Failure prognostics in a particle filtering framework ...Norme ISO 13381-1:2004 : " estimation of time to failure and risk for one or more existi ng and future failure modes" Main

Événement - date

Failure prognostics in a particle filtering framework – Application to a PEMFC stack

Marine JouinRafael Gouriveau, Daniel Hissel, Noureddine Zerhouni, Marie-Cécile Péra

FEMTO-ST Institute, UMR CNRS 6174, BesançonFCLAB Research Federation, FR CNRS 3539, Belfort

[email protected]

R e s e a r c h

Événement - date

Marine Jouin – Journées inter-GDRs– 12/06/2014 2R e s e a r c h

Motivations

– Fuel Cell : an alternative to traditional energies

Several application fields Transportation, µ-cogeneration, Portable devices powering, Aerospace.

No mobile parts = good reliability

Événement - date


Motivations




– Current limitations

Major limitation: lifespan still too short

Socio‐economic aspects Cost reduction of PEMFC system

Public acceptance

Technological boltsStable hydrogen supply with high purity

Hydrogen storage

Current Necessary Current Necessary8000 h ‐ transportation100 000 h ‐ stationary

PerformancesEfficency Durability

2000 – 3000 h35‐40%25‐30%

Événement - date


Motivations




– Current limitations

Major limitation: lifespan still too short

– Prognostics and Health Management (PHM) : a solution ?

Object : taking decision at the right time to optimize system use and avoid failures

Socio‐economic aspects Cost reduction of PEMFC system

Public acceptance

Technological boltsStable hydrogen supply with high purity

Hydrogen storage

Current Necessary Current Necessary8000 h ‐ transportation100 000 h ‐ stationary

PerformancesEfficency Durability

2000 – 3000 h35‐40%25‐30%

Événement - date


Failure prognostics in a particle filtering framework

1. Backgrounds

2. Feature extraction and aging modeling

3. Prognostics based on particle filters

4. Conclusion

Événement - date



1. Backgrounds- Prognostics and Health Management- Prognostics: a key element- PHM of PEMFC- First work and its limitations



4. Conclusion

Événement - date


1. Backgrounds

– Prognostics and Health Management (PHM)

Signals transformations: extraction / selection / descriptors generation

Data coming from sensors or transducers

System state of health, comparison of descriptors on-line / expected, detection and location of failuresCause of failure, isolation et

identification of the component responsible of the failure

Prediction of the future states of the system, RUL estimates

Recommended actions to accomplish the mission (maintenance, command)

Human-machine interface

Événement - date


1. Backgrounds

– Prognostics: a key element

Pronostic ≈ RUL estimates (Remaining Useful Life) Norme ISO 13381-1:2004 : " estimation of time to failure and risk for one or more existing and future failure

modes"

Main objectives

Estimation of the Remaining Useful Life(RUL)

Estimation of the probability of failure of the system at a given date

Taking into account uncertainty is a major issue Uncertainty / system Uncertainty / its use Uncertainty / sensors Uncertainty / prognostic model defined

RUL

t fail.0 tc RUL

t fail.0 tc RUL

t fail.0 tc0 tc

prob

state

S1

S2

S3

0,5

0,3

0,2 prob

state

S1

S2

S3

0,5

0,3

0,2

Événement - date


1. Backgrounds


Different approaches

Prognostic

Hybrid approaches

Data driven approachesModel-based approaches

• Transformation of raw data into behavioral models (learning)

• No degradation model a priori

• Good ability to catch nonlinearities

• Require a huge amount of data

• Analytical models of nonlinear phenomena

• Need a small quantity of data

• High computational cost

• Default / system specific

• Models hard to develop

• Benefit from the advantages of both approaches

• Better model learning

• Better uncertainty management

• Can be complex and computationally expensive

Événement - date


1. Backgrounds


Prognostics objective: illustration

time

Degradation level

Événement - date


1. Backgrounds



time

Degradation level

Failure threshold

Critical threshold before failure

Événement - date


1. Backgrounds



Degradation level

Starting point of prediction : tp

Failure threshold


time

Événement - date


1. Backgrounds



time

Degradation level

End of lifeCritical threshold reached


RUL

RUL pdf

Failure threshold


Learning

Événement - date


1. Backgrounds



time

Degradation level

End of lifeCritical threshold reached


RUL

RUL pdf

Failure threshold


time

RUL

tp1 tp2

RUL 1

RUL 2

Learning

Événement - date


1. Backgrounds

– PHM of PEMFC

M. Jouin, R. Gouriveau, D. Hissel, M-C. Péra, and N. Zerhouni, “Prognostics and health management of PEMFC state of the art and remaining challenges,” International Journal of Hydrogen Energy, vol. 38, no. 35, 15 307 – 15 317, 2013

Événement - date


1. Backgrounds

– PHM of PEMFC: challenges pointed out

Degradation, lossesand behavior

L1

L2

L3

L4

L5

L6

L7

Complex system

Data Acquisition

Data processing

Condition Assessment

Diagnostics

Decision Support

Human-Machine Interface

Observe

Model / Analyze

Decide Fault tolerant, self-adaptative and reconfigurable control system

Verification and validation procedures

Extended framework for detection and diagnostics approaches

Advanced prognostics models

Reliable, non-intrusive, non-damaging observation techniques

Easily implementable technology (cost, volume, online, etc.)

[1] M. Jouin, R. Gouriveau, D. Hissel, M-C. Péra, and N. Zerhouni, “Prognostics and health management of PEMFC state of the art and remaining challenges,” International Journal of Hydrogen Energy, vol. 38, no. 35, 15 307 – 15 317, 2013

Prognostics

Événement - date


1. Backgrounds

– PHM of PEMFC: challenges pointed out

Degradation, lossesand behavior

L1

L2

L3

L4

L5

L6

L7

Complex system

Data Acquisition

Data processing

Condition Assessment

Diagnostics

Prognostics

Decision Support

Human-Machine Interface

Observe

Model / Analyze

Decide Fault tolerant, self-adaptative and reconfigurable control system

Verification and validation procedures

Extended framework for detection and diagnostics approaches

Advanced prognostics models

Reliable, non-intrusive, non-damaging observation techniques

Easily implementable technology (cost, volume, online, etc.)

[1] M. Jouin, R. Gouriveau, D. Hissel, M-C. Péra, and N. Zerhouni, “Prognostics and health management of PEMFC state of the art and remaining challenges,” International Journal of Hydrogen Energy, vol. 38, no. 35, 15 307 – 15 317, 2013

Événement - date


1. Backgrounds

– First work and its limitations

Prognostics based on particle filters with simple empirical

aging models to predict the voltage degradation

3 models tested1. Linear2. Exponential3. Linear + logarithmic

Promising results with the 3rd but too much uncertainty on the results

Main limit = disturbances induced by characterizations

not taken into account

[2] M. Jouin, R. Gouriveau, D. Hissel, M-C. Péra, and N. Zerhouni, “Prognostics of PEM fuel cell in a particle filtering framework,” International Journal of Hydrogen Energy, vol. 39, no. 1, pp. 481 – 494, 2014

100 200 300 400 500 600 700 800 900-100

0

100

200

300

400

500

600

700

800

900

1000Predicted RUL comparison FC 2

Time in hours

RU

L

Linear model

Exponential model

Logarithmic model

Real RUL

90% confidence interval

Événement - date


1. Backgrounds

– First work and its limitations

Prognostics based on particle filters with simple empirical

aging models to predict the voltage degradation

3 models tested1. Linear2. Exponential3. Linear + logarithmic

Promising results with the 3rd but too much uncertainty on the results

Main limit = disturbances induced by characterizations

not taken into account

[2] M. Jouin, R. Gouriveau, D. Hissel, M-C. Péra, and N. Zerhouni, “Prognostics of PEM fuel cell in a particle filtering framework,” International Journal of Hydrogen Energy, vol. 39, no. 1, pp. 481 – 494, 2014

100 200 300 400 500 600 700 800 900-100

0

100

200

300

400

500

600

700

800

900

1000Predicted RUL comparison FC 2

Time in hours

RU

L

Linear model

Exponential model

Logarithmic model

Real RUL

90% confidence interval

PROBLEM ADDRESSED HERE

Événement - date



1. Backgrounds

2. Feature extraction and aging modeling- Principle- Modeling


4. Conclusion

Événement - date


2.Feature extraction and aging modeling

– Principle

1. Observation of the data

Different continuous aging parts separated by characterizations

Recovery observed after characterizations

Same trends of all the continuous aging parts but with an acceleration of the degradation

2. Selection of a model for continuous aging parts

Global model selected fromprevious work: P(t) = - a.ln(t) – b.t + c

Continuous aging parts

Characterizations

Événement - date



– Principle

3. Identification of coefficients a & b of the model on each part by robust least square fitting

0 20 40 60 80 100 120 140198

199

200

201

202

203

204

205

206

xdata

ydat

a

0 20 40 60 80 100 120198

199

200

201

202

203

204

205

xdata

ydat

a

0 10 20 30 40 50 60 70 80 90174

176

178

180

182

184

186

188

190

192

194

xdata

ydat

a

……

Part 1: a1 & b1

Part 2: a2 & b2

Part n: an & bn

Événement - date



– Principle

4. Feature extraction from ai & bi i є [1, n]

5. Extraction of the recovery

0 200 400 600 800 1000 1200 1400 1600 18003.2

3.25

3.3

3.35

3.4

3.45

temps

Vre

cup

fitted curveVrecup

Événement - date



– Modeling

Main objective: choosing models that are close to the data but can be justified by phenomena occurring within the stack

Global model for power aging: P(t) = - a.ln(t) – b.t + c

Models built thanks to feature extraction

for coefficient a aging: a(t) = a1.exp(a2.t) + a3.exp(a4.t)

for coefficient b aging: b(t) = b1.exp(b2.t) + b3

for recovery aging: R(t) = r1.exp(r2.t) + r3.exp(r4.t)

Événement - date



1. Backgrounds


3. Prognostics based on particle filters- Data available- Development hypotheses- Problem formalization and adaptation- Particle filtering approach- Results

4. Conclusion

Événement - date



– Data available

data set: power degradation through time aging tests on a stack of 5 cells, 100cm²

FC : 1750h at constant current solicitation of 60 A

FCi

t = 1750h

0.6A/cm²

Événement - date



– Development hypotheses

/ FC aging Degradation

– Irreversible with a long time constant– Not measurable directly (simply) deductible from another variable

Examples of possible candidates– Electrodes active surface area degradation – H2 crossover through the membrane

/ Functioning Constant current solicitation Constant operating conditions

/ Study framework Opening applicative limits: model

– Non-exact (unknown coefficients)– Non-stationary (time varying)– Non-linear– Non Gaussian noise

Aging observation through power evolution

Bayesian tracking

Événement - date



– Problem formalization

Formulation

Hidden state model Degradation state

Observation model Available measurements

Optimal Bayesian solution

Initial state distribution p(x0 | z0) ≡ p(x0) Obtaining of p(xk | z1:k) in 2 steps

1, ,k k k kx f x

,k k kz h x

1: 1 1 1 1: 1 1( / ) ( / ). ( / ).k k k k k k kp x z p x x p x z dx 1: 1

1:1: 1

( / ). ( / )( / )

( / )k k k k

k kk k

p z x p x zp x z

p z z

Événement - date



– Problem formalization

Formulation

Hidden state model Degradation state

Observation model Available measurements

Optimal Bayesian solution

Initial state distribution p(x0 | z0) ≡ p(x0) Obtaining of p(xk | z1:k) in 2 steps

1, ,k k k kx f x

,k k kz h x

1: 1 1 1 1: 1 1( / ) ( / ). ( / ).k k k k k k kp x z p x x p x z dx 1: 1

1:1: 1

( / ). ( / )( / )

( / )k k k k

k kk k

p z x p x zp x z

p z z

Problem adaptation

Modeling

Aging models developed earlier

Voltage and current measurements of the stack

Solving : particle filtering

Événement - date



– Particle filtering approach

Événement - date




Principle

Filters associations to include characterizations

Filter 1: power aging P

Filter 2: coefficient a

Filter 3: coefficient b

Filter 4: recovery R

Feature extraction

Filters initialization

Prognostics by PF

LEARNING

Raw data

Behavior predictionRUL

PREDICTION

Événement - date




Filters interactions

Threshold for learning or prognostics end not reached

Is a characterization scheduled ?

Update P with particles from models a, b & R

Filter 1

Yes

No

Filter 2

Filter 3

Filter4

P, a, b, R

t = t+1

Événement - date



– Results

Behavior prediction (1/2)

0 100 200 300 400 500 600 7000

0.5

1

1.5

t

a

0 100 200 300 400 500 600 7000

0.02

0.04

0.06

0.08

0.1

t

b

0 100 200 300 400 500 600 70050

100

150

200

250

t

R

0 100 200 300 400 500 600 7000

100

200

300

t

P

Learning of 500 hours

Prediction ended too early around 620 h

Événement - date



– Results

Behavior prediction (2/2)

0 200 400 600 800 1000 1200 1400 1600 18000

1

2

3

t

a

0 200 400 600 800 1000 1200 1400 1600 18000

0.05

0.1

0.15

0.2

t

b

0 200 400 600 800 1000 1200 1400 1600 1800190

195

200

205

210

t

R

0 200 400 600 800 1000 1200 1400 1600 1800

160

180

200

220

t

P

Learning of 1300 hours

Good prediction

Événement - date



– Results

Behavior prediction: discussion

MAPE during learning and prediction

Événement - date



– Results

Behavior prediction: discussion

Feature extraction change with the length of the learning: illustration on recovery and one coefficient of the power model

Événement - date



– Results

RUL estimates

Événement - date



1. Backgrounds



4. Conclusion

Événement - date


4. Conclusion

– Motivations Challenges

FC : technico-socio-economic stakes PHM : reliability / availability / costs – thematic of growing interest

Towards PHM of PEMFC : a lever to increase life duration

– Empirical modeling of aging Good way to represent power aging at constant current solicitation

Allows integrating recovery induced by characterizations

– Prognostics results Better prediction of power behavior

Less uncertainty in RUL estimates

But poor results if the learning is too short

– Planned expansion Take into account mission profiles / variable conditions by including the current in the models

Événement - date

Failure prognostics in a particle filtering framework – Application to a PEMFC stack

Marine JouinRafael Gouriveau, Daniel Hissel, Noureddine Zerhouni, Marie-Cécile Péra

FEMTO-ST Institute, UMR CNRS 6174, BesançonFCLAB Research Federation, FR CNRS 3539, Belfort

[email protected]

R e s e a r c h

ANR

PROPICE Summer School

Diagnostics and Prognostics of Fuel Cell Systems 01-04 July 2014, FCLAB, Belfort, France

https://propice.ens2m.fr/ecole-diag-pron-PAC.html Motivations and objectives

Fuel Cell Systems (FCS) appear to be a promising energy conversion device to face some of the economic and environmental challenges of modern society. However, even if this technology is close to being competitive, it is not yet ready to be considered for large scale industrial deployment: FCS still must be optimized, particularly by increasing their limited lifespan. Indeed, Proton Exchange Membrane Fuel Cell systems (PEMFC) usually have a life duration of around 2000 hours, whereas 6000 hours are required for some applications, including transportation... Enhancing FCS durability involves not only developing a better understanding of ageing phenomena but also requires the ability to emulate the behavior of the whole system to support the development of improvements to those systems. Prognostics and Health Management (PHM) of FCS is an emerging field of scientific and technological developments that has the potential to provide and enable improvements in the life management, use and support of Fuel Cell Systems. Objectives and program

The aim of this summer school is to provide a forum for researchers and practitioners to discuss PHM of Fuel Cell Systems, and identify actual and future research challenges in the area. Topics of “degradation mechanisms, diagnostic and prognostics of FCS”, as well as aspects related to the “social and economic challenges for a larger diffusion of FCS” will be addressed. Courses will combine:

� Academic and industrial lectures given by experts in the field; � Real case studies demonstrations with experimental manipulation on PEMFC platforms.

Program (see reverse side for more details) � Day 1: Introduction to Fuel Cell Technology � Day 2: Diagnostics and prognostics - backgrounds � Day 3: Socio-economic and industrial perspectives � Day 4: Case studies and demonstrations

Participants and registration

The school is open to both academics (from University) and professionals (from Industry). Scientists and practitioners interest in PHM technologies and application to Proton Exchange Membrane Fuel Cell (PEMFC) are encouraged to register. Registration fee (online registration, 200 €) includes:

� Summer School facilities; � Proceedings (hard copy); � Coffee breaks, daily lunches and gala dinner.

Documents

Failure prognostics in a particle filtering framework ...Norme ISO 13381-1:2004 : " estimation of time to failure and risk for one or more existi ng and future failure modes" Main