Passive sonar system-oriented full-complex-value underwater acoustic signal enhancement method and target neural network

Abstract

The invention provides a passive sonar system-oriented full-complex-value underwater acoustic signal enhancement method and a target neural network, which can be applied to the field of underwater acoustic signal processing. The full-complex-value underwater acoustic signal enhancement method comprises the following steps: converting an underwater radiation signal into an initial complex-value time-frequency spectrogram; performing down-sampling on the initial complex value time-frequency spectrogram, and outputting N complex value down-sampling features of N scales; respectively processing the N complex value down-sampling features based on a complex value convolution block attention mechanism, and outputting N first complex value attention features; processing the Nth complex value down-sampling feature based on a multi-dimensional complex value attention mechanism, and outputting a second complex value attention feature; and sequentially splicing the N first complex-valued attention features with the second complex-valued attention features, and performing up-sampling on the complex-valued attention features obtained by splicing after each time of splicing, thereby obtaining a de-noised target time-frequency spectrogram based on the complex-valued up-sampling features output by the complex-valued decoder group.

Description

Technical Field

[0001] This invention relates to the field of underwater acoustic signal processing, and more specifically to a method for enhancing fully complex underwater acoustic signals for passive sonar systems and a target neural network. Background Technology

[0002] For complex underwater environments, Underwater Acoustic Signal Enhancement (UWASE) technology can be used to suppress noise and improve the quality of acoustic signals in such environments. However, some UWASE implementations have limited enhancement effects on underwater acoustic signals. Furthermore, during noise suppression, weaker signals that differ from the noise are prone to over-elimination in areas of strong noise, leading to the incorrect removal of components of the target signal and resulting in signal distortion. Summary of the Invention

[0003] In view of the above problems, the present invention provides a method for enhancing fully complex underwater acoustic signals for passive sonar systems and a target neural network.

[0004] According to one aspect of the present invention, a method for enhancing fully complex-valued underwater acoustic signals for passive sonar systems is provided, comprising: converting an underwater radiation signal received by a passive sonar system into an initial complex-valued time-spectrum; the initial complex-valued time-spectrum is noisy; downsampling the initial complex-valued time-spectrum using a complex-valued encoder block group to output N complex-valued downsampled features at N scales; N being an integer greater than 1; processing the N complex-valued downsampled features respectively using a complex-valued attention gating module based on a complex-valued convolutional block attention mechanism to output N first complex-valued attention features; processing the Nth complex-valued downsampled feature using a complex-valued multi-dimensional attention module based on a multi-dimensional complex-valued attention mechanism to output a second complex-valued attention feature; and concatenating the N first complex-valued attention features sequentially with the second complex-valued attention feature using a complex-valued decoder block group, and upsampling the concatenated complex-valued attention feature after each concatenation, thereby obtaining a noise-removed target time-spectrum based on the complex-valued upsampled features output by the complex-valued decoder block group.

[0005] According to another aspect of the present invention, a target neural network is provided for implementing the above-described method for enhancing fully complex-valued underwater acoustic signals for passive sonar systems. The target neural network includes a complex-valued encoder block group, a complex-valued attention gating module, a complex-valued multi-dimensional attention module, and a complex-valued decoder block group. The target neural network is trained as follows: the sample complex-valued time-of-flight spectrum is input into an initial neural network to obtain a predicted complex-valued time-of-flight spectrum; based on a target loss function, the normalized mean square error between the predicted complex-valued time-of-flight spectrum and the label complex-valued time-of-flight spectrum corresponding to the sample complex-valued time-of-flight spectrum is calculated; the parameters of the initial neural network are adjusted based on the normalized mean square error to obtain the target neural network.

[0006] According to an embodiment of the present invention, the underwater radiation signal received by the passive sonar system is converted into an initial complex-valued time-of-flight spectrum. Then, a complex-valued encoder block group fully utilizes the complex-valued information of the initial complex-valued time-of-flight spectrum to downsample the spectrum, obtaining complex-valued downsampled features at different scales. Furthermore, a complex-valued attention gating module can process the corresponding complex-valued downsampled features at different scales using a complex-valued convolutional block attention mechanism. This allows for the enhancement of audio features other than noise by combining corresponding complex-valued information at different scales, thereby outputting the corresponding first complex-valued attention feature.

[0007] Then, the complex-valued multi-dimensional attention module processes the Nth complex-valued downsampled feature after the Nth downsampling based on a multi-dimensional complex-valued attention mechanism, further fully utilizing the complex-valued information of the Nth complex-valued downsampled feature to enhance audio features other than noise, and then outputs the corresponding second complex-valued attention feature.

[0008] Thus, by sequentially concatenating N first and second complex-valued attention features using a complex-valued decoder block group, and upsampling the concatenated complex-valued attention features after each concatenation, feature reconstruction can be performed sequentially at different scales to obtain the final target time-spectrum map.

[0009] Therefore, the entire process of this invention is modeled in the complex-valued domain. Furthermore, in this process, the invention combines a fully complex-valued target neural network and a multi-dimensional attention mechanism to fully utilize the amplitude and phase information of the spectrogram at the initial complex-valued stage. This maintains the integrity of the spectrogram structure and avoids the loss of amplitude and phase-related features during the process, thereby improving the enhancement performance of underwater radiation signals in audio form and significantly improving the accuracy of spectrogram reconstruction and waveform reconstruction. This enhances the recovery capability of weaker signals (excluding noise) in underwater radiation signals. Thus, it at least partially avoids the signal distortion problem caused by the removal of these weaker signals during the denoising process. Attached Figure Description

[0010] The above-mentioned contents, other objects, features and advantages of the present invention will become clearer from the following description of embodiments of the present invention with reference to the accompanying drawings, which will be described in conjunction with the drawings.

[0011] Figure 1 The diagram illustrates an application scenario of the fully complex underwater acoustic signal enhancement method according to an embodiment of the present invention.

[0012] Figure 2A A schematic diagram of a target neural network according to an embodiment of the present invention is shown.

[0013] Figure 2B A flowchart of a method for enhancing fully complex underwater acoustic signals according to an embodiment of the present invention is shown.

[0014] Figure 3 A schematic diagram of a target neural network according to another embodiment of the present invention is shown.

[0015] Figure 4 A schematic diagram of a complex-valued time-frequency module according to an embodiment of the present invention is shown.

[0016] Figure 5 A schematic diagram of a complex-valued convolutional block attention module according to an embodiment of the present invention is shown.

[0017] Figure 6 A schematic diagram of a target neural network according to another embodiment of the present invention is shown.

[0018] Figure 7A A schematic diagram of a complex value encoder block according to an embodiment of the present invention is shown.

[0019] Figure 7B A schematic diagram of a complex value decoder block according to an embodiment of the present invention is shown.

[0020] Figure 8 The diagram illustrates the effect of the target neural network and the baseline model in terms of spectrum and power spectral density according to an embodiment of the present invention.

[0021] Figure 9 A block diagram of an electronic device suitable for implementing a fully complex-valued underwater acoustic signal enhancement method for passive sonar systems, according to an embodiment of the present invention, is shown. Detailed Implementation

[0022] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0023] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0024] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0025] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0026] According to one aspect of the present invention, in order to improve the enhancement performance of underwater acoustic signals and alleviate distortion problems, the present invention proposes a fully complex-valued underwater acoustic signal enhancement method for passive sonar systems. The following is in conjunction with... Figure 1 The application scenarios of the fully complex underwater acoustic signal enhancement method of the present invention are described.

[0027] Figure 1 The diagram illustrates an application scenario of the fully complex underwater acoustic signal enhancement method according to an embodiment of the present invention.

[0028] like Figure 1As shown, application scenario 100 according to this embodiment may include a ship 110 and a passive sonar system 120. It should be understood that in an underwater environment, the signal quality is affected by various factors such as propagation distance, ocean background noise, and equipment interference. Therefore, the signal-to-noise ratio of the underwater radiation signal received by the passive sonar system from the ship is low, and the energy of the target signal in the underwater radiation signal is weak. Since underwater radiation signals can reflect important information such as ship structure and navigation status, it is necessary to improve the quality and identifiability of the underwater radiation signals received by the passive sonar system.

[0029] Therefore, the fully complex-valued underwater acoustic signal enhancement method of the present invention can be applied to low signal-to-noise ratio scenarios in complex marine environments and can be used to enhance the underwater radiation signals received by passive sonar systems from ships. This improves the quality and identifiability of the weak underwater radiation signals received by passive sonar systems.

[0030] Based on this, the present invention also provides a target neural network for implementing the above-mentioned fully complex underwater acoustic signal enhancement method. The following will be based on... Figure 1 The described scene, combined with Figure 2A and Figure 2B The method for enhancing fully complex underwater acoustic signals according to the present invention will be described in detail. It should be noted that the method for enhancing fully complex underwater acoustic signals according to the present invention can be executed by a passive sonar system, or by other electronic devices different from a passive sonar system; the present invention does not limit this.

[0031] Figure 2A A schematic diagram of a target neural network according to an embodiment of the present invention is shown. Figure 2B A flowchart of a method for enhancing fully complex underwater acoustic signals according to an embodiment of the present invention is shown.

[0032] like Figure 2A and Figure 2B As shown, the target neural network in this embodiment may include a complex-valued encoder block group, a complex-valued attention gating module, a complex-valued multi-dimensional attention module, and a complex-valued decoder block group. Furthermore, the fully complex-valued underwater acoustic signal enhancement method in this embodiment includes operations S210 to S250.

[0033] In operation S210, the underwater radiation signal received by the passive sonar system is converted into an initial complex time-varying spectrogram. This initial complex time-varying spectrogram contains noise.

[0034] In operation S220, the initial complex-time spectrogram is downsampled using a complex encoder block group, outputting N complex-downsampled features at N scales. N is an integer greater than 1.

[0035] In operation S230, the complex-valued attention gating module is used to process N complex-valued downsampled features based on the complex-valued convolutional block attention mechanism, and outputs N first complex-valued attention features.

[0036] In operation S240, the complex-valued multi-dimensional attention module is used to process the Nth complex-valued downsampled feature based on the complex-valued attention mechanism of multiple dimensions, and outputs the second complex-valued attention feature.

[0037] In operation S250, using the complex-valued decoder block group, N first complex-valued attention features are sequentially concatenated with second complex-valued attention features, and after each concatenation, the concatenated complex-valued attention features are upsampled, thereby obtaining the target time-spectrum map with noise removal based on the complex-valued upsampled features output by the complex-valued decoder group.

[0038] In embodiments of the present invention, the underwater radiated signal can be an audio signal received underwater by a passive sonar system. Based on this, considering that the frequency and energy of the line spectrum in ship radiated noise contain rich ship-related information and have distinguishable characteristics significantly different from environmental noise, the audio signal can be converted into an initial complex-valued time-frequency spectrum using a Short-Time Fourier Transform (STFT). Specifically, the non-stationary signal can be converted into multiple locally stationary short signals through "windowing and segmentation," and then a Fourier transform is performed on each short signal to obtain a two-dimensional "time-frequency" distribution of the signal, enabling simultaneous analysis of the signal's time-frequency characteristics. Based on this, the initial complex-valued time-frequency spectrum can characterize the relationship between the time and frequency domains of the audio signal. However, the present invention is not limited to this; the initial complex-valued time-frequency spectrum obtained by STFT can also carry the real and imaginary parts of the complex value. For example, the phase and amplitude of the initial complex-valued time-frequency spectrum can be calculated based on the real and imaginary parts of the complex value.

[0039] Furthermore, in this embodiment of the invention, the target neural network can be used to jointly model the amplitude and phase of the initial complex-valued spectrogram. Since the target neural network is obtained by jointly modeling the amplitude and phase of the complex-valued spectrogram, it reconstructs the initial complex-valued spectrogram based on the amplitude and phase of the initial complex-valued spectrogram, thereby removing the noise carried by the initial complex-valued spectrogram.

[0040] Specifically, the initial complex-time spectrogram can be input into a group of complex-value encoder blocks. This group of complex-value encoder blocks can include N complex-value encoder blocks. Each complex-value encoder block can be used to downsample the input features in the complex domain through complex-value convolution. Based on this, the N complex-value encoder blocks can output N complex-value downsampled features at different scales based on the initial complex-time spectrogram.

[0041] Furthermore, N complex-valued downsampled features can be input into the complex-valued attention gating module. The complex-valued convolutional block attention mechanism of this complex-valued attention gating module helps to capture the contextual dependencies between features in the complex domain, thereby better extracting audio features other than noise features from the complex-valued downsampled features, and outputting them as the first complex-valued attention feature.

[0042] Furthermore, the Nth complex-valued downsampled feature (e.g., the output of the Nth complex-valued encoder block out of N complex-valued encoder blocks) can be input into the complex-valued multi-dimensional attention module, so that the complex-valued multi-dimensional attention module processes the Nth complex-valued downsampled feature based on a multi-dimensional complex-valued attention mechanism. In this way, audio features other than noise features can be extracted from the Nth complex-valued downsampled feature from multiple dimensions and output as the second complex-valued attention feature.

[0043] Furthermore, N first complex-valued attention features and second complex-valued attention features can be input into a complex-valued decoder block group. This complex-valued decoder block group can include N complex-valued decoder blocks. Each complex-valued decoder block can be used to upsample the input features in the complex domain through complex-valued deconvolution to reconstruct the features. Based on this, since the N first complex-valued attention features correspond to different scales, the N first complex-valued attention features can be input into the corresponding complex-valued decoder blocks in reverse order of the complex-valued convolution performed by the N complex-valued encoder blocks. This allows the N complex-valued decoder blocks to sequentially concatenate the N first complex-valued attention features with the second complex-valued attention features, and upsample the concatenated complex-valued attention features after each concatenation to reconstruct the corresponding features at different scales and output the corresponding complex-valued upsampled features. Furthermore, the complex-valued upsampled features output by the Nth complex-valued decoder group can be used as the target time-frequency spectrogram.

[0044] Based on this, embodiments of the present invention convert the underwater radiation signal received by the passive sonar system into an initial complex-valued time-of-flight spectrum. Then, a complex-valued encoder block group fully utilizes the complex-valued information of the initial complex-valued time-of-flight spectrum to downsample the spectrum, obtaining complex-valued downsampled features at different scales. Furthermore, a complex-valued attention gating module can process the corresponding complex-valued downsampled features at different scales using a complex-valued convolutional block attention mechanism. This enhances audio features other than noise by combining corresponding complex-valued information at different scales, thereby outputting the corresponding first complex-valued attention feature.

[0045] Then, the complex-valued multi-dimensional attention module processes the Nth complex-valued downsampled feature after the Nth downsampling based on a multi-dimensional complex-valued attention mechanism, further fully utilizing the complex-valued information of the Nth complex-valued downsampled feature to enhance audio features other than noise, and then outputs the corresponding second complex-valued attention feature.

[0046] Thus, by sequentially concatenating N first and second complex-valued attention features using a complex-valued decoder block group, and upsampling the concatenated complex-valued attention features after each concatenation, feature reconstruction can be performed sequentially at different scales to obtain the final target time-spectrum map.

[0047] Therefore, the entire process of this invention is modeled in the complex-valued domain. Furthermore, in this process, the invention combines a fully complex-valued target neural network and a multi-dimensional attention mechanism to fully utilize the amplitude and phase information of the spectrogram at the initial complex-valued stage. This maintains the integrity of the spectrogram structure and avoids the loss of amplitude and phase-related features during the process, thereby improving the enhancement performance of underwater radiation signals in audio form and significantly improving the accuracy of spectrogram reconstruction and waveform reconstruction. This enhances the recovery capability of weaker signals (excluding noise) in underwater radiation signals. Thus, it at least partially avoids the signal distortion problem caused by the removal of these weaker signals during the denoising process.

[0048] Furthermore, considering the line spectrum characteristics of the underwater radiation signals received by the passive sonar system from the ship, the complex multidimensional attention module designed in this invention can process the deep features extracted by the encoder (i.e., the aforementioned complex-valued downsampling features), thereby further modeling multidimensional complex-valued attention features, capturing potential time-frequency dependencies, and highlighting key structural features. The following combines... Figure 3 Please provide an explanation.

[0049] Figure 3 A schematic diagram of a target neural network according to another embodiment of the present invention is shown.

[0050] like Figure 3 As shown, in this embodiment, the complex-valued multi-dimensional attention module includes a complex-valued time-frequency module and a first complex-valued convolutional block attention module. Based on this, the complex-valued multi-dimensional attention module processes the Nth complex-valued downsampled feature using a multi-dimensional complex-valued attention mechanism to output a second complex-valued attention feature. This can include: processing the Nth complex-valued downsampled feature sequentially using the complex-valued time-frequency module based on both a complex-valued time attention mechanism and a complex-valued frequency attention mechanism to obtain the complex-valued time-frequency attention feature; and processing the complex-valued time-frequency attention feature using the first complex-valued convolutional block attention module based on the complex-valued convolutional block attention mechanism to obtain the second complex-valued attention feature.

[0051] Specifically, the complex-valued time-frequency module can be based on a Conformer deep neural network. The Conformer deep neural network can be a neural network architecture combining convolutional neural networks and Transformer neural networks. However, it should be understood that the present invention is not limited to this. In this embodiment, the complex-valued time-frequency module can extend the self-attention mechanism and convolutional structure of the Conformer deep neural network to the complex-valued domain, thereby comprehensively mining long-term dependencies and local detail features in the complex-valued domain. Furthermore, the complex-valued time-frequency module can perform complex-valued self-attention modeling in both frequency and time dimensions, thereby capturing the temporal stability information and frequency distribution characteristics of ship radiation signals, and then outputting corresponding features.

[0052] The first complex-valued convolutional block attention module can be a Convolutional Channel and Spatial Attention Module (CCBAM). This module can include complex pooling layers and complex convolutional or complex linear layers. This allows for the implementation of complex-valued channel and spatial attention with fewer parameters. Furthermore, by modeling the importance of each dimension through channel and spatial attention, the first complex-valued convolutional block attention module can selectively enhance regions related to the linespectral characteristics of audio features other than noise, thereby suppressing redundant information and enhancing key features.

[0053] Furthermore, after inputting the initial complex-valued spectrogram into the complex-valued encoder block group, the complex-valued encoder block group can perform downsampling through multiple layers of complex convolution to extract the complex-valued downsampled features of the input spectrogram. Subsequently, the complex-valued multi-dimensional attention module can utilize a self-attention mechanism based on the time and frequency domains, along with a convolutional block attention mechanism based on the channel and feature dimensions. This not only enhances the network's ability to capture narrowband signal features in weak underwater signals but also promotes the effective separation of the target signal (i.e., the audio features other than noise) from the interference signal (i.e., the noise) while maintaining phase information consistency. Then, the complex-valued decoder block group can, based on the downsampling levels of the complex-valued encoder block group, correspondingly recover the audio features other than noise in the spectrogram through multiple layers of complex deconvolution. Furthermore, cross-layer feature fusion can be achieved through complex-valued attention-gated jump connections. In this way, the denoising process of the initial complex-valued spectrogram can be completed.

[0054] Furthermore, the complex-valued time-frequency module includes a complex-valued temporal attention module, a complex-valued frequency attention module, and a complex-valued convolution module. The following combines... Figure 4 Please provide a detailed explanation.

[0055] Figure 4A schematic diagram of a complex-valued time-frequency module according to an embodiment of the present invention is shown.

[0056] like Figure 4 As shown, the complex-valued time-frequency module may include, in sequence, a complex-valued feedforward module, a complex-valued temporal attention module, a complex-valued frequency attention module, a complex-valued convolution module, a complex-valued feedforward module, a complex-valued layer normalization module, and a fusion layer between these modules. It should be noted that... Figure 4 The “×1 / 2” in the text can be interpreted as “the output features need to be multiplied by the scaling factor of the residual connection”, which will not be elaborated here.

[0057] Building upon this, a complex-valued feedforward module can be used to perform nonlinear feature processing and information integration on the Nth complex-valued downsampled feature. The processed feature is then input into a complex-valued temporal attention module for further processing, yielding a complex-valued temporal attention feature. Subsequently, the Nth complex-valued downsampled feature is fused with the complex-valued temporal attention feature to obtain a fused temporal attention feature. Next, a complex-valued frequency attention module processes the fused temporal attention feature to obtain an intermediate complex-valued time-frequency attention feature. Then, the fused temporal attention feature is fused with the intermediate complex-valued time-frequency attention feature to obtain a fused time-frequency attention feature. Next, a complex-valued convolution module processes the fused time-frequency attention feature to obtain a convolutional time-frequency attention feature. Finally, the convolutional time-frequency attention feature is fused with the fused time-frequency attention feature to obtain a fused convolutional time-frequency attention feature. Finally, based on the fused convolutional time-frequency attention feature, a complex-valued time-frequency attention feature is obtained. For example, the complex-valued feedforward module can be used to process and integrate the fused convolutional time-frequency attention features. Then, the processed convolutional time-frequency attention features are input into the complex-valued layer normalization module to perform layer normalization operations in the complex-valued domain, thereby outputting the complex-valued time-frequency attention features.

[0058] Specifically, taking the complex-valued convolution module as an example, the complex-valued convolution module is used to process the fused time-frequency attention features to obtain the convolutional time-frequency attention features. This includes using the complex-valued convolution module to: calculate the real part product between the real part of the convolution kernel of the complex-valued convolution module and the real part of the fused time-frequency attention features; calculate the imaginary part product between the imaginary part of the convolution kernel of the complex-valued convolution module and the imaginary part of the fused time-frequency attention features; determine the real part of the complex-valued time-frequency attention features based on the difference between the real part product and the imaginary part product; and determine the imaginary part of the complex-valued time-frequency attention features based on the sum of the real part product and the imaginary part product.

[0059] For example, the complex convolution implemented based on this complex-valued convolution module can perform convolution operations using strictly defined complex multiplication-addition rules, enabling the model to establish an intrinsic functional mapping relationship between the real and imaginary parts. The specific calculation method of this complex-valued convolution module is as follows:

[0060] (1)

[0061] Among them, A=A r +jA i The features representing the input (e.g., the fused time-frequency attention features mentioned above), A r Let jA represent the real part of this feature. i This represents the imaginary part of the feature. Φ = Φ r +jΦ i Φ represents the kernel of complex convolution. r Let jΦ represent the real part of the convolution kernel. i This represents the imaginary part of the convolution kernel. Z=Z r +jZ i Z represents the output features (such as the time-frequency attention features of the convolution mentioned above). r Let jZ represent the real part of this feature. i This represents the imaginary part of the feature. j represents the imaginary unit.

[0062] For example, the aforementioned complex-valued feedforward module and complex-valued layer normalization module can calculate the real and imaginary parts of the corresponding input features respectively, so as to perform corresponding processing on the input features in the complex-valued domain, which will not be elaborated here.

[0063] For example, the complex-valued self-attention mechanism splits the query Q, key K, and value Q, which are complex values, into real and imaginary parts, specifically calculated using the following formula:

[0064] (2)

[0065] Where CompAtt(Q,K,V) represents the output features. Att(·) represents the original self-attention mechanism in the real domain. j represents the imaginary unit. Q r Q represents the real part query. i Indicates a query for the imaginary part. K r K represents the real part of the bond. i V represents the imaginary part key. r V represents the real part. i This represents the imaginary part.

[0066] Specifically, taking the complex-valued temporal attention module as an example, the Nth complex-valued downsampled feature is processed using the complex-valued temporal attention module to obtain the complex-valued temporal attention feature. This includes using the complex-valued temporal attention module to: determine the real part query, real part key, and real part value based on the real part of the Nth complex-valued downsampled feature; determine the imaginary part query, imaginary part key, and imaginary part value based on the imaginary part of the Nth complex-valued downsampled feature; process the real part query, real part key, and real part value based on the self-attention mechanism to obtain the first attention feature; process the imaginary part query, imaginary part key, and real part value based on the self-attention mechanism to obtain the second attention feature; process the imaginary part query, real part key, and imaginary part value based on the self-attention mechanism to obtain the third attention feature; and process the real part query, imaginary part key, and real part value based on the self-attention mechanism to obtain the third attention feature; and process the real part query, imaginary part key, and real part value based on the self-attention mechanism. The key and imaginary part value are used to obtain the fourth attention feature; the real part query, real part key, and imaginary part value are processed based on the self-attention mechanism to obtain the fifth attention feature; the real part query, imaginary part key, and real part value are processed based on the self-attention mechanism to obtain the sixth attention feature; the imaginary part query, real part key, and real part value are processed based on the self-attention mechanism to obtain the seventh attention feature; the imaginary part query, imaginary part key, and imaginary part value are processed based on the self-attention mechanism to obtain the eighth attention feature; the real part of the complex-valued time attention feature is obtained based on the first, second, third, and fourth attention features; the imaginary part of the complex-valued time attention feature is obtained based on the fifth, sixth, seventh, and eighth attention features. Thus, the complex-valued time attention feature can be obtained based on the real and imaginary parts of the complex-valued time attention feature. However, it should be understood that the embodiments of the present invention are not limited to this; the complex-valued frequency attention module can also be calculated in this way to obtain the corresponding complex-valued frequency attention module, which will not be elaborated here.

[0067] Based on this, the complex-valued time-frequency module of the present invention implements the feedforward module, attention module, and convolution module in the complex-valued domain, and extends the attention module to include both temporal and frequency attention, forming a dual-channel time-frequency attention. In this way, it captures the temporal stability information and frequency distribution characteristics of the underwater radiation signal from the ship, and then outputs corresponding features. This allows the first complex-valued convolution block attention module to more selectively enhance regions related to the line spectrum characteristics of audio features other than noise, thereby suppressing redundant information and enhancing key features.

[0068] Further, according to embodiments of the present invention, the complex-valued convolutional block attention module of the present invention (e.g., the aforementioned first complex-valued convolutional block attention module, etc.) may include corresponding: a complex-valued spatial average pooling layer, a complex-valued spatial max pooling layer, a complex-valued shared parameter multilayer perceptron, a first complex-valued Sigmoid (S-type) layer, a complex-valued channel average pooling layer, a complex-valued channel max pooling layer, a complex-valued two-dimensional convolutional layer, and a second complex-valued Sigmoid layer. The following is combined with... Figure 5 The complex-valued convolutional block attention module of this invention will be described in an embodiment.

[0069] Figure 5 A schematic diagram of a complex-valued convolutional block attention module according to an embodiment of the present invention is shown.

[0070] like Figure 5 As shown, the complex-valued convolutional block attention module of this embodiment may include a complex-valued spatial average pooling layer, a complex-valued spatial max pooling layer, a complex-valued shared parameter multilayer perceptron, a first complex-valued sigmoid layer, a complex-valued channel average pooling layer, a complex-valued channel max pooling layer, a splicing layer, a complex-valued two-dimensional convolutional layer, and a second complex-valued sigmoid layer.

[0071] Furthermore, the nth complex-valued downsampled feature U can be processed using a complex-valued space average pooling layer to obtain a complex-valued space average pooling feature. Similarly, the nth complex-valued downsampled feature U can be processed using a complex-valued space max pooling layer to obtain a complex-valued space max pooling feature. Subsequently, the complex-valued space average pooling feature can be processed using a complex-valued shared-parameter multilayer perceptron to obtain the first complex-valued channel attention weight, and the complex-valued space max pooling feature can be processed using the complex-valued shared-parameter multilayer perceptron to obtain the second complex-valued channel attention weight. Then, the first and second complex-valued channel attention weights are fused to obtain the fused complex-valued channel attention weight. Finally, the fused complex-valued channel attention weight is normalized using a first complex-valued sigmoid layer to obtain the normalized complex-valued attention weight G. C The normalized complex-valued attention weights G C The intermediate complex-valued fused feature U' is obtained by fusing the nth complex-valued downsampled feature with the intermediate complex-valued fused feature. Then, the intermediate complex-valued fused feature U' is processed using a complex-valued channel average pooling layer to obtain a complex-valued channel average pooling feature. Furthermore, the intermediate complex-valued fused feature U' can be processed using a complex-valued channel max pooling layer to obtain a complex-valued channel max pooling feature. Next, the complex-valued channel average pooling feature and the complex-valued channel max pooling feature are concatenated using a concatenation layer to obtain a concatenated complex-valued channel feature. Then, the concatenated complex-valued channel feature is processed using a complex-valued 2D convolutional layer to obtain a complex-valued 2D convolutional feature. Finally, the complex-valued 2D convolutional feature is normalized using a second complex-valued sigmoid layer to obtain a normalized complex-valued convolutional attention weight G. S Subsequently, the intermediate complex-valued fusion features U' and the normalized complex-valued convolutional attention weights G can be added. SThe fusion process yields the nth first complex-valued attention feature U". Thus, the complex-valued convolutional block attention module of this invention can extract channel attention and spatial attention sequentially through mechanisms such as complex-valued pooling, complex-shared parameter multilayer perceptron, and complex-valued two-dimensional convolution, thereby achieving attention feature extraction in the complex-valued domain. It should be noted that the various layer structures implemented in the complex-valued domain in the embodiments of this invention can achieve their respective functions by calculating the real and imaginary parts of the corresponding input features separately; these will not be elaborated upon here.

[0072] Furthermore, the following combination Figure 6 The complex value encoder group, complex value attention gating module and complex value decoder group of the present invention will be described.

[0073] Figure 6 A schematic diagram of a target neural network according to another embodiment of the present invention is shown.

[0074] like Figure 6 As shown, in this embodiment, the complex encoder block group may include N complex encoder blocks connected in sequence. The complex attention gating module includes N second complex convolutional block attention modules. The complex decoder group includes N complex decoder blocks connected in sequence. It should be noted that, for illustrative purposes, Figure 6 Only three complex value encoder blocks, the second complex value convolution block attention module, and the complex value decoder block are shown in the figure. It should be understood that the present invention does not limit these components.

[0075] Furthermore, for the nth complex-valued encoding block, the (n-1)th complex-valued downsampled feature can be processed using the nth complex-valued encoder block to output the nth complex-valued downsampled feature, where n is a positive integer less than or equal to N. Correspondingly, the nth complex-valued downsampled feature can be processed using the attention module of the nth second complex-valued convolutional block to obtain the nth first complex-valued attention feature.

[0076] Next, the (N-n+1)th complex-valued decoder block can be used to concatenate the nth complex-valued downsampled feature and the (n-1)th complex-valued upsampled feature to obtain the nth complex-valued concatenated feature. Then, the nth complex-valued concatenated feature can be upsampled to obtain the nth intermediate complex-valued upsampled feature. Following this, the nth intermediate complex-valued upsampled feature is processed using a complex-valued convolutional block attention mechanism to obtain the nth complex-valued upsampled feature. Based on this, the N complex-valued encoder blocks, N second complex-valued convolutional attention modules, and the complex-valued decoder block can be calculated in such a way that the Nth complex-valued decoder block outputs the Nth complex-valued upsampled feature, which can then be used as the target time-frequency spectrogram.

[0077] Based on this, the present invention designs a deep neural network (i.e., the target neural network) with a fully complex-valued encoder-decoder (i.e., the aforementioned complex-valued encoder block and complex-valued decoder block) as the backbone and a complex-valued multi-dimensional attention module as the bottleneck layer. The complex-valued encoder block group may include, for example, cascaded complex-valued encoder blocks. For example, the number of encoder blocks may be six. Each complex-valued encoder block can achieve downsampling through multiple layers of complex convolutions to perform multi-scale feature extraction, thereby improving the target neural network's ability to fully extract features of the target signal from the weak underwater radiation signal from the initial complex-valued spectrogram to higher-level features. The complex-valued decoder block group is generally symmetrical to the encoder structure, and each complex-valued decoder block can reconstruct the spectrogram through layer-by-layer upsampling via complex transposed convolutions. Simultaneously, between the complex-valued encoder blocks and complex-valued decoder blocks, each pair of complex-valued encoder blocks and complex-valued decoder blocks is skipped based on a second complex-valued convolutional attention module, so that high- and low-level features can be effectively fused. Furthermore, by introducing a complex-valued attention gating mechanism in the jump connection, background interference features can be suppressed and the narrow-band spectral features of weaker target signals can be highlighted, thereby avoiding the introduction of redundant or erroneous background information into the decoding process under complex non-stationary noise interference.

[0078] Furthermore, in this embodiment of the invention, the complex-valued encoder block can have structures such as complex numbers, complex-valued batch normalization, and complex-valued activation functions, and the complex-valued decoder block is similar. However, in addition to the symmetrical structure with the complex-valued encoder block, the complex-valued decoder block may also include a corresponding complex-valued convolutional block attention module to perform attention weighting on the input features. The specific structures of the complex-valued encoder block and the complex-valued decoder block in this embodiment of the invention are described below.

[0079] Figure 7A A schematic diagram of a complex value encoder block according to an embodiment of the present invention is shown.

[0080] like Figure 7A As shown, the complex encoder block includes a complex two-dimensional convolutional layer, a first complex batch normalization layer, and a first complex LReLU (Leaky Rectified Linear Unit) layer connected in sequence.

[0081] Building upon this, taking the calculation process of the nth complex-valued downsampled feature as an example, in this embodiment of the invention, the (n-1)th complex-valued downsampled feature can be processed using a first complex-valued two-dimensional convolutional layer to obtain intermediate downsampled features. Subsequently, the intermediate downsampled features can be processed using a first complex-valued batch normalization layer to obtain batch-normalized downsampled features. Afterward, the batch-normalized downsampled features can be normalized using a first complex-valued LReLU layer to obtain the nth complex-valued downsampled feature. Thus, the complex-valued encoder block can perform multi-scale feature extraction on the input features, thereby improving the target neural network's ability to fully extract features of the target signal from the weak underwater radiation signal from the initial complex-valued spectrogram to higher-level features.

[0082] Figure 7B A schematic diagram of a complex value decoder block according to an embodiment of the present invention is shown.

[0083] like Figure 7B As shown, the complex-valued decoder block includes a complex-valued two-dimensional deconvolution layer, a second complex-valued batch normalization layer, a second complex-valued LReLU layer, and a third complex-valued convolutional block attention module.

[0084] Based on this, the nth complex-valued concatenated feature can be processed using a complex-valued two-dimensional deconvolution layer to obtain an intermediate upsampled feature. Subsequently, the intermediate upsampled feature can be processed using a second complex-valued batch normalization layer to obtain a batch-normalized upsampled feature. Then, the batch-normalized upsampled feature is normalized using a second complex-valued LReLU layer to obtain the nth complex-valued upsampled feature. Finally, the nth intermediate complex-valued upsampled feature can be processed using a third complex-valued convolutional block attention module to obtain the nth complex-valued upsampled feature. Therefore, compared to the complex-valued encoder block, the complex-valued decoder block introduces this third complex-valued convolutional block attention module after the corresponding second complex-valued LReLU layer to further focus on the components of the target signal during the decoding stage. It should be noted that the specific structures of the second and third complex-valued convolutional block attention modules are similar to those described above for the first complex-valued convolutional block attention module, and will not be elaborated upon here.

[0085] Furthermore, the first and second complex batch normalization layers can be used to center the mean of the real and imaginary parts respectively, and to whiten the features based on the covariance matrix constructed from complex values, in order to preserve the correlation structure of the complex-valued representation. Complex-valued activation functions (such as the first and second complex-valued LReLU layers mentioned above) can be computed on the real and imaginary parts respectively to output the corresponding features.

[0086] According to an embodiment of the present invention, the target neural network is trained as follows: the spectrum of sample complex values ​​is input into the initial neural network to obtain the spectrum of predicted complex values; based on the target loss function, the normalized mean square error between the spectrum of predicted complex values ​​and the spectrum of labeled complex values ​​corresponding to the spectrum of sample complex values ​​is calculated; the parameters of the initial neural network are adjusted based on the normalized mean square error to obtain the target neural network.

[0087] Specifically, regarding network training constraints, considering the significant frequency domain characteristics and strong structure of underwater ship radiation signals, and the fact that downstream tasks are typically performed in the time-frequency domain, this invention proposes a normalized mean square error (nMSE) loss function for complex spectrograms as the target loss function for training the target neural network. This directly constrains the prediction error of the complex spectrograms. Compared to time-domain loss functions, nMSE allows the model to focus on low-energy but information-rich spectral components, thereby improving the reconstruction accuracy of weak signals and reducing over-attenuation, better preserving weak spectral line information in low signal-to-noise ratio environments. Spectral energy is obtained by summing the squares of the amplitudes over all time-frequency points. Since the energy differences between different target signals are significant, the nMSE loss normalizes the mean square error using the spectral energy of the clean signal, mitigating the impact of signal energy differences on model training stability.

[0088] Specifically, this invention uses the normalized mean square error as the target loss function to constrain the prediction accuracy of the target neural network on the complex spectrum. It can implicitly impose constraints on amplitude and phase errors, and is particularly suitable for weak signal enhancement in complex scenarios. The target loss function is as follows:

[0089] (3)

[0090] in, The spectrum represents the complex values ​​of the labels. This represents the spectrum when predicting complex values. This represents the Frobenius norm. Additionally, to mitigate the impact of signal energy differences on model training stability, a clean prediction complex-valued spectrogram can be used. The energy is normalized. The spectral energy is obtained by summing the squares of the amplitudes of all time-frequency points in the complex-valued spectrum.

[0091] In underwater acoustic signal enhancement tasks, some solutions struggle to simultaneously achieve weak target spectral line recovery and background interference suppression, and the enhancement effect needs further improvement. Therefore, this invention designs a fully complex-valued underwater acoustic signal enhancement method for passive sonar systems. This method uses a fully complex-valued neural network to jointly model amplitude and phase information to highlight target signal characteristics and reduce excessive attenuation.

[0092] Based on this, the present invention, by utilizing fully complex value modeling and a multidimensional attention mechanism, significantly improves signal quality in underwater acoustic signal enhancement tasks for passive sonar systems. This fully complex value modeling can completely combine modeling amplitude and phase information, enhancing the narrowband feature recovery of weak target signals, while effectively suppressing background interference, reducing excessive attenuation and false deletion, and promoting a more efficient role for the enhanced signal in downstream tasks, thus providing a better solution for underwater acoustic signal enhancement.

[0093] The following describes the effectiveness of the above-mentioned fully complex underwater acoustic signal enhancement method for passive sonar systems based on specific experimental data.

[0094] According to an embodiment of the present invention, this experiment uses a Hanning window with a window length of 512 points and a frame shift of 256 points to obtain the STFT spectrum for audio with a sampling rate of 16 kHz (kilohertz). The AdamW optimizer is used during training, with an initial learning rate set to 0.001 and an exponential decay learning rate adjustment strategy with a decay rate set to 0.99. The training cycle is 200 rounds. Simultaneously, to improve training stability, the L2 norm of the gradient is pruned to 5.0.

[0095] This experiment utilized the invented underwater ship noise database for a comprehensive evaluation. Recordings for this database were completed from autumn 2012 to summer 2013 in multiple port areas along the ocean coast, primarily in harbors and surrounding drowned valley waters, areas with heavy shipping traffic and diverse vessel types. The original database contained various ship categories and background noise, which were then categorized into five classes: Class A includes fishing boats, trawlers, aquaculture vessels, tugboats, and dredgers; Class B includes motorboats, pilot boats, and sailboats; Class C includes passenger ferries; Class D includes ocean-going liner ships and roll-on / roll-off ships; and Class E represents background noise. The recorded signals from actual ship radiation in the underwater ship noise database have a high signal-to-noise ratio and can be considered clean target signals.

[0096] This experiment selected 43 underwater radiation signal recordings from ships (including passenger ships and motorboats) and 3 background noise recordings (including wind noise, rain noise, and current noise). These recordings were randomly segmented into 3-second clean signal and noise segments, and then randomly mixed with each other within a signal-to-noise ratio range of -15 dB to -5 dB to generate noisy frequencies. This resulted in 75,000 3-second synthetic noisy frequencies, all uniformly resampled to 16 kHz. The dataset was divided into training, validation, and test sets at a ratio of 70% / 20% / 10% for model training and evaluation, ensuring fairness in the experiment and reliable verification of model performance.

[0097] In performance evaluation, multiple metrics were used to measure the overall quality of the enhanced signal, including: normalized mean square error (nMSE) in the time-frequency domain, signal distortion ratio (SDR) in the time domain, signal distortion ratio gain (SDRi), scale-invariant signal-to-noise ratio (SI-SNR), scale-invariant signal-to-noise ratio gain (SI-SNRi), and segmented signal-to-noise ratio (SegSNR). Except for nMSE, which is preferred to be as low as possible, all other metrics were preferred to be as high as possible. Furthermore, to further verify the positive impact of front-end enhancement on downstream tasks, this experiment designed a ship type recognition experiment as a representative back-end task. In the experiment, a ResNet classifier was used with the enhanced spectrogram as input to predict ship types, and classification accuracy (Acc) was used as the primary metric to reflect recognition performance.

[0098] Based on this, Table 1 below shows the performance comparison of the present invention with several baseline models on the ShipsEar dataset. The baseline models include three traditional underwater acoustic enhancement baseline models (e.g., EMD-TF, EMD-T, Wavelet), one deep learning-based speech enhancement baseline model (e.g., DCCRN), and two deep learning-based underwater acoustic enhancement baseline models (e.g., DBSA-Net, NAFSA-Net). The comparison items include model parameters (i.e., Para.), underwater acoustic quality assessment metrics, and accuracy of downstream classification tasks.

[0099] In some schemes, only wavelet thresholding decomposition shows slight improvement in SI-SNR and Acc compared to noisy frequencies. Clearly, traditional methods perform poorly overall and struggle to recover the target signal. Due to extremely low signal-to-noise ratios and complex noise characteristics, these methods fail to decompose noisy signals to obtain the dominant component of the target signal; thresholding often retains residual noise and incorrectly filters out the target signal. Depth-based methods significantly outperform traditional denoising methods across all metrics. Although DBSA-Net achieves good results on the SegSNR metric with fewer parameters, it performs poorly on SDR, SI-SNR, and Acc metrics. NAFSA-Net achieves the highest accuracy with the fewest parameters, but inevitably falls short on enhancement metrics, indicating that the trade-off between enhancement and classification still needs optimization. Compared to the best results of baseline methods, this invention achieves optimal results on SDR, SI-SNR, and SegSNR, while maintaining a high level of classification accuracy, demonstrating its excellent balance between enhancement and recognition. Despite having the largest model size, this method achieves a better trade-off between signal enhancement and target classification, with the increased number of parameters resulting in a significant performance improvement. Furthermore, since this method focuses on improving the signal quality received by passive sonar systems, the increase in model data size is within an acceptable range considering the model's superior performance.

[0100] Table 1. Comparison of the target neural network of this invention with the baseline model on the ShipsEar dataset.

[0101]

[0102] Table 2 below shows the performance of the nMSE loss function used in this invention and the commonly used loss function SI-SNR and its variants on the ShipsEar dataset. This experiment was conducted based on the proposed model, differing only in the loss function while maintaining other configurations. The results show that the SI-SNR loss function performs poorly on amplitude-related metrics such as nMSE, SDRi, and SegSNR. This poor performance can be attributed to the inability of the spectral mapping method to constrain amplitude during the enhancement process, leading to inaccurate reconstruction. To address this issue, we designed two weighted SI-SNR variants (e.g., wSI-SNR) that compensate for amplitude reconstruction by jointly using SDR and nMSE. Experimental results show that using SDR-weighted SI-SNR significantly improves these metrics, with only a slight decrease on SI-SNRi.

[0103] Furthermore, when nMSE-weighted SI-SNR is used, all metrics are further improved, indicating that nMSE is more suitable than SDR for amplitude constraints. Using nMSE as the loss function outperforms other methods in all metrics, achieving the lowest nMSE and the highest SDRi, SI-SNRi, and SegSNR. Notably, compared to directly using SI-SNR, nMSE can bring a higher SI-SNR improvement, mainly because nMSE can impose stronger and more direct constraints on spectral details, especially under low signal-to-noise ratio conditions, where the spectrum can more clearly express the information of underwater acoustic signals than the waveform. In addition, due to the transformation relationship between the spectrum and the waveform, enhancing the spectrum also naturally improves the similarity of the waveform. Therefore, nMSE not only performs well in terms of performance but also surpasses SI-SNR loss in improving SI-SNR. The above results highlight the significant advantages of the proposed loss function in spectral recovery, amplitude reconstruction, and noise cancellation.

[0104] Table 2 Comparison of the nMSE loss function used in this invention with commonly used loss functions and their variants

[0105]

[0106] Table 3 below shows the performance of the present invention and a scheme without partial attention modules on the ShipsEar dataset to evaluate the role of each attention module. The complex-valued encoder-decoder structure alone significantly improves signal enhancement performance and downstream recognition accuracy, demonstrating the effectiveness of the complex-valued structure and multi-scale feature extraction. The introduction of a multi-dimensional attention mechanism adaptively captures key information in multiple dimensions, including the time domain, frequency domain, channels, and feature maps, significantly enhancing feature representation capabilities. While CCBAM primarily improves signal quality, it may have a certain inhibitory effect on features relevant to the recognition task; the complex-valued time-frequency Conformer module, on the other hand, can effectively maintain and optimize recognition-related features while enhancing signal quality. By combining the two synergistically to form a multi-dimensional attention mechanism, the two achieve complementary optimization under the synergy of the multi-dimensional attention mechanism, enabling the present invention to achieve excellent performance in both enhancement and recognition.

[0107] Table 3 Comparison of the present invention with a scheme that does not include a partial attention module

[0108]

[0109] Figure 8 The diagram illustrates the effect of the target neural network and the baseline model in terms of spectrum and power spectral density according to an embodiment of the present invention.

[0110] like Figure 8 As shown, the present invention (i.e. Figure 8 The enhanced signal obtained in (b) yields the clearest and most reliable spectrum and power spectral density. Compared to the baseline method, this invention not only significantly suppresses noise but also avoids the problem of over-eliminating the target signal in the noise-dominated time-frequency region, especially in the low-frequency band and high-energy pulse signal region, thus preserving the target signal characteristics to the greatest extent. Furthermore, this invention can accurately distinguish and reconstruct the target signal on multiple spectral lines with similar frequencies, accurately recovering the number and amplitude of spectral lines, thereby achieving superior frequency resolution compared to the baseline method.

[0111] In this method, the complex-valued time spectrum is directly used as the network input, and all operations in the network are implemented in the complex domain. A complex-valued multi-dimensional attention module is designed to improve the reconstruction capability of the target signal, and a normalized mean square error is introduced as the loss function for underwater scenarios. Therefore, the distortion introduced by the enhancement process is reduced while improving the enhancement effect.

[0112] Figure 9 A block diagram of an electronic device suitable for implementing a fully complex-valued underwater acoustic signal enhancement method for passive sonar systems, according to an embodiment of the present invention, is shown.

[0113] like Figure 9As shown, an electronic device 900 according to an embodiment of the present invention includes a processor 901, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 902 or a program loaded from a storage portion 908 into a random access memory (RAM) 903. The processor 901 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 901 may also include onboard memory for caching purposes. The processor 901 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present invention.

[0114] RAM 903 stores various programs and data required for the operation of electronic device 900. Processor 901, ROM 902, and RAM 903 are interconnected via bus 904. Processor 901 executes various operations of the method flow according to embodiments of the present invention by executing programs in ROM 902 and / or RAM 903. It should be noted that the programs may also be stored in one or more memories other than ROM 902 and RAM 903. Processor 901 may also execute various operations of the method flow according to embodiments of the present invention by executing programs stored in said one or more memories.

[0115] According to an embodiment of the present invention, the electronic device 900 may further include an input / output (I / O) interface 905, which is also connected to a bus 904. The electronic device 900 may also include one or more of the following components connected to the input / output (I / O) interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to the input / output (I / O) interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 910 as needed so that computer programs read from it can be installed into the storage section 908 as needed.

[0116] The present invention also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of the present invention.

[0117] According to embodiments of the present invention, a computer-readable storage medium may be a non-volatile computer-readable storage medium, such as including, but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of the present invention, a computer-readable storage medium may include ROM 902 and / or RAM 903 and / or one or more memories other than ROM 902 and RAM 903 described above.

[0118] Embodiments of the present invention also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the fully complex-valued underwater acoustic signal enhancement method for passive sonar systems provided in the embodiments of the present invention.

[0119] When the computer program is executed by the processor 901, it performs the functions defined in the system / apparatus of this invention. According to embodiments of the invention, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0120] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 909, and / or installed from a removable medium 911. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0121] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 909, and / or installed from the removable medium 911. When the computer program is executed by the processor 901, it performs the functions defined in the system of this embodiment of the invention. According to embodiments of the invention, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0122] According to embodiments of the present invention, program code for executing the computer programs provided in the embodiments of the present invention can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0123] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0124] Those skilled in the art will understand that the features described in the various embodiments of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0125] The embodiments of the present invention have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of the invention. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of the invention, and all such substitutions and modifications should fall within the scope of the invention.

Claims

1. A method for enhancing fully complex underwater acoustic signals for passive sonar systems, characterized in that, include: Convert the underwater radiation signal received by the passive sonar system into an initial complex value spectrum; The initial complex value spectrum is noisy; The initial complex time-frequency spectrum is downsampled using a complex encoder block group, and N complex downsampled features at N scales are output; N is an integer greater than 1. The complex-valued attention gating module is used to process the N complex-valued downsampled features based on the complex-valued convolutional block attention mechanism, and outputs N first complex-valued attention features; The Nth complex-valued downsampled feature is processed using a complex-valued multi-dimensional attention module based on a multi-dimensional complex-valued attention mechanism, and the second complex-valued attention feature is output. Using a complex-valued decoder block group, the N first complex-valued attention features are sequentially concatenated with the second complex-valued attention features, and the concatenated complex-valued attention features are upsampled after each concatenation, thereby obtaining the target time-spectrum map with the noise removed based on the complex-valued upsampled features output by the complex-valued decoder group.

2. The method for enhancing fully complex underwater acoustic signals according to claim 1, characterized in that, The complex-valued multi-dimensional attention module includes a complex-valued time-frequency module and a first complex-valued convolutional block attention module; The Nth complex-valued downsampled feature is processed using a complex-valued multi-dimensional attention module based on a multi-dimensional complex-valued attention mechanism, outputting a second complex-valued attention feature, including: The Nth complex-valued downsampling feature is processed sequentially using the complex-valued time-frequency module based on the complex-valued time attention mechanism and the complex-valued frequency attention mechanism to obtain the complex-valued time-frequency attention feature; The complex-valued time-frequency attention features are processed using the first complex-valued convolutional block attention module based on the complex-valued convolutional block attention mechanism to obtain the second complex-valued attention features.

3. The method for enhancing fully complex underwater acoustic signals according to claim 2, characterized in that, The complex-valued time-frequency module includes a complex-valued time attention module, a complex-valued frequency attention module, and a complex-valued convolution module; The complex-valued time-frequency module processes the Nth complex-valued downsampling feature sequentially based on the complex-valued time attention mechanism and the complex-valued frequency attention mechanism to obtain the complex-valued time-frequency attention feature, including: The Nth complex-valued downsampled feature is processed using the complex-valued temporal attention module to obtain the complex-valued temporal attention feature; The Nth complex-valued downsampling feature is fused with the complex-valued temporal attention feature to obtain the fused temporal attention feature; The complex-valued frequency attention module is used to process the fused time attention features to obtain intermediate complex-valued time-frequency attention features; The fused time-frequency attention feature is fused with the intermediate complex-valued time-frequency attention feature to obtain the fused time-frequency attention feature; The complex-valued convolution module is used to process the fused time-frequency attention features to obtain convolutional time-frequency attention features; The convolutional time-frequency attention features are fused with the fused time-frequency attention features to obtain the fused convolutional time-frequency attention features; The complex-valued time-frequency attention features are obtained based on the fused convolutional time-frequency attention features.

4. The method for enhancing fully complex underwater acoustic signals according to claim 3, characterized in that, The Nth complex-valued downsampled feature is processed using the complex-valued temporal attention module to obtain the complex-valued temporal attention feature, including the use of the complex-valued temporal attention module: Based on the real part of the Nth complex-valued downsampling feature, determine the real part query, real part key, and real part value; Based on the imaginary part of the Nth complex-valued downsampling feature, determine the imaginary part query, imaginary part key, and imaginary part value; The real part query, real part key, and real part value are processed based on a self-attention mechanism to obtain the first attention feature; The second attention feature is obtained by processing the imaginary part query, imaginary part key, and real part value based on the self-attention mechanism. Based on the self-attention mechanism, the imaginary part query, real part key, and imaginary part value are processed to obtain the third attention feature; Based on the self-attention mechanism, the real part query, imaginary part key, and imaginary part value are processed to obtain the fourth attention feature; Based on the self-attention mechanism, the real part query, real part key, and imaginary part value are processed to obtain the fifth attention feature; The real part query, imaginary part key, and real part value are processed based on the self-attention mechanism to obtain the sixth attention feature; Based on the self-attention mechanism, the imaginary part query, real part key, and real part value are processed to obtain the seventh attention feature; Based on the self-attention mechanism, the imaginary part query, imaginary part key, and imaginary part value are processed to obtain the eighth attention feature; Based on the first attention feature, the second attention feature, the third attention feature, and the fourth attention feature, the real part of the complex-valued temporal attention feature is obtained; Based on the fifth attention feature, the sixth attention feature, the seventh attention feature, and the eighth attention feature, the imaginary part of the complex-valued temporal attention feature is obtained.

5. The method for enhancing fully complex underwater acoustic signals according to claim 3, characterized in that, The complex-valued convolution module is used to process the fused time-frequency attention features to obtain convolutional time-frequency attention features, including using the complex-valued convolution module: Calculate the real part product between the real part of the convolution kernel of the complex-valued convolution module and the real part of the fused time-frequency attention feature; Calculate the imaginary part product between the imaginary part of the convolution kernel of the complex-valued convolution module and the imaginary part of the fused time-frequency attention feature; The real part of the complex-valued time-frequency attention feature is determined based on the difference between the real part product and the imaginary part product. The imaginary part of the complex-valued time-frequency attention feature is determined based on the sum of the product of the real parts and the product of the imaginary parts.

6. The method for enhancing fully complex underwater acoustic signals according to any one of claims 1 to 5, characterized in that, The complex encoder block group includes N complex encoder blocks connected in sequence; the complex attention gating module includes N second complex convolutional block attention modules; the complex decoder group includes N complex decoder blocks connected in sequence. The initial complex time-frequency spectrum is downsampled using a complex encoder block group to output N complex downsampled features at N scales, including: processing the (n-1)th complex downsampled feature using the nth complex encoder block to output the nth complex downsampled feature; where n is a positive integer less than or equal to N; The N complex-valued downsampled features are processed by the complex-valued attention gating module based on the complex-valued convolutional block attention mechanism to output N first complex-valued attention features, including: processing the nth complex-valued downsampled feature by the nth second complex-valued convolutional block attention module to obtain the nth first complex-valued attention feature; Using a complex-valued decoder block group, the N first complex-valued attention features are sequentially concatenated with the second complex-valued attention features, and the concatenated complex-valued attention features are upsampled after each concatenation, including using the (N-n+1)th complex-valued decoder block: The nth complex value downsampling feature and the (n-1)th complex value upsampling feature are concatenated to obtain the nth complex value concatenated feature; Upsample the nth complex value concatenation feature to obtain the nth intermediate complex value upsampled feature; The nth intermediate complex-valued upsampled feature is processed based on the complex-valued convolutional block attention mechanism to obtain the nth complex-valued upsampled feature; wherein, the Nth complex-valued upsampled feature is used as the target time-spectrum map.

7. The method for enhancing fully complex underwater acoustic signals according to claim 6, characterized in that, The complex encoder block includes a complex two-dimensional convolutional layer, a first complex batch normalization layer, and a first complex LReLU layer connected in sequence. The nth complex encoder block processes the (n-1)th complex downsampled feature, outputting the nth complex downsampled feature, including: The first complex-valued two-dimensional convolutional layer is used to process the (n-1)th complex-valued downsampled feature to obtain the intermediate downsampled feature; The intermediate downsampled features are processed using the first complex-valued batch normalization layer to obtain batch-normalized downsampled features; The batch-normalized downsampled features are normalized using the first complex-valued LReLU layer to obtain the nth complex-valued downsampled feature.

8. The method for enhancing fully complex underwater acoustic signals according to claim 6, characterized in that, The complex-valued decoder block includes a complex-valued two-dimensional deconvolution layer, a second complex-valued batch normalization layer, a second complex-valued LReLU layer, and a third complex-valued convolutional block attention module; Upsampling the nth complex-valued concatenated feature to obtain the nth intermediate complex-valued upsampled feature includes: processing the nth complex-valued concatenated feature using the complex-valued two-dimensional deconvolution layer to obtain the intermediate upsampled feature; processing the intermediate upsampled feature using the second complex-valued batch normalization layer to obtain the batch-normalized upsampled feature; and normalizing the batch-normalized upsampled feature using the second complex-valued LReLU layer to obtain the nth complex-valued upsampled feature. The process of processing the nth intermediate complex-valued upsampled feature based on the complex-valued convolutional block attention mechanism to obtain the nth complex-valued upsampled feature includes: processing the nth intermediate complex-valued upsampled feature using the third complex-valued convolutional block attention module to obtain the nth complex-valued upsampled feature.

9. The method for enhancing fully complex underwater acoustic signals according to claim 6, characterized in that, The first complex-valued convolutional block attention module, the second complex-valued convolutional block attention module, and the third complex-valued convolutional block attention module each include the following: a complex-valued space average pooling layer, a complex-valued space max pooling layer, a complex-valued shared parameter multilayer perceptron, a first complex-valued sigmoid layer, a complex-valued channel average pooling layer, a complex-valued channel max pooling layer, a complex-valued two-dimensional convolutional layer, and a second complex-valued sigmoid layer. The nth complex-valued downsampled feature is processed using the attention module of the nth second complex-valued convolutional block to obtain the nth first complex-valued attention feature, including: The nth complex-valued downsampled feature is processed using the complex-valued space average pooling layer to obtain the complex-valued space average pooling feature; The nth complex-valued downsampled feature is processed using the complex-valued space max pooling layer to obtain the complex-valued space max pooling feature; The complex-valued shared parameter multilayer perceptron is used to process the complex-valued spatial average pooling features to obtain the first complex-valued channel attention weights; The complex-valued shared parameter multilayer perceptron is used to process the complex-valued space max pooling features to obtain the second complex-valued channel attention weights; The first complex-valued channel attention weight and the second complex-valued channel attention weight are fused to obtain the fused complex-valued channel attention weight; The fused complex-valued channel attention weights are normalized using the first complex-valued Sigmoid layer to obtain normalized complex-valued attention weights. The normalized complex-valued attention weights and the nth complex-valued downsampled feature are fused to obtain the intermediate complex-valued fused feature; The intermediate complex-valued fusion features are processed using the complex-valued channel average pooling layer to obtain the complex-valued channel average pooling features; The intermediate complex-valued fusion features are processed using the complex-valued channel max pooling layer to obtain the complex-valued channel max pooling features; The average pooling feature of the complex channel and the max pooling feature of the complex channel are concatenated to obtain the concatenated complex channel feature. The concatenated complex channel features are processed using the complex-valued two-dimensional convolutional layer to obtain complex-valued two-dimensional convolutional features; The complex two-dimensional convolutional features are normalized using the second complex-valued Sigmoid layer to obtain normalized complex-valued convolutional attention weights. The intermediate complex value fusion feature and the normalized complex value convolutional attention weight are fused to obtain the nth first complex value attention feature.

10. A target neural network for implementing the fully complex-valued underwater acoustic signal enhancement method for passive sonar systems as described in any one of claims 1 to 9, characterized in that, The target neural network includes the complex value encoder block group, the complex value attention gating module, the complex value multi-dimensional attention module, and the complex value decoder block group; The target neural network is trained in the following manner: Input the spectrum of the sample when it is complex valued into the initial neural network to obtain the spectrum of the predicted complex value. Based on the target loss function, the normalized mean square error between the predicted complex value spectrum and the label complex value spectrum corresponding to the sample complex value spectrum is calculated. The parameters of the initial neural network are adjusted based on the normalized mean square error to obtain the target neural network.

Images