Method for constructing an acoustic convergence zone influenced by a mesoscale eddy in the ocean based on a neural network

Abstract

The application discloses a method for constructing a sound convergence zone of a mesoscale eddy in the ocean based on a neural network, which comprises the following steps: obtaining an underwater sound speed field of the mesoscale eddy under a plurality of eddy parameters and calculating sound propagation loss under a plurality of sound source parameters; performing information coding on samples formed by the plurality of eddy parameters, the sound source parameters and the corresponding sound propagation loss to obtain training samples; building a neural network model, loading the training samples to train the neural network model to obtain an eddy-acoustic neural network model; inputting information coded from eddy parameters and sound source parameters of a target mesoscale eddy into the eddy-acoustic neural network model to predict sound propagation loss and obtain a predicted value of the sound propagation loss; and processing the predicted value of the sound propagation loss by a convergence zone calculation method to obtain sound convergence zone parameters of the target mesoscale eddy. The application aims to obtain the influence of the mesoscale eddy on the sound convergence zone and accurately obtain the variability characteristics of the convergence zone corresponding to the mesoscale eddy.

Description

Technical Field

[0001] This invention relates to the field of acoustic detection technology in ocean mesoscale eddy waters, specifically to a method for constructing acoustic clustering regions of ocean mesoscale eddy influences based on neural networks. Background Technology

[0002] Sound waves, as the only known form of energy suitable for long-distance propagation in seawater, occupy a central position in marine acoustic engineering. Their propagation characteristics directly determine the range and accuracy of underwater target detection, the efficiency of marine resource development, and the stability of underwater acoustic communication systems. They play an irreplaceable role in marine resource exploration, seabed engineering monitoring, sonar detection, and underwater information transmission, serving as a key technological support for the development and utilization of the ocean. The Acoustic Convergence Zone (CZ) is the caustic envelope region formed by the convergence of sound rays with different exit angles during propagation through refraction and reflection in the deep-sea acoustic channel. Within this region, sound energy is significantly enhanced, forming a relatively narrow but high-intensity "focusing zone." The propagation characteristics of ocean sound are closely related to the temperature and salinity structure of the marine environment, and the mesoscale eddies that are prevalent in ocean dynamics are the core dynamic factor regulating the ocean's temperature and salinity structure and thus reshaping the sound velocity field distribution. As a large-scale dynamic phenomenon accounting for 90% of the total kinetic energy of the ocean, mesoscale eddies can reach spatial scales of tens to hundreds of kilometers. By altering the vertical distribution and horizontal uniformity of seawater temperature and salinity, they directly regulate key parameters of the sound velocity profile, significantly impacting not only the direct sound propagation path and ocean acoustic channel efficiency but also becoming a core variable determining the energy focusing characteristics of acoustic convergence zones. Acoustic convergence zones, as special regions of highly concentrated energy in deep-sea sound propagation, have parameters such as distance (CZD), width (CZW), and intensity (CZS) directly related to the effective range of sonar detection and the reliability of underwater acoustic communication, making them a core concern in underwater acoustic engineering design and optimization. Especially in marine environments where mesoscale eddies are widely distributed, different types (cyclonic eddies, anticyclonic eddies) and different intensities (amplitudes) of eddies create differentiated sound velocity field distributions, leading to complex variations in convergence zone parameters. This characteristic makes the coupling relationship between mesoscale eddies and acoustic convergence zones a key research focus in the field of marine acoustics.

[0003] Currently, the prediction of ocean acoustic propagation loss and the analysis of convergence zone parameters mainly rely on traditional numerical models, including the Bellhop model based on ray theory, the Kraken model based on normal mode theory, and the RAM model based on parabolic equations. Although these models have clear physical meanings, they require extremely high precision in modeling the acoustic environment and consume massive amounts of computational resources. When mesoscale eddies induce enhanced heterogeneity in temperature and salinity structures, the spatiotemporal variability of the acoustic environment increases exponentially, leading to a sharp increase in the computational complexity of traditional numerical methods. This makes it impossible to meet the real-time prediction requirements of acoustic propagation characteristics in underwater exploration and communication scenarios, and it is also difficult to efficiently support large-sample analysis of the coupling mechanism between mesoscale eddies and convergence zones, severely restricting the effectiveness of engineering applications.

[0004] In recent years, machine learning technology has demonstrated enormous application potential in the field of marine acoustics due to its advantages of not relying on prior environmental knowledge and automatically mining correlations in multidimensional data. However, existing research still has significant technical shortcomings: on the one hand, most models focus on qualitative analysis of single sound field parameters and lack dedicated feature extraction modules designed for mesoscale eddy environments, making it difficult to accurately capture abrupt changes in sound field structure caused by eddies; on the other hand, there is a lack of quantitative analytical capabilities regarding the coupling relationship between eddies, sound sources, and convergence zones, failing to systematically reveal the differentiated control mechanisms of convergence zone parameters under different eddy conditions, and the models have poor dynamic environment adaptability, making it difficult to stably adapt to complex scenarios with different combinations of eddy intensities, polarities, and sound source parameters. Crucially, existing technologies lack dedicated sound propagation loss and convergence zone prediction models for mesoscale eddy environments, resulting in a lack of accurate and efficient technical support for underwater acoustic engineering, thus hindering improvements in underwater target detection accuracy, optimization of marine resource development efficiency, and reliability assurance of underwater acoustic communication systems. Therefore, finding a technical solution that can adapt to the dynamic environment of mesoscale eddies and accurately capture the key features of the sound field to achieve efficient prediction of sound propagation loss and convergence parameters is of great significance for improving the application level of marine acoustics engineering, and is also a technical problem that urgently needs to be solved in the field of marine acoustics. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for constructing acoustic clustering regions based on the influence of ocean mesoscale eddies, which aims to obtain the influence of mesoscale eddies on acoustic clustering regions and accurately obtain the variability characteristics of the clustering regions corresponding to mesoscale eddies.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A neural network-based method for constructing acoustic clusters of ocean mesoscale eddy influences includes the following steps:

[0008] S1, using the vortex acoustic model to obtain the mesoscale underwater acoustic velocity field of a variety of specified vortex parameters;

[0009] S2, for the sound velocity field under mesoscale eddy water, the sound propagation loss under various specified sound source parameters is calculated using the RAM acoustic toolbox;

[0010] S3 encodes information from samples consisting of various vortex parameters, sound source parameters, and corresponding sound propagation losses to obtain training samples for the neural network model.

[0011] S4. Build a neural network model and load training samples to train the neural network model so that it learns the mapping relationship between the input vortex parameters and sound source parameters, and the predicted value of the output sound propagation loss. After training is completed, the vortex-acoustic neural network model is obtained.

[0012] S5. The vortex parameters and sound source parameters of the target ocean mesoscale eddy are encoded and then input into the vortex-acoustic neural network model to predict the sound propagation loss and obtain the predicted value of the sound propagation loss.

[0013] S6. The predicted values ​​of acoustic propagation loss are processed using the convergence zone calculation method to obtain the acoustic convergence zone parameters of the target ocean mesoscale eddy.

[0014] Optionally, the vortex parameters in step S1 include vortex polarity. vortex amplitude and vortex radius vortex polarity Vortexes are classified into cyclonic vortices and anticyclonic vortices; the specified vortex parameters include those based on vortex amplitude. and vortex radius The range of variation, the interval step size, and the vortex polarity The types of vortex parameters are generated by combining different types; the sound source parameters in step S2 include the sound source frequency. Sound source depth and distance from the sound source The distance of the sound source This refers to the distance between the sound source and the center of the vortex. The specified various sound source parameters include those based on the sound source frequency. Sound source depth and distance from the sound source The range of variation and the interval step size are used to generate various sound source parameters.

[0015] Optionally, the information encoding in step S3 includes:

[0016] S3.1, vortex amplitude Normalization yields the normalized amplitude. , vortex radius Normalization yields the normalized radius. , depth of sound source Normalization yields the normalized sound source depth. The distance from the sound source Normalization yields the normalized source distance. Normalizing the sound propagation loss yields the normalized sound propagation loss. ;

[0017] S3.2, normalize the amplitude Normalized radius Normalized sound source depth Normalized sound source distance Encoding to a 64×64 matrix: for normalized amplitude and normalized radius For each sample, a constant matrix is ​​directly constructed with its scalar values ​​as constant elements, thus obtaining the amplitude constant matrix. and radius constant matrix ; Targeting normalized sound source depth and normalized distance from the sound source The range of values ​​is divided into 64 equal parts to obtain a two-dimensional grid, where the horizontal axis represents distance and the vertical axis represents depth; the position of each sample on the grid is then calculated. Thus, the sound source distance matrix corresponding to each sample is obtained. and sound source depth matrix ,in , All other elements are 0;

[0018] S3.3, the amplitude constant matrix of a single sample radius constant matrix Sound source distance matrix and sound source depth matrix The 3D input tensor of the neural network model is obtained by integration. 3D input tensor The size is 64×64×4; the normalized sound propagation loss of a single sample is calculated. The entire output is a 3D tensor label. Output tensor The size is 256×256×1, thus providing the 3D input tensor for each sample. and 3D output tensor labels To form training samples.

[0019] Optionally, in step S3.2, the position of each sample on the grid is calculated. The function expression is:

[0020] ;

[0021] ;

[0022] in, The position of the sample on the grid. The horizontal distance between the sound source and the center of the vortex. The depth of the sound source location. and These are taking the minimum value and taking the maximum value, respectively. This is the position index number of the two-dimensional grid.

[0023] Optionally, the neural network model built in step S4 consists of an encoder and a decoder, wherein the encoder is used to process the 3D input tensor. Encoding is performed to generate encoded features, and the decoder is used to decode the encoded features to generate a predicted value of the predicted acoustic propagation loss.

[0024] Optionally, the encoder includes four feature extraction units and a bottleneck layer. The first feature extraction unit is a residual convolutional block. The latter three feature units and the bottleneck layer each include a convolutional module, a batch normalization module, and a residual convolutional block. Adjacent feature extraction units are connected by downsampling modules. The decoder has four feature decoding units and an output layer. Each feature decoding unit includes a transposed convolutional module, a concatenation module, and a residual convolutional block. Adjacent feature decoding units are connected by upsampling modules. The output layer is a transposed convolutional module.

[0025] Optionally, the residual convolutional block includes a 1×1 convolutional module, two 3×3 convolutional modules, and a CBAM module. The outputs of the two 3×3 convolutional modules are both equipped with batch normalization and ReLU activation. The input features of the residual convolutional block are processed by the 1×1 convolutional module to generate a residual reference vector, which is then sequentially processed by the two 3×3 convolutional modules to perform unbiased convolution with a padding strategy of "same", batch normalization, and ReLU activation. Finally, the features are extracted by the CBAM module, added to the residual reference vector, and activated by the ReLU activation function to obtain the output features of the residual convolutional block.

[0026] Optionally, the acoustic convergence parameters of the target ocean mesoscale eddy in step S6 include the distance CZD, width CZW, and intensity CZS of the multi-order convergence region.

[0027] Optionally, the processing of the predicted sound propagation loss in step S6 using the convergence zone calculation method includes: calculating the average horizontal propagation loss above a specified depth using two-dimensional propagation loss, finding the minimum point of the average horizontal propagation loss curve, dividing the horizontal direction into intervals corresponding to multiple convergence zones, and assigning the minimum minimum point within each interval to the distance CZD of the corresponding convergence zone (in km); using the minimum minimum point within each interval as the convergence point, finding locations where the sound propagation loss is greater than a preset threshold from each convergence point to both sides, and identifying the farthest point within a specified distance from the convergence point as the boundary of that convergence zone, with the smaller one being the lower boundary and the larger one the upper boundary, and the distance between the two boundaries being the width CZW of that convergence zone (in km); within the convergence zone boundary, calculating the average propagation loss above a specified depth as the intensity CZS of that convergence zone (in dB).

[0028] Optionally, after step S6, the method further includes fitting the distance CZD, width CZW, and intensity CZS of the multi-order convergence region with respect to the vortex amplitude. vortex radius Sound source depth and distance from the sound source A cubic polynomial fit expression for four input parameters, used to determine the vortex amplitude. vortex radius Sound source depth and distance from the sound source The distance CZD, width CZW, and intensity CZS of the multi-order convergence region are directly calculated using the fitted cubic polynomial expression.

[0029] Compared with existing technologies, the present invention mainly achieves the following beneficial effects: The present invention utilizes a vortex acoustic model and a RAM acoustic toolbox to construct training samples of various vortex parameters, sound source parameters, and corresponding sound propagation losses in a numerical mode. By building a neural network model and loading the training samples, the neural network model is trained to learn the mapping relationship between the input vortex parameters and sound source parameters, and the predicted output sound propagation loss. This can effectively adapt to the dynamic environment of mesoscale vortices, accurately capture key features of the sound field to achieve efficient prediction of sound propagation loss and convergence parameters, obtain the influence of mesoscale vortices on acoustic convergence zones, and accurately obtain the variability features of the convergence zones corresponding to mesoscale vortices. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0031] Figure 2This is a schematic diagram of the vortex acoustic model in an embodiment of the present invention, where (a) is the normalized sound speed perturbation of a cyclone vortex and (b) is the normalized sound speed perturbation of an anticyclone vortex.

[0032] Figure 3 This is a schematic diagram of the distribution of sound source conditions and vortex environmental variables in an embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of the encoding matrix in an embodiment of the present invention, where (a) is the sound source distance matrix. and sound source depth matrix The structural schematic diagram, (b) is the amplitude constant matrix. and radius constant matrix A structural diagram.

[0034] Figure 5 This is a schematic diagram of the network structure of the neural network model built in an embodiment of the present invention.

[0035] Figure 6 These are three sets of samples of the sound propagation loss prediction results and convergence zone calculation results of the vortex-acoustic neural network model in this embodiment of the invention, where (a) represents the sound source depth. 70m, vortex amplitude (a) shows the vortex result at 0.05m, and (b) shows the sound source depth. 70m, vortex amplitude (c) shows the vortex results at 0.25m, and (d) shows the sound source depth. 150m, vortex amplitude The result is for the cyclone vortex at 0.25m, where (d) is the sound source depth. 70m, vortex amplitude The result for the anticyclone at 0.05m is shown in (e), where (e) represents the sound source depth. 70m, vortex amplitude The result is the anticyclone vortex at 0.25m, where (f) is the sound source depth. 150m, vortex amplitude The anticyclone result at 0.25m.

[0036] Figure 7 The following are the characteristics of the convergence zone under different vortex amplitude environments in the embodiments of the present invention, wherein (a) is the distance result of the first-order convergence zone, (b) is the distance result of the second-order convergence zone, (c) is the distance result of the third-order convergence zone; (d) is the width result of the first-order convergence zone, (e) is the width result of the second-order convergence zone, (f) is the width result of the third-order convergence zone; (g) is the intensity result of the first-order convergence zone, (h) is the intensity result of the second-order convergence zone, and (i) is the intensity result of the third-order convergence zone.

[0037] Figure 8The results of the ablation experiment of the vortex-acoustic neural network model in this embodiment of the invention are shown. Detailed Implementation

[0038] This invention aims to train a vortex-acoustic neural network using data samples generated by the RAM acoustic toolbox and vortex models, predict mesoscale vortex propagation loss using the neural network, and calculate acoustic clustering parameters based on the prediction results, thereby quickly obtaining accurate acoustic clustering variability characteristics corresponding to mesoscale vortices. To enable those skilled in the art to better understand the technical solution of this invention, the following will provide a more detailed description of the technical solution in conjunction with the accompanying drawings of the embodiments of this invention.

[0039] like Figure 1 As shown, the method for constructing acoustic clusters of oceanic mesoscale eddies based on neural networks in this embodiment includes the following steps:

[0040] S1, using the vortex acoustic model to obtain the mesoscale underwater acoustic velocity field of a variety of specified vortex parameters;

[0041] S2, for the sound velocity field under mesoscale eddy water, the sound propagation loss under various specified sound source parameters is calculated using the RAM acoustic toolbox;

[0042] S3 encodes information from samples consisting of various vortex parameters, sound source parameters, and corresponding sound propagation losses to obtain training samples for the neural network model.

[0043] S4. Build a neural network model and load training samples to train the neural network model so that it learns the mapping relationship between the input vortex parameters and sound source parameters, and the predicted value of the output sound propagation loss. After training is completed, the vortex-acoustic neural network model is obtained.

[0044] S5. The vortex parameters and sound source parameters of the target ocean mesoscale eddy are encoded and then input into the vortex-acoustic neural network model to predict the sound propagation loss and obtain the predicted value of the sound propagation loss.

[0045] S6. The predicted values ​​of acoustic propagation loss are processed using the convergence zone calculation method to obtain the acoustic convergence zone parameters of the target ocean mesoscale eddy.

[0046] The vortex parameters in step S1 include vortex polarity. vortex amplitude and vortex radius vortex polarity Vortexes are classified into cyclonic vortices and anticyclonic vortices; the specified vortex parameters include those based on vortex amplitude. and vortex radius The range of variation, the interval step size, and the vortex polarity The types of combinations generate various vortex parameters. For example, as an optional implementation, the vortex amplitude... The variation range is 0–0.3 m, with an interval of 0.05 m; vortex radius The variation range is 100–150 km, with an interval of 10 km. In step S1, the vortex acoustic model is a well-known model for calculating the mesoscale underwater sound velocity field of vortex parameters under numerical mode. Various existing vortex acoustic models can be used as needed, such as the synthetic vortex model obtained by synthetic analysis for the Kuroshio extension region.

[0047] In this embodiment, the eddies of the Kuroshio Extension (140°E–180°E, 30°N–40°N) are taken as the research object. The acoustic structure of mesoscale eddies with different amplitudes, radii, and polarities under the annual average in the Kuroshio Extension region is obtained through synthetic analysis. In step S1, when obtaining the underwater sound velocity field of mesoscale eddies under specified eddy parameters using the eddy acoustic model, the synthetic mesoscale eddy data are META, ARGO, and SODA, respectively. The eddy parameters include: eddy polarity (cyclonic, anticyclonic), eddy amplitude (0m:0.05m:0.3m), and radius (100km:10km:150km). The Mesoscale Eddy Trajectories Atlas (META 2.0) used in this embodiment is mainly for specialized research on oceanic mesoscale eddies. The original data for this product comes from daily altimeter sampling by two satellites from January 1, 1993 to March 7, 2020, and is further analyzed and processed according to the mesoscale eddy identification and tracking scheme provided by Chelton. The META2.0 mesoscale eddy dataset used in this embodiment contains sea surface characteristics of global ocean mesoscale eddies from January 1, 1993 to March 7, 2020. This includes the amplitude of the mesoscale eddy (the height difference between the extreme SLA value within the mesoscale eddy and the SLA of the profile defining the mesoscale eddy's perimeter), the type of mesoscale eddy (-1 for cyclones, +1 for anticyclones), the coordinates (latitude and longitude) of the mesoscale eddy center, the observation sequence number (the first detection date of the mesoscale eddy), the maximum average linear velocity, the radius, the observation date, and the trajectory identification number. The Argo buoy dataset used in this embodiment is a dataset released by the Coriolis Center that has undergone automatic quality control and processing. The selected dataset contains observational data of Argo buoys from May 3, 1998 to January 11, 2021, primarily including seawater pressure, temperature, salinity, conductivity, buoy number, coordinates, and other relevant measurement information. This embodiment uses SODA 3.4.2 climatological data from January 2000 to December 2019, which provides three-dimensional monthly average ocean state variables mapped to 50 vertical depth layers using a conventional 0.5° Mercator horizontal grid. The mesoscale eddy acoustic model structure obtained in step S1 of this embodiment is as follows: Figure 2 As shown, (a) represents the normalized sonic velocity perturbation of the cyclone, and (b) represents the normalized sonic velocity perturbation of the anticyclone, where R is the normalized radius (dimensionless) of the composite cyclone / anticyclone.

[0048] The sound source parameters in step S2 include the sound source frequency. Sound source depth and distance from the sound source The distance of the sound source This refers to the distance between the sound source and the center of the vortex. The specified various sound source parameters include those based on the sound source frequency. Sound source depth and distance from the sound source The variation range and interval step size are used to generate various sound source parameters. For example, as an optional implementation, the sound source frequency... Set to 100 Hz; sound source depth The range of variation is 50–250 m, with an interval of 10 m; sound source distance The variation range is -100 to 0 km, with an interval of 10 km. To better simulate the sound wave transmission characteristics under low-frequency, deep-sea, long-distance, and horizontally varying sound fields, step S2 uses the calculation results from the RAM acoustic toolbox as samples for neural network training. When calculating the propagation loss, in this embodiment, the parameters for mesoscale vortex propagation loss under specified vortex and sound source parameters calculated using the RAM acoustic toolbox in step S2 are: sound source depth (0 m: 10 m: 250 m), and distance from the sound source to the vortex center (-100 km: 10 km: 0 km) (the negative sign indicates that the sound source is behind the vortex center). The vortex parameters are the same as in step S1. In step S2, the RAM acoustic toolbox is a well-known tool for calculating sound propagation loss in numerical mode, so its implementation and usage details will not be described in detail here. In this embodiment, the depth coverage of the obtained mesoscale vortex propagation loss (numerical model) is 0–5000 m, the horizontal distance coverage is 200 km, and the grid resolution is 256×256. Figure 3 This is a schematic diagram of the distribution of sound source conditions and vortex environment variables in an embodiment of the present invention, where o is the vortex center, and the variable vortex polarity p is mainly reflected in the structure of the vortex itself, and is not a scalar.

[0049] The information encoding in step S3 includes:

[0050] S3.1, vortex amplitude Normalization yields the normalized amplitude. , vortex radius Normalization yields the normalized radius. , depth of sound source Normalization yields the normalized sound source depth. The distance from the sound source Normalization yields the normalized source distance. Normalizing the sound propagation loss yields the normalized sound propagation loss. ;

[0051] The normalization (max-min normalization) function expression used in this embodiment is as follows:

[0052] ;

[0053] in, for The normalization result, and This represents its minimum and maximum values; in addition, other normalization methods can be used as needed.

[0054] S3.2, normalize the amplitude Normalized radius Normalized sound source depth Normalized sound source distance Encoding to a 64×64 matrix: for normalized amplitude and normalized radius For each sample, a constant matrix is ​​directly constructed with its scalar values ​​as constant elements, thus obtaining the amplitude constant matrix. and radius constant matrix ; Targeting normalized sound source depth and normalized distance from the sound source The range of values ​​is divided into 64 equal parts to obtain a two-dimensional grid, where the horizontal axis represents distance and the vertical axis represents depth; the position of each sample on the grid is then calculated. Thus, the sound source distance matrix corresponding to each sample is obtained. and sound source depth matrix ,in , All other elements are 0; Figure 4 (a) in the figure represents the sound source distance matrix. and sound source depth matrix Structural diagrams (both have the same structure), (b) is the amplitude constant matrix. and radius constant matrix The structural diagrams (both have the same structure), where "Zero" represents 0;

[0055] S3.3, the amplitude constant matrix of a single sample radius constant matrix Sound source distance matrix and sound source depth matrix The 3D input tensor of the neural network model is obtained by integration. 3D input tensor The size is 64×64×4; the normalized sound propagation loss of a single sample is calculated. The entire output is a 3D tensor label. Output tensor The size is 256×256×1, thus providing the 3D input tensor for each sample. and 3D output tensor labels To form training samples.

[0056] In step S3.2, the position of each sample on the grid is calculated. The function expression is:

[0057] ;

[0058] ;

[0059] in, The position of the sample on the grid. The horizontal distance between the sound source and the center of the vortex. The depth of the sound source location. and These are taking the minimum value and taking the maximum value, respectively. This refers to the position index number of the two-dimensional grid, i.e.: .

[0060] like Figure 5 As shown, the neural network model built in step S4 consists of an encoder and a decoder. The encoder is used to process the 3D input tensor. Encoding is performed to generate encoded features, and the decoder is used to decode the encoded features to generate a predicted value of the predicted acoustic propagation loss.

[0061] like Figure 5 As shown, the encoder includes four feature extraction units and one bottleneck layer. The first feature extraction unit is a residual convolutional block. The last three feature extraction units and the bottleneck layer each include a convolutional module, a batch normalization module, and a residual convolutional block. Adjacent feature extraction units are connected by downsampling modules. The decoder has four feature decoding units and one output layer. Each feature decoding unit includes a transposed convolutional module, a concatenation module, and a residual convolutional block. Adjacent feature decoding units are connected by upsampling modules. The output layer is a transposed convolutional module.

[0062] like Figure 5As shown, the residual convolutional block includes a 1×1 convolutional module, two 3×3 convolutional modules, and a CBAM module. The outputs of the two 3×3 convolutional modules are both batch normalized and ReLU activated. The input features of the residual convolutional block are processed by the 1×1 convolutional module to generate a residual reference vector, which is then processed by the two 3×3 convolutional modules in sequence to perform unbiased convolution with a padding strategy of "same", batch normalization, and ReLU activation. Finally, the features are extracted by the CBAM module, added to the residual reference vector, and activated by the ReLU activation function to obtain the output features of the residual convolutional block. The CBAM module (Convolutional Block Attention Module) is a lightweight and general-purpose attention mechanism module. Its processing of the input feature map includes: performing global average pooling and global max pooling operations on the input feature map to obtain two 1×1×C channel description vectors; reducing the dimensionality of these vectors to C / 8 and then activating them with the ReLU activation function before increasing the dimensionality to C; summing the two vectors output from the shared fully connected layer; generating 1×1×C channel attention weights through the Sigmoid activation function; and sequentially integrating these weights with the original input feature map. The input feature map is obtained by multiplying the channels. Then, average pooling and max pooling are performed on the weighted feature map along the channel dimension to obtain two H×W×1 spatial description maps. These two H×W×1 spatial description maps are concatenated and passed through a 1×7 convolutional layer with padding using the "same" padding strategy, followed by a sigmoid activation function to generate H×W×1 spatial attention weights. These weights are then multiplied element-wise with the original input feature map to obtain the output features of the CBAM module, where C is the number of input channels, and H and W are the height and width of the input feature map. The "same" padding strategy ensures that the spatial dimensions (height and width) of the output feature map are consistent with those of the input feature map.

[0063] The downsampling module is used to perform 2×2, stride 2, padding strategy "same", unbiased convolution on the input feature map. After batch normalization and ReLU activation, the size is halved and the number of channels is doubled.

[0064] If the upsampling module uses bilinear interpolation, it doubles the feature map size using the existing 2D upsampling function UpSampling2D. If it uses transposed convolution, it performs a 2×2, stride-2, "same" padding strategy, unbiased transposed convolution followed by batch normalization and ReLU activation to double the size. The corresponding layer feature maps and upsampling results are aligned in size using the existing edge zero-padding function ZeroPadding2D, concatenated along the channel dimension, and then input into subsequent modules.

[0065] In step S4, building the neural network model specifically involves building a neural network model based on the Python language (parameters not trained), and loading training samples to train the neural network model so that it learns the mapping relationship between the input vortex parameters and sound source parameters, and the predicted value of the output sound propagation loss. This includes: S4.1, building a neural network model based on the Python language (parameters not trained); such as Figure 5 As shown, the input layer is first constructed to receive input data of the desired shape; then, the encoding path is built, and the input layer data is input into the residual convolution block (base_c=64) to obtain feature map x1, which is then passed through the downsampling module to obtain x2 (size 1 / 2, channel 128), x3 (size 1 / 4, channel 256), x4 (size 1 / 8, channel 512), and x5 (bottleneck layer, size 1 / 16, channel 1024); then, the decoding path is built, and x5 is fused with x4, x3, x2, and x1 through the upsampling module to obtain feature maps with 512, 256, 128, and 64 channels respectively; finally, the output layer is built, and the decoded feature map is transposed twice to restore it to the input size and the number of channels is reduced to 16, and then passed through 1×1 convolution and sigmoid activation to output a single-channel prediction result normalized to [0,1], thus completing the construction of the overall neural network model. S4.2, set the training parameters, including: sample split ratio, batch size, learning rate, and maximum number of training epochs. The sample split is random, with a split ratio of training set : validation set : test set = 0.85 : 0.1 : 0.05. The batch size is 64, and the initial learning rate is 2 × 10⁻⁶. -4 The maximum number of training rounds is 200; S4.3, load training samples to train the neural network, where... As input, As output, the loss function is chosen as mean squared error (MSE); S4.4, training is terminated when the deviation between the mean squared errors (MSE) of two adjacent iterations is less than the threshold ε or the number of iterations is equal to the preset maximum number of iterations maxit, and the vortex-acoustic neural network model (parameters have been trained) is obtained.

[0066] Step S5 encodes the vortex parameters and sound source parameters of the target ocean mesoscale eddy and inputs them into the vortex-acoustic neural network model to predict the sound propagation loss, obtaining the predicted value of the sound propagation loss. The neural network model can more quickly calculate the propagation loss under more complex mesoscale eddy environments and sound source conditions, thus obtaining more propagation loss samples. In this embodiment, the vortex parameters and sound source parameters of the target ocean mesoscale eddy are more numerous and denser, specifically: vortex amplitude and radius, sound source depth and distance from the vortex center. The vortex amplitude varies from 0 to 0.3 m with a step size of 0.01 m; the vortex radius varies from 100 to 150 km with a step size of 1 km; the sound source depth varies from 50 to 250 m with a step size of 1 m; and the distance between the sound source and the vortex center varies from -100 to 0 km with a step size of 1 km.

[0067] The acoustic convergence parameters of the target ocean mesoscale eddy in step S6 include the distance CZD (km), width CZW (km), and intensity CZS (dB) of the multi-order convergence zone. Step S6 involves processing the predicted sound propagation loss using a convergence zone calculation method. This includes: calculating the average horizontal propagation loss above a specified depth using two-dimensional propagation loss calculation, finding the minimum point of the average horizontal propagation loss curve, dividing the horizontal direction into intervals corresponding to multiple convergence zones, and assigning the minimum point within each interval as the distance CZD (in km) to the corresponding convergence zone. Using the minimum point within each interval as the convergence point, positions with sound propagation loss greater than a preset threshold are found on both sides of each convergence point. The point furthest from the convergence point within a specified distance is the boundary of that convergence zone, with the smaller point being the lower boundary and the larger point the upper boundary. The distance between the two boundaries is the width CZW (in km) of that convergence zone. Within the convergence zone boundary, the average propagation loss above a specified depth is calculated as the intensity CZS (in dB) of that convergence zone. Specifically, in this embodiment, the average horizontal propagation loss above a depth of 200 m is calculated using two-dimensional propagation loss calculation. The minimum points of the curves are then identified, dividing the horizontal direction into three intervals: 40–60 km, 100–130 km, and 160–180 km. The minimum points within each interval represent the distances CZD of the first, second, and third order convergence zones, respectively, in km. Using the minimum points corresponding to the distances to each convergence zone as centers, locations with sound propagation losses greater than 5 dB from the convergence point are searched to both sides. The point furthest from the convergence point within 20 km is the boundary of that order convergence zone; the smaller point is the lower boundary, and the larger point is the upper boundary. The distance between these two boundaries is the width CZW of that order convergence zone, in km. Within the convergence zone boundaries, the average propagation loss above a depth of 200 m is the intensity CZS of that order convergence zone, in dB. The convergence zone parameter results obtained in this embodiment are shown in the sample. Figure 6 As shown, (a) represents the depth of the sound source. 70m, vortex amplitude (a) shows the vortex result at 0.05m, and (b) shows the sound source depth. 70m, vortex amplitude (c) shows the vortex results at 0.25m, and (d) shows the sound source depth. 150m, vortex amplitude The result is for the cyclone vortex at 0.25m, where (d) is the sound source depth. 70m, vortex amplitude The result for the anticyclone at 0.05m is shown in (e), where (e) represents the sound source depth. 70m, vortex amplitude The result is the anticyclone vortex at 0.25m, where (f) is the sound source depth. 150m, vortex amplitude The anticyclone result at 0.25m.

[0068] As an optional implementation, step S6 further includes fitting the distance CZD, width CZW, and intensity CZS of the multi-order convergence region with respect to the vortex amplitude. vortex radius Sound source depth and distance from the sound source A cubic polynomial fit expression for four input parameters, used to determine the vortex amplitude. vortex radius Sound source depth and distance from the sound source The distance CZD, width CZW, and intensity CZS of the multi-order convergence region are directly calculated using the fitted cubic polynomial expression. The fitting results of the convergence region parameters with respect to the vortex amplitude obtained in step S7 of this embodiment are as follows: Figure 7As shown, the vortex radius was set to 100 km, the sound source was placed at the center of the vortex, and the vortex amplitude ranged from 0 to 0.3 m (where the step size of the RAM was 0.05 m and the step size of the neural network was 0.01 m). The statistical results are the average values ​​for sound source depths ranging from 50 to 200 m. The horizontal axis represents amplitude (unit: m); the first row of vertical axes represents distance (unit: km); the second row of vertical axes represents width (unit: km); and the third row of vertical axes represents intensity (unit: dB). From Figure 7 It can be observed that in a cyclonic vortex environment, as the vortex amplitude increases, the CZD (Covered Zeta) shows a monotonically decreasing trend across all orders, the CZW (Covered Zeta W) exhibits a linear expansion characteristic, and the CZS (Covered Zeta S) increases slightly. Further increasing the sound source depth will further promote the expansion of CZW and the increase of CZS. The core mechanism of this pattern lies in the fact that the divergent upwelling of the cyclonic vortex significantly enhances the negative sound velocity gradient of the upper seawater and raises the SOFAR axis (deep-sea acoustic channel axis), thus dominating the trend of convergence parameters. In contrast, in an anticyclonic vortex environment, the convergence parameters exhibit a clear near-field and far-field differentiation characteristic: as the vortex amplitude increases, the CZD still shows a monotonically increasing trend, but the CZW and CZS are mutually restrained by the dual mechanisms of "weakening of the sound velocity gradient" and "increased conjugate depth difference"—in the vortex amplitude range of 0.15–0.2 m, a shift in the dominant mechanism occurs. Its core driving mechanism is the convergent downflow of anticyclonic vortices: this process weakens the negative sound velocity gradient in the upper seawater and causes the SOFAR axis to sink, ultimately leading to differentiated responses in the convergence zone parameters. From an overall characteristic comparison, the CE corresponds to a smaller CZD (average shortening of 7.4 km), a larger CZW (average widening of 7.2 km), and a larger CZS (average increase of 2.9 dB). These two types of vortices, through differentiated modulation of the sound velocity profile and sound propagation path, ultimately result in significant differentiation in the convergence zone characteristics.

[0069] To verify the effectiveness of the vortex-acoustic neural network model in this embodiment, an ablation experiment was conducted on the vortex-acoustic neural network model. The ablation experiment results are as follows: Figure 8 As shown, "EANN" represents the result of the vortex-acoustic neural network model in this embodiment, and "Unet" represents the result of the traditional Unet network. The metric SSIM represents structural similarity, and the metric MSE represents mean squared error loss. From the scatter plot distribution characteristics, both networks exhibit a negative correlation trend: "MSE increases as SSIM decreases," indicating that the model prediction accuracy is significantly affected by the complexity of the vortex acoustic energy distribution. EANN's MSE is concentrated in the range of 0~0.4×10⁻⁶. -The SSIM in the ³ interval is close to a compact distribution of 0.95–1”, thus its numerical error in prediction results remains at a low level, while the spatial structure fidelity remains stable at a relatively good level; while the MSE distribution range of Unet is significantly expanded (0.4–1.8 × 10⁶). - (³), with SSIM dropping to a minimum of around 0.75. This characteristic intuitively reflects that Unet's performance is significantly affected by the complexity of sample features. As the nonlinearity of the sound field increases, its numerical error increases while its ability to reproduce spatial structures also decreases. EANN, on the other hand, effectively solves this problem, exhibiting more stable generalization performance. Box plot statistics further quantify the performance gap between the two models: in the MSE dimension, the average value of EANN is 2.2 × 10⁻⁶. -4 Compared to Unet's 9.1×10 -4 The error was reduced by 76%, and its box line interquartile range was narrower and the median was lower, indicating that the numerical error of EANN was significantly better than that of Unet in terms of both concentration and overall level. In the SSIM dimension, the average value of EANN was 0.96, which was 14% higher than that of Unet (0.84). Its box line interval was closer to the optimal threshold of 1, while the box line position of Unet was significantly lower and the distribution interval was wider, which fully confirms that EANN has a more outstanding ability to reproduce the fine spatial structure of mesoscale vortex sound field.

[0070] In summary, the method of this embodiment includes: obtaining the underwater sound velocity field of a mesoscale eddy under specified eddy parameters using a eddy acoustic model; calculating the mesoscale eddy propagation loss under specified eddy and sound source parameters using the RAM acoustic toolbox; encoding the eddy parameters, sound source parameters, and mesoscale eddy propagation loss to obtain neural network training samples; building and training a eddy-acoustic neural network using the Python programming language; subtracting the background sound velocity from the obtained mesoscale eddy disturbance sound velocity field to obtain sound velocity anomaly variation characteristics; inputting more and denser eddy and sound source parameters into the eddy-acoustic neural network to predict the sound propagation loss under the corresponding parameters, thus obtaining the mesoscale eddy propagation loss; processing the mesoscale eddy propagation loss results predicted by the neural network using a convergence zone calculation method to obtain the mesoscale eddy acoustic convergence zone parameters; combining the mesoscale eddy acoustic convergence zone parameters and their corresponding eddy and sound source parameters to analyze the statistical properties of the eddy-convergence zone parameters, obtaining the influence mechanism and fitting relationship of eddies on the acoustic convergence zone, and extracting the characteristics of the influence of oceanic mesoscale eddies on the acoustic convergence zone. This invention can construct universal underwater acoustic convergence zone characteristics of mesoscale eddies, providing theoretical support for the anisotropy of convergence zones under different characteristics of mesoscale eddies, providing a reference for the study of acoustic propagation effects in mesoscale eddy environments, and providing theoretical support for the study of the variability characteristics of underwater convergence zones of different mesoscale eddies.

[0071] Those skilled in the art will understand that the technical solutions provided by this invention may take the form of a method, system, or computer program product. Therefore, this invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention may take the form of a computer program product embodied on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce an implementation of the flowchart... Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for constructing acoustic clustering regions influenced by oceanic mesoscale eddies based on neural networks, characterized in that, Includes the following steps: S1. Obtain the mesoscale underwater acoustic field of a vortex under various specified vortex parameters using a vortex acoustic model. The vortex parameters include vortex polarity. vortex amplitude and vortex radius vortex polarity Vortexes are classified into cyclonic vortices and anticyclonic vortices; the specified vortex parameters include those based on vortex amplitude. and vortex radius The range of variation, the interval step size, and the vortex polarity The types of combinations generate various vortex parameters; S2, for the sound velocity field under mesoscale eddy water, the sound propagation loss under various specified sound source parameters is calculated using the RAM acoustic toolbox. These sound source parameters include the sound source frequency. Sound source depth and distance from the sound source The distance of the sound source This refers to the distance between the sound source and the center of the vortex. The specified various sound source parameters include those based on the sound source frequency. Sound source depth and distance from the sound source The range of variation and the interval step size are used to generate various sound source parameters; S3 encodes information from samples consisting of various vortex parameters, sound source parameters, and corresponding sound propagation losses to obtain training samples for the neural network model. S4. Build a neural network model and load training samples to train the neural network model so that it learns the mapping relationship between the input vortex parameters and sound source parameters, and the predicted value of the output sound propagation loss. After training is completed, the vortex-acoustic neural network model is obtained. S5. The vortex parameters and sound source parameters of the target ocean mesoscale eddy are encoded and then input into the vortex-acoustic neural network model to predict the sound propagation loss and obtain the predicted value of the sound propagation loss. S6, the predicted acoustic propagation loss is processed using a convergence zone calculation method to obtain the acoustic convergence zone parameters of the target ocean mesoscale eddy, including the distance CZD, width CZW, and intensity CZS of the multi-order convergence zones; the processing of the predicted acoustic propagation loss using the convergence zone calculation method includes: calculating the average horizontal propagation loss above a specified depth using two-dimensional propagation loss, finding the minimum point of the average horizontal propagation loss curve, dividing the horizontal direction into intervals corresponding to the multi-order convergence zones, and finding the minimum point within the interval corresponding to each order of convergence zone. CZD represents the distance of the corresponding convergence zone, in km. The minimum point within the interval corresponding to each convergence zone is taken as the convergence point. From each convergence point, the positions where the sound propagation loss is greater than the preset threshold are found on both sides. The point that is farthest from the convergence point within a specified distance range is the boundary of the convergence zone of that order. The smaller one is the lower boundary, and the larger one is the upper boundary. The distance between the two boundaries is the width of the convergence zone of that order, CZW, in km. Within the boundary of the convergence zone, the average propagation loss above the specified depth is calculated as the intensity of the convergence zone of that order, CZS, in dB.

2. The method for constructing acoustic clustering regions of ocean mesoscale eddy influences based on neural networks according to claim 1, characterized in that, The information encoding in step S3 includes: S3.1, vortex amplitude Normalization yields the normalized amplitude. , vortex radius Normalization yields the normalized radius. , depth of sound source Normalization yields the normalized sound source depth. The distance from the sound source Normalization yields the normalized source distance. Normalizing the sound propagation loss yields the normalized sound propagation loss. ; S3.2, normalize the amplitude Normalized radius Normalized sound source depth Normalized sound source distance Encoding to a 64×64 matrix: for normalized amplitude and normalized radius For each sample, a constant matrix is ​​directly constructed with its scalar values ​​as constant elements, thus obtaining the amplitude constant matrix. and radius constant matrix ; Targeting normalized sound source depth and normalized distance from the sound source The range of values ​​is divided into 64 equal parts to obtain a two-dimensional grid, where the horizontal axis represents distance and the vertical axis represents depth; the position of each sample on the grid is then calculated. Thus, the sound source distance matrix corresponding to each sample is obtained. and sound source depth matrix ,in , All other elements are 0; S3.3, the amplitude constant matrix of a single sample radius constant matrix Sound source distance matrix and sound source depth matrix The 3D input tensor of the neural network model is obtained by integration. 3D input tensor The size is 64×64×4; the normalized sound propagation loss of a single sample is calculated. Integrate into 3D output tensor labels 3D output tensor labels The size is 256×256×1, thus providing the 3D input tensor for each sample. and 3D output tensor labels To form training samples.

3. The method for constructing acoustic clustering regions of ocean mesoscale eddies based on neural networks according to claim 2, characterized in that, In step S3.2, the position of each sample on the grid is calculated. The function expression is: ； ； in, The position of the sample on the grid. The horizontal distance between the sound source and the center of the vortex. The depth of the sound source location. and These are taking the minimum value and taking the maximum value, respectively. This is the position index number of the two-dimensional grid.

4. The method for constructing acoustic clustering regions of oceanic mesoscale eddy influences based on neural networks according to claim 1, characterized in that, The neural network model built in step S4 consists of an encoder and a decoder. The encoder is used to process the 3D input tensor. Encoding is performed to generate encoded features, and the decoder is used to decode the encoded features to generate a predicted value of the predicted acoustic propagation loss.

5. The method for constructing acoustic clustering regions of ocean mesoscale eddy influences based on neural networks according to claim 4, characterized in that, The encoder includes four feature extraction units and a bottleneck layer. The first feature extraction unit is a residual convolutional block. The last three feature extraction units and the bottleneck layer each include a convolutional module, a batch normalization module, and a residual convolutional block. Adjacent feature extraction units are connected by downsampling modules. The decoder has four feature decoding units and an output layer. Each feature decoding unit includes a transposed convolutional module, a concatenation module, and a residual convolutional block. Adjacent feature decoding units are connected by upsampling modules. The output layer is a transposed convolutional module.

6. The method for constructing acoustic clustering regions of ocean mesoscale eddy influences based on neural networks according to claim 5, characterized in that, The residual convolutional block includes a 1×1 convolutional module, two 3×3 convolutional modules, and a CBAM module. The outputs of the two 3×3 convolutional modules are both batch normalized and ReLU activated. The input features of the residual convolutional block are processed by the 1×1 convolutional module to generate a residual reference vector, which is then processed by the two 3×3 convolutional modules in sequence. The residual vector is then subjected to unbiased convolution with a padding strategy of "same", batch normalization, and ReLU activation. Finally, the features are extracted by the CBAM module, added to the residual reference vector, and activated by the ReLU activation function to obtain the output features of the residual convolutional block.

7. The method for constructing acoustic clustering regions of ocean mesoscale eddy influences based on neural networks according to claim 1, characterized in that, Step S6 is followed by fitting the distance CZD, width CZW, and intensity CZS of the multi-order convergence region with respect to the vortex amplitude. vortex radius Sound source depth and distance from the sound source A cubic polynomial fit expression for four input parameters, used to determine the vortex amplitude. vortex radius Sound source depth and distance from the sound source The distance CZD, width CZW, and intensity CZS of the multi-order convergence region are directly calculated using the fitted cubic polynomial expression.

Images