You are on page 1of 41

Motivation Theory Applications Conclusions

Structural Health Monitoring - Theory and


Applications in ARTeMIS Modal

Starts at 09:00am
Szymon Greś
Structural Vibration Solutions A/S
Niels Jernes Vej 10
DK-9220 Aalborg, Denmark

21 January 2021

1
Motivation Theory Applications Conclusions

Outline

1 Motivation

2 Theory

3 Applications

4 Conclusions

2
Motivation Theory Applications Conclusions

Motivation

Why structural monitoring?


1 Poor condition of structures → risk of catastrophic failures
2 Real structural behavior → more accurate design models
3 Structures are remote and difficult to access → costly O&M

3
Motivation Theory Applications Conclusions

Vibration-based structural monitoring

What is it?
Integrity monitoring based on structural vibrations
Measurements often collected under operation conditions
of structures (wind, waves, traffic, rotating machinery)
Data analyzed by carefully designed algorithms (often)
linking it with structural models

Fields of research
Operational Modal Analysis (OMA)
Structural Health Monitoring (SHM)

4
Motivation Theory Applications Conclusions

Operational Modal Analysis

Context
Dynamic parameters of structure estimated from data
Measurements are the structural responses (e.g.
accelerations) observed at sensor positions

5
Motivation Theory Applications Conclusions

Structural Health Monitoring

Context
Looking for a change in vibration data (damage diagnosis)
Detection: is structure damaged?
Localization: where is the damage?
Quantification: what is the damage extent?
Remaining life prediction

6
Motivation Theory Applications Conclusions

Damage diagnosis approaches

Multitude of damage diagnosis methods in literature ...


Machine learning methods
Statistical evaluation of data-driven features (f , ϕ)
Model-driven approaches relying e.g. on model updating

Challenges
Computational efficiency
Appreciating uncertainty of models and features
Robustness towards changing environmental conditions

7
Motivation Theory Applications Conclusions

Need for a comprehensive framework


Computationally efficient
Known statistical properties
Able to link data with numerical models

8
Motivation Theory Applications Conclusions

Outline

1 Motivation

2 Theory

3 Applications

4 Conclusions

9
Motivation Theory Applications Conclusions

Some fault diagnosis basics


System parametrization θ
θ contains damage-sensitive parameters of the structure
θ is user-defined e.g. can relate to
modal parameters,
physical parameters e.g. element density, Youngs modulus.
θ∗ be the parameter in the reference (healthy) state
δ is the unknown change in the monitored parameter

Example
6 DOF chain-like system
k1 k2 k3 k4 k5 k6

m1 m2 m3 m4 m5 m6

c1 c2 c3 c4 c5 c6

10
Motivation Theory Applications Conclusions

Local approach for change detection

Framework (Benveniste et al. 1987)


Idea:
ζ: residual function
is Gaussian,
is a function of θ,
is a function of damage-sensitive system features.
Hypotheses
Is system healthy? Is system damaged?

H0 : θ = θ∗ H1 : θ = θ∗ + δ/ N
ζ ∼ N (0, Σζ ) ζ ∼ N (Jθζ∗ δ, Σζ )

where the Jacobian Jθζ∗ = ∂ζ/∂θ (θ∗ )


and Σζ can be approximated with first order Delta method
11
Motivation Theory Applications Conclusions

Statistical decision making


GLR test to decide between H0 and H1 hypotheses
 −1
ζ ζ T −1 ζ
t = ζ T Σ−1
ζ Jθ∗ Jθ∗ Σζ Jθ∗ Jθζ∗ T Σ−1
ζ ζ

under H0 , t follows a central χ2 (rank(Jθζ∗ )) distribution


under H1 , t follows a non-central χ2 (rank(Jθζ∗ )) distribution
with non-centrality parameter λ = δ T Jθζ∗ T Σ−1
ζ Jθ∗ δ
ζ

12
Motivation Theory Applications Conclusions

FDI framework for Gaussian residuals

In nutshell, this fault diagnosis framework boils down to


detection: global decision between H0 and H1 based on t
localization: testing k -th column of (Jθζ∗ )k for a change
quantification: computing δ based on Jθζ∗ and Σζ

Detection challenges
Computing the statistical uncertainty of ζ
Effect of changing environmental/operating conditions on ζ

13
Motivation Theory Applications Conclusions

Effect of changing operating conditions


Changes in excitation covariance Q
Q changes ⇒ data-driven features change
⇒ t changes even if structure does not change
⇒ false test results

Example
6DOF system subjected to damage and changes in Q
104 105
Q=I Q=I
Q = 2I
Q = randn(ndof,ndof)2
Q=I
Q=I
Q = 4I 104
3 Q = randn(ndof,ndof)2
10

103

102

102

1
10
101

100 100
reference healthy 5% damage 10% damage reference healthy 5% damage 10% damage

14
Motivation Theory Applications Conclusions

Design of fault detection method


Residual properties
1 ζ = 0 iff θ = θ∗ and ζ 6= 0 iff θ 6= θ∗
2 ζ ∼ asymptotically Gaussian distributed

Robustness towards changes in excitation conditions


under H0 : t ∼ χ2 (rank(Jθζ∗ ))
No false alarms: rank(Jθζ∗ ) = const iff Jθζ∗ depends only on
parameters estimated in the reference state

15
Motivation Theory Applications Conclusions

Design of fault detection method

Statistical damage detection

no yes
is FE model available?

θ : (f̂ , ϕ̂) θ : FE parameters


Jθζ∗ is purely data-driven Jθζ∗ links FE sensitivities with data

16
Motivation Theory Applications Conclusions

Design of fault detection method

Statistical damage detection

no yes
is FE model available?

θ : (f̂ , ϕ̂) θ : FE parameters


Jθζ∗ is purely data-driven Jθζ∗ links FE sensitivities with data

17
Motivation Theory Applications Conclusions

System modeling

Model: Linear Time Invariant (LTI) system


Mθ q̈(t) + C θ q̇(t) + Kθ q(t) = Fu(t) + ũ(t)
y(t) = La q̈(t) + Lv q̇(t) + Ld q(t) + v (t)
Mθ , C θ , Kθ : parametric mass, damping and stiffness matrices
q(t): displacements at degrees of freedom of a structure
u(t): inputs (excitation forces)
y(t): outputs (structural responses at sensor positions)
v (t): sensor noise

Vibration characteristics (modes)


Frequencies f , damping ratios ζ and mode shapes ϕ from

((µθ )2 Mθ + µC θ + Kθ )ϕθ = 0

18
Motivation Theory Applications Conclusions

State-space modeling

Discrete-time state space model


xk+1 = Aθ xk + B θ uk + B̃ ũk


y k = C θ xk + D θ uk + D̃ ũk + vk

y k : measured r outputs at time instant k


Aθ : state transition matrix
C θ : observation matrix
B θ : input matrix
D θ : feedthrough matrix (accelerations)
xk : state vector
vk : output noise

19
Motivation Theory Applications Conclusions

State-space modeling

Let’s drop θ for simplicity.


Discrete-time output-only state space model

xk+1 = A xk + Bu + B̃ ũ
k k
y k = C xk + Du
 + D̃ ũ + v
k k k

B̃ ũk and D̃ ũk : are the unknown and unmeasured inputs

 
0 I
τ C = Ld − La M−1 K Lv − La M−1 C
 
A = exp −1 −1
−M K −M C

(A − λI) φ = 0, ϕ = Cφ

20
Motivation Theory Applications Conclusions

Data-driven system identification

Recursion of xk+1
For k = 1 . . . N + p + q consider:

k = 1 : x2 = A x1 + B̃ ũ1
k = 2 : x3 = A x2 + B̃ ũ2
k = 3 : x4 = A x3 + B̃ ũ3

Recursion of yk
For k = 1 . . . N + p + q consider:

k = 1 : y 1 = C x1 + D̃ ũ1 + v1
k = 2 : y 2 = C x2 + D̃ ũ2 + v2
k = 3 : y 3 = C x3 + D̃ ũ3 + v3

21
Motivation Theory Applications Conclusions

Data-driven system identification


Important matrices
Data matrices Y + ∈ R(p+1)r ×N and Y − ∈ Rqr ×N
 ..   .. 
y y2 . yN y y q+2 . y N+q
 .1 .. .. ..   q+1 .. 
 .
 . . . . y y q+3 . y N+q+1 
  
 q+2
Y− =  , Y+ =  .

y .. . .. .. .. 
 q−1 yq . y N+q−2 
 .
 . . . 

.. ..
yq y q+1 . y N+q−1 y p+q+1 y p+q+2 . y p+q+N

Observability Γ ∈ Rpr ×n and lower block triangular Toeplitz


matrix H̃ ∈ Rpr ×pu
   
C D̃ 0 0 ... 0



 CA  C B̃

D̃ 0 ... 0 

Γ=

 ,
 CA2 H̃ =  CAB̃

C B̃ D̃ ... 0 

..
 ... ... ... . . . . . .
   
  .
CAp−1 CAp−2 B̃ CAp−3 B̃ CAp−4 B̃ . . . D̃

22
Motivation Theory Applications Conclusions

Data-driven system identification

Output-only data matrix relation


Data matrices Y − and Y + can be then expressed by recursion
of state-space equations

Y − = ΓX − + H̃ Ũ − + V −
Y + = ΓX + + H̃ Ũ + + V +

X − , X + : past and future state sequences


U − , U + : past and future input noise sequences
V − , V + : past and future measurement noise sequences

Task
Identify Γ from a matrix which is f (Y − , Y + )

23
Motivation Theory Applications Conclusions

Two output-only stochastic subspace algorithms

Covariance-driven algorithm
T T T
Hcov = Y + Y − = (ΓX + + H̃ Ũ + + V + )Y − = ΓX + Y − = ΓZcov

Orthogonal projection data-driven algorithm


T † − T † −
   
T T
Hdat = Y + /Y − = Y + Y − Y − Y − Y = ΓZcov Y − Y − Y
= ΓZdat

Classic damage sensitive features identified from data


Hcov , Hdat
Γ, A and C
collection of f , ζ and ϕ

24
Motivation Theory Applications Conclusions

Simple damage detection residual

Mahalanobis distance-based residual


√ θ∗ θ
ζ = N vec(Ĥref − Ĥtest )
⇒ Simple and well-known Mahalanobis distance
⇒ Accounts for uncertainty of both reference and test data

Assumption
Constant covariance Q of ambient excitation
w
w


No false alarms

25
Motivation Theory Applications Conclusions

Robust normalization for Mahalanobis residual

Task
θ so that Ĥθ∗ − Ĥθ is invariant to Q
Normalize Ĥtest ref test

Normalization scheme
Normalization property in the reference state (θ = θ∗ ):
θ∗ θ †
Ĥref − Ĥtest Ẑtest Ẑref ≈ 0

Factors Ẑtest and Ẑref are based on SVD

θ∗
 θ
  
Ĥref Ĥtest ≈ Ûs Ẑref Ẑtest

26
Motivation Theory Applications Conclusions

Robust residuals

Robust Mahalanobis distance-based residual (Gres et al. 2019)


√ 
θ∗ †

θ
ζmah = N vec Ĥref − Ĥtest Ẑtest Ẑref

where
under H0 : ζmah ∼ N (0, Σζmah ) (reference)
under H1 : ζmah ∼ N (Jθζ∗mah δ, Σζmah ) (faulty)

‘Good’ properties
No false alarms in healthy state
Covariance accounts test and reference data

27
Motivation Theory Applications Conclusions

Robust damage detection residuals

SVD of Hankel matrix


" #" #
i D̂ T
V̂s,ref
θ∗
h
s,ref 0
Ĥref = Ûs,ref Ûker,ref T ,
0 D̂ker,ref V̂ker,ref

Robust subspace residual (Döhler and Mevel, 2013)



T θ
ζsub = N vec(Ûker,ref Ĥtest )
where
under H0 : ζmah ∼ N (0, Σζsub ) (reference)
under H1 : ζmah ∼ N (Jθζ∗sub δ, Σζsub ) (faulty)

28
Motivation Theory Applications Conclusions

Outline

1 Motivation

2 Theory

3 Applications

4 Conclusions

29
Motivation Theory Applications Conclusions

Fault detection framework in application to SHM

Monitored structure Response measurements

Residual generation Statistical testing


30
Motivation Theory Applications Conclusions

Numerical validation on large FEM simulation

Monte Carlo simulation of an


offshore foundation model:
FE model: 8589 elements
Damage: thickness
reduction of 8 elements
Data: 10 sensors, 50Hz,
200,000 samples
Monte Carlo simulation
with 1000 realizations
Covariance of the ambient
excitation changing for
every data set

31
Motivation Theory Applications Conclusions

Numerical validation on large FEM simulation

32
Motivation Theory Applications Conclusions

Application: Z24 bridge benchmark

33
Motivation Theory Applications Conclusions

Damage detection test

Environmental load with forced vibration from shakers


Robust Mahalanobis distance damage detection test

95% quantile threshold


99% quantile threshold
106

105

104
reference healthy Lowering of pier, 20mm Lowering of pier, 40mm

34
Motivation Theory Applications Conclusions

ARTeMIS Modal: Dogna bridge

Four-span, 16m long concrete bridge in Italy


9 day progressive damage test campaign
50min of measurement analyzed

35
Motivation Theory Applications Conclusions

ARTeMIS Modal: Dogna bridge

10 accelerometers mounted on the deck


Two progressive damage cases

36
Motivation Theory Applications Conclusions

Outline

1 Motivation

2 Theory

3 Applications

4 Conclusions

37
Motivation Theory Applications Conclusions

Conclusions

Conclusions
Comprehensive framework for damage diagnosis
Implemented damage detection features
Automated and scalable
Good statistical properties
Robust towards changing excitation

38
Motivation Theory Applications Conclusions

Publications

Damage detection (selection)


M. Döhler, F. Hille, L. Mevel, and W. Rücker “Structural health monitoring with
statistical methods during progressive damage test of S101 Bridge”, Engineering
Structures, 2014.
M. Döhler, L. Mevel, F. Hille “Subspace-based damage detection under changes
in the ambient excitation statistics”, MSSP, 2014.
L. D. Avendaño-Valencia, E. N. Chatzi, D. Tcherniak, “Gaussian process models
for mitigation of operational variability in the structural health monitoring of wind
turbines”, MSSP, 2020
S. Greś, M.D. Ulriksen, M. Döhler, R.J. Johansen, P. Andersen, L. Damkilde,
“Statistical methods for damage detection applied to civil structures”,
EuroDyn2017, 2017.
D. Bernal “Kalman filter damage detection in the presence of changing process
and measurement noise”, MSSP, 2013.
M. D. Ulriksen, D. Tcherniak, P. H. Kirkegaard, and L. Damkilde “Operational
modal analysis and wavelet transformation for damage identification in wind
turbine blades”, SHM, 2016.

39
Motivation Theory Applications Conclusions

Publications

Damage localization (selection)


S. Allahdadian, M. Döhler, C. Ventura, L. Mevel, “Towards robust statistical
damage localization via model-based sensitivity clustering”, MSSP, 2019
M. Döhler, L. Mevel, Q. Zhang, “Fault detection, isolation and quantification from
Gaussian residuals with application to structural damage diagnosis”, Annual
Reviews in Control, 2016
D. Bernal, M. D. Ulriksen, “Subspace exclusion zones for damage localization”,
MSSP, 2019
M. D. Ulriksen, D. Bernal, L. Damkilde, “Shaped input distributions for structural
damage localization”, MSSP, 2018
A. Yan, J.-C. Golinval, “Structural damage localization by combining flexibility and
stiffness methods”, Engineering Structures, 2005
E. Balmés, M. Basseville, L. Mevel, H. Nasser, W. Zhou, “Statistical model-based
damage localization: A combined subspace-based and substructuring
approach”, SCHM, 2008

40
Motivation Theory Applications Conclusions

Thank you for the attention!

Szymon Greś
sg@svibs.com

41

You might also like