optimal estimation in systems health managementgorin/papers/ishm_seminar10_05... · 2005-11-22 ·...
TRANSCRIPT
Optimal Estimation in Optimal Estimation in Systems Health ManagementSystems Health Management
Dimitry GorinevskyConsulting Professor of EE
Information Systems Laboratory Stanford University
2November 2005
Acknowledgements
• Stanford: Sikandar Samar, Stephen Boyd
• Honeywell: Manny Nwadiogbu, Ossie Harris, HanifVhora, Suvo Ganguli, Shardul Deo, John Bain
3November 2005
Topic
• What is systems health management? – Estimation of system fault and damage states
from the operation data – Decision support: advisory information – Automated response: feedback control – Knowledge management: where and how we
get models for the analysis
4November 2005
Motivation
• Why system health management is important?– Maintenance servicing involves 30% of revenue
and 50% of profit across manufacturing industries (including hi-tech) and growing
– Emergence of autonomous systems. Autonomy requires automated contingency management
5November 2005
IVHM - Integrated Vehicle Health Management
• On-board fault diagnostics • Vehicles: space, air, ground, rail, marine
– Complex systems, many subsystems
6November 2005
Aircraft Fleet Maintenance
Airline EnterpriseJoint Strike Fighter
7November 2005
Semiconductor Manufacturing
• E-diagnostics and e-Manufacturing initiatives of SEMATECH, since 2001
• Maintenance of semiconductor capital equipment
8November 2005
Automotive Telematics
• Telematics– Embedded diagnostics and connectivity– Remote diagnostics and repair management
• GM’s OnStar: broadly licensed – IBM, Motorolla IS backbone– $600M revenue for 2006
• Toyota, BMW, Daimler• Everybody else follows
9November 2005
Key Problems
• Diagnostics– On-line: BIT/BITE, FDIR – Off-line: maintenance automation– Estimation and detection
• Prognostics– Trending for Condition Based Maintenance – OR: reliability-based maintenance – Prediction, statistical optimization
10November 2005
Some Prior/Other Work
• Related, but not discussed today• Discrete computational models, AI methods
– Model-based reasoning (graph models)– Case-based reasoning (machine learning)
• Pattern recognition – Discriminate discrete fault cases – Data driven approaches
• Diagnostics
11November 2005
Our Focus
• Parametric (continuous) data– Sensor data: raw or pre-processed
• Continuous faults – Deterioration trend for prognostics– Relaxation approach to discrete faults
• Controls and estimation approaches • Diagnostics and prognostics
12November 2005
Estimation Approaches
• Traditional univariate monitoring: SPC• Multivariate SPC • Multivariate fault estimation• Trending• Estimation based on embedded
constrained optimization
13November 2005
SPC - Statistical Process Control
• EPC (Engineering Process Control) -‘normal’ feedback control
• SPC - warning for an off-target manufacturing quality parameter
• Three main methods of SPC: – Shewhart chart (1920s)– EWMA (1940s)– CuSum (1950s)
14November 2005
SPC: EWMA
• Exponentially Weighted Moving Average• First order low pass filter
Detection threshold
)()1()1()1( txtyty λλ +−−=+
)2(2 λλ −= cZ
15November 2005
Estimation Approaches
• Traditional univariate monitoring: SPC• Multivariate SPC • Multivariate fault estimation• Trending• Estimation based on embedded
constrained optimization
16November 2005
Multivariate SPC: Hotelling's T2
• Multivariate normal distributed data
• Empirical parameter estimates
))()()((1,)(111
Tn
t
Tn
ttYtY
ntY
nµµµ −−≈Σ≈ ∑∑
==
),()( Σ≈ µNte
• The Hotelling's T2 statistics is
• T2 can be monitored using univariate SPC (almost)
( ) ( )µµ −Σ−= − )()( 12 tYtYT T
)()()( tftetY +=
17November 2005
-3 -2 -1 0 1 2 3
-2-1
01
23-2
-1
0
1
2
Multivariate SPC - Abnormality Detection
Multivariable envelope of scatter determined from normal operation data
Data point outside of this envelope implies abnormal behavior, a fault
18November 2005
• MSPC for cyber attack detection
• Y(t) consists of 284 audit events in Sun’s Solaris Basic Security Module
Multivariate SPC Example
Ye & Chen, Arizona State
19November 2005
Estimation Approaches
• Traditional univariate monitoring: SPC• Multivariate SPC • Multivariate fault estimation• Trending• Estimation based on embedded
constrained optimization
20November 2005
Multivariate Fault Estimation
• Fault parameters are explicitly modeled• Estimate and trend the fault parameters f
– MSPC detects anomaly for Y (unknown faults)
• Need models for fault signatures S(t)
)()()()( tetftStY +=
21November 2005
Engine Fleet Monitoring• Honeywell LF-507 Engine• Two-spool turbofan engine• Used in regional jets
Ganguli, Deo, & Gorinevsky, IEEE CCA’04
22November 2005
Parametric Faults - Sensitivities• Prediction residuals using detailed engineering model
• If f = 0 (nominal case) we should have Y = 0.• Nonzero residuals Y indicate faults
• A model is obtained by linearization (assume f is small)
)()()( teftStY +=
Flight cycleSensitivity
scatter = ‘noise’ ⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡=
driftsensor EGTleak band Bleed
iondeteriorat Turbinef
Fault parameters
),( fUFXY −=
23November 2005
Normalize Data: Prediction Model Residuals
• Ambient condition data– Altitude– Mach – Ambient Pressure– Air Temperature
• Engine input data– Speed of Spool #1
• Engine output data– Speed of Spool #2 – Exhaust gas temperature – Fuel Flow– Pressures, temperatures
Prediction Model Simulink Model
Ambient condition data
Engine input data
Engine output data predictions
Faults Actual engine output data,
Residuals
Collected Data
Y
X
24November 2005
Optimal Estimation
• Consider estimation at fixed time moment
• Bayes rule
• Optimal estimate – MAP/ML
eSfY += ),0(~ RNf ),0(~ 0QNe
min)( log)|( log →−−= fPfYPL
c fPfYPYfP ⋅= )()|()|(
min)()( 110 →+−− −− fRfSfYQSfY TT
25November 2005
Optimal Estimation
• Optimal estimate
• This is a GLS estimate • The prior for f effectively serves for
regularization. – important if S is ill-conditioned
[ ] YQSSQSRf oT
oT 1111ˆ −−−− +=
26November 2005
Fault Estimates
• Deployed: modeling, simulation, estimation
LF507 Engine
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
27November 2005
Estimation Approaches
• Traditional univariate monitoring: SPC• Multivariate SPC • Multivariate fault estimation• Trending• Estimation based on embedded
constrained optimization
28November 2005
Trending Functions• What is trending?
Filtering Smoothing
Present FuturePast
Prediction
29November 2005
Trend Estimation
• Data:
• Bayes rule
• Random walk model of the trend )()()( tetxty +=
)}(),...,1({}{)}(),...,1({}{
NxxXNyyY
==
Observed
Underlying trend
c XPXYPYXP ⋅⋅= })({}){|}({}){|}({
})({XP
}){|}({ XYP
)()()1( tvtxtx +=+ ),0(~ RNv
),0(~ QNe
30November 2005
Trend Estimation
• Observation model:
• Prior (trend model)
• ML loss index
• Batch least square estimation[ ] [ ] min)1()()()( 2121 →−−+−= −−∑ txtxRtxtyQL
( ) ( ) [ ]∏∏=
−
=
−==n
t
n
t
txtyQtxtyPXYP1
21
1
)()()(|)(}{|}{
( ) ( ) [ ]∏∏=
−
=
−−=−=n
t
n
t
txtxRtxtxPXP1
21
1
)1()()1(|)(}{
( ) min}{|}{log →−= YXPL
31November 2005
First-order Kalman Filter
• Recursive filter form of the LS KF• Stationary KF yields a recursive update
• The gain a comes from the Riccati equation• This is a EWMA filter – a basic SPC tool
)()(ˆ)1()1(ˆ taytxatx +−=+
32November 2005
Monotonic Regression - concept
• Monotonic deterioration, never an improvement
• Palmgren-Miner rule: additive damage accumulation, counting load cycles
• Modified random walk model
• Quite fundamental model!
0)(),()()1( ≥+=+ tvtvtxtx
Gorinevsky, IEEE ACC’04
33November 2005
Nonlinear Filtration
• Exponentially distributed driving noise
• Yields a batch ML estimate
• This a QP problem! – Can be solved using off-the-shelf solvers
• This is a nonlinear batch-mode filter
( )
),...,2(,0)1()(
min)1()())()(())()((21 1
1
1
Tttxtx
txtxqtxtyRtxtyJT
t
T
=≥−−
→−−+−−= −
=
−∑
⎩⎨⎧
<≥−
0,00,
~/
xxqe
vqx
)()()1( tvtxtx +=+
34November 2005
First-order Monotonic Regression
0 10 20 30 40 50 60 70 80-1
-0.5
0
0.5
1
1.5
2
2.51st ORDER MONOTONIC REGRESSION VS. EWMA FILTER
SAMPLE NUMBER
RAW DATAMONOTONIC REGRESSIONEWMA FILTERUNDERLYING TREND
35November 2005
2nd Order Random Walk
• Linear gaussian model
• Linear trend in the absence of the ‘noise’ v• ML estimate Hodrick-Prescott filter
)()()()()()()1(
)()()1(
2
2212
111
tetxtytvtxtdxtx
tvtxtx
+=++=+
+=+
[ ] [ ]2
221
1
21
ˆ
min)(ˆ)(ˆ)(21
xy
tyQtytyRJN
t
=
→∆+−= −
=
−∑
36November 2005
2nd Order Random Walk
• Hodrick-Prescott batch filter (trending)
0 50 100 150 200
0
2
4
6HP SMOOTHING, SECOND−ORDER REGRESSION MODEL
SAMPLE NUMBER
0 50 100 150 200
−0.1
0
0.1
0.2
SMOOTHING KERNELS FOR H−P FILTER
SAMPLE NUMBER
37November 2005
2nd-order Monotonic Regression
• Gaussian noise e• Exponential v1(t)≥ 0, describes primary damage • Exponential v2(t)≥ 0, secondary damage• Optimal MAP estimate yields a QP problem
)()()()()()()1(
)()()1(
2
2212
111
tetxtytvtxtdxtx
tvtxtx
+=++=+
+=+
[ ] [ ] [ ] min)()()()(21
21
11
1
22
1 →∆+∆+−= −−
=
−∑ txptxqtxtyRJN
t0)(,0)( 21 ≥∆≥∆ txtx
38November 2005
Monotonic Regression Trending of LF507 Performance Loss
39November 2005
Nonlinear Filtering
• Fault evolution models lead to nonlinear filtering problems
• A known systematic approach is particle filters – not scalable
• Embedded convex (QP) optimization is scalable
40November 2005
Shock and Wear Models
• Long studied, reliability theory motivation – e.g., Esary et al, Annals of Probability, 1973
• Shocks = jumps in deterioration parameter • Wear = gradual change on the parameter • 2nd-order monotonic regression
automatically provides estimates of the shocks and wear
41November 2005
Mono Regression Properties
• Accurate reproduction of jumps• Exponential noise well describes jump
processes.
0 50 100 150 200
0
2
4
6SECOND−ORDER MONOTONIC REGRESSION
SAMPLE NUMBER
42November 2005
Analysis of Mono Regression
• Increments as decision variables
• The QP problem can be presented as
• An equivalent of
⎥⎦
⎤⎢⎣
⎡
∆∆
=−
−
)()(
)(2
11
1
txqtxp
tX
min)1( 2 →+−−⋅ XBxAXY T1λ0 subject to ≥X
min→XT122)1( ,0 subject to ε≤−−≥ BxAXYX
43November 2005
Convex Relaxation of L0 Optimality
• Related recent work in signal processing:
is equivalent to
(under some technical conditions)• Sparse representation by overcomplete
dictionaries (basis pursuit regression)
min→XT1AXYX =≥ ,0 subject to
AXYX =≥ ,0 subject tomin)( components nonzero ofnumber →X
min1→X
min0→X
44November 2005
L1 Relaxation of L0 Optimality
• Donoho, Elad; Tropp; Candes - 2004/05 • Problem variations, e.g., noise
• Despite apparent similarity these results do not work for monotonic regression – Results are based on assuming that ‘mutual
coherence’ of A is small– Not so for a Toeplitz impulse response A
min→XT1 2 ,0 subject to ε≤−≥ AXYX
45November 2005
Mono Regression Sparsity
Theorem• Consider a 2nd order mono regression
• Assume that , where the jumps described by V are at least 2 steps apart
• Then the sparse solution is almost exactly recovered:
min→XT122)1( ,0 subject to ε≤−−≥ BxAXYX
1BvAVY +=
)(εOVU +=Gorinevsky 2005 (in preparation)
46November 2005
Estimation Approaches
• Traditional univariate monitoring: SPC• Multivariate SPC • Multivariate fault estimation• Trending• Estimation based on embedded
constrained optimization
47November 2005
Multivariable Model
• Model of fault evolution
)()()( teftStY +=
)()()()()1(
tCxtftvtAxtx
=+=+ ⎥
⎦
⎤⎢⎣
⎡=
1101
kA
Trend model
Exponentially distributed noise v• ML estimation is a QP
0)1()( ≥−− tAxtx[ ] [ ] min)1()()()()(
21
1
21 →−−+−= ∑=
−N
t
T tAxtxqtCxtStYRJ
48November 2005
ModelModel
Moving Horizon Estimation
• MHE
time
)()()( teftStY +=
)()()()()1(
tCxtftvtAxtx
=+=+
MHE Algorithm:Solve QP
Based on modelUsing the data
TrendLast point is the current estimate
49November 2005
Rocket Flight Control Example• Shuttle-class vehicle ascent
eir
bir
wivr
ir
Q
kgme
rbkr
ekr
θ
γ
α
δ
mg
vib
L
D
Ti
ω
ωe
le
Model
Gorinevsky, Samar, Bain, & Aaseng, IEEE Aerospace, 2005
50November 2005
Rocket Flight Control Example
0 20 40 60 80 100 120 140 1600
0.5
1
1.5ACCELERATION [ft/s2]
0 20 40 60 80 100 120 140 1600
1
2
3x 10
-3 FLIGHT ANGLE RATE [rad/s]
0 20 40 60 80 100 120 140 160-5
0
5
10x 10
-4 PITCH ACCELERATION [rad/s2]
0 20 40 60 80 100 120 140 160-5
0
5x 10
-3 ENGINE ANGLE RATE [rad/s]
TIME [s]
5 0 1 0 0 1 5 0-1 .4 0 5
-1 .4
-1 .3 9 5
-1 .3 9
D o w n ra n g e a n g le ( ra d )
5 0 1 0 0 1 5 0
5
1 0
1 5x 1 0
4 A l titu d e (f t)
5 0 1 0 0 1 5 0
2 0 0 0
4 0 0 0
6 0 0 0
8 0 0 0V e lo c ity (f t /s )
5 0 1 0 0 1 5 00 .40 .60 .8
11 .2
F lig h t p a th a n g le ( ra d )
5 0 1 0 0 1 5 00
5
1 0
1 5x 1 0
-3 E n g in e g im b a l a n g le ( ra d )
5 0 1 0 0 1 5 00
5
1 0
1 5
x 1 0-3E n g in e ro ta tio n a l ra te ( ra d /s )
5 0 1 0 0 1 5 00
5
1 0
1 5
x 1 0-3V e h ic le ro ta tio n a l ra te ( ra d /s )
5 0 1 0 0 1 5 0
0 .60 .8
11 .21 .4
P itc h a n g le (ra d )
5 0 1 0 0 1 5 0-0 .0 5
0
0 .0 5
G im b a l c o m m a n d ( ra d )
T im e (s )5 0 1 0 0 1 5 0
-0 .0 4-0 .0 2
00 .0 20 .0 4
G im b a l p o s itio n (ra d )
T im e (s )
Simulated telemetry data Model prediction residuals
51November 2005
Estimation Results
20 40 60 80 100 120 140 160
15
20
25
30GIMBAL SLUGGISHNESS, PERCENT
20 40 60 80 100 120 140 160
2
3
4THRUST REDUCTION, PERCENT
20 40 60 80 100 120 140 1600
2
4
6DRAG INCREASE, PERCENT
20 40 60 80 100 120 140 1600
0.2
0.4
0.6
0.8
PITCH MEASUREMENT OFFSET, PERCENT
Seeded Fault
Seeded Fault
Seeded Fault
Seeded Fault• Simulated flight
with seeded faults• Accurate
estimation achieved
• Estimates improve with time
• Constant, monotonic, and arbitrary faults
52November 2005
Aircraft Fault Estimator Design• Aerosonde UAV
Moving Horizon Estimation
Using Monotonic RegressionFault
Signatures
Fault Trends
Tuning ParametersPredictionResiduals
PredictionResiduals
Flight Data Model-
Based PredictionNo fault
Data Prediction ModelFault 1
Samar & Gorinevsky, 2005 (in preparation)
53November 2005
AeroSim simulation
Flight Control MPC
Guidance
Flight Dynamics, Wind Gusts
VHM Estimation
MS Flight Simulator
54November 2005
0 50 100 150 200 250 300
−0.20
0.20.4
SMOOTH (5sec smoothing) PREDICTION RESIDUALS
a x [m/s
2 ]
0 50 100 150 200 250 300−0.2
0
0.2
a y [m/s
2 ]
0 50 100 150 200 250 300
00.5
11.5
a z [m/s
2 ]
0 50 100 150 200 250 300−2
024
α x [rad
/s2 ]
0 50 100 150 200 250 300−0.6−0.4−0.2
00.20.4
α y [rad
/s2 ]
0 50 100 150 200 250 300
−0.20
0.2
TIME [s]
α z [rad
/s2 ]
Model Prediction Residuals
LinearAccelerations
AngularAccelerations
55November 2005
0 50 100 150 200 250 300
−505
1015
x 10−3
a x [m/s
2 ]
SENSITIVITY TO LEFT WING CL CHANGE %
0 50 100 150 200 250 300
−0.01
0
0.01
a y [m/s
2 ]
0 50 100 150 200 250 300
0.03
0.04
0.05
0.06
a z [m/s
2 ]0 50 100 150 200 250 300
−1.2
−1
−0.8
−0.6
α x [rad
/s2 ]
0 50 100 150 200 250 300
5
10
x 10−3
α y [rad
/s2 ]
0 50 100 150 200 250 300−0.12
−0.1−0.08−0.06−0.04
α z [rad
/s2 ]
TIME (s)
0 50 100 150 200 250 300
−4−2
024
x 10−3
a x [m/s
2 ]
SENSITIVITY TO LEFT WING CD CHANGE %
0 50 100 150 200 250 300−4−2
024
x 10−3
a y [m/s
2 ]
0 50 100 150 200 250 300−3−2−1
0
x 10−4
a z [m/s
2 ]
0 50 100 150 200 250 300
−4
−2
0x 10
−3
α x [rad
/s2 ]
0 50 100 150 200 250 300
3.5
4x 10
−3
α y [rad
/s2 ]
0 50 100 150 200 250 300
−0.044
−0.042
−0.04
−0.038
α z [rad
/s2 ]
TIME (s)
Fault Signatures (Sensitivities) left wing drag left wing lift
56November 2005
Fault Estimation
• 50 point MHE• 2Hz sampling
0 50 100 150 200 250 300
0
2
4
6RIGHT WING CL CHANGE %
0 50 100 150 200 250 300
0
2
4
6RIGHT WING CD CHANGE %
0 50 100 150 200 250 300
0
2
4
6LEFT WING CL CHANGE %
0 50 100 150 200 250 300
0
2
4
6LEFT WING CD CHANGE %
TIME [s]
0 200 400 600 800 1000 1200 1400 16000
500
1000
1500
2000
NO
RT
H P
OS
ITIO
N [m
]
EAST POSITION [m]
UAV TRAJECTORY IN NORTH−EAST PLANE
N-E trajectory
57November 2005
Vision: MHE of Faults • Related to MPC process control technology
– Estimation is dual to control– Closely related math– 1000 times faster computers
• Step response models in MPC
• Fault signature models
58November 2005
Conclusions
• ISHM estimation is the key to– service and maintenance automation– autonomous systems
• Estimation based on embedded constrained optimization – performs extremely well in ISHM problems– is a scalable and flexible technology– has great practical potential