A deep learning-based four-dimensional sound velocity field prediction method and system

Abstract

The application provides a four-dimensional sound velocity field prediction method and system based on deep learning, which comprises the following steps: after pre-processing historical sound velocity field data, the pre-processed data is input into a trained four-dimensional sound velocity field prediction model to predict the sound velocity field; the four-dimensional sound velocity field prediction model comprises an encoder, a multi-scale convolution module and a decoder; wherein the encoder is used to learn the space-time features in the sound velocity field data and gradually reduce the spatial size of the feature image; the multi-scale convolution module is used to extract and fuse features at different spatial scales; and the decoder is used to reversely reconstruct the reduced feature image, gradually enlarge the spatial size of the feature image and obtain the final prediction result. The application has the advantages that it has the abilities of local feature mining, global data correlation and high-dimensional data analysis, can effectively learn the multi-scale dependent features in the high-dimensional sound velocity field space-time data, and can more comprehensively understand and predict the four-dimensional sound velocity field.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of underwater acoustics engineering, ocean engineering, sonar technology, and the like, and particularly relates to a four-dimensional sound speed field prediction method and system based on deep learning. BACKGROUND

[0002] Accurate prediction of the ocean sound speed field (SSF) is the basis for realizing target detection, accurate positioning, reliable communication, and other deep-sea resource development, ocean environment monitoring, and underwater autonomous navigation tasks in complex marine environments. However, a large number of and complex marine dynamic processes cause the spatiotemporal variation of hydrological environmental parameters and the physical properties of the acoustic medium, which directly affects the horizontal-interlayer reflection and refraction of sound waves in the underwater propagation process, and causes serious interference to the prediction of the marine sound speed field. In the spatial dimension, the marine sound speed field is influenced by large-scale circulation, the thermocline, mesoscale eddies, fronts, and small-scale turbulence and internal waves, and presents a complex spatial correlation. In the time dimension, the interannual variation of seasonal temperature and salinity cycles, the daily periodic fluctuations of tides, and the high-frequency disturbance of ocean internal waves are coupled with each other, causing the sound speed to have a complex time sequence relationship. Therefore, establishing a four-dimensional (three-dimensional space + time) sound speed field dynamic prediction model that can decouple the complex physical processes between dynamic hydrological parameters and sound wave propagation processes, and has high accuracy and fast response speed in complex marine environments is the key to improving the level of deep-sea exploration.

[0003] Deep learning technology has great potential in the field of sound speed field prediction due to its significant advantages in deep feature extraction, high-dimensional and large-scale data processing, outstanding computing ability, strong generalization ability, high robustness and scalability, and end-to-end optimization learning.

[0004] Recurrent neural networks have shown good performance in time series prediction tasks. Zhang et al. used long short-term memory (LSTM) to build a time series prediction model and achieved prediction of sound speed profiles in the China offshore sea area. However, single time series networks face the dual challenges of gradient vanishing and rapid growth of computational complexity, making it difficult to adapt to large-scale calculations in high-dimensional marine environments. See reference (Zhang, Q., Wang, H., Dong, J., et al. Prediction of sea surface temperature using long short-term memory[J]. IEEE Geoscience and Remote Sensing Letters, 2017, 14(10): 1745-1749.) for this method. In addition, generative models provide a new approach to sound speed field prediction from the perspective of probability distribution. Gao et al. applied a conditional diffusion model to range-dependent sound speed field prediction, and for the first time applied the conditional distribution learning ability of the diffusion model to the spatio-temporal prediction task of sound speed field. See reference (Gao, Cheng, Zhang, Li. Unveiling the spatial-temporal dynamics: Diffusion-based learning of conditional distribution for range-dependent ocean sound speed field forecasting[J]. The Journal of the Acoustical Society of America, 2024, 156 (5): 3554-3573.) for this method. Diffusion models have shown significant advantages in sound speed field reconstruction and prediction tasks due to their theoretical rigor and generation quality. However, the sampling process requires multiple iterative forward propagation, leading to increased computational complexity and challenges in real-time application scenarios. SUMMARY

[0005] The present application aims to overcome the shortcomings of existing sound speed field prediction methods in long-distance spatial correlation and multi-scale feature extraction.

[0006] To achieve the above-mentioned purpose, the present application proposes a four-dimensional sound speed field prediction method based on deep learning, which comprises:

[0007] After preprocessing the historical sound speed field data, input the trained four-dimensional sound speed field prediction model to predict the sound speed field;

[0008] The four-dimensional sound velocity field prediction model is built based on a Swin Transformer and a U-Net architecture, and includes an encoder, a multi-scale convolution module and a decoder.

[0009] The encoder is configured to learn the space-time features in the sound velocity field data and gradually reduce the spatial size of the feature image.

[0010] The multi-scale convolution module is configured to extract and fuse features at different spatial scales.

[0011] The decoder is configured to reversely reconstruct the feature map output by the multi-scale convolution module, gradually enlarge the spatial size of the feature map, and obtain the final prediction result.

[0012] As an improvement of the above method, the historical sound velocity field data includes a number of longitude points, a number of latitude points, a number of depth points and a time sequence.

[0013] As an improvement of the above method, the preprocessing includes:

[0014] In view of the seasonal variation characteristics of the marine sound velocity field, a monthly grouping normalization strategy is adopted, and the normalization process is represented as:

[0015] ;

[0016] wherein, represents the normalized sound velocity data; represents the original sound velocity data; (x, y, z) is the spatial coordinate; t is the time index; m represents the month corresponding to the time t; and are the maximum and minimum values of all samples in the mth month, respectively.

[0017] As an improvement of the above method, the encoder includes a downsampling module, a Swin Transformer module and a patch merging module.

[0018] The encoder divides the input historical sound velocity field data into non-overlapping 3D image blocks, inputs the 3D image blocks into one downsampling module for downsampling, then inputs the 3D image blocks into one Swin Transformer module to extract the feature vectors of the image blocks, and then sequentially passes through multiple rounds of combinations of the downsampling module, the patch merging module and the Swin Transformer module for feature extraction and information fusion, so as to finally obtain the multi-scale feature information of the historical sound velocity field data.

[0019] As an improvement of the above method, the Swin Transformer module includes an embedding layer and a sliding window Transformer block.

[0020] The embedding layer comprises a patch segmentation unit and a linear mapping unit connected in sequence; wherein

[0021] The patch segmentation unit cuts the feature map generated by the encoder into a plurality of patch blocks;

[0022] The linear mapping unit linearly converts the segmented patch blocks into vector features;

[0023] The sliding window Transformer block comprises a window-based multi-head self-attention unit and a displacement window-based multi-head self-attention unit connected in sequence; wherein

[0024] The window-based multi-head self-attention unit is used to represent window multi-head self-attention, and the input feature map is segmented into a plurality of non-overlapping sub-windows, and self-attention calculation is performed on each sub-window;

[0025] The displacement window-based multi-head self-attention unit is used to represent moving window multi-head self-attention, and connections are established between different windows to realize information transmission between different windows

[0026] As an improvement of the above method, the decoder comprises a plurality of up-sampling layers and convolutional layers;

[0027] Each up-sampling layer in the decoder is connected to the corresponding SwinTransformer module in the encoder through a skip connection;

[0028] The decoder uses the up-sampling layer to enlarge the spatial size of the feature map to one time of the original, and then uses the skip connection to combine the output feature of the up-sampling layer with the encoder output feature of the same spatial size, to learn the multi-scale features in the sound velocity field sequence data, and then uses the convolutional layer to perform two layers of depth separable convolution, to reduce the channel number of the feature map to one half of the original size;

[0029] After continuous multiple up-sampling and convolution, one more up-sampling is performed, and finally a 1×1 standard convolution is used to generate the final prediction result.

[0030] As an improvement of the above method, the multi-scale convolutional module uses convolution kernels of different sizes for parallel calculation to generate feature maps of different scales, and then combines these feature maps, so that the features of each layer contain information from multiple scales.

[0031] As an improvement of the above method, the loss function of the four-dimensional sound velocity field prediction model is represented as:

[0032] ;

[0033] ;

[0034] ;

[0035] wherein, denotes a similarity measure; denotes a historical sound speed field; denotes a predicted target sound speed field; denotes a mapping function of sound field propagation; denotes a regularization term; denotes a weight balancing prediction accuracy and physical constraints; Ω denotes a spatial domain; p denotes position information, denotes a prediction result; denotes a mapping transformation.

[0036] The application also provides a four-dimensional sound speed field prediction system based on deep learning, which is realized based on the above method, and the system comprises:

[0037] a preprocessing module, configured to preprocess historical sound speed field data;

[0038] a prediction module, configured to input the preprocessed historical sound speed field data into a trained four-dimensional sound speed field prediction model to predict a sound speed field;

[0039] a four-dimensional sound speed field prediction model, configured to predict a sound speed field.

[0040] Compared with the prior art, the application has the following advantages:

[0041] The method of the application has the abilities of local feature mining, global data association and high-dimensional data analysis, and on the basis of maintaining the powerful spatial modeling capability of the Transformer, realizes unified modeling of the spatiotemporal evolution process of the sound speed field through time dimension coding, simultaneously extracts large-scale structural features of ocean circulation and local eddy fine-scale changes by using a hierarchical self-attention mechanism, effectively learns multi-scale dependent features in high-dimensional sound speed field spatiotemporal data, and more comprehensively understands and predicts the four-dimensional sound speed field. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 Fig. 1 shows a flowchart of a four-dimensional sound speed field prediction method based on deep learning;

[0043] Figure 2Fig. 1 shows the overall structure of the four-dimensional sound speed field prediction model; wherein, Downsampling: downsampling; PatchMerging: patch merging; Swin Transformer Block: Swin Transformer module; Multi-scaleconv: multi-scale convolution; Upsampling: upsampling; Skip connection: skip connection;

[0044] Fig. 3 (a) shows the prediction result map;

[0045] Fig. 3 (b) shows the real sound speed field map;

[0046] Figure 4 Fig. 4 shows the cross-section real value and prediction result in different latitude directions;

[0047] Figure 5 Fig. 5 shows the cross-section real value and prediction result in different longitude directions, and the difference between the prediction result and the real value;

[0048] Figure 6 Fig. 6 shows the cross-section real value and prediction result in different depth directions, and the difference between the prediction result and the real value. DETAILED DESCRIPTION

[0049] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings.

[0050] The purpose of the present application is to effectively solve the limitations of the existing sound speed field prediction method in long-distance space correlation and multi-scale feature extraction, and to propose a four-dimensional sound speed field prediction method and system based on deep learning by combining the U-Net model based on convolutional neural network and the window Transformer model based on multi-head self-attention mechanism.

[0051] Embodiment 1

[0052] As shown in Fig. 1, the four-dimensional sound speed field prediction method based on deep learning proposed by the present application comprises: Figure 1

[0053] Step 1: Obtain the sound speed field data of three-dimensional space and perform preprocessing to construct a normalized historical sound speed field dataset.

[0054] The preprocessing is aimed at the seasonal variation characteristics of the ocean sound speed field, and a monthly grouping normalization strategy is adopted. Let the original sound speed data be where t is the time index, and (x, y, z) is the spatial coordinate. The normalization process is defined as:

[0055] ;

[0056] ​where m = month(t) denotes the month corresponding to time t, and are the maximum and minimum values of all samples in the m-th month, respectively:

[0057] ;

[0058] ;

[0059] where, denotes the set of all time indices in the m-th month.

[0060] The preprocessed data is normalized by maximum and minimum values and divided into a training set and a test set, the preprocessed data is four-dimensional data, and the four-dimensional data is historical data of a sound velocity field, which includes longitude point number, latitude point number, depth point number and time sequence.

[0061] In specific embodiments, the sound velocity field data covering a region of 334km*334km*180m in a certain sea area is used, the latitude resolution is 4.40km, the longitude resolution is 8.79km, and the depth resolution is 20m. The data set contains 8760 three-dimensional data samples covering the whole year of 2022, and the dimension of each sample is 32*32*10. 80% of the data is used as the training set, and the remaining 20% of the data is used as the test set. For the missing value problem, the data set adopts the strategy of replacing the NaN value with the global mean value to avoid introducing false sound velocity gradient by interpolation:

[0062] ;

[0063] where, is the global mean value of all valid data points.

[0064] The training sample adopts a sliding window technique, and sets 7 days as the input sequence length, corresponding to multiple complete day cycles, and the multi-time scale change mode of the ocean sound velocity, including the stability and variability of the daily change, the comprehensive influence of the quasi-periodic tide phenomenon on the sound velocity profile. In addition, in order to improve the quality and availability of the data, the present application further preprocesses the original sound velocity data in combination with the characteristics of the ocean sound velocity, and sets the original sound velocity data , where t is the time index, and (x, y, z) is the spatial coordinate. The normalization process is defined as:

[0065] ;

[0066] where m = month(t) denotes the month corresponding to time t, and are the maximum and minimum values of all samples in the m-th month, respectively:

[0067] ;

[0068] ;

[0069] wherein, denotes the set of all time indices of the m-th month.

[0070] In the implementation, the 8760 hours of data of the whole year are marked by month, and by using the maximum and minimum values corresponding to each month, the real physical characteristics of the sound speed field are better recovered, and the data distortion caused by seasonal changes is avoided.

[0071] Step 2: Build a deep learning model based on Swin Transformer and U-Net architecture (four-dimensional sound speed field prediction model), and construct a weighted loss function.

[0072] The four-dimensional sound speed field prediction model takes four-dimensional data (T, x, y, z) composed of three-dimensional spatial scales (x, y, z) and one-dimensional time scale T as input, uses Swin Transformer to capture long-term dependencies, uses convolutional neural network to process local features, and at the same time through multi-scale convolution and skip connection, so that the model can utilize both the local detail changes of sound speed and the time series information in the entire historical time range, effectively learn the multi-scale dependency features in high-dimensional sound speed field spatio-temporal data, and more comprehensively understand and predict three-dimensional sound speed field.

[0073] The four-dimensional sound speed field prediction model is composed of two important parts, encoder and decoder, and the model structure diagram is as shown in Figure 2 The model takes historical three-dimensional spatial sound speed field image as input data, where B represents batch size, T represents time series length, X, Y and Z represent spatial size of sound speed field image. The encoder is responsible for learning the spatio-temporal features in the sound speed field data, and gradually reducing the spatial size of the feature image, and the decoder is responsible for reconstructing the feature image output by the encoder in reverse, and gradually enlarging the spatial size of the feature image to obtain the final extrapolation result.

[0074] The encoder part includes four down-sampling modules, four Swin Transformer modules and three patch merging modules. The encoder first divides the input historical sound speed field data into non-overlapping 3D image blocks, where the size of each dimension of the image block is P, and the i-th patch is denoted as where i ∈ {1,..., N}, N = X / P × Y / P × T / P is the total number of patches. Each patch is flattened and treated as a "token", and then a linear projection layer is used to project each token to a Z-dimensional feature representation:

[0075] ;

[0076] wherein, represents a linear projection matrix, and output .

[0077] After the linear projection layer, a few stages of patch merging are applied on the token z0, constantly learning the short-term dependencies and spatio-temporal features in the acoustic velocity field sequence data. Then, the Swin Transformer module is used to capture the long-term dependencies in the acoustic velocity field sequence data, and the two modules complement each other.

[0078] Specifically, the encoder divides the input historical acoustic velocity field data into non-overlapping 3D image blocks, inputs them into a down-sampling module for down-sampling, and then inputs them into a Swin Transformer module to extract the feature vectors of the pictures. Then, the combination of three rounds of down-sampling modules, patch merging modules and Swin Transformer modules is sequentially passed through to extract features and fuse information, and finally the multi-scale feature information of the historical acoustic velocity field data is obtained. In other embodiments, the number of combinations of down-sampling modules, patch merging modules and Swin Transformer modules can be adjusted according to actual needs.

[0079] The Swin Transformer module includes two key structures, an embedding layer and a sliding window Transformer block. First, the architecture uses a patch division unit in the embedding layer to divide the feature map generated by the down-sampling module into multiple patch blocks, and then uses a linear mapping unit to linearly convert the divided patch blocks into vector features. Finally, a sliding window Transformer block based on the self-attention mechanism is used to capture the long-term dependency features in the sequence, and the output feature map is input into the decoder for reverse reconstruction. The core of the sliding window Transformer block is the window-based multi-head self-attention W-MSA unit and the displacement window-based multi-head self-attention SW-MSA unit. Among them, W-MSA represents window multi-head self-attention, which is used to divide the input feature map into multiple non-overlapping sub-windows and perform self-attention calculation on each sub-window. SW-MSA represents a moving window multi-head self-attention, which is used to establish a connection between different windows, thereby realizing the information transmission between different windows. This operation reduces the computational complexity of the model from the quadratic relationship of the spatial size of the input acoustic velocity field history data to a linear relationship, significantly improving the training efficiency of the model. The calculation process of the Swin Transformer module is shown in the following formula:

[0080] ; ; ; ;

[0081] wherein, Represents the feature values ​​input from the previous layer. This represents the feature values ​​after passing through a window-based multi-head self-attention module. This represents the feature values ​​after passing through a multilayer perceptron. This represents the feature value after passing through the multi-head attention mechanism of the position window. The values ​​represent the feature values ​​after passing through the multilayer perceptron. W-MSA represents a window-based multi-head self-attention module, MLP represents a multilayer perceptron, SW-MSA represents a shift-window-based multi-head self-attention module, and LN represents layer normalization operation.

[0082] A multi-scale convolutional module is used between the encoder and decoder, employing convolutional kernels of different sizes for parallel computation to generate feature maps at different scales. These feature maps are then merged, ensuring that the features at each layer contain information from multiple scales.

[0083] The decoder consists of consecutive upsampling layers and convolutional layers. In the decoding stage, each upsampled feature map is connected to the corresponding feature map in the encoding path via skip connections, followed by two consecutive convolutional layers. The decoder first uses bilinear upsampling to double the spatial size of the feature map. Then, it uses skip connections to merge the upsampled output features with the encoder output features of the same spatial size to learn multi-scale features in the sound velocity field sequence data. Finally, it uses two depthwise separable convolutions to reduce the number of channels in the feature map to half its original size. After stacking four upsampling modules consecutively, a 1×1 standard convolution is used to generate the final extrapolated image sequence. In other embodiments, the number of upsampling and convolutional layers can be adjusted according to actual needs, but it must be the same as the number of downsampling, patch merging, and Swin Transformer modules in the encoder.

[0084] The loss function in the four-dimensional sound velocity field prediction model consists of two parts, one part Another part measures the similarity between the predicted sound speed field and the target sound speed field. It is a regularization term that ensures the smoothness of the sound velocity field variation:

[0085] ;

[0086] ;

[0087] in, Historical sound velocity field; Predict the target sound velocity field; : Mapping function for sound field propagation; : The weight of balancing prediction accuracy and physical constraints; Ω represents the spatial domain.

[0088] The similarity measure used in the loss function is the mean squared error, which is the mean of the squared differences between and :

[0089] ;

[0090] where p represents the position information, represents the prediction result, represents the mapping function from the initial state to the target state, and the mapping transformation.

[0091] Step 3: Based on the training set and the model, the normalized numerical values are input into the deep learning network for training to obtain a four-dimensional sound speed field prediction model; the test set is input into the four-dimensional sound speed field prediction model to adjust and optimize the prediction ability of the model, and the final four-dimensional sound speed field prediction model is obtained.

[0092] The present application uses Swin Transformer to capture long-term dependencies, uses convolutional neural networks to process local features, and simultaneously uses multi-scale convolution and skip connections to enable the model to utilize both local detail changes in sound speed and time series information over the entire historical time range.

[0093] Step 4: Use the four-dimensional sound speed field prediction model to predict the four-dimensional sound speed field.

[0094] Based on the input variables and the trained model, the four-dimensional sound speed field prediction result for the next 1h is obtained.

[0095] In this embodiment, the model is trained based on a certain sea area. The model uses the Adam optimizer, with an initial learning rate of 1e-3 and a batch size of 32. During training, the cosine annealing learning rate scheduling strategy is used, and the maximum number of training rounds is 200. To prevent overfitting, the early stopping strategy is used, and training is stopped when the validation loss does not improve for 10 consecutive rounds.

[0096] Figures 3(a) and 3(b) compare the model prediction results with the true sound speed field. It can be observed that the prediction results are highly consistent with the true values in terms of spatial structure, indicating that the model successfully captures the three-dimensional spatial features of the sound speed field. Figure 4 and Figure 5The cross-section comparison analysis in the horizontal dimension is shown, and each figure contains the true value, the predicted result and the prediction error. The predicted result well reproduces the numerical characteristics of the true sound speed field, including the high sound speed area in the surface layer, the sharp drop of sound speed in the thermocline, and the relatively stable area in the deep layer. The sound speed value ranges between 1505-1535 m / s, which is consistent with the actual observation characteristics of this sea area, indicating that the model has the ability to learn the basic physical laws of the ocean sound speed field. Due to the influence of ocean currents and water masses, some specific areas show relatively large prediction errors. These areas may be accompanied by complex ocean dynamics, making the spatiotemporal variation pattern of the sound speed field more complex and irregular. Figure 6 The cross-section comparison analysis in the depth direction is shown, and each figure contains the true value, the predicted result and the prediction error. In the vertical dimension, the sound speed distribution is mainly controlled by temperature stratification, salinity gradient and hydrostatic pressure. From the Figure 6 As shown in the depth direction difference figure, the largest prediction error is concentrated in the depth range of 60-100 meters, which corresponds to the typical depth of the ocean thermocline. The thermocline is the region with the largest temperature gradient in the ocean, and its sound speed changes very sharply, which can change by 10-20 m / s in a vertical distance of tens of meters. This strong nonlinear feature poses a great challenge to any prediction model.

[0097] Embodiment 2

[0098] The application also provides a four-dimensional sound speed field prediction system based on deep learning, which is realized based on the above method. The system comprises:

[0099] A preprocessing module for preprocessing historical sound speed field data;

[0100] A prediction module for inputting the preprocessed historical sound speed field data into the trained four-dimensional sound speed field prediction model to predict the sound speed field;

[0101] A four-dimensional sound speed field prediction model for predicting the sound speed field.

[0102] The application can also provide a computer device comprising at least one processor, memory, at least one network interface and user interface. The various components in the device are coupled together through a bus system. It can be understood that the bus system is used to realize the connection and communication between the components. In addition to the data bus, the bus system also includes power bus, control bus and state signal bus.

[0103] The user interface can include a display, a keyboard or a clicking device. For example, a mouse, a trackball, a touchpad or a touch screen, etc.

[0104] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (Read-Only Memory, ROM), a programmable read-only memory (Programmable ROM, PROM), an erasable programmable read-only memory (Erasable PROM, EPROM), an electrically erasable programmable read-only memory (Electrically EPROM, EEPROM) or a flash memory. The volatile memory can be a random access memory (Random Access Memory, RAM) used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (Static RAM, SRAM), dynamic random access memory (Dynamic RAM, DRAM), synchronous dynamic random access memory (Synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data Rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (Synchlink DRAM, SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DRRAM). The memory described herein is intended to include, but not limited to, these and any other suitable types of memory.

[0105] In some embodiments, the memory stores elements, executable modules or data structures, or a subset thereof, or an extended set thereof: an operating system and an application program.

[0106] Among them, the operating system includes various system programs, such as framework layer, core library layer, driver layer, etc., for implementing various basic services and processing hardware-based tasks. The application program includes various application programs, such as media player (Media Player), browser (Browser), etc., for implementing various application services. The program for implementing the method of the embodiments of the present disclosure can be included in the application program.

[0107] In the above-mentioned embodiments, the processor can also be used to execute the steps of the above-mentioned method by invoking the programs or instructions stored in the memory, in particular, the programs or instructions stored in the application program.

[0108] execute the steps of the above-mentioned method.

[0109] The method can be applied to a processor or implemented by the processor. The processor can be an integrated circuit chip having a signal processing capability. In implementation, the steps of the method can be completed by an integrated logic circuit of hardware in the processor or by an instruction in the form of software. The processor can be a general-purpose processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The methods disclosed above can be implemented or executed by the processor. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed above can be directly embodied as a hardware code executed by the processor or a combination of hardware and software modules in the processor. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, or other mature storage mediums in the art. The storage medium is located in the storage memory, and the processor reads information in the storage memory and combines the hardware to complete the steps of the method.

[0110] It can be understood that the embodiments described in the present application can be realized by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more Application Specific Integrated Circuits (ASICs), Digital Signal Processors (DSPs), Digital Signal Processing Devices (DSPDs), Programmable Logic Devices (PLDs), Field-Programmable Gate Arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for executing the functions described in the present application or a combination thereof.

[0111] For software implementation, the functions of the present application can be implemented by executing the functional modules (such as processes, functions, etc.) of the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0112] The application further provides a nonvolatile storage medium for storing the computer program. When the computer program is executed by a processor, each step in the above method embodiment can be implemented.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A deep learning-based four-dimensional sound speed field prediction method, the method comprising: inputting preprocessed historical sound speed field data into a trained four-dimensional sound speed field prediction model to predict the sound speed field; the four-dimensional sound speed field prediction model is built based on Swin Transformer and U-Net architecture, comprising an encoder, a multi-scale convolution module and a decoder; wherein, the encoder is used to learn the spatio-temporal features in the sound speed field data and gradually reduce the spatial size of the feature image; the multi-scale convolution module is used to extract and fuse features at different spatial scales; the decoder is used to reversely reconstruct the feature map output by the multi-scale convolution module, gradually enlarge the spatial size of the feature map, and obtain the final prediction result; the encoder comprises a downsampling module, a Swin Transformer module and a patch merging module; the encoder divides the input historical sound speed field data into non-overlapping 3D image blocks, inputs them into one downsampling module for downsampling, then inputs them into one Swin Transformer module to extract the feature vector of the image, and finally performs feature extraction and information fusion through multiple rounds of combination of the downsampling module, the patch merging module and the Swin Transformer module in sequence, to obtain the multi-scale feature information of the historical sound speed field data; the Swin Transformer module comprises an embedding layer and a sliding window Transformer block; the embedding layer comprises a patch segmentation unit and a linear mapping unit connected in sequence; wherein, the patch segmentation unit cuts the feature map generated by the encoder into multiple patch blocks; the linear mapping unit linearly converts the cut patch blocks into vector features; the sliding window Transformer block comprises a window-based multi-head self-attention unit and a displacement window-based multi-head self-attention unit connected in sequence; wherein, the window-based multi-head self-attention unit is used to represent window multi-head self-attention, divides the input feature map into multiple non-overlapping sub-windows, and performs self-attention calculation on each sub-window separately; the displacement window-based multi-head self-attention unit is used to represent moving window multi-head self-attention, establishes a connection between different windows to realize information transmission between different windows.

2. The deep learning-based four-dimensional acoustic velocity field prediction method of claim 1, wherein, the historical sound speed field data comprises longitude point number, latitude point number, depth point number and time sequence.

3. The deep learning-based four-dimensional acoustic velocity field prediction method of claim 1, wherein, the preprocessing comprises: for the seasonal variation characteristics of the ocean sound speed field, a monthly grouping normalization strategy is adopted, and the normalization process is represented as: ； wherein, denotes normalized sound speed data; denotes raw sound speed data; (x, y, z) are spatial coordinates; t is a time index; m denotes the month corresponding to time t; and are the maximum and minimum values of all samples in the mth month, respectively.

4. The deep learning-based four-dimensional acoustic velocity field prediction method of claim 1, wherein, the decoder comprises a plurality of upsampling layers and convolution layers; each upsampling layer in the decoder is connected to the corresponding Swin Transformer module in the encoder through a skip connection; The decoder uses the up-sampling layer to enlarge the spatial size of the feature map to twice the original size, then uses the jump connection to combine the output features of the up-sampling layer with the encoder output features of the same spatial size to learn the multi-scale features in the sound speed field sequence data, and then uses the convolution layer to perform two layers of depth separable convolution to reduce the channel number of the feature map to half the original size; After continuous multiple up-sampling and convolution, one more up-sampling is performed, and finally, a 1*1 standard convolution is used to generate the final prediction result.

5. The deep learning-based four-dimensional acoustic velocity field prediction method of claim 1, wherein, The multi-scale convolution module uses different size convolution kernels for parallel calculation to generate feature maps of different scales, and then combines these feature maps so that the features of each layer contain information from multiple scales.

6. The deep learning-based four-dimensional acoustic velocity field prediction method of claim 1, wherein, The loss function of the four-dimensional acoustic velocity field prediction model is represented as: ； ； ； wherein, represents a similarity measure; represents a historical sound speed field; represents a predicted target sound speed field; represents a mapping function for sound field propagation; represents a regularization term; represents a weight balancing prediction accuracy and physical constraints; Ω represents a spatial domain; p represents position information, represents a prediction result; represents a mapping transformation.

7. A deep learning based four-dimensional acoustic velocity field prediction system, implemented based on the method of any one of claims 1-6, characterized in that, The system comprises: a preprocessing module configured to preprocess historical sound speed field data; a prediction module configured to input the preprocessed historical sound speed field data into a trained four-dimensional sound speed field prediction model to predict the sound speed field; and a four-dimensional sound speed field prediction model configured to predict the sound speed field.

Images