SDU Mechanical Engineering

sdu.dk
State-space modal representations
for decomposition of multivariate

#sdudk
non-stationary signals

Luis D. Avendaño-Valencia Luis Enrique Avendaño Edilson Delgado-Trejos David Cuesta-Frau

(Presenting Speaker)
University of Southern Denmark Universidad Tecnológica de Pereira Instituto Tecnológico Metropolitano Universidad Politécnica de Valencia
Odense, Denmark Pereira, Colombia Medellín, Colombia Valencia, Spain

July 2021
SDU Mechanical Engineering

sdu.dk
Motivation: Modal decompositions in LTI systems
𝒚𝒚𝑡𝑡 ∈ ℝ𝑛𝑛 𝒒𝒒𝑡𝑡 ∈ ℝ𝑛𝑛

#sdudk
Excitation

LTI system Modal

Responses Modal components
decomposition

Modal representation of the response Mode shape matrix (mixing matrix)

𝚿𝚿 = 𝝍𝝍1 ⋯ 𝝍𝝍𝑀𝑀 𝜓𝜓𝑗𝑗,𝑚𝑚 : Contribution of 𝑚𝑚-th mode to 𝑗𝑗-th response
𝑀𝑀

𝒚𝒚𝑡𝑡 = 𝚿𝚿𝑇𝑇 𝒒𝒒𝑡𝑡 = � 𝝍𝝍𝑇𝑇𝑚𝑚 𝑞𝑞𝑚𝑚,𝑡𝑡 Modal components

𝑚𝑚=1 2 1 𝑚𝑚 = 𝑛𝑛
where 𝑞𝑞𝑚𝑚,𝑡𝑡 , 𝑞𝑞𝑛𝑛,𝑡𝑡 = � 𝜔𝜔𝑛𝑛,𝑚𝑚 , 𝜁𝜁𝑚𝑚
The response is the superposition of individual
𝑞𝑞𝑚𝑚,𝑡𝑡 = � 𝑎𝑎𝑚𝑚,𝑖𝑖 𝑞𝑞𝑚𝑚,𝑡𝑡−𝑖𝑖 + 𝑏𝑏𝑚𝑚 𝑢𝑢𝑡𝑡 0 𝑚𝑚 ≠ 𝑛𝑛
Natural frequencies and
oscillatory components (modal components) 𝑖𝑖=1 Orthonormality property damping ratios

July 2021
SDU Mechanical Engineering

sdu.dk
Motivation: Modal decompositions in LTV systems
𝒚𝒚𝑡𝑡 ∈ ℝ𝑛𝑛

#sdudk
𝒒𝒒𝑡𝑡 ∈ ℝ𝑀𝑀
𝐻𝐻1 (𝑓𝑓, 𝑡𝑡) 𝐻𝐻2 (𝑓𝑓, 𝑡𝑡) 𝐻𝐻𝑚𝑚1 (𝑓𝑓, 𝑡𝑡) 𝐻𝐻𝑚𝑚𝑚 (𝑓𝑓, 𝑡𝑡)

Excitation

LTV system Modal

Non-stationary Non-stationary
decomposition
responses modal components

Modal representation of the response Mode shape matrix (mixing matrix)*

𝚿𝚿 = 𝝍𝝍1 ⋯ 𝝍𝝍𝑀𝑀 𝜓𝜓𝑗𝑗,𝑚𝑚 : Contribution of 𝑚𝑚-th mode to 𝑗𝑗-th response
𝑀𝑀

𝒚𝒚𝑡𝑡 = 𝚿𝚿𝑇𝑇 𝒒𝒒𝑡𝑡 = � 𝝍𝝍𝑇𝑇𝑚𝑚 𝑞𝑞𝑚𝑚,𝑡𝑡 Modal components

𝑚𝑚=1 2 1 𝑚𝑚 = 𝑛𝑛
where 𝑞𝑞𝑚𝑚,𝑡𝑡 , 𝑞𝑞𝑛𝑛,𝑡𝑡 = � 𝜔𝜔𝑚𝑚,𝑡𝑡 , 𝐴𝐴𝑚𝑚,𝑡𝑡
The response is the superposition of individual
𝑞𝑞𝑚𝑚,𝑡𝑡 = � 𝑎𝑎𝑚𝑚,𝑖𝑖,𝑡𝑡 𝑞𝑞𝑚𝑚,𝑡𝑡−𝑖𝑖 + 𝑏𝑏𝑚𝑚,𝑡𝑡 𝑢𝑢𝑡𝑡 0 𝑚𝑚 ≠ 𝑛𝑛
Inst. Frequencies (IF) and
oscillatory components (modal components) 𝑖𝑖=1 Orthonormality property Inst. Amplitude (IA)

*We presently assume that the mode shape coefficients remain constant.

July 2021
SDU Mechanical Engineering

sdu.dk
State of the Art: LTV modal decomposition methods

#sdudk
Time-frequency methods7,9 • Ridge extraction from T-F distributions

Linear transformation • Principal or Independent Component Analysis

(covariance analysis)10 • Singular Spectrum Analysis

Subspace ID for non- • Generalization of LTI subspace ID methods for linear parameter-
stationary systems6 varying systems and general non-stationary systems

Time-dependent and • ‘Frozen’ modal analysis

Periodic AR/ARMA • Time-periodic modal analysis
models3,5,8

• Hilbert-Huang Transform (Empirical Mode Decomposition - EMD)

Related methods1,4 • Discrete Wavelet Transform (DWT) and wavelet packet analysis

July 2021
SDU Mechanical Engineering

sdu.dk
State of the Art: Diagonal state space representations
Block-diagonal time-dependent SS representation2:

#sdudk
𝛼𝛼𝑚𝑚,𝑡𝑡 𝛽𝛽𝑚𝑚,𝑡𝑡
𝑴𝑴1 𝜽𝜽𝑡𝑡 𝑴𝑴𝑚𝑚 𝜽𝜽𝑡𝑡 =
𝒛𝒛𝑡𝑡 = 𝑴𝑴 𝜽𝜽𝑡𝑡−1 𝒛𝒛𝑡𝑡−1 + 𝒖𝒖𝑡𝑡 −𝛽𝛽𝑚𝑚,𝑡𝑡 𝛼𝛼𝑚𝑚,𝑡𝑡
𝜽𝜽𝑡𝑡 = 𝜽𝜽𝑡𝑡−1 + 𝒗𝒗𝑡𝑡 𝑴𝑴2 𝜽𝜽𝑡𝑡
𝑴𝑴 𝜽𝜽𝑡𝑡 = 2nd order block
⋱
State/parameter evolution 𝑴𝑴𝑀𝑀 𝜽𝜽𝑡𝑡
equations 𝜆𝜆𝑚𝑚,𝑡𝑡 = 𝛼𝛼𝑚𝑚,𝑡𝑡 ± j𝛽𝛽𝑚𝑚,𝑡𝑡
Block-diagonal state transition matrix
Instantaneous eigenvalue

∗ ∗ 𝑇𝑇
State vector: 𝒛𝒛𝑡𝑡 = 𝑞𝑞1,𝑡𝑡 𝑞𝑞1,𝑡𝑡 ⋯ 𝑞𝑞𝑀𝑀,𝑡𝑡 𝑞𝑞𝑀𝑀,𝑡𝑡
𝑦𝑦𝑡𝑡 = 𝑪𝑪𝑇𝑇 𝒛𝒛𝑡𝑡 + 𝜀𝜀𝑡𝑡
𝑇𝑇
Parameter vector: 𝜽𝜽𝑡𝑡 = 𝛼𝛼1,𝑡𝑡 𝛽𝛽1,𝑡𝑡 ⋯ 𝛼𝛼𝑀𝑀,𝑡𝑡 𝛽𝛽𝑀𝑀,𝑡𝑡
Observation equation
(scalar response) Measurement vector: 𝑪𝑪 = 1 0 ⋯ 1 0 𝑇𝑇

2Avendaño et al., “Diagonal time dependent state space models for modal decomposition of non-stationary signals”, Signal Processing 147, pp. 208-223, 2018.

July 2021
SDU Mechanical Engineering

sdu.dk
State of the Art: Diagonal state space representations
Block-diagonal time-dependent SS representation2:

#sdudk
Modal form of the response
𝒛𝒛𝑡𝑡 = 𝑴𝑴 𝜽𝜽𝑡𝑡−1 𝒛𝒛𝑡𝑡−1 + 𝒖𝒖𝑡𝑡 Based on the previous definitions, each state is associated with a modal solution of the form:
𝜽𝜽𝑡𝑡 = 𝜽𝜽𝑡𝑡−1 + 𝒗𝒗𝑡𝑡
𝑡𝑡 𝑡𝑡
State/parameter evolution j𝜙𝜙𝑚𝑚,𝑡𝑡
𝑞𝑞𝑚𝑚,𝑡𝑡 = 𝐴𝐴𝑚𝑚,𝑡𝑡 e 𝐴𝐴𝑚𝑚,𝑡𝑡 = � 𝜆𝜆𝑚𝑚,𝑡𝑡 𝜙𝜙𝑚𝑚,𝑡𝑡 = � arg 𝜆𝜆𝑚𝑚,𝑡𝑡
equations
𝜏𝜏=1 𝜏𝜏=1

Inst. Amplitude (IA) Inst. Phase (IP)

𝑦𝑦𝑡𝑡 = 𝑪𝑪𝑇𝑇 𝒛𝒛𝑡𝑡 + 𝜀𝜀𝑡𝑡 Hence, the (scalar) response takes the modal form:
Observation equation 𝑀𝑀 𝑀𝑀
(scalar response) 𝑦𝑦𝑡𝑡 = � 𝑞𝑞𝑚𝑚,𝑡𝑡 + 𝜀𝜀𝑡𝑡 = � 𝐴𝐴𝑚𝑚,𝑡𝑡 ej𝜙𝜙𝑚𝑚,𝑡𝑡 + 𝜀𝜀𝑡𝑡
𝑚𝑚=1 𝑚𝑚=1

2Avendaño et al., “Diagonal time dependent state space models for modal decomposition of non-stationary signals”, Signal Processing 147, pp. 208-223, 2018.

July 2021
SDU Mechanical Engineering

sdu.dk
State of the Art: Diagonal state space representations
Block-diagonal time-dependent SS representation2:

#sdudk
Estimation of modal components via joint state/parameter estimation
𝒛𝒛𝑡𝑡 = 𝑴𝑴 𝜽𝜽𝑡𝑡−1 𝒛𝒛𝑡𝑡−1 + 𝒖𝒖𝑡𝑡 � 0 , 𝚺𝚺𝜃𝜃 ),
Given a time-series 𝑦𝑦1𝑁𝑁 ≔ {𝑦𝑦1 , 𝑦𝑦2 , … , 𝑦𝑦𝑁𝑁 }, initial conditions 𝒛𝒛0 ∼ 𝒩𝒩(�𝒛𝒛0 , 𝚺𝚺𝑧𝑧0 ) and 𝜽𝜽0 ∼ 𝒩𝒩(𝜽𝜽
𝜽𝜽𝑡𝑡 = 𝜽𝜽𝑡𝑡−1 + 𝒗𝒗𝑡𝑡 and noise statistics (hyperparameters) 𝒫𝒫 = 𝚺𝚺𝑢𝑢 , 𝚺𝚺𝑣𝑣 , 𝜎𝜎𝜀𝜀2 , filtered and smoothed estimates of the
0

State/parameter evolution state and parameter vector trajectories may be obtained by calculation of the densities:
equations
𝑝𝑝 𝒛𝒛𝑡𝑡 𝑦𝑦1𝑡𝑡 , 𝒫𝒫) = 𝒩𝒩 𝒛𝒛� 𝑡𝑡𝑡𝑡 , 𝑷𝑷𝑡𝑡𝑧𝑧𝑡𝑡 𝑝𝑝 𝒛𝒛𝑡𝑡 𝑦𝑦1𝑁𝑁 , 𝒫𝒫) = 𝒩𝒩 𝒛𝒛� 𝑁𝑁 𝑁𝑁
𝑡𝑡 , 𝑷𝑷𝑧𝑧𝑡𝑡

� 𝑡𝑡𝑡𝑡 , 𝑷𝑷𝑡𝑡𝜃𝜃
𝑝𝑝 𝜽𝜽𝑡𝑡 𝑦𝑦1𝑡𝑡 , 𝒫𝒫) = 𝒩𝒩 𝜽𝜽 � 𝑁𝑁
𝑝𝑝 𝜽𝜽𝑡𝑡 𝑦𝑦1𝑁𝑁 , 𝒫𝒫) = 𝒩𝒩 𝜽𝜽 𝑁𝑁
𝑡𝑡 , 𝑷𝑷𝜃𝜃𝑡𝑡
𝑡𝑡
𝑦𝑦𝑡𝑡 = 𝑪𝑪𝑇𝑇 𝒛𝒛𝑡𝑡 + 𝜀𝜀𝑡𝑡 Filtering density Smoothing density
Observation equation
(scalar response) Both densities may be calculated with the help of a (non-linear) Kalman filter and fixed-interval
smoother.

2Avendaño et al., “Diagonal time dependent state space models for modal decomposition of non-stationary signals”, Signal Processing 147, pp. 208-223, 2018.

July 2021
SDU Mechanical Engineering

sdu.dk
Study objectives

#sdudk
Block-diagonal SS Applications: Current methods limited to
representations are useful: • Mechanical/structural vibration scalar response.
signals (rotating machinery, wind
• Tracking of non-stationary
turbines) What about multivariate
oscillatory components (sinusoids
with time-dependent amplitude • Biological signals (vector) signals?
and frequency) (electrocardiography, • Previous application examples
• Extraction of instantaneous electroencephalography) naturally comprise multivariate
amplitudes and frequencies signals!
• New methodology may lead to a
generalization of modal analysis

July 2021
SDU Mechanical Engineering

sdu.dk
Study objectives

#sdudk
In this work

• We generalize the block-diagonal SS model to the multiple output

case
• A mode mixing matrix is introduced, but needs to be estimated
• An Expectation-Maximization algorithm for hyperparameter
estimation (including mixing matrix and noise statistics) is
specifically tailored

July 2021
SDU Mechanical Engineering

sdu.dk
Method summary: The stochastic model
Multiple output block-diagonal time-dependent SS representation:

#sdudk
𝒛𝒛𝑡𝑡 = 𝑴𝑴 𝜽𝜽𝑡𝑡−1 𝒛𝒛𝑡𝑡−1 + 𝒖𝒖𝑡𝑡 𝚿𝚿 = 𝝍𝝍1 ⋯ 𝝍𝝍𝑀𝑀 𝑦𝑦𝑗𝑗,𝑡𝑡 = 𝜓𝜓1,𝑗𝑗 𝑧𝑧1,𝑡𝑡 + 𝜓𝜓2,𝑗𝑗 𝑧𝑧2,𝑡𝑡 + ⋯ + 𝜓𝜓2𝑀𝑀,𝑗𝑗 𝑧𝑧2𝑀𝑀,𝑡𝑡 + 𝜀𝜀𝑗𝑗,𝑡𝑡
𝜽𝜽𝑡𝑡 = 𝜽𝜽𝑡𝑡−1 + 𝒗𝒗𝑡𝑡 Mode mixing matrix Form of the 𝑗𝑗-th response
State/parameter evolution
equations How to estimate 𝜳𝜳?

Use Expectation-Maximization (EM) framework

𝒚𝒚𝑡𝑡 = 𝚿𝚿 𝑇𝑇 𝒛𝒛𝑡𝑡 + 𝜺𝜺𝑡𝑡
 SS dimension remains the same
Modified observation  Computationally stable
equation (vector response)
 Noise statistics can be jointly estimated
• Multiple iteration through KF/KS
• Slow convergence…

July 2021
SDU Mechanical Engineering

sdu.dk
Method summary: EM optimization

#sdudk
Dynamic model Hyperparameter update*
𝒫𝒫𝑗𝑗+1 = 𝑓𝑓 𝒛𝒛� 𝑁𝑁 � 𝑁𝑁 𝑁𝑁 𝑁𝑁
𝑡𝑡 , 𝜽𝜽𝑡𝑡 , 𝑷𝑷𝑧𝑧𝑡𝑡 , 𝑷𝑷𝜃𝜃𝑡𝑡 | 𝒫𝒫𝑗𝑗
Extended KF / fixed-
interval smoother
Initial condition update

𝒛𝒛0 = 𝒛𝒛� 𝑁𝑁
0 𝚺𝚺𝒛𝒛0 = 𝑷𝑷𝑁𝑁
𝑧𝑧0
Smoothed
state/parameter densities � 𝑁𝑁
𝜽𝜽0 = 𝜽𝜽 0 𝚺𝚺𝜽𝜽0 = 𝑷𝑷𝑁𝑁
𝜃𝜃0

𝑝𝑝 𝒛𝒛𝑡𝑡 𝑦𝑦1𝑁𝑁 , 𝒫𝒫) = 𝒩𝒩 𝒛𝒛� 𝑁𝑁 𝑁𝑁

𝑡𝑡 , 𝑷𝑷𝑧𝑧𝑡𝑡

� 𝑁𝑁
𝑝𝑝 𝜽𝜽𝑡𝑡 𝑦𝑦1𝑁𝑁 , 𝒫𝒫) = 𝒩𝒩 𝜽𝜽 𝑁𝑁
𝑡𝑡 , 𝑷𝑷𝜃𝜃𝑡𝑡

*Eq. (12) of paper

Expectation Step Maximization Step

July 2021
SDU Mechanical Engineering

sdu.dk
Illustrative example: Numerical model
Original modes Measured signals

#sdudk
𝑞𝑞1,𝑡𝑡
𝑦𝑦1,𝑡𝑡

𝑞𝑞2,𝑡𝑡 𝒚𝒚𝑡𝑡 = 𝚿𝚿 T 𝒒𝒒𝑡𝑡 + 𝜺𝜺𝑡𝑡

𝑦𝑦2,𝑡𝑡
𝑞𝑞3,𝑡𝑡

Coefficients of the numerical model

Mode definition: j𝜙𝜙𝑚𝑚,𝑡𝑡
𝑞𝑞𝑚𝑚,𝑡𝑡 = 𝐴𝐴𝑚𝑚,𝑡𝑡 e 𝑚𝑚 𝐴𝐴𝑚𝑚,0 𝐴𝐴𝑚𝑚,1 𝑓𝑓𝑚𝑚,0 𝑓𝑓𝑚𝑚,1 𝛼𝛼𝑚𝑚 𝜙𝜙𝑚𝑚,0
𝐴𝐴𝑚𝑚,𝑡𝑡 = 𝐴𝐴𝑚𝑚𝑚 + 𝐴𝐴𝑚𝑚𝑚 cos 2𝜋𝜋𝛼𝛼𝑚𝑚 𝑇𝑇𝑠𝑠 𝑡𝑡
1 1.0 0.2 50 -20 1/2
𝜙𝜙𝑚𝑚,𝑡𝑡 = 𝜙𝜙𝑚𝑚𝑚 + 2𝜋𝜋 𝑓𝑓𝑚𝑚0 𝑡𝑡 + ∑𝑡𝑡𝜏𝜏=1 𝑓𝑓𝑚𝑚,1 sin 2𝜋𝜋𝛼𝛼𝑚𝑚 𝑇𝑇𝑠𝑠 𝑡𝑡 Rnd.
2 0.8 0.4 80 20 1/2
−𝜋𝜋, 𝜋𝜋
3 -0.2 1.0 120 20 1/4

July 2021
SDU Mechanical Engineering

sdu.dk
Illustrative example: Numerical model
Modal decomposition is challenging: Original IA and IF

#sdudk
2
Mode 1 Mode 2 Mode 3
 Underdetermined system: 3 modes, 2 signals 1.5

 Singularities in both IA and IF: 1

IA
• Amplitude of mode 3 temporarily vanishes
on periodic intervals
0.5

• IF trajectories of modes 1 and 2 0

0 0.5 1 1.5 2 2.5 3 3.5 4
periodically cross each other Time [s]

 Noise: Low SNR 10dB

200
Mode 1 Mode 2 Mode 3

150
 Ridge tracking methods have poor performance
on this type of signal2 100

IF [Hz]
50

0
0 0.5 1 1.5 2 2.5 3 3.5 4
Time [s]

2Avendaño et al., “Diagonal time dependent state space models for modal decomposition of non-stationary signals”, Signal Processing 147, pp. 208-223, 2018.

July 2021
SDU Mechanical Engineering

sdu.dk
Illustrative example: Modal decomposition results
Results after 100 Monte Carlo simulations – SNR 10 dB

#sdudk
Mode 1

Mode 3

Mode 2
Mode 2 EKF + Block-diagonal TD-SS model
• Single output
• EM algorithm for hyperparameter
optimization

Mode 3
Mode 1

July 2021
SDU Mechanical Engineering

sdu.dk
Illustrative example: Modal decomposition results
Results after 100 Monte Carlo simulations – SNR 10 dB

#sdudk
Mode 1

Mode 3

Mode 2
Mode 2 EKF + Block-diagonal TD-SS model
• Complete output set
• Manual hyperparameter adjustment
(trial and error)

Mode 3
Mode 1

July 2021
SDU Mechanical Engineering

sdu.dk
Illustrative example: Modal decomposition results
Results after 100 Monte Carlo simulations – SNR 10 dB

#sdudk
Mode 1

Mode 3

Mode 2
Mode 2 EKF + Block-diagonal TD-SS model
• Complete output set
• EM algorithm for hyperparameter
optimization

Mode 3
Mode 1

July 2021
SDU Mechanical Engineering

sdu.dk
Illustrative example: Modal decomposition results
Results: Noise SNR 10 dB Results: Noise SNR 20 dB

#sdudk
Mode 1
Mode 2
Mode 3

MO-Manual EKF + Diagonal SS model /

Multiple Output / Manual
hyperparameter adjustment
MO-EM EKF + Diagonal SS model /
Multiple Output / EM algorithm
for hyperparameter adjustment
SO-EM EKF + Diagonal SS model /
Single Output / EM algorithm
for hyperparameter adjustment

July 2021
SDU Mechanical Engineering

sdu.dk
Conclusions

#sdudk
Modal decomposition of Generalization for the A numerical model is used
multivariate non-stationary Multiple Output case was
performed in this work to illustrate the method
signals has been studied • Overall uncertainty reduction in IA and
• Generalization requires introduction of IF estimates by use of multiple outputs
• A parametric approach based on block- mixing matrix
diagonal time-variant state space • EM algorithm also improves the tracking
models is adopted • Mixing matrix and noise covariances performance
optimized with EM algorithm

July 2021
SDU Mechanical Engineering

sdu.dk
#sdudk
Thanks!

Check out also our GitHub repository to find the codes

presented here and other examples!
https://github.com/ldavendanov/NS-modal-decomposition

Contact information:
Luis David Avendaño-Valencia
Assistant Professor, University of Southern Denmark
e-mail: [email protected]
Campusvej 55, 5230 Odense M, DK.

July 2021
SDU Mechanical Engineering

sdu.dk
References

#sdudk
1. Alickovic et al., “Performance evaluation of EMD, DWT and wavelet packet decomposition for automated epileptic seizure detection and prediction”, Biomedical Signal
Processing and Control 39, pp. 94-102, 2018.
2. Avendaño et al., “Diagonal time dependent state space models for modal decomposition of non-stationary signals”, Signal Processing 147, pp. 208-223, 2018.
3. Avendaño-Valencia and Fassois, “Stationary and non-stationary random vibration modelling and analysis for an operating wind turbine”, Mechanical Systems and Signal
Processing 47(1-2), pp. 263-285, 2014.
4. Bao et al., “Time-varying system identification using a newly improved HHT algorithm”, Computers and Structures 87, pp. 1611-1623, 2009.
5. Chen et al., “Application of decoupled ARMA model to modal identification of linear time-varying system based on the ICA and assumption of ‘short-time linearly varying’”,
Journal of Sound and Vibration 499, p. 115997, 2021.
6. Cox and Toth, “Linear parameter-varying subspace identification: A unified framework”, Automatica 2021, p. 109296, 2021.
7. Iatsenko et al., “Extraction of instantaneous frequencies from ridges in time-frequency representations of signals”, Signal Processing 125, pp. 290-303, 2016.
8. Poulimenos and Fassois, “Parametric time-domain methods for non-stationary random vibration modeling and analysis: A critical survey and comparison”, Mechanical
Systems and Signal Processing 20(4), pp. 763-816, 2006.
9. Stankovic et al., “Time-frequency decomposition of multivariate multicomponent signals”, Signal Processing 142, pp. 468-479, 2018.
10. Trendafilova, “Singular spectrum analysis for the investigation of structural vibrations”, Engineering Structures 242, p. 112531, 2021.

July 2021