Physical cooperative neural network method for underwater acoustic propagation prediction based on error evaluation routing

Abstract

The present application belongs to the technical field of marine sound propagation loss fast modeling, and specifically discloses a physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing. The present application firstly constructs and trains a neural network-based underwater acoustic propagation loss prediction model; then, the horizontal propagation distance direction is divided into blocks at a fixed interval, and the SSIM and RMSE indexes of the underwater acoustic propagation loss prediction model output results are calculated based on the physical simulation results to construct an error evaluation dataset; an error evaluation model is built, and the error evaluation model is trained using marine sound field environment data and the error evaluation dataset; in the inference stage, dynamic intelligent routing is performed, the marine environment to be predicted is first input into the error evaluation model to obtain the prediction error of each distance block, and whether the preset threshold is met is judged to select the underwater acoustic propagation loss prediction model or the traditional physical simulation model to output the sound propagation loss field.

Description

Technical Field

[0001] This invention belongs to the field of rapid modeling technology for ocean acoustic propagation loss, and specifically relates to a physical cooperative neural network-based underwater acoustic propagation prediction method based on error evaluation routing. Background Technology

[0002] Existing underwater acoustic propagation losses are mainly calculated by solving simplified theories related to wave equations or Helmholtz equations. However, in tasks such as underwater target localization and parameter inversion, the need to perform gridded searches of a vast environmental parameter space and frequently iterate physical simulation models requires solving a large number of complex differential equations. This not only consumes huge computational and storage resources but also severely restricts the real-time performance and dynamic response capability of underwater acoustic tasks.

[0003] In recent years, novel artificial intelligence-based methods have offered entirely new solutions for the rapid modeling of underwater acoustic propagation loss fields, including end-to-end neural networks and physically constrained neural networks. However, most existing research focuses on designing AI models to "replace" physical simulation models. Because these methods rely on data-driven approaches, they struggle to identify out-of-distribution data and accurately model anomalous and abrupt environmental changes. Furthermore, trained AI models are essentially black-box models, lacking necessary physical interpretability, which significantly reduces their credibility in engineering applications.

[0004] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0005] The purpose of this invention is to propose a physical-cooperative neural network underwater acoustic propagation prediction method based on error assessment routing. This method introduces a risk assessment step into the neural network modeling of horizontal acoustic propagation loss field, so as to realize dynamic confidence assessment and diversion operation on the input marine acoustic field environmental data, thereby preserving the high-fidelity characteristics of the physical simulation model in complex and variable environments, and giving full play to the rapid prediction capability of the neural network model in simple and normal environments.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The physical cooperative neural network-based underwater acoustic propagation prediction method based on error evaluation routing includes the following steps: Step 1. Build and train a neural network-based underwater acoustic propagation loss prediction model, which takes ocean acoustic field environmental data as input and outputs the horizontal acoustic propagation loss field at a preset horizontal propagation distance. Step 2. Calculate the output of the underwater acoustic propagation loss prediction model based on the physical simulation results. and The index divides the preset horizontal propagation distance into blocks at fixed intervals along the horizontal propagation distance direction to evaluate the modeling effect of the underwater acoustic propagation loss prediction model at different distance blocks on the training dataset, and generates an error evaluation dataset for the training error evaluation model. Step 3. Construct an error assessment model based on a neural network. The error assessment model takes ocean acoustic field environmental data as input and outputs the values ​​at each distance block. and Indicators; training the constructed error assessment model; The error assessment model and the underwater acoustic propagation loss prediction model share the feature extraction part; Step 4. In the inference phase, dynamic intelligent routing is executed. The marine environment to be predicted is first input into the trained error assessment model. The error assessment model outputs the prediction error at each distance block. and index; If the prediction error of all distance blocks meets the pre-set threshold condition, the ocean acoustic field environment data to be predicted is sent into the trained underwater acoustic propagation loss prediction model, and the horizontal acoustic propagation loss field is predicted. If the prediction error of any distance block does not meet the threshold condition, it will automatically be routed to the traditional physical simulation model, which will be used to perform high-precision calculations and output the final reliable horizontal sound propagation loss field.

[0007] Furthermore, based on the aforementioned physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing, this invention also proposes a computer device comprising a memory and one or more processors.

[0008] Executable code is stored in memory. When the processor executes the executable code, it implements the steps of the above-described physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing.

[0009] Furthermore, based on the aforementioned physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing, this invention also proposes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the aforementioned physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing.

[0010] The present invention has the following advantages: As described above, this invention addresses the problem that existing neural network acoustic propagation loss modeling methods are often difficult to apply in engineering due to their dependence on training data. It proposes a physical-cooperative neural network underwater acoustic propagation prediction method based on error assessment routing. This method, based on the concept of out-of-distribution data detection in computer vision, introduces a risk assessment step into the neural network modeling of the horizontal acoustic propagation loss field. It performs dynamic confidence assessment and triage on the input marine acoustic field environmental data, then sends it to either the neural network prediction model or the traditional physical simulation model. The collaborative modeling strategy based on error assessment routing proposed in this invention aims to overcome the limitation of using neural networks to replace physics in underwater acoustic modeling, preserving the high-fidelity characteristics of physical simulation models in complex and variable environments, and fully leveraging the rapid prediction capabilities of neural network models in simple and normal environments. Experimental results show that the error assessment model achieves a 94.07% accuracy rate in identifying high-risk unreliable samples, effectively distinguishing reliable predicted samples (average RMSE 2.473dB) from unreliable samples (average RMSE 4.136dB), and reducing the overall computation time by more than 50%. Attached Figure Description

[0011] Figure 1 This is a flowchart of the physical cooperative neural network underwater acoustic propagation modeling method based on error evaluation routing in Embodiment 1 of the present invention; Figure 2 This is a model structure diagram of the evaluation-prediction architecture collaborative modeling in this embodiment of the invention; Figure 3 This is a schematic diagram of the sound field loss environment shown in the target physical simulation field in this embodiment of the invention; Figure 4 This is a schematic diagram of the predicted acoustic field environment output by a neural network-based underwater acoustic propagation loss prediction model. Figure 5 This is a schematic diagram of the evaluation error of the target physical simulation field and the prediction model of underwater acoustic propagation loss based on neural networks, and the block error of the prediction model. The bar chart represents the RMSE index, and the line represents the SSIM index. Detailed Implementation

[0012] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1 This embodiment describes a physical co-neural network-based underwater acoustic propagation prediction method based on error evaluation routing. The method first constructs a marine acoustic field environment dataset, dividing it into a training dataset and a test dataset. Based on the training dataset, a neural network-based underwater acoustic propagation loss prediction model is trained. This model can quickly output the horizontal acoustic propagation loss field at a preset horizontal distance according to the marine acoustic field environment and seabed topography. The test dataset is used to evaluate whether the model training has reached convergence. The output of the underwater acoustic propagation loss prediction model is calculated based on physical simulation results. and The system uses a pre-defined horizontal propagation distance to divide the data into blocks at fixed intervals along the horizontal propagation direction to evaluate the modeling performance of the underwater acoustic propagation loss prediction model at different distance blocks on the training dataset. A neural network error assessment model is constructed and trained. This model can quickly output the predicted SSIM and RMSE indices based on the ocean acoustic field environment and seabed topography, enabling prior error assessment of the prediction model's reliability. During the inference phase, dynamic intelligent routing is executed. The ocean environment to be predicted is first input into the error assessment model to obtain the prediction error for each distance block. If all distance blocks meet the threshold conditions, the ocean acoustic field environment data to be predicted is sent to the underwater acoustic propagation loss prediction model for rapid calculation. If any distance block does not meet the threshold, it is determined to be a difficult-to-predict environment, and automatic routing is used to a traditional physical simulation model for high-precision fallback calculation, outputting the final reliable acoustic propagation loss field.

[0013] like Figure 1 As shown, the physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing includes the following steps: Step 1. Build and train a neural network-based underwater acoustic propagation loss prediction model, which takes ocean acoustic field environmental data as input and outputs the horizontal acoustic propagation loss field at a preset horizontal propagation distance.

[0014] The marine acoustic field environmental data includes two parts: sound velocity profile data and seabed topography data.

[0015] The sound velocity profile data and seabed topography data together constitute the input data for the underwater acoustic propagation loss prediction model, while the horizontal acoustic propagation loss field serves as the supervision label used for training the underwater acoustic propagation loss prediction model.

[0016] The sound velocity profile data, seabed topography data, and horizontal sound propagation loss field are derived from three parts: the sound velocity profile input environment dataset, the topography input environment dataset, and the sound propagation loss output environment dataset, respectively.

[0017] The acoustic propagation loss output environment dataset is obtained by calculating the input environment using the physical simulation model Bellhop. Since the process of calculating acoustic propagation loss using the physical simulation model Bellhop is relatively conventional, it will not be described in detail here.

[0018] The underwater acoustic propagation loss prediction model constructed in this embodiment mainly includes feature extraction and feature mapping.

[0019] The feature extraction part includes a residual convolution module, which is used to extract nonlinear input feature information from the input ocean acoustic field environmental data. Its specific implementation is as follows: .

[0020] in This represents the marine acoustic field environmental input, including sound velocity profiles and topographic input. This represents the output features of the feature extraction part. Represents the SiLU activation function. represent Normalization This represents a convolutional layer.

[0021] The feature mapping part consists of a multilayer perceptron, which is used to fit the feature information into the output result. The specific implementation method is as follows: .

[0022] in To predict the horizontal sound propagation loss, Represents the SiLU activation function. Represents Layer normalization. and This is the learning matrix for the fully connected layer. and This represents the bias term.

[0023] This represents a randomly deactivated layer, which prevents the model from overfitting by randomly discarding some neurons.

[0024] It should be noted that the network structure of the underwater acoustic propagation loss prediction model is not the focus of this invention. Any network structure that includes feature extraction and feature mapping and can achieve underwater acoustic propagation loss prediction can be used as an alternative.

[0025] A dataset was constructed using sound velocity profile data, seabed topography data, and horizontal sound propagation loss field as samples. The dataset was then divided into training and testing datasets to enable the training and testing of the underwater sound propagation loss prediction model.

[0026] The underwater acoustic propagation loss prediction model is trained based on the training dataset. The underwater acoustic propagation loss prediction model can output the horizontal acoustic propagation loss field at a preset horizontal distance based on the ocean acoustic field environment data and seabed topography data.

[0027] Then, the test dataset was used to evaluate whether the underwater acoustic propagation loss prediction model training had reached convergence.

[0028] The training process uses the PyTorch framework and the Adam optimizer, with a learning rate of 0.001, a maximum iteration cycle limit of 1000, and 25 early stopping groups. If the error does not decrease after more than 25 groups on the test dataset, it means that the model has converged.

[0029] Step 2. Calculate the output of the underwater acoustic propagation loss prediction model based on the physical simulation results. and The index divides the preset horizontal propagation distance into blocks at fixed intervals along the horizontal propagation distance direction to evaluate the modeling effect of the underwater acoustic propagation loss prediction model at different distance blocks on the training dataset, and generates an error evaluation dataset for the training error evaluation model.

[0030] This invention relies on two evaluation indicators, Structural Similarity (SSIM) and Root Mean Square Error (RMSE), to divide the overall propagation distance into regions along the horizontal propagation path of underwater acoustics according to fixed spatial intervals. This allows for the evaluation of the modeling fitting effect and prediction accuracy of the underwater acoustic propagation loss prediction model in different regions of different propagation distances.

[0031] The underwater acoustic channel exhibits significant distance-dependent propagation characteristics. The sound propagation attenuation patterns, sound field distribution characteristics, multipath interference effects, and environmental interference levels vary considerably at different horizontal propagation distances. A unified global evaluation cannot accurately reflect the actual predictive performance of the model in different propagation ranges such as far-field, near-field, and mid-range.

[0032] Therefore, by independently calculating the error index in blocks according to the propagation distance, this invention can achieve refined accuracy assessment in different regions, accurately locate the prediction deviation range and error distribution pattern of the model in different propagation distance segments, and facilitate subsequent adaptive adjustment of routing weights based on the error assessment results. This enables the synergistic optimization of physical priors and neural networks, thereby specifically improving the overall prediction accuracy and scene adaptability of underwater acoustic propagation loss across the entire distance range.

[0033] After the underwater acoustic propagation loss prediction model outputs the horizontal acoustic propagation loss field at a preset horizontal propagation distance, the preset horizontal propagation distance is divided into blocks at fixed intervals to obtain multiple distance blocks within the preset horizontal propagation distance range.

[0034] Within each distance block, the results of physical simulation are compared with those of the underwater acoustic propagation loss prediction model. The SSIM and RMSE index values ​​are calculated using formulas to generate SSIM and RMSE error assessment datasets.

[0035] The ocean acoustic field environment data in the training dataset is used as the input to the error assessment model, and the SSIM and RMSE error assessment data are used as the output to train the error assessment model.

[0036] SSIM stands for Structural Similarity Index, and RMSE stands for Root Mean Square Error.

[0037] In this embodiment, the preset horizontal propagation distance is set according to the propagation characteristics of the target sea area; the fixed interval is divided into equal or non-equal intervals according to the distance dependence characteristics of underwater sound propagation.

[0038] For example, the preset horizontal propagation distance can be set to 100km; the entire preset horizontal propagation distance is divided into 10 blocks every 10km along the horizontal propagation distance to cover the entire preset horizontal propagation distance of 100km.

[0039] in and The formulas for calculating the indicators are as follows: The index is calculated region by region using a fixed-size sliding local calculation window, and the formula is as follows: .

[0040] in and These represent the horizontal acoustic propagation loss field obtained from underwater acoustic physics simulation and the acoustic propagation loss prediction data output from the underwater acoustic propagation loss prediction model, respectively. and Representative window and The average value.

[0041] and Representative window The variance of y Representative window and The covariance.

[0042] and These represent two tiny constants introduced to maintain computational stability.

[0043] in, This represents the dynamic range of values ​​for the elements in the sound propagation loss matrix, that is, the difference between the maximum and minimum values ​​of the elements in the sound propagation loss matrix. and Set as the standard empirical constant. , .

[0044] The continuous horizontal acoustic propagation loss field output by the underwater acoustic propagation loss prediction model and the physical simulation model is discretized and represented as a two-dimensional acoustic propagation loss matrix during computer processing.

[0045] Based on this sound propagation loss matrix The formula for calculating the indicator is: .

[0046] in, and Representing the first and second elements in the sound propagation loss matrix, respectively. The acoustic propagation loss values ​​obtained from the physical simulation solution corresponding to each element position and the acoustic propagation loss values ​​predicted by the underwater acoustic propagation loss prediction model. This represents the total number of elements in the sound propagation loss matrix. .

[0047] In actual calculations, in order to obtain the error characteristics of different distance segments, the preset horizontal propagation distance (e.g., 100km as mentioned above) is divided into regions along the horizontal propagation direction of underwater sound according to preset fixed distance intervals (e.g., 10km).

[0048] The specific implementation process is as follows: First, extract the first... The acoustic propagation loss matrix is ​​obtained from physical simulation and neural network prediction within each distance block.

[0049] Secondly, substitute the two extracted sound propagation loss matrices into the above... and The formula calculates the local structural similarity and root mean square error block by block, obtaining the value corresponding to that distance block. and .

[0050] Finally, all distance blocks are traversed sequentially, and the evaluation results corresponding to each distance block are spliced ​​together in spatial order to form an error index sequence, which is then stored as a supervision label data in the error evaluation dataset.

[0051] This error assessment dataset is used for training the error assessment model in step 3 below.

[0052] Step 3. Construct an error assessment model based on a neural network. The error assessment model takes ocean acoustic field environmental data as input and outputs the values ​​at each distance block. and Indicators; training the constructed error assessment model.

[0053] The error assessment model and the underwater acoustic propagation loss prediction model share the same feature extraction part, meaning they have the same feature extraction component. In this embodiment, the error assessment model includes feature extraction and feature mapping parts.

[0054] The feature extraction part of the error assessment model is the same as that of the underwater acoustic propagation loss prediction model.

[0055] The error assessment model and the neural network prediction model differ only in the dimension of the feature information output. The prediction model outputs the predicted sound propagation loss field, while the error assessment model outputs the predicted evaluation metrics SSIM and RMSE.

[0056] Error assessment models are used to determine the failure risk of a prediction model before it is built.

[0057] This error evaluation model also includes feature extraction and feature mapping. The feature extraction part mainly consists of residual convolution modules, which are used to extract nonlinear input feature information from the data. The specific implementation method is the same as that of the prediction model. .

[0058] in This represents the marine acoustic field environmental input, including sound velocity profiles and topographic input. This represents the output features of the feature extraction part. Represents the SiLU activation function. represent Normalization Represents a convolutional layer; The feature mapping module consists of a multilayer sensing mechanism. This module is used to perform nonlinear fitting mapping on the deep acoustic field features output by the feature extraction module, and finally outputs the prediction error evaluation value corresponding to each distance block region along the underwater acoustic propagation path, namely the SSIM evaluation index value and RMSE evaluation index value of the corresponding block region. The calculation expression is as follows: .

[0059] in The predicted values ​​of the SSIM index and RMSE index for each propagation distance block output by the model; Represents the SiLU activation function. Represents Layer normalization. and This is the learning matrix for the fully connected layer. and Represents the bias term. This represents a randomly deactivated layer, which prevents the model from overfitting by randomly discarding some neurons.

[0060] Step 4. Execute dynamic intelligent routing during the inference phase.

[0061] The marine environment to be predicted is first input into a trained error assessment model, which then outputs the prediction error at each distance block. and Indicators are used to assess the prior error of the predictive model's reliability.

[0062] During the inference phase, dynamic intelligent routing is executed, with a global veto rule: if any distance block does not meet the threshold, the system directly switches to the Bellhop physical simulation model and does not use the modeling results of the predictive neural network model.

[0063] If the prediction errors of all distance blocks meet the pre-set threshold conditions, the ocean acoustic field environment data to be predicted is sent to the trained underwater acoustic propagation loss prediction model, and the horizontal acoustic propagation loss field is predicted.

[0064] If the prediction error of any distance block does not meet the threshold condition, it will automatically be routed to the traditional physical simulation model, which will be used to perform high-precision calculations and output the final reliable horizontal sound propagation loss field.

[0065] set up and The judgment thresholds for the indicators are respectively and ,in The value range can be set to 0.85-0.9. The value range can be set to - .For example, , .

[0066] The pre-set threshold condition is: and .

[0067] in It focuses on measuring the similarity between the spatial distribution structure, contour morphology and interference texture of the sound field; The absolute deviation of the quantitative sound propagation loss index can intuitively reflect the accuracy of sound field amplitude prediction. The two indices are jointly verified from the two dimensions of numerical accuracy and spatial structure, which can avoid the one-sidedness of evaluation by a single index.

[0068] For the prediction error at each distance block, only when both conditions are met... and Only when the prediction error at a specific block is considered to meet the pre-set threshold condition is it considered that the prediction error at that specific block meets the pre-set threshold condition.

[0069] Furthermore, only when the prediction errors at all distance blocks meet the above threshold conditions will the ocean acoustic field environment data to be predicted be sent into the trained underwater acoustic propagation loss prediction model, and the horizontal acoustic propagation loss field be predicted.

[0070] Otherwise, it automatically routes to the traditional physical simulation model (in this embodiment, the Bellhop physical simulation model is used), performs high-precision calculations using the traditional physical simulation model, and outputs the final reliable horizontal sound propagation loss field.

[0071] like Figure 2 As shown, dynamic intelligent routing is executed during the inference phase. The ocean acoustic field environment data to be predicted is first input into the error evaluation model based on the neural network to obtain the prediction error of each distance block.

[0072] The prediction error of each distance block here is the prediction error corresponding to each distance block. and value.

[0073] If the prediction errors of all distance blocks meet the pre-set threshold conditions, the marine acoustic field environmental data to be predicted will be sent to the underwater acoustic propagation loss prediction model for prediction, and the model will predict the horizontal acoustic propagation loss field.

[0074] If the prediction error of any distance block does not meet the threshold condition, it will automatically be routed to the traditional physical simulation model to perform high-precision calculations and output the final reliable horizontal sound propagation loss field.

[0075] In this embodiment, it can be set to and A dual-indicator combined hard decision is adopted.

[0076] like Figure 3 The acoustic field loss environment shown in the target physical simulation field is presented. Figure 4 This demonstrates the predicted sound field environment obtained from the predictive neural network model. Figure 5 The errors of both methods are shown, as well as the error of the evaluation model's prediction.

[0077] in Figure 5 The bar chart represents the RMSE indicator, and the line chart represents the SSIM indicator.

[0078] Depend on Figure 5 The comparative analysis shows that the error of the underwater acoustic propagation loss prediction model will increase after 60km, indicating that the results of the underwater acoustic propagation loss prediction model after this distance are not reliable.

[0079] This invention utilizes a proposed neural network-based error assessment model to perform dynamic confidence assessment and triage on input marine acoustic environment data. The data is then fed into either a neural network prediction model for rapid forecasting or a traditional physical simulation model for high-fidelity, accurate forecasting, achieving collaborative modeling between physics and artificial intelligence. This architecture exhibits high model independence, requiring only lightweight fine-tuning to adapt to various existing neural networks, providing a novel solution for intelligent underwater acoustic forecasting.

[0080] Example 2 This embodiment 2 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing described in embodiment 1 above.

[0081] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0082] Example 3 This embodiment 3 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing in embodiment 1 above.

[0083] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0084] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A physical cooperative neural network-based underwater acoustic propagation prediction method based on error evaluation routing, characterized in that, Includes the following steps: Step 1. Build and train a neural network-based underwater acoustic propagation loss prediction model, which takes ocean acoustic field environmental data as input and outputs the horizontal acoustic propagation loss field at a preset horizontal propagation distance. Step 2. Calculate the output of the underwater acoustic propagation loss prediction model based on the physical simulation results. and The index divides the preset horizontal propagation distance into blocks at fixed intervals along the horizontal propagation distance direction to evaluate the modeling effect of the underwater acoustic propagation loss prediction model at different distance blocks on the training dataset, and generates an error evaluation dataset for the training error evaluation model. Step 3. Construct an error assessment model based on a neural network. The error assessment model takes ocean acoustic field environmental data as input and outputs the values ​​at each distance block. and Indicators; training the constructed error assessment model; The error assessment model and the underwater acoustic propagation loss prediction model share the feature extraction part; Step 4. In the inference phase, dynamic intelligent routing is executed. The marine environment to be predicted is first input into the trained error assessment model. The error assessment model outputs the prediction error at each distance block. and index; If the prediction error of all distance blocks meets the pre-set threshold condition, the ocean acoustic field environment data to be predicted is sent into the trained underwater acoustic propagation loss prediction model, and the horizontal acoustic propagation loss field is predicted. If the prediction error of any distance block does not meet the threshold condition, it will automatically be routed to the physical simulation model, and high-precision calculations will be performed using the traditional physical simulation model to output the final reliable horizontal sound propagation loss field.

2. The physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing according to claim 1, characterized in that, The marine acoustic field environmental data includes two parts: sound velocity profile data and seabed topography data. The sound velocity profile data and seabed topography data together constitute the input data of the underwater acoustic propagation loss prediction model, and the horizontal acoustic propagation loss field serves as the supervision label used for training the underwater acoustic propagation loss prediction model. The sound velocity profile data, seabed topography data, and horizontal sound propagation loss field are respectively derived from three parts: the sound velocity profile input environment dataset, the topography input environment dataset, and the sound propagation loss output environment dataset. The acoustic propagation loss output environment dataset is obtained by calculating the input environment using a physical simulation model.

3. The physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing according to claim 1, characterized in that, The underwater acoustic propagation loss prediction model includes feature extraction and feature mapping components; The feature extraction part of the underwater acoustic propagation loss prediction model includes a residual convolution module. This feature extraction part is used to extract nonlinear input feature information from the input ocean acoustic field environmental data. Its specific implementation is as follows: ； in This represents the marine acoustic field environmental input, including sound velocity profiles and topographic input. This represents the output features of the feature extraction part. Represents the SiLU activation function. represent Normalization Represents a convolutional layer; The feature mapping part consists of a multilayer perceptron, which is used to fit the feature information into the output result. The specific implementation method is as follows: ； in To predict the horizontal sound propagation loss, Represents the SiLU activation function. Represents Layer normalization. and This is the learning matrix for the fully connected layer. and Represents the bias term; This represents a randomly deactivated layer, which prevents the model from overfitting by randomly discarding some neurons.

4. The physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing according to claim 1, characterized in that, Step 2 specifically involves: After the underwater acoustic propagation loss prediction model outputs the horizontal acoustic propagation loss field at a preset horizontal propagation distance, the preset horizontal propagation distance is divided into blocks at fixed intervals to obtain multiple distance blocks within the preset horizontal propagation distance range. Within each distance block, the results of physical simulation are compared with those of the underwater acoustic propagation loss prediction model. The SSIM and RMSE index values ​​are calculated using formulas to generate SSIM and RMSE error assessment datasets. The ocean acoustic field environment data in the training dataset is used as the input to the error assessment model, and the SSIM and RMSE error assessment data are used as the output to train the error assessment model.

5. The physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing according to claim 4, characterized in that, In step 2, and The formulas for calculating the indicators are as follows: The index is calculated region by region using a fixed-size sliding local calculation window, and the formula is as follows: ； in and These represent the horizontal acoustic propagation loss field obtained from underwater acoustic physics simulation and the acoustic propagation loss prediction data output from the underwater acoustic propagation loss prediction model, respectively. and Representative window and The average value; and Representative window The variance of y Representative window and covariance; and It is a constant; The continuous horizontal acoustic propagation loss field output by both the underwater acoustic propagation loss prediction model and the physical simulation model is discretized into a two-dimensional acoustic propagation loss matrix during computer processing. Based on this acoustic propagation loss matrix, The formula for calculating the indicator is: ； in, and Representing the first and second elements in the sound propagation loss matrix, respectively. The acoustic propagation loss values ​​obtained from the physical simulation solution corresponding to each element position and the acoustic propagation loss values ​​predicted by the underwater acoustic propagation loss prediction model. This represents the total number of elements in the sound propagation loss matrix. ; The system divides the area into blocks along the horizontal propagation direction of underwater sound at preset fixed distance intervals, calculates the two evaluation indicators SSIM and RMSE for each block, and stores the SSIM and RMSE indicator data corresponding to each block.

6. The physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing according to claim 4, characterized in that, The preset horizontal propagation distance is set according to the propagation characteristics of the target sea area; The fixed intervals are divided into equal or non-equal intervals based on the distance-dependent characteristics of underwater acoustic propagation.

7. The physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing according to claim 1, characterized in that, The error assessment model includes feature extraction and feature mapping components; wherein the feature extraction module configured in the error assessment model adopts the same feature extraction structure and operation logic as the underwater acoustic propagation loss prediction model. The feature mapping module of the error evaluation model is composed of a multilayer sensing mechanism. It is used to perform nonlinear fitting mapping on the deep sound field features output by the feature extraction module, and output the prediction error evaluation value corresponding to each distance block region on the underwater acoustic propagation path, that is, the SSIM and RMSE evaluation index values ​​of the corresponding block region. The calculation expression is as follows: ； in The predicted values ​​of the SSIM index and RMSE index for each propagation distance block output by the model; Represents the SiLU activation function. Represents Layer normalization. and This is the learning matrix for the fully connected layer. and Represents the bias term. This represents a randomly deactivated layer, which prevents the model from overfitting by randomly discarding some neurons.

8. The physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing according to claim 1, characterized in that, In step 4, set and The judgment thresholds for the indicators are respectively and ; The pre-set threshold condition is: and .

9. A computer device, comprising a memory and one or more processors; characterized in that, The memory stores executable code, which, when executed by the processor, is used to implement the physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a program stored thereon; characterized in that, When executed by the processor, the program is used to implement the physical cooperative neural network underwater acoustic propagation prediction method based on error evaluation routing as described in any one of claims 1 to 8.

Images