Professional Documents
Culture Documents
1 Abstract: Aiming to tackle the challenges related to the characterization of modes in an acoustic
2 dispersive environment, we present a signal decomposition procedure, which separates modes
3 into individual components while preserving their integrity. With this approach, each mode can
4 be analyzed and processed individually, which carries opportunities for new insights into their
5 characterization possibilities. The proposed methodology is based on the eigenanalysis of the
6 autocorrelation matrix of the analyzed signal. When eigenvectors of this matrix are properly
7 linearly combined, each signal component can be separately reconstructed. A proper linear combi-
8 nation is determined based on the minimization of concentration measures calculated exploiting
9 time-frequency representations. In this paper, we engage a steepest-descent-like algorithm for the
10 minimization process. Numerical results support the theory and indicate the applicability of the
11 proposed methodology in the decomposition of acoustic signals in dispersive channels.
14 1. Introduction
Citation: M. Brajović Multivariate 15 Signals with time-varying spectral content, known as non-stationary signals, are
decomposition of Acoustic Signals in 16 analyzed using time-frequency signal (TF) signal analysis [1–17]. Some commonly used
Dispersive Channels. Mathematics 17 TF representations include Short-time Fourier transform (STFT) [1,3], pseudo-Wigner
2021, 1, 0. https://doi.org/ 18 distribution (WD) [1,9,12] and S-method (SM) [3]. Time-scale, multi-resolution analysis
19 using the wavelet transform is an additional approach to characterize non-stationary
Received: 20 signal behavior [4]. Various representations are primarily applied in the instantaneous
Accepted: 21 frequency (IF) estimation and related applications [8–15], since they concentrate the
Published: 22 energy of a signal component at and around the respective instantaneous frequency.
23 Concentration measures provide a quantitative description of signal concentration in
Publisher’s Note: MDPI stays neu- 24 the given representation domain [18], and can be used to assess the area of the time-
tral with regard to jurisdictional 25 frequency plane covered by a signal component.
claims in published maps and insti- 26 In order to characterize multicomponent signals, it is quite common to perform
tutional affiliations. 27 signal decomposition, which assumes that each individual component is extracted for
28 separate analysis, such as, for example, for the IF estimation. Decomposition techniques
Copyright: © 2021 by the authors.
29 for multicomponent signals are quite efficient if components do not overlap in the
Submitted to Mathematics for possi- 30 time-frequency plane [19–33]. The method originally presented in [20] can be used to
ble open access publication under 31 completely extract each component by using intrinsic relation between the PWD and
the terms and conditions of the 32 the SM. In the analysis of multicomponent signals, it is, however, common that various
Creative Commons Attribution (CC 33 components partially overlap in the time-frequency plane, making the decomposition
BY) license (https://creativecom- 34 process particularly challenging [19–33]. In this rather unfavorable scenario, overlapped
mons.org/licenses/by/ 4.0/). 35 components partially share the same domains of supports, and existing decomposition
36 techniques provide only partial results in the univariate case, limited to very narrow
37 signal classes. For example, linear frequency modulated signals are decomposed us-
38 ing the chirplet transform, Radon transform, or similar techniques [33], [24], whereas
39 sinusoidally modulated signals are separated using the inverse Radon transform [35].
40 However, these techniques cannot perform the decomposition when components have a
41 general, non-stationary form.
42 In the multivariate (multichannel) framework, it is assumed that the signals are
43 acquired using multiple sensors [20–44]. The sensors modify component amplitudes
44 and phases. However, the interdependence of values from various channels can be
45 utilized in the signal decomposition. This concept has been also exploited in the empiri-
46 cal mode decomposition (EMD) [39–43]. It has been previously shown that WD-based
47 decomposition is possible if signals are available in the multivariate form [21–23]. More-
48 over, the decomposition can be performed by directly engaging the eigenanalysis of
49 the auto-correlation matrix calculated for signals in the multivariate form [25–28]. It
50 should also be noted that the problem of multicomponent signal decomposition has
51 some similarities with the blind source separation [45–48]. However, the basic difference
52 is in the aim to extract each signal component in the decomposition framework, whereas
53 in the blind source separation, the aim is to separate signal sources (although one source
54 may generate several components). The mixing scheme from the blind source separation
55 framework is used in a recently proposed mode decomposition approach [49]. Another
56 line of the decomposition-related research includes mode decomposition techniques
57 which could be used for separation of modal responses and identification of progressive
58 changes in modal parameters [50].
59 Overlapped components pose a challenge in various applications, such as in biomed-
60 ical signal processing [44,51,52], radar signal processing [53] and processing of lamb
61 waves [54]. Popular approaches, such as the EMD and multivariate EMD (MEMD),
62 [39–43] cannot respond to the challenges posed by components overlapped in the time-
63 frequency plane and do not provide acceptable decomposition results in this particular
64 case [21]. Additionally, the applicability of these methods is highly influenced by am-
65 plitude variations of signal components. In this paper, we present a framework for the
66 decomposition of acoustic dispersive environment signals into individual modes based
67 on the multivariate decomposition of multicomponent non-stationary signals. Even
68 when simple signal forms are transmitted, acoustic signals in dispersive channels appear
69 in the multicomponent form with either very close or partially overlapped components.
70 Being reflected from the underwater surfaces and objects, each individual component
71 carries information about the underwater environment. That information is inaccessible
72 while the signal is in its multicomponent form. This makes analyzing acoustic signals
73 (mainly their localization and characterization) a challenging problem for research [55–
74 60]. The presented decomposition approach enables complete separation of components
75 and their individual characterization (e.g., IF estimation, based on which knowledge
76 regarding the underwater environment can be acquired).
77 We aim at solving this notoriously difficult practical problem by exploiting the
78 interdependencies of multiply acquired signals: such signals can be considered as multi-
79 variate and are subject to slight phase changes across various channels, occurring due
80 to different sensing positions and due to various physical phenomena, such as water
81 ripples, uneven seabed and changes in the seabed substrate. As each eigenvector of the
82 autocorrelation matrix of the input signal represents a linear combination of signal com-
83 ponents [25,27], slight phase changes across the various channels are actually favorable
84 for forming an undetermined set of linearly independent equations relating the eigen-
85 vectors and the components. Moreover, we have previously shown that each component
86 is a linear combination of several eigenvectors corresponding to the largest eigenvalues,
87 with unknown weights [25] (the number of these eigenvalues is equal to the number of
88 signal components). Among infinitely many possible combinations of eigenvectors, the
89 aim is to find the weights producing the most concentrated combination, as each indi-
Version October 29, 2021 submitted to Mathematics 3 of 30
90 vidual signal component (mode) is more concentrated than any linear combination of
91 components, as discussed in detail in [25]. Therefore, we engage concentration measures
92 [18] to set the optimization criterion and perform the minimization in the space of the
93 weights of linear combinations of eigenvectors.
94 We revisit our previous research from [21,25,27], and the main contributions are
95 two-fold. The decomposition principles of the auto-correlation matrix [25,27] are recon-
96 sidered. Instead of exploiting direct search [25] or a genetic algorithm [27], we show that
97 the minimization of concentration measure in the space of complex-valued coefficients
98 acting as weights of eigenvectors which are linearly combined to form the components,
99 can be performed using a steepest-descent-based methodology, originally used in the
100 decomposition from [21]. The second contribution is the consideration of a practical
101 application of the decomposition methodology.
102 The paper is organized as follows. After the Introduction, we present the basic
103 theory behind the considered acoustic dispersive environment in Section 2. Section
104 3 presents the principles of multivariate signal decomposition of dispersive acoustic
105 signals. The decomposition algorithm is summarized in Section 4. The theory is verified
106 on numerical examples and additionally discussed in Section 5. Whereas the paper ends
107 with concluding remarks.
142 of the transmitted signal [56,58,59,65]. Our aim is to decompose the received signal by
143 extracting each mode separately. Such extracted modes can be used in further processing,
144 such as IF estimation, characterization, and classification.
145 More general models assume a more complicated environment, where the boundary
146 of the bottom depends on the nature of the ocean, such as the roughness, depending
147 on the weather conditions and different environments in the ocean itself. These models
148 take into account the scattering of the transmitted signal as well. Our future work will
149 be oriented towards these models as well.
Figure 1. The considered underwater isovelocity setup. Water depth is D, transmitter depth is zt ,
the receiver is positioned at depth zr and the transmitter-receiver range is r.
with G p (zt ) and G p (zr ) being the modal functions of the p-th mode corresponding
to the transmitter
√ and the receiver [55,56,65], with the attenuation rate is At ( p, ω ) =
A( p, ω )/ r. Angular frequency is denoted by ω. The modes are dependent on
wavenumbers kr ( p, ω ) [55]
ω 2 π 2
k2r ( p, ω ) = − ( p − 0.5) . (2)
c D
156 The multicomponent structure of the transfer function is dependent on the number of
157 modes. The speed of sound propagation underwater is c = 1500 m/s.
The response to a monochromatic signal
This is the horizontal velocity of the corresponding phase for the p-th mode. The
energy propagation of the signal component is represented by the group velocity
dr (t) dω 1 1
u p (ω ) = = = dkr ( p,ω )
= q 2 . (6)
dt dkr ( p, ω ) d ω 2
dω dω c − ( p − 0.5) D
π
The received signal can be represented in the Fourier transform domain as a product
of the Fourier transform of the transmitted signal, S(ω ) and the transfer function H (ω )
of the channel in the normal-mode form, that is
X ( ω ) = S ( ω ) H ( ω ). (7)
In time domain, the received signal, x (n), is the convolution of the transmitted
signal, s(n) and the impulse response, h(n), from (1), i.e.
x ( n ) = s ( n ) ∗ h ( n ). (8)
158 In the following sections, we present an efficient methodology for the decomposition
159 of mode functions, which will make the problem of detecting and estimating the received
160 signal parameters straightforward.
P P
x (n) = ∑ s p (n) = ∑ A p (n)e jψp (n) . (11)
p =1 p =1
P jϕ
∑ p=1 α1p s p (n)e 1p
x (1) ( n ) a1 (n)e jφ1 (n)
P jϕ
x (n) a2 (n)e jφ2 (n) ∑ p=1 α2p s p (n)e 2p
(2)
x(n) =
..
= .. = .. , n = 1, 2, . . . , N, (12)
. .
.
x (C ) ( n ) aC (n)e jφC (n) jϕ
∑ Pp=1 αCp s p (n)e Cp
Matrix A of size C × P, which relates the signal in the c-th channel, x (c) (n) with signal
components, s p (n), in form of a linear combination
P P
x (c) ( n ) = ∑ acp s p (n) = ∑ αcp s p (n)e jφcp (16)
p =1 p =1
Version October 29, 2021 submitted to Mathematics 7 of 30
or
Xsen = AXcom . (19)
Now we can introduce the autocorrelation matrix R of the sensed signal, whose eigen-
vectors will be used in the multivariate decomposition framework:
H
R = Xsen Xsen , (20)
where (·) H denotes the Hermitian transpose. Individually, elements of this matrix are
products of x(n1 ) and x H (n1 ) at given instants n1 and n2 :
C
R ( n1 , n2 ) = x H ( n2 ) x ( n1 ) = ∑ x(i)∗ (n2 )x(i) (n1 ), (21)
i =1
179 where x(n1 ) = [ x (1) (n1 ) x (2) (n1 ) . . . x (C) (n1 )] T is the column vector of sensed values
180 at a given instant n1 . As it will be shown next, the eigenvectors of the autocorrelation
181 matrix, R, corresponding to the largest eigenvalues, consist of linear combinations of
182 signal components. This fact will be used to develop the algorithm for the extraction of
183 those components.
K
R = QΛQ H = ∑ λ p q p q Hp , (22)
p =1
185 with λ p being the eigenvalues and q p being the corresponding eigenvectors of matrix
186 R. Matrix Λ contains eigenvalues λ p , p = 1, 2, . . . , K on the main diagonal and zeros at
187 other positions. Matrix Q = [q1 , q2 , . . . , qK ] contains the eigenvectors q p as its columns.
188 Since R is symmetric, eigenvectors are orthogonal.
Version October 29, 2021 submitted to Mathematics 8 of 30
From definition (20) and based on relation Xsen = AXcom , autocorrrelation matrix R
can be rewritten as
P P
H
R = Xsen H
Xsen = Xcom A H AXcom = ∑ ∑ āij si s jH , (23)
i =1 j =1
where āij is used to denote elements of matrix A H A and si = [si (1), si (2), . . . si ( N )] H .
Elements of matrix R are
s1 ( n1 )
P P s2 ( n1 )
R(n1 , n2 ) = ∑ ∑ āij si (n1 )s∗j (n2 ) = s1∗ (n2 ), s2∗ (n2 ), . . . , s∗P (n2 ) AH A . . (24)
. .
i =1 j =1
s P ( n1 )
191 with M = min{C, P}. It will be further assumed that the number of sensors, C is such
192 that C ≥ P. In that case, there are M = P eigenvectors in (25). Two general cases can be
193 further discussed:
• Non-overlapped components. Note that the case when no components si and s j
overlap in the time-frequency plane implies that these components are orthogonal.
In that case, the right side of (25) becomes:
P P P P
R= ∑ si siH ∑ āij = κ p ∑ sp sH
p = ∑ λ p q p q Hp (26)
i =1 j =1 p =1 p =1
κ p s p = λ p q p , p = 1, 2, . . . , P. (27)
q1 = ξ 11 s1 + ξ 21 s2 + · · · + ξ P1 s P
q2 = ξ 12 s1 + ξ 22 s2 + · · · + ξ P2 s P
..
.
q M = ξ 1M s1 + ξ 2M s2 + · · · + ξ PM s P , (28)
198 infinitely many solutions of this undetermined system of equations, we aim at finding
199 those combinations which produce signal components. Moreover, since components
200 are partially overlapped, in the case when one component is detected, its contribution
201 should be removed from all equations (linear combinations of eigenvectors) in order to
202 avoid that it is detected again.
203 Obviously, for the detection of linear combinations of eigenvectors which represent
204 signal components, a proper detection criterion shall be established. Since non-stationary
205 signals can be suitably represented using time-frequency representations, and signal
206 components tend to be concentrated along their instantaneous frequencies, our criterion
207 will be based on time-frequency representations.
Time-frequency signal analysis provides a mathematical framework for a joint
representation of signals in time and frequency domains. If w(m) denotes a real-valued,
symmetric window function of length Nw , then signal s p (n) can be represented using
the STFT
Nw −1
STFTp (n, k) = ∑ w(m)s p (n + m)e− j2πmk/Nw , (30)
m =0
208 which renders the frequency content of the portion of signal around the each considered
209 instant n, localized by the window function w(n).
To determine the level of the signal concentration in the time-frequency domain, we
can exploit concentration measures. Among various approaches, inspired by the recent
compressed sensing paradigm, measures based on the `ρ norm of the STFT have been
used lately [18]
ρ
M STFTp (n, k) = kSTFT (n, k)kρ
= ∑ ∑ |STFT (n, k)|ρ = ∑ ∑ SPEC ρ/2 (n, k), (31)
n k n k
210 where SPEC (n, k) = |STFT (n, k)|2 represents the commonly used spectrogram, whereas
211 0 ≤ ρ ≤ 1. For ρ = 1, the `1 -norm is obtained.
212 We consider P components, s p (n), p = 1, 2, . . . , P. Each of these components has
213 finite support in the time-frequency domain, P p , with areas of support Π p , p = 1, 2, . . . , P.
214 Supports of partially overlapped components are also partially overlapped. Furthermore,
215 we will make a realistic assumption that there are no components that overlap completely.
216 Assume that Π1 ≤ Π1 ≤ · · · ≤ Π P .
Consider further the concentration measure M STFTp (n, k) of
y = η1 q1 + η2 q2 + · · · + ηP q P, (32)
217 for p = 0. If all components are present in this linear combination, then the concentration
218 measure kSTFT (n, k)k0 , obtained for p = 0 in (31), will be equal to the area of P1 ∪ P2 ∪
219 . . . PP .
220 If the coefficients η p , p = 1, 2, . . . , P are varied, then the minimum value of
221 the `0 -norm based concentration measure is achieved for coefficients η1 = γ11 , η2 =
222 γ21 , . . . , ηP = γP1 corresponding to the most concentrated signal component s1 (n), with
223 the smallest area of support, Π1 , since we have assumed, without the loss of generality,
224 that Π1 ≤ Π1 ≤ · · · ≤ Π P holds. Note that, due to the calculation and sensitivity issues
225 related with the `0 -norm, within the compressive sensing area, `1 -norm is widely used
226 as its alternative, since under reasonable and realistic conditions, it produces the same
227 results [25]. Therefore, it can be considered that the areas of the domains of support in
228 this context can be measured using the `1 -norm.
The problem of extracting the first component, based on eigenvectors of the auto-
correlation matrix of the input signal can be formulated as follows
229 Note that if β 11 = γ11 , β 21 = γ21 , . . . β P1 = γP1 holds, then the component is exact,
230 that is, s̄1 = s1 holds. In the case when the number of signal components is larger than
231 two, the concentration measure in (33) can have several local minima in the space of
232 unknown coefficients η1 , η2 , . . . , ηP , corresponding not only to individual components
233 but also to linear combinations of two, three or more components. Depending on the
234 minimization procedure, it can happen that the algorithm finds this local minimum, that
235 is, a set of coefficients producing a combination of components instead of an individual
236 component. In that case, we have not extracted successfully a component since s̄1 6= s1
237 in (34), but as it will be discussed next, this issue does not affect the final result, as the
238 decomposition procedure will continue with this local minimum eliminated.
q p − q1H q p q1
qp = q . (36)
1 − |q1H q p |2
240 This ensures that s̄1 is not repeatedly detected afterward. If s̄1 = s1 , then the
241 first component is found and extracted, whereas its projection on other eigenvectors is
242 removed.
The described procedure is then repeated iteratively, with linear combination y =
η1 q1 + η2 q2 + · · · + ηP q P with first eigenvector q1 = s̄1 and eigenvectors q p , p =
1, 2, . . . , P, modified according to (36). Upon detecting the second component (or linear
combination of a small number of components), s̄2 , the second eigenvector is replaced,
q1 = s̄2 , whereas its projections from remaining eigenvectors is removed using
q p − q2H q p q2
qp = q . (37)
1 − |q2H q p |2
243 The process repeats until all components are detected and extracted. These princi-
244 ples are incorporated into the decomposition algorithm presented in the next section.
where
x (1) ( 1 ) ... x (1) ( N )
(2)
x (1) ... x (2) ( N )
Xsen =
.. .. .. . (39)
. . .
x ( C ) (1) ... ( C
x (N))
P
1
qi =
C ∑ β pk q p (40)
p =1
275 if this situation happens, upon applying signal deflation by removing projection of the
276 linear combination of components from other eigenvectors, a linear dependence among
277 eigenvectors will still remain, and it will not allow Nu to fall to zero. This returns the
278 algorithm to Step 3, and the procedure for the component detection repeats, but with the
279 local minimum removed from the concentration measure since all the eigenvectors are
280 already updated in the previous cycle. Note that the component index k is reset to zero
281 in this case.
282 Moreover, it should be emphasized that, while the presented procedure produces P
283 eigenvectors, which is exactly equal to the given number of components, this number
284 is not always a priori known. In practical applications, it can be determined based on
285 eigenvalues of matrix R. As it will be illustrated in numerical examples, the largest
286 eigenvalues correspond to signal components [21,25]. The remaining eigenvalues cor-
287 respond to the noise. Therefore, a simple threshold can be used to calculate the exact
288 number of signal components. Namely, we simply count the number of eigenvalues
289 larger than a predefined threshold T, being a small positive constant. In the presence
290 of the noise, threshold T should be at least equal to the noise variance. The larger the
291 noise, the larger are the eigenvalues corresponding to the noise (thus, larger should the
292 threshold T be). Of course, the procedure works without the exact information about
293 the number of components: for p > P, eigenvectors q p contain only the noise after the
294 decomposition is finished.
P P
yr+ = ∑ β pk q p + ( β p0 k + ∆)q p0 = ∑ β pk q p + ∆q p0 = y + ∆q p0 (41)
p =1 p =1
p 6 = p0
is formed. For this signal, with energy normalized using the `2 -norm, that is
yr+ y + ∆q p0
= , (42)
+
k yr k2 ky + ∆q p0 k2
Output:
• Coefficients β 1k , β 2k , . . . , β Pk
Version October 29, 2021 submitted to Mathematics 14 of 30
yi+ yi + j∆q p0
= . (46)
kyi+ k2 ky + j∆q p0 k2
and
yi− yi − j∆q p0
= . (47)
k y − k2 ky − j∆q p0 k2
are formed, serving as a basis to calculate the corresponding concentration measures
and
Mi− =
STFT yi−
1 .
(49)
Now, based on the calculated concentration measures for variations of real and imaginary
parts, concentration measure gradient ∇ p is calculated and used to determine the
direction for the update of β p0 k
P
y= ∑ β pk q p ,
p =1
304 For coefficients β pk , p 6= i, the gradient is equal to zero, that is, ∇ p = 0, meaning that
305 these coefficients should not be updated.
Coefficient β pk is updated using the calculated gradient, in the steepest descent
manner
β pk ← β pk − ∇ p , for p = 1, 2, . . . , P. (52)
306 The process is repeated until ∑ Pp=1 |∇ p |2 becomes smaller than a predefined preci-
307 sion ε.
308 5. Results
For the visual presentation of the results, the discrete Winger distribution (pseudo-
Wigner distribution) will be used in our numerical examples. For a discrete signal x (n),
this second-order time-frequency representation is calculated according to
Nw −1 mk
WD (n, k) = ∑ w(m)w(−m) x (n + m) x ∗ (n − m)e− j4π Nw , (53)
m =0
315 vector of the pth extracted component, whereas s p is the actual (original) pth signal
316 component.
• Estimation results for the discrete IFs obtained from the two previous WDs are
compared by the means of Mean Squared Error (MSE) for each pair of components.
The IF estimate based on the WD of the original pth component, WD op (n, k), p =
1, 2, . . . , P is calculated as [3]
whereas the IF estimate based on the WD of the pth component extracted by the
proposed approach is calculated as
317 Since the extracted components do not have any particular order after the decom-
318 position is finished, the corresponding pairs of original and extracted components
319 are automatically determined using the following procedure:
320 1. For p in 1, 2, . . . , P repeat steps (a)–(f)
321 (a) Calculate kop (n) based on (54) for analytically defined component s p .
322 (b) Run decomposition algorithm. Use only eigenvectors corresponding
323 to the largest P eigenvalues. Each of these eigenvectors, q p , contain
324 an extracted signal component. If P is not given, estimate number of
325 components, P, as the number of eigenvalues, λ p , of matrix R, larger
326 than threshold T = σε2 + 10−4 .
327 (c) Initialize set E ← ∅, to store the errors between IFs estimated based
328 on the given original component s p and extracted (unordered) com-
329 ponents qi , i = 1, 2, . . . , P, being the outputs of the decomposition
330 procedure.
331 (d) For each extracted component, q1 , q2 , . . . , q P repeat steps i–iii:
i. Calculate the IF estimate kei (n) as:
ii. Calculate Mean Squared Error (MSE) between kop (n) and kei (n)
as
1 N −1 o 2
N n∑
e
MSE(i ) ← ( n ) − k ( n ) .
k p i
=0
N −1 2
1
∑
o
MSE p = k p (n) − kep (n) , p = 1, 2, . . . , P, (56)
N n =0
1 2 2
0.8
frequency
frequency
0.6 0 0
0.4
0.2 -2 -2
(a) 0 5 10 15
(b)
-100 -50 0 50 100
(c)
-100 -50 0 50 100
eigenvalue index time time
Figure 2. (a) Eigenvalues of autocorrelation matrix R, (b) Wigner distribution of the signal from
Example 1 and (c) Spectrogram of the signal from Example 1. Signal consists of P = 6 non-
stationary components. The signal is embedded in an intensive complex, Gaussian, zero-mean
noise with σε = 1. The number of channels is C = 128. The largest 6 eigenvalues correspond to
signal components.
337 where TWDo = 0.1 max{|WD op (n, k )|} is a threshold used to determine whether a com-
338 ponent is present at the considered instant n. If it is smaller than 10% of the maximal
339 value of the WD, it indicates that the component is not present.
341 −128 ≤ n ≤ 128 and N = 257. Phases ϑc , c = 1, 2, . . . , C, are random numbers with uni-
342 form distribution drawn from interval [−π, π ]. The signal is available in the multivariate
h iT
343 form x(n) = x (1) (n), x (2) (n), . . . , x (C) (n) , and is consisted of C channels since it is
344 embedded in a complex-valued, zero-mean noise ε(c) (n) with a normal distribution of its
345 real and imaginary part, N (0, σε2 ). Noise variance is σε2 , whereas A p = 1.2. Parameters
346 f p and φ p are FM parameters, while L p is used to define effective width of the Gaussian
347 amplitude modulation for each component.
348 We generate signal of the form (58) with P = 6 components, whereas the noise
349 variance is σε = 1. The respective number of channels is C = 128. The corresponding
350 autocorrelation matrix, R, is calculated, according to (20), and the presented decompo-
351 sition approach is used to extract the components. Eigenvalues of matrix R are given
352 in Fig. 2 (a). Largest 6 eigenvalues correspond to signal components, and they are
353 clearly separable from the remaining eigenvalues corresponding to the noise. WD and
354 spectrogram of the given signal (from one of the channels) are given in Fig. 2 (b), (c),
355 indicating that signal is not suitable for the classical TF analysis, since the components
356 are highly overlapped.
357 Each of eigenvectors of the matrix R is a linear combination of components, as
358 shown in Fig. 3. The presented decomposition approach is applied to extract the
359 components by linearly combining the eigenvectors from Fig. 3. The results are shown
360 in Fig. 4 (a)–(f). Although a small residual noise is present in the extracted components,
361 they highly match the original components, presented in Fig. 4 (g)–(l). The original
362 components in Fig. 4 (g)–(l) are not corrupted by the noise.
Version October 29, 2021 submitted to Mathematics 17 of 30
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
363 As a measure of quality, we engage MSE p given by (56), which is the error between
364 the IF estimation result based on the pth extracted signal component (shown in Fig.
365 4 (a)–(f)) versus the IF estimation calculated based on the WD of original, noise-free
366 component (from Fig. 4 (g)–(l)). The IF estimates and the corresponding MSEs are, for
367 each pair of components, presented in Fig. 5, for standard deviation of the noise σε = 1,
368 where the number of channels is C = 128.
369 Since MSE p given by (56) serves as a measure of the component extraction quality,
370 we evaluate the decomposition performance for various standard deviations of the noise,
371 σε ∈ {0.1, 0.4, 0.7, 1.0, 1.3, 1.9, 2.1} . Results are presented in Table 1. The presented MSEs
372 are calculated by averaging the results obtained based on 10 realizations of multichannel
373 signal of the form (58) with random phases ϑc , c = 1, 2, . . . , C and corrupted by random
374 realizations of the noise ε(c) (n)r, for each observed variance (standard deviation) of the
375 noise. Based on the results from Table 1, it can be concluded that each signal component
376 is successfully extracted for noise characterized by standard deviation up to σε = 1.3.
377 For stronger noise, only some components are successfully extracted. It shall be noted
378 that the performance of the algorithm depends also on the number of channels, C. For
379 the results from Table 1, the number of channels was set to C = 256. Larger value of C
380 increases the probability of successful decomposition, as investigated in [25].
381 Example 2. The decomposition algorithm is tested on a more complex signal of the
382 form (58), with P = 8 components, whereas the standard deviation of the noise is now
383 σε = 0.1. The number of channels is C = 128. After the input autocorrelation matrix, R,
384 is calculated, according to (20), eigendecomosition produced the eigenvalues given in
385 Fig. 6(a). Signal components overlap in the time-frequency domain, and therefore, the
386 corresponding Wigner distribution and spectrogram shown in Fig. 6(b) and (c) cannot be
387 used as adequate tools for their analysis. Fig. 7 indicates that the components are neither
388 visible in the time-frequency representation of any eigenvector corresponding to the
389 largest eigenvalues. This is in accordance with the fact that eigenvectors contain signal
390 components in the form of their linear combinations. Upon applying the presented
391 multivariate decomposition procedure on this set of eigenvectors, we obtain results
392 presented in Fig. 8. By comparing the results with Wigner distributions of individual,
393 noise-free components, shown in Fig. 9, comprising the considered multicomponent
394 signal, it can be concluded that the components are successfully extracted with preserved
395 integrity. This is additionally confirmed by the IF estimation results shown in Fig. 10,
Version October 29, 2021 submitted to Mathematics 18 of 30
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
Figure 4. Extracted and original signal components of the non-stationary multicomponent mul-
tichannel signal considered in Example 1. Panels (a)–(f) present extracted using the proposed
approach, whereas panels (g)–(l) show Wigner distributions calculated for individual components
of the original, noise-free signal.
396 where even lower MSE values for each component can be explained by the lower noise
397 level, as compared with results from the previous example.
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
Figure 5. Instantaneous frequency estimation for individual signal components based on: extracted
signal components (dashed black) and original signal components (solid white). MSEs between
the two IF estimates is provided for each component of the signal from Example 1. The noise
variance is σε2 = 1. Decomposition is based on C = 128 channels.
Eigenvalues WD of signal Spectrogram of signal
1 2 2
0.8
frequency
frequency
0.6 0 0
0.4
0.2 -2 -2
(a) 0 5 10 15
(b)
-100 -50 0 50 100
(c)
-100 -50 0 50 100
eigenvalue index time time
Figure 6. (a) Eigenvalues of autocorrelation matrix R, (b) Wigner distribution of the signal from
Example 2 and (c) Spectrogram of the signal from Example 2. Signal consists of P = 8 non-
stationary components. The signal is embedded in an intensive complex, Gaussian, zero-mean
noise with σε = 1. The number of channels is C = 128. The largest 8 eigenvalues correspond to
signal components.
for p = 4, 5, 6. Again, signal is defined for discrete indices −128 ≤ n ≤ 128 and N = 257
Phases ϑc , c = 1, 2, . . . , C, are random numbers with uniform distribution drawn from
interval [−π, π ]. The resulting multicomponent signal is formed in cth channel as:
6
∑ sp
(c) (c)
x p (n) = ( n ) + ε ( c ) ( n ), (63)
p =1
398 and is, as in previous examples, embedded in additive, white, complex-valued Gaussian
399 noise, now with variance σε = 1. The number of channels is C = 256. Eigenvalues of
400 the autocorrelation matrix R, WD and spectrogram are given in Fig. 11, again proving
401 the that the considered signal with heavily overlapped components cannot be analyzed
402 with these tools. Eigenvectors corresponding to the largest 6 eigenvalues are given in
403 Fig. 12. Extracted and original components can be visually compared in Fig. 13, again
Version October 29, 2021 submitted to Mathematics 20 of 30
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2
frequency
frequency
0 0
-2 -2
(g) (h)
-100 -50 0 50 100 -100 -50 0 50 100
time time
404 proving the efficiency of the approach, even in the case for components with a faster
405 varying frequency content. This is additionally confirmed by IF estimation results in Fig.
406 14. Larger estimation errors when faster sinusoidal frequency modulations are present
407 are related to poorer concentration of the WD in these cases [3].
P P
∑ sp ∑
(c)
x (c) ( n ) = (n) = At (m, ω0 ) exp( jω0 n − jk c ( p, ω0 )r ), (64)
p =1 p =1
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2
frequency
frequency
0 0
-2 -2
(g) (h)
-100 -50 0 50 100 -100 -50 0 50 100
time time
Table 1: Mean squared errors (MSE) between IF estimations based on extracted and
original components, for signal from Example 1 with P = 6 components. MSE p , p =
1, 2, . . . , 6 corresponds to the pth component. The results are presented for various values
of the standard deviation of the noise, σε . The results are averaged based on 10 random
realizations of signals with random phases and noise, for each considered value of σε .
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2
frequency
frequency
0 0
-2 -2
(g) (h)
-100 -50 0 50 100 -100 -50 0 50 100
time time
D = 20m and ϑc is a random variable drawn from interval [−0.25, 0.25] with uniform
distribution, therefore corresponding to depth variations of ±0.25 cm, modeling channel
depth changes due to surface waves or uneven seabed. The speed of sound propagation
underwater is c = 1500 m/s. The same results in this example are obtained for a more
precise speed, i.e. at c = 1480 m/s. The received multichannel signal is of the form
x (1) ( n )
(2)
x (n)
x(n) = .
. (66)
..
x (C ) ( n )
415 6. Discussion
416 Decomposition of non-stationary multicomponent signals has been a long-term,
417 challenging topic in time-frequency signal analysis [19–24,24–33,35]. Although decom-
418 position of non-overlapping components can be done using the S-method relations with
Version October 29, 2021 submitted to Mathematics 23 of 30
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2
frequency
frequency
0 0
-2 -2
Figure 10. Instantaneous frequency estimation for individual signal components based on: ex-
tracted signal components (dashed black) and original signal components (solid white). MSEs
between the two IF estimates is provided for each component of the signal from Example 2. The
noise variance is σε2 = 0.1. Decomposition is based on C = 128 channels.
419 the WD [20], this approach cannot be applied when the components partially overlap,
420 i.e. share the same domain of support in the time-frequency plane.
421 Other alternative methodologies are specialized for some specific signal classes,
422 and are efficient in the case of partially overlapped components [24,33,35]. In this sense,
423 chirplet and Radon transform-based decomposition is applicable for linear frequency
424 modulated signals [24,33]. Inverse Radon transform has produced excellent results in
425 separation of micro-Doppler appearing in radar signal processing, characterized by a
426 sinusoidal frequency modulation and periodicity [35]. However, outside the scope of
427 their predefined signal models, these techniques are inefficient in separation of non-
428 stationary signals characterized by some different laws of non-stationarity. Another
429 very popular concept, namely the EMD, has been also applied multivariate data [39–
430 43]. However, successful EMD-based multicomponent signal decomposition is possible
431 only for signals having nonoverlapping components in the TF plane. Amplitude
432 variations of components pose an additional challenge to the EMD-based decomposition.
433 The efficiency of the proposed method does not depend on the considered frequency
434 range, but only on the ability of a time-frequency representation to concentrate signal
435 components in the time-frequency plane. We use the STFT in concentration measure
436 (31) due to its ability to concentrate signal energy at the instantaneous frequency of
437 individual signal components. The decomposition approach studied in this paper
438 successfully extracts components highly overlapped in the time-frequency plane. The
439 method is not sensitive to the extent of overlap of signal components.
440 Since the modes appearing in the considered acoustic dispersive environment
441 framework are characterized by a non-linear (and non-sinusoidal) law of frequency
Version October 29, 2021 submitted to Mathematics 24 of 30
1 2 2
0.8
frequency
frequency
0.6 0 0
0.4
0.2 -2 -2
(a) 0 5 10 15
(b)
-100 -50 0 50 100
(c)
-100 -50 0 50 100
eigenvalue index time time
Figure 11. (a) Eigenvalues of autocorrelation matrix R, (b) Wigner distribution of the signal from
Example 3, and (c) Spectrogram of the signal from Example 3. Signal consists of P = 6 non-
stationary components. The signal is embedded in an intensive complex, Gaussian, zero-mean
noise with σε = 1. The number of channels is C = 256. The largest 6 eigenvalues correspond to
signal components.
WD of eigenvector WD of eigenvector WD of eigenvector
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
442 variations and have a partially overlapped support, neither of the mentioned univariate
443 techniques can produce acceptable decomposition results.
444 7. Conclusions
445 Characterization of modes in the acoustic dispersive environment is an ongoing
446 research topic. As modes are non-stationary and present in a multicomponent form in
447 the received signals, their separation (extraction) has been a challenging task. In this
448 paper, we have shown that the modes can be successfully extracted based on a multi-
449 variate decomposition technique that exploits the eigenanalysis of the autocorrelation
450 matrix of the received signal. This method, which utilizes concentration measures calcu-
451 lated based on time-frequency representations, separates the modes while completely
452 preserving their integrity, thus opening the possibility for their individual analysis. IF
453 estimations based on extracted components were highly accurate, even for a high level
454 of noise. Results indicate that the efficiency of the method is increased with the larger
455 number of sensors (channels). Our future work will be oriented towards the analysis of
456 the separated components. Instantaneous frequency estimation techniques developed
457 within the time-frequency signal analysis field can be applied directly on separated
458 modes, providing new insights and tools for the analysis of dispersive channels.
Version October 29, 2021 submitted to Mathematics 25 of 30
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
Figure 13. Extracted and original signal components of the non-stationary multicomponent
multichannel signal considered in Example 3. Panels (a)–(f) present extracted using the proposed
approach, whereas panels (g)–(l) show Wigner distributions calculated for individual components
of the original, noise-free signal.
459 Funding: This research was funded by COST action CA17137 - A network for Gravitational Waves,
460 Geophysics and Machine Learning.
461 References
462 1. B. Boashash, Time-Frequency Signal Analysis and Processing – A Comprehensive Reference, Else-
463 vier Science, Oxford, 2003.
464 2. P. Flandrin, Time-frequency/time-scale analysis, vol. 10. Academic press, 1998.
465 3. L. Stanković, M. Daković, T. Thayaparan,Time-Frequency Signal Analysis with Applications,
466 Artech House, Mar. 2013
467 4. A. Ouahabi (Ed.) Signal and image multiresolution analysis, John Wiley & Sons, 2012.
468 5. B. Boashash, “Estimating and interpreting the instantaneous frequency of a signal. I. Funda-
469 mentals,”Proceedings of the IEEE, vol. 80, no. 4, pp. 520-538, Apr 1992. doi: 10.1109/5.135376
470 6. S. Stanković, I. Orović, and E. Sejdić, Multimedia Signals and Systems: Basic and Advance
471 Algorithms for Signal Processing, Springer-Verlag, New York, 2015
472 7. A. Akan, O. K. Cura, “Time–frequency signal processing: Today and future, Digital Signal
473 Processing,”2021, 103216, https://doi.org/10.1016/j.dsp.2021.103216.
Version October 29, 2021 submitted to Mathematics 26 of 30
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
2 2 2
frequency
frequency
frequency
0 0 0
-2 -2 -2
Figure 14. Instantaneous frequency estimation for individual signal components based on: ex-
tracted signal components (dashed black) and original signal components (solid white). MSEs
between the two IF estimates is provided for each component of the signal from Example 3. The
noise variance is σε2 = 1. Decomposition is based on C = 256 channels.
Eigenvalues Wigner distribution of signal Spectrogram of signal
1.2 0 0
1 0.5 0.5
1 1
frequency
frequency
0.8
0.4 2 2
Figure 15. (a) Eigenvalues of autocorrelation matrix R and (b) time-frequency representation of
the considered acoustic signal from the dispersive environment. The number of sensors is C = 25,
whereas the number of modes is P = 5.
1 1 1
frequency
frequency
frequency
1.5 1.5 1.5
2 2 2
3 3 3
(a) -100 -50 0 50 100 (b) -100 -50 0 50 100 (c) -100 -50 0 50 100
time time time
Wigner distribution of eigenvector Wigner distribution of eigenvector
0 0
0.5 0.5
1 1
frequency
frequency
1.5 1.5
2 2
2.5 2.5
3 3
(d) -100 -50 0 50 100 (e) -100 -50 0 50 100
time time
495 15. J. Lerga, V. Sucic and B. Boashash, “An Efficient Algorithm for Instantaneous Frequency
496 Estimation of Nonstationary Multicomponent Signals in Low SNR,” EURASIP Journal on
497 Advances in Signal Processing, vol. 2011, pps. 16, 2011.
498 16. V. Katkovnik, L. Stanković, “Instantaneous frequency estimation using the Wigner distribu-
499 tion with varying and data driven window length,” IEEE Trans. on Signal Processing, Vol.46,
500 No.9, Sep.1998, pp.2315-2325.
501 17. V.N. Ivanović, M. Daković, L. Stanković, “Performance of Quadratic Time-Frequency Distri-
502 butions as Instantaneous Frequency Estimators,”IEEE Trans. on Signal Processing, Vol. 51, No.
503 1, Jan. 2003, pp.77-89
504 18. L. Stanković, “A measure of some time–frequency distributions concentration,” Signal Pro-
505 cessing, vol. 81, 2001, pp. 621–631.
506 19. L. Stanković, “A method for time-frequency signal analysis,”IEEE Transactions on Signal
507 Processing, Vol-42, No.1, Jan.1994.pp.225-229.
508 20. L. Stanković, T. Thayaparan, and M. Daković, “Signal Decomposition by Using the S-Method
509 with Application to the Analysis of HF Radar Signals in Sea-Clutter,”IEEE Transactions on
510 Signal Processing, Vol.54, No.11, Nov. 2006, pp.4332- 4342
511 21. L. Stanković, D. Mandic, M. Daković, and M. Brajović, “Time-frequency decomposition
512 of multivariate multicomponent signals,” Signal Processing, Volume 142, January 2018, pp.
513 468-479, http://dx.doi.org/10.1016/j.sigpro.2017.08.001
514 22. L. Stanković, M. Brajović, M. Daković, and D. Mandic, “Two-component Bivariate Signal
515 Decomposition Based on Time-Frequency Analysis,”22nd International Conference on Digital
516 Signal Processing IEEE DSP 2017, August 23-25, London, United Kingdom
517 23. M. Brajović, L. Stanković, M. Daković, and D. Mandic, “Additive Noise Influence on the
518 Bivariate Two-Component Signal Decomposition,” 7th Mediterranean Conference on Embedded
519 Computing, MECO 2018, Budva, Montenegro, June 2018.
520 24. G. Lopez-Risueno, J. Grajal and O. Yeste-Ojeda, “Atomic decomposition-based radar complex
521 signal interception,”IEE Proceedings - Radar, Sonar and Navigation, vol. 150, no. 4, pp. 323-31-,
522 1 Aug. 2003.
523 25. L. Stanković, M. Brajović, M. Daković, and D. Mandic, “On the Decomposition of Multichan-
524 nel Nonstationary Multicomponent Signals,”Signal Processing, Vol. 167, Feb. 2020, 107261,
525 DOI: https://doi.org/10.1016/j.sigpro.2019.107261
526 26. M. Brajović, I. Stanković, M. Daković, D. Mandic, and L. Stanković, “On the Number of
527 Channels in MulticomponentNonstationary Noisy Signal Decomposition,”25th International
528 Conference on Information Technology (IT 2021), Zabljak, Montenegro, 16 – 20 February, 2021
Version October 29, 2021 submitted to Mathematics 28 of 30
1 1 1
frequency
frequency
frequency
1.5 1.5 1.5
2 2 2
3 3 3
(a) -100 -50 0 50 100 (b) -100 -50 0 50 100 (c) -100 -50 0 50 100
time time time
Wigner distribution of component Wigner distribution of component
0 0
0.5 0.5
1 1
frequency
frequency
1.5 1.5
2 2
2.5 2.5
3 3
(d) -100 -50 0 50 100 (e) -100 -50 0 50 100
time time
Figure 17. Extracted modes of the signal received from the dispersive acoustic environment. The
decomposition is performed using the presented multivariate approach.
623 60. W. A. Kuperman and J. F. Lynch, “Shallow-water acoustics,” Physics Today, vol. 57, no. 10,
624 pp. 55–61, 2004.
625 61. M. Stojanovic, “Underwater acoustic communication,” Wiley Encyclopedia of Electrical and
626 Electronics Engineering, 2001.
627 62. M. Stojanovic and J. Preisig, “Underwater acoustic communication channels: Propaga- tion
628 models and statistical characterization,” IEEE Communications Magazine, vol. 47, no. 1, pp.
629 84–89, 2009.
630 63. C. Ioana, N. Josso, C. Gervaise, J. Mars, and Y. Stèphan, “Signal analysis approach for passive
631 tomography: applications for dispersive channels and moving configuration,” 2009.
632 64. E. de Sousa Costa, E. B. Medeiros, and J. B. C. Filardi, “Underwater acoustics modeling in
633 finite depth shallow waters,” in Modeling and measurement methods for acoustic waves
634 and for acoustic microdevices, IntechOpen, 2013.
635 65. G. V. Frisk, Ocean and seabed acoustics: a theory of wave propagation. Pearson Educa- tion,
636 1994.
637 66. J. Zhang and A. Papandreou-Suppappola, “Time-frequency based waveform and receiver de-
638 sign for shallow water communications,” in 2007 IEEE International Conference on Acoustics,
639 Speech and Signal Processing-ICASSP’07, vol. 3, pp. III–1149, IEEE, 2007.
640 67. Y. Jiang and A. Papandreou-Suppappola, “Discrete time-frequency characterizations of
641 dispersive linear time-varying systems,” IEEE Transactions on Signal Processing, vol. 55, no.
642 5, pp. 2066–2076, 2007.