Underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution

Abstract

The invention relates to an underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution, and the method comprises the steps: collecting an original sound pressure signal, carrying out the frequency domain feature enhancement, and generating a feature enhancement graph; the feature enhancement image is input into an underwater sound source positioning model, the distance and depth of a sound source are output, and the underwater sound source positioning model is used for carrying out feature extraction and target positioning on the feature enhancement image. Aiming at the complex challenges of strong underwater environment noise interference, obvious multipath propagation effect, time-frequency attenuation and the like, the invention provides a frequency domain feature enhancement technology to improve model positioning precision, and multi-band features of acoustic data are effectively extracted by fusing wavelet convolution and multi-scale convolution so as to accurately capture spatial distribution characteristics of the acoustic data.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater sound source positioning, and in particular to an underwater sound source positioning method based on frequency domain feature enhancement and multi-scale wavelet convolution. Background Art

[0002] Underwater sound source localization has long been a key challenge in the field of underwater acoustics. Matched field processing (MFP) is one of the most widely used traditional methods. MFP determines the spatial location of a sound source by comparing the correlation between the number of replica fields generated by an acoustic model and actual sea trial acoustic data, generating a two-dimensional fuzzy surface representing the underwater sound source's distance and depth. However, when key oceanographic environmental parameters used in theoretical modeling (such as the sound velocity profile, sediment parameters, and bottom layer parameters) deviate from the actual environment, MFP's localization performance degrades significantly. Inspired by matched field processing, Liu et al. proposed a feature matching method based on convolutional neural networks to estimate sound source depth in shallow water environments. Compared with traditional MFP methods, FM-CNN is more robust to environmental mismatches. While these studies have demonstrated the effectiveness of convolutional neural network methods for underwater sound source distance estimation, they primarily focus on the localization model itself and fail to fully account for interference issues such as noise, multipath propagation, and signal attenuation, which are common in real underwater environments. Failure to fully account for these factors can significantly weaken the model's robustness and accuracy in real-world scenarios.

[0003] In recent years, with the advancement of deep learning theory and the increase in computing resources, data-driven deep learning methods have been widely used in underwater sound source localization. These methods do not rely on complex sound field modeling and can automatically extract key features from large amounts of data, even when some environmental parameters are missing, resulting in excellent localization performance. Research on data-driven underwater sound source localization dates back to the 1990s, when shallow neural networks were first used to measure the distance to underwater acoustic targets. However, due to the limitations of computing resources at the time, data-driven methods did not make substantial progress. In recent years, with the rapid development of machine learning theory and the significant increase in computing power, data-driven methods have regained vitality in the field of underwater acoustic localization. To overcome the limitations of matched field processing methods, machine learning methods have been widely used in underwater acoustic localization. He et al. proposed a data model-driven hybrid multi-task learning model that combines a Transformer and a convolutional neural network (CNN) with a multi-level dilated convolutional fusion (MLF) module for underwater passive localization. Although this method proposed a model framework based on multi-task learning, the experimental training and testing were performed only on the same dataset, lacking adaptability to environmental changes. When applied to a real-world environment that differs from the training data, the model's performance may degrade significantly.

[0004] In order to solve the above problems, the present invention proposes an underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution. Summary of the Invention

[0005] The purpose of the present invention is to provide an underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution. In view of the complex challenges such as strong underwater environmental noise interference, significant multipath propagation effect and time-frequency attenuation, by integrating wavelet transform and convolution operations, it can not only effectively extract the multi-band features of acoustic data, but also accurately capture its spatial distribution characteristics.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for underwater sound source localization based on frequency domain feature enhancement and multi-scale wavelet convolution, comprising:

[0008] Collect the original sound pressure signal and perform frequency domain feature enhancement to generate a feature enhancement map;

[0009] The feature enhancement map is input into an underwater sound source localization model, and the distance and depth of the sound source are output, wherein the underwater sound source localization model is used to perform multi-feature extraction and target localization on the feature enhancement map.

[0010] Optionally, generating a feature enhancement map includes:

[0011] Performing Fourier transform processing on the original sound pressure signal to obtain frequency domain data;

[0012] Performing spectrum and energy analysis on the frequency domain data to extract energy features;

[0013] Normalizing the energy features, and calculating a sample covariance matrix based on the normalized energy features;

[0014] The sample covariance matrices are stacked to obtain the feature enhancement map.

[0015] Optionally, performing multi-feature extraction and target positioning on the feature enhancement map includes:

[0016] Inputting the feature enhancement map into a downsampling module and outputting a downsampled feature map;

[0017] Inputting the downsampled feature map into a parameter sharing layer for feature extraction to obtain multi-scale features, wherein the parameter sharing layer includes a wavelet convolution module and an InceptionWT module, and the InceptionWT module is obtained by combining depthwise separable convolution and wavelet transform;

[0018] Inputting the multi-scale features into the distance branch and the depth branch respectively to obtain distance features and depth features, wherein the distance branch and the depth branch are constructed by sequentially connecting module A, module B, a pooling layer and a multi-layer perceptron, and the module A and module B are constructed by parallel branches including convolution and / or maximum pooling;

[0019] The distance and depth of the sound source are output according to the distance feature and the depth feature.

[0020] Optionally, inputting the downsampled feature map into a parameter sharing layer for feature extraction includes:

[0021] Performing wavelet decomposition and reconstruction on the downsampled feature enhancement image through the wavelet convolution module to obtain a first feature;

[0022] The first feature is input into the InceptionWT module for feature extraction to obtain the multi-scale feature.

[0023] Optionally, inputting the first feature into the InceptionWT module for feature extraction includes: inputting the first feature into several depth-separable convolution branches and wavelet convolution branches for feature extraction, and fusing the extracted features to obtain the multi-scale feature.

[0024] Optionally, performing wavelet decomposition and reconstruction on the downsampled feature enhancement map by the wavelet convolution module includes:

[0025] Performing wavelet decomposition on the downsampled feature enhancement image to obtain a plurality of sub-band features;

[0026] After stacking the sub-band features to the channel dimension, depthwise separable convolution is performed to extract features and reconstruct them to obtain wavelet reconstructed data.

[0027] Optionally, obtaining the first feature includes: obtaining the first feature according to data after convolution of the feature enhancement map and the wavelet reconstruction data.

[0028] Optionally, the Loss function in the underwater sound source localization model includes:

[0029]

[0030] Among them, Loss represents the total loss, L r and L d are the losses of distance and depth single tasks respectively, w r and w d Represents the weight of a specific task, σ r and σ d represents the learned log variance.

[0031] The beneficial effects of the present invention are:

[0032] (1) A frequency domain-based feature enhancement method is proposed. By performing frequency domain conversion on the original signal and extracting key information, the core features in the signal are significantly enhanced, providing high-quality feature data for subsequent target positioning.

[0033] (2) We innovatively combine wavelet convolution with multi-scale structure to construct a multi-scale wavelet convolution structure, which not only effectively captures the low-frequency information of the signal, but also can simultaneously extract local and global time-frequency features, thereby significantly enhancing the adaptability of the model to complex marine environments.

[0034] (3) Through ablation experiments and comparative experiments, the effectiveness of the proposed feature extraction method and multi-scale wavelet convolution structure in underwater target positioning was systematically verified, demonstrating its significant advantages in improving positioning accuracy and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1This is a framework diagram of an underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to an embodiment of the present invention;

[0037] Figure 2 This is a feature enhancement flow chart of an embodiment of the present invention;

[0038] Figure 3 This is a diagram of the InceptionWT module framework of an embodiment of the present invention;

[0039] Figure 4 This is a framework diagram of module A and module B according to an embodiment of the present invention;

[0040] Figure 5 This is a diagram of an ocean experiment scene according to an embodiment of the present invention;

[0041] Figure 6 This is a diagram of ocean environment parameters according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] This embodiment provides a method for underwater sound source localization based on frequency domain feature enhancement and multi-scale wavelet convolution, including:

[0045] Collect the original sound pressure signal and perform frequency domain feature enhancement to generate a feature enhancement map;

[0046] The feature enhancement map is input into an underwater sound source localization model, which outputs the distance and depth of the sound source. The underwater sound source localization model is used to perform multi-feature extraction and target localization on the feature enhancement map.

[0047] Specifically, this embodiment converts the original signal into the frequency domain and extracts key feature information through a specific strategy to construct a sample covariance matrix, thereby effectively strengthening the core features in the signal, and verifying it using comparative experiments. In order to improve the accuracy of underwater target positioning, the model uses wavelet convolution and multi-scale structure for feature extraction. Wavelet convolution uses the ability to decompose data at different scales to significantly expand the model's receptive field. It can not only extract low-frequency information, but also accurately capture time-frequency features closely related to target positioning. The multi-scale convolution structure further improves the model's ability to capture local and global features, thereby enhancing its adaptability to the complexity and dynamic changes of the marine environment. Experimental results show that this method has significant advantages in improving positioning performance, and has demonstrated excellent positioning effects in practical applications, providing strong support for further optimizing underwater sound source positioning technology.

[0048] Furthermore, generating a feature enhancement map includes:

[0049] Perform Fourier transform on the original sound pressure signal to obtain frequency domain data;

[0050] Perform spectrum and energy analysis on frequency domain data to extract energy features;

[0051] Normalize the energy features and calculate the sample covariance matrix based on the normalized energy features;

[0052] Stack the sample covariance matrices to obtain the feature enhancement map.

[0053] Specifically, in response to complex challenges such as strong underwater environmental noise interference, significant multipath effect and signal attenuation, this embodiment designs a data enhancement method based on frequency domain features such as Figure 2 As shown, in order to improve the quality of underwater acoustic data and the accuracy of underwater target positioning. Since the time domain signal only reflects the change of the signal amplitude over time, this embodiment first converts the signal from the time domain to the frequency domain by performing FFT on the collected original time domain sound pressure signal, revealing the frequency components and amplitude information. After that, the energy distribution and main frequency components of the key frequency bands in the signal are extracted by performing an in-depth analysis of the frequency domain data. And taking into account the amplitude differences and inherent correlations between different collected samples, this embodiment normalizes the extracted frequency domain data, and constructs a sample covariance matrix based on the normalized data to quantify the statistical correlation between the frequency components. The specific implementation steps are as follows:

[0054] The original sound pressure data can be expressed as S = [s1, s2, ..., s t ], where s t It is represented as the sound pressure signal at time t. The time domain signal is processed by Fourier transform. The number of sampling points N is 1024. The implementation formula is as follows:

[0055]

[0056] Where p[k] represents the signal value of the kth frequency component in the frequency domain, S(n) represents the signal value of the nth sampling point, k is the frequency domain index, and n is the time domain index. Assuming that the number of vertical linear arrays of hydrophones is L, the frequency domain data p(k) is obtained by Fourier transforming the time domain signal S received by the hydrophone. Spectral and energy analysis are performed on the frequency domain data to extract the energy characteristics. Implementation formula:

[0057] E(k)=|p(k)| 2

[0058] Here, E(k) represents the energy at frequency k. To reduce the impact of random noise on the amplitude, normalization is performed during data preprocessing. Here, the L2 normalization is used to normalize the complex sound pressure data. The formula is as follows:

[0059]

[0060] in, Represents the energy size of the normalized frequency f. Then the sample covariance matrix is calculated from the vector obtained by the above formula, capturing the spatiotemporal correlation characteristics of the array received signal, providing input features for subsequent modeling, and obtaining the covariance matrix C(f):

[0061]

[0062] Where C(f) is the L×L sample covariance matrix (SCM). For broadband data, SCMs of different frequencies will be stacked in the third dimension of the matrix, so the network input feature data shape is L×L×F, where F is the number of frequencies.

[0063] Furthermore, multi-feature extraction and target positioning of the feature enhancement map include:

[0064] The feature enhancement map is input into the downsampling module, and the downsampled feature map is output;

[0065] The downsampled feature map is input into the parameter sharing layer for feature extraction to obtain multi-scale features. The parameter sharing layer includes a wavelet convolution module and an InceptionWT module. The InceptionWT module is obtained by combining depthwise separable convolution and wavelet transform.

[0066] Input the multi-scale features into the distance branch and the depth branch respectively to obtain the distance features and the depth features, wherein the distance branch and the depth branch are constructed by sequentially connecting module A, module B, pooling layer and multi-layer perceptron, and module A and module B are constructed by parallel branches including convolution and / or maximum pooling;

[0067] Output the distance and depth of the sound source based on the distance feature and depth feature.

[0068] Furthermore, the downsampled feature map is input into the parameter sharing layer for feature extraction, including:

[0069] The downsampled feature enhancement image is decomposed and reconstructed by wavelet convolution module to obtain the first feature;

[0070] The first feature is input into the InceptionWT module for feature extraction to obtain multi-scale features.

[0071] Furthermore, inputting the first feature into the InceptionWT module for feature extraction includes: inputting the first feature into several depth-wise separable convolution branches and wavelet convolution branches for feature extraction, and fusing the extracted features to obtain multi-scale features.

[0072] Specifically, for the underwater sound source localization problem, this embodiment proposes a localization algorithm based on wavelet convolution and multi-scale feature extraction, the core of which uses wavelet convolution and InceptionWT as the main feature extraction means, such as Figure 3 As shown in Figure 2, the InceptionWT module is based on a multi-branch design that combines depthwise separable convolution and wavelet transform to achieve multi-scale feature extraction. The module's input channel has four branches, three of which are depthwise separable convolutions and one is a wavelet convolution. The features of these four channels are then combined using a feature fusion mechanism.

[0073] The specific framework of the model is as follows Figure 1 As shown in the figure, the raw data is first converted to frequency domain data via Fourier transform, and then a normalized sample covariance matrix is constructed and input into the model for training. Two downsampling modules are then designed to gradually reduce the size of the feature map, reducing the number of network parameters and computational complexity. The downsampled feature maps are then input into the parameter-sharing layer where the wavelet convolution and InceptionWT modules reside for feature extraction. To enable the model to simultaneously predict the distance and depth of a sound source, two task-specific branches, the distance branch and the depth branch, are added after the parameter-sharing layer. These branches are used to extract more significant features for specific tasks, respectively. This ensures that the model achieves both prediction accuracy and inter-task synergy within the multi-task learning framework.

[0074] Distance branch: The distance branch involves distance feature extraction and output. It first passes through a module A to extract preliminary spatial features, then stacks module B three times to deeply explore multi-scale distance information. Subsequently, the feature map is compressed to a fixed size through AdaptiveAvgPool2d pooling, and finally mapped to a one-dimensional distance representation through a multi-layer perceptron (MLP) to obtain the final distance prediction value.

[0075] Depth branch: The depth branch includes distance feature extraction and output. It first passes through a module A to extract preliminary spatial features, then stacks module B three times to deeply explore multi-scale depth information, then uses AdaptiveAvgPool2d pooling to compress the feature map to a fixed size, and finally uses a multi-layer perceptron (MLP) to map it to a one-dimensional distance representation, thereby obtaining the final depth prediction value.

[0076] The distance branch and the depth branch have the same structure, but different feature information will be extracted during back propagation, which is more suitable for the current task.

[0077] like Figure 4 As shown in the figure, the structure of module A follows the Inception design principles and consists of four parallel branches: Branch 1 first compresses the channel using 1×1 convolutions, then extracts local features using 3×3 convolutions. Branch 2 similarly uses 1×1 dimensionality reduction before using 5×5 convolutions to capture a larger receptive field. Branch 3 first uses 1×1 dimensionality reduction, then uses 1×7 and 7×1 split convolutions to efficiently simulate a large convolutional receptive field, and finally uses 3×3 convolutions for further fusion. Branch 4 uses max pooling to preserve coarse-grained features before using 1×1 convolutions for dimensionality reduction. Finally, the four outputs are concatenated along the channels, allowing for the representation of features at multiple scales and shapes.

[0078] Module B structure: This further factorizes convolution, prioritizing lightweighting and channel interaction. It also consists of four branches: Branch 1 adjusts the channel and introduces nonlinearity through 1x1 convolution. Branch 2 further factorizes convolution, prioritizing lightweighting and channel interaction. It also consists of four branches: Branch 3 first uses 1x1 dimensionality reduction, then uses 3x3 feature extraction, and then uses crossed 3x1 and 1x3 convolutions to further enrich the top, bottom, left, and right spatial information. Branch 4 uses pooling to capture invariant features before performing dimensionality reduction. Finally, the four features are concatenated in the channel dimension, achieving lightweighting and diversified feature fusion.

[0079] Furthermore, the wavelet decomposition and reconstruction of the downsampled feature enhancement map through the wavelet convolution module include:

[0080] Perform wavelet decomposition on the downsampled feature enhancement image to obtain several sub-band features;

[0081] After stacking the sub-band features to the channel dimension, depth-wise separable convolution is performed to extract features and reconstruct them to obtain wavelet-reconstructed data.

[0082] Furthermore, obtaining the first feature includes: performing convolution on the feature enhancement map and data reconstructed by wavelet to obtain the first feature.

[0083] Specifically, the wavelet convolution module combines wavelet decomposition and reconstruction to perform multi-resolution decomposition on the downsampled feature enhancement map. First, the module decomposes the original signal into four sub-band data: LL, LH, HL, and HH, thereby extracting features in different frequency ranges, making it more adaptable to complex signals. In addition, by combining wavelet decomposition with convolution, the model can extract spatial structural features in the signal while extracting frequency features, obtaining multimodal features in the signal, thereby improving the positioning accuracy of the model. The specific implementation steps are as follows:

[0084] First, the current feature map is used as LL frequency data for wavelet decomposition to obtain four sub-band features:

[0085] [X LL ,X LH ,X HL ,X HH ]=WT(Input)

[0086] The four sub-band features are then stacked into the channel dimension to form a new feature map for depthwise separable convolution feature extraction:

[0087] X conv =reshape([X LL ,X LH ,X HL ,X HH ],C×4)

[0088] Y=DWconv 5×5 (X conv )

[0089] Among them, X conv is the feature after stacking the sub-band features, Y is the feature after depth-separable convolution, DWconv 5×5 It is a 5*5 depth-wise separable convolution.

[0090] Then, the currently stored high-frequency and low-frequency features are recombined and restored to a signal with a near-original resolution. By reconstructing the main information in the original signal, the effective signal can be enhanced and the model accuracy can be improved as follows:

[0091] Y iwt =IWT(Y)

[0092] Among them, IWT is inverse wavelet transform, Y iwt is the feature reconstructed by inverse wavelet transform.

[0093] Finally, the feature map is enhanced by inputting the data after wavelet reconstruction. The implementation method is as follows:

[0094] output=Conv 5×5 (input)+Y iwt

[0095] Furthermore, the Loss function includes:

[0096]

[0097] Among them, Loss represents the total loss, L r and L d are the losses of distance and depth single tasks respectively, w r and w d Represents the weight of a specific task, σ r and σ d represents the learned log variance.

[0098] Specifically, this embodiment adopts an uncertainty-weighted loss function. This method is derived from the work of Kendall et al., which calculates the loss of multiple tasks by using homoscedastic uncertainty. The loss function can be expressed as:

[0099]

[0100] Among them L i Represents the loss function of task i, w is the model weight, σ is the noise parameter, and by minimizing the noise parameter, the task loss of the basic balanced training process is achieved. By introducing the logarithmic variance σ r and σ d As a learnable parameter, it is converted into the weight of the task through the following formula:

[0101] w i =exp(-σ i )

[0102] where w i Expressed as different task weights, σ i is the logarithmic variance of different tasks. Therefore, the loss function can be rewritten as:

[0103]

[0104] Where Loss represents the total loss, L r and L d They are the losses of distance and depth single tasks respectively. The single task loss function uses square error, wr and w d Represents the weight of a specific task, σ r and σ d represents the learned log variance.

[0105] In this embodiment, the mean absolute error (MAE) is selected as the detection indicator of positioning performance, and its expression is as follows:

[0106]

[0107] Where N represents the number of test samples, y i Represents the true distance or depth value of the i-th sample, Represents the predicted distance or depth value of the i-th sample.

[0108] In this example, MAE_r (mean absolute error in distance) and MAE_d (mean absolute error in depth) are used as the main evaluation indicators of model performance. Comprehensive comparative experiments using these two indicators can objectively evaluate the model's performance in terms of prediction accuracy and systematic bias.

[0109] In order to verify the effectiveness of the method proposed in this embodiment, simulation experiments and sea trials are carried out as follows:

[0110] In the simulation experiments, this example used Saclant-Ex experimental environment parameters to generate a simulation dataset. Then, ablation experiments were designed using the simulation dataset to verify the contributions of the model components and between them. To verify the effectiveness and feasibility of the model, this example also conducted tests using sea trial data collected by the Saclant-Ex experiment and compared it with existing advanced methods.

[0111] Experimental environment:

[0112] This example is based on data collected during the Saclant-Ex experiment. The experiment was conducted by the SACLANT Center in October 1993 in a shallow area north of the island of Elba on the west coast of Italy. The experimental area was located on a flat seabed between the 120m and 140m isobaths, where acoustic propagation was under typical summer downward refraction conditions. During the experiment, the researchers deployed a vertical linear array to receive the acoustic signals, and used an auxiliary vessel to carry the acoustic signals along the Figure 5 The track shown by points A and B in the figure drags the moving sound source. The moving track of the sound source is parallel to the depth contour. The figure also shows the relative position of the vertical array and the buoy and the water depth. The distribution of the vertical linear array is shown in the figure. Figure 5As shown, the vertical linear array is located at a water depth of about 127m. The array consists of 48 hydrophones with a vertical spacing of 2m. The depth of the bottom hydrophone (hydrophone No. 1) is 112.7 meters, and the depth of the hydrophone closest to the water surface (hydrophone No. 48) is 18.7 meters. Each hydrophone collects data at a sampling rate of 1000Hz. In order to collect data that will help verify the coupled inversion method, the research center also collected a lot of acoustic-related data and environmental-related data. The underwater sound velocity profile and environmental parameters are as follows: Figure 6 As shown in Figure 2. The underwater sound velocity profile shows summer characteristics, with an almost constant velocity within a depth of about 60m. A strong thermocline exists from 60m to about 80m. In addition, the three-layer environmental parameters with sediment layer and seabed parameters are as follows: Figure 6 The sediment and seafloor parameters shown are average parameters determined by matching the predicted propagation losses with the measured propagation losses for the area.

[0113] To alleviate the data shortage, this example first used the aforementioned ocean environment parameters to simulate and generate a large amount of data. The simulated sound source frequency was set to 160-180 Hz with a 1 Hz increment; the sound source depth was set to 60-100 m with a 1 m step; and the sound source distance was set to 5-7 km with a 0.01 km step. The simulation environment parameters are shown in Table 1. Based on these parameters, the env and flp files were constructed. Simulation calculations were then performed using the KRAKEN acoustic modeling tool to generate a large amount of data to test the model's performance. A series of ablation experiments were then conducted to verify the effectiveness of each model component.

[0114] Afterwards, in order to verify the generalization ability of the model, this embodiment also used the sea trial data collected by the SACLANT Center for a comparative experiment. The experimental data was collected by the SACLANT Center using an auxiliary ship as a mobile sound source. The auxiliary ship was equipped with a pressure sensor to monitor the depth of the radiation source, and the position of the sound source was recorded using the global positioning system on the ship, providing a basis for calculating the distance between the sound source and the vertical linear array. The signal emitted by the source is a pseudo-random noise (PRN) signal with a center frequency of 370Hz and a duration of about 20 minutes. The signal is sent once a minute and each time lasts for 30 seconds. To ensure the reliability of the experiment, the collected acoustic data is divided into training, verification and test sets in a ratio of 8:1:1. The simulation parameters are shown in Table 1.

[0115] Table 1

[0116]

[0117] Ablation experiment:

[0118] In order to verify the contribution of the wavelet convolution module to the model performance, this embodiment uses simulation data to design an ablation experiment. The results of the wavelet convolution depth ablation experiment are shown in Table 2, where WTconv-1 indicates that the depth of the wavelet convolution is 1, which means that only one wavelet convolution process is performed. By gradually increasing the depth of the wavelet convolution layer and comparing the performance of models with different depths, the improvement of the model performance by the wavelet convolution module can be quantitatively evaluated. It can be seen in the table that as the depth of the wavelet convolution increases, the number of parameters and the amount of computation of the model also increase. Although there is a positive correlation between the depth of the model and the performance, this relationship is not linear. Therefore, in order to balance the relationship between accuracy and depth, this embodiment only performs ablation of four layers of wavelet convolution, and concludes that the accuracy of four layers of wavelet convolution is the best. Therefore, subsequent experiments are all based on four layers of wavelet convolution.

[0119] Table 2

[0120]

[0121] In order to explore the impact of each component of the wavelet convolution and inceptionNext modules on the performance of the model, this embodiment designed a series of ablation experiments, and the results of the ablation experiments of each module are shown in Table 3. The experiment uses a simulation data set for training and verification, and evaluates the effect of adding different modules to the baseline. First, a wavelet convolution module is added. This module enhances the receptive field of the convolutional neural network by using wavelet transform, thereby improving the model's response to low-frequency features. The input information is decomposed into components of different frequencies through wavelet convolution, and then the output is reconstructed using inverse wavelet transform to obtain a larger effective receptive field. Secondly, the model also introduces the InceptionNeXt module, which decomposes the large convolution kernel into 4 different branches, namely small kernel, orthogonal band kernel and identity mapping, which significantly reduces the memory access overhead while maintaining a large receptive field. The results show that adding different modules to the baseline significantly improves the model's performance. After adding wavelet convolution, the model's absolute errors in distance and depth in the test set were 0.01 km and 0.19 m, respectively. After adding the InceptionNeXt module, the model's errors were 0.02 km and 0.26 m. Furthermore, when both modules were added simultaneously, the model achieved optimal performance, with absolute errors of 0.01 km and 0.16 m in distance and depth, respectively. Compared to the baseline model, the ablated model variant, InceptionWT, showed significant performance improvements. In summary, this ablation experiment clarified the role and importance of each component in the model, providing a strong basis for subsequent research. These experimental results further validate the potential advantages of wavelet convolution and InceptionNeXt modules in underwater sound source localization.

[0122] Table 3

[0123]

[0124] Comparative experiment:

[0125] In order to verify the underwater positioning performance of the model in a real environment, this embodiment uses data measured by the Saclant-Ex experiment to verify the model. The sea trial data is first divided into training, verification and test sets at a ratio of 8:1:1, with a total data volume of 6,000. The experimental results are shown in Table 4. The results show that the underwater sound source localization method based on multi-scale wavelet convolution feature extraction performs superior performance in a real ocean environment. In terms of distance prediction, the absolute distance error of the model is 0.26 km, and in terms of depth prediction, the absolute depth error of the model is 0.67 m. Through comparative experiments with various methods, it can be seen that the method proposed in this embodiment has demonstrated excellent performance in the field of underwater sound source localization, verifying the effectiveness of wavelet convolution and multi-scale feature extraction in complex environments. Experimental results show that this method has significant advantages in practical applications, especially in terms of range and depth prediction. These achievements provide a more competitive solution for the field of underwater passive positioning and lay a solid foundation for subsequent research.

[0126] Table 4

[0127]

[0128] In order to verify the effectiveness of different feature extraction methods in underwater acoustic target positioning, this embodiment conducted a series of comparative experiments, and the results of the frequency domain feature enhancement comparative experiment are shown in Table 5. The experiment adopted a variety of feature extraction methods for model training. Since the emission frequency of the sound source is 170Hz, the data of the 160Hz-180Hz frequency band is selected as the Freq-specified feature. Freq-sum refers to the feature obtained by summing the data of each sub-band after dividing the 0-500Hz frequency band into multiple sub-bands. Freq-mean is the average value of the data of each sub-band in the 0-500Hz frequency band as a feature. Freq-max means extracting the maximum value of each sub-band in the 0-500Hz data as a feature. Freq-energy means extracting the maximum energy of each sub-band in the 0-500Hz data as a feature. After preprocessing, the extracted features are used as the input of the model for model training. Experimental results show that compared with other feature extraction methods, the Freq-energy feature has significant advantages in the positioning performance of underwater acoustic targets. The average absolute error and average depth error of this method are 0.12km and 0.43m respectively.

[0129] Table 5

[0130]

[0131] In summary, in order to solve the problem of sound source positioning in complex underwater environments, this embodiment innovatively proposes a method for underwater target positioning that combines frequency domain feature enhancement and multi-scale wavelet convolution. This positioning method designs a data enhancement method based on frequency domain features to address the complex challenges of underwater acoustic signals being strongly interfered by environmental noise, significant multipath propagation effects, and time-frequency attenuation characteristics. In addition, in order to overcome the defect of traditional single-task models in insufficient modeling of the correlation between sound source distance and depth, a multi-scale wavelet convolution and multi-task collaborative learning framework is constructed to achieve collaborative prediction of sound source distance and depth. Model framework such as Figure 1 As shown in the figure, the model first processes the input data through two downsampling modules. To enable information sharing and complementarity between tasks, the model designs a parameter-sharing layer. This layer, primarily composed of multi-scale wavelet convolution modules, is responsible for extracting key features from the acoustic signal. Finally, to achieve independent and accurate prediction of distance and depth, the model designs a range branch and a depth branch optimized for their respective objectives, effectively improving the prediction performance and robustness of underwater target localization tasks.

[0132] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for underwater sound source localization based on frequency domain feature enhancement and multi-scale wavelet convolution, characterized in that: include: Collect the original sound pressure signal and enhance the frequency domain characteristics. Generate feature enhancement map; The feature enhancement map is input into an underwater sound source localization model, and the distance and depth of the sound source are output, wherein the underwater sound source localization model is used to perform multi-feature extraction and target localization on the feature enhancement map.

2. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 1 is characterized in that: Generating a feature enhancement map includes: Performing Fourier transform processing on the original sound pressure signal to obtain frequency domain data; Performing spectrum and energy analysis on the frequency domain data to extract energy features; Normalizing the energy features, and calculating a sample covariance matrix based on the normalized energy features; The sample covariance matrices are stacked to obtain the feature enhancement map.

3. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 1 is characterized in that: Performing multi-feature extraction and target positioning on the feature enhancement map includes: Inputting the feature enhancement map into a downsampling module and outputting a downsampled feature map; Inputting the downsampled feature map into a parameter sharing layer for feature extraction to obtain multi-scale features, wherein the parameter sharing layer includes a wavelet convolution module and an InceptionWT module, and the InceptionWT module is obtained by combining depthwise separable convolution and wavelet transform; Inputting the multi-scale features into the distance branch and the depth branch respectively to obtain distance features and depth features, wherein the distance branch and the depth branch are constructed by sequentially connecting module A, module B, a pooling layer and a multi-layer perceptron, and the module A and module B are constructed by parallel branches including convolution and / or maximum pooling; The distance and depth of the sound source are output according to the distance feature and the depth feature.

4. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 3 is characterized in that: Inputting the downsampled feature map into a parameter sharing layer for feature extraction includes: Performing wavelet decomposition and reconstruction on the downsampled feature enhancement image through the wavelet convolution module to obtain a first feature; The first feature is input into the InceptionWT module for feature extraction to obtain the multi-scale feature.

5. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 4 is characterized in that: Inputting the first feature into the InceptionWT module for feature extraction includes: inputting the first feature into several depth-separable convolution branches and wavelet convolution branches for feature extraction, and fusing the extracted features to obtain the multi-scale feature.

6. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 4 is characterized in that: Performing wavelet decomposition and reconstruction on the downsampled feature enhancement map by the wavelet convolution module includes: Performing wavelet decomposition on the downsampled feature enhancement image to obtain a plurality of sub-band features; After stacking the sub-band features to the channel dimension, depthwise separable convolution is performed to extract features and reconstruct them to obtain wavelet reconstructed data.

7. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 6 is characterized in that: Obtaining the first feature includes: obtaining the first feature according to data after convolution of the feature enhancement map and the data reconstructed by the wavelet.

8. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 1 is characterized in that: The Loss function in the underwater sound source localization model includes: Among them, Loss represents the total loss, L r and L d are the losses of distance and depth single tasks respectively, w r and w d Represents the weight of a specific task, σ r and σ d represents the learned log variance.

Images