An underwater sound source positioning method based on frequency domain feature enhancement and multi-scale wavelet convolution

Abstract

The present application relates to a kind of underwater sound source positioning method based on frequency domain feature enhancement and multi-scale wavelet convolution, comprising: original sound pressure signal is collected and is carried out frequency domain feature enhancement, generates feature enhancement chart;The feature enhancement chart is input into underwater sound source positioning model, and the distance and depth of sound source are output, wherein the underwater positioning model is used to carry out feature extraction and target positioning to the feature enhancement chart.The present application is aimed at underwater environmental noise interference, multi-path propagation effect is significant and time-frequency attenuation and other complex challenges, proposes a kind of frequency domain feature enhancement technology to improve model positioning precision, and by fusing wavelet convolution and multi-scale convolution, the multi-band feature of acoustic data is effectively extracted, to accurately capture its spatial distribution characteristics.

Description

Technical Field

[0001] This invention relates to the field of underwater sound source localization technology, and in particular to an underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution. Background Technology

[0002] Underwater sound source localization has always 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 copies of the field calculated by the acoustic model and actual sea trial acoustic data, generating a two-dimensional fuzzy surface for the distance and depth of the underwater sound source. However, when key marine environmental parameters used for theoretical modeling (such as sound velocity profiles, sediment parameters, and bottom layer parameters) deviate from the actual environment, the localization performance of MFP will significantly decrease. Inspired by matched field processing, Liu et al. proposed a feature matching method based on convolutional neural networks to estimate the depth of sound sources in shallow water environments. Compared with the traditional MFP method, FM-CNN shows better robustness to environmental mismatch. While the above research demonstrates the effectiveness of convolutional neural network methods in underwater sound source distance estimation, it mainly focuses on the localization model itself and fails to fully consider the interference problems commonly present in the actual underwater environment, such as noise, multipath propagation, and signal attenuation. If these factors are not adequately considered, the robustness and accuracy of the model in real-world scenarios will be greatly weakened.

[0003] In recent years, with the development of deep learning theory and the increase in computing resources, data-driven deep learning methods have been widely used in underwater acoustic source localization. This method does not rely on complex sound field modeling and can automatically extract key features from a large amount of data even when some environmental parameters are missing, thus achieving good localization performance. Research on data-driven underwater acoustic source localization can be traced back to around the 1990s, when shallow neural networks were first used to achieve ranging of underwater acoustic targets. However, due to the limitations of computing resources at that time, data-driven methods did not make substantial progress. In recent years, with the rapid development of machine learning theory and the significant improvement in computing power, data-driven methods have been revitalized in the field of underwater acoustic localization. To overcome the limitations of matching field processing methods, machine learning methods have been widely used in the field of underwater acoustic localization. He et al. proposed a data model-driven hybrid multi-task learning model that combines Transformer and Convolutional Neural Networks (CNN) with Multi-Level Dilated Convolutional Fusion (MLF) modules for underwater passive localization. Although the above method proposes a model framework based on multi-task learning, the experiments only train and test on the same dataset, lacking adaptability to environmental changes. When applied to real-world environments that differ from the training data, the model's performance may deteriorate significantly.

[0004] To address the aforementioned problems, this 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 this invention is to provide an underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution. In response to the complex challenges of strong underwater environmental noise interference, significant multipath propagation effects, and time-frequency attenuation, the method can not only effectively extract multi-frequency features of acoustic data by fusing wavelet transform and convolution operations, but also accurately capture its spatial distribution characteristics.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] An underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution includes:

[0008] The raw sound pressure signal is acquired and frequency domain feature enhancement is performed to generate a feature enhancement map;

[0009] The feature enhancement map is input into the 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.

[0010] Optionally, generating feature-enhanced maps includes:

[0011] The original sound pressure signal is subjected to Fourier transform processing to obtain frequency domain data;

[0012] Spectrum and energy analysis are performed on the frequency domain data to extract energy features;

[0013] The energy characteristics are normalized, and the sample covariance matrix is ​​calculated based on the normalized energy characteristics.

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

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

[0016] The feature enhancement map is input into the downsampling module, which outputs the downsampled feature map.

[0017] 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.

[0018] The multi-scale features are input into the distance branch and the depth branch respectively to obtain distance features and depth features. The distance branch and the depth branch are constructed by sequentially connecting module A, module B, pooling layer and multilayer perceptron. Module A and module B are constructed by parallel branches including convolution and / or max pooling.

[0019] Based on the distance and depth features, the distance and depth of the sound source are output.

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

[0021] The wavelet convolution module is used to perform wavelet decomposition and reconstruction on the downsampled feature enhancement map to obtain the 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 depthwise separable convolutional branches and wavelet convolutional branches for feature extraction, and fusing the extracted features to obtain the multi-scale features.

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

[0025] Wavelet decomposition is performed on the downsampled feature enhancement map to obtain several sub-band features;

[0026] The sub-band features are stacked to the channel dimension and then subjected to depthwise separable convolution to extract features and reconstruct the data, obtaining the wavelet-reconstructed data.

[0027] Optionally, obtaining the first feature includes: obtaining the first feature by performing convolution on the feature enhancement map and then reconstructing it using the wavelet data.

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

[0029]

[0030] Where Loss represents the total loss, L r and L d The losses for distance and depth are respectively for individual tasks, w r and w d σ represents the weight of a specific task. r and σ d This represents the learned logarithmic variance.

[0031] The beneficial effects of this invention are as follows:

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

[0033] (2) Innovatively combining wavelet convolution with multi-scale structure, a multi-scale wavelet convolution structure was constructed. This structure not only effectively captures the low-frequency information of the signal, but also extracts local and global time-frequency features at the same time, thereby significantly enhancing the model's adaptability to complex marine environments.

[0034] (3) The effectiveness of the proposed feature extraction method and multi-scale wavelet convolution structure in underwater target localization was verified through ablation experiments and comparative experiments, demonstrating its significant advantages in improving localization accuracy and robustness. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[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 flowchart illustrating the feature enhancement process according to an embodiment of the present invention;

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

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

[0040] Figure 5 This is a marine experimental scenario diagram according to an embodiment of the present invention;

[0041] Figure 6 This is a diagram of marine environmental parameters according to an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

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

[0045] The raw sound pressure signal is acquired and frequency domain feature enhancement is performed to generate a feature enhancement map;

[0046] The feature-enhanced map is input into the underwater sound source localization model, which outputs the distance and depth of the sound source. The underwater sound source localization model is used to extract multiple features from the feature-enhanced map and locate the target.

[0047] Specifically, this embodiment performs frequency domain transformation on the original signal and extracts key feature information to construct a sample covariance matrix using a specific strategy, thereby effectively strengthening the core features in the signal. Comparative experiments are used to verify this. To improve the accuracy of underwater target localization, the model employs wavelet convolution and a multi-scale structure for feature extraction. Wavelet convolution utilizes data decomposition capabilities at different scales to significantly expand the model's receptive field, enabling the extraction of low-frequency information and accurate capture of time-frequency features closely related to target localization. The multi-scale convolution structure further enhances the model's ability to capture local and global features, thereby strengthening its adaptability to the complexity and dynamic changes of the marine environment. Experimental results show that this method has significant advantages in improving localization performance and demonstrates excellent localization results in practical applications, providing strong support for further optimization of underwater sound source localization technology.

[0048] Furthermore, generating feature-enhanced maps includes:

[0049] The original sound pressure signal is subjected to Fourier transform processing to obtain frequency domain data;

[0050] Perform spectral and energy analysis on frequency domain data to extract energy characteristics;

[0051] The energy characteristics are normalized, and the sample covariance matrix is ​​calculated based on the normalized energy characteristics.

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

[0053] Specifically, addressing the complex challenges of strong underwater environmental noise interference, significant multipath effects, and signal attenuation, this embodiment designs a data augmentation method based on frequency domain features, such as... Figure 2 As shown, this method aims to improve the quality of underwater acoustic data and the accuracy of underwater target positioning. Since time-domain signals only reflect the change in signal amplitude over time, this embodiment first performs an FFT on the acquired raw time-domain sound pressure signal to convert the signal from the time domain to the frequency domain, revealing the frequency components and amplitude information. Then, through in-depth analysis of the frequency domain data, the energy distribution and main frequency components of key frequency bands in the signal are extracted. Furthermore, considering the amplitude differences and inherent correlations between different acquired 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 each frequency component. The specific implementation steps are as follows:

[0054] The raw sound pressure data can be represented as S=[s1,s2,…,s t ], where s t Let be the sound pressure signal at time t. Perform a Fourier transform on the time-domain signal, with N being 1024 sampling points. The formula is as follows:

[0055]

[0056] Where p[k] represents the signal value of the k-th frequency component in the frequency domain, S(n) represents the signal value of the n-th sampling point, k is the frequency domain index, and n is the time domain index. Assuming the number of vertical linear arrays in the hydrophone is L, the frequency domain data p(k) is obtained by performing a Fourier transform on the time domain signal S received by the hydrophone. Spectrum and energy analysis are then performed on the frequency domain data to extract energy characteristics. Implementation formula:

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

[0058] Where E(k) represents the energy level at frequency k. Subsequently, to reduce the excessive influence of random noise on the amplitude, a normalization operation was performed during data preprocessing. Here, the L2 norm was used to normalize the complex sound pressure data. The formula is as follows:

[0059]

[0060] in, This represents the energy magnitude at the normalized frequency f. Then, the sample covariance matrix is ​​calculated from the vector obtained in the above equation, capturing the spatiotemporal correlation characteristics of the array-received signal and providing input features for subsequent modeling, resulting in the covariance matrix C(f):

[0061]

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

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

[0064] Input the feature enhancement map into the downsampling module, and output the downsampled feature map;

[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] Multi-scale features are input into the distance branch and the depth branch respectively to obtain distance features and depth features. The distance branch and the depth branch are constructed by sequentially connecting module A, module B, pooling layer and multilayer perceptron. Module A and module B are constructed by parallel branches including convolution and / or max pooling.

[0067] Based on distance and depth characteristics, the distance and depth of the sound source are output.

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

[0069] The first feature is obtained by performing wavelet decomposition and reconstruction on the downsampled feature enhancement map using a wavelet convolution module.

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

[0071] Furthermore, the first feature is input into the InceptionWT module for feature extraction, which includes: inputting the first feature into several depthwise separable convolutional branches and wavelet convolutional branches for feature extraction, and fusing the extracted features to obtain multi-scale features.

[0072] Specifically, for the problem of underwater sound source localization, this embodiment proposes a localization algorithm based on wavelet convolution and multi-scale feature extraction. The core algorithm employs wavelet convolution and InceptionWT as the main feature extraction methods, such as... Figure 3 As shown, the InceptionWT module is based on a multi-branch design, fusing depthwise separable convolution and wavelet transform to achieve multi-scale feature extraction. The module has a total of four input channels, three of which are depthwise separable convolutions and the other is a wavelet convolution. Subsequently, a feature fusion mechanism is used to fuse the features from the four channels.

[0073] The specific framework of the model is as follows: Figure 1 As shown in the figure, the original data is first converted into frequency domain data through 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 maps, thereby reducing the number of network parameters and computational cost. The downsampled feature maps are input into the parameter-sharing layer containing wavelet convolution and the InceptionWT module for feature extraction. To enable the model to simultaneously predict the distance and depth of the sound source, two task-specific branches are added after the parameter-sharing layer: a distance branch and a depth branch. These branches are used to extract more significant features in specific tasks, ensuring that the model possesses both prediction accuracy and inter-task synergy within a multi-task learning framework.

[0074] Distance branch: The distance branch includes distance feature extraction and output. First, it goes through a module A to extract preliminary spatial features. Then, it stacks three modules B to mine multi-scale distance information. Subsequently, it compresses the feature map to a fixed size through AdaptiveAvgPool2d pooling. Finally, it is mapped to a one-dimensional distance representation through a multilayer perceptron (MLP) to obtain the final distance prediction value.

[0075] The deep branch includes distance feature extraction and output. First, it goes through a module A to extract preliminary spatial features. Then, it stacks three modules B to further mine multi-scale depth information. Subsequently, it compresses the feature map to a fixed size through AdaptiveAvgPool2d pooling. Finally, it is mapped to a one-dimensional distance representation through a multilayer perceptron (MLP) to obtain the final depth prediction value.

[0076] The distance branch and depth branch have the same structure, but different feature information is extracted during backpropagation, making them more suitable for the current task.

[0077] like Figure 4 As shown, Module A's structure continues the Inception design concept, containing four parallel branches: Branch 1 first uses 1×1 convolution to compress channels, then uses 3×3 convolution to extract local features; Branch 2 similarly first uses 1×1 to reduce dimensionality, then uses 5×5 convolution to capture a larger receptive field; Branch 3 first uses 1×1 to reduce dimensionality, then uses 1×7 and 7×1 split convolutions to efficiently simulate a large convolutional receptive field, and finally uses 3×3 convolution for further fusion; Branch 4 uses max pooling to retain coarse-grained features and then uses 1×1 convolution for dimensionality reduction. The four outputs are finally concatenated along the channels, taking into account the representation of features of multiple scales and different shapes.

[0078] Module B structure: Further factorizing the convolution, emphasizing lightweight design and channel interaction, also consists of four branches: Branch 1 adjusts channels and introduces non-linearity through 1x1 convolution; Branch 2 further factorizes the convolution, emphasizing lightweight design and channel interaction, also consisting of four branches: Branch 3 first uses 1x1 for dimensionality reduction, then uses 3x3 to extract features, and subsequently enriches the spatial information in all directions through interleaved 3x1 and 1x3 convolutions; Branch 4 uses pooling to capture invariant features before dimensionality reduction. Finally, the four features are concatenated along the channel dimension, achieving lightweight design and diverse feature fusion.

[0079] Furthermore, wavelet decomposition and reconstruction of the downsampled feature enhancement map using a wavelet convolution module includes:

[0080] Wavelet decomposition is performed on the downsampled feature enhancement map to obtain several sub-band features;

[0081] After stacking the subband features to the channel dimension, depthwise separable convolution is performed to extract features and reconstruct the data to obtain wavelet-reconstructed data.

[0082] Furthermore, obtaining the first feature includes: obtaining the first feature based on the data after convolution of the feature enhancement map and wavelet reconstruction.

[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-bands: LL, LH, HL, and HH, thereby extracting features from different frequency ranges and enhancing its adaptability to complex signals. Furthermore, by combining wavelet decomposition with convolution, the model can extract not only frequency features but also spatial structure features, acquiring multimodal features and improving the model's localization accuracy. The specific implementation steps are as follows:

[0084] First, the current feature map is used as LL frequency data and wavelet decomposition is performed to obtain 4 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, which is then used for depthwise separable convolution to extract features.

[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 Y represents the features after stacking subband features, and DWconv represents the features after depthwise separable convolution. 5×5 It is a 5x5 depthwise separable convolution.

[0090] Then, the signal with approximately the original resolution is recovered by recombining and restoring the high-frequency and low-frequency features stored at present. 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] Where IWT is the inverse wavelet transform, Y iwt These are the features reconstructed by the inverse wavelet transform.

[0093] Finally, feature map enhancement is performed using the input and wavelet-reconstructed data, as follows:

[0094] output = Conv 5×5 (input)+Y iwt

[0095] Furthermore, the Loss function includes:

[0096]

[0097] Where Loss represents the total loss, L r and L d The losses for distance and depth are respectively for individual tasks, w r and w d σ represents the weight of a specific task. r and σ d This represents the learned logarithmic variance.

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

[0099]

[0100] Where L i Let w represent the loss function for task i, w be the model weights, and σ be the noise parameter. Minimizing the noise parameter achieves a basic balance in the task loss during training. This is achieved by introducing the log-variance σ. r and σ d These parameters are learnable and can be converted into task weights using the following formula:

[0101] w i =exp(-σ i )

[0102] Where w i Represented as different task weights, σ i Let be the log-variance for different tasks. Therefore, the loss function can be rewritten as:

[0103]

[0104] Where Loss represents the total loss, L r and L d The losses for distance and depth are respectively calculated for individual tasks. The loss function for each individual task uses the squared error, w.r and w d σ represents the weight of a specific task. r and σ d This represents the learned logarithmic variance.

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

[0106]

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

[0108] In this embodiment, MAE_r (mean absolute error of distance) and MAE_d (mean absolute error of depth) are used as the main evaluation metrics for model performance. Comprehensive comparative experiments using these two metrics allow for an objective assessment of the model's performance in terms of prediction accuracy and systematic bias.

[0109] To verify the effectiveness of the method proposed in this embodiment, simulation experiments and sea trials are conducted below:

[0110] In the simulation experiment, this embodiment uses Saclant-Ex experimental environment parameters to generate a simulation dataset. Then, an ablation experiment was designed using the simulation dataset to verify the contributions of each component of the model and between components. To verify the effectiveness and feasibility of the model, this embodiment also used sea trial data collected from the Saclant-Ex experiment for testing and compared it with existing state-of-the-art methods.

[0111] Experimental environment:

[0112] This embodiment is based on data collected from the Saclant-Ex experiment. This experiment was conducted by the Saclant Center in October 1993 in shallow waters north of the island of North Elba off the west coast of Italy. The experimental area was on a flat seabed between the 120m and 140m isobaths, where acoustic propagation was under typical summer downward refraction conditions. In the experiment, researchers deployed a vertical linear array to receive acoustic signals and used an auxiliary vessel along... Figure 5 The sound source is dragged along the tracks shown at points A and B. The track of the sound source is parallel to the isobaths. The figure also shows the relative positions of the vertical array and the buoy, as well as the water depth. The distribution of the vertical linear array is as follows: Figure 5As shown, the vertical linear array is located at a water depth of approximately 127m and consists of 48 hydrophones spaced 2m apart vertically. The bottom hydrophone (hydrophone 1) is at a depth of 112.7m, and the hydrophone closest to the surface (hydrophone 48) is at a depth of 18.7m. Each hydrophone collects data at a sampling rate of 1000Hz. Furthermore, to collect data that helps validate the coupled inversion method, the research center also collected a significant amount of acoustic and environmental data. Underwater sound velocity profiles and environmental parameters are shown below. Figure 6 As shown, the underwater sound velocity profile exhibits summer characteristics, with almost constant velocity up to a depth of approximately 60 meters, and a strong thermocline existing between 60 and 80 meters. Furthermore, three layers of environmental parameters, including sedimentary layers and seafloor parameters, are obtained as follows: Figure 6 As shown, the sediment and seafloor parameters are averaged parameters determined by matching the predicted propagation loss in the region with the measured propagation loss.

[0113] To alleviate the problem of insufficient data, this embodiment first uses the aforementioned marine environmental parameters to perform data simulation to generate a large amount of data. The frequency of the simulated sound source is set to 160-180Hz with an increment of 1Hz; the sound source depth is set to 60-100m with a step size of 1m; and the sound source distance is set to 5-7km with a step size of 0.01km. The simulation environment parameters are shown in Table 1. Based on these parameters, env and flp files were constructed. Subsequently, the KRAKEN acoustic modeling tool was used for simulation calculations to generate a large amount of data to test the model performance. A series of ablation experiments were then conducted to verify the effectiveness of each component of the model.

[0114] To verify the model's generalization ability, this embodiment also utilized sea trial data collected by the SACLANT Center for comparative experiments. This experimental data was obtained by the SACLANT Center using an auxiliary vessel as a mobile sound source. The auxiliary vessel was equipped with pressure sensors to monitor the depth of the radioactive source and used its GPS to record the source's location, providing a basis for calculating the distance between the sound source and the vertical linear array. The source emitted a pseudo-random noise (PRN) signal with a center frequency of 370 Hz, lasting approximately 20 minutes, transmitted once per minute for 30 seconds each time. To ensure the reliability of the experiment, the collected acoustic data was divided into training, validation, and test sets in an 8:1:1 ratio. Simulation parameters are shown in Table 1.

[0115] Table 1

[0116]

[0117] Ablation experiment:

[0118] To verify the contribution of the wavelet convolution module to model performance, this embodiment designed an ablation experiment using simulation data. The results of the wavelet convolution depth ablation experiment are shown in Table 2, where WTconv-1 represents a wavelet convolution depth of 1, meaning only one wavelet convolution process was performed. By gradually increasing the depth of the wavelet convolution layers and comparing the performance of models with different depths, the improvement of model performance by the wavelet convolution module can be quantitatively evaluated. The table shows that as the depth of the wavelet convolution increases, the number of model parameters and computational cost also increase. Although there is a positive correlation between model depth and performance, this relationship is not linear. Therefore, to balance the relationship between accuracy and depth, this embodiment only performed ablation with four layers of wavelet convolution, and it was found that four layers of wavelet convolution provided the best accuracy. Therefore, subsequent experiments were all based on four layers of wavelet convolution.

[0119] Table 2

[0120]

[0121] To investigate the impact of wavelet convolution and the various components of the InceptionNext module on model performance, this embodiment designed a series of ablation experiments. The results of the ablation experiments for each module are shown in Table 3. The experiments used a simulated dataset for training and validation, evaluating the effects of adding different modules to the baseline. First, a wavelet convolution module was added. This module enhances the receptive field of the convolutional neural network by utilizing wavelet transform, improving the model's response to low-frequency features. Wavelet convolution decomposes the input information into components of different frequencies, and then inverse wavelet transform is used to reconstruct the output, resulting in a larger effective receptive field. Second, the model also introduced the InceptionNeXt module, which decomposes the large convolutional kernel into four different branches: a small kernel, an orthogonal band kernel, and an identity mapping. While maintaining a large receptive field, this significantly reduces memory access overhead. The results show that adding different modules to the baseline significantly improves the model's performance. After adding wavelet convolution, the absolute errors of distance and depth on the test set are 0.01 km and 0.19 m, respectively. After adding the InceptionNeXt module, the errors are 0.02 km and 0.26 m. Furthermore, the model achieves optimal performance when both modules are added simultaneously, with absolute errors of 0.01 km and 0.16 m. Compared to the baseline model, the ablation variant InceptionWT shows a significant performance improvement. In conclusion, this ablation experiment clarifies the role and importance of each component in the model, providing strong evidence for future research. These experimental results further validate the potential advantages of wavelet convolution and the InceptionNeXt module in underwater sound source localization.

[0122] Table 3

[0123]

[0124] Comparative experiment:

[0125] To verify the model's underwater positioning performance in real-world environments, this embodiment uses data from the Saclant-Ex experiment for model validation. The sea trial data was first divided into training, validation, and test sets in an 8:1:1 ratio, with a total of 6000 data points. The experimental results are shown in Table 4. The results demonstrate that the underwater sound source localization method based on multi-scale wavelet convolution feature extraction exhibits superior performance in real-world marine environments. In distance prediction, the model's absolute distance error is 0.26 km, and in depth prediction, the model's absolute depth error is 0.67 m. Comparative experiments with various methods show that the proposed method in this embodiment demonstrates excellent performance in underwater sound source localization, validating the effectiveness of wavelet convolution and multi-scale feature extraction in complex environments. Experimental results indicate that this method has significant advantages in practical applications, especially in range and depth prediction. These achievements provide a more competitive solution for underwater passive positioning and lay a solid foundation for future research.

[0126] Table 4

[0127]

[0128] To verify the effectiveness of different feature extraction methods in underwater acoustic target localization, this embodiment conducted a series of comparative experiments. The results of the frequency domain feature enhancement comparison experiments are shown in Table 5. Multiple feature extraction methods were used for model training. Since the sound source's emission frequency is 170Hz, data from the 160Hz-180Hz frequency band were selected as Freq-specified features. 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 uses the average value of the data in each sub-band within the 0-500Hz frequency band as a feature. Freq-max represents the maximum value of each sub-band in the 0-500Hz data as a feature, and Freq-energy represents the maximum energy value of each sub-band in the 0-500Hz data as a feature. After preprocessing, the extracted features were used as input to the model for training. Experimental results show that Freq-energy features have significant advantages in the localization performance of underwater acoustic targets compared with other feature extraction methods. The mean absolute error and mean depth error of this method are 0.12 km and 0.43 m, respectively.

[0129] Table 5

[0130]

[0131] In summary, this embodiment innovatively proposes an underwater target localization method that combines frequency domain feature enhancement and multi-scale wavelet convolution to address the problem of sound source localization in complex underwater environments. This method addresses the complex challenges of strong environmental noise interference, significant multipath propagation effects, and time-frequency attenuation characteristics of underwater acoustic signals by designing a data augmentation method based on frequency domain features. Furthermore, to overcome the shortcomings of traditional single-task models in modeling 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. The model framework is as follows: Figure 1 As shown, the model first processes the input data through two downsampling modules. Then, to achieve information sharing and complementarity between tasks, a parameter-sharing layer is designed. This layer mainly consists of multi-scale wavelet convolution modules responsible for extracting key features from the acoustic signal. Finally, to achieve independent and accurate prediction of distance and depth, the model designs distance and depth branches optimized for their respective targets, effectively improving the prediction performance and robustness of underwater target localization tasks.

[0132] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope 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: The raw sound pressure signal is acquired and its frequency domain features are enhanced. Generate feature-enhanced maps; The feature enhancement map is input into the 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. Performing multi-feature extraction and target localization on the feature enhancement map includes: The feature enhancement map is input into the downsampling module, which outputs the downsampled feature map. 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. The multi-scale features are input into the distance branch and the depth branch respectively to obtain distance features and depth features. The distance branch and the depth branch are constructed by sequentially connecting module A, module B, pooling layer and multilayer perceptron. Module A and module B are constructed by parallel branches including convolution and / or max pooling. Based on the distance and depth features, the distance and depth of the sound source are output; The downsampled feature map is input into the parameter sharing layer for feature extraction, which includes: The wavelet convolution module is used to perform wavelet decomposition and reconstruction on the downsampled feature enhancement map to obtain the first feature. The first feature is input into the InceptionWT module for feature extraction to obtain the multi-scale feature; The first feature is input into the InceptionWT module for feature extraction, which includes: inputting the first feature into several depthwise separable convolutional branches and wavelet convolutional branches for feature extraction, and fusing the extracted features to obtain the multi-scale features; The wavelet decomposition and reconstruction of the downsampled feature enhancement map by the wavelet convolution module includes: Wavelet decomposition is performed on the downsampled feature enhancement map to obtain several sub-band features; The sub-band features are stacked to the channel dimension and then subjected to depthwise separable convolution to extract features and reconstruct the data, obtaining the wavelet-reconstructed data.

2. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 1, characterized in that, The generation of feature enhancement maps includes: The original sound pressure signal is subjected to Fourier transform processing to obtain frequency domain data; Spectrum and energy analysis are performed on the frequency domain data to extract energy features; The energy characteristics are normalized, and the sample covariance matrix is ​​calculated based on the normalized energy characteristics. 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, characterized in that, Obtaining the first feature includes: obtaining the first feature by performing convolution on the feature enhancement map and then reconstructing it using the wavelet.

4. The underwater sound source localization method based on frequency domain feature enhancement and multi-scale wavelet convolution according to claim 1, characterized in that, The Loss function in the underwater sound source localization model includes: ； Where Loss represents the total loss. and The losses for distance and depth are respectively for individual tasks. and Represents the weight of a specific task. and This represents the learned logarithmic variance.

Images