pem fuel cell fault detection and identification using...

Eur. Phys. J. Appl. Phys. 54, 23412 (2011)DOI: 10.1051/epjap/2011100277

THE EUROPEANPHYSICAL JOURNAL

APPLIED PHYSICS

Regular Article

PEM fuel cell fault detection and identification using differentialmethod: simulation and experimental validation

E. Frappe1,a, A. De Bernardinis1,b, O. Bethoux2,c, D. Candusso1,3, F. Harel1,3, C. Marchand2, and G. Coquery1

1 IFSTTAR LTN/SPEE Labs, 25 allee des Marronniers, 78000 Versailles-Satory, France2 Laboratoire de Genie Electrique de Paris (LGEP)/SPEE-Labs, SUPELEC, Univ Paris-Sud, Univ Pierre et Marie CurieParis 6, CNRS (UMR 8507), 11 rue Joliot Curie, Plateau de Moulon, 91192 Gif-sur-Yvette Cedex, France3 FC LAB/IFSTTAR, Rue E. Thierry-Mieg, Technopole, 90010 Belfort, France

Received: 11 July 2010 / Received in final form: 17 November 2010 / Accepted: 10 April 2011Published online: 18 May 2011 – c© EDP Sciences

Abstract. PEM fuel cell performance and lifetime strongly depend on the polymer membrane and MEAhydration. As the internal moisture is very sensitive to the operating conditions (temperature, stoichiom-etry, load current, water management. . . ), keeping the optimal working point is complex and requiresreal-time monitoring. This article focuses on PEM fuel cell stack health diagnosis and more precisely onstack fault detection monitoring. This paper intends to define new, simple and effective methods to getrelevant information on usual faults or malfunctions occurring in the fuel cell stack. For this purpose, theauthors present a fault detection method using simple and non-intrusive on-line technique based on thespace signature of the cell voltages. The authors have the objective to minimize the number of embeddedsensors and instrumentation in order to get a precise, reliable and economic solution in a mass marketapplication. A very low number of sensors are indeed needed for this monitoring and the associated algo-rithm can be implemented on-line. This technique is validated on a 20-cell PEMFC stack. It demonstratesthat the developed method is particularly efficient in flooding case. As a matter of fact, it uses directly thestack as a sensor which enables to get a quick feedback on its state of health.

1 Introduction

Environmental issues have increased the demand for lesspolluting energy generation technologies. A hydrogen fuelcell (FC) directly converts the electrochemical energy ofhydrogen into electricity and only produces heat and wa-ter. Proton Exchange Membrane Fuel Cell (PEMFC) is anattractive technology because of its high power density, itssolid membrane and its low operating temperature allow-ing fast startups and immediate response to changes inthe demand of power [1]. However, the membrane has tobe fully water saturated in order to enhance its ionic con-ductivity. As present cells operate at a rated temperaturerange between 60 ◦C and 80 ◦C, stack water managementis a PEMFC key issue. Flooding, due to an excess of wa-ter in the cells, inhibits gas transport to the reaction sitesand reduces the active surface area of the catalysts. Theadverse effect is a significant rise of the activation andconcentration losses leading sometimes to a catastrophicdecrease of the cell efficiency [2]. On the contrary, a dryingsituation results in an increase of the membrane resistivity

a e-mail: [email protected] e-mail: [email protected] e-mail: [email protected]

also reducing cell efficiency [3–5]. A long operation withlow water content reduces the FC lifetime [6].

Any fault induces a voltage drop [5]. Hence, cell voltageis a good failure indicator, easy to implement. Moreovera voltage sensor is a high bandwidth and accurate device.Nevertheless voltage monitoring makes fault identificationdifficult to achieve. A first approach is to use the frequencysignature of each failure: dehydration adversely affects theionic conductivity which is a low time response phenom-enon, whereas flooding negatively diminishes gas diffusionperformance which is a very low time response observ-able fact. Hence, electrochemical impedance spectroscopy(EIS) provides more complete information about the FCstate of health thanks to fuel cell impedance measure[7–9]. However, EIS is a relatively slow method since itneeds time to perform a full frequency spectrum [10]. EISalso requires a stabilized working point which is difficultto obtain in real operation. Furthermore, it involves ex-pensive and sometimes voluminous devices to perform theidentification. Nevertheless, partial frequency impedancemay be fast executed and gives interesting but limitedinformation. Hinaje et al. [11] use the HF current rip-ple generated by the power converter, which is commonlyused for FC system electric power conditioning. Using thisprinciple, they perform a membrane resistance measure.

23412-p1

http://dx.doi.org/10.1051/epjap/2011100277

http://www.epj.org

http://www.epj.org

http://www.epj.org

The European Physical Journal Applied Physics

Subsequently drying is easily detected by resistance in-crease but flooding cannot be identified. Even though EISremains a promising “perturb and observe” technique, itbecomes difficult to implement in stack with large num-ber of cells. As a matter of fact, individual cell voltagesmust be monitored within a stack since a failure (flood-ing, drying, poisoning. . . ) can only occur in a few cells.Although some methods exist [12,13], the use of a hugenumber of sensors is subject to failure. This makes volt-age monitoring a challenge for high voltage stacks (longstacks).

Power fuel cells (Power FCs) are currently developedfor applications like traction for electrical vehicle, auxil-iary power units (APU) or even battery charger in somehigh power vehicles or fuel cell hybrid electric vehicles(FCHEVs). For example, the French research projectSPACT-80 designed, manufactured and tested in real con-ditions a robust and durable air/H2 80 kW PEM fuel cell-based system, specifically developed for railway and roadapplications [14–16]. The GENEPAC project was run inorder to design and build a fuel cell for automotive ap-plications. It focuses on the development of a compact80 kW fuel cell stack with high output dynamic perfor-mances [17]. Recently, Toyota revealed its new FCHEVpowered by an 80 kW fuel cell [18]. All these high powerPEM stacks are made of large specific areas and com-posed of about a hundred cells. They are sometimes usedas a twin-stack or a multi-stack fuel cell generator [19] be-cause PEMFC power increase is limited by technical andmechanical constraints [20]. As a matter of fact, stackswith large membrane electrode assembly (MEA) and animportant cell number are difficult to operate because ofheterogeneous fluidic distribution between cells. This dis-crepancy affects FC performances [21] and makes watermanagement difficult to operate efficiently. Consequentlyon-line monitoring is particularly important for safe op-eration. In order to get a precise, reliable and economicsolution in a mass market application, the objective isto minimize the number of embedded sensors, the priceof instrumentation and the complexity of the relatedalgorithm.

In this perspective, a new fault detection method isproposed and studied in this article. It still relies on volt-age measurement but is no longer based on frequency sig-nature. It intends to make use of the small disparity oftemperature and the non-uniform distribution of reactantsalong the stack. The fact is that many authors pointed outthe cell discrepancies throughout the stack which dependon the operating conditions [22]. Some authors show thatcell voltage is lower in the cells furthest from the fuel in-let of the stack (nearest to the air inlet) due to unevengas distribution or water flooding [23]. Different experi-ments also point out that the central interior of the stackis hotter than its exterior, and subsequently drying mostlyoccurs in center cells [24,25]. In fact, Ramousse et al. [26]highlight that water in the cells is highly dependent onthe temperature; as a result flooding occurs in cooler cellsand drying in hotter cells. Based on these remarks, wesuggest measuring the voltage of group of cells in the in-let, the center and the outlet of the stack. It will allow the

detection of a fault when a voltage difference occurs andalso permit identification using the shape of this voltagenon-uniformity.

In this paper, we propose a differential voltage method,which only needs a few voltage sensors located judiciouslyalong the FC stack and supervising only some particularcells or groups of cells. It can be performed on-line with-out disturbing the FC electric load neither with a smallsignal (EIS) nor with current interrupt (CI). Moreover,the measurements are associated with a simple algorithmwhich permits to get rid of many signal disturbances; forinstance, load changes and slow dynamic phenomena(aging effects. . . ) do not affect the accuracy of thedetection.

The paper is organized as follows. First a dynamicmodeling of a cell is presented which allows performingthe simulations and studying the behavior of the PEMfuel cell when a fault occurs. Then, the paper focuses onthe fault detection and identification method applied tothe multi-cell stack. The originally developed approachis first simulated in drying and flooding cases. In a sec-ond step, the method is experimented and validated on a20-cell PEMFC test bench in the particular case of stackflooding. Finally the advantage of the method is discussedand compared to EIS and CI techniques.

2 Fuel cell modeling and polarizationsignature

2.1 Fuel cell modeling: application to a 20-cell PEMFC

Some work has already been reported in the literature,static and dynamic modeling based on empirical equationswas given in [1,27–30]. The voltage drop across the fuel cell(FC) can be written as a function of the activation, ohmicand concentration overvoltages given by equation (1) forone cell:

Vcell = E − ηact − ηohm − ηconc (1)

with E the electromotive force given using the Nernstequation:

E = 1.229 − 8.5 × 10−4(Tfc − 298.15) + 4.3085 × 10−5

×Tfc(ln(PH2)) + 0.5 ln(PO2). (2)

Tfc is the fuel cell stack temperature; PH2 and PO2 arethe partial pressures of hydrogen and oxygen respectively(bar).

The activation voltage losses represent the fact thatsome energy is needed to generate a reaction product.They are obtained by the Tafel equation [1]:

ηact = A ln((J + Jn)/J0) (3)

with

A = (RTfc)/(2αF ). (4)

23412-p2

E. Frappe et al.: PEM fuel cell fault detection and identification using differential method

J is the fuel cell current density, Jn the leakage currentdensity, J0 the exchange current density, R the perfectgas constant, α the charge transfer coefficient and F theFaraday constant.

The ohmic voltage losses are due to the resistance forboth electronic and ionic currents. They result in a slowand linear voltage drop with an increasing current. Themain parameter of this voltage drop is the membraneresistance Rmem:

ηohm = RmemJ. (5)

The concentration voltage losses are due to the reductionof the reactant concentration at the electrode surface in-duced by reactants’ consumption. This irreversibility be-comes significant at high current density and the relatedovervoltage is obtained empirically:

ηconc = m exp(nJ). (6)

m and n are constants depending on the construction ofthe cell. It is useful to formulate the real voltage of the cellwhen there is no current; this voltage is named the opencircuit voltage (OCV). Using equations (2) and (3), theopen current voltage Eocv can be computed as follows:

Eocv = 1.229 − 8.5 × 10−4(Tfc − 298.15)

+ 4.3085 × 10−5 × Tfc(ln(PH2)+ 0.5 ln(PO2)) + A ln(J0) (7)

ηact2 = A ln(J + Jn). (8)

More precisely the cell voltage becomes:

Vcell = Eocv − ηact2 − ηohm − ηconc. (9)

Concentration of charges (electrons and H+ ions) on theactive surface (electrode and electrolyte surface) greatlyinfluences the probability of reaction. This build-up ofcharges takes time to establish, drives non-faradic cur-rent and depends on the load current. Consequentlyduring current transient, overvoltage does not change in-stantaneously. This phenomenon is complex and knownas charge double layer; it can be modeled by a capacitor[28,31].

In the dynamic model, the double layer capacitor isplaced in parallel to the sources representing electrochem-ical voltage losses (ηact2 and ηconc). That is the reasonwhy we have to modify the model elements. As a matterof fact, the fuel cell current density J is shared between thedouble layer capacitor and ηact2, ηconc. Then the currentin the voltage loss branch (Jf ) needs to be calculated. Forthat purpose, we proceed as follows.

ηact2 is replaced by a current source, called Jf ,controlled by the voltage ηact2, which is calculated withKirchhoff’s voltage law:

ηact2 + ηconc = Vcdl, (10)ηact2 = Vcdl − ηconc. (11)

Jf is calculated thanks to equation (8)

Jf = exp(ηact2/A) − Jn. (12)

Fig. 1. Dynamic FC model: representation of the double layerphenomenon.

0 10 20 30 40 50 60 70 808

10

12

14

16

18

20

Fuel Cell current (A)

Fuel

Cel

l vol

tage

(V)

Polarization curve of 20-cell PEMFC

model

experimental

Fig. 2. (Color online) Polarization curves of the experimentedPEMFC stack and its related model.

ηconc is a voltage source controlled by Jf . The ohmic lossis modeled by a resistor. The model representation withthe double layer capacitor is given in Figure 1.

The cell model that is tuned with parameters is givenin Table 1. In order to have the fuel cell voltage, Vcell

is multiplied by the number of cells. Figure 2 shows thepolarization curve of the simulated model compared with apolarization curve obtained by experimental measurementon a 20-cell stack. The experimental polarization curve isobtained as follows: first fuel cell is set to nominal point(stack temperature = 60 ◦C, H2/O2 stoichiometry = 2/4,H2/O2 hygrometry = 14/22%). Then current is reducedfrom 60 to 0 A in 10 min (ramp current = −0.1 A/s).As shown, the simulated curve fits in accordance with theexperimental curve.

2.2 Polarization signature

The work described in this article focuses on two kindsof faults: membrane drying and cell flooding. A drying,due to insufficient water content in the membrane,increases its resistance [25,32]. In case of a flooding,

23412-p3


Table 1. Parameters used in the FC analytical electrical model.

α Charge transfer coefficient 0.44m Concentration coefficient 2.11 × 10−5

n Concentration coefficient 8 × 10−3

Scell Cell surface 100 cm2

Rmem Membrane resistance 300 × 10−6 kΩ cm2

Jn Leakage current density 3 mA/cm2

J0 Exchange current density 4 × 10−4 mA/cm2

Cdl Double layer capacity 50 mF/cm2 [10]F Faraday’s constant 9.65 × 104 C/molR Perfect gas constant 8.314472 J/mol/K

Tfc Fuel cell temperature 333.15 KPH2 Hydrogen partial pressure 0.93 barPO2 Oxygen partial pressure 0.17 bar

0 20 40 60 80 100 120 140

0.2

0.4

0.6

0.8

1

Fuel Cell current (A)

Cel

l vol

tage

(V)

polarization curve of a cell

(C) Flooded cell

(A) Normal cell

(B) Dried cell

Fig. 3. (Color online) Different polarization curves of a cellobtained by simulation.

del Real et al. [27] explain that the water forms a thinfilm blocking part of the active fuel cell area. This phe-nomenon results in a lower apparent active area [33,34]and thus a higher current density which increases over-voltages.

In general faults do not occur in the entire stack butonly affect a few cells [21]. The impact on a single cellvoltage is shown in Figure 3. Plot (A) shows the cell po-larization curve in rated conditions. Plot (B) displays thepolarization curve of the same cell when membrane dry-ing occurs: the dried membrane resistance is increased by1.5 compared to the rated value [11]. Plot (C) illustratesthe polarization curve of this cell when membrane flood-ing happens: the active area was reduced to 80%, this willact on the cell current density according to this relationJ = Ifc/(Scell × k) with k = 0.8 the degree of reducedactive area [27]. As already described in [35], the voltagedrop is similar either in the drying case or in floodingcircumstances. The V-I characteristics become differentonly when the cell voltage is lower than its minimum volt-age (0.4 V). The fuel cell should not be used under thisvoltage [36].

3 Fault detection for a power stack

3.1 Monitoring for a power stack

Stack-level investigations offer more opportunities thansingle cell level analysis. In particular, references [22,26,37] show that flooding or drying only affects some local-ized cells or groups of cells. In case of a high dew pointtemperature, humidified air quickly condenses at the inletof the stack. Furthermore water produced by the electro-chemical reaction can easily accumulate at the outlet ofthe stack. Eckl et al. [25] demonstrate that, during flood-ing experiments on a 20-cell stack, the most perturbedcells were the cells 3 and 5. Hernandez et al.’s experiments[37] pointed out, with a 20-cell stack, voltage degradationdue to flooding in cells 4–6, 18, and 20. In Corbo et al.’sexperiments on a 34-cell stack [38], flooding has a greateffect on the 6th and 28th cells. Conversely the center ofthe stack is hotter than its extreme parts; the presence ofliquid water is lower. In Park and Caton’s drying experi-ments on an 8-cell stack, the most affected cells were the4th, 5th and 6th [32]. Eckl et al. give the same observationwith a 20-cell stack [25], where cell voltages V9, V11 andV13 were deeply disturbed by a drying.

Hence a flooding may occur in the inlet and outlet ofthe stack while a drying may occur in the center of thestack. That is the reason why, instead of monitoring allthe cells, it could be judicious to monitor only a groupof representative cells: cells located at the inlet and theoutlet in order to detect a flooding, and cells in the centerof the stack for detecting a drying.

3.2 Fault detection and identification (FDI) strategyusing the differential method: approach by simulation

Referring to the previous studies mentioned in the arti-cle, it appears relevant to instrument with voltage sensorsthree main areas of the stack: the inlet, the outlet andthe center (Fig. 4). In this case, the detection principle isbased on the monitoring of a differential between the volt-age in the center of the stack Vcenter and the inlet/outletvoltages (respectively Vinlet, Voutlet). This principle allows

23412-p4


Cell

Vinlet

Voutlet

Vcenter

Stack

Fig. 4. (Color online) Principle of monitored cells for a fuelcell stack.

to generate a new state-of-health indicator of the fuel cellstack. In the case of no fault, all voltages are constant ordrop similarly due to load variation: the differential volt-ages equal zero. If a drying appears, only Vcenter drops andthe two differential voltages become positive. However, ifa flooding occurs, the inlet and/or the outlet voltage dropsleading to a decrease of one or two differential voltages.

Using the differential method, the FC stack can be as-similated to a voltage sensor. Its characteristic is to pro-vide two key signals representative of the state of health.The feedback information is simple, fast and based on thereal-time operating conditions without adding any exter-nal disturbance. Moreover the method can be applied alsowhen the operating conditions are slightly modified.

As a first approach, the proposed detection method isvalidated by numerical simulations.

3.2.1 Fuel cell model enhancement

In order to study the differential method, the FC stack ismodeled in the following way: firstly, FC is segmented intofive segments of 3, 6, 3, 5, 3 cells distributed from the inletto the outlet of the stack. Previous cell model is used tocalculate each segment voltage Vst,i. Then electrical modelof segment i is coupled with thermal one as represented inFigure 5.

The temperature of one stack segment is calculated asfollows [27,39]:

mstCst∂Tst

∂t=

•Hreact − Qcool − Qconv

−Qrad − PEl + Qcond. (13)

mstCst is the thermal capacity of the system, Tst is thestack segment temperature. With

•Hreact the energy of the

electrochemical reaction:•Hreact =

•mH2ΔhH2 +

•mO2ΔhO2 − •

mH2O

(h0

fΔh)H2O

.

(14)

Thermal modelSegment i

Tst, i

V st, iP O2

P H2 ElectricalmodelSegment i

Ist

TH2,TO2,mcoolant

Thermal and electrical modelSegment i+1

V st, i+1Ist

Fig. 5. Segmented FC stack model (global overview).

h0f is the mass specific enthalpy of formation and

•m

denotes the mass flow of the species:

•mH2 = MH2Ncell

Ist

2F, (15)

•mO2 = MO2Ncell

Ist

4F, (16)

•mH2O = MH2ONcell

Ist

2F. (17)

M is the molecular mass (kg/mol), Ncell is the stack seg-ment number of cell, F the Faraday constant and Ist thecurrent stack.

Δh is the mass specific enthalpy difference from thepresent state to the reference state:

ΔhH2 = CpH2 (TH2 in − T0) , (18)ΔhO2 = CpO2 (Tair in − T0) , (19)ΔhH2O = CpH2O (Tair out − T0) . (20)

Cp are the specific heat energies, TH2 in, Tair in and Tair out

are the inlet H2 gas temperature, the inlet air gas temper-ature and the outlet air gas temperature.

Outlet gas temperature is calculated thanks to:

Tair out = 2T stackair − Tair in, (21)

where

T stackair = Tst −

•Hreact − PEl

kT. (22)

Here it is assumed that the temperature difference be-tween the FC and the cathode air is proportional to thetotal waste heat of the reaction. The constant proportion-ality kT can be determined experimentally.

Qconv and Qrad are the heat transfer to the surround-ing area by a convective and a radiative heat flow:

Qconv = hambAamb (Tst − Tamb) , (23)Qrad = εσAamb

(T 4

st − T 4amb

). (24)

hamb is the heat transfer coefficient of the FC, Aamb de-notes the outer surface of the body, ε is the emissivity ofthe body, σ is the Stefan-Boltzmann constant and Tamb

represents the temperature of the environment.The electric power of the system equals the product of

the stack segment voltage with the current:

PEl = VstIst. (25)

23412-p5


Qcool represents the heat transfer rate from the body tothe coolant:

Qcool = hcoolAcool (Tst − Tcool) (26)

with Acool the heat transfer area, Tcool the temperatureof the coolant inside the FC and hcool the heat transfercoefficient given by:

hcool = Kh

( •mcoolant

). (27)

This heat transfer coefficient is a function of the coolantmass flow

•mcoolant, the coefficient Kh has to be determined

experimentally.The temperature of the coolant inside the FC is calcu-

lated as follows:

mcoolCcool∂Tcool

∂t= Δ

•Hcool + Qcool. (28)

mcoolCcool is the thermal capacity of the coolant systemin the FC segment.Δ

•Hcool is the enthalpy difference of the coolant flow:

Δ•Hcool =

•mcoolantCpH2O (Tcool in − Tcool) (29)

with Tcool in the inlet coolant temperature.Lastly conductive effects between one segment to its

neighbor are given by:

Qconv =

⎧⎨⎩

Dδ (T2 − Tst) for segment 1,Dδ (Tn+1 + Tn−1 − 2Tst) for segment n = 2, 3, 4,Dδ (T4 − Tst) for segment 5.

(30)D is the transfer coefficient and δ the thickness betweentwo segments. All parameters required for the thermalmodel are summarized in Table 2.

Temperature along the stack is plotted in Figure 6. H2

and air gas temperature is set to 40 ◦C, ambient temper-ature is 25 ◦C. Coolant inlet temperature is 50 ◦C. FCcurrent is set to 30 A.

As expected, the center of the FC is hotter thanextremity.

A temperature raise of 0.7 ◦C in the center of the stackincreases the vapor partial pressure by 2.4%. Then lessliquid water should be present in the center of the fuelcell.

3.2.2 Simulation results

Figure 7 shows the detection method in the case of a flood-ing. Flooding occurs at 3 s in the inlet and outlet of thestack, and water is slowly accumulated in the cells untilthe active area is reduced to 80%, while inlet and outletmonitored voltage drop slowly. The global stack voltagedoes not drop significantly. On the contrary, monitoringthe three groups of cells makes it possible to detect theearly fault. The differential voltages (Vintlet − Vcenter) and(Voutlet −Vcenter) are negative and are falling slowly as thewater content increases.

2 4 6 8 10 12 14 16 18 2059.4

59.5

59.6

59.7

59.8

59.9

60

60.1

60.2

Cell number

FC stack temperature

Tem

p (°

C)

Fig. 6. (Color online) Fuel cell stack temperature profile versuscell number.

10

15

20stack voltage

Vol

t

-0.5

0

0.5differential voltage Vinlet-Vcenter

Vol

t

0 2 4 6 8 10 12 14 16-0.5

0

0.5differential voltage Voutlet-Vcenter

Vol

t

time (s)

Fig. 7. (Color online) Principle of monitored cells: floodingcase.

Figure 8 shows the case of a drying affecting the cen-ter cells. In this case, because the center cells are drying,their resistances increase gradually up to a 1.5 factor re-sulting in a voltage drop of center cells only. Thus, the twodifferential voltages are positive and increase. In this casetoo, the global stack voltage does not drop significantly,but thanks to the differential voltages, the fault can bedetected in a sensitive way.

Simulation results demonstrate the validity of the tech-nique on well-known faults. However, it should distinguishbetween a real fault and a disturbance in order to preventfalse alarm.

Figure 9 shows the case of a load variation of a healthystack in order to validate the detection principle duringtransient. In this occurrence, all cells are perturbed in thesame way and thus their voltage drops are equal leadingto a constant differential voltage.

23412-p6


Table 2. Parameters used in the FC thermal model.

mst Mass of the stack segment 1.4 kgCst Stack specific heat 1300 J kg−1 K−1

MH2,O2,H2O Molecular mass of H2, O2 and H2O kg/molNcell Number of cells in the segment

CpH2,O2,H2O Specific heat energy of H2, O2 and H2O J kg−1 K−1

T0 Reference temperature 298.15 KKT Constant proportionality 340 W K−1

D Heat transfer coefficient between two FC segments 0.35 W m K−1

δ Thickness between two FC segments 5 × 10−3 mhamb Heat transfer coefficient of FC 3.9 W m−2 K−1

Aamb Convection exchange surface 0.023 m2

ε Emissivity of the body 0.9σ Stefan-Boltzmann constant 5.678 × 10−8 W m−2 K−4

Acool Heat transfer area of coolant 0.011 m2

Kh Heat transfer coefficient 2.16 × 105 W kg−2 K−1

mcoolCcool Thermal capacity of the coolant system 110 J K−1

•mcoolant Coolant mass flow 0.08 kg s−1

H0f Mass specific enthalpy of formation −1.58 × 107 J K−1

10

15

20stack voltage

Vol

t

-0.5

0


Vol

t

0 2 4 6 8 10 12 14 16-0.5

0

0.5differential voltage Voutlet-Vcenter

Vol

t

time (s)

Fig. 8. (Color online) Principle of monitored cells: drying case.

Thanks to the measurement of the three voltagesrepresenting the three key parts of a stack, this fault de-tection method reveals to be very sensitive. It allows de-tecting early fuel cell failure, before it completely affectsthe entire stack behavior. It requires a very low numberof sensors, is a non-intrusive technique and also does notperturb the stack functioning. Moreover, it could be easilyimplemented in embedded applications.

4 Experimental validation

Experimentations have been performed on a real FC stackin the FC LAB in Belfort. The PEMFC is a 500 W, 20-cell,100 cm2 stack from ZSW-UBZM manufacturer (Fig. 10).The objectives of these experiments are:– To control that inlet and outlet voltage cells are nega-

tively influenced during flooding.

2040

stack curent

Am

ps

121416

stack voltage

Vol

t

-0.50


Vol

t

2 3 4 5 6 7 8 9 10-0.5

00.5

differential voltage Voutlet-Vcenter

Vol

t

time (s)

Fig. 9. (Color online) Principle of monitored cells: healthystack with current variation.

– To visualize which cells are flooded in each case.– To choose cells representative of FC state of health and

instrument them.– To apply the differential method.

Many flooding experiments are performed at differentoperating conditions: different stack temperatures and hy-grometry rates. Some experimental results are summa-rized in Table 3. It presents the cells with the highervoltage degradation due to flooding during all the experi-ments. White table cell corresponds to cell voltage above0.6 V, colored table cell means that cell voltage drops be-tween 0.5 and 0.6 V. Finally, colored cell marked with across signifies cell voltage drops under 0.5 V.

This table points out that cells 4, 5, 6 and 18, 20 arealways influenced by flooding. These results are in goodagreement with previous observations. Ramousse et al.[26] experiment multiple floodings on a 65-cell stack: they

23412-p7


Table 3. (Color online) Most impacted cell voltages during tests.

Cell 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20Operating conditions

1 Temp 50 °C HR = 100% I = 30 A x x2 Temp 50 °C HR = 100% I = 40 A x x3 Temp 50 °C x x4 Temp 50 °C HR = 100% I = 40 A5 Temp 40 °C HR = 45% I = 40 A x x6 Temp 40 °C HR = 56% I = 40 A x x x7 Temp 40 °C HR = 75% x x x8 Temp 40 °C HR = 96% x x9 Temp 40 °C HRvar x x

Legend: Vcell

< 0.6 V HR: air inlet hygrometry x V

cell < 0.5 V HRvar: variable air hygrometry from 25% to 100%

I: fuel cell current Temp: fuel cell temperature

I = 10 A

I = 40 A

I = 40 A

I = 40 A

HRvar

Fig. 10. (Color online) Picture of the instrumented 20-cellPEMFC stack under test.

observe that the worst voltage degradation always affectsthe same cells. They show that this performance mismatchis due to a non-uniform thermal distribution. The gas flowdistribution also plays an important role in the floodingprocess [40]. In theory, all cells have the same air flow dis-tribution but in reality bipolar plates are slightly different[41]. Moreover, the clamping pressure, due to stack assem-bly, modifies the air flow and plays an important role inFC performance and flooding phenomenon [21,42]. Thus,because of these material differences, some cells are moreprone to flooding than others. Then we can consider thatthese cells are representative of FC state of health.

During the tests, in many occurrences, cell voltage 13presents voltage degradation. This is not in agreementwith previous observations, but an explanation could bethe FC aging. The stack has been previously tested forabout 500 h under different operating conditions. In nor-mal conditions, this cell voltage was already lower thanothers; we suspect a possible degradation of its material.

For example an increase of its channel friction wouldimply a lower gas flow in this cell and then an importantflow disparity [41]. But we cannot exclude possible flood-ing space cyclicity where, in long cell assembly, some cellscould be affected by flooding in a cyclic location alongthe stack. Tests on other stacks should be performed inparticular on large stack to validate this hypothesis.

As explained in the previous section, it is judiciousto monitor only three groups of cell voltage in order todetect a flooding or a drying. The choice of the numberof measured cells is in fact a trade-off between easy sen-sor implementation by measuring sufficient voltage am-plitude and the need of a good signal/noise ratio. In ourcase, three cells are monitored in the inlet (V4+V5+V6named Vinlet), three cells in the center (V11+V12+V13named Vcenter) and three cells in the outlet of the stack(V18+V19+V20 named Voutlet).

For the next step of this study, Figures 11–13 focus onthe particular experiment number 5, performed under aflooding fault with the following conditions:

– Temperature of the stack: 40 ◦C.– FC current: 40 A.– Air inlet hygrometry rate: 45%.

Each cell voltage is depicted in Figure 11. Firstlyflooded cell voltages drop smoothly because of water ac-cumulation in the gas diffusion layer (GDL), and then thevoltages behave regarding the characteristic of the flood-ing signature. Indeed, water droplet formation blocks thegas channel and causes the voltage drop owing to localgas starvation. Meanwhile, the inlet pressure increases.Finally gas pressure evacuates the droplet; as a resultthe voltage increases until the next droplet is generated.Water droplet formation and extraction, associated withlocal gas starvation [43], cause cell voltage oscillations.

At t = 0.12 h, the cell voltage 18 drops under thesafety level set to 380 mV. Thus, so as not to degradeirreversibly the catalysts of FC materials [44,45], the stackis disconnected.

23412-p8


0 0.02 0.04 0.06 0.08 0.1 0.12 0.14

0.4

0.5

0.6

0.7

0.8

0.9

1

Time [h]

Vol

tage

s [V

]

Individual cell voltages evolution during the test

Cell 5

Cell 18

Fig. 11. (Color online) Evolution of individual cell voltages.

0 2 4 6 8 10 12 14 16 18 20

0.4

0.5

0.6

0.7

0.8

0.9

1individual cell voltage evolution during the test

Cell number

Vol

tage

s [V

]

OCV

Fig. 12. (Color online) Dispersion of individual cell voltages.

The cell voltage discrepancy is shown in Figure 12.This figure illustrates the voltage drop during all the ex-periments. Thus two areas can be identified:

– When FC is delivering current: cell voltages are under0.7 V.

– When FC is disconnected: cell voltages are on OCV(i.e., above 0.9 V).

The voltage drop of the cell number 4, 5, 13, 18, and20 is clearly observed.

The two differential voltages are plotted in Figure 13 inthe two upper subplots. Because the flooding takes placein the inlet and outlet cells, the differential measures arenegative and slightly drop according to the flooding de-gree. This figure points out that the differential measurescould detect the fault with a threshold set to −0.2 V, be-fore the stack is disconnected because of a low cell voltage.This behavior is in agreement with the simulation study.It should be noted, that a fuel cell stack is affected by

-0.4-0.2

0Vinlet-Vcenter

Vol

t

-0.4-0.2

0

Voutlet-Vcenter

Vol

t

-0.050

0.05

Vinlet-Vcenter with high pass filter

Vol

t

0 0.02 0.04 0.06 0.08 0.1 0.12 0.14-0.05

00.05

Voutlet-Vcenter with high pass filter

Vol

t

time (h)

Fig. 13. (Color online) Differential and filtered differentialmeasures for cell voltages with their thresholds.

many degradation processes that are not yet well under-stood. Nevertheless the aging phenomenon leads to a slowvoltage decrease [14,46] and an inhomogeneous voltagedistribution along the stack even in nominal conditions.Consequently an aged fuel cell will present a mismatch dif-ference voltage in normal operation mode. Hence, the abil-ity to discriminate between aging and malfunction leadsto define a high level threshold. However this tuning hastwo drawbacks: on the one hand it may delay the detectionalarm, on the other hand it may not be accurate enoughto perform the detection.

Figure 11 also points out the erratic voltage behaviorowing to the water droplet. That is the reason why wesuggest to filter differential measures with a high pass fil-ter in order to get rid of the slow voltage derivation andonly visualize the voltage oscillation due to water propa-gation. As droplet formation and extraction are slow phe-nomena, high pass cut-off frequency is set to 0.8 mHz.Filtered measures are displayed in Figure 13 in the twolower subplots. They highlight an increase of the oscilla-tion magnitude along with the flooding. This could be anew criterion to improve flooding detection in terms of ra-pidity and robustness regarding rejection of perturbation(load variation and aging effect). For instance a 0.04 Vthreshold allows a faster detection compared to the safetylevel procedure which requires each cell monitoring. Thefiltered data reveal that the proposed detection method istwice faster.

To conclude, the differential method is a non-intrusivefault detection technique. It requires only three voltagesensors, is easy to implement and is cost-effective.Namely it allows to detect precisely and quickly flood-ing phenomenon (illustrated experimentally) without anysophisticated instrumentation or complex data post-processing. Study needs to be improved by performingEIS in order to prove that the PEMFC is under flood-ing condition. Further investigations have to be done oncell number 13 about its low voltage and its irregularbehavior.

23412-p9


5 Conclusion and perspectives

The objective of the paper was to investigate voltage mon-itoring for the detection of common faults occurring in aPEMFC stack. To enhance stack lifetime it is importantto realize fast fault detection identification (FDI) in orderto remain as close as possible to the optimal operatingconditions. Additionally, FDI implementation should besimple, reliable and non-intrusive. In mass market appli-cations, the cost is also an important criterion.

Monitoring cell voltages gives information on the stateof health but its implementation in a power stack can berather complex because of the high number of cells. Sub-sequently, space distribution discrepancies throughout theFC stack when a fault occurs provide promising and en-couraging results regarding cost issues, low number of sen-sors and quality of delivered information. The proposeddifferential method, using judiciously placed sensors, wastested during flooding conditions. It gives a fast and rel-evant feedback on the fault detection and identification.Its main advantage relies on its principle: indeed it di-rectly uses the stack as a sensor and takes advantage ofits electrical signals while operating.

Other techniques like electrochemical impedance spec-troscopy (EIS) or current interrupt (CI) also monitor cellvoltages, but are founded on a “perturb and observe” ap-proach. On the one hand the CI method strongly disturbsthe output power delivered to the load and hence requiresan important auxiliary storage device. On the other handEIS is based on small AC signal injection and weakly per-turbs the stack power. Its main drawback relies on signalanalysis which is based on steady state assumption andneeds time to be performed. For the drying case, the mem-brane resistance monitoring makes sense. It can be eval-uated with a high frequency disturbing signal inherentlygenerated by the interface DC-DC converter for instance.That is the reason why EIS is effective and fast in thisoccurrence. However flooding circumstances are initiatedby MEA water droplet and involves mass-transportationphenomenon. This fluidic phenomenon has a high timeresponse and thus needs very small frequency superim-posed AC perturbation. Moreover signal processing en-abling flooding detection is also time-consuming. It makesEIS flooding recognition long and for this reason very sen-sitive to load changes.

On the contrary, our study demonstrates that the dif-ferential technique does not modify the output power andprovides fast detection because it relies directly on thestack space signature. Tests performed on a 20-cellPEMFC stack validate this original differential methodespecially for flooding. Moreover, the differential methodcan also be implemented on-line for embedded applica-tions (vehicle, demonstrator platform. . . ) to give a quickfeedback on the fuel cell state of health.

As a perspective, the method should be experimentedon other fuel cell stacks to be widely validated: on re-cently manufactured or non-aged stacks and particularlyon larger multi-cell stacks. Thus, it should be interest-ing to investigate possible space cyclicity throughout alarger fuel cell stack. The differential method using voltage

sensors could be improved by adding thermal sensors inthe bipolar plates to get a thermal cartography of thestack, in order to improve the information content. More-over it should be interesting to perform 3D simulationscoupling together fluidic, thermal and water phenomena.At last, it would be appropriate to compare EIS and thedifferential method in terms of detection quickness andload perturbation rejection. It might lead to combine EISfor drying detection and the differential method forflooding.

References

1. J. Larminie, A. Dicks, Fuel Cell Systems Explained (JohnWiley & Sons, Chichester, 2003)

2. H. Li, Y. Tang, Z. Wang, Z. Shi, S. Wu, D. Song, J. Zhang,K. Fatih, J. Zhang, H. Wang, Z. Liu, R. Abouatallah,A. Mazza, J. Power Sources 178, 103 (2008)

3. S. Knights, J. Power Sources 127, 127 (2004)4. X. Xue, J. Tang, N. Sammes, Y. Ding, J. Power Sources

162, 388 (2006)5. N. Yousfi-Steiner, P. Mocoteguy, D. Candusso, D. Hissel,

A. Hernandez, A. Aslanides, J. Power Sources 183, 260(2008)

6. B. Wahdame, D. Candusso, F. Harel, X. Francois, M.-C.Pera, D. Hissel, J.-M. Kauffmann, J. Power Sources 182,429 (2008)

7. S. Wasterlain, D. Candusso, D. Hissel, F. Harel, P.Bergman, P. Menard, M. Anwar, J. Power Sources 195,984 (2010)

8. J. Wu, X. Yuan, H. Wang, M. Blanco, J. Martin, J.Zhang, Int. J. Hydrogen Energy 33, 1735 (2008)

9. C. Brunetto, A. Moschetto, G. Tina, Electr. Power Syst.Res. 79, 17 (2009)

10. S. Wasterlain, Ph.D. thesis, Universite de Franche-Comte,Belfort, France, 2010

11. M. Hinaje, I. Sadli, J.-P. Martin, P. Thounthong, S. Rael,B. Davat, Int. J. Hydrogen Energy 34, 2718 (2009)

12. D. Webb, J. Power Sources 103, 54 (2001)13. G. Mulder, F. Deridder, P. Coenen, D. Weyen, A.

Martens, Int. J. Hydrogen Energy 33, 5728 (2008)14. N. Guillet, S. Didierjean, A. Chenu, C. Bonnet, P. Carre,

B. Wahdame, L. Dumercy, X. Francois, L. Girardot, F.Harel, D. Hissel, S. Besse, S. Boblet, V. Chaudron, A.De Bernardinis, G. Coquery, S. Escribano, N. Bardi, Sci-entific and Technological Progress Toward the Develop-ment of an 80 kWe PEM Fuel Cell System for TransportApplications, in EVS’07 23rd Electrical Vehicles Sympo-sium, Anaheim, California, USA, 2007

15. C. Bonnet, S. Didierjean, N. Guillet, S. Besse, T.Colinart, P. Carre, J. Power Sources 182, 441 (2008)

16. A. De Bernardinis, F. Harel, L. Girardot, D. Candusso,D. Hissel, X. Francois, G. Coquery, S. Besse, Experimen-tal Dynamic Performance of a 30 kW 90Cell PEFC Stackunder Transportation Load Cycle Constraints, in EFCF2009, Lucerne, Switzerland, 2009, pp. 1–16

17. J.-P. Poirot-Crouvezier, F. Roy, GENEPAC Project:Realization of a fuel cell stack prototype dedicated to theautomotive application, in WHEC, Lyon, France, 2006,pp. 1–15

23412-p10


18. K. Sekizawa, N. Kitamura, K. Manabe, Y. Nonobe, M.Kizaki, K. Kojima, ECS Trans. 33, 1947 (2010)

19. A. De Bernardinis, M.-C. Pera, J. Garnier, D. Hissel,G. Coquery, J.-M. Kauffmann, Energy Convers. Manage.49, 2367 (2008)

20. P. Lin, P. Zhou, C.W. Wu, J. Power Sources 194, 381(2009)

21. P. Rodatz, F. Buchi, C. Onder, L. Guzzella, J. PowerSources 128, 208 (2004)

22. M. Miller, A. Bazylak, J. Power Sources 196, 601 (2011)23. W.H. Zhu, R.U. Payne, D.R. Cahela, B.J. Tatarchuk,

J. Power Sources 128, 231 (2004)24. J. Jang, H. Chiu, W. Yan, W. Sun, J. Power Sources 180,

476 (2008)25. R. Eckl, W. Zehtner, C. Leu, U. Wagner, J. Power Sources

138, 137 (2004)26. J. Ramousse, K.P. Adzakpa, Y. Dube, K. Agbossou,

M. Fournier, A. Poulin, M. Dostie, J. Fuel Cell Sci.Technol. 7, 041006 (2010)

27. A.J. del Real, A. Arce, C. Bordons, J. Power Sources 173,310 (2007)

28. V. Vishnyakov, Vacuum 80, 1053 (2006)29. C. Wang, M.H. Nehrir, S.R. Shaw, IEEE Trans. Energy

Convers. 20, 442 (2005)30. D. Candusso, F. Harel, A. De Bernardinis, X. Francois,

M.-C. Pera, D. Hissel, P. Schott, G. Coquery, J.-M. Kauff-mann, Int. J. Hydrogen Energy 31, 1019 (2006)

31. G. Fontes, C. Turpin, S. Astier, A large signal dynamiccircuit model of a H2/O2 PEM fuel cell: description,parameter identification and exploitation, in Proc. ofFDFC 2008, Nancy, France, 2008

32. Y. Park, J. Caton, J. Power Sources 179, 584 (2008)

33. P.K. Das, X. Li, Z.-S. Liu, Int. J. Hydrogen Energy 35,2403 (2010)

34. J.P. Owejan, T.A. Trabold, D.L. Jacobson, D.R. Baker,D.S. Hussey, M. Arif, Int. J. Heat Mass Transfer 49, 4721(2006)

35. N. Fouquet, C. Doulet, C. Nouillant, G. Dauphin-Tanguy,B. Ould-Bouamama, J. Power Sources 159, 905 (2006)

36. F. Barbir, PEM Fuel Cells (Springer, Berlin, 2005)37. A. Hernandez, R. Outbib, D. Hissel, Journal Europeen

des Systemes Automatises 42, 1225 (2008)38. P. Corbo, F. Migliardini, O. Veneri, Int. J. Hydrogen

Energy 32, 4340 (2007)39. E.A. Muller, A.G. Stefanopoulou, Analysis, Modeling,

and Validation for the Thermal Dynamics of a Poly-mer Electrolyte Membrane Fuel Cell System, in 3rd Int.Conf. on Fuel Cell Science, Engineering and Technology,Ypsilanti, USA (ASME, New York, 2005), pp. 389–404

40. J.P. Owejan, T.A. Trabold, J.J. Gagliardo, D.L.Jacobson, R.N. Carter, D.S. Hussey, M. Arif, J. PowerSources 171, 626 (2007)

41. P.A.C. Chang, J. St-Pierre, J. Stumper, B. Wetton, J.Power Sources 62, 340 (2006)

42. S.-D. Yim, B.-J. Kim, Y.-J. Sohn, Y.-G. Yoon, G.-G.Park, W.-Y. Lee, C.-S. Kim, Y.C. Kim, Curr. Appl. Phys.10, S59 (2010)

43. M. Gerard, J.-P. Poirot-Crouvezier, D. Hissel, M.-C.Pera, Int. J. Hydrogen Energy 35, 12295 (2010)

44. A. Taniguchi, T. Akita, K. Yasuda, Y. Miyazaki, J. PowerSources 130, 42 (2004)

45. A. Taniguchi, T. Akita, K. Yasuda, Y. Miyazaki, Int. J.Hydrogen Energy 33, 2323 (2008)

46. S. Knights, J. Power Sources 127, 127 (2004)

23412-p11

pem fuel cell fault detection and identification using...

Documents