Marine sound velocity profile prediction method based on fusion weighted EOF-LSTM

Abstract

The invention belongs to the technical field of underwater acoustic communication, and relates to an ocean sound velocity profile prediction method based on fusion weighted EOF-LSTM, and the method comprises the steps: constructing a historical sound velocity profile data matrix; based on the year difference and the month difference with the target year and month, obtaining a comprehensive time distance, and constructing a Gaussian weighting function; performing normalization processing, calculating a weighted average sound velocity profile, extracting a spatial change mode of the sound velocity profile, realizing dimension reduction and feature extraction, and calculating a principal component; constructing a time window sample by using the principal component time sequence; using LSTM network training to obtain a principal component prediction value, reconstructing a sound velocity profile, and obtaining a target year monthly sound velocity profile prediction value; and extracting a real sound velocity profile value of the target year, and performing error comparison with the predicted value. And a weighting function is constructed for the average sound velocity profile for weighting, an optimal time step length is searched for LSTM input data, an error is used as a reference iterative optimization weight parameter, and effective reduction of a prediction error is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater acoustic communication, and particularly relates to a fusion weighted EOF-LSTM ocean sound speed profile prediction method and device, computer equipment and a storage medium. BACKGROUND

[0002] The sound speed profile is prior information for ocean acoustic field calculation and analysis, and its application covers multiple aspects from ocean environment research, geophysical exploration, acoustic theory verification to underwater acoustic communication, underwater navigation and positioning, ocean resource exploration and exploitation, and the like, and will promote ocean exploration to higher precision and wider field.

[0003] At present, the statistical regression methods for sound speed profile prediction mainly include EOF (Empirical Orthogonal Function) and neural network. Since the actual measurement of underwater sound speed profile is costly and requires expensive instruments, the statistical regression method for sound speed profile can improve convenience and reduce construction cost, but the prediction error is relatively large due to the inherent characteristics of the model and method. SUMMARY

[0004] To solve the above technical problems, the present application provides a fusion weighted EOF-LSTM ocean sound speed profile prediction method, which adopts the following technical solution:

[0005] S1, extracting the sound speed profile data of a preset spatial position in Argo database for m years, constructing a C=N*M historical sound speed profile data matrix, interpolating the sound speed profile data to realize the sound speed value every 1 meter interval, wherein C is the sound speed profile data matrix, N is the profile quantity (time series length), and M is the depth layer number (point number after interpolation) of each profile;

[0006] S2, based on the year difference Y dist and the month difference M dist of the target year and month, obtaining a comprehensive time distance

[0007] T dist = alpha * Y dist + beta * M dist , constructing a Gaussian weight function wherein sigma controls the weight decay range, alpha=1, beta=1, and sigma=8, wherein alpha is the year difference weight coefficient, beta is the month difference weight coefficient, sigma is a parameter for controlling the weight decay range, and omega is the Gaussian weight value;

[0008] S3, normalizing the Gaussian weight function, calculating the weighted average sound speed profile C, using the covariance matrix R, extracting the spatial variation mode of the sound speed profile according to RV=DV, realizing dimension reduction and feature extraction, wherein D is the eigenvalue matrix, V is the eigenvector matrix, the eigenvalues are arranged in descending order, the first K modes with cumulative contribution rate greater than or equal to 0.99 are selected, the EOF mode matrix EOFs is obtained, and the principal components wherein EOFs is a mode matrix composed of the first K eigenvectors, C is the weighted average sound speed profile, PC is the principal component matrix, and K is the number of selected modes, that is, the principal component dimension;

[0009] S4, constructing a time window sample for the principal component time sequence, setting a time step t, and inputting the principal component PC of each time window for t consecutive time steps and outputting the principal component PC of the t+1 time step, wherein t is the time step;

[0010] S5, training the constructed input and output by using the LSTM network to obtain the principal component prediction value of the future time step reconstructing the sound speed profile to obtain the predicted sound speed profile value of the target year and month, wherein is the predicted principal component;

[0011] S6, extracting the real sound speed profile value of the target year, comparing the prediction value with the error, taking the root mean square error RMSE and the mean absolute error MAE as the judgment standard, and constantly iterating a, β, σ and t, outputting the final prediction result and the final root mean square error RMSE and the final mean absolute error MAE, wherein y i is the i-th real value, is the i-th predicted value, and n is the sample number.

[0012] Preferably, the step S1, extracting the sound speed profile data of the preset spatial position in the Argo database in m years, constructing a C=N*M historical sound speed profile data matrix, interpolating the sound speed profile data to realize the sound speed value every 1 meter, wherein C is the sound speed profile data matrix, N is the profile quantity, and M is the depth layer number of each profile, the step specifically comprises:

[0013] S11, extracting the sound speed profile data of the preset spatial position in m years from the global ocean observation system Argo database;

[0014] S12, cleaning the extracted sound speed profile data, and constructing a C=N*M historical sound speed profile data matrix, wherein C is the sound speed profile data matrix, N is the profile quantity (time sequence length), and M is the depth layer number (the number of points after interpolation) of each profile; ​

[0015] S13 interpolated the sound velocity profile data to achieve a sound velocity value every 1 meter.

[0016] Preferably, S2 is based on the year difference Y with the target year and month. dist and the difference in months M dist The overall time distance T is obtained. dist =α·Y dist +β·M dist Construct Gaussian weight function The steps for controlling the weight decay range, where σ controls the weight decay range, and initialization is α = 1, β = 1, σ = 8, where α is the year difference weight coefficient, β is the month difference weight coefficient, σ is the parameter controlling the weight decay range, and ω is the Gaussian weight value, specifically include:

[0017] S21, Extract the year and month timestamp T corresponding to each historical sound velocity profile data. hist And the target prediction year and month T target Calculate the year difference Y respectively dist Difference between months M dist Y dist =|Y target -Y hist |,M dist =|M target -M hist |,

[0018] S22, Calculate the overall time distance T dist =α·Y dist +β·M dist And use the Gaussian kernel function to calculate the Gaussian weight corresponding to each historical data point. Initialize α = 1, β = 1, σ = 8, where α is the year difference weight coefficient, β is the month difference weight coefficient, σ is the parameter that controls the weight decay range, and ω is the Gaussian weight value.

[0019] Preferably, in step S3, the Gaussian weighting function is normalized to calculate the weighted average sound velocity profile. Using the covariance matrix R, and based on RV = DV, the spatial variation modes of the sound velocity profile are extracted to achieve dimensionality reduction and feature extraction. Here, D is the eigenvalue matrix, and V is the eigenvector matrix. The eigenvalues ​​are arranged in descending order, and the top K modes with a cumulative contribution rate ≥ 0.99 are selected to obtain the EOF mode matrix EOFs. Principal components are then calculated. Wherein, EOFs is the modality matrix composed of the first K eigenvectors. For the weighted average sound velocity profile, PC is the principal component matrix, and K is the number of selected modes, i.e., the steps for determining the principal component dimensions specifically include:

[0020] S31, normalize the Gaussian weights so that their sum is 1, the normalization formula is: where ω i is the original weight of the i-th sample, ω i ' is the normalized weight, and N is the total number of samples;

[0021] S32, calculate the weighted average sound speed profile using the normalized weights C i represents the i-th sound speed profile vector;

[0022] S33, calculate the covariance matrix R, where C is the original sound speed profile data matrix, is the weighted average sound speed profile, W is a diagonal matrix with normalized weights ω i ' as diagonal elements, solve the characteristic equation RV = DV, where D is the eigenvalue matrix and V is the eigenvector matrix;

[0023] S34, sort the eigenvalues from large to small: λ1≥λ2≥…≥λ M , and calculate the variance contribution rate and cumulative contribution rate of each mode: Select the first K modes with cumulative contribution rate ≥ 0.99 to form the EOF mode matrix EOFs = [v1, v2, …, v K ], and calculate the principal components

[0024] Preferably, S4, construct a time window sample from the principal component time series, set the time step t, and input the principal components PC (dimension K x t) of the continuous t time steps for each time window, and output the principal components PC (K x 1) of the t+1 time step, where t is the time step.

[0025] S41, preprocess the principal component matrix PC, and normalize each principal component: where PC (i) represents the i-th principal component, μ i and σ i represent the mean and standard deviation of the component over the entire time series, respectively;

[0026] S42, convert the one-dimensional principal component time series into a sample set with time sequence relationship using a sliding window, set the time step parameter t, and for a time series of length T, generate T-t training samples through the sliding window, the input part of each sample is the principal component data of continuous t time steps: X i = [PC i , PC i+1 , …, PCi+t-1 ] T (Shape: K x t), and the corresponding output part is the principal component value of the t+1th time step: Y i = PC i+t (Shape: K x 1);

[0027] S43, divide the generated window samples into training set, validation set and test set.

[0028] Preferably, S5, the constructed input and output are trained using an LSTM network to obtain principal component prediction values of future time steps Using reconstructing the sound speed profile to obtain the target year sound speed profile prediction value, wherein The step of predicting the principal component value specifically includes:

[0029] S51, construct an LSTM network architecture suitable for principal component time series prediction, the network input is the time window sample, the shape is sample number x time step length t x feature dimension K, and the output is the principal component prediction value of the future time step, the shape is sample number x feature dimension K;

[0030] S52, using the trained LSTM model, predicting the principal component of the future time step, for the target year month t+1, the input of the principal component data of the recent t time steps is: Where f LSTM represents the trained LSTM model, is the predicted principal component value;

[0031] S53, using reconstructing the sound speed profile to obtain the target year sound speed profile prediction value.

[0032] Preferably, S6, extract the real sound speed profile value of the target year, compare the error with the predicted value, take the root mean square error RMSE and the mean absolute error MAE as the evaluation standard, and constantly iterate α, β, σ, t, output the final prediction result and the final root mean square error RMSE and the final mean absolute error MAE, Where, y i is the ith real value, is the ith predicted value, and n is the sample number.

[0033] S61, extract the real sound speed profile value of the target year, and perform data alignment processing on the real sound speed profile value;

[0034] S62, compare the error with the predicted value, take the root mean square error RMSE and the mean absolute error MAE as the evaluation standard, wherein y i is the i-th real value, is the i-th predicted value, and n is the number of samples;

[0035] S63, based on the error analysis result, the year difference weight coefficient a, the month difference weight coefficient b, the parameter s of the control weight decay range and the LSTM time step t are systematically adjusted, iterative optimization is performed, optimization convergence judgment is made and final output is performed.

[0036] To solve the above technical problems, the application also provides a fusion weighted EOF-LSTM ocean sound speed profile prediction device, which adopts the technical scheme as follows:

[0037] The extraction module is used for extracting the sound speed profile data of a preset spatial position m years in the Argo database, constructing a C=N*M historical sound speed profile data matrix, interpolating the sound speed profile data to realize the sound speed value every 1 meter interval, wherein C is the sound speed profile data matrix, N is the profile quantity (time series length), and M is the depth layer number (point number after interpolation) of each profile;

[0038] The construction module is used for constructing a Gaussian weight function based on the year difference Y dist and the month difference M dist of the target year and month, obtaining a comprehensive time distance T dist = a*Y dist + b*M dist , and constructing the Gaussian weight function wherein s controls the weight decay range, a=1, b=1 and s=8 are initialized, a is the year difference weight coefficient, b is the month difference weight coefficient, s is the parameter of the control weight decay range, and w is the Gaussian weight value;

[0039] The processing module is used for normalizing the Gaussian weight function, calculating the weighted average sound speed profile C, extracting the spatial variation mode of the sound speed profile according to RV=DV by using the covariance matrix R, realizing dimension reduction and feature extraction, wherein D is the eigenvalue matrix, V is the eigenvector matrix, the eigenvalues are arranged in descending order, the first K modes with a cumulative contribution rate greater than or equal to 0.99 are selected, an EOF mode matrix EOFs is obtained, and the principal components are calculated. wherein EOFs is a mode matrix composed of the first K eigenvectors, is the weighted average sound speed profile, PC is the principal component matrix, and K is the number of selected modes, that is, the principal component dimension;

[0040] The sample module is used for constructing a time window sample of the principal component time series, setting a time step t, inputting the principal component PC (dimension K*t) of each time window for t consecutive time steps, and outputting the principal component PC (K*1) of the t+1 time step, wherein t is the time step.

[0041] A prediction module is configured to train the constructed input and output by using the LSTM network to obtain the principal component prediction value of the future time step By using Reconstructing the sound velocity profile to obtain the sound velocity profile prediction value of the target year and month, wherein, The predicted principal component is;

[0042] An output module is configured to extract the real sound velocity profile value of the target year, compare the prediction value with the error, take the root mean square error (RMSE) and the mean absolute error (MAE) as the evaluation criteria, and iteratively output the final prediction result and the final root mean square error (RMSE) and the final mean absolute error (MAE), Wherein, y i The i-th real value is, The i-th prediction value is, and n is the sample number.

[0043] In order to solve the above technical problems, the present application also provides a computer device, which adopts the technical scheme as follows: a memory and a processor, the memory stores computer readable instructions, and the processor executes the computer readable instructions to realize the steps of the above-mentioned EOF-LSTM ocean sound velocity profile prediction method based on fusion weighting.

[0044] In order to solve the above technical problems, the present application also provides a computer readable storage medium, which adopts the technical scheme as follows: the computer readable storage medium stores computer readable instructions, and the computer readable instructions are executed by the processor to realize the steps of the above-mentioned EOF-LSTM ocean sound velocity profile prediction method based on fusion weighting.

[0045] Compared with the prior art, the present application has the following beneficial effects: based on the traditional EOF, the weight function of the average sound velocity profile is weighted, the optimal time step of the LSTM input data is found, the weight parameters are iteratively optimized based on the error as the benchmark, and the prediction error is effectively reduced. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the schemes in the present application, the following will briefly introduce the drawings needed in the description of the embodiments of the present application. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creating any creative labor.

[0047] Figure 1 is a flowchart of an embodiment of the present application based on the EOF-LSTM ocean sound velocity profile prediction method based on fusion weighting;

[0048] Figure 2 is a comparison effect diagram of the method of the present application and the traditional EOF, LSTM, and EOF-LSTM methods with real data prediction.

[0049] Figure 3 is an error distribution and error comparison diagram of the method of the present application and the traditional EOF, LSTM, and EOF-LSTM methods.

[0050] Figure 4 is a comparison effect diagram of the method of the present application and the traditional EOF, LSTM, and EOF-LSTM methods with real data prediction.

[0051] Figure 5 is an error distribution and error comparison diagram of the method of the present application and the traditional EOF, LSTM, and EOF-LSTM methods.

[0052] Figure 6 is a structural schematic diagram of an embodiment of the fusion weighted EOF-LSTM marine sound speed profile prediction device of the present application;

[0053] Figure 7 is a structural schematic diagram of an embodiment of the computer device of the present application. DETAILED DESCRIPTION

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs; the terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application; the description and claims of the application and the above-mentioned drawing figures are not to be construed as including any non- patentable material unless and except when intentionally included in the following description. The terms "comprising," "having," "including," and "containing" used herein are meant to be open-ended and non-limiting.

[0055] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the application. The appearances of the phrase "in an embodiment" in various places in the specification are not necessarily all referring to the same embodiment, nor are they necessarily all directed to the same embodiment, or to a single alternative embodiment.

[0056] In order for those skilled in the art to better understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings.

[0057] It should be noted that the fusion weighted EOF-LSTM based marine sound speed profile prediction method provided in the embodiments of the present application is generally executed by a server / terminal device, and accordingly, the fusion weighted EOF-LSTM based marine sound speed profile prediction apparatus is generally arranged in the server / terminal device.

[0058] It should be understood that the number of terminal devices, networks and servers is only illustrative. Any number of terminal devices, networks and servers can be provided according to the implementation needs.

[0059] Example One

[0060] Please refer to Figure 1 , a flowchart of one embodiment of the fusion weighted EOF-LSTM based marine sound speed profile prediction method of the present application is shown. The fusion weighted EOF-LSTM based marine sound speed profile prediction method comprises the following steps:

[0061] Step S1, extracting the sound speed profile data of the preset spatial position for m years in the Argo database, constructing

[0062] C=N*M historical sound speed profile data matrix, interpolating the sound speed profile data to realize the sound speed value every 1 meter interval, wherein C is the sound speed profile data matrix, N is the profile quantity (time series length), and M is the depth layer number (point number after interpolation) of each profile.

[0063] In the present embodiment, the electronic device (such as a server / terminal device) on which the fusion weighted EOF-LSTM based marine sound speed profile prediction method runs can receive the fusion weighted EOF-LSTM based marine sound speed profile prediction request through a wired connection mode or a wireless connection mode. It should be noted that the wireless connection mode can include but is not limited to 3G / 4G / 5G connection, WiFi connection, Bluetooth connection, WiMAXX connection, Zigbee connection, UWB (ultra wideband) connection, and other now known or future developed wireless connection modes.

[0064] In the present embodiment, step S1 can further comprise the following steps:

[0065] S11, extracting the sound speed profile data of the preset spatial position for m years from the global ocean observation system Argo database.

[0066] The Argo database is a global ocean observation system that continuously collects ocean environmental data through buoys placed in the global ocean. These data include seawater temperature, salinity, depth and other parameters, which are of great value for studying the marine environment and predicting the sound speed profile.

[0067] When extracting data, SQL query language can be used to filter conditions and set specific spatial locations such as m years and time ranges to obtain the required data set from the massive database. These data not only contain detailed information of sound velocity profile, but also provide rich historical data for subsequent model training.

[0068] S12, clean the extracted sound velocity profile data, and construct a C=N×M historical sound velocity profile data matrix, where C is the sound velocity profile data matrix, N is the profile number (time series length), and M is the depth layer number (interpolated point number) of each profile.

[0069] The extracted sound velocity profile data is checked for completeness, and records with missing values or incomplete key parameter records are removed. Consistency check is performed to compare data from different sources or collected at different times, identify and correct logically inconsistent data points. Furthermore, statistical methods such as calculating mean and standard deviation are used to identify and remove outliers far from the normal range.

[0070] The cleaned sound velocity profile data is sorted by time series and depth to ensure that the data of each profile is complete and ordered.

[0071] For example, you can use the NumPy library in Python to convert the sorted sound velocity profile data into a two-dimensional array form, and construct a C=N×M historical sound velocity profile data matrix, where C is the sound velocity profile data matrix, N is the profile number (time series length), and M is the depth layer number (interpolated point number) of each profile.

[0072] By constructing a data matrix, the originally scattered data points are integrated into a structured data set, which facilitates subsequent data analysis and model training. At the same time, this data organization form is also more conducive to feature extraction and prediction using EOF and LSTM algorithms.

[0073] S13, the sound velocity profile data is interpolated to achieve a sound velocity value every 1 meter.

[0074] Due to the actual observation, the sound velocity values at different depths may not be evenly distributed. In order to facilitate subsequent analysis and processing, the sound velocity profile data is interpolated to achieve a sound velocity value every 1 meter.

[0075] For example, linear interpolation algorithm is used to process the sound velocity profile data. This algorithm is simple and efficient, and can well preserve the characteristics of the original data.

[0076] For example, use the interpolation function in the SciPy library in Python to interpolate the sound velocity data of each profile, ensuring that each profile has corresponding sound velocity values at different depths.

[0077] The data interpolation processing not only makes the data points of each profile more dense and uniform, but also improves the consistency and comparability of the data. This is very beneficial for subsequent spatial feature extraction using EOF and time series prediction using LSTM.

[0078] Step S2, based on the year difference Y dist and month difference M dist of the target year and month, obtain the comprehensive time distance T dist = a·Y dist + b·M dist , construct the Gaussian weight function where σ controls the weight decay range, and the initial values are a = 1, b = 1, and σ = 8, where a is the year difference weight coefficient, b is the month difference weight coefficient, σ is the parameter controlling the weight decay range, and ω is the Gaussian weight value.

[0079] In this embodiment, step S2 can further include the following steps:

[0080] S21, extract the year and month timestamp T hist corresponding to each historical sound velocity profile data and the target prediction year and month T target , respectively calculate the year difference Y dist and month difference M dist : Y dist = |Y target -Y hist |, M dist = |M target -M hist |.

[0081] The core of this step is to determine the difference in time and month between the target data point and the reference data point. For a given target year and month and a reference year and month, the year difference is the difference between the two year values, and the month difference is the difference between the two month values. For example, if the target year and month is August 2025 and the reference year and month is May 2023, then the year difference Y dist = 2025 - 2023 = 2, and the month difference M dist = 8 - 5 = 3.

[0082] The calculation of the year difference and the month difference is the basis for subsequent construction of the comprehensive time distance. The year difference reflects the changes of the data in the long time scale. Different years may have differences in the ocean sound velocity profile due to periodic or trend changes in factors such as climate and marine environment. The month difference reflects the changes of the data in the seasonal scale. The marine environment is affected by factors such as solar radiation and ocean current in different months, and has obvious seasonal characteristics, which will be directly reflected in the ocean sound velocity profile. By accurately calculating the year difference and the month difference, the deviation of the target data and the reference data in time and season can be quantified, providing accurate data input for the construction of the comprehensive time distance.

[0083] S22, calculating the comprehensive time distance T dist = a · Y dist + b · M dist and calculating the Gaussian weight corresponding to each historical data by using the Gaussian kernel function Initialize a = 1, b = 1, and s = 8, where a is the year difference weight coefficient, b is the month difference weight coefficient, s is the parameter for controlling the weight decay range, and w is the Gaussian weight value.

[0084] The formula fuses the year difference and the month difference into a comprehensive index by linear weighting, which is used to measure the distance between the target data and the reference data in the time and seasonal comprehensive dimensions.

[0085] The larger the s is, the slower the weight decays with the distance; otherwise, the faster the weight decays.

[0086] By linearly fusing the year and month differences, a comprehensive time distance index is constructed, which overcomes the limitations of a single time scale.

[0087] The Gaussian function is used to map the distance into a weight value, so that the data points closer to the target time have higher weights, which preserves the time locality feature and provides a basis for subsequent weighted EOF analysis.

[0088] In step S3, the Gaussian weight function is normalized to calculate the weighted average sound velocity profile C, and the spatial variation mode of the sound velocity profile is extracted by using the covariance matrix R according to RV = DV, which realizes dimension reduction and feature extraction. Wherein, D is the eigenvalue matrix, V is the eigenvector matrix, the eigenvalues are arranged in descending order, the first K modes with cumulative contribution rate ≥0.99 are selected, the EOF mode matrix EOFs is obtained, and the principal component PC is calculated. Wherein, EOFs is the mode matrix composed of the first K eigenvectors, is the weighted average sound velocity profile, PC is the principal component matrix, and K is the number of selected modes, i.e. the principal component dimension.

[0089] In this embodiment, step S3 can further include the following steps:

[0090] S31, normalize the Gaussian weights so that their sum is 1, with the normalization formula: where ω i is the original weight of the i-th sample, ω i ′ is the normalized weight, and N is the total number of samples.

[0091] To avoid the bias caused by the inconsistent distribution of weights among different samples, all weights are normalized so that their sum is 1.

[0092] To ensure that the weights of all historical data form a probability distribution, avoid the model bias caused by different absolute weight sizes.

[0093] To improve the stability and interpretability of the weighted EOF decomposition, provide more normalized input features for the LSTM network.

[0094] S32, calculate the weighted average sound speed profile using the normalized weights C i represents the i-th sound speed profile vector.

[0095] In implementation, you can use Python's NumPy library for efficient matrix and vector operations.

[0096] Weighted average profile Reflects the background sound speed field under the influence of time weights, represents the baseline state of the sound speed profile, and provides a reference for subsequent anomaly calculation and feature extraction.

[0097] S33, calculate the covariance matrix R, where C is the original sound speed profile data matrix, is the weighted average sound speed profile, and W is a diagonal matrix with normalized weights ω i ′ as diagonal elements, solve the eigenvalue equation RV = DV, where D is the eigenvalue matrix and V is the eigenvector matrix.

[0098] In implementation, you can use the eigenvalue decomposition function in numerical computing libraries (such as SciPy or NumPy) to efficiently implement this process.

[0099] The weighted covariance matrix captures the spatial variation characteristics of the sound speed profile, reflecting the covariant relationship between different depth layers.

[0100] Eigen decomposition converts data to a new coordinate system, with the eigenvector representing the spatial variation pattern of the sound speed profile (EOF mode), and the eigenvalue reflecting the variance size of each mode explanation.

[0101] S34, sort the eigenvalues from large to small: λ1≥ λ2≥ … ≥ λ M And calculate the variance contribution rate and cumulative contribution rate of each mode: Select the first K modes with cumulative contribution rate ≥ 0.99 to form the EOF mode matrix EOFs = [v1, v2, …, v K ] and calculate the principal components

[0102] By selecting the main mode, the data dimensionality is reduced, the calculation complexity is reduced, and the most important change information in the data is retained.

[0103] The principal component matrix PC represents the projection of the original sound speed profile data on the main EOF mode, converts high-dimensional spatial data into low-dimensional time series, and provides effective input features for subsequent LSTM network processing.

[0104] Step S4, construct a time window sample from the principal component time series, set the time step t, and input the principal component PC (dimension K x t) of each time window for t consecutive time steps. The output is the principal component PC (K x 1) of the t+1 time step, where t is the time step.

[0105] In this embodiment, step S4 can further include the following steps:

[0106] S41, pre-process the principal component matrix PC, and normalize each principal component: Where PC (i) represents the i-th principal component, μ i and σ i represent the mean and standard deviation of the component over the entire time series, respectively.

[0107] Because the numerical range of different principal components may differ greatly, in order to avoid unstable gradients during training or some features dominating model training, each principal component needs to be normalized.

[0108] In specific implementation, the StandardScaler class in the Scikit-learn library of Python can be used, or direct vectorization calculation using NumPy can be used to ensure efficient and consistent processing.

[0109] Through step S41, the dimensional differences of different principal components can be eliminated, avoiding the excessive influence of components with large numerical ranges on model training; the stability and convergence speed of LSTM training can be improved, providing numerical stability guarantee for subsequent gradient descent optimization.

[0110] S42, convert the one-dimensional principal component time series into a sample set with time sequence relationship by using a sliding window, set the time step parameter t, for a time series with length T, generate T-t training samples through the sliding window, the input part of each sample is the principal component data of the continuous t time steps: X i = [PC i , PC i+1 ,..., PC i+t-1 ] T (shape: Kxt), and the corresponding output part is the principal component value of the t+1 time step: Y i = PC i+t (shape: Kxl).

[0111] In implementation, NumPy array operations and slicing functions of Python can be used to generate all window samples through loop or vectorization.

[0112] By using step S42, the time series data is converted into a supervised learning problem, the input-output correspondence relationship is clear, the time-dependent pattern and short-term evolution law of the sound speed profile change can be captured, and the input data format with clear time structure is provided for the LSTM network.

[0113] S43, divide the generated window samples into training set, validation set and test set.

[0114] Considering the time sequence characteristics of time series data, forward validation segmentation method can be used to divide the data set in time sequence to avoid future information leakage.

[0115] According to the proportion of 70%-15%-15%, the training set: the first 70% of the samples are used for model parameter training; the validation set: the middle 15% of the samples are used for hyperparameter tuning and early stopping; the test set: the last 15% of the samples are used for final model evaluation.

[0116] At the same time, the data format is adjusted to the input format required by the LSTM network:

[0117] Input shape: (number of samples, time step t, feature dimension K);

[0118] Output shape: (number of samples, feature dimension K).

[0119] For example, Dataset and DataLoader classes in Keras or PyTorch framework of Python can be used to realize batch data loading and memory management.

[0120] By employing step S43, the fairness and reliability of model evaluation can be ensured, overfitting can be avoided, performance monitoring and early stopping mechanisms can be provided during model training, memory usage can be optimized and training efficiency can be improved, and appropriate data support can be provided for different model stages.

[0121] It should be noted that the choice of time step t needs to be adjusted according to the time variation characteristics of the sound velocity profile: smaller t values ​​(such as 3-6) are suitable for capturing short-term changes; larger t values ​​(such as 12-24) are suitable for capturing seasonal changes; the optimal time step can be determined through autocorrelation function analysis.

[0122] In practice, grid search or Bayesian optimization methods can be used to find the optimal t value to balance model complexity and prediction performance.

[0123] Step S5: Use the constructed input and output to train an LSTM network to obtain the principal component predictions for future time steps. use Reconstructing the sound velocity profile yields predicted sound velocity profile values ​​for the target year and month, where... These are the predicted principal components.

[0124] In this embodiment, step S5 may further include the following steps:

[0125] S51, construct an LSTM network architecture suitable for principal component time series prediction. The network input is the time window samples, with a shape of sample number × time step t × feature dimension K. The output is the principal component prediction value for the future time step, with a shape of sample number × feature dimension K.

[0126] The LSTM network architecture includes:

[0127] Input layer: Receives time-series data of shape (t, K);

[0128] One or more LSTM layers: Each layer contains several LSTM units to capture dependencies at different time scales;

[0129] Dropout layer: Prevents overfitting and improves the model's generalization ability;

[0130] Fully connected layer: maps the LSTM output to a K-dimensional principal component space.

[0131] The optimization algorithm can be either Adam or RMSprop, and the learning rate is adjusted based on the performance on the validation set. Early stopping is used during training to prevent overfitting, preserving the model parameters that best perform on the validation set.

[0132] By adopting step S51, the complex nonlinear evolution law of the principal component time series is learned, the long-term dependence relationship and short-term fluctuation characteristics of the sound speed profile change are captured, and accurate principal component prediction values are provided for sound speed profile prediction.

[0133] S52, using the trained LSTM model, the principal component of the future time step is predicted, and for the target year and month t+1, the principal component data of the recent t time steps is input as: Where f LSTM represents the trained LSTM model, is the predicted principal component value.

[0134] To further quantify the prediction uncertainty, Monte Carlo Dropout can be used: enable Dropout multiple times during prediction to obtain the mean and variance of the prediction distribution; Bootstrap aggregation: train multiple LSTM models to reduce the prediction variance through ensemble learning; Bayesian neural network: directly model the weight uncertainty to provide probability prediction.

[0135] By adopting step S52, accurate prediction values of the principal component of the future time step can be provided, the uncertainty of the prediction result can be quantified, the reliability evaluation for decision-making can be provided, and the stability and accuracy of the prediction can be improved through the integration method.

[0136] S53, using reconstruct the sound speed profile to obtain the predicted value of the target year and month sound speed profile.

[0137] After reconstruction, appropriate post-processing optimization can also be performed: physical constraint application: clip the predicted value according to the physical range of sound speed (usually 1450-1550 m / s); smoothing processing: use Savitzky-Golay filter or moving average to smooth the profile and eliminate unreasonable fluctuations; error correction: establish a correction model based on historical prediction errors to improve prediction accuracy.

[0138] By adopting step S53, the low-dimensional principal component prediction can be reconstructed into a complete sound speed profile, the prediction result can be ensured to meet the physical reality and smoothness requirements, and the accuracy and practicality of the final prediction result can be improved.

[0139] Step S6, extract the real sound speed profile value of the target year, compare the error with the predicted value, take the root mean square error RMSE and the mean absolute error MAE as the evaluation standard, and constantly iterate α, β, σ, t, output the final prediction result and the final root mean square error RMSE and the final mean absolute error MAE, Where y i is the i-th real value, is the i-th predicted value, and n is the sample size.

[0140] In the present embodiment, step S6 can further include the steps of:

[0141] S61, extracting the real sound speed profile values of the target year, and performing data alignment processing on the real sound speed profile values.

[0142] Data alignment includes: spatiotemporal matching: using latitude and longitude grid matching and target timestamp alignment to ensure that each prediction point has a corresponding real observation value; depth layer alignment: using the same interpolation method as in the preprocessing stage (such as linear interpolation) to interpolate the real data to standard depth layers; quality control: removing abnormal observation values and missing data to ensure the integrity and reliability of the evaluation data set, etc.

[0143] In specific implementation, Pandas and NumPy libraries of Python can be used for data operation, and GeoPandas can be used for spatial data matching.

[0144] By using step S61, accurate and reliable benchmark data can be provided for model evaluation, and the comparability of prediction values and real values in the spatiotemporal and depth dimensions can be ensured, and a fair and objective model performance evaluation basis can be established.

[0145] S62, error comparison with the prediction values, taking root mean square error RMSE and mean absolute error MAE as evaluation criteria, where y i is the i-th real value, is the i-th prediction value, and n is the sample size.

[0146] Other auxiliary evaluation indicators can also be calculated, such as the coefficient of determination (R 2 ): measuring the correlation between prediction values and real values; average deviation (Bias): evaluating the size of systematic error; skill score (Skill Score): improvement degree relative to the baseline model

[0147] The metrics module in machine learning libraries such as Scikit-learn can be used to ensure the standardization and efficiency of the calculation.

[0148] By using step S62, the prediction error can be quantified, the model performance can be objectively evaluated, and the prediction ability difference of the model at different depth layers can be identified, so as to provide a clear optimization target and direction for parameter optimization.

[0149] S63, based on the error analysis results, systematically adjusting the year difference weight coefficient a, the month difference weight coefficient b, the parameter s controlling the weight decay range, and the LSTM time step t, performing iterative optimization, making optimization convergence judgment, and performing final output.

[0150] The following optimization strategies can be used: grid search: systematically traversing various parameter combinations in the parameter space; random search: random sampling in the parameter space, suitable for high-dimensional parameter optimization; Bayesian optimization: based on Gaussian process and other probabilistic models, efficient search for optimal parameters; cross-validation: use k-fold cross-validation to ensure the robustness of the optimization results.

[0151] The objective function of the optimization process is to minimize the RMSE or weighted loss function on the validation set: where w1 and w2 are weight coefficients, which can be adjusted according to application requirements.

[0152] In specific implementation, optimization convergence criteria can be established to ensure the effectiveness and efficiency of the optimization process. For example, set the convergence condition: relative improvement is less than the threshold: maximum number of iterations limit and computing resource limit, etc.

[0153] The final output is: the optimal parameter combination: alpha * , beta * , sigma * , t * , and the final prediction result: Performance indicators: final RMSE and MAE values; optimization process record: parameter and error changes at each iteration

[0154] By using step S63, the optimization process can be terminated at the appropriate time, avoiding overfitting and waste of computing resources, providing complete optimization records for result analysis, and outputting the final prediction model that can be used for business applications.

[0155] Figure 2 is the comparison effect diagram of the method of the present application and the traditional EOF, LSTM, and EOF-LSTM methods with the real data prediction. The prediction results of the traditional EOF, LSTM, EOF-LSTM, and the fusion weighted EOF-LSTM method proposed by the present application are compared with the real data, and the prediction results are highly consistent with the measured data in terms of spatial structure and time evolution characteristics.

[0156] Figure 3 is the error distribution and error comparison diagram of the method of the present application and the traditional EOF, LSTM, and EOF-LSTM methods. The error analysis is performed using the prediction results and the traditional methods, including the prediction error distribution characteristics and the root mean square error (RMSE) and mean absolute error (MAE) quantitative indicators of the four methods (EOF, LSTM, EOF-LSTM, and weighted EOF-LSTM) at each depth. From the quantitative indicators, the present application achieves high-precision prediction results.

[0157] Figure 4is a test chart of the method of the present application and the traditional EOF, LSTM, EOF-LSTM method and the real data prediction comparison effect diagram.

[0158] Figure 5 is a test chart of the method of the present application and the traditional EOF, LSTM, EOF-LSTM method prediction error distribution and error comparison chart.

[0159] The embodiment is implemented, and beneficial effects are that: the weight function is constructed based on the traditional EOF and weighted on the average sound velocity profile, the optimal time step is found for the LSTM input data, the weight parameter is iteratively optimized taking the error as the benchmark, and the prediction error is effectively reduced.

[0160] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in a distributed computing environment, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0161] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by computer-readable instructions instructing related hardware, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments. The storage medium can be a non-volatile storage medium such as a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0162] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0163] Example Two

[0164] Further reference Figure 6 As a response to the above Figure 1 The present invention provides an embodiment of a fusion-weighted EOF-LSTM ocean acoustic profile prediction device, which, in accordance with the implementation of the method shown, provides an embodiment of the device. Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0165] like Figure 6 As shown, the ocean acoustic velocity profile prediction device 40 based on fusion weighted EOF-LSTM described in this embodiment includes: an extraction module 71, a construction module 72, a processing module 73, a sample module 74, a prediction module 75, and an output module 76. Wherein:

[0166] Extraction module 71 is used to extract sound velocity profile data for a preset spatial location m years from the Argo database, construct a historical sound velocity profile data matrix of C = N × M, and interpolate the sound velocity profile data to achieve sound velocity values ​​at intervals of 1 meter. Here, C is the sound velocity profile data matrix, N is the number of profiles (time series length), and M is the depth layer of each profile (number of points after interpolation).

[0167] Module 72 is used to construct modules based on the year difference Y with respect to the target year and month. dist and the difference in months M dist The overall time distance T is obtained. dist =α·Y dist +β·M dist Construct Gaussian weight function σ controls the weight decay range, initialized with α = 1, β = 1, σ = 8, where α is the year difference weight coefficient, β is the month difference weight coefficient, σ is the parameter controlling the weight decay range, and ω is the Gaussian weight value.

[0168] Processing module 73 is used to normalize the Gaussian weighting function and calculate the weighted average sound velocity profile. Using the covariance matrix R, and based on RV = DV, the spatial variation modes of the sound velocity profile are extracted to achieve dimensionality reduction and feature extraction. Here, D is the eigenvalue matrix, and V is the eigenvector matrix. The eigenvalues ​​are arranged in descending order, and the top K modes with a cumulative contribution rate ≥ 0.99 are selected to obtain the EOF mode matrix EOFs. Principal components are then calculated. Wherein, EOFs is the modality matrix composed of the first K eigenvectors. For the weighted average sound velocity profile, PC is the principal component matrix, and K is the number of selected modes, i.e., the principal component dimension.

[0169] Sample module 74 is used to construct time window samples from principal component time series. The time step is set to t. The input for each time window is the principal component PC (with dimension K×t) for t consecutive time steps, and the output is the principal component PC (K×1) for the (t+1)th time step, where t is the time step.

[0170] Prediction module 75 is used to train an LSTM network using the constructed inputs and outputs to obtain the principal component predictions for future time steps. use Reconstructing the sound velocity profile yields predicted sound velocity profile values ​​for the target year and month, where... The predicted principal components;

[0171] Output module 76 is used to extract the true sound velocity profile values ​​for the target year, compare them with the predicted values, and use the root mean square error (RMSE) and mean absolute error (MAE) as evaluation criteria. It iterates through α, β, σ, and t to output the final prediction result, the final RMSE, and the final MAE. Among them, y i For the i-th true value, Let be the i-th predicted value, and n be the number of samples.

[0172] The beneficial effects of implementing this embodiment are: by constructing a weighting function based on the traditional EOF average sound velocity profile, finding the optimal time step for the LSTM input data, and iteratively optimizing the weight parameters with the error as a benchmark, the prediction error is effectively reduced.

[0173] Example Three

[0174] To address the aforementioned technical problems, embodiments of the present invention also provide a computer device. Please refer to [link / reference needed]. Figure 7 , Figure 7 This is a basic structural block diagram of the computer device in this embodiment.

[0175] The computer device 8 includes a memory 81, a processor 82, and a network interface 83, which are communicatively connected via a system bus. It should be noted that the computer device 8 is only shown with the components of the memory 81, the processor 82, and the network interface 83, but it should be understood that not all of the illustrated components are required to be implemented, and more or fewer components can be alternatively implemented. As understood by one skilled in the art, the computer device herein refers to a device capable of automatically performing numerical calculation and / or information processing according to pre-set or stored instructions, and the hardware thereof includes, but is not limited to, a microprocessor, an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), an embedded device, and the like.

[0176] The computer device can be a desktop computer, a notebook computer, a palm computer, a cloud server, or the like. The computer device can interact with a user through a keyboard, a mouse, a remote controller, a touchpad, a voice control device, or the like.

[0177] The memory 81 includes at least one type of readable storage medium, such as a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, or the like), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disk, or the like. In some embodiments, the memory 81 can be an internal storage unit of the computer device 8, such as a hard disk or a memory of the computer device 8. In other embodiments, the memory 81 can also be an external storage device of the computer device 8, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, or the like. Of course, the memory 81 can also include both the internal storage unit and the external storage device of the computer device 8. In the present embodiment, the memory 81 is generally used to store an operating system and various application software installed in the computer device 8, such as computer readable instructions based on the fused weighted EOF-LSTM ocean sound speed profile prediction method, and the like. In addition, the memory 81 can also be used to temporarily store various data that have been output or will be output.

[0178] The processor 82 can be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip in some embodiments. The processor 82 is generally used to control the overall operation of the computer device 8. In the present embodiment, the processor 82 is configured to execute computer readable instructions stored in the memory 81 or process data, such as computer readable instructions of the fusion weighted EOF-LSTM ocean sound speed profile prediction method.

[0179] The network interface 83 can include a wireless network interface or a wired network interface, and is generally used to establish a communication connection between the computer device 8 and other electronic devices.

[0180] The present embodiment has the beneficial effect that the average sound speed profile is weighted based on the traditional EOF to construct a weight function, the optimal time step is found for the LSTM input data, and the weight parameters are iteratively optimized based on the error as a reference to effectively reduce the prediction error.

[0181] Example Four

[0182] The present application also provides another embodiment, that is, a computer readable storage medium storing computer readable instructions, the computer readable instructions being executable by at least one processor to cause the at least one processor to perform the steps of the fusion weighted

[0183] EOF-LSTM ocean sound speed profile prediction method as described above.

[0184] The present embodiment has the beneficial effect that the average sound speed profile is weighted based on the traditional EOF to construct a weight function, the optimal time step is found for the LSTM input data, and the weight parameters are iteratively optimized based on the error as a reference to effectively reduce the prediction error.

[0185] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by means of software and the necessary general hardware platform, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for causing a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the various embodiment methods of the present application.

[0186] Obviously, the above-described embodiments are only some embodiments but not all the embodiments of the present application, the preferred embodiments of the present application are shown in the drawings, but do not limit the patent scope of the present application. The present application can be implemented in many different forms, and conversely, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions recorded in the foregoing specific embodiments, or make equivalent replacements to some technical features therein. Any equivalent structure made by using the content of the present application specification and drawings, directly or indirectly applied to other related technical fields, is also within the patent protection scope of the present application.

Claims

1. A method for predicting ocean acoustic velocity profiles based on fused weighted EOF-LSTM, characterized in that, Includes the following steps: S1. Extract sound velocity profile data for a preset spatial location over a year from the Argo database, construct a historical sound velocity profile data matrix of C = N × M, and interpolate the sound velocity profile data to achieve sound velocity values ​​at intervals of 1 meter. Here, C is the sound velocity profile data matrix, N is the number of profiles, and M is the depth layer of each profile. S2, based on the year difference Y from the target year and month. dist and the difference in months M dist The overall time distance T is obtained. dist =α·Y dist +β·M dist Construct Gaussian weight function σ controls the weight decay range, initialized with α = 1, β = 1, σ = 8, where α is the year difference weight coefficient, β is the month difference weight coefficient, σ is the parameter controlling the weight decay range, and ω is the Gaussian weight value. S3. The Gaussian weighting function is normalized, and the weighted average sound velocity profile _C_ is calculated. Using the covariance matrix R, according to RV = DV, the spatial variation modes of the sound velocity profile are extracted to achieve dimensionality reduction and feature extraction. Here, D is the eigenvalue matrix and V is the eigenvector matrix. The eigenvalues ​​are arranged in descending order, and the top K modes with a cumulative contribution rate ≥ 0.99 are selected to obtain the EOF mode matrix EOFs. The principal components are then calculated. Wherein, EOFs is the modality matrix composed of the first K eigenvectors. For the weighted average sound velocity profile, PC is the principal component matrix, and K is the number of selected modes, i.e., the principal component dimension. S4, construct time window samples from the principal component time series, set the time step t, input the principal components PC for t consecutive time steps for each time window, and output the principal components PC for the (t+1)th time step, where t is the time step; S5. The constructed input and output are used to train an LSTM network to obtain the principal component predictions for future time steps. use Reconstructing the sound velocity profile yields predicted sound velocity profile values ​​for the target year and month, where... The predicted principal components; S6 extracts the actual sound velocity profile values ​​for the target year and compares them with the predicted values. Using the root mean square error (RMSE) and mean absolute error (MAE) as evaluation criteria, iterates through α, β, σ, and t to output the final prediction result, the final RMSE, and the final MAE. Among them, y i For the i-th true value, Let be the i-th predicted value, and n be the number of samples.

2. The ocean acoustic velocity profile prediction method based on fusion weighted EOF-LSTM according to claim 1, characterized in that, The steps in S1, which involve extracting sound velocity profile data for a preset spatial location over a year from the Argo database, constructing a historical sound velocity profile data matrix of C = N × M, and interpolating the sound velocity profile data to achieve sound velocity values ​​at 1-meter intervals, where C is the sound velocity profile data matrix, N is the number of profiles, and M is the depth layer of each profile, specifically include: S11, extract sound velocity profile data for a predetermined spatial location over a year from the Argo global ocean observation system database; S12, clean the extracted sound velocity profile data and construct a C = N × M historical sound velocity profile data matrix, where C is the sound velocity profile data matrix, N is the number of profiles, and M is the depth layer of each profile. S13 interpolated the sound velocity profile data to achieve a sound velocity value every 1 meter.

3. The ocean acoustic velocity profile prediction method based on fusion weighted EOF-LSTM according to claim 1, characterized in that, S2 is based on the year difference Y with the target year and month. dist and the difference in months M dist The overall time distance T is obtained. dist =α·Y dist +β·M dist Construct Gaussian weight function The steps for controlling the weight decay range, where σ controls the weight decay range, and initialization is α = 1, β = 1, σ = 8, where α is the year difference weight coefficient, β is the month difference weight coefficient, σ is the parameter controlling the weight decay range, and ω is the Gaussian weight value, specifically include: S21, Extract the year and month timestamp T corresponding to each historical sound velocity profile data. hist And the target prediction year and month T target Calculate the year difference Y respectively dist Difference between months M dist Y dist =|Y target -Y hist |,M dist =|M target -M hist |, S22, Calculate the overall time distance T dist =α·Y dist +β·M dist And use the Gaussian kernel function to calculate the Gaussian weight corresponding to each historical data point. Initialize α = 1, β = 1, σ = 8, where α is the year difference weight coefficient, β is the month difference weight coefficient, σ is the parameter that controls the weight decay range, and ω is the Gaussian weight value.

4. The ocean acoustic velocity profile prediction method based on fusion weighted EOF-LSTM according to claim 3, characterized in that, In step S3, the Gaussian weighting function is normalized, and the weighted average sound velocity profile is calculated. Using the covariance matrix R, and based on RV = DV, the spatial variation modes of the sound velocity profile are extracted to achieve dimensionality reduction and feature extraction. Here, D is the eigenvalue matrix, and V is the eigenvector matrix. The eigenvalues ​​are arranged in descending order, and the top K modes with a cumulative contribution rate ≥ 0.99 are selected to obtain the EOF mode matrix EOFs. Principal components are then calculated. Wherein, EOFs is the modality matrix composed of the first K eigenvectors. For the weighted average sound velocity profile, PC is the principal component matrix, and K is the number of selected modes, i.e., the steps for determining the principal component dimensions specifically include: S31, normalize the Gaussian weights so that their sum is 1. The normalization formula is: Where ω i ω is the original weight of the i-th sample. i ′ is the normalized weight, and N is the total number of samples; S32, calculate the weighted average sound velocity profile using the normalized weights. C i Represents the i-th sound speed profile vector; S33, Calculate the covariance matrix R. Where C is the original sound velocity profile data matrix. For the weighted average sound speed profile, W is the normalized weight ω i Let ' be a diagonal matrix with diagonal elements. Solve the characteristic equation RV = DV, where D is the eigenvalue matrix and V is the eigenvector matrix. S34, sort the eigenvalues ​​in descending order: λ1≥λ2≥…≥λ M And calculate the variance contribution rate and cumulative contribution rate for each mode: Select the top K modes that result in a cumulative contribution rate ≥ 0.99 to form the EOF mode matrix EOFs = [v1, v2, ..., v K ] Calculate principal components 5. The ocean acoustic velocity profile prediction method based on fusion weighted EOF-LSTM according to claim 4, characterized in that, S4 involves constructing time window samples from the principal component time series, setting a time step t, and inputting the principal components PC (with dimensions K×t) for each time window over t consecutive time steps. The output is the principal component PC (K×1) for the (t+1)th time step. Specifically, the step where t is the time step includes: S41, preprocess the principal component matrix PC, and standardize each principal component component: PC (i) μ represents the i-th principal component. i and σ i These represent the mean and standard deviation of the component over the entire time series, respectively. S42, a sliding window is used to transform the one-dimensional principal component time series into a sample set with temporal relationships. The time step parameter t is set. For a time series of length T, Tt training samples are generated through the sliding window. The input part of each sample is the principal component data of t consecutive time steps: X i =[PC i PC i+1 ,...,PC i+t-1 ] T The corresponding output is the principal component value at time step t+1: Y i =PC i+t ; S43 divides the generated window samples into training set, validation set and test set.

6. The ocean acoustic velocity profile prediction method based on fusion weighted EOF-LSTM according to claim 5, characterized in that, In step S5, the constructed input and output are trained using an LSTM network to obtain the principal component prediction values ​​for future time steps. use Reconstructing the sound velocity profile yields predicted sound velocity profile values ​​for the target year and month, where... The specific steps for predicting principal component values ​​include: S51, Construct an LSTM network architecture suitable for principal component time series prediction. The network input is the time window samples, with a shape of sample number × time step t × feature dimension K. The output is the principal component prediction value for the future time step, with a shape of sample number × feature dimension K. S52, using a trained LSTM model, predicts the principal components for future time steps. For the target year and month t+1, the input principal component data for the most recent t time steps is: Where f LSTM This represents the trained LSTM model. These are the predicted principal component values; S53, utilizing Reconstruct the sound velocity profile to obtain the predicted sound velocity profile values ​​for the target year and month.

7. The ocean acoustic velocity profile prediction method based on fusion weighted EOF-LSTM according to claim 6, characterized in that, In step S6, the actual sound velocity profile values ​​for the target year are extracted and compared with the predicted values. The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation criteria. The steps are iterated continuously using α, β, σ, and t to output the final prediction result, the final RMSE, and the final MAE. Among them, y i For the i-th true value, The steps for predicting the i-th value, where n is the number of samples, specifically include: S61, extract the true sound velocity profile values ​​for the target year, and perform data alignment processing on the true sound velocity profile values. S62 compares the error with the predicted value, using the root mean square error (RMSE) and mean absolute error (MAE) as the evaluation criteria. Among them, y i For the i-th true value, Let n be the i-th predicted value, and n be the number of samples. S63, based on the error analysis results, systematically adjusts the year difference weight coefficient α, the month difference weight coefficient β, the parameter σ that controls the weight decay range, and the LSTM time step t, performs iterative optimization, makes optimization convergence judgment, and outputs the final result.

8. A device for predicting ocean acoustic velocity profiles based on fusion-weighted EOF-LSTM, characterized in that, include: The extraction module is used to extract sound velocity profile data for a preset spatial location m years from the Argo database, construct a historical sound velocity profile data matrix of C = N × M, and interpolate the sound velocity profile data to achieve sound velocity values ​​at intervals of 1 meter. Here, C is the sound velocity profile data matrix, N is the number of profiles (time series length), and M is the depth layer of each profile (number of points after interpolation). Build a module for using the year difference Y from the target year and month. dist and the difference in months M dist The overall time distance T is obtained. dist =α·Y dist +β·M dist Construct Gaussian weight function σ controls the weight decay range, initialized with α = 1, β = 1, σ = 8, where α is the year difference weight coefficient, β is the month difference weight coefficient, σ is the parameter controlling the weight decay range, and ω is the Gaussian weight value. The processing module is used to normalize the Gaussian weighting function and calculate the weighted average sound velocity profile. Using the covariance matrix R, and based on RV = DV, the spatial variation modes of the sound velocity profile are extracted to achieve dimensionality reduction and feature extraction. Here, D is the eigenvalue matrix, and V is the eigenvector matrix. The eigenvalues ​​are arranged in descending order, and the top K modes with a cumulative contribution rate ≥ 0.99 are selected to obtain the EOF mode matrix EOFs. Principal components are then calculated. Wherein, EOFs is the modality matrix composed of the first K eigenvectors. For the weighted average sound velocity profile, PC is the principal component matrix, and K is the number of selected modes, i.e., the principal component dimension. The sample module is used to construct time window samples from the principal component time series. The time step is set to t. The input for each time window is the principal component PC (with dimension K×t) for t consecutive time steps, and the output is the principal component PC (K×1) for the (t+1)th time step, where t is the time step. The prediction module is used to train an LSTM network using the constructed inputs and outputs to obtain the principal component predictions for future time steps. use Reconstructing the sound velocity profile yields predicted sound velocity profile values ​​for the target year and month, where... The predicted principal components; The output module extracts the actual sound velocity profile values ​​for the target year, compares them with the predicted values, and uses the root mean square error (RMSE) and mean absolute error (MAE) as evaluation criteria. It iterates through α, β, σ, and t to output the final prediction result, the final RMSE, and the final MAE. Among them, y i For the i-th true value, Let be the i-th predicted value, and n be the number of samples.

9. A computer device comprising a memory and a processor, the memory storing computer-readable instructions, wherein the processor, when executing the computer-readable instructions, implements the steps of the fusion-weighted EOF-LSTM ocean acoustic profile prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the ocean acoustic profile prediction method based on any one of claims 1 to 7.

Images