Pulse signal reconstruction method based on deep learning in non-Gaussian environment

Abstract

The invention discloses a pulse signal reconstruction method based on deep learning in a non-Gaussian environment, and relates to a pulse signal reconstruction method based on deep learning. The invention provides a pulse signal reconstruction method based on deep learning in a non-Gaussian environment, and aims to solve the problems that the existing marine environmental noise often has a non-Gaussian statistical characteristic, and a traditional method based on Gaussian distribution hypothesis is low in pulse signal reconstruction accuracy under the background of non-Gaussian interference and low input signal to interference plus noise ratio. The pulse signal reconstruction method based on deep learning in the non-Gaussian environment comprises the following specific steps: constructing a training set; constructing a deep neural network model; training the constructed deep neural network model based on the training set to obtain a trained deep neural network model; and inputting the received underwater sound pulse signal into the trained deep neural network model, and outputting a reconstructed pulse signal by the trained deep neural network model. The method is applied to the field of pulse signal reconstruction.

Description

Technical Field

[0001] The present invention relates to a pulse signal reconstruction method based on deep learning. Background Art

[0002] The marine environment contains a variety of interference sources, such as the sounds of marine organisms such as fish and shrimp, ship navigation, and seabed drilling. The interference signals generated by these high-energy interference sources often have non-Gaussian statistical characteristics [1-4] ([1] Y. Li, X. Ma, L. Wang, and Y. Liu, "Using deep learning to de-noise and reconstruct pulse signals in non-Gaussian environment," Journal of Applied Acoustics, vol. 40, no. 1, pp. 131-141, Jan., 2021, doi: 10.11684 / j.issn.1000-310X.2021.01.016. [2] X. Mo, H. Wen, and Y. Yang, "A parameter estimation method of α stable distribution and its application in the statistical modeling of ice-generated noise," ActaAcustica, vol.48, no.2, pp.319-326, Mar., 2023, doi:10.15949 / j.cnki.0371-0025.2023.02.001. [3] G. Song, X. Guo, and L. Ma, "αstabledistribution model in oceanambient noise," Acta Acustica, vol.44, no.2, pp.177-188, Mar., 2019, doi:10.15949 / j.cnki.0371-0025.2019.02.004. [4] F. Traverso, G. Vernazza, and A. Trucco, "Simulation of non-White and non-Gaussian underwater ambient noise,” Oceans-Yeosu, 2012, pp. 1-10.) and leads to a significant decrease in the signal-to-noise ratio of the received signal. Since the statistical characteristics of non-Gaussian interference significantly deviate from the normal distribution and are accompanied by non-stationary characteristics, this causes its second-order moment to diverge, resulting in the traditional method based on the Gaussian distribution assumption [5-6] ([5] Z. Zhao, Q. Li, Z. Xia, and D.Shang, "A Single-HydrophoneCoherent-Processing Method for Line-Spectrum Enhancement," Remote Sens., vol.15, no.3, pp.659, Jan., 2023, doi:10.3390 / rs15030659. [6] C.Xing, Y.Wu, L.Xie, and D.Zhang, "Asparse dictionary learning-based denoising method for underwateracoustic sensors,”Appl.Acoust.,vol.180,pp.108140.1–108140.13,Apr.,2021,doi:10.1016 / j.apacoust.2021.108140.)When faced with non-Gaussian interference, its weight iteration process may not converge, and it is impossible to effectively achieve interference suppression and signal reconstruction. The signal processing performance of traditional methods is significantly reduced[7-9]([7]H.Yang,Y.Cheng,and G.Li,“A denoising method for ship radiated noise based onSpearman variational mode decomposition,spatial-dependence recurrence sampleentropy,improved wavelet threshold denoising,and Savitzky-Golay filter," Alex.Eng.J., vol.60, no.3, pp.3379-3400, Jan., 2021, doi:10.1016 / j.aej.2021.01.055. [8] J.Wang, J.Li, S.Yan, W.Shi, X.Yang, Y.Guo, and T.Gulliver, "Anovel underwateracoustic signal denoising algorithm for Gaussian / Non-Gaussian impulsivenoise," IEEE Trans.Veh.Technol., vol.70, no.1, pp.429-445, Jan., 2021, doi:10.1109 / TVT.2020.3044994. [9] Y. Li, and L.Wang, “A novel noise reduction technique for underwater acoustic signals based on complete ensemble empirical mode decomposition with adaptive noise, minimum mean square variance criterion and least mean square adaptive filter,” Def. Technol., vol. 16, no. 3, pp. 543-554, Jun., 2020, doi: 10.1016 / j.dt.2019.07.020.). In addition, non-Gaussian interference has a more significant spike pulse characteristic in the time waveform, which is similar to the characteristics of the target pulse signal, making the reconstruction of the pulse signal more difficult. How to make full use of the characteristics of underwater acoustic pulse signals and solve the problem of pulse signal reconstruction under the background of low input signal-to-interference-noise ratio and non-Gaussian interference remains to be solved. .

[0003] Deep neural networks have excellent automatic feature extraction and learning capabilities. They can deeply explore and understand the potential structures and features in complex signals. Through multiple layers of nonlinear transformations, they can capture fine information that is difficult to capture with traditional methods. In recent years, signal denoising and reconstruction methods based on deep neural networks have been widely used [10-15] (

[10] X. Wang, Y. Zhao, X. Teng, and W. Sun, "A stacked convolutional sparse denoising autoencoder model for underwater heterogeneous information data," Appl. Acoust., vol. 167, Oct., 2020, Art. no. 107391, doi: 10.1016 / j.apacoust.2020.107391.

[11] Y. Song, F. Liu, and T. Shen, "A novel noise reduction technique for underwater acoustic signals based on dual-path recurrent neural network," IET Commun., vol. 17, no. 2, pp. 135-144, Oct., 2022, doi: 10.1049 / cmu2.12518.

[12] X. Zhou, and K.Yang, "Adenoising representationframework for underwater acoustic signal recognition," J.Acoust.Soc.Am., vol.147, no.4, pp.EL377-EL383, Apr., 2020, doi:10.1121 / 10.0001130.

[13] D.Ju, C.Chi, Z.Li, Y.Li, C.Zhang, and H.Huang, "Deep-learning-based line enhancer for passivesonar systems," IET Radar Sonar Nav., vol.16, no.3, pp.589-601, Dec., 2021, doi:10.1049 / rsn2.12205.

[14] J.Yin, W.Luo, L.Li, X.Han, L.Guo, and J.Wang, “Enhancementof underwater acoustic signal based on denoising automatic-encoder,” Journal on Communications, vol. 40, no. 10, pp. 119-126, Oct., 2019, doi: 10.11959 / j.issn.1000-436x.2019181.

[15] R. Zaheer, I. Ahmad, Q. Viet Phung, and D. Habibi, “Blind Source Separation and Denoising of Underwater Acoustic Signals,” IEEE Access, vol. 12, pp. 80208-80222, Jun., 2024, doi: 10.1109 / ACCESS.2024.3410276.) is developing rapidly. Denoising techniques based on deep learning are mainly divided into two categories: direct mapping and mask separation methods. The direct mapping method takes the pure signal as the learning target and directly recovers the pure signal from the noisy signal through an end-to-end neural network model. For example, the reference

[16] (

[16] AANair, and K.Koishida, "Cascaded time+time-frequency Unet for speech enhancement: jointly addressing clipping, codec distortions, and gaps," in ICASSP, 2021, pp. 7153-7157.) applies the UNet network

[17] (

[17] O. Ronneberger, P. Fischer, and T. Brox, "U-Net: convolutional networks for biomedical image segmentation," in MICCAI, 2015, pp. 234-241.) to signal denoising. The U-shaped network structure with jump connections effectively integrates shallow and deep features, significantly improving the signal denoising effect. Furthermore, the CBDNet network

[18] (

[18] S. Guo, Z. Yan, K. Zhang, W. Zuo, and M. Wang, “Toward convolutional blind denoising of realphotographs,” in CVPR, 2019, pp. 1712-1722.) combines the fully convolutional network with the UNet network. The former is responsible for estimating the noise component in the input data, while the latter uses the noise estimation value obtained by the fully convolutional network for denoising. The RIDNet network

[19] (

[19] S. Anwar, and N. Barnes, "Real image denoising with feature attention," in ICCV, 2019, pp. 3155-3164.) introduces a residual structure for feature extraction and combines it with the attention mechanism to denoise the input data, demonstrating the potential of the attention mechanism in signal processing. The mask separation method focuses on estimating the mask between the noisy signal and the pure signal, using the mask to denoise the input signal and restore a clearer acoustic signal. For example, reference

[20] (

[20] Y. Song, F. Liu, and T. Shen, “Method of Underwater Acoustic Signal Denoising Based on Dual-Path Transformer Network,” IEEE Access, vol. 12, pp. 81483-81494, Nov., 2024, doi: 10.1109 / ACCESS.2022.3224752.) uses a dual-path Transformer network to construct a mask, and uses the sequence modeling capability of the Transformer network to capture long-term dependencies, successfully achieving underwater acoustic signal denoising in the time domain. In order to further improve the denoising performance of the model, the literature

[21] (

[21] J. Fan, J. Yang, X. Zhang, and C. Zheng, "Monaural speech enhancement using U-net fused with multi-head self-attention," Acta Acustica, vol. 47, no. 6, pp. 703-716, Nov., 2022, doi: 10.15949 / j.cnki.0371-0025.2022.06.007.) proposed the TUNet model, which combines the multi-scale feature fusion capability of the UNet network with the dual-path Transformer structure of the multi-head attention mechanism, and uses the Transformer structure to construct a mask separation structure to achieve single-channel signal denoising based on the time domain. . Summary of the invention

[0004] The purpose of the present invention is to solve the problem that the existing ocean environmental noise often has non-Gaussian statistical characteristics, and the traditional method based on the Gaussian distribution assumption has low accuracy in pulse signal reconstruction under the background of non-Gaussian interference and low input signal-to-interference-noise ratio, and proposes a pulse signal reconstruction method based on deep learning in a non-Gaussian environment.

[0005] The specific process of the pulse signal reconstruction method based on deep learning in a non-Gaussian environment is as follows:

[0006] Step 1: Construct a training set;

[0007] Step 2: Build a deep neural network model;

[0008] Step 3: Train the constructed deep neural network model based on the training set to obtain a trained deep neural network model;

[0009] Step 4: Input the received underwater acoustic pulse signal into the trained deep neural network model, and the trained deep neural network model outputs the reconstructed pulse signal.

[0010] The beneficial effects of the present invention are:

[0011] Affected by natural phenomena or human interference, marine environmental noise often has non-Gaussian statistical characteristics. In view of the problem that the performance of traditional methods based on the Gaussian distribution assumption deteriorates under the background of non-Gaussian interference and low input signal-to-noise ratio, the present invention proposes a dual-path Transformer network signal reconstruction method based on the time-frequency domain attention mechanism. The network of the present invention introduces dilated convolutional layers and jump connection structures in the encoder and decoder structures to fuse the shallow and deep features in the input data, and uses the attention mechanism combined with the Transformer network structure to obtain a separator to realize mask feature extraction. In view of the difference in distribution characteristics of pulse signals and non-Gaussian interference in the time domain and frequency domain, the network of the present invention reconstructs the pulse signal on a global and local scale. The results show that when the input signal-to-noise ratio is low, the network of the present invention has good performance in the reconstruction task of pulse signals under non-Gaussian interference and Gaussian noise environments, and shows good generalization. The simulation results show that when the input signal-to-interference ratio and signal-to-noise ratio are -10dB, the signal-to-interference-to-noise ratio gain of the network of the present invention is 38.65dB and 37.32dB respectively, and the signal-to-distortion ratio is 5.80dB and 4.80dB respectively, both of which are better than the comparison algorithm. The experimental results show that the method of the present invention can effectively realize the reconstruction of underwater acoustic pulse signals in actual physical scenes, which reflects the feasibility and effectiveness of the method of the present invention.

[0012] In order to further improve the signal reconstruction performance, the present invention proposes a dual-path Transformer network signal reconstruction method based on the time-frequency domain attention mechanism to solve the signal reconstruction problem of underwater acoustic pulse signals under the background of non-Gaussian interference with low input signal-to-interference-noise ratio. In the time-frequency domain, classic underwater acoustic pulse signals such as single rectangular pulse (Continuous Wave, CW) signals, linear frequency modulation (Linear Frequency Modulation, LFM) pulse signals and hyperbolic frequency modulation pulse signals (Hyperbolic Frequency Modulation, HFM) all have limited pulse width and bandwidth, and valuable feature information is concentrated in a specific time-frequency domain area. Unlike the former, non-Gaussian pulse interference has a narrower pulse width and a wider bandwidth. The present invention makes full use of the above signal characteristics to design a dual-path Transformer network based on the time-frequency domain attention mechanism, aiming to use the network of the present invention to distinguish non-Gaussian interference from pulse signals in the time-frequency domain, and finally achieve high-precision pulse signal reconstruction under the background of non-Gaussian interference. The innovative points of the present invention are as follows:

[0013] (1) Traditional CNN networks are limited by the limited receptive field of the convolutional layer and can often only capture local information. Therefore, the network of the present invention uses dilated convolution combined with the Transformer network structure to expand the receptive field and better capture the global information in the input data.

[0014] (2) Due to the information concentration characteristics of underwater acoustic pulse signals in the time and frequency domain, the network of the present invention pays more attention to the application of the attention mechanism. The present invention designs a new attention mechanism to increase the model's time and frequency domain analysis capabilities and the ability to extract signal features in specific time and frequency domain regions, and effectively suppress non-Gaussian interference.

[0015] (3) Compared with the comparative method, the method of the present invention aims to target the time-frequency characteristics of the pulse signal, and utilizes the dilated convolution, attention mechanism, and jump connection structure combined with the Transformer network structure to simultaneously improve the network's information extraction capabilities at both global and local scales, so as to better realize the signal reconstruction of the underwater acoustic pulse signal under the background of non-Gaussian interference.

[0016] The present invention simulates two common pulse signals, namely CW signal and LFM pulse signal, and superimposes them with α stable distribution interference and Gaussian white noise to obtain a training set for network training, and uses the simulated HFM pulse signal to verify the generalization of the network of the present invention. The present invention verifies the feasibility and effectiveness of the network in the task of pulse signal reconstruction through simulation data and experimental data. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the present invention;

[0018] Figure 2 It is a physical scene model diagram;

[0019] Figure 3 Construct a deep neural network model structure diagram for the present invention;

[0020] Figure 4 The structure diagram of the encoder and decoder of the present invention, Skip Connection is a skip connection structure;

[0021] Figure 5 This is a structural diagram of the skip connection structure of the present invention, where Skip_Connection_Input is the input of the skip connection structure, and Skip_Connection_Output is the output of the skip connection structure;

[0022] Figure 6 This is a structural diagram of the DPSAT Block of the present invention;

[0023] Figure 7 It is the Transformer structure diagram;

[0024] Figure 8 This is the structural diagram of the attention mechanism;

[0025] Fig. 9 The signal processing result diagram when the desired signal is a CW signal, (a) input signal time domain, (b) desired signal time domain, (c) network reconstructed signal time domain of the present invention;

[0026] Fig.10 The signal processing result diagram when the desired signal is an LFM signal, (a) input signal time domain, (b) desired signal time domain, (c) network reconstructed signal time domain of the present invention;

[0027] Fig.11 The signal processing result diagram when the desired signal is an HFM signal, (a) input signal time domain, (b) desired signal time domain, (c) network reconstructed signal time domain of the present invention;

[0028] Fig.12 is a curve diagram showing the relationship between input signal-to-interference ratio and output signal-to-interference-noise ratio gain;

[0029] Fig.13 It is the relationship curve between input signal-to-interference ratio and signal distortion ratio;

[0030] Fig.14 is a curve diagram showing the relationship between input signal-to-noise ratio and output signal-to-interference-noise ratio gain;

[0031] Fig.15 It is the relationship curve diagram of input signal-to-noise ratio and signal-to-distortion ratio;

[0032] Fig.16It is the time domain waveform of the measured data;

[0033] Fig.17 The signal processing result diagram of the measured data, (a) the signal time domain reconstructed by the network of the present invention, (b) the signal time domain reconstructed by UNet, (c) the signal time domain reconstructed by CBDNet, and (d) the signal time domain reconstructed by RIDNet. DETAILED DESCRIPTION

[0034] Specific implementation method 1: The specific process of the pulse signal reconstruction method based on deep learning in a non-Gaussian environment in this implementation method is as follows:

[0035] Step 1: Construct a training set;

[0036] Step 2: Build a deep neural network model;

[0037] Step 3: Train the constructed deep neural network model based on the training set to obtain a trained deep neural network model;

[0038] Step 4: Input the received underwater acoustic pulse signal into the trained deep neural network model, and the trained deep neural network model outputs the reconstructed pulse signal.

[0039] This paper proposes a dual-path Transformer network model based on the time-frequency domain attention mechanism, which aims to achieve accurate reconstruction of unknown pulse signal waveforms in the context of complex marine environments, especially under the condition of coexistence of non-Gaussian distribution interference and Gaussian white noise. The network of the present invention integrates dilated convolution, attention mechanism, jump connection structure and Transformer network structure, effectively realizing the feature extraction of local and global features of pulse signals in the time-frequency domain, and finally realizing accurate modeling of pulse signals in complex marine background environments. The key conclusions are as follows:

[0040] (1) Under different input signal-to-interference ratio and signal-to-noise ratio conditions, the proposed method outperforms other comparison algorithms in two key indicators: signal distortion ratio and signal-to-interference ratio gain. This result shows that the proposed method has higher robustness and accuracy when processing pulse signals in complex noise environments.

[0041] (2) By introducing dilated convolution, attention mechanism, skip connection structure and Transformer network structure, the method of the present invention significantly enhances the feature extraction and fusion capabilities of pulse signals in the time and frequency domains, thereby improving the reconstruction accuracy of pulse signals.

[0042] (3) Through the processing and analysis of measured data, the method of the present invention further verifies its feasibility and effectiveness in practical applications. Compared with other comparative methods, the method of the present invention shows more superior performance when processing measured signals in complex marine background environments, which fully demonstrates its wide applicability and good generalization ability.

[0043] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that: in step 1, a training set X is constructed; the specific process is:

[0044] Physical scene models such as Figure 2 shown.

[0045] The receiving hydrophone not only receives the desired pulse signal from the transmitting transducer, but also receives the ocean environmental noise and non-Gaussian interference from sound sources such as marine life, ship navigation and seabed drilling.

[0046] Assume that the ocean environmental noise is Gaussian white noise and the interference is α stable distribution noise;

[0047] The acoustic signal is emitted from the transmitting transducer and propagates to the receiving hydrophone. The time-frequency domain representation of the signal received by the receiving hydrophone is:

[0048] X(t,f)=S(t,f)+N(t,f)+I(t,f) (1)

[0049] Among them, is the time-frequency domain representation of the signal received by the receiving hydrophone; is the time-frequency domain representation of the desired pulse signal emitted by the transmitting transducer; is the time-frequency domain representation of the interference; is the time-frequency domain representation of the ocean environmental noise;

[0050] t=1,…,T,f=1,…,F,T and F are the total number of time frames and frequency frames represented by the input signal X in the time-frequency domain respectively.

[0051] The other steps and parameters are the same as those in the first embodiment.

[0052] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that: in step two, a deep neural network model is constructed; the specific process is:

[0053] The overall network structure is as follows Figure 3 As shown;

[0054] The deep neural network model includes an encoder, a separator and a decoder;

[0055] The working process of the deep neural network model is:

[0056] The time-frequency representation of the signal received by the hydrophone is As the input signal of the deep neural network model, the number of input signal channels is 2, which are the real and imaginary parts of the short-time Fourier transform of the signal; T and F are the total number of time frames and frequency frames represented by the input signal in the time-frequency domain, respectively; is a real number;

[0057] The input signal is passed through the encoder to extract the signal feature Encoder_Out, and the signal feature Encoder_Out is passed through the separator to obtain the signal mask Mask;

[0058] The product of the signal feature Encoder_Out and the signal mask Mask is used as the input of the decoder, and the decoder outputs the reconstructed signal

[0059] The other steps and parameters are the same as those in the first or second embodiment.

[0060] Specific implementation method 4: This implementation method is different from any one of the specific implementation methods 1 to 3 in that: the encoder includes a first dilation block and a ninth unit in sequence;

[0061] The first dilation block includes a first unit, a second unit, a third unit, a fourth unit, a fifth unit, a sixth unit, a seventh unit, an eighth unit, a first skip connection structure, a second skip connection structure and a third skip connection structure.

[0062] Each of the first unit, the second unit, the third unit, the fourth unit, the fifth unit, the sixth unit, the seventh unit, the eighth unit, and the ninth unit includes, in sequence: a dilated convolution layer (Dilated Conv) and a ReLU activation function layer;

[0063] The working process of the encoder is as follows:

[0064] Input Data Input the first unit, the first unit outputs feature A;

[0065] Feature A is input into the second unit, and the second unit outputs feature B;

[0066] Feature A is input into the first skip connection structure (Skip Connection), and the first skip connection structure (SkipConnection) outputs feature C;

[0067] Feature B is input into the third unit, and the third unit outputs feature D;

[0068] Feature B is input into the second skip connection structure (Skip Connection), and the second skip connection structure (SkipConnection) outputs feature E;

[0069] Feature D is input into the fourth unit, and the fourth unit outputs feature F;

[0070] Feature D is input into the third skip connection structure (Skip Connection), and the third skip connection structure (SkipConnection) outputs feature G;

[0071] Feature F is input into the fifth unit, and the fifth unit outputs feature H;

[0072] Feature H and the third skip connection structure (Skip Connection) output feature G are input into the sixth unit, and the sixth unit outputs feature I;

[0073] Feature I and the second skip connection structure (Skip Connection) output feature E input into the seventh unit, and the seventh unit outputs feature J;

[0074] Feature J and the first skip connection structure (Skip Connection) output feature C are input into the eighth unit, and the eighth unit outputs feature K;

[0075] Feature K is input into the ninth unit, and the ninth unit outputs feature L, which is the encoder output feature Encoder_Out;

[0076] The expression is:

[0077] Encoder_Out=ReLU(Dilated_Conv(Dilation_Block(X))) (2)

[0078] Among them, Dilation_Block() is the dilation module (Dilation Block), Dilated_Conv is the dilated convolution layer (Dilated Conv), and ReLU is the ReLU activation function layer.

[0079] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0080] Specific implementation mode five: This implementation mode is different from any one of specific implementation modes one to four in that: each of the first skip connection structure (Skip Connection), the second skip connection structure (Skip Connection), and the third skip connection structure (Skip Connection) includes a first convolution layer, a first ReLU activation function layer, a first maximum pooling layer, a second convolution layer, a first Tanh activation function layer, a second maximum pooling layer, a third convolution layer, and a first Sigmiod activation function layer;

[0081] The working process of each skip connection structure (SkipConnection) in the first skip connection structure (Skip Connection), the second skip connection structure (SkipConnection), and the third skip connection structure (Skip Connection) is as follows:

[0082] The input feature A′ is sequentially input into the first convolutional layer and the first ReLU activation function layer, and the first ReLU activation function layer outputs feature B′;

[0083] The input feature A′ is sequentially input into the first maximum pooling layer, the second convolutional layer, and the first Tanh activation function layer, and the first Tanh activation function layer outputs the feature C′;

[0084] The input feature A′ is sequentially input into the second maximum pooling layer, the third convolutional layer, and the first Sigmaod activation function layer, and the first Sigmaod activation function layer outputs the feature D′;

[0085] Perform dot product of feature B′, feature C′, and feature D′ to obtain feature E′;

[0086] Feature E′ is used as the output feature of the skip connection structure.

[0087] The other steps and parameters are the same as those in Specific Implementation 1 to 4-1.

[0088] Specific implementation six: This implementation is different from any one of the specific implementations one to five in that: the separator includes a normalization layer (LayerNorm), a fourth convolutional layer, a first dual-path self-attention mechanism Transformer module (DPSATBlock), a second dual-path self-attention mechanism Transformer module (DPSAT Block), a fifth convolutional layer, a first PReLU activation function layer, a sixth convolutional layer, a second Tanh activation function layer, a seventh convolutional layer, a second Sigmiod activation function layer, an eighth convolutional layer, and a second ReLU activation function layer;

[0089] The working process of the separator is:

[0090] The encoder output feature Encoder_Out is input into the normalization layer (LayerNorm) and the fourth convolutional layer in turn, and the fourth convolutional layer outputs the feature F′;

[0091] The feature F′ is input into the first dual-path self-attention mechanism Transformer module (DPSAT Block), and the first dual-path self-attention mechanism Transformer module (DPSAT Block) outputs the feature G′;

[0092] The feature G′ is input into the second dual-path self-attention mechanism Transformer module (DPSAT Block), and the second dual-path self-attention mechanism Transformer module (DPSAT Block) outputs the feature H′;

[0093] The feature H′ is sequentially input into the fifth convolutional layer and the first PReLU activation function layer, and the first PReLU activation function layer outputs the feature I′;

[0094] Feature I′ is input into the sixth convolutional layer and the second Tanh activation function layer in sequence, and the second Tanh activation function layer outputs feature J′;

[0095] Feature I′ is sequentially input into the seventh convolutional layer and the second Sigmiod activation function layer, and the second Sigmiod activation function layer outputs feature K′;

[0096] Perform dot product of feature J′ and feature K′ to obtain feature L′;

[0097] The feature L′ is sequentially input into the eighth convolutional layer and the second ReLU activation function layer, and the second ReLU activation function layer outputs the feature M′;

[0098] The feature M′ is the separator output signal mask Mask.

[0099] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0100] Specific implementation seven: This implementation is different from any one of specific implementations one to six in that: each of the first dual-path self-attention mechanism Transformer module (DPSAT Block) and the second dual-path self-attention mechanism Transformer module (DPSAT Block) includes a time domain sub-block (Time-domain Sub-block) and a frequency domain sub-block (Frequency-domain Sub-block), a fifteenth convolutional layer, and a second PReLU activation function layer;

[0101] The time-domain sub-block includes: a ninth convolutional layer, a third ReLU activation function layer, a first Transformer structure, a third maximum pooling layer, a tenth convolutional layer, a third Tanh activation function layer, an eleventh convolutional layer, and a third Sigmiod activation function layer;

[0102] The frequency-domain sub-block includes: a twelfth convolutional layer, a fourth ReLU activation function layer, a second Transformer structure, a fourth maximum pooling layer, a thirteenth convolutional layer, a fourth Sigmiod activation function layer, a fourteenth convolutional layer, and a fourth Tanh activation function layer;

[0103] The working process of each dual-path self-attention mechanism Transformer module (DPSATBlock) in the first dual-path self-attention mechanism Transformer module (DPSAT Block) and the second dual-path self-attention mechanism Transformer module (DPSAT Block) is as follows:

[0104] 1) Feature Z′ is input into the ninth convolutional layer and the third ReLU activation function layer in sequence, and the third ReLU activation function layer outputs feature N′;

[0105] The feature Z′ is sequentially input into the first Transformer structure and the third maximum pooling layer, and the third maximum pooling layer outputs the feature O′;

[0106] The feature O′ is input into the tenth convolutional layer and the third Tanh activation function layer in sequence, and the third Tanh activation function layer outputs the feature P′;

[0107] The feature O′ is sequentially input into the eleventh convolutional layer and the third Sigmiod activation function layer, and the third Sigmiod activation function layer outputs the feature Q′;

[0108] Perform point multiplication on feature N′, feature P′, and feature Q′ to obtain feature R′; feature R′ is the output feature of the time-domain sub-block;

[0109] 2) The feature Z′ is sequentially input into the twelfth convolutional layer and the fourth ReLU activation function layer, and the fourth ReLU activation function layer outputs the feature S′;

[0110] The feature Z′ is sequentially input into the second Transformer structure and the fourth maximum pooling layer, and the fourth maximum pooling layer outputs the feature T′;

[0111] The feature T′ is sequentially input into the thirteenth convolutional layer and the fourth Sigmiod activation function layer, and the fourth Sigmiod activation function layer outputs the feature U′;

[0112] The feature T′ is sequentially input into the fourteenth convolutional layer and the fourth Tanh activation function layer, and the fourth Tanh activation function layer outputs the feature V′;

[0113] Perform point multiplication on feature S′, feature U′, and feature V′ to obtain feature W′; feature W′ is the output feature of the frequency-domain sub-block;

[0114] 3) Input the output feature R′ of the time-domain sub-block and the output feature W′ of the frequency-domain sub-block into the fifteenth convolutional layer and the second PReLU activation function layer in sequence. The second PReLU activation function layer outputs the feature Y′, which is used as the output feature of each dual-path self-attention mechanism Transformer module (DPSAT Block).

[0115] The other steps and parameters are the same as those in Specific Implementations 1 to 6-1.

[0116] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that the working process of the separator is expressed as follows:

[0117] The encoder output feature Encoder_Out is input into the normalization layer (LayerNorm) and the fourth convolutional layer in turn, and the fourth convolutional layer outputs the feature DPSAT_In; it is expressed as:

[0118] DPSAT_In=Conv(Layer_Norm(Encoder_Out)) (3)

[0119] Among them, Layer_Norm represents the normalization layer (LayerNorm);

[0120] DPSAT_In is sequentially input into the R-layer dual-path self-attention mechanism Transformer module (DPSAT Block) to obtain the feature DPSAT_Out, R = 2; it is expressed as:

[0121] DPSAT_Out=DPSAT_Block R (…(DPSAT_Block1(DPSAT_In))) (4)

[0122] Among them, DPSAT_Block R Denotes the R-th layer dual-path self-attention mechanism Transformer module (DPSATBlock), and DPSAT_Block1 denotes the 1-th layer dual-path self-attention mechanism Transformer module (DPSAT Block);

[0123] Each dual-path self-attention mechanism Transformer module (DPSAT Block) includes a frequency-domain sub-block (Frequency-domain Sub-block) and a time-domain sub-block (Time-domain Sub-block), such as Figure 6 As shown;

[0124] The working process of the frequency-domain sub-block is as follows:

[0125] DPSAT_F_ReLU=ReLU(Conv(DPSAT_Out r-1 )) (5)

[0126] DPSAT_F=Transformer(DPSAT_Out r-1 [:,f,:]),f=1,…,F (6)

[0127] DPSAT_F_Tanh=Tanh(Conv(Maxpool(DPSAT_F))) (7)

[0128] DPSAT_F_Sigmoid=Sigmoid(Conv(Maxpool(DPSAT_F))) (8)

[0129] DPSAT_F_sub=DPSAT_F_ReLU·DPSAT_F_Tanh·DPSAT_F_Sigmoid (9)

[0130] Among them, DPSAT_F_ReLU represents DPSAT_Out in the frequency-domain sub-block (Frequency-domain Sub-block) r-1 The features output after inputting the convolution layer and the ReLU activation function layer in sequence; r = 1, 2; when r = 1, DPSAT_Out r-1 =DPSAT_In;

[0131] DPSAT_Out r-1 [:,f,:] represents DPSAT_Out r-1 Characteristics of each frequency frame in DPSAT_F; DPSAT_Out r-1 The output of each frequency frame feature after passing through the Transformer structure;

[0132] DPSAT_F_Tanh means the output of DPSAT_F after passing through the maximum pooling layer, convolution layer, and Tanh activation function layer in sequence; DPSAT_F_Sigmoid means the output of DPSAT_F after passing through the maximum pooling layer, convolution layer, and Sigmoid activation function layer in sequence;

[0133] DPSAT_F_sub represents the dot product of DPSAT_F_ReLU, DPSAT_F_Tanh, and DPSAT_F_Sigmoid;

[0134] The working process of the time-domain sub-block is as follows:

[0135] DPSAT_T_ReLU=ReLU(Conv(DPSAT_Out r-1 )) (10)

[0136] DPSAT_T=Transformer(DPSAT_Οut r-1 [:,:,t]),t=1,…,T (11)

[0137] DPSAT_T_Tanh=Tanh(Conv(Maxpool(DPSAT_T))) (12)

[0138] DPSAT_T_Sigmoid=Sigmoid(Conv(Maxpool(DPSAT_T))) (13)

[0139] DPSAT_T_sub=DPSAT_T_ReLU·DPSAT_T_Tanh·DPSAT_T_Sigmoid (14)

[0140] Among them, DPSAT_T_ReLU represents DPSAT_Out in the time-domain sub-block (Time-domain Sub-block) r-1 The features output after inputting the convolution layer and the ReLU activation function layer in sequence; r = 1, 2; when r = 1, DPSAT_Out r-1 =DPSAT_In;

[0141] DPSAT_Οut r-1 [:,:,t] means DPSAT_Out r-1 Characteristics of each time frame in; DPSAT_T represents DPSAT_Out r-1 The output of each time frame feature after passing through the Transformer structure;

[0142] DPSAT_T_Tanh means the output of DPSAT_T after passing through the maximum pooling layer, convolution layer, and Tanh activation function layer in sequence; DPSAT_T_Sigmoid means the output of DPSAT_T after passing through the maximum pooling layer, convolution layer, and Sigmoid activation function layer in sequence;

[0143] DPSAT_T_sub represents the dot product of DPSAT_T_ReLU, DPSAT_T_Tanh, and DPSAT_T_Sigmoid;

[0144] DPSAT_Out r =PReLU(Conv[DPSAT_F_sub,DPSAT_T_sub]) (15)

[0145] DPSAT_Out r Represents the output features of the r-th layer dual-path self-attention mechanism Transformer module (DPSATBlock); Figure 3 , the final Mask is expressed as

[0146] Separator_Feature=PReLU(Conv(DPSAT_Out)) (16)

[0147] Mask=ReLU(Conv(Tanh(Conv(Separator_Feature))·Sigmoid(Conv(Separator_Feature)))) (17)

[0148] Among them, Separator_Feature represents the output of DPSAT_Out after passing through the convolution layer and the PReLU activation function layer in sequence; · represents the dot product.

[0149] The other steps and parameters are the same as those in Specific Implementation 1 to 7-1.

[0150] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that: the decoder includes a second dilation block (Dilation Block), an eighteenth unit in sequence;

[0151] The second dilation block includes a tenth unit, an eleventh unit, a twelfth unit, a thirteenth unit, a fourteenth unit, a fifteenth unit, a sixteenth unit, a seventeenth unit, a fourth skip connection structure, a fifth skip connection structure and a sixth skip connection structure;

[0152] Each of the tenth unit, the eleventh unit, the twelveth unit, the thirteenth unit, the fourteenth unit, the fifteenth unit, the sixteenth unit, and the seventeenth unit includes, in sequence: a dilated convolution layer (Dilated Conv) and a ReLU activation function layer;

[0153] The eighteen units include: a dilated convolution layer (Dilated Conv), a Tanh activation function layer;

[0154] The working process of the decoder is:

[0155] Multiply the encoder output feature Encoder_Out by the separator output Mask to get the feature

[0156] The second unit output feature B is multiplied by the output Mask of the separator to obtain the feature

[0157] The third unit output feature D is multiplied by the output Mask of the separator to obtain the feature

[0158] The fourth unit output feature F is multiplied by the output Mask of the separator to obtain the feature

[0159] The fifth unit output feature H is multiplied by the output Mask of the separator to obtain the feature

[0160] The sixth unit output feature I is multiplied by the output Mask of the separator to obtain the feature

[0161] The seventh unit output feature J is multiplied by the output Mask of the separator to obtain the feature

[0162] The output feature K of the eighth unit is multiplied by the output Mask of the separator to obtain the feature

[0163] feature Input the tenth unit, the tenth unit outputs the feature

[0164] feature Input a fourth skip connection structure (Skip Connection), and the fourth skip connection structure (SkipConnection) outputs a feature δ;

[0165] feature and Features Input unit 11, unit 11 output features

[0166] feature Input the fifth skip connection structure (Skip Connection), and the fifth skip connection structure (SkipConnection) outputs the feature δ′;

[0167] feature and Features Input the twelfth unit, the twelfth unit outputs the feature

[0168] feature Input the sixth skip connection structure (Skip Connection), the sixth skip connection structure (SkipConnection) outputs the feature

[0169] feature and Features Input unit 13, unit 13 output features

[0170] feature and Features Input unit 14, unit 14 output features

[0171] feature feature And feature δ″ is input into the fifteenth unit, and the fifteenth unit outputs feature

[0172] feature feature And feature δ′ is input into the sixteenth unit, and the sixteenth unit outputs feature

[0173] feature feature And feature δ is input into the seventeenth unit, and the seventeenth unit outputs feature

[0174] feature Input unit 18, unit 18 output features feature Obtaining an estimate of the desired signal for the decoder;

[0175] The working process of each skip connection structure (SkipConnection) in the fourth skip connection structure (Skip Connection), the fifth skip connection structure (SkipConnection), and the sixth skip connection structure (Skip Connection) is as follows:

[0176] The input feature A′ is sequentially input into the first convolutional layer and the first ReLU activation function layer, and the first ReLU activation function layer outputs feature B′;

[0177] The input feature A′ is sequentially input into the first maximum pooling layer, the second convolutional layer, and the first Tanh activation function layer, and the first Tanh activation function layer outputs the feature C′;

[0178] The input feature A′ is sequentially input into the second maximum pooling layer, the third convolutional layer, and the first Sigmaod activation function layer, and the first Sigmaod activation function layer outputs the feature D′;

[0179] Perform dot product of feature B′, feature C′, and feature D′ to obtain feature E′;

[0180] Feature E′ is used as the output feature of the skip connection structure.

[0181] The other steps and parameters are the same as those in Specific Implementation 1 to 8-1.

[0182] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that: in step three, the constructed deep neural network model is trained based on the training set to obtain a trained deep neural network model;

[0183] The specific process is:

[0184] The time-frequency domain signal representation X of the hydrophone receiving signal in step 1 is normalized to obtain normalized data; the expression is:

[0185]

[0186] The normalized data is used as the input of the deep neural network model constructed in step 2, and the expected signal is used as the output of the deep neural network model;

[0187] Set the loss function to

[0188]

[0189] Where MSE() represents mean square error, S represents the expected signal, represents the expected signal estimate output by the deep neural network model, L(θ) represents the loss function, and θ represents the deep neural network model parameters;

[0190]

[0191] Among them, g θ (·) is the mapping function between the neural network input and output;

[0192] The network uses the Adaptive Moment Estimation (Adam) method to update the parameters in the network;

[0193]

[0194] Until convergence, a trained deep neural network model is obtained.

[0195] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 9.

[0196] Encoders and Decoders

[0197] The structure of encoder and decoder is as follows Figure 4 As shown. The encoder and decoder have similar structures, which are obtained by superimposing a series of dilated convolutional layers (Dilated Conv) and ReLU activation functions. In the present invention, the number of convolutional layer filters is 64. Non-Gaussian interference manifests itself as complex spectral characteristics and time dynamic changes in the time-frequency domain. The capture of global information helps to reveal the inherent laws of signals in time series and frequency distribution. The dilated convolutional layer (Dilated Conv) effectively increases the receptive field without changing the size of the output feature map, so as to better capture the global information in the input data. The use of the dilated convolutional layer (Dilated Conv) helps to identify and distinguish pulse signals from non-Gaussian interference in the time-frequency domain. However, the superposition of a series of convolutional layers makes the neural network tend to fit low-frequency features first and then high-frequency features, and at the same time brings the problem of gradient disappearance. The present invention adopts Figure 5The skip connection structure (Skip Connection) shown in the figure performs dynamic feature selection on the skip connection input features. The network uses this skip connection structure to connect the high-frequency features and low-frequency features extracted by the encoder and decoder, and establishes a direct connection between the encoder and decoder as a whole. Through this multi-scale feature fusion mechanism, the network can more comprehensively capture the local details and global structure in the input data, while effectively alleviating the problem of gradient disappearance. In general, the network makes full use of the high- and low-frequency information of the input signal, and more effectively restores the detailed structure of the pulse signal while identifying and distinguishing non-Gaussian interference and pulse signals as a whole.

[0198] like Figure 3-Figure 4 As shown, the input data can be expressed as The feature map is obtained through the encoder, that is, the encoder output Encoder_Out. Encoder_Out is multiplied by the output Mask of the separator and the estimated value of the desired signal is obtained through the decoder

[0199] Separator

[0200] In the separator part, the input DPSAT_In of the DPSAT Block is passed through R layers of DPSAT Block to further extract the features of the input signal to obtain a mask. In the present invention, R=2. The structure of the rth layer DPSAT Block is as follows: Figure 6 shown.

[0201] DPSAT Block includes a time-domain sub-block and a frequency-domain sub-block, which captures the time-domain information and frequency information in the model input features in both the time dimension and the frequency dimension. DPSAT Block uses the Transformer structure to capture the global information in the input signal, effectively distinguishes the distribution differences between non-Gaussian interference and pulse signals in the time-frequency domain, and finally suppresses non-Gaussian interference to retain the pulse signal information. Subsequently, DPSAT Block uses the maximum pooling layer MaxPool to nest Tanh and Sigmoid activation functions on the time-domain sub-block and the frequency-domain sub-block to realize dynamic feature selection in the time-frequency domain, and multiplies the input signal after the ReLU activation function to further extract the time-frequency domain features of the input signal, which is an improvement of the present invention on the attention mechanism. This mechanism increases the weight of the desired signal area, reduces the weight of the background noise and non-Gaussian interference area, strengthens the selection ability of the network of the present invention for important pulse features, and effectively suppresses noise and non-Gaussian interference while extracting the desired signal features. At the end of the DPSAT Block, the extracted time-frequency domain features are passed through the convolution layer Conv and the PReLU activation function to obtain the output DPSAT_Out of the rth layer DPSAT Block. r .

[0202] like Figure 3 , Encoder_Out is obtained through the normalization layer (Layer Norm) and the convolution layer Conv to obtain DPSAT_In, and the final DPSAT_Out is obtained by DPSAT_In through the R layer DPSAT Block layer. The structure of each DPSAT Block layer is as follows Figure 6 Each DPSAT Block layer mainly includes a time domain word block and a frequency domain word block, and the structures of the time domain sub-block (Time-domain Sub-block) and the frequency domain sub-block (Frequency-domain Sub-block) are similar.

[0203] The Transformer structure in the DPSAT Block also includes the attention mechanism SelfAttention. The detailed structure of the Transformer structure and the attention mechanism is as follows Figure 7-Figure 8 shown.

[0204] In the Transformer structure, Transformer_In is obtained by combining the attention mechanism with the feedforward neural network, residual connection and layer normalization to obtain Transformer_Out. The attention mechanism effectively combines the information at different positions in the input data by processing multiple attention distributions in parallel to better improve the feature extraction ability of the model. The addition of the attention mechanism to the SelfAttention attention mechanism obtains the parallel query matrix Q, key matrix K, and value matrix V by transforming Self_Attention_In h times linearly, where h=2 in the present invention. The output of each head is calculated through parallel attention, the results of each head are merged and the final Self_Attention_Out is obtained through the Linear layer.

[0205] train

[0206] During the training process, the Adaptive Moment Estimation (Adam) method is used to update the parameters in the network. The number of training times is 100 times. The update process is shown in the following formula.

[0207]

[0208] Wherein, m represents the first-order momentum of the gradient, v represents the second-order momentum of the gradient, β1 and β2 are the exponential decay rates of the first-order momentum and the second-order momentum, respectively, lr is the learning rate, and ε is a minimum value to prevent the denominator of formula (24) from being 0 during the parameter update process. In the present invention, β1=0.9, β2=0.999, lr=5e-5, ε=1e-8 are taken. As the network parameters are updated, the network output value gradually approaches the label and finally obtains the reconstructed signal.

[0209] Sample Set Settings

[0210] The network input signal contains pulse signals, Gaussian white noise and α-stable distribution interference. The duration of each signal is 1s and the sampling rate is 10kHz. In the training, there are 360 ​​sets of training data and 360 sets of test data. The input signal-to-noise ratio (SNR) ranges from 0dB to 10dB, with 2dB as the basic interval unit, and the input (SIR) ranges from -10dB to 0dB, with 2dB as the basic interval unit. In the training and test data, the frequency of the CW signal is randomly distributed in [1000, 4500] Hz, the initial frequency and cutoff frequency of the LFM pulse signal are randomly distributed in [1000, 4500] Hz, the pulse width of the pulse signal is randomly distributed in [5, 200] ms, and the start time of the pulse signal appears at any position within [0, 1] s.

[0211] Evaluation Criteria

[0212] In order to evaluate the performance of the network, the signal to interference and noise ratio gain and the signal to distortion ratio (SDR) are used to evaluate the performance of the network, thereby verifying the effectiveness of the network of the present invention.

[0213] ①. Signal-to-interference-noise ratio gain

[0214] The signal-to-noise ratio is defined as the ratio of the average power of the desired signal to the average power of the interference and noise in the signal. The signal-to-noise ratio gain is the difference between the signal-to-noise ratio of the reconstructed signal and the signal-to-noise ratio of the signal before reconstruction. The signal-to-noise ratio gain reflects the signal reconstruction capability of the deep neural network.

[0215]

[0216] SINR gain =SINR out -SINR in (26)

[0217] Among them, P s represents the average power of the desired signal, which is the average power of the pulse signal in the present invention, P in Represents the average power of interference and noise. SINR out Represents the signal-to-interference-noise ratio (SINR) of the reconstructed signal in It shows the signal-to-interference-noise ratio of the signal before reconstruction.

[0218] ②.Signal distortion ratio

[0219] An important indicator for evaluating the degree of distortion of the reconstructed signal is the signal distortion ratio, which is defined as the ratio of the square of the binary norm of the expected signal to the square of the binary norm of the difference between the expected signal and the reconstructed signal.

[0220]

[0221] The following examples are used to verify the beneficial effects of the present invention:

[0222] Simulation Analysis

[0223] This section uses a dual-path Transformer network based on the time-frequency domain attention mechanism to process the input signal, thereby verifying the feasibility and generalization of the network of the present invention in the task of reconstructing pulse signals in a non-Gaussian interference environment. The simulation conditions are as follows. The pulse width of the pulse signal is 100ms, the frequency of the CW signal is 2.5kHz, the starting frequency of the LFM pulse signal is 1kHz, the ending frequency is 1.45kHz, the starting frequency of the HFM pulse signal is 1000Hz, and the ending frequency is 2000Hz. Set the input signal-to-noise ratio to 10dB and the signal-to-interference ratio to -10dB. The characteristic index α of the α-stable distribution interference is 1.8, the skew index β is 0, the scale parameter γ is 1, and the position parameter a is 0. Figure 9-11 They are the time domain waveforms of the input signal, the expected signal and the reconstructed signal when the expected signal is a CW signal, an LFM pulse signal and an HFM pulse signal respectively.

[0224] like Figure 9-10 As shown in the figure, before the input signal is processed by the signal reconstruction network, the desired signal in the time domain is almost completely submerged in the background noise and interference. When the input signal is processed by the dual-path Transformer network based on the time-frequency domain attention mechanism, the reconstructed signal effectively retains the characteristics of the original pulse signal and most of the interference is eliminated. The network of the present invention effectively removes the α-stable distribution interference and Gaussian white noise, and the effect of offsetting interference is more significant, which effectively enhances the target pulse signal.

[0225] pass Fig.11 It can be seen that when the expected signal type is different from the training data, the dual-path Transformer network based on the time-frequency domain attention mechanism can still well restore the expected signal that is almost submerged in the background noise, and achieve an almost distortion-free reconstruction effect. It is sufficient to see that the network of the present invention has a certain generalization ability. The experimental results prove the wide application potential of the network in the task of pulse signal reconstruction.

[0226] In order to systematically analyze the signal reconstruction performance of the network of the present invention, ablation experiments are performed to verify the roles of the jump connection structure, dilated convolution, Transformer network structure and attention mechanism in the method of the present invention. The simulation conditions are the same as above, and four comparison models are introduced here. Comparison model 1: Delete the attention mechanism in the DPSAT Block module; Comparison model 2: Delete the Transformer network structure in the network of the present invention; Comparison model 3: Replace the jump connection structure in the encoder and decoder modules in the network of the present invention with a normal jump connection structure; Comparison model 4: Replace the dilated convolution layer in the encoder and decoder modules in the network of the present invention with a normal convolution layer. Table 1-2 shows the signal-to-noise ratio gain and signal distortion ratio of the network of the present invention and the comparison model under different input signal-to-interference ratios and signal-to-noise ratios.

[0227] Table 1 Signal-to-interference-noise ratio gain and signal-to-distortion ratio of the network of the present invention and the comparative model under different signal-to-interference ratios

[0228]

[0229] Table 2 Signal-to-interference-noise ratio gain and signal-to-distortion ratio of the network of the present invention and the comparative model under different signal-to-noise ratio conditions

[0230]

[0231]

[0232] It can be seen from the ablation experiment results that the network model designed by the present invention shows significant advantages over the four comparison models in terms of pulse signal reconstruction, especially in terms of transient signal modeling ability. The dilated convolution and Transformer structures enable the network to capture longer-distance dependencies by increasing the receptive field, which is crucial for reconstructing subtle structures and transient changes in signals. The introduction of the attention mechanism and the jump connection structure further enhances the network's ability to select important features. The attention mechanism dynamically adjusts the weights of different feature areas, allowing the network to focus more on key information that is beneficial to signal reconstruction, thereby effectively suppressing noise and interference. The jump connection structure realizes dynamic selection of input features and fusion of global features and local features. In the ablation experiment, when the above four structures are removed, the signal-to-noise ratio gain and signal-to-distortion ratio of the network both decrease to varying degrees, which fully demonstrates the important role of the network of the present invention in improving the accuracy of pulse signal reconstruction. It is worth noting that the network of the present invention performs better than the comparison model under different input signal-to-noise ratio and input signal-to-interference ratio conditions, which means that no matter how the signal environment changes, the network of the present invention can maintain high stability and robustness. In general, the signal to noise ratio gain and signal distortion ratio of the network of the present invention under different signal to noise ratios and signal to interference ratios are better than those of the comparison model, which is enough to show the effectiveness of the expanded convolutional layer, attention mechanism, Transformer structure and jump connection structure.

[0233] In order to further investigate the effectiveness of the method of the present invention, the present invention further compares and analyzes the performance of the proposed algorithm with three deep neural network denoising algorithms, namely UNet, CBDNet and RIDNet, and plots the relationship curves of the signal-to-noise ratio gain with the input signal-to-interference ratio and the input signal-to-noise ratio, and the relationship curves of the signal-to-distortion ratio with the input signal-to-interference ratio and the input signal-to-noise ratio, as shown in Figure 2. Figure 12-Figure 15 shown.

[0234] Figure 12-Figure 15It shows that the input signal-to-interference ratio and signal-to-noise ratio have a certain influence on the signal reconstruction performance of the network of the present invention. As the input signal-to-interference ratio and input signal-to-noise ratio increase, the signal-to-interference-noise ratio gain shows a downward trend, and the signal-to-distortion ratio increases significantly with the increase of the input signal-to-interference ratio. After analysis, as the input signal-to-interference ratio and input signal-to-noise ratio increase, the performance of the dual-path Transformer network based on the time-frequency domain attention mechanism is less affected by interference and noise, and the degree of signal distortion is reduced. Therefore, the input signal-to-interference ratio and the input signal-to-noise ratio are required not to be too low to ensure a better signal reconstruction effect. In summary, under a wide range of input signal-to-interference-noise ratios, the method of the present invention has shown good stability and robustness, and can effectively reduce the interference and noise components in the input signal while retaining the pulse signal characteristics in the input signal. The signal-to-interference-noise ratio gain and signal-to-distortion ratio of the method of the present invention are better than those of the other three comparison methods. The reason is that the method of the present invention fully combines the local and global characteristics of the input signal, so that the network can fully take into account the change characteristics of the signal in the local area and the overall trend in the global characteristics during the signal reconstruction process, thereby achieving a more comprehensive and accurate modeling of the pulse signal.

[0235] In order to investigate the feasibility of the method in the actual physical scene, this section uses the dual-path Transformer network based on the time-frequency domain attention mechanism to process the measured data to further verify the effectiveness and feasibility of the method. In this experiment, the period of the CW signal is 1s, the pulse width is 10ms, the center frequency is 3kHz, and the sampling rate is 10kHz. Fig.16 is the time domain waveform of the measured signal, Fig.17 (a)-(d) are respectively the time domain waveforms of the reconstructed signals obtained by processing the measured data using the method of the present invention, UNet, CBDNet and RIDNet. Table 3 shows the statistical results of the signal-to-interference-noise ratio gain and signal-to-distortion ratio of the network of the present invention and the comparison network when processing the measured data.

[0236] Table 3 Statistical results of signal-to-interference-noise ratio gain and signal-to-distortion ratio of the network of the present invention and the comparative network when processing measured data

[0237]

[0238] Depend on Figure 16-17 It can be seen that the pulse signal in the measured data is not only significantly disturbed by the non-Gaussian distributed pulse noise, but the background noise also shows a non-balanced characteristic in the frequency domain, which is specifically manifested as the noise intensity in the low-frequency band is significantly higher than that in the high-frequency band. In this complex environment, the signal reconstruction effect of the method of the present invention on the experimental data is good, which further verifies the generalization of the method of the present invention, and the dual-path Transformer network based on the time-frequency domain attention mechanism can be applied to the processing of measured data. Figure 16-Figure 17It can be seen that the three methods of UNet, CBDNet, and RIDNet can also effectively reconstruct pulse signals when processing experimental data, but the method of the present invention performs better in accurately distinguishing target signals from background interference. The dual-path Transformer network based on the time-frequency domain attention mechanism eliminates almost all interference and noise in the reconstructed signal through its unique time-frequency domain attention mechanism. It can be seen from Table 3 that the dual-path Transformer network based on the time-frequency domain attention mechanism has better superiority in processing measured signals, which further verifies the feasibility and effectiveness of the method of the present invention.

[0239] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A pulse signal reconstruction method based on deep learning in a non-Gaussian environment, characterized by: The specific process of the method is: Step 1: Construct a training set; Step 2: Build a deep neural network model; Step 3: Train the constructed deep neural network model based on the training set to obtain a trained deep neural network model; Step 4: Input the received underwater acoustic pulse signal into the trained deep neural network model, and the trained deep neural network model outputs the reconstructed pulse signal.

2. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 1, characterized in that: In the step 1, a training set is constructed; the specific process is as follows: The receiving hydrophone not only receives the desired pulse signal from the transmitting transducer, but also receives the ocean environment noise and interference; Assume that the ocean environmental noise is Gaussian white noise and the interference is α stable distribution noise; The acoustic signal is emitted from the transmitting transducer and propagates to the receiving hydrophone. The time-frequency domain representation of the signal received by the receiving hydrophone is: X(t,f)=S(t,f)+N(t,f)+I(t,f) (1) Wherein, X is the time-frequency domain representation of the signal received by the receiving hydrophone; S is the time-frequency domain representation of the desired pulse signal emitted by the transmitting transducer; I is the time-frequency domain representation of the interference; N is the time-frequency domain representation of the ocean environment noise; t=1,…,T,f=1,…,F,T and F are the total number of time frames and frequency frames represented by the input signal X in the time-frequency domain respectively.

3. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 2, characterized in that: In the step 2, a deep neural network model is constructed; the specific process is as follows: The deep neural network model includes an encoder, a separator, and a decoder; The working process of the deep neural network model is: The time-frequency domain representation of the signal received by the hydrophone is As the input signal of the deep neural network model, the number of input signal channels is 2, which are the real and imaginary parts of the short-time Fourier transform of the signal; T and F are the total number of time frames and frequency frames represented by the input signal in the time-frequency domain, respectively; is a real number; The input signal is passed through the encoder to extract the signal feature Encoder_Out, and the signal feature Encoder_Out is passed through the separator to obtain the signal mask Mask; The product of the signal feature Encoder_Out and the signal mask Mask is used as the input of the decoder, and the decoder outputs the reconstructed signal 4. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 3 is characterized in that: The encoder comprises a first expansion module and a ninth unit in sequence; The first expansion module includes a first unit, a second unit, a third unit, a fourth unit, a fifth unit, a sixth unit, a seventh unit, an eighth unit, a first jump connection structure, a second jump connection structure, and a third jump connection structure; Each of the first unit, the second unit, the third unit, the fourth unit, the fifth unit, the sixth unit, the seventh unit, the eighth unit, and the ninth unit sequentially comprises: an expanded convolution layer and a ReLU activation function layer; The working process of the encoder is: Input Data Input the first unit, the first unit outputs feature A; Feature A is input into the second unit, and the second unit outputs feature B; Feature A is input into the first skip connection structure, and the first skip connection structure outputs feature C; Feature B is input into the third unit, and the third unit outputs feature D; Feature B is input into the second jump connection structure, and the second jump connection structure outputs feature E; Feature D is input into the fourth unit, and the fourth unit outputs feature F; Feature D is input into the third jump connection structure, and the third jump connection structure outputs feature G; Feature F is input into the fifth unit, and the fifth unit outputs feature H; Feature H and the third skip connection structure output feature G are input into the sixth unit, and the sixth unit outputs feature I; Feature I and the second skip connection structure output feature E are input into the seventh unit, and the seventh unit outputs feature J; Feature J and the first skip connection structure output feature C are input into the eighth unit, and the eighth unit outputs feature K; Feature K is input into the ninth unit, and the ninth unit outputs feature L, which is the encoder output feature Encoder_Out; The expression is: Encoder_Out=ReLU(Dilated_Conv(Dilation_Block(X))) (2) Among them, Dilation_Block() is the dilation module (Dilation Block), Dilated_Conv is the dilated convolution layer (Dilated Conv), and ReLU is the ReLU activation function layer.

5. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 4, characterized in that: Each of the first jump connection structure, the second jump connection structure, and the third jump connection structure includes a first convolution layer, a first ReLU activation function layer, a first maximum pooling layer, a second convolution layer, a first Tanh activation function layer, a second maximum pooling layer, a third convolution layer, and a first Sigmiod activation function layer; The working process of each jump connection structure in the first jump connection structure, the second jump connection structure, and the third jump connection structure is as follows: The input feature A′ is sequentially input into the first convolutional layer and the first ReLU activation function layer, and the first ReLU activation function layer outputs feature B′; The input feature A′ is sequentially input into the first maximum pooling layer, the second convolutional layer, and the first Tanh activation function layer, and the first Tanh activation function layer outputs the feature C′; The input feature A′ is sequentially input into the second maximum pooling layer, the third convolutional layer, and the first Sigmaod activation function layer, and the first Sigmaod activation function layer outputs the feature D′; Perform dot product of feature B′, feature C′, and feature D′ to obtain feature E′; Feature E′ is used as the output feature of the skip connection structure.

6. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 5, characterized in that: The separator includes a normalization layer, a fourth convolution layer, a first dual-path self-attention mechanism Transformer module, a second dual-path self-attention mechanism Transformer module, a fifth convolution layer, a first PReLU activation function layer, a sixth convolution layer, a second Tanh activation function layer, a seventh convolution layer, a second Sigmiod activation function layer, an eighth convolution layer, and a second ReLU activation function layer; The working process of the separator is: The encoder output feature Encoder_Out is input into the normalization layer and the fourth convolution layer in sequence, and the fourth convolution layer outputs the feature F′; The feature F′ is input into the first two-way self-attention mechanism Transformer module, and the first two-way self-attention mechanism Transformer module outputs the feature G′; The feature G′ is input into the second dual-path self-attention mechanism Transformer module, and the second dual-path self-attention mechanism Transformer module outputs the feature H′; The feature H′ is sequentially input into the fifth convolutional layer and the first PReLU activation function layer, and the first PReLU activation function layer outputs the feature I′; Feature I′ is input into the sixth convolutional layer and the second Tanh activation function layer in sequence, and the second Tanh activation function layer outputs feature J′; Feature I′ is sequentially input into the seventh convolutional layer and the second Sigmiod activation function layer, and the second Sigmiod activation function layer outputs feature K′; Perform dot product of feature J′ and feature K′ to obtain feature L′; The feature L′ is sequentially input into the eighth convolutional layer and the second ReLU activation function layer, and the second ReLU activation function layer outputs the feature M′; The feature M′ is the separator output signal mask Mask.

7. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 6, characterized in that: Each of the first two-way self-attention mechanism Transformer module and the second two-way self-attention mechanism Transformer module includes a time domain sub-block and a frequency domain sub-block, a fifteenth convolutional layer, and a second PReLU activation function layer; The time domain sub-block includes: a ninth convolution layer, a third ReLU activation function layer, a first Transformer structure, a third maximum pooling layer, a tenth convolution layer, a third Tanh activation function layer, an eleventh convolution layer, and a third Sigmiod activation function layer; The frequency domain sub-block includes: a twelfth convolution layer, a fourth ReLU activation function layer, a second Transformer structure, a fourth maximum pooling layer, a thirteenth convolution layer, a fourth Sigmiod activation function layer, a fourteenth convolution layer, and a fourth Tanh activation function layer; The working process of each of the first two-way self-attention mechanism Transformer module and the second two-way self-attention mechanism Transformer module is as follows: 1) Feature Z′ is sequentially input into the ninth convolutional layer and the third ReLU activation function layer, and the third ReLU activation function layer outputs feature N′; The feature Z′ is sequentially input into the first Transformer structure and the third maximum pooling layer, and the third maximum pooling layer outputs the feature O′; The feature O′ is input into the tenth convolutional layer and the third Tanh activation function layer in sequence, and the third Tanh activation function layer outputs the feature P′; The feature O′ is sequentially input into the eleventh convolutional layer and the third Sigmiod activation function layer, and the third Sigmiod activation function layer outputs the feature Q′; Perform point multiplication on feature N′, feature P′, and feature Q′ to obtain feature R′; feature R′ is the output feature of the time domain sub-block; 2) The feature Z′ is sequentially input into the twelfth convolutional layer and the fourth ReLU activation function layer, and the fourth ReLU activation function layer outputs the feature S′; The feature Z′ is sequentially input into the second Transformer structure and the fourth maximum pooling layer, and the fourth maximum pooling layer outputs the feature T′; The feature T′ is sequentially input into the thirteenth convolutional layer and the fourth Sigmiod activation function layer, and the fourth Sigmiod activation function layer outputs the feature U′; The feature T′ is sequentially input into the fourteenth convolutional layer and the fourth Tanh activation function layer, and the fourth Tanh activation function layer outputs the feature V′; Perform point multiplication on feature S′, feature U′, and feature V′ to obtain feature W′; feature W′ is the sub-block output feature in the frequency domain; 3) Input the time domain sub-block output feature R′ and the frequency domain sub-block output feature W′ into the fifteenth convolutional layer and the second PReLU activation function layer in sequence. The second PReLU activation function layer outputs the feature Y′, which is used as the output feature of each dual-path self-attention mechanism Transformer module.

8. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 7, characterized in that: The working process expression of the separator is: The encoder output feature Encoder_Out is input into the normalization layer and the fourth convolution layer in sequence, and the fourth convolution layer outputs the feature DPSAT_In; it is expressed as: DPSAT_In=Conv(Layer_Norm(Encoder_Out)) (3) Among them, Layer_Norm represents the normalization layer; DPSAT_In is sequentially input into the R-layer dual-path self-attention mechanism Transformer module to obtain the feature DPSAT_Out, R=2; it is expressed as: DPSAT_Out=DPSAT_Block R (…(DPSAT_Block1(DPSAT_In))) (4) Among them, DPSAT_Block R Denotes the R-th layer dual-path self-attention mechanism Transformer module, and DPSAT_Block1 denotes the 1-th layer dual-path self-attention mechanism Transformer module; Each dual-path self-attention mechanism Transformer module includes a frequency domain sub-block and a time domain sub-block; The working process of the frequency domain sub-block is as follows DPSAT_F_ReLU=ReLU(Conv(DPSAT_Out r-1 )) (5) DPSAT_F=Transformer(DPSAT_Out r-1 [:,f,:]),f=1,…,F (6) DPSAT_F_Tanh=Tanh(Conv(Maxpool(DPSAT_F))) (7) DPSAT_F_Sigmoid=Sigmoid(Conv(Maxpool(DPSAT_F))) (8) DPSAT_F_sub=DPSAT_F_ReLU·DPSAT_F_Tanh·DPSAT_F_Sigmoid (9) Among them, DPSAT_F_ReLU represents DPSAT_Out in the frequency domain sub-block r-1 The features output after inputting the convolution layer and the ReLU activation function layer in sequence; r = 1, 2; DPSAT_Out r-1 [:,f,:] represents DPSAT_Out r-1 Characteristics of each frequency frame in DPSAT_F; DPSAT_Out r-1 The output of each frequency frame feature after passing through the Transformer structure; DPSAT_F_Tanh means the output of DPSAT_F after passing through the maximum pooling layer, convolution layer, and Tanh activation function layer in sequence; DPSAT_F_Sigmoid means the output of DPSAT_F after passing through the maximum pooling layer, convolution layer, and Sigmoid activation function layer in sequence; DPSAT_F_sub represents the dot product of DPSAT_F_ReLU, DPSAT_F_Tanh, and DPSAT_F_Sigmoid; The time domain sub-block working process is as follows DPSAT_T_ReLU=ReLU(Conv(DPSAT_Out r-1 )) (10) DPSAT_T=Transformer(DPSAT_Οut r-1 [:,:,t]),t=1,…,T (11) DPSAT_T_Tanh=Tanh(Conv(Maxpool(DPSAT_T))) (12) DPSAT_T_Sigmoid=Sigmoid(Conv(Maxpool(DPSAT_T))) (13) DPSAT_T_sub=DPSAT_T_ReLU·DPSAT_T_Tanh·DPSAT_T_Sigmoid (14) Among them, DPSAT_T_ReLU represents DPSAT_Out in the time domain sub-block r-1 The features output after inputting the convolution layer and the ReLU activation function layer in sequence; DPSAT_Οut r-1 [:,:,t] means DPSAT_Out r-1 Characteristics of each time frame in; DPSAT_T represents DPSAT_Out r-1 The output of each time frame feature after passing through the Transformer structure; DPSAT_T_Tanh means the output of DPSAT_T after passing through the maximum pooling layer, convolution layer, and Tanh activation function layer in sequence; DPSAT_T_Sigmoid means the output of DPSAT_T after passing through the maximum pooling layer, convolution layer, and Sigmoid activation function layer in sequence; DPSAT_T_sub represents the dot product of DPSAT_T_ReLU, DPSAT_T_Tanh, and DPSAT_T_Sigmoid; DPSAT_Out r =PReLU(Conv[DPSAT_F_sub,DPSAT_T_sub]) (15) DPSAT_Out r Represents the output features of the r-th layer dual-path self-attention mechanism Transformer module; r = 1, 2; The final Mask is expressed as Separator_Feature=PReLU(Conv(DPSAT_Out)) (16) Mask=ReLU(Conv(Tanh(Conv(Separator_Feature))·Sigmoid(Conv(Separator_Feature)))) (17) Among them, Separator_Feature represents the output of DPSAT_Out after passing through the convolution layer and the PReLU activation function layer in sequence; · represents the dot product.

9. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 8, characterized in that: The decoder comprises, in sequence, a second expansion module and an eighteenth unit; The second expansion module includes a tenth unit, an eleventh unit, a twelve unit, a thirteenth unit, a fourteenth unit, a fifteenth unit, a sixteenth unit, a seventeenth unit, a fourth jump connection structure, a fifth jump connection structure, and a sixth jump connection structure; Each of the tenth unit, the eleventh unit, the twelveth unit, the thirteenth unit, the fourteenth unit, the fifteenth unit, the sixteenth unit, and the seventeenth unit includes, in sequence: an expanded convolution layer and a ReLU activation function layer; The eighteen units include, in sequence: an expanded convolution layer, a Tanh activation function layer; The working process of the decoder is: Multiply the encoder output feature Encoder_Out by the separator output Mask to get the feature The second unit output feature B is multiplied by the output Mask of the separator to obtain the feature The third unit output feature D is multiplied by the output Mask of the separator to obtain the feature The fourth unit output feature F is multiplied by the output Mask of the separator to obtain the feature The fifth unit output feature H is multiplied by the output Mask of the separator to obtain the feature The sixth unit output feature I is multiplied by the output Mask of the separator to obtain the feature The seventh unit output feature J is multiplied by the output Mask of the separator to obtain the feature The output feature K of the eighth unit is multiplied by the output Mask of the separator to obtain the feature feature Input the tenth unit, the tenth unit outputs the feature feature Input a fourth jump connection structure, and the fourth jump connection structure outputs a feature δ; feature and Features Input unit 11, unit 11 output features feature Input the fifth skip connection structure, and the fifth skip connection structure outputs the feature δ′; feature and Features Input the twelfth unit, the twelfth unit outputs the feature feature Input the sixth skip connection structure, and the sixth skip connection structure outputs a feature δ″; feature and Features Input unit 13, unit 13 output features feature and Features Input unit 14, unit 14 output features feature feature And feature δ″ is input into the fifteenth unit, and the fifteenth unit outputs feature feature feature And feature δ′ is input into the sixteenth unit, and the sixteenth unit outputs feature feature feature And feature δ is input into the seventeenth unit, and the seventeenth unit outputs feature feature Input unit 18, unit 18 output features feature Obtaining an estimate of the desired signal for the decoder; The working process of each jump connection structure in the fourth jump connection structure, the fifth jump connection structure, and the sixth jump connection structure is as follows: The input feature A′ is sequentially input into the first convolutional layer and the first ReLU activation function layer, and the first ReLU activation function layer outputs feature B′; The input feature A′ is sequentially input into the first maximum pooling layer, the second convolutional layer, and the first Tanh activation function layer, and the first Tanh activation function layer outputs the feature C′; The input feature A′ is sequentially input into the second maximum pooling layer, the third convolutional layer, and the first Sigmaod activation function layer, and the first Sigmaod activation function layer outputs the feature D′; Perform dot product of feature B′, feature C′, and feature D′ to obtain feature E′; Feature E′ is used as the output feature of the skip connection structure.

10. The pulse signal reconstruction method based on deep learning in a non-Gaussian environment according to claim 9, characterized in that: In step 3, the constructed deep neural network model is trained based on the training set to obtain a trained deep neural network model; the specific process is: The time-frequency domain signal representation X of the hydrophone receiving signal in step 1 is normalized to obtain normalized data; the expression is: The normalized data is used as the input of the deep neural network model constructed in step 2, and the expected signal is used as the output of the deep neural network model; Set the loss function to Where MSE() represents mean square error, S represents the expected signal, represents the expected signal estimate output by the deep neural network model, L(θ) represents the loss function, and θ represents the deep neural network model parameters; Among them, g θ (·) is the mapping function between the neural network input and output; The network uses adaptive moment estimation method to update the parameters in the network; Until convergence, a trained deep neural network model is obtained.

Images