Private large model training method based on afno-based sound propagation loss calculation

Abstract

The application discloses a private large model training method based on AFNO sound propagation loss calculation, and belongs to the field of ship electronic information technology, comprising the following steps: generating high-fidelity sound field data based on four typical marine environments through a traditional acoustic physical model; constructing an AFNO-based sound propagation loss calculation model architecture, and performing self-adaptive training by using the high-fidelity sound field data to generate a private large model; setting evaluation indexes and obtaining a test set to complete model evaluation of the private large model; and the application combines the efficient frequency domain calculation capability of AFNO with the physical characteristics of the traditional acoustic method, provides a private large model, and realizes optimization of sound propagation loss calculation.

Description

Technical Field

[0001] This invention belongs to the field of ship electronic information technology, specifically relating to a method for training a private large model for calculating acoustic propagation loss based on AFNO. Background Technology

[0002] Globally, the complex and ever-changing marine environment and the need for high-precision seabed topography have presented unprecedented challenges to acoustic propagation calculations. The propagation path of sound waves in complex seabed topography is significantly influenced by various factors such as seabed undulations, sediment layer thickness, and seabed type. High-precision acoustic propagation loss calculations can provide accurate sound field predictions for the design of underwater communication systems and sonar detection equipment, thereby improving their performance and reliability. However, traditional acoustic modeling methods often suffer from low computational efficiency and insufficient adaptability when dealing with complex environments. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a method for training a private large model for calculating acoustic propagation loss based on AFNO, in order to solve the above-mentioned technical problems.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for training a private large model for calculating acoustic propagation loss based on AFNO, comprising:

[0006] Based on four typical marine environments, high-fidelity sound field data is generated using traditional acoustic physics models;

[0007] A sound propagation loss calculation model architecture based on AFNO is constructed, and high-fidelity sound field data is used for adaptive training to generate a private large model.

[0008] Set evaluation metrics and obtain a test set to complete the model evaluation of the private large model.

[0009] Furthermore, the process for generating high-fidelity sound field data includes:

[0010] Step 1: Define simulation parameters based on four typical marine environments; the number of simulation parameters is several, including sound source depth, frequency, and sound velocity profile;

[0011] Step 2: Initialize the .env file;

[0012] Step 3: Modify the frequency parameters in the outer loop;

[0013] Step 4: Modify the sound source depth parameters in the inner loop;

[0014] Step 5: Generate the .env file;

[0015] Step 6: Select a traditional acoustic physical model; among which, traditional acoustic physical models include BELLHOP, KRAKEN, and RAM;

[0016] Step 7: Generate a .shd file based on the selected traditional acoustic physics model;

[0017] Step 8: Integrate the results and save them as a .mat file to complete the generation of high-fidelity sound field data.

[0018] Furthermore, steps 1-8, when processing data, also include outlier cleaning, normalization, and data augmentation.

[0019] Furthermore, the acoustic propagation loss calculation model architecture based on AFNO is a frequency-spatial domain hybrid calculation architecture, including an input layer, a Fourier transform module, an adaptive weighting module, a nonlinear activation layer, a residual connection, and an output layer.

[0020] The input layer is used to convert high-dimensional time-domain data into frequency-domain data.

[0021] The Fourier transform module is used to extract frequency domain features;

[0022] The adaptive module is used to adjust the weights in the frequency space;

[0023] The output layer is used to project the frequency domain features back into the time domain and generate the final prediction result;

[0024] Residual connections are used throughout the entire network.

[0025] Furthermore, the adaptive training process includes:

[0026] The AFNO-based sound propagation loss calculation model predicts sound propagation loss by inputting a dependent variable, where the dependent variable is high-fidelity sound field data.

[0027] During training, the frequency domain weights are dynamically adjusted using adaptive Fourier transform, and training is completed by combining mixed precision training and a dual-path loss function. The dual-path loss function is designed based on the similarity of frequency energy and spatial structure.

[0028] After each training iteration, the parameters in the model network are updated through backpropagation based on the gradient descent optimization algorithm to achieve model optimization. The core objective of training is to minimize the error between the model's predicted values ​​and the actual observed data.

[0029] Furthermore, the four typical marine environments include shallow sea horizontal environment, shallow sea wedge environment, annular seamount environment, and deep sea horizontal environment; for these four typical marine environments, the adaptive training process also includes comparative training:

[0030] In shallow sea environments, multipath reflection feature modeling is optimized by jointly training the AFNO frequency domain adaptive mechanism with the traditional acoustic physics model.

[0031] In shallow sea wedge environments, the model's adaptability to gradually changing terrain is optimized by simulating the reflection loss of the slope topography and seabed sediment parameters.

[0032] In a ring-shaped seamount environment, we learn the multipath scattering characteristics of sound waves by the seamount structure to enhance the model's accuracy in predicting the sound field of complex obstacles.

[0033] In deep-sea horizontal environments, the acoustic ray bending effect caused by water layer stratification is captured to ensure accurate modeling of the sound field in the convergence zone.

[0034] The beneficial effects of this invention are as follows:

[0035] This invention proposes a method for training a private large model for calculating sound propagation loss based on the Adaptive Fourier Neural Operator (AFNO) by combining the general capabilities of large models with expertise in the field of acoustics. This method can integrate data from multiple sensors, simulate and predict the propagation characteristics of sound waves in complex environments, process and learn data from complex seabed topography, and simulate the propagation loss of sound waves under different conditions. At the same time, through the personalized private large model, this invention significantly improves the accuracy and robustness of acoustic models, providing strong technical support for optimizing the design of acoustic systems.

[0036] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0037] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0039] Figure 1 This is a flowchart of a method for training a private large model based on AFNO to calculate acoustic propagation loss in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the high-fidelity sound field data generation process in a private large model training method for calculating sound propagation loss based on AFNO in an embodiment of the present invention.

[0041] Figure 3 This is a diagram of the AFNO model architecture in a private large model training method for calculating acoustic propagation loss based on AFNO, as described in an embodiment of the present invention. Detailed Implementation

[0042] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0043] like Figure 1 As shown, this invention proposes a method for training a private large model based on AFNO for calculating acoustic propagation loss, including:

[0044] S101. Based on four typical marine environments, high-fidelity sound field data is generated through traditional acoustic physics models.

[0045] S102. Construct an AFNO-based acoustic propagation loss calculation model architecture and use high-fidelity sound field data for adaptive training to generate a private large model.

[0046] S103. Set evaluation metrics and obtain test sets to complete the model evaluation of the private large model;

[0047] The working principle of the above technical solution is as follows: The core of this application is to construct an efficient sound propagation loss calculation model based on the adaptive Fourier neural operator (AFNO) for four typical marine scenarios: shallow sea horizontal environment, shallow sea wedge environment, deep sea horizontal environment and annular seamount environment.

[0048] Specifically, based on four typical marine environments (shallow sea level / wedge, deep sea level, and ring seamount), high-fidelity sound field data is generated using physical models (BELLHOP, RAM, KRAKEN). This data is then combined with outlier cleaning, normalization, and data augmentation to construct a standardized training set adapted to complex terrain. Next, a frequency-spatial hybrid computing architecture is designed, dynamically adjusting frequency domain weights using adaptive Fourier transform. This is combined with hybrid precision training and a dual-path loss function (frequency domain energy + spatial structure similarity) to achieve efficient fitting of physical laws. Finally, the model accuracy is validated using MSE / MAE metrics, adapting to shallow / deep / complex terrain scenarios, optimizing sonar detection and noise control, and supporting lightweight deployment and real-time prediction.

[0049] Among them, the frequency-spatial hybrid computing architecture utilizes AFNO's fast Fourier transform to process global sound field features in the frequency domain (such as multipath reflection in shallow sea horizontal environments and sound fields in deep sea convergence zones), and combines spatial convolution to capture local terrain details (such as the gradual terrain of shallow sea wedge slopes and the complex scattering effects of ring seamounts). Through an adaptive mechanism, the computing weights of the frequency domain and spatial domain are dynamically allocated, which significantly improves the computing efficiency under complex terrain.

[0050] In one specific embodiment, after the model is trained, its performance is evaluated using a test set. Evaluation metrics include mean squared error (MSE), mean absolute error (MAE), and error distribution. For the 100×100 dataset, the MSE is 14.68 dB and the MAE is 2.75 dB; for the 500×500 dataset, the MSE is 13.92 dB and the MAE is 2.67 dB. The error distribution shows that 90% of the prediction errors are within ±10 dB (AFNO100 achieves 81.58% accuracy, and AFNO500 achieves 82.88% accuracy). In shallow sea horizontal environments, the prediction results are highly consistent with BELLHOP simulations (error <3 dB), but in shallow sea wedge environments, the model lacks sufficient analysis of terrain details (such as slope gradients). In deep-sea Gaussian seamount environments, the sound field texture matches the ground truth by 85%, and the diffraction path prediction error is <2 dB.

[0051] Furthermore, in practical applications, the AFNO model can be widely used in underwater communication optimization, target detection enhancement, and environmental noise control. By predicting the propagation loss of sound waves in complex terrain, it optimizes the deployment location and frequency selection of sonar systems, improving the effective detection range of submarines and underwater equipment. Simultaneously, it dynamically adjusts noise suppression strategies by combining real-time sound field simulation. To support edge device deployment, the model uses quantization technology (FP32→INT8) to increase inference speed by 3 times (target <1 second / 100 samples). In addition, the model continuously optimizes through a feedback learning mechanism, dynamically updating model parameters based on user-measured data (such as sonar echoes and seabed topography scans). Samples with prediction errors >15dB are retrained to iteratively optimize the model's generalization ability. Ultimately, the AFNO model achieves a prediction error <3dB in deep-sea and complex terrain scenarios, meeting engineering application requirements and providing an efficient and accurate solution for calculating sound propagation loss in marine acoustics engineering.

[0052] Furthermore, it supports users in customizing private models based on measured data from specific sea areas (such as seabed sediment thickness and seabed sound velocity profiles); for example, for shallow sea environments with multiple reflection paths, high-frequency sonar data can be fused to optimize model details; for ring-shaped seamount areas, side-scan sonar and topographic data can be combined to train models to accurately analyze sound waves.

[0053] The beneficial effects of the above technical solution are as follows: Through the above technical solution, the AFNO model uses Fast Fourier Transform (FFT) for efficient calculation in the frequency domain and uses an adaptive mechanism to dynamically adjust the processing ratio of the frequency domain and the spatial domain, effectively improving the computational efficiency in complex acoustic environments; it supports users to customize training models based on specific scenarios, integrate sensor data and seabed topographic information to meet the needs of different scenarios, and ensure computational accuracy and data confidentiality; it can be widely used in the field of marine acoustics, such as sonar system design, underwater communication optimization, marine resource exploration, and environmental noise prediction, providing efficient tools for acoustic simulation and optimization, and improving the performance and reliability of acoustic equipment.

[0054] like Figure 2 As shown, in one embodiment, the process for generating high-fidelity sound field data includes:

[0055] Step 1: Define simulation parameters based on four typical marine environments. The number of simulation parameters includes key parameters such as sound source depth (50m-500m, 50m interval), frequency (50Hz-5000Hz, 50Hz interval) and sound velocity profile (e.g., Munk model) to ensure that the dataset can comprehensively reflect the sound propagation characteristics under different marine environments.

[0056] In one specific embodiment, each model simulates 1000 samples, and the total dataset contains 4000 samples, occupying 5.25GB of space (500×500 size). The generation time is approximately 8 hours. Some dataset attributes are as follows:

[0057]

[0058] Step 2: Initialize the .env file;

[0059] Step 3: Modify the frequency parameters in the outer loop;

[0060] Step 4: Modify the sound source depth parameters in the inner loop;

[0061] Step 5: Generate the .env file;

[0062] Step 6: Select a traditional acoustic physical model; among which, traditional acoustic physical models include BELLHOP, KRAKEN, and RAM;

[0063] Step 7: Generate a .shd file based on the selected traditional acoustic physics model;

[0064] Step 8: Integrate the results and save them as a .mat file to complete the generation of high-fidelity sound field data;

[0065] It is worth noting that steps 1-8, when processing data, also include outlier cleaning, normalization, and data augmentation to complete the construction of a standardized training set adapted to complex terrain.

[0066] like Figure 3 As shown, in one embodiment, the acoustic propagation loss calculation model architecture based on AFNO is a frequency domain-spatial domain hybrid calculation architecture, including an input layer, a Fourier transform module, an adaptive weight module, a nonlinear activation layer, a residual connection, and an output layer.

[0067] The input layer is used to convert high-dimensional time-domain data into frequency-domain data.

[0068] The Fourier transform module is used to extract frequency domain features;

[0069] The adaptive module is used to adjust the weights in the frequency space;

[0070] The output layer is used to project the frequency domain features back into the time domain and generate the final prediction result;

[0071] Residual connections are used throughout the entire network;

[0072] The working principle and beneficial effects of the above technical solution are as follows: The Adaptive Fourier Neural Operator (AFNO) is a deep learning model that is particularly suitable for handling frequency-related scientific and engineering problems, especially in the modeling of complex physical processes, such as fluid dynamics, weather forecasting, and acoustic simulation. AFNO combines the mathematical properties of Fourier transform with the adaptive capabilities of neural networks, enabling efficient frequency domain analysis of input signals and capturing the periodicity and frequency characteristics in the data, thereby improving the modeling capabilities for high-dimensional data and complex nonlinear systems.

[0073] The acoustic propagation loss calculation model based on AFNO provided in this application includes the following key modules: an input layer, a Fourier transform module, an adaptive weight module, a nonlinear activation layer, residual connections, and an output layer. The input layer converts high-dimensional time-domain data into frequency-domain data, the Fourier transform module extracts frequency-domain features, the adaptive module adjusts the weights in the frequency space, and the output layer projects the frequency-domain features back to the time domain to generate the final prediction result. Residual connections run throughout the entire network, effectively alleviating the gradient vanishing problem and improving training stability.

[0074] AFNO employs 2D Fourier transform, with the frequency filtering parameter controlled at 32 or lower to preserve efficient frequency domain feature representation; the activation function is generally ReLU or GELU, and the optimizer is mostly Adam or AdamW, trained in conjunction with cosine annealing or exponential decay learning rate strategies; commonly used loss functions include mean squared error (MSE) or L1 loss, used to capture the difference between the prediction and the true solution.

[0075] Compared to traditional convolutional neural networks (CNNs), AFNOs have the advantage of efficiently capturing long-range dependencies and handling complex nonlinear problems. At the same time, their frequency domain operations greatly reduce computational overhead, making them an efficient, flexible, and powerful scientific computing tool. In practical applications, AFNOs can be used to solve large-scale data problems that require accurate modeling of physical laws, such as climate simulation and fluid dynamics modeling.

[0076] In one embodiment, the adaptive training process includes:

[0077] The AFNO-based sound propagation loss calculation model predicts sound propagation loss by inputting a dependent variable, where the dependent variable is high-fidelity sound field data.

[0078] During training, the frequency domain weights are dynamically adjusted using adaptive Fourier transform, and training is completed by combining mixed precision training and a dual-path loss function. The dual-path loss function is designed based on the similarity of frequency energy and spatial structure.

[0079] After each training iteration, the parameters in the model network are updated through backpropagation based on the gradient descent optimization algorithm to achieve model optimization; the core objective of training is to minimize the error between the model's predicted values ​​and the actual observed data.

[0080] The working principle and beneficial effects of the above technical solution are as follows: The AFNO model predicts sound propagation loss by input dependent variables (such as sound source depth, frequency, sound speed, etc.); the core objective of training is to minimize the error between the model's predicted value and the actual observed data, so as to ensure that the model's prediction results are as close as possible to the real physical phenomena. AFNO uses optimization algorithms such as gradient descent to update the parameters in the network through backpropagation and continuously optimize the model.

[0081] As training progresses, the model gradually learns how to identify the propagation characteristics of sound waves under different frequency and spatial conditions, and adjusts parameters to adapt to diverse acoustic scenarios. With sufficient training data, A FNO can fully learn the deep correlation between input and output, exhibiting strong generalization ability and high prediction accuracy. With the support of large-scale data, the model can extract more information from different acoustic situations and effectively adapt to new and unseen acoustic scenarios.

[0082] In one embodiment, four typical marine environments include shallow horizontal environment, shallow wedge environment, annular seamount environment, and deep horizontal environment; for different environmental characteristics, the model is trained comparatively by combining the physical laws of traditional acoustic methods (such as Bellhop).

[0083] In shallow sea environments, the multipath reflection feature modeling is optimized by jointly training the AFNO frequency domain adaptive mechanism with the traditional acoustic physics model, thereby achieving efficient calculation of sound propagation loss.

[0084] In shallow sea wedge environments, the model's adaptability to gradually changing terrain is optimized by simulating the reflection loss of the slope topography and seabed sediment parameters.

[0085] In a ring-shaped seamount environment, we learn the multipath scattering characteristics of sound waves by the seamount structure to enhance the model's accuracy in predicting the sound field of complex obstacles.

[0086] In the deep-sea horizontal environment, capture the sound ray bending effect caused by water layer stratification to ensure accurate modeling of the sound field in the convergence zone;

[0087] The beneficial effects of the above technical solution are as follows: through comparative training, the model parameters gradually fit the physical characteristics of traditional acoustic methods, realizing high-precision prediction of sound propagation loss, with a prediction error of less than 3dB to meet engineering requirements.

[0088] This invention leverages the efficient frequency domain computation capabilities of the Adaptive Fourier Neural Operator (AFNO) and combines it with the physical characteristics of traditional acoustic methods (such as Bellhop) to optimize the calculation of sound propagation loss through comparative training. The AFNO model performs efficient computation in the frequency domain using Fast Fourier Transform and dynamically adjusts the processing ratio between the frequency and spatial domains using an adaptive mechanism to adapt to the needs of complex acoustic environments. During training, the output of the AFNO model is compared with the calculation results of Bellhop, and the model parameters are optimized through a loss function to gradually fit the physical characteristics of traditional methods. Furthermore, this invention supports private model training, allowing users to customize the training model according to specific scenarios, ensuring computational accuracy and data confidentiality.

[0089] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.

Claims

1. A method for training a private large model for calculating acoustic propagation loss based on AFNO, characterized in that, include: Based on four typical marine environments, high-fidelity sound field data is generated using traditional acoustic physics models; A sound propagation loss calculation model architecture based on AFNO is constructed, and high-fidelity sound field data is used for adaptive training to generate a private large model. Set evaluation metrics and obtain test sets to complete the model evaluation of the private large model; The acoustic propagation loss calculation model architecture based on AFNO is a frequency domain-spatial domain hybrid calculation architecture, including an input layer, a Fourier transform module, an adaptive weight module, a nonlinear activation layer, a residual connection, and an output layer. The input layer is used to convert high-dimensional time-domain data into frequency-domain data. The Fourier transform module is used to extract frequency domain features; The adaptive module is used to adjust the weights in the frequency space; The output layer is used to project the frequency domain features back into the time domain and generate the final prediction result; Residual connections are used throughout the entire network.

2. The method for training a private large model for calculating acoustic propagation loss based on AFNO according to claim 1, characterized in that, The process of generating high-fidelity sound field data includes: Step 1: Define simulation parameters based on four typical marine environments; the number of simulation parameters is several, including sound source depth, frequency, and sound velocity profile; Step 2: Initialize the .env file; Step 3: Modify the frequency parameters in the outer loop; Step 4: Modify the sound source depth parameters in the inner loop; Step 5: Generate the .env file; Step 6: Select a traditional acoustic physical model; among which, traditional acoustic physical models include BELLHOP, KRAKEN, and RAM; Step 7: Generate a .shd file based on the selected traditional acoustic physics model; Step 8: Integrate the results and save them as a .mat file to complete the generation of high-fidelity sound field data.

3. The method for training a private large model based on AFNO for calculating acoustic propagation loss according to claim 2, characterized in that, Steps 1-8, when processing data, also include outlier cleaning, normalization, and data augmentation.

4. The method for training a private large model for calculating acoustic propagation loss based on AFNO according to claim 1, characterized in that, The adaptive training process includes: The AFNO-based sound propagation loss calculation model predicts sound propagation loss by inputting a dependent variable, where the dependent variable is high-fidelity sound field data. During training, the frequency domain weights are dynamically adjusted using adaptive Fourier transform, and training is completed by combining mixed precision training and a dual-path loss function. The dual-path loss function is designed based on the similarity of frequency energy and spatial structure. After each training iteration, the parameters in the model network are updated through backpropagation based on the gradient descent optimization algorithm to achieve model optimization. The core objective of training is to minimize the error between the model's predicted values ​​and the actual observed data.

5. The method for training a private large model for calculating acoustic propagation loss based on AFNO according to claim 4, characterized in that, The four typical marine environments include shallow horizontal environment, shallow wedge environment, annular seamount environment, and deep horizontal environment; for these four typical marine environments, the adaptive training process also includes comparative training: In shallow sea environments, multipath reflection feature modeling is optimized by jointly training the AFNO frequency domain adaptive mechanism with the traditional acoustic physics model. In shallow sea wedge environments, the model's adaptability to gradually changing terrain is optimized by simulating the reflection loss of the slope topography and seabed sediment parameters. In a ring-shaped seamount environment, we learn the multipath scattering characteristics of sound waves by the seamount structure to enhance the model's accuracy in predicting the sound field of complex obstacles. In deep-sea horizontal environments, the acoustic ray bending effect caused by water layer stratification is captured to ensure accurate modeling of the sound field in the convergence zone.

Images