Professional Documents
Culture Documents
Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct
A R T I C L E I N F O A B S T R A C T
Keywords: This study aims to demonstrate some capabilities of singular spectrum analysis (SSA) for the purposes of analysis
Structural vibrations of vibrating/dynamic structures. SSA is a powerful method for time series analysis, which encompasses elements
Signal analysis of classical time series analysis, multivariate statistics, principal component analysis, dynamical systems and
Singular spectrum analysis
signal processing. This paper introduces the application of SSA for vibration response time series and investigates
Vibration signal decomposition
Vibration signal reconstruction
the decomposition of the structural vibrations into harmonics, trend and noise components and their
reconstruction.
After an introduction to SSA and its application for structural vibration signals, the paper investigates a simple
model of a dynamic system and attempts to analyse it using classical methods of structural dynamics and SSA.
The results and the capabilities of the two methods are then compared and discussed. Furthermore, the paper
considers an experimental case of a layered GFRP composite plate which is subjected to free vibration and the
recorded signals are analysed using conventional structural dynamics methods and SSA. The results demonstrate
that SSA can be used as a powerful tool for analysis of the vibratory behaviour especially for structures with more
complex and/or nonlinear dynamic behaviour. It is capable of extracting the vibratory/oscillation patterns even
in the case of nonlinear oscillations.
https://doi.org/10.1016/j.engstruct.2021.112531
Received 26 October 2020; Received in revised form 15 April 2021; Accepted 9 May 2021
Available online 23 May 2021
0141-0296/Crown Copyright © 2021 Published by Elsevier Ltd. All rights reserved.
I. Trendafilova Engineering Structures 242 (2021) 112531
structural and machinery monitoring. There are already quite some different random effects/influences of the signal.
studies that suggest its application for structural and for machinery SSA can be used for signal denoising and this can be achieved by
damage and fault diagnosis and assessment [7–10,17]. But to our removing the random components of the decomposed signal. This
knowledge there are no studies investigating the analysis of the vibra feature of SSA has been exploited in a couple of vibration and other
tion response through SSA. The publication of Golinval [16] considers signal analysis papers [10,11,12,17]. In [9] SSA is used to denoise the
the application of principal component analysis (PCA) to examine the measured machine vibration signal and then to perform fault diagnosis
forced vibration of a linear multi-DOF system but again with the ulti with other methods. It should be mentioned that most of the methods
mate objective for structural damage assessment. This study suggests that suggest using SSA for machinery and structural monitoring diag
that a more in depth and rigorous analysis of the decomposition and the nosis actually suggest combining SSA with another method to achieve
reconstruction stages that SSA offers can bring more insight in the the final goal of the work [9–11]. Several papers can be mentioned that
analysis of the nonlinear vibration response signals and can be subse suggest using SSA in combination with other methods. [10] suggests to
quently developed for analysis of other important related problems combine SSA with empirical mode decomposition and ANFIS for pur
including damage assessment within changing environmental poses of bearing fault detection. The authors of [14] advise to use SSA
conditions. together with ARMAX modelling for structural damage assessment.
Accordingly, the idea of this study is to introduce the application of While the authors of [13] suggest the use of recursive SSA for purposes
SSA [32,33] to the measured vibration response and use it for purposes of real-time structural damage detection. The authors of [12] exploit
structural vibration analysis, simulation and prediction of the vibratory SSA and compare the residue between the original and the reconstructed
behaviour. The author is on the opinion that SSA holds much more signal using statistical metrics for purposes of structural damage
potential for application to structural vibrations, besides its use solely assessment. Loh et al. [15] used SSA to decompose the response signal
for purposes of structural damage assessment. This study aims to obtained from reinforced concrete frames aiming to monitor permanent
demonstrate that for a number of applications it can be used as a tool for deformations in that structure. The damage was extracted exploring the
modal analysis, specifically for nonlinearly vibrating structures and high-frequency variability of SSA. They developed a method to contin
especially when the spectral and the modal decomposition cannot be uously monitor the structural quality of the Taiwan Arc. SSA and
applied as the vibration signals have periodic but nonharmonic nonlinear principal component analysis (NPCA) were used to extract
components. residual deformation within the structure. In [18] Traore, Pantera et al
SSA has been used for purposes of signal analysis mainly in different apply SSA for nuclear reactors monitoring using acoustic emission sig
physical, social and climatological studies [15–19]. The purposes of nals. The signals are denoised with SSA and then analysed in order to
using SSA in most of these applications are pretty much similar – to distinguish between the signals from the healthy (non-corrupted) and
decompose the measured signal(s) into simpler easily interpretable the corrupted signals from a damaged reactor. The authors find that SSA
signals, which can be related to the behaviour of the system being provides very good separability between the two signal types. As was
analysed, and thus extract some information about it [32,33]. In some mentioned above SSA possesses two stages decomposition and recon
applications it is suggested to combine SSA with another procedure like struction. The second stage of SSA has not been much used even for the
e.g. AR modelling in [18] in order to improve the decomposition and the purposes of damage and fault diagnosis. The reconstruction stage gives
identification of seasonal/periodic components in the measured time the possibility to recover part or all of the features of the initially
series. In structural dynamics and vibration analysis SSA can be very measured raw signal. It is true that the most popular application is for
useful for decomposing especially complex and nonlinear vibration noise cancellation, but the reconstruction can be used for other purposes
response signals, in identifying the existing harmonics, trends and noise as well as in the reconstructed all or just some of the components can be
components and in reconstructing the noise-free signal. retained. This might be potentially useful for structural vibration signal
SSA can be successfully applied to any seemingly complex time series analysis.
with a potential structure in order to recover this structure [28–32]. In This study introduces and explores the application of SSA for the
doing so SSA is capable of 1) finding trends within the signal; 2) analysis, the decomposition and the reconstruction of the structural vi
smoothing in the sense of noise and other unwanted components elim bration signal. It demonstrates how SSA can be used to decompose the
ination; 3) extraction of periodic components with different magnitudes vibration signal into harmonics and it shows that on some occasions,
and identifying their frequencies/periods; 4) extraction of cycles with specifically for nonlinearly vibrating structures, this might be more
different periods; 5) finding structure and recovering the structure of the beneficial, and it presents a more powerful tool than the spectral anal
signal. Another important feature of SSA is that it is a non-parametric ysis and decomposition of the vibration signal.
method and as such does not require a model of the signal being ana The rest of the paper is organised as follows. It starts with a brief
lysed or any model of the dynamic system that produced it. introduction of SSA methodology and its main steps in the next para
SSA is applied in two main stages and with three main aims: graph. Then the application of SSA for structural vibration analysis is
decomposition, reconstruction and prediction [25–27]. The aim of the briefly discussed. After that a numerical example is offered where SSA is
first stage, the decomposition, is to decompose the original series into a applied for the purposes of the analysis of a simple nonlinear 2 DOF
sum of a small number of independent and interpretable components structure. The paper then proceeds with the application of SSA for the
such as a slowly varying trend, oscillatory components and a structure- purposes of analysis of the vibration of a composite beam. Eventually the
less noise. And the capabilities of this stage will be applied for the study draws some conclusions regarding the possible applications of SSA
purposes of vibration analysis. The second stage can be used to recon for the purposes of analysis of structural vibrations specifically when the
struct the original signal using a selected number of components. For vibration signal is non-linear/non-harmonic.
instance, one can select only a certain number of components which
contain the important signal features, like its trend and oscillatory 2. The methodology of singular spectrum analysis and its
components and discard the noise (structure-less) components thus application for the case of structural vibrations.
reconstructing a noise free signal. Some studies consider signal predic
tion as a third stage, but this is done on the basis of the linear recurrent 2.1. The SSA methodology
formula used for the prediction. Similarly to the reconstruction, it can be
applied with selected components from the decomposed signal, it can Usually the vibratory/dynamic behaviour of a structure is repre
include only the deterministic components, which carry some informa sented by its acceleration, velocity or displacement signal. This can be
tion about the signal structure including its trend and periodicities and introduced by a real-valued nonzero time series YT = (y1 , ..., yT ) of a
exclude components, which look structure-less and are responsible for certain sufficient length T. SSA is a time series analysis technique, which
2
I. Trendafilova Engineering Structures 242 (2021) 112531
is based on multivariate statistics. Its main purpose is to decompose the the left and right singular vectors of the trajectory matrix. The collection
original signal into a sum of signals, so that each component in this sum (√λi, Ui, Vi) is the i-th eigentriple of the matrix X ,√λi (i = 1, …, d) are
can be interpreted as either a trend, periodic or quasi-periodic compo the singular values of the matrix X and the set {√λi} is called the
nent (possibly amplitude-modulated), and noise. This is followed by a spectrum of the trajectory matrix X. If all the eigenvalues have multi
reconstruction of the original series. Thus, SSA technique consist of two plicity one, then the expansion (3) is uniquely defined. It can be shown
complementary stages: decomposition and reconstruction, both of that the SVD (3) is optimal in the sense that among all the matrices X(r)
∑
which include two separate steps. At the first stage the original signal is of rank r < d, the matrix ri=1 Xi provides the best approximation to the
decomposed into a smaller number of simple and more meaningful trajectory matrix X , so that ‖X − X(r)‖ is minimum. Note that ‖X‖2 =
signals and at the second stage the original series is reconstructed as a ∑d 2
i=1 λi and ‖Xi ‖ =λi for i = 1, … , d. Thus, we can consider the ratio λi /
sum of all or some of these simple signals, which can be used for pur ∑d
poses of forecasting new data points. Below a brief introduction on the i=1 λi as the characteristic of the contribution of the matrix Xi to the
∑ ∑
methodology of the SSA technique is provided, which mainly follows expansion of X (Eq. (3)). Consequently ri=1 λi / di=1 λi , the sum of the
[26, chap.1 and 2]. first r ratios, is the characteristic of the optimal approximation of the
trajectory matrix by the matrices of rank r.
2.1.1. Stage 1 Decomposition
This first stage consists of embedding and singular value decompo 2.1.2. Stage 2: Reconstruction
sition (SVD). Each of these is briefly introduced below. The reconstruction consists of grouping which is followed by diag
onal averaging.
2.1.1.1. Embedding. Embedding can be regarded as a mapping that
transfers a one-dimensional 2.1.2.1. Grouping. The grouping splits the elementary matrices Xi into
time series Y T = (y1 ,...,yT )into the multi-dimensional series X1 … XK several groups and sums the matrices within each group. The matrix XI
with vectors corresponding to certain group of indexesI = {i1 , i2 , ⋯., ik } is the matrix
XI = Xi1 + Xi2 + ⋯ + Xik . Then the trajectory matrix can be represented
Xi = (yi , ..., yi− L+1 ) ∈ RL , whereK = T − L + 1. (1)
as:
At this stage the measured acceleration signal is mapped onto K
X = XI1 + XI2 + ⋯ + XIm (4)
lagged vectors Xi, each of which has a window length L. The single
parameter of the embedding is the window length L, an integer such that The procedure of choosing the groups XI1 , XI2 , ⋯, XIm is called eigen
2 ≤ L ≤ T. The result of the embedding is the trajectory matrix triple grouping. The contribution of each group into the expansion of the
trajectory matrix is measured by the sum of the corresponding eigen
X = [X1 , ..., X K ] = (Xij )i, j = 1, ⋯., L, K (2)
values λi .
It should be noted that the trajectory matrix X is a Hankel matrix,
that is all the elements along the diagonal i + j = const are equal. 2.1.2.2. Diagonal averaging. This step transfers each matrix Xi into a
Embedding is a standard procedure in time series analysis and the future time series, which is an additive component to the original time series.
analysis depends on the purpose of the investigation. The choice of the Each k-th element of the resulting series is obtained by averaging all xij
window length L is a trade off between two considerations: quality of components over all i and j so that i + j = k + 2. This is the same process
extracted information and the statistical degree of confidence in this as the so-called Hankelization of the elementary matrices. The result of
information. The first consideration requires as wide as possible win this process is the Hankel matrix corresponding to an elementary matrix,
dow, while the second one requires many repetitions of the window, that which is the trajectory matrix of the series obtained from the diagonal
is T/L as large as possible [17,18]. averaging. The result of the Hankelization of a matrix Z is the Hankel
In vibration analysis one can impose a lower limit on the window matrix HZ, which is the trajectory matrix corresponding to the series
length depending on the desired minimum resolution frequency. Thus, if obtained as a result of the diagonal averaging. Note that the Hankeli
one is interested to observe a phenomenon at frequencies in the range of zation is an optimal procedure in the sense that the matrix HZ is the
f (at least equal to f) the minimum window length should be set at L ≥ ffs , nearest to Z (with respect to the matrix norm) among all Hankel matrices
where fs is the sampling frequency of the signal. of the corresponding size [15, chap. 6, sec. 2]. In its turn, the Hankel
matrix HZ uniquely defines the series by relating the value in the di
agonals to the values in the series. By applying the Hankelization pro
2.1.1.2. Singular value decomposition (SVD). The second stage of SSA,
cedure to all matrix components of (4), we obtain another expansion:
the SVD, subjects the trajectory matrix to singular value decomposition
and represents it as a sum of rank-one bi-orthogonal elementary X = XI1 + XI2 + ⋯ + XIm (5)
matrices. Denote by λ1, … , λL the eigenvalues of XX’ in decreasing order
of magnitude (λ1 ≥ …λL ≥ 0) and by U1, … , UL the orthonormal system Where XI1 = H.X. This is equivalent to the decomposition of the initial
(that is, (Ui, Uj) = 0 for i = j (the orthogonality property) and ‖Ui ‖ = 1 signal YT = (y1 , ⋯., yT ) into a sum of m signals:
(the unit norm property)) of the eigenvectors of the matrix XX’ corre
sponding to these eigenvalues. (Ui, Uj) is the inner product of the vectors ∑
m
yt = yt (k) (6)
Ui and Uj and ‖Ui ‖ is the norm of the vector Ui. If we denote by Vi = k=1
X’Ui/√λi, then the SVD of the trajectory matrix can be written as: ( )
(k)
Where YT = y1 , ⋯, yT corresponds to the matrix XIk .
(k) (k)
X = X1 + ... + Xd , (3)
3
I. Trendafilova Engineering Structures 242 (2021) 112531
acceleration(mm/s2)
The above process is applied in two main stages- decomposition and
reconstruction. At the decomposition stage, the measured acceleration
signal is embedded into a trajectory matrix Xij, i,j = 1,…., L,K using the
embedding window L, where K = T − L + 1 and T is the initial signal
length (step 1). Then SVD is applied (step 2). At this stage one is in
possession of L eigenvalues λi and the left and the right-hand side
decomposition matrices U and V. It should be noted that the principal
components (PC’s) of the decomposed time domain signal have a tem
poral structure and are defined over the same time interval as the initial
signal. The reconstruction stage (step 3) provides a series of L recon time(sec)
structed components, which can be used to reconstruct the initial signal
Fig. 2. Time domain signal for the 2-DOF system.
as per equation (6). The reconstructed components obviously have
temporal structure as the one of the initial signal. The full system of L
from the trend components the damping corresponding to the different
reconstructed components can be used for a full reconstruction of the
elementary oscillations or to the different modes can be estimated using
initial series, but it is also possible to use only several reconstructed
the log decrement method.
components to represent the initial signal (steps 4–7). Using the eigen
Normally, two very close or equal eigenvalues λi will define /corre
values and the reconstructed components one is able to identify the
spond to a periodic component. The rest of the eigenvalues are
periodic components, the trends and the noise components.
responsible for the trend of the signal and the noise. The first up to 10
In general, one is looking to denoise the measured signal, as all
RC’s should be enough for a good reconstruction of the initial signal, as
measured signals contain some noise, and to decompose it into simple
the noise components, which are in general not needed, are usually
elementary signals corresponding to single frequency oscillations. SSA
connected to the higher eigenvalues, above the 10th one. The periodic
has been proven to do both of these [23] and for a number of applica
RC’s would have periodic structure and it is possible to identify their
tions SSA is used to de-noise univariate time series and to study their
period either from their temporal representation or using their perio
spectral profile. In spectrum sensing applications, as structural vibration
dograms or spectra [18]. If the aim is to decompose the noise free signal,
analysis, the eigenvalue profile is used to determine the frequency
then the RC’s corresponding to the noise components can be removed.
bands, which are occupied by the signal and/or estimate the noise floor
The periodic components will correspond to the main linear and/or
of the frequency band. In a more general perspective SSA allows to
nonlinear modes of vibration of the structure analysed. They will
decompose the signal into a number of components which reflect
represent the harmonic/linear modes of vibration as well as the non-
distinct frequency bands of the original signal. The SSA-based subspace
harmonic / nonlinear ones. Consequently this temporal decomposition
model corresponds to an orthogonal matrix whose column vectors form
could be used in place of a modal decomposition as it is used to represent
an eigen-basis of the multidimensional space created by the embedding.
the signal as a sum of periodic components including linear and
Relevant components of the vibration signal are then obtained by pro
nonlinear ones. This is a potentially important application as a number
jecting it onto this eigen basis, while irrelevant components can be
of vibrating structures demonstrate nonlinear behaviour, which results
omitted. In this particular application one measures a free decay signa
in periodic but not sinusoidal vibration components, which generally
and is interested to extract from it the elementary oscillatory motions
cannot be identified in the signal spectrum. Then applying SSA would be
including the nonlinear ones. Similar periodicity extractions using SSA
more appropriate as it will identify both –the linear as well as the
have been done for some climatological and meteorological applications
nonlinear periodicities in the structural vibratory behaviour.
with the aim of identifying the seasonal/periodic components [24]. SSA
The following two sections &3 and &4 give two demonstrations for
actually decomposes the measured vibration signal into harmonics and
the application of SSA for the purposes decomposition and identification
the reconstruction provided by (6) was proven to give the best possible
of periodic/ modal components for two systems with nonlinearities: a
approximation of the signal in the presence of noise [2 2]. Additionally,
4
I. Trendafilova Engineering Structures 242 (2021) 112531
a)
b)
5
I. Trendafilova Engineering Structures 242 (2021) 112531
Table 2 paragraph: First normal frequency analysis through using the FFT is
Lay-up configuration of the carbon-fibre beam. applied and then the application of SSA on the time domain signals is
Number of Layers 3 1 6 1 3 discussed.
Degrees 0/90 +45/− 45 0/90 − 45/+45 0/90
much smaller.
Thus, solely from the measured time domain vibration response we
are able to recover all the information about the initial vibration signal
and from there extract information about the vibrating system. The
signal is decomposed into 1) trend, which is related to the first, the
fourth and possibly the fifth eigen triples, 2) oscillatory components
which are defined by the third and fourth and the sixth and the seventh
eigen triples and 3) noise and other structureless components, which
Acceleration (m/s2)
however have very little effect on the signal behaviour, and are pre
sented by the further eigen triples. From the measured time domain
vibration signal and the results of the SSA one can derive that the system
has two vibration modes -one linear and one non-linear. The frequencies
and the amplitudes of these modes (the two harmonic components) can
be determined using the reconstructed signals RC’s (see Appendix A,
Figure A1.3) by using their spectra or periodograms. And one can
confirm that the frequencies of these two harmonics, the linear one and
the nonlinear one, are exactly the same as in our simple mass-spring-
damper system.
Fig. 4 presents the complete reconstruction of the initial signal and Fig. 6. Time domain signal measured on the composite beam.
the reconstructed signal using the first 8 components. The complete
reconstruction reconstructs the whole signal fully including the trend,
the vibrational modes/the harmonic components and the noise. The first
8 components present a very good approximation of the initial signal,
because this reconstruction actually contains all the important features
which give information about the vibrating system, but without the
noise and the very small amplitude vibrations.
This exercise with a simple vibrating system demonstrates the
importance of the reconstruction phase of SSA for the case of signals
from vibrating structures. It allows to reconstruct all or part of the signal,
which contains particular features. For example, one can only recon
struct the vibration modes of the system and find the linear and the
nonlinear ones. The reconstruction can be also used for purposes of noise
cleaning. Then the noise components should be identified and then
removed.
6
I. Trendafilova Engineering Structures 242 (2021) 112531
Acceleration mm/s2
7
I. Trendafilova Engineering Structures 242 (2021) 112531
Acceleration (m/s 2)
Acceleration (m/s 2)
Time (seconds) Time (seconds)
Fig. 10. Signal reconstruction with RCs 1–6 and with RCs 1–10.
by RC’s 4 and 5. So these two are most likely contributing the second
frequency of 20 Hz but represent a nonlinear/non sinusoidal vibration
pattern. Thus, it can be presumed that this oscillation is possibly
nonlinear with a period of 0.05 s. Thirdly we were able to identify a third
oscillatory motion, which is defined by the 8th and the 9th reconstructed
components, but the contribution of this oscillatory motion is rather
small as compared to the amplitude of the measured signal. The fre
Acceleration (m/s2)
8
I. Trendafilova Engineering Structures 242 (2021) 112531
Subsequently SSA is applied to both cases. In both cases the use of SSA is demonstrated for the case of the composite beam vibrations where the
able to identify additional periodic components which eventually lead to particular modes of vibration are reconstructed, analysed and their
a much better signal reconstruction. frequencies are found (see &4). The importance of the reconstruction is
As it has been mentioned, SSA has been used for a number of pur also stressed by the fact that the signal and accordingly the vibrations
poses, in different disciplines and for a variety of different applications can be reconstructed with a certain precision. The whole reconstruction
including social, climatological and meteorological applications. is the most precise one and contains all the features of the signal. But it is
Recently some applications of SSA to more complex systems like the also possible to use a reconstruction with a certain number of recon
application for continuous monitoring of the Taiwan arc [15] and ap structed components, which can focus on specific features of the vi
plications for monitoring of nuclear safety experiments [18] have been bration. The full reconstruction as compared to reconstructions with a
published. The applications to more complex structural systems like limited component number is discussed for the the case of the 2-DOF
systems made of composite materials and reinforced concrete frames system (&3) and also for the case of the composite beam (&4).
[5–7,15–17] make a new advancement in the applications of SSA to real From the obtained numerical and experimental results, it can be said
and more complex systems and structures. that SSA is able to decompose the free vibration signal into periodic
Recently a number of applications of SSA for purposes of structural motions/components (sinusoidal and non-sinusoidal ones), trends and
health monitoring and for machinery condition monitoring have been noise. The frequencies of the oscillatory motions can be identified using
published [5–18], which make contribution to the engineering appli the corresponding reconstructed components. This can be used as a
cations of SSA. A lot of the mentioned publications use SSA for the useful tool for determining the vibration modes and modal frequencies
purposes of noise cleaning before the application of other methods for including the nonlinear ones, specifically for cases of nonlinearly
purposes of monitoring and damage assessment [8–10,12,13,25]. After vibrating structures. Subsequently the signal can be reconstructed as a
identifying and removing the noise components of the decomposed sum of oscillations removing the identified noise components. This can
signal, it is reconstructed without the structure less noise. This publi be used to “clean” the structureless noise of a recorded free vibration
cation puts a stress on the application of all the stages of SSA for the signal.
purposes of analysis of structural vibrations, without focussing on ap Thus, it can be concluded that SSA can be used as a powerful analysis
plications for monitoring and damage diagnosis. It draws the attention and reconstruction tool for free vibration signal analysis. Unlike the
to the application of the reconstruction stage for purposes of analysis signal spectrum it is able to identify the linear and the nonlinear periodic
and determination of the individual modes of vibration. This is a very motions/components of the vibration and clean the structureless noise
useful application of SSA for structural vibration and, as this study of the signal.
demonstrates, it can be used for analysis of the vibrations and extracting
all of their important features like periodic vibrations, linear and
nonlinear modes and trends. A stress is put on the importance of the Declaration of Competing Interest
reconstruction stage, which can be used not solely por noise cleaning
and removing the structure-less components. As it is demonstrated in The authors declare that they have no known competing financial
this study the reconstruction can be used for the reconstruction and the interests or personal relationships that could have appeared to influence
analysis of the whole vibration but also for different parts of it. For the work reported in this paper.
example, one may focus on a particular mode of the vibration, which can
be reconstructed and then analysed individually in terms of linearity or Appendix A. SSA results for the simulated spring-mass-damper
nonlinearity, types of nonlinearity, frequency and period of vibration system
and also its contribution to the whole vibration behaviour. This is
See Figs. A11-A13
Fig. A11. Eigen values λ and eigenvectors RHO for the spring-mass-damping system.
9
I. Trendafilova Engineering Structures 242 (2021) 112531
10
I. Trendafilova Engineering Structures 242 (2021) 112531
Fig. A22. The first 20 Principal components (PC’s) for the signal measured on the composite beam.
11
I. Trendafilova Engineering Structures 242 (2021) 112531
Fig. A23. The first 20 Reconstructed components (RC’s) measured on the composite beam.
12
I. Trendafilova Engineering Structures 242 (2021) 112531
[27] Jigljavski Anatoly. Singular Spectrum Analysis or time series. Springer; 2014. [31] Biglieri Ezio, Yao Kung. Some properties of singular value decomposition and their
[28] Hassani Hossein. Singular Spectrum Analysis: Methodology and Comparison. applications to digital signal processing. Signal Process 1989;18:277–89.
J Data Sci 2007;5:239–57. [32] A.M. Tome, D. Malafaia, A.R. Teixeira, E.W. Lang , On the use of Singular Spectrum
[29] Hassani Hossein, Ghodsib Zara. A glance at the applications of Singular Spectrum Analysis, Cornell University, Electrical Engineering and Systems Science > Signal
Analysis in gene expression data. Biomol Detect Quantif 2015;4:17–21. Processing,arXiv.org > eess > arXiv:1807.10679, p. 1–23.
[30] Bozzoa Enrico, Carniel Roberto, Fasino Dario. Relationship between Singular [33] Joliffe IT. Principal component analysis, Springer series in statistics. Springer;
Spectrum Analysis and Fourier analysis: Theory and application to the monitoring 2002.
of volcanic activity. Comput Math Appl 2010;60:812820.
13