Sound velocity distribution forecasting method based on Mama neural network

Abstract

The invention discloses a sound velocity distribution forecasting method based on a Mama neural network, and belongs to the technical field of ocean observation. According to the method, the historical profile data of the forecasting task area is used as a reference, the historical profile data are arranged according to the time dimension, and then the sound velocity distribution of the forecasting area at the future moment is forecasted. According to the method, the problem of prediction of underwater sound velocity profile distribution in the time dimension is solved, sound velocity distribution at future moments can be accurately predicted in real time, and sound velocity distribution at specified time resolution intervals in the future is estimated; the problems of insufficient time resolution of ocean sound velocity distribution estimation and lack of future sound velocity forecasting function are solved, and important guiding significance is provided for ocean scientific research.

Description

Technical Field

[0001] The invention specifically relates to a sound speed distribution prediction method based on a Mamba neural network, and belongs to the technical field of ocean observation. Background Art

[0002] The complex underwater environment can easily cause errors in the data of sound field measurement, and the high acquisition cost of sound velocity profile data leads to insufficient training sample data, which leads to low model accuracy. In order to improve the anti-interference ability of the sound velocity field construction model, Wuhan University proposed an auto-encoding feature-mapping neural network (AEFMNN) structure, which significantly improved the anti-interference ability by denoising and reconstructing the sound field data through the auto-encoding network, but the application of this method is limited by the high equipment research and development cost. In order to achieve the expansion of sound velocity data in the time dimension and then predict future data, the Shandong University team proposed a sound velocity profile prediction method based on radial basis function (RBF) neural network. This method uses historical temperature, salinity profiles and sampling position information to construct the target area temperature and salinity field network, uses historical profile data and sampling position information to construct the sound velocity field network, and constrains the sound velocity field network through the temperature and salinity field network. Although this method does not require actual sound field observation data, the construction result tends to the sample mean, which is difficult to accurately describe the temporal and spatial changes of the sound velocity profile. In order to reduce the dependence on real-time measurement of underwater temperature, salinity and sound speed data, Ocean University of China introduced a time-dimensional sound speed prediction method based on long short-term memory (LSTM) neural network. This method first arranges the sound speed profile data according to the time dimension, then performs layered processing in the spatial dimension, trains and predicts the network on each depth layer, and finally combines the sound speed prediction data of each layer to form a complete sound speed profile data. This method gets rid of the dependence on data information such as salinity, temperature and location, and can quickly predict future sound speed profile data, but there is still room for improvement in model prediction accuracy and real-time performance.

[0003] In summary, most of the current methods for constructing the sound velocity field are based on the spatial dimension and can make relatively accurate predictions, but there is currently little research on the time dimension, and there is still much room for improvement in the prediction accuracy and real-time performance of the model. Summary of the invention

[0004] The purpose of the present invention is to provide a sound speed distribution prediction method based on Mamba neural network to make up for the deficiencies of the prior art.

[0005] Most of the existing underwater sound speed profile distribution estimation methods work based on real-time sonar observation data, which makes it difficult to achieve real-time estimation and prediction of sound speed distribution. Underwater sound speed profile methods that are not based on sonar observation data have insufficient prediction accuracy and cannot accurately capture the periodic variation law of sound speed distribution. The present invention proposes a sound speed distribution prediction method based on the Mamba neural network. The method uses the historical profile data of the forecast task area as a reference, arranges the historical profile data according to the time dimension, and then predicts the sound speed distribution of the forecast area at future times, so as to solve the problem of underwater sound speed profile prediction in the time dimension and achieve real-time and accurate estimation and prediction of ocean sound speed.

[0006] To achieve the above object, the present invention adopts the following technical solutions: A method for predicting sound speed distribution based on Mamba neural network, the specific steps are as follows: S1: Determination of sound speed distribution forecast locations and extraction of spatial historical sound speed datasets; S2: The time resolution, forecast start time and maximum forecast time length of the sound speed distribution forecast are determined; S3: Secondary screening of the spatial historical sound speed dataset, as well as standardized interpolation of depth and time dimensions; S4: dividing the sound velocity distribution sample data processed in S3 into a training set, a validation set and a test set, and performing data preprocessing; S5: constructing a Mamba neural network sound speed distribution prediction model, and performing model training and verification; the Mamba neural network sound speed distribution prediction model includes an input layer, a Mamba block and an output layer; S6: Denormalize the predicted output data output by the model; S7: Calculate the root mean square error RMSE. If the RMSE value has no significant change compared with the previous round, proceed to S8, otherwise return to S5; S8: Use the trained Mamba neural network sound speed distribution prediction model to perform test predictions and output the prediction results.

[0007] Furthermore, in S1, the spatial coordinates of the task area are predicted by the speed of sound , the historical sound speed profile data of the target area in the GDCSM_Argo dataset are selected and arranged vertically according to the spatial dimension and horizontally according to the time dimension to form a historical dataset as the reference data of the sound speed profile data.

[0008] Furthermore, in S2, the prediction time resolution is determined according to the sound speed distribution prediction task, using Indicates; forecast starting point, using Indicates; the maximum forecast time length, using express.

[0009] Furthermore, in S3, according to the time resolution of the sound speed distribution forecast determined in S2, data is dynamically selected from the spatial historical sound speed data set in S1 to form the time-standardized data of the historical sound speed distribution as a training reference data set; wherein the time range covered by the training data contains at least 3 whole-cycle data, such as when the forecast time resolution is "month", the required sample data should contain data of no less than 3 years before the forecast time. The selected data is sequentially interpolated by standardized linear layers in the vertical depth direction, and the difference is standardized by the time dimension in the horizontal direction. Among them, the standardized linear layered interpolation method in the vertical depth direction is as follows: from 0 to 10 meters, one layer is divided every 5 meters; from 10 to 180 meters, one layer is divided every 10 meters; from 180 to 460 meters, one layer is divided every 20 meters; from 460 to 500 meters, it is divided into a single layer; from 500 to 1300 meters, one layer is divided every 50 meters; from 1300 to 1900 meters, one layer is divided every 100 meters; the depth exceeding 1900 meters is one layer, and a total of 58 layers are divided; the screened data are interpolated in the time dimension, and the interpolation time interval is equal to the forecast time resolution.

[0010] Furthermore, the S3 is specifically: S3-1: Historical Normalized Sound Speed ​​Profile Dataset , expressed as: (1); in, Expressed as The entire data set consists of The sound velocity profile data are arranged according to the time dimension, and the time interval and time resolution consistent; and Sound velocity profile data Further expressed as: (2); The speed of sound , unit: m / s, Expressed as Layer sound velocity profile data, China Layer data, and in formula (2) is transposed; S3-2: Standardized sound velocity profile data set according to The first The layer data is expanded to show: (3); The standardized screening and collection of the spatial historical sound speed data set is thus completed, and S4 is entered.

[0011] Further, the S4 includes: S4-1: Divide the training set, validation set and test set: According to the natural variation law of sound speed data, the data variation time period is set to , historical normalized sound velocity profile dataset The front The sound speed distribution data is used as the training set. To front The sound speed distribution data is used as the validation set, and finally The sound speed distribution data is used as the test set, where ; S4-2: Data preprocessing: Standardized spatial historical sound speed dataset Center front The columns are used as training sets, as shown in formula (4), which is expressed as ; forward To front The column is used as the validation set, as shown in formula (6), which is expressed as ; at last The columns are used as the test set, as shown in formula (7), which is expressed as ; First, the training set According to the normalization process in the matrix row direction, we can get , then according to the training set The parameters standardized in the validation set and test set Perform standardization operations to obtain and ; Then the standardized training set is further divided into training input training set And output training set , a sliding window is used for the training set, and the sliding window size is set to be consistent with the change time period. , the sliding step is 1, that is, use the previous The data in the next column is predicted by the column data. After training once, slide one column to the right, and so on. The training input set And output training set As shown in formula (8) and (9) respectively: (4); (5); (6); (7); (8); (9); Then, the sound velocity distribution sample data preprocessing is completed and the process proceeds to S5.

[0012] Further, the S5 includes: S5-1: Construction of neural network model: According to the different depth layers, a Mamba neural network is built on each depth layer and the sound speed distribution prediction is performed; For each depth layer of the Mamba network, the input time series passes through the input layer, Mamba block and output layer in sequence; the prediction result is an output sound speed value time series with a length of 1, and the difference between the output sound speed value time series and the input time series is 12 time resolutions; the linear layer in the output layer adjusts the output result to a 1-dimensional time series, and Tanh represents the Tanh activation function: (10); This ensures that the output result is projected within the range of [-1,1]; Mamba introduces a selectivity mechanism based on the State Space Model (SSM) and adjusts the SSM parameters in an input-dependent manner, allowing the model to selectively transmit or forget information based on the current data.

[0013] Take N layers of Mamba blocks as hidden layers, and each hidden layer contains M neurons.

[0014] The overall calculation process of the Mamba block can be described as: (12); in represents the system input, represents the hidden state vector at the current moment, Represents the output results at discrete time moments; represents the output matrix; and They are respectively expressed as the discretized state transfer matrix and input matrix, and further expressed as: (13); (14); In the formula is the discretized time step; (2) Model training During model training, the Mamba network on each depth layer is trained separately; before model training, the network training learning rate on each depth layer is determined And the number of training iterations ,in Represents the depth layer, and its value is .

[0015] First depth layer Taking the network as an example, in each iteration of the training process, the input dimension length of the model is Time series of sound speed values , the corresponding model outputs a time series of sound speed values ​​with a length of 1 ,m is , corresponding to the input and output For sequence; training Depth Layer In each iteration, the input and output of the model are the training input set And output training set No. Row data, that is and ,in and They represent the first row of data of the training input set and output set, that is, the data of the first depth layer. The training method of the remaining depth layers is similar.

[0016] (3) Model validation First, use the training set The End The column data is used as the validation input of the model to predict the data of a future time resolution, that is, the sequence data of length 1. Then, the standardized training set is used The End Column data and validation set The first column of data is concatenated into a length of The sequence data of is used as input to predict the data of a future time resolution, that is, the sequence data of length 1. The standardized training set and validation set Slide back one position and splice to form a length of As input, we predict the data of a future time resolution in sequence until we have traversed the entire standardized validation set, totaling When validating the first depth layer, the input of the model is the training set The first line ends Column Data , the first data predicted by the sound speed verification data at the next moment can be obtained ,in Represents the first result of the first depth layer prediction. It is necessary to predict a data change time period, that is, to predict Step 1. Continue to input predictions into the model and use the standardized training set data At the end of the first line Column Data and a standardized validation set The first column of data is concatenated into a length of The sequence is used as input to obtain the second data of the sound speed verification data prediction , and so on, until the verification of Prediction results , the predictions of the remaining depth layers are similar to the first depth layer. Represents the first depth layer Network speed verification data predicts the future The speed of sound value with a time resolution.

[0017] The validation prediction result of the first depth layer is of length The time series of the speed of sound can be expressed as: (15); The same is true for the other depth layers, so the final sound velocity data set for verification and prediction is , which can be expressed as: (16); After the model is verified, proceed to S6.

[0018] Furthermore, in S6, since the data preprocessing is standardized before model training, the predicted sound speed output data is denormalized (the inverse operation of the original standardization operation), that is, the predicted data is restored to the scale range of the original data, which is convenient for comparison and analysis with the verification data.

[0019] Prediction results after denormalization , expressed as: (17); in, Representative The depth layer The predicted sound speed value with a time resolution of ; After the denormalization operation, the process goes to S7.

[0020] Further, in S7, the root mean square error (RMSE) between the predicted sound velocity value data of each depth layer and the verification data is calculated. The root mean square error of the layer depth layer is specifically expressed as: (18).

[0021] Further, the S8 includes: First, use the entire validation set (Length is ) as the input of the model prediction to predict the data of a future time resolution, that is, the sequence data of length 1. Then, use the validation set The End Column data and test set The first column of The time series is used as the input of the model prediction to predict the data of a future time resolution, that is, the sequence data of length 1. and test set Slide back one position and splice into a length of The time series is used as the input of the model prediction to predict the data of a future time resolution in sequence until the entire test set is traversed ,total When predicting the first depth layer, the input of the model is the validation set of total Column Data , the first data of the sound speed prediction at the next moment can be obtained ,in Represents the first result of the first depth layer prediction. It is necessary to predict a data change time period, that is, to predict Step. Continue to make predictions and use the validation set After Column Data and test set The first column of data is concatenated into a length of The time series is used as the model input to obtain the second data of the sound speed data prediction at the next moment. , and so on, until the verification of Prediction results , the predictions of the remaining depth layers are similar to the first depth layer. Represents the first depth layer The future of network speed data prediction The speed of sound value with a time resolution of .

[0022] The prediction result of the first depth layer is a length of The time series of the speed of sound is expressed as: (19); The same is true for the other depth layers, so the final predicted sound speed data set is , expressed as: (20); Then the prediction result is transferred to S6 for denormalization operation to obtain the final prediction result.

[0023] Compared with the prior art, the present invention has at least the following beneficial effects:

[0024] The present invention solves the problem of predicting the underwater sound speed profile distribution in the time dimension, can make real-time and accurate predictions of the sound speed distribution at future moments, and estimate the sound speed distribution at future specified time resolution intervals, solving the problems of insufficient time resolution of ocean sound speed distribution estimation and lack of future sound speed prediction function, and provides important guiding significance for marine scientific research. The present invention can make relatively accurate predictions of the sound speed distribution at future specified time resolution intervals in the target area, and the prediction efficiency is also high, providing important guiding significance for marine scientific research, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a schematic diagram of the sound speed distribution prediction method based on Mamba neural network.

[0026] Figure 2 This is a schematic diagram of the Mamba neural network model.

[0027] Figure 3 This is a diagram of the Mamba block structure.

[0028] Figure 4 It is a schematic diagram comparing the predicted full-sea-depth sound speed distribution and the actual full-sea-depth sound speed distribution.

[0029] Figure 5 This is a schematic diagram comparing the predicted trend of the 4th depth layer (corresponding to a sea depth of 20 meters) with the original trend. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in combination with specific embodiments and with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] Example 1 A method for predicting sound velocity distribution based on Mamba neural network, such as Figure 1 As shown, the specific implementation steps are as follows: Step 1: Determine the location for sound speed distribution forecast and extract spatial historical sound speed dataset The spatial location of the sound speed prediction task area can be specified according to the actual needs of the user.

[0032] Predicting the spatial coordinates of the mission area using the speed of sound Taking this as an example, the sound speed profile data measured in the target area for a total of 5 years and 60 months from 2019 to 2023 are selected from the GDCSM_Argo dataset and arranged vertically according to the spatial dimension and horizontally according to the time dimension to form a historical dataset as the reference data for the sound speed profile data.

[0033] Step 2: Determine the time resolution, forecast start time and maximum forecast time length of the sound speed distribution forecast Determine the forecast time resolution based on the sound speed distribution forecast task For 1 month; forecast starting point, use is January 2023; the maximum forecast time length is For 1 year.

[0034] Step 3: Secondary screening of the spatial historical sound speed dataset and standardized interpolation of depth and time dimensions The time resolution of the sound speed distribution prediction determined in step 2 is , dynamically select data from the spatial historical sound speed dataset in step 1 to form the time-standardized data of the historical sound speed distribution as the training reference dataset. The time range covered by the training data contains at least 3 full-cycle data. In this example, the forecast time resolution is selected as "month", so the required sample data should contain data no less than 3 years before the forecast time. A total of 5 years and 60 months of data from 2019 to 2023 in this spatial area is used as the spatial historical reference dataset. The forecast time is from January 2023 to December 2023. Before that, there is a complete 4-year full-cycle data as reference data, which meets the training requirements.

[0035] The selected data are interpolated in accordance with the standardized linear layering in the vertical depth direction, and the standardized difference in the horizontal direction is in accordance with the time dimension. The standardized linear layering interpolation method in the vertical depth direction is as follows: from 0 to 10 meters, one layer is divided every 5 meters; from 10 to 180 meters, one layer is divided every 10 meters; from 180 to 460 meters, one layer is divided every 20 meters; from 460 to 500 meters, one layer is divided into a single layer; from 500 to 1300 meters, one layer is divided every 50 meters; from 1300 to 1900 meters, one layer is divided every 100 meters; the depth exceeding 1900 meters is one layer, and a total of 58 layers are divided. The selected data are interpolated in the time dimension, and the interpolation time interval is equal to the forecast time resolution.

[0036] Let the historical normalized sound speed profile dataset , expressed as: (1); in, Expressed as The entire data set contains sound velocity profile data. There are 60 in total, arranged in chronological order from January 2019 to December 2023, a total of 5 years (60 months). Sound velocity profile data It can be further expressed as: (2);

[0037] The speed of sound , unit: m / s. Expressed as Layer sound velocity profile data, There are 58 layers of data in total. And in formula (2) is transposed.

[0038] The standardized sound velocity profile data set is expanded according to the sound velocity value data of 58 rows and 60 columns to show the specific representation: (3); The standardized screening and collection of the spatial historical sound speed data set is now completed, and we proceed to step 4.

[0039] Step 4: Preprocess the sound velocity distribution sample data The sound speed distribution sample data is mainly divided into three parts: training set, validation set and test set, and the data is preprocessed.

[0040] (1) Divide the training set, validation set, and test set According to the natural variation law of sound speed data, the data variation time period is set to 12, historical normalized sound velocity profile data set The first 36 sound speed distribution data in are used as training sets, the first 37 to the first 48 sound speed distribution data are used as validation sets, and the last 12 sound speed distribution data are used as test sets.

[0041] (2) Specific preprocessing of the data set Standardized spatial historical sound speed dataset The first 36 columns are used as training sets, expressed as ; The first 37 to 48 columns are used as the validation set, expressed as ; The last 12 columns are used as the test set, expressed as ; Training set: ; Validation set: ; Test set: .

[0042] First, the training set According to the normalization process in the matrix row direction, we can get , then according to the training set The parameters standardized in the validation set and test set Perform standardization operations to obtain and ; The standardized training set: .

[0043] Then the standardized training set is further divided into training input training set And output training set , the training set is input by sliding window, the sliding window size is set to 12, and the sliding step is 1, that is, the first 12 columns of data are used to predict the next column of data, and after training once, slide one column to the right, and so on. Training input set And output training set As shown in formula (4) and (5) respectively: (4); (5); The sound velocity distribution sample data preprocessing is then completed, and the process proceeds to step 5.

[0044] Step 5: Construct a model for predicting sound velocity distribution using the Mamba neural network The construction of the Mamba neural network model is divided into three parts: neural network model construction, model training and model verification.

[0045] (1) Construction of neural network model According to the different depth layers, a Mamba neural network is built on each depth layer and the sound speed distribution prediction is performed.

[0046] The prediction result is an output sound speed value time series with a length of 1, and the time difference between the output sound speed value time series and the input time series is 12 months.

[0047] Take 6 layers of Mamba blocks as hidden layers, and each hidden layer contains 32 neurons.

[0048] The Mamba neural network structure is as follows Figure 2As shown in the figure, for each depth layer of the Mamba network, the input time series passes through the input layer, Mamba block and output layer in sequence; the prediction result is an output sound speed value time series with a length of 1, and the difference between the output sound speed value time series and the input time series is 12 time resolutions; the linear layer in the output layer adjusts the output result to a 1-dimensional time series, and Tanh represents the Tanh activation function: (10); This ensures that the output result is projected within the range of [-1,1]; Mamba introduces a selectivity mechanism based on the State Space Model (SSM) and adjusts the SSM parameters in an input-dependent manner, allowing the model to selectively transmit or forget information based on the current data.

[0049] Take N layers of Mamba blocks as hidden layers, and each hidden layer contains M neurons.

[0050] Mamba block structure is as follows Figure 3 As shown in the figure, "+" and "×" represent addition and multiplication operations respectively. Represents the SiLU activation function: (11); SiLU can provide better performance in many tasks by smoothing nonlinear characteristics.

[0051] The overall calculation process of the Mamba block can be described as: (12); in represents the system input, represents the hidden state vector at the current moment, Represents the output results at discrete time moments; represents the output matrix; and They are respectively expressed as the discretized state transfer matrix and input matrix, and further expressed as: (13); (14); In the formula is the discretized time step.

[0052] (2) Model training When training the model, the Mamba network on each depth layer is trained separately.

[0053] Main parameter settings for model training: learning rate is 0.001, the number of training iterations is 200, Represents the depth layer, and its value is .

[0054] At the 1st depth layer For example, in each iteration of the training process, the model's input dimension is a time series of sound speed values ​​with a length of 12. , the corresponding model outputs a time series of sound speed values ​​with a length of 1 ,m is , corresponding to the input and output For sequence; train the second depth layer In each iteration, the input and output of the model are the training input set And output training set The second row of data, that is and , where m is . and Represents the second row of data of the training input set and output set, that is, the data of the second depth layer. The training method of the remaining depth layers is similar.

[0055] (3) Model validation First, use the training set The last 12 columns of data are used as the validation input of the model to predict the data of the future time resolution, that is, the sequence data of length 1. Then, the standardized training set is used The last 11 columns of data and the validation set The first column of data is concatenated into a sequence data of length 12 as input to predict the data of a future time resolution, that is, the sequence data of length 1. The standardized training set is sequentially and validation set Slide back one position and concatenate to form a sequence data of length 12 as input to predict the data of one time resolution in the future in sequence until the entire standardized validation set is traversed, totaling 12 sets of sequences.

[0056] When validating the first depth layer, the input of the model is the training set The first row and last 12 columns of data , the first data predicted by the sound speed verification data at the next moment can be obtained ,in Represents the first result of the first depth layer prediction. It is necessary to predict a data change time period, that is, to predict 12 steps. Continue to input predictions into the model and pass the standardized training set data The last 11 columns of the first row of data and a standardized validation set The first column of data is concatenated into a sequence of length 12 as input, and the second data predicted by the sound speed verification data can be obtained. , and so on, until the verification of Prediction results , the predictions of the remaining depth layers are similar to the first depth layer.

[0057] The verification prediction result of the first depth layer is a time series of sound speed values ​​with a length of 12, which can be expressed as: (6); The same is true for the other depth layers, so the final sound velocity data set for verification and prediction is , which can be expressed as: (7); After the model is verified, go to step 6.

[0058] Step 6: Denormalize the predicted output data Since the data preprocessing was standardized before model training, the predicted sound speed output data was denormalized (the inverse operation of the original standardization operation), that is, the predicted data was restored to the scale range of the original data, which is convenient for comparative analysis with the verification data.

[0059] Prediction results after denormalization , which can be expressed as: (8); After the denormalization operation is completed, go to step 7.

[0060] Step 7: Calculate the root mean square error (RMSE) Calculate the root mean square error (RMSE) between the predicted sound velocity data and the verification data at each depth layer. The root mean square error of the layer depth layer can be specifically expressed as: (9); If the RMSE value has no significant change compared to the previous round, proceed to step 8, otherwise return to step 5.

[0061] Step 8: Model prediction First, use the entire validation set (length is 12) as the input of the model prediction to predict the data of a future time resolution, that is, the sequence data of length 1. Then, use the validation set The last 11 columns of data and test set The first column of is concatenated into a time series of length 12 as the input of the model prediction to predict the data of a future time resolution, that is, the sequence data of length 1. and test set Slide back one position and splice it into a time series of length 12 as the input of the model prediction, and predict the data of the next month in sequence until the entire test set is traversed. , a total of 12 sequences.

[0062] When predicting the first depth layer, the input of the model is the validation set There are 12 columns of data , the first data of the sound speed prediction at the next moment can be obtained ,in Represents the first result of the first depth layer prediction. It is necessary to predict a data change time period, that is, to predict 12 steps. Continue to predict and use the validation set The last 11 columns of data and test set The first column of data is concatenated into a length of The time series is used as the model input to obtain the second data of the sound speed data prediction at the next moment. , and so on, until the 12th prediction result , the predictions of the remaining depth layers are similar to the first depth layer.

[0063] The prediction result of the first depth layer is a time series of sound speed values ​​with a length of 12, which can be expressed as: (10); The same is true for the other depth layers, so the final predicted sound speed data set is , which can be expressed as: (11); Then the prediction result is transferred to step 6 for denormalization to obtain the final prediction result.

[0064] (12).

[0065] In order to verify the accuracy of the predicted sound speed profile, the predicted first full-sea-depth sound speed profile data was compared with the actual first full-sea-depth sound speed profile data.

[0066] Since the time-standardized data set of historical sound speed distribution is not full-sea depth data, but data of 58 depth layers, the predicted result is also data of 58 depth layers. The linear interpolation method is used to interpolate the predicted data and actual data within the full sea depth range. The full sea depth here is 1~1975 meters.

[0067] Prediction data of full ocean depth sound velocity distribution after interpolation processing And the actual data of the full depth sound velocity distribution , which can be expressed as: , (13); in, Represents a sea depth of 1975 meters.

[0068] The root mean square error between the predicted data of the full-sea-depth sound velocity distribution and the actual data of the full-sea-depth sound velocity distribution is: (14).

[0069] The root mean square error (RMSE_all) between the predicted data of the full-sea-depth sound velocity distribution and the actual data of the full-sea-depth sound velocity distribution after multiple model verifications is within 1 m / s, and the average RMSE_all is about 0.69 m / s. Figure 4 As shown in the figure, it can be seen that the predicted sound speed distribution is highly similar to the actual sound speed distribution. Compared with the Lstm network, this method can predict the future sound speed distribution more accurately.

[0070] To ensure the time rigor, a series of sound speed distribution data is consistent in time, so the 4th layer (corresponding to 20 meters) data is randomly selected between depth layers 1 to 58 to test whether this method can accurately capture the periodic law of sound speed distribution. Figure 5 It can be seen that this method can accurately capture the periodic law of sound speed distribution.

[0071] To further explore the high accuracy of this method, this method is compared with the existing LSTM method. The results are shown in the following table:

[0072] Table 1 Comparison of prediction errors of two methods .

[0073] Table 1 gives the results of the two methods for predicting the sound speed data in January and May 2023 at 58 depth layers. It can be seen that the proposed method has better performance and improves the prediction accuracy.

[0074] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects disclosed in the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting sound velocity distribution based on Mamba neural network, characterized in that: The method comprises the following steps: S1: Determination of sound speed distribution forecast locations and extraction of spatial historical sound speed datasets; S2: The time resolution, forecast start time and maximum forecast time length of the sound speed distribution forecast are determined; S3: Secondary screening of the spatial historical sound speed dataset, as well as standardized interpolation of depth and time dimensions; S4: dividing the sound velocity distribution sample data processed in S3 into a training set, a validation set and a test set, and performing data preprocessing; S5: constructing a Mamba neural network sound speed distribution prediction model, and performing model training and verification; the Mamba neural network sound speed distribution prediction model includes an input layer, a Mamba block and an output layer; S6: Denormalize the predicted output data output by the model; S7: Calculate the root mean square error RMSE. If the RMSE value has no significant change compared with the previous round, proceed to S8, otherwise return to S5; S8: Use the trained Mamba neural network sound speed distribution prediction model to perform test predictions and output the prediction results.

2. The method for predicting sound velocity distribution according to claim 1, characterized in that: In S1, the spatial coordinates of the task area are predicted by the speed of sound , the historical sound speed profile data of the target area in the GDCSM_Argo dataset are selected and arranged vertically according to the spatial dimension and horizontally according to the time dimension to form a historical dataset as the reference data of the sound speed profile data.

3. The method for predicting sound velocity distribution according to claim 1, characterized in that: In S2, the prediction time resolution is determined according to the sound speed distribution prediction task, and the prediction time resolution is used Indicates; forecast starting point, using Indicates; the maximum forecast time length, using express.

4. The method for predicting sound velocity distribution according to claim 1, characterized in that: In S3, according to the sound speed distribution forecast time resolution determined in S2, data is dynamically selected from the spatial historical sound speed data set in S1 to form historical sound speed distribution time standardized data as a training reference data set; wherein the time range covered by the training data includes at least 3 integer change cycle data, and the screened data are sequentially interpolated according to the standardized linear layered interpolation in the vertical depth direction, and the difference is standardized according to the time dimension in the horizontal direction; the screened data are interpolated in the time dimension, and the interpolation time interval is equal to the forecast time resolution.

5. The method for predicting sound velocity distribution according to claim 1, characterized in that: The S3 is specifically: S3-1: Historical Normalized Sound Speed ​​Profile Dataset , expressed as: (1)； in, Expressed as The entire data set consists of The sound velocity profile data are arranged according to the time dimension, and the time interval and time resolution consistent; and Sound velocity profile data Further expressed as: (2)； The speed of sound , unit: m / s, Expressed as Layer sound velocity profile data, China Layer data, and in formula (2) is transposed; S3-2: Standardized sound velocity profile data set according to The first The layer data is expanded to show: (3)； The standardized screening and collection of the spatial historical sound speed data set is thus completed, and S4 is entered.

6. The method for predicting sound velocity distribution according to claim 1, characterized in that: The S4 includes: S4-1: Divide the training set, validation set and test set: According to the natural variation law of sound speed data, the data variation time period is set to , historical normalized sound velocity profile dataset The front The sound speed distribution data is used as the training set. To front The sound speed distribution data is used as the validation set, and finally The sound speed distribution data is used as the test set, where ; S4-2: Data preprocessing: Standardized spatial historical sound speed dataset Center front The columns are used as training sets, as shown in formula (4), which is expressed as ; forward To front The column is used as the validation set, as shown in formula (6), which is expressed as ; at last The columns are used as the test set, as shown in formula (7), which is expressed as ; First, the training set According to the normalization process in the matrix row direction, we can get , then according to the training set The parameters standardized in the validation set and test set Perform standardization operations to obtain and ; Then the standardized training set is further divided into training input training set And output training set , a sliding window is used for the training set, and the sliding window size is set to be consistent with the change time period. , the sliding step is 1, that is, use the previous The data in the next column is predicted by the column data. After training once, slide one column to the right, and so on. The training input set And output training set As shown in formula (8) and (9) respectively: (4)； (5)； (6)； (7)； (8)； (9)； After the sound velocity distribution sample data preprocessing is completed, enter S5.

7. The method for predicting sound velocity distribution according to claim 1, characterized in that: The S5 includes: S5-1: Construction of neural network model: According to the different depth layers, a Mamba neural network is built on each depth layer, and the sound speed distribution prediction is performed; the prediction result is an output sound speed value time series with a length of 1, and the difference between the output sound speed value time series and the input time series is 12 time resolutions; the Mamba block calculation process is described as: (12)； in represents the system input, represents the hidden state vector at the current moment, Represents the output results at discrete time moments; represents the output matrix; and They are respectively expressed as the discretized state transfer matrix and input matrix, expressed as: (13)； (14)； In the formula is the discretized time step; S5-2: Model training Determine the network training learning rate at each depth layer before model training And the number of training iterations ,in Represents the depth layer, and its value is ; First Depth Layer In each iteration of the network training process, the input dimension length of the model is Time series of sound speed values , the corresponding model outputs a time series of sound speed values ​​with a length of 1 ,m is , corresponding to the input and output For sequence; training Depth Layer In each iteration, the input and output of the model are the training input set And output training set No. Row data, that is and ,in and Represent the first row of data of the training input set and output set, that is, the data of the first depth layer; S5-3: Model Validation First, use the training set The End The column data is used as the validation input of the model for prediction, predicting the data of a future time resolution, that is, the sequence data of length 1; then, the standardized training set is used The End Column data and validation set The first column of data is concatenated into a length of The sequence data of is used as input to predict the data of a future time resolution, that is, the sequence data of length 1; the standardized training set is and validation set Slide back one position and splice to form a length of As input, we predict the data of a future time resolution in sequence until we have traversed the entire standardized validation set, totaling Group sequence; The validation prediction result of the first depth layer is of length The time series of the speed of sound is expressed as: (15)； The final sound speed dataset for verification prediction is , expressed as: (16)； After the model verification is completed, go to S6.

8. The method for predicting sound velocity distribution according to claim 1, characterized in that: In S6, the prediction result after denormalization processing , expressed as: (17)； in, Representative The depth layer The predicted sound speed value with a time resolution of .

9. The method for predicting sound velocity distribution according to claim 1, characterized in that: In S7, the root mean square error RMSE between the predicted sound velocity value data of each depth layer and the verification data is calculated. The root mean square error of the layer depth layer is specifically expressed as: (18)。 10. The method for predicting sound velocity distribution according to claim 1, characterized in that: The S8 includes: First, the entire validation set As the input of the model prediction, we predict the data of a future time resolution, that is, the sequence data of length 1; then, we use the validation set The End Column data and test set The first column of The time series is used as the input of the model prediction to predict the data of a future time resolution, that is, the sequence data of length 1; the validation set and test set Slide back one position and splice into a length of The time series is used as the input of the model prediction to predict the data of a future time resolution in sequence until the entire test set is traversed ,total When predicting the first depth layer, the input of the model is the validation set of total Column Data , get the first data of the sound speed data prediction at the next moment ,in Represents the first result of the first depth layer prediction; it is necessary to predict a data change time period, that is, to predict Step; continue to predict and use the validation set After Column Data and test set The first column of data is concatenated into a length of The time series is used as the model input to obtain the second data of the sound speed data prediction at the next moment , and so on, until the verification of Prediction results , the predictions of the remaining depth layers are similar to the first depth layer; among them, Represents the first depth layer The future of network speed data prediction The sound velocity value of the time resolution; the prediction result of the first depth layer is the length of The time series of the speed of sound is expressed as: (19)； The final predicted sound speed dataset is , expressed as: (20)； The prediction result is transferred to S6 for denormalization operation to obtain the final prediction result.

Images