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

Abstract

The invention belongs to the technical field of noise control, and particularly discloses a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, and the method comprises the following steps: collecting air channel data of an environment where a UUV comes out through a microphone, and collecting synchronous underwater acoustic channel data through a hydrophone; hydrophone noise signals and microphone noise signals of the UUV in the underwater environment are obtained, and preprocessing operation is carried out; the preprocessed hydrophone noisy signals and microphone noisy signals are input into a feature extraction module in parallel, and multi-scale feature extraction and fusion are carried out; and performing time sequence prediction and recovery on the signal after multi-scale feature fusion through a time sequence signal recovery module. According to the UUV self-noise suppression method based on the dual-channel collaborative noise reduction neural network, the noise reduction efficiency and accuracy are improved, the influence of environmental noise on the working performance of the UUV is reduced, and reliable guarantee is provided for stable operation of the UUV in a complex noise environment.

Description

Technical Field

[0001] The present invention relates to the field of noise control technology, and in particular to a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network. Background Art

[0002] In the application scenarios of underwater unmanned vehicles (UUVs), self-noise can seriously interfere with the detection performance of the sonar system, placing higher demands on noise control technology. Traditional noise control technology has many limitations: 1) Spectral subtraction estimates the noise spectrum of a noisy signal by using its amplitude or energy spectrum, then subtracts the noise spectrum from it to obtain a clean signal. This method is simple in principle and computationally fast, making it suitable for use in steady-state noise environments. However, the accuracy of noise spectrum estimation is limited, making it prone to artifacts in complex noise conditions and ineffective in handling dynamically changing noise.

[0003] 2) The least mean square error (LMS) algorithm estimates the gradient vector through the steepest descent method, which is simple and fast. It can adapt to non-stationary noise, but is sensitive to the power of the input signal, resulting in slow convergence and the problem of steady-state imbalance in time-varying noise environments.

[0004] 3) The recursive least squares (RLS) algorithm updates the filter weights recursively, which can converge quickly and adapt to time-varying noise. Its advantage is fast convergence, but it is computationally intensive and is suitable for non-real-time systems. It may also cause the algorithm to diverge when processing pathological autocorrelation matrices.

[0005] In summary, traditional noise control technology is difficult to meet actual engineering needs in the application of UUV self-noise reduction due to poor adaptability to complex environments, excessive computing resource requirements, or insufficient algorithm stability. It is urgent to explore more efficient and reliable noise reduction technology solutions. Summary of the Invention

[0006] The purpose of this invention is to provide a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, aiming to solve the problem that the interference signal generated by the self-noise of the UUV affects the detection performance of the sonar system in a complex underwater environment. Combined with 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 to effectively improve the noise reduction performance. This method realizes self-noise suppression in complex underwater acoustic environments by introducing real-time data from the air channel, significantly improving the sonar detection capability of the UUV.

[0007] To achieve the above objectives, the present invention provides a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, comprising the following steps: S1. Use a microphone to collect air channel data of the environment in which the UUV is located, and use a hydrophone to collect synchronous underwater acoustic channel data; S2, obtaining the noisy hydrophone signal and the noisy microphone signal of the UUV in the underwater environment and performing preprocessing operations; S3, inputting the pre-processed hydrophone noisy signal and microphone noisy signal into the feature extraction module in parallel to extract and fuse multi-scale features; S4. Through the time series signal recovery module, the signal after multi-scale feature fusion is predicted and restored.

[0008] Preferably, in S2, the noisy signal collected by the given hydrophone is Noisy signal collected by microphone

[0009] , the preprocessing operations are as follows: The sampling rate is increased to 192kHz through double-stage Sinc interpolation upsampling. The calculation formula is: (1); in, It is a 2x upsampling operator based on the Smith-Gossett algorithm, and uses a 56-order Hanning window weighted sinc function to implement anti-aliasing interpolation.

[0010] Preferably, in S3, the feature extraction module is composed of a weight-sharing encoder combined with a channel-temporal joint attention CSA module. The feature extraction module adopts a dual-path weight-sharing encoder structure. Specifically: The pre-processed hydrophone noisy signal and microphone noisy signal are input into the 5-stage convolution encoder in parallel. The mathematical expression is: (2); Where, For the Layer shared weight parameters, It is a hierarchical feature.

[0011] Preferably, the multi-scale feature extraction and fusion process in S3 is: In each encoder layer, the preprocessed signal first undergoes one-dimensional convolution for feature extraction, and then passes through the ReLU activation function: (3); in, is the number of input channels, the first layer , is the hidden layer dimension, n is the level index, the convolution kernel size , step length , the timing resolution is compressed to ; The extracted features are gated by the GLU gated linear unit: (4); The multi-source feature fusion mechanism MSFusioner is introduced, and combined with the channel-temporal joint attention CSA module, global information aggregation of input audio features is first performed through adaptive average pooling. Then, two one-dimensional convolutional layers are used to reduce and restore the features to form channel attention weights. The calculation formula of channel attention is as follows: (5); A temporal attention path is established to capture the temporal information in the audio signal. This path reduces the channel dimension to 1 through a normal convolution, a convolution DConv with a dilated convolution kernel, and a one-dimensional convolution layer. The calculation formula of temporal attention is as follows: (6); By combining channel attention and temporal attention, the CSA module generates a weight matrix that integrates multi-scale features and is used to weight the input signal. The output of the CSA module is: (7); in, F 1 is the input hydrophone signal.

[0012] Preferably, in S4, the signal after multi-scale feature fusion is subjected to time series prediction and recovery by a time series signal recovery module, the time series signal recovery module being composed of a bidirectional long short-term memory network combined with a decoder; In the decoding stage, As feature input, the fused multi-scale features It is sent to the decoder as a skip connection for joint decoding operation, and finally outputs the noise-reduced signal.

[0013] Preferably, the dual-channel collaborative denoising neural network includes a feature extraction module and a time series signal recovery module. The feature extraction module is composed of a weight-sharing encoder combined with a channel-time series attention mechanism CSA, which is used to extract and fuse multi-source features of the air channel data collected by the microphone and the underwater acoustic channel data collected synchronously by the hydrophone; The time series signal recovery module consists of a bidirectional long short-term memory network combined with a decoder, which is used to perform time series prediction and recovery on the signal after multi-scale feature fusion.

[0014] Therefore, the present invention adopts the above-mentioned UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, and the beneficial effects are as follows: (1) Significantly improved noise reduction performance: Compared to traditional single-channel noise reduction methods, the dual-channel collaborative noise reduction network (DUAL-HydroNet) proposed in this paper can more effectively capture noise characteristics by simultaneously processing the air-medium microphone signal and the water-medium hydrophone signal, thereby significantly improving the noise reduction effect. With the synergistic effect 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.

[0015] (2) Enhanced signal recovery stability: By introducing a multi-source feature fusion mechanism (MSFusioner) and combining it with a channel-temporal joint attention module (CSA), the present invention can deeply model and fuse multi-dimensional feature information during the noise reduction process. Compared with traditional single-signal noise reduction methods, this technology not only improves the time domain waveform and spectrum 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.

[0016] (3) Effective integration of cross-media features: This invention innovatively expands the feature dimensions that can be used for noise reduction by synchronously collecting signals from air-medium microphones and water-medium hydrophones, and uses deep neural networks to fuse the features of air and water acoustic signals, thereby achieving effective integration of cross-media features. This not only enhances the recognition ability of noise features, but also significantly improves the noise reduction performance through the fusion mechanism.

[0017] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a dual-channel collaborative noise reduction network structure diagram of an embodiment of a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network of the present invention; Figure 2 Schematic diagram of a feature extraction module of an embodiment of a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network of the present invention; Figure 3 This 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 denoising neural network of the present invention; Figure 4 This is a structural diagram of a timing signal recovery module of an embodiment of a UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network of the present invention. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0020] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0021] The present invention combines deep learning technology to provide a dual-channel collaborative noise reduction neural network (DUAL-HydroNet). Figure 1 As shown in , the network includes a feature extraction module and a timing signal recovery module, as Figure 2 As shown, the feature extraction module is composed of a weight-sharing encoder Encoder combined with Figure 3 The channel-temporal attention CSA module shown is used to extract and fuse multi-source features from the air channel data collected by the microphone and the underwater acoustic channel data collected synchronously by the hydrophone. Figure 4 As shown in the figure, the time series signal recovery module consists of a bidirectional long short-term memory network combined with a decoder, which is used to perform time series prediction and recovery on the signal after multi-scale feature fusion.

[0022] A UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network includes the following steps: S1. Use a microphone to collect air channel data of the environment in which the UUV is located, and use a hydrophone to collect synchronous underwater acoustic channel data.

[0023] S2. Obtain the noisy hydrophone signal and the noisy microphone signal of the UUV in the underwater environment and perform preprocessing operations.

[0024] In one possible implementation, given a noisy signal acquired by a hydrophone Noisy signal collected by microphone , the preprocessing operations are as follows: The sampling rate is increased to 192kHz through double-stage Sinc interpolation upsampling. The calculation formula is: (1); in, It is a 2x upsampling operator based on the Smith-Gossett algorithm, and uses a 56-order Hanning window weighted sinc function to achieve anti-aliasing interpolation. After this operation, the time domain signal length is extended to , effectively improving the resolution of high-frequency components.

[0025] S3. The pre-processed hydrophone noisy signal and microphone noisy signal are input into the feature extraction module in parallel to extract and fuse multi-scale features.

[0026] The feature extraction module adopts a dual-path weight sharing encoder structure, specifically: The pre-processed hydrophone noisy signal and microphone noisy signal are input into the 5-stage convolution encoder in parallel. The mathematical expression is: (2); Where, For the Layer shared weight parameters, It is a hierarchical feature.

[0027] In each encoder layer, the preprocessed signal first undergoes one-dimensional convolution for feature extraction, and then passes through the ReLU activation function. The specific process is shown in the following formula: (3); in, is the number of input channels, the first layer , is the hidden layer dimension, is the level index, the convolution kernel size , step length , the timing resolution is compressed to .

[0028] The extracted features are then gated by the GLU gated linear unit: (4).

[0029] By introducing a gating mechanism to adjust and filter the input features, the feature response strength of different channels is automatically adjusted to enhance useful information and suppress invalid information.

[0030] The multi-source feature fusion mechanism MSFusioner is introduced, combined with the channel-temporal joint attention (CSA) module to first aggregate global information from the input audio features through adaptive average pooling. Next, two one-dimensional convolutional layers are used to reduce and restore the features, forming channel attention weights. This process can be viewed as enhancing the response of important channels by learning global dependencies between features. The channel attention is calculated as follows: (5); To capture the timing information in audio signals, the CSA module establishes an independent temporal attention path. This path is implemented through the combined operation of three convolutional layers: a standard convolution, a convolution with a dilated kernel (DConv), and a one-dimensional convolution layer to reduce the channel dimension to 1. This path is designed based on the convolution decomposition strategy. The dilated convolution effectively expands the receptive field, enabling the model to capture longer temporal dependencies.

[0031] The calculation formula of temporal attention is as follows: (6); By combining the above-mentioned channel attention and temporal attention, the CSA module can generate a weight matrix that integrates multi-scale features to weight the input signal. That is, the output of the CSA module is: (7); in, F 1 is the input hydrophone signal.

[0032] S4. Through the time series signal recovery module, the signal after multi-scale feature fusion is predicted and restored.

[0033] In the decoding stage, As feature input, the fused multi-scale features It is sent to the decoder as a skip connection for joint decoding operation, and finally outputs the noise-reduced signal.

[0034] Numerical simulation experiments and pool experiments: To verify the performance of the proposed dual-channel collaborative noise reduction system for suppressing self-noise in unmanned underwater vehicles (UUVs), numerical simulations and water tank experiments were conducted to comprehensively evaluate the system's noise reduction effectiveness under various signal types, signal-to-noise ratio (SNR) conditions, and actual underwater environments. This experiment integrates signals collected by microphones and hydrophones using the DUAL-HydroNet model and incorporates the channel-sequential connection attention (CSA) mechanism to achieve efficient noise suppression. The following details the experimental setup, process, simulation test results, and water tank experiment results, highlighting the system's technical advantages.

[0035] A. Numerical Simulation Experiment

[0036] A.1 Experimental Setup

[0037] Numerical simulation experiments are designed to verify the noise reduction performance of the DUAL-HydroNet model under different signal-to-noise ratios (SNRs) and signal types. The experimental dataset includes the following three signal types, representing the self-noise characteristics of UUVs under typical operating conditions: Single-frequency signal: The fixed frequency is 1kHz, simulating the single-frequency noise generated by UUV propellers or mechanical components.

[0038] Pulse signal: The frequency is 1kHz, simulating transient impact noise.

[0039] FM signal: The frequency range is 50Hz to 5kHz, simulating the broadband noise of UUV under complex operating conditions.

[0040] To simulate different noise environments, the dataset was divided into three groups based on signal-to-noise ratios: 5dB, 3dB, and -1dB. Each signal type contained 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 independence between model training and evaluation. Data preprocessing included amplitude normalization, downsampling, and segmentation to generate a standardized format suitable for neural network input. Downsampling used the Smith-Gossett algorithm, upsampling the sampling rate from 48,000 Hz to 192,000 Hz to improve the resolution of high-frequency components.

[0041] The experimental platform uses a high-performance computing cluster equipped with NVIDIA A100 GPUs, running the Python 3.8 environment, and implementing the DUAL-HydroNet model using the PyTorch framework. Model training uses a batch size of 32, a learning rate of 0.001, and the Adam optimizer for 50 epochs to ensure convergence.

[0042] A.2 Signal Processing and Analysis Methods

[0043] Signal processing uses time-domain and frequency-domain analysis to extract the characteristics of the UUV's self-noise signals captured by microphones and hydrophones, characterizing the propagation characteristics of acoustic signals in air and underwater environments. DEMON (demodulated noise) spectrum analysis is used to extract low-frequency components from high-frequency modulated signals and identify key frequency components relevant to noise reduction. The analysis process is as follows: 1. Time domain analysis: Perform time domain waveform analysis on microphone and hydrophone signals to observe the random fluctuations and periodic characteristics of the signals.

[0044] 2. Frequency domain analysis: Generate a spectrum through fast Fourier transform (FFT) to identify the main frequency components.

[0045] 3.DEMON spectrum analysis: Demodulate the high-frequency modulated signal, extract the low-frequency envelope information, and generate a low-frequency spectrum diagram.

[0046] The analysis results show that the hydrophone signal exhibits significant spectral peaks at 100Hz, 250Hz, 600Hz, and 1300Hz, reflecting the primary noise frequency components of the UUV at 1800rpm. The microphone signal has a dense distribution below 1500Hz, indicating the influence of the air medium on signal propagation. The calculated structural similarity index (SSIM) is 0.8682, indicating a high degree of similarity in the spectral structure of the microphone and hydrophone signals, validating the theoretical basis of dual-channel collaborative processing.

[0047] A.3 Simulation Test Results

[0048] The DUAL-HydroNet model was compared with traditional denoising methods (spectral subtraction, improved spectral subtraction, least mean squares (LMS), recursive least squares (RLS)) and other neural network methods (SEGAN, Denoiser). The main evaluation metric was signal-to-noise ratio (SNR) improvement. The test results are as follows: Single-frequency signal (1kHz): Under the condition of SNR=-1dB, DUAL-HydroNet achieves an SNR improvement of 32.45dB, which is better than the 6.5–8.2dB of traditional methods and 21.78dB of SEGAN.

[0049] Pulse signal (1kHz): DUAL-HydroNet achieves an SNR improvement of 34.12dB under the SNR=3dB condition, significantly better than RLS (7.89dB) and Denoiser (28.45dB).

[0050] FM signal (50Hz–5kHz): At SNR=5dB, DUAL-HydroNet achieves an SNR improvement of 37.01dB, far exceeding the 6–8dB of traditional methods, 23.09dB of SEGAN, and 29.69dB of Denoiser.

[0051] Comparative spectrum analysis demonstrates that DUAL-HydroNet effectively preserves the energy distribution of both low-frequency (50Hz–1kHz) and high-frequency (3kHz–5kHz) signals in wideband signal processing, avoiding the amplitude attenuation and phase distortion common to traditional methods. Time-domain waveform analysis demonstrates that the model performs exceptionally well when processing frequency-modulated signals with dynamic frequency variations, successfully capturing both the periodic and transient characteristics of the signals.

[0052] B. Pool Experiment

[0053] B.1 Experimental Setup

[0054] Tank experiments were conducted in a controlled underwater environment to validate the DUAL-HydroNet model's noise reduction performance in real-world scenarios. The experimental facility consisted of a 10m × 5m × 3m tank equipped with a dual-channel data acquisition system (microphone and hydrophone). The UUV operated at 1800 rpm, collecting signals with the propellers both on and off to simulate different noise scenarios. The microphone and hydrophone sampled synchronously at a uniform 192,000 Hz sampling rate for 10 minutes.

[0055] Data preprocessing uses an amplitude thresholding method to remove outliers and eliminate the effects of equipment interference and environmental noise. The processed signals are divided into internal signals (collected by microphones) and external underwater signals (collected by hydrophones), and time- and frequency-domain analysis is performed separately. The experimental equipment is rigorously calibrated to ensure stable and consistent data acquisition.

[0056] B.2 Data Analysis

[0057] Time-frequency analysis confirmed that the microphone and hydrophone signals exhibited random fluctuations and periodicity, consistent with typical UUV self-noise characteristics. DEMON spectrum analysis revealed significant peaks in the hydrophone signal at 100Hz, 250Hz, 600Hz, and 1300Hz, consistent with simulation results. The microphone signal exhibited a dense distribution of frequencies below 1500Hz, reflecting the propagation characteristics of the air medium.

[0058] The DUAL-HydroNet model was applied to water tank experimental data to denoise single-frequency, pulse, and frequency-modulated signals. Time-domain waveform and frequency-domain analysis demonstrated that the model effectively preserved signal energy distribution in both the low-frequency (50Hz–1kHz) and high-frequency (3kHz–5kHz) ranges, avoiding the amplitude and phase distortion common in traditional methods. The channel-series attention mechanism (CSA) significantly enhanced the model's ability to capture multi-scale temporal features through adaptive average pooling and dilated convolution.

[0059] B.3 Experimental Results

[0060] The results of the water tank experiments further validated the conclusions of the simulation experiments. DUAL-HydroNet outperformed traditional noise reduction algorithms in all test signal types. The specific results are as follows: Single-frequency signal (1kHz): With the propellers on, DUAL-HydroNet achieves an SNR improvement of 31.89dB, which is better than the 7.12dB of improved spectral subtraction and the 22.45dB of SEGAN.

[0061] Pulse signal (1kHz): With the propellers turned off, DUAL-HydroNet achieves an SNR improvement of 33.67dB, which is better than RLS's 7.56dB and Denoiser's 28.12dB.

[0062] Frequency-modulated signals (50Hz–5kHz): In complex noise environments, DUAL-HydroNet achieves an SNR improvement of 36.78dB, significantly better than the 6–8dB of traditional methods and the 23–30dB of other neural network methods.

[0063] Spectral comparison and magnified waveform analysis demonstrate the stability and robustness of DUAL-HydroNet in wideband signal processing, particularly in the low- and high-frequency ranges. The model successfully preserves the periodicity and dynamic characteristics of the signal, significantly reducing background noise and providing a highly efficient solution for underwater acoustic signal processing for UUVs.

[0064] C. Conclusion

[0065] Through numerical simulations and water tank experiments, this paper demonstrates the efficiency and robustness of the DUAL-HydroNet model in suppressing UUV self-noise. The dual-channel collaborative processing approach integrates microphone and hydrophone data and incorporates a channel-time-linked attention mechanism to achieve precise feature extraction and broadband noise reduction.

[0066] Experimental results demonstrate that the system significantly improves the signal-to-noise ratio across a wide range of signal types and noise conditions, outperforming both traditional and existing neural network approaches. Numerical simulations and tank experiments demonstrate the innovative and practical value of this invention in underwater acoustic signal processing, providing a reliable technical solution for enhancing the stealth and combat effectiveness of UUVs.

[0067] Therefore, the present invention adopts the aforementioned UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network. This dual-channel collaborative noise reduction network (DUAL-HydroNet) performs joint noise reduction on synchronously collected air and hydroacoustic channel signals. By effectively integrating multi-source features, the method significantly outperforms traditional single-channel noise reduction schemes in noise suppression, avoiding the target signal weakening problem that may occur due to single-channel signal processing. By combining the signal characteristics of the hydrophone and microphone and utilizing the channel-temporal joint attention module (CSA), it can effectively fuse audio feature information across media, enhancing the multi-channel signal processing capabilities. Compared with traditional single-channel noise reduction methods, this method can fully utilize the advantages of different signal sources, improving noise reduction accuracy and robustness.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network, characterized in that: The following steps are involved: S1. Use a microphone to collect air channel data of the environment in which the UUV is located, and use a hydrophone to collect synchronous underwater acoustic channel data; S2, obtaining the noisy hydrophone signal and the noisy microphone signal of the UUV in the underwater environment and performing preprocessing operations; S3, inputting the pre-processed hydrophone noisy signal and microphone noisy signal into the feature extraction module in parallel to extract and fuse multi-scale features; S4. Through the time series signal recovery module, the signal after multi-scale feature fusion is predicted and restored.

2. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 1 is characterized in that: In S2, given the noisy signal collected by the hydrophone Noisy signal collected by microphone , the preprocessing operations are as follows: The sampling rate is increased to 192kHz through double-stage Sinc interpolation upsampling. The calculation formula is: （1）； in, It is a 2x upsampling operator based on the Smith-Gossett algorithm, and uses a 56-order Hanning window weighted sinc function to implement anti-aliasing interpolation.

3. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 2 is characterized in that: In S3, the feature extraction module consists of a weight-sharing encoder combined with a channel-temporal joint attention (CSA) module. The feature extraction module adopts a dual-path weight-sharing encoder structure. Specifically: The pre-processed hydrophone noisy signal and microphone noisy signal are input into the 5-stage convolution encoder in parallel. The mathematical expression is: （2）； Where, For the Layer shared weight parameters, It is a hierarchical feature.

4. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 3 is characterized in that: The extraction and fusion process of multi-scale features in S3 is as follows: In each encoder layer, the preprocessed signal first undergoes one-dimensional convolution for feature extraction, and then passes through the ReLU activation function: （3）； in, is the number of input channels, the first layer , is the hidden layer dimension, is the level index, the convolution kernel size , step length , the timing resolution is compressed to ; The extracted features are gated by the GLU gated linear unit: （4）； The multi-source feature fusion mechanism MSFusioner is introduced, and combined with the channel-temporal joint attention CSA module, global information aggregation of input audio features is first performed through adaptive average pooling. Then, two one-dimensional convolutional layers are used to reduce and restore the features to form channel attention weights. The calculation formula of channel attention is as follows: （5）； A temporal attention path is established to capture the temporal information in the audio signal. This path reduces the channel dimension to 1 through a normal convolution, a convolution DConv with a dilated convolution kernel, and a one-dimensional convolution layer. The calculation formula of temporal attention is as follows: （6）； By combining channel attention and temporal attention, the CSA module generates a weight matrix that integrates multi-scale features and is used to weight the input signal. The output of the CSA module is: （7）； in, F 1 is the input hydrophone signal.

5. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 4 is characterized in that: In S4, the time series signal recovery module is used to perform time series prediction and recovery on the signal after multi-scale feature fusion. The time series signal recovery module consists of a bidirectional long short-term memory network combined with a decoder. In the decoding stage, As feature input, the fused multi-scale features It is sent to the decoder as a skip connection for joint decoding operation, and finally outputs the noise-reduced signal.

6. The UUV self-noise suppression method based on a dual-channel collaborative noise reduction neural network according to claim 5 is characterized in that: The dual-channel collaborative denoising neural network includes a feature extraction module and a time series signal recovery module. The feature extraction module consists of a weight-sharing encoder combined with a channel-time series attention mechanism (CSA). It is used to extract and fuse multi-source features from the air channel data collected by the microphone and the underwater acoustic channel data collected synchronously by the hydrophone. The time series signal recovery module consists of a bidirectional long short-term memory network combined with a decoder, which is used to perform time series prediction and recovery on the signal after multi-scale feature fusion.

Images