Electrochemical Impedance Spectroscopy Options for Proton Exchange Membrane Fuel Cell Diagnostics by JORGE IGNACIO VALENZUELA B.A.Sc.. Electrical and Computer Engineering, University of British Columbia, 1993 B.Sc.. Physiology, University of British Columbia, 1988 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF APPLIED SCIENCE in The Faculty of Graduate Studies (Electrical and Computer Engineering) THE UNIVERSITY OF BRITISH COLUMBIA November 2007 Jorge Ignacio Valenzuela, 2007 ii Abstract Electrochemical impedance spectroscopy (EIS) has been exploited as a rich source of Proton Exchange Membrane Fuel Cell (PEMFC) diagnostic information for many years. Several investigators have characterized different failure modes for PEMFCs using EIS and it now remains to determine how this information is to be obtained and used in a diagnostic or control algorithm for an operating PEMFC. This work utilizes the concept of impedance spectral fingerprints (ISF) to uniquely identify between failure modes in an operating PEMFC. Three well documented PEMFC failure modes, carbon monoxide (CO) poisoning, dehydration, and flooding were surveyed, modelled, and simulated in the time domain and the results were used to create a database of ISFs. The time domain simulation was realized with a fractional order differential calculus state space approach. A primary goal of this work was to develop simple and cost effective algorithms that could be included in a PEMFC on-board controller. To this end, the ISF was discretized as coarsely as possible while still retaining identifying spectral features using the Goertzel algorithm in much the same way as in dual tone multi-frequency detection in telephony. This approach generated a significant reduction in computational burden relative to the classical Fast Fourier Transform approach. The ISF database was used to diagnose simulated experimental PEMFC failures into one of five levels of failure: none (normal operation), mild, moderate, advanced, and extreme from one of the three catalogued failure modes. The described ISF recognition algorithm was shown to correctly identify failure modes to a lower limit of SNR = -1dB. iii Table of Contents Abstract.................................................................................................................................ii Table of Contents.................................................................................................................iii List of Tables ....................................................................................................................... iv List of Figures....................................................................................................................... v Acronyms, Abbreviations, and Symbols ...........................................................................viii Acknowledgements.............................................................................................................. ix 1 Introduction................................................................................................................... 1 2 Fuel cell review............................................................................................................. 4 3 Fault conditions .......................................................................................................... 14 3.1 Dehydration ........................................................................................................ 14 3.2 Flooding.............................................................................................................. 17 3.3 Carbon monoxide poisoning............................................................................... 19 4 Fuel cell time domain simulation ............................................................................... 25 4.1 Non-integer order system response .................................................................... 25 5 Electrochemical Impedance Spectroscopy ................................................................. 31 5.1 Obtaining the impedance spectra........................................................................ 32 5.2 Multi-frequency EIS techniques and noise......................................................... 34 5.3 Multi-frequency EIS techniques and non-stationarity........................................ 36 5.4 Non-linearity and the multi-frequency stimulus waveform................................ 38 6 Goertzel algorithm based spectral analysis................................................................. 40 6.1 The basic Goertzel algorithm ............................................................................. 40 6.2 Computation burden (memory and operations).................................................. 43 6.3 Goertzel filter bank ............................................................................................. 44 7 Spectral fingerprint-based diagnostics........................................................................ 50 7.1 Spectral fingerprint evolution............................................................................. 53 7.2 Failure mode identification via spectral fingerprinting ...................................... 56 8 Conclusions ................................................................................................................ 66 9 Suggestions for future work........................................................................................ 68 References........................................................................................................................... 69 Appendix 1. Simulation of non-integer order transfer function ...................................... 73 Appendix 2. Equivalent circuit values from the literature .............................................. 79 Appendix 3. Matlab program listings.............................................................................. 82 iv List of Tables Table 1.1 State of and contributions to the art in this investigation .................................... 3 Table 3.1 Effects of dehydration on anode charge transfer resistance and membrane resistance..................................................................................................................... 16 Table 3.2 Effects of flooding on cathode diffusion and charge transfer parameters ......... 18 Table 3.3 Effects of Carbon Monoxide contamination on anode diffusion parameters.... 22 Table 6.1 Computational and memory requirements of FFT and Goertzel algorithms..... 44 Table 7.1 Computational and memory requirements of spectral fingerprint diagnostic algorithm..................................................................................................................... 53 Table 7.2 Experimental failure mode detection challenges............................................... 65 Table A2.1 Equivalent circuit parameters from the literature ........................................... 81 v List of Figures Figure 1.1 Scope of investigation ........................................................................................ 2 Figure 2.1 Basic PEMFC operation..................................................................................... 5 Figure 2.2 Graphical representation of the Butler-Volmer equation................................... 6 Figure 2.3 Charge double layers in the PEMFC [5] ............................................................ 8 Figure 2.4 Simple PEMFC equivalent circuit...................................................................... 9 Figure 2.5 Schematic polarization curve for PEMFC ....................................................... 10 Figure 2.6 Polarization curve [12]. Operating points a – g described in Appendix 2 ...... 11 Figure 2.7 Bode plots of fitted equivalent circuit in Appendix 2 for operating points a, b, & c .............................................................................................................................. 11 Figure 2.8 Bode plots of fitted equivalent circuit in Appendix 2 for operating points c, d, e, f & g ........................................................................................................................ 12 Figure 2.9 Nyquist plots of fitted equivalent circuit in Appendix 2 for operating points a & b. ............................................................................................................................. 12 Figure 2.10 Nyquist plots of fitted equivalent circuit in Appendix 2 for operating points c, d & e. .......................................................................................................................... 13 Figure 3.1 Water management in PEMFC ........................................................................ 15 Figure 3.2 Effect of dehydration on PEMFC: Bode plot (a→e = normal→dehydrated) .. 16 Figure 3.3 Effect of dehydration on PEMFC: Nyquist plot (a→e = normal→dehydrated) .................................................................................................................................... 17 Figure 3.4 Effect of flooding on PEMFC: Nyquist plot (a→e = normal→flooded) ......... 19 Figure 3.5 Effect of flooding on PEMFC: Bode plot (a→e = normal→flooded) ............. 19 Figure 3.6 Effect of CO contamination on PEMFC: Bode plot (a→e = normal→contaminated) .............................................................................................. 23 Figure 3.7 Effect of CO contamination on PEMFC: Nyquist plot (a→e = normal→contaminated) .............................................................................................. 23 Figure 4.1 Analytic and numeric solutions for Dα sin(x) with 2 times oversampling factor .................................................................................................................................... 26 Figure 4.2 Analytic and numeric solutions for Dα sin(x) with 10 times oversampling factor ........................................................................................................................... 27 Figure 4.3 Analytic and numeric solutions for Dα sin(x) with 100 times oversampling factor ........................................................................................................................... 27 Figure 4.4 Equivalent circuit for integer order software comparison............................... 28 Figure 4.5 Comparison between commercial integer order simulation software and fractional order simulation software. .......................................................................... 28 Figure 4.6 Equivalent circuit for non-integer order software comparison ....................... 29 vi Figure 4.7 Comparison between analytical non-integer order frequency response versus Fourier transform of fractional order simulation software time domain output: Nyquist plot. ............................................................................................................... 29 Figure 4.8 Comparison between analytical non-integer order frequency response versus Fourier transform of fractional order simulation software time domain output: Bode plot. ............................................................................................................................. 30 Figure 5.1 The frequency domain effect of time domain signal truncation ...................... 37 Figure 6.1 Frequency response and z-plane pole zero location: basic Goertzel algorithm42 Figure 6.2 Realization of basic Goertzel algorithm........................................................... 42 Figure 6.3 Frequency response and z-plane pole zero location: modified Goertzel algorithm..................................................................................................................... 43 Figure 6.4 Realization of modified Goertzel algorithm..................................................... 43 Figure 6.5 Computational requirements of FFT and modified Goertzel algorithms for N = 4×106 .......................................................................................................................... 44 Figure 6.6 Implementation of spectral analysis with Goertzel filter bank ........................ 45 Figure 6.7 Time domain representation of current stimulus and voltage response signals46 Figure 6.8 Impedance calculated via FFT and Goertzel algorithms.................................. 46 Figure 6.9 Noise performance of FFT and Goertzel algorithms ....................................... 47 Figure 6.10 Implementation of spectral analysis with Decimated Goertzel filter bank .... 48 Figure 6.11 Noise performance of Decimated Goertzel algorithm ................................... 49 Figure 7.1 Non-FRA-based PEMFC diagnostics .............................................................. 50 Figure 7.2 Discrete frequency points requirement ............................................................ 52 Figure 7.3 Non FRA based PEMFC diagnostics using Goertzel and EIS fingerprint recognition .................................................................................................................. 52 Figure 7.4 Five spectral fingerprints at increasing stages of CO PEMFC poisoning (SNR=0dB)................................................................................................................. 54 Figure 7.5 Five spectral fingerprints at increasing stages of CO PEMFC poisoning (SNR=–1.5dB)............................................................................................................ 54 Figure 7.6 Five spectral fingerprints at increasing stages of PEMFC dehydration (SNR=0dB)................................................................................................................. 55 Figure 7.7 Five spectral fingerprints at increasing stages of PEMFC dehydration (SNR=– 1.3dB) ......................................................................................................................... 55 Figure 7.8 Five spectral fingerprints at increasing stages of PEMFC flooding (SNR=0dB) .................................................................................................................................... 56 Figure 7.9 Five spectral fingerprints at increasing stages of PEMFC flooding (SNR=– 1.5dB) ......................................................................................................................... 56 Figure 7.10 Spectral fingerprint matching during normal PEMFC operation (SNR=0dB) .................................................................................................................................... 57 Figure 7.11 Spectral fingerprint matching during PEMFC CO poisoning (SNR=0dB) ... 58 vii Figure 7.12 Spectral fingerprint matching during PEMFC CO poisoning (SNR=-1.5dB)59 Figure 7.13 Spectral fingerprint matching during PEMFC Dehydration (SNR=0dB)...... 60 Figure 7.14 Spectral fingerprint matching during PEMFC Dehydration (SNR=-1.3dB) .61 Figure 7.15 Spectral fingerprint matching during PEMFC Flooding (SNR=0dB) ........... 62 Figure 7.16 Spectral fingerprint matching during PEMFC Flooding (SNR=-1.5dB) ....... 63 Figure 7.17 Spectral fingerprint matching during PEMFC Dehydration (SNR=0dB)...... 64 Figure A1.1 One electrode of PEMFC .............................................................................. 73 viii Acronyms, Abbreviations, and Symbols EIS Electrochemical impedance spectroscopy FRA Frequency response analyzer STFT Short term Fourier transform FFT Fast Fourier transform DSP Digital signal processing PEMFC Proton exchange membrane fuel cell TPB Triple phase boundary PFTE Polytetrafluoroethylene T Temperature F Faraday’s constant R Universal gas constant αa Anodic transfer coefficients io Exchange current density i Current density Rct,a Charge transfer resistance, anode BV Butler-Volmer Aeff Effective area dH Helmholtz layer thickness εo Permittivity of free space εr Relative permittivity Cdl Double layer capacitance Rel Electrolyte resistance Rs Series resistance s Laplace variable A(i) Warburg resistance (also denoted as σ) δ Diffusion film thickness τ Time constant of diffusion Deff Effective diffusion constant λc Stoichiemtric coefficient, cathode θx Adsorption fraction species x GL Grünwald-Letnikov SNR Signal to noise ratio IIR Infinite impulse response DFT Discrete Fourier transform CNLS Complex non-linear least squares ix Acknowledgements I would like to thank my supervisors, Dr. W.G. Dunford and Dr. W. Merida for their patience and guidance throughout this project. My colleagues, Tatiana Romero and Javier Gazzarri have been constant sources of fruitful discussion and challenging questions. I offer my sincere thanks to them for their participation and friendship. For most of my adult life Chris Parks, Klaus Kallesøe and Gordon White have been the brothers I wish my parents had begat. I thank you all for your great friendship and diligent efforts to maintain it despite my lack of availability over the last few years. My parents and sister deserve a great deal of thanks for everything they have done for me during this and my previous degrees. Without question, if I stand tall, it is only because I can stand on their shoulders. Finally, and certainly not least, I would like to thank my wife, Maria Alicia Silva. She has been a constant source of love, inspiration, support, and good humour throughout the last 9 years of my life. Without her, none of this would have been possible or even worthwhile. 1 1 Introduction Fuel cells convert chemical energy from hydrogen, methanol, diesel, gasoline, etc. into electrical energy. While fuel cells are capable of very high conversion efficiencies in both stationary and mobile applications, the automotive application is the most anticipated application due to the elimination of harmful atmospheric emissions. Apart from the obvious questions about hydrogen generation and the materials issues that plague fuel cells, questions arise with regard to how the system is going to be tested and controlled in the field. Many different electrochemical techniques have been used to study fuel cells. These include current interrupt, cyclic voltammetry, electrochemical impedance spectroscopy (EIS), Raman spectroscopy, etc. Of these, the impedance spectrum provides one of the richest sources of information about electrochemical systems. In addition, the non- invasive nature of EIS makes it a valuable diagnostic tool. Complex impedance values are usually obtained by exciting the electrochemical system with a single frequency sinusoidal voltage or current signal and measuring the current or voltage respectively. This same process is carried out sequentially by a frequency response analyzer (FRA) for each frequency of interest until the desired operating frequency range scan is complete. The EI spectrum is then fitted to an equivalent circuit or distinct frequency bands are compared to determine the state of health of the fuel cell (Figure 1.1). This approach is unsuitable for embedded diagnostics use for two reasons: first, the equipment involved is cumbersome and expensive; second, the measurement is lengthy and often, the state of the fuel cell evolves during the measurement. There are techniques available to mitigate this latter problem while still using FRA-based EI spectroscopy [1, 2], but they involve additional testing and digital signal processing that further increases the experimental duration and computation burden. Despite the fact that fuel cell non-stationarity imposes a measurement challenge, the information useful to designers and to real-time control systems is precisely that which is embedded within the time-varying impedance data. For example, as the fuel cell membrane dehydrates, it loses efficiency and drops power output. At the same time, the 2 polymer electrolyte membrane impedance increases and can thus be used as a signal to increase the humidification. Many researchers have approached this problem from the perspective of the Short Time Fourier Transform (STFT). This approach employs the Fast Fourier Transform (FFT) of time domain system input-output data multiplied by an appropriate short time window to obtain the impedance spectrum at discrete slices of time. As with FRA-based EIS, this spectrum is then fitted to equivalent circuits or distinct frequency bands are compared. Unfortunately, this approach while more suitable than traditional FRA-based EIS diagnostics, still imposes a sizeable computational burden on the diagnostic equipment. FC freq-domain impedance model FRA-EIS STFT-EIS Goertzel-EISFC time domain model Impedance spectral fingerprint analysis FC equivalent circuit model Equivalent circuit fitting Bandpass filter output ratios Figure 1.1 Scope of investigation This investigation extends the work of Popkirov, Schindler, Darowicki and others on the STFT by employing an alternate digital signal processing (DSP) method to obtain the electrochemical impedance spectrum: the Goertzel algorithm (highlighted in bold in Figure 1.1). Further, the EI spectrum thus obtained is analyzed in much the same way as fingerprints are used to identify criminals within the justice system. Like the use of fingerprinting in criminal processes, this requires (i) uniqueness of the fingerprint, and (ii) the prior characterization of the EI spectrum for each failure mode; the fingerprint database. To evaluate this approach, equivalent circuit models for the fuel cell under normal and fault conditions are obtained from the literature. These equivalent circuit models are used in two ways (Figure 1.1): First, they give rise to a reference frequency domain impedance spectrum. Since this spectrum is calculated via a closed form analytic expression, it does not suffer from any bias or variance imposed by a spectral estimation algorithm. Second, they provide the parameters for a time domain state space model that 3 allows us to simulate the time domain response of the fuel cell to an EIS stimulus signal. The spectral estimates of the simulated time domain response obtained via Goertzel algorithm can be evaluated against the theoretical reference spectrum. It is hypothesized that these improved approaches should be more suitable for use in a real-time, on-board, impedance-based fuel cell diagnostic module than the traditional STFT approach because of the decreased computational burden, elimination of complex least squares equivalent circuit fitting procedures, and increased specificity of spectral content analysis. In the following chapter, we review the required fuel cell specific information such that we can put subsequent impedance models into physical context. In chapter 3 we describe the fuel cell failure modes to be considered and their respective equivalent circuit impacts. In chapter 4 we create a time-domain fuel cell simulator that allows us to create datasets for testing of our EIS diagnostic routines under controlled and specified fault conditions. In chapter 5, we explore EIS in both its development and diagnostic role. In chapter 6, we apply the Goertzel algorithm to the EIS problem and compare the results with the classical STFT method. In chapter 7, we develop a multi-dimensional spectral fingerprint control space based on the EIS data from chapter 6. Table 1.1 State of and contributions to the art in this investigation State of the art Contribution Chapter Use of the Short Time Fourier Transform (STFT) to calculate experimental EI spectral results Use of the Goertzel algorithm (borrowed from telephony) to reduce computational burden 6 Integer and fractional1 order time domain simulation of PEMFC behaviour Fractional order time domain simulation of PEMFC behaviour 4 Up to 4 EIS frequency band diagnostic algorithms. EIS spectral fingerprint based diagnostics 7 1 There are errors in the derivation of the published fractional order simulation algorithm. We have had personal communications with the article author highlighting the error and we have included the complete, correct derivation in Appendix 1. 4 2 Fuel cell review The polymer electrolyte membrane fuel cell (PEMFC) will be used as a model for the present investigation, but the signal processing approach discussed in this paper can be applied to any fuel cell system including solid oxide, molten carbonate, direct methanol, formic acid, phosphoric acid, etc. A brief summary of the operation of the PEMFC is provided to lay the foundation for the equivalent circuit approximations employed later. For greater detail concerning losses, thermodynamics, heat management, kinetics, catalysis, water balance, etc., the reader is referred to other references [3-5]. Figure 2.1 shows the basic elements of the PEMFC. We can consider it to be a system in which the controlled oxidation of hydrogen via an electrochemical process results in useful electrical work according to the chemical reactions (2.1) through (2.3). ( ) ( )2 2 aqgH H +⇀↽ 2e−+ 0.000 SHEV (2.1) ( ) ( ) 1 22 2 aqgO H ++ + 2e−+ ( )2 @1.229 SHEg STPH O V⇀↽ (2.2) ( ) ( ) ( ) 21 2 2 22 1.229 e g g gH O H O V − + ⇀↽ (2.3) The fuel cell itself consists of three basic elements: the anode, the electrolyte, and the cathode. The anode contains a gas diffusion layer through which the humidified gas travels and encounters the triple phase boundary (TPB) where the fuel gas, solid catalyst (usually platinum), and aqueous electrolyte phases meet. It is at this point that the hydrogen molecule gives up two electrons that travel through the external circuit (the electrolyte is ionically conductive but not electronically conductive). The resulting protons diffuse along their concentration gradient and migrate along the electric field Figure 2.3 through the electrolyte to the cathode TPB where they combine with oxygen and electrons to form water. The resulting cell potential is 1.229V at standard temperature and pressure. The PEMFC employs a sulfonated polytetrafluoroethylene (PTFE) membrane (e.g., Nafion by Dupont) as the electrolyte. This membrane must be sufficiently hydrated that ionic conduction of the protons is possible. At the same time, the anode and cathode TPBs must be sufficiently dry that the gas phase is not occluded from the solid catalyst. 5 Humidification is controlled by oxidant gas flow (higher air flow results in greater evaporation at the cathode thus drying out the membrane) and by direct humidification of the fuel and/or oxidant gas streams. Figure 2.1 Basic PEMFC operation The Butler-Volmer (BV) equation (2.4) governs the kinetics of the hydrogen oxidation (2.1) and oxygen reduction (2.2) reactions 0 a c s s F F RT RTi i e e α αη η− = − (2.4) where T is the temperature in K, F is Faraday’s constant, R is the universal gas constant, αa and αc are the anodic and cathodic transfer coefficients, respectively, i0 is the exchange current density, i is the current density through the fuel cell, and ηs is the surface (or activation) overpotential [6]. The BV current-voltage relationship is plotted in Figure 2.2. A detailed examination of the kinetic (Eq.(2.4) and Figure 2.2) and physical (Figure 2.3) properties of the electrode interface reveals several elements that we will re-cast into an electronic model (Figure 2.4) for ease of interpretation and prediction. The inverse of the slope of the voltammogram (Figure 2.2) gives the first element in our electronic model or equivalent circuit: the polarization resistance or charge transfer resistance, Rct (Figure 2.4). Specifically, the value of Rct (2.6) can be calculated from the 6 small signal ( )0.05s voltsη < approximation of the BV equation (2.5). This expression is obtained by ignoring the second order and greater terms in the Taylor series expansion of the BV equation 0 0 Linear region Fu el ce ll cu rr en t d en si ty (A /c m 2 ) Overpotential (V) Figure 2.2 Graphical representation of the Butler-Volmer equation ( )0 a c si Fi RT α α η= + (2.5) giving ( )0 s ct a c RTR i i F η α α = = + (2.6) This element represents an electrochemical reaction that traverses the electrode electrolyte interface and therefore represents a pathway for DC. Since i0 varies proportionally with reaction rate, a small Rct implies a kinetically favourable reaction. The charge transfer resistance is calculated at the null potential (open circuit) where it is maximal and decreases with current density. When the fuel cell is polarized (i.e., a load current is being drawn), the small signal approximation fails and we turn to the Tafel approximation 0ln ln a s Fi i RT α η= + (2.7) for anodic ( )20 , 0s volts i A cmη −> > ⋅ polarization [6]. From this expression, we obtain 1s ct a d RTR di F i η α = = (2.8) 7 indicating that the charge transfer resistance decreases with increasing load current. This demonstrates that impedance spectra are functions of current density and any spectral fingerprint-matching algorithm must take this into account. While the electrodes and the electrolyte are both close to electroneutral, the region near the electrode-electrolyte interface is not (Figure 2.3). The mobile charges are electrostatically attracted to the oppositely charged species in the adjacent phase. These charges accumulate at the interface and, owing to the charge separated by their solvated ionic radii2, an electric field is established. This charge separation can be modelled as a capacitor. Since the charge distribution is not along a perfect parallel plate, but rather a rough surface, the deep invaginations of which limit the penetration of mobile ions due to size and diffusion lag, the capacitor model departs from the ideal giving rise to the so- called constant phase element3. For the purposes of this investigation, however, we shall treat this element as an ideal capacitor o r eff dl H A C d ε ε = (2.9) where Aeff is the surface area, which is different from the nominal geometric electrode area owing to the catalyst loading, availability, and roughness of the electrode surface, εo and εr are the permittivity of free space and the relative permittivity respectively, and dH is the Helmholtz layer thickness4. This element is in parallel with the charge transfer resistance because current can traverse the electrode-electrolyte interface by either mechanism Figure 2.4. 2 These species are usually weakly bound to water molecules further increasing their effective ionic radii. In fact, the electric field in the double layer is so strong that it aligns the water molecules according to their dipole moment in a very uniform layer. 3 The parallel plate capacitor analogy is further complicated by the fact that the cations (H+) are mobile, while the anions (hydrated sulfonyl group on Nafion polymer) is immobile. The PEMFC cathode double layer is therefore somewhat of a mystery. 4 Gouy and Chapman argued this layer would be diffuse owing to a statistical distribution of ionic species near the metal-liquid interface. Later, Stern showed that there would be a charge-free region defined by the often solvated ionic radius of the ionic species involved. 8 Figure 2.3 Charge double layers in the PEMFC [5] The voltage at each point along the cell width is depicted schematically below Figure 2.3. The double layers are depicted as having similar widths, but different accumulated charge densities. However, it must be stressed that the true shape of these potential curves is unknown and is an area of active research5. Additionally, the slopes in the electrodes and the electrolyte are non-zero owing to the non-zero resistance of these regions. The hydrogen ions diffuse and migrate across the electrolyte through an aqueous sulfonated PTFE matrix. The small channels formed by the hydrophilic regions of the PTFE polymer chain permit proton transport, but the tortuousness of the proton conduction path gives rise to a resistance, Rel. This resistance is a function of ionic concentration (or water content since the solute is fixed), temperature, and stack clamp pressure. This is a purely linear resistance and the lost energy is dissipated as heat. Since this resistance is in series with all the other elements of the fuel cell, we lump it together 5 Additionally, the distribution of anions is more complex than it seems in the diagram because the sulfonyl groups in Nafion are not mobile, but fixed. The ionic species that populate the double layer boundary is therefore not certain. See Appendix 2. 9 with the other pure resistances (bipolar plates and gas diffusion electrodes) and rename it Rs in our equivalent circuit (Figure 2.4). Finally, we must consider mass transfer effects. Consider the reactant species involved in the reactions at the anode or the cathode. At the cathode, the oxygen must diffuse in the gas phase and the hydrogen ions must diffuse in the aqueous phase to the triple phase boundary (TPB) where the electrons can join them and, in the presence of the platinum catalyst, form water. Hydrogen gas diffusion at the anode is very fast and does not limit the reaction rate to any significant extent. Water, however, is required to transport the products of the hydrogen reduction reaction at the anode and its availability is often limited by electro-osmotic drag mediated dehydration at the anode. It is the diffusion of water that most affects the value of the Warburg impedance under normal conditions [7]. These mass transfer effects can be modelled using the analogy between a resistive capacitive transmission line and Fick’s law. The result is the finite Warburg impedance ( ) ( ) ( )tanhW sZ s A i s τ τ = (2.10) where s is the Laplace variable, and A(i) is a non-linear function of current density, i [8, 9] and 2 effDτ δ= (2.11) is the time constant of diffusion with δ representing the thin film thickness6 and Deff, the effective diffusion coefficient. Rct Cdl ZW Rs Rct Cdl ZW Anode Electrolyte Cathode Figure 2.4 Simple PEMFC equivalent circuit 6 Not to be confused with the Helmholtz layer thickness, dH 10 While the impedance of the anode electrode-electrolyte interface is neglected in many investigations, we have retained these elements in our equivalent circuit model because the anode is significantly affected by some of the fault conditions that we will simulate. It is important to note that most of these equivalent circuit parameters change with DC operating point in addition to changing with fault conditions. Three principal regions in the PEMFC polarization curve (Figure 2.5) can be identified: the Activation, Ohmic, and Mass Transfer Limited regions. The evolution of impedance with current density can be observed by noting the overvoltages (voltage losses) associated with each of these regions. In the activation region, cathode activation impedance dominates and the Ohmic and Warburg impedance are negligible. In the ohmic region, activation impedance drops and is roughly constant, ohmic impedance is constant and Warburg impedances are still negligible. Lastly, in the Mass Transfer Limited region, the Warburg impedance begins to grow rapidly [3, 5, 10]. Most PEMFCs are operated at or near peak power, near the end of the ohmic region. Cell current density (A·cm2) Ce ll vo lta ge (V ) η cathode η anode η ohmic η mass transfer Ohmic region ← Activation region Mass transfer region → ← Power Figure 2.5 Schematic polarization curve for PEMFC In this study, the equivalent circuit of Figure 2.4 has been fitted to EIS data from the literature [9, 11-13] to provide a model with which to examine the signal processing routines described herein (see Appendix 2). The frequency response (Figure 2.7 and Figure 2.8) of the fitted equivalent circuit is plotted (Figure 2.6) at several operating points in the activation (a,b), ohmic (c,d,e), and the beginning of the mass transfer limited regions (f,g). 11 0 100 200 300 400 500 600 0 0.2 0.4 0.6 0.8 1 Ce ll vo lta ge (V ) a b c d e f g 0 80 160 240 320 400 Po w er (m W ·cm − 2 ) Current Density (mA·cm2) Figure 2.6 Polarization curve [12]. Operating points a – g described in Appendix 2 10−2 10−1 100 101 102 Im pe da nc e (Ω ) a b c 10−3 10−2 10−1 100 101 102 103 104 −80 −60 −40 −20 0 Frequency (Hz) Ph as e (de gre es) Figure 2.7 Bode plots of fitted equivalent circuit in Appendix 2 for operating points a, b, & c 12 0.01 0.02 0.03 0.04 0.05 0.06 Im pe da nc e (Ω ) c d e f g 10−3 10−2 10−1 100 101 102 103 104 −40 −30 −20 −10 0 Frequency (Hz) Ph as e (de gre es) Figure 2.8 Bode plots of fitted equivalent circuit in Appendix 2 for operating points c, d, e, f & g 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 6 0 0.5 1 1.5 2 2.5 3 Real(Z) (Ω) − Im ag (Z ) ( Ω ) a b c d e Figure 2.9 Nyquist plots of fitted equivalent circuit in Appendix 2 for operating points a & b. 13 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0 0.005 0.01 Real(Z) (Ω) − Im ag (Z ) ( Ω ) a b c d e Figure 2.10 Nyquist plots of fitted equivalent circuit in Appendix 2 for operating points c, d & e. In summary, a conceptual PEMFC equivalent circuit model was created from the literature and the electrochemical origins of the classical Randles cell components were described. This model goes beyond those typically reported in the literature in that it includes all potential equivalent circuit elements rather than a reduced set as in many investigations. In this manner, variation of any parameter, regardless of the relative magnitude of its spectral effect, can be evaluated. Finally, values for the equivalent circuit parameters were derived from data reported in the literature and it was shown that the theoretical EI spectra coincide well with spectra reported in the literature (Appendix 2). 14 3 Fault conditions We are concerned with faults that begin to manifest electrically within minutes of the onset of the fault condition. Long-term degradation processes that occur on a time scale of thousands of hours of operation are not considered. Long term deterioration, however, cause gradual increases of both Rct,a and Rct,c owing to both reversible and irreversible mechanisms. This complicates fault detection for on-board implementations of the methods discussed in this investigation in that “normal” fuel cell operation must be continually redefined or measured. Schulze et al. have shown that there are both reversible and irreversible mechanisms involved in the long-term degradation of the fuel cell. The irreversible degradation is caused by agglomeration of the platinum catalyst [7]. Having suffered such degradation, the presence of carbon monoxide at the fuel cell anode is more pronounced (see section 3.3) because the effective platinum loading is reduced. For the purposes of this investigation, this effect will be treated as negligible since the increase in Rct,a owing to long term irreversible degradation is fourfold over a 1000 hour period, whereas the CO poisoning effect on Rct,a is two orders of magnitude over a few hours [14]. The reversible degradation process is mostly associated with membrane hydration state and recovers after shut down and re-start and will not be considered further. 3.1 Dehydration Water management is one of the most critical elements of PEM fuel cell design [3, 15-17]. Water can enter the PEMFC via the anode and cathode gas streams and is produced at the cathode by the oxygen reduction reaction (Eqn. (2.2) & Figure 3.1). It exits the PEMFC via evaporation, primarily at the cathode. Water management is further complicated by electro-osmotic drag of water from the anode to the cathode. This means that despite a net production of water in the PEMFC and the counteracting diffusion and convection processes, the anode becomes progressively dehydrated. Further, electro-osmotic drag increases linearly with current density such that the problem is more difficult to manage at high loads. The contrary is true of the cathode. It will progressively flood, especially at high load currents, unless water is removed. Dehydration has several deleterious effects 15 on durability and reliability that we seek to prevent by careful monitoring of hydration [18, 19]. Two main effects of dehydration manifest themselves in PEMFCs. Anode Electrolyte Cathode H2O Diffusion H2O Convection H2O Electro-osmotic dragH2O in H2O out H2O in H2O out H2O from ORR Figure 3.1 Water management in PEMFC First, as the cell dehydrates, the proton conduction channels in the PTFE matrix constrict increasing Rs. As mentioned earlier, the increased resistance produces losses that are dissipated as heat. In extreme cases, parts of the membrane become so heated that membrane drying accelerates which, in turn, further exacerbates the dehydration problem. The heat permanently damages the proton conduction channels in the membrane and can even result in catastrophic failure of the PEMFC due to perforation of the membrane resulting in an inability to maintain reactant separation. Second, as the electrolyte is dehydrated the anode TPB begins to dry out and either destroys TPBs by eliminating the liquid phase or makes TPB “islands” (i.e., H+ is produced, but it cannot reach the electrolyte) [16]. This drop in catalytic efficacy is manifest as an increase in Rct,a in our equivalent circuit. Some representative results from the literature are summarized in Table 3.1. 16 Table 3.1 Effects of dehydration on anode charge transfer resistance and membrane resistance % ∆ in Rct,a Rs Investigator Notes 16.7 9.3 [13] A=28cm2, j=500mA·cm2, Pa=Pc=1atm, T=75°C, (dry cathode and anode), H2/O2 24.1 16.1 [17] A=50cm2, j=200mA·cm2, T=80°C, ∆RH=74%, Nafion 117, H2/O2 flow=200cm3·min-1, 16.1 64 [20] Hydrogenics series 500 stack (1 cell), j=400mA·cm2, Pa,Pc=121.3kPa, T=65°C, dry H2 & O2 gas feed, λa=1.2, λc=2 31.3 71.9 [21] A=150cm2, j=467mA·cm2, Pa=Pc=1atm, T=60°C, RHa=10%, RHc=15%, λa=1.2, λc=4, H2:Air Andreaus recommends minimizing electrolyte thickness and membrane equivalent weight in order to maximize the back diffusion of water thus preventing dehydration [13]. For a given fuel cell design, however, the control strategy is limited to modifying humidification of gas feeds or stoichiometric ratios thus requiring minimal lag humidity detection which is the object of this study. 0.01 0.02 0.03 0.04 0.05 Im pe da nc e (Ω ) a b c d e 10−3 10−2 10−1 100 101 102 103 104 −30 −25 −20 −15 −10 −5 0 Frequency (Hz) Ph as e (de gre es) Figure 3.2 Effect of dehydration on PEMFC: Bode plot (a→e = normal→dehydrated) 17 0 0.005 0.01 0.015 0.02 0.025 0.03 0.035 0.04 0.045 0 0.005 0.01 Real(Z) (Ω) − Im ag (Z ) ( Ω ) a b c d e Figure 3.3 Effect of dehydration on PEMFC: Nyquist plot (a→e = normal→dehydrated) Using a 0 to 50% linear increase in Rs and 0 to 20% linear increase in Rct,a we simulate progressive dehydration with time under dehydrating conditions (e.g., dry gas feeds). Bode plots of spectra from dehydrated PEMFCs are characterized by an increase of cell impedance magnitude across all frequencies and an increase (less capacitive) in phase angle in the mid to high frequencies. The dehydration effect is equally visible in the Nyquist plot, which shows a right lateral shift by the amount of the membrane resistance increase. The anode charge transfer resistance does not manifest on the graph because the cathode charge transfer resistance is orders of magnitude larger and masks the anode contribution to the overall PEMFC faradaic impedance. The spectra shown in Figure 3.2 and Figure 3.3 are consistent with published results in the references in Table 3.1. 3.2 Flooding Under high load conditions, fuel cell current density increases and the production of water at the cathode by oxygen reduction reaction (2.2) and by electro-osmotic drag increases (Figure 3.1). Unchecked, this can lead to flooding of the GDL. The initial stage of flooding is characterized by a slow decrease in output voltage as liquid water begins to accumulate in the GDL and constrict the gas flow channels thus increasing the diffusion related equivalent circuit parameters Ac(i) and τc. Additionally, some TPBs are lost as the gas phase becomes occluded from the solid catalyst and liquid electrolyte thus raising Rct,c. This process occurs over minutes [16, 22-24]. When the GDL channels become completely blocked, the subsequent stage of flooding manifests as a drastic voltage drop over the course of seconds [20-22]. This fault condition is especially damaging if allowed to develop fully because the fuel cell can go into reversal (-ve output voltage). This can result in electrolysis instead of oxygen 18 reduction, localized heating and permanent damage to the cell. The reversal mechanism can damage both the membrane and the catalyst layer [25] and lead to long term degradation effects [7]. Detection of the flooding condition by simple voltage monitoring (first stage) has been proposed, but it is not as fast or sensitive as an EIS-based diagnostic when operating at high current levels [21, 22]. Recovery from this latter stage of flooding failure is effected by a high pressure pulse of reactant gas to clear the channel [5]. However, preventing this extent of failure by increasing oxidant gas (typically air) flow to exaggerate evaporation while the flooding is still in the initial stages is the primary goal of humidification control. There are few studies in the literature devoted to EIS of flooding PEMFCs. Fewer still fit the EI spectral data to the Randles cell. Two such results are presented in Table 3.2 [15, 20]7. Table 3.2 Effects of flooding on cathode diffusion and charge transfer parameters % ∆ in Ac(j) τc Rct,c Investigator Notes 817 8.07 104 [21] A=150cm2, j=467mA·cm2, Pa=Pc=1atm, T=60°C, RHa=70%, RHc=50%, λa=1.2, λc=4, H2:Air 1200 - 150 [20] Hydrogenics series 500 stack (1 cell), j=400mA·cm2, Pa,Pc=121.3kPa, T=65°C, λa=1.2, λc=2→1.18 We use a 0 to 900% linear increase in Ac(j) and 0 to 100% linear increase in Rct,c to simulate progressive cathode gas diffusion channel flooding with time under flooding conditions (e.g., reduced λc). Bode plots of spectra from dehydrated PEMFCs are characterized by an increase of low frequency cell impedance magnitude and a decrease in phase angle (more capacitive) in the mid to low frequencies. The flooding effect can be related to equivalent circuit parameter changes in the Nyquist plot, which shows the growth of the high frequency loop dominated by Rct,c and Cdl,c as well as a growth of the diffusion resistance, Ac(j), characterized by an increase in the size of the low frequency loop with its associated 45° slope on the left edge. The spectra shown in Figure 3.4 and Figure 3.5 are consistent with published results in the references in Table 3.2. 7 Merida did not report an increase in Rct,c as observed in other studies, but the impairment of diffusion (↑Ac(j) & ↑τc) was observed (15. Merida, W.R., Diagnosis of PEMFC stack failures via electrochemical impedance spectroscopy, in Mechanical Engineering. 2002, University of Victoria: Victoria.) 19 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 0.11 0.12 0.13 0.14 0.15 0.16 0 0.01 0.02 0.03 Real(Z) (Ω) − Im ag (Z ) ( Ω ) a b c d e Figure 3.4 Effect of flooding on PEMFC: Nyquist plot (a→e = normal→flooded) 10−2 10−1 Im pe da nc e (Ω ) a b c d e 10−3 10−2 10−1 100 101 102 103 104 −40 −30 −20 −10 0 Frequency (Hz) Ph as e (de gre es) Figure 3.5 Effect of flooding on PEMFC: Bode plot (a→e = normal→flooded) 3.3 Carbon monoxide poisoning The use of pure hydrogen to power fuel cell vehicles is problematic for three reasons. First, on-board storage is not yet as simple, reliable, and inexpensive as comparable hydrocarbon storage and it has a lower gravimetric and volumetric power density. Second, widespread distribution of hydrogen is not yet available although it can be argued that if the market manifests a need, the energy companies will follow. Third, pure hydrogen is not readily available in nature and must be obtained either from hydrocarbon sources by reforming reactions or from renewable sources which, at present, lack power 20 density. For at least the early stages of fuel cell commercialization, then, hydrogen supply will be largely from existing hydrocarbons such as methane via the endothermic reforming reaction ( )2 22n m mC H n H O nCO n H+ → + + (3.1) This carbon monoxide (CO) contaminates the fuel cell platinum catalyst via the exothermic water gas shift reaction 2 2 2CO H O CO H+ → + (3.2) by occupying platinum metal (M in (3.3)) catalyst sites to the exclusion of hydrogen [26, 27]. 2 2 2 2 2 2 | | | | | | | | | | | 2 adjacent sites dissolved linear adsorbed adsorbed H O CO activated complexadjacent sites H O CO H O CO M M M M CO H O CO H O e CO OH e H M M M M M M CO OH CO e H M M M − − + − + + + → + + → → → + + + upcurlybracketleftupcurlybracketmidupcurlybracketright upcurlybracketleftupcurlybracketmidupcurlybracketrightupcurlybracketleftupcurlybracketmidupcurlybracketright ⋯ ⋯ ⋯ ⋯ (3.3) The rate of hydrogen oxidation in the presence of CO is given by ( )2 2 21H CO H COi i θ= − (3.4) where θx is the adsorption fraction of species x. Since the hydrogen oxidation reaction (2.1) is so much faster than CO oxidation, it controls the surface potential when θCO is small. When θCO increases, the overpotential grows until it reaches that which is required to oxidize CO [27, 28]. 21 Careful hydrogen reformer design reduces the concentration of CO in reformate gas feeds from parts per hundred to parts per million [29, 30], but this amount is still sufficient to limit the hydrogen reaction rate due to the preferential adsorption of CO over hydrogen together with the lack of a spontaneous CO removal mechanism. Thus, while the overpotential of the CO-free anode (hydrogen oxidation) reaction is typically orders of magnitude lower than the cathode (oxygen reduction) reaction, CO poisoning increases the anode overpotential by two orders of magnitude resulting in considerable reduction in output power [27]. Several CO-poisoning mitigation strategies have been studied including: elevation of fuel cell temperature thus increasing the rate of the water gas shift reaction, incorporating catalyst alloys designed to favour the water gas shift reaction, and bleeding O2 to oxidize the CO to CO2 thus liberating the catalytic site for hydrogen oxidation. The first of these strategies cannot be used with the low temperature range Nafion membrane and require special high temperature PE membranes [31]. The second of these CO-tolerant schemes involves the incorporation of Ru or Mo into the Pt catalyst and has been shown to allow CO concentrations up to 100 ppm without sacrificing output power [26, 32]. Both solutions are fixed at the design stage. The third of these three CO-poisoning mitigation strategies, however, allows a control system to respond to varying fuel conditions dynamically, creating a need to monitor the level of CO contamination [20]. Several equivalent circuit models have been fitted to EIS data obtained from fuel cells exposed to CO contamination [14, 28, 31, 33-36]. These models have been conceived to reflect the PEMFC behaviour while contaminated in order to understand the mechanisms for the performance degradation associated with CO. As a result, they incorporate an inductive element that models the behaviour of the PEMFC once the output voltage has dropped sufficiently for CO to be oxidized at the anode. This is not a valid operational condition8 and ideally, the CO contamination should be detected and corrected before reaching this stage. The modelling task for detection thus becomes more difficult as we cannot include the inductive behaviour in our diagnostic criteria. 8 Because the output voltage is negligible, power output drops to less than 50% full power. 22 The effect of CO contamination on the PEMFC equivalent circuit (Figure 2.4) is that Rct,a, Rct,c, Aa(j), Ac(j), τa, and τc all grow considerably compared to non-contaminated values. However, the increase in cathode parameters only occurs near open circuit. This is because once current is flowing, there is sufficient oxidant available to oxidize the CO and liberate the catalysis sites [28]. At the anode, on the other hand, it has been shown that the diffusion impedance Aa(j) grows in the first stages of contamination until the anode overpotential increases to such an extent that CO oxidation reaction (3.2) can proceed. At this point, the anodic relaxation time constant, τa, starts to grow along with the charge transfer resistance Rct,a [14]. Since we are concerned only with the stage of CO contamination associated with progressive exclusion of hydrogen from the catalysis sites, we focus on the change in the Aa(j) parameter. This data (Table 3.3) is limited to 1 significant figure as it is derived graphically from small figures in published reports. Nevertheless, the precision is more than adequate for our purposes. Table 3.3 Effects of Carbon Monoxide contamination on anode diffusion parameters % ∆ Aa(j) Investigator Notes 700 [14] A=23cm2, j=268mA·cm2, Pa=Pc=2bar, T=80°C, humidified anode, dry cathode, λa=dead end, λc=8, H2+100ppmCO:O2, Nafion 117, LPt=0.4mg·cm-2, Pt/C=20% 900 [28] A=5cm2, Vcell=0.65-0.82V, T=50°C, 90ml·min-1 2%CO+H2:Air, LPt=0.09mg·cm-2, Pt/C=35mg·cm-2 600 [33] A=1cm2, j=0mA·cm2 i.e., OCP, T=50°C, H2+100ppmCO:H2 i.e., just anode, Nafion 112, LPt=1.7mg·cm-2, Pt/Vulcan XC-72=20% 900 [31] A=5cm2, bias=150mV, T=80°C, RHa=RHc=100%, flow rate anode and cathode=80cm3·min-1, H2+100ppmCO:O2, NZTP membrane, LPt=0.5mg·cm-2, Pt/C=46.8% 500 [20] Hydrogenics series 500 stack (1 cell), j=400mA·cm2, Pa,Pc=121.3kPa, T=65°C, dry H2+100ppm CO & O2 gas feed, λa=1.2, λc=2, 23 10−2 10−1 100 Im pe da nc e (Ω ) a b c d e 10−3 10−2 10−1 100 101 102 103 104 −60 −50 −40 −30 −20 −10 0 Frequency (Hz) Ph as e (de gre es) Figure 3.6 Effect of CO contamination on PEMFC: Bode plot (a→e = normal→contaminated) 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0 0.05 0.1 Real(Z) (Ω) − Im ag (Z ) ( Ω ) a b c d e Figure 3.7 Effect of CO contamination on PEMFC: Nyquist plot (a→e = normal→contaminated) We use a 0 to 800% linear increase in Aa(i) to simulate progressive contamination with time for an approximately CO100ppm+H2 feed. Bode plots of spectra from CO contaminated PEMFCs are characterized by an increase in low frequency cell impedance magnitude and a decrease in phase angle (more capacitive) in the mid frequencies. Additionally, the Nyquist plot shows the characteristic growth of the anode low frequency 24 arc with the 45° slope at the left edge. The spectra shown in Figure 3.6 and Figure 3.7 are consistent with published results in the references in Table 3.3. In summary, three PEMFC failure modes (CO poisoning, dehydration, and flooding) have been described and the equivalent circuit consequences of each have been detailed. With this data and the equivalent circuit models developed in Chapter 2, the PEMFC can be simulated in the time domain to provide EI spectra with which to evaluate the proposed PEMFC diagnostic algorithms. 25 4 Fuel cell time domain simulation Fuel cell models can be largely divided into two groups. First, there are the theoretical, steady state models derived from fundamental thermodynamic, kinetic, and chemical considerations [8, 37, 38]. Second, we have the semi-empirical models that fit experimental parameters to an electronic equivalent circuit for the fuel cell; said equivalent circuit having been derived from an understanding of the underlying electrochemical processes as shown in section 2. Since we would like to simulate certain fuel cell fault states, the equivalent circuit consequences of which have been described in the literature, we have adopted the latter approach. In this chapter, we discuss how the equivalent circuit model developed in Chapter 2 was used to obtain time domain EIS voltage response data. This was accomplished with a state space approach based on the fractional order calculus. 4.1 Non-integer order system response Simulating the time domain convolution of the stimulus waveform and the fuel cell impedance transfer function is a trivial task if we ignore the Warburg impedance since the transfer function becomes purely integer order. If we include the Warburg impedance, however, we have a fractional order differential equation that must be solved using special techniques. While the fundamentals of the fractional calculus are beyond the scope of this work, several excellent references are available [39, 40] and the numeric approach used in this paper has been documented9 by other authors [9]. For the present, we limit ourselves to show the validity of the numerical fractional derivative using one frequency of the multi-frequency stimulus waveform; the function, sin(At). An analytical expression for the fractional order derivative of sin(At) is [Tseng, Miller] ( ) ( )sin sin 2 D At A Atα α pi α = + (4.1) 9 Owing to errors in the published account we have included the correct detailed derivation in Appendix 1. 26 wherein we have introduced the notation ( ) ( )( ) ( )( )dD f t f tdt α α α = with α ∈ℝ . Numerically, we can calculate the same result using the Grünwald-Letnikov (GL) derivative (see Appendix 1) ( ) ( ) ( ) ( )0 1 end start start end t t t t t k k D f t w f t k t t α α α − ∆ = ≈ − ∆ ∆ ∑ (4.2) where tstart and tend can be considered 0 and t respectively, x is the floor function and 0 1 1 0 11 1, 2,k k k w k w w w k k α α α αα − = = = + = − = ⋯ (4.3) Figure 4.1 through Figure 4.3 show the fractional order derivatives of sin(At) that form a continuum between the well known integer order derivatives ( )0 sin sinD At At= and ( )1 sin cosD At At= . The analytical and numeric approximations for several oversampling factors show that the steady state error in equation (4.2) can be made arbitrarily small as ∆t approaches 0. 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) α D α sin (A t) a n al yt ic 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) D α sin (at ) n u m er ic 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) D iff er en ce = a na ly tic − n um er ic Figure 4.1 Analytic and numeric solutions for Dα sin(x) with 2 times oversampling factor 27 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) α D α sin (A t) a n al yt ic 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) D α sin (at ) n u m er ic 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) D iff er en ce = a na ly tic − n um er ic Figure 4.2 Analytic and numeric solutions for Dα sin(x) with 10 times oversampling factor 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) α D α sin (A t) a n al yt ic 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) D α sin (at ) n u m er ic 0 1 2 3 40 0.5 1 −1 −0.5 0 0.5 1 time (pi units) D iff er en ce = a na ly tic − n um er ic Figure 4.3 Analytic and numeric solutions for Dα sin(x) with 100 times oversampling factor As discussed in Appendix 1, the GL derivative can be further manipulated to provide a state space, time domain simulation of the fuel cell voltage response to the current stimulus (see section 5.4 for an explanation of the current stimulus waveform) based on the equivalent circuit described in Figure 2.4. We consider the fractional order simulation software, written based on the development in Appendix 1, to have been correctly implemented because of three results: 1. The GL derivative, which forms the core of the software, has been compared numerically to analytical results for Dα sin(x) in Figure 4.1 through Figure 4.3. When the function being differentiated is sampled at high enough rates, the numeric approximation agrees closely with the analytic result. 2. Since commercial linear system simulation software (Matlab – lsim function) does not operate with non-integer order systems, we compare the GL derivative-based state space simulator against the Matlab lsim function for the integer order system 28 transfer function depicted in Figure 4.4. The results (Figure 4.5) show a zero mean RMS difference of less than 1%. Rct Cdl Figure 4.4 Equivalent circuit for integer order software comparison −5 0 5 10 Input waveform Cu rre nt (A ·cm − 2 ) 0 5 10 15 20 Time domain response without Warburg (lsimMatlab) V ol ta ge (m V) 0 5 10 15 20 Time domain response without Warburg (foSim) V ol ta ge (m V) 0 2 4 6 8 10 12 14 16 18 20 −40 −20 0 20 40 Difference between lsim and foSim V ol ta ge (µ V ) DifferenceRMS = 8.1204µV Time (msec) Figure 4.5 Comparison between commercial integer order simulation software and fractional order simulation software. 3. We cannot compare the simulated with the theoretical time domain voltage response for a fractional order system because we have no theoretical time domain 29 source for this information. Instead, we compare a proxy – the Fourier transform, with an analytical result. Since the Fourier transform of the signal is a unique frequency domain representation with exactly the same degrees of freedom as the time domain representation, the time and frequency representations can be considered identical in terms of information content. This test, then, is the most complete of the three tests in that it evaluates all the elements of the fractional order system simulation software rather than just a subset. Rct Cdl ZW Rs Figure 4.6 Equivalent circuit for non-integer order software comparison Figure 4.7 and Figure 4.8 show the analytical and simulated frequency responses of one electrode in series with the series membrane resistance shown in Figure 4.6 with the following parameter values: Rct = 8mΩ, A(j) = 5mΩ, Cdl = 0.2F, τ = 90msec, Rs = 4mΩ.. 0 2.5 5 7.5 10 12.5 15 17.5 20 0 2.5 5 Re(Z eq) (mΩ) − Im (Z eq ) ( mΩ ) analytical simulated Figure 4.7 Comparison between analytical non-integer order frequency response versus Fourier transform of fractional order simulation software time domain output: Nyquist plot. Both Figure 4.7 and Figure 4.8 show that the simulation software closely approximates the analytical frequency response for the given equivalent circuit. The largest error is manifest as a slight phase error in the lower frequencies and can be attributed to the start-up transient of the state space simulation kernel shown in Figure 4.1 through Figure 4.3. This effect is most evident for the lowest 30 frequencies because there are relatively fewer full cycles included in the simulation at these frequencies such that the effect becomes more noticeable. At other, higher frequencies, the start-up transient represents a negligible fraction of the total cycles available for Fourier analysis. The RMS error, again less than 1%, is not considered severe or even problematic since the absolute output of the simulated fuel cell is not relevant, but rather the change from one simulation instance to the next as the system parameters evolve that is of significance. 0.005 0.01 0.015 Im pe da nc e (Ω ) analytical simulated 10−3 10−2 10−1 100 101 102 103 104 −30 −25 −20 −15 −10 −5 0 Frequency (Hz) Ph as e (de gre es) Figure 4.8 Comparison between analytical non-integer order frequency response versus Fourier transform of fractional order simulation software time domain output: Bode plot. In summary, the state space time domain simulation of the linear system defined by the equivalent circuit shown in Figure 2.4 has been shown to produce results that are consistent with theory and a commercial software package (Matlab for integer order transfer function only). Additionally, it has been shown that by increasing the oversampling ratio, the non-integer order derivative approximation involved in this simulation can be made arbitrarily exact. 31 5 Electrochemical Impedance Spectroscopy Much of the elucidation of the electrochemical and physical mechanisms of PEMFCs and electrochemical systems in general (e.g., batteries, corrosion of metal, biological excitable cells, semiconductors) has taken place via the use of Electrical Impedance Spectroscopy (EIS) as an experimental tool [1, 2, 41-43]. The technique rests fundamentally on the notion that electrical phenomena associated with physical or electrochemical processes display the same characteristic time constants as the underlying processes. Because of this link between the electrochemistry and the electrical circuit behaviour of the system, electrochemical systems are often modelled with electrical circuit elements and can be investigated using electronic methods. Consider, for example, the electrochemical system modelled by the electrical circuit shown in Figure 2.4. The impedance of this circuit is a well-defined characteristic obtained from the Laplace transform of the differential equations relating the voltage and current through the circuit and solving for the ratio Zeq(s)=V(s)/I(s) (see Appendix 1 for detail) ( ) ( )( )( ) ( ) ( )( ) 1 11 12 22 2 1 11 12 22 2 , 0, 0, , 0, 0, , , 0, 0, , , 0, 0,1 1 ct a a a a ct c c c c eq s dl a ct a a a a dl c ct c c c c R s R s Z s R sC R s sC R s σ ω ω σ ω ω σ ω ω σ ω ω − − − − + + + + = + + + + + + + + (5.1) Since impedance is a function of the complex Laplace variable s=jω, impedance can be expressed in a rectangular ( )( ) ( )( )Re Z j Im Z jjω ω+ or polar ( ) ( )j Z jZ j e ωω ∠ coordinate system. The rectangular representation gives rise to the Cole-Cole plot (i.e., a complex plane or Nyquist plot with the imaginary axis inverted). Looking at solely the cathode impedance (Figure 2.9 and Figure 2.10), this plot does not show the frequency dependence explicitly, but the frequency grows from right to left. The polar representation leads to the Bode plot (Figure 2.7 and Figure 2.8), the magnitude and phase of the impedance are plotted against frequency thus showing the frequency dependence explicitly. Assuming that in this example, the equivalent circuit was, in fact, representative of the underlying physical process, we would be able to determine the equivalent circuit parameters from all of the above spectra. The details of how to do this are to be found in 32 any undergraduate electrical circuit analysis text and/or references [1, 2, 43] and shall not be repeated here since we are primarily concerned with the empirical matching of spectra rather than the significance of the spectra per se. Moreover, real EI spectra are functions of distributed circuit elements and sequential coupled electrochemical reaction steps rather than ideal elements (e.g., capacitor vs. constant phase element) and single activation time constants (e.g., oxidation vs. adsorption, dissociation then oxidation). Under these conditions, the equivalent circuit would be far more complex and the above-mentioned methods would fail to extract equivalent circuit parameters from the EI spectrum in a meaningful way. Instead, complex non-linear least squares fitting is employed, but neither technique is required for the approach discussed in this paper. An additional complication arises when obtaining equivalent circuit models from EIS data in that several different equivalent circuits can represent the same frequency response. Thus, any equivalent circuit representation may be ambiguous. Further, provided that enough component equivalent circuits are employed, the least squares approximation can be made arbitrarily exact. However, the resulting equivalent circuit would be devoid of physical meaning. This does not preclude the use of the equivalent circuit modeling approach, but rather demands that investigators develop a solid intuition for the real behaviour of the system based on physical models. Subsequently, EIS can help to assign numeric values to the parameters of such mathematical models based on physical understanding, but EIS cannot provide the physical understanding itself. The above describes the EIS technique in an idealized context. We have not yet considered how the above spectra are obtained in practice. This can be done in either the frequency or the time domain. 5.1 Obtaining the impedance spectra A single frequency sinusoidal voltage is applied to the system and the resulting current measured (potentiostatic EIS). Conversely, a current can be applied and the voltage measured (galvanostatic EIS). Either way, these measurements provide the data required to calculate the impedance according to the generalized Ohm’s law expression ( )( ) ( )V j Z j I jω ω ω= (5.2) 33 The inverse Fourier transform of a product in the frequency domain is a convolution in the time domain (5.3). In this representation, if the current stimulus were a perfect impulse, for which the Fourier transform is unity (i.e., contains all frequencies), the voltage term would be equivalent to the impedance term. ( ) ( ) ( )v t z i t dtτ τ∞ −∞ = ⋅ −∫ (5.3) So now we have the impedance represented in two domains such that we can evaluate it experimentally. In the frequency domain, we measure the ratio of magnitudes of the perturbation and response signals and their phase difference to obtain one impedance datum in the spectrum. This procedure must be repeated for each frequency for which impedance information is sought. Since all signal energy is concentrated into one discrete frequency while system and measurement noise is distributed continuously amongst all frequencies, narrow band measurement circuitry or phase sensitive detection can be used to obtain excellent quality data. Conversely, in the time domain, we apply a current impulse and obtain the voltage response (or vice versa). Subsequently, we divide the Fourier transform of the voltage response by the Fourier transform of the current stimulus (or vice versa) giving the impedance (or admittance) spectrum for all frequencies in one measurement. From an experimental standpoint it is attractive to adopt this latter approach since the entire spectrum is obtained very quickly. However, this is not generally done because of the practical impossibility of producing an ideal impulse. Additionally, the inclusion of all frequencies in the measurement makes noise rejection difficult. Both approaches are opposite extremes of the same continuum. As we reduce the number of included frequencies from infinity in the case of the impulse response (time domain) approach, we begin to concentrate signal energy into fewer and fewer frequencies while the noise power remains widely distributed among all frequencies. In the limit of this restriction, we have all of the signal energy contained in one single frequency as in the case of the frequency domain approach. Some point in between these two extremes provides an attractive compromise between noise rejection and efficiency of measurement. 34 This begs the question as to why not simply increase the stimulus signal power far beyond the noise level? The answer lies in a basic assumption inherent in transfer function analysis, linearity. Increasing the perturbation signal strength beyond the energy range of thermodynamic noise (25 mV) allows higher order harmonics to creep into the response signal thus corrupting the measurement [1, 2, 44]. Both frequency domain and time domain experiments can be carried out only as fast as the lowest frequency in the stimulus frequency set permits. As the frequencies of interest extend to lower frequencies, the EIS experiment takes longer to realize. Over a long experiment, the dynamic characteristics of the system under test can change thus violating another fundamental assumption of EIS, system time invariance or stationarity. Researchers have adopted several techniques to investigate non-stationary electrochemical systems that can be broadly classed into two groups: first, those that allow the system to evolve while taking measurements and correct for the evolution of the system later (e.g., time course interpolation) and second, techniques that preclude system evolution as much as possible by shortening the time required to obtain the EI spectrum (e.g., multi- frequency stimulus approach) thus allowing the assumption of local stationarity for the duration of the experiment. 5.2 Multi-frequency EIS techniques and noise Blanc et al. employed a multi-frequency stimulus approach by using a white noise input10 into an electrochemical system and cross correlating the output with the input. Since the spurious noise and the polarization signals are uncorrelated with the input, their contributions go to zero and the output is simply the correlation between the white noise and the system impulse response h(t). The impedance is then just the Fourier transform of the impulse response. They were able to reduce measurement time by one order of magnitude [45, 46], but EI spectra obtained in the time domain with a white noise stimulus signal have been shown to be inferior to the frequency domain EI Spectra in terms of noise rejection [47]. 10 Actually Blanc et al. used a pseudo random white noise generated by a binary sequence. An ideal white noise signal is unrealizable because it requires infinite power. 35 A key shortcoming of the approach put forward by Blanc et al. is that stimulus signal power is distributed continuously across all frequencies in the stimulus signal. For a given stimulus amplitude, the power spectral density is therefore minimal. A more effective alternative employs the superposition of k discrete sine waves distributed across the frequencies of interest (5.4). For the same stimulus amplitude, the spectral power density is now greater. At the same time, the noise power spectral density remains constant so the signal to noise ratio (SNR) improves. Additionally, the discrete distribution of signal energy in the frequency domain allows frequency selective detection algorithms to reject noise falling in the bandwidths between the discrete sine waves thus further improving SNR [48]. Noise performance tests similar to those demonstrating the superiority of the frequency domain over the time domain approach have demonstrated orders of magnitude improvement with a structured noise11 stimulus signal as compared to white noise [44, 49]. Finally, the structured noise perturbation signal allows for the reduction or elimination of harmonics of the lower frequencies from the stimulus discrete frequency set such that experimental signals observed at frequencies outside of this set can be either safely considered a product of the non-linearity of the system and rejected, or used as a metric of system non-linearity [44]. A rule of thumb is that perturbation frequency overlap with harmonics of lower frequencies should be avoided for at least n ≤ 3, where n is the integer multiple of the fundamental frequency [1]. ( ) ( ) , 1 sin 2 5 50 N o k k k k i t i f t kpi θ = = + ≤ ≤∑ (5.4) Further optimizations of the multi-frequency time domain EIS technique have been proposed to improve signal to noise ratio. First, for potentiostatic EIS, a response sensitive perturbation method is used to equalize measured signal power across all frequencies. This entails measuring the spectrum a first time, then applying the inverse spectral profile to the stimulus signal by adjusting the io,k parameters in (5.4) and measuring the spectrum a second time. The measured response current in the second 11 The term “structured noise” as opposed to “white noise” was coined by Popkirov when referring to the sum of a discrete set of sine waves as a perturbation signal. 36 experiment should display approximately equal power at each tested frequency. This improves the noise rejection capacity of the system by a factor of 4-5 [44]. Second, the phases, θk, of the individual sine waves (5.4) are adjusted such that the peak-to-peak sum is minimized. This allows the signal amplitude to be increased while still limiting the stimulus excursions into the non-linear region. A 25% improvement in SNR was obtained [49]. Some investigators have used a linear chirp signal as the stimulus waveform [50], but this has been shown to suffer from the same problem as that encountered when using white noise as the stimulus, namely, that the signal and noise occupy the same frequencies. The opportunity for noise rejection is therefore lost. The key contribution from these investigators was instead their exploration of several different data windows: Blackman, Hanning, Hamming, and Gaussian for the EIS experiment. Each of these windows reduces the effects of spectral leakage to some extent at the expense of resolution loss or spectral smearing [51]. 5.3 Multi-frequency EIS techniques and non-stationarity With the exception of the time domain correlation approach employed by Blanc et al., most multi-frequency methods utilize Fourier transformation of the stimulus and response waveforms to take calculation (5.3) from the time domain into the frequency domain (5.2). Linear, time-invariant systems form the backbone of Fourier analysis, but PEMFCs are neither linear nor time invariant [3, 5]. Because EIS experiments are kept short to minimize the impact of non-stationarity, these methods also create spectral leakage and smearing (loss of resolution) in the frequency domain. Spectral leakage and smearing both result in the appearance of signal energy at unexpected frequencies. We can see how the shortening of the stimulus signal in the time domain affects the frequency domain by looking at the truncation of the infinite sum that makes up the Discrete Time Fourier Transform (DTFT). This operation is described mathematically as ( ) ( ) [ ) [ ), 0,2 or ,j j n n X e x n eω ω ω pi ω pi pi +∞ − =−∞ = ∈ ∈ − +∑ (5.5) The frequency variable ω actually ranges from -∞ to +∞, but since the frequency-domain spectrum is periodic with a period of 2π, only the unique range is relevant. When the 37 discrete time domain signal is not of infinite length, we are forced to truncate the summation to accommodate the reduced degrees of freedom offered by the length N input sequence. This gives the Discrete Fourier Transform (DFT). The DFT is defined as ( ) ( ) 010 0 0 2 , 0,1, 2, , 1 N jnk n X k x n e where k N and N ω piω ω − − = = = − =∑ ⋯ (5.6) for an length N discrete time series. In the limit, as N approaches infinity, the DFT approaches the DTFT. The DFT is typically calculated using the fast Fourier transform (FFT) originally developed by Tukey and Cooley in 1956 [52]. s[n] Ti m e do m ai n S(f) Fr eq ue nc y do m ai n W(f) w[n] w[n] V(f) × * = = Figure 5.1 The frequency domain effect of time domain signal truncation The limitations of the DFT (or FFT) begin to appear when we have limited data from which to make our spectral estimate. For example, for an infinite complex sinusoidal signal, the DTFT shows a single peak with no signal energy anywhere else in the spectrum (Figure 5.1). Now, if we multiply the time domain signal with a rectangular window we must convolve the respective Fourier transforms in the frequency domain. This example illustrates the origins of spectral smearing and spectral leakage when using the DFT with real data and, in particular, with short time series. Spectral leakage is attributable to the sidelobes of the sinc pulse in Figure 5.1. These sidelobes introduce signal energy into adjacent frequency bins. Spectral smearing is a result of the width of the main lobe in that signal energy that is supposed to be restricted exclusively to fo, 38 begins to spread out or “smear” across adjacent frequencies as the signal duration decreases. This is a classical DSP result: one cannot obtain arbitrary precision in both the frequency and time domains. These two domains are inextricably linked via Heisenberg’s inequality such that the minimum time-bandwidth product is always greater than or equal to unity. The net effect of shortening the stimulus signal to accommodate non-stationary systems is that it creates imprecision in the impedance estimate via the two mechanisms discussed above. Depending on what is more important to the investigation, sidelobe height or main lobe width can be arbitrarily manipulated via the use of the appropriate window (Hanning, Hamming, Kaiser, Gaussian, Blackman, etc.) However, decreasing sidelobe height increases main lobe width (decreases resolution) and vice versa. While Popkirov and Schindler used sequential, rectangular data windows in their investigations, Wiegand et al. make use of sliding windows (i.e. Welch periodogram method) to reduce the variation in the spectral estimate by overlapping successive data sections together with a wavelet transform12 rather than a Fourier transform in their biological EIS investigations [42]. 5.4 Non-linearity and the multi-frequency stimulus waveform In both FRA-based and multi-frequency EIS experiments, non-linear system dynamics can cause stimulus signal energy to appear at harmonics of the lower frequencies. Typically, the amount of signal energy is negligible after the third harmonic [1]. For this reason, the multi-frequency stimulus waveform should exclude the first three harmonics. Another cause of non-linear behaviour in both FRA-based and multi-frequency EIS experiments is amplitude of the stimulus beyond 25mV. It is, therefore, desirable to 12 The wavelet transform is particularly suitable for this application (patch clamp of ionic channels in biological tissue) because the events are discrete and hence time resolution is required and possible. The fuel cell EIS application is analogous to voltage clamp experiments of whole muscle tissue where we are not looking at one channel but a multitude of them. Instead of using individual channel analysis methods, then, we employ ensemble approach that necessarily leads us to the Fourier transform. However, investigating a “single channel” fuel cell using the same approach could generate considerable useful mechanistic information. 39 minimize the peak-to-peak amplitude of the multi-frequency signal while maximizing the power of the stimulus signal at each frequency to maximize the signal to noise ratio. Optimization methods used to adjust the phases such that the overall voltage envelope remains less than 25mV include Monte Carlo and the downhill simplex method [49]. Closed form, analytical approaches to selecting the phase of each frequency component have been explored, but only for harmonically related signals [53]. In this investigation, the phases are not optimized because the system under test is purely linear. However, it is important to note that the techniques discussed herein are not dependent on the system being linear. In summary, EIS theory has been described and the experimental methods for obtaining EIS data detailed. Of the various ways of obtaining EIS data, multi-frequency EIS has been shown to be the most appropriate for investigation of non-stationary systems and methodologies for increasing SNR and controlling non-linearity in the EIS measurement were discussed. 40 6 Goertzel algorithm based spectral analysis13 To implement impedance spectral analysis for diagnostic or control purposes on-board the operating fuel cell, several objectives must be achieved: • The impedance spectrum of the time-variant electrochemical system must be obtained with sufficient time resolution to capture the sought after electrochemical events or conditions. • The impedance spectrum computation burden must be minimized. • The impedance spectrum computation must be robust in the face of noise. • The impedance spectrum measurement must encompass a wide frequency range to accommodate the wide distribution of time constants involved in PEMFC electrochemical mechanisms. One alternative for obtaining the impedance spectra of time-variant electrochemical systems is the short term Fourier transform (STFT) as discussed in chapter 5 and applied for many years to electrochemical systems by, amongst others, Blanc, Popkirov, and Darowicki. The shortcomings of this approach are threefold: 1. There are considerable redundant calculations, particularly if the upper frequency limit is large. 2. Spectral leakage and smearing exacerbate the non-linearity problem. 3. The system under test is time variant and the STFT assumes time-invariance. The Goertzel algorithm has been used for many years in digital telephony for dual tone multi-frequency (DTMF) detection. A bank of filters based on the Goertzel approach could replace the STFT and improve upon the first two of the three of the STFT limitations. 6.1 The basic Goertzel algorithm Starting with the definition of the DFT 13 A version of this chapter will be submitted for publication. 41 ( ) ( ) 2 1 0 0,1,2 1 N N km N m j N X k x m W k N W e pi − = − = = − = ∑ … we multiply by 2 2 1 kN NjkN j k NW e e pi pi−− − = = = to give ( ) ( ) ( ) ( ) ( )1 1 1 0 0 0 N N N k N mkN km km kN N N N N m m m X k W x m W x m W x m W − − − − − − − = = = = = =∑ ∑ ∑ which results in an expression that has the same form as the convolution ( ) ( ) ( )1 0 N k n m k N m y n x m W − − − = = ∑ of the finite data sequence x(m) and the FIR filter impulse response ( ) knk Nh n W −= While the yk(n) sequence is defined for all n, we are only interested in the filter output when n=N such that ( ) ( )k n NX k y n == The z-transform of the filter hk(n) is ( ) 0 1 1 1 kn n k N n k N k N H z W z W z z z W ∞ − − = − − − = = − = − ∑ Which is a first order infinite impulse response (IIR) filter with a complex pole, a zero at the origin, and a frequency response as shown in Figure 6.1. Owing to the IIR nature of the filter, the Goertzel frequency response displays a sharp peak, rapid roll off, and negligible sidelobes that result in less spectral smearing and leakage relative to the FIR filter (DFT or FFT). 42 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 −100 −75 −50 −25 0 Normalized frequency (×pi rad/sample) M ag ni tu de (d B) Magnitude −90 −45 0 45 90 Ph as e (de gre es) Phase −1 0 1 −1 0 1 Real(z) Im ag (z) Figure 6.1 Frequency response and z-plane pole zero location: basic Goertzel algorithm We can implement this filter with the following IIR structure x(n) yk(n) pi k -j2 Ne Z-1 + Figure 6.2 Realization of basic Goertzel algorithm This structure highlights an important benefit of the Goertzel algorithm, namely: that data is processed as it becomes available. There is no batch processing as in the FFT. The lag time between sampling end and result availability is therefore reduced. Now, if we multiply by the complex conjugate pole, we create a second order IIR filter with purely real coefficients in the denominator. ( ) ( )( ) ( ) 1 1 1 1 1 1 1 0 21 1 2 21 1 1 22 2 1 0 2 1 1 11 1 1 11 1 1 1 1 2cos 2 1 2 2 k k N N k k N N k k k N N N k k k k kk k N N N N NN N j j kj j N N W z W z W zH z W z W z W z W z W zW z W z e z e z z ze e z W z pi pi pi pi pi − − − − − − − − − − − − − − − − − − − − − − − − − = = = − − − − + − − − − = = − + + − + 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 −100 −75 −50 −25 0 Normalized frequency (×pi rad/sample) M ag ni tu de (d B) Magnitude −90 −45 0 45 90 Ph as e (de gre es) Phase −1 0 1 −1 0 1 Real(z) Im ag (z) 43 Figure 6.3 Frequency response and z-plane pole zero location: modified Goertzel algorithm The magnitude frequency response is identical to the original Goertzel filter because the extra pole and zero cancel out. This z-plane transfer function can be represented as a discrete difference equation ( ) ( ) ( ) ( ) ( ) ( ) ( )2 2cos 2 1 2 1 pi pi − = − − − + = − − k k k kj N k k k k v n v n v n x n N y n v n e v n which can be implemented using the realization shown in Figure 6.2. vk(n) is calculated for every n, but yk(n) is not calculated for n≠N; only for n=N. Thus we have a largely real calculation with one complex multiplication and addition at the end. + vk(n)x(n) yk(n) pi− 2 -e kj Npi k2cos 2 N -1 + + Z-1 Z-1 Figure 6.4 Realization of modified Goertzel algorithm 6.2 Computation burden (memory and operations) The FFT operation involves ( ) 2log2N N complex multiplications and 2logN N complex additions [51]. By contrast, the modified Goertzel algorithm requires only 1N + multiplications and 2 1N + additions per frequency bin. Further, all but one multiplication and one addition in the Goertzel algorithm are real. Comparing the FFT and Modified Goertzel algorithms in terms of real multiplications (Figure 6.5), we note that the Goertzel algorithm is more convenient than the FFT if the number of frequency bins 22logM N< in terms of multiplication effort and 21.5logM N< in terms of addition effort. 44 In terms of memory requirements, the Goertzel algorithm requires far less storage of coefficients compared to the FFT (Table 6.1) since the FFT memory requirement grows linearly with N while the Goertzel algorithm grows logarithmically with N. Table 6.1 Computational and memory requirements of FFT and Goertzel algorithms Real multiplications Real additions Memory (# coefficients) FFT 22 logN N 23 logN N N Modified Goertzel* 4N + 2 4N + 2M * per frequency bin (M) 100 105 1010 R ea l m ul itl pl ic at io ns 100 101 102 100 105 1010 Frequency bins R ea l a dd iti on s FFT Goertzel Goertzel with decimation Figure 6.5 Computational requirements of FFT and modified Goertzel algorithms for N = 4×106 6.3 Goertzel filter bank The DFT size, N, discrete frequency bin, k, sampling interval, T, and Goertzel filter passband centre frequency, fk, are related as k s k kf f NT N = = (6.1) 45 Using (6.1), the Goertzel algorithm can be used as described to determine the fuel cell system response to the stimulus waveform at each frequency of the stimulus waveform. For example, a 10 second signal sample with a 25kHz maximum stimulus frequency results in 5×104 discrete samples if we sample at exactly the Nyquist frequency. In practice, the signal is usually oversampled by several times to ease the anti-aliasing filter requirements making N even larger. Nevertheless, for these simple sampling conditions, the Goertzel algorithm based analysis can incorporate up to M = 37 discrete frequencies and still be computationally simpler than the FFT approach. This gives rise to the analysis system shown in Figure 6.6 for each of the current stimulus and voltage response signals. 1 1 = s k f N F x(n) Goertzel filter bank X(f1) X(f2) X(f3) X(f4) X(fM) 2 2 = s k f N F 3 3 = s k f N F 4 4 = s k f N F = M M s k f N F … Figure 6.6 Implementation of spectral analysis with Goertzel filter bank Using the foregoing approach with M = 15, we compare the FFT-based impedance versus the Goertzel-based impedance in Figure 6.8. In both cases, the system under test is the simple equivalent circuit of Figure 4.4. While the FFT data is shown only at the test frequencies, the FFT algorithm makes all the unnecessary calculations to obtain intermediate frequencies regardless as shown in Figure 6.5. The FFT and Goertzel spectral results are identical to the theoretical impedance over the tested frequency range. In addition, noise performance of the Goertzel algorithm is identical to the FFT. Figure 6.9 shows impedance spectral estimate error using the Goertzel algorithm decreasing at the same rate with increasing SNR as for the FFT derived impedance spectral estimate error. 46 −0.4 −0.2 0 0.2 0.4 Cu rre nt st im ul us (m A) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −10 −5 0 5 10 15 V ol ta ge re sp on se (m V) Time (sec) Figure 6.7 Time domain representation of current stimulus and voltage response signals 100 101 102 103 104 −100 −80 −60 −40 −20 0 Ph as e (de gre es) Frequency (Hz) −30 −20 −10 0 10 20 30 40 Im pe da nc e (dB Ω ) Theoretical FFT Goertzel Figure 6.8 Impedance calculated via FFT and Goertzel algorithms 47 0 2 4 6 8 M ag ni tu de e rro r ( mΩ ) FFT Goertzel −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 0.1 0.2 0.3 0.4 0.5 Ph as e er ro r ( de gre es) Signal to Noise Ratio (dB) Figure 6.9 Noise performance of FFT and Goertzel algorithms Recognizing that the lower frequencies of the multi-frequency stimulus waveform and corresponding PEMFC response are oversampled allows us to further enhance the computational efficiency of the Goertzel algorithm. If we decimate the signal prior to processing (Figure 6.10), we obtain a further savings as shown in Figure 6.5. The highest frequency in the multi-frequency signal is used directly without decimation while the remainder are decimated in relation to their respective oversampling ratio. In this way, the most heavily decimated frequencies are the lowest frequencies and consequently also enjoy the greatest computational savings. This approach is not recommended, however, for systems wherein the noise considerations are significant because the impedance spectral estimate error using the Goertzel algorithm increases considerably with decimation unless the SNR is high (Figure 6.11). There are three improvements or adjustments to be made to the described experimental protocol to reduce the variance in the Goertzel spectral estimate: First, the classical result in Fourier analysis that signal blocks of longer duration reduce variance by allowing the zero-mean noise to average out in the FIR convolution [51]. While not shown explicitly, the same is true in the case of the Goertzel algorithm. This implies that experiments must 48 be designed with sufficiently short durations to capture the non-stationary behavioural features of the PEMFC, but no shorter. Second, phase optimization as discussed in Section 5.4 allows maximal stimulus signal strength to be employed at each frequency without entering the non-linear regime. Third, response sensitive perturbation as discussed in Section 5.2 can be used to concentrate stimulus signal energy in those frequencies where the response signal is weak [49]. These three approaches are not implemented in this study so as to make the comparison between the FFT and Goertzel derived spectral estimates as transparent as possible and focused solely on the choice of underlying algorithm. 1 1 = s k f N F x(n)Decimated Goertzel filter bank X(f1) X(f2) X(f3) X(f4) X(fM) 2 2 = s k f N F 3 3 = s k f N F 4 4 = s k f N F = M M s k f N F … 1D↓ 2D↓ 3D↓ 4D↓ Figure 6.10 Implementation of spectral analysis with Decimated Goertzel filter bank 49 0 100 200 300 400 500 M ag ni tu de e rro r ( mΩ ) D1 − No decimation D2 D3 D4 D5 − Max decimation −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 0 2 4 6 8 10 Ph as e er ro r ( de gre es) Signal to Noise Ratio (dB) Figure 6.11 Noise performance of Decimated Goertzel algorithm In summary, the Goertzel algorithm was described and shown to be uniquely applicable to the acquisition of EIS data in a non-stationary PEMFC diagnostic application. In particular, it was shown that the Goertzel algorithm requires orders of magnitude less computation and negligible memory for the time to frequency domain conversion as compared to the equivalent FFT based approach. In respect of accuracy, and noise performance, the Goertzel algorithm was shown to be identical to the FFT. Decimation of the time domain signal to further reduce the computation burden was shown to be not advisable except under the rather unrealistic conditions of minimal system noise. 50 7 Spectral fingerprint-based diagnostics14 Impedance data can be processed within a multi-dimensional control space to determine the state of health of the operating fuel cell. That is, the impedance measurement can be used as a global sensor, augmenting or replacing several humidity, temperature, pressure, flow, concentration, etc. sensors. Other authors have proposed using the impedance at a few select frequencies with an heuristic (degrees of freedom = 2) [15] or fuzzy logic (degrees of freedom = 3) [54] algorithm or complex non-linear least squares equivalent circuit fitting (degrees of freedom = 4) [21] to distinguish between a finite set of fault conditions (Figure 7.1). However, other fault conditions may mimic one or more of the fault conditions from this finite set owing to the limited degrees of freedom of the test strategy. Multi- frequency AC stimulus FCUT STFT Selected frequencies Equivalent circuit fitting Diagnostic or control algorithm Diagnostic or control algorithm Rm, Rct, Cdl, ZW, etc. v(t) Z(f) Z(f1), Z(f2), ...Z(fN<3) i(t) Window v(t)·w(t) Figure 7.1 Non-FRA-based PEMFC diagnostics It is hypothesized that each fuel cell fault condition will give rise to a unique impedance spectrum that identifies it in much the same way as a fingerprint does a human being if we create a test strategy with degrees of freedom matched to the complexity of the queried fault condition. A spectral fingerprint concept with potentially unlimited degrees of freedom and lesser computation burden than complex non-linear least squares (CNLS) equivalent circuit fitting is therefore proposed and evaluated to further enhance the diagnostic ability of impedance spectroscopy in operating fuel cells. For example, a simple test for proper hydration has been devised based on the uniqueness of spectral characteristics associated with the dehydration and flooding failure modes. 14 A version of this chapter will be submitted for publication. 51 Impedance is calculated at a discrete frequency in each of two frequency bands: 10-100Hz and 1-10kHz and the changes in each band, relative to a reference impedance spectrum allow discrimination between one failure mode and another [15, 54]. This type of data has, in part, formed the basis of a PEMFC fuzzy logic control system. The control system devised in this study monitored system impedance at 1 kHz for signs of growth indicating dehydration, while at the same time monitoring individual cell voltages for abrupt voltage drops indicating flooding. The fuzzy logic control algorithm achieved the desired result, namely: using the airflow fan speed to control the humidification of the membrane in response to varying power demands [54]. Other researchers have found similar results supporting the diagnostic strategy proposed above, although without incorporating them into a control algorithm [13, 17, 20]. Despite the fact that the simple measurement of cell voltage was used as the sensor in the Schumacher study, there is still considerable benefit to using EIS to detect PEMFC flooding since it has been shown that the onset of flooding is accompanied by a small change (<21%) in output voltage while equivalent circuit element values associated with low frequency impedance increase nine fold [21] over the same interval. Because the long-term goal is to control all aspects of PEMFC operation and not solely humidification, the simplified approaches proposed by Merida, Schumacher, Fouquet, and others may not provide sufficient data for complete observability. Specifically, the use of few (M ≤ 4) frequencies to define the failure mode fingerprint does not provide sufficient degrees of freedom when multiple failure modes need to be evaluated and in particular, when spectral fingerprints are very similar. According to Table 6.1, we could use up to 37 discrete frequencies in our fingerprint-matching scheme before the computational savings of the Goertzel algorithm over the FFT is lost. In order to guarantee the uniqueness of the EIS fingerprint for each failure mode in a PEMFC diagnostic or control system, we need to use a large enough discrete frequency set to obtain sufficient impedance spectral information to define the shape of the spectrum in the frequency domain without excluding features necessary for fingerprint identification. For the failure modes and stimulus durations examined in this study, this lower limit has been set by inspection. For example, in Figure 7.2, spectral features begin to disappear if the number of discrete frequencies is less than 15. As a 52 worst-case scenario, therefore, 15 discrete frequencies were used for all fingerprint- matching tests in this investigation. 0 5 10 15 20 25 30 35 40 0 2 4 6 8 10 Re(Zeq) (mΩ) Im (Z eq ) ( mΩ ) Mfreq = 100 Mfreq = 32 Mfreq = 15 Mfreq = 6 Figure 7.2 Discrete frequency points requirement The overall fingerprint detection scheme, then, is shown in Figure 7.3. The multi- frequency AC stimulus is applied to the PEMFC model as discussed in chapters 4 and 5, the resulting response signal is converted into the frequency domain with the Goertzel algorithm as discussed in chapter 6 and the spectral fingerprint matching is effected by calculating the minimum Euclidean distance between the experimental EI spectrum and the reference spectra (fingerprints). Multi- frequency AC stimulus FCUT Goertzel Fingerprint matching Diagnostic or control algorithm v(t) Z(fk)|k=1,2,3...15 FDH, FFL, FCOi(t) Window v(t)·w(t) Figure 7.3 Non FRA based PEMFC diagnostics using Goertzel and EIS fingerprint recognition For the discrete experimental spectra ( ) ( ) ( )1 2, ,expt expt expt MZ j Z j Z jω ω ω ⋯ and the reference spectra ( ) ( ) ( )1 2, ,ref ref ref MZ j Z j Z jω ω ω ⋯ we have ( ) ( ) 1,2,3Z expt k ref kd Z j Z j k Mω ω= − = … (7.1) where the subscript ref refers to the CO poisoning, flooding, and dehydration reference spectra. The minimum dZ corresponds to the matched fingerprint. The identified failure mode is then used in the diagnostic or control routines as defined by the system designer. In the following figures, dZ is inverted and normalized resulting in a peak fingerprint 53 matching metric coinciding with the minimum Euclidean distance for ease of visual interpretation. The fingerprint-matching algorithm need not carry out this step. Table 7.1 Computational and memory requirements of spectral fingerprint diagnostic algorithm Real multiplications Real additions/subtractions Memory ( )( )2 1 1M F S − + ( ) ( )( )3 1 1 1M F S− − + ( )( )2 1 1M F S − + 390 572 390 The computational cost of this analysis is slight. The operations involved are shown in Table 7.1 together with the actual number of calculations for the example discussed in section 7.1 where M is the number of discrete frequencies used in the spectral fingerprint, F is the number of failure modes, and S is the number of distinct states into which each failure mode has quantized. It is evident that the cost of this algorithm is directly proportional to M, F, and S and these, in turn, are proportional to the fineness of the spectral fingerprint database, and the sampling rate of the frequency spectrum. Regardless, the computational burden is negligible when compared against that of the conversion from the time to frequency domain (Table 6.1). The memory requirements are also insignificant as they consume only a very small part of the memory savings of the Goertzel algorithm over the FFT based conversion from the time to frequency domain. 7.1 Spectral fingerprint evolution To demonstrate that the identified spectral fingerprints can be detected, we first demonstrate their uniqueness. A series of fingerprints at different stages of CO poisoning, dehydration, or flooding have been developed as described in chapters 3 and 4. These are shown in Figure 7.4 through Figure 7.9 as filled circles. While there is a slight numeric error in the lower frequencies as compared to the theoretical result, this is associated with the PEMFC simulation (section 4.1) and not the transformation into the frequency domain with the Goertzel algorithm (section 6.3). The important result is that different stages of PEMFC failure can be distinguished. As Figure 7.4 through Figure 7.9 show, even in the face of significant noise (SNR = 0dB), this is indeed the case. Note that the zero mean signal noise is mostly eliminated by the averaging effect of the long convolution that is inherent in the Goertzel algorithm. It is only when the SNR is less than –1.5dB that we see a significant effect on the spectral fingerprint. 54 0 50 100 150 200 250 300 350 0 50 100 150 Re(Zeq) (mΩ) Im (Z eq ) ( mΩ ) Increasing CO poisoning theory fingerprint experimental Figure 7.4 Five spectral fingerprints at increasing stages of CO PEMFC poisoning (SNR=0dB) 0 50 100 150 200 250 300 350 0 50 100 150 Re(Zeq) (mΩ) Im (Z eq ) ( mΩ ) Increasing CO poisoning theory fingerprint experimental Figure 7.5 Five spectral fingerprints at increasing stages of CO PEMFC poisoning (SNR=–1.5dB) CO poisoning manifests as a large increase in the low frequency impedance arc. As discussed in section 3.3, if the fault condition were allowed to persist for longer periods, the impedance would begin to curve below the real axis. The objective however, is to detect and correct the fault condition prior to this extreme level of failure. We limit ourselves, therefore, to the early onset spectral fingerprint for the detection exercise. 55 0 5 10 15 20 25 30 35 40 45 50 0 2 4 6 8 10 Re(Zeq) (mΩ) Im (Z eq ) ( mΩ ) Increasing Dehydration theory fingerprint experimental Figure 7.6 Five spectral fingerprints at increasing stages of PEMFC dehydration (SNR=0dB) 0 5 10 15 20 25 30 35 40 45 50 0 2 4 6 8 10 Re(Zeq) (mΩ) Im (Z eq ) ( mΩ ) Increasing Dehydration theory fingerprint experimental Figure 7.7 Five spectral fingerprints at increasing stages of PEMFC dehydration (SNR=–1.3dB) Dehydration failure is characterized by a shift in the direction of the positive real axis (Figure 7.6 and Figure 7.7). This is the result of constriction of the PE membrane proton channels giving rise to a purely real increase in impedance without affecting any of the reactive elements. Different dehydration stages are self-similar as a result of the translational spectral fingerprint evolution. Consequently, we expect it to be more difficult to resolve between adjacent stages of dehydration as compared to failure modes for which the spectral fingerprint undergoes translation and scaling. The large increase in extremely low frequency impedance predicted by the theoretical spectral fingerprint is not manifest in the experimental flooding spectral fingerprint because the sample was too short. For longer time samples, we could obtain greater low frequency information, but the high frequency arc still provides sufficient data to distinguish one state from another. 56 0 50 100 150 0 10 20 30 40 Re(Zeq) (mΩ) Im (Z eq ) ( mΩ ) Increasing Flooding theory fingerprint experimental Figure 7.8 Five spectral fingerprints at increasing stages of PEMFC flooding (SNR=0dB) 0 50 100 150 0 10 20 30 40 Re(Zeq) (mΩ) Im (Z eq ) ( mΩ ) Increasing Flooding theory fingerprint experimental Figure 7.9 Five spectral fingerprints at increasing stages of PEMFC flooding (SNR=–1.5dB) The experimental results (open circles) are virtually identical with fingerprint data for SNR greater than or equal to 0 (Figure 7.4 c.f. Figure 7.5). A simple examination of Figure 7.5, Figure 7.7, and Figure 7.9 shows that distinguishing one fingerprint from another is quite difficult when SNR drops below –1.3dB. 7.2 Failure mode identification via spectral fingerprinting Using the simple Euclidean distance fingerprint recognition algorithm. (7.1), we can show that each of these fingerprints is distinguishable numerically from the rest within reasonable limits of noise contamination (Figure 7.11, Figure 7.13, and Figure 7.15). Once system noise exceeds EIS signal strength, however, detection becomes increasingly difficult (Figure 7.12, Figure 7.14, and Figure 7.16). 57 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 Stage of failure Fi ng er pr in t m at ch Figure 7.10 Spectral fingerprint matching during normal PEMFC operation (SNR=0dB) There is some correlation between the normal fingerprint and the dehydration fingerprint evident in Figure 7.10 as well as Figure 7.13. This is a result of the simple translation of the fingerprint without scaling or rotation. In essence, we have the same fingerprint, but in a different location. This makes detection of different stages of dehydration more ambiguous, but does not affect the ability to distinguish between dehydration and other failure modes. 58 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 a Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 b Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 c Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 d Stage of failure Fi ng er pr in t m at ch Figure 7.11 Spectral fingerprint matching during PEMFC CO poisoning (SNR=0dB) When system SNR is greater than or equal to 0dB, Figure 7.11 shows mild (panel a), moderate (panel b), advanced (panel c), or extreme (panel d) CO poisoning being easily and unambiguously detected. When SNR drops to –1.5dB, the early stages of CO poisoning develop a similarity with advanced or extreme flooding (Figure 7.12). Later stages of CO poisoning are quite dissimilar to all stages of flooding so the difficulty resolves itself. Further, the diagnostic algorithm must also consider that flooding is unlikely to progress from none to extreme without passing through mild, moderate, and advanced. 59 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 a Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 b Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 c Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 d Stage of failure Fi ng er pr in t m at ch Figure 7.12 Spectral fingerprint matching during PEMFC CO poisoning (SNR=-1.5dB) 60 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 a Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 b Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 c Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 d Stage of failure Fi ng er pr in t m at ch Figure 7.13 Spectral fingerprint matching during PEMFC Dehydration (SNR=0dB) Early stages of dehydration are very similar to normal operation (Figure 7.13 and Figure 7.14) as can be seen by the elevated normal “fence” along the x-axis. This problem is particularly difficult under noisier conditions (Figure 7.14). As the dehydration progresses however, the similarity decreases and the detection becomes more reliable. 61 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 a Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 b Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 c Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 d Stage of failure Fi ng er pr in t m at ch Figure 7.14 Spectral fingerprint matching during PEMFC Dehydration (SNR=-1.3dB) 62 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 a Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 b Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 c Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 d Stage of failure Fi ng er pr in t m at ch Figure 7.15 Spectral fingerprint matching during PEMFC Flooding (SNR=0dB) When system SNR is greater than or equal to 0dB, Figure 7.15 shows the various stages of flooding being easily and unambiguously detected. As for CO poisoning and dehydration, though, when SNR drops to –1.5dB, the latter stages of flooding develop a similarity with early CO poisoning (Figure 7.16). In contrast to the CO poisoning ambiguity, it is not unreasonable for the diagnostic algorithm to conclude that early stage CO poisoning is occurring coincident with latter stage flooding. This failure mode could, therefore mask an incipient CO poisoning problem. 63 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 a Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 b Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 c Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 d Stage of failure Fi ng er pr in t m at ch Figure 7.16 Spectral fingerprint matching during PEMFC Flooding (SNR=-1.5dB) Figure 7.10 through Figure 7.16 show data resulting from trials where the experimental PEMFC state coincides with a fingerprint database entry. This is an optimal situation and the PEMFC cannot be relied upon to behave as conveniently in practice. However, the algorithm still produces valid results when the experimental PEMFC state falls between fingerprint database entries. 64 CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 a Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 b Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 c Stage of failure Fi ng er pr in t m at ch CO poisoning Dehydration Flooding none mildmoderate advancedextreme 0 0.5 1 d Stage of failure Fi ng er pr in t m at ch Figure 7.17 Spectral fingerprint matching during PEMFC Dehydration (SNR=0dB) Figure 7.17 shows trials conducted with PEMFC membrane resistance varying between moderate and advanced dehydration. Panel a represents moderate dehydration, panel b represents ¼ the evolution between moderate and advanced dehydration, panel c represents ½ the distance between moderate and advanced dehydration, and panel d represents ¾ the distance between moderate dehydration and advanced dehydration. In all cases except ½ the distance between moderate dehydration and advanced dehydration, the algorithm selects the nearest state. When exactly ½ the distance between moderate dehydration and advanced dehydration occurs, the algorithm appears to conclude that moderate and advanced states are equally close. However, this is an artefact of the imprecise nature of the 3D graph, and the minimum Euclidean distance in this case was indeed determined to be to moderate dehydration (noise pushed the fingerprint away from 65 the exact midpoint and marginally in the moderate direction). As the failure mode spectral fingerprint database is made finer (e.g., 10 steps between no failure and extreme failure instead of 5), the precision of state detection increases, but computation also grows as shown in Table 7.1. This trade-off must be evaluated based on the intended usage of the detection algorithm results. It is worthwhile to note that this method is inherently empirical in that the underlying mechanism of the evolution of the failure mode is not relevant to its detection. It is sufficient that we can replicate the failure mode during system configuration so as to be able to recognize it when it occurs during normal operation. Additionally, experimental lead inductance and parasitic capacitance often create EIS spectral features not associated with the PEMFC internal operation. However, with the present detection scheme, as long as the measuring setup remains fixed, the only changing element is the PEMFC. And we can map this and subsequently detect the PEMFC changes. In summary, we have introduced a method for diagnosing PEMFC failure modes based on their EI spectra and a database of spectral fingerprints. The method has been shown to be viable in the presence of considerable system noise (SNR ≥ -1dB) and quite easy to implement with little computational cost beyond the initial time to frequency domain conversion. With the exception of the challenges discussed in Table 7.2 that only become significant under extremely noisy conditions the spectral fingerprint approach has been shown to be accurate and reliable. Table 7.2 Experimental failure mode detection challenges CO poisoning Dehydration Flooding Early stages easily confused with extreme flooding when SNR low Evolution of failure mode is simple translation; no rotation or scaling – therefore difficult to determine failure mode level exactly when SNR low Extreme cases easily confused with early onset CO poisoning when SNR low 66 8 Conclusions Three common and important failure modes for PEMFCs, CO poisoning, dehydration, and flooding, were characterized from the literature and modelled using a two-electrode Randles approach. This model provided quick and easy access to impedance spectra allowing simple and expedient experimentation with digital signal processing algorithms for PEMFC failure mode detection. The model behaviour closely mimics real PEMFC behaviour reported in the literature. The PEMFC model was used in a state space simulation to provide time domain EI spectroscopy experimental data. Frequency domain transformations of the time domain data provided by the state space approach agreed closely enough with the theoretical PEMFC frequency response to permit evaluation of a novel pattern recognition-based failure mode detection algorithm. Both the PEMFC model and the state space EIS simulation have usefulness beyond this investigation and can be re-used for other investigations where rapid analysis of PEMFC behaviour based on published equivalent circuit parameters is required. The ultimate goal of this work is to develop an EIS-based state of health sensor for operating fuel cells in which cost, power, and complexity are key design criteria. As a result, an important objective of this investigation was to minimize the computational burden so that small, inexpensive, low power consumption digital signal processors can meet with all of the design constraints while providing effective, real time state of health detection. The Goertzel algorithm, long employed in DTMF detection for telephony, was proposed as an alternative to traditional STFT (FFT) approaches and was demonstrated to offer significant reductions in computation burden, memory size, and computation lag time with no accuracy penalty. The Goertzel algorithm is very apt for the multi-frequency EIS approach because it only calculates the required data without spending computing resources on unnecessary calculations unlike the STFT approach. Decimation of the time domain signal prior to the frequency domain transformation was shown to be ill-advised owing to noise constraints. 67 Other multi-frequency EIS improvements such as response sensitive perturbation and phase optimization were not explored because they have been extensively documented by other authors (see Chapter 5). However, it is important to note that any on-board PEMFC state of health sensor, should contemplate these refinements to provide operation across the greatest possible range of SNR. The spectral fingerprint concept was introduced and validated for three common and important PEMFC failure modes, CO poisoning, dehydration, and flooding. The pattern recognition algorithm employed is computationally very simple and exploits the massive concentration of EI spectral data into small fingerprints so that it imposes no significant additional burden on the digital signal processor. It was shown that the spectral fingerprint approach readily distinguishes between different stages of the described failure modes when SNR is greater than –1dB. Under circumstances where SNR falls below this value, the algorithm should not be relied upon as it suffers from failure mode ambiguity particularly between extreme flooding and mild CO poisoning as well as between all states of dehydration. 68 9 Suggestions for future work Failure modes may be easier to detect (more robust to noise) using the present detection algorithm if a greater number of discrete frequencies are used in the stimulus. It would be useful to determine the relationship between frequency bins and detection robustness for a real operating PEMFC by varying the number of frequency bins while monitoring the fingerprint recognition metric. These experiments must be done with a multi-frequency approach because of non-stationarity (similar to [20]). The trade-off to be made is computation burden vs. how critical is the failure mode. For example, it is imperative that flooding be detected early because once established, flooding proceeds at an alarmingly fast rate and can damage the cell permanently. While the same can be said of dehydration with respect to long-term damage, the onset is far more gradual. It would be worthwhile, then, to spend the extra computing power to obtain early detection for flooding. Once the basic technique is proven on operating PEMFCs, the next question becomes, how many different failure modes can we distinguish. To this end, we need to determine the relative importance of other contaminants (e.g., SOx, NOx, H2S, NH3, CH4, Fe3+, Ni2+, Cu2+, Cr3+, Al, Si, K, other sulphur organics, and other hydrocarbons, see [27]) and characterize them in terms of the Randles cell model used in this paper. With this model in hand, the same simulation protocol can be used to simulate the failure modes and determine if sufficient degrees of freedom exist to uniquely identify one failure mode from another. Following this, the trade-off discussed above must be evaluated for these new failure modes. With all the above in hand, and assuming everything works as expected, the next step is to integrate the EIS fault sensor output into a control algorithm. While a state space or PID approach is possible, in all likelihood it would present an excessive computational burden thus squandering the effort made to reduce the computation in the main detection algorithm. A fuzzy logic approach (see [54]) would probably suffice. 69 References 1. Macdonald, J.R., Impedance spectroscopy: emphasizing solid materials and systems. 1987, New York: John Wiley & Sons. 2. Ivers-Tiffée, E., Weber, A., Schichlein, H., Electrochemical impedance spectroscopy, in Handbook of fuel cells, W. Vielstich, Gasteiger, H. A., Lamm, A., Editor. 2003, John Wiley & Sons. 3. Larminie, J., Dicks, A., Fuel cell systems explained. 2003, Chichester: John Wiley & Sons. 4. Vielstich, W., Gasteiger, H. A., Lamm, A., Handbook of fuel cells, ed. W. Vielstich, Gasteiger, H. A., Lamm, A. Vol. 1. 2003: John Wiley & Sons. 5. Li, X., Principles of fuel cells. 2006, New York: Taylor & Francis. 572. 6. Oldham, K.B. and J.C. Myland, Fundamentals of electrochemical science. 1994, San Diego, Calif.: Academic Press. xxii, 474. 7. Schulze, M., et al., Combined electrochemical and surface analysis investigation of degradation processes in polymer electrolyte membrane fuel cells. Electrochimica Acta, 2007. 52(6): p. 2328-2336. 8. Bautista, M., Y. Bultel, and P. Ozil, Polymer electrolyte membrane fuel cell modelling d.c. and a.c. solutions. Chemical Engineering Research & Design, 2004. 82(A7): p. 907-917. 9. Iftikhar, M.U., et al., Dynamic modeling of proton exchange membrane fuel cell using non-integer derivatives. Journal of Power Sources, 2006. 160(2): p. 1170- 1182. 10. Ciureanu, M. and R. Roberge, Electrochemical impedance study of PEM fuel cells. Experimental diagnostics and modeling of air cathodes. Journal of Physical Chemistry B, 2001. 105(17): p. 3531-3539. 11. Wang, Y. and C.Y. Wang, Transient analysis of polymer electrolyte fuel cells. Electrochimica Acta, 2005. 50(6): p. 1307-1315. 12. Wagner, N., Characterization of membrane electrode assemblies in polymer electrolyte fuel cells using a.c. impedance spectroscopy. Journal of Applied Electrochemistry, 2002. 32(8): p. 859-863. 13. Andreaus, B., A.J. McEvoy, and G.G. Scherer, Analysis of performance losses in polymer electrolyte fuel cells at high current densities by impedance spectroscopy. Electrochimica Acta, 2002. 47(13-14): p. 2223-2229. 14. Wagner, N. and E. Gulzow, Change of electrochemical impedance spectra (EIS) with time during CO-poisoning of the Pt-anode in a membrane fuel cell. Journal of Power Sources, 2004. 127(1-2): p. 341-347. 15. Merida, W.R., Diagnosis of PEMFC stack failures via electrochemical impedance spectroscopy, in Mechanical Engineering. 2002, University of Victoria: Victoria. 16. Andreaus, B. and G.G. Scherer, Proton-conducting polymer membranes in fuel cells - humidification aspects. Solid State Ionics, 2004. 168(3-4): p. 311-320. 17. Abe, T., et al., Study of PEFCs by AC impedance, current interrupt, and dew point measurements - I. Effect of humidity in oxygen gas. Journal of the Electrochemical Society, 2004. 151(1): p. A101-A105. 18. Knights, S.D., et al., Aging mechanisms and lifetime of PEFC and DMFC. Journal of Power Sources, 2004. 127(1-2): p. 127-134. 70 19. St Pierre, J., et al., Relationships between water management, contamination and lifetime degradation in PEFC. Journal of New Materials for Electrochemical Systems, 2000. 3(2): p. 99-106. 20. Le Canut, J.M., R.M. Abouatallah, and D.A. Harrington, Detection of membrane drying, fuel cell flooding, and anode catalyst poisoning on PEMFC stacks by electrochemical impedance spectroscopy. Journal of the Electrochemical Society, 2006. 153(5): p. A857-A864. 21. Fouquet, N., et al., Model based PEM fuel cell state-of-health monitoring via ac impedance measurements. Journal of Power Sources, 2006. 159(2): p. 905-913. 22. Hakenjos, A., et al., Simultaneous electrochemical impedance spectroscopy of single cells in a PEM fuel cell stack. Journal of Power Sources, 2006. 154(2): p. 360-363. 23. Freire, T.J.P. and E.R. Gonzalez, Effect of membrane characteristics and humidification conditions on the impedance response of polymer electrolyte fuel cells. Journal of Electroanalytical Chemistry, 2001. 503(1-2): p. 57-68. 24. Springer, T.E., et al., Characterization of polymer electrolyte fuel cells using AC impedance spectroscopy. Journal of the Electrochemical Society, 1996. 143(2): p. 587-599. 25. Wei, Z.D., et al., MnO2-Pt/C composite electrodes for preventing voltage reversal effects with polymer electrolyte membrane fuel cells. Journal of Power Sources, 2006. 160(1): p. 246-251. 26. Baschuk, J.J. and X.G. Li, Carbon monoxide poisoning of proton exchange membrane fuel cells. International Journal of Energy Research, 2001. 25(8): p. 695-713. 27. Cheng, X., et al., A review of PEM hydrogen fuel cell contamination: Impacts, mechanisms, and mitigation. Journal of Power Sources, 2007. 165(2): p. 739-756. 28. Hu, J.M., et al., Kinetics investigation of H-2/CO electrooxidation in PEFCs by the combined use of equivalent circuit fitting and mathematical modeling of the faradaic impedance. Electrochimica Acta, 2004. 49(28): p. 5227-5234. 29. Rosa, F., et al., Design of a diesel reformer coupled to a PEMFC. Catalysis Today, 2006. 116(3): p. 324-333. 30. Kawamura, Y., et al., Multi-layered microreactor system with methanol reformer for small PEMFC. Journal of Chemical Engineering of Japan, 2005. 38(10): p. 854-858. 31. Jiang, R.C., H.R. Kunz, and J.M. Fenton, Electrochemical oxidation of H-2 and H- 2/CO mixtures in higher temperature (T-cell > 100 degrees C) proton exchange membrane fuel cells: Electrochemical impedance spectroscopy. Journal of the Electrochemical Society, 2005. 152(7): p. A1329-A1340. 32. Ciureanu, M., H. Wang, and Z.G. Qi, Electrochemical impedance study of membrane-electrode assemblies in PEM fuel cells. II. Electrooxidation of H-2 and H-2/Co mixtures on Pt/Ru-based gas-diffusion electrodes. Journal of Physical Chemistry B, 1999. 103(44): p. 9645-9657. 33. Ciureanu, M. and H. Wang, Electrochemical impedance study of anode CO- poisoning in PEM fuel cells. Journal of New Materials for Electrochemical Systems, 2000. 3(2): p. 107-119. 34. Kim, J.D., et al., Effect of CO gas and anode-metal loading on H-2 oxidation in proton exchange membrane fuel cell. Journal of Power Sources, 2001. 103(1): p. 127-133. 71 35. Schiller, C.A., et al., Relaxation impedance as a model for the deactivation mechanism of fuel cells due to carbon monoxide poisoning. Physical Chemistry Chemical Physics, 2001. 3(11): p. 2113-2116. 36. Yang, M.C. and C.H. Hsueh, Impedance analysis of working PEMFCs in the presence of carbon monoxide. Journal of the Electrochemical Society, 2006. 153(6): p. A1043-A1048. 37. Bernardi, D.M. and M.W. Verbrugge, A Mathematical-Model of the Solid- Polymer-Electrolyte Fuel-Cell. Journal of the Electrochemical Society, 1992. 139(9): p. 2477-2491. 38. Bieberle, A. and L.J. Gauckler, State-space modeling of the anodic SOFC system Ni, H-2-H2O vertical bar YSZ. Solid State Ionics, 2002. 146(1-2): p. 23-41. 39. Miller, K.S. and B. Ross, An introduction to the fractional calculus and fractional differential equations. 1993, New York: Wiley. xiii, 366 p. 40. Podlubny, I., Fractional differential equations : an introduction to fractional derivatives, fractional differential equations, to methods of their solution and some of their applications. Mathematics in science and engineering ; v. 198. 1999, San Diego ;Toronto: Academic Press. xxiv, 340. 41. Huet, F., A review of impedance measurements for determination of the state-of- charge or state-of-health of secondary batteries. Journal of Power Sources, 1998. 70(1): p. 59-69. 42. Wiegand, G., K.R. Neumaier, and E. Sackmann, Fast impedance spectroscopy: General aspects and performance study for single ion channel measurements. Review of Scientific Instruments, 2000. 71(6): p. 2309-2320. 43. Gabrielli, C., Identification of electrochemical processes by frequency response analysis. 1998, Solartron. 44. Popkirov, G.S. and R.N. Schindler, A New Impedance Spectrometer for the Investigation of Electrochemical Systems. Review of Scientific Instruments, 1992. 63(11): p. 5366-5372. 45. Blanc, G., et al., Measurement of Electrode Impedance in a Wide Frequency Range Using a Pseudo-Random Noise. Electrochimica Acta, 1975. 20(8): p. 599- 601. 46. Blanc, G., et al., Study of Deterministic and Stochastic Behavior of Electrochemical Interface by Means of Numerical Correlation. Journal of Electroanalytical Chemistry, 1975. 62(1): p. 59-94. 47. Gabrielli, C., F. Huet, and M. Keddam, Comparison of Sine Wave and White-Noise Analysis for Electrochemical Impedance Measurements. Journal of Electroanalytical Chemistry, 1992. 335(1-2): p. 33-53. 48. Creason, S.C. and D.E. Smith, Fourier-Transform Faradaic Admittance Measurements .2. Ultra-Rapid, High Precision Acquisition of Frequency-Response Profile. Journal of Electroanalytical Chemistry, 1972. 40(1): p. 1-&. 49. Popkirov, G.S. and R.N. Schindler, Optimization of the Perturbation Signal for Electrochemical Impedance Spectroscopy in the Time-Domain. Review of Scientific Instruments, 1993. 64(11): p. 3111-3115. 50. Darowicki, K. and P. Slepski, Influence of the analyzing window on electrode impedance measurement by the continuous frequency scanning method. Journal of Electroanalytical Chemistry, 2002. 533(1-2): p. 25-31. 72 51. Proakis, J.G. and D.G. Manolakis, Digital signal processing : principles, algorithms, and applications. 2nd ed. 1992, New York, Toronto: Macmillan ;Maxwell Macmillan Canada ; Maxwell Macmillan International. xviii, 969 p. 52. Cooley, J.W. and J.W. Tukey, An Algorithm for Machine Calculation of Complex Fourier Series. Mathematics of Computation, 1965. 19(90): p. 297-301. 53. Schroeder, M.R., Synthesis of Low-Peak-Factor Signals and Binary Sequences with Lol Autocorrelation. Ieee Transactions on Information Theory, 1970. IT16(1): p. 85-+. 54. Schumacher, J.O., et al., Control of miniature proton exchange membrane fuel cells based on fuzzy logic. Journal of Power Sources, 2004. 129(2): p. 143-151. 73 Appendix 1. Simulation of non-integer order transfer function The electrical behaviour of the PEMFC can be modelled approximately by the equivalent circuit (Figure 2.4) discussed in Section 2. In this section we explore the time domain voltage response of the anode or cathode equivalent circuit [9]. The overall cell voltage response is the sum of the anode, cathode and membrane responses. ( ) ( ) ( ) ( )cell anode membrane cathodev t v t v t v t= + + (A1.1) R C ZW Figure A1.1 One electrode of PEMFC Starting with one electrode of the PEMFC, we use the Laplace transform to obtain the impedance of these circuit elements in terms of the complex variable s. ( ) ( ) ( )( ) ( ) ( ) ( ) 0 0 1 , , C f eq f c C f Z s Z s Z s Z s R Z s Z s Z s sCs σ ω ω = = + = + + such that ( ) ( )( )( ) 11 22 11 22 0 0 1 0 0 0 1 eq R s Z s variable s s sC R s σω ω ω σω ω − − + + = ∆ = + + + + Now if we make a change of variable 1 0s s ω= + ( ) ( ) ( ) 1 1 2 2 1 1 2 2 0 1 1 0 1 0 0 1 1 eq R sZ s rearrange s C R s σω ω ω σω − − + − = − + + and rearrange, we obtain ( ) 1 1 2 2 1 2 31 1 1 2 2 2 2 0 1 1 0 1 1 0 1 0 0 1 1eq R s C CZ s multiply by s Rs s R s C σ ω ω σω ω σω − − + − = + − − + 74 which, when multiplied by 1 2 1s , results in ( ) 1 1 2 2 3 31 1 2 2 2 2 0 1 1 0 0 0 1 0 1 1 1eq R s C CZ s R s s Rs C σ ω ω σω ω σω + − = − + − + + which we then simplify as ( ) ( ) 1 2 1 2 31 2 2 1 3 2 2 0 11 0 1 0 0 1 1 1 1 A A sV s I s B B s B s B s ω ω + − = − + + + Now, we cross multiply ( ) ( ) ( ) ( )31 12 2 21 3 1 2 22 1 0 0 1 1 1 1 1 0 0 1V s B B s B s B s I s A A sω ω− + + + = − + and use the Laplace transform pairs ( ){ } ( ) ( ) ( ){ } ( ) atL e f t F s a L D f t s F sα α − = + = to obtain ( ) ( ) ( )( ) ( ) ( )( ) ( ) ( )( ) ( ) ( ) ( )( ) 31 220 0 0 0 1 3 2 2 1 20 0 1 2 1 0 1 0 t t t t t t B v t e B D v t e B D v t e B D v t e A i t e A D i t e ω ω ω ω ω ω + + + = + Now if we set ( ) ( ) ( ) ( )0 0t tf t v t e g t i t eω ω= = we obtain ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )31 122 21 3 1 2 22 1 0 1 0B f t B D f t B D f t B D f t A g t A D g t+ + + = + Now we set up the state equations 75 ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )( )( ) ( ) 1 2 1 1 1 2 2 2 1 1 1 1 2 2 2 2 1 1 2 2 1 3 3 1 4 f t x t D x t x t D x t D D x t x t D x t D D D x t x t = = = = = = giving ( ) ( ) ( ) ( ) ( ) ( ) ( )121 3 1 2 22 0 1 2 1 3 4 0B x t B x t B x t B x t A g t A D g t+ + + = + So we have ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) 1 2 1 2 1 2 2 1 3 2 4 3 x t D x t x t D x t x t D x t = = = (A1.2) and ( ) ( ) ( ) ( ) ( ) ( ) ( )( )121 1 2 2 3 2 4 0 1 2 1 3 0 1 x t B x t B x t B x t A g t A D g t B = − − − + + (A1.3) or, in matrix form ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( )( ) 1 2 1 2 1 2 1 2 3 3 3 3 2 2 2 2 2 1 1 3 2 2 0 4 3 30 1 0 1 0 0 0 0 1 0 1 x t x t x t x t x t x t A g t A D g t x t x t B x tB B BB B B = = + + − − − Now, considering the Grünwald – Letnikov fractional derivative ( ) ( ) ( ) ( ) ( ) ( ) ( ) 0 0 0 1lim 1 1 end start start end end start t t t k t t t k t t t k k D f t y t k t kt w y t k t t α α α α α − ∆ ∆ → = − ∆ = = − − ∆ ∆ ≈ − ∆ ∆ ∑ ∑ (A1.4) 76 where tstart and tend can be considered 0 and t respectively, x is the floor function and kw α is the product of ( )1 k− and the binomial coefficient ( )( ) ( )( )1 2 1 ! n k k α α α αα − − − − = ⋯ (A1.5) ( ) ( ) 1 ! 1k k k α α α Γ + = Γ − + generalized for real α [39, 40]. Since the factorial and the gamma function both exceed the machine number limit for approximately k > 170, which reduces the accuracy of our estimate, we employ a recursive calculation for kw α that is not subject to these limitations based on expression (A1.5). 0 1 1 0 11 1,2, 1k k k w k w w w k N k α α α αα − = = = + = − = − ⋯ (A1.6) We now assume a small enough ∆t giving ( ) 0 1 end start start end t t t end start end start t t k k t t t tD y t w y k t t tt α α α − ∆ = − − ∆ = − ∆ ∆ ∆ ∆ ∑ Substituting end startt tn t − = ∆ , separating the first element of the sum, and modifying the summation limits accordingly, we have ( ) ( ) ( )( ) ( )( )1 1 0 start end n t t k k D y n t y n t w y n k t t α α α = ∆ = − ∆ + − ∆ ∆ ∑ or ( ) ( ) ( ) ( )( ) 1 start endt t D y n t y n t Q n t t α α ∆ = ∆ + ∆ ∆ where 77 ( ) ( )( ) 1 n k k Q n t w y n k tα = ∆ = − ∆∑ so that finally we have ( ) ( ) [ ] [ ]( ) 1[ ] 0,1, 2,3 end startt ty n y n Q n n tt α α − = + = ∆ ∆ … (A1.7) where we now use y[n] and Q[n] as abbreviations for y(n∆t) and Q(n∆t) which are the discrete time versions of y(t) and Q(t) respectively. Additionally, we introduce the notation y(α)[n]to represent the discrete non-integer derivative of the function y(t). Now we can rewrite equations (A1.2) and (A1.3) ( ) [ ] [ ] ( ) [ ] [ ] ( ) [ ] [ ] [ ] [ ] [ ] ( ) [ ] 1 2 1 2 1 11 1 2 22 2 3 3 3 3 3 2 2 2 2 2 1 2 2 3 0 01 3 1 2 3 0 0 x n x n x n x n B AB AB x n x n x n x n g n g n B B B B B − = − = + + + = + and, using result (A1.7), we obtain ( ) [ ] [ ]( ) [ ] ( ) [ ] [ ]( ) [ ] 1 2 1 2 1 1 2 2 2 3 1 0 1 0 x n Q n x n t x n Q n x n t + − = ∆ + − = ∆ ( ) [ ] [ ]( ) [ ] [ ] [ ] [ ] ( ) [ ] [ ]( ) 1 2 1 2 3 3 3 2 2 2 1 2 1 2 3 3 2 2 0 1 3 3 1 2 3 0 1 1 g BB B x n Q n x n x n x n B B Bt AA g n g n Q n B B t + + + + ∆ = + + ∆ or, in matrix form 78 ( ) ( ) ( ) [ ] [ ] [ ] ( ) [ ] [ ] [ ] ( ) [ ] [ ] [ ]( ) 1 2 1 1 2 2 1 2 1 2 1 2 3 31 2 22 3 3 3 2 2 2 1 1 2 2 3 0 30 1 1 1 0 1 10 1 1 g t x n Q n x n Q n t t x n AA tQ n g n g n Q nBB B B B B B Bt − ∆ − − = − ∆ ∆ ∆ − + + + + ∆ Mx = Q Now we take the inverse of M to solve for -1x = M Q where [ ] [ ] [ ] ( ) ( ) ( ) ( ) ( ) ( ) ( ) 0 1 1 2 2 2 3 3 3 n t x n x n t v n t e x n x n t x n t x n x n t x n t ω ∆ ∆ ∆ = ∆ = ∆ ∆ ∆ x = , ( ) ( )( ) ( )( ) ( )( ) ( ) ( ) ( ) ( ) ( ) ( )( ) ( )( ) 1 2 1 2 1 2 1 2 11 1 020 02 2 3 3 2 2 1 1 2 1 0 3 1 1 1 n k k n k k n n n k tn t n t k k k k w x n k t w x n k t t AA t w x n k t i n t e i n t e w i n k t e B B ωω ω = = − ∆∆ ∆ = = = − − ∆ − − ∆ ∆ ∆ − − ∆ + ∆ + ∆ + − ∆ ∑ ∑ ∑ ∑ Q and x(0) = 0. 79 Appendix 2. Equivalent circuit values from the literature Since the model for the PEMFC that we are employing has not been used in its entirety by other investigators15, we infer parameters from different studies trying to keep the experimental conditions and dimensions as similar as possible. For this study, it is the ability of the algorithms discussed herein to detect and identify changes occurring within the model via EIS that is of significance rather than the fidelity of this model to any particular fuel cell. EIS data is available for experiments carried out on a PEMFC at 80°C with a Nafion 117 membrane, symmetric E-TEK electrodes (20% Pt/C, 0.4mg·cm-2, 200µm thickness, 23 cm2 geometric surface area) impregnated with 1mg·cm-2 Nafion suspension, humidified anode H2 and dry cathode O2 gas feed at 2 bar absolute, dead end H2 flow and O2 stoichiometry of λ=8 [12]. The EIS data can be fit to the equivalent circuit of Figure 2.4 using the following approach: 1. Rs can be easily inferred from the high frequency end of the magnitude Bode plot. At high frequency, the double layer capacitors behave as short circuits leaving Rs as the sole real impedance in the equivalent circuit. 2. Rct,a and Rct,c. These values can be inferred from the two extremes of the magnitude Bode plot. , ,ct total left right ct total sZ Z Z R R= − = − (A2.1) In the same study, io,anode and io,cathode were obtained for H2:H2 and O2:O2 operation. Using these io values together with the , , , ,ct total ct total ct anode ct cathodeR Z R R= = + values and relationship (2.6), we can obtain an approximate value for Rct,a and Rct,c over the range of experimental current densities. Note that this approach fails at high 15 Instead they have used subsets of this circuit depending on which element represented the behaviour under investigation. For example, catalysis studies tend to ignore the Warburg impedance focusing instead on the Rct-Cdl elements. Other studies concerning O2 diffusion include the Warburg impedance but ignore the anode contribution. 80 current densities because the diffusion limitation begins to appear. In such cases, we extrapolate the horizontal section of the magnitude bode plot to find ,ct totalZ . 3. Now we turn to the double layer capacitances, Cdl,a and Cdl,c. We can calculate Cdl,total from the pole frequency in the magnitude Bode plot together with Rct,total. Now, in order to discover the respective anode and cathode contribution we note , , , , , ,, , o r anode eff anode dl anode H anode o r cathode eff cathodedl cathode H cathode A C d AC d ε ε ε ε = (A2.2) and ( ) 1 1 1 3 2 2 2 2 3 3 2 1 13 11 1 4 4 3 2 2 4 4 3 2 2 part part eff geom molar surf molar grav surf eff geom grav A A L M V D A U mg mol particle cm cm cm cm cm mg cm mol particle d dV D D M A d dA A L M D M pi pi pi pi − − − − − − − − = ⋅ ⋅ ⋅ ⋅ ⋅ ⋅ = ⋅ ⋅ ⋅ ⋅ ⋅ = = = = ⋅ ⋅ ⋅ ⋅ ⋅ ( ) 2 16 eff geom grav U A A L D U d − ⋅ = ⋅ ⋅ ⋅ ⋅ (A2.3) where A, L, M, V, D, and U represent area, loading, molar mass, volume, density, and utilization, respectively. Since the anode and cathode electrodes are symmetric, Ageom, L, d, Dgrav, and U are the same and therefore Aeff,anode = Aeff,cathode. , , , , , , dl anode r anode H cathode dl cathode r cathode H anode C d C d ε ε = (A2.4) The anode and cathode double layer compositions, however, are quite different and εr,anode ≠ εr,cathode and dH,anode ≠ dH,cathode. The ionic species at the anode are hydrated protons forming a layer of aligned water molecules for which the dielectric constant is 6 rather than the unaligned water dielectric constant of 80 (ref: Modern electrochemistry). At the cathode surface, the precise ionic 81 composition is unknown, but the tight monolayer of aligned water molecules is unlikely. Further, any anions lining the cathode surface will be large compared to the proton such that this double layer is likely to be several times thicker and more diffuse than that of the anode. The effects of thickness and dielectric constant are opposite in (A2.4) so for simplicity, we set Cdl,anode = Cdl,cathode resulting in , , , 2dl anode dl cathode dl totalC C C= = . 4. The dominant diffusion processes at each electrode influencing the Warburg impedances involve water transport on the anode side and oxygen gas transport on the cathode side [13]. The time constants for these processes have been shown to be on the order of 10 seconds for the anode and 0.01 to 0.1 seconds for the cathode [11]. We are then left with determining the current density dependent A(j) parameter from (2.10). We fit this last parameters by trial and error with the constraint that Aa(j) has been reported as 3 orders of magnitude smaller than Ac(j) in a symmetrical fuel cell [9]. Table A2.1 Equivalent circuit parameters derived from the literature Cell voltage (V) 0.984 0.900 0.797 0.697 0.597 0.497 0.397 Current density (mA·cm-2) 0.004 5.814 93.00 254.4 405.0 516.8 622.1 Power density (mW·cm-2) 0.004 5.233 74.12 177.3 241.8 256.8 246.9 Rct,a (µ Ω) 342.2 15.20 1.600 1.100 1.000 1.200 1.400 Rct,c (Ω) 5.704 0.253 0.026 0.018 0.017 0.020 0.023 Rs (mΩ) 10.00 10.00 10.00 10.00 10.00 10.00 10.00 Cdl,a (F) 0.100 0.100 0.100 0.100 0.100 0.100 0.100 Cdl,c (F) 0.100 0.100 0.100 0.100 0.100 0.100 0.100 Aa (µΩ) 0.000 0.000 0.000 0.000 5.000 10.00 15.00 Ac (mΩ) 0.000 0.000 0.000 0.000 5.000 10.00 15.00 τa (sec) 0.100 0.100 0.100 0.100 0.100 0.100 0.100 τc (sec) 10.00 10.00 10.00 10.00 10.00 10.00 10.00 The resulting equivalent circuit parameters are collected in Table A2.1 and the corresponding impedance plots are shown in Figure 2.7 and Figure 2.8. 82 Appendix 3. Matlab program listings These are rather extensive to list and are available from the author in electronic form (ivsm at telus dot net).
- Library Home /
- Search Collections /
- Open Collections /
- Browse Collections /
- UBC Theses and Dissertations /
- Electrochemical impedance spectroscopy options for...
Open Collections
UBC Theses and Dissertations
Featured Collection
UBC Theses and Dissertations
Electrochemical impedance spectroscopy options for proton exchange membrane fuel cell diagnostics Valenzuela, Jorge Ignacio 2007
pdf
Page Metadata
Item Metadata
Title | Electrochemical impedance spectroscopy options for proton exchange membrane fuel cell diagnostics |
Creator |
Valenzuela, Jorge Ignacio |
Publisher | University of British Columbia |
Date Issued | 2007 |
Description | Electrochemical impedance spectroscopy (EIS) has been exploited as a rich source of Proton Exchange Membrane Fuel Cell (PEMFC) diagnostic information for many years. Several investigators have characterized different failure modes for PEMFCs using EIS and it now remains to determine how this information is to be obtained and used in a diagnostic or control algorithm for an operating PEMFC. This work utilizes the concept of impedance spectral fingerprints (ISF) to uniquely identify between failure modes in an operating PEMFC. Three well documented PEMFC failure modes, carbon monoxide (CO) poisoning, dehydration, and flooding were surveyed, modelled, and simulated in the time domain and the results were used to create a database of ISFs. The time domain simulation was realized with a fractional order differential calculus state space approach. A primary goal of this work was to develop simple and cost effective algorithms that could be included in a PEMFC on-board controller. To this end, the ISF was discretized as coarsely as possible while still retaining identifying spectral features using the Goertzel algorithm in much the same way as in dual tone multi-frequency detection in telephony. This approach generated a significant reduction in computational burden relative to the classical Fast Fourier Transform approach. The ISF database was used to diagnose simulated experimental PEMFC failures into one of five levels of failure: none (normal operation), mild, moderate, advanced, and extreme from one of the three catalogued failure modes. The described ISF recognition algorithm was shown to correctly identify failure modes to a lower limit of SNR = 1dB. |
Extent | 3308007 bytes |
Subject |
impedance diagnostic |
Genre |
Thesis/Dissertation |
Type |
Text |
FileFormat | application/pdf |
Language | eng |
Date Available | 2008-01-07 |
Provider | Vancouver : University of British Columbia Library |
Rights | Attribution-NonCommercial-NoDerivatives 4.0 International |
DOI | 10.14288/1.0066203 |
URI | http://hdl.handle.net/2429/266 |
Degree |
Master of Applied Science - MASc |
Program |
Electrical and Computer Engineering |
Affiliation |
Applied Science, Faculty of Electrical and Computer Engineering, Department of |
Degree Grantor | University of British Columbia |
GraduationDate | 2008-05 |
Campus |
UBCV |
Scholarly Level | Graduate |
Rights URI | http://creativecommons.org/licenses/by-nc-nd/4.0/ |
AggregatedSourceRepository | DSpace |
Download
- Media
- 24-ubc_2008_spring_valenzuela_jorge_ignacio.pdf [ 3.15MB ]
- Metadata
- JSON: 24-1.0066203.json
- JSON-LD: 24-1.0066203-ld.json
- RDF/XML (Pretty): 24-1.0066203-rdf.xml
- RDF/JSON: 24-1.0066203-rdf.json
- Turtle: 24-1.0066203-turtle.txt
- N-Triples: 24-1.0066203-rdf-ntriples.txt
- Original Record: 24-1.0066203-source.json
- Full Text
- 24-1.0066203-fulltext.txt
- Citation
- 24-1.0066203.ris
Full Text
Cite
Citation Scheme:
Usage Statistics
Share
Embed
Customize your widget with the following options, then copy and paste the code below into the HTML
of your page to embed this item in your website.
<div id="ubcOpenCollectionsWidgetDisplay">
<script id="ubcOpenCollectionsWidget"
src="{[{embed.src}]}"
data-item="{[{embed.item}]}"
data-collection="{[{embed.collection}]}"
data-metadata="{[{embed.showMetadata}]}"
data-width="{[{embed.width}]}"
async >
</script>
</div>
Our image viewer uses the IIIF 2.0 standard.
To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0066203/manifest