NOISE-ASSISTED MULTIVARIATE VARIATIONAL MODE DECOMPOSITION
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
NOISE-ASSISTED MULTIVARIATE VARIATIONAL MODE DECOMPOSITION Charilaos A. Zisou? Georgios K. Apostolidis? Leontios J. Hadjileontiadis?† ? School of Electrical & Computer Eng., Aristotle University of Thessaloniki, GR-54124, Greece † Department of Electrical & Computer Eng., Khalifa University, PO Box 127788, UAE ABSTRACT amplitudes, i.e. narrow-band signals. The very first method proposed was the empirical mode decomposition (EMD) [2] The variational mode decomposition (VMD) is a widely which analyzes an input signal into a set of narrow-band applied optimization-based method, which analyzes non- components via an iterative sifting process which is driven by stationary signals concurrently. Correspondingly, its recently the signal’s extrema and does not require any a priori defined proposed multivariate extension, i.e., MVMD, has shown basis. More recently, a different data-driven approach that has great potentials in analyzing multichannel signals. However, drawn significant attention is the variational mode decompo- the requirement of presetting the number of extracted com- sition (VMD) which formulates an optimization problem and ponents K diminishes the analytic property of both VMD minimizes the collective bandwidth (CBW) of the signal’s and MVMD methods. This work combines MVMD with the components subject to the full reconstruction constraint. Sev- noise injection paradigm to propose an efficient alternative eral data-driven signal decomposition methods can be found for both VMD and MVMD, i.e., the noise-assisted MVMD in the bibliography, many of which have been proposed as (NA-MVMD), that aims at relaxing the requirement of pre- extensions of the original EMD algorithm [3, 4, 5, 6] as well setting K, as well as improving the quality of the resulting as alternative approaches, such as the iterative filtering [7], decomposition. The noise is injected by adding noise vari- the empirical wavelet transform [8], the synchrosqueezed ables/channels to the initial signal to excite the filter bank transform [9] and the swarm decomposition [10]. Typically, property of VMD/MVMD on white Gaussian noise. More- the aforementioned methods are designed to analyze univari- over, an alternative approach of updating center frequen- ate signals. Nevertheless, multivariate non-stationary signal cies is proposed, which uses the centroid of the generalized analysis using data-driven methods attracts great interest from cross–spectrum instead of a simple average of the individ- the scientific areas related to multichannel/multisensor signal ual spectral centroids, showing faster convergence. The processing. Although a multivariate signal can be processed NA–MVMD is applied to both univariate and multivariate channel-wise, the alignment of frequencies across channels, synthetic signals, showing improved analytical ability, noise i.e. the mode-alignment, is not guaranteed. Thus, multivari- intolerance, and less sensitivity in selecting the K parameter. ate extensions of the univariate data-driven decomposition Index Terms— multivariate analysis, variational mode methods have been proposed, e.g. the multivariate EMD decomposition, multichannel signals, non-stationary signals (MEMD) [11, 12] and the multivariate VMD (MVDM) [13], among others [14, 15, 16]. 1. INTRODUCTION This work focuses on improving the VMD and MVMD Over the last two decades, data-driven signal decomposition methods. The main concept of MVMD, similarly to VMD, methods have proven to be useful tools in non-stationary is to formulate an optimization problem to extract a collec- signal analysis and have been widely used in a plethora of tion of components that exhibit minimum CBW subject to problems. The objective of these methods is to output a set full reconstruction of the signal of every channel of the input of components whose Hilbert transform accurately represents multivariate data. At every optimization step an ensemble of the instantaneous dynamics, i.e. instantaneous amplitude and components for each channel k is resulted via applying band- frequency, of an input signal. It is well known that the Hilbert pass filtering around floating dominant frequencies ωk , which transform can disentangle the instantaneous amplitude and are updated online during the process. The floating dominant frequency, only in the case of non-overlapping frequency frequency for each component is calculated as the average of spectra (Bedrosian theorem) [1]. Thus all existing data-driven the spectral centroids across all channels. methods aim at extracting components with slow-varying The main issue of both VMD and MVMD methods is that they are highly parameter-sensitive. Specifically, the decom- Thanks to Projects: Care4MyHeart with ADEK Award Number AARE18-135 and PROTEIN Grant no.817732 within the H2020 Research position result is heavily influenced by the number of com- and Innovation Program for funding this work. (C. Zisou and G. Apostolidis ponents K that must be defined beforehand and arbitrary pre- contributed equally to this work.). Code: https://github.com/chariszisou setting of K may cause mode splitting. Thus, scholars have
approached this problem with a focus on estimating the op- Algorithm 1 MVMD timal parameter K [17, 18, 19, 20]. The aim of this work is procedure MVMD(x(t), K, α, τ , ) not to propose another parameter K estimation approach, but (1) (1) Initialize {ûk,c }, {ωk }, λ̂(1) , n ← 0 to make the original methods independent from K, under cer- repeat tain circumstances. An additional objective is to improve the n←n+1 original methods’ decomposition capability, i.e. to enhance for k = 1 : K do the resulting component orthogonality. To meet these objec- for c = 1 : C do Update component ûk,c : tives, we exploit the filter bank property of both VMD and (n+1) ûk,c (ω) ← MVMD on white Gaussian noise (WGN) [13, 21]. (n) The new method, the noise-assisted MVMD (NA–MVMD) P (n+1) P (n) λ̂c (ω) x̂c (ω) − ik ûi,c (ω) + combines the MVMD with the noise injection paradigm pro- 2 (n) posed in [6]. In short, NA–MVMD creates a noise injected 1 + 2α(ω − ωk )2 multivariate signal by adding new channels that consist of end for independent realizations of WGN to the input data channels end for and then applies the standard MVMD to the new multivari- for k = 1 : K do Update center frequency ωk : (n+1) ate signal. Finally, in the update of the floating dominant (n+1) 1 PR∞ |ûk,c (ω)|2 dω frequencies, the average spectral centroid across all chan- ωk ← ω P ∞ (n+1) dω C c 0 R |û (ω)|2 k,c nels is replaced by the spectral centroid of the generalized c 0 cross-spectrum [22]. To prove the potential of NA–MVMD, end for experiments are conducted using both univariate and multi- for c = 1 : C do Update λ̂c : variate synthetic signals. (n+1) (n) P (n+1) λ̂c (ω) ← λ̂c (ω) + τ (x̂c (ω) − ûk,c (ω)) k end for 2. MULTIVARIATE VMD P P kûk,c (n+1) (n) − ûk,c k22 until convergence: k c (n)
3. NOISE-ASSISTED MVMD number of extracted components as well as investigating the orthogonality condition. The NA–MVMD method is a refinement of MVMD. The new The synthetic signals generated for this experiment were elements that NA–MVMD includes are the adoption of the realizations of a stochastic class of non-stationary multi- noise injection paradigm proposed in [6] and the substitution component signals. This class of signals was defined as of the center frequency update calculation. The NA–MVMD XM approach can be applied to both univariate and multivariate s(n) = sm (n) + η(n), n ∈ Z, (10) m=1 signals. The noise injection aims at exciting the filter bank character of the original MVMD method to provide a distinct sm (n) = wLm (n − dm ) · Am · cos(ωm n), n ∈ Z, (11) processing behavior regardless of the input signal characteris- where M is the number of generated components, η(n) is the tics. The noise injection is performed by adding channels that added WGN, wLm [n − dm ] is a Hanning window of time lo- consist of independent realizations of WGN, cation dm and length Lm , Am is the amplitude and ωm is the frequency of mth component. For this experiment M , dm and x0 (t) = [x(t), η1 (t), · · · , ηN (t)], (6) Lm were fixed, i.e. M = 3, dm = {250, 125, 375}, Lm = where x0 (t) is the new multivariate signal, x(t) is the input {500, 125, 100}, whereas, Am and ωm were randomly drawn univariate or multivariate signal and ηi (t), ∀i = 1, · · · , N , from uniform distributions, i.e. P DF (Am ) = U (0.5, 1.5) are WGN signals. Then, x0 (t) replaces x(t) in Algorithm and P DF (ωm ) = U (0, π), ∀m. 1. Moreover, the center frequencies update part of MVMD In order to compare both methods in a generalized man- changes from the average of spectral centroids across all ner, multiple simulations were executed for a sufficient channels to the centroid of the generalized cross-spectrum amount of iterations, i.e. 500, for three different signal-to- (GCS) [22]. For a set of N signals {zi (t)}N noise ratios (SNR). Every iteration consisted of the following i=1 , the GCS is steps: 1) a multi-component signal (10) was generated, 2) γ(ω) = (1/(N − 1))(Λmax (Σz (ω)) − 1), (7) the signal was decomposed for different values of K, i.e. K = {3 : 10} using both VMD and NA–MVMD, 3) the Sz1 z1 (ω) . . . Sz1 zN (ω) decompositions were characterized as successful or not, and, Σz (ω) = .. .. .. (8) finally, 4) their orthogonality index (OI) was calculated. Suc- . . . SzN z1 (ω) . . . SzN zN (ω) cess was defined as a binary variable (1/0). A decomposition was classified as successful, if all input components (11) where Σz (ω) contains all the pairwise complex cross– appeared exactly once in the set of the output components. spectra, Szi,zj (ω) ∀i, j = 1, · · · , N , and Λmax (Σx (ω)) The absence or the splitting of a component was considered is the largest eigenvalue of the matrix Σz (ω). It is evident incorrect. This definition is very strict, e.g. if 2 out of 3 that for the case of N = 2 the GCS is reduced to the standard components are correctly extracted, the success is set to 0. At cross–spectrum. Now, let γk (ω) be the GCS of the set of each scenario, after all iterations, the success rate SR was cal- k th -level components {ûk,c (ω)}C c=1 at an arbitrary iteration culated as the number of successful decompositions divided of Algorithm 1 the update of center frequency ωk becomes by the total number of repetitions. Furthermore, the OI was Z ∞ calculated based on the definition in [2], where for an input γk (ω) ωk = ω R∞ dω. (9) signal s(n) and output components um (n), ∀m = 1, · · · , M , 0 0 γk (ω)dω X XM XM OI = ( ui (n) · uj (n)/s2 (n)) (12) The reason for this replacement is that the new approach is n i=1 j=1 a more elegant representation of the components’ collective Except for K, the rest of Algorithm 1 parameters are set as spectrum and it is a direct extension of the VMD method. α = 1000, τ = 10−4 and = 10−8 . Following (6), one assisting noise channel is added. Fig. 1 illustrates the analysis 4. EXPERIMENTAL RESULTS of an excerpt signal from the simulations, revealing the main shortcoming of VMD, i.e. the mode splitting. To evaluate its performance, NA-MVMD was applied, along Fig. 2 depicts the success of both the proposed and the with VMD and MVMD, to non-stationary univariate and mul- original method on decomposing the generated synthetic sig- tivariate synthetic signals. nals versus the parameter K. For high values of SNR, VMD shows a spike around K = 3, highlighting the necessity of 4.1. Dependence of the parameter K, influence of noise choosing the correct number of K a priori, since VMD fails and orthogonality property at other K. In contrary, the NA–MVMD method successfully The performance of the NA–MVMD method was tested on analyzes signals independently of K making the a priory se- synthetic univariate multi-component signals and compared lection of K less restrictive. Moreover, NA–MVMD behavior with the well-known VMD method. Particular emphasis is remains unchanged for different values of SNR showing good given on testing the influence of the parameter K, i.e. the noise intolerance.
2 SR of NA-MVMD vs. parameter K 1 s(n) 0 SNR=20dB SNR=15dB -2 0.8 0 100 200 300 400 500 SNR=10dB n (samples) NA-MVMD VMD 0.6 SR 1 1 0.4 u1 0 0 -1 -1 1 1 0.2 u2 0 0 -1 -1 0 1 1 2 3 4 5 6 7 8 9 10 u3 0 0 number of components (K) -1 -1 1 1 (a) u4 0 0 -1 -1 SR of VMD vs. parameter K 0 100 200 300 400 500 0 100 200 300 400 500 1 n (samples) n (samples) SNR=20dB SNR=15dB 0.8 SNR=10dB Fig. 1. An excerpt signal from the experiment with SN R = 0.6 20dB, ωm = {0.2π, 0.6π, 0.7π} and Am = {0.9, 1.2, 1.4} SR and its extracted components applying the NA–MVMD and 0.4 the VMD methods. This example shows the mode splitting 0.2 that occurs in VMD (see components u1 and u2 ) and, the re- 0 spective rectification using NA–MVMD. 2 3 4 5 6 7 8 9 10 number of components (K) Table 1 depicts the OI versus SNR for both NA–MVMD (b) and VMD methods for those K that each method has the best Fig. 2. The SR of NA–MVMD and VMD with respect to K. SR. It is evident that NA–MVMD performs better in terms of the orthogonality property. Table 2. Median and standard deviation of the total number Table 1. The OI×10−2 (mean and standard deviation) of NA- of steps until convergence for the GCS-based (proposed) and MVMD and VMD for different SNR, using the best perform- the average-based (original) center frequency update. ing K, i.e K=8 for the NA-MVMD and K=3 for the VMD. SNR (dB) 20 10 5 0 SNR (dB) 20 15 10 GCS 16 ± 6 54 ± 48 63 ± 61 79 ± 79 NA–MVMD 0.06 ± 0.03 0.07 ± 0.03 0.1 ± 0.03 Average 21 ± 2 73 ± 45 80 ± 60 114 ± 106 VMD 0.17 ± 0.13 0.13 ± 0.08 0.17 ± 0.08 4.2. Convergence Analysis 5. CONCLUDING REMARKS In this experiment, a comparative evaluation regarding the center frequency update (see Algorithm 1) was conducted. In this work, we presented the NA–MVMD method which The comparison involved two approaches, the averaging- augments the recently proposed MVMD method incorporat- based and the GCS-based, i.e. the original and the proposed ing the noise injection paradigm, i.e. adding new channels one, respectively. For this purpose a synthetic trivariate signal with WGN. The NA–MVMD fully exploits the filter bank was generated, x(n) = [x1 (n), x2 (n), x3 (n)] + η(n) with, property of MVMD on WGN. The new method allows the processing of both univariate and multivariate signals mak- x1 (n) = 0.5 sin(ω1 n) + 0.2 sin(ω2 n) + 0.1 sin(ω3 n), ing it a generalization of both VMD and MVMD. To evaluate x2 (n) = 0.45 cos(ω1 n) + 0.3 sin(ω3 n) + 0.2 sin(ω4 n), its performance, NA–MVMD was applied to synthetic multi- component signals showing better analytical ability, noise in- x3 (n) = 0.4 sin(ω1 n) + 0.25 sin(ω2 n), tolerance and less sensitivity in selecting the K parameter. where ω1 = 0.1π, ω2 = 0.2π, ω3 = 0.3π, ω4 = 0.8π, Moreover, an alternative center frequency update, i.e. a core η(n) is WGN and n ∈ Z. MVMD was applied using both calculation in MVMD algorithm, based on the GCS was pro- approaches, for different values of SNR. At each SNR level, posed allowing faster convergence. the process was repeated for 100 iterations and the number Nonetheless, the NA-MVMD should be subjected to of steps until convergence was monitored. The results are more investigation regarding 1) the better understanding of presented in Table 2 showing that the proposed approach pro- the noise injection parameters, i.e. number of noisy channels vides faster convergence especially in high noise levels. This and the noise amplitude, 2) the application to real-life univari- is explained by the fact that the averaging, as a smoothing op- ate and/or multivariate signals and 3) the combination of the eration, is in general less sensitive in cases of high variance. proposed paradigm with the other modifications/extensions of the original VMD and MVMD methods.
6. REFERENCES [12] Danilo P Mandic, Naveed ur Rehman, Zhaohua Wu, and Norden E Huang, “Empirical mode decomposition- [1] Edward Bedrosian, “A product theorem for hilbert trans- based time-frequency analysis of multivariate signals: forms,” Proc. IEEE, vol. 51, pp. 868–869, 1963. The power of adaptive data analysis,” IEEE signal pro- cessing magazine, vol. 30, no. 6, pp. 74–86, 2013. [2] Norden E Huang and et al., “The empirical mode de- composition and the hilbert spectrum for nonlinear and [13] Naveed ur Rehman and Hania Aftab, “Multivariate vari- non-stationary time series analysis,” Proceedings of the ational mode decomposition,” IEEE Transactions on Royal Society of London. Series A: mathematical, phys- Signal Processing, vol. 67, no. 23, pp. 6039–6052, 2019. ical and engineering sciences, vol. 454, no. 1971, pp. [14] Antonio Cicone, “Multivariate fast iterative filtering 903–995, 1998. for the decomposition of nonstationary signals,” arXiv [3] Zhaohua Wu and Norden E Huang, “Ensemble empiri- preprint arXiv:1902.04860, 2019. cal mode decomposition: a noise-assisted data analysis [15] Omkar Singh and Ramesh Kumar Sunkaria, “An empir- method,” Advances in adaptive data analysis, vol. 1, no. ical wavelet transform based approach for multivariate 01, pp. 1–41, 2009. data processing application to cardiovascular physiolog- ical signals,” Bio-Algorithms and Med-Systems, vol. 14, [4] Jia-Rong Yeh, Jiann-Shing Shieh, and Norden E Huang, no. 4, 2018. “Complementary ensemble empirical mode decompo- sition: A novel noise enhanced data analysis method,” [16] Alireza Ahrabian and et al., “Synchrosqueezing-based Advances in adaptive data analysis, vol. 2, no. 02, pp. time-frequency analysis of multivariate data,” Signal 135–156, 2010. Processing, vol. 106, pp. 331–341, 2015. [5] Marı́a E Torres and et al., “A complete ensemble empir- [17] Xian-Bo Wang, Zhi-Xin Yang, and Xiao-An Yan, ical mode decomposition with adaptive noise,” in 2011 “Novel particle swarm optimization-based variational IEEE international conference on acoustics, speech and mode decomposition method for the fault diagnosis of signal processing. IEEE, 2011, pp. 4144–4147. complex rotating machinery,” IEEE/ASME Transactions on Mechatronics, vol. 23, no. 1, pp. 68–79, 2017. [6] Naveed ur Rehman and et al., “Emd via memd: multi- [18] Qiming Chen, Xun Lang, Lei Xie, and Hongye Su, variate noise-aided computation of standard emd,” Ad- “Detecting nonlinear oscillations in process control loop vances in Adaptive Data Analysis, vol. 5, no. 02, pp. based on an improved vmd,” IEEE Access, vol. 7, pp. 1350007, 2013. 91446–91462, 2019. [7] Luan Lin, Yang Wang, and Haomin Zhou, “Iterative [19] Jijian Lian, Zhuo Liu, Haijun Wang, and Xiaofeng filtering as an alternative algorithm for empirical mode Dong, “Adaptive variational mode decomposition decomposition,” Advances in Adaptive Data Analysis, method for signal processing based on mode character- vol. 1, no. 04, pp. 543–560, 2009. istic,” Mechanical Systems and Signal Processing, vol. [8] Jerome Gilles, “Empirical wavelet transform,” IEEE 107, pp. 53–77, 2018. transactions on signal processing, vol. 61, no. 16, pp. [20] Yujie Zhao and et al., “A modified variational mode 3999–4010, 2013. decomposition method based on envelope nesting and multi-criteria evaluation,” Journal of Sound and Vibra- [9] Ingrid Daubechies, Jianfeng Lu, and Hau-Tieng Wu, tion, vol. 468, pp. 115099, 2020. “Synchrosqueezed wavelet transforms: An empirical mode decomposition-like tool,” Applied and computa- [21] Konstantin Dragomiretskiy and Dominique Zosso, tional harmonic analysis, vol. 30, no. 2, pp. 243–261, “Variational mode decomposition,” IEEE transactions 2011. on signal processing, vol. 62, no. 3, pp. 531–544, 2013. [10] Georgios Apostolidis and Leontios Hadjileontiadis, [22] David Ramirez, Javier Via, and Ignacio Santamaria, “Swarm decomposition: A novel signal analysis using “A generalization of the magnitude squared coherence swarm intelligence,” Signal Processing, vol. 132, pp. spectrum for more than two signals: definition, proper- 40–50, 2017. ties and estimation,” in 2008 IEEE International Con- ference on Acoustics, Speech and Signal Processing. [11] Naveed Rehman and Danilo P Mandic, “Multivariate IEEE, 2008, pp. 3769–3772. empirical mode decomposition,” Proceedings of the [23] Dimitri P Bertsekas, Constrained optimization and La- Royal Society A: Mathematical, Physical and Engineer- grange multiplier methods, Academic press, 2014. ing Sciences, vol. 466, no. 2117, pp. 1291–1302, 2010.
You can also read