An underwater passive positioning method combining multi-scale feature extraction and transfer learning

Abstract

The present invention discloses an underwater passive positioning method that combines multi-scale feature extraction and transfer learning, comprising: obtaining acoustic wave data of a sound source to be located in a simulation environment; performing feature extraction on the acoustic wave data of the sound source to be located to obtain acoustic wave characteristics of the sound source; inputting the acoustic wave characteristics of the sound source into a positioning prediction model to predict the distance and depth of the sound source, wherein the positioning prediction model is obtained by training a training set, wherein the training set is waveform data of sound pressure at different frequencies; and performing domain transfer of the acoustic wave positioning problem of the sound source in the simulation environment to a real sea trial environment to perform underwater sound source positioning. The present invention can fully capture the acoustic diversity and dynamic change characteristics of the marine environment, improving the model's adaptability to the real environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater passive positioning, and in particular relates to an underwater passive positioning method combining multi-scale feature extraction and transfer learning. Background Art

[0002] In recent years, with the continuous advancement of underwater acoustic signal processing technology, underwater passive positioning, as a key branch of the underwater acoustics field, has garnered widespread attention. This technology determines the position of underwater targets by analyzing the sound pressure signals received by a hydrophone array. However, due to the complexity of the underwater propagation environment, such as multipath effects, time delays, and interference from ambient noise, estimating the distance and depth of the sound source remains a challenging task. Despite this, underwater passive positioning technology has demonstrated significant value in fields such as ocean exploration, seabed resource exploration, marine environmental monitoring, and military applications, becoming a research hotspot in the field of underwater acoustic signal processing.

[0003] Currently, research focuses on improving the accuracy of sound source localization, enhancing the model's adaptability to complex environments, and developing advanced algorithms that can cope with changing ocean conditions. These research areas not only aim to overcome the technical challenges posed by the underwater acoustic environment, but also play a significant role in further enhancing the reliability and practicality of underwater passive positioning technology.

[0004] Within the existing technology landscape, underwater passive positioning methods are becoming increasingly diverse, encompassing traditional matching field processing, machine learning positioning, and emerging deep learning-based positioning algorithms. Each method possesses specific theoretical foundations, applicable scenarios, and technical characteristics, and has achieved significant success in diverse underwater environments.

[0005] Matched field processing (MFP) methods determine the location of sound sources by comparing the correlation between copied field data calculated using an acoustic propagation model and measured acoustic data. However, this method relies heavily on the accuracy of the acoustic propagation model, and its performance can significantly degrade when there is a mismatch between the copied field and the actual acoustic environment. Furthermore, the complexity and dynamic nature of the underwater environment (such as fluctuations in the sound velocity profile, irregularities in the seafloor topography, and background noise) further exacerbate the model mismatch problem, limiting the applicability of MFP methods in highly variable environments.

[0006] The rapid development of machine learning and deep learning technologies has shown great potential in the field of underwater passive positioning. Deep learning-based algorithms have made significant progress, particularly in handling complex acoustic environments and improving positioning accuracy. However, the application of deep learning methods still faces a significant challenge: the scarcity of labeled data. Current research primarily relies on the prior knowledge of acoustic experts to generate samples, but this data scarcity significantly limits the widespread application of these methods.

[0007] Therefore, there is an urgent need for an underwater passive positioning method that combines multi-scale feature extraction and transfer learning. Summary of the Invention

[0008] To solve the above technical problems, the present invention proposes an underwater passive positioning method that combines multi-scale feature extraction and transfer learning, which can improve the model's ability to extract acoustic features and adapt to real environments, thereby expanding the application of this technology in practical applications.

[0009] The present invention provides an underwater passive positioning method combining multi-scale feature extraction and transfer learning, comprising:

[0010] Acquire the sound wave data of the sound source to be located in the simulation environment;

[0011] Extracting features from the sound wave data of the sound source to be located to obtain sound wave features of the sound source;

[0012] Inputting the sound source sound wave characteristics into a positioning prediction model to predict the sound source distance and sound source depth, wherein the positioning prediction model is obtained by training a training set, and the training set is waveform data of sound pressure at different frequencies;

[0013] The sound source acoustic wave localization problem in the simulation environment is domain-migrated to a real sea trial environment to perform underwater sound source localization.

[0014] Optionally, obtaining the training set includes:

[0015] Simulate the ocean environment and the waveform data of sound pressure at different frequencies within the preset bandwidth;

[0016] Superimposing the waveform data to obtain simulation waveform data;

[0017] The simulation waveform data is preprocessed to obtain the training set.

[0018] Optionally, performing feature extraction on the sound wave data of the sound source to be located to obtain the sound wave features of the sound source includes:

[0019] Performing Fourier transform on the simulation waveform data to obtain complex sound pressure data;

[0020] Normalizing the complex sound pressure data using an L2 normal form to obtain the characteristic signal;

[0021] A covariance operation is performed on the characteristic signal samples to obtain the sound wave characteristics of the sound source.

[0022] Optionally, a method for performing feature extraction on the sound wave data of the sound source to be located to obtain the sound wave features of the sound source is:

[0023] p(f)=[p1(f),p2(f),...,p L (f)] T

[0024] Among them, p(f) is the array signal with frequency f, p L (f) is the complex sound pressure signal of channel L at frequency f, where L is the number of vertical linear arrays of hydrophones;

[0025] The method for normalizing the complex sound pressure data using the L2 norm is:

[0026]

[0027] in, is the normalized array signal;

[0028] Then the vector obtained by the above formula is normalized for covariance processing, preliminary feature extraction is performed on the data, and the covariance matrix C(f) is constructed:

[0029]

[0030] Where C(f) is the L×L complex covariance matrix (SCM), for The transposed conjugate matrix of .

[0031] Optionally, the positioning prediction model includes: an input module, a parameter sharing module, a feature processing module and an output module;

[0032] The input module is used to input the feature data to be located;

[0033] The parameter sharing module is used for feature sharing between different tasks;

[0034] The feature processing module is used to perform distance feature processing and depth feature processing;

[0035] The output module is used to output the sound source distance and sound source depth.

[0036] Optionally, the parameter sharing module includes: a first unit, a second unit, a convolutional attention mechanism unit and a third unit;

[0037] The first unit is used to realize multi-scale feature extraction of the model;

[0038] The second unit is used to reduce the spatial dimension of the feature map while retaining key features at multiple scales;

[0039] The convolutional attention mechanism unit is used to distinguish key features in the channel dimension and locate important areas in the spatial dimension;

[0040] The third unit is used to perform multi-scale feature extraction again based on the features processed by the convolutional attention mechanism unit to achieve feature sharing between different tasks.

[0041] Optionally, the feature processing module includes: a parallel distance feature processing submodule and a depth feature submodule;

[0042] The distance feature processing submodule is used to process distance features;

[0043] The depth feature processing submodule is used to process depth features.

[0044] Optionally, the distance feature processing submodule includes: a fourth unit, a fifth unit and a global average pooling unit;

[0045] The fourth unit is used to perform multi-scale feature extraction and dimensionality reduction in the deep network;

[0046] The fifth unit is used to capture more complex local and global features and enhance the ability to combine features between branches;

[0047] The global average pooling unit is used to perform adaptive two-dimensional average pooling on the processed feature map, and finally adjust the output feature map to a specified target size.

[0048] Optionally, the deep feature processing submodule includes a fourth unit, a fifth unit, a convolutional attention mechanism unit and a global average pooling unit;

[0049] The fourth unit is used to perform multi-scale feature extraction and dimensionality reduction in the deep network;

[0050] The fifth unit is used to capture more complex local and global features and enhance the ability to combine features between branches;

[0051] The convolutional attention mechanism unit is used to distinguish key features in the channel dimension and locate important areas in the spatial dimension;

[0052] The global average pooling unit is used to perform adaptive two-dimensional average pooling on the processed feature map, and finally adjust the output feature map to a specified target size.

[0053] Compared with the prior art, the present invention has the following advantages and technical effects:

[0054] The present invention proposes a novel architecture for underwater target positioning in shallow sea time-varying environments. By integrating multiple advanced technologies, this architecture effectively addresses the dual challenges of lack of experimental data and environmental complexity.

[0055] Based on a multi-scale feature extraction strategy, the present invention can comprehensively capture the acoustic diversity and dynamic change characteristics of the marine environment by analyzing and extracting acoustic features at different scales, thereby improving the model's adaptability to the real environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0057] Figure 1 This is a flow chart of an underwater passive positioning method combining multi-scale feature extraction and transfer learning according to an embodiment of the present invention;

[0058] Figure 2 This is a simulation experiment scene diagram of an embodiment of the present invention;

[0059] Figure 3 is a marine environment map according to an embodiment of the present invention;

[0060] Figure 4 It is a diagram of the prediction model structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0061] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0062] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0063] To address the problem of data scarcity, generating simulated data has become one of the solutions. By combining ocean environment parameters with acoustic tools, a large amount of simulated data can be generated for model training, and domain migration can be achieved with the help of transfer learning technology, thereby gradually reducing dependence on real data. On this basis, based on the strategy of multi-scale feature extraction, by analyzing and extracting acoustic features at different scales, it is possible to fully capture the acoustic diversity of the ocean environment and its dynamic change characteristics. Compared with traditional feature extraction methods, multi-scale feature extraction can not only provide more comprehensive acoustic information, but also improve the robustness and accuracy of the model in complex underwater acoustic environments. The application of multi-scale feature extraction methods has become an important technical means to solve the challenges of complex underwater acoustic environments.

[0064] This embodiment proposes an underwater passive positioning method that combines multi-scale feature extraction and transfer learning, such as Figure 1As shown, the specific steps include:

[0065] Acquire the sound wave data to be located;

[0066] Input the sound wave data to be located into the positioning prediction model to predict the distance and depth of the sound source, wherein the positioning prediction model is obtained by training a training set, which is waveform data of sound pressure at different frequencies;

[0067] Domain migration of sound source distance and depth is performed for underwater positioning.

[0068] Specifically, to address the problem of underwater passive positioning, this embodiment proposes a novel underwater positioning architecture that uses multi-task learning technology to collaboratively predict the distance and depth of underwater sound sources. The framework diagram of this method is shown in the figure below. Figure 1 As shown in Figure 1, it consists of four key steps: data simulation, data preprocessing, model training, and transfer learning.

[0069] Furthermore, obtaining a training set includes:

[0070] Simulate the ocean environment and the waveform data of sound pressure at different frequencies within the preset bandwidth;

[0071] Superimpose the waveform data to obtain simulation waveform data;

[0072] Preprocess the simulation waveform data to obtain the training set.

[0073] Specifically, firstly, we simulate the experimental scene by using known environmental parameters, such as acoustic profile, sediment layer parameters and seabed parameters, etc. Figure 2 As shown in the figure, it shows the parameters of the simulated sound source and the arrangement of the receiving hydrophones. The ocean environment diagram is as follows Figure 3 As shown, the env file and flp file are written through the above-mentioned ocean environment, and then the Kraken toolbox is used to perform acoustic modeling of the ocean environment. The waveform data of the simulated sound pressure at different frequencies within a certain bandwidth are superimposed to obtain the hydrophone receiving waveform in the time domain, generating a large amount of simulation data for model training.

[0074] Furthermore, the simulation waveform data is preprocessed to obtain a training set including:

[0075] Performing Fourier transform on the simulation waveform data to obtain complex sound pressure data;

[0076] The complex sound pressure data is normalized using the L2 norm to obtain the training set.

[0077] Specifically, because the target signals received by each hydrophone in the hydrophone array are correlated, while the noise received is largely uncorrelated, we can use this property to perform preliminary feature extraction. First, we calculate the sound pressure data received by the hydrophones and use the resulting normalized sample covariance matrix to account for signal correlation and noise uncorrelation. This method can be used as a preliminary feature extraction for the raw data, effectively highlighting the target signal characteristics while suppressing the influence of uncorrelated noise. The specific steps are as follows:

[0078] The sound pressure data is expressed as p = [p1, p2, ..., p f ,…,p F ], where p F The original sound pressure data of the sensor at frequency F is obtained by Fourier transform. The sound pressure model p at frequency f f for:

[0079] p f =S(f)g(f,r)+ε

[0080] Among them, S(f) is the sound source term, g(f,r) is the Green's function, and ε is the noise.

[0081] Assuming that the number of vertical linear arrays of hydrophones is L, the complex sound pressure data is obtained by Fourier transforming the time domain signal p(t) received by the hydrophone, which is expressed as:

[0082] p(f)=[p1(f),p2(f),...,p L (f)] T

[0083] Considering that the intensity of the same sound source varies at different distances and depths, this embodiment normalizes the data during preprocessing to improve model accuracy and convergence speed and more accurately locate the sound source. This mitigates the potential impact of random noise on amplitude. The L2 normalization is used to normalize the complex sound pressure data.

[0084] The L2 normal form equation is as follows:

[0085]

[0086] Then the vector obtained by the above formula is normalized covariance processing, and preliminary feature extraction is performed on the data to obtain the covariance matrix C(f):

[0087]

[0088] Where C(f) is the L×L complex covariance matrix (SCM). The real and imaginary parts of the matrix are then separated to obtain L×L×2 single-frequency features. For broadband data, SCMs of different frequencies will be stacked in the third dimension of the matrix. Therefore, the network input feature data shape is L×L×2F, where F is the number of frequencies.

[0089] Furthermore, the positioning prediction model includes: an input module, a parameter sharing module, a feature processing module and an output module;

[0090] An input module, used for inputting data to be located;

[0091] Parameter sharing module, used for feature sharing;

[0092] Feature processing module, used for distance feature processing and depth feature processing;

[0093] Output module, used to output sound source distance and sound source depth.

[0094] Specifically, in order to effectively handle the complex situation where the target sound source is constantly moving, this embodiment regards the sound source localization problem as a regression task. The positioning algorithm model proposed in this embodiment adopts a multi-task architecture, and its model structure diagram is shown in the figure below. Figure 4 As shown in the figure, the entire network consists of four parts. The input features first enter the model from the input stream, then pass through the parameter sharing layer for feature sharing, and then pass through the distance feature branch and depth feature branch to predict the sound source distance and depth respectively.

[0095] Furthermore, the parameter sharing module includes: InceptionA unit (first unit), InceptionD unit (second unit), Convolutional Attention Mechanism unit (CBAM) and InceptionB unit (third unit);

[0096] The InceptionA unit adopts a multi-branch design. Branch 1 uses 1x1 convolution to achieve dimensionality reduction and extract local features. Branches 2 and 3 then use 3x3 and 5x5 convolution operations respectively, enabling the model to capture features of different receptive fields. Finally, the pooled features are extracted by combining maximum pooling and 1x1 convolution, achieving efficient multi-scale feature extraction.

[0097] The InceptionD unit is mainly used to reduce the spatial dimension of the feature map while retaining key information for efficient downsampling. This unit uses 3x3 convolution in the branch for complex downsampling to reduce the size of the feature map while retaining features at multiple scales;

[0098] Convolutional attention mechanism unit, which includes spatial attention mechanism and channel attention mechanism. This design enables the model to adaptively learn and adjust the importance of features, distinguishing key features in the channel dimension and locating important areas in the spatial dimension;

[0099] The InceptionB unit achieves efficient multi-scale feature extraction and reduces computational complexity through a multi-branch design combined with vertical and horizontal decomposition convolution operations. The difference from the InceptionA unit is that branches 2 and 3 use 1x7 and 7x1 convolution decompositions to extract long strip features, thereby increasing the receptive field while reducing computational complexity.

[0100] Furthermore, the feature processing module includes: a distance feature processing submodule and a depth feature submodule in parallel;

[0101] Distance feature processing submodule, used to process distance features;

[0102] The deep feature processing submodule is used to process deep features.

[0103] Furthermore, the distance feature processing submodule includes: InceptionE unit (fourth unit), InceptionC (fifth unit) unit and global average pooling unit (AdaptiveAvgPool2d);

[0104] The InceptionE unit uses a multi-branch design. Branch 1 uses 1x1 convolution for dimensionality reduction followed by 3x3 convolution for downsampling. Branch 2 further extracts features through vertical and horizontal decomposition convolution operations combined with 3x3 convolution while reducing computational complexity. Branch 3 uses maximum pooling to achieve feature downsampling and retain key information. This unit performs multi-scale feature extraction and dimensionality reduction in deep networks.

[0105] The InceptionC unit adopts a multi-branch design, in which branch 1 uses 1x1 convolution to extract local features, branch 2 uses 1x3 and 3x1 convolutions to achieve efficient feature extraction and diversified representation, and branch 3 uses 3x3 convolution followed by 1x3 and 3x1 decomposition convolution to increase the receptive field. Through the horizontal and vertical convolution decomposition operations between branches, it captures more complex local and global features and enhances the feature combination capability between branches.

[0106] The global average pooling unit is used to perform adaptive two-dimensional average pooling on the input feature map, and finally adjust the output feature map to the specified target size.

[0107] Furthermore, the deep feature processing submodule includes: InceptionE unit, InceptionC unit, convolutional attention mechanism unit (CBAM) and global average pooling unit (AdaptiveAvgPool2d);

[0108] The InceptionE unit uses a multi-branch design. Branch 1 uses 1x1 convolution for dimensionality reduction followed by 3x3 convolution for downsampling. Branch 2 further extracts features through vertical and horizontal decomposition convolution operations combined with 3x3 convolution while reducing computational complexity. Branch 3 uses maximum pooling to achieve feature downsampling and retain key information. This unit performs multi-scale feature extraction and dimensionality reduction in deep networks.

[0109] The InceptionC unit is used to adopt a multi-branch design, in which branch 1 uses 1x1 convolution to extract local features, branch 2 uses 1x3 and 3x1 convolutions to achieve efficient feature extraction and diversified representation, and branch 3 uses 3x3 convolution followed by 1x3 and 3x1 decomposition convolution to increase the receptive field. Through the horizontal and vertical convolution decomposition operations between branches, it captures more complex local and global features and enhances the ability to combine features between branches. The convolutional attention mechanism unit includes a spatial attention mechanism and a channel attention mechanism. This design enables the model to adaptively learn and adjust the importance of features, distinguishing key features in the channel dimension and locating important areas in the spatial dimension. This enables the deep feature processing submodule to automatically adjust the importance of depth-related features, improving model prediction accuracy.

[0110] The global average pooling unit is used to perform adaptive two-dimensional average pooling on the input feature map, and finally adjust the output feature map to the specified target size.

[0111] Specifically, compared to the traditional method of simply deepening the number of model layers, this embodiment adopts the method of increasing the network width by using the Inception parallel module to provide a more balanced and efficient deep learning architecture. The algorithm introduces convolution kernels of different sizes in each Inception parallel module, so that the model can extract features of different scales in parallel in the same layer, and splice the features of each branch together through feature fusion. This multi-scale feature extraction and fusion method helps the model capture various acoustic features in the marine environment, so as to better adapt to the complexity and variability of the marine environment. The Inception parallel module mainly includes an input layer and multiple convolution branches. Among them, the 1x1 convolution branch is mainly used for dimensionality reduction operations and cross-channel information integration. In addition, by decomposing the large convolution kernel by multiple small convolution kernels, the number of parameters is reduced while maintaining the receptive field.

[0112] In order to enable the model to adjust the importance of different features, this method also introduces an attention mechanism. The entire attention mechanism process is as follows:

[0113]

[0114] Where x is the input feature, x' is the enhanced feature, M c is the channel attention module, M s is the spatial attention module, is element-wise multiplication.

[0115] The channel attention and spatial attention modules are specifically implemented as follows:

[0116] M c (x)=sigmoid(MLP(Avgpool(x))+MLP(Maxpool(x))

[0117] The above formula represents the channel attention mechanism, which shares the parameters in the MLP.

[0118] M s (x))=sigmoid(Conv 7×7 ([Avgpool(x); Maxpool(x)]))

[0119] The above formula represents the spatial attention mechanism, Conv 7×7 Indicates a convolution operation with a convolution kernel of 7×7.

[0120] CBAM's dual attention mechanism design allows the model to adaptively learn and adjust the importance of features, distinguishing key features in the channel dimension and locating important areas in the spatial dimension. This approach significantly enhances the model's ability to capture complex feature interactions, thereby improving the model's accuracy and robustness in complex environments.

[0121] The hyperparameter settings during training in this embodiment are shown in Table 1:

[0122] Table 1

[0123] optimizer Adam Learning rate (lr) 0.001 batch_size 32 epoch 60

[0124] Performance evaluation indicators:

[0125] In this embodiment, the mean absolute error (MAE) is selected as the detection indicator of positioning performance, and its expression is as follows:

[0126]

[0127] Where N represents the number of test samples, y iRepresents the true distance or depth value of the i-th sample, Represents the predicted distance or depth value of the i-th sample.

[0128] 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. By conducting comprehensive comparative experiments with these two indicators, we can objectively evaluate the performance of the model in terms of prediction accuracy and systematic bias.

[0129] Transfer Learning:

[0130] Because the model was trained solely on simulated data, which cannot fully represent the complex, time-varying ocean environment, its generalization ability is relatively weak when applied to the real world. To address this performance degradation in the real world, the architecture cleverly utilizes transfer learning techniques. This allows the model to transfer knowledge learned from simulated data to real-world scenarios while adapting to the specific characteristics of the real world.

[0131] First, a base model was trained using a simulated dataset, enabling it to learn the fundamental features of distance and depth estimation tasks in a simulated environment. This step aims to establish a foundational knowledge system, capture common features, and lay the foundation for subsequent transfer learning. Subsequently, to make the model more applicable to real-world ocean environments, real-world sea trial data was divided into training, validation, and test data in a ratio of 6:1:3. The model was fine-tuned using the training data from the real-world data to achieve domain transfer of the model to the real environment. This process aims to retain the effective features acquired during simulation training while gradually adjusting the model to the real environment, thereby improving the model's robustness and generalization ability.

[0132] Table 2

[0133]

[0134] Finally, the model's performance was validated using simulation and sea trial datasets. The simulation dataset was generated based on environmental parameters collected by the SACLANT Center in shallow waters north of Elba Island on the west coast of Italy in October 1993. It was divided into training, validation, and test sets. The sea trial mobile source data was collected by the SACLANT Center using an auxiliary vessel on the afternoon of October 27, 1993, to collect mobile source data. The depth of the source buoy was indicated, and this dataset was used as the test set for the sea trial. The simulation results are shown in Table 2. The test results for the simulation dataset show that, compared to the baseline model, the proposed method improves the mean absolute error (MAE) of distance and depth estimation by 0.005 km and 0.1 m. Furthermore, the model was validated using transfer learning techniques using the sea trial dataset. The experimental results show that the proposed method achieves a mean absolute error of 0.194 km and 1.75 m for distance and depth prediction, respectively. These results demonstrate that the proposed method exhibits excellent performance in practical applications, particularly in range and depth prediction.

[0135] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. An underwater passive positioning method combining multi-scale feature extraction and transfer learning, characterized in that: include: Acquire the sound wave data of the sound source to be located in the simulation environment; Extracting features from the sound wave data of the sound source to be located to obtain sound wave features of the sound source; Inputting the sound source sound wave characteristics into a positioning prediction model to predict the sound source distance and sound source depth, wherein the positioning prediction model is obtained by training a training set, and the training set is waveform data of sound pressure at different frequencies; The positioning prediction model includes: an input module, a parameter sharing module, a feature processing module and an output module; The input module is used to input feature data to be located; The parameter sharing module is used for feature sharing between different tasks; The feature processing module is used to perform distance feature processing and depth feature processing; The output module is used to output the sound source distance and sound source depth; The parameter sharing module includes: a first unit, a second unit, a convolutional attention mechanism unit and a third unit; The first unit is used to realize multi-scale feature extraction of the model; The second unit is used to reduce the spatial dimension of the feature map while retaining key features at multiple scales; The convolutional attention mechanism unit is used to distinguish key features in the channel dimension and locate important areas in the spatial dimension; The third unit is used to perform multi-scale feature extraction again based on the features processed by the convolutional attention mechanism unit to achieve feature sharing between different tasks; The sound source acoustic wave localization problem in the simulation environment is domain-migrated to a real sea trial environment to perform underwater sound source localization.

2. The underwater passive positioning method combining multi-scale feature extraction and transfer learning according to claim 1 is characterized in that: Obtaining the training set includes: Simulate the ocean environment and the waveform data of sound pressure at different frequencies within the preset bandwidth; Superimposing the waveform data to obtain simulation waveform data; The simulation waveform data is preprocessed to obtain the training set.

3. The underwater passive positioning method combining multi-scale feature extraction and transfer learning according to claim 2 is characterized in that: Extracting features from the sound wave data of the sound source to be located to obtain the sound wave features of the sound source includes: Performing Fourier transform on the simulation waveform data to obtain complex sound pressure data; Normalizing the complex sound pressure data using an L2 normal form to obtain a characteristic signal; A covariance operation is performed on the characteristic signal to obtain the sound wave characteristics of the sound source.

4. The underwater passive positioning method combining multi-scale feature extraction and transfer learning according to claim 3 is characterized in that: The method of extracting features from the sound wave data of the sound source to be located and obtaining the sound wave features of the sound source is as follows: p(f)=[p1(f),p2(f),...,p L (f)] T Among them, p(f) is the array signal with frequency f, p L (f) is the complex sound pressure signal of channel L at frequency f, where L is the number of vertical linear arrays of hydrophones; The method for normalizing the complex sound pressure data using the L2 norm is: in, is the normalized array signal; Then the vector obtained by the above formula is normalized for covariance processing, preliminary feature extraction is performed on the data, and the covariance matrix C(f) is constructed: Where C(f) is the L×L complex covariance matrix SCM, for The transposed conjugate matrix of .

5. The underwater passive positioning method combining multi-scale feature extraction and transfer learning according to claim 1, characterized in that: The feature processing module includes: a distance feature processing submodule and a depth feature processing submodule in parallel; The distance feature processing submodule is used to process distance features; The depth feature processing submodule is used to process depth features.

6. The underwater passive positioning method combining multi-scale feature extraction and transfer learning according to claim 5, characterized in that: The distance feature processing submodule includes: a fourth unit, a fifth unit and a global average pooling unit; The fourth unit is used to perform multi-scale feature extraction and dimensionality reduction in the deep network; The fifth unit is used to capture more complex local and global features and enhance the ability to combine features between branches; The global average pooling unit is used to perform adaptive two-dimensional average pooling on the processed feature map, and finally adjust the output feature map to a specified target size.

7. The underwater passive positioning method combining multi-scale feature extraction and transfer learning according to claim 5, characterized in that: The deep feature processing submodule includes a fourth unit, a fifth unit, a convolutional attention mechanism unit and a global average pooling unit; The fourth unit is used to perform multi-scale feature extraction and dimensionality reduction in the deep network; The fifth unit is used to capture more complex local and global features and enhance the ability to combine features between branches; The convolutional attention mechanism unit is used to distinguish key features in the channel dimension and locate important areas in the spatial dimension; The global average pooling unit is used to perform adaptive two-dimensional average pooling on the processed feature map, and finally adjust the output feature map to a specified target size.

Images