Underwater acoustic signal denoising method and system based on generative adversarial network

Abstract

The application discloses a method and system for underwater acoustic signal denoising based on a generative adversarial network. Through the division and cooperation of the feature separation branch and the signal denoising branch, the network has the advantages of interpretability and modularity. Meanwhile, the separation loss function of the multi-loss function is combined, which not only ensures the efficient separation of the target signal and the interference feature, but also strengthens the physical authenticity of the underwater acoustic signal through the true and false discrimination of the discriminator. The designed signal denoising branch guarantees real-time processing through the design of the causal convolution module and the adaptive forget gate module, and can cope with the complexity of the time-varying channel to ensure real-time noise reduction application.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic communication technology, specifically to a method and system for denoising underwater acoustic signals based on generative adversarial networks. Background Technology

[0002] As a core carrier of underwater information transmission, underwater acoustic signals are widely used in various scenarios such as marine resource exploration, underwater vehicle communication, marine environmental monitoring, and underwater sonar detection. However, the underwater acoustic channel is a typical complex time-varying channel. The non-uniformity of sound velocity in the seawater medium, multipath propagation effects, Doppler frequency shift, and various interferences in the marine environment such as ship mechanical noise, ocean current turbulence noise, underwater biological noise, and equipment thermal noise can severely contaminate the underwater acoustic signal during transmission and reception. The signal-to-noise ratio of noisy underwater acoustic signals is low, and effective target information is masked by noise, which greatly reduces the accuracy of subsequent signal processing, analysis, and transmission, becoming a core bottleneck restricting the efficiency of underwater information interaction.

[0003] Generative Adversarial Networks (GANs), through their adversarial training mechanism between generators and discriminators, can learn the true distribution characteristics of signals and demonstrate excellent performance in denoising tasks in fields such as image and speech. However, when applied to underwater acoustic signal denoising, existing solutions mostly directly transplant general network structures without customizing them to address the non-stationarity, time-varying nature of underwater acoustic signals, and the engineering requirements of low computing power and real-time processing underwater. This results in problems such as mismatch between feature extraction and underwater acoustic signal characteristics, loss of target signal during denoising, and poor model real-time performance, failing to meet the high-precision and high-robustness denoising requirements of actual underwater scenarios.

[0004] Therefore, there is an urgent need for a generative adversarial network denoising method that is customized for underwater acoustic signals and channel characteristics. Through precise feature extraction and separation and network structure design adapted to time-varying channels, it can effectively suppress composite noise while preserving the key information of the target underwater acoustic signal to the maximum extent. It can balance denoising accuracy, robustness and engineering feasibility, and solve the technical problems of poor adaptability, poor denoising effect and insufficient real-time performance of existing underwater acoustic signal denoising methods. Summary of the Invention

[0005] To address these issues, this invention proposes a method and system for denoising underwater acoustic signals based on generative adversarial networks.

[0006] According to one aspect of the present invention, a method for denoising underwater acoustic signals based on generative adversarial networks is proposed, comprising the following steps:

[0007] S1, acquire the target underwater acoustic signal and the corresponding noisy underwater acoustic signal after transmission through the underwater acoustic channel, and input the noisy underwater acoustic signal and the target underwater acoustic signal into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator.

[0008] S2, the generator includes a feature separation branch and a signal denoising branch. The feature separation branch includes a feature extraction sub-branch and a feature separation sub-branch. The feature extraction sub-branch performs multi-scale feature extraction on the noisy underwater acoustic signal and outputs high-dimensional mixed features. The feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output target signal features. The signal denoising branch generates a denoised underwater acoustic signal based on the target signal features. The discriminator distinguishes between the target underwater acoustic signal and the denoised underwater acoustic signal, and performs adversarial training with the generator based on the discrimination result.

[0009] S3, Noise reduction is performed on the noisy underwater acoustic signal based on the trained generator.

[0010] Specifically, in S2, the feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output the target signal features and environmental interference features. The feature separation sub-branch includes two parallel-connected feature separation modules. The two parallel-connected feature separation modules are pre-trained using the target signal standard features and the environmental interference standard features to obtain the initial network parameters. Then, the network parameters are updated based on the feature separation results of the feature separation sub-branch through a separation loss function.

[0011] Specifically, the separation loss function includes:

[0012] A weighted combination of a feature loss function that calculates the mean square error of the target signal features and environmental interference features with their corresponding standard features, a non-Gaussian loss function that calculates the negative entropy of the target signal features and environmental interference features, and an independence loss function that calculates the mutual information between the target signal features and environmental interference features.

[0013] Specifically, the separation loss function Defined as: ,in, , and These represent adjustable weights;

[0014] The feature loss function is defined as follows:

[0015] ,in, and These represent the corresponding weight coefficients. Indicates the number of training samples in a single batch. and They represent the first The target signal characteristics and environmental interference characteristics of each sample and They represent the first Standard characteristics of the target signal and standard characteristics of environmental interference for each sample;

[0016] The non-Gaussian loss function is defined as follows: ,in, Representing computational features Negative entropy, defined ,in Indicates the calculation of nonlinear functions The mathematical expectation, Representation and Features Homoscedastic Gaussian random variables and They represent the first Gaussian random variables representing the target signal characteristics and Gaussian random variables representing the environmental interference characteristics of each sample;

[0017] The independence loss function is defined as follows: ,in, Representing computational features and characteristics The mutual information value.

[0018] Specifically, the feature extraction sub-branch frames the noisy underwater acoustic signal according to symbols to obtain the corresponding time-series matrix. Then, it performs a short-time Fourier transform to obtain the corresponding time-frequency real part matrix and time-frequency imaginary part matrix. The time-series matrix, the time-frequency real part matrix, and the time-frequency imaginary part matrix are concatenated to obtain a dual-modal input. The dual-modal input is processed sequentially through three multi-scale convolutional layers and one max-pooling layer with a 2×2 pooling kernel to obtain the high-dimensional hybrid features. The first multi-scale convolutional layer uses a 3×3 convolutional kernel to extract symbol detail features, the second multi-scale convolutional layer uses a 5×5 convolutional kernel to extract multipath trend features, and the third multi-scale convolutional layer uses a 7×7 convolutional kernel to extract channel global features. Each multi-scale convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function.

[0019] Specifically, the signal denoising branch includes a causal convolution module, an adaptive forget gate module, and a deconvolution processing module. The causal convolution module performs feature extraction and processing based on the target signal features. The adaptive forget gate module generates historical feature retention weights based on the real-time estimated channel change rate and performs time-series modeling on the features output by the causal convolution module at the current time and the features output at the previous time to generate optimized time-series features. The deconvolution processing module generates the denoised underwater acoustic signal based on the optimized time-series features.

[0020] Specifically, the causal convolution module includes two causal convolutional layers. The first causal convolutional layer has a stride of 1, a left padding of 1, and a kernel size of 3×3. It is used to extract boundary features from the target signal features. The second causal convolutional layer has a stride of 1, a left padding of 1, and a kernel size of 3×3. It is used to further extract local correlation features. Each causal convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function, and a residual connection structure is introduced.

[0021] The adaptive forget gate obtains the real-time estimated channel change rate, and calculates the historical feature retention weight based on the Sigmoid function and the channel change rate. The features output by the causal convolution module at the current time are used as the current features. The features output by the causal convolution module at the previous time step are used as historical features. ,based on The optimized temporal features are generated and output to the deconvolution processing module;

[0022] The deconvolution processing module includes two deconvolution layers. The first deconvolution layer has a stride of 2, padding=1, and a kernel size of 3×3. The second deconvolution layer has a stride of 1, padding=1, and a kernel size of 3×3. Each deconvolution layer is followed by a batch normalization layer and a LeakyReLU activation function.

[0023] According to one aspect of the present invention, an underwater acoustic signal denoising system based on generative adversarial networks is proposed, comprising the following modules according to any one of the first aspects:

[0024] The data acquisition module is configured to acquire the target underwater acoustic signal and the corresponding noisy underwater acoustic signal after transmission through the underwater acoustic channel, and input the noisy underwater acoustic signal and the target underwater acoustic signal into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator.

[0025] The network training module is configured such that the generator includes a feature separation branch and a signal denoising branch. The feature separation branch includes a feature extraction sub-branch and a feature separation sub-branch. The feature extraction sub-branch performs multi-scale feature extraction on the noisy underwater acoustic signal and outputs high-dimensional mixed features. The feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output target signal features. The signal denoising branch generates a denoised underwater acoustic signal based on the target signal features. The discriminator distinguishes between the target underwater acoustic signal and the denoised underwater acoustic signal, and performs adversarial training with the generator based on the discrimination results.

[0026] The intelligent noise reduction module is configured to reduce the noise of the noisy underwater acoustic signal based on a trained generator.

[0027] According to one aspect of the present invention, a computer program product is provided having a computer program stored thereon, which, when executed by a processor, performs the method as described in the first aspect.

[0028] According to one aspect of the present invention, an electronic device is provided, comprising: one or more processors; a storage device for storing one or more programs; wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0029] The advantages of this invention are: by dividing and coordinating the feature separation branch and the signal denoising branch, the network has the advantages of interpretability and modularity; at the same time, by combining the separation loss function with multiple loss functions, it not only ensures the efficient separation of the target signal and interference features, but also strengthens the physical authenticity of the underwater acoustic signal through the discriminator's true and false discrimination; the designed signal denoising branch ensures real-time processing through the design of causal convolution module and adaptive forget gate module, which can cope with the complexity of time-varying channels and ensure real-time noise reduction applications. Attached Figure Description

[0030] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0031] Figure 1 A flowchart illustrating a method for denoising underwater acoustic signals based on generative adversarial networks according to the present invention is shown.

[0032] Figure 2 A schematic diagram of an underwater acoustic signal denoising system based on a generative adversarial network according to the present invention is shown.

[0033] Figure 3 A schematic diagram of a computer system architecture suitable for implementing the embodiments of this application is shown. Detailed Implementation

[0034] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0035] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] Figure 1 A method for denoising underwater acoustic signals based on generative adversarial networks is shown, including the following steps:

[0037] S1, acquire the target underwater acoustic signal and the corresponding noisy underwater acoustic signal after transmission through the underwater acoustic channel, and input the noisy underwater acoustic signal and the target underwater acoustic signal into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator.

[0038] S2, the generator includes a feature separation branch and a signal denoising branch. The feature separation branch includes a feature extraction sub-branch and a feature separation sub-branch. The feature extraction sub-branch performs multi-scale feature extraction on the noisy underwater acoustic signal and outputs high-dimensional mixed features. The feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output target signal features. The signal denoising branch generates a denoised underwater acoustic signal based on the target signal features. The discriminator distinguishes between the target underwater acoustic signal and the denoised underwater acoustic signal, and performs adversarial training with the generator based on the discrimination result.

[0039] S3, Noise reduction is performed on the noisy underwater acoustic signal based on the trained generator.

[0040] In a specific embodiment, the feature extraction sub-branch frames the noisy underwater acoustic signal according to symbols. The original timing signal at the receiving end is a one-dimensional time-domain signal with a length of L = m×t, where m represents the number of symbols and t represents the number of sampling points per symbol. The one-dimensional timing signal is framed according to symbol boundaries, directly dividing it into an m×t two-dimensional timing matrix. Then, a short-time Fourier transform is performed to obtain the corresponding real and imaginary time-frequency matrices. The timing matrix, the real and imaginary time-frequency matrices are concatenated to obtain an m×t×3-dimensional dual-modal input.

[0041] The dual-modal input is processed sequentially through three multi-scale convolutional layers and one max-pooling layer with a 2×2 pooling kernel to obtain the high-dimensional hybrid features.

[0042] The core information of the underwater acoustic signal (such as phase transitions, symbol start / end boundaries, and sampling point-level noise) is concentrated within a single symbol. The receptive field of the 3×3 small convolutional kernel just covers 1~2 sampling points, so fine features are not lost. The first multi-scale convolutional layer has a stride of 1, padding=1, an input dimension of m×t×3, and an output dimension of m×t×32.

[0043] The multipath effect in underwater acoustic channels can cause signal delays to overlap within 2-3 adjacent symbols (resulting in nonlinear mixing). The receptive field of a 5×5 convolutional kernel can cover the associated features of 2-3 symbols, accurately extracting the local trends caused by multipath. The second multi-scale convolutional layer has a stride of 1, padding=2, an input dimension of m×t×32, and an output dimension of m×t×64.

[0044] Marine environmental factors (ocean currents, noise) can cause frequency shifts and changes in the overall signal-to-noise ratio (SNR). These global features are distributed across multiple symbol ranges. A 7×7 large convolutional kernel can cover 4-5 symbols in its receptive field, capturing global channel characteristics. The third multi-scale convolutional layer has a stride of 1, padding of 3, an input dimension of m×t×64, and an output dimension of m×t×64. The max-pooling layer has a stride of 2, resulting in a high-dimensional hybrid feature map with a final output dimension of (m / 2)×(t / 2)×64, achieving accurate capture of multi-scale features of underwater acoustic signals. Each multi-scale convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function.

[0045] In a specific embodiment, the feature separation sub-branch in S2 receives the high-dimensional mixed features and then performs feature separation to output the target signal features and environmental interference features. The feature separation sub-branch includes two parallel-connected feature separation modules, which are pre-trained using the target signal standard features and environmental interference standard features respectively to obtain initial network parameters.

[0046] This application proposes a generator architecture that combines feature separation with signal denoising for underwater acoustic signal denoising, representing a fundamental improvement over existing GAN denoising methods. Current generator architectures typically learn the mapping from noisy signals to clean signals directly, essentially operating as a "black box" model lacking explicit modeling of interference components. In contrast, this application sets up a feature separation branch, decoupling the mixed signal features into target signal features and environmental interference features. This gives the generator interpretability and modularity, overcoming the technical problem of information loss due to insufficient separation accuracy by separating the features before denoising.

[0047] In a specific embodiment, the parameters are initialized using prior features of the known target signal. First, the network parameters of the first feature separation module are initialized with the feature vector of the pure target signal, making it inherently sensitive to the phase and amplitude characteristics of the target signal. Then, the parameters of the second feature separation module are initialized with prior features of ocean noise, such as the frequency domain characteristics of ship propeller noise and ocean current noise, making the second feature separation module inherently sensitive to this type of irregular noise. Then, the network parameters are updated based on the feature separation results of the feature separation sub-branch using a separation loss function. Each feature separation module only updates its network parameters.

[0048] When designing the separation loss function, the deviation between the output and the standard features is calculated using cross-entropy loss. The smaller the deviation, the more accurate the separation. Furthermore, since the target signal is artificially generated and possesses non-Gaussian properties, while environmental noise is mostly Gaussian with low non-Gaussianity, this helps distinguish between the target and interference. In addition, the target signal features and environmental interference features are independent and can be strengthened through constraints. Therefore, in a specific embodiment, the separation loss function specifically includes: a feature loss function based on calculating the mean square error between the target signal features and environmental interference features and their corresponding standard features, a non-Gaussianity loss function based on calculating the negative entropy of the target signal features and environmental interference features, and a weighted combination of an independence loss function based on calculating the mutual information between the target signal features and environmental interference features.

[0049] Specifically, the separation loss function Defined as: ,in, , and These represent adjustable weights;

[0050] The feature loss function is defined as follows: ,in, and These represent the corresponding weight coefficients. Indicates the number of training samples in a single batch. and They represent the first The target signal characteristics and environmental interference characteristics of each sample and They represent the first Standard characteristics of the target signal and standard characteristics of environmental interference for each sample;

[0051] Let represent the non-Gaussianity loss function. Non-Gaussianity is measured by negative entropy; the larger the negative entropy, the stronger the non-Gaussianity. The artificially designed target non-Gaussian signal has a fixed phase and amplitude pattern, its statistical distribution deviates from a Gaussian distribution, and its negative entropy value is large. Environmental interference (mostly Gaussian noise) has a very small negative entropy value. Therefore, it is defined as: ,in, Representing computational features Negative entropy, defined ,in Indicates the calculation of nonlinear functions The mathematical expectation, Representation and Features Homoscedastic Gaussian random variables and They represent the first Gaussian random variables representing the target signal characteristics and Gaussian random variables representing the environmental interference characteristics of each sample.

[0052] The independence loss function is defined as follows: ,in, Representing computational features and characteristics The mutual information value.

[0053] In one embodiment, the system further includes an interference feature feedback suppression branch, which includes a lightweight feature encoder and a sigmoid function. The lightweight feature encoder receives environmental interference features output by the two parallel-connected feature separation modules, generates a weight mask map based on the sigmoid function, and then reprocesses the high-dimensional mixed features output by the feature extraction sub-branch based on the weight mask map. The processed high-dimensional mixed features are then input into the feature separation sub-branch for feature separation.

[0054] This application incorporates a temporal modeling module in its signal denoising branch, combining causal convolution and an adaptive forget gate, specifically optimized for the time-varying characteristics and real-time processing requirements of underwater acoustic channels. Existing generative adversarial networks (GANs) typically employ ordinary convolution, neglecting temporal causality and thus failing to support streaming processing. In contrast, this application introduces a temporal modeling module, ensuring that the output time depends solely on historical information, meeting the real-time processing needs of underwater acoustic communication. Crucially, the adaptive forget gate introduced in this application dynamically adjusts the retention weights of historical features based on the real-time estimated channel change rate. When the channel undergoes abrupt changes, the model can quickly discard outdated information, preventing error accumulation. This channel-state-aware temporal modeling mechanism is an innovative design for addressing the drastic time-varying characteristics of underwater acoustic channels.

[0055] In a specific embodiment, the signal denoising branch includes a causal convolution module, an adaptive forget gate module, and a deconvolution processing module. The causal convolution module performs feature extraction and processing based on the target signal features. The adaptive forget gate module generates historical feature retention weights based on the real-time estimated channel change rate and performs temporal modeling on the features output by the causal convolution module at the current time and the features output at the previous time, generating optimized temporal features. The deconvolution processing module generates the denoised underwater acoustic signal based on the optimized temporal features.

[0056] The causal convolution module receives target signal features with a dimension of (m / 2)×(t / 2)×1. This module comprises two causal convolutional layers. The first layer has a stride of 1, left padding of 1, and a kernel size of 3×3, used to extract boundary features from the target signal. The second layer also has a stride of 1, left padding of 1, and a kernel size of 3×3, used to further extract local correlation features. Each causal convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function, and a residual connection structure is introduced. Through two 3×3 causal convolutional layers, the number of channels increases from 1 to 64, accurately extracting boundary features such as symbol phase transitions and amplitude abrupt changes. It also captures local temporal correlations caused by multipath effects, providing highly discriminative features for subsequent denoising. The introduction of residual connections (input directly superimposed on output) adapts to long-term dependency modeling of underwater acoustic signals, avoids gradient decay during deep network training, and ensures the integrity of feature transfer.

[0057] The adaptive forget gate obtains the real-time estimated channel change rate. The channel change rate can be estimated in real time using algorithms such as Kalman filtering, and then the historical feature retention weights are calculated based on the Sigmoid function and the channel change rate. The features output by the causal convolution module at the current time are used as the current features. The features output by the causal convolution module at the previous time step are used as historical features. ,based on Generate optimized time-series features when the channel rate of change... During mutations, redundant historical features are quickly discarded; when the channel change rate... When stable, effective historical features are retained to aid noise reduction, adapting to the time-varying characteristics of the underwater acoustic channel and suppressing feature distortion caused by multipath and Doppler. The optimized temporal features are then output to the deconvolution processing module.

[0058] The deconvolution processing module includes two deconvolution layers. The first deconvolution layer has a stride of 2, padding=1, and a kernel size of 3×3, which enlarges the (m / 2)×(t / 2)×64 dimension feature map to m×t×32, restoring the temporal dimension of the symbols. The second deconvolution layer has a stride of 1, padding=1, and a kernel size of 3×3, which outputs the target underwater acoustic signal format in m×t×1 dimension, preserving the physical authenticity of the signal phase and amplitude. A batch normalization layer and a LeakyReLU activation function are added after each deconvolution layer.

[0059] like Figure 2As shown, according to one aspect of the present invention, an underwater acoustic signal denoising system based on generative adversarial networks is proposed, comprising the following modules according to any one of the first aspects:

[0060] The data acquisition module 201 is configured to acquire the target underwater acoustic signal and the corresponding noisy underwater acoustic signal after transmission through the underwater acoustic channel, and input the noisy underwater acoustic signal and the target underwater acoustic signal into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator.

[0061] The network training module 202 is configured such that the generator includes a feature separation branch and a signal denoising branch. The feature separation branch includes a feature extraction sub-branch and a feature separation sub-branch. The feature extraction sub-branch performs multi-scale feature extraction on the noisy underwater acoustic signal and outputs high-dimensional mixed features. The feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output target signal features. The signal denoising branch generates a denoised underwater acoustic signal based on the target signal features. The discriminator distinguishes between the target underwater acoustic signal and the denoised underwater acoustic signal, and performs adversarial training with the generator based on the discrimination results.

[0062] The intelligent noise reduction module 203 is configured to reduce the noise of the noisy underwater acoustic signal based on a trained generator.

[0063] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system 300 suitable for implementing electronic devices according to embodiments of the present application. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0064] like Figure 3 As shown, the computer system 300 includes a central processing unit (CPU) 301, which performs various appropriate actions and processes based on programs stored in read-only memory (ROM) 302 or programs loaded from storage section 309 into random access memory (RAM) 304. The RAM 304 also stores various programs and data required for the operation of the system 300. The CPU 301, ROM 302, ROM 303, and RAM 304 are interconnected via a bus 305. An input / output (I / O) interface 306 is also connected to the bus 305.

[0065] The following components are connected to I / O interface 306: an input section 307 including a keyboard, mouse, etc.; an output section 308 including a liquid crystal display (LCD) and speakers, etc.; a storage section 309 including a hard disk, etc.; and a communication section 310 including a network interface card such as a LAN card and a modem, etc. The communication section 310 performs communication processing via a network such as the Internet. A drive 311 is also connected to I / O interface 306 as needed. A removable medium 312, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 311 as needed so that computer programs read from it can be installed into storage section 309 as needed.

[0066] Specifically, according to embodiments of this disclosure, the processes described above with reference to the flowcharts are implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program is downloaded and installed from a network via communication section 310, and / or installed from removable medium 312. When the computer program is executed by central processing unit (CPU) 301, it performs the functions defined in the methods of this application.

[0067] It should be noted that the computer-readable storage medium of this application is a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium is, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium is any tangible medium that contains or stores a program used by or in connection with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium includes a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals take various forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium or any computer-readable storage medium other than a computer-readable storage medium may transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wireline, optical fiber, RF, etc., or any suitable combination thereof.

[0068] Computer program code for performing the operations of this application is written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code executes entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer is connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or connected to an external computer (e.g., via the Internet using an Internet service provider).

[0069] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram represents a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually execute substantially in parallel, and they may sometimes execute in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, is implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0070] The modules described in the embodiments of this application are implemented in software or hardware.

[0071] In another aspect, this application also provides a computer-readable storage medium included in the electronic device described in the above embodiments; it also exists independently and is not assembled into the electronic device. The computer-readable storage medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: S1, acquire a target underwater acoustic signal and a corresponding noisy underwater acoustic signal transmitted through an underwater acoustic channel, and input the noisy underwater acoustic signal and the target underwater acoustic signal into a generative adversarial network for training, wherein the generative adversarial network includes a generator and a discriminator;

[0072] S2, the generator includes a feature separation branch and a signal denoising branch. The feature separation branch includes a feature extraction sub-branch and a feature separation sub-branch. The feature extraction sub-branch performs multi-scale feature extraction on the noisy underwater acoustic signal and outputs high-dimensional mixed features. The feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output target signal features. The signal denoising branch generates a denoised underwater acoustic signal based on the target signal features. The discriminator distinguishes between the target underwater acoustic signal and the denoised underwater acoustic signal, and performs adversarial training with the generator based on the discrimination result.

[0073] S3, Noise reduction is performed on the noisy underwater acoustic signal based on the trained generator.

[0074] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for denoising underwater acoustic signals based on generative adversarial networks, characterized in that, Includes the following steps: S1, acquire the target underwater acoustic signal and the corresponding noisy underwater acoustic signal after transmission through the underwater acoustic channel, and input the noisy underwater acoustic signal and the target underwater acoustic signal into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator. S2, the generator includes a feature separation branch and a signal denoising branch. The feature separation branch includes a feature extraction sub-branch and a feature separation sub-branch. The feature extraction sub-branch performs multi-scale feature extraction on the noisy underwater acoustic signal and outputs high-dimensional mixed features. The feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output target signal features and environmental interference features. The feature separation sub-branch includes two parallel-connected feature separation modules. These two modules are pre-trained using target signal standard features and environmental interference standard features to obtain initial network parameters. Then, the network parameters are updated based on the feature separation results of the feature separation sub-branch using a separation loss function. The separation loss function specifically includes: a feature loss function based on the mean square error of the target signal features and environmental interference features with their corresponding standard features, a non-Gaussian loss function based on the negative entropy of the target signal features and environmental interference features, and a weighted combination of an independence loss function based on the mutual information between the target signal features and environmental interference features. Defined as: ,in, , and These represent adjustable weights; The feature loss function is defined as follows: ,in, and These represent the corresponding weight coefficients. Indicates the number of training samples in a single batch. and They represent the first The target signal characteristics and environmental interference characteristics of each sample and They represent the first Standard characteristics of the target signal and standard characteristics of environmental interference for each sample; The non-Gaussian loss function is defined as follows: ,in, Representing computational features Negative entropy, defined ,in Indicates the calculation of nonlinear functions The mathematical expectation, Representation and Features Homoscedastic Gaussian random variables and They represent the first Gaussian random variables representing the target signal characteristics and Gaussian random variables representing the environmental interference characteristics of each sample; The independence loss function is defined as follows: ,in, Representing computational features and characteristics The mutual information value; the signal denoising branch generates a denoised underwater acoustic signal based on the target signal features; the discriminator distinguishes between the target underwater acoustic signal and the denoised underwater acoustic signal, and performs adversarial training with the generator based on the discrimination result; the signal denoising branch includes a causal convolution module, an adaptive forget gate module, and a deconvolution processing module; the causal convolution module performs feature extraction and processing based on the target signal features; the adaptive forget gate module generates historical feature retention weights based on the real-time estimated channel change rate, performs time-series modeling on the features output by the causal convolution module at the current time and the features output by the previous time, and generates optimized time-series features; the deconvolution processing module generates the denoised underwater acoustic signal based on the optimized time-series features; S3, Noise reduction is performed on the noisy underwater acoustic signal based on the trained generator.

2. The underwater acoustic signal denoising method based on generative adversarial networks according to claim 1, characterized in that, The feature extraction sub-branch frames the noisy underwater acoustic signal according to symbols to obtain the corresponding time-series matrix. Then, it performs a short-time Fourier transform to obtain the corresponding time-frequency real part matrix and time-frequency imaginary part matrix. The time-series matrix, the time-frequency real part matrix, and the time-frequency imaginary part matrix are concatenated to obtain a dual-modal input. The dual-modal input is processed sequentially through three multi-scale convolutional layers and one max-pooling layer with a 2×2 pooling kernel to obtain the high-dimensional mixed features. The first multi-scale convolutional layer uses a 3×3 convolutional kernel to extract symbol detail features, the second multi-scale convolutional layer uses a 5×5 convolutional kernel to extract multipath trend features, and the third multi-scale convolutional layer uses a 7×7 convolutional kernel to extract channel global features. Each multi-scale convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function.

3. The underwater acoustic signal denoising method based on generative adversarial networks according to claim 1, characterized in that, The causal convolution module includes two causal convolutional layers. The first causal convolutional layer has a stride of 1, a left padding of 1, and a kernel size of 3×3. It is used to extract boundary features from the target signal. The second causal convolutional layer has a stride of 1, a left padding of 1, and a kernel size of 3×3. It is used to further extract local correlation features. Each causal convolutional layer is followed by a batch normalization layer and a LeakyReLU activation function, and a residual connection structure is introduced. The adaptive forget gate obtains the real-time estimated channel change rate, and calculates the historical feature retention weight based on the Sigmoid function and the channel change rate. The features output by the causal convolution module at the current time are used as the current features. The features output by the causal convolution module at the previous time step are used as historical features. ,based on The optimized temporal features are generated and output to the deconvolution processing module; The deconvolution processing module includes two deconvolution layers. The first deconvolution layer has a stride of 2, padding=1, and a kernel size of 3×3. The second deconvolution layer has a stride of 1, padding=1, and a kernel size of 3×3. Each deconvolution layer is followed by a batch normalization layer and a LeakyReLU activation function.

4. A hydroacoustic signal denoising system based on generative adversarial networks, characterized in that, The method according to any one of claims 1 to 3 comprises the following modules: The data acquisition module is configured to acquire the target underwater acoustic signal and the corresponding noisy underwater acoustic signal after transmission through the underwater acoustic channel, and input the noisy underwater acoustic signal and the target underwater acoustic signal into a generative adversarial network for training. The generative adversarial network includes a generator and a discriminator. The network training module is configured such that the generator includes a feature separation branch and a signal denoising branch. The feature separation branch includes a feature extraction sub-branch and a feature separation sub-branch. The feature extraction sub-branch performs multi-scale feature extraction on the noisy underwater acoustic signal and outputs high-dimensional mixed features. The feature separation sub-branch receives the high-dimensional mixed features and performs feature separation to output target signal features and environmental interference features. The feature separation sub-branch includes two parallel-connected feature separation modules. These two modules are pre-trained using target signal standard features and environmental interference standard features to obtain initial network parameters. Then, the network parameters are updated based on the feature separation results of the feature separation sub-branch using a separation loss function. Specifically, the separation loss function includes: a feature loss function based on the mean square error of the target signal features and environmental interference features relative to their corresponding standard features; a non-Gaussian loss function based on the negative entropy of the target signal features and environmental interference features; and a weighted combination of an independence loss function based on the mutual information between the target signal features and environmental interference features. Defined as: ,in, , and These represent adjustable weights; The feature loss function is defined as follows: ,in, and These represent the corresponding weight coefficients. Indicates the number of training samples in a single batch. and They represent the first The target signal characteristics and environmental interference characteristics of each sample and They represent the first Standard characteristics of the target signal and standard characteristics of environmental interference for each sample; The non-Gaussian loss function is defined as follows: ,in, Representing computational features Negative entropy, defined ,in Indicates the calculation of nonlinear functions The mathematical expectation, Representation and Features Homoscedastic Gaussian random variables and They represent the first Gaussian random variables representing the target signal characteristics and Gaussian random variables representing the environmental interference characteristics of each sample; The independence loss function is defined as follows: ,in, Representing computational features and characteristics The mutual information value; the signal denoising branch generates a denoised underwater acoustic signal based on the target signal features; the discriminator distinguishes between the target underwater acoustic signal and the denoised underwater acoustic signal, and performs adversarial training with the generator based on the discrimination result; the signal denoising branch includes a causal convolution module, an adaptive forget gate module, and a deconvolution processing module; the causal convolution module performs feature extraction and processing based on the target signal features; the adaptive forget gate module generates historical feature retention weights based on the real-time estimated channel change rate, performs time-series modeling on the features output by the causal convolution module at the current time and the features output by the previous time, and generates optimized time-series features; the deconvolution processing module generates the denoised underwater acoustic signal based on the optimized time-series features; The intelligent noise reduction module is configured to reduce the noise of the noisy underwater acoustic signal based on a trained generator.

5. A computer program product, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-3.

6. An electronic device, comprising: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 3.

Images