A spatio-temporal double-blind hydrophone array signal self-supervised denoising method and system

Abstract

The application discloses a spatio-temporal double-blind hydrophone array signal self-supervised denoising method and system, and relates to the technical field of hydrophone array signal processing and deep learning. The method obtains time-domain observation signals of multiple hydrophone array elements and forms an input tensor in a spatial order, and maps the input tensor to a high-dimensional feature space through a feature extraction head; a time-domain blind spot mode, a spatial domain blind channel mode and a full context mode are respectively constructed through a mode conditioning context mask module; and each mode output is obtained through a shared deep feature enhancement network and a feature reconstruction tail. During training, the time-domain blind spot output is used as a pseudo-reference after gradient stopping, and the spatial domain blind channel output and the full context output are constrained; during inference, the three outputs are fused to obtain a denoised signal. The method does not require a pure reference signal, can suppress time-domain random noise and channel-independent noise, and can maintain the waveform continuity and spatial coherence of the array signal.

Description

Technical Field

[0001] This invention relates to the fields of hydrophone array signal processing and deep learning technology, specifically to a spatiotemporal double-blind self-supervised denoising method and system for hydrophone array signals. Background Technology

[0002] Hydrophone arrays, as the core sensing unit of passive underwater acoustic systems, can sample the spatial distribution of sound fields. By utilizing the time delay, amplitude, and coherence information between channels, the system can achieve target detection, tracking, identification, localization, and long-term monitoring. However, the quality of actual array recordings is often compromised by factors such as environmental noise, electrical noise, flow noise, platform self-noise, fluid disturbances, and transient interference. Furthermore, the scarcity of clean reference signals limits the application of supervised denoising methods in real underwater measurement environments, prompting researchers to turn to self-supervised learning schemes that rely solely on noisy observation signals.

[0003] Unlike conventional single-channel signal denoising, hydrophone array signal denoising not only needs to suppress noise interference but also must maintain the waveform continuity, inter-channel phase relationship, time delay structure, and spatial coherence of the target acoustic signal. These physical properties directly determine the reliability of subsequent tasks such as direction finding, positioning, and target monitoring. If the denoising process disrupts the phase relationship or spatial coherence between array elements, even if the noise level of a single-channel signal is reduced, it may lead to deviations in the subsequent array processing results.

[0004] Traditional underwater acoustic denoising techniques are primarily based on statistical signal processing, filtering theory, and time-frequency analysis. These methods typically offer good interpretability and computational efficiency, but their performance is highly dependent on prior assumptions about the statistical characteristics of signal and noise, and requires manual parameter adjustments for different operating conditions. In complex and variable underwater acoustic environments, noise types, propagation paths, and background interference exhibit significant uncertainties, making it difficult for traditional methods to maintain stable denoising performance across various scenarios. Furthermore, existing methods often struggle to simultaneously address temporal random noise and spatial channel-independent noise, often resulting in problems such as noise residue, loss of signal detail, or damage to the array's spatial structure.

[0005] In recent years, deep learning methods have shown great potential in the field of underwater acoustic signal denoising due to their strong nonlinear modeling capabilities and data-driven feature learning capabilities. Existing deep denoising methods mostly employ supervised learning, using noisy observation signals as input and clean or near-clean signals as training targets. While these methods can achieve good denoising performance in some tasks, they typically rely on a large number of noisy-clean paired samples, which are difficult to obtain in real underwater measurement scenarios. Even if training data is constructed through simulation or laboratory conditions, its distribution may differ from the actual deployment environment, thus affecting the model's generalization ability.

[0006] Noise in hydrophone array signals is not limited to a single form. On the one hand, environmental random noise may exhibit weak predictability in the time dimension, making it suitable for suppression through temporal blind spot prediction. On the other hand, element self-noise, local circuit noise, or channel-specific interference may be relatively independent among different hydrophone elements, making it difficult to fully eliminate such noise by relying solely on temporal blind spot prediction. Therefore, how to simultaneously utilize the temporal predictability and array spatial correlation of the signal under conditions without clean labels to suppress temporal random noise and spatial channel-independent noise, while preserving the waveform details, phase relationships, and spatial coherence structure of the array signal, is a key technical challenge in current intelligent denoising of hydrophone arrays. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a spatiotemporal double-blind self-supervised denoising method and system for hydrophone array signals, which solves the problems of existing hydrophone array signal denoising methods relying on a clean reference signal, having difficulty simultaneously suppressing temporal random noise and spatial channel-independent noise, and easily damaging the spatial coherence structure of the array.

[0008] The technical solution of this invention is as follows: On one hand, the present invention provides a spatiotemporal double-blind self-supervised denoising method for hydrophone array signals, comprising the following steps: The time-domain observation signals collected by multiple hydrophone array elements are acquired, and the collected time-domain observation signals are constructed into an input tensor according to the spatial arrangement order of the hydrophone array elements.

[0009] A spatiotemporal double-blind self-supervised denoising network for hydrophone array signals is constructed. The spatiotemporal double-blind self-supervised denoising network for hydrophone array signals is used to denoise the input tensor to obtain the denoised hydrophone array signal.

[0010] The spatiotemporal double-blind hydrophone array signal self-supervised denoising network was trained using the input tensor to obtain the trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network.

[0011] The input tensor is denoised using a trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network to obtain the denoised hydrophone array signal.

[0012] Furthermore, the spatiotemporal dual-blind hydrophone array signal self-supervised denoising network includes a feature extraction head, a pattern conditional context masking module, a shared deep feature enhancement network, a feature reconstruction tail, and a fusion module.

[0013] The feature extraction head is used to process the input tensor Mapping to a high-dimensional feature space yields a feature channel count of Shallow feature representation.

[0014] The mode conditional context masking module takes shallow feature representation as input, performs mask convolution operation on the input shallow feature representation through three mode branches, and applies different binary mask matrices to the convolution kernel weights of each mode branch to obtain features under different modes.

[0015] The shared deep feature enhancement network is used to perform deep feature extraction and nonlinear denoising on features under different modes to obtain deep features under different modes.

[0016] The feature reconstruction tail is used to map deep features from different modes back to the original audio space, respectively obtaining temporal blind spot mode outputs. Airspace blind channel mode output and full context mode output .

[0017] The fusion module is used to output the temporal blind spot mode. Airspace blind channel mode output and full context mode output The signals are then fused to obtain the final denoised hydrophone array signal.

[0018] Furthermore, the feature extraction head consists of at least one convolutional layer and a corresponding activation function.

[0019] Furthermore, the three modes include a temporal blind spot mode. Blind Channel Mode in Airspace and full context mode ; The time-domain blind spot mode adopts The binary mask matrix, where odd number and In the temporal blind spot mode, the binary mask matrix is ​​set to zero at the center position in the time dimension, and 1 at the other positions; The airspace blind channel mode adopts The binary mask matrix, where odd number and The binary mask matrix of the spatial blind channel mode is set to zero at the spatial position or center row corresponding to the channel of the current hydrophone array element to be reconstructed; The full context mode uses an all-one binary mask matrix.

[0020] Furthermore, the shared deep feature enhancement network is composed of It consists of a multi-level feature channel attention residual dilated convolution module; The single feature channel attention residual dilated convolution module includes a linear projection and activation layer, a dilated convolution module, an SE Block feature channel attention mechanism module, and residual connections; The linear projection and activation layer are composed of It consists of convolutional layers and ReLU activation functions; The dilated convolution module takes the linear projection and the output of the activation layer as input. In the temporal blind spot mode and the spatial blind channel mode, standard dilated convolution is used; in the full context mode, dense convolution with the same effective receptive field is used. The multi-level feature channel attention residual dilated convolution module is divided into two stages: the pre-stage... The level feature channel attention residual dilated convolution module adopts The convolution specification; after The level feature channel attention residual dilated convolution module adopts The convolution specification; The SE Block feature channel attention mechanism module takes the output of the dilated convolution module as input, first extracts channel-level feature descriptors through global average pooling, then learns the relative importance of each feature channel and generates attention weights through two fully connected layers; finally, the attention weights are applied to the output of the dilated convolution module channel by channel. The residual connection is used to sum the input of the current feature channel attention residual dilated convolution module with the output features processed by the activation layer, dilated convolution module, and SE Block feature channel attention mechanism module after linear projection.

[0021] Furthermore, during the training process, a multi-mode parallel self-supervised training strategy is adopted. The same noisy input tensor is fed into the temporal blind spot mode, spatial blind channel mode, and full context mode respectively after passing through the feature extraction head to obtain features under different modes. The features under the temporal blind spot mode are used as pseudo-reference signals after the gradient stopping operation, and the consistency constraints are applied to the features under the spatial blind channel mode and the features under the full context mode. The stopping gradient operation is represented as follows: (3); in, This is a false reference signal. This indicates that the gradient operation has been stopped. The self-supervised training loss during the training process is: (4); in, For self-supervised training loss, For temporal blind spot reconstruction loss, For airspace blind channel consistency loss, For full context consistency regularization loss, and All are balance coefficients greater than 0.

[0022] The temporal blind spot reconstruction loss is defined as: (4); in, Represents the mathematical expectation. express Norm.

[0023] The spatial blind channel consistency loss is defined as: (5); The full context consistency regularization loss is defined as: (6).

[0024] On the other hand, the present invention also provides a spatiotemporal double-blind hydrophone array signal self-supervised denoising system for implementing a spatiotemporal double-blind hydrophone array signal self-supervised denoising method, comprising: The data acquisition module is used to acquire time-domain observation signals collected by multiple hydrophone array elements and construct the acquired time-domain observation signals into an input tensor according to the spatial arrangement order of the hydrophone array elements.

[0025] The model building module is used to construct a spatiotemporally double-blind self-supervised denoising network for hydrophone array signals.

[0026] The training module is used to train the spatiotemporal double-blind hydrophone array signal self-supervised denoising network using the input tensor, so as to obtain the trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network.

[0027] The denoising module is used to denoise the input tensor using a trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network to obtain the denoised hydrophone array signal.

[0028] Thirdly, this application proposes an electronic device comprising: one or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the aforementioned spatiotemporal dual-blind hydrophone array signal self-supervised denoising method.

[0029] Fourthly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the aforementioned spatiotemporal dual-blind hydrophone array signal self-supervised denoising method.

[0030] Fifthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned spatiotemporal dual-blind hydrophone array signal self-supervised denoising method.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces a pattern-conditional context masking module to construct temporal blind spot mode, spatial blind channel mode and full context mode under the same network architecture, enabling the network to learn complementary features from temporal neighborhood, spatial neighborhood and full local context respectively, and achieve spatiotemporal joint denoising under the condition of no pure label.

[0032] 2. This invention can simultaneously suppress different types of underwater noise interference. Among them, the temporal blind spot mode helps suppress temporal random noise, the spatial blind channel mode helps suppress array element channel independent noise or local channel interference, and the full context mode helps compensate for waveform detail and signal energy loss caused by blind constraints.

[0033] 3. This invention adopts a self-supervised learning framework, which does not require a clean reference signal or noisy-clean paired samples. It can better adapt to the application conditions in real underwater measurement environments where clean labels are difficult to obtain, and improve the engineering applicability of the intelligent denoising method for hydrophone array signals.

[0034] 4. This invention employs an asymmetric gradient blocking consistency constraint strategy, which uses the output of the temporal blind spot mode as a stable pseudo-reference signal to guide the spatial blind channel mode and the full context mode to perform consistency learning, thereby reducing the risk of the full context mode degenerating into an identity mapping.

[0035] 5. In the inference stage, this invention adopts a multi-mode output fusion method, which can comprehensively utilize the noise suppression capability of the temporal blind spot mode, the spatial correlation modeling capability of the spatial blind channel mode, and the detail fidelity capability of the full context mode, thereby obtaining a denoising result that takes into account noise suppression, waveform continuity and spatial coherence. Attached Figure Description

[0036] Figure 1 This is a flowchart of the spatiotemporal double-blind hydrophone array signal self-supervised denoising method in an embodiment of the present invention.

[0037] Figure 2 This is a diagram illustrating the overall architecture of the spatiotemporal dual-blind hydrophone array signal self-supervised denoising network in an embodiment of the present invention.

[0038] Figure 3 This is a structural diagram of the CA-RDC module in an embodiment of the present invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are used to illustrate the technical solutions of the present invention, and not to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can adjust the number of network layers, convolutional kernel size, number of feature channels, dilation rate, loss function form, boundary processing method, and fusion weights.

[0040] Example 1: This invention proposes a spatiotemporally dual-blind self-supervised denoising method for hydrophone array signals. It utilizes the differences in spatial and temporal correlation between signal and noise to preserve signal components and suppress noise components. By introducing a mode-conditional context masking mechanism, three paths are constructed: a temporal blind spot mode, a spatial blind channel mode, and a full context mode. Furthermore, a feature channel attention mechanism is combined to enhance feature extraction capabilities, achieving high-quality spatiotemporal joint denoising under unlabeled conditions. This method can handle environmental random noise and element channel-independent noise in hydrophone array signals and helps maintain the temporal waveform and array spatial structure of the target acoustic signal. Figure 1 As shown, the specific steps include: Step 1: Acquire time-domain observation signals from multiple hydrophone array elements, and construct an input tensor from the acquired time-domain observation signals according to the spatial arrangement order of the hydrophone array elements.

[0041] The input tensor is Where 1 indicates that the input is in the form of a single feature map. Indicates the number of elements in the hydrophone array. The number of time-domain sampling points is indicated. The time-domain observation signals acquired by the multiple hydrophone array elements are serialized according to the physical arrangement order of the hydrophone array elements in the array, so that the hydrophone array element index in the input tensor corresponds to the spatial position of the hydrophone array element, thereby preserving the spatial correlation, time delay relationship, and phase relationship between the time-domain observation signals acquired by different hydrophone array elements. The hydrophone array elements are set on an underwater platform, which includes an underwater robot, a towed platform, a fixed observation platform, or other underwater acoustic measurement platforms.

[0042] Step 2: Construct a spatiotemporal double-blind self-supervised denoising network for hydrophone array signals; the spatiotemporal double-blind self-supervised denoising network for hydrophone array signals is used to denoise the input tensor to obtain the denoised hydrophone array signal.

[0043] like Figure 2 As shown, the spatiotemporal dual-blind hydrophone array signal self-supervised denoising network includes a feature extraction head, a pattern conditional context masking module, a shared deep feature enhancement network, a feature reconstruction tail, and a fusion module.

[0044] The feature extraction head is used to process the input tensor Mapping to a high-dimensional feature space yields a feature channel count of The shallow feature representation; the feature extraction head consists of at least one convolutional layer and a corresponding activation function.

[0045] In this embodiment, the input tensor is first processed by a kernel size of... The convolutional layer, without changing the hydrophone-temporal two-dimensional resolution, maps from a single feature map to... A high-dimensional feature space with multiple feature channels is then nonlinearly mapped using the ReLU activation function. This design is used to complete shallow feature extraction and provide sufficient representation space for subsequent blind prediction and context modeling.

[0046] The mode conditional context masking module takes shallow feature representation as input, performs mask convolution operation on the input shallow feature representation through three mode branches, and applies different binary mask matrices to the convolution kernel weights of each mode branch to obtain features under different modes.

[0047] The convolutional kernel weights in different modes are represented as follows: (1); in, For mode parameters, For convolution kernel weights, Represents the original convolution kernel weights. Representation pattern The corresponding binary mask matrix, It represents the Hadamardi (or Hadama) stack.

[0048] The three modes include the temporal blind-spot mode. Spatial Blind-channel mode and Full-context mode .

[0049] The temporal blind spot mode is used to mask the current temporal sampling point and use temporal neighborhood information for prediction, forcing the network to predict the target temporal sampling point based on the temporal neighborhood information, thereby suppressing temporal noise with strong randomness; the temporal blind spot mode adopts The binary mask matrix, where odd number and In this design, the binary mask matrix of the temporal blind spot pattern is set to zero at its center position and 1 at other positions, thus severing the direct connection between the current temporal sampling point and itself. This design allows the network to avoid directly using the observation value of the current temporal sampling point when predicting it, and instead uses the temporal neighborhood information before and after the current temporal sampling point for prediction and reconstruction. This satisfies the blind spot prediction or J-invariant self-supervised denoising idea and helps to suppress temporal noise with strong randomness. In this embodiment, The binary mask matrix can be represented as . The spatial blind channel mode is used to shield the current hydrophone element channel to be reconstructed and to use the information of adjacent hydrophone element channels for prediction, forcing the network to reconstruct the signal based on the spatial correlation of adjacent hydrophone element channels, thereby suppressing channel-independent noise or element self-noise; the spatial blind channel mode adopts The binary mask matrix, where odd number and In the spatial blind channel mode, the binary mask matrix is ​​set to zero at the spatial position or center row corresponding to the current hydrophone element channel to be reconstructed. This allows the network to avoid directly using the observation information of the current hydrophone element channel when reconstructing the signal, and instead uses the spatial related information of adjacent hydrophone element channels for prediction and reconstruction. Since the target acoustic signal usually has a certain spatial coherence between adjacent hydrophone elements, and the self-noise, electrical noise, or local channel interference of some hydrophone elements are relatively independent between channels, this mode helps to suppress channel-independent noise and maintain the array spatial structure. In this embodiment, The binary mask matrix can be represented as , This is a transpose.

[0050] The full-context mode is used to compensate for waveform detail and signal energy loss caused by temporal blind spot mode and spatial blind channel mode by utilizing the complete local context including the current time-domain sampling point. For hydrophone elements located at the boundary of the hydrophone array or time-domain sampling points located at the time boundary, local context is constructed by zero-padding, mirror padding, copy padding, or using only the effective neighborhood. The boundary processing method does not change the basic idea of ​​self-supervised prediction using temporal and spatial neighborhoods. The full-context mode uses a binary mask matrix of all 1s, enabling the network to extract features using the complete local context information including the current time-domain sampling point to compensate for waveform detail or signal energy loss caused by blind constraints in temporal blind spot mode and spatial blind channel mode. To avoid the full-context mode degenerating into direct input copying in self-supervised training, the present invention further constrains it by consistency loss through asymmetric gradient blocking. In this embodiment, the binary mask matrix of the full-context mode can be represented as follows: .

[0051] The shared deep feature enhancement network is used to perform deep feature extraction and nonlinear denoising on features under different modes to obtain deep features under different modes.

[0052] The shared deep feature enhancement network is composed of It consists of a Channel-Attention Residual Dilated Convolution (CA-RDC) module.

[0053] like Figure 3 As shown, the single feature channel attention residual dilated convolution module includes a linear projection and activation layer, a dilated convolution module, an SE Block (Squeeze-and-Excitation Block) feature channel attention mechanism module, and residual connections.

[0054] The linear projection and activation layer are composed of The system consists of convolutional layers and ReLU activation functions, used to fuse latent feature channel information while maintaining the hydrophone-temporal two-dimensional resolution.

[0055] The dilated convolution module takes the linear projection and the output of the activation layer as input to expand the receptive field and extract temporal and spatial context features. In the temporal blind spot mode and the spatial blind channel mode, standard dilated convolution is used to expand the effective receptive field while keeping the number of parameters stable. In the full context mode, dense convolution (all-1 convolution) with the same effective receptive field range is used to enhance the expression of local structural details.

[0056] The The multi-level feature channel attention residual dilated convolution module is divided into two stages: the pre-stage... The level feature channel attention residual dilated convolution module adopts The convolution specification is used to capture the local and long-range correlations of the signal along the time axis; then... The level feature channel attention residual dilated convolution module adopts The convolution specification is used to fuse temporal features with array spatial dimension features to enhance the suppression of channel-independent noise and maintain array spatial consistency.

[0057] The execution logic of the dilated convolution module in this embodiment is as follows: with a two-dimensional dilation rate... of Taking convolution as an example, the effective coverage form of the standard dilated convolution kernel is: ,in, The first component represents the dilation rate in the direction of the hydrophone array elements, and the second component represents the dilation rate in the time direction. 1 indicates the location where the convolution weights are applied, and 0 indicates the location of the interval introduced by the dilated convolution. In full-context mode, a dense kernel is used within the same effective coverage area. The dense kernel is obtained through equivalent Dense convolution or a full-1 mask within the same effective receptive field can be used to reduce the loss of local structural information.

[0058] To more effectively filter useful acoustic features under self-supervised conditions lacking pure labels, the CA-RDC module embeds an SE Block (Squeeze-and-Excitation Block) feature channel attention mechanism module. This module achieves adaptive enhancement of effective acoustic components by modeling the dependencies between feature channels.

[0059] Specifically, the SE Block feature channel attention mechanism module takes the output of the dilated convolution module as input, first extracts channel-level feature descriptors through global average pooling, and then learns the relative importance of each feature channel and generates attention weights through two fully connected layers. Specifically, the first fully connected layer compresses the feature channel dimension to its original size. And it incorporates the ReLU activation function, in this embodiment The second fully connected layer restores the feature channel dimension to its uncompressed dimension and then connects it to the Sigmoid function, generating a range of... The channel attention weights are then applied to the output of the dilated convolution module channel by channel to achieve feature reweighting. In the self-supervised denoising process, different feature channels often correspond to different frequency bands or different spatiotemporal components. The SE Block feature channel attention mechanism module can adaptively enhance the effective features containing the target acoustic signal structure, while suppressing feature responses with strong noise residue, thereby improving the signal-to-noise separation capability in the deep feature space.

[0060] The residual connection is used to sum the input of the current feature channel attention residual dilated convolution module with the output features processed by the activation layer, dilated convolution module, and SE Block feature channel attention mechanism module after linear projection, so as to achieve the fusion of shallow features and deep abstract features, suppress the gradient vanishing problem in the training process of deep network, and maintain the detailed information features of the hydrophone array signal in the nonlinear transformation process.

[0061] In self-supervised denoising, different feature channels often correspond to different frequency bands or different spatiotemporal components. SEBlock can adaptively enhance the effective features containing the target acoustic signal structure while suppressing feature responses with strong noise residue, thereby improving the signal-to-noise separation capability in deep feature spaces.

[0062] The feature reconstruction tail is used to map deep features from different modes back to the original audio space, respectively obtaining temporal blind spot mode outputs. Airspace blind channel mode output and full context mode output .

[0063] The feature reconstruction tail is composed of It consists of convolutional layers, in which the first layer... Convolutional layers use Convolutional layers and the ReLU activation function achieve progressive compression and information integration of feature channel dimensions; the last convolutional layer uses... The convolutional layer maps features to a single-channel output and does not set an activation function to preserve the dynamic range and temporal waveform of the original hydrophone array signal.

[0064] The fusion module is used to output the temporal blind spot mode. Airspace blind channel mode output and full context mode output The signals are then fused to obtain the final denoised hydrophone array signal.

[0065] The fusion adopts either equal-weighted arithmetic mean fusion or weighted fusion, wherein equal-weighted arithmetic mean fusion is expressed as: (2); in, This represents the final denoised hydrophone array signal.

[0066] In this embodiment, the shared deep feature enhancement network is composed of The network consists of a multi-level feature channel attention residual dilated convolution module, and its parameters are as follows: , , , , , , .

[0067] Step 3: Train the spatiotemporal double-blind hydrophone array signal self-supervised denoising network using the input tensor to obtain the trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network.

[0068] During training, this invention employs a multi-mode parallel self-supervised training strategy. The same noisy input tensor is processed by a feature extraction head and then input into the temporal blind spot mode, spatial blind channel mode, and full context mode respectively, yielding features under different modes. The features under the temporal blind spot mode are used as pseudo-reference signals after a stopping gradient operation. Consistency constraints are applied to the features under the spatial blind channel mode and the full context mode, ensuring that the full context mode preserves details while avoiding direct learning of identity mappings.

[0069] The stopping gradient operation is represented as follows: (3); in, This is a false reference signal. The terminator, or detach operation, indicates the cessation of gradient operations. During the backpropagation of the spatial blind channel consistency loss and the full context consistency loss, this operation blocks the gradient path corresponding to the temporal blind point mode output, preventing its parameters from being updated. Thus, the temporal blind point mode is primarily driven by its own blind point reconstruction loss, while the spatial blind channel mode and the full context mode learn a denoised representation consistent with the blind point prediction result under the constraint of a pseudo-reference signal, thereby forming a stable multi-mode collaborative training mechanism.

[0070] The self-supervised training loss during the training process is: (4); in, For self-supervised training loss, For temporal blind spot reconstruction loss, For airspace blind channel consistency loss, For full context consistency regularization loss, and All are balance coefficients greater than 0.

[0071] The temporal blind spot reconstruction loss is defined as: (4); in, This represents the mathematical expectation, used here to calculate the expected value versus the target value across all training samples. Average distance loss, express Norm.

[0072] This loss function uses noisy observations as a self-supervised objective, training the network to predict the current temporal sampling point based on contextual information under blind spot constraints. Since the current temporal sampling point itself is masked in the temporal blind spot mode, the network has difficulty directly replicating the current noisy observation, thus helping to learn the effective signal structure in the temporal context.

[0073] The spatial blind channel consistency loss is defined as: (5); This loss is used to constrain the output of the spatial blind channel mode to be consistent with the output of the temporal blind point mode after gradient blocking, so that the spatial blind channel mode can suppress channel-independent noise by utilizing the information of adjacent array elements while maintaining consistency with the temporal blind point denoising results.

[0074] The full context consistency regularization loss is defined as: (6); This loss is used to constrain the output of the full-context mode to be consistent with the output of the temporal blind spot mode, which is blocked by gradients. Since the full-context mode has access to the complete local context, including the center sampling point, it may tend to learn a direct copy of noisy inputs if unconstrained.

[0075] This invention constrains the full context mode by using the stopping gradient pseudo-reference signal output by the time-domain blind spot mode, thereby reducing the risk of degrading to an identity mapping while restoring waveform details and signal energy.

[0076] Step 4: Use the trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network to denoise the input tensor to obtain the denoised hydrophone array signal.

[0077] Through the above training and reasoning process, this invention can comprehensively utilize the temporal predictability and spatial correlation of hydrophone array signals under the conditions of no pure reference signal and no noisy-pure paired samples to achieve synergistic suppression of environmental random noise and independent noise of array element channels, while maintaining the waveform continuity, inter-channel phase relationship and spatial coherence structure of the target acoustic signal.

[0078] The number of network layers, convolutional kernel size, number of feature channels, two-dimensional dilation rate, loss function form, balance coefficient, boundary processing method, and fusion weights in the above embodiments are only used to illustrate the specific implementation of the present invention and do not constitute a limitation on the scope of protection of the present invention. Under the technical concept of the present invention, any method that combines temporal blind spot mode, spatial blind channel mode, and full context mode, and achieves hydrophone array signal denoising through mode-conditional context masking, self-supervised training, and multi-mode inference fusion, should fall within the scope of protection of the present invention.

[0079] Example 2: A spatiotemporal double-blind self-supervised denoising system for hydrophone array signals, used to implement a spatiotemporal double-blind self-supervised denoising method for hydrophone array signals, comprising: The data acquisition module is used to acquire time-domain observation signals collected by multiple hydrophone array elements and construct the acquired time-domain observation signals into an input tensor according to the spatial arrangement order of the hydrophone array elements.

[0080] The model building module is used to construct a spatiotemporally double-blind self-supervised denoising network for hydrophone array signals.

[0081] The training module is used to train the spatiotemporal double-blind hydrophone array signal self-supervised denoising network using the input tensor, so as to obtain the trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network.

[0082] The denoising module is used to denoise the input tensor using a trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network to obtain the denoised hydrophone array signal.

[0083] Example 3: This embodiment proposes an electronic device, including: one or more processors, and a memory, wherein the memory is used to store instructions, and when the instructions are executed by the one or more processors, the one or more processors execute the aforementioned spatiotemporal dual-blind hydrophone array signal self-supervised denoising method.

[0084] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements a spatiotemporal dual-blind hydrophone array signal self-supervised denoising method as described in the embodiments. It is understood that the electronic device may also include input / output (I / O) interfaces and communication components.

[0085] The processor is used to execute all or part of the steps in the spatiotemporally dual-blind self-supervised denoising method for hydrophone array signals as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0086] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the spatiotemporal dual-blind hydrophone array signal self-supervised denoising method described in the above embodiments.

[0087] Example 4: This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0088] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the spatiotemporal double-blind hydrophone array signal self-supervised denoising method described in various embodiments of this application.

[0089] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory, etc.), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the aforementioned spatiotemporal dual-blind hydrophone array signal self-supervised denoising method.

[0090] Example 5: This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned spatiotemporal double-blind self-supervised denoising method for hydrophone array signals.

[0091] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0092] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0093] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.

Claims

1. A spatiotemporal double-blind self-supervised denoising method for hydrophone array signals, characterized in that, Includes the following steps: Acquire time-domain observation signals from multiple hydrophone array elements, and construct an input tensor from the acquired time-domain observation signals according to the spatial arrangement order of the hydrophone array elements; A spatiotemporal double-blind self-supervised denoising network for hydrophone array signals is constructed. The spatiotemporal double-blind self-supervised denoising network for hydrophone array signals is used to denoise the input tensor to obtain the denoised hydrophone array signal. The spatiotemporal double-blind hydrophone array signal self-supervised denoising network was trained using the input tensor to obtain the trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network. The input tensor is denoised using a trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network to obtain the denoised hydrophone array signal.

2. The spatiotemporal dual-blind hydrophone array signal self-supervised denoising method according to claim 1, characterized in that, The spatiotemporal dual-blind hydrophone array signal self-supervised denoising network includes a feature extraction head, a pattern conditional context masking module, a shared deep feature enhancement network, a feature reconstruction tail, and a fusion module. The feature extraction head is used to process the input tensor Mapping to a high-dimensional feature space yields a feature channel count of Shallow feature representation; The mode conditional context masking module takes shallow feature representation as input, performs mask convolution operation on the input shallow feature representation through three mode branches, and applies different binary mask matrices to the convolution kernel weights of each mode branch to obtain features under different modes. The shared deep feature enhancement network is used to perform deep feature extraction and nonlinear denoising on features under different modes to obtain deep features under different modes. The feature reconstruction tail is used to map deep features from different modes back to the original audio space, respectively obtaining temporal blind spot mode outputs. Airspace blind channel mode output and full context mode output ; The fusion module is used to output the temporal blind spot mode. Airspace blind channel mode output and full context mode output The signals are then fused to obtain the final denoised hydrophone array signal.

3. The spatiotemporal dual-blind hydrophone array signal self-supervised denoising method according to claim 2, characterized in that, The feature extraction head consists of at least one convolutional layer and a corresponding activation function.

4. The spatiotemporal dual-blind hydrophone array signal self-supervised denoising method according to claim 2, characterized in that, The three modes include the temporal blind spot mode. Blind Channel Mode in Airspace and full context mode ; The time-domain blind spot mode adopts The binary mask matrix, where It is an odd number and In the temporal blind spot mode, the binary mask matrix is ​​set to zero at the center position in the time dimension, and 1 at the other positions; The airspace blind channel mode adopts The binary mask matrix, where It is an odd number and The binary mask matrix of the spatial blind channel mode is set to zero at the spatial position or center row corresponding to the channel of the current hydrophone array element to be reconstructed; The full context mode uses an all-one binary mask matrix.

5. The spatiotemporal dual-blind hydrophone array signal self-supervised denoising method according to claim 2, characterized in that, The shared deep feature enhancement network is composed of It consists of a multi-level feature channel attention residual dilated convolution module; The single feature channel attention residual dilated convolution module includes a linear projection and activation layer, a dilated convolution module, an SEBlock feature channel attention mechanism module, and residual connections; The linear projection and activation layer are composed of It consists of convolutional layers and ReLU activation functions; The dilated convolution module takes the linear projection and the output of the activation layer as input. In the temporal blind spot mode and the spatial blind channel mode, standard dilated convolution is used; in the full context mode, dense convolution with the same effective receptive field is used. The multi-level feature channel attention residual dilated convolution module is divided into two stages: the pre-stage... The level feature channel attention residual dilated convolution module adopts The convolution specification; after The level feature channel attention residual dilated convolution module adopts The convolution specification; The SE Block feature channel attention mechanism module takes the output of the dilated convolution module as input, first extracts channel-level feature descriptors through global average pooling, then learns the relative importance of each feature channel and generates attention weights through two fully connected layers; finally, the attention weights are applied to the output of the dilated convolution module channel by channel. The residual connection is used to sum the input of the current feature channel attention residual dilated convolution module with the output features processed by the activation layer, dilated convolution module, and SE Block feature channel attention mechanism module after linear projection.

6. The spatiotemporal dual-blind hydrophone array signal self-supervised denoising method according to claim 1, characterized in that, During training, a multi-mode parallel self-supervised training strategy is adopted. The same noisy input tensor is fed into the temporal blind spot mode, spatial blind channel mode and full context mode respectively after passing through the feature extraction head to obtain features under different modes. The features under the temporal blind spot mode are used as pseudo-reference signals after stopping gradient operation to constrain the consistency of features under the spatial blind channel mode and features under the full context mode. The gradient stopping operation is represented as follows: （3）； in, This is a false reference signal. This indicates that the gradient operation has been stopped. The self-supervised training loss during the training process is: （4）； in, For self-supervised training loss, For temporal blind spot reconstruction loss, For airspace blind channel consistency loss, For full context consistency regularization loss, and All are balance coefficients greater than 0; The temporal blind spot reconstruction loss is defined as: （4）； in, Represents the mathematical expectation. express Norm; The spatial blind channel consistency loss is defined as: （5）； The full context consistency regularization loss is defined as: （6）。 7. A spatiotemporal double-blind self-supervised denoising system for hydrophone array signals, used to implement the spatiotemporal double-blind self-supervised denoising method for hydrophone array signals as described in any one of claims 1-6, characterized in that, include: The data acquisition module is used to acquire time-domain observation signals collected by multiple hydrophone array elements, and construct the acquired time-domain observation signals into an input tensor according to the spatial arrangement order of the hydrophone array elements; The model building module is used to construct a spatiotemporal double-blind self-supervised denoising network for hydrophone array signals; The training module is used to train the spatiotemporal double-blind hydrophone array signal self-supervised denoising network using the input tensor, so as to obtain the trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network. The denoising module is used to denoise the input tensor using a trained spatiotemporal double-blind hydrophone array signal self-supervised denoising network to obtain the denoised hydrophone array signal.

8. An electronic device, characterized in that, include: One or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform a spatiotemporally dual-blind hydrophone array signal self-supervised denoising method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed, cause the processor to perform a spatiotemporal dual-blind hydrophone array signal self-supervised denoising method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program or instructions that, when executed by a processor, implement the spatiotemporal dual-blind hydrophone array signal self-supervised denoising method as described in any one of claims 1-6.

Images