A UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network

Abstract

The application belongs to the technical field of noise control, and specifically discloses a UUV self-noise suppression method based on a double-channel cooperative noise reduction neural network, which comprises the following steps: first, using a microphone to collect air channel data of the environment where the UUV is located, and simultaneously using a hydrophone to collect synchronous underwater acoustic channel data; obtaining the hydrophone noise signal and the microphone noise signal of the UUV in the underwater environment, and performing a pretreatment operation; inputting the pretreated hydrophone noise signal and the pretreated microphone noise signal into a feature extraction module in parallel, performing multi-scale feature extraction and fusion; and through a time sequence signal recovery module, performing time sequence prediction and recovery on the signal after multi-scale feature fusion. The UUV self-noise suppression method based on the double-channel cooperative noise reduction neural network improves the noise reduction efficiency and accuracy, reduces the influence of environmental noise on the working performance of the UUV, and provides reliable protection for the stable operation of the UUV in a complex noise environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of noise control, in particular to a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network. BACKGROUND

[0002] In the application scenario of underwater unmanned vehicles (UUVs), self-noise can seriously interfere with the detection performance of sonar systems, and higher requirements are put forward for noise control technology. Traditional noise control technology has many limitations:

[0003] 1) The spectral subtraction method estimates the noise spectrum from the amplitude spectrum or energy spectrum of the noisy signal, and then subtracts the noise spectrum to obtain the clean signal. Its principle is simple and the calculation speed is fast, and it is suitable for stable noise environment. However, the accuracy of noise spectrum estimation is limited, and false tones are easily generated in complex noise conditions, and it cannot effectively deal with dynamically changing noise.

[0004] 2) The least mean square error (LMS) algorithm estimates the gradient vector by the steepest descent method, which is simple and fast; it can adapt to non-stationary noise, but it is sensitive to the power of the input signal, resulting in slow convergence speed, and there is a problem of steady-state misadjustment in the time-varying noise environment.

[0005] 3) The recursive least squares RLS algorithm updates the filter weights in a recursive manner, which can quickly converge and adapt to time-varying noise; its advantage is fast convergence, but the computational complexity is large, which is suitable for non-real-time systems, and may cause algorithm divergence when dealing with ill-conditioned autocorrelation matrix.

[0006] In summary, traditional noise control technology in the application of UUV self-noise reduction is difficult to meet the actual engineering requirements due to poor adaptability to complex environments, high demand for computational resources, or insufficient algorithm stability, and it is urgent to explore more efficient and reliable noise reduction technology solutions. SUMMARY

[0007] The purpose of the present application is to provide a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, which aims to solve the problem of interference signal caused by self-noise affecting the detection performance of sonar systems in complex underwater environments for UUVs, combining deep learning technology, an innovative dual-channel collaborative noise reduction network (DUAL-HydroNet) is proposed, and a multi-source feature fusion mechanism (MSFusioner) is introduced, which effectively improves the noise reduction performance; This method realizes self-noise suppression in complex underwater acoustic environment by introducing air channel real-time data, which significantly improves the sonar detection capability of UUV.

[0008] To achieve the above purpose, the present application provides a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, comprising the following steps:

[0009] S1, using a microphone to collect air channel data of the environment where the UUV is located, and using a hydrophone to collect synchronous underwater acoustic channel data;

[0010] S2, obtaining the hydrophone noisy signal and the microphone noisy signal of the UUV in the underwater environment, and performing a preprocessing operation;

[0011] S3, parallelly inputting the preprocessed hydrophone noisy signal and the microphone noisy signal into a feature extraction module to perform multi-scale feature extraction and fusion;

[0012] S4, performing timing prediction and recovery on the multi-scale feature fused signal through a timing signal recovery module.

[0013] Preferably, in S2, the given hydrophone collected noisy signal and the microphone collected noisy signal

[0014] The preprocessing operation specifically includes:

[0015] The sampling rate is raised to 192kHz through double-stage Sinc interpolation upsampling, and the calculation formula is:

[0016] (1);

[0017] wherein, is a 2x upsampling operator based on the Smith-Gossett algorithm, and a 56-order Hanning window weighted sinc function is used to realize anti-aliasing interpolation.

[0018] Preferably, in S3, the feature extraction module is composed of a weight-shared encoder Encoder combined with a channel-time joint attention CSA module, and the feature extraction module adopts a double-path weight-shared encoder structure, specifically:

[0019] The preprocessed hydrophone noisy signal and the microphone noisy signal are parallelly inputted into a 5-level convolutional encoder, and the mathematical expression is:

[0020] (2);

[0021] wherein, is the shared weight parameter of the layer, is the hierarchical feature.

[0022] Preferably, in S3, the multi-scale feature extraction and fusion process is:

[0023] In each layer of the encoder, the preprocessed signal is first subjected to one-dimensional convolution for feature extraction, and then subjected to a ReLU activation function:

[0024] (3);

[0025] wherein, is the input channel number, the first layer , is the hidden layer dimension, n is the layer index, the convolution kernel size , the step size , the time resolution is compressed to by the stride operation;

[0026] The extracted features are subjected to gating processing by the GLU gating linear unit:

[0027] (4);

[0028] A multi-source feature fusion mechanism MSFusioner is introduced, which combines a channel-time joint attention CSA module to first perform global information aggregation on the input audio features by adaptive average pooling; then, two one-dimensional convolution layers are used for feature dimension reduction and recovery to form channel attention weights, and the calculation formula of the channel attention is as follows:

[0029] (5);

[0030] A time attention path is established to capture the time information in the audio signal, and this path reduces the channel dimension to 1 through a normal convolution, a convolution DConv with an extended convolution kernel and a one-dimensional convolution layer, and the calculation formula of the time attention is as follows:

[0031] (6);

[0032] Through the combination of channel attention and time attention, the CSA module generates a weight matrix that fuses multi-scale features, which is used to weight the input signal, and the output of the CSA module is:

[0033] (7);

[0034] wherein, F 1 is the input hydrophone signal.

[0035] Preferably, in S4, the signal after multi-scale feature fusion is subjected to time prediction and recovery by a time signal recovery module, and the time signal recovery module is composed of a bidirectional long short-term memory network combined with a decoder;

[0036] In the decoding stage, as a feature input, the fused multi-scale features As a jump connection is sent into a decoder, a joint decoding operation is performed, and finally a noise-reduced signal is output.

[0037] Preferably, the dual-channel collaborative noise reduction neural network comprises a feature extraction module and a time sequence signal recovery module, the feature extraction module is composed of a weight-shared encoder Encoder combined with a channel-time sequence attention mechanism CSA, and is used for extracting and fusing multi-source features of air channel data collected by a microphone and underwater acoustic channel data collected by a hydrophone in synchronization;

[0038] The time sequence signal recovery module is composed of a bidirectional long short-term memory network combined with a decoder, and is used for time sequence prediction and recovery of the signal after multi-scale feature fusion.

[0039] Therefore, the UUV self-noise suppression method based on the dual-channel collaborative noise reduction neural network has the following beneficial effects:

[0040] (1) The noise reduction performance is significantly improved: compared with the traditional single-channel noise reduction method, the dual-channel collaborative noise reduction network (DUAL-HydroNet) proposed in the application can more effectively capture noise features by simultaneously using air medium microphone signals and water medium hydrophone signals for joint processing, thereby significantly improving the noise reduction effect. Under the synergistic action of different signal sources, the signal-to-noise ratio is improved by more than 40dB, far exceeding traditional noise reduction algorithms and speech noise reduction networks.

[0041] (2) The signal recovery stability is enhanced: by introducing a multi-source feature fusion mechanism (MSFusioner) combined with a channel-time sequence joint attention module (CSA), the application can deeply model and fuse multi-dimensional feature information during the noise reduction process; compared with the traditional single signal noise reduction method, this technology not only improves the time domain waveform and spectral stability of the signal, but also ensures the integrity and stability of the target signal, effectively avoiding the weakening of the target signal during the noise reduction process.

[0042] (3) Cross-medium feature effective integration: the application innovatively synchronously collects air medium microphone and water medium hydrophone signals, expands the feature dimension that can be used for noise reduction, and fuses the features of air and underwater acoustic signals by using a deep neural network, thereby realizing effective integration of cross-medium features, enhancing the recognition ability of noise features, and significantly improving the noise reduction performance through the fusion mechanism.

[0043] The technical solutions of the application will be further described in detail below with reference to the drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a dual-channel collaborative noise reduction network structure diagram of an embodiment of the UUV self-noise suppression method based on the dual-channel collaborative noise reduction neural network.

[0045] Figure 2 is a feature extraction module schematic diagram of an embodiment of a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network of the application;

[0046] Figure 3 is a channel-time joint attention CSA structure diagram of an embodiment of a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network of the application;

[0047] Figure 4 is a time series signal recovery module structure diagram of an embodiment of a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network of the application. DETAILED DESCRIPTION

[0048] The technical solutions of the application are further described below through the drawings and embodiments.

[0049] Unless otherwise defined, the technical terms or scientific terms used in the application should be understood as the usual meaning understood by those skilled in the art to which the application belongs.

[0050] The application combines deep learning technology and provides a dual-channel collaborative noise reduction neural network (DUAL-HydroNet), as shown in Figure 1 The network includes a feature extraction module and a time series signal recovery module, as shown in Figure 2 The feature extraction module is composed of a weight-shared encoder Encoder combined with a channel-time attention CSA module, as shown in Figure 3 , for extracting and fusing multi-source features of air channel data collected by a microphone and synchronous underwater acoustic channel data collected by a hydrophone. As shown in Figure 4 The time series signal recovery module is composed of a bidirectional long short-term memory network combined with a decoder, for time series prediction and recovery of the signal after multi-scale feature fusion.

[0051] A UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network includes the following steps:

[0052] S1, using a microphone to collect air channel data of the environment where the UUV is located, and simultaneously using a hydrophone to collect synchronous underwater acoustic channel data.

[0053] S2, obtaining the hydrophone noisy signal and the microphone noisy signal of the UUV in the underwater environment, and performing a preprocessing operation.

[0054] In one possible implementation, given the noisy signal collected by the hydrophone and the noisy signal collected by the microphone, the preprocessing operation is specifically:

[0055] The sampling rate is raised to 192 kHz by two-stage Sinc interpolation upsampling, and the calculation formula is:

[0056] (1);

[0057] wherein, is a 2x upsampling operator based on the Smith-Gossett algorithm, and a 56-order Hanning window weighted Sinc function is used to realize anti-aliasing interpolation. After this operation, the time domain signal length is extended to , effectively improving the resolution of high frequency components.

[0058] S3, the preprocessed hydrophone noise signal and the microphone noise signal are input into the feature extraction module in parallel, and multi-scale features are extracted and fused.

[0059] The feature extraction module adopts a dual-path weight sharing encoder structure, specifically:

[0060] The preprocessed hydrophone noise signal and the microphone noise signal are input into a 5-level convolutional encoder in parallel, and the mathematical expression is:

[0061] (2);

[0062] wherein, is the shared weight parameter of the layer, is the hierarchical feature.

[0063] In each layer of the encoder, the preprocessed signal is first subjected to one-dimensional convolution for feature extraction, and then subjected to a ReLU activation function, and the specific process is shown in the following formula:

[0064] (3);

[0065] wherein, is the number of input channels, the first layer , is the hidden layer dimension, is the hierarchical index, the convolution kernel size , the step size , and the time resolution is compressed to by the cross-step operation.

[0066] Then the extracted features are subjected to GLU gate linear unit for gating processing:

[0067] (4).

[0068] By introducing a gating mechanism to regulate and filter the input features, the feature response strength of different channels is automatically adjusted to enhance useful information and suppress invalid information.

[0069] A multi-source feature fusion mechanism MSFusioner is introduced, which combines a channel-time joint attention CSA module. First, the input audio features are globally aggregated by adaptive average pooling (Adaptive Average Pooling). Then, two one-dimensional convolution layers are used for feature dimension reduction and recovery to form channel attention weights. This process can be regarded as enhancing the response of important channels by learning the global dependency between features. The calculation formula of channel attention is as follows:

[0070] (5);

[0071] To capture the timing information in the audio signal, the CSA module establishes an independent timing attention path to capture the timing information in the audio signal. This path is realized through the joint operation of three convolution layers, which are an ordinary convolution, a convolution with an extended convolution kernel DConv and a one-dimensional convolution layer, which reduces the channel dimension to 1. The design of this path is derived from the convolution decomposition strategy, which effectively expands the receptive field by extending the convolution, enabling the model to capture longer temporal dependencies.

[0072] The calculation formula of timing attention is as follows:

[0073] (6);

[0074] Through the combination of the above channel attention and timing attention, the CSA module can generate a weight matrix that integrates multi-scale features, which is used to weight the input signal. That is, the output of the CSA module is:

[0075] (7);

[0076] where, F 1 is the input hydrophone signal.

[0077] S4, through the timing signal recovery module, the multi-scale feature fused signal is timing predicted and recovered.

[0078] In the decoding stage, As a feature input, the fused multi-scale feature is sent into the decoder as a skip connection for joint decoding operation, and finally the denoised signal is output.

[0079] Numerical simulation experiment and pool experiment:

[0080] To verify the performance of the proposed dual-channel collaborative noise reduction system in the self-noise suppression of unmanned underwater vehicles (UUVs), numerical simulation experiments and pool experiments were conducted to comprehensively evaluate the noise reduction effect of the system under various signal types, signal-to-noise ratio (SNR) conditions, and actual underwater environments. The DUAL-HydroNet model integrates the signals collected by microphones and hydrophones, and combines the channel-temporal connection attention mechanism (CSA) to achieve efficient noise suppression. The following describes the experimental setup, process, simulation test results, and pool experiment results in detail, highlighting the technical advantages of the system.

[0081] A. Numerical simulation experiment

[0082] A.1 Experimental setup

[0083] The numerical simulation experiment aims to verify the noise reduction performance of the DUAL-HydroNet model under different signal-to-noise ratios (SNRs) and signal types. The experimental data set includes the following three signal types, representing the self-noise characteristics of UUVs under typical operating conditions:

[0084] Single-frequency signal: fixed frequency of 1 kHz, simulating the single-frequency noise generated by the UUV propeller or mechanical components.

[0085] Impulse signal: frequency of 1 kHz, simulating transient impact noise.

[0086] Frequency-modulated signal: frequency range of 50 Hz to 5 kHz, simulating the wideband noise of UUVs under complex operating conditions.

[0087] To simulate different noise environments, the data set is divided into three groups according to the signal-to-noise ratio: 5 dB, 3 dB, and -1 dB. Each type of signal contains 1003 samples, divided into a training set (501 samples), a validation set (100 samples), and a test set (402 samples) in a ratio of 5:1:4 to ensure the independence of model training and evaluation. Data preprocessing includes amplitude normalization, downsampling, and segmentation operations to generate standardized formats suitable for neural network input. Downsampling uses the Smith-Gossett algorithm to upsample the sampling rate from 48000 Hz to 192000 Hz to improve the resolution of high-frequency components.

[0088] The experimental platform uses a high-performance computing cluster equipped with NVIDIA A100 GPUs, running Python 3.8 environment, and implements the DUAL-HydroNet model using the PyTorch framework. The model training is set to a batch size of 32, a learning rate of 0.001, and uses the Adam optimizer for 50 cycles to ensure convergence.

[0089] A.2 Signal processing and analysis method

[0090] Signal processing extracts UUV self-noise signal features captured by microphones and hydrophones through time and frequency domain analysis, characterizing the propagation characteristics of acoustic signals in air and underwater environments. DEMON (de-modulated noise) spectral analysis is used to extract low-frequency components from high-frequency modulated signals, identifying key frequency components related to noise reduction. The analysis process is as follows:

[0091] 1. Time domain analysis: Time domain waveform analysis of microphone and hydrophone signals to observe random fluctuations and periodic characteristics of signals.

[0092] 2. Frequency domain analysis: Generate frequency spectrum through Fast Fourier Transform (FFT) to identify main frequency components.

[0093] 3. DEMON spectral analysis: Demodulate high-frequency modulated signals to extract low-frequency envelope information and generate low-frequency spectrum.

[0094] Analysis results show that hydrophone signals exhibit significant spectral peaks at 100Hz, 250Hz, 600Hz and 1300Hz, reflecting the main noise frequency components of UUV at 1800rpm speed. Microphone signals have a dense frequency distribution below 1500Hz, indicating the influence of air medium on signal propagation. The Structural Similarity Index (SSIM) result is 0.8682, indicating that microphone and hydrophone signals have high similarity in spectral structure, verifying the theoretical basis of dual-channel cooperative processing.

[0095] A.3 Simulation test results

[0096] DUAL-HydroNet model is compared with traditional noise reduction methods (spectral subtraction, improved spectral subtraction, least mean square (LMS), recursive least squares (RLS)) and other neural network methods (SEGAN, Denoiser) in terms of performance, with the main evaluation index being signal-to-noise ratio (SNR) improvement. Test results are as follows:

[0097] Single frequency signal (1kHz): DUAL-HydroNet achieves SNR improvement of 32.45dB under SNR=-1dB, which is 6.5-8.2dB better than traditional methods and 21.78dB better than SEGAN.

[0098] Impulse signal (1kHz): DUAL-HydroNet achieves SNR improvement of 34.12dB under SNR=3dB, significantly better than RLS (7.89dB) and Denoiser (28.45dB).

[0099] Frequency-modulated signal (50Hz-5kHz): Under the condition of SNR=5dB, DUAL-HydroNet achieves SNR improvement to 37.01dB, far exceeding the 6-8dB of traditional methods, 23.09dB of SEGAN, and 29.69dB of Denoiser.

[0100] Spectrum contrast analysis shows that DUAL-HydroNet can effectively preserve the energy distribution of low-frequency (50Hz-1kHz) and high-frequency (3kHz-5kHz) signals in wideband signal processing, avoiding the amplitude attenuation and phase distortion commonly seen in traditional methods. Time-domain waveform analysis shows that the model performs well in processing dynamic frequency-changing frequency-modulated signals, successfully capturing the periodicity and transient characteristics of the signals.

[0101] B. Pool experiment

[0102] B.1 Experimental setup

[0103] The pool experiment was conducted in a controlled underwater environment to verify the noise reduction performance of the DUAL-HydroNet model in actual scenarios. The experimental facility includes a pool with dimensions of 10m x 5m x 3m, equipped with a dual-channel data acquisition system (microphone and hydrophone). The UUV runs at a speed of 1800rpm, and signals are collected under the conditions of propeller on and off, simulating different noise scenarios. The microphone and hydrophone are sampled synchronously, with a sampling rate of 192000Hz and a collection time of 10 minutes.

[0104] Data preprocessing uses the amplitude threshold method to remove outliers, eliminating the influence of equipment interference or environmental noise. The processed signals are divided into internal signals (microphone collection) and external underwater signals (hydrophone collection), which are analyzed in time and frequency domains. The experimental equipment is strictly calibrated to ensure the stability and consistency of data collection.

[0105] B.2 Data analysis

[0106] Through time-frequency domain analysis, it is confirmed that the microphone and hydrophone signals have random fluctuations and periodic characteristics, consistent with the typical characteristics of UUV self-noise. DEMON spectrum analysis shows that the hydrophone signal has significant spectral peaks at 100Hz, 250Hz, 600Hz, and 1300Hz, consistent with the simulation results. The microphone signal has a dense frequency distribution below 1500Hz, reflecting the propagation characteristics of the air medium.

[0107] The DUAL-HydroNet model is applied to pool experiment data, and is used for noise reduction processing of single-frequency signals, pulse signals and frequency-modulated signals. Time-domain waveform and frequency-domain analysis show that the model effectively retains the signal energy distribution in the low frequency (50Hz-1kHz) and high frequency (3kHz-5kHz) range, avoiding the common amplitude and phase distortion in traditional methods. The channel-time sequence association attention mechanism (CSA) significantly enhances the model's ability to capture multi-scale time sequence features through adaptive average pooling and dilated convolution.

[0108] B.3 Experimental results

[0109] The pool experiment results further verify the conclusions of the simulation experiment. DUAL-HydroNet is superior to traditional noise reduction algorithms in all types of test signals, and the specific results are as follows:

[0110] Single-frequency signal (1kHz): Under the condition of propeller opening, DUAL-HydroNet achieves an SNR improvement of 31.89dB, which is 7.12dB better than the improved spectral subtraction and 22.45dB better than SEGAN.

[0111] Pulse signal (1kHz): Under the condition of propeller closing, DUAL-HydroNet achieves an SNR improvement of 33.67dB, which is 7.56dB better than RLS and 28.12dB better than Denoiser.

[0112] Frequency-modulated signal (50Hz-5kHz): In a complex noise environment, DUAL-HydroNet achieves an SNR improvement of 36.78dB, which is significantly better than the 6-8dB of traditional methods and the 23-30dB of other neural network methods.

[0113] Spectrum comparison and amplified waveform analysis show that DUAL-HydroNet has stability and robustness in wideband signal processing, especially in low and high frequency ranges. The model successfully retains the periodicity and dynamic change characteristics of the signal, significantly reducing background noise and providing an efficient solution for UUV underwater acoustic signal processing.

[0114] C. Conclusion

[0115] Through numerical simulation and pool experiment, the invention verifies the efficiency and robustness of the DUAL-HydroNet model in UUV self-noise suppression. The dual-channel cooperative processing method integrates microphone and hydrophone data, and combines the channel-time sequence association attention mechanism to achieve accurate feature extraction and wideband noise reduction.

[0116] The experimental results show that the system significantly improves the signal-to-noise ratio under various signal types and noise conditions, and is superior to the traditional method and the existing neural network method. The comprehensive results of numerical simulation and pool experiment prove the innovation and practical value of the application in underwater acoustic signal processing, and provide a reliable technical scheme for improving the concealment and combat effectiveness of UUV.

[0117] Therefore, the application adopts the above-mentioned UUV self-noise suppression method based on a dual-channel cooperative noise reduction neural network, adopts a dual-channel cooperative noise reduction network (DUAL-HydroNet), performs joint noise reduction through synchronously collected air channel and underwater acoustic channel signals, effectively integrates multi-source features, and is significantly superior to a traditional single-channel noise reduction scheme in noise suppression effect, thereby avoiding the problem of weakening of a target signal that may be caused by single-channel signal processing; through combination of the signal features of a hydrophone and a microphone, a channel-time sequence joint attention module (CSA) is used to effectively fuse cross-medium audio feature information, thereby enhancing the multi-channel signal processing capability. Compared with the traditional single-channel noise reduction method, the method can fully utilize the advantages from different signal sources, and improve the noise reduction precision and robustness.

[0118] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application and not to limit it, although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the application, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the application.

Claims

1. A UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, characterized in that, The method comprises the following steps: S1, using a microphone to collect air channel data of the environment where the UUV is located, and using a hydrophone to collect synchronous underwater acoustic channel data; S2, obtaining the hydrophone noisy signal and the microphone noisy signal of the UUV in the underwater environment, and performing a preprocessing operation; S3, inputting the preprocessed hydrophone noisy signal and the microphone noisy signal into a feature extraction module in parallel, and performing multi-scale feature extraction and fusion; S4, performing timing signal recovery on the multi-scale feature fused signal through a timing signal recovery module; In S2, the noise signal collected by the given hydrophone the noise signal collected by the microphone The preprocessing operation is specifically: The sampling rate is raised to 192kHz through double-stage Sinc interpolation upsampling, and the calculation formula is: （1）； wherein, is a 2x upsampling operator based on the Smith-Gossett algorithm, with a 56th order Hanning window weighted sinc function for anti-aliasing interpolation; In S4, the timing signal recovery module is composed of a bidirectional long short-term memory network combined with a decoder, and is used for timing prediction and recovery of the multi-scale feature fused signal. In the decoding stage, As the feature input, the fused multi-scale features As the skip connection, it is sent into the decoder to perform joint decoding operation and finally output the denoised signal.

2. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 1, characterized in that, In S3, the feature extraction module is composed of a weight-shared encoder Encoder combined with a channel-time joint attention CSA module, and adopts a double-path weight-shared encoder structure, specifically: The preprocessed hydrophone noisy signal and the microphone noisy signal are input into a 5-level convolutional encoder in parallel, and the mathematical expression is: （2）； In the formula, is the first layer shared weight parameter, is the hierarchical feature.

3. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 2, characterized in that, In S3, the multi-scale feature extraction and fusion process is: In each layer of the encoder, the preprocessed signal is first subjected to one-dimensional convolution for feature extraction, and then subjected to a ReLU activation function: （3）； wherein, is the input channel number, the first layer , is the hidden layer dimension, is the layer index, the convolution kernel size , the step size , the time resolution is compressed to by the stride operation; The extracted features are subjected to gate processing through a GLU gate linear unit: （4）； A multi-source feature fusion mechanism MSFusioner is introduced, and a channel-time joint attention CSA module is combined to first perform global information aggregation on the input audio features through adaptive average pooling; then, two one-dimensional convolution layers are used for feature dimension reduction and recovery to form channel attention weights, and the calculation formula of the channel attention is as follows: （5）； A timing attention path is established to capture the timing information in the audio signal, and the path reduces the channel dimension to 1 through a normal convolution, a convolution DConv with an extended convolution kernel and a one-dimensional convolution layer, and the calculation formula of the timing attention is as follows: （6）； Through the combination of channel attention and timing attention, the CSA module generates a weight matrix for fusing multi-scale features, which is used for weighting the input signal, and the output of the CSA module is: （7）； wherein, is an input hydrophone signal feature.

4. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 3, characterized in that, The double-channel cooperative denoising neural network comprises a feature extraction module and a timing signal recovery module, the feature extraction module is composed of a weight-shared encoder Encoder combined with a channel-time attention mechanism CSA, and is used for multi-source feature extraction and fusion of the air channel data collected by the microphone and the synchronous underwater acoustic channel data collected by the hydrophone; The timing signal recovery module is composed of a bidirectional long short-term memory network combined with a decoder, and is used for timing prediction and recovery of the multi-scale feature fused signal.

Images