A Deep-Sea Underwater Acoustic Propagation Loss Prediction Method Based on Deep Learning

Abstract

This application belongs to the field of ocean exploration technology, and specifically relates to a deep-sea underwater acoustic propagation loss prediction method based on deep learning, including obtaining a one-dimensional sound velocity profile; preprocessing the one-dimensional sound velocity profile to obtain a two-dimensional velocity model, and converting the two-dimensional velocity model into spatio-temporal feature data with positional encoding; using an attention mechanism to calculate the correlation between spatio-temporal feature data to obtain the first output of multiple parallel attention heads, and performing feature mapping after fusing the first output to form a second output; performing a convolutional dimensionality reduction operation on the features of the second output to obtain dimensionality-reduced features; performing channel weighted fusion on the dimensionality-reduced features and source information; mapping the dimensionality-reduced features back to two-dimensional spatial information features, and cross-using convolution and transposed convolution in combination with source information, and performing convolution on the combined features to output the underwater acoustic propagation loss distribution. It can accurately predict key details such as interference fringes of propagation loss and is suitable for complex deep-sea waveguide environments.

Description

Technical Field

[0001] This application belongs to the field of ocean exploration technology, and particularly relates to a method for predicting deep - sea underwater acoustic propagation loss based on deep learning. Background Art

[0002] Underwater acoustic propagation refers to the process of sound waves propagating in water. Especially in the ocean environment, underwater acoustic propagation is widely used in fields such as ocean exploration, marine bioacoustics, and ocean communication. Underwater acoustic propagation loss is a phenomenon that describes the energy attenuation of sound waves during propagation due to reasons such as absorption, scattering, and refraction. Low - frequency sound waves, due to their low frequency and slow attenuation, have the characteristic of long - distance propagation, so their application in the deep - sea environment is of great significance. However, traditional underwater acoustic propagation loss models mainly rely on models based on physical principles, such as ray models, normal mode models, parabolic equation models, etc. These models usually require a large number of environmental parameter inputs, have poor adaptability to different ocean environments, high computational complexity, and are difficult to adapt to complex and variable ocean environments. Summary of the Invention

[0003] An embodiment of this application provides a method for predicting deep - sea underwater acoustic propagation loss based on deep learning, which solves the problems of high computational complexity and difficulty in adapting to complex and variable ocean environments.

[0004] This application is implemented as follows:

[0005] A method for predicting deep - sea underwater acoustic propagation loss based on deep learning includes:

[0006] Obtain a one - dimensional sound speed profile;

[0007] Process the one - dimensional sound speed profile based on a trained deep - learning network and output the underwater acoustic propagation loss distribution; including:

[0008] Pre - process the one - dimensional sound speed profile to obtain a two - dimensional velocity model, and convert the two - dimensional velocity model into spatio - temporal feature data with positional encoding;

[0009] Use the attention mechanism to calculate the correlation between spatio - temporal feature data to obtain the first output of multiple parallel attention heads, and perform feature mapping on the fused first output to form a second output;

[0010] Perform a convolution dimensionality reduction operation on the features of the second output to obtain dimensionality - reduced features;

[0011] Perform channel - weighted fusion of the dimensionality - reduced features with the sound source information;

[0012] Map the dimensionality - reduced features back to two - dimensional spatial information features, cross - use convolution and transposed convolution in combination with the sound source information, and perform convolution on the combined features to output the underwater acoustic propagation loss distribution.

[0013] Further, preprocess the one-dimensional sound speed profile to obtain a two-dimensional velocity model, and convert the two-dimensional velocity model into spatio-temporal feature data with positional encoding, including: expanding the one-dimensional sound speed profile according to the horizontal distance to form a two-dimensional velocity model, and adding positional encoding to the two-dimensional velocity model after dimensionality reduction to form spatio-temporal feature data.

[0014] Further, adding positional encoding to the two-dimensional velocity model after dimensionality reduction to form spatio-temporal feature data includes:

[0015] Generating a -dimensional absolute positional encoding vector for each position of the two-dimensional velocity model after dimensionality reduction, and the absolute positional encoding vectors form an absolute positional encoding matrix, and the absolute positional encoding vector is represented by a sine function;

[0016] Performing convolution operations on the two-dimensional velocity model after dimensionality reduction using convolution kernels of different scales to generate multi-scale local feature maps, and splicing and reducing the dimensions of the multi-scale local feature maps to form a relative positional encoding matrix;

[0017] Splicing the absolute positional encoding matrix and the relative positional encoding matrix in the channel dimension, and merging them with the two-dimensional velocity model after dimensionality reduction to form spatio-temporal feature data.

[0018] Further, calculating the correlation between spatio-temporal feature data using an attention mechanism to obtain the first output of multiple parallel attention heads, and performing feature mapping on the fused first output to form a second output, including:

[0019] Using global attention to model the long-distance sound field interference of spatio-temporal feature data to obtain global features;

[0020] Using local window attention to restore the global features to a two-dimensional structure, dividing the two-dimensional structure into multiple local windows to obtain multiple local window features, and calculating attention independently within each local window;

[0021] Splicing the attention of all local windows back to the spatial dimension of the global features to obtain local window features;

[0022] Weightedly fusing the global features and the local window features to form a second output.

[0023] Further, dynamically fusing the global features and the local window features through a gating mechanism to weightedly fuse and form a second output.

[0024] Further, performing channel weighted fusion on the dimensionality-reduced features and the sound source information, including:

[0025] Normalizing the coordinate information of the sound source information;

[0026] Convert the frequency information of the sound source information to a logarithmic scale;

[0027] Map the sound source information to a high-dimensional vector through a lightweight MLP;

[0028] Generate a channel weight vector from the high-dimensional vector through a fully connected network;

[0029] Multiply the dimensionality-reduced features by the channel weight vector in each channel.

[0030] Compared with the prior art, the present application has at least the following beneficial effects:

[0031] The method of the embodiment of the present application reduces the prediction deviation caused by insufficient position information in complex terrains, avoids the problem of information forgetting in long-distance predictions, and improves the prediction accuracy of the global underwater acoustic field distribution.

[0032] It can accurately predict key details such as interference fringes of propagation loss and is suitable for complex deep-sea waveguide environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of a method for predicting deep-sea underwater acoustic propagation loss based on deep learning according to an embodiment of the present application;

[0034] Figure 2 It is a flowchart of processing a one-dimensional sound speed profile based on a trained deep learning network according to an embodiment of the present application;

[0035] Figure 3 It is a schematic diagram of the network structure of a trained deep learning network according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0036] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0037] A method for predicting deep - sea underwater acoustic propagation loss based on deep learning in an embodiment of this application is used for predicting deep - sea underwater acoustic propagation loss. Based on a trained deep - learning network, for the training process of the deep - learning network, first, a parabolic equation model is used to model the underwater acoustic propagation path. The parabolic equation model is a commonly used method for underwater acoustic propagation modeling, which simplifies the three - dimensional wave equation into a parabolic partial differential equation along the main propagation direction, thereby reducing the computational complexity and efficiently simulating long - distance acoustic propagation. The simulation data is divided into a training set, a validation set, and a test set. The training set is used for training the deep - learning network, the validation set is used to adjust the hyperparameters of the deep - learning network, and the test set is used to evaluate the performance of the final deep - learning network. The input data includes the sound - speed profile and the sound - source information. The sound - speed profile is the distribution of the sound - wave propagation speed in seawater, and the sound - source information includes the position and type of the underwater sound source. The output data is the underwater acoustic propagation loss distribution.

[0038] For example, the training dataset and the test dataset are simulated and generated in a channel with a continental slope. Its maximum range distance is 100 kilometers and the depth is 5 kilometers. The ocean surface is assumed to be a pressure - release boundary, and the seabed is an acoustic - elastic half - space, where the sound speed is 1550 m / s and the density is 1 g / cm³. The calculation of the underwater acoustic propagation loss distribution is to calculate and generate the real underwater acoustic propagation loss distribution using the RAM code. The RAM code uses the parabolic equation model (PE) method to obtain the results.

[0039] The dataset consists of three parts: a training set (Training dataset), a validation set (Validation set), and a test set (Testing dataset). The Argo buoy data from 2004 to 2008 for a total of 5 years is used. 1000 are selected from the Argo - SSP dataset in the Luzon Strait as the sound - speed profiles, and the longitudinal sampling number is 128 points. The sound - speed profile is replicated and extended 256 times horizontally to form a 128 * 256 array as the velocity model for training input. Each velocity model is different because the time and location of the buoy data records selected are different, and the ocean environment is different, so the calculated and estimated sound - speed profiles are different, and the calculated velocity models are different. Here, the velocity model refers to the description of the spatial distribution of the sound - wave propagation speed in the water medium.

[0040] Three sound - source points are randomly placed at the horizontal initial position and excited by sound sources of 10 Hz, 30 Hz, and 50 Hz respectively. Each velocity model is used to simulate the seismic records of three different source positions, and finally 3000 pairs of velocity model - underwater acoustic propagation loss distribution samples are obtained. These 3000 pairs of samples are divided into a training set and a validation set according to a ratio of 9:1. The test set uses the same method as the training set to generate 300 pairs of velocity model - underwater acoustic propagation loss distribution test sets, and these test sets are not included in the training set.

[0041] Based on the above-mentioned training set, validation set, and test set, a deep learning network is trained to obtain the trained deep learning network.

[0042] The trained deep learning network provided by the embodiments of the present application includes an encoder and a decoder. The specific structure and functions are described in the embodiments of a method for predicting deep-sea underwater acoustic propagation loss based on deep learning.

[0043] See Figure 1 As shown in the flowchart of a method for predicting deep-sea underwater acoustic propagation loss based on deep learning according to the embodiments of the present application. A method for predicting deep-sea underwater acoustic propagation loss based on deep learning provided includes:

[0044] S1 Obtain a one-dimensional sound speed profile;

[0045] A one-dimensional sound speed profile refers to the functional relationship between the sound speed and the ocean water depth. Usually, the depth is the independent variable and the sound speed is the dependent variable. It belongs to a cross-sectional graph that describes the variation law of the sound speed in the vertical direction. The shape of the sound speed profile is affected by the physical properties of the medium, such as temperature, salinity, density, etc.

[0046] The one-dimensional sound speed profile can be obtained by direct measurement methods, such as using a sound speed profiler, a hydrophone array, etc. For example, by arranging a hydrophone array in seawater, measuring the propagation time of sound waves, and combining the known sound source position, the sound speed profile is inverted. It can also be an inversion based on a physical model. For example, the ray acoustic model and the normal mode model are obtained by establishing a velocity model and performing inversion calculations using the velocity model.

[0047] S2 Process the one-dimensional sound speed profile based on the trained deep learning network and output the underwater acoustic propagation loss distribution;

[0048] See Figure 3 As shown in the schematic diagram of the network structure of a trained deep learning network. The trained deep learning network includes an encoder and a decoder. The encoder consists of a preprocessing module 1, an attention mechanism module 2, and a first convolution module 3. The decoder includes a deconvolution module 5, a hybrid module 6, and a second convolution module 7. A splicing module 4 is arranged between the encoder and the decoder. The input of the splicing module 4 is the output data of the encoder and the sound source information. The output data of the encoder and the sound source information are spliced and then output to the decoder. The decoder combines the output data of the encoder and the sound source information, re-integrates the features, and maximally retains the detailed content and the spatial feature distribution, and outputs the final underwater acoustic propagation loss distribution.

[0049] In some embodiments, the one-dimensional sound speed profile is processed based on the trained deep learning network and the underwater acoustic propagation loss distribution is output. See Figure 2The flowchart for processing the one-dimensional sound velocity profile based on the trained deep learning network is specifically as follows:

[0050] S21 Preprocess the one-dimensional sound velocity profile to obtain a two-dimensional velocity model, and convert the two-dimensional velocity model into spatio-temporal feature data with positional encoding;

[0051] This process is implemented by the preprocessing module 1. The one-dimensional sound velocity profile is expanded according to the horizontal distance to form a two-dimensional velocity model. For the convenience of understanding, the following will be described in combination with specific data feature dimensions. It can be understood that the specific data does not limit the protection scope. First, through the convolutional layer, the shape of the data of the two-dimensional velocity model is represented as (B, 1, 128, 256), where B represents the batch size, 1 represents the feature dimension, and 128 * 256 represents the size of the data in the input two-dimensional velocity model. The following data in the same format represents the same meaning and will not be explained one by one. The two-dimensional velocity model is downsampled to (B, 768, 8, 16). Then, a flattening operation (Flatten) is performed, and the downsampled two-dimensional velocity model is added with positional encoding to form spatio-temporal feature data. By adding positional encoding, the two-dimensional velocity model combines spatio-temporal features.

[0052] S22 Use the attention mechanism to calculate the correlation between spatio-temporal feature data, obtain the first output of multiple parallel attention heads, and perform feature mapping on the fused first output to form a second output;

[0053] The attention mechanism is to parallelly set multiple attention heads through the attention mechanism module 2. Each attention head calculates the correlation between spatio-temporal feature data through a query matrix, a key matrix, and a value matrix. Among them, the attention scores are calculated according to the query matrix and the key matrix to obtain the attention scores of multiple attention heads; the softmax activation function can be used to obtain the attention weights according to the attention scores and then perform weighted averaging on the value matrix to obtain the output of each attention head; finally, the outputs of all attention heads are connected in parallel. The attention mechanism module 2 sequentially passes the connected output of the attention heads through a linear layer, a GELU activation function layer, and a dropout layer for feature mapping. Multiple attention heads can simultaneously capture the details of the near-field sound field and the global features of long-distance propagation and the local details of the near-field sound field, which can improve the reconstruction accuracy of complex interference patterns.

[0054] S23 Perform a convolutional dimensionality reduction operation on the features of the second output to obtain dimensionality-reduced features;

[0055] It can be understood that the convolutional dimensionality reduction can be gradually reduced through one or more convolutions, so as to prepare for adding sound source information after dimensionality reduction.

[0056] S24 Perform channel weighted fusion on the dimensionality-reduced features and the sound source information;

[0057] The dimensionality-reduced features and the sound source features of the sound source information are concatenated through a concatenation module 4 and output to a decoder as an overall feature;

[0058] S25 maps the dimensionality-reduced features back to two-dimensional spatial information features, cross-uses convolution and transposed convolution in combination with the sound source information, and convolves the combined features to output the underwater acoustic propagation loss distribution.

[0059] A decoder is used to execute the above-mentioned step S24. In one embodiment, the decoder includes a transposed convolution module, a hybrid module 6, and a second convolution module 7. Between the encoder and the decoder is a concatenation module 4, and the input of the decoder is an overall feature obtained by concatenating the dimensionality-reduced features and the sound source features of the sound source information through the concatenation module 4.

[0060] The transposed convolution module in the decoder is responsible for upsampling the output of the concatenation module 4. After the intermediate feature data is input, it is upsampled to a data dimension of (B, 512, 8, 4) through two layers of deconvolution modules 5, and gradually mapped back to two-dimensional spatial information features. The hybrid module 6 realizes the fusion of the upsampled features by cross-using convolution and transposed convolution.

[0061] Maximize the retention of detail content and spatial feature distribution, and then use two layers of transposed convolution layers in the hybrid module 6 to increase the data dimension to (B, 8, 64, 256). The second convolution module 7 includes multiple convolutional feature reorganization layers to obtain the final underwater acoustic propagation loss distribution with a data dimension of (B, 1, 128, 256). Among them, the convolutional feature reorganization layer refers to a network layer that reorganizes or adjusts the input features through convolution.

[0062] The method in the embodiment of the present application reduces the prediction deviation caused by insufficient position information in complex terrain, avoids the problem of information forgetting in long-distance prediction, and improves the prediction accuracy of the global underwater acoustic field distribution.

[0063] In one embodiment, S21 preprocesses the one-dimensional sound speed profile to obtain a two-dimensional velocity model, and converts the two-dimensional velocity model into spatio-temporal feature data with position encoding. The position encoding uses multi-scale position encoding to improve the spatio-temporal feature learning ability of the sound speed layer, including absolute position encoding and relative position encoding.

[0064] Among them, for each position of the two-dimensional velocity model after dimensionality reduction processing, a dimensionality-reduced dimensional absolute position encoding vector is generated to form an absolute position encoding matrix, and the absolute position encoding vector is represented by a sine function;

[0065] For example: the two-dimensional velocity model after dimensionality reduction processing is two-dimensional matrix data , where denotes the real number space, denotes the size of the feature dimension, for each position , and are the coordinates in the depth and range directions respectively, generating an absolute position encoding vector with a dimension size of . The sine function used is:

[0066] ,

[0067] where , is the dimension index, is usually set to 256. The output is the absolute position encoding matrix .

[0068] Perform convolution operations on the two-dimensional velocity model after dimensionality reduction using convolution kernels of different scales to generate multi-scale local feature maps, and splice the multi-scale local feature maps and then reduce the dimension to form a relative position encoding matrix;

[0069] The relative position encoding input is the same two-dimensional matrix data as the above input , using convolution kernels of different scales, such as , and to perform convolution operations on the two-dimensional matrix data to generate multi-scale local feature maps: , is the feature map obtained by performing convolution operations with a convolution kernel of size , is the two-dimensional convolution operation with a convolution kernel of size , represents the size of the convolution kernel, and the number of output channels of each convolution kernel is . After splicing the feature maps of different scales, perform convolution dimensionality reduction through a convolution kernel of :

[0070] , is the feature map obtained by reducing the dimension of the spliced data, is the two-dimensional convolution operation with a convolution kernel of size , represents the splicing operation of the feature maps.

[0071] Obtain the relative position encoding matrix . Concatenate the absolute position encoding matrix and the relative position encoding matrix in the channel dimension to obtain , ​It is the feature map after splicing. In one embodiment, the weights of the absolute position encoding and the relative position encoding are adjusted through a learnable gating mechanism: , are the weights of the absolute position encoding and the relative position encoding, is the multi-scale position encoding obtained by weighting the absolute position encoding and the relative position encoding.

[0072] ,

[0073] Among them, is the Sigmoid function, is the element-wise multiplication. The absolute position encoding matrix and the relative position encoding matrix are concatenated in the channel dimension and merged with the two-dimensional velocity model after dimensionality reduction to form spatio-temporal feature data.

[0074] In the embodiment of the present application, through multi-scale position encoding: combining absolute position encoding for capturing the global spatial order and relative position encoding for modeling the local sound speed change relationship. It improves the representation ability of the spatial continuity of the sound speed profile, especially in complex terrains, and reduces the prediction deviation caused by insufficient position information.

[0075] In one embodiment, step S22 uses an attention mechanism to calculate the correlation between spatio-temporal feature data to obtain the first output of multiple parallel attention heads, and after fusing the first output, performs feature mapping to form the second output, including:

[0076] Using global attention to model the long-distance sound field interference of spatio-temporal feature data to obtain global features;

[0077] Using local window attention to restore the global features to a two-dimensional structure, dividing the two-dimensional structure into multiple local windows to obtain multiple local window features, and calculating attention independently within each local window;

[0078] Concatenating the attention of all local windows back to the spatial dimension of the global features to obtain local window features;

[0079] Weightedly fusing the global features and the local window features to form the second output.

[0080] The spatio-temporal feature data then applies a dropout operation, and the output data dimension is (B, 128, 768). The feature is summarized through the attention mechanism. The core idea is to simultaneously capture the global dependence relationship and local detail features of sound wave propagation through a multi-branch structure and dynamic interaction. It is divided into three parts: global attention is responsible for modeling long-distance sound field interference, such as the propagation path within a 100-kilometer range. Local window attention focuses on the details of the near-field sound field, such as the energy attenuation near the sound source. The cross-scale gating fusion module is used to dynamically adjust the weights of global features and local features.

[0081] Capture the global dependencies of long-distance sound wave propagation through multi-head self-attention. Introduce local window attention, divide the input into local windows, calculate the attention within the local windows, and capture the details of the near-field sound field. Achieve cross-scale interaction by dynamically fusing the outputs of global features and local features through a gating mechanism. At the same time, optimize the global features of long-distance propagation and the local details of the near-field sound field, significantly improving the reconstruction accuracy of complex interference patterns.

[0082] The above attention mechanism plus the Dropout layer constitute a residual processing block, and through a linear layer, a GELU activation function layer, and a dropout layer, perform a re-feature mapping on the features of the multi-head attention to improve the feature extraction ability of the multi-head attention, and obtain the final output data dimension of (B, 768, 8, 16).

[0083] In one embodiment, perform channel weighted fusion of the dimensionality-reduced features and the sound source information, including:

[0084] Normalize the coordinate information of the sound source information;

[0085] Convert the frequency information of the sound source information to a logarithmic scale;

[0086] Map the sound source information to a high-dimensional vector through a lightweight MLP;

[0087] Generate a channel weight vector through a fully connected network for the high-dimensional vector;

[0088] Multiply the dimensionality-reduced features by the channel weight vector in each channel.

[0089] Through the first convolution module 3 in the encoder, perform three-layer convolution dimensionality reduction operations to change the data shape from (B, 768, 8, 16) to (B, 768, 4, 8), then to (B, 768, 2, 4), and finally to (B, 1024, 1, 1), which is to prepare for adding the sound source information.

[0090] The sound source information is mainly coordinate information and frequency information. Normalize the coordinate information, and convert the frequency information to a logarithmic scale. Then map the sound source information to a high-dimensional vector through a lightweight MLP (Lightweight Multilayer Perceptron):

[0091] ,

[0092] where is the embedding dimension, is the feature map after the high-dimensional mapping of the sound source information, are the sound source positions in the depth kernel and range directions respectively, It represents the logarithmic processing of the sound source frequency, It represents that the size of the feature dimension after mapping is . The generation of the channel weight vector adopts a fully connected network:

[0093] ,

[0094] The obtained channel weight vector has a dimension of , and represent specific dimension values, is the activation function, and the value is scaled to [0,1] by the activation function. It represents a linear transformation. The channel weight vector is expanded to . The output obtained by multiplying channel by channel is:

[0095] ,

[0096] represents the feature output of the encoder part. This operation dynamically adjusts the activation intensity of each channel, highlighting the sound source features. The output of the encoder is the mixture of the dimensionality-reduced features and the sound source information with a shape of (B, 1024, 1, 1).

[0097] Through the adaptive weighted fusion of the sound source information: The sound source features (position, frequency) are passed through independent lightweight to generate a weight matrix, which is fused with the output of the encoder through channel weighting, rather than directly concatenating. Automatically adjusts the influence weight of the sound source information on the propagation loss distribution, avoids interference from invalid features, and improves the adaptability to multi-source scenarios.

[0098] Furthermore, in one embodiment, the final underwater acoustic propagation loss distribution with a data dimension of (B, 1, 128, 256) is obtained through the decoder.

[0099] In order to verify the effectiveness of the method of the embodiment of the present application, a large number of simulation experiments were carried out. The experimental results show that the method of the embodiment of the present application has high accuracy when predicting the deep-sea extremely low-frequency underwater acoustic propagation loss. Compared with traditional deep learning, the prediction error of the method provided by each embodiment of the present application is significantly reduced, it can well extract global information, significantly improves the ability to capture complex seismic patterns, and thus realizes more accurate and reliable values. In addition, in terms of key details such as interference fringes in long-distance propagation, these information can be better retained, and it has good adaptability to data of different sound source frequencies, grazing angles, and seabed topographies.

[0100] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

Claims

1. A deep - sea underwater acoustic propagation loss prediction method based on deep learning, characterized in that, Including: Obtain a one-dimensional sound speed profile; Process the one-dimensional sound speed profile based on a trained deep learning network and output the underwater acoustic propagation loss distribution, including: Preprocess the one-dimensional sound speed profile to obtain a two-dimensional velocity model, and convert the two-dimensional velocity model into spatio-temporal feature data with positional encoding; Use the attention mechanism to calculate the correlation between spatio-temporal feature data to obtain the first output of multiple parallel attention heads, and perform feature mapping on the fused first output to form a second output; Perform a convolution dimensionality reduction operation on the features of the second output to obtain dimensionality-reduced features; Perform channel weighted fusion on the dimensionality-reduced features and the sound source information, including: normalizing the coordinate information of the sound source information, converting the frequency information of the sound source information to a logarithmic scale, mapping the sound source information to a high-dimensional vector through a lightweight MLP, generating a channel weight vector through a fully connected network, and multiplying the dimensionality-reduced features by the channel weight vector in each channel; Map the dimensionality-reduced features back to two-dimensional spatial information features, cross-use convolution and transposed convolution to combine with the sound source information, and perform convolution on the combined features to output the underwater acoustic propagation loss distribution.

2. The method for predicting deep - sea underwater acoustic propagation loss based on deep learning according to claim 1, wherein, Preprocess the one-dimensional sound speed profile to obtain a two-dimensional velocity model, and convert the two-dimensional velocity model into spatio-temporal feature data with positional encoding, including: expanding the one-dimensional sound speed profile according to the horizontal distance to form a two-dimensional velocity model, and adding positional encoding to the two-dimensional velocity model after dimensionality reduction processing to form spatio-temporal feature data.

3. The method for predicting deep-sea underwater acoustic propagation loss based on deep learning according to claim 2, wherein Adding positional encoding to the two-dimensional velocity model after dimensionality reduction processing to form spatio-temporal feature data, including: Generate a -dimensional absolute position encoding vector for each position of the two-dimensional velocity model after dimensionality reduction. The absolute position encoding vectors form an absolute position encoding matrix, and the absolute position encoding vectors are represented by sine functions; Performing convolution operations on the dimensionality-reduced two-dimensional velocity model using convolution kernels of different scales to generate multi-scale local feature maps, splicing the multi-scale local feature maps and then reducing the dimension to form a relative positional encoding matrix; Splicing the absolute positional encoding matrix and the relative positional encoding matrix in the channel dimension, and merging with the dimensionality-reduced two-dimensional velocity model to form spatio-temporal feature data.

4. A method for predicting deep - sea underwater acoustic propagation loss based on deep learning according to claim 1, characterized in that, Using the attention mechanism to calculate the correlation between spatio-temporal feature data to obtain the first output of multiple parallel attention heads, and performing feature mapping on the fused first output to form a second output, including: Using global attention to model the long-distance sound field interference of spatio-temporal feature data to obtain global features; Using local window attention to restore the global features to a two-dimensional structure, dividing the two-dimensional structure into multiple local windows to obtain multiple local window features, and independently calculating attention within each local window; Splicing the attention of all local windows back to the spatial dimension of the global features to obtain local window features; Weightedly fusing the global features and the local window features to form a second output.

5. The method for predicting deep - sea underwater acoustic propagation loss based on deep learning according to claim 4, wherein, Weightedly fusing the global features and the local window features to form a second output through gated mechanism dynamic fusion.

Images