Multi-spatial-scale seawater sound velocity prediction method and system

Abstract

The invention discloses a multi-spatial-scale seawater sound velocity prediction method and system, and relates to the technical field of ocean observation, and the method comprises the steps: obtaining current target prediction sea area data; performing format conversion and normalization processing on the current target prediction sea area data to obtain input data; constructing an STA-ConvLSTM sound velocity prediction model according to the space-time attention mechanism, the convolutional neural network and the long short-term memory network; and inputting the input data into the STA-ConvLSTM sound velocity prediction model to obtain the seawater sound velocity data of the target predicted sea area. By constructing the STA-ConvLSTM sound velocity prediction model, the problem of multi-spatial scale prediction of the seawater sound velocity is solved, the prediction of the sound velocity profile, the sound velocity section and the sound velocity structural body is realized, and the precision of seawater sound velocity prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of ocean observation technology, and in particular to a multi-spatial-scale seawater sound velocity prediction method and system. Background Art

[0002] As a core parameter reflecting ocean acoustic characteristics, the speed of sound in the ocean determines the characteristics of sound propagation in the ocean and is a prerequisite for the research and application of underwater acoustic technology. Obtaining accurate information on the speed of sound in the ocean is of great significance for improving the accuracy and efficiency of underwater detection and communication.

[0003] Currently, seawater sound velocity is primarily acquired through observation and prediction. Limited by the complex ocean environment and high observation costs, the density of seawater sound velocity observation sites remains insufficient. Sound velocity prediction methods, as an effective supplement and extension, have gradually become a research hotspot. Ocean numerical model forecasts are based on ocean dynamics theory and physical and chemical processes, and their accuracy is significantly affected by the amount of computation required. Statistical prediction methods, which analyze historical data to determine the distribution characteristics of seawater sound velocity, require a large amount of historical data and offer poor prediction results under complex conditions.

[0004] In recent years, advances in computer technology and deep learning theory have provided new insights into seawater sound velocity prediction. This has led to a deeper understanding of the feature extraction of sound velocity data, shifting prediction requirements from a single sound velocity profile to a multi-dimensional sound velocity field. Therefore, it is necessary to design a multi-scale sound velocity prediction method that automatically assigns spatiotemporal weights based on the spatiotemporal characteristics of seawater sound velocity. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-spatial-scale seawater sound velocity prediction method and system to solve the multi-spatial-scale prediction problem of seawater sound velocity and realize the prediction of sound velocity profile, sound velocity section and sound velocity structure.

[0006] To achieve the above object, the present invention provides a multi-spatial-scale seawater sound velocity prediction method, comprising the following steps: S1. Obtaining the current target predicted sea area data; the current target predicted sea area data includes: time, spatial location, seawater temperature, seawater salinity and seawater depth; S2. Perform format conversion and normalization processing on the current target predicted sea area data to obtain input data; S3. Based on the spatiotemporal attention mechanism, convolutional neural network and long short-term memory network, a STA-ConvLSTM sound speed prediction model is constructed; S4. Input the input data into the STA-ConvLSTM sound speed prediction model to obtain the seawater sound speed data of the target predicted sea area; the seawater sound speed data of the target predicted sea area includes: a sound speed profile, a sound speed section and a sound speed structure.

[0007] Preferably, S3 constructs a STA-ConvLSTM sound speed prediction model based on the spatiotemporal attention mechanism, convolutional neural network, and long short-term memory network, specifically including: S301, obtaining historical target prediction sea area data; S302, calculating the seawater sound velocity data of the historical target predicted sea area based on the empirical formula of sound velocity according to the historical target predicted sea area data; S303, performing format conversion and normalization processing on the historical target prediction sea area data and the historical target prediction sea area seawater sound speed data to obtain the historical target prediction sea area input data and the historical target prediction sea area seawater sound speed input data; S304. Construct a training set, a validation set, and a test set based on the historical target prediction sea area input data and the historical target prediction sea area seawater sound speed input data; S305. Construct a convolutional layer according to a convolutional neural network and a long short-term memory network; the convolutional layer includes: a first ConvLSTM layer and a second ConvLSTM layer; S306. Construct a temporal attention layer and a spatial attention layer according to the spatiotemporal attention mechanism; S307, connecting the convolution layer, the temporal attention layer, and the spatial attention layer to obtain an initial STA-ConvLSTM sound speed prediction model; S308, inputting the training set and the validation set into the initial STA-ConvLSTM sound speed prediction model for training to obtain a trained initial STA-ConvLSTM sound speed prediction model; S309, inputting the test set into the trained initial STA-ConvLSTM sound speed prediction model, testing and evaluating the trained initial STA-ConvLSTM sound speed prediction model, and obtaining a trained initial STA-ConvLSTM sound speed prediction model; S310: Determine the trained initial STA-ConvLSTM sound speed prediction model as the STA-ConvLSTM sound speed prediction model.

[0008] Preferably, in S305, constructing a convolutional layer according to a convolutional neural network and a long short-term memory network specifically includes: By using filtering and convolution operations, the input data of the historical target prediction sea area is convolved in time and space to obtain the local characteristics and time characteristics of the input data of the historical target prediction sea area; A nonlinear activation function is used to activate local features and temporal features respectively to obtain local activation features and temporal activation features; Perform element-wise multiplication of the local activation feature and the temporal activation feature to obtain the first output feature data and the first ConvLSTM layer; Performing feature extraction on the first output feature data by using filtering and convolution operations to obtain second output feature data and a second ConvLSTM layer; Connect the first ConvLSTM layer and the second ConvLSTM layer to obtain a convolutional layer.

[0009] Preferably, constructing the temporal attention layer and the spatial attention layer according to the spatiotemporal attention mechanism in S306 specifically includes: extracting features of different time steps through a first convolution operation to obtain time step feature data; the first convolution operation is to use a 1×1 convolution layer; Based on the fully connected layer, the first gating factor is calculated to determine the importance of the temporal attention weight in the time step feature data; According to the importance of the temporal attention weight, the temporal attention weight is calculated using the first convolution operation and the nonlinear activation function; Normalize the temporal attention weight to obtain the normalized temporal attention weight; Multiply the normalized temporal attention weight by the second output feature data and the first gating factor element by element, construct a temporal attention layer, and obtain the third output feature data; Through the second convolution operation, features at different spatial positions are extracted to obtain spatial position feature data; the second convolution operation uses a 7×7 convolution layer; Based on the fully connected layer, the second gating factor is calculated to determine the importance of the spatial attention weight in the spatial position feature data; According to the importance of the spatial attention weight, the spatial attention weight is calculated using the second convolution operation and the nonlinear activation function; Normalize the spatial attention weight to obtain the normalized spatial attention weight; The normalized spatial attention weight is sequentially multiplied element-wise with the third output feature data and the second gating factor to construct a spatial attention layer to obtain the fourth output feature data.

[0010] Preferably, in S308, the training set and the validation set are input into the initial STA-ConvLSTM sound speed prediction model for training to obtain the trained initial STA-ConvLSTM sound speed prediction model, which specifically includes: When the initial STA-ConvLSTM sound speed prediction model is trained using the training set and the validation set, the mean square error is used as the loss value, and the early stopping method monitoring model is used to detect the loss value. The training is stopped until the decrease in the loss value meets the set requirements, and the trained initial STA-ConvLSTM sound speed prediction model is obtained; the set requirement is that the loss value does not decrease in 20 rounds of calculation.

[0011] Preferably, in S309 , the trained initial STA-ConvLSTM sound speed prediction model is tested and evaluated, specifically by using root mean square error and mean absolute error for evaluation.

[0012] Preferably, the calculation formula of the root mean square error is: ; Among them, RMSE is the root mean square error, is the number of training sets, For the Input data of seawater sound speed in historical target prediction sea area, For the The seawater sound speed prediction data of historical target prediction sea areas; The calculation formula of the mean absolute error is: ; Among them, MAE is the mean absolute error.

[0013] Preferably, in S2, the current target predicted sea area data is format converted and normalized to obtain input data, which specifically includes: Convert the current target prediction sea area data into the input data format of the STA-ConvLSTM model to obtain 6D tensor data; the 6D tensor data includes: number of samples, time steps, depth, number of rows, number of columns and number of features; The maximum and minimum normalization method is used to normalize the 6D tensor data to obtain the input data.

[0014] Preferably, the calculation formula of the input data is: ; in, is the input data, is 6D tensor data, is the maximum value in the 6D tensor data, is the minimum value in the 6D tensor data.

[0015] The present invention also provides a multi-spatial-scale seawater sound velocity prediction system, comprising: A data acquisition module is used to acquire the current target predicted sea area data; the current target predicted sea area data includes: time, spatial position, sea water temperature, sea water salinity and sea water depth; The data processing module is used to perform format conversion and normalization on the current target prediction sea area data to obtain input data; A model building module, which is used to build a STA-ConvLSTM sound speed prediction model based on the spatiotemporal attention mechanism, convolutional neural network, and long short-term memory network; The prediction module is used to input the input data into the STA-ConvLSTM sound speed prediction model to obtain the seawater sound speed data of the target predicted sea area; the seawater sound speed data of the target predicted sea area includes: sound speed profile, sound speed section and sound speed structure.

[0016] In summary, the multi-spatial-scale seawater sound speed prediction method and system of the present invention have the following beneficial effects compared with traditional technologies: the present invention constructs a STA-ConvLSTM sound speed prediction model, which integrates the attention mechanism, convolutional neural network and long short-term memory network, can effectively extract the spatiotemporal characteristics of seawater sound speed data, and realize the automatic allocation of calculation weights, effectively improving the accuracy of seawater sound speed prediction; at the same time, it can also perform multi-spatial-scale seawater sound speed prediction, including sound speed profiles, sound speed sections and sound speed structures, to meet the application needs of different scenarios.

[0017] The technical method of the present invention is further described in detail below through the accompanying drawings and examples. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of a flow chart of a multi-spatial-scale seawater sound velocity prediction method of the present invention; Figure 2 A structural diagram of constructing a STA-ConvLSTM sound speed prediction model for the present invention and using the STA-ConvLSTM sound speed prediction model to predict the sound speed of seawater; Figure 3 Schematic diagram of the STA-ConvLSTM sound speed prediction model of the present invention; Figure 4 A flow chart for format conversion and normalization processing of target prediction sea area data, historical target prediction sea area data, and seawater sound velocity data of historical target prediction sea areas of the present invention; Figure 5 This is the sound velocity structure prediction error curve obtained based on the STA-ConvLSTM seawater sound velocity prediction method; Figure 6 This is a module diagram of a multi-spatial-scale seawater sound velocity prediction system. DETAILED DESCRIPTION

[0019] The technical method of the present invention is further described below through the accompanying drawings and embodiments. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and values ​​described in these embodiments do not limit the scope of this application.

[0020] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0021] Technologies, systems, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0022] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0023] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0024] like Figure 1 As shown, the present invention provides a multi-spatial-scale seawater sound velocity prediction method, comprising: Step S1: Acquire the current target predicted sea area data, wherein the current target predicted sea area data includes: time, spatial location, seawater temperature, seawater salinity and seawater depth.

[0025] Step S2: Perform format conversion and normalization processing on the current target predicted sea area data to obtain input data.

[0026] Step S2, such as Figure 4 As shown, specifically including: The current target prediction sea area data is converted into the input data format of the STA-ConvLSTM model to obtain 6D tensor data. The 6D tensor data includes: number of samples, time steps, depth, number of rows, number of columns, and number of features.

[0027] Among them, the number of rows and columns is used to represent the spatial dimensional characteristics of the sound speed prediction. When performing sound speed profile prediction, the row (rows) is 1 and the column (cols) is 1, that is, rows=1, cols=1; when performing sound speed cross-section prediction, rows=1, cols≥2; when performing sound speed structure prediction, rows≥2, cols≥2.

[0028] The 6D tensor data is normalized using the maximum and minimum normalization method to obtain the input data. The calculation formula of the input data is: ; in, is the input data, is 6D tensor data, is the maximum value in the 6D tensor data, is the minimum value in the 6D tensor data.

[0029] Step S3: Based on the spatiotemporal attention mechanism, convolutional neural network and long short-term memory network, a STA-ConvLSTM sound speed prediction model is constructed.

[0030] Step S3, as Figure 2 As shown, specifically including: Step S301: Acquire historical target prediction sea area data.

[0031] Furthermore, data from a portion of a country's oceans, collected from the global ocean Argo gridded dataset, was selected as the historical target ocean data for prediction. This data includes the spatial location (latitude and longitude), seawater temperature, salinity, and depth of the country's oceans. The horizontal resolution is 1°×1°, the depth range is 0–1500 m, and the data is divided into 58 vertical layers of varying spacing. The vertical resolution ranges from 5–100 m from the sea surface to the seafloor, and the monthly average data spans 216 months.

[0032] Step S302: Based on the historical target prediction sea area data, the seawater sound speed data of the historical target prediction sea area is calculated based on the sound speed empirical formula.

[0033] The calculation formula for the historical data of seawater sound speed in the historical target sea area is as follows: ; in, C Represents the historical target predicted sea area seawater sound speed data, in m / s, T Indicates the seawater temperature of the historical target forecast sea area, in ℃. S Indicates the seawater salinity of the historical target forecast sea area, in ‰. Z Indicates the depth of the historical target predicted sea area, in meters.

[0034] Step S303 : performing format conversion and normalization processing on the historical target prediction sea area data and the historical target prediction sea area seawater sound speed data to obtain the historical target prediction sea area input data and the historical target prediction sea area seawater sound speed input data.

[0035] Step S304: construct a training set, a validation set, and a test set based on the historical target prediction sea area input data and the historical target prediction sea area seawater sound speed input data.

[0036] The process of format conversion and normalization of the historical target prediction sea area data and the historical target prediction sea area seawater sound speed data in step S303 is similar to the process of format conversion and normalization of the current target prediction sea area data to obtain input data in step S2.

[0037] Furthermore, the present invention can select corresponding spatial scale features according to the sound speed information requirements to be predicted, including sound speed profiles, sound speed cross sections and sound speed structures.

[0038] Taking the sound velocity structure prediction as an example, the sound velocity structure composed of the four points A (114.0°E, 10.5°N), B (114.0°E, 11.0°N), C (114.5°E, 11.0°N) and D (114.5°E, 10.5°) is selected as the prediction object, and the corresponding historical target prediction sea area data and historical target prediction sea area seawater sound velocity data are extracted for format conversion and normalization. The specific process is as follows Figure 4 As shown in FIG, the multi-scale spatial sound velocity data is gradually increased in dimension to form N sequences of six-dimensional spatiotemporal data, and then the data set is divided and normalized.

[0039] The sound speed spatiotemporal data are spliced, and a sliding window is used to traverse the dataset with a window width set to 12 months. The data format is converted into a 6D tensor with dimensions of (193, 12, 58, 2, 2, 1), which respectively represent 193 data samples, 12-month time steps, 58 vertical layers, 2 rows of sound speed structures, 2 columns of sound speed structures, and only one feature of sound speed in the feature dimension.

[0040] The maximum and minimum normalization method is used to reduce the processed data, that is, the input data, to the range [0,1].

[0041] Furthermore, the present invention divides the normalized data into a training set (155, 12, 58, 2, 2, 1), a validation set (19, 12, 58, 2, 2, 1) and a test set (19, 12, 58, 2, 2, 1) in a ratio of 8:1:1.

[0042] Step S305: Construct a convolutional layer based on the convolutional neural network and the long short-term memory network, wherein the convolutional layer includes: a first ConvLSTM layer and a second ConvLSTM layer.

[0043] The specific contents of step S305 include: By using filtering and convolution operations, the input data of the historical target prediction sea area are convolved in time and space to obtain the local characteristics and time characteristics of the input data of the historical target prediction sea area.

[0044] The nonlinear activation function is used to activate the local features and temporal features respectively to obtain local activation features and temporal activation features.

[0045] The local activation feature and the temporal activation feature are element-wise multiplied to obtain the first output feature data and the first ConvLSTM layer.

[0046] The first output feature data is subjected to feature extraction by adopting filtering processing and convolution operation to obtain second output feature data and obtain a second ConvLSTM layer.

[0047] Connect the first ConvLSTM layer and the second ConvLSTM layer to obtain a convolutional layer.

[0048] Step S306: Construct the temporal attention layer and the spatial attention layer according to the spatiotemporal attention mechanism.

[0049] The specific content of step S306 is as follows: Figure 3 As shown, including: Through the first convolution operation, features of different time steps are extracted to obtain time step feature data; the first convolution operation adopts a 1×1 convolution layer.

[0050] Based on the fully connected layer, the first gating factor is calculated to determine the importance of the temporal attention weight in the time step feature data.

[0051] According to the importance of the temporal attention weight, the first convolution operation and the nonlinear activation function are used to calculate the temporal attention weight to obtain the importance of each time step.

[0052] The temporal attention weights are normalized so that their sum is 1 to obtain the normalized temporal attention weights.

[0053] The normalized temporal attention weight is sequentially multiplied element-wise with the second output feature data and the first gating factor to construct a temporal attention layer and obtain the third output feature data. This strengthens the influence of key time steps and controls the degree of attention information transmission.

[0054] Through the second convolution operation, features at different spatial locations are extracted to obtain spatial location feature data. The second convolution operation uses a 7×7 convolution layer.

[0055] Based on the fully connected layer, the second gating factor is calculated to determine the importance of the spatial attention weight in the spatial position feature data.

[0056] According to the importance of the spatial attention weight, the second convolution operation and the nonlinear activation function are used to calculate the spatial attention weight to obtain the importance of each spatial position.

[0057] The spatial attention weights are normalized so that their sum is 1 to obtain the normalized spatial attention weights.

[0058] The normalized spatial attention weights are sequentially multiplied element-wise with the third output feature data and the second gating factor to construct a spatial attention layer, and the fourth output feature data is obtained. This strengthens the influence of key spatial positions and controls the degree of attention information transmission.

[0059] Step S307: Connect the convolution layer, the temporal attention layer, and the spatial attention layer to obtain an initial STA-ConvLSTM sound speed prediction model.

[0060] Step S308: Input the training set and the validation set into the initial STA-ConvLSTM sound speed prediction model for training to obtain a trained initial STA-ConvLSTM sound speed prediction model.

[0061] The specific content of step S308 is: when the initial STA-ConvLSTM sound speed prediction model is inputted for training using the training set and the validation set, the mean square error is used as the loss value, and the early stopping method monitoring model is used to detect the loss value until the decrease in the loss value meets the set requirements, then the training is stopped to obtain the trained initial STA-ConvLSTM sound speed prediction model. The set requirement is that the loss value does not decrease in 20 rounds of calculation, that is, setting "patience" = 20 can effectively prevent the trained STA-ConvLSTM sound speed prediction model from overfitting. The present invention is trained by the back propagation algorithm, and the weight of the model is updated according to the error of the training data. Multiple cycles are passed during the training process to ensure that the model converges to the optimal performance.

[0062] The calculation formula of the loss value is: ; Among them, MSE is the mean square error, is the number of training sets, For the Input data of seawater sound speed in historical target prediction sea area, For the The seawater sound speed prediction data of the historical target prediction sea area.

[0063] Step S309: Input the test set into the trained initial STA-ConvLSTM sound speed prediction model, test and evaluate the trained initial STA-ConvLSTM sound speed prediction model, and obtain a trained initial STA-ConvLSTM sound speed prediction model.

[0064] In step S309, the trained initial STA-ConvLSTM sound speed prediction model is tested and evaluated, specifically using the root mean square error and mean absolute error for evaluation.

[0065] The calculation formula of the root mean square error is: ; Here, RMSE is the root mean square error.

[0066] The formula for calculating the mean absolute error is: ; Among them, MAE is the mean absolute error.

[0067] Step S310: Determine the trained initial STA-ConvLSTM sound speed prediction model as the STA-ConvLSTM sound speed prediction model.

[0068] Furthermore, the STA-ConvLSTM sound speed prediction model is used to predict the seawater sound speed at multiple spatial scales. According to the sound speed information requirements to be predicted, the corresponding spatial scale features are selected, including sound speed profile, sound speed section and sound speed structure.

[0069] Furthermore, the STA-ConvLSTM sound speed prediction model structure constructed by the present invention is as follows: Figure 3 As shown in the figure, the STA-ConvLSTM sound speed prediction model has seven layers: an input layer, two convolutional layers, a temporal attention layer, a spatial attention layer, a fully connected layer, and an output layer. By testing different network structures, we obtained optimized parameter values: a batch size of 16, a ReLU activation function, a convolution kernel size of 7×7, a number of filters of 32, and a stacking layer of 2. The details of each layer are as follows: Input layer: The model receives target prediction sea area data through the input layer, providing input for subsequent layers. The input shape is (number of samples, time steps, number of depths, number of rows in the sound velocity structure, number of columns in the sound velocity structure, number of features), and the specific data is (130, 12, 58, 2, 2, 1).

[0070] Convolutional layers: The input layer is followed by convolutional layers, including the first and second ConvLSTM layers. The first ConvLSTM layer uses 64 filters and a 7×7 convolution kernel to capture the local features and temporal correlations of the input data by performing convolutions in both time and space. The ReLU activation function then introduces nonlinearity to the model. "Padding" (adding extra pixels to the edges of the input data) is set to "same" to keep the output size the same as the input size, and "return_sequences" (a Boolean parameter used to control the output behavior of a long short-term memory network or other recurrent neural network layer) is set to "True" to retain the output at all time steps.

[0071] The second ConvLSTM layer is similar to the first ConvLSTM layer and also uses 64 filters and 7×7 convolution kernels to further extract the features of the input data, enhancing the STA-ConvLSTM sound speed prediction model's ability to capture temporal and spatial information.

[0072] Temporal Attention Layer: The output of the second ConvLSTM layer is used as the input of the temporal attention layer to focus on the importance of different time steps.

[0073] Spatial Attention Layer: A spatial attention layer is followed by the temporal attention layer to better capture the information of key spatial locations.

[0074] Fully connected layer: After extracting spatiotemporal features, a fully connected layer is added to concatenate the original convolutional layer output and the spatial attention layer output along the channel dimension. This combines the original features with the attention-weighted features, allowing the model to retain the original information while also capturing key information through the attention mechanism.

[0075] Output layer: A 2D convolutional layer with a single filter and a 7×7 kernel is used to map the concatenated feature maps into a sound velocity profile prediction, which serves as the output layer. The default activation function for the output layer is linear, and the output value directly represents the prediction result. "padding" is set to "same" to keep the output size the same as the input size, and "data_format" is set to "channels_last (in channels_last mode, the data dimension order is (number of samples, rows or height, columns or width, number of channels))" to maintain the same channel order as the input data.

[0076] Furthermore, the STA-ConvLSTM sound speed prediction model needs to determine the batch number, activation function, convolution kernel size, number of filters and number of stacked layers. The initial batch number is set to 16, the activation function is ReLU, the convolution kernel size is 3×3, the number of filters is 32, and the number of stacked layers is 2. The optimizer selects the Adam algorithm, the number of iterations is uniformly set to 200, and the mean square error is used as the loss and evaluation indicator. The optimized parameter values ​​are obtained by testing different network structures.

[0077] Step S4: Input the input data into the STA-ConvLSTM sound speed prediction model to obtain the seawater sound speed data of the target predicted sea area, wherein the seawater sound speed data of the target predicted sea area includes: sound speed profile, sound speed section and sound speed structure.

[0078] Furthermore, the present invention uses the STA-ConvLSTM sound speed prediction model to predict the seawater sound speed from January to December of a certain year. The results are as follows: Figure 5 As shown, it can be seen that the present invention can achieve accurate prediction of the sound velocity structure.

[0079] To further verify the effectiveness and superiority of the present invention, the above process is synchronously executed using the ConvLSTM model, and the prediction results of the STA-ConvLSTM sound speed prediction model and the ConvLSTM model are compared, as shown in Table 1: Table 1 Prediction results of STA-ConvLSTM sound speed prediction model and ConvLSTM model ;

[0080] According to Table 1, compared with the ConvLSTM model, the STA-ConvLSTM sound speed prediction model shows lower prediction error and better prediction effect.

[0081] The present invention provides a multi-spatial-scale seawater sound speed prediction method that can extract the temporal and spatial characteristics of seawater sound speed historical data, thereby achieving accurate prediction of seawater sound speed. Furthermore, by setting the dimensions of the sound speed data, multi-spatial-scale prediction of seawater sound speed can be achieved, including sound speed profiles, sound speed cross sections, and sound speed structures. The method can be widely used in many fields such as marine military, underwater detection and communication, and scientific research.

[0082] A multi-spatial-scale seawater sound velocity prediction system, such as Figure 6 As shown, including: The data acquisition module is used to obtain the current target predicted sea area data; the current target predicted sea area data includes: time, spatial position, sea water temperature, sea water salinity and sea water depth.

[0083] The data processing module is used to perform format conversion and normalization on the current target predicted sea area data to obtain input data.

[0084] A model building module is used to build a STA-ConvLSTM sound speed prediction model based on the spatiotemporal attention mechanism, convolutional neural network, and long short-term memory network.

[0085] The prediction module is used to input the input data into the STA-ConvLSTM sound speed prediction model to obtain the seawater sound speed data of the target prediction sea area. The target prediction seawater sound speed data includes: sound speed profile, sound speed section, and sound speed structure.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical method of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical method to deviate from the spirit and scope of the technical method of the present invention.

Claims

1. A multi-spatial-scale seawater sound velocity prediction method, characterized in that: The following steps are involved: S1. Obtaining the current target predicted sea area data; the current target predicted sea area data includes: time, spatial location, seawater temperature, seawater salinity and seawater depth; S2. Perform format conversion and normalization processing on the current target predicted sea area data to obtain input data; S3. Based on the spatiotemporal attention mechanism, convolutional neural network and long short-term memory network, a STA-ConvLSTM sound speed prediction model is constructed; S4. Input the input data into the STA-ConvLSTM sound speed prediction model to obtain the seawater sound speed data of the target predicted sea area; the seawater sound speed data of the target predicted sea area includes: a sound speed profile, a sound speed section and a sound speed structure.

2. The multi-spatial-scale seawater sound velocity prediction method according to claim 1, characterized in that: In S3, the STA-ConvLSTM sound speed prediction model is constructed based on the spatiotemporal attention mechanism, convolutional neural network, and long short-term memory network. The specific steps include: S301, obtaining historical target prediction sea area data; S302, calculating the seawater sound velocity data of the historical target predicted sea area based on the empirical formula of sound velocity according to the historical target predicted sea area data; S303, performing format conversion and normalization processing on the historical target prediction sea area data and the historical target prediction sea area seawater sound speed data to obtain the historical target prediction sea area input data and the historical target prediction sea area seawater sound speed input data; S304. Construct a training set, a validation set, and a test set based on the historical target prediction sea area input data and the historical target prediction sea area seawater sound speed input data; S305. Construct a convolutional layer according to a convolutional neural network and a long short-term memory network; the convolutional layer includes: a first ConvLSTM layer and a second ConvLSTM layer; S306. Construct a temporal attention layer and a spatial attention layer according to the spatiotemporal attention mechanism; S307, connecting the convolution layer, the temporal attention layer, and the spatial attention layer to obtain an initial STA-ConvLSTM sound speed prediction model; S308, inputting the training set and the validation set into the initial STA-ConvLSTM sound speed prediction model for training to obtain a trained initial STA-ConvLSTM sound speed prediction model; S309, inputting the test set into the trained initial STA-ConvLSTM sound speed prediction model, testing and evaluating the trained initial STA-ConvLSTM sound speed prediction model, and obtaining a trained initial STA-ConvLSTM sound speed prediction model; S310: Determine the trained initial STA-ConvLSTM sound speed prediction model as the STA-ConvLSTM sound speed prediction model.

3. The multi-spatial-scale seawater sound velocity prediction method according to claim 2, characterized in that: In S305, a convolutional layer is constructed based on a convolutional neural network and a long short-term memory network, specifically including: By using filtering and convolution operations, the input data of the historical target prediction sea area is convolved in time and space to obtain the local characteristics and time characteristics of the input data of the historical target prediction sea area; A nonlinear activation function is used to activate the local features and the time features respectively to obtain local activation features and time activation features; the nonlinear activation function includes: a ReLU function and a Sigmoid function; Perform element-wise multiplication of the local activation feature and the temporal activation feature to obtain the first output feature data and the first ConvLSTM layer; Performing feature extraction on the first output feature data by using filtering and convolution operations to obtain second output feature data and a second ConvLSTM layer; Connect the first ConvLSTM layer and the second ConvLSTM layer to obtain a convolutional layer.

4. The multi-spatial-scale seawater sound velocity prediction method according to claim 3, characterized in that: In S306, the temporal attention layer and the spatial attention layer are constructed according to the spatiotemporal attention mechanism, specifically including: Through the first convolution operation, features of different time steps are extracted to obtain time step feature data; the first convolution operation adopts a 1×1 convolution layer; Based on the fully connected layer, the first gating factor is calculated to determine the importance of the temporal attention weight in the time step feature data; According to the importance of the temporal attention weight, the temporal attention weight is calculated using the first convolution operation and the nonlinear activation function; Normalize the temporal attention weight to obtain the normalized temporal attention weight; Multiply the normalized temporal attention weight by the second output feature data and the first gating factor element by element, construct a temporal attention layer, and obtain the third output feature data; Through the second convolution operation, features at different spatial positions are extracted to obtain spatial position feature data; the second convolution operation uses a 7×7 convolution layer; Based on the fully connected layer, the second gating factor is calculated to determine the importance of the spatial attention weight in the spatial position feature data; According to the importance of the spatial attention weight, the spatial attention weight is calculated using the second convolution operation and the nonlinear activation function; Normalize the spatial attention weight to obtain the normalized spatial attention weight; The normalized spatial attention weight is sequentially multiplied element-wise with the third output feature data and the second gating factor to construct a spatial attention layer to obtain the fourth output feature data.

5. The multi-spatial-scale seawater sound velocity prediction method according to claim 2, characterized in that: In S308, the training set and the validation set are input into the initial STA-ConvLSTM sound speed prediction model for training to obtain the trained initial STA-ConvLSTM sound speed prediction model, which specifically includes: When the initial STA-ConvLSTM sound speed prediction model is trained using the training set and the validation set, the mean square error is used as the loss value, and the early stopping method monitoring model is used to detect the loss value. The training is stopped until the decrease in the loss value meets the set requirements, and the trained initial STA-ConvLSTM sound speed prediction model is obtained; the set requirement is that the loss value does not decrease in 20 rounds of calculation.

6. The multi-spatial-scale seawater sound velocity prediction method according to claim 2, characterized in that: In S309 , the trained initial STA-ConvLSTM sound speed prediction model is tested and evaluated, specifically using the root mean square error and mean absolute error for evaluation.

7. The multi-spatial-scale seawater sound velocity prediction method according to claim 6, characterized in that: The calculation formula of the root mean square error is: ； Among them, RMSE is the root mean square error, is the number of training sets, For the Input data of seawater sound speed in historical target prediction sea area, For the The seawater sound speed prediction data of historical target prediction sea areas; The calculation formula of the mean absolute error is: ； Among them, MAE is the mean absolute error.

8. The multi-spatial-scale seawater sound velocity prediction method according to claim 1, characterized in that: In S2, the current target predicted sea area data is format converted and normalized to obtain input data, including: Convert the current target prediction sea area data into the input data format of the STA-ConvLSTM model to obtain 6D tensor data; the 6D tensor data includes: number of samples, time steps, depth, number of rows, number of columns and number of features; The maximum and minimum normalization method is used to normalize the 6D tensor data to obtain the input data.

9. The multi-spatial-scale seawater sound velocity prediction method according to claim 8, characterized in that: The calculation formula for the input data is: ； in, is the input data, is 6D tensor data, is the maximum value in the 6D tensor data, is the minimum value in the 6D tensor data.

10. A multi-spatial-scale seawater sound velocity prediction system, characterized in that: include: Data acquisition module, used to obtain the current target predicted sea area data; The current target predicted sea area data includes: time, spatial location, seawater temperature, seawater salinity and seawater depth; The data processing module is used to perform format conversion and normalization on the current target prediction sea area data to obtain input data; A model building module, which is used to build a STA-ConvLSTM sound speed prediction model based on the spatiotemporal attention mechanism, convolutional neural network, and long short-term memory network; The prediction module is used to input the input data into the STA-ConvLSTM sound speed prediction model to obtain the seawater sound speed data of the target predicted sea area; the seawater sound speed data of the target predicted sea area includes: sound speed profile, sound speed section and sound speed structure.

Images