You are on page 1of 12

Traitement statistique de signaux

pour la Surveillance d’Intégrité de Structures

Michèle Basseville, Laurent Mevel


Albert Benveniste, Maurice Goursat

IRISA, Campus de Beaulieu

http://www.irisa.fr/sisthem/ -- basseville@irisa.fr

1 2

3 4
Traitement statistique

de l’information multi-capteurs

sur la base de modèles physiques

? pour la surveillance et le diagnostic

de systèmes dynamiques complexes

Problématique privilégiée :

surveillance de systèmes soumis à des vibrations

5 6

Problème : surveillance en fonctionnement Model-based statistical approach - Key issues


• The excitation is typically:
• 9. Modelling
– natural, not controlled.
– not measured:
∗ buildings, bridges, offshore structures, • 11. Output-only covariance-driven subspace identification

∗ rotating machinery (e.g. steam flowing),


• 13. Merging multiple measurements setups
∗ cars, trains, aircrafts.
– nonstationary (e.g., turbulent).
• 19. Robustness to nonstationary excitation
• The data contain slow and fast non-stationarities

• 21. Structural monitoring: damage detection and diagnostics


• How to merge multiple measurements setups
e.g. in case of moving sensors?
30. Examples: Z24 bridge & flutter monitoring

• How to detect and localize small damages ?


45. Free toolbox
7 8
Vibrations : problème de structure propre Vibrations : problème ARMA non-stationnaire










Xk+1 = F Xk + Vk









M Z̈(s) + C Ż(s) + K Z(s) = ν(s) 



Yk = H Xk
 


FE model: 









 Y (s) = L Z(s)
p q p
Yk = Ai Yk−i+ Bj (k) Vk−j , H Fp = Ai H F p−i
X X X
(M µ2 + Cµ + K) Ψµ = 0 , ψµ = L Ψµ
i=1 j=0 i=1









Xk+1 = F Xk + Vk

State space: 


 • Monitor AR part, with nonstationary MA part.
Yk = H Xk





F ϕλ = λ ϕλ , φλ = H ϕλ • Modal changes not visible on spectra (1%).
 

δµ = λ , ∆  Λ 
Parameter: e ψ µ = φλ} ; θ = 



| {z } | {z

vec Φ 
• Likelihood : no hope !
modes mode shapes
9 10

Output-only covariance-driven subspace identification Implementation


  

E Xk YkT 

 ∆








Xk+1 = F Xk + Vk G=

  

R̂0 R̂1 R̂2 ...



  
 
   




Yk = H Xk ∆
Ri = E Yk Yk−i T  N

R̂1

R̂2 R̂3 ...










R̂i = 1/N
X T ,
Yk Yk−i Ĥ =








 | {z }
R̂2 R̂3 R̂4 ...
  




ok if stationary ! | {z
k=1 }






 
ok when nonstationary ! .. .. ... .. 

 
R0 
 R1 R2 . . .   

∆ R



 R2 R3 . . . 




H
HF

 Ĥ ≈ Ô Cˆ
H= 1

, O=




, C=
∆  
   G FG F G ...2
R2 

 R3 R4 . . .  


HF 2
...



SVD(Ĥ) + truncation −→ Ô −→ (Ĥ, F̂ ) −→ (λ̂, ϕ̂λ)


..  .. ... .. 
 
∆1 0
Ĥ = U ∆ W T = U W T ; Ô = U ∆1/2
Ri = H F i G =⇒ H = O C
 
0 ∆0 1

Op↑(H, F ) = Op(H, F ) F

H −→ O −→ (H, F ) −→ (λ, ϕλ)


det(F − λ I) = 0 , F Φλ = λ Φλ, ϕλ = H Φλ

11 12
Model-based statistical approach - Key issues Merging multiple measurements setups (1)
     
(0,1) (0,2) (0,J )

 Yk 


 Yk 


 Yk 

• Modelling



 (1)







 (2)




... 


 (J )





Yk  
Yk  
Yk 

| {z } | {z } | {z }
Record 1 Record 2 Record J
• Output-only covariance-driven subspace identification



 (j) (j) (j)









Xk+1 = F Xk + Vk



• Merging multiple measurements setups 








(0,j) (j)





Yk = H0 X k (the reference)






• Robustness to nonstationary excitation 




(sensor pool noj)



 (j) (j)




 Yk = Hj X k
• Structural monitoring: damage detection and diagnostics (0,j)T (0,j)T
0,j ∆ (0,j) j ∆ (j)
Ri = E Yk Yk−i , Ri = E Yk Yk−i
Examples: Z24 bridge & flutter monitoring
(j) T (j 0) (j) T
Yk−i (j 6= j 0) not available
(j)
E Yk Yk−i not used, E Yk
Free toolbox
13 14

Merging multiple measurements setups (2)


Stationary excitation
Merging multiple measurements setups (3)
(j) ∆ (j) (0,j)T
covVk =Q, G=E Xk Yk
j Reference sensors: using the redundancy
= H0 F i G = Ri = Hj F i G
0,j ∆
Ri Ri0 ,
  !

= H0 F i Gj ,
 
0,j ∆



Ri0 





H0 


Ri G= G1 G2 . . . GJ

Ri1  
H1 
Riπ = = H Fi G ,
∆   ∆  




..




H = 

 .. 


     

= H0 F i G
 
  ∆  0,1

RiJ
 
HJ

Ri0 = Ri
0,2
Ri
0,J
. . . Ri 

 
Nonstationary excitation R00


R10 R20 . . . 
 
(0,j)T ∆ R1
0 R20 R30 . . . 

(j) ∆ (j) 
covVk = Qj , Gj = E Xk Yk H0 = 

  = O(H0 , F ) C(F, G) −→ C(F, G )
j
R20




R30 R40 . . . 
Ri
0,j
= H0 F i Gj ,
j
Ri = Hj F i Gj
..  .. ... .. 

Hint: right renormalization of the covariances.


15 16
Merging multiple measurements setups (4) Merging multiple measurements setups (5)
Moving sensors: normalizing the data (0,j)
 
• Build the J Hankel matrices H0,j = Hank(Ri )
 R0j R1j R2j ... 
 
j R1j R2j R3j ∆
Ri = Hj F i Gj , Hj =
∆ 



... 


 = O(Hj , F ) C(F, Gj ) • H0 = interleaving the block-columns of the H0,j ’s



R2j R3j R4j ... 


... ... ... ... ∆
• SVD (H0) + truncation −→ C(F, G) = C
 
 
−1

H̄j = Hj 

 C T (F, Gj ) 
C(F, Gj ) C T (F, Gj )  C(F, G1) • Partition C = ( C1 C2 . . . CJ )

j ∆ • Compute (CjT ( Cj CjT )−1 C1)


H̄j ↔ R̄i = Hj F i G1 (1 ≤ j ≤ J )
j
  • Build the J Hankel matrices Hj = Hank(Ri )

R̄i0 
 
∆ R̄i1
R̄i0 = Ri = H0 F i G1
∆ 0,1
Hj (CjT ( Cj CjT )−1 C1); H̄0 =
 
R̄i = 



...




, ∆
• Renormalize: H̄j = ∆
H0,1
 
R̄iJ

  • H̄ = interleaving the block-rows of the (J + 1) H̄j ’s

R̄0 R̄1 R̄2 ... 
 
∆  R̄1 R̄2 R̄3 ... 
H̄ = 


 R̄2 R̄3 R̄4 ...




= O(H, F ) C(F, G1) • Apply the subspace algorithm to H̄.
 
... ... ... ...
17 18

Model-based statistical approach - Key issues Robustness to nonstationary excitation

Approximate factorization of covariances : R̂i ≈ H F i Ĝ


• Modelling
Consistency : T −1 F̂ T → F, Ĥ → H; (λ̂, ϕ̂λ) → (λ, ϕλ)
• Output-only covariance-driven subspace identification

Holds true when merging multiple setups as well.


• Merging multiple measurements setups

Theory and experience show that the combination of:


• Robustness to nonstationary excitation
• the key factorization property of the covariances,
• Structural monitoring: damage detection and diagnostics
• the averaging operation in their computation,
Examples: Z24 bridge & flutter monitoring

Free toolbox allows to cancel out nonstationarities in the excitation


19 20
Model-based statistical approach - Key issues Structural monitoring : Eigenstructure monitoring


• Modelling 







Xk+1 = F Xk + Vk F ϕλ = λ ϕλ








Yk = H Xk φλ = H ϕλ



• Output-only covariance-driven subspace identification 

 

∆  Λ  modes
• Merging multiple measurements setups Canonical parameter : θ = 





vec Φ mode shapes
 
 Φ 
• Robustness to nonstationary excitation 

 Φ∆ 


Observability in modal basis : Op+1(θ) = 
 ... 

 
 
Φ∆p
• Structural monitoring: damage detection and diagnostics
System parameter characterization:
Examples: Z24 bridge & flutter monitoring
Hp+1,q and Op+1(θ) have the same left kernel.
Free toolbox
21 22

Eigenstructure monitoring - Contd Eigenstructure monitoring - Contd


 
θ0 : reference parameter, known (or identified) R̂0 


R̂1 R̂2 ... 


 
R̂1 

R̂2 R̂3 ... 

Fresh data −→ R̂i −→ Ĥ =
 
 
Yk : N -size sample of new measurements R̂2



 R̂3 R̂4 ...




 
 
..  .. ... .. 

Test statistics ζN (θ0): Which estimating function ?  





Φ 


 



Φ∆ 


Observability
Nominal model : O(θ0) = 


Φ∆2


 in modal basis
Local approach (small deviations) 





 .. 


Test H0 : θ = θ0 against H1 : θ = θ0 + δθ/ N !H=OC ! ker ĤT ?
= ker O T (θ0)

23 24
Eigenstructure monitoring - Contd The test statistics is asymptotically Gaussian

System parameter characterization: Mean sensitivity (Jacobian) J (θ0) and covariance Σ(θ0)

Hp+1,q and Op+1(θ) have the same left kernel.










N( 0, Σ(θ0)) under Pθ 0




ζN (θ0) → 
N ( J (θ0) δθ, Σ(θ0))









under Pθ + √δθ

 0 N
∃ S, S T S = Is, S T Op+1(θ0) = 0; say S(θ0)

(GLR) χ2-test for modal monitoring


θ0 ↔ (Ri0)i characterized by: S T (θ0) Ĥ0p+1,q = 0
T Σ−1 J (J T Σ−1 J )−1 J T Σ−1 ζ
ζN ≥h
N

The subspace-based residual ζN (θ0):


(GLR) Directional χ2-test for modal diagnosis
∆ √
 

ζN (θ0) = N vec  S T (θ0) Ĥp+1,q 


T Σ−1 J (J T Σ−1 J )−1 J T Σ−1 ζ
ζN ≥h
i i i i N
25 26

On-board damage diagnostics: projecting changes


Jacobian ?
Model-based monitoring - Generalization clustering
. .
. .. .
. . .....
Any estimating function can play the role of a residual .. .

... .
Warning: .. .
The prediction error is OK for sensor faults, . ...
FE domain . . . .
NOT for structural damages !
changes
modal domain
changes

27 28
Damage diagnostics: (local) sensitivity approach Examples

 
δM  
• Example 1 - Z24 bridge
ζ ∼ N (J δθ, Σ), δθ = I J(M0?,K0?) 



δK  

(M0?, K0?) : design model


• Example 2 - Beam within thermal chamber
J(M ?,K ?)
Jacobian : (δM , δK) 0 0 → (δµ, δψµ)
• Example 3 - Aircraft flutter monitoring
Reduction: I matching computed/identified modes

 
M  
Problem : dim 


  dim θ • Test cases processed so far
K  

http://www.irisa.fr/sisthem/sisthem-testcases.pdf
Hint: Cluster the vectors (δµ, δψµ) using the χ2-metric

29 30

Processed test cases (1) - Simulators and laboratory test-beds Processed test cases (2) - Real structures

Task Structure Our algos tested at Partner Task Structure Our algos tested at Partner

I & D 18 masses & springs Sisthem CNEXO I & D Offshore platforms CNEXO, Sisthem Elf Norway
I & D Rotating shaft Sisthem EDF I & D Turbo-alternator Sisthem EDF
I Triangular structure AS&I Dataid I Subway carrying bogies AS&I Dataid Ratp
I & D In-flight rocket Sisthem Ecole Centrale P. I Washing machine AS&I Dataid
I & D Steelquake simulator Sisthem I & D Paris MS760 aircraft Sopemea, Sisthem
I & D Flutter simulator Sisthem V.U.Brussels I & D Sport car LMS
I & D In-flight data Sisthem Airbus France I Helicopter PZL-Swidnik
I Launcher EADS L.V. CNES I & D Z24 bridge LMS, Sisthem COST F3
I & D Bridge deck Sisthem Ecole Centrale P. I & D Slat track LMS
I & D Cantilever beam Sisthem EDF I & D Ariane5 launcher, ground tests Sisthem EADS L.V.
I & D Clamped & free aluminum beam LMS I & D Ariane5, flights 501 & 502 CNES, Sisthem EADS L.V.
I & D Steel frame breadboard model LMS I Launcher EADS Launch Vehicles CNES
I & D Sport car scale model LMS I Airbus A320 - In-flight Airbus France, Sisthem Airbus Fr.
I & D Steelquake model Sisthem COST F3 I Airbus A340 - In-flight Sisthem Airbus Fr.
I & D Composite plate U. Siegen I Rafale - Ground tests Sopemea, Sisthem Dassault
I & D Reticular structure LMS I Rafale - In-flight Sopemea, Dassault, Sisthem Dassault
I & D Pre-stressed beam Sisthem LCPC I Falcon - In-flight Sopemea, Dassault, Sisthem Dassault
I Large scale pre-flexed beam LCPC, Sisthem I Bradford stadium Sisthem U.Sheffield
I & D Aircraft model Sisthem U. Liège I Manchester stadium Sisthem U.Sheffield
I & D Composite plastic engine oil pan LMS, Sisthem
31 32
Example 1 - Z24 bridge
Identified first four natural frequencies / Test values
(Results with four sensors)

• A benchmark of the BRITE/EURAM project


SIMCES and of the European COST action F3 Mode 1 2 3 4 χ2

Safe Freq.(Hz) 3.88 5.01 9.80 10.30 8.80 · 10e2


• Response to traffic excitation under the bridge
measured over one year in 139 points
Dam. DS1 Freq.(Hz) 3.87 5.06 9.79 10.32 8.00 · 10e5

• Two damage scenarios (DS1 and DS2): Dam. DS2 Freq.(Hz) 3.76 4.93 9.74 10.25 3.96 · 10e6
pier settlements of 20mm and 80mm.

33 34

Example 2 - Beam within thermal chamber


• A laboratory test-case provided by LCPC

• Vertical clamped beam, subject to increasing temperatures

• Local damage: horizontal clamped spring attached to the


Test values over nine months (log-scale). beam, with tunable stiffness and height

• Thermal effect modelling, analytical computations for temperatur


updated modal parameters

Histograms of the test values for each of the nine months.


35 36
60 154

59
safe − theory
safe − identify(cosmad)
damaged − identify(cosmad)
safe − theory
safe − identify(cosmad)
damaged − identify(cosmad)
Robustness to temperature effects
152

58

57
150 New results - 3 solutions

Second − Frequency
First − Frequency
56 148
• Empirical kernel, merging many temperature conditions
55
146

• Simplified model of the temperature effect


54

144
53

52 142
Analytical kernel updating
20 30 40 50 60 70 80 90 100 110 120 20 30 40 50 60 70 80 90 100 110 120
ε − αθ(µm/m) ε − αθ(µm/m)
x x

288 466
safe − theory
safe − identify(cosmad)
damaged − identify(cosmad)
464
safe − theory
safe − identify(cosmad)
damaged − identify(cosmad)
• Simplified model of the temperature effect
Statistical nuisance rejection techniques
286

462
284

Fourth − Frequency
Third − Frequency

7 6 6
460 6
x 10
14
x 10
5
x 10

safe safe safe


282 damaged damaged damaged
4.5
12
5
458
4

10
280 4 3.5

456
8 3

E(χ2)

E(χ )

E(χ2)
2
3
2.5
278 6
454
2
2
4
1.5
276 452
20 30 40 50 60 70 80 90 100 110 120 20 30 40 50 60 70 80 90 100 110 120 1
ε − αθ(µm/m) ε − αθ(µm/m) 2 1
x x

0.5
0 0
20 30 40 50 60 70 80 90 100 110 120 20 30 40 50 60 70 80 90 100 110 120 20 30 40 50 60 70 80 90 100 110 120
ε − αθ(µm/m) ε − αθ(µm/m) ε − αθ(µm/m)
x x x

Original χ2-test After empirical merging After kernel updating


First 4 frequencies vs. thermal constraint.
Computed (black) and identified (blue: safe, red: damaged) Safe (blue) and damaged (red)
37 38

Monitoring on-line a damping coefficient


Example 3 - Aircraft flutter monitoring
• Write the subspace-based test-statistics ζ
• Aero-elastic flutter: critical unstability phenomenon as a cumulative sum

• Flight flutter testing procedure • Test ρ ≥ ρc against ρ < ρc


Non local ! Use a different asymptotics for ζ
• Objective: on-line in-flight exploitation of test data

• Introduce a minimum change magnitude


• On-line flight flutter monitoring problem:
(actual change magnitude unknown)
monitoring some specific damping coefficient

• Run two CUSUM tests in parallel


• Decide whether ρ < ρc, critical value
(actual change direction unknown)

39 40
Monitoring a damping coefficient - Details
n−p √

ζ n(ρ0) = J (ρ0)T Σ(θ0)−1 ζn(θ0) = Zk (ρ0) / n
X

k=q Monitoring a damping coefficient - Details - Contd

Decreasing mean Increasing mean


 
− T 
Zk (ρ0) = J (ρ0)T Σ(θ0)−1 vec U (θ0)T Yk,p+1
∆ +
Yk,q 
∆ −1/2 Pn−p
Sn(−) = Σ k=q (Zk + νm) Sn
∆ −1/2 Pn−p
(+) = Σ k=q (Zk − νm)

∆ (−) ∆ (+)
Tn(−) = maxq≤k≤n−p Sk Tn(+) = minq≤k≤n−p Sk

Zk ∼ N (ν, .) , Zk ’s independent − =
gn ∆
Tn(−) − Sn(−) gn ∆
+ = Sn(+) − Tn(+)

− ≥ γ−
gn gn
+ ≥ γ+

H0 : ν ≥ + νm/2 H1 : ν < − νm/2


g g
and

41 42

Test for ρc = ρ(1)


c .

Test for ρc = ρ(2)


c < ρc .
(1)

Ariane booster launcher during a launch scenario on the ground − reflects ρ < ρ . Top: g + reflects ρ > ρ .
Bottom: −gn c n c
Automatic identification. Each symbol: processing 5 sec. data.
43 44
Functions

Toolbox COSMAD • Output-only subspace-based identification


• Input-output subspace-based identification
COvariance Subspace Modal Analysis and Diagnosis
• Automated subspace modal analysis and tracking
• Automated recursive subspace modal analysis
http://www.irisa.fr/sisthem/cosmad/
• Subspace-based moving sensors data fusion
• Damage detection
A toolbox to be used within Scilab
• Damage monitoring
a free matrix computing environment
• Modal diagnosis

with support of Eurêka projects SINOPSYS and FliTE


• Damage localization
• Optimal sensors positioning for monitoring
45 46

Experimented for and/or within References


• M. Basseville, A. Benveniste, M. Goursat, L. Hermans, L. Mevel, H. Van der
• LMS International Auweraer, Output-only subspace-based structural identification, from theory to
industrial testing practice. ASME Jal Dynamic Systems Measurement and Con-
• EADS Space Transportation and CNES trol, Special Issue on Identification of Mechanical Systems, V.123, N.4, Dec.
2001, pp.668-676.
• L. Mevel, M. Basseville, A. Benveniste, M. Goursat, Merging sensor data from
• Dassault Aviation multiple measurement setups for nonstationary subspace-based modal analysis.
Jal Sound and Vibration, V.249, N.4, Jan. 2002, pp.719-741.
• L. Mevel, A. Benveniste, M. Basseville, M. Goursat, Blind subspace-based eigen-
• Eurocopter structure identification under nonstationary excitation using moving sensors.
IEEE Trans. Signal Processing, V.50, N.1, Jan. 2002, pp.41-48.
• SOPEMEA (Test Lab.) • M. Basseville, M. Abdelghani, A. Benveniste, Subspace-based fault detection
algorithms for vibration monitoring. Automatica, V.36, N.1, Jan. 2000, pp.101-
109.
• LCPC/SMI (Civil Eng. Inst. / Metrology dept.) • M. Basseville, L. Mevel, M. Goursat, Statistical model-based damage detection
and localization, subspace-based residuals and damage-to-noise sensitivity ratios.
Jal Sound and Vibration, V.275, N.3-5, Aug. 2004, pp.769-794.
• L. Mevel, M. Basseville, A. Benveniste, Fast in-flight detection of flutter onset -
Current research topics and developments A statistical approach. AIAA Jal Guidance, Control, and Dynamics, V.28, N.3,
May 2005, pp.431-438.
• M. Basseville, A. Benveniste, M. Goursat, L. Mevel, Subspace-based algorithm-
• Flutter monitoring (aircrafts) s for structural identification, damage detection, and sensor data fusion. Jal
Applied Signal Processing, Special Issue on Advances in Subspace-Based Tech-
• Rejecting temperature effects (civil eng.) niques for Signal Processing and Communications, under minor revision, Feb.
2006.
47 48

You might also like