Anda di halaman 1dari 47

Prognostics

Prognostics Center of Excellence


TEAM MEMBERS Kai Goebel Ph.D. (Center Lead) Vadim Smelianskyi, Ph.D. (Center Co-Director) Scott Poll (ADAPT) Edward Balaban Jose Celaya, Ph.D. Matt Daigle, Ph.D. Bhaskar Saha, Ph.D. Sankalita Saha, Ph.D. Abhinav Saxena, Ph.D. Phil Wysocki
12/09/2009

NASA Ames Research Center, Moffett Field CA 94035

Kai Goebel

April 27, 2010

http://prognostics.nasa.gov
DIAGNOSTICS & PROGNOSTICS
1

Prognostics CoE
The Prognostics Center of Excellence (PCoE) at Ames Research Center provides an umbrella for prognostic technology development, specifically addressing technology gaps within the application areas of aeronautics and space exploration. En route to becoming a national asset Expertise in prediction technology and

uncertainty management for systems


health monitoring

12/09/2009

DIAGNOSTICS & PROGNOSTICS

Prognostics
4.2 4 3.8

E (measured)

Prediction points
E (from PF)

health (V) voltage

3.6 3.4 3.2 3 2.8 2.6 0

EOD

EOD pdfs
tEOD

500

1000

1500

2000

2500

3000

3500

time (secs)

Definition: Predict damage progression of a fault based on current and future operational and environmental conditions to estimate the time at which a component no longer fulfils its intended function within desired bounds (Remaining Useful Life)
12/09/2009

DIAGNOSTICS & PROGNOSTICS

Key Ingredients for Prognostics


Run-to-failure data
Measurement data Ground truth data Operational conditions
Load profiles Environmental conditions

Failure threshold

Physics of Failure models


For each fault in the fault catalogue

Uncertainty information
Sources of uncertainty Uncertainty characterization

12/09/2009

DIAGNOSTICS & PROGNOSTICS

Prognostics Algorithms
Data Driven Algorithms
Gaussian Process Regression Relevance Vector Machine Neural Networks Polynomial Regression

Model Based Algorithms Hybrid Algorithms


Particle Filters
Classical PF Rao-Blackwellized PF Risk Sensitive PF

Kalman Filters
Classical KF Extended KF

12/09/2009

DIAGNOSTICS & PROGNOSTICS

Metrics Example
Establishing a common ground to compare different prognostic efforts Metrics must not only measure accuracy and precision but also the convergence of both properties Data/time requirement of algorithms (prognostic horizon) before they produce consistent predictions is also important

Particle Filter Prediction


1

- Metric ( = 0.2, = 0.5)


40 End of Life (EOL) 80% accuracy zone actual RUL RVM-PF RUL GRP RUL Baseline Statistics RUL

C/1 capacity (mAh); RUL pdf (biased by 0.7)

0.95 0.9

Real data

35 30

0.85 0.8

25

RUL
RUL threshold RUL pdfs 0 10 20 30 40 50 60 70

0.75 0.7 0.65

20 15 10

0.6

Time (weeks) Prediction Spread (weeks)


l l l

Prediction Horizon (weeks) 32 16 32

4.23 3.25 5.34

0 30

35

40

45

50

55

60

65

Time (weeks)

12/09/2009

DIAGNOSTICS & PROGNOSTICS

Uncertainty Management
Model System complexity Insufficient knowledge Usage Load profile Temperature Noise Internal, external Electrical, mechanical, thermal Sensors

Sources of Uncertainty

Uncertainty Management

Training data based extrapolation Probabilistic state space model Online model adaptation Noise modeling Probabilistic regression Hyperparameters to prevent overfitting

12/09/2009

DIAGNOSTICS & PROGNOSTICS

Ames Research Center

Modeling, Algorithms, Metrics

Application Example: Valves

12/09/2009

DIAGNOSTICS & PROGNOSTICS

Valve Prognostics
Apply model-based prognostics to pneumatic valves Develop high-fidelity simulation model
Progressive damage models include seal wear (internal and external leaks), spring degradation, and increase in friction.

Investigate performance under different circumstances using prognostic performance metrics for comparison
Impact of using different filters Effects of increased sensor noise Effects of increased process noise and model uncertainty Feasibility of different sensor sets (e.g. continuous position sensor vs. discrete open/closed indicators) Computational Architecture

Source: M. Daigle and K. Goebel, "Model-based Prognostics with Fixed-lag Particle Filters Accepted for publication at PHM09
12/09/2009

DIAGNOSTICS & PROGNOSTICS

Problem Formulation
Prognostics goal
Compute EOL = time point at which component no longer meets specified performance criteria Compute RUL = time remaining until EOL

System model
Output

State

Parameters

Input

Process Noise

Sensor Noise

Define condition that determines if EOL has been reached EOL and RUL defined as

Compute

and/or

12/09/2009

DIAGNOSTICS & PROGNOSTICS

10 5/

Prognostics Architecture
1
System receives inputs, produces outputs
yk Fault Detection Isolation & Identification F yk

Estimate current state and parameter values


p(xk,k|y0:k) Prediction p(EOLk|y0:k) p(RULk|y0:k)

uk

System

Damage Estimation

Identify active damage mechanisms

Predict EOL and RUL as probability distributions

12/09/2009

DIAGNOSTICS & PROGNOSTICS

11 5/

Case Study
Apply framework to pneumatic valve
Complex mechanical devices used in many domains including aerospace Failures of critical valves can cause significant effects on system function Top

Pneumatic valve operation

Pneumatic Port Return Spring Piston

Valve opened by opening bottom port to supply pressure and top Bottom port to atmosphere Pneumatic Port Valve closed by opening bottom port to atmosphere and top port Fluid Flow to supply pressure Return spring ensures valve will close upon loss of supply pressure
12/09/2009

Plug

DIAGNOSTICS & PROGNOSTICS

5 2 12 /

Case Study
Faults
External leaks at ports & internal leaks across piston Friction buildup due to lubrication breakdown, sliding wear, buildup of particulate matter Spring degradation

Defining EOL
Limits defined for open and close times of valves

Top Pneumatic Port Return Spring Piston

E.g., main fill valve opens in 20 seconds (26 req.), closes in 15 (20 req.)

Limits placed on valve leakage rates Bottom (pneumatic gas) Pneumatic Port Valve must be able to fully close upon fail-safe Valve is at EOL when any of above conditions Fluid Flow violated (defines CEOL)
Function of amount of damage, parameterized in model

Plug

12/09/2009

DIAGNOSTICS & PROGNOSTICS

13 5/

Physics-based Modeling
Valve state defined by
Valve position Valve velocity Gas mass above piston Gas mass below piston

State derivatives given by


Velocity Acceleration Gas flow above piston Gas flow below piston

Inputs given by

Fluid pressure (left) Fluid pressure (right) Input pressure at top port Input pressure at bottom port
DIAGNOSTICS & PROGNOSTICS
14 5/

12/09/2009

Physics-based Modeling: Forces


Piston movement governed by sum of forces, including
Pneumatic gas: Process fluid: Weight: Spring: Friction: Contact forces:
Valve Stroke Length

12/09/2009

DIAGNOSTICS & PROGNOSTICS

15 5/

Physics-based Modeling: Flows


Gas flows determined by choked/non-choked orifice flow equations:

Fluid flow determined by orifice flow equation:

12/09/2009

DIAGNOSTICS & PROGNOSTICS

16 5/

Pneumatic Valve Modeling


Possible sensors include
Valve position Gas pressure (top) Gas pressure (bottom) Fluid flow Open indicator Closed Indicator

where,

12/09/2009

DIAGNOSTICS & PROGNOSTICS

17 5/

Modeling Damage

Increase in friction Based on sliding wear equation Describes how friction coefficient changes as function of friction force, piston velocity, and wear coefficient

Degradation of spring Assume form similar to sliding wear equation Describes how spring constant changes as function of spring force, piston velocity, and wear coefficient

Growth of internal leak Based on sliding wear equation Describes how leak size changes as function of friction force, piston velocity, and wear coefficient

Growth of external leak Based on environmental factors such as corrosion Assume a linear change in absence of known model

12/09/2009

DIAGNOSTICS & PROGNOSTICS

18 5/

Damage Progression
4 3.5
Friction Coefficient (Ns/m)
6 x 10 Damage Progression of Friction Coefficient

x 10 1.1

10 9
Spring Constant (N/m)

x 10

Damage Progression of Spring Constant x 10 6.94 6.92 6.9 6.88 6.86 6.84 40 41
4

3 2.5 2 1.5 1 0.5 0 0 20 40 60 Time (cycles) 80 100 1.05 40 41

8 7 6 5 4

20

40 60 Time (cycles)

80

100

x 10

-6

Damage Progression of Internal Leak 3 2.5 2 1.5 1 0.5 0

-5 x 10 Damage Progression of Top External Leak

1.5

1 7 0.5 6.9 6.8 0

x 10

-7

40 0 20 40 60 Time (cycles)

41 80 100

Top External Leak Area (m2)

Internal Leak Area (m2)

20

40 60 Time (cycles)

80

100

12/09/2009

DIAGNOSTICS & PROGNOSTICS

19 5/

Damage Estimation
Wear parameters are unknown, and must be estimated along with system state
Position Velocity Gas mass above piston Gas mass below piston Friction coefficient Spring rate Internal leak area External leak area (top) External leak area (bottom) Friction wear Spring wear Internal leak wear External leak wear (top) External leak wear (bottom)
DIAGNOSTICS & PROGNOSTICS
20 5/

Augment system state with unknown parameters and use state observer

12/09/2009

Particle Filters
Employ particle filters for joint state-parameter estimation
Represent probability distributions using set of weighted samples Help manage uncertainty (e.g., sensor noise, process noise, etc.) Similar approaches have been applied successfully to actuators, batteries, and other prognostics applications
w

State represented with discrete probability distribution


x

Distribution evolves in time


t

12/09/2009

DIAGNOSTICS & PROGNOSTICS

21 5/

Damage Estimation with PF


Particle filters (PFs) are state observers that can be applied to general nonlinear processes with non-Gaussian noise
Approximate state distribution by set of discrete weighted samples:
Suboptimal, but approach optimality as N

Parameter evolution described by random walk:


Selection of variance of random walk noise is important Variance must be large enough to ensure convergence, but small enough to ensure precise tracking

PF approximates posterior as

12/09/2009

DIAGNOSTICS & PROGNOSTICS

22 5/

Sampling Importance Resampling


Begin with initial particle population Predict evolution of particles one step ahead Compute particle weights based on likelihood of given observations Resample to avoid degeneracy issues
Degeneracy is when small number of particles have high weight and the rest have very low weight Avoid wasting computation on particles that do not contribute to the approximation

x Initial Particle Population Compute Weights Predict Evolution Resample

12/09/2009

DIAGNOSTICS & PROGNOSTICS

23 5/

Prediction
Particle filter computes

Prediction n steps ahead approximated as

Similarly, EOL prediction approximated as

General idea
Propagate each particle forward until EOL reached (condition on EOL evaluates to true) Use particle weights for EOL weights
12/09/2009

DIAGNOSTICS & PROGNOSTICS

24 5/

Prediction

Hypothesized inputs

Friction progression EOL prediction

12/09/2009

DIAGNOSTICS & PROGNOSTICS

25

25 5/

Validation of Methodology

Internal Leak EOL Predictions

Estimate of wear parameter converges after a few cycles, after this, leak area can be tracked well.

EOL predictions all contain true EOL, and get more accurate and precise as EOL is approached.

12/09/2009

DIAGNOSTICS & PROGNOSTICS

26

26 5/

- Performance
Plot summarizes performance of internal leak prognosis Over 50% of probability mass concentrated within the bounds at all prediction points except at 20 and 30 cycles
Mean RULs are within the bounds at these points
=0.1, =0.5

For =0.122, metric is satisfied at all points

12/09/2009

DIAGNOSTICS & PROGNOSTICS

27

27 5/

Prediction Performance
- metric for spring damage prediction, where =0.1, =0.4, M={x,f,pt,pb} - metric for spring damage prediction, where =0.1, =0.4, M={open,closed}

Both cases have similar accuracy, but the case with continuous measurements has much better precision, as the metric evaluates to true for all but one point.

12/09/2009

DIAGNOSTICS & PROGNOSTICS

28

28 5/

Discussion
Different sensor sets have comparable estimation and prediction accuracy, but some differences are observed Wide differences observed in precision of estimation and prediction Results reveal that some sensors are more useful for certain faults than others
Flow measurement can be dropped with little effect For friction and spring faults, sensor sets with position measurement perform best For leak faults, sensor sets with pressure measurements perform best Helps decide importance of sensors based on which faults are most important

Overall performance still reasonable with higher levels of noise


Sensor sets with continuous measurements impacted most

12/09/2009

DIAGNOSTICS & PROGNOSTICS

29

29 5/

Ames Research Center

Application Example: Electronics

12/09/2009

DIAGNOSTICS & PROGNOSTICS

30

Prognostics for Electronics


Line Replaceable Unit: Power Controller

Component: Power Transistor

12/09/2009

DIAGNOSTICS & PROGNOSTICS

31

Motivation
Electronic components have increasingly critical role in onboard, autonomous functions for
Vehicle controls, Communications, Navigation, Radar systems

Future aircraft systems will rely more on electric & electronic components
More electric aircraft Next Generation Air Traffic System (NGATS)

Move toward lead-free electronics and microelectromechanical devices (MEMS) Assumption of new functionality increases number of electronics faults with perhaps unanticipated fault modes Needed
Understanding of behavior of deteriorated components to develop capability to anticipate failures/predict remaining RUL
12/09/2009

DIAGNOSTICS & PROGNOSTICS

32

Current Research Efforts


Thermal overstress aging of MOSFETs and IGBTs Electrical overstress aging testbed (isothermal) Modeling of MOSFETs Identification of precursors of failure for different IGBT technologies* Prognostics for output capacitor in power supplies+ Effects of lightning events of MOSFETS Effects of ESD events of MOSFETS and IGBTs Effects of radiation on MOSFETS and IGBTs
In collaboration with
* University of Maryland + Vanderbilt University

12/09/2009

DIAGNOSTICS & PROGNOSTICS

33

Accelerated aging system


A platform for aging, characterization, and scenario simulation of gate controlled power transistors. The platform supports:
Thermal cycling Simulation of operation conditions Isothermal aging

In situ state monitoring is supported at varying gate and drain voltage levels.

12/09/2009

DIAGNOSTICS & PROGNOSTICS

34

Experiment on IGBT
Collector-emitter voltage turn-OFF transient
Collector-Emmiter Voltage (V) Collector-Emmiter and Gate (V)
10 Vce 5 Vg T = ~ 330C Vd = 4v Vg = 10v
Switching Transient peak(V)
11.5 11 10.5 Degradation time (s) 10000 9000 8000 7000 6000 5000 4000 3000 2000 7.5 1000 7 326 326.5 327 327.5 328 328.5 329 temperature (C) 329.5 330 330.5 331

10 9.5 9 8.5 8

4.15

4.2

4.25 time (s)

4.3

4.35 x 10 180 min 150 min 110 min 65 min 30 min


-5

8.2 8 7.8 7.6 7.4 4.22 4.222 4.224 4.226 time (s) 4.228 4.23

4.232 x 10
-5

Potential degradation Indicator

Turn-OFF collector-emitter voltage transient decreased significantly with both increases in temperature and thermal overstress aging time
35

12/09/2009

DIAGNOSTICS & PROGNOSTICS

Electronics Aging
Offstate ICE 0.6 0.55 0.5 0.5 0.45 0.4 0.4 0.3
IGBT Waveforms ICE

Offstate ICE measurements curve fit measurements curve fit

Offstate ICE 1 0.9 measurements curve fit

0.7

0.6

0.8 0.7 0.6 0.5 0.4


8 7 IGBT Waveforms ICE VGE

0.35 0.3 0.25

VGE

0.2

0.3 0.2

6 5 4

3 2

0.1
33.84 33.86 33.88 33.9 33.92 Time (mins) 33.94 33.96 33.98 34 34.02

0.1
85.44 85.45 85.46 85.47 Time (min) 85.48 85.49

0 86.94

86.96

86.98

87

0 33.82

87.02 87.04 87.06 Time (mins)

87.08

87.1

87.12

5.07

5.08

5.09

5.1

5.11 5.12 Time (min)

5.13

5.14

5.15

5.16

86.88

86.89

86.9 86.91 Time (min)

86.92

86.93

0 -0.5

x 10

Coeff of Polyfit Term P 1t3 8

x 10

Coeff of Polyfit Term P 2t2


Particle Filter Parameter Estimation
1

6 -1 -1.5 -2 -2.5 0 -3 -3.5 0 -2 4


P1

0.5

-0.5

-1

-1.5

-2

-2.5 0

10

20

30

40

50

Prediction @ 51.875 mins

60

70

80

Time (mins)

10

20

30

40 50 Time (mins)

60

70

80

90

10

20

30

40 50 Time (mins)

60

70

80

90

12/09/2009

x 10

Coeff of Polyfit Term P 3t 20

x 10

10

Coeff of Polyfit Term P 4

DIAGNOSTICS & PROGNOSTICS

36

Ames Research Center

Application Example: EMA

12/09/2009

DIAGNOSTICS & PROGNOSTICS

37

Electro-Mechanical Actuators
Physical Modeling

EMA
5 metric ton load capacity
Data collection in laboratory

Model Verification

and flight conditions


309

Temperature Plot: Baseline Run 2


Motor TC Tsurf (Model)

Temperature (K)

306.5

0.5

304

301.5

-0.5

12/09/2009

Mission Statement

299 0

500

1000

1500

2000

2500

3000

3500

-1 4000

Time (s)

Error (K)

DIAGNOSTICS & PROGNOSTICS

38

Flyable Electro-mechanical Actuator (FLEA) Testbed


One load actuator and two test actuators (nominal and faulty), switcheable in flight Sensor suite includes accelerometers, current sensors, position sensors, temperature sensors and a load cell

Real-time flight surface loads simulation and data recording Data collection flights performed on C-17 (DFRC) and planned on UH-60 (ARC)
12/09/2009

DIAGNOSTICS & PROGNOSTICS

39

Ames Research Center

Demonstration

Application Example: Energy Storage Devices


12/09/2009

DIAGNOSTICS & PROGNOSTICS

40

Prognostics HIL Testbed


Demonstrate prognostic algorithm performance
Fast Inexpensive Control of several run-to-failure parameters Interesting dynamics

Evaluate different prediction algorithms and uncertainty management schemes

12/09/2009

DIAGNOSTICS & PROGNOSTICS

41

Modeling Batteries
Freq. Domain Time Domain

1/C

Increasing Voltage E

IR drop

Eo

m=1/(rctCnf)
-> oo re re +2S2Cnf

Activation polarization Concentration polarization

Polarization Curve

re +rct

Increasing Current I
Im(Z)

Electrolyte Weakening
IncreasingOhmic ResistanceResistance Ohmic Increasing
Im(Z)

Plate Sulphation
Increasing Charge Transfer Resistance Increasing Charge Transfer Resistance

0 0

0 0

Re(Z)

Re(Z)

12/09/2009

DIAGNOSTICS & PROGNOSTICS

42

Modeling SOC
Objective: Predict when Li-ion battery voltage will dip below 2.7V indicating endof-discharge (EOD) Approach
Model non-linear electro-chemical phenomena that explain the discharge process Learn model parameters from training data Let the PF framework fine tune the model during the tracking phase Use the tuned model to predict EOD
Eo Eo-E Eo-E

mt sd

voltage

Eo-E
mt: mass transfer sd: self discharge rd: reactant depletion
o

rd

E=E -E -E -E
sd rd

mt

12/09/2009

DIAGNOSTICS & PROGNOSTICS

time

43

Modeling SOL
Objective: Predict when Li-ion battery capacity will fade by 30% indicating life (EOL) Approach
Model self-recharge and Coulombic efficiency that explain the aging process Learn model parameters from training data Let the PF framework fine tune the model during a few initial cycles Use the tuned model to predict EOL

discharge

self-recharge

R. Huggins, Advanced Batteries: Materials Science Aspects, 1st ed., Springer, 2008.

voltage

from measurements

from model

time
12/09/2009

DIAGNOSTICS & PROGNOSTICS

44

Battery Testbed
Cells are cycled through charge and discharge under different load and environmental conditions set by the electronic load and environmental chamber respectively Periodically EIS measurements are taken to monitor the internal condition of the battery DAQ system collects externally observable parameters from the sensors Switching circuitry enables cells to be in the charge, discharge or EIS health monitoring state as dictated by the aging regime

BHM

EIS: Electro-chemical Impedance Spectroscopy


12/09/2009

DIAGNOSTICS & PROGNOSTICS

45

Prognostics in Action

12/09/2009

DIAGNOSTICS & PROGNOSTICS

46

References
A. Saxena, J. Celaya, B. Saha, S. Saha, K. Goebel, Metrics for Offline Evaluation of Prognostic Performance, International Journal of Prognostics and Health Management, 001, 2010 M. Daigle and K. Goebel, Model-based Prognostics under Limited Sensing, Proceeding of IEEE Aerospace conference 2010, 2010 A. Saxena, K. Goebel, Requirements Specification for Prognostics Performance - An Overview, Proceedings of AIAA@Infotech 2010, 2010 B. Saha, K. Goebel, and J. Christophersen, Comparison of Prognostic Algorithms for Estimating Remaining Useful Life of Batteries, Transactions of the Royal UK Institute on Measurement & Control, special issue on Intelligent Fault Diagnosis & Prognosis for Engineering Systems, pp. 293-308, 2009 B. Saha, K. Goebel, S. Poll and J. Christophersen, Prognostics Methods for Battery Health Monitoring Using a Bayesian Framework, IEEE Transactions on Instrumentation and Measurement, Vol. 58, No. 2, pp. 291-296, 2009 J. Celaya, N. Patil, S. Saha, P. Wysocki, and K. Goebel, Towards Accelerated Aging Methodologies and Health Management of Power MOSFETs, Proceedings of Annual Conference of the PHM Society 2009, 2009 B. Saha, K. Goebel, Modeling Li-ion Battery Capacity Depletion in a Particle Filtering Framework, Proceedings of Annual Conference of the PHM Society 2009, 2009 K. Goebel, B. Saha, A. Saxena, J. Celaya, and J. Christophersen, Prognostics in Battery Health Management, I&M Magazine, Vol. 11, No. 4, pp. 33-40, 2008 K. Goebel, B. Saha, and A. Saxena, A Comparison of Three Data-driven Techniques for Prognostics, Proceedings of MFPT 2008 2008

12/09/2009

DIAGNOSTICS & PROGNOSTICS

47