Available online at www.sciencedirect.

com

ScienceDirect
Available online at www.sciencedirect.com
Structural Integrity Procedia 00 (2019) 000–000
www.elsevier.com/locate/procedia
ScienceDirect
Procedia Structural Integrity 37 (2022) 410–416

ICSI 2021 The 4th International Conference on Structural Integrity

Prospects of Structural Damage Identification Using Modal

Analysis and Anomaly Detection
Deniss Mironovs1*, Sandris Ručevskis1, Kārlis Dzelzītis1
1
Institute of Materials and Structures, Riga Technical University, Kalku street 1, Riga, Latvia, LV-1050, [email protected]

Abstract

Structural health monitoring (SHM) has significant importance in providing reliability, safety, and economical sensibility for
structures. High sensitivity and precision of SHM techniques can be quite demanding. Some SHM testing procedures require
structures, for example, wind turbine blades or aircraft structures, to be taken out of operation, which increases costs and causes
downtime. Other techniques include sophisticated signal processing and data analysis, which require skilled personnel. An
alternative is to use a form of system identification technique called Operational Modal Analysis, during which an operator
performs structural vibration measurements without disrupting normal operation of structure. The aim of the current study is to
develop a machine learning algorithm, which is able to identify damaged states of a structure, based on the assessment of its modal
parameters. The assessment of the structural health is performed using machine learning technique called Anomaly Detection. The
proposed method establishes a description of normality using features representing undamaged conditions and then test for
abnormality which indicates presence of damage in structure. Structure under test is one-meter-long laminated composite beam.
The finite element model of the beam is developed using eight-node shear-deformable shell elements. Damage in the beam is
introduced as delamination between the plies representing 5% of the total area of the beam. Numerical modal analysis is carried
out to obtain modal frequencies and corresponding mode shapes for the first three bending modes of both the healthy and damaged
beams. The results of the numerical experiments show possibilities of the proposed structural health monitoring method to identify
damage in composite structures under varying modal parameters.

© 2022 The Authors. Published by Elsevier©B.V.

2022 The Authors. Published by ELSEVIER B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of Pedro Miguel Guimaraes Pires Moreira
Peer-review under responsibility of Pedro Miguel Guimaraes Pires Moreira
Keywords: Structural Health Monitoring; Operational Modal Analysis; Anomaly Detection

* Corresponding author. Tel.: +371 26008121.

E-mail address: [email protected]

2452-3216 © 2022 The Authors. Published by ELSEVIER B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of Pedro Miguel Guimaraes Pires Moreira

2452-3216 © 2022 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of Pedro Miguel Guimaraes Pires Moreira
10.1016/j.prostr.2022.01.103
2 Deniss
Deniss Mironovs
Mironovs/ et al. / Procedia
Structural Structural00
Integrity Procedia Integrity
(2019) 37 (2022) 410–416
000–000 411

1. Introduction

In the past 30 years with the development of non-destructive testing methods structural health monitoring has been
rapidly developing in efficiency, practicality, and precision. There are numerous SHM approaches, vibration-based is
being one of them (Trendafilova et al. (2008), Weijtjens et al. (2017)). Vibration-based methods are usually based on
modal assessment of an investigated structure. It relies on the fact that if a change is introduced into a structure, then
this change modifies its physical properties and thus affects also modal parameters – frequency, mode shape and
damping (Maia et al (1997)).
A lot of research is focused on high precision of processing methods and high added value, for example damage
localisation. Katunin et al (2015) in their laboratory measurements show satisfactory reliability of vibration-based
SHM approaches for damage localization. Promising approaches are presented in Janeliukstis et al (2020), where
defect localisation is successfully utilized for a composite plate in a laboratory environment. Although it is
advantageous to have full information on structural state, still laboratory tested approaches have difficulties in
application to a real-life situation, which includes small differences in structures of the same type and series,
significant variability in operation conditions (loads, external forces, temperature) and other unpredictable factors.
Operational modal analysis OMA (Brincker and Ventura (2015)) is a branch of conventional Modal Analysis. It
comes from the name that OMA measurements are performed during structures operation. This is very advantageous,
as OMA potentially enables SHM without disrupting operations, eliminating downtime, increasing efficiency of
operations. OMA approach would be a desirable choice for structural health monitoring of structures under natural
excitation conditions, like wind turbines or helicopter blades. However, structures like these experience variable
operating conditions, as mentioned earlier, which have to be taken into account when analysing modal parameters.
One other issue that operators face in structural health monitoring is the huge amount of data (vibrational responses,
records of loads, rotating speed, temperature, etc.), especially for a series of typical structures. Human resources are
necessary to analyse how modal parameters change for a specific structure, taking into account different operating
conditions. This analysis can be substituted by an automated process.
This paper introduces application of machine learning technique called anomaly detection (Zimek and Schubert
(2017)) for analysis of modal parameters. The aim of the analysis is to detect sets of modal parameters which represent
an anomaly compared to the majority, which is an indication of damage. The proposed method establishes a
description of normality using features representing undamaged conditions and then test for abnormality which
indicates presence of damage in structure.

2. Theoretical considerations

2.1. Modal parameters estimation

Operational modal analysis is based on the experimental modal analysis (EMA) theory with slight modification. In
EMA the vibrational output response can be transformed into cross power spectra matrix between degrees-of-freedom:
𝐺𝐺𝑥𝑥𝑥𝑥 (𝜔𝜔) = 𝐻𝐻(𝜔𝜔)𝐺𝐺𝑓𝑓𝑓𝑓 (𝜔𝜔)𝐻𝐻(𝜔𝜔)𝐻𝐻 . (1)
where 𝐻𝐻(𝜔𝜔) is the matrix of frequency response functions and 𝐺𝐺𝑓𝑓𝑓𝑓 (𝜔𝜔) is the force energy matrix, and 𝐻𝐻(𝜔𝜔)𝐻𝐻 is the
Hermitian transform of the FRF matrix. The response is usually measured using accelerometers or strain gauges.
Recent developments by Mironov et al (2020) show that it is equally possible to use industrial piezo film sensors
which measure dynamic signals proportional to deformation velocity. The forces are obtained from load cells of other
force gauges (for example impact hammers). In OMA, however, force is not measured, but assumed to be Gaussian
in nature and evenly distributed across the excited system, i.e., force is assumed to be white noise. Thus, vibrational
responses are assumed to be systems FRFs:
𝐺𝐺𝑥𝑥𝑥𝑥 (𝜔𝜔) ∝ 𝐻𝐻(𝜔𝜔)𝐻𝐻(𝜔𝜔)𝐻𝐻 . (2)
Power spectra matrix 𝐺𝐺𝑥𝑥𝑥𝑥 is then decomposed at each frequency using Singular Value Decomposition (SVD),
𝐺𝐺𝑥𝑥𝑥𝑥 (𝜔𝜔) = 𝑈𝑈(𝜔𝜔) 𝑆𝑆(𝜔𝜔) 𝑉𝑉 𝐻𝐻 (𝜔𝜔), (3)
412 Deniss Mironovs et al. / Procedia Structural Integrity 37 (2022) 410–416
Author name / Structural Integrity Procedia 00 (2019) 000–000 3

where 𝑈𝑈(𝜔𝜔) represent singular vectors, while 𝑆𝑆(𝜔𝜔) is a diagonal matrix of singular values (Chauhan (2015)). SVD is
analogous to Principal Component Analysis (PCA) which give principal components (eigenvectors) of presented data
(Sen et al (2019)) with corresponding energy factors (eigenvalues). In modal analysis eigenvectors represent mode
shapes and eigenvalues represent corresponding frequencies.
The above mentioned approach in operational modal analysis is called Frequency Domain Decomposition, first
introduced by Brincker et al (2000) and widely used today, especially its enhanced version EFDD. Alternative modal
parameter estimation approach is based in time domain and is called Stochastic Subspace Identification (SSI),
explained by Brincker and Andersen (2006), where in case of covariance driven SSI, singular value decomposition is
applied to a Hankel matrix, formed from vibrational data covariance functions matrices in time domain. In data driven
SSI, Hankel matrix is formed from raw time data.
When eigenvectors and eigenvalues are extracted from raw data, it is a matter of choosing the right peak of power
spectrum, which is meant to represent the mode. Unfortunately, pick peaking is unreliable, there is a strong human
factor involved, it depends on signal-to-noise ratio, also harmonics can be mistaken for modes. Above mentioned
approaches develop mathematical polynomial models that best fit vibrational data (Hu et al (2012)), so modal
parameters are estimated from models instead.

2.2. Anomaly detection

Anomaly detection, also known as Novelty detection, as shown by Yan et al (2005) and Hu (2011) is a statistical
method for signaling whether a new data set is similar to established reference data set. This is a type of unsupervised
machine learning algorithm. The analysis first step is to create a representation of a normal condition of a structure.
The second step is to observe subsequent data to see if they are significantly different from the normal condition.
For this study anomaly detection data is formed as a matrix of modal frequencies and mode shape vectors. Other
potential factors to add to this matrix are load, temperature, humidity, wind speed, etc. The matrix is subjected to
calculation of mean values for each row and standard uncertainty for the presented data set of 𝑚𝑚 samples. One
important assumption is that the studied structures modal characteristics and operational conditions influence are
linear and thus have Gaussian distribution. A probability of each measured sample is calculated using multivariate
Gaussian distribution equation (Do (2008)):
1 1
(𝑥𝑥−𝜇𝜇)𝑇𝑇 Σ−1 (𝑥𝑥−𝜇𝜇)
𝑝𝑝(𝑥𝑥, 𝜇𝜇, Σ) = 𝑛𝑛 1 𝑒𝑒 −2 (4)
(2𝜋𝜋) 2 |Σ|2
where 𝑥𝑥, 𝜇𝜇, Σ are given sample vector, mean value vector and covariance matrix respectively, 𝑛𝑛 is the number of
features. To satisfy multivariate condition for number of samples to be significantly larger than number of features, a
12th degree polynomial approximation can be used for mode shape vectors.

3. Simulation

The study is performed in the following way. A finite element model of a cantilever carbon fiber reinforced polymer
composite beam with the spatial dimensions of 1×0.1 m a total thickness of 2 mm is constructed in Ansys software.
There are 20 layers in the beam, which are layered orthogonally relative to each other. The laminate lay-up of the
beam is (0/90)5s with a ply thickness of 0.1 mm. The FE model of the beam is developed using 100×10 eight-node
shear-deformable shell elements and the clamped boundary condition is applied at one edge of the beam. The material
properties of the model are displayed in table 1. Modal analysis is performed for this structure and 5 modes are
identified with representative modal parameters – frequency and mode shape. Mode shape is formed of strain data for
each node, which forms strain matrix 𝑁𝑁 × 2 for the whole beam, where 𝑁𝑁 is the number of nodes equal to 1111 and
number 2 represents X and Y strain directions. Strain matrix is normalized and approximated with 12th degree
polynomial using 202 elements.

Table 1. Material properties of the FE model.

direction x / xy plane direction y / yz plane direction z / xz plane
 Deniss Mironovs et al. / Procedia Structural Integrity 37 (2022) 410–416 413
4 Deniss Mironovs/ Structural Integrity Procedia 00 (2019) 000–000

Young’s modulus E, GPa 137.82 7.1796 137.82

Poisson’s ratio υ 0.33357 0.3 0.33357
Shear modulus G, GPa 2.9597 6.9269 2.9597
Density ρ, kg/m 3
1549.7

The result is formed into a matrix Y with 5 rows, corresponding to the number of calculated modes. Each row has
the frequency value as the first element and a mode shape vector for elements 2 to 203. Potentially this matrix can be
populated with measured data of environmental and operational factors, but the further data processing stays identical.
In order to simulate numerous cases of modal analysis measurements under stable environmental and operating
conditions, a Gaussian noise of ± 0.2 % was added to reference state modal parameters matrix Y, thus creating 5000
unique samples.
Damaged state is modelled as delamination with a degree of 5% of beam area. There are four damaged states, with
the same degree of delamination positioned at different locations along the beam’s length. The total amount of
generated samples is 5004.
For anomaly detection algorithm 3000 training samples are chosen out of total amount of samples, all being
representative of healthy state. The mean matrix Μ(𝑌𝑌) is calculated between these training samples. Each training
samples multivariate probability 𝑝𝑝𝑡𝑡𝑡𝑡 (𝑌𝑌) is then calculated, where the mean matrix has the maximum probability. The
distribution of 𝑝𝑝𝑡𝑡𝑡𝑡 (𝑌𝑌) is naturally Gaussian.
Next, cross-validation CV samples are formed – another 1000 healthy states and two damaged states, 1002 samples
combined. Multivariate probability 𝑝𝑝𝐶𝐶𝐶𝐶 (𝑌𝑌) is calculated for 1002 cross-validation samples. Scholars show that one
can set the threshold 𝜀𝜀 between states at a ±3𝜎𝜎 limit, i.e. standard deviation with a factor of 3, which gives 99.7 %
confidence in true results. Alternatively, for higher precision, 𝑝𝑝𝐶𝐶𝐶𝐶 (𝑌𝑌) is compared to a state label vector, where 0
denotes healthy and 1 – damaged states. This labeling allows to automatically establish a threshold between healthy
and damaged state, using F1 score calculation.
When the probability model is created and a healthy state threshold is established, a test sample set is formed out
of remaining 1000 healthy states and 2 damaged states, thus ensuring independent data for objective algorithm
validation. Probability 𝑝𝑝𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 (𝑌𝑌) is calculated and compared to the threshold. For 𝑝𝑝𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 (𝑌𝑌) < 𝜀𝜀 the sample is
considered to be damaged and vice versa.

a b

Fig. 1. a) Composite beam, reference state; b) healthy state 4th mode, 3rd bending shape.

4. Results and discussion

The presented algorithm has been able to correctly identify both damaged states from test sample set for modes 1,
2 and 4, which are all bending modes. For modes 3 and 5, which are torsional modes, the mode shape vector is formed
414 Deniss Mironovs et al. / Procedia Structural Integrity 37 (2022) 410–416
Author name / Structural Integrity Procedia 00 (2019) 000–000 5

by FE program ANSYS in orthogonal way (see Fig. 2c), so 12th order polynomial does not fit this vector and does not
represent subtle changes in the mode shape. This calls for another type of mode shape formation. What is more, the
damage is located symmetrically in the center of the beam axis, so the delamination has little influence on the torsional
strain.
One might argue with the decision to approximate mode shapes, however, this is dictated by properties of
multivariate Gaussian mathematics, i.e. covariance matrix is non-invertible if 𝑛𝑛 > 𝑚𝑚 (see Eq. (4)). It is possible,
however, to implement single variable Gaussian distribution for each feature and multiply these probabilities. Multiple
variables distribution product has large scale (sometimes up to 10 15) due to large data set, which complicates
computations. This is overcome by taking logarithm from probabilities, thus giving scale of tens and hundreds (for
example see Fig. 3c).

Delamination
region

a b

c d

Fig. 2. a) Damaged state 4th mode; b) Damaged state 5th torsional mode; c) Two mode shape vectors differently formed (X and Y directions
combined); d) Comparison between interpolated 4th mode shape vectors (only X direction plotted).

As a workaround for this issue, probabilities of individual parameters (frequency or mode shape) can be compared
between samples, see Fig. 3(a) for example. Probability histogram of test samples 𝑝𝑝𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 (𝑌𝑌) for 1st mode is shown in
Fig. 3(b), which clearly shows distinction between healthy and damaged states. More optimal representation of results
is seen in Fig. 3(c), as 𝑝𝑝𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 (𝑌𝑌) is plotted against frequency axis. The evaluation of state based only on modal
frequency changes would be incorrect, as for some cases damaged state frequency lies within healthy state frequency
distribution. Multivariable data set, formed as a combined frequency and mode shape matrix Y, allows to include all
possible changes in the analysis, thus giving a sample with completely different probability (Fig. 3(c)).
It would be advantageous to preprocess mode shape vectors and obtain single scalar, for example, by using Modal
Assurance Criterion (Pastor et al (2012)), or a relatively short vector, which describe slightest changes in original
mode shapes vectors. This way one can also fully utilize multivariate Gaussian distribution computation, which
captures correlations between features automatically, as opposed to single variable Gaussian distribution.
 Deniss Mironovs et al. / Procedia Structural Integrity 37 (2022) 410–416 415
6 Deniss Mironovs/ Structural Integrity Procedia 00 (2019) 000–000

5. Conclusion

A CFRP laminated composite beam structure was modelled in ANSYS software package. Modal analysis of the
beam was performed resulting in 5 modes, represented by modal frequencies and mode shape vectors. Gaussian noise
was added to the modal data set, creating 5000 independent samples. Additionally, different delamination was
introduced into the beam model, giving 4 extra modal data samples. A machine learning algorithm called anomaly
detection was implemented to 1) train the statistical model how the reference healthy state is represented by
frequencies and mode shapes; 2) set the threshold between known healthy and damaged states; 3) validate the
algorithm by using independent testing data set of 1002 samples (1000 healthy, 2 damaged).

a b

f, Hz p

f, Hz
Fig. 3. a) Probability of frequency values for healthy and damaged states for 5 th mode; b) multivariate 𝑝𝑝𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 (𝑌𝑌) for 1st mode; c) multivariate
𝑝𝑝𝑡𝑡𝑡𝑡𝑡𝑡𝑡𝑡 (𝑌𝑌) for 1st mode as function of frequency.

The algorithm has shown promising results, as it correctly distinguishes healthy and damaged states, even in cases
where modal frequencies are indistinguishable between mentioned states. It has been noted that FE modelling program
output format for mode shapes can influence the result. More research is planned for the topic, specifically to add
mode shape preprocessing to reduce number of variables in the data set for faster and more reliable algorithm
execution, as well as for reduction of probability scale.

Acknowledgement

This work has been supported by the European Regional Development Fund within the Activity 1.1.1.2 “Post-
doctoral Research Aid” of the Specific Aid Objective 1.1.1 “To increase the research and innovative capacity of
scientific institutions of Latvia and the ability to attract external financing, investing in human resources and
416 Deniss Mironovs et al. / Procedia Structural Integrity 37 (2022) 410–416
Author name / Structural Integrity Procedia 00 (2019) 000–000 7

infrastructure” of the Operational Programme “Growth and Employment” (No.1.1.1.2/VIAA/3/19/414) and by Riga
Technical University's Doctoral Grant program.

References

Trendafilova, I., Cartmell, M.P., Ostachowicz, W., 2008. Vibration-based damage detection in an aircraft wing scaled model using principal
component analysis and pattern recognition. Journal of Sound and Vibration 313, Issues 3–5, Pages 560-566.
Weijtjens, W., Verbelen T., Capello, E., Devriendt, C., 2017. Vibration based structural health monitoring of the substructures of five offshore wind
turbines. Procedia Engineering 199, 2294-2299.
Maia, N. M. M., Silva, J. M. M., et al, 1997. Theoretical and Experimental Modal Analysis. Research Studies Press, Baldock.
Katunin, A., Dragan, K., Dziendzikowski, M., Damage identification in aircraft composite structures: A case study using various non-destructive
testing techniques. Composite Structures 127:1-9.
Janeliukstis, R., Rucevskis, S., Chate, A., 2020. Condition monitoring with defect localisation in a two-dimensional structure based on linear
discriminant and nearest neighbour classification of strain features. Nondestructive Testing and Evaluation, Volume 35, Issue 1.
Brincker, R., Ventura, C., 2015. Introduction to Operational Modal Analysis, Wiley.
Zimek, A., Schubert, E., 2017. Outlier Detection. In: Liu L., Özsu M. (eds) Encyclopedia of Database Systems. Springer, New York, NY.
Mironov, A., Priklonskiy, A., Mironovs, D., Doronkin, P., 2020. Application of Deformation Sensors for Structural Health Monitoring of Transport
Vehicles. In: Kabashkin I., Yatskiv I., Prentkovskis O. (eds) Reliability and Statistics in Transportation and Communication. RelStat 2019.
Lecture Notes in Networks and Systems, vol 117, Springer.
Chauhan, S., 2015. Parameter estimation algorithms in operational modal analysis: a review. 6th International Operational Modal Analysis
Conference, Gijon, Spain.
Sen, D., Erazo, K., Zhang, W., Nagarajaiah, S., Sun, L., 2019. On the effectiveness of principal component analysis for decoupling structural
damage and environmental effects in bridge structures. Journal of Soundand Vibration, Volume 457, pages 280-298.
Brincker, R., Zhang, L., Andersen, P., 2000. Modal identification from ambient responses using domain decomposition. Proceedings, IMAC.
Brincker, R., Andersen, P., 2006. Understanding Stochastic Subspace Identification. In: Proc. 24th International Modal Analysis Conference
(IMAC), St. Louis (MO), USA.
Hu, SL.J., Bao, X., Li, H., 2012. Improved polyreference time domain method for modal identification using local or global noise removal
techniques. Science China Physics, Mechanics and Astronomy, 55, 1464–1474.
Hu, W-H., 2011. Operational modal analysis and continuous dynamic monitoring of footbridges. PhD thesis, Faculty of Engineering, University of
Porto.
Yan, A-M., Kerschen, G., De Boe, P., Golinval, J-C., 2005. Structural damage diagnosis under varying environmental conditions—Part I: A linear
analysis. Mechanical Systems and Signal Processing 19, pages 847–864.
Do, C.B., 2008. The Multivariate Gaussian Distribution. CS229 – Machine Learning. Stanford University.
Pastor, M., Binda, M., Harčarik, T., 2012. Modal Assurance Criterion. Procedia Engineering 48 (2012) 543 – 548, Elsevier.