Self-adaptive multivariate optimization underwater acoustic signal denoising method

Abstract

The invention relates to the technical field of underwater acoustic signal processing, in particular to a self-adaptive multivariate optimization underwater acoustic signal denoising method for a complex ocean environment, which can keep the integrity of a target signal while finely removing complex ocean noise. The method comprises the following steps: step 1, signal decomposition and parameter optimization based on BVMD; step 2, carrying out normalized KNN-MI preliminary screening; step 3, carrying out RCMFDE secondary screening; and 4, carrying out BWSTD noise reduction and signal reconstruction. According to the method, the complex ocean background noise can be effectively removed through fine screening of the noise and the target signal IMFs, high signal-to-noise ratio adaptability is achieved, and signal distortion is avoided through sufficient reservation of the target signal. The method is especially suitable for signal de-noising processing in underwater target detection, marine environment monitoring and other applications.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater acoustic signal processing, and particularly relates to an adaptive multi - element optimized underwater acoustic signal denoising method for complex marine environments. The method aims to achieve fine suppression of complex marine background noise interference through multi - level parameter optimization, signal decomposition, and noise suppression techniques, while ensuring the integrity of the target echo signal, providing reliable signal processing support for applications such as underwater target detection and marine environment monitoring. Background Technique

[0002] Underwater acoustic signal denoising plays a key role in underwater acoustic research. With the rapid development of fields such as marine environment monitoring, deep - sea exploration, robot navigation, and military reconnaissance, effectively removing noise from complex underwater acoustic signals and restoring effective signals has become an important research direction in the field of acoustic signal processing. In the development process, underwater acoustic signal denoising methods have evolved continuously, experiencing a process from traditional denoising methods, to empirical mode decomposition (EMD) and variational mode decomposition (VMD) methods, then to multi - method fusion, and continuously optimizing by combining emerging artificial intelligence methods.

[0003] In the early stage, underwater acoustic signal denoising mainly relied on traditional signal processing techniques such as filtering methods and wavelet transform. Low - pass filtering can remove high - frequency noise when the main frequency of the signal is lower than the noise frequency, but its effect is limited in complex noise environments and is prone to signal distortion. Wavelet transform can analyze signals at multiple scales, making the denoising process more refined. However, improper threshold setting will cause artifacts and distortion problems, and it is difficult to adaptively select basis functions and decomposition layers. With the increase in the complexity of underwater acoustic signals and the types of noise, the limitations of traditional denoising methods have become increasingly prominent, making it difficult to effectively separate noise from signals and even resulting in signal distortion or loss of details.

[0004] Facing the non - stationary and non - linear characteristics of underwater acoustic signals, traditional denoising methods are gradually unable to meet the requirements, and new signal decomposition methods have emerged. Empirical mode decomposition (EMD) can adaptively decompose complex signals into a series of intrinsic mode functions (IMFs), effectively dealing with non - stationary signals, but there are problems of mode mixing and end - point effects. To solve these problems, many improved algorithms have been proposed by scholars. Variational mode decomposition (VMD) decomposes signals into multiple IMF components with quasi - orthogonality to achieve effective signal separation. However, the number of IMFs K and the penalty factor α need to be preset before decomposition. If the two are set improperly, it will affect the decomposition efficiency. VMD decomposes signals based on variational theory, overcoming some of the disadvantages of EMD and having advantages in non - stationary signal processing. Therefore, it has been widely used in the research of underwater acoustic signal denoising. Numerous studies have combined various intelligent algorithms with VMD, such as grey wolf algorithm, improved whale algorithm, black widow optimization algorithm (BWOA), etc., to optimize VMD parameters and improve the denoising effect.

[0005] With the development of wavelet transform and VMD methods, researchers have attempted to fuse multiple denoising methods to further improve the denoising effect. The improved wavelet threshold method reduces signal distortion by optimizing the threshold function; the combination of methods such as multi-scale entropy analysis and mutual information analysis improves the separation accuracy of noise and effective signals; the method of fusing and optimizing VMD and improved wavelet threshold denoising optimizes VMD with the help of a triangular topology optimization algorithm and processes IMF components through a specific screening method. Experiments have shown that its denoising effect is better than that of traditional filtering algorithms; there are also studies that have proposed a hybrid feature extraction model based on multiple algorithms, which has shown effectiveness and superiority in the feature extraction and classification of underwater acoustic signals.

[0006] In summary, the underwater acoustic signal denoising technology is constantly evolving, and continuous breakthroughs have been made in terms of denoising effect, computational efficiency, and application scope. However, in the actual marine environment, noise interference is complex and diverse. In addition to ocean background noise, there are also platform mechanical noise, biological noise, and multi-path effects. Therefore, it is still necessary to further explore more efficient denoising methods. This paper fuses multiple improved denoising methods and proposes an adaptive multi-optimization underwater acoustic signal denoising method, aiming to remove complex and diverse noise interferences and obtain relatively pure target underwater acoustic signals. Summary of the Invention

[0007] The purpose of the present invention is to provide an adaptive multi-optimization underwater acoustic signal denoising method to solve the problem of severe noise interference in complex marine environments. Its process framework is as Figure 1 shown. Underwater acoustic signals contain various components with different frequencies, and it is difficult to directly distinguish useful signals from noise. By signal decomposition, complex signals can be split into relatively simple intrinsic mode functions (IMFs), which facilitates subsequent processing. Among the decomposed IMFs, some contain the target signal, while others are noise components. Through screening and classification, the target signal components can be identified and the pure noise components can be removed. Although screened, there is still residual noise in the noise-dominated components. We need to further denoise and finally reconstruct the complete and high-quality target signal.

[0008] The present invention uses the boundary optimization variational mode decomposition (BVMD) method to decompose the original signal into different intrinsic mode functions (IMFs), and combines the black widow optimization algorithm (BWOA) to adaptively optimize the parameters of the variational mode decomposition (VMD), so as to improve the accuracy and effectiveness of the decomposition. Then, the obtained intrinsic mode function (IMFs) components are initially screened by the normalized K-nearest neighbor mutual information (KNN-MI) to distinguish signal components and noise components. Then, the IMFs classified as noise are secondarily screened by the refined composite multi-scale fluctuation dispersion entropy (RCMFDE) to distinguish noise-dominated components and pure noise, and the pure noise is removed, and only the noise-dominated component part is retained. Finally, the noise-dominated IMFs are denoised by Bayesian wavelet soft thresholding (BWSTD). After removing the residual noise, they are reconstructed with the signal components obtained by the initial screening to obtain the final denoised signal, so as to ensure the integrity of the signal and achieve an ideal denoising effect. The specific steps are as follows:

[0009] Step 1: Signal decomposition and parameter optimization based on BVMD

[0010] The signals collected underwater contain various noises and interferences, and the components are complex. Directly using them for target detection and tracking is inefficient and inaccurate. Decomposing the signal into many intrinsic mode functions (IMFs) is because signals with different frequency components are intertwined in the original mixed state, making it difficult to distinguish useful information and noise. Through decomposition, complex signals can be split into relatively simple IMFs with specific frequency characteristics, making the characteristics of the signal clearer and more definite. Among these IMFs, some mainly contain target signal components, while others are mainly noise. Subsequently, these IMFs can be screened and classified to accurately identify the target signal components and avoid mistakenly removing useful signals as noise during the processing.

[0011] Signal decomposition and parameter optimization based on the boundary optimization variational mode decomposition (BVMD) is to use the black widow optimization algorithm (BWOA) to optimize the parameters of the variational mode decomposition (VMD), making the decomposition process more accurate and effective. It can adaptively determine the decomposition parameters, improve the accuracy and effectiveness of signal decomposition, and make the obtained IMFs better reflect the true characteristics of the signal.

[0012] Select the sample entropy (SampleEntropy, SampEn) as the fitness function to evaluate the quality of VMD decomposition. For a one-dimensional time series of length N

[0013]

[0014] The definition of sample entropy measures its complexity and uncertainty by calculating the frequency of similar patterns in the time series. Then, SampEn is defined as

[0015]

[0016] Among them, represents the distance metric between positions m and n in the time series, where x t is the time series, and m and n are the indices of the time series.

[0017] Next, calculate the matching ratio of the embedding vectors with length in the time series

[0018]

[0019] Among them, is the matching ratio between all embedding vectors under the given threshold r. d is the distance between two sequences, r is a set distance threshold, N is the sequence length, is the embedding dimension. Further, calculate the ratio at different scales

[0020]

[0021] Finally, obtain the expression of sample entropy

[0022]

[0023] Sample entropy reflects the complexity of the time series at different scales. The smaller its value, the stronger the regularity of the sequence, and the larger the value, the higher the complexity of the sequence. In BVMD, sample entropy is used as the fitness function to evaluate the effect of each VMD decomposition, and the goal is to minimize the sample entropy to obtain better signal decomposition.

[0024] The sub-steps of BVMD are as follows:

[0025] (1) Initialize the parameters of BWOA, set the population size as N pop , the upper limit of the number of iterations MaxIter, the value range of the number of modes K, and the range of the penalty factor α.

[0026] (2) Perform VMD decomposition on the signal to obtain the preliminary IMFs.

[0027] (3) Obtain the fitness by calculating the sample entropy SampEn.

[0028] (4) According to the update rules of the black widow optimization algorithm, update the positions of the spider population and optimize the parameters K and a.

[0029] (5) Continuously repeat steps 2, 3, and 4 until the set maximum number of iterations is reached.

[0030] (6)Finally, a set of optimal VMD parameters (K, a) is obtained.

[0031] (7)Update the VMD parameters K and a, and re - decompose the signal by VMD.

[0032] (8)Based on the optimized parameters, perform the VMD decomposition of the signal again.

[0033] (9)Finally, K IMFs are obtained as the output of the algorithm.

[0034] Step 2: Preliminary screening by normalized KNN - MI

[0035] The obtained K IMFs contain target and noise signals. Use normalized KNN - MI to preliminarily screen and classify them into signal components and noise components. Calculate the average value of the normalized KNN - MI values of the K IMFs and set it as the threshold. If the normalized KNN - MI value of a certain IMF component is greater than then it is determined as a signal - dominant IMF component; if it is less than then it is determined as a noise IMF component.

[0036] For each IMF component X obtained by decomposition, calculate its k - nearest neighbor mutual information with the original signal f(t) According to the k - nearest neighbor mutual information estimation method

[0037]

[0038] where N represents the total number of samples, k is the number of the nearest neighbors, n x and n y represent the number of the nearest neighbors of the random variables X and Y respectively, and ψ(N) is the digamma function used to adjust the number of the nearest neighbors.

[0039] Normalize the calculated mutual information

[0040]

[0041] and strictly limit it within the range [0, 1]. Here, H(X) is the entropy of the random variable X

[0042]

[0043] where p(x) is the probability of X, and x is the possible value of X.

[0044] Step 3: Secondary screening by RCMFDE

[0045] The noise components after the initial screening of normalized KNN-MI will be further screened by RCMFDE, which are divided into noise-dominated components and pure noise. Calculate the RCMFDE values of all noise IMF components and find their average value. As the threshold, the IMF components greater than will be removed as pure noise, and those less than will enter the subsequent processing as noise-dominated IMF components.

[0046] For the noise IMF components, calculate their refined composite multiscale fluctuation dispersion entropy (RCMFDE). Let the noise IMF component be u = {u1, u2, …, u L}}. For a given scale factor τ, calculate the k-th coarse-grained sequence of the time series u

[0047]

[0048] Calculate the probability mean of the β-th dispersion pattern in the coarse-grained sequence

[0049]

[0050] where is the dispersion pattern corresponding to the k-th coarse-grained sequence under the scale factor τ probability.

[0051] Define the Shannon entropy of the average of such sequence dispersion pattern probabilities as the RCMFDE value E RCMFD (x, m, c, d, τ)

[0052]

[0053] where the embedding dimension n is usually chosen as 2 or 3, the time delay d = 1, and the number of classes c takes an integer between [3, 8]. Step 4: BWSTD denoising and signal reconstruction

[0054] Perform further BWSTD denoising on the noise-dominated components obtained after the second screening by RCMFDE, and reconstruct the signal with the denoised signal and the signal components after the initial screening of normalized KNN-MI to obtain the final complete denoised signal.

[0055] First, use wavelet transform to convert the screened noise-dominated signal into the wavelet domain to obtain the wavelet coefficients c j,n at various scales j and frequency bands n. This step can be expressed as: x(t) → {c j,n}, with the aim of decomposing the signal at different frequencies and time scales, just like decomposing a complex mixture into different components, so as to analyze and process the wavelet coefficients of each component subsequently.

[0056] Calculate the standard deviation σ of the signal. Calculate according to the statistical characteristics of the wavelet coefficients.

[0057]

[0058] Here, x i can be analogized to the wavelet coefficient c j,n , is their mean value, and N is the number of samples. Calculating the standard deviation is to understand the degree of dispersion of the signal in the wavelet domain and provide an important reference index for subsequent determination of the threshold using the Bayesian method.

[0059] Calculate the posterior probability distribution P(c j,n |x) of the wavelet coefficients according to the Bayesian equation

[0060]

[0061] where A represents a certain state of the wavelet coefficient (such as whether it belongs to noise or signal), and B represents the observed value of the signal (i.e., the current wavelet coefficient c j,n ). P(B|A) is the likelihood function, representing the probability of observing B under the given state A; P(A) is the prior probability of A, reflecting the guess or known information about the state A of the wavelet coefficient before observing the signal x; P(B) is the probability of observing B (which can be estimated through the statistical characteristics of the signal and the distribution of the wavelet coefficients). Through the calculation of this Bayesian equation, the prior information of the signal and the current wavelet coefficient situation can be comprehensively considered to obtain a more accurate description of the probability distribution of the wavelet coefficients.

[0062] Determine the average value of the posterior probability distribution P(c j,n |x) as the wavelet threshold λ

[0063]

[0064] where M is the number of samples participating in the calculation of the average value (here it may be all the wavelet coefficients or a part of the representative coefficients). Selecting the average value as the threshold is a reasonable strategy based on statistics, which can determine a relatively appropriate segmentation point according to the overall situation of the posterior probability distribution and divide the wavelet coefficients into two parts that may be signal and may be noise.

[0065] For the wavelet coefficient c j,n , use the soft threshold function for processing

[0066]

[0067] When |c j,n | ≥ λ, the wavelet coefficients are shrunk by sgn(c j,n )(|c j,n | - λ), where sgn(c j,n ) is the sign function, which takes 1 or -1 according to the sign of c j,n . In this way, the influence of noise can be reduced while the main features of the signal are retained. When |c j,n | < λ, the wavelet coefficients are directly set to 0 to further remove possible noise components.

[0068] Inverse transform is performed on the processed wavelet coefficients c' j,n to obtain the denoised signal That is This step is to convert the signal processed in the wavelet domain back to the original time domain, enabling us to obtain the final denoised result for use in subsequent signal analysis and applications. Finally, the noise-dominated IMF components and signal-dominated IMF components after BWSTD denoising are reconstructed to obtain the final denoised signal.

[0069] Compared with the prior art, the present invention not only improves the denoising effect of underwater acoustic signals, but also has the following advantages:

[0070] (1) It is applicable to complex underwater acoustic environments, can effectively remove interferences such as ocean noise and ship noise, and improve signal quality;

[0071] (2) It is applicable to signal denoising under different signal-to-noise ratio conditions, and can maintain a good denoising effect both in high signal-to-noise ratio and low signal-to-noise ratio environments;

[0072] (3) It is applicable to maintaining signal details and integrity. Through optimized signal decomposition and denoising strategies, signal features are retained while denoising, avoiding the problem of signal distortion in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 Flow framework diagram of the efficient adaptive multi - optimization signal denoising method

[0074] Figure 2 Waveform diagram of the noisy signal

[0075] Figure 3 Iteration curve of the BWOA - optimized VMD parameters

[0076] Figure 4 Waveform diagrams after BVMD decomposition (a) Time domain diagram (b) Frequency spectrum diagram

[0077] Figure 5KNN–MI value and its normalization graph

[0078] Figure 6 Time-domain waveforms after denoising by three methods: (a) Wavelet threshold denoising; (b) VMD denoising; (c) Denoising by the method of the present invention Specific implementation manner

[0079] The following takes the Lorenz signal as a pure signal and the simulation experiment of adding white Gaussian noise as an example, and combines the accompanying drawings to illustrate the specific implementation manner of the present invention: Select the typical chaotic signal of the Lorenz signal as the pure signal for the simulation experiment

[0080]

[0081] Where x, y, and z are the state variables of the system, usually representing the three components of the system. σ is the Prandtl number, usually set to 10. ρ is the Rayleigh number, usually set to 28. β is a constant, usually set to 8 / 3

[0082] Set the step size to 0.01, and the starting points are x0 = -1, y0 = 0, z0 = 1. Select 2048 sampling points of the x component of the Lorenz signal as the simulated underwater acoustic signal, and add Gaussian white noise of -5dB, 0dB, and 5dB for the simulation experiment. Take the analysis of the 5dB noisy signal as an example. The waveform diagram of the noisy signal is as Figure 2 shown. At a signal-to-noise ratio of 5dB, there is already obvious interference to the signal quality, which will affect the clarity and recognizability of the signal, and can represent the underwater acoustic signal of general noise interference

[0083] Execute Step 1: Signal decomposition and parameter optimization based on BVMD

[0084] First, use BWOA to optimize the initial parameters of the VMD to decompose the noise signal, and set the population number N pop = 30, the upper limit of the number of iterations MaxIter = 30, the value range of the number of modes K is [5, 13], and the value range of the penalty factor α is [1000, 10000]. The VMD parameter optimization process of the simulation signal is as Figure 3 shown. The best parameter combination obtained by the algorithm iteration is K = 12, α = 5370, and the best fitness value of SampEn is 7.45479

[0085] Then, perform VMD decomposition on the noisy signal using the optimized parameter configuration to obtain 12 modal components, and their time-domain morphology and spectrum are as Figure 4As shown. The observation results show that by improving the BVMD method, complex noise signals are successfully separated, and components with similar frequencies are clearly assigned to independent frequency bands. The spectrum of each component is pure and free of aliasing, showing a single frequency characteristic and avoiding the phenomenon of spectrum confusion.

[0086] Execute Step 2: Normalized KNN-MI Preliminary Screening

[0087] Calculate the normalized KNN-MI values of each IMF component and the input signal f(t) to distinguish the signal-dominant IMF components and noise IMF components. The results are shown in Table 1.

[0088] Table 1 Normalized KNN-MI Values of Each IMF Component

[0089]

[0090] Perform normalization on it to obtain the average value which is 0.0755. Use this as the threshold to screen IMF components. The normalization is as Figure 5 shown. Take the IMF1 and IMF2 components greater than the threshold as the signal-dominant IMF components, and take the IMF3-IMF12 less than the threshold as the noise IMF components.

[0091] Execute Step 3: RCMFDE Secondary Screening

[0092] Calculate the RCMFDE values of IMF3-IMF12 to distinguish the noise IMF components into pure noise IMF and noise-dominant IMF components. The calculation results are shown in Table 2.

[0093] Table 2 RCMFDE Values of Each IMF Component

[0094]

[0095] Obtain the average value which is 2.1286. Use this as the threshold.

[0096] Execute Step 4: BWSTD Denoising and Signal Reconstruction

[0097] IMF3, IMF4, and IMF5 are less than the average value. Take them as the noise-dominant IMF components and perform the next step of BWSTD denoising. IMF6-IMF12 are greater than the average value. Discard them as pure noise. Reconstruct the signal-dominant IMF components and the noise-dominant IMF components after BWSTD denoising to obtain the final denoised signal.

[0098] To verify the effectiveness of the proposed algorithm, three algorithms, namely VMD denoising, wavelet threshold denoising method, and the algorithm proposed in this paper, are compared and analyzed. As an advanced signal decomposition method, VMD can handle non-stationary signals well and has been widely used in signal denoising and feature extraction. The performance of VMD has been verified in multiple fields. Therefore, as a benchmark method, it can effectively reflect the advantages of the new algorithm. Although the wavelet threshold denoising method is relatively classical, it is prone to signal distortion when dealing with complex noise. Therefore, it is also used as one of the comparison methods to comprehensively evaluate the performance of the method proposed in this paper. The signal waveforms before and after denoising are as Figure 6 shown.

[0099] As can be seen from Figure 6 (a), after wavelet threshold denoising, although the noise in the signal is reduced, the difference between the denoised waveform and the original signal is relatively large, especially in the high-frequency part, where obvious distortion of the signal appears. And after wavelet threshold denoising, the waveforms of some signals are distorted, and the distortion is relatively serious, especially in some detail parts, where the shape of the signal is quite different from the original signal. This is mainly because the wavelet threshold denoising method has limited ability to handle complex noise during the denoising process, especially when dealing with parts of the signal with more details or faster signal changes, it is easy to cause waveform distortion. Therefore, although the wavelet threshold denoising method can effectively reduce noise, it has certain limitations in signal preservation.

[0100] By observing Figure 6 (b), it can be seen that after the Lorenz signal is denoised by VMD, although most of the noise components have been removed, some waveforms are slightly rough, indicating that this process slightly damages the signal during filtering; while after wavelet threshold denoising, although the noise is reduced, the waveform of the signal is severely damaged by distortion, and the waveform of the signal is distorted in some places; the algorithm proposed in this paper filters out most of the noise, and the waveform of the signal remains smooth without distortion, and can effectively filter out high-frequency noise and most of the low-frequency noise.

[0101] As can be seen from Figure 6 (c), the method proposed in this paper can well maintain the smoothness and details of the signal while removing most of the noise. After denoising, the waveform of the signal is highly consistent with the original signal, without obvious distortion or aberration, which indicates that the method proposed in this paper can effectively retain the original form of the signal while removing noise. Especially in the removal of high-frequency noise and the suppression of low-frequency noise, the method proposed in this paper not only successfully filters out the noise, but also avoids over-smoothing or distortion of the signal, demonstrating strong denoising ability and signal preservation ability. By comparing with the VMD and wavelet threshold denoising methods, it can be seen that the method proposed in this paper avoids the distortion and roughness problems that appear in other methods while maintaining the smoothness of the signal, demonstrating its comprehensive advantages in noise removal and signal recovery.

[0102] To comprehensively evaluate the performance of the three denoising strategies, the Signal to Noise Ratio (SNR) and the Root Mean Square Error (RMSE) are used as the key evaluation criteria. Table 3 summarizes the denoising efficacy of the three methods, and through intuitive comparison, reveals the performance of different algorithms in terms of SNR and RMSE.

[0103] Table 3 Denoising Evaluation Metrics for the Three Methods

[0104]

[0105] As can be seen from Table 3, according to the tabular data, the performance differences of the three denoising methods at different noise levels are obvious. For the low-noise environment (SNR = 5), the SNR of the VMD method is 14.854 and the RMSE is 1.728, showing good performance and ideal denoising effect; the SNR of the wavelet threshold method is 14.749 and the RMSE is 1.733, with performance similar to that of VMD; while the algorithm in this paper shows the best performance, with an SNR of 15.117 and an RMSE of only 1.247, and the denoising effect is the most superior.

[0106] In the medium-noise environment (SNR = 0), the SNR of VMD drops to 8.241 and the RMSE increases to 9.962, and the denoising effect decreases significantly; the SNR of the wavelet threshold method is 10.169 and the RMSE is 4.245, showing better performance than the VMD method; the SNR of the algorithm in this paper is 9.983 and the RMSE is 2.761. Although it does not reach the highest SNR, it shows better performance in terms of RMSE than the other two methods, and the denoising effect is relatively ideal.

[0107] In the high-noise environment (SNR = -5), the SNR of VMD drops to 2.338 and the RMSE surges to 35.677, showing the worst performance; the SNR of the wavelet threshold method is 4.472 and the RMSE is 15.092. Although there is some improvement, it is still not ideal; while the SNR of the algorithm in this paper is 6.473 and the RMSE is 3.774, showing the best performance, and the denoising effect is significantly better than the other two methods.

[0108] Generally speaking, the denoising method of the present invention shows the best performance at all noise levels. Especially under high-noise conditions, it has obvious advantages and can provide better denoising effects, while the VMD method shows the worst performance in environments with higher noise.

[0109] The present invention also provides a computer program including software instructions, which implement the performance evaluation method when executed by a computer, including but not limited to personal computers, minicomputers, mainframes, workstations, network or distributed computing environments, separate or integrated computer platforms, or communication with charged particle tools or other imaging devices, etc. Aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer and can be used to configure and operate the computer to execute the processes described herein when the storage medium or device is read by the computer. In addition, the machine-readable code, or portions thereof, can be transmitted via a wired or wireless network. When such media include instructions or programs that implement the steps described above in combination with a microprocessor or other data processor, the inventions described herein include these and other different types of non-transitory computer-readable storage media. The present invention also includes the computer itself when programmed according to the methods and techniques of the present invention.

[0110] As described above, these are only the preferred embodiments of the present invention. The present invention is not limited to the above-described embodiments. As long as the same means are used to achieve the technical effects of the present invention, they should fall within the protection scope of the present invention. Within the protection scope of the present invention, various different modifications and variations can be made to its technical solutions and / or implementation manners.

Claims

1. Adaptive multivariate optimization underwater acoustic signal denoising method, characterized in that: The following steps are involved: Step 1: Signal decomposition and parameter optimization based on BVMD; Step 2: Normalized KNN-MI preliminary screening; Step 3: RCMFDE secondary screening; Step 4: BWSTD noise reduction and signal reconstruction.

2. The method for denoising underwater acoustic signals by adaptive multivariate optimization according to claim 1 is characterized in that: The step 1: signal decomposition and parameter optimization based on BVMD, including: Sample Entropy (SampEn) is selected as the fitness function to evaluate the quality of VMD decomposition; for a one-dimensional time series of length N The definition of sample entropy measures the complexity and uncertainty by calculating the frequency of similar patterns in the time series, then SampEn is defined as in, represents the distance measure between positions m and n in the time series, x t is a time series, m and n are the indices of the time series; Next, calculate the length of the time series The matching ratio of the embedding vector in, is the matching ratio between all embedded vectors under a given threshold r; d is the distance between two sequences, r is a set distance threshold, N is the sequence length, is the embedding dimension; further, calculate Ratios at different scales Finally, the expression of sample entropy is obtained Sample Entropy It reflects the complexity of the time series at different scales. The smaller the value, the stronger the regularity of the sequence, and the larger the value, the higher the complexity of the sequence. In BVMD, sample entropy is used as a fitness function to evaluate the effect of each VMD decomposition. The goal is to minimize sample entropy to obtain better signal decomposition. The sub-steps of BVMD are as follows: (1) Initialize the parameters of BWOA, and set the population size to N pop , the upper limit of the number of iterations MaxIter, the value range of the number of modes K, and the range of the penalty factor α; (2) Perform VMD decomposition on the signal to obtain preliminary IMFs; (3) Obtain the fitness by calculating the sample entropy SampEn; (4) According to the update rules of the black widow optimization algorithm, the position of the spider population is updated and the parameters K and a are optimized; (5) Repeat steps 2, 3, and 4 until the maximum number of iterations is reached. (6) Finally, a set of optimal VMD parameters (K, a) is obtained; (7) Update the VMD parameters K and a, and re-perform VMD decomposition on the signal; (8) Based on the optimized parameters, VMD decomposition of the signal is performed again; (9) Finally, K IMFs are obtained as the output of the algorithm.

3. The method for denoising underwater acoustic signals by adaptive multivariate optimization according to claim 1, characterized in that: The step 2: normalized KNN-MI preliminary screening includes: The obtained K IMFs contain target and noise signals, which are initially classified into signal components and noise components using normalized KNN-MI; the average value of the normalized KNN-MI values ​​of the K IMFs is calculated And set it as the threshold; if the normalized KNN-MI value of an IMF component is greater than It is judged as the signal-dominant IMF component; if it is less than It is determined to be the noise IMF component; For each IMF component X obtained by decomposition, calculate its k-nearest neighbor mutual information with the original signal f(t) According to the k-nearest neighbor mutual information estimation method Where N is the total number of samples, k is the number of nearest neighbors, and n x and n y They represent the number of nearest neighbors of random variables X and Y respectively, and ψ(N) is the digamma function used to adjust the number of nearest neighbors; Normalize the calculated mutual information It is strictly limited to the range [0, 1]; here H(X) is the entropy of the random variable X Where p(x) is the probability of X, and x is a possible value of X.

4. The method for denoising underwater acoustic signals by adaptive multivariate optimization according to claim 1, characterized in that: Step 3: RCMFDE secondary screening, including: The noise components after the normalized KNN-MI preliminary screening are subjected to RCMFDE secondary screening and divided into noise-dominant components and pure noise; the RCMFDE values ​​of all noise IMF components are calculated and their average value is calculated. As the threshold, The IMF component is removed as pure noise, which is less than The noise-dominated IMF component enters the subsequent processing stage; For the noise IMF component, calculate its refined composite multi-scale fluctuation dispersion entropy (RCMFDE); let the noise IMF component be u={u1,u2,…,u L }, for a given scale factor τ, calculate the kth coarse-grained sequence of the time series u Calculate the probability mean of the βth scatter mode in the coarse-grained sequence in is the scattering pattern corresponding to the kth coarse-grained sequence under the scale factor τ probability; The Shannon entropy of the probability average of this sequence distribution pattern is defined as the RCMFDE value E RCMFD (x,m,c,d,τ); The embedding dimension m is usually chosen to be 2 or 3, the delay d=1, and the number of categories c is an integer between [3,8].

5. The method for adaptive multivariate optimization of underwater acoustic signal denoising according to claim 1 is characterized in that: The step 4: BWSTD noise reduction and signal reconstruction includes: The noise dominant component obtained after the secondary screening of RCMFDE is further subjected to BWSTD denoising to obtain the denoised signal and the signal component after the initial screening of normalized KNN-MI is reconstructed to obtain the final complete denoised signal. The filtered noise-dominated signal is converted into the wavelet domain using wavelet transform to obtain the wavelet coefficients c of various scales j and frequency bands n. j,n ; This step can be expressed as: x(t)→{c j,n }, the purpose is to decompose the signal at different frequency and time scales, just like decomposing a complex mixture into different components, so that the wavelet coefficients of each component can be analyzed and processed later; Calculate the standard deviation σ of the signal; calculate based on the statistical characteristics of the wavelet coefficients; According to the Bayesian equation, the posterior probability distribution P(c j,n |x) A represents a state of the wavelet coefficient, specifically whether it is noise or signal, and B represents the observed value of the signal, specifically the current wavelet coefficient c j,n ; P(B|A) is the likelihood function, which indicates the probability of observing B given state A; P(A) is the prior probability of A, which reflects the guess or known information about the wavelet coefficient state A before the signal x is observed; P(B) is the probability of observing B; Determine the posterior probability distribution P(c j,n The average value of |x) is used as the wavelet threshold λ Where M is the number of samples involved in calculating the mean; For the wavelet coefficient c j,n , using soft threshold function for processing When |c j,n |≥λ, through sgn(c j,n )(|c j,n |-λ) shrinks the wavelet coefficients, where sgn(c j,n ) is a symbolic function, which is based on c j,n The positive or negative value of is 1 or -1, so that the influence of noise can be reduced while retaining the main characteristics of the signal; and when |c j,n When |<λ, the wavelet coefficients are directly set to 0 to further remove possible noise components; The processed wavelet coefficient c′ j,n Perform inverse transform to get the denoised signal Right now This step is to convert the signal processed in the wavelet domain back to the original time domain, so that we can get the final denoising result for use in subsequent signal analysis and applications; finally, the noise-dominated IMF component and the signal-dominated IMF component after BWSTD denoising are reconstructed to obtain the final denoised signal.

6. A computer device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement any one of the methods described in claims 1 to 5 when executing the computer program.

7. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by a processor, any one of the methods described in claims 1 to 5 is implemented.

Images