A Unified Nonlinear Adaptive Approach for Detection … Unified Nonlinear Adaptive Approach for...

A Unified Nonlinear Adaptive Approach for Detection and Isolation of Engine FaultsDetection and Isolation of Engine Faults

Liang Tang Xiaodong (Frank) ZhangJonathan DeCastroImpact Technologies, LLC

Wright State University

www.impact-tek.com © 2009, Impact Technologies, LLC – FAR 52.227-14(h) Rights in Data 1

2009 NASA GRC Propulsion Controls and Diagnostics (PCD) WorkshopDecember 8 – 10, Cleveland, Ohio

OutlineOutline

• Background/Motivationsg• Our approach• Simulation results• Future work

www.impact-tek.com 2© 2009, Impact Technologies, LLC – FAR 52.227-14(h) Rights in Data

Motivations of This Programot at o s o s og a

Many existing EHM methods have the following disadvantages:

• deal with component faults and sensor faults separatelyhigh false alarm rate or unnecessary maintenance.

• involve linearizing the engine dynamics and blending the linear health monitors through gain scheduling

overly time-consuming and complicatedoverly time consuming and complicated

• designed at a nominal health, or non-degraded, condition lose their effectiveness as the real engine degrades over timelose their effectiveness as the real engine degrades over time

• use fixed thresholds for residual evaluation false alarm especially in transient scenarios


p y

Our ApproachOu pp oac

faults degradation

Control inputsEngine

a systematic and unified framework for sensor faults and component faults

Key Innovations

A bank of nonlinear adaptivef lt di ti ti t

for sensor faults and component faults

a direct nonlinear approach to dealing with nonlinear engine models and faults

fault diagnostic estimators

Transferable Belief Model (TBM)

Diagnostic residuals

nominal nonlinear onboard engine model enhanced by an adaptive neural network

Fault sensitivity and robustness further enhanced by adaptive thresholdsTransferable Belief Model (TBM)

based residual evaluation

Diagnostic decisionNonlinear adaptive

y p

An architecture that naturally supports distributed parallel processing


pfault diagnostic method

The Developed Diagnosis/Isolation Architecturep g

NeuralNetwork


Assumptionsssu pt o s

• Only one single subsystem (sensor/actuator/component) may be faulty at a time;faulty at a time;• Consistent with NASA IVHM goal• The approach can potentially be extended to cover multiple

simultaneous faults

• Engine health degradation is not considered as a fault, whereas a fault is an abnormal and abrupt event;

N ti t d t t ti• No assumptions on steady state operation


Fault Isolation Logicau t so at o og c

Fault detection logic: The FDI logic indicates that “a subsystem fault is detected” if the following condition is met:

• At least one component of the diagnostic residuals generated by the FDE exceeds its corresponding adaptive threshold.

Fault Isolation Logic: The FDI logic indicates that “a fault in subsystem s is isolated” if the following two conditions are both met:

All th di ti id l t d b FIE i b l th i• All the diagnostic residuals generated by FIE s, remains below their corresponding adaptive thresholds.• At least one component of the diagnostic residuals generated by each remaining FIE exceed its corresponding threshold.

* A subsystem can be either a sensor, an actuator or a component.

g p g


Adaptive ThresholdAdaptive Threshold

• The residual may change with the ti i t t l i t dtime-variant control inputs and dynamic operating conditions of the engine.

• Adaptive thresholds are derived based on the nonlinear adaptive estimation/observer algorithms and the bound on modeling uncertainty.the bound on modeling uncertainty.

• The thresholds automatically adapt to the changing engine operating

diti t h th b tconditions to enhance the robustness and fault sensitivity of the FDI scheme.


Adaptive Real-time Nonlinear Engine Modelp g

• Objective: to compensate for modeling uncertainty such as normal aging and engine-to-engine variation

control inputs sensor outputscontrol inputs sensor outputs aging and engine-to-engine variation

• A neural network is trained during flight to capture changes in engine dynamics as a result of normal aging Nominal engine model

+ +-Nominal engine model+ +-FDIE

dynamics as a result of normal aging.

• The training is done separately in parallel to the fault diagnostic estimators

Nominal engine model

Neural Network

+

t k i ht

Nominal engine model

Neural Network

+

t k i htestimators.

• The weights of neural network in the on-board engine model are updated after a number of flights

network weightsadaptive laws

adaptive real-time reference engine model

network weightsadaptive laws

adaptive real-time reference engine model

after a number of flights.

Two NNs: one being trained, another one with fixed weights used in FDIEs.


Nonlinear FDI: System Model

• Known nominal dynamics:

C SS ( )

• Modeling uncertainties: • State equation:

Enhanced C-MAPSS (healthy)

• State equation: • Output equation:

F lt d l• Fault models: • Component/actuator faults: • sensor faults:


Nonlinear FDI: Fault Models• Fault time profiles :

• Sensor fault • F is the known fault distribution vector, • θ(t) is the unknown magnitude of the time--varying sensor biasθ(t) is the unknown magnitude of the time varying sensor bias.

• Process fault :For the purpose of fault isolation, we consider a fault class given byp p g y

•

•

estimated by the p-th FIEpiθ


Adaptive thresholds are appropriately designed such that only residualsof the corresponding FIE always remain below.

Nonlinear FDI: Assumptions

• Assumption 1. The modeling uncertainties are unstructured but bounded by given known functions, i.e.,

| ( , , ) | ( , , )x u t y u tη η≤ | ( , , ) | ( , , )d x u t d y u t≤

Based on worst case scenario of normal engine degradation in a

M d li / t ( d th i b d ) d l d

Based on worst case scenario of normal engine degradation in anumber of flights

• Modeling/measurement errors (and their bounds) are modeled as a function of state and input.

• Enables the implementation of adaptive thresholds.• Improves FDI performance for transient scenarios

• Assumption 2. The rates of change of the component fault parameter vector and the sensor bias are uniformly bounded respectively


vector and the sensor bias are uniformly bounded, respectively.

Fault Detection Method• Fault detection estimator

• Fault detection decision schemeA fault is detected when at least one component of the modules of the output estimation error exceeds its corresponding threshold given byestimation error exceeds its corresponding threshold given by

Implemented as:

• Method based on nonlinear adaptive estimate techniques• Reduced to Kalman Filtering approach in case of a linear system• Adaptive threshold is a function of states/outputs and inputs


Adaptive threshold is a function of states/outputs and inputs

Fault Isolation Method• Sensor fault isolation estimators

• Adaptive Thresholds for fault isolationAdaptive Thresholds for fault isolation

• Fault parameters (severity) are estimated by the FIEsAd ti th h ld i f ti f t t / t t d i t


• Adaptive threshold is a function of states/outputs and inputs

Engine Application Using C-MAPSSg e pp cat o Us g C SS

Fan

BypassNozzleInlet

VBV

HPC HPT

CoreNozzle

FuelVSV

VBV

Combustor

LPC LPT

AmbientConditions

Core Shaft

Fan Shaft

20 2421 4130 48 501310 70 90

(state)(state)(state)


Simulation Results


Structure of the Fault Detection and Isolation E ti t (I l t d i Ph I C MAPSS)Estimators (Implemented in Phase I on C-MAPSS)

Engineu y

FDEFault Detection

Decision Scheme(w/ adaptivethresholds)

FIE #1 (Nf)

thresholds)activation

FIE #2 (Nc)

FIE #3 (T24)Fault Isolation

Decision Scheme

FIE #4 (Wf)

FIE #5 (VSV)

(w/ adaptivethresholds)


FIE #6 (HPC)

Fault Isolation Logic (Nf Sensor Bias Example)

00FIE1 (Nf) OR

100

FDE OR1

trigger00

FIE1 (Nf)

100

FIE2 (Nc) OR0 Nf f lt

Isolation indicator

0

110

FIE3 (T24) OR

0

1

1

Nf sensor fault

Isolation indicator:

Only one “0”: fault isolated100

FIE4 (Wf) OR1

1

Only one “0”: fault isolatedNo “0”: unknown faultMultiple “0”s: fault not isolable

100

FIE5 (VSV) OR

1

1


100

FIE6 (HPC) OR

Fault Scenarios & Implementationau t Sce a os & p e e tat o

• 6 faults (bias)Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) (ver. 2.0) Plot

resluts

• Sensors• Nf• Nc T 2_sens

Nc_sens

Nf_sens

imevec .DVSV' datavec .DVSV0:123

[timevec .Wf' datavec .Wf']0:326

[timevec .VBV' datavec .VBV']0:342

[timevec .TRA ' datavec .TRA ']0:543

0:56

mevec .DTamb ' datavec .DTamb0:55

[timevec .Alt ' datavec .Alt ']0:54 altitude

dTamb

Nf _sens

Nc_sens

T2_sens

TRA

DVSV

VBV

Wf cmd

• T24• Actuators

• WfVSV

OL / CL switch

OL Wf cmd

CL Wf cmdWf cmd

Ps30_sens

P2_sens

T 48 _sens

T 24 _sens

T 48 sens

T 24 _sens

T 2_sens

Nf_sens

Nc_sens

[timevec .MN' datavec .MN']0:56 Mach

Wf cmd

VSV cmd

VBV cmd

T24_sens

T48_sens

P2_sens

Ps30_sens

Nf

Nc

T2

T24

T48

VSV cmd

• VSV• Component

• HPCHealth Parameters

P50 _sensT 48 _sens

Ps30_sens

P2_sens

P50 _sens

Nf _sens

Nc_sens

T2_sensaltitude

T24 sens

Engine

VBV cmd P50_sens

Controller

T48

Ps30

P2

P50

VBV cmd

Simulation Outputs Output Scopes Initializer

T24_sensT48_sens

dTamb

P2_sensPs30_sens

P50_sens

MachWf cmd

VSV cmd


FDI

VBV cmd

Sensor Noise and Dynamics ModuleSe so o se a d y a cs odu e

Sensor Variance SLS Full value

Gaussian Noise Variance:

Nf 0.0 RPM2 2388

Nc 0.0 RPM2 9048

T2 1.0 degree R2 519 • Variance values provided by GRCT2 1.0 degree R 519

T24 5.0 degree R2 642

T48 5.0 degree R2 2070

P2 0.001 PSI2 14.6

• Small variance due to raw data processing/filtering

• Algorithm works with larger noise

Ps30 8.0 PSI2 523

P50 0.08 PSI2 19

Time constant = 1; time step = 0 015Parameters for sensor dynamics: Time constant = 1; time step = 0.015Parameters for sensor dynamics:


Transient and Degradation Scenarioa s e t a d eg adat o Sce a o

• PLA is ramped from 60o to 69o in 1 second. After steady-state operation, PLA is ramped back to 60o in 1 second. The altitude is 25000 ft, and the Mach number is 0.62.

• Simulated engine degradation: 0.2% degradations in LPC efficiency, LPC flow capacity, HPC efficiency, and HPC flow capacity


Fault Scenario 1: Nf sensor fault

20

40 residual-Nfthreshold-Nf

20

25F IE 1 (N f) b ias es t im at ion

• Bias = 20 (~1%); fault time = 35;Nf sensor bias estimation (from FIE1)Fault detected (Td=35.1)

0 10 20 30 40 50 60 70 800

0

10

20 residual-Ncthreshold-Nc 10

15

0 10 20 30 40 50 60 70 800

0 10 20 30 40 50 60 700

5

10

15FDE residuals

residual T24threshold

35 40 45 50 55 60 65 70 75 800

5

Thresholds change as operating condition changes0 10 20 30 40 50 60 70

0

10

20

30FIE 2 (NC)

residual-Nfthreshold-Nf

0

10

20

FIE 1 (Nf)


X

35 40 45 50 55 60 65 70 75 800

35 40 45 50 55 60 65 70 75 800

50

100

150

residual-Ncthreshold-Nc

35 40 45 50 55 60 65 70 75 800

35 40 45 50 55 60 65 70 75 800

20

40

60



35 40 45 50 55 60 65 70 75 80

35 40 45 50 55 60 65 70 75 800

5

10


35 40 45 50 55 60 65 70 75 80

35 40 45 50 55 60 65 70 75 800

5

10


FIE 3 (T24) FIE 4 (Wf)

Fault Scenario 1: Nf sensor fault (cont’d)X

35 40 45 50 55 60 65 70 75 800

20

40FIE 3 (T24)


20

35 40 45 50 55 60 65 70 75 800

20

40FIE 4 (Wf)

Nf: residualNc: threshold

40

XX

35 40 45 50 55 60 65 70 75 800

10

20


20 residual-T24

35 40 45 50 55 60 65 70 75 800

20

40

Nc: residualNc: threshold

10

id l T24

35 40 45 50 55 60 65 70 75 800

10

residual-T24threshold-T24

FIE 5 (VSV)

35 40 45 50 55 60 65 70 75 800

5

10


XX

35 40 45 50 55 60 65 70 75 800

20

40FIE 6 (HPC)

Nf: residualNf: threshold

FIE 635 40 45 50 55 60 65 70 75 80

0

20

40FIE 5 (VSV)


40

XX

35 40 45 50 55 60 65 70 75 800

20

40FIE 6


35 40 45 50 55 60 65 70 75 800

20

40

Nc:residualNc: threshold


35 40 45 50 55 60 65 70 75 800

5

10


35 40 45 50 55 60 65 70 75 800

5

10


Fault Scenario 2: Wf actuator faultBi 200 ( 3%) f lt ti 35

5


• Bias = -200 (~3%); fault time = 35 Wf actuator bias estimation (from FIE4)Fault detected (Td=36.1)

5 0F IE 4 (W f) b ia s e s t im a t io n

0 10 20 30 40 50 60 70 800

0 10 20 30 40 50 60 70 800

20

40


-1 0 0

-5 0

0

0 10 20 30 40 50 60 70 80

0 10 20 30 40 50 60 700

5

10

15FDE residuals

residual T24threshold 3 5 4 0 4 5 5 0 5 5 6 0 6 5 7 0 7 5 8 0

-2 0 0

-1 5 0

35 40 45 50 55 60 65 70 75 800

10

20

FIE 1 (Nf)


35 40 45 50 55 60 65 70 75 800

100

200FIE 2 (NC)

residual-Nfthreshold-NfX

35 0 5 50 55 60 65 0 5 80

35 40 45 50 55 60 65 70 75 800

20

40

60


35 0 5 50 55 60 65 0 5 80

35 40 45 50 55 60 65 70 75 800

5

10x 10

9


X

www.impact-tek.com 24© 2009, Impact Technologies, LLC – FAR 52.227-14(h) Rights in Data35 40 45 50 55 60 65 70 75 80

0

5

10


35 40 45 50 55 60 65 70 75 800

5

10


Fault Scenario 2: Wf actuator fault (cont’d)

35 40 45 50 55 60 65 70 75 800

5

10

15FIE 4 (Wf)


35 40 45 50 55 60 65 70 75 800

5

10FIE 3 (T24)


X

35 40 45 50 55 60 65 70 75 800

20

40


35 40 45 50 55 60 65 70 75 800

10

20

30


35 40 45 50 55 60 65 70 75 800

5

10


35 40 45 50 55 60 65 70 75 800

10

20


0

10

20FIE 5 (VSV)


35 40 45 50 55 60 65 70 75 800

5

10FIE 6 (HPC)


XX

35 40 45 50 55 60 65 70 75 800

35 40 45 50 55 60 65 70 75 800

20

40


35 0 5 50 55 60 65 0 5 80

35 40 45 50 55 60 65 70 75 800

20

40FIE 6



5

10


35 40 45 50 55 60 65 70 75 800

5

10


Fault Scenario 3: HPC component fault• Efficiency loss = 0 02; Flow capacity loss = 0 02; fault time = 35

5


Efficiency loss 0.02; Flow capacity loss 0.02; fault time 35Fault severity estimation (from FIE6)Fault detected (Td=35.7)

0F IE 6 (H P C ) b ias es t im at ion

0 10 20 30 40 50 60 70 800

0

20

40 residual-Ncthreshold-Nc -0 .015

-0 .01

-0 .005

0 10 20 30 40 50 60 70 800

0 10 20 30 40 50 60 700

5

10

15FDE residuals


35 40 45 50 55 60 65 70 75 80-0 .03

-0 .025

-0 .02

35 40 45 50 55 60 65 70 75 800

10

20FIE 2 (NC)


35 40 45 50 55 60 65 70 75 800

10

20

FIE 1 (Nf)


X35 0 5 50 55 60 65 0 5 80

35 40 45 50 55 60 65 70 75 800

50

100

150


35 40 45 50 55 60 65 70 75 80

35 40 45 50 55 60 65 70 75 800

20

40

60


X


0

5

10


35 40 45 50 55 60 65 70 75 800

5

10


FIE 4 (Wf)FIE 3 (T24)

Fault Scenario 3: HPC faultX

35 40 45 50 55 60 65 70 75 800

5

10

15FIE 4 (Wf)


40

35 40 45 50 55 60 65 70 75 800

5

10FIE 3 (T24)


40

X X

35 40 45 50 55 60 65 70 75 800

20


10

residual T24

35 40 45 50 55 60 65 70 75 800

20


20 residual-T24

FIE 6 (HPC)15FIE 5 (VSV)

35 40 45 50 55 60 65 70 75 800

5

0


35 40 45 50 55 60 65 70 75 800

10


35 40 45 50 55 60 65 70 75 800

5

10FIE 6 (HPC)


40FIE 6

35 40 45 50 55 60 65 70 75 800

5

10

15


40

X

35 40 45 50 55 60 65 70 75 800

20

40


10

residual T24

35 40 45 50 55 60 65 70 75 800

20


10

residual-T24


35 40 45 50 55 60 65 70 75 800

5


35 40 45 50 55 60 65 70 75 800

5

threshold-T24

Adaptive Engine Model with Neural Networks

(1) The NN learns(1) The NN learns (2) NN becomesa part of

(3) Updated model

(4) Updated residual signals

FDE

(4) Updated residual signals

u y+-

FIEs -+


Note that the NN being trained online is NOT used for FDIE.

Neural Networks Enhancement –Si l ti R ltSimulation Results

• Simulation Scenario:FanFlow loss: 0.3%; LPCFlow loss: 0.3%; HPCEff loss: 0.3%; HPCFlow loss: 0.3%

7comparions of fault detection residuals - Nf

7comparions of fault detection residuals - Nc

5

6

residual with NNresidual without NN

5

6

residual with NNresidual without NN

2

3

4

2

3

4

0 10 20 30 40 50 60 70 800

1

time (sec)

0 10 20 30 40 50 60 70 80

0

1

time (sec)


Minimal detectable/isolable fault sizeMinimal detectable/isolable fault size

Fault TypeMinimum Detectable Fault Size (% of SLS)

Minimum Isolable Fault Size (% of SLS)

Estimated Fault Size

Nf sensor bias 5 (0.21%) 10 (0.42%) 11.2

Nc sensor bias 16 (0.18%) 25 (0.28%) 23.7

T24 sensor bias 5 (0.78%) 11 (1.6%) 11.1

Wf actuator bias ‐120 (0.49%) ‐125 (0.52%) ‐110

VSV actuator bias ‐0.20 (3.39%) ‐0.25 (4.24%) ‐0.21

HPC efficiency & flow capacity loss ‐0.009 (0.9%) ‐0.015 (1.5%) ‐0.016

Note:1. Tested on C-MAPSS simulator2. Tested with the selected transient scenario only


3. Tested with the selected noise level only4. Extensive tests will be conducted in Phase 2

Future Workutu e o

• Extend the fault modes to cover more components and sensors• Further development of on-board Neural Networks-based adaptive

engine model• Further system development to cover the full flight envelopey p g p• Comparative studies with other fault diagnostics methods

• Kalman filter• Sliding mode observerSliding mode observer

• Extensive software-in-the-loop testing and performance evaluation• Further development on OEM engine model

R l ti h d i th l t ti d f l ti• Real-time hardware-in-the-loop testing and performance evaluation with OEM support


Related publicationsp

• X. Zhang, M. M. Polycarpou, and T. Parisini, "Design and analysis of a fault isolation scheme for a class of uncertain nonlinear systems " IFAC Annual Reviews in Control pp 107-121 2008

Liang Tang, Xiaodong Zhang, Jonathan DeCastro and Donald L. Simon, A Unified Nonlinear Adaptive Approach for Detection and Isolation of Engine Faults, submitted to ASME TURBO EXPO 2010, June 14-18, 2010, Glasgow, UK.

uncertain nonlinear systems, IFAC Annual Reviews in Control, pp. 107 121, 2008.• X. Zhang, T. Parisini, and M. M. Polycarpou, "A sensor bias fault isolation scheme for a class of nonlinear

systems", IEEE Transactions on Automatic Control, vol. 50, no. 3, pp. 370- 376, March 2005.• X. Zhang, T. Parisini, and M. M. Polycarpou, "Adaptive fault-tolerant control of nonlinear systems: a diagnostic

information-based approach ", IEEE Transactions on Automatic Control, vol. 49, no.8, pp. 1259-1274, August 2004, (Nominated for the Axelby Best Paper Award).

• X Zhang M M Polycarpou and T Parisini "A robust detection and isolation scheme for abrupt and incipientX. Zhang, M. M. Polycarpou, and T. Parisini, A robust detection and isolation scheme for abrupt and incipient faults in nonlinear systems", IEEE Transactions on Automatic Control, vol. 47, no. 4, pp. 576-593, April 2002.

• X. Zhang, M. M. Polycarpou, and T. Parisini, "Robust fault isolation of a class of nonlinear input-output systems", International Journal of Control, vol. 74, no.13, pp. 1295-1310, September 2001.

• DeCastro, J.A., Litt, J.S., and Frederick, D.K., “A Modular Aero-Propulsion System Simulation of a Large Commercial Aircraft Engine”, Proceedings of the 44th AIAA/ASME/SAE/ASEE Joint Propulsion Conference & Exhibit, Hartford, CT, July 21–23, 2008Exhibit, Hartford, CT, July 21 23, 2008

• DeCastro, J.A., “Rate-Based Model Predictive Control of Turbofan Engine Clearance,” J. Propulsion and Power, Vol. 23, No. 4, July–August 2007.

• L. Tang, M. Roemer, S. Bharadwaj, and C. Belcastro, “An Integrated Aircraft Health Assessment and Fault Contingency Management System, AIAA Guidance,” Navigation and Control Conference and Exhibit, Honolulu, Hawaii, 18 - 21 Aug 2008.

• L. Tang, M. J. Roemer, and G. J. Kacprzynski, “Real-Time Health Estimation And Automated FaultL. Tang, M. J. Roemer, and G. J. Kacprzynski, Real Time Health Estimation And Automated Fault Accommodation For Propulsion Systems,” GT2007-27862, Proceedings of ASME Turbo Expo 2007 Power for Land, Sea and Air, May 14-17, 2007, Montreal, Canada.

• L. Tang, M. J. Roemer, G. J. Kacprzynski, and J. Ge, “Dynamic Decision Support and Automated Fault Accommodation for Jet Engines,” IEEE Aerospace Conference, Big Sky, Montana. Mar. 3-10, 2007.

• L. Tang, A. Saxena, M. E. Orchard, G. J. Kacprzynski, G. Vachtsevanos, and A. Patterson-Hine, “Simulation-based Design and Validation of Automated Contingency Management for Propulsion Systems,” IEEE


g g y g p y ,Aerospace Conference, Big Sky, Montana. Mar. 3-10, 2007.

• M. J. Roemer, L. Tang, G. J. Kacprzynski, J. Ge, and G. Vachtsevanos, “Simulation-Based Health and Contingency Management,” IEEE Aerospace Conference, Big Sky, Montana. Mar. 4-11, 2006.

Thank you!Thank you!

Acknowledgement:

This work was supported by NASA under SBIR Phase I Contract NNX09CC71P The authors acknowledge theContract NNX09CC71P. The authors acknowledge the contributions of Don Simon and Jeffrey Armstrong from NASA Glenn Research Center and Dr. Al Volponi from Pratt & Whitney for his support on this program.


A Unified Nonlinear Adaptive Approach for Detection … Unified Nonlinear Adaptive Approach for...

Documents

Transcript of A Unified Nonlinear Adaptive Approach for Detection … Unified Nonlinear Adaptive Approach for...