ee392m fault diagnostics systems introduction gorinevsky fault diagnostics systems 1 ee392m fault...

2009Gorinevsky

Fault Diagnostics Systems 1

EE392m EE392m Fault Diagnostics Systems Fault Diagnostics Systems

Introduction Introduction Dimitry Gorinevsky Consulting Professor

Information Systems Laboratory

E392m ● Spring 2009

2009Gorinevsky


Course Subject

• Engineering of fault diagnostics systems • Embedded computer interacting with real world

– Detect abnormal operation – Fault tolerance: more than 80% of critical control code

• Operations and maintenance – More than 50% of the system lifetime costs – Troubleshooting support – Condition-based maintenance – CBM

2009Gorinevsky


Prerequisites and Course Place

• The subject is not covered in other courses• Prerequisites (helpful but not necessary)

– Stat 116; EE263 or Eng 207a; EE278 or Eng 207b

• The course is about technical approaches that are actually used in fault diagnostics applications – Survived demands of real life – Used and supported by BS-level engineers in industry – Should be accessible to a Stanford grad student

2009Gorinevsky


Course Mechanics • Class website: www.stanford.edu/class/ee392M/• Weekly seminars

– Follow website announcements • Guest lecturers from diverse industries

– Co-sponsored by NASA • Travel support for lecturers

– Lecture notes will be posted as available• Attendance • Reference texts

– Isermann; Chiang, Russel, & Braatz; Patton, Clark & Frank– Different coverage – Contact me if you have a specific interest

2009Gorinevsky


On-line (Embedded) Functions

• Embedded system, anomaly warnings – BIT – Built-in-Test– BITE – Built-in-Test Equipment

• FDIR – Fault Detection Identification and Recovery

• FT-RM– Fault Tolerance and Redundancy Management

2009Gorinevsky


Off-line Functions• Reliability

– FMECA- Failure Mode, Effects, and Criticality Analysis – Design time analysis – open loop

• Maintenance and Support – Diagnostics for maintenance– Troubleshooting support– Test equipment – CBM – Condition Based Maintenance– Pre-testing – disk drives

2009Gorinevsky


Fault Diagnostics in Industry• Space systems • Defense systems: aviation, marine, and ground • Commercial aerospace

– Aircraft, jet engines • Ground vehicles

– Locomotives, trucks, cars• High-tech

– Networks and IT systems – Disk drives– Server farms

• Process control– IC Manufacturing– Refineries– Power plants

• Oil and gas drilling

2009Gorinevsky


Guest Lecture Overview

Automotive powertrain

Military aircraft systems

Computing systems

Jet engines

Avionics of commercial aircraft

IC Manufacturing processes

Networks and IT systems

Introduction and overview

Diagnostics Application

2-Jun-0910

FordKolmanovsky26-May-099

LockheedBodden19-May-098

SunUrmanov12-May-097

5-May-096

GE InfrastructureAdibhatla28-Apr-095

HoneywellFelke21-Apr-094

IntelTuv14-Apr-093

VMTurbo (EMC) Rabover7-Apr-092

StanfordGorinevsky31-Mar-091

From LectorDate#

2009Gorinevsky


Diagnostics Methods Overview

• Shewhart chart (Control chart)• Multivariable SPC, T2

• Model-based estimation – Least squares estimation

• Integrated diagnostics – Cascaded design

2009Gorinevsky


Abnormality Detection - SPC

• SPC - Statistical Process Control – monitoring of manufacturing processes– warning for off-target quality

• Main SPC method – Shewhart Chart (1920s)

• Also see – EWMA (1940s)– CuSum (1950s) – Western Electric Rules (1950s)

2009Gorinevsky


SPC: Shewhart Control Chart• W.Shewhart, Bell Labs, 1924• Statistical Process Control (SPC) • UCL = mean + 3·σ• LCL = mean - 3·σ

Walter Shewhart (1891-1967)

sample3 6 9 1212 15

mean

qual

ity v

aria

ble

Lower Control Limit

Upper Control Limit

2009Gorinevsky


Shewhart Chart, cont’d• Quality variable assumed randomly

changing around a steady state value• Detection: y(t) > UCL=mean+3·σ• For normal distribution, false alarm

probability is less than 0.27%

P(e > 3) = 1-Φ(3) = 0.1350·10-2

P(e < 3) = Φ(-3) = 0.1350·10-2

σµ0)()( −

=tyte

UCLLCL

2009Gorinevsky


Shewhart Chart –Hypothesis Test• Null hypothesis

– given mean and covariance

• Fault hypothesis– a different mean

• Hypothesis testing

),(~)(: 200 σµNtyH

),(~)(: 2011 σµµ ≠NtyH

},{)(

10 HHXtyY

∈=given

find

00 )( pHP =

11)( pHP =

2009Gorinevsky


Bayesian Formulation • Data:

• Bayes rule

• Observation model:• Prior model:• Maximum A posteriori Probability estimate

XY Observation

Underlying State

c XPXYPYXP ⋅⋅= )()|()|()|( XYP

)(XP

( ) ( )( )4444 34444 21

indexposterior log

log|logminarg−−

−−=L

XPXYPX

Rev. Thomas Bayes(1702-1761)

2009Gorinevsky


Hypothesis Testing

• Null hypothesis

• Fault hypothesis

• Log-likelihood ratio

• Declare fault if

( ) ( )20221|log µσ

−=− yXYP ( ) 0loglog pXP −=−

( ) ( ) 02

1|log 212 =−=− µ

σyXYP ( ) 1loglog pXP −=−

( ) ( ) 0log2

110

20210 >−−=−=Λ ppyLL µ

σ

110 )1(log2 ppy −>− σµ

10 1 pp −=

σ30113.01 ⇒=p

100

1

)|()|(log LL

YHPYHP

−==Λ

2009Gorinevsky


Shewhart Chart: Use Examples

• SPC in manufacturing

• Fault monitoring

• Fault tolerance – sensor integrity monitoring

Sensor

Reference-+

Fault

|∆y| < 3σ Normalyes

no

2009Gorinevsky






2009Gorinevsky


Multivariable SPC• Univariate process

• Two independent univariate processes

( ))(1)(

)1,(1 222

cccFczP

Φ−+−Φ=−=>

( ) ( ) 22

222 ~22

2

22

21

1

11 χσµ

σµ

321321z

y

z

yz −− +=

22

02 ~ χσµ

⎟⎠⎞

⎜⎝⎛ −

=yz

( ) )2;(1 222 cFczP −=>

Chi-squared CDF

2009Gorinevsky


Multivariable SPC

• Two correlated univariate processes y1(t) and y2(t) cov(y1,y2) = Q, Q-1= LTL

• Uncorrelated linear combinationsz(t) = L·[y(t)-µ]

• Declare fault (anomaly) if ( ) ( ) 2

212 ~ χµµ −−= − yQyz T

⎥⎦

⎤⎢⎣

⎡=

2

1

yy

y

⎥⎦

⎤⎢⎣

⎡=

2

1

µµ

µ

( ) )2;(1 222 cFczP −=>

( ) ( ) 21 cyQy T >−− − µµ

2009Gorinevsky


Multivariate SPC - Hotelling's T2

• Empirical parameter estimates

( )

( )µµµ

µ

−≈−−=

≈=

∑

∑

=

=

ytytyn

Q

XEtyn

Tn

t

T

n

t

cov))()()((1ˆ

)(1ˆ

1

1

• Hotelling's T2 statistics is

• T2 can be trended as a univariate SPC variable

( ) ( )µµ −−= − )(ˆ)( 12 tyQtyT T

Harold Hotelling(1895-1973)

2009Gorinevsky






2009Gorinevsky


Least Squares Estimation

• Linear observation model:

• Fault signature model– Columns of C are fault signatures– Could be obtained from physics model

• secant method– Could be identified from data:

• regression, data mining

• Estimate – regularized least squares

vCXY +=

( ) YCrICCX TT 1ˆ −+=

Carl Friedrich Gauss (1777-1855)

2009Gorinevsky


Bayesian Estimation

• Observation model: P(Y|X)

• Prior models: P(X)

• MAP estimate:

vCXY +=

( ) YCQRCQCX TT 1111ˆ −−−− +=

),0(~ QNv

),0(~ RNX

3214434421)(log

221

)|(log

221

11minargˆ

XP

R

XYP

QXCXYX

−−

−− +−=

...)(log 121 +=− − XRXXP T

2009Gorinevsky


Model-based Residuals

Plant Model

input variables

Fault Diagnostics

prediction residual

+-

output variables

Plant Data Collected on-line

Input data

Output data

• Compute model-based prediction residualY = Yraw - f(U,X)

• If X = 0 (nominal case) we should have Y = 0. • Residuals Y reflect faults

– Sensor fault model - additive output change– Actuator fault model - additive input change

2009Gorinevsky


Example: Jet Engine Model

• Nonlinear jet engine model– static map

• Residuals

• Linearized model

vCXY += ⎥⎦

⎤⎢⎣

⎡=

driftsensor EGTleak band Bleed

iondeteriorat TurbineX

),( XUfYY raw −=

XXUfC

∂∂

−=),(

2009Gorinevsky


Example: Fault Estimates• Maintenance decision support tool

HoneywellLF507 EngineFleet

Estimates of (fault)performance parameter deterioration

1550 1600 1650 1700 1750 1800 1850 1900-1

-0.50

0.51

1.52

2.53

HP

Spo

ol D

eter

iora

tion

(%)

Sample No.

1550 1600 1650 1700 1750 1800 1850 1900-10-505

1015202530

Ble

ed B

and

Leak

age

(%)

Sample No.

1550 1600 1650 1700 1750 1800 1850 1900-40

-20

0

20

40

60

EG

T S

enso

r Offs

et (d

eg C

)

Sample No.

6-10-03

Ganguli, Deo, & Gorinevsky, IEEE CCA’04

2009Gorinevsky






2009Gorinevsky


Cascaded Design

• Increasing complexity and integration of system

• Slower time scale • Simple inner loop models• Examples

– Control systems – Estimation and data fusion – Fault diagnostics systems

Unit

Subsystem

System

UnitUnit

2009Gorinevsky


Integrated System Diagnostics• Complex integrated systems • Examples

– Aerospace vehicle, e.g. B777– Large scale computer network– Medical equipment

Subsystem 1 Diagnostics

Subsystem nDiagnostics

Integrated Diagnostic System

Decision Support Interface

Subsystem 2 Diagnostics …

2009Gorinevsky


Discrete Fault Signatures

0000000

#0Null

001011#7011000#6010001#5000001#4100101#3010110#2100110#1

#6#5#4#3#2#1Root Cause Symptom Code

kk BY =Model of root cause fault k :

2009Gorinevsky


Estimation Algorithm

• Diagnosis problem:Given data Y, diagnose root cause k

• Solution:

• Minimal Hamming distance• Justifications

– Case-based reasoning (table of fault cases)– Model-based reasoning (fault signature model)– Bayesian

1minarg kBYk −=

2009Gorinevsky


Bayesian Justification

• Data

• Observation model

{ }nn BHBHHXY :,,:,0: 1

10 K=

11)|( pHbyP kkjj −==

)|( XYP1)|( pHbyP k

kjj =≠

( )444 3444 21

2/1for ,0

110

1

1

)1log(log||)1log(<>=

−−⋅−−−⋅− ∑p

n

j

kjj ppbypn

1)|(log k

k bywcHYP −+=−

yj follows the model

deviates from the model

=−+−−=− ∑∑≠= k

jjkjj byby

k ppHYP 11 log)1log()|(log

2009Gorinevsky


Bayesian Justification

• Prior model

• MAP Estimate

)(XP ( )11)( += nHP k

( )dbywck k

k+−+=

1minarg

dHP k =− )(log

( ) ( )( )kk HPHYPk log|logminarg −−=

1minarg k

kbyk −=

2009Gorinevsky


Conclusions

• Basic diagnostics estimation methods – Are known for long time – Used in on-line systems for less time– Can be explained in several ways, e.g., Bayesian

• Engineering of fault diagnostics systems– Is new and current – Will be discussed in guest lectures– Not just diagnostics algorithms

2009Gorinevsky


Guest Lectures: Approaches

Automotive

Aircraft

Computing

Jet Engines

Aircraft

IC Manufacture

Network

Application

Model-basedFordKolmanovsky26-May-099

Fault toleranceLockheedBodden19-May-098

Multivariate SPC, MLSunUrmanov12-May-097

Multivariate estimationGE InfraAdibhatla28-Apr-095

Integrated diagnostics HoneywellFelke21-Apr-094

Fault signature IDIntelTuv14-Apr-093

Integrated diagnostics EMCRabover7-Apr-092

ApproachFrom LectorDate#

ee392m fault diagnostics systems introduction gorinevsky fault diagnostics systems 1 ee392m fault...

Documents