Construction method, model and application of seabed reverberation suppression model based on generative adversarial network

Abstract

The application relates to a seabed reverberation suppression model construction method and model based on a generative adversarial network and application, and belongs to the technical field of underwater sound. The method first constructs a deep learning data set, including two parts of seabed reverberation simulation and target echo signal simulation, and uses simulation data to construct a training set and a test set. Secondly, the specific architecture of the generative adversarial network is built, including four parts of a data processing module, a generative adversarial module, a feature enhancement module and a signal reconstruction module. The application uses the generative adversarial network for seabed reverberation suppression from the angle of a time-frequency domain, and effectively improves a signal-to-reverberation ratio.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater acoustics, and particularly relates to a seabed reverberation suppression model construction method and model based on a generative adversarial network and application. BACKGROUND

[0002] When an active sonar detects and identifies a target, it may be affected by three kinds of background interference: reverberation, ocean ambient noise and platform self-noise. Among them, reverberation is formed by the accumulation of the sound signals emitted by the active sonar at the receiving point due to the sea surface, seawater and seabed scattering effects in the propagation process. The lower the frequency, the stronger the intensity and the farther the propagation distance of the transmitted signal, the more significant the reverberation. In the process of long-distance propagation of sound waves, the seabed reverberation caused by the rough scattering of the seabed interface and the volume scattering of the sediment is more significant and is the main source of reverberation. With the development of modern sonar systems towards low frequency, high power and long distance, seabed reverberation has become the main background interference in most cases when the active sonar is working. How to effectively suppress seabed reverberation and improve target detection efficiency has always been a hot and difficult topic in the study of underwater acoustics.

[0003] At present, the reverberation suppression methods at home and abroad are developing rapidly. In the early stage, corresponding reverberation suppression algorithms were formed around traditional signal processing methods, but the traditional methods have limited understanding of the characteristics of target echoes in reverberation suppression, and the inaccuracy of target feature analysis and extraction will affect the reverberation suppression results. With the development of artificial intelligence, deep learning-based reverberation suppression methods are emerging. These networks have good performance and achieve results that traditional methods cannot achieve. Therefore, the anti-reverberation technology is gradually developing towards deep learning.

[0004] Among the many deep learning methods, recursive neural networks, convolutional neural networks, deep neural networks and generative adversarial networks can all achieve signal suppression functions. Among these neural networks, the generative adversarial network has a strong generation function and strong adaptability to complex and diverse signals. It can be trained directly without establishing an accurate physical model, avoiding complex intermediate steps and greatly simplifying the training process. At the same time, the generative adversarial network can learn general patterns and features and has strong generalization ability. As the first adversarial neural network model widely used in real tasks, the superiority of the generative adversarial network has been proven in many fields and has good effects on indoor speech reverberation suppression.

[0005] The existing seabed reverberation suppression method based on the generative adversarial network of time domain waveform has limitations in processing signals and has the problem of insufficient generalization ability. SUMMARY

[0006] The present application is aimed at the problem of sea bottom reverberation suppression and the problem of insufficient generalization ability of time domain signal processing, and proposes a construction method of a sea bottom reverberation suppression model based on a time-frequency feature enhanced generative adversarial network (TFE-GAN). The method first constructs a deep learning data set, including two parts of sea bottom reverberation simulation and target echo signal simulation, and uses the simulation data to construct a training set and a test set. Secondly, the specific architecture of the generative adversarial network is built, including four parts of a data processing module, a generative adversarial module, a feature enhancement module and a signal reconstruction module. The superiority of the network is verified by comparing the TFE-GAN method with the non-negative matrix reverberation suppression method; the effectiveness of the feature enhancement module in the TFE-GAN method is verified by the ablation experiment. Finally, the sea bottom reverberation suppression effect of the TFE-GAN method is analyzed, and the sea bottom reverberation suppression under different signal-to-reverberation ratios is clarified. The present application uses the time-frequency feature enhanced generative adversarial network for sea bottom reverberation suppression, and effectively improves the signal-to-reverberation ratio.

[0007] The present application adopts the following technical solutions:

[0008] A construction method, model and application of a sea bottom reverberation suppression model based on a generative adversarial network, specifically including the following steps:

[0009] Step one, simulate reverberation signals and target echo signals, superimpose to form a mixed signal, and form a data pair with the simulated target echo signal to provide data preparation for the network;

[0010] Step two, perform short-time Fourier transform on the mixed signal and the pure target echo signal to generate a signal spectrum diagram, which contains an amplitude spectrum and a phase spectrum; the amplitude spectrum is introduced into the subsequent network training, and the original phase spectrum of the mixed signal is retained;

[0011] Step three, train the TFE-GAN network architecture, which includes a generator and a discriminator; the generator is composed of an encoder-feature enhancement module-decoder three-level architecture, the encoder and the decoder are composed of VGG modules, the feature enhancement module integrates four key components of a channel attention branch, a weak signal enhancement branch, a spatial attention branch and an adaptive enhancement mechanism, and realizes sea bottom reverberation suppression and target echo signal extraction through the generator; the discriminator adopts four VGG modules and introduces a skip connection, which is used to distinguish the actual pure target echo signal from the reverberation suppression signal generated by the generator;

[0012] Step four, the loss function in the generator contains three parts, namely the adversarial loss function, the amplitude spectrum reconstruction loss function and the focal spectrum loss function, and the comprehensive loss function is formed by setting different weights for the three parts to optimize the output signal and adjust the network parameter training; the discriminator discriminates the input signal of the generator and the pure target echo signal through the discriminator loss function to achieve the comprehensive seabed reverberation suppression effect;

[0013] Step five, combine the amplitude spectrum output by the generator with the original phase spectrum, and use the inverse short-time Fourier transform to reconstruct the time-domain signal.

[0014] Further, the encoder is composed of four VGG modules, each module containing two 3x3 convolution layers, LeakyReLU activation function and optional batch normalization layer; the encoding process starts from the first VGG module, which maps the single-channel amplitude spectrum to a 64-dimensional feature space, and this module does not use batch normalization to preserve the original signal features; then, the second, third and fourth VGG modules sequentially increase the feature dimension to 128, 256 and 512 channels respectively, forming a complete encoder structure containing 8 convolution layers; all convolution layers use padding=1 setting.

[0015] Further, the feature enhancement module: the channel attention branch sequentially passes through an adaptive average pooling layer, a convolution layer with channel number from 512 to 32, a RELU function, a convolution layer with channel number from 32 to 512, and a sigmoid function to obtain the channel attention weight; the spatial attention branch passes through a 7x7 convolution layer with channel number from 512 to 1, and then passes through a sigmoid function to obtain the spatial attention weight; the weak signal enhancement branch sequentially passes through a 3x3 convolution layer (unchanged channel number), a LeayReLU function, a 3x3 convolution layer (unchanged channel number), and a sigmoid function to obtain the weak signal enhancement weight; the adaptive enhancement mechanism calculates the signal intensity to obtain a Gaussian-type enhancement factor; finally, a residual connection is added to achieve the overall feature enhancement effect.

[0016] Further, the decoder is also composed of four VGG modules, and the decoding process starts from the 512-channel deep features output by the feature enhancement module, first enters the first decoding VGG module, reduces the channel number from 512 to 256, and keeps the spatial dimension unchanged; the subsequent decoding process introduces a jump connection mechanism, splices the features of the corresponding layers of the encoder and the current features of the decoder in the channel dimension, and forms feature complementation; the second decoding VGG module receives the spliced features and reduces the channel number from 512 to 128; the third decoding VGG module processes the spliced 256-channel features and outputs a 64-channel feature map; the fourth decoding VGG module continues to process the spliced 128-channel features and generates a 32-channel feature map; finally, the output layer uses a 1*1 convolution to reduce the feature channel number from 32 to 1, and passes through a ReLU activation function to ensure that the output is non-negative, as the enhanced amplitude spectrum.

[0017] Further, the weight coefficients of the generator adversarial loss function, the amplitude spectrum reconstruction loss function and the focal spectrum loss function are 0.2, 0.6 and 0.2 respectively; the generator comprehensive loss function L G is the weighted sum of the adversarial loss function, the amplitude spectrum reconstruction loss function and the focal spectrum loss function:

[0018] L G = 0.2·L adv + 0.6·L recon + 0.2L focal

[0019] In the formula, L G is the comprehensive loss function, L adv is the adversarial loss function, L recon is the amplitude spectrum reconstruction loss function, and L focal is the focal spectrum loss function.

[0020] A seabed reverberation suppression model based on a generative adversarial network, the model comprises a data processing module, a generative adversarial module and a signal reconstruction module;

[0021] The data processing module runs steps one and two in the method;

[0022] The generative adversarial module runs steps three and four in the method;

[0023] The signal reconstruction module runs step five in the method.

[0024] The application also provides an application of the model constructed by the method, and the application is to process a mixed signal by using the model to suppress seabed reverberation.

[0025] Compared with the prior art, the application has the following beneficial effects:

[0026] The present application is directed to the problem of bottom reverberation background interference in the working process of current active sonar system, and proposes a method for suppressing bottom reverberation. Considering the rough scattering of the seabed, a bottom reverberation model is constructed from the physical mechanism, the time-domain waveform simulation of the bottom reverberation is carried out, and the statistical characteristic analysis of the simulation signal is carried out; the bright spot model is used to simulate the target echo signal; the deep learning method is combined with the reverberation suppression, and a bottom reverberation suppression method based on time-frequency feature enhancement of the generative adversarial network is proposed, which realizes the effective suppression of the bottom reverberation. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a technical scheme comprehensive architecture diagram;

[0028] Figure 2 is a data processing module flow chart;

[0029] Figure 3 is a feature enhancement module architecture diagram;

[0030] Figure 4 is a bottom reverberation suppression result diagram under high signal-to-mixture ratio conditions, (a) mixed signal (b) no feature enhancement module processing result (c) feature enhancement module processing result;

[0031] Figure 5 is a bottom reverberation suppression result diagram under low signal-to-mixture ratio conditions, (a) mixed signal (b) target echo signal (c) feature enhancement module processing result (d) no feature enhancement module processing result;

[0032] Figure 6 is a bottom reverberation suppression result diagram when SRR=10dB, (a) mixed signal (b) target echo signal (c) no feature enhancement module processing result (d) feature enhancement module processing result;

[0033] Figure 7 is a bottom reverberation suppression result diagram when SRR=5dB, (a) mixed signal (b) target echo signal (c) no feature enhancement module processing result (d) feature enhancement module processing result;

[0034] Figure 8 is a bottom reverberation suppression result diagram when SRR=0dB, (a) mixed signal (b) target echo signal (c) feature enhancement module processing result;

[0035] Figure 9 is a bottom reverberation suppression result diagram when SRR=-3dB, (a) mixed signal (b) target echo signal (c) feature enhancement module processing result;

[0036] Figure 10is the reverberation suppression result figure when SRR = -5dB, (a) mixed signal (b) target echo signal (c) feature enhancement module processing result. DETAILED DESCRIPTION

[0037] The technical solutions of the present application are further explained below.

[0038] The implementation process of the seabed reverberation suppression method proposed in the present application mainly includes six steps: (1) a single base seabed reverberation model is established using normal mode theory and small slope approximation theory; (2) the accuracy of the simulation signal is verified by simulating and analyzing the statistical characteristics of the seabed reverberation time domain waveform using the seabed reverberation model; (3) the target echo signal is simulated using the bright spot model, and a mixed signal is constructed according to a certain signal-to-mixed ratio, and a data set is formed together with the pure target echo signal; (4) the TFE-GAN network is trained, the training set is used to train the network, and the network parameters are continuously adjusted to improve the network performance through the comprehensive loss function; (5) ablation experiments are carried out to verify the effectiveness of the feature enhancement module; (6) the data under different signal-to-mixed ratios are analyzed to verify the reverberation suppression effect of the network.

[0039] Figure 1 The technical solution is a comprehensive architecture diagram, and the implementation process of the present application will be described below.

[0040] (1) a single base seabed reverberation model is established using normal mode theory and small slope approximation theory.

[0041] The reverberation field at the receiver is calculated by normal mode theory

[0042]

[0043] In the formula, r is the distance; z is the depth; Z m (z s ) is the eigenfunction of the mth order normal mode at the sound source z s ; Z n (z r ) is the eigenfunction of the nth order normal mode at the receiver depth z r ; M and N are the order of the normal mode; A m (z b ) and A n (z b ) are the incident wave amplitudes of the mth and nth order normal modes at the seabed depth z b ; and and are the phases caused by the propagation of the mth and nth order normal modes at the seabed depth z b ; k m and k n are the real parts of the horizontal wave numbers of the mth and nth order normal modes; Km and K n are eigenvalues of the mth and nth order normal modes, respectively; g mn is the ratio of the scattered pressure of a plane wave per unit area to the incident pressure of the plane wave.

[0044] Bottom roughness scattering function σ mn calculated by the small slope approximation theory

[0045]

[0046] where k0 is the wave number in seawater; I K is the Kirchhoff integral; Δk z = k sz - k iz represents the vertical wave number difference, which describes the wave number change of the incident normal mode and the scattered normal mode in the vertical direction; ΔK = K s - K i represents the horizontal wave number difference, which describes the wave number change of the incident normal mode and the scattered normal mode in the horizontal direction.

[0047]

[0048] where V ww represents the reflection coefficient of the seabed interface; θ i is the grazing angle of the incident wave; θ s is the grazing angle of the scattered wave; a ρ represents the ratio of the seabed density to the seawater density; a p represents the ratio of the seabed longitudinal sound speed to the seawater sound speed.

[0049] (II) Simulation of seabed reverberation signals and analysis of statistical characteristics.

[0050] The scattered sound field can be regarded as the channel transfer function of reverberation. After Fourier transform, if the frequency spectrum of the transmitted signal is S(w), the time-domain reverberation signal caused by the scattering element of the seabed under the action of this signal can be obtained

[0051]

[0052] where w is the angular frequency of the transmitted signal, t is time, and P rev is the reverberation signal.

[0053] After the simulation of the seabed reverberation signal, statistical characteristic analysis of the reverberation signal was performed. The theoretical formula of the instantaneous value V(t) of the reverberation signal is

[0054]

[0055] where W is the probability density, is the variance of the instantaneous value of the reverberation signal, and V is the instantaneous value of the reverberation signal.

[0056] The theoretical formula of reverberation signal envelope is

[0057]

[0058] In the formula, W is the probability density, is the variance of reverberation signal amplitude, and E is the envelope amplitude of reverberation signal.

[0059] Finally, the correlation of the reverberation signal is analyzed by the autocorrelation function. If a reverberation is x(t), its autocorrelation function is

[0060]

[0061] In the formula, δ is the normalization coefficient; τ represents the time delay.

[0062] (Three) The bright spot model is used to simulate the target echo signal, and the mixed signal is formed by superimposing the sea bottom reverberation signal.

[0063] Suppose the target is composed of N scattering points, and the echo signal s i (t) of each scattering point can be represented as

[0064]

[0065] In the formula, α is the reflection coefficient of the scattering point; w0is the center frequency of the signal; φ i is the initial phase of each scattering point; τ is the propagation time delay, which is calculated according to the distance between the scattering point and the receiving point.

[0066] The target echo signal S(t) is the superposition of all scattering point echo signals

[0067]

[0068] The simulated sea bottom reverberation signal and the target echo signal are used to construct the training and test data sets. The training set contains 7200 data pairs, and the test set contains 1350 data pairs, covering the following parameters:

[0069] The training set is composed of 36 linear frequency modulation signals, including 4 center frequencies (300 Hz, 500 Hz, 800 Hz, 1000 Hz, etc.), 3 bandwidths (50 Hz, 100 Hz, 150 Hz), and 3 pulse widths (0.5s, 1s, 1.5s). The target echo considers two cases (single bright spot, double bright spot), and a total of 72 pure target echo signals are generated. Each target echo generates 20 random reverberation instances, which are mixed under 5 signal-to-reverberation ratio conditions (10 dB, 5 dB, 0 dB, -3 dB, -5 dB) to form a total of 7200 training data sets.

[0070] The test set consists of 27 linear frequency modulation signals, including 3 center frequencies (300 Hz, 500 Hz, 800 Hz, etc.), 3 bandwidths (50 Hz, 100 Hz, 150 Hz), and 3 pulse widths (0.5 s, 1 s, 1.5 s). The target echo considers two cases (single bright spot, double bright spot), a total of 54 pure target echo signals are generated. Five random reverberation instances are generated for each target echo, and 5 signal-to-reverberation ratio conditions (10 dB, 5 dB, 0 dB, -3 dB, -5 dB) are mixed to form a total of 2700 pairs of training data sets.

[0071] (iv) Training the TFE-GAN network using the training set.

[0072] ① First, the data processing module, the processing flow is as shown in Figure 2 The short-time Fourier transform is performed on the reverberation signal and the pure target echo signal to generate the amplitude spectrum and the phase spectrum. The amplitude spectrum is imported into the subsequent training process and is dynamically compressed by logarithmic operation. The phase spectrum is reserved for subsequent time-domain waveform recovery:

[0073]

[0074] where STFT x (τ,f) represents the short-time Fourier transform of the signal x(t); τ represents the sliding position of the window function on the time axis, which determines the time point of local analysis of the signal; f represents the frequency, which reflects the frequency component of the signal in the local time; e -j2πft is the complex exponential function of f, which is used to transform the windowed signal from the time domain to the frequency domain, and the spectral component at time τ and frequency f is calculated by integration.

[0075] ② Next is the generative adversarial module, which consists of a generator and a discriminator. The generator consists of an encoder, a feature enhancement module, and a decoder. The amplitude spectrum processed by the data processing module is imported into the network for training. When the amplitude spectrum is input into the generator, the processed amplitude spectrum is obtained by feature extraction. The specific process is as follows:

[0076] Encoder: It is composed of four VGG modules, each of which contains two 3x3 convolutional layers, LeakyReLU activation functions, and optional batch normalization layers. The encoding process starts with the first VGG module, which maps the single-channel amplitude spectrum to a 64-dimensional feature space. This module does not use batch normalization to preserve the original signal features. Subsequently, the second, third, and fourth VGG modules sequentially increase the feature dimensions to 128, 256, and 512 channels, respectively, forming a complete encoder structure containing eight convolutional layers. All convolutional layers use padding = 1 settings to ensure that the feature map space size remains unchanged throughout the encoding process, thereby gradually extracting deep abstract features of the amplitude spectrum through a series of convolution, batch normalization, and activation operations.

[0077] Feature enhancement module: It is composed of four key components: channel attention branch, weak signal enhancement branch, spatial attention branch, and adaptive enhancement mechanism, as shown in the architectural diagram Figure 3 The channel attention branch sequentially passes through an adaptive average pooling layer, a convolution layer with a channel number from 512 to 32, a RELU function, a convolution layer with a channel number from 32 to 512, and a sigmoid function to obtain the channel attention weight. The spatial attention branch passes through a 7x7 convolution layer with a channel number from 512 to 1, and then through a sigmoid function to obtain the spatial attention weight. The weak signal enhancement branch sequentially passes through a 3x3 convolution layer (with unchanged channel number), a LeayReLU function, a 3x3 convolution layer (with unchanged channel number), and a sigmoid function to obtain the weak signal enhancement weight. The adaptive enhancement mechanism mainly calculates the Gaussian-type enhancement factor based on the signal intensity. Finally, a residual connection is added to achieve the overall feature enhancement effect.

[0078] Decoder: also consists of four VGG modules, the decoding process starts from the 512-channel deep features output by the feature enhancement module, first enters the first decoding VGG module, reduces the channel number from 512 to 256, and keeps the spatial dimension unchanged. The subsequent decoding process introduces a jump connection mechanism, which splices the features of the corresponding layers of the encoder and the current features of the decoder in the channel dimension to form complementary features. The second decoding VGG module receives the spliced features (256+256=512 channels) and reduces them to 128 channels; the third decoding VGG module processes the spliced features (128+128=256 channels) and outputs a 64-channel feature map; the fourth decoding VGG module continues to process the spliced features (64+64=128 channels) and generates a 32-channel feature map. Through this layer-by-layer decoding and feature fusion method, the entire decoder effectively combines high-level semantic information and shallow detail information, thereby retaining the key features of the signal. Finally, the output layer uses a 1x1 convolution to reduce the number of feature channels from 32 to 1, and uses a ReLU activation function to ensure that the output is non-negative, serving as the enhanced amplitude spectrum.

[0079] A multi-level generator synthesis loss function is designed. The generator adversarial loss function, the amplitude spectrum reconstruction loss function and the focus spectrum loss function are combined to optimize the output signal, and the comprehensive seabed reverberation suppression effect is realized.

[0080] Adversarial loss function L adv : the generator minimizes the mean square error between the generated sample discrimination result and the all-1 label to deceive the discriminator into thinking that the generated sample is real data

[0081]

[0082] Amplitude spectrum reconstruction loss function: in this embodiment, the amplitude spectrum reconstruction loss function adopts the L1 loss function. Compared with the L2 loss function, the L1 loss function has stronger robustness to abnormal values, and can avoid the problem of unstable training caused by a small number of abnormal points. In addition, in the generation of the spectrum graph, it can better restore the details and structure of the spectrum graph. By minimizing the absolute error between the de-reverberation amplitude spectrum output by the generator and the pure target echo signal amplitude spectrum, it ensures that the generator can accurately reconstruct the main structure and energy distribution of the amplitude spectrum

[0083]

[0084] In the formula, G(z i ) is the de-reverberation amplitude spectrum output by the generator, and y i is the pure target echo signal amplitude spectrum.

[0085] Focal spectrum loss function: In order to solve the problem that the target echo signal is too weak to be identified, the focal spectrum loss is introduced. This loss function gives higher weight to the low-energy area (containing useful information) to focus on the area with low true amplitude. The focal spectrum loss L focal is as follows

[0086]

[0087] where M pred and M ture represent the predicted and true amplitude spectrum respectively; α = 0.25 is the balance factor; γ = 2 is the focusing parameter.

[0088] Generator comprehensive loss function: In order to balance the various loss functions of the generator, weight coefficients are added for adjustment. Through various strategies such as prior knowledge and gradual adjustment, the weight coefficients are finally determined as 0.2, 0.6 and 0.2. The generator comprehensive loss function L G is the weighted sum of the adversarial loss function, the amplitude spectrum reconstruction loss function and the focal spectrum loss

[0089] L G = 0.2·L adv + 0.6·L recon + 0.2L focal

[0090] Discriminator loss function L D : By minimizing the mean square error between the discrimination result of the real sample (pure target echo signal amplitude spectrum) and the all-1 label, and the mean square error between the discrimination result of the generated sample (de-reverberation signal amplitude spectrum) and the all-0 label, the discriminator is forced to accurately distinguish between real and generated samples

[0091]

[0092] where E represents the mathematical expectation; z ~ p z , z represents a random vector input into the generator, which is sampled from the probability distribution p z ; D represents the discriminator; D(x) represents the output of the discriminator for real data samples; G represents the generator; G(z) represents the generated fake sample after receiving the vector; x ~ p date , x represents a real data sample, which comes from the real data probability distribution p date .

[0093] ③ Finally, the signal reconstruction module, the specific implementation process is: the amplitude spectrum processed by the generator is fused with the original signal phase spectrum to form a complete de-reverberation signal spectrum, and the inverse short-time Fourier transform is used to restore the time-domain signal.

[0094] The formula of the inverse short-time Fourier transform is as follows:

[0095]

[0096] where x[n] is the reconstructed time-domain signal; x(m, k) is the complex spectrum of the mth frame and the kth frequency bin; N is the number of points of the discrete Fourier transform; M is the total number of frames; w[n] is a window function; and H is a frame shift.

[0097] (V) Ablation experiments are carried out to verify the effectiveness of the feature enhancement module.

[0098] The network with and without the feature enhancement module is trained respectively, and detection analysis is carried out under different signal-to-reverberation ratios. The following indicators are introduced for evaluation: signal-to-reverberation ratio improvement, mean square error, correlation coefficient, peak signal-to-noise ratio improvement, and visual analysis of the spectrum and time-domain waveform.

[0099] Table 1 Comparison results of relevant indicators under high signal-to-reverberation ratio conditions

[0100]

[0101] Under high signal-to-reverberation ratio conditions, from Table 1 and Figure 4 It can be seen that the network with the feature enhancement module is cleaner in amplitude spectrum and has better peak signal-to-noise ratio improvement. However, in the time-domain waveform, it is worse than the network without the feature enhancement module. Therefore, in the target positioning and detection, the advantage of the feature enhancement module is significant, and it can achieve ideal reverberation suppression effect; while in the application scenario focusing on the amplitude information of the time-domain signal, the module can be considered to be closed.

[0102] Under low signal-to-reverberation ratio conditions, as shown in Figure 5 the network with the feature enhancement module can still effectively identify the target echo signal and accurately suppress the reverberation under low signal-to-reverberation ratio conditions; while the network without the feature enhancement module cannot correctly identify the target echo signal and will mistake the target echo as reverberation in the reverberation suppression process.

[0103] In summary, under high signal-to-reverberation ratio conditions, the reverberation suppression effect of the network with and without the feature enhancement module is good, the network with the feature enhancement module performs better in amplitude spectrum and has cleaner processing, but due to the feature enhancement that increases the target echo capability, the reconstructed time-domain waveform amplitude is slightly distorted; under low signal-to-reverberation ratio conditions, the feature enhancement module must be used, but due to the amplitude distortion, only the reverberation suppression situation in the amplitude spectrum is concerned at this time.

[0104] (VI) Data under different signal-to-reverberation ratio conditions are analyzed to verify the reverberation suppression effect of the network.

[0105] There are 540 groups of signals in each group of signal-to-mixed ratio (10 dB, 5 dB, 0 dB, -3 dB, -5 dB). In the high signal-to-mixed ratio conditions (10 dB, 5 dB), the network with and without the feature enhancement module is used for processing, and in the low signal-to-mixed ratio conditions (0 dB, -3 dB, -5 dB), only the network with the feature enhancement module is used for processing.

[0106] Table 2 Comparison results of relevant indicators under different signal-to-mixed ratio conditions

[0107]

[0108] Figure 6-10 The reverberation suppression results under different signal-to-mixed ratio conditions are given, and the quantitative results are given in Table 2. It can be seen that in the case of signal-to-mixed ratio 10 dB and 5 dB, the average signal-to-mixed ratio of TFE-GAN is improved by 5.07 dB and 5.97 dB, the target echo recognition rate is 100%, the amplitude spectrum suppression is clean, and the time domain signal structure restoration degree is high. In the case of signal-to-mixed ratio 0 dB, the network still has good reverberation suppression effect in the amplitude spectrum, but due to the feature enhancement module, the signal amplitude is increased, and the network with the feature enhancement module performs worse than the network without the feature enhancement module in the signal-to-mixed ratio improvement, with an average signal-to-mixed ratio improvement of 2.80 dB. In the case of low signal-to-mixed ratio -3 dB and -5 dB, the network still has good reverberation suppression effect in the amplitude spectrum.

Claims

1. A method for constructing a seabed reverberation suppression model based on a generative adversarial network, characterized in that: The method specifically comprises the following steps: Step 1: Superimpose the simulated reverberation signal and the target echo signal to form a mixed signal, and form a data pair with the simulated target echo signal to provide data preparation for the network; Step 2: Perform short-time Fourier transform on the mixed signal and the pure target echo signal to generate a signal spectrum diagram, which contains the amplitude spectrum and the original phase spectrum; import the amplitude spectrum into subsequent network training, and retain the original phase spectrum of the mixed signal; Step 3: Train the architecture of a generative adversarial network based on time-frequency feature enhancement, which includes a generator and a discriminator. The generator consists of a three-level architecture: encoder-feature enhancement module-decoder. The encoder and decoder are composed of VGG modules. The feature enhancement module integrates four key components: channel attention branch, weak signal enhancement branch, spatial attention branch, and adaptive enhancement mechanism. The generator realizes seabed reverberation suppression and target echo signal extraction. The discriminator uses four VGG modules and introduces jump connections to distinguish between pure target echo signals and reverberation-suppressed signals generated by the generator. The feature enhancement module: the channel attention branch passes through an adaptive average pooling layer, a convolution layer with the number of channels from 512 to 32, a Relu function, a convolution layer with the number of channels from 32 to 512, and a sigmoid function in sequence to obtain the channel attention weight; the spatial attention branch passes through a 7×7 convolution layer with the number of channels from 512 to 1, and then passes through a sigmoid function to obtain the spatial attention weight; the weak signal enhancement branch passes through a 3×3 convolution layer, a LeayReLU function, a 3×3 convolution layer, and a sigmoid function in sequence to obtain the weak signal enhancement weight; the adaptive enhancement mechanism obtains a Gaussian enhancement factor by calculating the signal strength; and finally, a residual connection is added to achieve the overall feature enhancement effect; Step 4: The loss function in the generator consists of three parts: the adversarial loss function, the amplitude spectrum reconstruction loss function, and the focal spectrum loss function. Different weights are set for these three parts to form a comprehensive loss function to optimize the output signal and adjust the network parameter training. The discriminator uses the discriminator loss function to distinguish the generator input signal from the pure target echo signal to achieve seabed reverberation suppression. Step 5: Combine the amplitude spectrum output by the generator with the original phase spectrum and use the inverse short-time Fourier transform to reconstruct the time domain signal.

2. The method according to claim 1, characterized in that The encoder consists of four VGG modules, each of which contains two 3×3 convolutional layers, a LeakyReLU activation function, and an optional batch normalization layer. The encoding process starts with the first VGG module, which maps the single-channel amplitude spectrum to a 64-dimensional feature space. This module does not use batch normalization to preserve the original signal characteristics. Subsequently, the second, third, and fourth VGG modules increase the feature dimension to 128, 256, and 512 channels, respectively, forming a complete encoder structure containing 8 convolutional layers. All convolutional layers use a padding = 1 setting.

3. The method according to claim 1, characterized in that The decoder is also composed of four VGG modules. The decoding process starts with the 512-channel deep features output by the feature enhancement module. It first enters the first decoding VGG module to reduce the number of channels from 512 to 256, keeping the spatial dimension unchanged. The subsequent decoding process introduces a skip connection mechanism to splice the features of the corresponding layer of the encoder with the current features of the decoder in the channel dimension to form feature complementarity. The second decoding VGG module receives the concatenated features and reduces the number of channels from 512 to 128; the third decoding VGG module processes the concatenated 256-channel features and outputs a 64-channel feature map; the fourth decoding VGG module continues to process the concatenated 128-channel features and generates a 32-channel feature map; finally, the output layer uses 1×1 convolution to reduce the number of feature channels from 32 to 1, and uses the ReLU activation function to ensure that the output is non-negative as the enhanced amplitude spectrum.

4. The method according to claim 1, wherein The weight coefficients of the generator adversarial loss function, amplitude spectrum reconstruction loss function and focal spectrum loss function are 0.2, 0.6 and 0.2 respectively; the generator comprehensive loss function L G That is the weighted sum of the adversarial loss function, the amplitude spectrum reconstruction loss function, and the focal spectrum loss function: L G =0.2·L adv +0.6 L recon +0.2L focal Where, L G is the comprehensive loss function, L adv To counter the loss function, L recon is the amplitude spectrum reconstruction loss function, L focal is the focal spectrum loss function.

5. A seabed reverberation suppression system based on generative adversarial networks, characterized in that: The system includes a data processing module, a generative adversarial module and a signal reconstruction module; The data processing module executes step 1 and step 2 of the method according to any one of claims 1 to 4; The generative adversarial module executes steps 3 and 4 of the method according to any one of claims 1 to 4; The signal reconstruction module executes step 5 of the method according to any one of claims 1 to 4.

6. Use of the system according to claim 5, characterized in that: The system is used to process the mixed signal to suppress seabed reverberation.

Images