A reverberation suppression method and system based on self-supervised learning and a storage medium

Abstract

The application provides a reverberation suppression method and system based on self-supervised learning, and a storage medium, comprising: collecting signals received by an active sonar array element and preprocessing; calculating the direction with the maximum sonar array element correlation of the preprocessed signal to determine the target direction, and compensating the phase shift between each array element and the target direction; based on the compensated array element data, performing reverberation suppression on each array element data through a self-supervised reverberation suppression network model, and outputting a dereverberation result; based on the dereverberation result, performing beamforming on each array element channel data, and outputting the result after reverberation suppression. The application solves the problem of poor sonar reverberation suppression effect.

Description

Technical Field

[0001] This invention relates to the field of sonar signal processing technology, and in particular to a reverberation suppression method, system and storage medium based on self-supervised learning. Background Technology

[0002] Reverberation suppression is a fundamental issue in active sonar signal processing. The scattered echoes produced when the acoustic signals emitted by an active sonar transmitter encounter a target object of interest in the water are called target echoes. However, these emitted signals also produce scattered echoes when they encounter other scattering objects of interest in the water, such as fish, bubbles, the bottom, and the surface; these scattered echoes are collectively called reverberation. Because reverberation is correlated with target echoes, it affects the quality of the sonar echo signal and severely interferes with sonar detection of underwater targets. Therefore, reverberation suppression is an essential part of active sonar signal processing. Active sonar receivers use an array of multiple hydrophones to receive echo signals. The transducers in the hydrophones convert the received scattered echoes into electrical signals. These multi-channel data then undergo bandpass filtering, automatic gain control, sampling, quadrature demodulation, matched filtering, and other processing steps before beamforming and synthesizing a single data output.

[0003] Reverberation suppression methods for sonar receivers mainly fall into two categories: those based on single-pulse firing data and those based on multi-pulse firing data. Methods based on single-pulse firing data can suppress reverberation in the time domain, transform domain, and spatial domain. Classical time-domain methods for reverberation suppression include AR pre-whitening, subspace decomposition, and reverberation statistical modeling. These time-domain methods have good reverberation suppression effects under specific conditions, but their models require strong assumptions and are difficult to adapt to the complex and variable marine environment. Transform-domain methods for reverberation suppression include fractional Fourier transform, Doppler effect methods, and non-negative matrix factorization. These methods suppress reverberation in the transform domain, but often introduce errors during the inverse transform process. Spatial-domain methods for reverberation suppression mainly include time-varying beamforming and spatiotemporal adaptive processing. These methods require prior information such as target direction, which is often unavailable in practice. Reverberation suppression methods based on multi-pulse firing data are applied to the de-reverberation of sonar videos with moving targets, but their application scope is relatively limited.

[0004] With the rapid development of deep learning, more and more scholars are applying it to the field of sonar signal processing. Existing deep learning-based reverberation suppression methods mostly use supervised learning, that is, constructing demeveraging models in a data-driven manner. To address the problem of scarce sonar data, existing supervised learning methods generate targets and reverberation through model simulation and synthesize them into simulated data. Models trained on this synthesized data can achieve good reverberation suppression results under conditions where the reverberation statistical model is known. However, when applied to data with unknown reverberation statistical characteristics, the reverberation suppression performance of these models will be greatly limited. Currently, the application of self-supervised learning in sonar signal processing is limited to sonar image denoising; it has not yet been applied to sonar reverberation suppression, and existing reverberation suppression methods lack generalization ability in complex marine environments. Summary of the Invention

[0005] This invention provides a reverberation suppression method, system, and storage medium based on self-supervised learning to solve the problem of poor reverberation suppression in existing sonar systems.

[0006] This invention provides a reverberation suppression method based on self-supervised learning, comprising:

[0007] Collect and preprocess the signals received by the active sonar array elements;

[0008] The direction with the highest correlation between the sonar array elements is calculated from the preprocessed signal and determined as the target direction. The phase shift between each array element and the target direction is compensated.

[0009] Based on the compensated array data, a self-supervised reverberation suppression network model is used to suppress the reverberation of each array data element, and the de-reverberation result is output.

[0010] Based on the dereverberation results, beamforming is performed on the data from each array element channel to complete the output of the reverberation-suppressed result.

[0011] According to the reverberation suppression method based on self-supervised learning provided by the present invention, the step of acquiring and preprocessing the signal received by the active sonar array elements specifically includes:

[0012] Collect signals received by active sonar array elements;

[0013] Bandpass filtering limits the bandwidth of received data and suppresses out-of-band noise.

[0014] The received carrier signal is converted into a baseband signal through quadrature demodulation;

[0015] Matched filtering is used to convolve the transmitted baseband signal with the data received by each array element channel, thereby enhancing the part of the received signal that is related to the transmitted signal and suppressing the noise part of the received signal that is unrelated to the transmission.

[0016] According to the reverberation suppression method based on self-supervised learning provided by the present invention, the method of calculating the direction with the highest correlation between sonar array elements in the preprocessed signal and determining it as the target direction, and compensating for the phase shift between each array element and the target direction, specifically includes:

[0017] Analysis of the correlation between target echo and reverberation between array elements revealed that the correlation of reverberation between array elements is much lower than that of target echo.

[0018] Calculate the direction with the highest correlation within the common observation area of ​​all array elements and determine it as the target direction of interest;

[0019] Phase compensation is performed on the data of each channel based on the principle of aligning the array receiving direction with the direction of greatest correlation, so that the target echo in the received data of each channel remains in phase.

[0020] According to the present invention, a reverberation suppression method based on self-supervised learning is provided, wherein the reverberation suppression of each matrix element is performed on the compensated matrix element data through a self-supervised reverberation suppression network model, and the de-reverberation result is output, specifically including:

[0021] Input the metadata of each array into the self-supervised reverberation suppression network model;

[0022] The self-supervised reverberation suppression network model is used to suppress reverberation in the metadata of each array, and the de-reverberation result is output.

[0023] According to the reverberation suppression method based on self-supervised learning provided by the present invention, the specific training process of the self-supervised reverberation suppression network model is as follows:

[0024] Data from multiple channels received after each pulse firing of the active sonar are fed into a reverberation suppression network for training.

[0025] The loss function is set to maximize the correlation between data between channels. During training, an online learning strategy is used to train on the data received after each pulse firing. Multiple iterations are performed on the data of each pulse firing until the iteration converges. Inference is performed after training is completed.

[0026] Alternatively, the data after beamforming can be used as constraints, and the reverberation suppression network can be input with multiple channels of data. After training, inference can be performed to generate a self-supervised reverberation suppression network model.

[0027] According to the present invention, a reverberation suppression method based on self-supervised learning is provided, which performs beamforming on the data of each array element channel based on the dereverberation result, and outputs the result after reverberation suppression, specifically including:

[0028] The data from each channel is weighted and summed using a preset beamforming model, and the multiple data streams are combined into one output stream.

[0029] Using a Chebyshev window during weighting allows for the narrowest main lobe width while maintaining a constant side lobe height, enabling beamforming and outputting the result after reverberation suppression.

[0030] The present invention also provides a reverberation suppression system based on self-supervised learning, the system comprising:

[0031] The signal preprocessing module is used to acquire and preprocess the signals received by the active sonar array elements;

[0032] The phase compensation module is used to calculate the direction with the greatest correlation between the sonar array elements and determine it as the target direction from the preprocessed signal, and to compensate for the phase shift between each array element and the target direction.

[0033] The self-supervised reverberation suppression network module is used to suppress the reverberation of each array data element based on the compensated array data element through the self-supervised reverberation suppression network model, and output the de-reverberation result.

[0034] The beamforming module is used to perform beamforming on the data of each array element channel based on the dereverberation result, and output the result after reverberation suppression.

[0035] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the reverberation suppression method based on self-supervised learning as described above.

[0036] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the reverberation suppression method based on self-supervised learning as described above.

[0037] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the reverberation suppression method based on self-supervised learning as described above.

[0038] This invention provides a reverberation suppression method, system, and storage medium based on self-supervised learning. Through a phase compensation module, the target echo in each array element's data is aligned in phase. Based on the characteristic that the correlation of reverberation between array elements is much smaller than the correlation of target echoes, and the self-supervised training approach, reverberation suppression of sonar received data can be achieved using only reverberated array element data, without requiring clean, reverberation-free target data or an accurate reverberation model. Compared to traditional methods and supervised learning methods, this invention can adapt to and suppress reverberation variations under different marine environments. Through online learning, this invention can perform real-time training and inference, ensuring the real-time performance and adaptability of reverberation suppression using neural networks, thus improving the reverberation suppression effect. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0040] Figure 1 This is one of the flowcharts of a reverberation suppression method based on self-supervised learning provided by the present invention;

[0041] Figure 2 This is the second flowchart of a reverberation suppression method based on self-supervised learning provided by the present invention;

[0042] Figure 3 This is the third flowchart of a reverberation suppression method based on self-supervised learning provided by the present invention;

[0043] Figure 4 This is the fourth flowchart of a reverberation suppression method based on self-supervised learning provided by the present invention;

[0044] Figure 5 This is the fifth flowchart of a reverberation suppression method based on self-supervised learning provided by the present invention;

[0045] Figure 6 This is a schematic diagram of the module connection of a reverberation suppression system based on self-supervised learning provided by the present invention;

[0046] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention;

[0047] Figure 8 This is a schematic diagram of a reverberation suppression system architecture based on self-supervised learning provided by the present invention;

[0048] Figure 9 This is a schematic diagram of the encoder-decoder structure provided by the present invention;

[0049] Figure 10 This is a schematic diagram of the original data after pulse firing beamforming provided by the present invention;

[0050] Figure 11 This is a schematic diagram of the data processed by the PCI method provided by the present invention;

[0051] Figure 12 This is a schematic diagram of the data processed by the SSE method provided by the present invention;

[0052] Figure 13This is a schematic diagram of the data processed by the PCI-SVM method provided by the present invention;

[0053] Figure 14 This is a schematic diagram of the data processed by the self-supervised learning reverberation suppression system provided by the present invention;

[0054] Figure 15 This is the original side-scan sonar image after beamforming provided by the present invention;

[0055] Figure 16 The present invention provides a side-scan sonar image after reverberation suppression;

[0056] Figure 17 This is a schematic diagram of the 219th pulse firing data of the original side-scan sonar image provided by the present invention;

[0057] Figure 18 This is a schematic diagram of the 219th pulse firing data of the side-scan sonar image after reverberation suppression provided by the present invention;

[0058] Figure label:

[0059] 110: Signal preprocessing module; 120: Phase compensation module; 130: Self-supervised reverberation suppression network module; 140: Beamforming module;

[0060] 710: Processor; 720: Communication interface; 730: Memory; 740: Communication bus. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0062] The following is combined with Figures 1-5 A reverberation suppression method based on self-supervised learning according to the present invention includes:

[0063] S100: Acquire the signals received by the active sonar array elements and perform preprocessing;

[0064] S200: Calculate the direction with the highest correlation between the sonar array elements after preprocessing the signal and determine it as the target direction, and compensate for the phase shift between each array element and the target direction.

[0065] S300: Based on the compensated array data, the reverberation suppression of each array data element is performed through a self-supervised reverberation suppression network model, and the de-reverberation result is output.

[0066] S400. Based on the reverberation result, beamforming is performed on the data of each array element channel to complete the output of the result after reverberation suppression.

[0067] This invention overcomes the limitations of current active sonar reverberation suppression technologies, which require specific model assumptions and threshold parameter adjustments in different environments. It addresses the bottleneck of insufficient generalization ability of existing reverberation suppression methods in complex marine environments, significantly improving the signal-to-mixing ratio of sonar received data to aid subsequent target detection and identification. The self-supervised learning-based active sonar reverberation suppression system proposed in this invention can achieve adaptive reverberation suppression for different environmental reverberation levels through real-time pulse-by-pulse firing training using only multi-channel sonar received data with reverberation. It achieves state-of-the-art performance in both subjective visual effects and objective metrics.

[0068] The signals received by the active sonar array elements are acquired and preprocessed, specifically including:

[0069] S101. Acquire signals received by the active sonar array elements;

[0070] S102. Bandpass filtering is used to limit the bandwidth of received data and suppress out-of-band noise.

[0071] S103. Convert the received carrier signal into a baseband signal through quadrature demodulation;

[0072] S104. By using matched filtering, the transmitted baseband signal is convolved with the data received by each array element channel to enhance the part of the received signal related to the transmitted signal and suppress the noise part of the received signal unrelated to the transmission.

[0073] This invention preprocesses the sampled active sonar array element received signal, specifically for linear frequency modulated (LFM) signals, including bandpass filtering, quadrature demodulation, and matched filtering. Bandpass filtering uses a digital bandpass filter to filter the data from each array element channel. Quadrature demodulation converts the carrier signal into a baseband signal using I and Q-channel demodulation to obtain the main signal information. Matched filtering convolves the transmitted baseband signal with the received data from each array element channel, enhancing the portion of the received signal related to the transmitted signal and suppressing noise unrelated to the transmitted signal.

[0074] The direction with the highest correlation between the sonar array elements is calculated from the preprocessed signal and determined as the target direction. The phase shift between each array element and the target direction is compensated, specifically including:

[0075] S201. Through the analysis of the correlation between target echo and reverberation between array elements, it was determined that the correlation of reverberation between array elements is much lower than that of target echo.

[0076] S202. Calculate the direction with the highest correlation within the common observation area of ​​all array elements and determine it as the target direction of interest;

[0077] S203. Based on the principle of aligning the array receiving direction with the direction of greatest correlation, phase compensation is performed on the data of each channel to ensure that the target echo in the received data of each channel remains in phase.

[0078] In this invention, the phase of the target echo in each array element channel data is aligned. After the data received by each array element undergoes orthogonal demodulation and matched filtering, the direction with the highest correlation among the channel data is calculated within the possible receiving directions, and this direction is considered the most likely target direction. Specifically, beamforming scans are performed on the possible receiving directions, and the direction with the highest correlation is selected based on the principle of maximizing the energy of the beamforming results for each channel data. Then, phase compensation is performed on the channel data according to this direction with the highest correlation. At this point, the phase of the target echo in each array element channel data is considered aligned, and the array data with aligned target echo is output. In practical implementation, the direction with the highest correlation can be defaulted to zero degrees.

[0079] Based on the compensated array data, a self-supervised reverberation suppression network model is used to suppress the reverberation of each array data element, outputting the de-reverberation result, specifically including:

[0080] Input the metadata of each array into the self-supervised reverberation suppression network model;

[0081] The self-supervised reverberation suppression network model is used to suppress reverberation in the metadata of each array, and the de-reverberation result is output.

[0082] The specific training process of the self-supervised reverberation suppression network model is as follows:

[0083] S301. The data from multiple channels received after each pulse firing of the active sonar are fed into the reverberation suppression network for training.

[0084] S302. Set the loss function to maximize the correlation between channel data. During training, use an online learning strategy to train on the data received after each pulse firing. Perform multiple iterations on the data of each pulse firing until the iteration converges. After training is completed, perform inference.

[0085] S303, or the data after beamforming can be used as constraints, the reverberation suppression network can be input with multiple channels of data, and inference can be performed after training to generate a self-supervised reverberation suppression network model.

[0086] refer to Figure 8In this invention, an encoder-decoder architecture network is used to suppress reverberation on active sonar received data through self-supervised training. A 1D-Unet network can be used. Specifically, the data from each pulse received by the active sonar, firing multiple channels, is fed into the reverberation suppression network for training. The loss function is set to maximize the correlation between data between channels. This invention does not require clean, reverberation-free data; reverberation suppression is achieved simply through self-supervised training between multiple channels. During training, an online learning strategy is used for pulse-by-pulse training. Multiple iterations are performed on the data from each pulse. After convergence, inference is performed, and the most correlated data between each channel is output.

[0087] For the network design of this module, please refer to the appendix. Figure 9 This module employs an encoder-decoder architecture, specifically using a one-dimensional U-Net network for feature extraction and signal reconstruction. This example includes 5 convolutional layers and 5 deconvolutional layers, with direct connections between the deconvolutional layers and their corresponding convolutional layers. All convolutional and deconvolutional layers use 11×11 convolutional kernels, employ ReLU as the activation function, and utilize batch normalization layers. Except for the last convolutional layer, each convolutional layer is followed by a pooling layer, which downsamples the data by a factor of two. Therefore, the feature map size in the encoder gradually decreases, while the number of channels in the feature map increases. Similarly, except for the last deconvolutional layer, each deconvolutional layer is followed by an upsampling layer, so the feature map size in the decoder gradually increases, while the number of channels in the feature map decreases. Both the training and inference networks in this module use the aforementioned one-dimensional U-Net network. In practical implementations, this network can be replaced with other encoder-decoder architectures, such as the SRN (Scale-Recurrent Network) used for deblurring.

[0088] For network initialization, Kaming initialization is used for convolutional layer parameters, which effectively avoids gradient vanishing or gradient explosion during backpropagation and accelerates network convergence. For training and inference design, this invention adopts an online learning approach, performing pulse-by-pulse training and inference. N array data sequences received by the sonar array within the same time period are grouped into N(N-1) pairs. During training, each iteration uses one pair of data sequences; the data sequence of one array element is directly input into the one-dimensional U-Net network, while the data of the other element serves as a constraint for the network. After training all N(N-1) pairs of data sequences, this training process is repeated until iterative convergence. Besides this training method, array element channel data can also be used as training input, with the beamforming results of each channel data serving as training constraints, iterating until training convergence. After iterative convergence, model inference is performed, inputting the data of the N array elements sequentially into the trained network to obtain the reverberation-suppressed data of the N array elements. This pulse-by-pulse firing training method effectively addresses the issue of varying reverberation distributions caused by changes in the marine environment. The reverberation suppression network can learn real-time reverberation changes based on the principle of maximum correlation, achieving lifelong learning. In contrast, current traditional methods and supervised learning methods only offer good reverberation suppression under specific conditions and cannot generalize to diverse marine environments. Besides the above training method, beamforming data can also be used as constraints, with data from N channels as input.

[0089] Regarding the loss function design, based on the characteristic that the correlation of target echoes in the received data of array elements is greater than the reverberation correlation, the correlation coefficient between the data of two array elements is used as the loss function. If X... p X represents the received data vector (column vector) of the p-element array. q (Column vector) represents the received data vector of the q-element array. X represents p The mean of a vector. X represents q The mean of the vector, |·| represents the L2 norm, and C is a constant, which can be 1 + 1 × 10⁻⁶. -9 Then the loss function for training data between the two array elements can be expressed as:

[0090]

[0091] in This is the correlation coefficient between the data from the two datasets. Taking the negative logarithm of the correlation coefficient increases the penalty for data with low correlation coefficients. In implementation, the loss function can also be designed in other forms, as long as it is optimized in the direction of increasing the correlation between the two channels.

[0092] Based on the dereverberation results, beamforming is performed on the data from each array element channel to complete the output of the reverberation-suppressed result, specifically including:

[0093] S401. The data from each channel is weighted and summed using a preset beamforming model, and the multiple data streams are combined into one output stream.

[0094] S402. When using a Chebyshev window for weighting, the main lobe width can be minimized while maintaining a certain side lobe height, thus enabling beamforming and outputting the result after reverberation suppression.

[0095] Referring to Appendix Table 1, reverberation suppression performance was tested on a synthetic dataset using different reverberation suppression algorithms. The synthetic dataset was synthesized from real seabed reverberation data and simulated targets, consisting of 10 pings, 36 array element channels, and 2000 sampling points per channel. The table compares this method with traditional methods. It shows that when the initial data reverberation is high and the signal-to-mixing ratio is low (initial signal-to-mixing ratio 7.24 dB), the reverberation suppression effect of traditional methods is generally poor. However, the self-supervised reverberation suppression network method in this invention achieves a signal-to-mixing ratio improvement of over 10 dB, far exceeding that of traditional algorithms.

[0096] Table 1

[0097]

[0098] See attached document Figure 10-14 , respectively represent the original data after one of the ping beams of the synthetic dataset is formed, the data after processing by the PCI method, the data after processing by the SSE method, the data after processing by the PCI-SVM method, and the data after processing by this system. From the visual effect, it can be seen that the self-supervised reverberation suppression network of the present invention has achieved excellent reverberation suppression effect on real seabed reverberation with unknown distribution, and significantly highlights the target echo signal.

[0099] See attached document Figure 15-18 This series of data consists of seabed side-scan sonar data obtained from real-world measurements at sea, with cylindrical targets placed on the seabed. On this real dataset, reverberation suppression was performed on multiple areas containing real cylindrical targets, achieving an average signal-to-mixing ratio improvement of 9.46 dB. (Attached) Figure 15 This is the original side-scan sonar image after beamforming. The part pointed to by the arrow in the image is the actual cylindrical target. (Attached) Figure 16 These are side-scan sonar images after reverberation suppression using this system. It can be seen that the contrast of the cylindrical target area is significantly improved compared to its surroundings, and its outline is clearer. Reverberation in other areas has also been effectively suppressed. (Attached) Figure 17 It is attached Figure 15 The 219th pulse firing data shows that the highest pulse was an echo from a cylindrical target, while the others were seabed reverberation or multipath interference from targets. (Attached) Figure 18 It is attached Figure 16 The 219th pulse firing data, visually, shows that seabed reverberation is greatly suppressed, multipath interference from the target is effectively suppressed, and the true cylindrical target echo is highlighted, achieving a very good reverberation suppression effect. It can be seen that this invention can significantly improve the signal-to-mixing ratio of sonar data and enhance the visual recognizability of sonar data, achieving excellent results in both objective indicators and subjective quality.

[0100] refer to Figure 6 The present invention also discloses a reverberation suppression system based on self-supervised learning, the system comprising:

[0101] The signal preprocessing module 110 is used to acquire and preprocess the signals received by the active sonar array elements;

[0102] Phase compensation module 120 is used to calculate the direction with the greatest correlation between the sonar array elements and determine it as the target direction from the preprocessed signal, and to compensate for the phase shift between each array element and the target direction.

[0103] The self-supervised reverberation suppression network module 130 is used to suppress the reverberation of each array data element based on the compensated array data element through the self-supervised reverberation suppression network model, and output the de-reverberation result.

[0104] The beamforming module 140 is used to perform beamforming on the data of each array element channel based on the dereverberation result, and output the result after dereverberation suppression.

[0105] Among them, the signal preprocessing module 110 collects the signals received by the active sonar array elements;

[0106] Bandpass filtering limits the bandwidth of received data and suppresses out-of-band noise.

[0107] The received carrier signal is converted into a baseband signal through quadrature demodulation;

[0108] Matched filtering is used to convolve the transmitted baseband signal with the data received by each array element channel, thereby enhancing the part of the received signal that is related to the transmitted signal and suppressing the noise part of the received signal that is unrelated to the transmission.

[0109] The phase compensation module 120, through the analysis of the correlation between target echo and reverberation between array elements, determines that the correlation of reverberation between array elements is much lower than that of target echo.

[0110] Calculate the direction with the highest correlation within the common observation area of ​​all array elements and determine it as the target direction of interest;

[0111] Phase compensation is performed on the data of each channel based on the principle of aligning the array receiving direction with the direction of greatest correlation, so that the target echo in the received data of each channel remains in phase.

[0112] The self-supervised reverberation suppression network module 130 inputs the metadata of each matrix into the trained self-supervised reverberation suppression network model;

[0113] The self-supervised reverberation suppression network model is used to suppress reverberation in the metadata of each array, and the de-reverberation result is output.

[0114] Data from multiple channels received after each pulse firing of the active sonar are fed into a reverberation suppression network for training.

[0115] The loss function is set to maximize the correlation between data between channels. During training, an online learning strategy is used to train on the data received after each pulse firing. Multiple iterations are performed on the data of each pulse firing until the iteration converges.

[0116] After iterative convergence, a self-supervised reverberation suppression network model is generated.

[0117] The beamforming module 140 uses a preset beamforming model to perform weighted summation on the data from each channel, combining multiple data streams into one output stream.

[0118] Using a Chebyshev window during weighting allows for the narrowest main lobe width while maintaining a constant side lobe height, enabling beamforming and outputting the result after reverberation suppression.

[0119] This invention discloses a reverberation suppression system based on self-supervised learning. Through a phase compensation module, the system aligns the target echo phase in each array element's data. Based on the characteristic that the correlation of reverberation between array elements is much smaller than the correlation of target echoes, and the self-supervised training approach, reverberation suppression of sonar received data can be achieved using only reverberated array element data, without requiring clean, reverberation-free target data or an accurate reverberation model. Compared to traditional methods and supervised learning methods, this invention can adapt to and suppress reverberation variations under different marine environments. Through online learning, this invention can perform real-time training and inference, ensuring the real-time performance and adaptability of reverberation suppression using neural networks, thus improving the reverberation suppression effect.

[0120] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include a processor 710, a communications interface 720, a memory 730, and a communication bus 740. The processor 710, communications interface 720, and memory 730 communicate with each other via the communication bus 740. The processor 710 can call logical instructions in the memory 730 to execute a reverberation suppression method based on self-supervised learning. This method includes: acquiring signals received by active sonar array elements and performing preprocessing.

[0121] The direction with the highest correlation between the sonar array elements is calculated from the preprocessed signal and determined as the target direction. The phase shift between each array element and the target direction is compensated.

[0122] Based on the compensated array data, a self-supervised reverberation suppression network model is used to suppress the reverberation of each array data element, and the de-reverberation result is output.

[0123] Based on the dereverberation results, beamforming is performed on the data of each array element channel to complete the output of the result after reverberation suppression.

[0124] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, the computer program being executed by a processor, the computer being able to execute a reverberation suppression method based on self-supervised learning provided by the above methods, the method including: acquiring signals received by active sonar array elements and performing preprocessing;

[0126] The direction with the highest correlation between the sonar array elements is calculated from the preprocessed signal and determined as the target direction. The phase shift between each array element and the target direction is compensated.

[0127] Based on the compensated array data, a self-supervised reverberation suppression network model is used to suppress the reverberation of each array data element, and the de-reverberation result is output.

[0128] Based on the dereverberation results, beamforming is performed on the data of each array element channel to complete the output of the result after reverberation suppression.

[0129] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform a reverberation suppression method based on self-supervised learning provided by the above methods, the method comprising: acquiring signals received by active sonar array elements and performing preprocessing;

[0130] The direction with the highest correlation between the sonar array elements is calculated from the preprocessed signal and determined as the target direction. The phase shift between each array element and the target direction is compensated.

[0131] Based on the compensated array data, a self-supervised reverberation suppression network model is used to suppress the reverberation of each array data element, and the de-reverberation result is output.

[0132] Based on the dereverberation results, beamforming is performed on the data of each array element channel to complete the output of the result after reverberation suppression.

[0133] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0134] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0135] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A reverberation suppression method based on self-supervised learning, characterized in that, include: Collect and preprocess the signals received by the active sonar array elements; The direction with the highest correlation between the sonar array elements is calculated from the preprocessed signal and determined as the target direction. The phase shift between each array element and the target direction is compensated. Based on the compensated array data, a self-supervised reverberation suppression network model is used to suppress the reverberation of each array data element, and the de-reverberation result is output. Based on the dereverberation results, beamforming is performed on the data of each array element channel to complete the output of the result after reverberation suppression. The process of acquiring and preprocessing the signals received by the active sonar array elements specifically includes: Collect signals received by active sonar array elements; Bandpass filtering limits the bandwidth of received data and suppresses out-of-band noise. The received carrier signal is converted into a baseband signal through quadrature demodulation; By using matched filtering, the transmitted baseband signal is convolved with the data received by each array element channel to enhance the part of the received signal that is related to the transmitted signal and suppress the noise part of the received signal that is unrelated to the transmission. The specific training process of the self-supervised reverberation suppression network model is as follows: Data from multiple channels received after each pulse firing of the active sonar are fed into a reverberation suppression network for training. The loss function is set to maximize the correlation between data between channels. During training, an online learning strategy is used to train on the data received after each pulse firing. Multiple iterations are performed on the data of each pulse firing until the iteration converges. Inference is performed after training is completed. Alternatively, the data after beamforming can be used as constraints, and the reverberation suppression network can be input with multiple channels of data. After training, inference can be performed to generate a self-supervised reverberation suppression network model.

2. The reverberation suppression method based on self-supervised learning according to claim 1, characterized in that, The process of calculating the direction with the highest correlation between sonar array elements from the preprocessed signal and determining it as the target direction, and compensating for the phase shift between each array element and the target direction, specifically includes: Analysis of the correlation between target echo and reverberation between array elements revealed that the correlation of reverberation between array elements is much lower than that of target echo. Calculate the direction with the highest correlation within the common observation area of ​​all array elements and determine it as the target direction of interest; Phase compensation is performed on the data of each channel based on the principle of aligning the array receiving direction with the direction of greatest correlation, so that the target echo in the received data of each channel remains in phase.

3. The reverberation suppression method based on self-supervised learning according to claim 1, characterized in that, The compensated array data is subjected to reverberation suppression through a self-supervised reverberation suppression network model, and the de-reverberation result is output, specifically including: The metadata of each array is input into a self-supervised reverberation suppression network model; the reverberation suppression of the metadata of each array is performed through the self-supervised reverberation suppression network model, and the de-reverberation result is output.

4. The reverberation suppression method based on self-supervised learning according to claim 1, characterized in that, Based on the dereverberation results, beamforming is performed on the data from each array element channel to complete the output of the reverberation-suppressed result, specifically including: The data from each channel is weighted and summed using a preset beamforming model, and the multiple data streams are combined into one output stream. Using a Chebyshev window during weighting allows for the narrowest main lobe width while maintaining a constant side lobe height, enabling beamforming and outputting the result after reverberation suppression.

5. A reverberation suppression system based on self-supervised learning, characterized in that, The system includes: The signal preprocessing module is used to acquire and preprocess the signals received by the active sonar array elements; The phase compensation module is used to calculate the direction with the greatest correlation between the sonar array elements and determine it as the target direction from the preprocessed signal, and to compensate for the phase shift between each array element and the target direction. The self-supervised reverberation suppression network module is used to suppress the reverberation of each array data element based on the compensated array data element through the self-supervised reverberation suppression network model, and output the de-reverberation result. The beamforming module is used to perform beamforming on the data of each array element channel based on the dereverberation result, and output the result after reverberation suppression. The process of acquiring and preprocessing the signals received by the active sonar array elements specifically includes: Collect signals received by active sonar array elements; Bandpass filtering limits the bandwidth of received data and suppresses out-of-band noise. The received carrier signal is converted into a baseband signal through quadrature demodulation; By using matched filtering, the transmitted baseband signal is convolved with the data received by each array element channel to enhance the part of the received signal that is related to the transmitted signal and suppress the noise part of the received signal that is unrelated to the transmission. The specific training process of the self-supervised reverberation suppression network model is as follows: Data from multiple channels received after each pulse firing of the active sonar are fed into a reverberation suppression network for training. The loss function is set to maximize the correlation between data between channels. During training, an online learning strategy is used to train on the data received after each pulse firing. Multiple iterations are performed on the data of each pulse firing until the iteration converges. Inference is performed after training is completed. Alternatively, the data after beamforming can be used as constraints, and the reverberation suppression network can be input with multiple channels of data. After training, inference can be performed to generate a self-supervised reverberation suppression network model.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the reverberation suppression method based on self-supervised learning as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the reverberation suppression method based on self-supervised learning as described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the reverberation suppression method based on self-supervised learning as described in any one of claims 1 to 4.

Images