Underwater acoustic signal denoising method and device, computer device and readable storage medium

Abstract

Embodiments of the present application provide a method and device for underwater acoustic signal noise reduction, computer equipment and a readable storage medium. The method comprises: obtaining a to-be-de-noised time-domain signal of a to-be-de-noised underwater acoustic signal, and extracting features of the to-be-de-noised time-domain signal through a target model to obtain a to-be-de-noised feature tensor; calculating local position-dependent features and local time-dependent features between each local feature block and adjacent feature blocks in parallel through a local gated recurrent unit attention mechanism of the target model to obtain local noise features; calculating global noise features of the to-be-de-noised feature tensor through a global gated recurrent unit attention mechanism of the target model; reconstructing a noise time-domain signal based on a noise mask matrix generated by fusion of the local noise features and the global noise features through a decoder of the target model; and filtering noise from the to-be-de-noised time-domain signal based on the noise time-domain signal to obtain a target waveform signal. In this way, the accuracy of the reconstructed waveform signal can be improved.

Description

Technical Field

[0001] This application relates to the field of underwater acoustic signal processing technology, and in particular to an underwater acoustic signal noise reduction method, apparatus, computer equipment, and readable storage medium. Background Technology

[0002] Underwater acoustic signals are sound wave signals that propagate in the underwater environment. They can be used in various applications such as underwater communication, detection, navigation, and monitoring, including underwater robots, submarines, and sonar systems. They are of great significance for improving the intelligence level of underwater systems, promoting marine scientific research, and protecting the marine environment. However, due to the complexity of the underwater environment, underwater acoustic signals are often interfered with by various noises during propagation, such as water currents, marine life, and the noise of the equipment itself. These noises reduce the clarity and reliability of the signal, affecting its accurate reception and interpretation. Therefore, noise reduction processing of underwater acoustic signals is necessary to ensure the effectiveness and accuracy of underwater communication and detection systems.

[0003] In related technologies, a continuous time-domain waveform is typically segmented into multiple short-time intervals using a Short-Time Fourier Transform (STFT), and a Fourier Transform is performed on each segment to convert the time-domain signal into a time-frequency graph for signal reconstruction. However, in processing underwater acoustic signals using STFT, phase estimation and reconstruction algorithms are required. Factors such as multipath propagation, signal scattering, reverberation, and noise interference in the underwater acoustic environment can all lead to inaccurate phase information extracted from the noise background, resulting in incorrect phase estimation and consequently, inaccurate reconstructed waveform signals. Summary of the Invention

[0004] The main objective of this application is to provide a method, apparatus, computer device, and readable storage medium for underwater acoustic signal noise reduction, which can improve the accuracy of the reconstructed waveform signal.

[0005] To achieve the above objectives, a first aspect of this application proposes a method for underwater acoustic signal noise reduction, the method comprising:

[0006] The time-domain signal of the underwater acoustic sound to be denoised is obtained, and the feature of the time-domain signal to be denoised is extracted through the target model to obtain the feature tensor to be denoised, wherein the feature tensor to be denoised includes multiple local feature blocks.

[0007] By using the local door-controlled recurrent unit attention mechanism of the target model, the local positional dependency features and local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised are calculated in parallel to obtain the local noise features;

[0008] By using the global gated recurrent unit attention mechanism of the target model, the global positional dependency features and global temporal dependency features of the tensor of the feature to be denoised are calculated to obtain the global noise features;

[0009] The noise time-domain signal of the target noise is reconstructed from the time-domain signal to be denoised by the decoder of the target model based on the noise mask matrix generated by fusing the local noise features and the global noise features.

[0010] Based on the noise time-domain signal, noise filtering is performed on the time-domain signal to be denoised to obtain the target waveform signal.

[0011] Accordingly, a second aspect of this application provides an underwater acoustic signal noise reduction device, the device comprising:

[0012] The acquisition module is used to acquire the time-domain signal of the underwater acoustic signal to be denoised, and to extract features from the time-domain signal to be denoised through the target model to obtain the feature tensor to be denoised, wherein the feature tensor to be denoised includes multiple local feature blocks.

[0013] The first calculation module is used to calculate the local positional dependency features and local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised in parallel through the local departmental control recurrent unit attention mechanism of the target model, so as to obtain the local noise features.

[0014] The second calculation module is used to calculate the global positional dependency features and global temporal dependency features of the tensor of the feature to be denoised through the global gated recurrent unit attention mechanism of the target model, so as to obtain the global noise features.

[0015] The reconstruction module is used to reconstruct the noise time-domain signal of the target noise from the time-domain signal to be denoised, based on the noise mask matrix generated by fusing the local noise features and the global noise features, through the decoder of the target model.

[0016] The filtering module is used to filter noise from the time-domain signal to be denoised based on the noise time-domain signal to obtain the target waveform signal.

[0017] In some implementations, the first computing module is further configured to:

[0018] By using each local attention head of the local departmental control recurrent unit attention mechanism of the target model, the local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised are calculated in parallel to obtain the local temporal features corresponding to the multiple local feature blocks.

[0019] Multiple local temporal features corresponding to multiple local attention heads are concatenated to obtain a local temporal feature tensor;

[0020] The local positional dependency features of the local temporal feature tensor are calculated by the local gated layer of the local gated recurrent unit attention mechanism to obtain local noise features.

[0021] In some implementations, the first computing module is further configured to:

[0022] The local temporal feature tensor and the feature tensor to be denoised are fused to obtain the first feature;

[0023] The local position-dependent features of the first feature are calculated through the local gate control layer of the local gate control recurrent unit attention mechanism to obtain the second feature;

[0024] The second feature is nonlinearly transformed using the leakage linear rectification function of the local departmental control loop unit attention mechanism to obtain the third feature;

[0025] The local noise features are obtained by fusing the first feature and the third feature.

[0026] In some embodiments, the second computing module is further configured to:

[0027] By using each global attention head of the global gated recurrent unit attention mechanism of the target model, the global temporal dependency features of the tensor of the feature to be denoised are calculated, and the global temporal features corresponding to each global attention head are obtained.

[0028] Multiple global temporal features corresponding to multiple global attention heads are concatenated to obtain a global temporal feature tensor;

[0029] The global positional dependency features of the global temporal feature tensor are calculated through the global gating layer of the global gating recurrent unit attention mechanism to obtain the global noise features.

[0030] In some embodiments, the underwater acoustic signal noise reduction device further includes a training module for:

[0031] The sample time-domain signal of the underwater sound to be denoised is acquired, and the sample time-domain signal is feature extracted by a preset model to obtain a sample time-domain signal tensor, wherein the sample time-domain signal tensor includes multiple sample local feature blocks.

[0032] By using the local control loop unit attention mechanism of the preset model, the sample local position dependency features and sample local time dependency features between each sample local feature block and its adjacent feature blocks in the sample time domain signal tensor are calculated in parallel to obtain the sample local noise features.

[0033] By using the global gated recurrent unit attention mechanism of the preset model, the global positional dependence features and global temporal dependence features of the sample time-domain signal tensor are calculated to obtain the global noise features of the sample.

[0034] The decoder of the preset model reconstructs the predicted noise time-domain signal of the sample noise from the sample time-domain signal based on the sample noise mask matrix generated by fusing the sample local noise features and the sample global noise features.

[0035] Based on the predicted noise time-domain signal, noise filtering is performed on the sample time-domain signal to obtain the predicted waveform signal;

[0036] The target loss is determined based on the predicted noise time-domain signal and the predicted waveform signal;

[0037] The target model is obtained by training the preset model based on the target loss.

[0038] In some implementations, the training module is further configured to:

[0039] Obtain the time-domain signal of the sample noise and the sample waveform signal corresponding to the underwater acoustic noise to be denoised.

[0040] Calculate the corresponding first similarity based on the predicted noise time-domain signal and the sample noise time-domain signal;

[0041] Calculate the corresponding second similarity based on the predicted waveform signal and the sample waveform signal;

[0042] The target loss is determined based on the first similarity and the second similarity.

[0043] In some embodiments, the acquisition module is further configured to:

[0044] The convolutional neural network of the target model is used to extract features from the time-domain signal to be denoised, thereby obtaining the corresponding features to be denoised.

[0045] The noise reduction feature is divided into multiple local feature blocks according to the preset division length and division number;

[0046] The multiple local feature blocks are spliced ​​together to obtain the feature tensor to be denoised.

[0047] Accordingly, a third aspect of the present application provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the underwater acoustic signal noise reduction method of any one of the embodiments of the first aspect of the present application.

[0048] Accordingly, a fourth aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the underwater acoustic signal noise reduction method of any one of the embodiments of the first aspect of this application.

[0049] This application acquires the time-domain signal of the underwater acoustic signal to be denoised, and extracts features from the time-domain signal using a target model to obtain a feature tensor to be denoised, wherein the feature tensor includes multiple local feature blocks. Using the local gated recurrent unit attention mechanism of the target model, the local positional dependency and local temporal dependency features between each local feature block and its adjacent feature blocks in the feature tensor to be denoised are calculated in parallel to obtain local noise features. Using the global gated recurrent unit attention mechanism of the target model, the global positional dependency and global temporal dependency features of the feature tensor to be denoised are calculated to obtain global noise features. Using the decoder of the target model, the noise time-domain signal of the target noise is reconstructed from the time-domain signal to be denoised based on the noise mask matrix generated by fusing the local and global noise features. Based on the noise time-domain signal, noise filtering is performed on the time-domain signal to be denoised to obtain the target waveform signal. Therefore, by combining local gated recurrent unit attention (ROU) mechanisms and global gated recurrent unit attention (GU) mechanisms, the model can simultaneously capture local and global dependencies in sequence data. It can accurately capture transient noise and short-term signal details through the local GUU attention mechanism, while modeling long-term dependencies through the global GUU attention mechanism, thus accurately identifying and utilizing complex patterns in the signal. This allows for more effective preservation of the target signal's features during noise reduction. Furthermore, this application directly reconstructs a clear signal waveform in the time domain and learns and extracts noise patterns only, facilitating noise filtering rather than enhancing the target waveform signal. This avoids the phase and amplitude decoupling problem in traditional methods, effectively improving the accuracy of the reconstructed waveform signal. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the underwater acoustic signal noise reduction system provided in the embodiments of this application;

[0051] Figure 2 This is a flowchart of the underwater acoustic signal noise reduction method provided in the embodiments of this application;

[0052] Figure 3 This is a detailed structural diagram of a self-attention module provided in an embodiment of this application;

[0053] Figure 4 This is an overall structural diagram of the model provided in the embodiments of this application;

[0054] Figure 5These are illustrations of the noise reduction effect of this application provided in the embodiments of this application;

[0055] Figure 6 This is a schematic diagram of the functional modules of the underwater acoustic signal noise reduction device provided in the embodiments of this application;

[0056] Figure 7 This is a schematic diagram of the hardware structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0058] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0059] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0060] Underwater acoustic signals are sound wave signals that propagate in the underwater environment. They can be used in various applications such as underwater communication, detection, navigation, and monitoring, including underwater robots, submarines, and sonar systems. They are of great significance for improving the intelligence level of underwater systems, promoting marine scientific research, and protecting the marine environment. However, due to the complexity of the underwater environment, underwater acoustic signals are often interfered with by various noises during propagation, such as water currents, marine life, and the noise of the equipment itself. These noises reduce the clarity and reliability of the signal, affecting its accurate reception and interpretation. Therefore, noise reduction processing of underwater acoustic signals is necessary to ensure the effectiveness and accuracy of underwater communication and detection systems.

[0061] In related technologies, a continuous time-domain waveform is typically segmented into multiple short-time intervals using a Short-Time Fourier Transform (STFT), and a Fourier Transform is performed on each segment to convert the time-domain signal into a time-frequency graph for signal reconstruction. However, in processing underwater acoustic signals using STFT, phase estimation and reconstruction algorithms are required. Factors such as multipath propagation, signal scattering, reverberation, and noise interference in the underwater acoustic environment can all lead to inaccurate phase information extracted from the noise background, resulting in incorrect phase estimation and consequently, inaccurate reconstructed waveform signals.

[0062] Based on this, embodiments of this application provide a method, apparatus, computer device, and readable storage medium for underwater acoustic signal noise reduction, which can improve the accuracy of the reconstructed waveform signal.

[0063] The underwater acoustic signal noise reduction method, apparatus, computer equipment, and readable storage medium provided in the embodiments of this application are specifically described through the following embodiments. First, the underwater acoustic signal noise reduction system in the embodiments of this application is described.

[0064] Please refer to Figure 1 In some embodiments, this application provides an underwater acoustic signal noise reduction system, including a terminal 11 and a server 12.

[0065] For example, terminal 11 can be an underwater acoustic signal acquisition and preprocessing device, which may include a hydrophone, an embedded computer device, communication equipment, etc. Terminal 11 can acquire underwater acoustic signals in real time; terminal 11 can also perform preprocessing on the acquired signals locally, such as format conversion, simple filtering, etc., to reduce the amount of data that needs to be transmitted later, and send the preprocessed signal to server 12 for further processing via wired or wireless means.

[0066] Furthermore, the server 12 can be a high-performance server or cloud computing resource, which can train a preset model based on samples, and after obtaining the target model through training, it can calculate local and global noise features in parallel on the signal sent by the terminal 11, generate a noise time-domain signal using a decoder, obtain the target waveform signal through time-domain subtraction, and finally return the target waveform signal to the terminal 11. In this way, the efficiency and accuracy of the reconstructed waveform signal can be improved.

[0067] The underwater acoustic signal noise reduction method in this application can be illustrated by the following examples.

[0068] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent will be obtained first. Furthermore, the collection, use, and processing of this data will comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user will be obtained through pop-ups or redirects to confirmation pages. Only after obtaining the user's separate permission or consent will the necessary user-related data for the normal operation of the embodiments of this application be obtained.

[0069] In this embodiment, the description will focus on the underwater acoustic signal noise reduction device, which can be integrated into a computer device. See [link to relevant documentation]. Figure 2 , Figure 2 This is a flowchart illustrating the steps of the underwater acoustic signal noise reduction method provided in this application embodiment. Taking the underwater acoustic signal noise reduction device specifically integrated into a terminal or server as an example, the specific process when the processor on the terminal or server executes the program instructions corresponding to the underwater acoustic signal noise reduction method is as follows:

[0070] Step 101: Obtain the time-domain signal of the underwater acoustic signal to be denoised, and extract features from the time-domain signal to be denoised using the target model to obtain the feature tensor to be denoised, wherein the feature tensor to be denoised includes multiple local feature blocks.

[0071] In some implementations, in order to remove noise and recover the original signal more effectively, the time-domain signal to be denoised can be extracted from the acquired underwater acoustic signal to be denoised, and the target model can be used to perform deep feature extraction to obtain a feature tensor to be denoised that can represent the essential features of the signal, so as to facilitate more detailed local feature analysis and global feature integration in the future.

[0072] Among them, the underwater sound to be noise-reduced can be the original underwater sound collected from the underwater environment (such as the marine environment) and includes noise (such as wind, rain, water flow, marine life, etc.) and the sound emitted by the detection target (such as ships, submarines, etc.).

[0073] Among them, the time-domain signal to be denoised can be the time-series data corresponding to the underwater acoustic signal to be denoised, which is a form that can be directly used for digital signal processing.

[0074] The target model can be an end-to-end underwater noise sensing model, used to extract the features of the target noise from the time-domain signal to be denoised and generate the target noise, so that the target noise can be filtered from the time-domain signal to be denoised in the future.

[0075] The feature tensor to be denoised can be a high-dimensional data structure obtained by performing a convolution operation on the input time-domain signal to be denoised by the target model, dividing it according to a preset partitioning method, and then splicing it together. It contains the key feature information of the target noise.

[0076] Among them, the local feature block can be a number of sub-parts into which the feature tensor to be denoised is divided. Each sub-part (each local feature block) represents the local characteristics of the time-domain signal to be denoised in a specific time period, which facilitates the subsequent learning of local and global dependencies.

[0077] For example, a hydrophone or similar device can be used to collect the time-domain signal of the underwater sound to be denoised. The underwater sound to be denoised includes the sound from the ship (or other target to be detected) and background noise (such as the sound of wind, rain, water flow and marine life). The sampling rate and sampling time of the time-domain signal to be denoised can be set according to the actual situation. For example, the sampling rate can be set to 16kHz and the sampling time can be set to 3 seconds.

[0078] For example, a pre-trained target model can be used to process the time-domain signal x∈R to be denoised. T Assuming the input time-domain signal to be denoised has a length of T = 48000 sample points, after a one-dimensional convolution operation, the output feature dimension is F = 256. The time step remains constant, and the convolution kernel size k is set to 16. This allows us to capture the local structural characteristics of the signal. The calculation formula is as follows: W = H(conv1d(x));

[0079] Where W∈R F×T The features to be denoised are obtained after convolution operations, and H(·) represents the ReLU activation function.

[0080] Furthermore, after the above convolution operation, a feature to be denoised with shape [F, T] is obtained. To facilitate subsequent local and global dependent feature processing, the feature to be denoised can be divided into multiple local feature blocks according to a preset partitioning method, and then these multiple local feature blocks can be concatenated to obtain the feature tensor to be denoised. The preset partitioning method can be to partition the feature to be denoised according to a preset partitioning length and partitioning hop count.

[0081] By acquiring the time-domain signal to be denoised and dividing it to obtain the feature tensor to be denoised, it is easier to extract local and global dependent features in the subsequent process.

[0082] In some implementations, to enhance the ability to identify local patterns and dynamic changes in complex underwater acoustic environments, the features to be denoised can be divided into multiple local feature blocks according to a specific partitioning method. This facilitates the formation of a feature tensor that integrates local details and global contextual information, providing rich contextual information for subsequent noise removal and signal recovery. For example, step 101, "extracting features from the time-domain signal to be denoised using the target model to obtain the feature tensor to be denoised," may include:

[0083] (101.1) The convolutional neural network of the target model is used to extract features of the time-domain signal to be denoised, and the corresponding features to be denoised are obtained.

[0084] (101.2) Divide the features to be denoised into multiple local feature blocks according to the preset division length and division number;

[0085] (101.3) Concatenate multiple local feature blocks to obtain the feature tensor to be denoised.

[0086] Among them, the convolutional neural network can be a one-dimensional convolutional neural network (Conv1D) in the target model used to extract features from the time-domain signal to be denoised. It can learn the high-dimensional representation of the input signal and capture the spatial structure characteristics of the input signal through a series of convolutional kernels.

[0087] Among them, the features to be denoised can be the feature representation obtained after processing the time-domain signal to be denoised through a convolutional neural network, which effectively captures the important information in the time-domain signal to be denoised.

[0088] The preset division length can be a fixed length K set for each block when dividing the feature to be denoised into multiple local feature blocks, which determines the time step or the number of data points contained in each local feature block.

[0089] The number of hops can be the time step or data point interval P between adjacent blocks when dividing the feature to be denoised into multiple local feature blocks. It affects the degree of overlap between local feature blocks and the target model's ability to capture sequence information.

[0090] For example, a pre-trained target model can be used to process the time-domain signal x∈R to be denoised. T Assuming the length of the input time-domain signal to be denoised is T = 48000 sample points, using a one-dimensional convolutional neural network (e.g., kernel size 16, stride 1, output channels 256) to perform a one-dimensional convolution operation on the signal, the output feature dimension is F = 256. The time step remains unchanged, and the kernel size k is set to 16. This allows us to capture the local structural characteristics of the signal. The calculation formula is as follows: W = H(conv1d(x));

[0091] Where W∈R F×T The features to be denoised are obtained after convolution operations, and H(·) represents the ReLU activation function.

[0092] Furthermore, after the above convolution operation, a feature to be denoised with a shape of [F,T] can be obtained, for example, the shape can be [256, 47984] (due to the reduction in time step caused by the convolution operation). If the length of the local feature block (i.e., the preset partition length) K = 256 and the partition number P = 128, the feature to be denoised is divided into multiple local feature blocks. Then, the coverage range of the first local feature block is [0, 256), the coverage range of the second local feature block is [128, 384), and so on, multiple local feature blocks can be obtained.

[0093] In some implementations, overlap can be set between adjacent local feature blocks to ensure data continuity and avoid information fragmentation. For example, 50% overlap can be set, that is, a preset division length of 256 and a division jump number of 128 can be set.

[0094] Furthermore, all local feature blocks can be concatenated along the time dimension to form a feature tensor to be denoised. For example, if the shape of each local feature block is [256, 256], then the shape of the concatenated feature tensor to be denoised is [256, 374×256], i.e., [256, 95744]. It is understood that the preset segmentation length and the number of segmentation jumps can be set according to actual conditions, and this application does not impose specific limitations.

[0095] In some implementations, convolutional kernels of different sizes can be used to capture features in different frequency ranges. For example, convolutional kernels of sizes 16, 32, and 64 can be used simultaneously to capture short-term dynamic changes, medium-term changes, and long-term changes of the time-domain signal to be denoised, respectively. The output results can be integrated by concatenation or weighted summation to form richer features to be denoised.

[0096] By employing the above methods, the target model can be able to analyze different parts of the signal in detail and enhance the understanding of local patterns, thereby laying a solid foundation for subsequent high-quality underwater acoustic signal processing.

[0097] Step 102: Through the local departmental control recurrent unit attention mechanism of the target model, the local positional dependency features and local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised are calculated in parallel to obtain the local noise features.

[0098] In some implementations, in order to extract information that can describe the characteristics of local noise, the local positional dependence and local temporal dependence features between each local feature block and its neighboring feature blocks in the feature tensor to be denoised can be computed in parallel through the local departmental control recurrent unit attention mechanism in the target model, thereby identifying the local patterns and temporal dynamic changes within the signal, so as to remove noise more accurately and restore the original signal.

[0099] Among them, the local gated recurrent unit attention mechanism can be an attention mechanism designed in this application for processing sequential data. It combines the advantages of gated recurrent units and attention mechanisms, and focuses on capturing local dependencies in sequential data.

[0100] Among them, the adjacent feature blocks can be local feature blocks that are physically or temporally adjacent to the local feature blocks that need to be weighted in the feature tensor to be denoised.

[0101] Among them, local position-dependent features can be the features (such as transient impulse noise or high-frequency interference) extracted when analyzing each local feature block and its adjacent feature blocks under the action of the local door-controlled recurrent unit attention mechanism, which reflect the relative positional relationship between the two and are used to characterize the spatial correlation between different local feature blocks.

[0102] Among them, local time-dependent features can be the features extracted for each local feature block and its adjacent feature blocks under the action of the local departmental recurrent unit attention mechanism, which reflect the time series dependency between the two (such as the persistence or periodic fluctuation of low-frequency noise), and are used to characterize the dynamic change law of different local feature blocks in the time dimension.

[0103] Among them, the local noise features can be a comprehensive representation extracted from the feature tensor to be denoised after being processed by the local departmental recurrent unit attention mechanism. This representation can describe the local positional dependence features and local temporal dependence features within each local feature block and between it and its adjacent feature blocks.

[0104] In some implementations, each local attention head of the local controlled recurrent unit attention mechanism can compute the local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of features to be denoised in parallel. In this way, each local attention head can obtain a local temporal feature based on the tensor of features to be denoised.

[0105] Furthermore, the local temporal features corresponding to all local attention heads can be concatenated along the feature dimension to obtain a local temporal feature tensor. If the output feature dimensions of each head are the same, they can be directly concatenated; if the output feature dimensions of different heads are different, the dimensions need to be aligned first (e.g., through linear transformation) before concatenation. This allows for the integration of information from multiple perspectives, resulting in a richer and more comprehensive final feature representation.

[0106] In some implementations, since the position encoding part in the Transformer encoder is not applicable to acoustic sequences, in order to accurately obtain the position dependency information of the feature tensor to be denoised, this application deletes the position encoding part in the Transformer encoder and replaces the first fully connected layer of the feedforward network with a local gate control layer, that is, a gated recurrent unit (GRU). The local position dependency features of the local temporal feature tensor are calculated through the local gate control layer to obtain the local noise features.

[0107] By employing the above methods, not only are the local patterns and temporal dynamic changes within the signal to be denoised effectively captured, but the ability to identify noise characteristics in complex underwater acoustic environments is also enhanced. By integrating the interrelationships between local feature blocks, this method significantly improves the accuracy of noise removal and the effect of target signal recovery, laying a solid foundation for subsequent high-quality underwater acoustic signal processing.

[0108] In some implementations, to accurately capture local patterns and temporal dynamic changes within the signal, the tensor of the feature to be denoised can be processed using the local attention head and local gate control layer in the local gate control recurrent unit attention mechanism. This ultimately yields local noise features that focus on local noise characteristics, thereby improving the accuracy of subsequent noise removal and signal recovery. For example, step 102 may include:

[0109] (102.1) Through each local attention head of the local departmental control recurrent unit attention mechanism of the target model, the local time dependency features between each local feature block and adjacent feature blocks in the tensor of the feature to be denoised are calculated in parallel to obtain the local temporal features corresponding to multiple local feature blocks.

[0110] (102.2) Concatenate the local temporal features corresponding to multiple local attention heads to obtain the local temporal feature tensor;

[0111] (102.3) The local position dependence features of the local temporal feature tensor are calculated by the local gate control layer through the local gate control recurrent unit attention mechanism to obtain the local noise features.

[0112] In this context, the local attention head can be the basic unit responsible for parallel processing of the relationships between local feature blocks in the local gated recurrent unit attention mechanism. Each local attention head focuses on learning local temporally dependent features within a specific subspace, that is, the characteristics of how its local feature blocks and adjacent feature blocks influence each other over time.

[0113] Among them, the local temporal features can be obtained by processing the local attention head, which reflects the comprehensive feature representation of the local temporal dependence between each local feature block and its adjacent feature blocks. Each local attention head can obtain a local temporal feature after processing the feature tensor to be denoised.

[0114] The local temporal feature tensor can be a high-dimensional data structure formed by concatenating the local temporal features output by multiple local attention heads. It integrates the learning results of all local attention heads on the temporal dependencies between local feature blocks, providing richer contextual information for subsequent analysis.

[0115] Among them, the local control layer can be a component located in the local control recurrent unit attention mechanism, used to further process the local temporal feature tensor, extract local position-dependent features, capture and integrate the relative positional relationships between local feature blocks, enhance the model's ability to understand local details, and provide support for accurately identifying local noise.

[0116] Please refer to Figure 3 , Figure 3 The structural diagram of the self-attention module designed for this application is shown. The self-attention module can be represented by MV-MHSA. Each self-attention module contains a local gated recurrent unit attention mechanism and a global gated recurrent unit attention mechanism. This will be combined with... Figure 3 Here are examples of the processes from (102.1) to (102.3).

[0117] In some implementations, if the number of local attention heads is 4 (this is just an example; there may actually be more or fewer local attention heads), different local attention heads may capture the temporal dependencies of the feature tensors to be denoised in different representation subspaces. For example, local attention head 1 may focus on high-frequency noise components, local attention head 2 may focus on low-frequency target signals, local attention head 3 may focus on dynamic changes in a short period of time, and local attention head 4 may focus on long-term trends of signals. Alternatively, local attention head 1 may focus on N adjacent feature blocks to the left and right of the local feature block whose weight needs to be calculated, local attention head 2 may focus on M adjacent feature blocks to the left of the local feature block whose weight needs to be calculated, local attention head 3 may focus on X adjacent feature blocks to the right of the local feature block whose weight needs to be calculated, local attention head 4 may focus on G adjacent feature blocks to the left and right of the local feature block whose weight needs to be calculated, and so on. The specific spatial scale of focus can be adjusted according to the actual situation, and this application embodiment does not limit this.

[0118] For example, suppose the feature tensor X to be denoised is a tensor of shape (N, L, D), where N is the number of samples, L is the sequence length (i.e., the number of local feature blocks), and D is the feature dimension. If there are query, key, and value matrices corresponding to the query matrix, key matrix, and value matrix, respectively, then for each local attention head i, the calculation process is as follows:

[0119]

[0120] Here, head i This represents the local temporal features, specifically the output of the i-th local attention head, which captures the local temporal dependencies between each local feature block and its neighboring feature blocks within the tensor of the features to be denoised; W i Q Represents the query matrix, W i K W represents the bond matrix. i V Represents a value matrix; d k The dimension of the key vector is represented; softmax is the function that transforms the input feature tensor to be denoised into a probability distribution.

[0121] Furthermore, the local temporal features corresponding to multiple local attention heads can be concatenated to obtain a local temporal feature tensor. Specifically, assuming there are h local attention heads in total, and the local temporal features output by each local attention head are: head1, head2, ..., head hBy concatenating these local temporal features together, a local temporal feature tensor MultiHead(Q,K,V) can be formed:

[0122] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W o ;

[0123] Among them, W o This is another learnable weight matrix used to transform the concatenated data back to the original dimensions, ultimately yielding a local temporal feature tensor.

[0124] Furthermore, local positional dependencies can be calculated on the local temporal feature tensor using a local gated layer with a local gated recurrent unit attention mechanism. Specifically, for the features of the local temporal feature tensor at each time step, redundant noise features for the current time step can be filtered out by the reset gate of the local gated layer (e.g., suppressing historical noise residues unrelated to the current time step). Then, the input of the current time step and the historical features of the corresponding historical time steps can be dynamically fused by the update gate of the local gated layer (e.g., strengthening the transient correlation of burst noise). This yields local noise features, which can then accurately extract local positional dependencies, thus obtaining the final local noise features.

[0125] By combining the key information of the original signal with the deep-seated dynamic characteristics of the time series, the model's ability to capture complex patterns and its noise reduction effect can be significantly improved, making the model more efficient and accurate in processing complex underwater acoustic signals.

[0126] In some implementations, to more accurately identify and separate local noise components, the local position-dependent features of the local temporal features can be further calculated using the local gate control layer of the local gate control recurrent unit attention mechanism to obtain more accurate local noise features. For example, (102.3) may include:

[0127] (102.3.1) The local temporal feature tensor and the feature tensor to be denoised are fused to obtain the first feature;

[0128] (102.3.2) The local position-dependent features of the first feature are calculated by the local gate control layer of the local gate control recurrent unit attention mechanism to obtain the second feature;

[0129] (102.3.3) The second feature is transformed nonlinearly by the leakage linear rectification function of the attention mechanism of the local departmental control loop unit to obtain the third feature;

[0130] (102.3.4) The local noise features are obtained by fusing the first and third features.

[0131] The first feature can be a feature representation obtained by fusing the feature tensor to be denoised and the local temporal feature tensor through residual connection. It combines the key features of the original signal and the temporal sequence relationship between local feature blocks, providing rich contextual information for subsequent analysis.

[0132] The second feature can be the result of processing the first feature through the local gate control layer of the local gate control recurrent unit attention mechanism, and is used to characterize the relative positional relationship between different features in the first feature and their impact on local noise.

[0133] The Leaky Rectified Linear Function (LEX) can be an activation function that addresses the issue of ReLU units being completely inactive with negative inputs. Specifically, when the input is positive, Leaky ReLU behaves the same as ReLU, with the output equal to the input; however, for negative inputs, Leaky ReLU does not completely suppress them to zero but multiplies them by a small positive slope to increase the model's learning and expressive capabilities.

[0134] The third feature can be the output result of the second feature after a nonlinear transformation by a leakage linear rectifier function.

[0135] Please refer to Figure 3 In some implementations, since the position encoding part of the original Transformer encoder is not suitable for acoustic sequences, this application removes the position encoding part when designing the model and replaces the first fully connected layer of the feedforward network with a local control layer to learn position information instead of the position encoding module.

[0136] Furthermore, local dependencies between spatial locations can be modeled using a local gated layer within the local gated recurrent unit attention mechanism. Specifically, the reset gate in the local gated layer can suppress local noise patterns irrelevant to the current frame, avoiding redundant interactions in global attention computation. The update gate dynamically adjusts the weights of local features (e.g., sudden changes in noise intensity), retaining only key features dependent on local locations and filtering out irrelevant information. Moreover, the weights of the reset and update gates can be dynamically adjusted to adapt to rapid changes in non-stationary noise. For example, when noise power abruptly changes, the update gate can quickly enhance the current input weights, avoiding interference from historical noise features.

[0137] In some implementations, the leaky linear rectifier function can be designed as follows:

[0138]

[0139] Here, α is a small positive number that can be set according to the actual situation, such as 0.01, which determines the magnitude of the negative gradient. Leaky ReLU assigns a non-zero gradient to negative inputs, allowing neurons that were originally suppressed in ReLU to continue participating in the training of the preset model, thereby improving the efficiency of training the preset model and the expressive power and generalization performance of the trained target model.

[0140] In some implementations, the denoising feature tensor X obtained after processing the temporal signal to be denoised by the target model can be acquired. The specific method for acquiring the denoising feature tensor has been described above and will not be repeated here. Then, the denoising feature tensor X is processed by the four local attention heads of the local controlled recurrent unit attention mechanism to obtain the local temporal features output by each local attention head. The multiple local temporal features corresponding to the four local attention heads are then concatenated to obtain the local temporal feature tensor, i.e., MultiHead. For example, if the output dimension of a single head is T×64, the concatenated dimension is T×256.

[0141] Please refer to Figure 3 Furthermore, a normalization layer (LayerNorm) can be used to normalize the tensor X of the noise feature to be denoised, i.e., X1 = LayerNorm(X), to suppress abnormal amplitude fluctuations of noise. Then, the normalized tensor X1 of the noise feature to be denoised is fused (Add) with local temporal features to obtain the first feature, i.e., X2 = LayerNorm(X1 + MultiHead), to preserve the details of the original signal while introducing contextual information extracted by the attention mechanism, thereby alleviating signal distortion caused by phase decoupling. Afterward, the local position-dependent features of the first feature X2 are calculated by the local gate control layer of the local gate control recurrent unit attention mechanism to obtain the second feature, i.e., GRU(X2), to dynamically correct the spatiotemporal correlation of noise features. Furthermore, the second feature can be nonlinearly transformed using the leakage linear rectification function (LeakyReLU) of the local control loop unit to obtain the third feature FFN, i.e., FFN = LeakyReLU(GRU(X2)). Finally, the first and third features are residually fused using the residual connection results of the target model to obtain the local noise feature, i.e., Output = LayerNorm(X2 + FFN). This effectively prevents information loss.

[0142] By employing the above methods, not only are important feature details preserved, but the learning ability and stability of the model are also enhanced by introducing nonlinear transformation and residual connection techniques. In this way, the key information of the original signal and the deep-seated dynamic characteristics of the sequence can be effectively combined to enhance the model's ability to capture complex patterns, thereby improving the noise reduction effect and the accuracy of signal restoration. This approach can be widely applied to high-precision noise suppression scenarios such as underwater communication and sonar detection.

[0143] Step 103: Using the global gated recurrent unit attention mechanism of the target model, calculate the global positional dependency features and global temporal dependency features of the tensor of the feature to be denoised, and obtain the global noise features.

[0144] In some implementations, in order to identify the overall pattern and long-term dynamic changes of the signal, the global gated recurrent unit attention mechanism in the target model can be used to extract information that can describe the global noise characteristics of the entire signal, thereby improving the accuracy and completeness of noise identification and thus better recovering the original signal.

[0145] Among them, the global gated recurrent unit attention mechanism can be used to analyze and learn the overall structure of the feature tensor to be denoised and its internal time dynamic changes.

[0146] Among them, global location-dependent features can be used to characterize the spatial correlation between various local feature blocks within the entire time domain signal to be denoised, which helps to understand the overall structure of the time domain signal to be denoised. For example, in the time domain signal to be denoised, noise in a certain frequency band may affect the energy distribution of other frequency bands.

[0147] Among them, global time-dependent features can be used to characterize the dynamic changes of the entire time-domain signal to be denoised over a long time scale, helping to identify and separate global noise components, such as the periodic changes of noise or the temporal continuity of the time-domain signal to be denoised.

[0148] Global noise features can be a comprehensive characterization of noise components by the target model after integrating global location-dependent and time-dependent information. Global noise features can include global characteristics such as the spectral distribution and temporal energy variation of noise, and are used to identify noise from mixed signals.

[0149] In some implementations, multiple global attention heads within a globally gated recurrent unit attention mechanism can be used to compute, in parallel, the global temporal dependency features between each global feature block in the tensor of features to be denoised and other feature blocks in the entire feature sequence. Each global attention head can obtain a global temporal feature based on the tensor of features to be denoised. It is understood that the number of global attention heads can be set according to actual needs; for example, 8, 9, 12, etc., can be set.

[0150] Furthermore, along the feature dimension, the global temporal features corresponding to all global attention heads can be concatenated to obtain a global temporal feature tensor. If the output feature dimensions of each global attention head are the same, they can be directly concatenated; if the output feature dimensions of different global attention heads are different, the dimensions need to be aligned first (e.g., through linear transformation) before concatenation. In this way, information from multiple perspectives can be integrated, resulting in a richer and more comprehensive final feature representation.

[0151] In some implementations, similar to the local gated recurrent unit attention mechanism, since the position encoding part in the Transformer encoder is not applicable to acoustic sequences, in order to accurately obtain the global position dependency information of the feature tensor to be denoised, this application deletes the position encoding part in the Transformer encoder and replaces the first fully connected layer of the feedforward network with a global gated layer, i.e., a gated recurrent unit (GRU). The global position dependency features of the global temporal feature tensor are calculated through the global gated layer to obtain the global noise features.

[0152] The above methods not only effectively capture the global patterns and temporal dynamics within the signal to be denoised, but also enhance the ability to identify noise characteristics in complex underwater acoustic environments. This helps to significantly improve the accuracy of noise removal and the effect of target signal recovery, laying a solid foundation for subsequent high-quality underwater acoustic signal processing.

[0153] In some implementations, the aim is to compute the global temporal dependency features of the tensor of features to be denoised in parallel using the global attention head in the global gating recurrent unit attention mechanism of the target model, and then concatenate these global temporal features into a global temporal feature tensor. Next, the global gating layer is used to further analyze the global positional dependency features, ultimately extracting global noise features that can describe the noise characteristics of the entire signal. This process helps to capture the overall pattern and long-range temporal dynamic changes of the signal, thereby improving the accuracy of noise removal and signal recovery. Step 103 may include:

[0154] (103.1) Through each global attention head of the global gated recurrent unit attention mechanism of the target model, calculate the global temporal dependency features of the tensor of the feature to be denoised, and obtain the global temporal features corresponding to each global attention head;

[0155] (103.2) Concatenate multiple global temporal features corresponding to multiple global attention heads to obtain a global temporal feature tensor;

[0156] (103.3) The global position dependency features of the global temporal feature tensor are calculated by the global gating layer through the global gating recurrent unit attention mechanism to obtain the global noise features.

[0157] Among them, the global attention head can be the basic unit in the global gated recurrent unit attention mechanism responsible for parallel processing of the overall structure of the tensor of the feature to be denoised and its internal time dynamic relationship.

[0158] Among them, the global temporal features can be obtained by processing the global attention head. They are used to reflect the long-term time series correlation of different local feature blocks in the tensor of the feature to be denoised within the entire range of the signal to be denoised, which helps to understand the overall temporal dynamic change law of the signal to be denoised.

[0159] Among them, the global temporal feature tensor can be a high-dimensional data structure formed by concatenating the global temporal features output by multiple global attention heads.

[0160] The global gating layer can be a component located within the global gating recurrent unit attention mechanism, used to further process the global temporal feature tensor and extract global position-dependent features. The global gating layer can effectively capture and integrate the relative positional relationships between different local feature blocks through gating mechanisms, enhancing the target model's understanding of the global structure.

[0161] Please refer to Figure 3 , Figure 3 The structural diagram of the self-attention module designed for this application is shown. The self-attention module can be represented by MV-MHSA. Each self-attention module contains a local gated recurrent unit attention mechanism and a global gated recurrent unit attention mechanism. This will be combined with... Figure 3 Here are examples of the processes from (103.1) to (103.3).

[0162] In some implementations, if the number of global attention heads is 8 (this is just an example; in reality, there may be more or fewer global attention heads), assume that the feature tensor X to be denoised is a tensor of shape (N, L, D), where N is the number of samples, L is the sequence length (i.e., the number of global feature blocks), and D is the feature dimension. If there are queries, keys, and values ​​corresponding to query matrices W... i Q Key matrix W i K Sum matrix W i V , where i represents the i-th global attention head.

[0163] Furthermore, the dot product of the query matrix and the key matrix can be calculated, scaled, and then the Softmax function can be applied to obtain the attention weights. Multiplying the attention weights by the value matrix yields the global temporal features output by each global attention head for the tensor of the feature to be denoised. The specific calculation process is as follows:

[0164]

[0165] Here, head i This represents the global temporal features, i.e., the output of the i-th global attention head, which captures the global temporal dependencies between each global feature block and its neighboring feature blocks in the feature tensor to be denoised; W i Q Represents the query matrix, W i K W represents the bond matrix. i V Represents a value matrix; d k The dimension of the key vector is represented; softmax is the function that transforms the input feature tensor to be denoised into a probability distribution.

[0166] Furthermore, the global temporal features corresponding to multiple global attention heads can be concatenated to obtain a global temporal feature tensor. Specifically, assuming there are h global attention heads in total, and the global temporal features output by each global attention head are: head1, head2, ..., head h By concatenating these global temporal features together, a global temporal feature tensor MultiHead(Q,K,V) can be formed:

[0167] MultiHead(Q,K,V)=Concat(head1,head2,...,head h W o ;

[0168] Among them, W o This is another learnable weight matrix used to transform the concatenated data back to the original dimensions, ultimately yielding the global temporal feature tensor.

[0169] Furthermore, global position-dependent features of the global temporal feature tensor can be calculated through a global gating layer using a global gating recurrent unit attention mechanism to obtain global noise features. Specifically, for the features of the global temporal feature tensor at each time step, redundant noise features for the current time step can be filtered out by the reset gate of the global gating layer (e.g., suppressing historical noise residues unrelated to the current time step). Then, the input of the current time step and the historical features of the corresponding historical time steps can be dynamically fused by the update gate of the global gating layer (e.g., strengthening the transient correlation of burst noise) to obtain global noise features. In this way, global position-dependent features can be accurately extracted, thus obtaining the final global noise features.

[0170] In some implementations, to more accurately identify and separate global noise components, the global position dependence features of the global temporal features can be further calculated using the global gating layer of the global gated recurrent unit attention mechanism to obtain more accurate global noise features. For example, (103.3) may include: fusing the global temporal feature tensor and the feature tensor to be denoised to obtain a fourth feature; calculating the global position dependence features of the fourth feature using the global gating layer of the global gated recurrent unit attention mechanism to obtain a fifth feature; performing a nonlinear transformation on the fifth feature using the leakage linear rectification function of the global gated recurrent unit attention mechanism to obtain a sixth feature; and fusing the fourth and sixth features to obtain the global noise features.

[0171] Please refer to Figure 3 In some implementations, global dependencies between spatial locations can be modeled using a global gating layer of a global gating recurrent unit attention mechanism. Specifically, a reset gate in the global gating layer can suppress global noise patterns irrelevant to the current frame, avoiding redundant interactions in global attention calculations. An update gate dynamically adjusts the weights of global features (e.g., changes in the intensity of sudden noise), retaining only key features dependent on global location and filtering out irrelevant information. Furthermore, the weights of the reset and update gates can be dynamically adjusted to adapt to rapid changes in non-stationary noise. For example, when noise power abruptly changes, the update gate can quickly enhance the current input weights, avoiding interference from historical noise features.

[0172] As is understandable, the relevant content regarding the leakage of the linear rectified function has already been discussed above and will not be repeated here. Furthermore, the eight global attention heads of the global gated recurrent unit attention mechanism can be used to process the denoising feature tensor X, obtaining the global temporal features output by each global attention head. These global temporal features from the eight global attention heads are then concatenated to obtain the global temporal feature tensor, i.e., MultiHead. For example, if the output dimension of a single head is T×64, the concatenated dimension is T×512.

[0173] Please refer to Figure 3 Furthermore, a normalization layer (LayerNorm) can be used to normalize the tensor X of the noise feature to be denoised, i.e., X1 = LayerNorm(X), to suppress abnormal amplitude fluctuations of noise. Then, the normalized tensor X1 of the noise feature to be denoised is fused (Add) with the global temporal feature to obtain the fourth feature, i.e., X2 = LayerNorm(X1 + MultiHead), to preserve the details of the original signal while introducing the context information extracted by the attention mechanism, thereby alleviating the signal distortion caused by phase decoupling. Afterward, the global position-dependent feature of the fourth feature X2 is calculated by the global gated layer of the global gated recurrent unit attention mechanism to obtain the fifth feature, i.e., GRU(X2), to dynamically correct the spatiotemporal correlation of the noise feature. Furthermore, the fifth feature can be nonlinearly transformed using the Leaky ReLU function of the global gated loop unit to obtain the sixth feature FFN, i.e., FFN = Leaky ReLU(GRU(X2)). Finally, the fourth and sixth features are residually fused using the residual connection results of the target model to obtain the global noise feature, i.e., Output = LayerNorm(X2 + FFN). This effectively prevents information loss.

[0174] By using multiple global attention heads in parallel to capture long-term dependencies and global patterns throughout the sequence, the model can more accurately identify and separate noise. This makes the model more efficient and accurate in handling low signal-to-noise ratio scenarios, significantly improving the accuracy of noise reduction and recognition.

[0175] Step 104: Using the decoder of the target model, the noise time-domain signal of the target noise is reconstructed from the time-domain signal to be denoised based on the noise mask matrix generated by fusing local noise features and global noise features.

[0176] In some implementations, in order to accurately estimate the noise components and separate them from the original signal, a decoder in the target model can be used to reconstruct the noise time-domain signal of the target noise from the time-domain signal to be denoised, so as to provide a basis for subsequent noise removal and restoration of the clean target waveform signal.

[0177] The decoder can be the part of the target model used to convert high-dimensional feature representations (such as noise mask matrices) back to time-domain signals. The decoder can consist of a series of deconvolutional or transposed convolutional layers.

[0178] The noise mask matrix can be formed by adding local noise features and global noise features, normalizing the features, and then compressing them to [0,1].

[0179] The target noise can be a signal representation that contains only noise components, obtained after being processed by the decoder.

[0180] The noise time-domain signal can be a specific noise signal reconstructed by the decoder based on the noise mask matrix, represented in time series form. The noise time-domain signal is used to accurately characterize the noise components in the time-domain signal to be denoised, and can be used in the subsequent noise filtering process to remove noise from the original signal and restore a clear target waveform signal.

[0181] Please refer to Figure 4 In some implementations, the target model includes multiple self-attention modules, each containing a local gated recurrent unit attention mechanism and a global gated recurrent unit attention mechanism. The results of multiple self-attention modules are fused module by module to obtain local noise features and global noise features. The local noise features and global noise features are concatenated along the channel dimension to obtain the mask matrix M, which can be [B,T,2C].

[0182] In some implementations, the noise mask matrix W can be obtained by multiplying the feature W to be denoised by the noise mask matrix M. The noise mask matrix can be used to characterize the noise-dominated feature region. The specific formula is as follows: Wm = M * W.

[0183] As is understandable, the method for obtaining the features to be denoised has been introduced above. That is, the time-domain signal to be denoised (e.g., dimension [B,T]) is input into the encoder of the target model, and the features to be denoised are obtained by processing it through convolution kernels and ReLU activation functions. This part has been elaborated above and will not be repeated here.

[0184] Furthermore, based on the noise mask matrix, Wm can be mapped back to the time domain through transpose convolution (TransposeConv1D) to generate the time-domain noise signal of the target noise. The specific formula is as follows: S t =TransposeConv1D(W m ).

[0185] By reconstructing the noise time-domain signal from the noise-to-denoise time-domain signal using the learned noise mask matrix, the phase distortion problem of traditional time-frequency domain methods can be bypassed. The noise features can be directly learned and processed, which significantly improves the noise separation accuracy and is beneficial for subsequent noise filtering to obtain the target waveform signal.

[0186] Step 105: Based on the noise time-domain signal, perform noise filtering on the time-domain signal to be denoised to obtain the target waveform signal.

[0187] In some implementations, in order to accurately estimate and remove noise components from the original signal, the reconstructed noise time-domain signal can be used to filter the noise in the time-domain signal to be denoised, thereby obtaining a clean target waveform signal to ensure the effectiveness and accuracy of the acquired underwater acoustic signal.

[0188] The target waveform signal can be a clean signal extracted from the time domain signal to be denoised after noise filtering. It represents the ideal output of the original time domain signal to be denoised after removing noise interference, retains the key features and information of the target signal, and has a high signal-to-noise ratio and clarity.

[0189] In some implementation methods, the target waveform signal, such as the sound of a ship or marine life, can be directly obtained by subtracting the target noise from the time-domain signal to be denoised in the underwater acoustic environment, without the interference of background noise. This facilitates further analysis, identification, or other applications of the target waveform signal, such as underwater target detection, localization, and tracking, demonstrating good practicality.

[0190] This application acquires the time-domain signal of the underwater acoustic signal to be denoised, and extracts features from the time-domain signal using a target model to obtain a feature tensor to be denoised, wherein the feature tensor includes multiple local feature blocks. Using the local gated recurrent unit attention mechanism of the target model, the local positional dependency and local temporal dependency features between each local feature block and its adjacent feature blocks in the feature tensor to be denoised are calculated in parallel to obtain local noise features. Using the global gated recurrent unit attention mechanism of the target model, the global positional dependency and global temporal dependency features of the feature tensor to be denoised are calculated to obtain global noise features. Using the decoder of the target model, the noise time-domain signal of the target noise is reconstructed from the time-domain signal to be denoised based on the noise mask matrix generated by fusing the local and global noise features. Based on the noise time-domain signal, noise filtering is performed on the time-domain signal to be denoised to obtain the target waveform signal. Therefore, by combining local gated recurrent unit attention (ROU) mechanisms and global gated recurrent unit attention (GU) mechanisms, the model can simultaneously capture local and global dependencies in sequence data. It can accurately capture transient noise and short-term signal details through the local GUU attention mechanism, while modeling long-term dependencies through the global GUU attention mechanism, thus accurately identifying and utilizing complex patterns in the signal. This allows for more effective preservation of the target signal's features during noise reduction. Furthermore, this application directly reconstructs a clear signal waveform in the time domain and learns and extracts noise patterns only, facilitating noise filtering rather than enhancing the target waveform signal. This avoids the phase and amplitude decoupling problem in traditional methods, effectively improving the accuracy of the reconstructed waveform signal.

[0191] In some implementations, to obtain a target model with better performance, a preset model can be trained using the time-domain signal of the sample underwater acoustic data to be denoised. This enables the model to accurately learn, identify, and separate noise features, and recover a clean predicted waveform signal, ensuring that the model can still effectively identify noise features when faced with complex signals. For example, the target model can be trained in the following way:

[0192] (A.1) Obtain the sample time-domain signal of the underwater acoustic noise to be denoised, and extract features from the sample time-domain signal using a preset model to obtain the sample time-domain signal tensor, wherein the sample time-domain signal tensor includes multiple sample local feature blocks;

[0193] (A.2) By using the local control loop unit attention mechanism of the preset model, the sample local position dependency features and sample local time dependency features between each sample local feature block and the sample adjacent feature blocks in the sample time domain signal tensor are calculated in parallel to obtain the sample local noise features;

[0194] (A.3) By using the global gated recurrent unit attention mechanism of the preset model, the global positional dependence feature and global time dependence feature of the sample time domain signal tensor are calculated to obtain the global noise feature of the sample;

[0195] (A.4) Using the decoder of the preset model, based on the sample noise mask matrix generated after fusing the sample local noise features and sample global noise features, the predicted noise time domain signal of the sample noise is reconstructed from the sample time domain signal;

[0196] (A.5) Based on the predicted noise time-domain signal, noise filtering is performed on the sample time-domain signal to obtain the predicted waveform signal;

[0197] (A.6) Determine the target loss based on the predicted noise time-domain signal and the predicted waveform signal;

[0198] (A.7) Train the preset model based on the target loss to obtain the target model.

[0199] The underwater sound samples to be denoised can be raw underwater sound collected from an underwater environment (e.g., a marine environment) containing noise (such as wind, rain, currents, marine life, etc.) and sounds emitted by the target being detected (such as ships, submarines, etc.). During the training of the preset model, each underwater sound sample to be denoised corresponds to a sample noise time-domain signal and a sample waveform signal. Alternatively, the underwater sound samples to be denoised can also be obtained directly from a preset training dataset.

[0200] Among them, the sample time-domain signal can be the time series data corresponding to the underwater acoustic sound of the sample to be denoised, which is a form that can be directly used for digital signal processing.

[0201] The preset model can be an initial model structure set in the early stage of training. By training the preset model, the target model can be obtained.

[0202] Among them, the sample time-domain signal tensor can be a high-dimensional data structure obtained by performing convolution operation on the input sample time-domain signal through a preset model, dividing it according to a preset partitioning method, and then splicing it together. It contains key feature information of sample noise.

[0203] Among them, the sample local feature block can be a number of sub-parts into which the sample time domain signal tensor is divided. Each sub-part (each sample local feature block) represents the local characteristics of the sample time domain signal in a specific time period, which is convenient for subsequent learning of local and global dependencies.

[0204] Among them, the adjacent feature blocks of a sample can be local feature blocks of a sample that are physically or temporally adjacent to the local feature block of the sample that needs to be weighted in the current sample time domain signal tensor.

[0205] Among them, the sample local position-dependent features can be the features (such as transient impulse noise or high-frequency interference) extracted when analyzing each sample local feature block and its adjacent feature blocks under the action of the local gate-controlled recurrent unit attention mechanism, which reflect the relative positional relationship between the two. These features are used to characterize the spatial correlation between different sample local feature blocks.

[0206] Among them, the sample local time-dependent features can be the features (such as the persistence or periodic fluctuation of low-frequency noise) extracted for each sample local feature block and its adjacent feature blocks under the action of the local gate-controlled recurrent unit attention mechanism, which reflect the time series dependency between the two. These features are used to characterize the dynamic change law of different sample local feature blocks in the time dimension.

[0207] Among them, the sample local noise features can be a comprehensive representation extracted from the sample time-domain signal tensor after being processed by the local departmental recurrent unit attention mechanism. This representation can describe the local positional dependence features and local time dependence features within each sample local feature block and between it and its adjacent feature blocks.

[0208] Among them, the global location-dependent features of the sample can be used to characterize the spatial correlation between local feature blocks of each sample within the entire time domain signal range of the sample, which helps to understand the overall structure of the sample time domain signal. For example, in the sample time domain signal, noise in a certain frequency band may affect the energy distribution of other frequency bands.

[0209] Among them, the global time-dependent features of the sample can be used to characterize the dynamic changes of the entire sample time-domain signal over a long time scale, helping to identify and separate global noise components, such as the periodic changes of noise or the temporal continuity of the sample time-domain signal.

[0210] The global noise features of a sample can be a comprehensive representation of noise components by a pre-defined model that integrates global location-dependent and time-dependent information. These global noise features can include global characteristics such as the noise's spectral distribution and temporal energy variations, and are used to identify noise from mixed signals.

[0211] The sample noise mask matrix can be formed by adding the sample local noise features and the sample global noise features, normalizing the features, and then further compressing them to [0,1].

[0212] The predicted noise time-domain signal can be a specific noise signal reconstructed by the decoder based on the sample noise mask matrix, represented in time series form. The predicted noise time-domain signal is used to accurately characterize the noise components in the sample time-domain signal and can be used in the subsequent noise filtering process to remove noise from the original signal and recover a clear predicted waveform signal.

[0213] The predicted waveform signal can be a clean signal estimate obtained by filtering noise based on the predicted noise time domain signal, representing the ideal output after removing noise interference.

[0214] The target loss can be a loss value calculated based on the differences between the predicted noise time-domain signal and the actual sample noise time-domain signal, as well as between the predicted waveform signal and the actual sample waveform signal, and is used to guide the optimization process of the preset model.

[0215] For example, a comprehensive underwater acoustic noise reduction dataset covering various scenarios and conditions can be constructed for training and evaluating models. The underwater acoustic noise reduction data in this application is aggregated based on the public underwater acoustic signal dataset ShipsEar, generated using two types of ship signals: passenger ships and ROROs (Ro-Ro ships), and two types of underwater environmental noise signals (wind noise, rain noise, and current noise). The underwater acoustic noise reduction dataset mixes ship radiated noise signals with randomly selected environmental noise signals to synthesize a total of 12,274 sample time-domain signals with extremely low signal-to-noise ratios [-15dB, -10dB], [-10dB, -5dB], and [-5dB, 0dB]. 60% of these are used for training, 20% for training and validation, and 20% for evaluation (the proportions can be adjusted according to actual conditions). All sample time-domain signals are resampled to 16kHz and clipped to 3s. Analysis of the time-domain spectrum and power spectrum of the sample reveals that the passenger ship has relatively clear spectral lines in the 0-500 Hz and 500-1 kHz ranges. However, at low signal-to-noise ratios, the passenger ship's signal is almost completely masked, while noise exists in both low and high frequencies. Therefore, it is necessary to train the pre-defined model to enable it to accurately identify noise.

[0216] In some implementations, steps (A.1) to (A.5) involve processing the sample time-domain signal using a preset model to obtain a sample noise mask matrix, followed by noise filtering of the sample time-domain signal based on the sample noise mask matrix to obtain a predicted waveform signal. This process is identical to the process described above, which involves processing the denoised time-domain signal using a target model to obtain a noise mask matrix, followed by noise filtering of the denoised time-domain signal based on the noise mask matrix to obtain a target waveform signal. Both have the same algorithm architecture and operation steps, differing only in the stage of the signal processing model (preset model and target model, respectively) and the nature of the signal (offline sample data to be labeled and real-time input signal data, respectively). Therefore, the processing flow described above can be referenced, and will not be elaborated upon here.

[0217] In some implementations, the target loss can be determined by combining the noise prediction term and the signal prediction term. Specifically, this application designs a weighted noise loss function (r-nSI-SNR) to simultaneously consider the accuracy of noise prediction and the accuracy of signal prediction obtained from noise estimation based on the similarity between the predicted noise time-domain signal and the sample noise time-domain signal, as well as the predicted waveform signal and the sample waveform signal.

[0218] Furthermore, the gradient of the loss function can be propagated back to the parameters of the preset model using the backpropagation algorithm, updating the weights and biases of the preset model, and iteratively optimizing: repeating the forward propagation, target loss calculation, and backpropagation steps until the target loss of the preset model converges to a satisfactory level. For example, if the target loss is lower than the preset value (such as 0.01) for three consecutive times, or if the number of training iterations reaches the preset number (such as 500 or 1500), the training of the preset model can be stopped, and the optimized target model can be obtained.

[0219] By using target loss, the pre-defined model can continuously adjust its parameters to minimize the difference between predicted noise and real noise, while maximizing the similarity between predicted waveform and real waveform, enabling the model to run more stably and efficiently in practical applications.

[0220] In some implementations, to enable the model to learn noise features more accurately and reconstruct the noise signal, the target loss can be determined by comparing the similarity between the predicted noise time-domain signal and the sample noise time-domain signal, as well as between the predicted waveform signal and the sample waveform signal. This improves model performance, guides the optimization direction, and subsequently enables accurate recovery of the clean target signal. For example, (A.6) may also include:

[0221] (A.6.1) Obtain the time-domain signal of the sample noise and the sample waveform signal corresponding to the underwater acoustic noise to be denoised;

[0222] (A.6.2) Calculate the corresponding first similarity based on the predicted noise time-domain signal and the sample noise time-domain signal;

[0223] (A.6.3) Calculate the corresponding second similarity based on the predicted waveform signal and the sample waveform signal;

[0224] (A.6.4) Determine the target loss based on the first similarity and the second similarity.

[0225] Among them, the sample noise time-domain signal can be a time series signal containing only noise components extracted from the underwater acoustic samples to be denoised. It is a known real noise signal, used to compare with the predicted noise time-domain signal generated by the preset model in order to evaluate the accuracy of the preset model in estimating the noise components.

[0226] Among them, the sample waveform signal can be the ideal clean signal after removing noise from the underwater acoustic sample to be denoised, that is, the real target signal. It is used to compare with the predicted waveform signal generated by the preset model in order to evaluate the accuracy of the preset model in recovering the waveform signal.

[0227] The first similarity can be a metric derived by comparing the similarity between the predicted noise time-domain signal and the sample noise time-domain signal. The first similarity can be calculated using methods such as mean squared error, cosine similarity, or other appropriate similarity metrics.

[0228] The second similarity can be a metric derived by comparing the similarity between the predicted waveform signal and the sample waveform signal. It can also be calculated using methods such as mean square error and cosine similarity.

[0229] For example, the sample noise time-domain signal and sample waveform signal corresponding to the sample underwater acoustic noise to be denoised can be obtained by means of signal separation technology (such as blind source separation technology), signal processing method (such as filtering, spectral subtraction or statistical method), expert knowledge, and pre-setting for the sample underwater acoustic noise to be denoised. It is only necessary to ensure that the sample noise time-domain signal and sample waveform signal are clean and accurate signals separated from the sample underwater acoustic noise to be denoised. This application does not limit the specific acquisition method.

[0230] In some implementations, if the predicted noise time-domain signal obtained by the preset model is n′, the sample noise time-domain signal is n, the predicted waveform signal is mix-n′, and the sample waveform signal is s, then the target loss r-nSISNR can be calculated using the following formula:

[0231] r-nSISNR=αSISNR(n',n)+βSISNR(mix-n',s);

[0232] Where α is a weighting coefficient used to balance the importance of noise prediction in target loss; β is a weighting coefficient used to balance the importance of signal prediction in target loss; SISNR(n',n) represents the first similarity and SISNR(mix-n',s) represents the second similarity.

[0233] By calculating the target loss in the above manner, the preset model can better learn the characteristics of noise from the mixed signal and accurately separate the clean target signal, thereby improving the noise reduction performance of the model.

[0234] Please refer to Figure 3 and Figure 4 In some implementations, combined with Figure 3 and Figure 4 This section will introduce the overall process of this application. First, refer to... Figure 4After obtaining the underwater acoustic signal to be denoised, the time-domain signal to be denoised can be encoded by a one-dimensional convolutional layer (Conv1D) and a rectified linear unit (ReLU) to generate the denoised features. The denoised features are then divided into multiple local feature blocks and spliced ​​together to obtain the denoised feature tensor. In this way, the phase decoupling problem of STFT can be avoided by using time-domain convolution.

[0235] Furthermore, the tensor of the feature to be denoised can be input into a multi-view noise learning module composed of multiple self-attention modules (MV-MHSA). Each self-attention module contains a local gated recurrent unit attention mechanism and a global gated recurrent unit attention mechanism to extract local position / time dependent features (such as transient noise) and global position / time dependent features (such as background noise), respectively, and obtain the corresponding local noise features and global noise features.

[0236] Furthermore, the noise mask matrix (i.e., the noise mask matrix) obtained by fusing local and global noise features using components such as Parametrically Modified Linear Unit (PReLU), 2D convolutional layers (Conv2D), and hyperbolic tangent function (Tanh) in the target model can be reconstructed by a decoder (1D transposed convolution) to obtain the target waveform signal. The target waveform signal is then obtained by subtracting the noise signals in the time domain. This end-to-end time-domain processing framework bypasses the phase estimation problem of time-frequency domain methods and, combined with a multi-view self-attention mechanism, separates the target waveform signal from the noise, significantly improving the signal-to-noise ratio and signal fidelity in low signal-to-noise ratio scenarios.

[0237] Furthermore, Figure 3 This is a detailed architecture diagram of a self-attention module (MV-MHSA). After the tensor of the feature to be denoised is input into the local gated recurrent unit attention mechanism and the global gated recurrent unit attention mechanism respectively, the target model can extract noise features in parallel in multiple subspaces through the multi-head mechanism, thereby enhancing its adaptability to complex noise patterns.

[0238] Furthermore, the outputs of the local gated recurrent unit attention mechanism and the global gated recurrent unit attention mechanism can be fused by superposition (Add) and layer normalization (LayerNorm), and combined with the Leaky ReLU function to suppress gradient vanishing. Finally, a noise mask matrix is ​​generated by the Sigmoid function to identify the noise-dominant region (with weights of 0-1). Finally, the noise time-domain signal is extracted by multiplying the feature to be denoised and the noise mask matrix.

[0239] Please refer to Figure 5 In some implementations, a RORO signal from the test set of the underwater acoustic noise reduction dataset is taken as an example. Figure 5The power spectrum of the mixed signal shows that the time-domain signal to be denoised ( Figure 5 The power spectrum and target noise (marked as Figure A in the figure) Figure 5 The power spectrum of the signal (marked as Figure C) is extremely similar to that of the real waveform signal (Figure C). Figure 5 The target noise (marked as Figure B) is almost completely submerged in noise and cannot be distinguished. This application obtains a noise mask matrix through a target model, reconstructs the time-domain noise signal of the target noise based on the noise mask matrix, and obtains the target waveform signal (by time-domain subtraction) based on the time-domain noise signal. Figure 5 The waveform (marked as D in the image) is extremely similar to the actual target waveform signal. Calculations show that after denoising using the target model, the signal-to-interference plus noise ratio (SISNR) of the time-domain signal to be denoised is improved by 21.7 dB (dB is decibel), and the signal-to-noise ratio (SDR) is improved by 22.3 dB. This demonstrates that the target model can recover clean signal and noise components even at low SDR levels.

[0240] Please see Figure 6 This application also provides an underwater acoustic signal noise reduction device, which can implement the above-mentioned underwater acoustic signal noise reduction method. The underwater acoustic signal noise reduction device includes:

[0241] The acquisition module 61 is used to acquire the time-domain signal of the underwater acoustic signal to be denoised, and to extract features from the time-domain signal to be denoised through the target model to obtain the feature tensor to be denoised, wherein the feature tensor to be denoised includes multiple local feature blocks.

[0242] The first calculation module 62 is used to calculate the local positional dependency features and local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised through the local door-controlled recurrent unit attention mechanism of the target model, so as to obtain the local noise features.

[0243] The second calculation module 63 is used to calculate the global positional dependency features and global temporal dependency features of the tensor of the feature to be denoised through the global gated recurrent unit attention mechanism of the target model, so as to obtain the global noise features.

[0244] The reconstruction module 64 is used to reconstruct the noise time-domain signal of the target noise from the time-domain signal to be denoised, based on the noise mask matrix generated by fusing local noise features and global noise features, through the decoder of the target model.

[0245] The filtering module 65 is used to filter noise from the noise time-domain signal to be denoised, and obtain the target waveform signal.

[0246] The specific implementation of this underwater acoustic signal noise reduction device is basically the same as the specific embodiment of the underwater acoustic signal noise reduction method described above, and will not be repeated here. Subject to meeting the requirements of the embodiments of this application, the underwater acoustic signal noise reduction device may also be equipped with other functional modules to implement the underwater acoustic signal noise reduction method described above.

[0247] This application also provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described underwater acoustic signal noise reduction method. This computer device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0248] Please see Figure 7 , Figure 7 The hardware structure of a computer device according to another embodiment is illustrated. The computer device includes:

[0249] The processor 71 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application.

[0250] The memory 72 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 72 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 72 and is called and executed by the processor 71 using the underwater acoustic signal noise reduction method of the embodiments of this application.

[0251] Input / output interface 73 is used to implement information input and output;

[0252] The communication interface 74 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0253] Bus 75 transmits information between various components of the device (e.g., processor 71, memory 72, input / output interface 73, and communication interface 74);

[0254] The processor 71, memory 72, input / output interface 73, and communication interface 74 are connected to each other within the device via bus 75.

[0255] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described underwater acoustic signal noise reduction method.

[0256] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0257] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0258] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0259] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0260] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0261] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0262] It should be understood that in this application, "at least one" and "several" refer to one or more, and "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0263] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0264] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0265] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0266] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0267] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for denoising underwater acoustic signals, characterized in that, The method includes: The time-domain signal of the underwater acoustic sound to be denoised is obtained, and the feature of the time-domain signal to be denoised is extracted through the target model to obtain the feature tensor to be denoised, wherein the feature tensor to be denoised includes multiple local feature blocks. By using the local door-controlled recurrent unit attention mechanism of the target model, the local positional dependency features and local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised are calculated in parallel to obtain the local noise features; The global noise features are obtained by calculating the global positional dependency features and global temporal dependency features of the tensor of the feature to be denoised through the global gated recurrent unit attention mechanism of the target model. The noise time-domain signal of the target noise is reconstructed from the time-domain signal to be denoised by the decoder of the target model based on the noise mask matrix generated by fusing the local noise features and the global noise features. Based on the noise time-domain signal, noise filtering is performed on the time-domain signal to be denoised to obtain the target waveform signal.

2. The underwater acoustic signal noise reduction method according to claim 1, characterized in that, The local noise features are obtained by using the local door-controlled recurrent unit attention mechanism of the target model to compute in parallel the local positional dependency features and local temporal dependency features between each local feature block and its neighboring feature blocks in the feature tensor to be denoised, including: By using each local attention head of the local departmental control recurrent unit attention mechanism of the target model, the local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised are calculated in parallel to obtain the local temporal features corresponding to the multiple local feature blocks. Multiple local temporal features corresponding to multiple local attention heads are concatenated to obtain a local temporal feature tensor; The local positional dependency features of the local temporal feature tensor are calculated by the local gated layer of the local gated recurrent unit attention mechanism to obtain local noise features.

3. The underwater acoustic signal noise reduction method according to claim 2, characterized in that, The local positional dependency features of the local temporal feature tensor are calculated by the local gated layer through the local gated recurrent unit attention mechanism to obtain local noise features, including: The local temporal feature tensor and the feature tensor to be denoised are fused to obtain the first feature; The local position-dependent features of the first feature are calculated through the local gate control layer of the local gate control recurrent unit attention mechanism to obtain the second feature; The second feature is nonlinearly transformed using the leakage linear rectification function of the local departmental control loop unit attention mechanism to obtain the third feature; The local noise features are obtained by fusing the first feature and the third feature.

4. The underwater acoustic signal noise reduction method according to claim 1, characterized in that, The global noise features are obtained by calculating the global positional and temporal dependencies of the feature tensor to be denoised using the global gated recurrent unit attention mechanism of the target model, including: By using each global attention head of the global gated recurrent unit attention mechanism of the target model, the global temporal dependency features of the tensor of the feature to be denoised are calculated, and the global temporal features corresponding to each global attention head are obtained. Multiple global temporal features corresponding to multiple global attention heads are concatenated to obtain a global temporal feature tensor; The global positional dependency features of the global temporal feature tensor are calculated through the global gating layer of the global gating recurrent unit attention mechanism to obtain the global noise features.

5. The underwater acoustic signal noise reduction method according to claim 1, characterized in that, The target model is trained in the following way: The sample time-domain signal of the underwater sound to be denoised is acquired, and the sample time-domain signal is feature extracted by a preset model to obtain a sample time-domain signal tensor, wherein the sample time-domain signal tensor includes multiple sample local feature blocks. By using the local control loop unit attention mechanism of the preset model, the sample local position dependency features and sample local time dependency features between each sample local feature block and its adjacent feature blocks in the sample time domain signal tensor are calculated in parallel to obtain the sample local noise features. By using the global gated recurrent unit attention mechanism of the preset model, the global positional dependence features and global temporal dependence features of the sample time-domain signal tensor are calculated to obtain the global noise features of the sample. The decoder of the preset model reconstructs the predicted noise time-domain signal of the sample noise from the sample time-domain signal based on the sample noise mask matrix generated by fusing the sample local noise features and the sample global noise features. Based on the predicted noise time-domain signal, noise filtering is performed on the sample time-domain signal to obtain the predicted waveform signal; The target loss is determined based on the predicted noise time-domain signal and the predicted waveform signal; The target model is obtained by training the preset model based on the target loss.

6. The underwater acoustic signal noise reduction method according to claim 5, characterized in that, The step of determining the target loss based on the predicted noise time-domain signal and the predicted waveform signal includes: Obtain the time-domain signal of the sample noise and the sample waveform signal corresponding to the underwater acoustic noise to be denoised. Calculate the corresponding first similarity based on the predicted noise time-domain signal and the sample noise time-domain signal; Calculate the corresponding second similarity based on the predicted waveform signal and the sample waveform signal; The target loss is determined based on the first similarity and the second similarity.

7. The underwater acoustic signal noise reduction method according to claim 1, characterized in that, The step of extracting features from the time-domain signal to be denoised using the target model to obtain the feature tensor to be denoised includes: The convolutional neural network of the target model is used to extract features from the time-domain signal to be denoised, thereby obtaining the corresponding features to be denoised. The noise reduction feature is divided into multiple local feature blocks according to the preset division length and division number; The multiple local feature blocks are spliced ​​together to obtain the feature tensor to be denoised.

8. A device for reducing noise in underwater acoustic signals, characterized in that, The device includes: The acquisition module is used to acquire the time-domain signal of the underwater acoustic signal to be denoised, and to extract features from the time-domain signal to be denoised through the target model to obtain the feature tensor to be denoised, wherein the feature tensor to be denoised includes multiple local feature blocks. The first calculation module is used to calculate the local positional dependency features and local temporal dependency features between each local feature block and its adjacent feature blocks in the tensor of the feature to be denoised in parallel through the local departmental control recurrent unit attention mechanism of the target model, so as to obtain the local noise features. The second calculation module is used to calculate the global positional dependency features and global temporal dependency features of the tensor of the feature to be denoised through the global gated recurrent unit attention mechanism of the target model, so as to obtain the global noise features. The reconstruction module is used to reconstruct the noise time-domain signal of the target noise from the time-domain signal to be denoised, based on the noise mask matrix generated by fusing the local noise features and the global noise features, through the decoder of the target model. The filtering module is used to filter noise from the time-domain signal to be denoised based on the noise time-domain signal to obtain the target waveform signal.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the underwater acoustic signal noise reduction method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the underwater acoustic signal noise reduction method according to any one of claims 1 to 7.

Images