Application of Wavelet Transform for Determining Diagnostic Signs Volodymyr Eremenko1[0000-0002-4330-7518], Artur Zaporozhets2[0000-0002-0704-4116], Volodymyr Isaenko3[0000-0001-8010-8844], Kateryna Babikova3[0000-0002-5053-1999] 1 National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnical Institute”, Kyiv, Ukraine 2 Institute of Engineering Thermophysics of NAS of Ukraine, Kyiv, Ukraine 3 National Aviation University, Kyiv, Ukraine a.o.zaporozhets@nas.gov.ua Abstract. It is proposed to apply the wavelet transform to localize in time the frequency components of the information signals in this article. The wavelet transform allows to fulfil time-frequency analysis of signals, which is very im- portant for studying the structure of a composite material from the mode compo- sition of free oscillations. The proposed approach to the development of infor- mation signals using wavelet transform makes it possible to further study the na- ture of the occurrence of free oscillations and the propagation of acoustic waves in individual layers of composites and to study the change in the structure of composites from the changes in the three-dimensional wavelet spectrum. Keywords: wavelet transform, MHAT wavelet, Morlet wavelet, diagnostic signs, information signal, composite material, free oscillation method 1 Introduction Information signals obtained in the process of diagnosing composite materials by low- frequency acoustic methods belong to the class of single-pulse signals with locally con- centrated features. These are, for example, signals of free oscillations, whose modes have not only frequency, but also temporal distribution, signals of impulse impedance method change frequency and current carrier phase for one radio pulse, signals of low- speed impact method locally change their shape depending on material defectiveness. For such signals, the task of identifying diagnostic signs is significantly more diffi- cult than for signals in which the information component is evenly distributed over the observation interval [1, 2]. This is explained by the fact that diagnostic signs are focused on small time intervals or fragments of signal realization, and the signal itself has a rather complex form that cannot be described by a formal constructive model. The most common methods for isolating diagnostic features of such signals are [3, 4, 5, 6]:  methods for evaluating the integral characteristics ‒ the center of mass of the pulse, the similarity coefficients, etc. These methods have high noise immunity, but have very little sensitivity to local changes in signal parameters;  methods of decomposition of signals with an orthogonal basis. In general, they pro- vide information about the shape of the pulsed signal, but provide only an integral representation of its components throughout the entire domain of definition and are not sensitive to local variations in characteristics;  methods of structural analysis using signal segmentation as a sequence of separate fragments. Based on segmentation signal clustering is performed, chains of clusters are used to structure the signal;  methods for representing signals in phase space ‒space, which is determined by a finite set of state parameters. The disadvantage is the need for multiple repetition of the pulse signal and the analysis of multidimensional data arrays;  heuristic methods, in particular, methods using neural network technologies, which allow to select informative fragments of signals and to make a comparison with the "reference" ones. Classical methods for processing signals of low-frequency acoustic diagnostic meth- ods, in particular, the method of decomposing signals with an orthogonal basis (gener- alized Fourier transform), which give an integral representation of the signal compo- nents in the entire domain of their definition, are ineffective [7]. Therefore it was pro- posed to apply the wavelet transform for localizing the frequency components of infor- mation signals in time [8]. The signals of free oscillations are damped and consist of several modes, then each mode, depending on which layer of the composite it is excited, will have its own am- plitude, frequency and attenuation coefficient. That is, knowing the time behavior of each mode, one can draw conclusions about the structure of the controlled zone of the composite [9]. The continuous wavelet transform allows to carry out a time-frequency analysis [10]. These studies can be used in many areas, including the monitoring of complex technical systems [11, 12]. 2 Main Part The wavelet transform combines two types of transformations ‒ direct and inverse, which, respectively, translate the studied function f(t) into a set of W (a, b) f wavelet coefficients and vice versa [13]. The direct wavelet transform is performed according to the rule:  1 1  x b  W (a, b) f  C  a *  a   f (t )dt , (1)   where a and b are the parameters that determine, respectively, the scale and offset of the function ψ, which is called the analyzing wavelet; Cψ is normalization coefficient. The basic, or maternal, wavelet ψ forms with the help of stretch marks and landslides a family of functions ψ (t–b/a). Having a known set of coefficients Wψ(a, b)f, we can restore the original form of the function f(t):   1 1  t b  da  db f (t )  C  a   a   W (a, b) f  a2 . (2)    The direct (1) and the inverse (2) transforms depend on some function  (t )  L2 ( R ) which is called the basic wavelet. In practice, the only restriction on its choice is the condition for the finiteness of the normalizing coefficient [14]: ˆ ( ) ˆ ( )   2 C   d   2 d   , (3)   0  where ˆ ( ) is Fourier image of the  ( ) wavelet:  1 ˆ ( )    (t )e  it dt. (4) 2  This condition satisfies many functions, so it is possible to choose the type of wavelet that is most suitable for solving a specific problem. In particular, for analyzing damped harmonic oscillations, it is more expedient to select wavelets, which are also damped oscillations. The article deals with the MHAT-wavelet and Morlet wavelet. The condition (3) means that the Fourier transform of the wavelet is zero at zero frequency, i.e. ˆ ( )   0  0 . In another case the denominator of the fraction in the in- tegral (3) is equal to zero, while the numerator has a nonzero value, and the Сψ coeffi- cient ceases to be finite. In turn, this requirement can be presented in another form. Since the Fourier trans-  formˆ ( ) at zero frequency has the form   (t )dt , it can write the following:   1 ˆ ( )    (t )e  it dt. (5) 2  A characteristic feature of the analyzing wavelets is time-frequency localization. This means that wavelets ψ(t) and their Fourier transforms ˆ   differ significantly from zero only at small time intervals and frequencies, and differ very little from zero (or simply equal to zero) outside these intervals. The quantitative measure of the localization of a function  (t )  L2 (R ) can be its center t and radius  t2 :  1 t   t  (t ) dt, 2 (6)  2   1  t  t   (t ) dt. 2 t2  2 (7)  2  In this case, the effective wavelet width is assumed to be 2 t . Nowadays a large number of basic wavelet functions are known [15, 16]. As men- tioned above, the members of any family of wavelets must satisfy condition (7) one of such families, the Gauss and wavelets. The functions of this family are derived from the Gaussian exponent: d n t 2 2 g n (t )  (1) n 1 e , n  N. (8) dt t The normalization coefficient takes the value Cgn  2 (n  1)! , 0  n   . (9) The most widely used Gaussian wavelets of small orders. The properties of Gaussian wavelets are discussed in detail in [17]. 2.1 MHAT Wavelet This wavelet is a second-order wavelet of the family of Gaussian wavelets (Fig. 1) and is formed by a double differentiation of the Gauss function: g2 (t )  (1  t 2 )et 2 . 2 (10) (а) (b) Fig. 1. Second order Gaussian wavelet (MHAT) (a) and its Fourier transform (b) The Fourier transform of this wavelet is ˆ    2 2 e 2 2 . (11) The graph of this function is shown in Fig. 1 (b). MHAT wavelets are well localized in both the time and frequency domains. The centers and localization radii in both areas have the following meanings; t  0 ; t  1.08 ,   1.51 ,   0.49 . 2.2 Morlet Wavelet The analytical representation of the Morlet wavelet and its Fourier transform is given by the following expressions:   t   et  eik t  e k  4  , 2 2 2 2 0 0   (12) ˆ      e  k   4  e  k   4  . 2 2 2 2 0 0 (13)   The Morlet wavelet is a plane wave modulated by a Gausian. The parameter α specifies the width of the Gausian, the parameter k0 is the frequency of the plane wave. Usually it choose  2  2 and k0  2 . With these values with sufficient accuracy it can be taken [18]:   t   e  t  e i 2 t , 2 2 (14) ˆ      e  2   4 . 2 2 (15) Graphs of these functions are shown in Fig. 2. (a) (b) Fig. 2. The real part of the Morlet wavelet (a) and its Fourier transform (b) The center and localization radius of the Morlet wavelet in the time domain are deter- mined by the corresponding values; t  0 , t   2 . In the Morlet wavelet, only zero moment is equal zero. The definitions of the integral wavelet transform introduced above cannot be used in practice, since in digital processing of results the main transformation objects are not functions defined on the entire time axis, but discrete signals whose length are always finite [19]. For this reason, instead of the above theoretical concepts, their practical counterparts (assessments) should be introduced. We assume that the signal is given by the function values with a constant step Δt: f k  f  tk  , tk  t  k , k  0, N  1 . (16) To estimate the wavelet transform of this sequence, we use the following expression: 1 N 1  t b  W ( a, b)   n( a, b) k  0 f k *  k  a  , (17) 2 N 1  1  tk  b  n( a, b)   e B a  . (18) k 0 where B=2 for MHAT wavelet and B=α2 for Morlet wavelet In the transition from (1) to (17), the multiplier a from the denominator of the formula (1) is replaced as follows   t  b 2  e  a2 B dt  a B , (19) the discrete approximation of which is function (18). This made it possible to eliminate the dependence of the amplitudes of the harmonic components on the parameter a, which usually makes it difficult to correctly estimate their relative intensities from the graphic image of the wavelet spectra. In addition, the function n(a,b) as an approximation allows to “equalize” for different values of the scale factor a the number of samples of the original function involved in the calculation [20]. Calculating the wavelet transform in the scale-offset coordinates is somewhat incon- venient for perception, since the scale ai specified with constant pitch compresses the high-frequency region and the components of the signal under study that belong to this region become difficult to distinguish. Therefore, it is proposed by replacing  i  1 ai to switch to νi value that is analogous to the frequency in the Fourier transform [21]. Then a pair of wavelet transforms of the function f(t) with (17) will look like this: 1  *  (t  b)  f (t )dt , n( , b)  W ( , b) f  (20)  B   (t  b)  W ( , b) f  d  db , C   f (t )  (21)   t  b 2  2  where n( , b)   e B dt .  The amplitude value of the discrete signal wavelet function will be calculated by the following equations: N 1 f k *  (tk  b j ) i  , 1 W ( i , b j )   n( i , b j ) k  0 (22)   N 1  1 ( t  b ) 2 n( i , b j )   e B k j i . (23) k 0 2.3 Discretization of Arguments Each wavelet has its own shape and characteristic size, which for a fixed value of the scale factor is determined by the value da  2t a , (24) where Δt is the wavelet radius. The function W(a,b) (22) determines the correlation between the analyzing wavelet located at a point b and at a certain part of a signal of da length with a center at a point b. The module of this function takes the greatest value in the case when the size of the wavelet coincides with the size of the "current" signal detail. In the case of polyhar- monic functions, the natural measure of the scale of its details is the period of the har- monic components, while the measure of the wavelet length da is determined by the value a of the scale factor. For a polyharmonic function defined on a grid with a step Δt=const, the range of periods of harmonics is determined by the quantities Pmin  2t , Pmax  ( N  1)t . In accordance with this, the largest and smallest values of the scale factor are selected from the condition of matching the size of the wavelet and the limiting periods of har- monious components 2t amin  Pmin , 2t amax  Pmax , (25) where we get amin  t t , amax  ( N  1)t 2t . (26) Note that these values are taken in cases where it is necessary to perform a signal analysis in the full scale range. Often, however, it is advisable to examine the signal in a narrower range of scales. In this case, the value amin and amax choose from other con- siderations. a  amin We propose a discretization step a  max of the scales, after which we de- Na  1 fine the discrete values of the scale factors ai  amin  a  i . Since the width of the spectral line increases with increasing scale, sometimes the value of the parameter is presented on a logarithmic scale. If the calculation of the wavelet transform is carried out according to (22), (23), then the minimum and maximum value of the quantity ν can be calculated using equations   min (25). Values νi are calculated with a constant step   max by the formula Na  1  i   min    i . In the simplest case, the boundaries of the landslide range are defined as follows bmin  0 , bmax  ( N  1)t , and discrete values of the displacements can be calculated by the following formulas: b j  bmin  b  j , (27) bmax  bmin b  . (28) Nb  1 With this method of discretization of the parameter b near the boundaries bmin and bmax of magnitude W(ai,bj) will be calculated with errors, since it is impossible to use the entire length of the analyzing wavelet near the boundaries. To exclude marginal effects, it is necessary to calculate the wavelet transform only for landslide values that are remote from the boundaries by an amount equal to the current radius of the wavelet aΔt. With this approach, formula (26) is transformed as follows b j  bmin  b  j , j  J a*  1,..., Nb  J a*  1 , (29) where J a* is the radius of the wavelet, expressed in units Δb, which corresponds to the current scale value a (for the Morlet wavelet)  a  J a*   i  (30)  2b  . In the case of application instead of the scale of the formula we get:    J*    (31)  2b i  . In these formulas, the rounding operation to the nearest integer is indicated by square brackets. The set of nodes of the discrete grid, which is defined by formulas (25) and (29), is called the probability triangle. Note that very often the edge effects are ignored, and the results of the wavelet analysis are simply represented in the rectangular area of the nodes (25) and (27). 2.4 Construction of a Plurality of Diagnostic Features The selection of each individual mode of oscillation is an important step in the study of the properties and nature of the destruction of composite materials. Attenuation of the components of free oscillations carries information about the quality factor of the con- trolled zone of the composite. This allows to investigate defects that are not associated with delamination, such as fatigue and impact damage to the surface. The proposed approach with the use of wavelet transform makes it possible to further study the nature of the propagation of acoustic waves in individual layers of composites and to investigate the change in the structure of composites by the revealed changes in the wavelet spectrum of free vibrations. Next we consider the wavelet transform of signals received in the intact and damaged area of a cellular panel with a thickness of 20 mm. The amplitude spectra of these sig- nals are shown in Fig. 3, 4. Fig. 3. Amplitude spectrum of the signal of free oscillations of the intact area of the cellular panel Fig. 4. Amplitude spectrum of a signal of free oscillations of a zone with a defect of 20 mm radius of a cellular panel The estimated amplitude spectra preliminarily determine the frequency range within which the wavelet transform will be performed — zone I in Fig. 3, 4. Fig. 5 shows plots of the amplitude wavelet spectra of these signals, calculated by a Morlet wavelet in the selected frequency range. (a) (b) Fig. 5. Graphs of amplitude wavelet functions of signals of free oscillations: a ‒ intact zone, b ‒ zone with damage of 20 mm radius From these figures it can be seen that, for example, the 1 st, 2nd and 3rd modes of oscil- lations change the nature of the attenuation with the appearance of a defect. In fig. 6 shows the restored third mode oscillations of the benign zone and the damaged zone. Its attenuation coefficient changes with the appearance of a defect from 4.72 to 2.56. For faster decision making on the presence or absence of a defect in the controlled area of a composite material, it is proposed to compare the amplitude wavelet spectra of free oscillations of the reference and controlled areas, calculated with the same shift. For example, in fig. 7(a) amplitude wavelet spectra of a signal of free oscillations of a benign zone with shifts b j  b1, j  30, b j  b 2, j  50, b j  b3, j  70 are presented. Fig. 7(b) shows similar spectra of free oscillations of a zone with a defect diameter of 20 mm. In other words, these spectra are actually a cross section of the graphs in Fig. 5. Fig. 6. The third damping component of free vibrations of the benign zone (S1) and the damaged zone (S2) (a) (b) Fig. 7. Plots of amplitude wavelet spectra of free vibrations of the intact zone (a) and zones with a separation of 20 mm (b) at displacements b1