Regional ocean sound propagation field millisecond-level prediction method based on deep neural network

Abstract

The invention relates to a regional ocean sound propagation field millisecond-level prediction method based on a deep neural network, and the method comprises the steps: constructing two factors which have the greatest influence on a sound propagation field for a target region: a sound velocity profile and a terrain; an environment information data set of an area sound propagation field is constructed by constructing an annual sound velocity profile of a target area node and building a large number of two-dimensional terrain models by extending around the node by 360 degrees, and the sound propagation fields of the area under different environment conditions are calculated by using a traditional sound field calculation model to serve as a training set; feature simplification parameters of the sound velocity profile and the terrain in each environment sample are extracted to serve as guide information of a corresponding sound propagation field, and a deep neural network is built to learn distribution rules of the sound propagation fields under different environment conditions in the area; and finally, the trained model can quickly predict the corresponding sound propagation field in the region according to different environmental condition characteristic batches.

Description

Technical Field

[0001] This invention belongs to the fields of marine engineering and underwater acoustics engineering, as well as the fields of marine acoustics and deep learning technology. It relates to a millisecond-level prediction method for regional marine sound propagation fields based on deep neural networks. Specifically, it is a rapid prediction method for regional sound propagation fields based on regional environmental information feature extraction and deep neural networks. It is suitable for dynamic and rapid sound field modeling on platforms or in simulation environments, and provides an innovative solution for rapid sound field calculation, design and optimization. Background Technology

[0002] Rapid and accurate prediction of underwater sound propagation fields is crucial for sonar detection performance evaluation and acoustic concealment calculation. Classical methods for calculating underwater sound propagation are mostly based on the wave equation theoretical framework; however, traditional numerical calculations, including ray tracing, finite element method, and parabolic equation methods, are time-consuming and difficult to meet the real-time requirements of underwater sound field calculations in modern engineering. In recent years, with the development of machine learning, data-driven methods have become another effective approach to solving traditional underwater sound field modeling problems; however, they suffer from overly simplistic scenarios and difficulty in extending to practical engineering applications. A detailed analysis of these two types of methods for calculating underwater sound fields is as follows: Most classical methods for underwater sound wave propagation, based on physical field models, are derived from the sound wave equation. However, closed-form solutions to the wave equation are often difficult to obtain in marine environments. Therefore, the wave equation typically employs multiple approximate solutions, most of which can be considered variations of four basic methods: the ray method, normal modes, parabolic equations, and wave number integrals. These models generally require a large number of parameters, involving multiple factors such as sound velocity profiles, seabed topography, and ocean current conditions. The complexity of parameter settings increases the difficulty of using these models. Moreover, although computational methods based on acoustic field physical models are accurate, they usually require long computation times and high computational resource requirements, which limits their efficiency in simulated combat scenarios.

[0003] Furthermore, leveraging the rapid development of deep learning, data-driven end-to-end mapping of sound fields and physical neural networks have emerged as fast and efficient methods for calculating underwater sound fields in recent years. Compared to traditional methods, data-driven approaches eliminate the need to solve complex equations, efficiently simulating sound propagation processes by learning patterns from prior data. However, physical neural networks lack the convenience of end-to-end prediction. Currently, other data-driven underwater sound propagation field prediction methods suffer from limitations such as overly simplistic and idealized scenarios, limited generalization ability, insufficient adaptability to complex and dynamic marine environments, and difficulty in extending to practical engineering applications.

[0004] Therefore, there is an urgent need to find new computational methods that can rapidly generate underwater acoustic propagation fields under corresponding environmental conditions within an end-to-end framework, thereby providing support for the rapid assessment of sonar detection and communication capabilities. Summary of the Invention

[0005] Technical problems to be solved To overcome the shortcomings of existing technologies, this invention proposes a millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks. Current research using deep learning and data-driven approaches to solve traditional underwater acoustic propagation calculations fails to address lightweight and rapid prediction, resulting in overly simplistic scenarios, limited generalization capabilities, and insufficient adaptability to complex and dynamic marine environments. This research addresses both lightweight design and practical application by proposing a method based on regional environmental information feature extraction, using a deep neural network model to predict the corresponding acoustic propagation field. By efficiently learning the acoustic propagation loss patterns under diverse two-dimensional topography and sound velocity profiles, rapid end-to-end prediction is achieved, significantly improving real-time performance and robustness in practical engineering applications.

[0006] Technical solution A millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks is characterized by: constructing an environmental information dataset for the regional acoustic propagation field by establishing year-round sound velocity profile data for the target area and building a two-dimensional terrain model extending 360° around the location of the sound velocity profile; using a traditional sound field calculation model to calculate the sound propagation field under different environmental conditions in the area as a training set; then extracting simplified parameters of the sound velocity profile and terrain features from each environmental sample as guiding information for the corresponding sound propagation field; building a deep neural network to learn the distribution pattern of the sound propagation field under different environmental conditions in the area; and finally, the trained model rapidly predicts the corresponding sound propagation field in batches for different environmental conditions within the target area. The specific steps are as follows: Step 1: Construction of regional environmental dataset: Using the location information of the annual sound velocity profile of the target area as nodes, a two-dimensional terrain model is established by extending the distance in a 360° direction, and a dataset of environmental information covering the entire year of the region is constructed. The annual environmental information dataset includes an annual sound velocity profile dataset and a topographic profile dataset; Step 2: Extract features from the annual sound velocity profile dataset to obtain M environmental information sound velocity profile features; Step 3: Extract features from the terrain profile set to obtain N environmental information terrain features; Step 4: Construct the sound propagation field dataset: Input the environmental information dataset generated in Step 1 into the traditional sound field calculation model, and calculate the sound propagation loss matrix K*K under the corresponding sound speed profile and terrain conditions; the calculation results are used as the training set of the model, and the environmental information features extracted in Steps 2 and 3 are assigned the corresponding sound propagation field as feature labels; in addition, the terrain vector corresponding to each sound field calculation will also be recorded as additional conditional information for model training. Step 5: Build and train a deep neural network model: Based on the sound speed profile features of dimension M and the seabed topographic features of dimension N extracted in steps 2 and 3 respectively, build an end-to-end deep neural network model to directly predict the high-resolution sound propagation loss field from environmental parameters. The model has a size of K*K. The designed deep neural network consists of three parts: a fully connected layer, an upsampling module, and an output layer. The input is a feature vector of dimension M+N, and the output is a single-channel K*K normalized sound propagation loss matrix. Fully connected layer: Receives M+N dimensional input features, maps them to high-dimensional feature vectors through linear transformation, and then reshapes them into an initial high-dimensional feature map, which serves as the input to the subsequent upsampling module, realizing the effective encoding of low-dimensional environmental parameters into the latent sound field representation; The upsampling module connecting the fully connected layers uses multiple residual-enhanced transposed convolutional blocks to upsample the high-dimensional feature map step by step. Each block combines transposed convolution upsampling feature refinement, batch normalization and LeakyReLU activation, and implements residual connections through matched shortcut paths to alleviate deep gradient vanishing. The output layer uses a single-channel transposed convolution. The output of the upsampling module is processed by the single-channel transposed convolution and the Sigmoid activation output to obtain a normalized acoustic propagation loss matrix of size (k*k). The acoustic propagation loss matrix expresses a fast and efficient end-to-end prediction from physical parameters to high-resolution propagation attenuation distribution. Step 6: Train the constructed network to learn the dataset within the target region and learn the distribution law of sound propagation field under different environmental characteristics; the loss function is designed as the root mean square error between the prediction result of the calculation model and the sound propagation field result calculated by the traditional sound field calculation model.

[0007] The annual sound velocity profile of the target area in step 1 is obtained from an existing dataset or through field measurements tailored to the needs of the target area.

[0008] The environmental information sound velocity profile features in step 2 include: 1) Initial sound velocity of the sound velocity mixing layer in the sound velocity profile; 2) Initial sound velocity of the transition layer; 3) Depth value at the starting position of the transition layer; 4) Sound velocity gradient in the transition layer; 5) Minimum sound speed in the sound velocity profile; 6) Depth at the point of minimum sound speed in the sound velocity profile, i.e., the depth value of the deep-sea acoustic channel axis; 7) Temperature gradient value of the isothermal layer; 8) Maximum sound speed value in the sound speed profile 9) There are nine characteristics of the depth at the maximum sound speed value in the sound speed profile.

[0009] The environmental information terrain features in step 3 include: 1) Maximum depth: The maximum depth value in the terrain profile; 2) Minimum depth: The minimum depth value in the topographic profile; 3) Average depth: The average of all depth values ​​in the topographic profile, reflecting the overall depth level of the profile; 4) Median depth: The median of the topographic profile depth data, reflecting the middle value of the depth distribution. Compared with the mean, it is less sensitive to outliers. 5) Location of the mountain peak: The horizontal distance between the location of the mountain peak and the starting point in the topographic profile; 6) Maximum absolute slope: The maximum absolute value of the slope between adjacent points in a topographic profile, reflecting the steepest change in terrain; 7) Average slope: The average of all slopes in the topographic profile, reflecting the overall tilt trend of the terrain; 8) Flatness: The standard deviation of the depth values ​​of the terrain profile, reflecting the flatness of the terrain; 9) Number of oscillations: The number of times the slope sign changes in the topographic profile, reflecting the frequency of topographic undulations or fluctuations; 10) Skewness: Indicates whether the overall topographic profile is from shallow to deep or from deep to shallow, and the degree thereof; 11) Kurtosis: Kurtosis of topographic profile depth data measures the "sharpness" of the depth distribution.

[0010] The interval between adjacent points in the topographic profile of step 3 is 1 kilometer.

[0011] The formula for calculating the flatness value in step 3 is as follows: ,in This represents the number of depth values ​​in the terrain profile. The first in the topographic profile One depth value; The average depth of the terrain profile.

[0012] The formula for calculating the skewness value in step 3 is as follows: ,in This represents the number of depth values ​​in the terrain profile. The first in the topographic profile One depth value; The average depth of the topographic profile; This represents the standard deviation of the topographic profile.

[0013] The formula for calculating the kurtosis value in step 3 is as follows: ,in This represents the number of depth values ​​in the terrain profile. The first in the topographic profile One depth value; The average depth of the topographic profile; This represents the standard deviation of the topographic profile.

[0014] The traditional sound field calculation model uses Bellhop.

[0015] In step 6, the specific calculation range of the loss function is designed according to the requirements, targeting either the global acoustic propagation loss matrix or the portion above the seabed topography. The loss function is calculated by taking the root mean square error of the two at a fixed reception depth. Alternatively, different weights can be added to these two different error calculation methods to unify them as the loss function. The number of iterations and the batch_size hyperparameter are set according to the dataset size and training hardware conditions.

[0016] This method integrates multi-dimensional environmental information, including sound velocity profile features and terrain features, as input parameters, and can directly predict the sound propagation field corresponding to different environmental information features in millisecond-level batch end-to-end.

[0017] This method is based on big data training of the model. It does not require understanding and analyzing the complex laws of ocean sound propagation. It only uses existing sound field calculation data. The model analyzes and learns the distribution law of sound field between different sound speed profiles and topography, so as to realize the rapid prediction of sound propagation field under different environmental conditions within the coverage area. It has low computational load, simple process, and relatively accurate results.

[0018] Beneficial effects This invention proposes a millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks. It constructs an environmental information dataset for the regional acoustic propagation field by considering the two factors that have the greatest impact on the acoustic propagation field in the target area: sound velocity profile and topography. This dataset is built by processing the annual sound velocity profiles of nodes in the target area and establishing numerous two-dimensional topographic models extending 360° around the nodes. Traditional sound field calculation models are used to calculate the acoustic propagation field under different environmental conditions in the region as a training set. Then, simplified parameters of the sound velocity profile and topography features in each environmental sample are extracted as guiding information for the corresponding acoustic propagation field. A deep neural network is then built to learn the distribution patterns of the acoustic propagation field under different environmental conditions in the region. Finally, the trained model can quickly predict the corresponding acoustic propagation field in batches for different environmental conditions within the region.

[0019] The beneficial effects of this invention are specifically reflected in: (1) This invention provides a practically operable method for the real-time calculation of underwater sound propagation field in engineering applications. Taking the experiment as an example, the underwater sound propagation field matrix of 128*128 dimensions in the corresponding sea area is predicted by extracting 17 parameters from the features of the sound velocity profile and the topographic profile. Compared with the traditional sound field calculation model, the calculation speed is faster and the sound propagation field can be predicted in batches at the millisecond level. The speed is improved by up to 300 times and the results are more accurate.

[0020] (2) This invention uses sound velocity profile features and terrain features as guiding conditions and uses the sound propagation loss matrix for training. After training, the corresponding sound propagation loss matrix can be directly generated according to the required sound velocity profile features and terrain features, realizing end-to-end sound field modeling and prediction. Its advantage is that the trained model is simple to use. Compared with the traditional sound field calculation model, it does not require professional and complex underwater acoustic knowledge and cumbersome usage rules. Attached Figure Description

[0021] Figure 1 A flowchart of a method for predicting regional sound propagation fields based on deep neural networks. Figure 2 Area selection and terrain conditions Figure 3: Environmental information feature extraction. (a) Sound velocity profile feature extraction; (b) Topographic profile feature extraction; Figure 4 The model structure diagram consists of a fully connected layer, an upsampling module, and a final output layer.

[0022] Figure 5 The error distribution of the prediction results on 1000 test samples after the model training is completed. From left to right, the comparison is as follows: the overall RMSE of the test set bellhop calculation results and the model prediction results; the comparison of the test set bellhop calculation results and the model prediction results at a receiving depth of 360m; and the comparison of the test set bellhop calculation results and the model prediction results at a receiving depth of 960m.

[0023] Figure 6 The following is a display of the prediction results of the model for four test samples. The prediction results for the four test samples are shown in the following columns from left to right: the first column shows the calculation results of the Bellhop model, the second column shows the prediction results of the deep neural network based on environmental information features, the third column shows the absolute error between the two, and the fourth column shows a comparison of the sound source depth error curves. Detailed Implementation

[0024] The present invention will now be further described in conjunction with the embodiments and accompanying drawings: A millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks is characterized by: constructing an environmental information dataset for the target region that includes the two factors that have the greatest impact on the acoustic propagation field: sound velocity profile and topography. This dataset is built by processing the annual sound velocity profile of nodes in the target region and establishing a large number of two-dimensional topographic models extending 360° around the nodes. The acoustic propagation field under different environmental conditions in the region is calculated using a traditional sound field calculation model as a training set. Then, the simplified parameters of the sound velocity profile and topography features in each environmental sample are extracted as guiding information for the corresponding acoustic propagation field. A deep neural network is built to learn the distribution law of the acoustic propagation field under different environmental conditions in the region. Finally, the trained model can quickly predict the corresponding acoustic propagation field in batches for different environmental conditions in the region.

[0025] The entire method and process can be divided into the following five steps: Step 1, Regional Environmental Dataset Construction: Use existing datasets or conduct field measurements to obtain the annual sound velocity profile of the target area. Use the sound velocity profile location information as nodes to extend the distance in a 360° direction to build a two-dimensional terrain model and construct a dataset that can cover the environmental features of the area. Step 2, Environmental Information Sound Velocity Profile Feature Extraction: Based on the target area environmental information dataset constructed in Step 1, the following features are extracted from the sound velocity profile: 1) Initial sound velocity of the sound velocity mixing layer in the sound velocity profile; 2) Initial sound velocity of the transition layer; 3) Depth value at the starting position of the transition layer; 4) Sound velocity gradient in the transition layer; 5) Minimum sound speed in the sound velocity profile; 6) Depth at the point of minimum sound speed in the sound velocity profile, i.e., the depth value of the deep-sea acoustic channel axis; 7) Temperature gradient value of the isothermal layer; 8) Maximum sound speed value in the sound speed profile 9) There are nine characteristics of the depth at the maximum sound speed value in the sound speed profile.

[0026] Step 3: Environmental Information Terrain Feature Extraction: Based on the target area environmental information dataset constructed in Step 1, the following features are extracted from the terrain profile: 1) Maximum depth: The maximum depth value in the terrain profile; 2) Minimum depth: The minimum depth value in the topographic profile; 3) Average depth: The average of all depth values ​​in the topographic profile, reflecting the overall depth level of the profile; 4) Median depth: The median of the topographic profile depth data, reflecting the middle value of the depth distribution. Compared with the mean, it is less sensitive to outliers. 5) Location of the mountain peak; 6) Maximum absolute slope: The maximum absolute value of the slope between adjacent points in a topographic profile, reflecting the steepest change in terrain; 7) Average slope: The average of all slopes in the topographic profile, reflecting the overall tilt trend of the terrain; 8) Flatness: The standard deviation of the depth values ​​of the terrain profile, reflecting the flatness of the terrain; Flatness is represented by flatness values:

[0027] in The number of depth values ​​in the topographic profile; subscript. Indicates the order of the terrain sequence species; The average depth of the topographic profile; The first in the topographic profile One depth value; 9) Number of oscillations: The number of times the slope sign changes in the topographic profile, reflecting the frequency of topographic undulations or fluctuations; 10) Skewness: Indicates whether the overall topographic profile is from shallow to deep or from deep to shallow, and the degree thereof; The skewness is expressed as a skewness value:

[0028] in The number of depth values ​​in the topographic profile; subscript. Indicates the order of the terrain sequence species; The average depth of the topographic profile; The first in the topographic profile One depth value; The standard deviation of the topographic profile; 11) Kurtosis: Kurtosis of topographic profile depth data measures the "sharpness" of the depth distribution. (High kurtosis indicates a more concentrated distribution, while low kurtosis indicates a flatter distribution).

[0029] The kurtosis is expressed as a kurtosis value:

[0030] in The number of depth values ​​in the topographic profile; subscript. Indicates the order of the terrain sequence species; The average depth of the topographic profile; The first in the topographic profile One depth value; The standard deviation of the topographic profile; Step four, construct the sound propagation field dataset: Input the environmental information dataset generated in step one into the traditional sound field calculation model Bellhop. Except for the sound velocity profile and terrain settings, other parameters are set according to the actual situation of the target area and individual choices. Calculate the sound propagation loss matrix under the corresponding sound velocity profile and terrain conditions. These calculation results serve as the model's training set and are assigned corresponding labels based on the environmental information features extracted in steps two and three. Furthermore, the terrain vector corresponding to each sound field calculation will also be recorded to serve as additional conditional information for model training.

[0031] Step 5, build and train a deep neural network model: combine the sound speed profile and terrain feature latitude extracted from the environmental information, and design the network according to the size of the calculated and saved sound propagation loss matrix. The network consists of a fully connected layer, an upsampling module, and an output layer.

[0032] Based on the sound velocity profile features of dimension M and the seabed topography features of dimension N extracted in steps 2 and 3 respectively, an end-to-end deep neural network model is constructed to directly predict the high-resolution sound propagation loss field from environmental parameters, with a size of K*K. The designed deep neural network consists of three parts: a fully connected layer, an upsampling module, and an output layer. The input is a feature vector of dimension M+N, and the output is a single-channel K*K normalized sound propagation loss matrix. Fully connected layer: Receives M+N dimensional input features, maps them to high-dimensional feature vectors through linear transformation, and then reshapes them into an initial high-dimensional feature map, which serves as the input to the subsequent upsampling module, realizing the effective encoding of low-dimensional environmental parameters into the latent sound field representation; The upsampling module connecting the fully connected layers uses multiple residual-enhanced transposed convolutional blocks to upsample the high-dimensional feature map step by step. Each block combines transposed convolution upsampling feature refinement, batch normalization and LeakyReLU activation, and implements residual connections through matched shortcut paths to alleviate deep gradient vanishing. The output layer uses a single-channel transposed convolution. The output of the upsampling module is processed by the single-channel transposed convolution and the Sigmoid activation output to obtain a normalized acoustic propagation loss matrix of size (k*k). The acoustic propagation loss matrix expresses a fast and efficient end-to-end prediction from physical parameters to high-resolution propagation attenuation distribution. The neural network model of the embodiment is shown in [example]. Figure 4As shown, taking Batch_size=1 as an example, the fully connected layer is a linear layer that expands the input dimension from 16 to 2048. Subsequently, a reshaping operation converts the expanded multidimensional environmental feature output into 4×4×128. The following upsampling module performs upsampling processing on four residual augmented transposed convolutional block sequences. The internal structure of each residual augmented transposed convolutional block sequence is main path → shortcut path → output set. The main path operation involves upsampling and channel transformation using a transposed convolution (kernel size 4, stride 2, padding 1), followed by batch normalization and LeakyReLU activation. Then, a standard convolution (kernel size 3, stride 1, padding 1) further refines the features, followed by batch normalization again. The shortcut path operation uses a transposed convolution (kernel size 4, stride 2, padding 1) and batch normalization to directly pass input features and implement residual connections, mitigating the gradient vanishing problem in deep networks. Output activation involves applying LeakyReLU activation after residual summation to ensure non-linear transformation. The final output layer is a single-channel transposed convolution (kernel size 4, stride 2, padding 1), followed by activation using the Sigmoid activation function.

[0033] In this embodiment, the feature dimensions of the sound velocity profile and terrain extracted from the environmental information are combined. Based on the size of the calculated and stored sound propagation loss matrix, a network is designed, consisting of a fully connected layer, an upsampling module, and an output layer. First, the fully connected layer maps the multi-dimensional input features (the sound velocity profile and seabed terrain features extracted in steps two and three) into a high-dimensional feature map of the corresponding sound propagation field. The upsampling module consists of multiple sequences of residual-enhanced transposed convolutional blocks. These blocks upsample the high-dimensional feature map step-by-step. Each block internally combines transposed convolution upsampling feature refinement, batch normalization, and LeakyReLU activation, and uses matched shortcut paths to achieve residual connections to alleviate deep gradient vanishing. Finally, the output layer outputs a normalized sound propagation loss matrix of the corresponding size through a single-channel transposed convolution and sigmoid activation. This method enables fast and efficient end-to-end prediction of propagation attenuation distribution from physical parameters to high resolution, while maintaining training stability and physical interpretability. The loss function can be designed according to specific needs. It can be designed as the root mean square error between the prediction result of the calculation model and the sound propagation field result calculated by the traditional sound field calculation model. The specific calculation range can be designed to target the global sound propagation loss matrix or the part above the seabed topography. Alternatively, it can be designed to calculate the root mean square error of the two at a fixed receiving depth as the loss function. Different weights can also be added to these two different error calculation methods to unify them as a loss function.

[0034] Step 6: Train the constructed network to learn the dataset within the target region and learn the distribution law of sound propagation field under different environmental characteristics; the loss function is designed as the root mean square error between the prediction result of the calculation model and the sound propagation field result calculated by the traditional sound field calculation model.

[0035] For specific data, taking this experiment as an example, the selected sea area is the Kuroshio Current region in the western Pacific Ocean, between 124.5° and 126.5° east longitude and 17° and 19° north latitude. (See...) Figure 2 By analyzing temperature, salinity, and depth data from 882 locations in the sea area throughout 2021, a total of 321,930 annual sound velocity profiles were obtained. Subsequently, a two-dimensional topographic profile model was established by extending 130 km at a random 360° angle around these 882 locations each day, resulting in 321,930 environmental samples containing both topographic and sound velocity profile data. Steps two, three, and four were then performed, generating a training set of 279,845 training samples. This experiment extracted 16-dimensional features from the environmental information, calculated a sound propagation field with a data size of 128*128, and designed a deep neural network consisting of a fully connected layer, an upsampling module, and a final output layer. Figure 4 Taking Batch_size=1 as an example, the specific structure is as follows: the fully connected layer is a linear layer that expands the input dimension from 16 to 2048. Subsequently, a reshaping operation converts the expanded multidimensional environmental feature output into 4×4×128. The following upsampling module performs upsampling processing on four residual augmented transposed convolutional block sequences. The internal structure of each residual augmented transposed convolutional block sequence is main path → shortcut path → output set. The main path operation involves upsampling and channel transformation using a transposed convolution (kernel size 4, stride 2, padding 1), followed by batch normalization and LeakyReLU activation. Then, a standard convolution (kernel size 3, stride 1, padding 1) further refines the features, followed by batch normalization again. The shortcut path operation uses a transposed convolution (kernel size 4, stride 2, padding 1) and batch normalization to directly pass input features and implement residual connections, mitigating the gradient vanishing problem in deep networks. Output activation involves applying LeakyReLU activation after residual summation to ensure non-linear transformation. The final output layer is a single-channel transposed convolution (kernel size 4, stride 2, padding 1), followed by activation using the Sigmoid activation function. The loss function used in training is the root mean square error of the global matrix between the sound propagation field predicted by the model and the sound propagation field calculated by the Bellhop model. The batch size is 512, and the training is iterated for 300 generations. The trained model is then tested on the validation set. For details, see [link to details]. Figure 5 as follows: The error distribution of the prediction results on 1000 test samples after the model training is completed is as follows, from left to right: comparison of the overall RMSE of the test set bellhop calculation results and the model prediction results, comparison of the test set bellhop calculation results at a reception depth of 360m and the model prediction results, and comparison of the test set bellhop calculation results at a reception depth of 960m and the model prediction results.

[0036] Figure 6 This is a display of the prediction results of the model for four test samples. The prediction results for the four test samples are shown in the following columns from left to right: the first column shows the calculation results of the Bellhop model, the second column shows the prediction results of the deep neural network based on environmental information features, the third column shows the absolute error between the two, and the fourth column shows a comparison of the sound source depth error curves.

[0037] Table 1. Detailed error data for the test set.

[0038] Table 1: After the model training is completed, 1000 samples are pre-selected as the test set during the training construction in this region and not used for training. The following is the error analysis of 100 test samples. (b) Comparison of the overall RMSE between the test set bellhop calculation results and the model prediction results: the average RMSE is 3.71dB, the maximum RMSE is 32.22dB, and the minimum RMSE is 1.37dB. (c) Comparison of the test set bellhop calculation results at a receiving depth of 360m with the model prediction results: the average RMSE is 5.63dB, the maximum RMSE is 30.75dB, and the minimum RMSE is 1.70dB. (d) Comparison of the test set bellhop calculation results at a receiving depth of 960m with the model prediction results: the average RMSE is 3.73dB, the maximum RMSE is 1.08dB, and the minimum RMSE is 32.66dB.

[0039] Table 2: To verify the prediction speed of the KuS-Net model, the Bellhop model was used for comparison. The Bellhop model calculated 30,720 samples, taking a total of 7 hours and 15 minutes, with a calculation time of approximately 1.5 seconds per sample. The KuS-Net model, on the other hand, can perform parallel predictions on multiple samples (depending on the computer and GPU memory size). Speed ​​tests were conducted on these 30,720 samples using different batch sizes. Detailed data on the prediction speed of the KuS-Net model is provided. (In actual testing, even with the same data and batch size, there were several instances of slightly different times, but the differences were small, all within 5 minutes.) Table 2 Model Prediction Speed

Claims

1. A millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks, characterized in that: An environmental information dataset for the sound propagation field of the target area is constructed by establishing year-round sound velocity profile data and building a two-dimensional terrain model extending 360° around the location of the sound velocity profile. A traditional sound field calculation model is used to calculate the sound propagation field under different environmental conditions in this area as the training set. Then, simplified parameters of the sound velocity profile and terrain features in each environmental sample are extracted as guiding information for the corresponding sound propagation field. A deep neural network is built to learn the distribution pattern of the sound propagation field under different environmental conditions in this area. Finally, the trained model can quickly predict the corresponding sound propagation field in batches within the target area for different environmental conditions. The specific steps are as follows: Step 1: Construction of regional environmental dataset: Using the location information of the annual sound velocity profile of the target area as nodes, a two-dimensional terrain model is established by extending the distance in a 360° direction, and a dataset of environmental information covering the entire year of the region is constructed. The annual environmental information dataset includes an annual sound velocity profile dataset and a topographic profile dataset; Step 2: Extract features from the annual sound velocity profile dataset to obtain M environmental information sound velocity profile features; Step 3: Extract features from the terrain profile set to obtain N environmental information terrain features; Step 4: Construct the sound propagation field dataset: Input the environmental information dataset generated in Step 1 into the traditional sound field calculation model, and calculate the sound propagation loss matrix K*K under the corresponding sound speed profile and terrain conditions; the calculation results are used as the training set of the model, and the environmental information features extracted in Steps 2 and 3 are assigned the corresponding sound propagation field as feature labels; in addition, the terrain vector corresponding to each sound field calculation will also be recorded as additional conditional information for model training. Step 5: Build and train a deep neural network model: Based on the sound speed profile features of dimension M and the seabed topographic features of dimension N extracted in steps 2 and 3 respectively, build an end-to-end deep neural network model to directly predict the high-resolution sound propagation loss field from environmental parameters. The model has a size of K*K. The designed deep neural network consists of three parts: a fully connected layer, an upsampling module, and an output layer. The input is a feature vector of dimension M+N, and the output is a single-channel K*K normalized sound propagation loss matrix. Fully connected layer: Receives M+N dimensional input features, maps them to high-dimensional feature vectors through linear transformation, and then reshapes them into an initial high-dimensional feature map, which serves as the input to the subsequent upsampling module, realizing the effective encoding of low-dimensional environmental parameters into the latent sound field representation; The upsampling module connecting the fully connected layers uses multiple residual-enhanced transposed convolutional blocks to upsample the high-dimensional feature map step by step. Each block combines transposed convolution upsampling feature refinement, batch normalization and LeakyReLU activation, and implements residual connections through matched shortcut paths to alleviate deep gradient vanishing. The output layer uses a single-channel transposed convolution. The output of the upsampling module is processed by the single-channel transposed convolution and the Sigmoid activation output to obtain a normalized acoustic propagation loss matrix of size (k*k). The acoustic propagation loss matrix expresses a fast and efficient end-to-end prediction from physical parameters to high-resolution propagation attenuation distribution. Step 6: Train the constructed network to learn the dataset within the target region and learn the distribution law of sound propagation field under different environmental characteristics; the loss function is designed as the root mean square error between the prediction result of the calculation model and the sound propagation field result calculated by the traditional sound field calculation model.

2. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1, characterized in that: The annual sound velocity profile of the target area in step 1 is obtained from an existing dataset or through field measurements tailored to the needs of the target area.

3. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1, characterized in that: The environmental information sound velocity profile features in step 2 include 1) Initial sound velocity of the sound velocity mixing layer in the sound velocity profile; 2) Initial sound velocity of the transition layer; 3) Depth value at the starting position of the transition layer; 4) Sound velocity gradient in the transition layer; 5) Minimum sound speed in the sound velocity profile; 6) Depth at the point of minimum sound speed in the sound velocity profile, i.e., the depth value of the deep-sea acoustic channel axis; 7) Temperature gradient value of the isothermal layer; 8) Maximum sound speed value in the sound speed profile 9) There are nine characteristics of the depth at the maximum sound speed value in the sound speed profile.

4. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1, characterized in that: The environmental information terrain features in step 3 include: 1) Maximum depth: The maximum depth value in the terrain profile; 2) Minimum depth: The minimum depth value in the topographic profile; 3) Average depth: The average of all depth values ​​in the topographic profile, reflecting the overall depth level of the profile; 4) Median depth: The median of the topographic profile depth data, reflecting the middle value of the depth distribution. Compared with the mean, it is less sensitive to outliers. 5) Location of the mountain peak: The horizontal distance between the location of the mountain peak and the starting point in the topographic profile; 6) Maximum absolute slope: The maximum absolute value of the slope between adjacent points in a topographic profile, reflecting the steepest change in terrain; 7) Average slope: The average of all slopes in the topographic profile, reflecting the overall tilt trend of the terrain; 8) Flatness: The standard deviation of the depth values ​​of the terrain profile, reflecting the flatness of the terrain; 9) Number of oscillations: The number of times the slope sign changes in the topographic profile, reflecting the frequency of topographic undulations or fluctuations; 10) Skewness: Indicates whether the overall topographic profile is from shallow to deep or from deep to shallow, and the degree thereof; 11) Kurtosis: Kurtosis of topographic profile depth data measures the "sharpness" of the depth distribution.

5. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1 or 4, characterized in that: The interval between adjacent points in the topographic profile of step 3 is 1 kilometer.

6. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1 or 4, characterized in that: The flatness in step 3 is calculated using the flatness value formula: ,in This represents the number of depth values ​​in the terrain profile. The first in the topographic profile One depth value; The average depth of the terrain profile.

7. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1 or 4, characterized in that: The skewness in step 3 is calculated using the skewness value formula: ,in This represents the number of depth values ​​in the terrain profile. The first in the topographic profile One depth value; The average depth of the topographic profile; This represents the standard deviation of the topographic profile.

8. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1 or 4, characterized in that: The kurtosis in step 3 is calculated using the following formula: ,in This represents the number of depth values ​​in the terrain profile. The first in the topographic profile One depth value; The average depth of the topographic profile; This represents the standard deviation of the topographic profile.

9. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1, characterized in that: The traditional sound field calculation model uses Bellhop.

10. The millisecond-level prediction method for regional ocean acoustic propagation fields based on deep neural networks according to claim 1, characterized in that: In step 6, the specific calculation range of the loss function is designed according to the requirements, targeting either the global acoustic propagation loss matrix or the portion above the seabed topography. The loss function is calculated by taking the root mean square error of the two at a fixed reception depth. Alternatively, different weights can be added to these two different error calculation methods to unify them as the loss function. The number of iterations and the batch_size hyperparameter are set according to the dataset size and training hardware conditions.

Images