbers and RM2>RM1. In this case of RM , the inequality holds in general as RL C >> RL G >> R M .

(8)

The difference permeance ΔC21(f) (=1/RM1 - 1/RM2) becomes simple as Δ C 21 ( f ) ≈

1 [RM 2 RL G

2

− R M 1 ].

(9)

Here RLC is common for the two probes because the excitation current flows in common with the two probes, and so does RLG. In general case of complex variable of RM , ΔC21(f) is exactly expressed by Δ C 21 ( f ) ≈

1 RM 2 [ RL 2 RL 2 + R M

− 2

RM 1 ] . RL 2 + R M 1

(10)

302

K. Yamada, J. Luo, and M. Enokizono

Here, ΔC(f) is always positive in the case of real numbers and RM2>RM1 because Eq. (10) is increasing function with increasing RMj. Now, the frequency dispersions in the material is variously defined by the relaxation approximations of eddy current or of the magnetic wall movements etc as R M ( f ) ≈ 2π f

E

R M 0 exp( − 2 π f τ M ) .

(11)

Here RM0 stands for a complex number depending on the magnetic susceptibility and the conductance of the material, E, the power number (E=1-2: for eddy current loss). The relaxation time constant might be determined well by the skin depth (d) and the sample thickness (D). Further, in the frequency range of relaxations of any kinds, τM could be complex number to give rise the phase shifts between I(ω) and VD(ω). Fig.12 shows the experimental results of the magnetic material as a sample, which shows the phase shift between I( ω) and VD(ω) as expected before in the case without rectifying diodes in Fig.8.

Fig. 12. The observed phase shifts observed by the device.

Here, we found the almost 170 degree phase shift between that of iron and that of cupper in the difference signal (VD). The phase shift will be correctly determined by using a Lock-in amplifier.

4 Conclusions We would like to emphasize here that the analytical results of MR are almost independent of the sample thicknesses in MR ratio due to the division by the resistance in zero-field. This is an out-standing feature in NDE, with the demerit of the necessary high precision. In MBN observations at the surface, the amplitudes of VN reflect the true inhomogeneity inside of the material by using the probe in this experiment. Therefore, the tools developed here are available in the factory except noise problems. The most important feature in ac magnetic probe is the non-contacting method to the sample surface. Because of the atomic power station, all the material should not be injured by the diagnosis. The method of MBN observation, the reference is always

Nondestructive Evaluations of Iron-based Materials by Using AC and DC

303

necessary to compare because the value of expectations is relatively determined. Therefore as a conclusion, NDE might be performed by cross-checks and compare with the material without degradation.

References [1] [2] [3] [4]

[5] [6]

[7]

[8]

[9]

[10]

[11]

[12] [13] [14]

[15] [16] [17]

Jiles, D.C.: Review of magnetic methods for NDE. NDT Int. 21(5), 311–319 (1988) Kronmuler, H.: Canadian J. Phys. 45 (1967) Shoji, S.: Doctor Thesis, Saitama Univ. (March 1999) (in Japanese) Yamada, K., Shoji, S., Yamaguchi, K., Tanaka, Y.: Fractal Dimension of Magnetic Noises: A Diagnosis Tool for Nondestructive Testing. In: Kose, Sievert, J. (eds.). Studies in Applied Electromagnetic and Mechanics, vol. (13), pp. 153–156. IOS Press, Tokyo (1998) Yamada, K., Saitoh, T.: Observation of Barkhausen effect in ferromagnetic amorphous ribbon by sensitive pulsed magnetometer. J. Mag. Magn. Mater. 104, 341–343 (1991) Yamada, K., Shoji, S., Tanaka, Y., Uno, Y., Takeda, H., Uesaka, M., Miya, K.: Nondestructive Evaluation of Iron-based Material by Magnetic Sensor. In: Proc. of the 2nd Int. Workshop on Advanced Mechatronics (IWAM 1997), Nagasaki, pp. 114–119 (1997) Yamada, K., Shoji, S., Tanaka, Y., Uno, Y., Takeda, H., Toyooka, S., Spurapedi, Isobe, Y., Ara, K., Uesaka, M., Miya, K.: Nondestructive Cross Evaluation of Iron-based Material by Magnetic Sensors and by Laser Speckle Interferometry. J. Magn. Soc. Jpn. 23, 718–720 (1999) Hagiwara, N., Fukuda, N., Masuda, T., Yamada, K.: Nondestructive Evaluation of Plastic Strain in Pipeline Using Barkhausen Noises. In: Proc. Workshop on Magnetism and Lattice Imperfections, Hanamaki, Iwate, pp. 97–100 (April 2000) Uesaka, M., Sukegawa, T., Miya, K., Takahashi, S., Echigoya, J., Yamada, K., Kasai, N., Morishita, K., Ara, K., Ebine, N., Isobe, Y.: Round-robin test work for magnetic nondestructive evaluation of structural materials in nuclear power plants. In: Proc. Workshop on Magnetism and Lattice Imperfections, Hanamaki, Iwate, pp. 59–65 (April 2000) Yamada, K., Yamaguchi, K., Toyooka, S., Isobe, Y.: Magnetic and Optical Nondestructive Evaluation for Iron-based Materials. In: Green, R.E., et al. (eds.) Nondestructive Characterization of Materials X, pp. 333–340. Elsevier Publ., Karuizawa (2000) Yamada, K., Yamaguchi, K., Takeda, H., Tonooka, S., Masuda, T., Hagiwara, N.: Nondestructive cross evaluations iron-based material by optical and magnetic diagnosis tools. Invited paper for 3rd Int. Workshop on Advanced Mechatronic (IWAM 1999), Chunchon, Korea, pp. XXV–XXX (December 1999) Kasuya, T.: Prog. Theo. Phys. 16, 45–57 (1956) Yoshida, K.: Phys. Rev. 106, 893–898 (1957) Yamaguchi, K., Yamada, K., Isobe, Y.: Monte-Carlo Simulation of Magnetization Processes including Lattice Imperfections. In: Proc. Int. Meeting on the relationship between Lattice Imperfection and Magnetism, Hanamaki, Iwate, pp. 69–71 (April 2000) Yamaguchi, K., Yamada, K.: Simulation of Spin system for Nondestructive Evaluations of Iron-based Materials to be published by Proc. EMMA 2000, Kiev (May 2000) Liu, B.: Doctor Thesis, Saitama University (March 2006) Yamada, K., et al.: To be presented in Magda Conf. in Hokkaido, Hokkaido (November 22-23, 2010)

STACK: Sparse Timing of Algorithms Using Computational Knowledge Vasanth Iyer1 , S. Sitharama Iyengar2, Garmiela Rama Murthy1 , Kannan Srinathan1 , Mandalika B. Srinivas3 , and Regeti Govindarajulu1 1

International Institute of Information Technology, Hyderabad, India-500 032 [email protected], [email protected], [email protected], [email protected] 2 Louisiana State University, Baton Rouge, LA 70803, USA [email protected] 3 Birla Institute of Technology & Science, Hyderabad Campus, Hyderabad-500078, India [email protected]

Abstract. Research in computational aspects and algorithm optimizations help design tools to acceleration the execution of algorithms. Cost and availability of FPGA design boards have driven number of computations per second close to the general-purpose model of CPUs. In this chapter, we study the effects of algorithms with the knowledge of the underlying computing model for getting consistent and coherent view of the sensed data. The computing model uses uniprocessor, multiprocessor and acceleration using pipeline and data-path forwarding with Byzantine fault-tolerance. The pre-processing approach of the modified algorithm for sparse sensing gives better consistency and the application based calibration allowing coherent view of the data and at the same time reduces the total power consumption. This is analogous to the needle in a hay stack. The STACK implementation runs 4 times faster than the normal program based optimizations for static and dynamic scheduling. Keywords: Sensor-centric data fusion; Algorithm design; Compressed Sensing (DCS); Sensor Fusion; Pre- and Post processing of Sensors; Cross-layer Protocols; Processor Computational Models.

1

Introduction

We study the optimization of data-centric algorithms, which need to process data locally and aggregate and optimize in a distributed way. Most of the time these families of algorithms are compared measure based on computational complexities in time and space, which allow studying its scalability. The lifetime of sensor network [1,8,13] is also an important measure to benchmark the performance of these algorithms. When using this measure as they are prone to error than a general deployment error correction needs to be included in terms of a Byzantine agreement algorithm. S.C. Mukhopadhyay et al. (Eds.): New Developments and Appl. in Sen. Tech., LNEE 83, pp. 305–320. Springer-Verlag Berlin Heidelberg 2011 springerlink.com

306

V. Iyer et al.

The computational model for the programs shown in Figure 1, Algorithm 1 and 2 assume sequential execution. The values of Flag and turn in algorithm decide which process enters the critical section allowing shared data modeling [3,4] for consistency [3] and atomic operations. Message passing model is used in a distributed power-aware topology; this is shown in Algorithm 3. The data coherency [3] cannot be achieved due to – unknown number of processors – independent inputs at different locations – several programs executing at once, staring at different times, and operating at different speeds – processor nondeterminism – uncertain message delivery times – unknown message ordering – processor and communication failures The value of UID is pre assigned and the value of the leader is based on having a reliable message passing FIFO available at each processing node element (PE). Computational Model for Consistency: – – – – –

Pre-processing - reduce computation per application Dynamic Range - register bit usage Pipe-line (1-CPI) fine grain instruction level parallelization Double buffer for datapath support Dynamic power dissipation

The computational model is affected by software compiler optimizations, having sequential consistency [3], this is called program order. For embedded systems, most of the optimization is targeted for space as the target has resource limitations. In- order sequential consistency is performed by the compiler by using a window and appropriate scheduling of instructions to enhance performance. The pipe-line optimizations are done dynamically (out-of-order) [3] execution, which furthers performance which is not possible in the earlier case such as memory access and register allocations. Due to this data hazards are possible due to data coherency needs. Processor and hardware support needs to support the high-level construct of the language in context with sequential consistency [1]. The non-blocking memory operations can be implemented using specific address ranges and explicit addresses can be assigned for shared memory operations. The processor can distinguish it by physical and virtual address space during dynamic execution. Unused op-code fields can be used to implement the latter case. Even in applications where reasonable amounts of parallelism are available, significant serial, or nearly serial, code segments may occur. A problem where 90% of the program can obtain a speed-up of 10, and the remaining 10% has no parallelism, will actually run only about 5 times faster. In the hardware implementation, the same clock runs the memory that allows access to memory complete in one clock cycle. We show that the STACK model of execution gives superior performance than the sequential consistency model for data-path based algorithms, which effect sampling rate and power.

STACK: Sparse Timing of Algorithms Using Computational Knowledge

1.1

307

Computational Models

The modern processors use efficient Instruction Set Architecture (ISA) such as RISC processors. The performance measures defined as below P erf = 1.2

clock speed instruction count ∗ cycles per instruction

(1)

Van Numen

RISC [12] machines attempt to maximize performance by producing improvements in clock speed (factors of 2-5, typically) and major improvements in the cycles per instruction (factors of 5 to 10). They allow a slight increase in the instruction count (less than a factor of 2). 1.3

Distributed Architecture

A synchronous network system consists of a collection of computing elements located at the nodes of a directed network graph. We refer to these computing elements as processing elements (PES), which suggests that they are pieces of sensor hardware. In order to define synchronous network system start with a directed graph G = (V, E). We use the letter n to denote | V |, the number of nodes in the network digraph. 1.4

FPGA

The hardware implementation uses a data driven pipelining, the clock rate can be calculate as CLK = Active event Q + Longest delay + Setup time + ClkSkew

(2)

1 Capacitive Load × V oltage2 × F req Switch (3) 2 The clock cycles shared between execution units and the memory unit, which allows transferring data from memory in one clock cycle. The actual implementation, which needs to calculate the longest delay while completing a calculation determines the clock period, it tends to be longer to accommodate all the variations. P ower =

1.5

Computational IP Cores

FPGA design framework [7] allows using many available algorithms in a form of a library protected by intellectual rights. Pre-processing of data needs many steps which can use these algorithm libraries, such as Scaling [7], Interpolation [7] and Mixing [7]for video oriented data streams.

308

V. Iyer et al.

Algorithm 1. Process P1 1: statesi : 2: flag[0] = 0; 3: flag[1] = 0; 4: turn; 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18:

Algorithm 2. Process P2 1: statesi : 2: flag[0] = 0; 3: flag[1] = 0; 4: turn;

msgsi : P0: flag[0] = 1; P1: flag[1] = 1; turn = 1;

msgsi : P1: flag[0] = 1; P1: flag[1] = 1; turn = 0;

== 1) do

while (flag[0] == 1 && turn == 1) do

5: 6: 7: 8: while (flag[1] == 1 && turn 9:

// busy wait

end while // critical section ... // end of critical section flag[0] = 0;

10: 11: 12: 13: 14: 15: 16: 17: 18:

// busy wait

end while // critical section ... // end of critical section flag[1] = 0;

Algorithm 3. Process P1,P2,...Pn 1: Each process begins by sending its UID to its clockwise neighbor. 2: Each process checks its UID (u) against the one it just received (v), 3: if v > u then 4: the process sends v on to the next process 5: end if 6: if v = u then 7: the process is chosen and sends out a leader message 8: end if Fig. 1. Program for atomicity

STACK: Sparse Timing of Algorithms Using Computational Knowledge

Overlaping sensor intervals

( a ) Spa r s e s e ns o r r e a dings

309

Overlaping sensor intervals

(b) Averaging cost function

Fig. 2. Sensor-centric Data aggregation using atomic averaging

1.5.1 Scaling The raw data need to be scaled to a metric before an application can further process, this family of algorithms use register size, integer with fixed bits or floating point scaling to ensure metric. 1.5.2 Interpolation Sampling rate helps to define how to represent the input signal and interpolate it, when calculating its dynamic range. 1.5.3 Mixing Data stream which represents images often needs an algorithm to fuse data, it uses a pre-processing step on every pixel before the fusion can be performed.

2 2.1

Data Pre-processing – Non Blocking Double Buffering

The synchronization methods used by IPC [5] are generally fall into the category of waited and non-waited. If the algorithm is designed with data-path approach using buffering then less computationally complex algorithm can be implemented which is wait free. The porting issues, which is processor dependent, is the availability of atomic operations to access and operate on memory variables atomically. Algorithm 4, DoubleBuffer() shows how the value of latest is shared between many Readers and continuously updated by one writer. Line 4 used the flag pointing to the buffer pairs which contain the latest update, and the non-blocking write uses number of readers (N + 1) if all are in use, its new value as shown in line 12. The design of double buffering allows to simultaneously have multiple readers access the same buffer. Writer can only interfere with the reader when they both choose to use the same row. This can occur in two cases. The first case can occur when a reader is interrupted after it has chosen a row (after line 3 in Algorithm 4),

310

V. Iyer et al.

but before it updates the use count. The writer then executes, and can potentially choose the same row as the reader. The second case occurs when the writer is interrupted after it has chosen a row (line 9 in Algorithm 4). If this row happens to be Latest, then the reader can also choose to read from the same row. So, it is possible for the readers and the writer to select the same row i. However, the reader will read from the buffer indicated by C1[i], while the writer will use the opposite one. As the writer updates C1[i] only after the complete message is written, and the reader always increments the use count before reading Cl[i], we can guarantee that the writer and readers cannot interfere with each other in this algorithm, even if they happen to use the same row. Algorithm 4. DoubleBuffer 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15:

3

Reader() ridx = Latest inc ReaderCnt[rindx] cl = Cl[ridx] read Buff[ridx][cl] dec ReaderCnt[ridx] Writer() for (ii=Latest;;ii++) if ReaderCnt(ii mod NRows]==0 then break; end if cl= not Cl[ii] write Buff[ii][cl]; Cl[ii]=c; Latest=ii;

Algorithms for Data Coherency

As different sensors are connected to each node, the nodes have to periodically measure the values for the given parameters which are correlated. The inexpensive sensors may not be calibrated, and need processing of correlated data, according to intra and inter sensor variations. The pre-processing algorithms allow to accomplish two functions, one to use minimal number of measurement at each sensor, and the other to represent the signal in its loss-less sparse representation, which allows application level views. Figure 2(a) and 2(b) shows the cluster tree of inter-sensor intervals for a batch of sensors from Table 1. 3.1

Compressed Sensing (CS)

The pre-processing steps used by CS allows to sample at rates lower than the Nyquist rate, it recovers the original signal by rescaling and interpolation, which are save as compressed co-efficient during pre-processing.

STACK: Sparse Timing of Algorithms Using Computational Knowledge

311

– Sampling rate of i.i.ds [8,13] – Aggregation window of the ensemble – The delta ranges possibly measured for the complete ensemble - calibration The signal measured if it can be represented at a sparse representation, then this technique is called the sparse basis as shown in equation (4), of the measured signal. The technique of finding a representation with a small number of significant coefficients is often referred to as Sparse Coding. When sensing locally many techniques have been implemented such as the Nyquist rate [1], which define the minimum number of measurements needed to faithfully reproduce the original signal. Using CS it is further possible to reduce the number of measurement for a set of sensors with correlated measurements [8,9].   x= ϑ(n)Ψn = ϑ(nk )Ψnk , (4)

Consider a real-valued signal x ∈ RN indexed as x(n), n ∈ 1, 2, ..., N. Suppose that the basis Ψ = [Ψ1 , ..., ΨN ] provides a K-sparse representation of x; that is, where x is a linear combination of K vectors chosen from, Ψ, nk are the indices of those vectors, and ϑ(n) are the coefficients; the concept is extendable to tight frames. Alternatively, we can write in matrix notation x = Ψ ϑ, where x is an N × 1 column vector, the sparse basis matrix is N × N with the basis vectors Ψn as columns, and ϑ(n) is an N × 1 column vector with K nonzero elements. Using  . p to denote the ℓp norm, we can write that  ϑ p = K; we can also write the set of nonzero indices Ω1, ..., N , with |Ω| = K. Various expansions, including wavelets [6], Gabor bases [6], curvelets [6], are widely used for representation and compression of natural signals, images, and other data. Algorithms 5, 6 and 7 allow applications to select how it views the state of the nature. FloodMinVal() algorithm uses a k-agreement [5], which allow to represent real-time floating values with lower-bound intervals. Figure 2, uses an averaging algorithm from R-Systems, general tree structure to calculate the dendrogram for the data-set in Table 1. The algorithm gives a good lower bound which has the values between (1.6, 2.25). The power-aware algorithms use clustering which allows to read correlated value, due to lack of calibration in small sensors, the algorithm should be able to update only higher confidence values seen. The algorithm FloodMinRange() allows maximizing the coherency of the previous algorithm FloodMinVal() by calibrating between overlapping ranges, which are active during the cluster formation. FloodMinRange(), Line 12 allows to find the minimum of the overlapping ranges, which is a better representation of the signals value in terms of the sensors current sampling and density of coverage. To check the validity and the effectiveness, we use regression analysis using off-line statistical methods in section 5. 3.2

Coherency Cost Function of Sparse Representation

A single measured signal of finite length, which can be represented in its sparse representation, by transforming into all its possible basis representations. The number of basis for the for each level j can be calculated from the equation as

312

V. Iyer et al.

Aj+1 = A2j + 1

(5) 2

2

So staring at j = 0, A0 = 1 and similarly, A1 = 1 + 1 = 2, A2 = 2 + 1 = 5 and A3 = 52 + 1 = 26 different basis representations. Let us define a framework to quantify the sparsity of ensembles of correlated signals x1 , x2, ..., xj and to quantify the measurement requirements. These correlated signals can be represented by its basis from equation (5). The collection of all possible basis representation is called the sparsity model. x = Pθ

(6)

Where P is the sparsity model of K vectors (K 1, goto step 4. The marked basis has the lowest possible cost value, which is the value currently assigned to the top element.

314

3.4

V. Iyer et al.

Sensor Centric Algorithm

DCS allows to enable distributed coding algorithms to exploit both intra-and inter-signal correlation structures. In a sensor network deployment, a number of sensors measure signals that are each individually sparse in the some basis and also correlated [6,9] from sensor to sensor. If the separate sparse basis are projected onto the scaling and wavelet [8] functions of the correlated sensors(common coefficients), then all the information is already stored to individually recover each of the signal at the joint decoder. This does not require any pre-initialization between sensors. The expanded wavelet optimization and its cost-functions are shown in Figure 3(a) and 3(b). 3.4.1 Joint Sparsity Representation For a given ensemble X, we let PF (X) ⊆ P denote the set of feasible location matrices P ∈ P for which a factorization X = P Θ exits. We define the joint sparsity levels of the signal ensemble as follows. The joint sparsity level D of the signal ensemble X is the number of columns of the smallest matrix P ∈ P. In these models each signal xj is generated as a combination of two components: (i) a common component zC , which is present in all signals, and (ii) an innovation component zj , which is unique to each signal. These combine additively, giving xj = zC + zj , j ∈ ∀ (7) X = PΘ

(8)

We now introduce a bipartite graph G = (VV , VM , E), as shown in Figure 4, that represent the relationships between the entries of the value vector and its measurements. The common and innovation components KC and Kj , (1 < j < J), as well as the joint sparsity D = KC + KJ . Signal 4.7 6.7 2.7 1.6 3.2

0

3.0 4.5 1.5 1.8 2.8 0.8 0

5.5 2.1 1.6 3.7 1.6 1.8

3.8 2.6 1.7 -1.1 1.7

0

0

0

0

0

0

-1 0.5 1.6 -0.7 -0.1

0 1

0 0.07 0 -0.2 -0.4 -0.7 1.17 0.57 0 -0.4 0

3.2 0.5 0.3 1.4 0.9 0.9 0.04 0.04 0.1 -0.3 0.2 -0.1 0.28 0.28-0.2 -0.2 1.9 1.3 0.8 -0.5 0.9

0 0.03 0 -0.1 0.2 -0.4 0.6 0.28 0 +

(a) Sparse sensor readings

+

-0.2

0

+

(b) Basis representation with maximal overlap

Fig. 3. Sensor-centric Data Fusion during the aggregation step using sparse STACK model.

STACK: Sparse Timing of Algorithms Using Computational Knowledge Value vector coefficient

315

Measurement (1,1)

1

(1,2) 2 (2,1) 3

(2,2)

4

(j,Mj)

D Vv

Vm

Fig. 4. Bipartite graphs representing fused aggregation for data coherency

The set of edges E is defined as follows: – The edge E is connected for all Kc if the coefficients are not in common with Kj . – The edge E is connected for all Kj if the coefficients are in common with Kj . A further optimization can be performed to reduce the number of measurement made by each sensor, the number of measurement is now proportional to the maximal overlap of the inter sensor ranges and not a constant as shown in equation (4). This is calculated by the common coefficients Kc and Kj , if there are common coefficients in Kj then one of the Kc coefficient is removed and the common Zc is added, these change does not effecting the reconstruction of the original measurement signal x. 3.5

Distributed Fused Parameter Dictionary

The sample sensor measurements of Table 1 and its transformed basis are shown in Figure 3 (a) and Figure 3 (b), illustrate all its possible basis representations. The cast-function [7] searches to find an optimal (grey rectangles) best basis matching the least number of coefficients to represent the signal without overlaps. The lowest range is calculated by selecting consecutive significant coefficients (1.3, 1.7), which determine the maximal overlap for the sensor intervals. This best basis dictionary is stored in the hashed location of the application’s search tree.

316

4 4.1

V. Iyer et al.

STACK Model Validation Lower Bound Validation Using Covariance

The Figure 3(b) shows lower bound of the overlapped sensor i.i.d. of S1 − S8 , as shown it is seen that the lower bound is unique to the temporal variations of S2 . In our analysis we will use a general model which allows to detect sensor faults. The binary model can result from placing a threshold on the real-valued readings of sensors. Let mn be the mean normal reading and mf the mean event reading for a sensor. A reasonable threshold for distinguishing between the two m +m possibilities would be 0.5( n 2 f ). If the errors due to sensor faults and the fluctuations in the environment can be modeled by Gaussian distributions with mean 0 and a standard deviation σ, the fault probability p would indeed be symmetric. It can be evaluated using the tail probability of a Gaussian [9], the Q-function [9], as follows:   m +m   (0.5( n 2 f ) − mn ) mf − mn p=Q =Q (9) σ 2σ From the measured i.i.d. value sets we need to determine if they have any faulty sensors. This can be shown from equation (9) that if the correlated sets can be distinguished from the mean values then it has a low probability of error due to sensor faults, as sensor faults are not correlated. Using the statistical analysis package R, we determine the correlated matrix of the sparse sensor outputs as shown This can be written in a compact matrix form if we observe that for this case the co-variance matrix is diagonal, this is, ⎛ ⎞ ρ1 0 .. 0 ⎜ ⎟ ⎜ 0 ρ2 .. 0 ⎟ ⎟ Σ=⎜ (10) ⎜ : : ց : ⎟ ⎝ ⎠ 0 0 .. ρd The correlated co-efficient are shown matrix (11) the corresponding diagonal elements are highlighted. Due to overlapping reading we see the resulting matrix shows that S1 and S2 have higher index. The result sets is within the desired bounds of the previous analysis using DWT. Here we not only prove that the sensor are not faulty but also report a lower bound of the optimal correlated result sets, that is we use S2 as it is the lower bound of the overlapping ranges. Table 1. Sparse representation of sensor values

Sensors S1 i.i.d.1 2.7 i.i.d.2 4.7 i.i.d.3 6.7

S2 0 1.6 3.2

S3 1.5 3 4.5

S4 0.8 1.8 2.8

S5 3.7 4.7 5.7

S6 0.8 1.6 2.4

S7 2.25 3 3.75

S8 1.3 1.8 2.3

STACK: Sparse Timing of Algorithms Using Computational Knowledge

⎛− ⎞ → 4.0 3.20 3.00 2.00 2.00 1.60 1.5 1.0 ⎜ ⎟ −→ ⎜ 3.2 − 2.56 2.40 1.60 1.60 1.28 1.20 0.80 ⎟ ⎜ ⎟ −−−→ ⎜ 3.0 2.40 2.250 1.50 1.50 1.20 1.125 0.75 ⎟ ⎜ ⎟ ⎜ ⎟ −→ ⎜ 2.0 1.60 1.50 − 1.00 1.00 0.80 0.75 0.5 ⎟ ⎜ ⎟ Σ=⎜ ⎟ −−→ ⎜ 2.0 1.60 1.50 1.00 1.00 0.80 0.75 0.5 ⎟ ⎜ ⎟ −→ ⎜ 1.6 1.28 1.20 0.80 0.80 − 0.64 0.60 0.4 ⎟ ⎜ ⎟ ⎜ ⎟ −−−−→ ⎜ 1.5 1.20 1.125 0.75 0.75 0.60 0.5625 0.375 ⎟ ⎝ ⎠ −−−→ 1.0 0.80 0.750 0.50 0.50 0.40 0.375 0.250

317

(11)

The sensor data are correlated due to variations in deployment they are difficult to calibrate. The pre-processing of data will needs to correct the coefficients before applying the fusion function. As this constraint the design of the algorithm, we need custom sensor hardware. FPGA allows to design data-paths(see Table 2), which keep the algorithm design independent of any pre-processing step. Once such example is double buffering which allows slow data stream from flash memory to settle before the next data-set is applied.

5

Algorithm Acceleration

Table 2 shows all the optimizations available for algorithm performance tuning. The efficiency of the algorithm[4] depends on the Instruction count and CPI [12]. These parameters are determined by the processor design, once the type of tool sets is chosen then further optimization of the program size and speed are targeted. The notion of throughput of algorithm execution and its design dependency on the data-path demands is studied for all the computation models. 5.1

Static Program Order Model

The compiler uses basic block optimization techniques, which can use Instruction Level Parallelism (ILP). Optimization of this type uses infinite resources such as window size and base register counts. The expected throughput may be reduced, as it is dependent on the target. Program consistency and shared data coherency is available in the higher level program construct and can be addressed with simplicity. Some data-path optimizations are possible, memory dependencies, which are architecture dependent, are not addressed at this level, this step allow for fine grain instruction level parallelism. 5.2

Dynamic Out-of-Order Model

The performance measures computed with processor-based architecture and optimization from Table 2, Figure 5 illustrates that there is increase in clock rate

318

V. Iyer et al.

if the pipeline [1,12] is used. In our case, the pipeline depth is varied during re-configuration. The simulation results show that total increase is around 10% using dynamic optimizations. As long theyre available registers, the ILP can execute in parallel by increasing the issue rate gradually. RISC uses Tomusulu’s algorithm [12] for register reservation to hide any load and read latencies. Processor based IPC mechanism and data cache allows to accelerate the data-path depending on the architecture. General purpose data forwarding, pipeline optimizations including branch predictions and speculation using outof-order execution are supported in the hardware. The pipeline utilizes temporal parallelism. 5.3

STACK

Temporal and spacial data coherency is achieved by using the pre-processing and compressed sensing techniques. The algorithm allows to compressed signal without any information loss and further enhances the working data range of the algorithm. As there are no delays in load and store instruction in the hardware register language [7], it can further accelerate 40% of the instructions as shown in Figure 5(a), which uses dynamic range for bit growth and register allocation as shown in Figure 5(b) and equation (12). fine grain parallelism. At the same time, it can take advantage of instruction level parallelism, which is pipeline optimization technique used in the previous models. k