An underwater sound field prediction method, medium and system based on multi-source data fusion and agent architecture

Abstract

The application provides a kind of underwater sound field prediction method, medium and system based on multi-source data fusion and agent architecture, belongs to underwater sound field prediction technical field, the application is processed through the abnormal value of marine observation data, high-dimensional multi-parameter coupling data is mapped to low-dimensional manifold space by adopting manifold dimension reduction algorithm and is interpolated and expanded, the station environment data is reconstructed using a physical constraint variational encoder, the reconstructed data is constructed as a graph structure to input the marine environment graph reconstruction model to generate a three-dimensional marine environment field by topological interpolation prediction. An underwater sound field prediction agent is constructed, the quantitative features of the three-dimensional marine environment field are extracted, the heterogeneous computing model is adaptively scheduled based on the preset strategy, and finally the original calculation results of each model are integrated and converted into the standardized sound field data unified in the system. The system realizes high-fidelity reconstruction of the marine environment field and adaptive optimization of the sound field prediction model, and completes the full-link automation from data processing to accurate prediction.

Description

Technical Field

[0001] This invention belongs to the field of underwater acoustic field prediction technology. Specifically, it relates to an underwater acoustic field prediction method, medium, and system based on multi-source data fusion and intelligent agent architecture. Background Technology

[0002] Underwater sound field prediction is a key technology in the field of marine acoustics. Traditional methods acquire environmental data such as sound velocity profiles, temperature, salinity, depth, and ocean currents from marine observation buoys, and then use ray acoustics or normal mode models to calculate the sound field distribution. Existing technologies mainly use spatial interpolation methods such as Kriging interpolation and inverse distance weighting to augment data from sparse stations, and combine them with parabolic or wave equation solvers for sound field calculations. These methods are widely used in underwater target detection, submarine communication, and marine monitoring. However, in current marine observation networks, due to the limited density of station deployment and the strong nonlinear coupling characteristics of marine environmental parameters, traditional Euclidean spatial interpolation methods struggle to accurately capture the complex relationships between multiple parameters such as temperature, salinity, sound velocity, and ocean currents. In regions with drastic changes in sound velocity gradients, such as strata, non-physical oscillations and discontinuities are easily generated, and the reconstructed environmental field often violates fundamental physical constraints such as mass conservation and energy conservation. In other words, current underwater acoustic field prediction technologies face two main bottlenecks: First, at the data foundation level, measured marine environmental data are typically sparsely distributed, limiting the accuracy of reconstructing the three-dimensional environmental field (especially the sound velocity profile). Traditional interpolation or assimilation methods struggle to accurately reconstruct small- to medium-scale marine features, resulting in inherent errors in the input data for acoustic field prediction, directly impacting the accuracy and reliability of the final prediction results. Second, at the tool and platform level, current mainstream underwater acoustic computational models (such as Bellhop, RAM, and Kraken) are mostly independent software, lacking interoperability. In practical applications, users must manually select models, prepare input files, and perform calculations independently based on their professional experience. This process is cumbersome and highly dependent on manual intervention, lacking a unified platform capable of intelligent scheduling and automated execution based on environmental characteristics and task requirements. This leads to low prediction efficiency and a high technical barrier. Summary of the Invention

[0003] In view of this, the present invention provides an underwater acoustic field prediction method, medium and system based on multi-source data fusion and intelligent agent architecture, which can solve the technical problem of insufficient accuracy of underwater acoustic field prediction due to low accuracy of marine environmental field reconstruction under sparse monitoring station conditions in the prior art, and build an underwater acoustic field prediction platform that can be intelligently scheduled and automatically executed by intelligent agents according to environmental characteristics and task requirements.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides an underwater acoustic field prediction method based on multi-source data fusion and an intelligent agent architecture. This method collects acoustic velocity profile observation data, ocean current velocity field observation data, and temperature, salinity, and depth (TDM) observation data from an ocean observation buoy system. The acoustic velocity profile observation data is layered by depth; the ocean current velocity field observation data undergoes Helmholtz decomposition; and the TDM observation data undergoes outlier removal and time series alignment. A manifold dimensionality reduction algorithm is used to map the multi-parameter coupled data to a low-dimensional manifold space and perform interpolation augmentation. A physically constrained variational encoder reconstructs the station environment data. The reconstructed station environment data is then used to construct a graph-structured topological network, which is input into an ocean environment atlas reconstruction model to generate three-dimensional ocean environment field data. An underwater acoustic field prediction intelligent agent is constructed to extract the quantitative features of the three-dimensional ocean environment field. Based on a preset strategy, heterogeneous computing models such as BELLHOP, KRAKEN, or RAM are adaptively scheduled. Finally, the original calculation results of each model are integrated and converted into unified, standardized acoustic field data within the system. This invention constructs an underwater acoustic field multi-agent collaborative system, which realizes high-fidelity reconstruction of the marine environmental field and adaptive optimization of the acoustic field prediction model, and completes the full-link automation from data processing to accurate prediction.

[0005] Specifically, the ocean current velocity field observation data is decomposed into irrotational potential flow components and divergent eddy current components. The velocity potential function and stream function are extracted and scalar interpolation is performed. The interpolated potential flow velocity components and eddy current velocity components are obtained by calculating the gradient of the interpolated velocity potential function and the curl of the interpolated stream function.

[0006] Among them, the manifold dimensionality reduction algorithm adopts the parameterized uniform manifold approximation and projection method. It constructs a weighted adjacency graph between high-dimensional data points to represent the local topological structure of the data, and uses the cross-entropy loss function to optimize the low-dimensional embedding so that it maintains the local neighborhood relationship in the high-dimensional space.

[0007] The physical constraint variational encoder includes an encoder network and a decoder network. During training, it simultaneously optimizes the reconstruction loss and the physical constraint loss. The physical constraint loss includes a mass conservation term constrained by the divergence norm of the reconstruction velocity field and an energy conservation term constrained by the total energy deviation of the temperature and salinity field.

[0008] Among them, the marine environment map reconstruction model is a multi-layer graph attention network architecture. The graph attention layer contains multiple attention heads that learn the information propagation weights between nodes and their neighbors. The physical encoder module encodes the geostrophic equilibrium equation and the hydrostatic equilibrium equation into global feature vectors of the graph.

[0009] In the training dataset for the marine environmental map reconstruction model, 60% to 80% of the station location coordinates are randomly selected as input nodes, and the remaining station location coordinates are used as prediction target nodes. The Pearson correlation coefficients of temperature and salinity fields between node pairs are calculated as the correlation components of the edge weight parameters.

[0010] In the training of the marine environmental map reconstruction model, the loss function is defined as a weighted sum of data fitting loss, physical constraint loss and contrastive learning loss, and a dynamic weight adjustment strategy is adopted to automatically balance the weight coefficients according to the magnitude of each loss term.

[0011] Among them, the calculation of the physical consistency adjustment function value is achieved by extracting sound velocity profile data to calculate the number of statistical gradient anomaly layers in the vertical sound velocity gradient distribution, extracting temperature and salinity field data to calculate the number of statistical stability anomaly layers in the density field distribution, and extracting ocean current velocity field to calculate the number of statistical conservation anomaly points in the divergence field and vorticity field.

[0012] The intelligent agent analyzes the input three-dimensional mesh data (sound speed, water depth, and seabed topography) and extracts core features through mathematical calculations: calculating the average water depth and variance of the region; analyzing the vertical gradient structure of the sound speed profile, quantifying the intensity of the cascade and the characteristics of the sound channel; and evaluating the horizontal inhomogeneity index of the sound speed field. These calculations transform the complex field data into a set of quantitative and comparable feature scalars, providing a basis for model decision-making.

[0013] The matching strategy is based on a pre-built expert knowledge base, which defines the mapping rules between different feature combinations (such as "shallow sea-strong negative gradient-high frequency") and the optimal computational model (such as BELLHOP). The decision engine compares the real-time feature vectors with the rule entries in the knowledge base, performs fast search and matching, and outputs a definite model invocation instruction, thus realizing rule-based automated selection.

[0014] The system standardizes and encapsulates models such as Bellhop, Kraken, and RAM. Upon invocation, the agent automatically generates input files (e.g., .env, .prm) conforming to the selected model's format requirements based on the policy matching results, and then invokes the corresponding executable program. The core logic is "scenario-specific model selection": Bellhop theory is suitable for high-frequency / deep-sea scenarios; Kraken mode theory is suitable for low-frequency / horizontally uniform scenarios; and RAM is used to handle complex horizontally varying problems.

[0015] The standardization module designs a dedicated resolver for the raw outputs (sound ray trajectories, modal amplitudes, and complex sound pressure fields) of each model. Through coordinate mapping, unit unification, and interpolation algorithms, all results are transformed and resampled onto a standard 3D spatial grid defined within the system, ultimately outputting a unified propagation loss (dB) matrix in absolute coordinates. This logic ensures the comparability of heterogeneous results and the consistency of interfaces with downstream applications.

[0016] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the aforementioned underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture.

[0017] A third aspect of the present invention provides an underwater acoustic field prediction system based on multi-source data fusion and intelligent agent architecture, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0018] This invention maps high-dimensional ocean data to a low-dimensional manifold space using a manifold dimensionality reduction algorithm and performs interpolation in this space, avoiding the curse of dimensionality in traditional Euclidean space interpolation. It effectively captures the nonlinear coupling relationships of parameters such as temperature, salinity, sound velocity, and ocean currents, while preserving the intrinsic geometric structure of the data. A physically constrained variational encoder embeds mass and energy conservation constraints during the reconstruction process, ensuring that the reconstructed data conforms to ocean physical laws. The ocean environment map reconstruction model achieves adaptive propagation of information between stations through a graph attention mechanism. The physical encoder module encodes geostrophic equilibrium and hydrostatic equilibrium equations as global features, making the prediction results naturally satisfy dynamic constraints. A contrastive learning strategy enhances the model's generalization ability under different station distributions. A pre-defined strategy matching principle enables adaptive selection of the target model from multiple heterogeneous underwater acoustic field prediction models. In summary, this invention solves the technical problem mentioned in the background art, where low accuracy of ocean environment field reconstruction under sparse station conditions leads to insufficient accuracy in underwater acoustic field prediction, and provides adaptive optimization of acoustic field prediction models. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2 This is a system architecture diagram from Example 2.

[0021] Figure 3 This is a flowchart of the core workflow of the system in Example 2.

[0022] Figure 4 This is a spatial distribution diagram of sound field propagation loss in Example 2.

[0023] Figure 5 This is a cross-sectional view of the sound velocity in Example 2.

[0024] Figure 6 This is a diagram showing the assessment results of the rejection distribution pattern in the convergence zone in Example 2.

[0025] Figure 7 This is a graph showing the propagation loss distribution pattern assessment results in Example 2. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a flowchart of an underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture provided by the first aspect of the present invention. This method includes the following steps:

[0028] S01. Collect multi-source marine environmental observation data from the marine observation buoy system. The multi-source marine environmental observation data includes sound velocity profile observation data, ocean current velocity field observation data, and temperature, salinity, and depth (TSD) observation data. The sound velocity profile observation data is layered according to the depth direction into surface mixed layer sound velocity data, interlayer sound velocity data, and deep layer sound velocity data. The ocean current velocity field observation data is decomposed by Helmholtz to obtain potential flow component data and eddy current component data. The TSD observation data is processed by outlier removal and time series alignment to obtain aligned TSD data.

[0029] S02. High-dimensional data compression is performed on the surface mixed layer sound velocity data, the interlayer sound velocity data, the deep layer sound velocity data, the potential flow component data, the eddy current component data, and the aligned temperature, salinity, and depth data using a manifold dimensionality reduction algorithm. The multi-parameter coupled data of temperature, salinity, sound velocity, and ocean current velocity are mapped to a low-dimensional manifold space. In the low-dimensional manifold space, the data between stations are interpolated and expanded to obtain expanded manifold space data. The expanded manifold space data is then encoded and decoded and reconstructed using a physical constraint variational encoder to obtain the reconstructed station environment data.

[0030] S03. Construct the reconstructed station environmental data into a graph structure topology network, using the station location coordinates as graph nodes, and constructing edge weight parameters using spatial distance parameters and data correlation parameters between stations. Input the data into the marine environmental map reconstruction model for topological interpolation prediction, and generate complete three-dimensional marine environmental field data in the target prediction area.

[0031] S04. The agent initiates the feature extraction engine to perform multi-dimensional quantitative analysis of the input field, calculates spatial statistical features, extracts the average water depth and standard deviation of the calculation area, and compares them with a threshold of 200 meters to label "shallow sea" or "deep sea" scenarios. Subsequently, the sound velocity field structure is analyzed: the sound velocity profile along the main axis of the sound source-receiver is analyzed, its vertical gradient is calculated, and the maximum positive / negative gradient values ​​and their depths are identified to quantify the intensity of structures such as surface sound channels or deep-sea sound channels. The variance of gradient changes is calculated as an indicator of profile complexity. Simultaneously, horizontal non-uniformity is evaluated: at key depth layers (such as the sound source depth), the horizontal gradient field of sound velocity is calculated, and its root mean square value is statistically analyzed as a measure of the severity of horizontal changes. Finally, all quantified features are encapsulated into a structured feature vector.

[0032] S05. The agent standardizes and encapsulates heterogeneous models such as BELLHOP, KRAKEN, and RAM. Upon invocation, the agent automatically converts the 3D environment data into a specific format configuration file required by the model, based on the input specifications of the target model, and sets physical parameters and computational grids adapted to the current scene. The core scheduling logic follows the principle of "model selection based on environment": ray models are suitable for high-frequency or deep-sea problems; normal mode models are optimized for low-frequency and horizontally uniform environments; and parabolic equation models are specifically designed to handle mid-to-long-range propagation problems with complex horizontal refraction and inhomogeneity.

[0033] S06. To unify the processing of heterogeneous outputs from various models, the agent utilizes a standardized transformation module. This module incorporates dedicated resolvers for the original results of different models (such as ray sets, modal functions, and complex sound pressure fields). Through coordinate system-one mapping, physical dimension normalization, and high-precision spatial interpolation algorithms, all results are transformed and resampled onto a standard three-dimensional spatial grid defined within the system. Ultimately, a unified gridded propagation loss field with standard metadata and coordinate reference is output, ensuring the comparability of multi-source results and providing a consistent data interface for downstream analysis and visualization.

[0034] The specific steps of performing Helmholtz decomposition on the ocean current velocity field observation data include: decomposing the ocean current velocity field observation data into irrotational potential flow components and divergent eddy current components; extracting the velocity potential function from the irrotational potential flow component and the stream function from the divergent eddy current component; performing scalar interpolation on the velocity potential function and the stream function at the station location coordinates; obtaining the interpolated potential flow velocity component by calculating the gradient of the interpolated velocity potential function; obtaining the interpolated eddy current velocity component by calculating the curl of the interpolated stream function; and reconstructing the complete ocean current velocity field by vector superposition of the interpolated potential flow velocity component and the interpolated eddy current velocity component. The reconstructed complete ocean current velocity field strictly satisfies the physical conservation conditions of divergence and irrotation. The potential flow component data is the interpolated potential flow velocity component, and the eddy current component data is the interpolated eddy current velocity component.

[0035] The principle of the manifold dimensionality reduction algorithm is to perform nonlinear dimensionality reduction on high-dimensional ocean data by using a parameterized uniform manifold approximation and projection method. The parameterized uniform manifold approximation and projection method represents the local topological structure of the data by constructing a weighted adjacency graph between high-dimensional data points. The cross-entropy loss function is used to optimize the low-dimensional embedding so that it maintains the local neighborhood relationship in the high-dimensional space. The Euclidean distance of the data points in the low-dimensional manifold space is approximately equal to the geodesic distance of the data points along the manifold in the high-dimensional space. The parameterized mapping function realizes the fast dimensionality reduction projection of the new data points.

[0036] The physical constraint variational encoder comprises an encoder network and a decoder network. The encoder network maps high-dimensional ocean data to low-dimensional latent variables, and the decoder network reconstructs high-dimensional ocean data from the low-dimensional latent variables. During training, the reconstruction loss and physical constraint loss are optimized simultaneously. The physical constraint loss includes a mass conservation term and an energy conservation term. The mass conservation term is constrained by calculating the divergence norm of the reconstructed velocity field, and the energy conservation term is constrained by calculating the total energy deviation of the temperature-salinity field. The physical constraint loss term is added to the total loss function using the Lagrange multiplier method.

[0037] Manifold dimensionality reduction algorithms solve the curse of dimensionality problem when interpolating high-dimensional sparse data in traditional Euclidean space. By mapping the data to a low-dimensional manifold space, the nonlinear coupling relationship between marine environmental parameters is effectively captured, avoiding the high computational complexity of solving the covariance matrix. Interpolation operations in the low-dimensional manifold space can maintain the intrinsic geometric structure and physical continuity of the data. The reconstructed high-dimensional data satisfies both the station observation constraints and the laws of marine physics, significantly improving the interpolation accuracy and computational efficiency of environmental fields under sparse station conditions.

[0038] The marine environmental mapping reconstruction model is structured as a multi-layer graph attention network architecture. The input layer receives graph node feature matrices and adjacency matrices. The graph node feature matrices include the longitude, latitude, and depth of the station location coordinates, as well as temperature, salinity, sound speed, and ocean current speed parameters. The edge weight parameters of the adjacency matrix are jointly determined by the spatial distance parameters and data correlation parameters between nodes. The graph attention layer contains multiple attention heads, each of which learns the information propagation weights between a node and its neighboring nodes. The feature information of neighboring nodes is adaptively aggregated through the attention mechanism. The physical encoder module encodes the geostrophic equilibrium equation and the hydrostatic equilibrium equation into a global feature vector of the graph. The global feature vector of the graph is concatenated and fused with the node features in each layer of graph convolution. After three layers of graph attention convolution, the updated node feature representation is output. Virtual nodes are added at the target prediction area location. The virtual nodes obtain predicted features through attention aggregation with surrounding real station nodes. The output layer decodes the predicted features of the virtual nodes to obtain the predicted values ​​of marine environmental parameters at the target prediction area location.

[0039] The steps for establishing the training dataset for the marine environmental mapping reconstruction model specifically include: collecting marine observation data from many years of history as the original data source; extracting station location coordinates and corresponding marine environmental parameters for each time slice; randomly selecting 60% to 80% of the station location coordinates as input nodes and the remaining station location coordinates as prediction target nodes; calculating the spatial distance parameters between all node pairs when constructing the graph structure topology network; connecting node pairs with spatial distance parameters smaller than the correlation scale as edges, wherein the correlation scale is determined according to the mesoscale eddy characteristic scale of the sea area; calculating the Pearson correlation coefficients of the temperature field and salinity field between node pairs as the correlation component of the edge weight parameter; the edge weight parameter is obtained by weighted summation of the distance component and the correlation component; normalizing the data from different sea areas to unify their numerical range; and generating training sample pairs containing the graph structure topology network, node features, edge weight parameters, and prediction targets.

[0040] The specific steps of training the marine environment map reconstruction model include: initializing the weight parameters of the graph attention network using the Xavier initialization method; defining the loss function as a weighted sum of data fitting loss, physical constraint loss, and contrastive learning loss; the data fitting loss being the mean square error between the predicted and actual values; the physical constraint loss including a sound velocity gradient constraint term and a density gradient constraint term; the sound velocity gradient constraint term being measured by calculating the deviation between the vertical gradient of the predicted sound velocity field and the gradient of the empirical formula; and the density gradient constraint term being measured by calculating the deviation between the density gradient corresponding to the predicted temperature and salinity field and the hydrostatic equilibrium condition using the state equation; and the contrastive learning loss being measured by narrowing the gap between the predicted and actual values. The network weight parameters are updated using the feature representation of the prediction results of different station subsets under sea state. The learning rate is initially set to 0.001 and decays exponentially with the training rounds. Each training batch contains 16 to 32 map samples. During training, a dynamic weight adjustment strategy is used to automatically balance the weight coefficients according to the magnitude of each loss term. When the physical constraint loss decreases by less than 1% for 10 consecutive rounds, the weight coefficient of the physical constraint loss is increased. When the data fitting loss decreases by less than 1% for 10 consecutive rounds, the weight coefficient of the data fitting loss is increased. The training termination condition is that the validation set loss no longer decreases for 20 consecutive rounds or the maximum training round of 500 rounds is reached.

[0041] The marine environmental atlas reconstruction model achieves adaptive propagation and fusion of information between stations through a graph attention mechanism. Compared with traditional spatial interpolation methods, it can better capture the spatial heterogeneity and anisotropy of the marine environmental field. By introducing a physical encoder module, the marine dynamic equations are embedded into the graph neural network, so that the model prediction results naturally satisfy the geostrophic equilibrium and hydrostatic equilibrium constraints. The contrastive learning strategy enhances the robustness and generalization ability of the model under different station distribution patterns. The dynamic weight adjustment strategy effectively solves the gradient conflict problem between data-driven loss and physical constraint loss, enabling the model to accurately fit the observation data and maintain physical consistency, significantly improving the accuracy and reliability of marine environmental field reconstruction under sparse station conditions.

[0042] The calculation steps for the physical consistency adjustment function value are as follows: extracting sound velocity profile data from the underwater sound field prediction results data, calculating the vertical sound velocity gradient distribution, and statistically analyzing the absolute values ​​of the sound velocity gradients in the vertical sound velocity gradient distribution that are greater than 0.05. The number of depth layers is recorded as the gradient anomaly layer number. Temperature and salinity field data are extracted from the underwater acoustic field prediction data. Density field distribution is calculated using the equation of state. The vertical stability of the density field distribution is checked, and the number of depth layers with density inversions is recorded as the stability anomaly layer number. The ocean current velocity field is extracted from the underwater acoustic field prediction data to calculate the norm distribution of the divergence and eddy current fields. The divergence norm greater than a certain value is counted. or vorticity norm greater than The number of grid points is denoted as the number of conserved anomalies. The number of gradient anomaly layers, the number of stability anomaly layers, and the number of conserved anomalies are divided by the corresponding total number of layers or the total number of grid points to obtain three normalized anomaly ratios. The weighted average of the three normalized anomaly ratios is used to obtain the comprehensive anomaly degree. The value of the physical consistency adjustment function is 1 minus the comprehensive anomaly degree.

[0043] The core of intelligent agents processing three-dimensional marine environmental fields lies in transforming continuous physical fields into discrete, computable decision features. The principle is to treat the ocean as a multivariable coupled dynamic system, with acoustic characteristics implicit in the spatial distribution and gradient changes of hydrological parameters. The process begins with structural deconstruction of the input high-fidelity three-dimensional mesh data (including sound velocity, temperature, salinity, water depth, and seabed topography). The agent first performs spatial statistics and scene classification, calculating the average water depth and its standard deviation for the entire simulation area, and clearly delineating basic acoustic scenes such as deep sea or shallow sea based on water depth thresholds. Then, vertical structural analysis is performed, extracting sound velocity profiles along the principal axis of sound propagation. By analyzing the curvature and inflection points of these profiles with depth, the agent automatically identifies and quantifies the intensity, depth range, and gradient steepness of key structures such as surface acoustic channels, deep-sea acoustic channels, or homogeneous layers. Simultaneously, the horizontal inhomogeneity assessment module is activated. On a selected representative depth plane, it calculates the density of sound velocity contour lines and the spatial variation of gradient vectors, thereby quantifying the potential refraction impact of mesoscale processes such as ocean eddies and fronts on the sound wave propagation path. Ultimately, all these analytical results—including water depth category, profile complexity index, and intensity of horizontal variation—along with the mission's own acoustic frequency and bandwidth parameters, are encoded into a structured multidimensional feature vector. This vector no longer represents a specific physical field but rather a highly condensed digital summary of the overall environmental acoustic propagation conditions, providing a precise and computable basis for subsequent model selection.

[0044] The principle behind the intelligent agent's model scheduling decision-making lies in establishing a dynamic reasoning system that maps "environmental features" to "optimal computing tools" to simulate and surpass expert experience. This is achieved through a hybrid decision engine that integrates logical reasoning based on explicit knowledge with implicit performance prediction based on historical data. In implementation, rule matching is performed first: the engine compares the aforementioned feature vectors with a pre-built expert knowledge base. This knowledge base encapsulates classic theories and practical experience in acoustics in the form of "if-then" rules, such as "if the sea is shallow and the frequency is low, then the normal mode theory is more applicable." This step provides fast and interpretable preliminary candidate solutions. In parallel, a data-driven performance prediction model is activated. This model, trained on a large number of historical computational cases, learns more complex, non-linear feature-performance relationships, such as predicting the ratio of computational accuracy to time consumption of a model under a specific combination of tiered terrain and complex topography. The decision core fuses and arbitrates the outputs of these two paths based on confidence. In "accuracy-first" mode, the prediction model has a higher weight; in "fast response" mode, the result from the rule base may be directly adopted. Once the optimal model (such as the parabolic equation model RAM) is selected, the agent immediately initiates automated configuration and invocation. Based on the specific data format and parameter template required by the model, it automatically converts the 3D environmental field data into a standard input file, sets the computational grid and boundary conditions, and seamlessly starts the computation process through an encapsulated backend interface. This instantly transforms abstract decisions into concrete computational tasks, achieving a closed loop from environmental cognition to model execution without human intervention.

[0045] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the aforementioned underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture.

[0046] A third aspect of the present invention provides an underwater acoustic field prediction system based on multi-source data fusion and intelligent agent architecture, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0047] The specific implementation methods of the above steps are described in detail below.

[0048] The specific implementation of step S01 involves reading real-time multi-source marine environmental observation data from the data acquisition module of the marine observation buoy system distributed in the forecast sea area. The data acquisition module automatically records sound velocity profile observation data, ocean current velocity field observation data, and temperature, salinity, and depth observation data at time intervals of 10 to 30 minutes. First, the sound velocity profile observation data is processed in layers according to the depth dimension. Data within a depth range of 0 to 50 meters is classified as surface mixed layer sound velocity data, and data within a depth range of 50 to 200 meters with an absolute sound velocity gradient greater than 0.02 is further classified. The data were categorized as stratified sound velocity data, and data below 200 meters in depth were categorized as deep sound velocity data. Then, Helmholtz decomposition was performed on the ocean current velocity field observation data to separate the velocity field into a potential current component with zero curl and an eddy current component with zero divergence. The velocity potential function and stream function were obtained by solving the Poisson equation. Quality control was performed on the temperature, salinity, and depth observation data to remove outliers that exceeded the physically reasonable range. The physically reasonable range is a temperature of -2°C to 35°C, a salinity of 28 to 40, and a depth error of less than 5 meters. The time series from different observation equipment were timestamped to unify all data to the Coordinated Universal Time (UTC) standard. The purpose of this layered preprocessing is to organize the raw observation data in a structured manner according to the ocean physical characteristics, providing a physically constrained data basis for subsequent dimensionality reduction and interpolation processing.

[0049] The specific implementation of step S02 involves organizing the surface mixed layer sound velocity data, stratospheric sound velocity data, deep layer sound velocity data, potential current component data, eddy current component data, and aligned temperature, salinity, and depth data obtained in step S01 into a high-dimensional feature vector. Each station location corresponds to a multi-dimensional vector containing temperature parameters, salinity parameters, sound velocity parameters, and ocean current velocity parameters. A parameterized uniform manifold approximation and projection method is used to construct a k-nearest neighbor graph between the station data points. The number of neighbors k in the k-nearest neighbor graph is between 15 and 30. The high-dimensional data is mapped to a low-dimensional manifold space of 8 to 16 dimensions by optimizing the cross-entropy loss function. The cross-entropy loss function measures the difference in similarity distribution between data point pairs in the high-dimensional space and the low-dimensional space. Local weighted interpolation is performed on the blank areas between stations in the low-dimensional manifold space. The interpolation weight is determined by the geodesic distance in the manifold space; the closer the distance, the higher the weight. The larger the weight, the more expanded the manifold space data is obtained. This expanded manifold space data is then input into a physical constraint variational encoder. The encoder network consists of three fully connected layers, with 64, 32, and 16 neurons in each layer, respectively. The activation function is a modified linear unit. The decoder network structure is symmetrical to the encoder. During training, the reconstruction loss and physical constraint loss are optimized simultaneously. The reconstruction loss is the mean square error between the predicted and true values. The physical constraint loss includes velocity field divergence constraints and temperature-salinity-energy constraints. The weight coefficients of the physical constraint terms are set to 0.3 to 0.5 using the Lagrange multiplier method. The reconstructed station environmental data output by the decoder maintains the physical consistency of the original high-dimensional space. The manifold dimensionality reduction algorithm reduces the data dimensionality while preserving the nonlinear coupling relationships between marine environmental parameters, avoiding the structural information lost by traditional linear dimensionality reduction methods.

[0050] The specific implementation of step S03 involves constructing a graph-structured topological network based on the spatial location of the reconstructed station environmental data obtained in step S02. The coordinates of each station location serve as graph nodes. The great circle distance between any two nodes is calculated as the spatial distance parameter. The Pearson correlation coefficients of the temperature and salinity parameters corresponding to any two nodes are calculated as data correlation parameters. The edge weight parameter is calculated by multiplying the negative exponential function of the spatial distance parameter by the data correlation parameter. Only edges with edge weight parameters greater than a threshold of 0.1 are retained to reduce computational load. The graph node feature matrix and adjacency matrix are input into the marine environmental atlas reconstruction model. The first graph attention layer of the model contains eight attention heads. Each attention head calculates a linear transformation of node features using a learnable weight matrix, and then calculates the attention coefficient between the features of the central node and its neighboring nodes. The attention coefficient is obtained by concatenating the features. The graph attention layer is obtained by using a single-layer neural network and exponential normalization. The second and third layers contain 4 and 2 attention heads, respectively. Residual connections are used between each layer to avoid gradient vanishing. The physical encoder module encodes the Coriolis force and pressure gradient balance relationship represented by the geostrophic equilibrium equation and the vertical pressure gradient and gravity balance relationship represented by the hydrostatic equilibrium equation into a 32-dimensional global feature vector. The global feature vector and the feature of each node are concatenated in the channel dimension and then input into the next layer. Virtual nodes are added at the target prediction area according to the latitude and longitude grid. The virtual nodes aggregate information from the surrounding real station nodes through the graph attention mechanism. The output layer applies a two-layer fully connected network to decode the predicted features of the virtual nodes to obtain the predicted values ​​of temperature parameters, salinity parameters, sound speed parameters, and ocean current speed parameters. The purpose of the graph neural network interpolation is to achieve high-precision environmental field reconstruction by utilizing the spatial correlation and physical constraints between stations.

[0051] The specific implementation of step S04 involves extracting quantized feature vectors from the three-dimensional ocean environmental field data obtained in step S03. Spatial statistical feature calculations are performed: the agent reads gridded water depth data and calculates the average water depth of the entire task area. If this value is less than 200 meters, it is marked as a "shallow sea" scene; otherwise, it is marked as a "deep sea" scene. Simultaneously, the standard deviation of the water depth is calculated to reflect the degree of topographic relief. Vertical structural analysis of sound velocity is performed: along the propagation axis from the sound source to the receiver, a profile of sound velocity variation with depth is extracted. The vertical gradient of this profile is calculated using the central difference method. The maximum positive gradient value and the maximum negative gradient value, along with their corresponding depths, are scanned and recorded. These two extreme points are used to quantify the existence and intensity of surface or deep-sea sound channels. Furthermore, the variance of the entire profile gradient is calculated as a complexity index characterizing the degree of nonlinear variation in sound velocity stratification. To assess horizontal inhomogeneity, a two-dimensional gradient of the sound velocity field in the horizontal direction is calculated at the depth layer where the sound source is located (if the sound source is underwater) or at a reference depth layer of 50 meters below the sea surface. The root mean square value of the gradient field at all grid points is then used as the core metric for the intensity of horizontal variability in the environment. After calculation, the agent encapsulates the average water depth, scene label, water depth standard deviation, extreme values ​​of the sound velocity gradient and their corresponding depths, profile complexity index, horizontal inhomogeneity index, and the sound wave frequency set for the task into a structured nine-dimensional feature vector, which is then output to the decision engine.

[0052] The specific implementation of step S05 involves the agent automatically selecting and calling the optimal sound field calculation model based on the feature vector. After receiving the feature vector, the decision engine initiates hybrid policy matching: first, it queries an expert rule base based on an XML architecture, which defines explicit rules such as "if the scene is 'deep sea' and the frequency is higher than 1000 Hz, then the ray model is selected first," for rapid filtering; simultaneously, a lightweight gradient boosting decision tree prediction model is triggered in parallel. This model, trained on historical datasets, can predict the prediction accuracy and efficiency scores of each candidate model under the current features. The decision core performs a weighted fusion of the rule matching results and prediction scores, and finally selects the execution model. Subsequently, the agent executes the standardized encapsulated model call: based on the selected model (such as BELLHOP, KRAKEN, or RAM), it calls the corresponding configuration file template. For example, for a RAM model, the agent automatically clips the three-dimensional sound velocity field into two-dimensional slices along the propagation principal axis, generating an .env environment file that conforms to the RAM format; it generates a .bty seabed topography file based on the water depth data in the feature vector; and it automatically calculates and sets the initial angle, grid step size, and maximum distance parameters required for the parabolic equation calculation based on the propagation distance and frequency. After all input files are ready, the agent automatically executes the corresponding model executable program (such as ram.exe) in silent mode through the encapsulated background command-line interface and submits the calculation task to the system's calculation queue for management.

[0053] The specific implementation of step S06 involves uniformly converting the raw results generated by each model into the system's standard format. After the agent detects that the model calculation is complete, it activates a dedicated resolver: For the .ray file output by the BELLHOP model, the resolver reads the coordinate sequence and amplitude of each sound ray, and uses a sound ray density inversion algorithm to interpolate the discrete sound ray trajectories onto a regular distance-depth grid, initially calculating the propagation loss; for the .mod file output by the KRAKEN model, the resolver extracts the mode functions, horizontal wavenumbers, and coefficients of each normal mode, resynthesizes the complex sound pressure field at a set receiving distance according to the normal mode summation formula, and then converts it into propagation loss; for the binary .shd file output by the RAM model, the resolver directly reads its complex sound pressure matrix. Subsequently, a unified grid reconstruction is performed: all the initially processed propagation loss data, along with their original irregular grid coordinates, are sent to a standard resampling module. This module targets a predefined 3D spatial grid with a unified geographic projection coordinate system and a fixed resolution (e.g., 10 meters horizontally and 1 meter vertically). It employs a bilinear interpolation algorithm to resample all data onto this standard grid. Finally, the agent writes all data, grid coordinates, and metadata containing information such as data source, frequency, and water depth range into a standard format file, generating the final standardized sound field product for the system's internal systems.

[0054] It should be noted that one of the key technical ideas of this invention is to use a manifold dimensionality reduction algorithm combined with a physically constrained variational encoder to process high-dimensional sparse marine environmental data. Traditional Kriging interpolation methods face the curse of dimensionality in solving the covariance matrix in high-dimensional spaces, with computational complexity increasing cubically with the data dimension, and it is difficult to capture the nonlinear coupling relationship between multiple parameters such as temperature, salinity, and sound speed. This invention maps high-dimensional data to a low-dimensional manifold space through parameterized uniform manifold approximation and projection methods. In the manifold space, the Euclidean distance of the data points reflects the geodesic distance in the original space, preserving the intrinsic geometric structure of the data. Interpolation in the low-dimensional space avoids the high-dimensional sparsity problem. The physically constrained variational encoder forces the satisfaction of mass and energy conservation constraints during the decoding and reconstruction process, ensuring that the reconstructed data conforms to the laws of marine physics. Compared with traditional methods, this significantly improves the interpolation accuracy and computational efficiency under sparse station conditions. The second key technical approach is to use a graph neural network architecture to achieve topological interpolation prediction of the marine environmental field. Traditional spatial interpolation methods, such as inverse distance weighting or spline interpolation, only consider the geometric distance between stations, ignoring the physical correlation and anisotropy of the marine environmental field. This invention constructs station data into a graph structure and adaptively learns the information propagation weights between nodes through a graph attention mechanism. Stations with high correlation can effectively transmit information even if they are far apart. The physical encoder embeds ocean dynamic equations such as geostrophic equilibrium and hydrostatic equilibrium into the global features of the graph, so that the model prediction naturally meets basic physical constraints. The contrastive learning strategy enhances the model's generalization ability to different station distribution patterns. The dynamic weight adjustment mechanism solves the gradient conflict problem between data fitting and physical constraints, significantly improving the accuracy and physical consistency of environmental field reconstruction. The third key technical approach is to use a hybrid strategy-based intelligent scheduling mechanism for sound field model to perform sound field calculation. Traditional sound field prediction relies on manual experience to select calculation models, which has problems such as low efficiency, strong subjectivity, and difficulty in ensuring theoretical fit. This invention constructs an intelligent agent for underwater acoustic field forecasting based on a hybrid strategy. First, it performs automated feature extraction and quantification of the three-dimensional marine environmental field, generating structured feature vectors. The intelligent scheduling engine integrates rule matching based on explicit expert knowledge with performance prediction driven by historical data, adaptively selecting the theoretically optimal computational model (such as BELLHOP, KRAKEN, or RAM) through weighted decision-making. The scheduling core strictly adheres to the principle of "model selection based on context," ensuring that heterogeneous tools such as ray models, normal mode models, and parabolic equation models can be accurately invoked in scenarios where their respective theoretical advantages lie. This mechanism eliminates human error in the process, achieving a fully automated closed loop from environmental perception to model execution, significantly improving the accuracy, efficiency, and reliability of acoustic field forecasting in complex marine environments.The three key technological approaches described above work synergistically. The manifold dimensionality reduction algorithm and graph neural network jointly solve the problem of high-precision environmental field reconstruction of sparse station data, providing accurate sound velocity field input for the agent to perform underwater sound field calculations. The intelligent scheduling engine integrates rule matching with display expert knowledge to achieve a fully automatic closed loop from environmental perception to model execution. The physical consistency adjustment function dynamically adjusts the model training strategy according to the forecast results, achieving an adaptive balance between data-driven and physical constraints. The three work together to form a complete technical chain from data acquisition and preprocessing to environmental field reconstruction and sound field forecasting, which has significant advantages over traditional methods in terms of forecast accuracy, stability, and computational efficiency.

[0055] Specifically, the principle of this invention is as follows: This invention transforms the environmental field reconstruction problem under sparse station conditions into a topological inference problem combining manifold learning and graph neural networks. The manifold dimensionality reduction algorithm utilizes a parameterized uniform manifold approximation method, constructing a weighted adjacency graph to represent the local topological structure of the data, ensuring that the low-dimensional embedding maintains the geodesic distance relationship in the high-dimensional space, theoretically solving the curse of dimensionality in high-dimensional sparse data interpolation. The physically constrained variational encoder embeds conservation laws into the optimization objective through the Lagrange multiplier method, ensuring that the reconstructed data meets both observational constraints and physical consistency. The graph attention mechanism of the marine environmental atlas reconstruction model adaptively allocates information propagation weights based on the spatial distance between nodes and data correlation. The physical encoder transforms the dynamic equations into global constraints of the graph, making the model prediction mathematically equivalent to solving a boundary value problem of a partially differential equation with physical constraints. The underwater acoustic field prediction agent, by integrating a rule base built based on explicit expert knowledge, performs real-time underwater acoustic field calculation tool matching and decision-making on perceived marine environmental characteristics, autonomously driving the entire process from data analysis and model optimization to calculation execution, forming a closed-loop intelligent scheduling that requires no human intervention.

[0056] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0057] The specific implementation of step S01 is as follows: Multi-source marine environmental observation data is collected from the ocean observation buoy system. This multi-source data includes sound velocity profile observation data, ocean current velocity field observation data, and temperature, salinity, and depth (TDM) observation data. The sound velocity profile observation data is layered according to depth into surface mixed layer sound velocity data, interlayer sound velocity data, and deep layer sound velocity data. The layering criterion is determined based on the sound velocity gradient threshold; when the absolute value of the sound velocity gradient is less than 0.02... When the absolute value of the sound velocity gradient is greater than 0.1, it is determined to be a surface mixed layer. The area within the first layer is classified as a mesogenic layer, while the remaining area is classified as a deep layer. The formula for calculating the sound velocity gradient is as follows: In the formula, For the first Layer sound velocity, unit: ; For the first Layer depth, in units of ; For the first Layer sound velocity, unit: ; For the first Layer depth, in units of The observed ocean current velocity field data were subjected to Helmholtz decomposition, which decomposed the ocean current velocity field into an irrotational potential flow component and a divergent eddy current component. Let the ocean current velocity field be... Then the Helmholtz decomposition is expressed as In the formula, This is the ocean current velocity vector field, in units of... ; The velocity potential function is expressed in units of 1000 kJ / m². ; It is a vector potential, with units of . ; The potential flow component, in units of ; The eddy current component is expressed in units of 1. ; Gradient operator, unit: ; Curl operator, unit is The formula for calculating the velocity potential function extracted from the irrotational potential flow component is as follows: The formula for calculating the stream function extracted from the non-dispersion eddy current component is as follows: In the formula, For the Laplace operator, the unit is . ; The vertical component of the stream function is given by units of . ; The vertical component of the curl of the velocity field, in units of ; Here is the divergence operator, in units of Scalar interpolation was performed on the velocity potential function and stream function at the station coordinates. The interpolated potential-current velocity component was obtained by calculating the gradient of the interpolated velocity potential function, and the interpolated eddy current velocity component was obtained by calculating the curl of the interpolated stream function. The interpolated potential-current velocity component and the interpolated eddy current velocity component were then vector-superimposed to reconstruct the complete ocean current velocity field. Outlier removal was performed on the temperature, salinity, and depth (TDM) observation data using a three-standard-deviation criterion. Data points deviating from the mean by more than three standard deviations were marked as outliers and removed. Time series alignment was performed by unifying the observation times of different stations to a standard time grid using linear interpolation. The linear interpolation formula is as follows: In the formula, Interpolation time Observed values; and They are time points and Observed values; The target interpolation time, in units of ; and For adjacent observation times, the unit is 1. .

[0058] The specific implementation of step S02 is to perform high-dimensional data compression on the surface mixed layer sound velocity data, the interlayer sound velocity data, the deep layer sound velocity data, the potential current component data, the eddy current component data, and the aligned temperature, salinity, and depth data using a parameterized uniform manifold approximation and projection method. Let the high-dimensional ocean data point set be... ,in Indicates the first Data points, The original data dimension is dimensionless. The total number of data points is dimensionless. The parameterized uniform manifold approximation and projection method represents the local topological structure by constructing a weighted adjacency graph. The adjacency weight is calculated using the following formula: In the formula, For nodes and nodes The adjacency weights between them are dimensionless. For data points and The Euclidean distance between them, in units of ; This is a local scale parameter, in units of Typically, this value is taken as 0.5 times the average nearest neighbor distance. Low-dimensional embedding optimization uses the cross-entropy loss function, which is expressed as... In the formula, Cross-entropy loss, dimensionless; Data points in low-dimensional space and The similarity is dimensionless and is calculated using the following formula: ,in For low-dimensional embedded coordinates, For targets with low dimensionality and no dimension, an empirical value is 5 to 10. Data points in low-dimensional space and The Euclidean distance between them is dimensionless. The normalization constant is dimensionless and is calculated using the following formula: The summation iterates through all node pairs. Interpolation augmentation is performed on the inter-station data in a low-dimensional manifold space using radial basis function interpolation, with the interpolation function expressed as follows: In the formula, The interpolation function value; For the first One interpolation coefficient; These are radial basis functions, typically in Gaussian form. ,in Radial distance, dimensionless. is the radial basis function scaling parameter, dimensionless, with an empirical value of 0.5; The number of known data points, dimensionless; The low-dimensional coordinates of the point to be interpolated; For the first Low-dimensional coordinates of known data points. The encoder network of the physically constrained variational encoder will convert high-dimensional ocean data... Mapping to low-dimensional latent variables The decoder network extracts latent variables from low-dimensional variables. Reconstructing high-dimensional ocean data The total loss function is expressed as In the formula, The total loss function value is dimensionless. The reconstruction loss is dimensionless and is calculated using the following formula: ,in For the original high-dimensional data, To reconstruct high-dimensional data; The loss is due to physical constraints and is dimensionless. This is the weighting coefficient, dimensionless, and defaults to 1. Physical constraint losses include mass conservation terms and energy conservation terms, calculated using the following formula: In the formula, To reconstruct the velocity field, the unit is... ; The unit is velocity. The empirical value is 0.1; The unit is length. The experience value is 10000; To reconstruct the total energy of the thermo-salinity field, the unit is... ; To observe the total energy of the temperature-salinity field, the unit is... ; It is an energy scale, with units of 1000 kilometres per second. The empirical value is 1000. The formula for calculating the total energy of a thermo-salinous field is: In the formula, The density of seawater is expressed in units of 1000 kJ / m³. ; Specific heat capacity, unit: The experience value is 4000; Temperature, unit: ; For reference temperature, the unit is... The experience value is 300; This is the acceleration due to gravity, in units of 1. The value is 9.8; These are depth coordinates, in units of ; Salinity, in units of ; For reference salinity, the unit is... The experience value is 35; This is the integral volume, in units of ; For volume infinitesimal elements, the unit is . .

[0059] The specific implementation method of step S03 is the same as described above, and will not be repeated in detail here.

[0060] The specific implementation of step S04 is as follows: the agent initiates an automatic feature quantization process on the input three-dimensional meshed environmental data. First, spatial statistical feature calculation is performed: the agent reads the water depth values ​​of all grid points, and the regional average water depth... The calculation formula is the arithmetic mean of all water depth values, in meters; the standard deviation of water depth. The calculation formula is the relative water depth value. The standard deviation is expressed in meters. The scene labeling strategy is: if If the value is less than 200, it is marked as "shallow sea"; otherwise, it is marked as "deep sea". Then, a vertical structural analysis of the sound velocity is performed: the sound velocity profile is extracted along the line connecting the sound source point and the center point of the receiver array. Its vertical gradient The central difference method is used for calculation, and the formula is as follows: The unit is per second per meter Scan the entire profile and identify... The maximum positive value and its corresponding depth and the maximum negative value and its corresponding depth Profile complexity index The calculation formula is Variance across the entire depth range. Finally, a horizontal non-uniformity assessment is performed: at the sound source depth layer. Above, calculate the sound velocity field. In the East and North and South Horizontal gradient components in the direction and Horizontal gradient magnitude Intensity of horizontal change The calculation formula is The root mean square value over the entire horizontal region. Finally, the structured feature vector is output. ,in The frequency of the sound source.

[0061] The specific implementation of step S05 is that the agent, based on the feature vector... The system automatically selects and invokes the optimal sound field calculation model. The decision engine employs a hybrid strategy matching approach: its built-in expert rule base includes rules such as "if..." For 'deep sea' and frequency >1000 If the output is "BELLHOP model", then the explicit rule is "BELLHOP model". Simultaneously, a lightweight prediction model based on Gradient Boosting Decision Tree (GBDT) is invoked in parallel, with the input being... The output is a precision and efficiency score vector for the candidate model set. The decision fusion unit performs weighted voting, and its core function is... ,in Representative candidate models, To score points, This sets the rule weight (default is 0.4). After selecting a model, the normalized encapsulation call is initiated: for example, if the RAM model is selected, the agent automatically extracts two-dimensional slices from the three-dimensional sound velocity field along the propagation principal axis. Generate a .env file, formatted as "Depth (meters) Sound Speed ​​(m / s)" per line. Generate a .bty seabed file based on the water depth data, formatted as "Horizontal Distance (meters) Depth (meters)" per line. The core parameter automatic configuration strategy is: calculate the grid step size. The default is wavelength. One-tenth of the maximum calculated distance User-defined initial launch angle The settings are configured according to the task requirements. The calculations are ultimately performed in a user-friendly, headless mode via system commands.

[0062] The specific implementation of step S06 is that the agent converts the original calculation results of each model into a unified internal standard format. First, a dedicated parser is started: for the ASCII format .ray file output by the BELLHOP model, the parser reads the sequence number, start angle, and coordinate point sequence of each vocal ray according to the file structure. and transmission time The trajectory data is composed of various modes. For the .mod file output by the KRAKEN model, the parser reads the total number of modes. Horizontal wavenumbers of each mode Modal functions at discrete depth The value on and excitation coefficient For the binary .shd file output by the RAM model, the parser interprets it according to the dimensions (distance points) defined in its header file. Depth points (and data type, directly read into the complex sound pressure matrix) Subsequently, a unified grid-based reconstruction is performed: all data is transformed to a unified geographic coordinate system (such as UTM or WGS-84). The system's internal standard grid is defined as: horizontal axis Starting from the sound source point, at fixed intervals An increasing sequence (e.g., 10 meters); vertical axis For the distance from the sea surface to the maximum water depth, at fixed intervals An increasing sequence (e.g., 1 meter). Propagation loss. The unified calculation formula is ,in The reference sound pressure is located 1 meter from the sound source. A bilinear interpolation algorithm is used to convert the irregular outputs of each model... Value resampling to At the grid points. Finally, the standard grid coordinates... , Propagation loss field after resampling The data, along with metadata including the data source, frequency, and sound source depth, are written together into a NetCDF format file to complete the standardized output.

[0063] It should be noted that the variables involved in this embodiment are explained in detail in Tables 1 and 2.

[0064] Table 1. Variable Explanation Table (Part 1)

[0065]

[0066] Table 2. Variable Explanation Table (Part Two)

[0067]

[0068] To better understand and implement this invention, a specific application scenario, Example 2, is provided below: To solve the problem of underwater sound field prediction under complex hydrological conditions in a certain sea area, technicians built an underwater sound field prediction system based on multi-source data fusion and an intelligent agent architecture. The water depth in this sea area ranges from 45m to 180m, with significant seabed topography and multiple layered structures. Traditional sound field prediction methods suffer from sparse station data and strong spatiotemporal variability of environmental parameters, resulting in prediction accuracy that is difficult to meet practical application requirements. Using the technical solution of this invention, technicians achieved fully automated processing from data acquisition and processing to sound field prediction output through the collaborative work of intelligent agents.

[0069] The system architecture adopts a four-layer design, such as Figure 2 As shown, from top to bottom, the layers are the interaction layer, application layer, model / algorithm layer, and data layer. The interaction layer provides a WebUI visual interface, allowing users to upload ocean observation data and set forecast parameters. The sound source frequency is set to 150Hz, the forecast distance range is 0 to 12km, and the depth range is 0 to 180m. The application layer constructs a cluster of three intelligent agents: a data generation agent, a sound field forecasting agent, and a result analysis agent. These agents work collaboratively through an intelligent scheduler. The model / algorithm layer integrates the BELLHOP ray acoustic model, the KRAKEN normal mode model, and the RAM parabolic equation model, and is configured with an improved spatiotemporal kriging algorithm and a GAN data augmentation algorithm. The data layer establishes a marine environmental database, storing sound velocity profile data, seabed sediment parameters, and sea surface state data for the area over the past three years.

[0070] During the data acquisition phase, technicians obtained multi-source marine environmental observation data from 12 ocean observation buoy systems. Sound velocity profile data covered a depth range of 0 to 180 m, with a sampling interval of 5 m and a measurement time interval of 3 hours. Ocean current velocity field observation data was acquired using an acoustic Doppler current profiler, with a spatial resolution of 10 m and a temporal resolution of 1 hour. Temperature, salinity, and depth (CTD) observation data were acquired by a CTD sensor, with a temperature measurement range of 8°C to 24°C, a salinity measurement range of 30 to 35, and a depth measurement accuracy of 0.5 m. After outlier removal, it was found that some stations had missing data in the strata region at depths of 60 m to 90 m, with a missing rate reaching 28%.

[0071] In the data preprocessing stage, technicians first stratified the sound velocity profile observation data according to depth. The surface mixed layer corresponds to a depth range of 0 to 35 m. The sound velocity in this layer is relatively uniform, with an average sound velocity of 1518 m / s and a standard deviation of 3.2 m / s. The interlayer sound velocity data corresponds to a depth range of 35 m to 95 m. In this layer, the sound velocity gradient changes drastically, with a maximum gradient of 0.18. The negative gradient layer appears at depths between 72m and 84m. Deep sound velocity data corresponds to depths between 95m and 180m, where sound velocity increases slowly with depth, with an average sound velocity gradient of 0.012. Helmholtz decomposition was performed on the observed ocean current velocity field data to obtain potential current component data and eddy current component data. The average velocity of the potential current component at the surface is 0.35 m / s, with the main flow direction being northeast. The eddy current component is most significant near the mezzanine, with a maximum vorticity value of [value missing]. After time series alignment processing, the temporal resolution of temperature data was unified to 1 hour, and the spatial interpolation error of salinity data was controlled within 0.08.

[0072] like Figure 3As shown, the core workflow of the system includes five stages: data input preprocessing, data augmentation quality assessment, model scheduling and computation, result fusion analysis, and interactive output feedback. In the data augmentation stage, the data augmentation sub-agent in the data generation agent calls a manifold dimensionality reduction algorithm to compress the high-dimensional ocean data. Using a parameterized uniform manifold approximation and projection method, the four-dimensional parameters (temperature, salinity, sound speed, and current velocity) are coupled and mapped to a two-dimensional manifold space, preserving the original data's topological structure. Interpolation augmentation is performed on the data from 12 stations in the low-dimensional manifold space, generating environmental parameter data for 36 virtual stations. A physically constrained variational encoder encodes and decodes the augmented manifold space data, mapping the high-dimensional ocean data into 16-dimensional latent variables, and the decoder network reconstructs the complete environmental parameters from the latent variables. During training, the divergence norm constraint threshold of the mass conservation term is set to... The total energy deviation threshold of the temperature-salinity field in the energy conservation term was set to 2%, and after 120 rounds of iterative training, the reconstruction error converged to 0.6%.

[0073] In the environmental field reconstruction phase, technicians constructed a graph-structured topological network from the reconstructed station environmental data. The graph network consisted of 48 nodes, including 12 real station nodes and 36 virtual predicted nodes. The node feature matrix contained 8 dimensions: depth, temperature, salinity, sound speed, eastward current velocity, northward current velocity, potential current component, and eddy current component. The edge weight parameters of the adjacency matrix were jointly determined by spatial distance and data correlation parameters. The spatial distance parameter was calculated using a Gaussian kernel function, with a correlation scale set to 8 km. The data correlation parameter was obtained by calculating the Pearson correlation coefficient of the temperature and salinity fields between node pairs, with a correlation coefficient threshold of 0.65. The marine environmental atlas reconstruction model adopted a three-layer graph attention network architecture, with each layer containing 4 attention heads and a hidden layer dimension of 64. The physical encoder module encoded the geostrophic equilibrium equation and hydrostatic equilibrium equation into 32-dimensional global feature vectors, which were then concatenated and fused with the node features in each layer's graph convolution. The model was trained using the Xavier initialization method with an initial learning rate of 0.001 and a batch size of 24. After 280 training rounds, the validation set loss converged. The model generated complete 3D marine environmental field data in the target prediction area, with a spatial resolution of 500m×500m×5m and a coverage area of ​​12km×8km×180m.

[0074] In the feature extraction stage, the agent loads 3D meshed environmental data covering a target sea area of ​​12km × 8km × 180m with a resolution of 500m × 5m. Based on water depth data from 45m to 180m across the entire area, the agent calculates the average water depth of the region to be 112.5m, with a standard deviation of 38.7m, and automatically labels it as a "shallow sea" acoustic scene according to a 200m threshold rule. Subsequently, the vertical profile of sound velocity is extracted and analyzed along the propagation axis from the sound source to the 12km receiver. Calculations show that the vertical gradient of sound velocity exhibits a maximum negative value of -0.18 in the stratified region between depths of 55m and 70m. The gradient variance of the entire profile was calculated to be 0.0071. This reflects a moderately complex vertical layered structure. Horizontal inhomogeneity of the sound velocity field was assessed at the sound source depth (set to 50 meters), and the root mean square value of the horizontal gradient field was calculated to be 0.0053 (m / s) / m, indicating that the horizontal variation of the ocean structure at this depth is relatively gentle. Finally, the agent encapsulated all the above quantitative indicators and the sound source frequency (150Hz) into a structured feature vector [112.5, 38.7, 'Shallow', -0.18, 65, 0.0071, 0.0053, 150], fully characterizing the environmental acoustic features of the current forecasting task and providing accurate numerical input for subsequent intelligent model decision-making.

[0075] During the model scheduling phase, the agent's decision engine receives feature vectors from S04. The engine's built-in expert rule base first performs matching: rule "If the scene is 'shallow sea' and the frequency..." The message "≤500Hz, then prioritize normal mode models" is triggered, pointing to the KRAKEN model. Simultaneously, a lightweight GBDT prediction model trained on historical performance data is launched in parallel. After receiving this feature vector, it outputs accuracy and efficiency prediction scores for each candidate model, with KRAKEN scoring highest in accuracy. The decision fusion unit weights the rule matching results and prediction scores (in this example, the weights are...). (Set to 0.3), ultimately determining KRAKEN as the optimal computational model. After model selection, the agent automatically executes the encapsulation and invocation process: it extracts two-dimensional sound velocity profiles and seabed topography profiles along the main propagation direction from the three-dimensional environment field, generating .env environment files and .bty seabed topography files according to KRAKEN format requirements; based on the task distance (12km) and frequency (150Hz), it automatically configures the maximum computation distance to 15000 meters, the upper frequency limit to 200Hz, and the maximum number of modes to 200, and sets an adaptive depth mesh. After configuration, the system silently calls the kraken.exe program in the background through the subprocess module, passing in the parameter file. The entire process requires no manual operation by technical personnel, achieving closed-loop automation from environmental perception to model execution.

[0076] During the standardization phase, the agent listens to and captures the output file after the KRAKEN model calculation is completed. A dedicated parser first reads the .mod file, extracting the complex amplitude, horizontal wavenumber, and depth-dependent mode functions of the first 85 effective normal modes. Subsequently, it reads the complex acoustic pressure field data from the .shd binary file. Next, the standardized module initiates the processing flow: based on the formula... The complex sound pressure field is converted into a propagation loss field, where the reference sound pressure is... The value is taken at a distance of 1 meter from the sound source. To unify the data interface, this module resamples the original, non-uniformly distributed TL data onto a predefined standard 3D spatial grid with a resolution of 10 meters (horizontal) × 1 meter (vertical) using a bilinear interpolation algorithm. The resulting standardized sound field product clearly shows two significant sound convergence zones with propagation losses below 82 dB at distances of approximately 4.1 km and 8.7 km from the sound source, with center depths around 72 meters and 115 meters, respectively. In the near-field range of 0.8 km to 2.5 km, the sound field structure is complex due to multiple reflections from the seabed, resulting in drastic changes in the propagation loss gradient. This standardized grid data, along with complete metadata, is written into a NetCDF file, which can be directly used by downstream visualization and analysis modules. Figure 4 Two significant acoustic convergence zones were clearly identified, and their central locations (approximately 4.1 km and 8.7 km) and depths (approximately 72 m and 115 m) could be quantified, which is completely consistent with the numerical results. Meanwhile, Figure 4 The complex near-field (0.8-2.5 km) sound field structure and sound shadow zone distribution shown in the data corroborate the numerical calculation's description of "drastic changes in the propagation loss gradient." This standardized grid data, along with complete metadata, was written into a file. For example... Figure 5 As shown, the vertical sound velocity gradient distribution clearly reveals a significant negative gradient cascade structure. Calculations show that the maximum negative gradient value is -0.18. It appeared at a depth of 65 meters, which is consistent with Figure 5 The lowest point of the gradient curve precisely corresponds to this depth, which the agent quantitatively identifies as the core location of the main gradient layer. The gradient variance of the entire profile is calculated to be 0.0071. This is confirmed numerically. Figure 5 The degree of drastic change in the curve reflects the moderate complexity of the vertical stratification. Meanwhile, the horizontal inhomogeneity assessment result at the sound source depth layer is 0.0053 (m / s) / m, indicating that the horizontal change is gradual.

[0077] In the forecast results output phase, the results analysis agent extracted the distribution characteristics of propagation loss with distance and depth. For example... Figure 6 , Figure 7As shown, at a distance of 5 km and a depth of 100 m from the sound source, the propagation loss is 73 dB, located at the edge of the first convergence zone. At a distance of 10 km and a depth of 150 m from the sound source, the propagation loss is 94 dB, located in the transition region between the two convergence zones. The sound field mapping sub-agent generates propagation loss cloud maps, sound ray trajectory maps, sound velocity profile maps, and convergence zone distribution maps. All charts are displayed interactively, allowing users to view detailed parameters at any location via mouse operation. The report generation sub-agent, combining a preset template with a large ocean knowledge model, generates a forecast report including data source descriptions, environmental parameter statistics, calculation method descriptions, forecast result analysis, and uncertainty assessment. The intelligent question-answering sub-agent, based on the RAG architecture, responds to user inquiries about the impact of hierarchical structures on sound propagation. After retrieving relevant literature fragments from the domain knowledge base, it generates detailed explanations of the physical mechanisms and quantitative impact analyses.

[0078] This invention represents a significant technological advancement over traditional underwater acoustic field prediction methods. Traditional methods rely on uniform grid interpolation and fixed-parameter models, which cannot effectively handle high-dimensional nonlinear coupling problems under sparse station conditions, leading to insufficient prediction accuracy in stratified and convergent regions. This invention maps high-dimensional ocean data to a low-dimensional manifold space using a manifold dimensionality reduction algorithm, effectively capturing the nonlinear coupling relationships between temperature, salinity, sound velocity, and ocean current velocity, avoiding the curse of dimensionality problem of traditional Euclidean space interpolation. The physically constrained variational encoder simultaneously optimizes reconstruction loss and physical constraint loss during data reconstruction, ensuring that the reconstructed data satisfies both station observation constraints and the laws of mass and energy conservation. The ocean environmental atlas reconstruction model achieves adaptive propagation of information between stations through a graph attention mechanism, better capturing the spatial heterogeneity and anisotropy of the ocean environmental field compared to traditional spatial interpolation methods. The physical encoder module embeds the geostrophic equilibrium equation and hydrostatic equilibrium equation into the graph neural network, ensuring that the environmental field reconstruction results naturally satisfy ocean dynamic constraints. The symplectic geometric integral algorithm fundamentally guarantees energy conservation and phase space volume conservation during ray tracking, avoiding the systematic energy drift error accumulated by traditional numerical integration methods over long distances, and significantly improving the prediction accuracy of convergence zone location and intensity. The block-based adaptive grid strategy significantly reduces the number of grids and computational resource consumption while maintaining prediction accuracy, enabling high-frequency long-distance sound field prediction on a single-node computer. The agent-based collaborative architecture automates the entire process from data processing to model scheduling to result analysis, reducing manual intervention and reliance on experience, and improving the intelligence level and efficiency of the prediction system.

[0079] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for underwater acoustic field prediction based on multi-source data fusion and intelligent agent architecture, characterized in that, Data on sound velocity profiles, ocean current velocity fields, and temperature, salinity, and depth (TMD) were collected from an ocean observation buoy system. The sound velocity profile data was layered by depth, the ocean current velocity field data underwent Helmholtz decomposition, and the TMD data were processed for outlier removal and time series alignment. A manifold dimensionality reduction algorithm was used to map the multi-parameter coupled data to a low-dimensional manifold space and perform interpolation augmentation. A physically constrained variational encoder was used to reconstruct the station's environmental data. The reconstructed environmental data was then used to construct a graph-structured topological network, which was input into an ocean environmental atlas reconstruction model to generate three-dimensional ocean environmental field data. An underwater acoustic field prediction agent was constructed. After receiving the three-dimensional environmental field, the agent extracted multi-dimensional feature vectors, such as water depth and sound velocity profile complexity, for quantitative analysis. Based on its built-in knowledge-based and rule-based decision engine, the agent adaptively selected the optimal prediction model according to a feature matching strategy and automatically executed it. The agent then fused the heterogeneous outputs of each model into a standardized gridded propagation loss field using a unified conversion algorithm, completing the integrated computation process.

2. The underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture according to claim 1, characterized in that, The manifold dimensionality reduction algorithm employs a parameterized uniform manifold approximation and projection method. It constructs a weighted adjacency graph between high-dimensional data points to represent the local topological structure of the data, and uses the cross-entropy loss function to optimize the low-dimensional embedding so that it maintains the local neighborhood relationships in the high-dimensional space.

3. The underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture according to claim 1, characterized in that, The physical constraint variational encoder consists of an encoder network and a decoder network. During training, it simultaneously optimizes the reconstruction loss and the physical constraint loss. The physical constraint loss includes a mass conservation term constrained by the divergence norm of the reconstruction velocity field and an energy conservation term constrained by the total energy deviation of the temperature and salinity field.

4. The underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture according to claim 1, characterized in that, The marine environmental atlas reconstruction model is a multi-layer graph attention network architecture. The graph attention layer contains multiple attention heads that learn the information propagation weights between nodes and their neighbors. The physical encoder module encodes the geostrophic equilibrium equation and the hydrostatic equilibrium equation into global feature vectors of the graph.

5. The underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture according to claim 1, characterized in that, Feature extraction and quantification analysis are performed on the environmental field data to obtain multidimensional feature vectors; based on the multidimensional feature vectors and preset strategy matching principles, a target model is adaptively selected from multiple heterogeneous underwater acoustic field prediction models; the target model is automatically called and executed to obtain the original calculation results; the original calculation results are converted into unified standardized acoustic field data within the system.

6. The underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture according to claim 1, characterized in that, The optimal forecasting model is one of BELLHOP, KRAKEN, or RAM.

7. The underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture according to claim 1, characterized in that, The intelligent agent will output heterogeneous data from each model, including sound rays, modalities, and sound pressure fields.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the underwater acoustic field prediction method based on multi-source data fusion and intelligent agent architecture as described in any one of claims 1-7.

9. An underwater acoustic field prediction system based on multi-source data fusion and intelligent agent architecture, characterized in that, The system comprises the computer-readable storage medium of claim 8, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images