Deep-sea underwater sound propagation loss prediction method based on deep learning

Abstract

The invention belongs to the technical field of ocean exploration, and particularly relates to a deep-sea underwater sound propagation loss prediction method based on deep learning, and the method comprises the steps: 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 spatial-temporal characteristic data with position codes; calculating correlation between the spatial-temporal characteristic data by adopting an attention mechanism to obtain first outputs of a plurality of parallel attention heads, fusing the first outputs, and performing characteristic mapping to form a second output; performing convolution dimension reduction operation on the second output feature to obtain a dimension reduction feature; performing channel weighted fusion on the dimension reduction features and the sound source information; and mapping the dimension reduction features back to the two-dimensional space information features, combining the two-dimensional space information features with sound source information by using convolution and transpose convolution in a cross manner, and performing convolution on the combined features to output underwater sound propagation loss distribution. Key details such as interference fringes of propagation loss can be accurately predicted, and the method is suitable for a complex deep-sea waveguide environment.

Description

Technical Field

[0001] The present application belongs to the field of ocean exploration technology, and specifically relates to a method for predicting deep sea water acoustic propagation loss based on deep learning. Background Art

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

[0003] The embodiment of the present application provides a method for predicting deep sea water acoustic propagation loss based on deep learning, which solves the problem of high computational complexity and difficulty in adapting to complex and changeable marine environments.

[0004] This application is implemented as follows: A deep sea water acoustic propagation loss prediction method based on deep learning, comprising: Obtain a one-dimensional sound velocity profile; The one-dimensional sound velocity profile is processed based on the trained deep learning network, and the underwater acoustic propagation loss distribution is output; including: Preprocessing the one-dimensional sound velocity profile to obtain a two-dimensional velocity model, and converting the two-dimensional velocity model into spatiotemporal characteristic data with position coding; An attention mechanism is used to calculate the correlation between spatiotemporal feature data to obtain first outputs of multiple parallel attention heads, and the first outputs are fused and feature mapped to form a second output; Performing a convolution dimensionality reduction operation on the features of the second output to obtain a reduced dimensionality feature; Perform channel-weighted fusion of dimensionality reduction features and sound source information; The reduced-dimensional features are mapped back to two-dimensional spatial information features, and are combined with the sound source information using cross-convolution and transposed convolution. The combined features are convolved to output the distribution of underwater acoustic propagation loss.

[0005] Furthermore, the one-dimensional sound velocity profile is preprocessed to obtain a two-dimensional velocity model, and the two-dimensional velocity model is converted into spatiotemporal feature data with position coding, including: expanding the one-dimensional sound velocity profile according to the horizontal distance to form a two-dimensional velocity model, and performing dimensionality reduction processing on the two-dimensional velocity model and adding the position coding to form the spatiotemporal feature data.

[0006] Furthermore, the two-dimensional velocity model is processed by dimension reduction and position coding is added to form spatiotemporal feature data, including: Generate a dimensional absolute position encoding vector, the absolute position encoding vector forms an absolute position encoding matrix, and the absolute position encoding vector is represented by a sine function; The convolution operation is performed on the reduced-dimensional two-dimensional velocity model using convolution kernels of different scales to generate a multi-scale local feature map, and the multi-scale local feature map is concatenated and reduced in dimension to form a relative position encoding matrix; The absolute position encoding matrix and the relative position encoding matrix are spliced ​​in the channel dimension, and combined with the two-dimensional velocity model processed by the dimension reduction process to form spatiotemporal feature data.

[0007] Furthermore, the attention mechanism is used to calculate the correlation between the spatiotemporal feature data to obtain the first outputs of multiple parallel attention heads, and the first outputs are fused and feature mapped to form the second output, including: Global attention is used to model the long-distance sound field interference on the spatiotemporal feature data to obtain global features; The local window attention is used to restore the global features to a two-dimensional structure, and the two-dimensional structure is divided into multiple local windows to obtain multiple local window features, and the attention is calculated independently in each local window; Splice the attention of all local windows back to the spatial dimension of the global feature to obtain the local window feature; The global features and local window features are weighted and fused to form the second output.

[0008] Furthermore, the global features and the local window features are weightedly fused through dynamic fusion via a gating mechanism to form a second output.

[0009] Furthermore, the dimension reduction features are fused with the sound source information in a channel-weighted manner, including: Normalizing the coordinate information of the sound source information; Converting the frequency information of the sound source information into a logarithmic scale; Map the sound source information to a high-dimensional vector through a lightweight MLP; Generate channel weight vectors from high-dimensional vectors through a fully connected network; Multiply the reduced features with the channel weight vector at each channel.

[0010] Compared with the prior art, the present invention has at least the following beneficial effects: The method of the embodiment of the present application reduces prediction deviations caused by insufficient position information in complex terrain, avoids the problem of information forgetting in long-distance predictions, and improves the prediction accuracy of the global water acoustic field distribution.

[0011] 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

[0012] Figure 1 A flowchart of a method for predicting deep sea water acoustic propagation loss based on deep learning according to an embodiment of the present application; Figure 2 This is a flow chart of processing a one-dimensional sound velocity profile based on a trained deep learning network according to an embodiment of the present application; Figure 3 A schematic diagram of the network structure of a trained deep learning network according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the present application more clear, the present application is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and are not used to limit the present application.

[0014] A deep-sea acoustic propagation loss prediction method based on deep learning in an embodiment of the present application is used for deep-sea acoustic propagation loss prediction. It is based on a trained deep learning network. The training process of the deep learning network first uses a parabolic equation model to model the acoustic propagation path. The parabolic equation model is a commonly used method for modeling acoustic propagation. It 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 sound propagation. The simulation data is divided into a training set, a validation set, and a test set, wherein the training set is used for deep learning network training, 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 a sound velocity profile and sound source information. The sound velocity profile is the distribution of sound wave propagation speed in seawater, and the sound source information includes the location and type of the acoustic source. The output data is the distribution of acoustic propagation loss.

[0015] For example, the training and test datasets are simulated in a channel with a continental slope. It has a maximum range of 100 km and a depth of 5 km. 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 underwater acoustic propagation loss distribution is calculated using the RAM code to generate a realistic underwater acoustic propagation loss distribution, which uses the parabolic equation model (PE) method to obtain the results.

[0016] The dataset consists of three parts: training dataset, validation set and testing dataset. The Argo buoy data from 2004 to 2008 were used. 1,000 sound velocity profiles were selected from the Argo-SSP dataset in the Luzon Strait, and the number of longitudinal sampling points was 128. The sound velocity profile was replicated and expanded 256 times in the horizontal direction to form a 128*256 array as the velocity model for training input. Since the selected buoy data were recorded at different times and places and the ocean environment was different, the sound velocity profiles calculated for each velocity model were different, and the calculated velocity models were different. The velocity model here refers to the description of the spatial distribution of the propagation velocity of sound waves in the water medium.

[0017] Three sound source points were randomly placed at the horizontal initial position, and sound sources of 10Hz, 30Hz and 50Hz were used for excitation respectively. Each velocity model was used to simulate earthquake records at three different source positions, and finally 3000 pairs of velocity model-water acoustic propagation loss distribution samples were obtained. These 3000 pairs of samples were divided into training set and validation set in a ratio of 9:1. The test set used the same method as the training set to generate 300 pairs of velocity model-water acoustic propagation loss distribution test sets, which were not included in the training set.

[0018] Based on the above training set, validation set and test set, a deep learning network is trained to obtain a trained deep learning network.

[0019] The trained deep learning network provided in the embodiment of the present application includes an encoder and a decoder, and the specific structure and function are described in an embodiment of a deep sea water acoustic propagation loss prediction method based on deep learning.

[0020] See also Figure 1 As shown, a flowchart of a method for predicting deep sea water acoustic propagation loss based on deep learning according to an embodiment of the present application is provided. A method for predicting deep sea water acoustic propagation loss based on deep learning is provided, comprising: S1 obtains the one-dimensional sound velocity profile; A one-dimensional sound velocity profile is a function of the speed of sound as it changes with the depth of the ocean. It usually takes depth as the independent variable and the speed of sound as the dependent variable. It is a cross-sectional graph that describes the law of change of the speed of sound in the vertical direction. The shape of the sound velocity profile is affected by the physical properties of the medium, such as temperature, salinity, density, etc.

[0021] The one-dimensional sound velocity profile can be obtained by direct measurement, such as using a sound velocity profiler, a hydrophone array, etc. For example, by arranging a hydrophone array in the seawater, measuring the propagation time of the sound wave, and combining the known sound source position to invert the sound velocity profile. It can also be obtained by inversion based on a physical model, for example: the ray acoustic model and the simple normal wave model are obtained by establishing a velocity model and using the velocity model inversion calculation.

[0022] S2 processes the one-dimensional sound velocity profile based on the trained deep learning network and outputs the distribution of underwater acoustic propagation loss; See also Figure 3 A schematic diagram of the network structure of a trained deep learning network is shown, wherein the trained deep learning network includes an encoder and a decoder, wherein the encoder is composed of a preprocessing module 1, an attention mechanism module 2, and a first convolution module 3, and the decoder includes a deconvolution module 5, a mixing module 6, and a second convolution module 7. A splicing module 4 is provided between the encoder and the decoder, and the input of the splicing module 4 is the output data and the sound source information of the encoder. The output data of the encoder and the sound source information are spliced ​​and output to the decoder, which combines the output data of the encoder and the sound source information, and reintegrates the features, retains the detailed content and the spatial feature distribution to the maximum extent, and outputs the final underwater sound propagation loss distribution.

[0023] In some embodiments, the one-dimensional sound velocity profile is processed based on the trained deep learning network, and the water sound propagation loss distribution is output, see Figure 2 The flowchart shown is for processing a one-dimensional sound velocity profile based on a trained deep learning network, specifically including: S21 preprocesses the one-dimensional sound velocity profile to obtain a two-dimensional velocity model, and converts the two-dimensional velocity model into spatiotemporal characteristic data with position coding; This process is implemented through the preprocessing module 1, which expands the one-dimensional sound velocity profile according to the horizontal distance to form a two-dimensional velocity model. For ease of understanding, the following will be described in conjunction with the specific data feature dimensions. It can be understood that the specific data is not used as a limitation on the protection scope. First, through the convolution 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 have 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 is performed, and the reduced-dimensional two-dimensional velocity model is added to the position code to form spatiotemporal feature data. The two-dimensional velocity model is combined with spatiotemporal features by adding the position code.

[0024] S22 uses an attention mechanism to calculate the correlation between the spatiotemporal feature data, obtains first outputs of multiple parallel attention heads, and performs feature mapping on the first outputs after fusing them to form a second output; The attention mechanism is to set multiple attention heads in parallel through the attention mechanism module 2. Each attention head calculates the correlation between the spatiotemporal feature data through the query matrix, the key matrix and the value matrix. The attention score is 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 weight according to the attention score and then the value matrix is ​​weighted averaged to obtain the output of each attention head. Finally, the outputs of all attention heads are connected in parallel. The attention mechanism module 2 maps the outputs of the connected attention heads through the linear layer, the GELU activation function layer and the random inactivation layer in turn. Multiple attention heads can simultaneously capture the details of the close-range sound field, the global features of long-distance propagation and the local details of the close-range sound field, which can improve the reconstruction accuracy of complex interference patterns.

[0025] S23 performs a convolution dimension reduction operation on the feature of the second output to obtain a dimension reduction feature; It is understandable that convolution dimensionality reduction can be performed step by step through one or more convolutions to prepare for adding sound source information after dimensionality reduction.

[0026] S24 performs channel-weighted fusion of the dimension reduction features and the sound source information; The dimension reduction feature and the sound source feature of the sound source information are spliced ​​through the splicing module 4 and output as an overall feature to the decoder; S25 maps the reduced-dimensional features back to two-dimensional spatial information features, and cross-uses convolution and transposed convolution to combine them with the sound source information, and convolves the combined features to output the distribution of underwater sound propagation loss.

[0027] A decoder is used to perform the above step S24. In one embodiment, the decoder includes a transposed convolution module, a mixing module 6 and a second convolution module 7. Between the encoder and the decoder is a splicing module 4. The input of the decoder is an overall feature obtained by splicing the dimension reduction feature and the sound source feature of the sound source information through the splicing module 4.

[0028] The transposed convolution module in the decoder is responsible for upsampling the output of the splicing module 4. After the intermediate feature data is input, it is upsampled to a data dimension of (B, 512, 8, 4) through a two-layer deconvolution module 5, and gradually mapped back to two-dimensional spatial information features. The hybrid module 6 achieves the fusion of the upsampled features by cross-using convolution and transposed convolution.

[0029] The detailed content and spatial feature distribution are retained to the maximum extent, and then the two-layer transposed convolution layer in the hybrid module 6 is used to increase the data dimension to (B, 8, 64, 256). The second convolution module 7 includes multiple convolution feature reorganization layers to obtain the final underwater acoustic propagation loss distribution with a data dimension of (B, 1, 128, 256). Among them, the convolution feature reorganization layer refers to a network layer that reorganizes or adjusts the input features through convolution.

[0030] The above-mentioned method of 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 water acoustic field distribution.

[0031] In one embodiment, S21 preprocesses the one-dimensional sound velocity profile to obtain a two-dimensional velocity model, and converts the two-dimensional velocity model into spatiotemporal feature data with position coding. The position coding adopts multi-scale position coding to improve the ability to learn the spatiotemporal features of the sound velocity layer, including absolute position coding and relative position coding.

[0032] Among them, a dimension size of dimensional absolute position encoding vector to form an absolute position encoding matrix, wherein the absolute position encoding vector is represented by a sine function; For example: The two-dimensional velocity model processed by dimensionality reduction is a two-dimensional matrix data ,in represents the real number space, Indicates the feature dimension size, for each position , and They are the coordinates of the depth and range directions, generating a dimension size of The absolute position encoding vector , the sine function used is: , in , is the dimension index, Usually set to 256. The output is an absolute position encoding matrix .

[0033] The convolution operation is performed on the reduced-dimensional two-dimensional velocity model using convolution kernels of different scales to generate a multi-scale local feature map, and the multi-scale local feature map is concatenated and reduced in dimension to form a relative position encoding matrix; Relative position encoding input is the same as the above input two-dimensional matrix data , using convolution kernels of different scales, such as , and For two-dimensional matrix data Perform convolution operations to generate multi-scale local feature maps: , Is the size of The feature map obtained by convolution operation with the convolution kernel, Is the size of The convolution kernel performs a two-dimensional convolution operation. Represents the size of the convolution kernel, and the number of output channels of each convolution kernel is After concatenating feature maps of different scales, Convolution kernel is used for convolution dimensionality reduction: , It is the feature map obtained by reducing the dimension of the concatenated data. It's size The convolution kernel performs a two-dimensional convolution operation. Indicates the concatenation operation of feature maps.

[0034] Get the relative position encoding matrix The absolute position encoding matrix and the relative position encoding matrix are concatenated in the channel dimension to obtain , is the concatenated feature map. In one embodiment, the weights of the absolute position encoding and the relative position encoding are adjusted by a learnable gating mechanism: , is the weight of absolute position encoding and relative position encoding, It is a multi-scale position coding obtained by weighting the absolute position coding and the relative position coding.

[0035] , in, is the Sigmoid function, The absolute position encoding matrix and the relative position encoding matrix are concatenated in the channel dimension and combined with the two-dimensional velocity model processed by the dimension reduction process to form the spatiotemporal feature data.

[0036] In the embodiment of the present application, multi-scale position coding is used: absolute position coding is used to capture the global spatial order and relative position coding is used to model the local sound speed change relationship. The ability to characterize the spatial continuity of the sound speed profile is improved, especially in complex terrain, to reduce the prediction deviation caused by insufficient position information.

[0037] In one embodiment, step S22 uses an attention mechanism to calculate the correlation between spatiotemporal feature data to obtain first outputs of multiple parallel attention heads, and performs feature mapping on the first outputs to form a second output, including: Global attention is used to model the long-distance sound field interference on the spatiotemporal feature data to obtain global features; The local window attention is used to restore the global features to a two-dimensional structure, and the two-dimensional structure is divided into multiple local windows to obtain multiple local window features, and the attention is calculated independently in each local window; Splice the attention of all local windows back to the spatial dimension of the global feature to obtain the local window feature; The global features and local window features are weighted and fused to form the second output.

[0038] The spatiotemporal feature data is then subjected to a random dropout operation, and the output data dimension is (B, 128, 768). The features are summarized through the attention mechanism. The core idea is to capture the global dependencies and local detail features of sound wave propagation through multi-branch structures and dynamic interactions. It is divided into three parts: global attention is responsible for modeling long-distance sound field interference, such as a propagation path within a range of 100 kilometers. Local window attention focuses on close-range sound field details, such as energy attenuation near the sound source. The cross-scale gated fusion module is used to dynamically adjust the weights of global and local features.

[0039] The global dependency of long-distance sound wave propagation is captured through multi-head self-attention. Local window attention is introduced to divide the input into local windows, and attention is calculated within the local windows to capture the details of the close-range sound field. The output of global features and local features are dynamically fused through the gating mechanism to achieve cross-scale interaction. At the same time, the global features of long-distance propagation and the local details of the close-range sound field are optimized, significantly improving the reconstruction accuracy of complex interference patterns.

[0040] The above attention mechanism plus the Dropout layer constitute a residual processing block, and through the linear layer, GELU activation function layer and random inactivation layer, the features of the multi-head attention are remapped to improve the feature extraction ability of the multi-head attention, and the final output data dimension is (B, 768, 8, 16).

[0041] In one embodiment, the dimension reduction feature is fused with the sound source information by channel weighting, including: Normalizing the coordinate information of the sound source information; Converting the frequency information of the sound source information into a logarithmic scale; Map the sound source information to a high-dimensional vector through a lightweight MLP; Generate channel weight vectors from high-dimensional vectors through a fully connected network; Multiply the reduced features with the channel weight vector at each channel.

[0042] The first convolution module 3 in the encoder performs a three-layer convolution dimension reduction operation to reduce 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), in preparation for adding sound source information.

[0043] The sound source information mainly consists of coordinate information and frequency information. The coordinate information is normalized and the frequency information is converted to a logarithmic scale. Then, the sound source information is mapped to a high-dimensional vector through a lightweight MLP (Lightweight Multilayer Perceptron): , in is the embedding dimension, It is the feature map after high-dimensional mapping of the sound source information. are the sound source positions in the depth kernel and range directions, Indicates logarithmic processing of the sound source frequency. Indicates that the feature dimension size after mapping is The channel weight vector is generated using a fully connected network: , The obtained channel weight vector The dimension is , and Indicates the specific dimension value. is the activation function, and the value is scaled to [0,1] by the activation function. Represents a linear transformation. Expand to The output obtained by channel-by-channel multiplication is: , Represents the feature output of the encoder part. This operation dynamically adjusts the activation strength of each channel to highlight the sound source characteristics. The output of the encoder is a mixture of dimensionality reduction features and sound source information. The shape is (B, 1024, 1, 1).

[0044] Adaptive weighted fusion of sound source information: The sound source features (position, frequency) are integrated through independent lightweight Generate a weight matrix and perform channel-weighted fusion with the encoder output instead of direct concatenation. Automatically and dynamically adjust the weight of the impact of sound source information on the distribution of propagation loss to avoid invalid feature interference and improve adaptability to multi-sound source scenarios.

[0045] Furthermore, in one embodiment, a final underwater acoustic propagation loss distribution with a data dimension of (B, 1, 128, 256) is obtained through a decoder.

[0046] 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 in predicting the propagation loss of deep-sea ultra-low frequency underwater acoustics. Compared with traditional deep learning, the prediction error of the method provided by each embodiment of the present application is significantly reduced, global information can be extracted well, and the ability to capture complex seismic patterns is significantly improved, thereby achieving more accurate and reliable numerical values. In addition, in terms of key details such as interference fringes of long-distance propagation, this information can be better retained, and it has good adaptability to data of different sound source frequencies, grazing angle conditions and seabed topography.

[0047] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A method for predicting deep sea water acoustic propagation loss based on deep learning, characterized in that: include: Obtain a one-dimensional sound velocity profile; The one-dimensional sound velocity profile is processed based on the trained deep learning network, and the underwater acoustic propagation loss distribution is output; including: Preprocessing the one-dimensional sound velocity profile to obtain a two-dimensional velocity model, and converting the two-dimensional velocity model into spatiotemporal characteristic data with position coding; An attention mechanism is used to calculate the correlation between spatiotemporal feature data to obtain first outputs of multiple parallel attention heads, and the first outputs are fused and feature mapped to form a second output; Performing a convolution dimensionality reduction operation on the features of the second output to obtain a reduced dimensionality feature; Perform channel-weighted fusion of dimensionality reduction features and sound source information; The reduced-dimensional features are mapped back to two-dimensional spatial information features, and are combined with the sound source information using cross-convolution and transposed convolution. The combined features are convolved to output the distribution of underwater acoustic propagation loss.

2. The method for predicting deep sea water acoustic propagation loss based on deep learning according to claim 1, characterized in that: The one-dimensional sound velocity profile is preprocessed to obtain a two-dimensional velocity model, and the two-dimensional velocity model is converted into time-space characteristic data with position coding, including: expanding the one-dimensional sound velocity profile according to the horizontal distance to form a two-dimensional velocity model, and performing dimension reduction processing on the two-dimensional velocity model and adding the position coding to form the time-space characteristic data.

3. A method for predicting deep sea water acoustic propagation loss based on deep learning according to claim 2, characterized in that: After the dimensionality reduction processing of the two-dimensional velocity model, position coding is added to form spatiotemporal feature data, including: Generate a dimensional absolute position encoding vector, the absolute position encoding vector forms an absolute position encoding matrix, and the absolute position encoding vector is represented by a sine function; The convolution operation is performed on the reduced-dimensional two-dimensional velocity model using convolution kernels of different scales to generate a multi-scale local feature map, and the multi-scale local feature map is concatenated and reduced in dimension to form a relative position encoding matrix; The absolute position encoding matrix and the relative position encoding matrix are spliced ​​in the channel dimension, and combined with the two-dimensional velocity model processed by the dimension reduction process to form spatiotemporal feature data.

4. A method for predicting deep sea water acoustic propagation loss based on deep learning according to claim 1, characterized in that: The attention mechanism is used to calculate the correlation between the spatiotemporal feature data to obtain the first outputs of multiple parallel attention heads, and the first outputs are fused and feature mapped to form the second output, including: Global attention is used to model the long-distance sound field interference on the spatiotemporal feature data to obtain global features; The local window attention is used to restore the global features to a two-dimensional structure, and the two-dimensional structure is divided into multiple local windows to obtain multiple local window features, and the attention is calculated independently in each local window; Splice the attention of all local windows back to the spatial dimension of the global feature to obtain the local window feature; The global features and local window features are weighted and fused to form the second output.

5. A method for predicting deep sea water acoustic propagation loss based on deep learning according to claim 4, characterized in that: The global features and local window features are weightedly fused through a gating mechanism to form the second output.

6. A method for predicting deep sea water acoustic propagation loss based on deep learning according to claim 1, characterized in that: The dimension reduction features are fused with the sound source information in a channel-weighted manner, including: Normalizing the coordinate information of the sound source information; Converting the frequency information of the sound source information into a logarithmic scale; Map the sound source information to a high-dimensional vector through a lightweight MLP; Generate channel weight vectors from high-dimensional vectors through a fully connected network; Multiply the reduced features with the channel weight vector at each channel.

Images