A model training method, a seabed terrain prediction method and a model training device

Abstract

The present invention discloses a model training method, a seabed terrain prediction method and a model training device. The method includes: obtaining an initial training data set; obtaining verification data; performing data dimension augmentation on gravity data in the initial training data set through two-dimensional empirical mode decomposition, and then sorting to obtain test data; using wavelength band depth data as an output label, inputting the test data into a preset neural network for network training to obtain a depth prediction result; based on a single batch of verification data, combining the depth prediction result to perform distance-based depth error evaluation processing to obtain a depth error; performing backpropagation on the depth error, and combining gradient descent to correct the parameters of the preset neural network until a preset condition is satisfied, so as to obtain a seabed terrain depth prediction model. The present invention does not require a large amount of computation, is applicable to seabed terrain prediction in large areas, and does not require parameter tuning, can better depict the terrain, and can be widely applied to the field of data processing technology.

Description

[0001] This application is a divisional application of the application with the application date of December 06, 2023, application number 202311668062.5, and invention title "Model Training Method, Seabed Topography Prediction Method and Model Training Device". Technical Field

[0002] The present invention relates to the technical field of data processing, and in particular, to a model training method, a seabed topography prediction method and a model training device. Background Art

[0003] Currently, the mainstream technical solution for seabed topography prediction takes gravity data as input and calculates the seabed topography through various inversion algorithms (such as the Parker formula inversion algorithm). This method has a huge amount of calculation and large data errors.

[0004] The disadvantages of the prior art include: 1) Large amount of calculation. The inversion calculation involves a large number of matrix calculations. As the amount of data increases, the amount of calculation also rises sharply, and it is impossible to process the topography prediction of a large area. 2) Many parameters and complex parameter tuning. General inversion algorithms design multiple parameters (such as depth, density difference, high-pass filtering, low-pass filtering, window function threshold, etc.), and the parameter tuning is complex. 3) Large errors and inability to depict a fine model. Depicting the seabed topography requires continuous parameter tuning, and it is difficult to obtain a fine topography. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems in the related art to some extent. Therefore, the present invention provides a model training method, a seabed topography prediction method and a model training device, which can efficiently perform model training and obtain an accurate seabed topography depth prediction model.

[0006] On the one hand, an embodiment of the present invention provides a model training method, including:

[0007] Obtain an initial training data set; obtain verification data; the initial training data set includes matching gravity data and wavelength band depth data; the verification data includes multibeam depth data and seismic depth data;

[0008] Through fast Fourier transform, perform data dimension expansion on the gravity data in the initial training data set, and then organize and obtain test data;

[0009] Use the wavelength band depth data as the output label, input the test data into a preset neural network for network training, and obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer and an output layer, and the hidden layer includes multiple fully connected layers;

[0010] Based on the verification data of a single batch, combined with the depth prediction results, perform distance-based depth error evaluation processing to obtain the depth error;

[0011] Backpropagate the depth error, and combine gradient descent to correct the parameters of the preset neural network until the preset conditions are met to obtain the seabed terrain depth prediction model; the preset conditions include one or more combinations of the verification data used reaching the preset number of batches, the depth error being less than the preset threshold, and the network training reaching the preset number of iterations.

[0012] Optionally, through fast Fourier transform, perform data dimension augmentation on the gravity data in the initial training dataset, and then organize it to obtain the test data, including:

[0013] Organize the gravity data into a two-dimensional matrix;

[0014] Perform two-dimensional Fourier decomposition on the two-dimensional matrix to obtain the frequency domain information corresponding to the gravity data;

[0015] Decompose the frequency domain information into the real part of the frequency domain and the imaginary part of the frequency domain, and then organize it in combination with the gravity data to obtain the test data.

[0016] Optionally, the test data includes gravity data and the real part of the frequency domain and the imaginary part of the frequency domain obtained by processing the gravity data; the input layer includes a first input layer and a second input layer, and the hidden layer includes a first hidden layer and a second hidden layer; input the test data into the preset neural network for network training to obtain the depth prediction result, including:

[0017] Input the gravity data in the time domain format into the first input layer, and then perform the first multi-layer full connection processing on the result of the first input layer through the first hidden layer to obtain the first feature;

[0018] Input the real part of the frequency domain and the imaginary part of the frequency domain in the frequency domain format into the second input layer, and then perform the second multi-layer full connection processing on the result of the second input layer through the second hidden layer to obtain the second feature;

[0019] Obtain the depth prediction result through the output layer based on the first feature and the second feature.

[0020] Optionally, based on the verification data of a single batch, combined with the depth prediction results, perform distance-based depth error evaluation processing to obtain the depth error, including:

[0021] Based on the verification data of a single batch, obtain multiple gravity data points through the positions corresponding to the verification data; the gravity data points represent the positions corresponding to the gravity data; and then determine the distance data between the positions corresponding to the verification data and each gravity data point;

[0022] According to the gravity data corresponding to the gravity data points, obtain the corresponding predicted depth from the depth prediction results;

[0023] Based on the predicted depths corresponding to the distance data and each gravity data point, obtain the estimated depth of the position corresponding to the verification data; furthermore, combine the verification data of a single batch to obtain an error depth sequence;

[0024] Calculate the root mean square value based on the error depth sequence to obtain the depth error.

[0025] Optionally, the method further includes at least one of the following:

[0026] When no gravity data point is obtained through the position corresponding to the verification data, filter the verification data;

[0027] When the position corresponding to the verification data coincides with a certain gravity data point, use the predicted depth obtained from the gravity data corresponding to this gravity data point as the estimated depth of the position corresponding to the verification data.

[0028] Optionally, perform backpropagation on the depth error and correct the parameters of the preset neural network in combination with gradient descent, including:

[0029] Use the error backpropagation algorithm to perform backpropagation on the gradient information of the depth error in the preset neural network, and implement gradient descent to correct the parameters of the preset neural network based on the adaptive moment estimation method.

[0030] On the other hand, an embodiment of the present invention provides a seabed terrain prediction method, including:

[0031] Obtain the gravity data of each position in the seabed area to be recognized;

[0032] Through fast Fourier transform, expand the data dimension of the gravity data, and then organize to obtain the input data;

[0033] Input the input data into the seabed terrain depth prediction model to perform seabed terrain depth prediction to obtain the predicted depths of each position in the seabed area to be recognized; wherein, the seabed terrain depth prediction model is trained by the previous model training method;

[0034] According to the predicted depths of each position in the seabed area to be recognized, obtain the seabed terrain of the seabed area to be recognized.

[0035] On the other hand, an embodiment of the present invention provides a model training device, including:

[0036] The first module is used to obtain the initial training data set; obtain the verification data; the initial training data set includes the matched gravity data and wavelength band depth data; the verification data includes multi-beam depth data and seismic depth data;

[0037] The second module is used to perform data dimension augmentation on the gravity data in the initial training dataset through fast Fourier transform, and then organize and obtain the test data;

[0038] The third module is used to take the wavelength band depth data as the output label, input the test data into a preset neural network for network training, and obtain the depth prediction result; the preset neural network includes an input layer, a hidden layer, and an output layer, and the hidden layer includes multiple fully connected layers;

[0039] The fourth module is used to perform distance-based depth error evaluation processing on the basis of a single batch of validation data in combination with the depth prediction result to obtain the depth error;

[0040] The fifth module is used to perform backpropagation on the depth error and correct the parameters of the preset neural network in combination with gradient descent until the preset conditions are met, and obtain the seabed terrain depth prediction model; the preset conditions include one or more combinations of the validation data used reaching the preset number of batches, the depth error being less than the preset threshold, and the network training reaching the preset number of iterations.

[0041] Optionally, the model training device further includes at least one of the following:

[0042] The sixth module is used to filter the validation data when no gravity data point is obtained at the position corresponding to the validation data;

[0043] The seventh module is used to, when the position corresponding to the validation data coincides with a certain gravity data point, use the predicted depth obtained from the gravity data corresponding to the gravity data point as the estimated depth of the position corresponding to the validation data.

[0044] On the other hand, an embodiment of the present invention provides a seabed terrain prediction device, including:

[0045] The eighth module is used to obtain the gravity data of each position in the seabed area to be recognized;

[0046] The ninth module is used to perform data dimension augmentation on the gravity data through fast Fourier transform, and then organize and obtain the input data;

[0047] The tenth module is used to input the input data into the seabed terrain depth prediction model to predict the seabed terrain depth of each position in the seabed area to be recognized; wherein, the seabed terrain depth prediction model is trained by the previous model training method;

[0048] The eleventh module is used to obtain the seabed terrain of the seabed area to be recognized according to the predicted depth of each position in the seabed area to be recognized.

[0049] On the other hand, an embodiment of the present invention provides a model training method, including:

[0050] Obtain an initial training dataset; obtain validation data; the initial training dataset includes matched gravity data and depth data; the validation data includes multi-beam depth data and seismic depth data;

[0051] Through two-dimensional empirical mode decomposition, perform data dimension augmentation on the gravity data in the initial training dataset, and then organize and obtain test data;

[0052] Using the depth data as the output label, input the test data into a preset neural network for network training to obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer, and an output layer, and the hidden layer includes multiple fully connected layers;

[0053] Based on a single batch of validation data, combine the depth prediction result to perform distance-based depth error evaluation processing to obtain a depth error;

[0054] Perform backpropagation on the depth error, and combine gradient descent to correct the parameters of the preset neural network until the preset conditions are met to obtain a seabed terrain depth prediction model; the preset conditions include one or more combinations of the validation data used reaching a preset number of batches, the depth error being less than a preset threshold, and the network training reaching a preset number of iterations.

[0055] Optionally, through two-dimensional empirical mode decomposition, perform data dimension augmentation on the gravity data in the initial training dataset, and then organize and obtain test data, including:

[0056] Organize the gravity data into a two-dimensional matrix;

[0057] Project the two-dimensional matrix into the two-dimensional frequency domain through two-dimensional empirical mode decomposition, and then decompose the two-dimensional matrix into multiple independent modes to obtain a residual component and several intrinsic mode function components;

[0058] Organize and obtain test data according to the intrinsic mode function components, the residual component, and the gravity data.

[0059] Optionally, the test data includes gravity data, a residual component, and several intrinsic mode function components obtained by processing the gravity data; the input layer includes a first input layer and a second input layer, and the hidden layer includes a first hidden layer and a second hidden layer; input the test data into the preset neural network for network training to obtain a depth prediction result, including:

[0060] Input the gravity data into the first input layer, and then perform first multi-layer fully connected processing on the result of the first input layer through the first hidden layer to obtain a first feature;

[0061] Input the intrinsic mode function components and the residual component into the second input layer, and then perform second multi-layer fully connected processing on the result of the second input layer through the second hidden layer to obtain a second feature;

[0062] The depth prediction result is obtained by the output layer according to the first feature and the second feature.

[0063] Optionally, based on the verification data of a single batch, a distance-based depth error evaluation process is performed in combination with the depth prediction result to obtain a depth error, including:

[0064] Based on the verification data of a single batch, multiple gravity data points are obtained through the positions corresponding to the verification data; the gravity data points represent the positions corresponding to the gravity data; and then the distance data between the positions corresponding to the verification data and each gravity data point is determined;

[0065] According to the gravity data corresponding to the gravity data points, the corresponding predicted depth is obtained from the depth prediction result;

[0066] According to the distance data and the predicted depths corresponding to each gravity data point, the estimated depth of the position corresponding to the verification data is obtained; and then, in combination with the verification data of a single batch, an error depth sequence is obtained;

[0067] Based on the error depth sequence, a root mean square value calculation is performed to obtain a depth error.

[0068] Optionally, the method further includes at least one of the following:

[0069] When no gravity data points are obtained through the position corresponding to the verification data, the verification data is filtered;

[0070] When the position corresponding to the verification data coincides with a certain gravity data point, the predicted depth obtained from the gravity data corresponding to the gravity data point is used as the estimated depth of the position corresponding to the verification data.

[0071] Optionally, the depth error is backpropagated and the parameters of the preset neural network are corrected in combination with gradient descent, including:

[0072] Using the error backpropagation algorithm, the gradient information of the depth error is backpropagated in the preset neural network, and gradient descent is implemented based on the adaptive moment estimation method to correct the parameters of the preset neural network.

[0073] On the other hand, the model training device provided by the embodiments of the present invention is equally applicable to the above model training method. Specifically:

[0074] The first module is used to obtain an initial training data set; obtain verification data; the initial training data set includes matching gravity data and depth data; the verification data includes multi-beam depth data and seismic depth data;

[0075] The second module is used to perform data dimension augmentation on the gravity data in the initial training dataset through two-dimensional empirical mode decomposition, and then organize and obtain the test data;

[0076] The third module is used to take the depth data as the output label, input the test data into a preset neural network for network training, and obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer, and an output layer, and the hidden layer includes multiple fully connected layers;

[0077] The fourth module is used to perform distance-based depth error evaluation processing on the basis of a single batch of validation data in combination with the depth prediction result to obtain a depth error;

[0078] The fifth module is used to perform backpropagation on the depth error and correct the parameters of the preset neural network in combination with gradient descent until a preset condition is met to obtain a seabed terrain depth prediction model; the preset condition includes one or more combinations of the validation data used reaching a preset batch number, the depth error being less than a preset threshold, and the network training reaching a preset number of iterations.

[0079] Optionally, the model training device further includes at least one of the following:

[0080] The sixth module is used to filter the validation data when no gravity data point is obtained at the position corresponding to the validation data;

[0081] The seventh module is used to, when the position corresponding to the validation data coincides with a certain gravity data point, use the predicted depth obtained from the gravity data corresponding to the gravity data point as the estimated depth of the position corresponding to the validation data.

[0082] On the other hand, an embodiment of the present invention provides a seabed terrain prediction method, including:

[0083] Obtain the gravity data of each position in the seabed area to be recognized;

[0084] Perform data dimension augmentation on the gravity data through two-dimensional empirical mode decomposition, and then organize and obtain the input data;

[0085] Input the input data into the seabed terrain depth prediction model to perform seabed terrain depth prediction to obtain the predicted depth of each position in the seabed area to be recognized; wherein, the seabed terrain depth prediction model is trained by the above model training method;

[0086] Obtain the seabed terrain of the seabed area to be recognized according to the predicted depth of each position in the seabed area to be recognized.

[0087] On the other hand, the seabed terrain prediction device provided by the embodiment of the present invention is equally applicable to the above seabed terrain prediction method. Specifically:

[0088] The eighth module is used to obtain the gravity data at each position in the seabed area to be recognized;

[0089] The ninth module is used to perform data dimension augmentation on the gravity data through two-dimensional empirical mode decomposition, and then organize and obtain the input data;

[0090] The tenth module is used to input the input data into the seabed terrain depth prediction model to predict the seabed terrain depth at each position in the seabed area to be recognized; wherein, the seabed terrain depth prediction model is trained by the previous model training method;

[0091] The eleventh module is used to obtain the seabed terrain of the seabed area to be recognized according to the predicted depth at each position in the seabed area to be recognized.

[0092] On the other hand, an embodiment of the present invention provides an electronic device, including: a processor and a memory; the memory is used to store a program; the processor executes the program to implement any one of the above model training methods or seabed terrain prediction methods.

[0093] On the other hand, an embodiment of the present invention provides a computer storage medium, in which a program executable by a processor is stored, and the program executable by the processor is used to implement any one of the above model training methods or seabed terrain prediction methods when executed by the processor.

[0094] In an embodiment of the present invention, an initial training data set is first obtained; verification data is obtained; the initial training data set includes matched gravity data and wavelength band depth data; the verification data includes multi-beam depth data and seismic depth data; through fast Fourier transform, data dimension augmentation is performed on the gravity data in the initial training data set, and then test data is sorted out; using the wavelength band depth data as the output label, the test data is input into a preset neural network for network training to obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer and an output layer, and the hidden layer includes multiple fully connected layers; based on a single batch of verification data, a distance-based depth error evaluation process is performed in combination with the depth prediction result to obtain a depth error; the depth error is backpropagated, and the parameters of the preset neural network are corrected in combination with gradient descent until a preset condition is met, and a seabed terrain depth prediction model is obtained; the preset conditions include one or a combination of more of the verification data used reaching a preset number of batches, the depth error being less than a preset threshold, and the network training reaching a preset number of iterations. In the embodiment of the present invention, by using gravity data for network training of the neural network, the overall computational amount is small, the model can be reused. After training the model with a large amount of gravity data, the model can be directly reused without a large computational amount later, which is suitable for seabed terrain prediction in large areas. Moreover, the neural network does not have dozens of parameters such as high-pass filtering and low-pass filtering, and belongs to a black box model without interpretability, so there is no need to adjust parameters. At the same time, it can achieve more refined terrain characterization. The neural network has good non-linear simulation ability and can better finely characterize the terrain. Description of the Drawings

[0095] The drawings are used to provide a further understanding of the technical solutions of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation to the technical solutions of the present invention.

[0096] Figure 1 It is a schematic diagram of an implementation environment for model training provided by an embodiment of the present invention;

[0097] Figure 2 It is a schematic flowchart of a model training method provided by an embodiment of the present invention;

[0098] Fig. 3(a) is a schematic diagram of gravity anomaly GA data provided by an embodiment of the present invention;

[0099] Fig. 3(b) is a schematic diagram of terrain data provided by an embodiment of the present invention;

[0100] Figure 4 It is a BP neural network topology diagram provided by an embodiment of the present invention;

[0101] Figure 5 It is a schematic diagram of the data flow of the BP neural network provided by an embodiment of the present invention;

[0102] Figure 6 Schematic diagram of the depth error calculation principle based on distance provided by an embodiment of the present invention;

[0103] Figure 7 Schematic flowchart of another model training method provided by an embodiment of the present invention;

[0104] Figure 8 Another BP neural network topology diagram provided by an embodiment of the present invention;

[0105] Figure 9 Schematic overall flowchart of a model training provided by an embodiment of the present invention;

[0106] Figure 10 Schematic overall flowchart of another model training provided by an embodiment of the present invention;

[0107] Figure 11 Schematic flowchart of a seabed terrain prediction method provided by an embodiment of the present invention;

[0108] Figure 12 Schematic flowchart of a process for predicting seabed terrain based on a trained model provided by an embodiment of the present invention;

[0109] Figure 13 Schematic flowchart of another seabed terrain prediction method provided by an embodiment of the present invention;

[0110] Figure 14 Schematic flowchart of another process for predicting seabed terrain based on a trained model provided by an embodiment of the present invention;

[0111] Figure 15 Schematic structural diagram of a model training device provided by an embodiment of the present invention. Detailed implementation manners

[0112] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

[0113] It should be noted that although functional module division is performed in the system schematic diagram and the logical sequence is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from the module division in the system or the sequence in the flowchart. Terms such as "first / S100" and "second / S200" in the description, claims and the above accompanying drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence.

[0114] References to "embodiments" in this application mean that specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the invention. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment each time, nor are they independent or alternative embodiments mutually exclusive of other embodiments. Those skilled in the art will explicitly and implicitly understand that the embodiments described in this application can be combined with other embodiments.

[0115] For the convenience of understanding the technical solution of the present invention, first, the professional names that may appear in the technical solution of the present invention are explained:

[0116] Seabed topography: Seabed topography can be divided into three basic topographies: continental margin, ocean basin, and mid-ocean ridge. The continental margin is the transitional zone between the land and the ocean floor. The mid-ocean ridge is a continuous and huge ridge on the ocean floor. The ocean basin is the broad zone between the mid-ocean ridge and the continental margin. The seabed topography mentioned in this application is to calculate and predict the depth of seawater, so as to obtain the entire seabed topography.

[0117] BP (Back Propagation) neural network: It is a multi-layer feedforward neural network trained according to the error backpropagation algorithm and is one of the most widely used neural network models. The calculation process of the BP neural network consists of a forward calculation process and a backward calculation process. In the forward propagation process, the input pattern is processed layer by layer from the input layer through the hidden unit layer and then transferred to the output layer. The state of each layer of neurons only affects the state of the next layer of neurons. If the desired output cannot be obtained in the output layer, it will turn to the backward propagation, and the error signal will be returned along the original connection path. By modifying the weights of each neuron, the error signal is minimized.

[0118] Bouguer gravity anomaly GA (Gravity anomalies): The observed gravity value (g_obs) at the measurement point, after latitude correction (g_lat), intermediate layer correction (g_mid), and terrain correction (g_ter), the obtained gravity value is called "Bouguer correction of the gravity value", and then subtracting the normal gravity value. The Bouguer gravity anomaly reflects the influence of abnormal mass inside the earth on the gravity measurement result. That is to say, the Bouguer gravity anomaly is mainly caused by the undulations of the Mohorovicic discontinuity, Conrad discontinuity, sedimentary basement interface, geological structures, and ore bodies, etc., where the material densities are uneven. Therefore, the Bouguer gravity anomaly has certain geological significance. It is commonly used to study the mass distribution inside the earth's crust. The gravity anomaly mentioned in the present invention has the same meaning as the Bouguer gravity anomaly. Generally speaking, the gravity anomaly can be obtained by shipborne measurement and satellite altimetry.

[0119] Vertical Gravity Gradient Anomalies (VGGA): Gravity gradient anomalies are divided into horizontal gradient anomalies and vertical gradient anomalies. The horizontal gradient anomalies Wxz and Wyz can be obtained by two methods. One is obtained by torsion balance measurement after terrain correction and normal correction; the other is obtained by taking the derivatives of the gravity anomaly in the x and y directions. This application mainly uses the vertical gravity gradient anomaly. Generally speaking, the gravity anomaly can be obtained by shipborne measurement and satellite altimetry.

[0120] Fast Fourier Transform (FFT): FFT is an efficient algorithm for the Discrete Fourier Transform (DFT), called the fast Fourier transform. The Fourier transform is one of the most basic methods in time-domain and frequency-domain transform analysis. This invention uses the Fourier transform to separate signals in the time domain and frequency domain. The frequency-domain data obtained by FFT is complex, divided into a real part and an imaginary part. From a personal understanding, the real part and the imaginary part represent different information.

[0121] RMS: Abbreviation for Root Mean Square, which is the result of taking the square root after dividing the sum of the squares of N terms by N, that is, the result of the root mean square.

[0122] Bidimensional Empirical Mode Decomposition (BEMD): It projects the data to be processed onto the two-dimensional frequency domain and then finely decomposes it into multiple independent modes. Simply put, EMD is mainly used for one-dimensional data processing, and BEMD is extended for processing two-dimensional data. BEMD can decompose the original data into multiple independent modes according to the data composition.

[0123] It can be understood that the model training method provided by the embodiments of the present invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various types of terminals or servers. When the computer device in the embodiment is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal is a smart phone, a tablet computer, a laptop computer, a desktop computer, etc., but it is not limited thereto.

[0124] As Figure 1 shown, it is a schematic diagram of an implementation environment provided by an embodiment of the invention. Refer toFigure 1 , the implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be network-connected by wireless or wired means to complete data transmission and exchange.

[0125] The server 101 can be an independent physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0126] In addition, the server 101 can also be a node server in a blockchain network. Among them, the blockchain is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, and encryption algorithms.

[0127] The terminal 102 can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The terminal 102 and the server 101 can be directly or indirectly connected through wired or wireless communication means, and the embodiments of the present invention do not limit this here.

[0128] Exemplarily based on Figure 1 the shown implementation environment, the embodiments of the present invention provide a model training method. Taking the application of this model training method in the server 101 as an example for illustration, it can be understood that this model training method can also be applied to the terminal 102.

[0129] Referring to Figure 2 , Figure 2 is the flowchart of the model training method applied to the server provided by the embodiments of the present invention. The execution subject of this model training method can be any of the foregoing computer devices. Referring to Figure 2 , this method includes the following steps:

[0130] S100. Obtain an initial training data set; obtain verification data;

[0131] It should be noted that the initial training data set includes matching gravity data and wavelength band depth data; the verification data includes multi-beam depth data and seismic depth data.

[0132] S200. Through fast Fourier transform, perform data dimension augmentation on the gravity data in the initial training data set, and then organize and obtain test data;

[0133] It should be noted that in some embodiments, step S200 may include: organizing the gravity data into a two-dimensional matrix; performing two-dimensional Fourier decomposition on the two-dimensional matrix to obtain the frequency-domain information corresponding to the gravity data; decomposing the frequency-domain information into a frequency-domain real part and a frequency-domain imaginary part, and then organizing the test data by combining the gravity data.

[0134] Exemplarily, in some specific embodiments, the processes of steps S100 and S200 may obtain test data and verification data by preprocessing the gravity data and depth data. The specific steps are as follows:

[0135] Regarding the test data, the input data is mainly gravity data, including gravity anomalies GA (Gravity anomalies), vertical gravity gradient anomalies VGGA (Vertical Gravity Gradient Anomalies), wavelength band depth WBD (Wavelength Band Depth), etc. These quantities come from geological ship tests or satellite altimetry, and the format is in the form of (longitude, latitude, gravity anomaly GA, vertical gravity gradient anomaly VGGA, wavelength band depth WBD). Since the neural network training data requires clear input and output labels, the wavelength band depth WBD here is the label data. As mentioned before, the model obtained by training only with GA and VGGA data is not fine enough, so more dimensional data is needed for supplementation.

[0136] In order to obtain data in more dimensions, the present invention proposes an idea for separating time-domain and frequency-domain signals based on FFT (Fast Fourier Transform). Taking the decomposition of gravity anomaly GA as an example, the gravity anomaly GA data is organized into a two-dimensional matrix, and then two-dimensional Fourier decomposition (2FFT) is performed to obtain the frequency-domain information corresponding to the gravity anomaly GA. Considering that the frequency-domain information is in complex form and the subsequent error calculation needs to be performed in the real domain, the complex matrix is separated into real and imaginary parts respectively to obtain the gravity anomaly frequency-domain real part GAFR matrix (Gravity Anomalies Frequency Domain Real Part) and the gravity anomaly frequency-domain imaginary part GAFI matrix (Gravity Anomalies Frequency Domain Imaginary Part). The same approach is applied to the vertical gravity gradient anomaly VGGA. It is organized into a two-dimensional matrix and then 2FFT decomposition is carried out to obtain the vertical gravity gradient anomaly frequency-domain real part VGGAFR matrix (Vertical Gravity Gradient Anomalies Frequency Domain Real Part) and the vertical gravity gradient anomaly frequency-domain imaginary part VGGAFI matrix (Vertical Gravity Gradient Anomalies Frequency Domain Imaginary Part). Finally, data in the form of (longitude, latitude, GA, VGGA, GAFR, GAFI, VGGAFR, VGGAFI, wavelength band depth WBD) can be obtained. This format of data can be directly used for network model training. Among them, the test data directly input into the network model are GA, VGGA, GAFR, GAFI, VGGAFR, VGGAFI. Longitude and latitude are used as position data to assist in subsequent distance-based depth error evaluation processing, and the wavelength band depth WBD is used as the output label.

[0137] Here, the physical meaning of using FFT to decompose data is explained: GA and VGGA can directly reflect the terrain depth. The existing inversion algorithms use GA as the input and then calculate the depth data. Referring to Fig. 3(a), it is the GA data, and Fig. 3(b) is the terrain data. The two graphs are very similar. The frequency-domain information obtained through FFT decomposition can directly reflect the tortuosity of the terrain (which is widely used in the fields of signal separation and filtering). This is the most important reason for using FFT here. Therefore, the present invention proposes to jointly use the time-domain information (GA and VGGA) and the frequency-domain information (GAFR, GAFI, VGGAFR, and VGGAFI) for depth prediction (since the frequency-domain information image has no obvious features, no examples are given here). Moreover, through a large number of tests, the present invention has found and designed a BP neural network structure with better performance, which is also related to the topology and time-frequency domain.

[0138] Regarding the verification data, the present invention does not extract the verification data from the preprocessed data, but uses the data obtained from other channels as the verification data. It is mainly multi-beam depth data (Multi-beam Depth) or seismic depth data (Seismic Depth), which are collectively referred to as verification data later, and the format is in the form of (longitude, latitude, depth). The reason for doing this is that the data obtained from the multi-beam and seismic channels are more accurate, and at the same time, it can ensure a better refined description of the BP neural network model.

[0139] S300. Using the wavelength-band depth data as the output label, input the test data into a preset neural network for network training to obtain the depth prediction result;

[0140] It should be noted that the preset neural network includes an input layer, a hidden layer, and an output layer. The hidden layer includes multiple fully connected layers; the test data includes gravity data and the real part and imaginary part of the frequency domain processed based on the gravity data; the input layer includes a first input layer and a second input layer, and the hidden layer includes a first hidden layer and a second hidden layer; in some embodiments, inputting the test data into the preset neural network for network training to obtain the depth prediction result may include: inputting the gravity data in the time-domain format into the first input layer, and then performing the first multi-layer fully connected processing on the result of the first input layer through the first hidden layer to obtain the first feature; inputting the real part and imaginary part of the frequency domain in the frequency-domain format into the second input layer, and then performing the second multi-layer fully connected processing on the result of the second input layer through the second hidden layer to obtain the second feature; obtaining the depth prediction result through the output layer according to the first feature and the second feature.

[0141] Exemplarily, in some specific embodiments, the steps for training the BP neural network based on the test data are as follows:

[0142] It should be noted that the present invention does not make theoretical improvements to the BP neural network itself, but puts forward unique ideas in aspects such as data processing and model training. Based on the prediction direction of the seabed topography, through a large number of tests and theoretical calculations, the present invention proposes a network topology model with good adaptability. Specifically, as shown in Figure 4 the topological diagram of the BP neural network shown

[0143] As shown in Figure 4 , the gravity data is preprocessed through the previous steps to obtain input data such as GA, VGGA, GAFR, GAFI, VGGAFR, and VGGAFI. Distinguishing by time domain and frequency domain, the time-domain data is connected to the input layer (2) (i.e., the first input layer), and the frequency-domain data is connected to the input layer (4) (i.e., the second input layer), and then connected to the corresponding fully connected layer, which is mainly used for feature extraction training. The topological key point proposed by the present invention is to separately train and calculate the time domain and frequency domain in the input layer and the hidden layer, and then merge the vectors at the output layer, and finally output uniformly (for details, refer to Figure 5 ). The reason for this design: The time-domain data has a strong similarity with the seabed topography data image, so it can be approximately understood as a linear relationship. The frequency-domain data mainly reflects the tortuosity degree of the seabed topography data, and the image does not have a strong similarity, so it can be approximately understood as a non-linear relationship. Through different types of topological designs, implementations, and tests, it is found that the model designed in this way performs better.

[0144] Figure 5 It is the data flow diagram corresponding to the BP neural network. The 6 inputs such as the input data GA, VGGA, GAFR, GAFI, VGGAFR, and VGGAFI, the time-domain data and the frequency-domain data are separated in the input layer and the hidden layer, and unified output is completed through vector splicing at the output layer.

[0145] The topological design proposed by the present invention starts from the linear and non-linear relationships between data. Through theoretical calculations and a large number of tests, the data accuracy and performance of the model output are better.

[0146] S400. Based on the verification data of a single batch, combined with the depth prediction result, perform distance-based depth error evaluation processing to obtain the depth error;

[0147] It should be noted that in some embodiments, step S400 may include: obtaining a plurality of gravity data points based on the verification data of a single batch through the positions corresponding to the verification data; the gravity data points representing the positions corresponding to the gravity data; and then determining the distance data between the positions corresponding to the verification data and each gravity data point; obtaining the corresponding predicted depth from the depth prediction result according to the gravity data corresponding to the gravity data point; obtaining the estimated depth of the position corresponding to the verification data according to the distance data and the predicted depths corresponding to each gravity data point; and then combining the verification data of a single batch to obtain an error depth sequence; calculating the root mean square value based on the error depth sequence to obtain the depth error.

[0148] Among them, in some embodiments, the method may further include at least one of the following: when no gravity data point is obtained through the position corresponding to the verification data, filtering the verification data; when the position corresponding to the verification data coincides with a certain gravity data point, using the predicted depth obtained from the gravity data corresponding to the gravity data point as the estimated depth of the position corresponding to the verification data.

[0149] Exemplarily, in some specific embodiments, the error is calculated based on the verification data of a single batch. Step S300 is trained based on all data, but the model performance is verified by calculating the batch data error. This is because there is a large amount of depth data such as multi-beam and seismic data. If too much data is used for prediction, the calculation efficiency will be too low without batch verification, resulting in the model training, verification, and output being too slow. Generally, the batch size BatchSize is 512, which can be configured. If it is configured as the number of the maximum depth data, the verification process degenerates to all depth data participating in the verification.

[0150] Different from directly using gravity data to divide the test set and the verification set, the present invention obtains depth data from other channels as verification data, which is also one of the core points of the present invention. The format of the verification data is in the form of (longitude, latitude, depth), and the longitude and latitude of the verification data are not necessarily the same as those of the gravity data. Therefore, the present invention uses a depth error evaluation method based on distance.

[0151] Such as Figure 6 , GA11 to GA34 represent gravity data, and D1, D2, and D3 represent depth data. For each verification data of (longitude, latitude, depth), four gravity data points closest to itself are found in the depth data according to the longitude and latitude, and then the distances of the corresponding points are calculated. The closer the distance, the greater the influence, and the calculation mainly falls into three cases.

[0152] Case1: Taking D1 as an example, the gravity data points closest to D1 are GA11, GA12, GA21, and GA22. The distances between D1 and these four points are calculated using the longitude and latitude, which are l1, l2, l3, and l4 respectively. Then the estimated depth of D1 is:

[0153]

[0154] where Depth 11 represents the wavelength band depth WBD of GA11, and the others are similar. PredictDepth represents the depth predicted based on these four points. If the longitude and latitude of D1 are exactly the same as those of a certain point in the gravity data, then PredictDepth is the wavelength band depth WBD of that point.

[0155] Case2: Taking D2 as an example, the gravity data points closest to D2 are GA22 and GA23. Calculate the distances from D2 to these four points using the longitude and latitude, which are l5 and l6 respectively. Then the estimated depth of D2 is:

[0156]

[0157] where Depth 22 represents the wavelength band depth WBD of GA22, and the others are similar. PredictDepth represents the depth predicted based on these two points. If the longitude and latitude of D2 are exactly the same as those of a certain point in the gravity data, then PredictDepth is the wavelength band depth WBD of that point.

[0158] Case3: Taking D3 as an example, since the longitude and latitude of D3 are outside the gravity data, D3 does not participate in the calculation. During data preprocessing, data like D3 that exceeds the gravity longitude and latitude range will be filtered out and not participate in the calculation.

[0159] The formula for calculating the distance based on the longitude and latitude of two points is as follows, where lng1 and lat1 represent the longitude and latitude of point 1, lng2 and lat2 represent the longitude and latitude of point 2, asin and sqrt represent the inverse cosine operation and the square root operation respectively, Distance is the finally calculated distance, and 6371 represents the radius of the earth, with the unit of km.

[0160] dlon = lng2 - lng1

[0161] dlat = lat2 - lat1

[0162]

[0163] After calculating the predicted depths of all verification data one by one, the error depth sequence = predicted depth sequence - true depth sequence. Calculate the RMS (Root Mean Square) based on the error depth sequence, and this is the error. The present invention makes neural network adjustments based on this error, which is one of the core inventive points.

[0164] S500. Perform backpropagation on the depth error, and combine gradient descent to correct the parameters of the preset neural network until the preset conditions are met, thereby obtaining a seabed terrain depth prediction model.

[0165] It should be noted that the preset conditions include one or more combinations of the verification data used reaching the preset batch number, the depth error being less than the preset threshold, and the network training reaching the preset number of iterations. In some embodiments, performing backpropagation on the depth error and combining gradient descent to correct the parameters of the preset neural network may include: using the error backpropagation algorithm to perform backpropagation on the gradient information of the depth error in the preset neural network, and implementing gradient descent based on the adaptive moment estimation method to correct the parameters of the preset neural network.

[0166] Exemplarily, in some specific embodiments, the correction or adjustment of the parameters of the network model may include the following process:

[0167] Error backpropagation: Use the error backpropagation algorithm to perform backpropagation on the gradient information of the error loss function.

[0168] Gradient descent to adjust the neural network parameters. Adjust the neural network parameters based on the Adam algorithm. The Adam algorithm, namely the Adaptive Moment Estimation method, can calculate the adaptive learning rate for each parameter. Compared with other adaptive learning rate algorithms, the Adam algorithm has a faster convergence speed, more effective learning effect, and can correct problems existing in other optimization techniques, such as the disappearance of the learning rate, slow convergence, or large fluctuations in the loss function caused by high-variance parameter updates.

[0169] Among them, in some embodiments, the method further includes: calculating the errors of all batches of verification data. Assuming that BatchSize is 512 and the total amount is 5120, then there are a total of 10 batches of data for calculating errors and adjusting the model. If all batches of data are traversed, calculate the overall error based on the verification data; otherwise, process the incomplete verification data in sequence; among them, calculate the overall error based on the verification data. When the training of the batch data is completed, based on the current model, all depth data participate in the calculation of the overall error, and the calculation method is the same as that in the specific embodiment of step S400.

[0170] Until the preset conditions are met, obtain a seabed terrain depth prediction model, where the preset conditions include but are not limited to:

[0171] Meet the minimum error: If the calculated error RMS (which can be judged based on the error obtained from the verification data of a single batch or the overall error obtained from the verification data of all batches, with the overall error judgment being the preferred implementation) is less than the minimum MinIterRMS (which can be configured, with a default value of 1), it indicates that the currently adjusted model is excellent enough to be directly used; otherwise, the number of iterations for network training is judged. MinIterRMS has physical significance, representing the average error depth. Assuming the depth unit involved in the calculation is km and MinIterRMS is 1, it means that the average error of the output depth matrix is required to be between [-1 km, 1 km].

[0172] Reach the maximum number of iterations: If the number of iterations exceeds the maximum limit MaxIterCnt, the final seabed terrain depth prediction model is obtained, indicating that the best seabed terrain depth prediction model has been found; otherwise, the network model continues to be trained and adjusted.

[0173] On the other hand, as Figure 7 shown, Figure 7 is a flowchart of another model training method applied to a server provided by an embodiment of the present invention. The execution subject of this model training method can be any of the aforementioned computer devices. Referring to Figure 7 , the method includes the following steps:

[0174] U100. Obtain an initial training data set; obtain verification data;

[0175] It should be noted that the initial training data set includes matching gravity data and depth data; the verification data includes multi-beam depth data and seismic depth data.

[0176] U200. Through two-dimensional empirical mode decomposition, perform data dimension expansion on the gravity data in the initial training data set, and then organize and obtain test data;

[0177] It should be noted that in some embodiments, step U200 may include: organizing the gravity data into a two-dimensional matrix; projecting the two-dimensional matrix into the two-dimensional frequency domain through two-dimensional empirical mode decomposition, and then decomposing the two-dimensional matrix into multiple independent modes to obtain a residual component and several intrinsic mode function components; organizing and obtaining test data based on the intrinsic mode function components, the residual component, and the gravity data.

[0178] Exemplarily, in some specific embodiments, the processes of steps U100 and U200 can obtain test data and verification data by preprocessing the gravity data and depth data. The specific steps are as follows:

[0179] Regarding the test data, the input data is mainly gravity data, including gravity anomaly GA, depth, etc. These quantities are from geological ship tests or satellite altimetry, and the format is like (longitude, latitude, gravity anomaly GA, depth). Since the neural network training data requires clear input and output labels, the depth here is the label data. As mentioned before, the model obtained by only relying on GA data training is not fine enough, so more dimensional data is needed to supplement.

[0180] In order to obtain more dimensional data, the present invention proposes an idea of modal separation based on BEMD This is one of the core points of the present invention, and there is no similar approach in the existing work. Finally, data in the form of (longitude, latitude, GA, IMF1, IMF2,..., IMF N , Res, depth) can be obtained, and this format of data can be directly used for network model training. Specifically:

[0181] Decompose the gravity data based on BEMD. Taking the decomposition of gravity anomaly GA as an example, organize the gravity anomaly GA data into a two-dimensional matrix, and decompose the gravity data based on the BEMD method. Finally, multiple IMF (Intrinsic Mode Function) components and Res residual components are obtained: Where GA represents the gravity anomaly data, IMF i represents the separated modal data (i.e., the intrinsic mode function components), and Res represents the residual data (i.e., the residual components). The core idea of BEMD is to decompose the data GA to be processed into IMF i in the order of the data composition components, and IMF i and IMF i+1 are in a sequential relationship, and Res is the residual left after the decomposition. Here, the physical meaning of using BEMD to decompose the mode is explained: GA is composed of the superposition of multiple different signals, and BEMD is to separate different signals. The IMF i of different modes reflects the degree of terrain tortuosity from different angles. Therefore, the present invention proposes an idea of separating modal data based on BEMD and predicting the terrain with the help of a BP neural network. After a large number of tests, it is proved that the model's ability to finely depict the terrain is significantly improved.

[0182] Regarding the verification data, the present invention does not extract verification data from the preprocessed data, but uses data obtained from other channels as verification data. It is mainly multi-beam depth data or seismic depth data, which are hereinafter referred to as verification data or validation data, and the format is like (longitude, latitude, depth). The reason for doing this is that the data obtained from multi-beam and seismic channels is more accurate, and at the same time, it can ensure better fine-grained description of the BP neural network model.

[0183] U300 uses depth data as output labels, inputs test data into a preset neural network for network training, and obtains a depth prediction result; the preset neural network includes an input layer, a hidden layer, and an output layer, and the hidden layer includes multiple fully connected layers;

[0184] It should be noted that the test data includes gravity data, residual components, and several intrinsic mode function components obtained by processing the gravity data; the input layer includes a first input layer and a second input layer, and the hidden layer includes a first hidden layer and a second hidden layer; in some embodiments, inputting the test data into the preset neural network for network training to obtain a depth prediction result may include: inputting the gravity data into the first input layer, and then performing a first multi-layer fully connected process on the result of the first input layer through the first hidden layer to obtain a first feature; inputting the intrinsic mode function components and residual components into the second input layer, and then performing a second multi-layer fully connected process on the result of the second input layer through the second hidden layer to obtain a second feature; obtaining a depth prediction result through the output layer according to the first feature and the second feature.

[0185] Exemplarily, in some specific embodiments, training a BP neural network based on test data, the specific steps are as follows:

[0186] It should be noted that the present invention does not make theoretical improvements to the BP neural network itself, but proposes unique ideas in aspects such as data processing and model training. Based on the direction of seabed terrain prediction, through a large number of tests and theoretical calculations, the present invention proposes a network topology model with better adaptability. Specifically as Figure 8 shown in the BP neural network topology diagram.

[0187] Such as Figure 8 , the gravity data is preprocessed through step 1 to obtain input data in the form of (longitude, latitude, GA, IMF1, IMF2,..., Res, depth). The original data is connected to the input layer (1), and the decomposed modal data is connected to the input layer (2), and then connected to the corresponding fully connected layer. The fully connected layer is mainly used for feature extraction training. The topological key point proposed by the present invention is that the original data and modal data are separately trained and calculated in the input layer and the hidden layer, and the vectors are merged again at the output layer, and finally unified output. The reason for such a design: There is a strong similarity between the gravity data GA and the seabed terrain data image, so it can be approximately understood as a linear relationship. The modal data mainly reflects the tortuosity degree of the seabed terrain data, and there is no strong similarity in the image, so it can be approximately understood as a non-linear relationship, and there is an order dependence relationship between the IMF i and IMF i+1 in the modal data. Through different types of topological designs, implementations, and tests, it is found that the model designed in this way performs better.

[0188] In summary, the topological design proposed by the present invention starts from the linear and non-linear relationships between data. After theoretical calculations and a large number of tests, the data accuracy and performance output by the model are better.

[0189] U400. Based on the verification data of a single batch, combined with the depth prediction result, perform distance-based depth error evaluation processing to obtain the depth error.

[0190] U500. Backpropagate the depth error, and combine gradient descent to correct the parameters of the preset neural network until the preset conditions are met, obtaining the seabed terrain depth prediction model; the preset conditions include one or a combination of more of the verification data used reaching the preset number of batches, the depth error being less than the preset threshold, and the network training reaching the preset number of iterations.

[0191] It should be noted that the specific process principles of steps U400 and U500 are the same as the process and the technical principles of the specific embodiments of the previous model training method and its steps S400 and S500, and will not be elaborated here.

[0192] To explain the principle of the technical solution of the present invention in detail, the overall process of the model training method of the present invention will be described below in conjunction with some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present invention and should not be regarded as a limitation of the present invention.

[0193] The training process of the entire BP neural network model is as Figure 9 or Figure 10 shown, and the detailed steps are as follows:

[0194] As Figure 9 shown, a model training process, step 1: Preprocess the gravity data and depth data to obtain test data and calibration data (i.e., verification data).

[0195] Regarding the test data, the input data is mainly gravity data, including gravity anomalies GA (Gravity anomalies), vertical gravity gradient anomalies VGGA (Vertical Gravity Gradient Anomalies), wavelength band depth WBD (Wavelength Band Depth), etc. These quantities come from geological ship tests or satellite altimetry, and the format is in the form of (longitude, latitude, gravity anomaly GA, vertical gravity gradient anomaly VGGA, wavelength band depth WBD). Since the neural network training data requires clear input and output labels, here the wavelength band depth WBD is the label data. As mentioned before, the model obtained by training only with GA and VGGA data is not fine enough, so more dimensional data is needed to supplement.

[0196] In order to obtain data in more dimensions, the present invention proposes an idea for separating time-domain and frequency-domain signals based on FFT (fast Fourier transform). Taking the decomposition of gravity anomaly GA as an example, the gravity anomaly GA data is organized into a two-dimensional matrix, and then two-dimensional Fourier decomposition (2FFT) is performed to obtain the frequency-domain information corresponding to the gravity anomaly GA. Considering that the frequency-domain information is in complex form and subsequent error calculations need to be carried out in the real number domain, the complex matrix is separated into real and imaginary parts respectively to obtain the gravity anomaly frequency-domain real part matrix GAFR (Gravity Anomalies FrequencyDomain Real Part) and the gravity anomaly frequency-domain imaginary part matrix GAFI (Gravity Anomalies FrequencyDomain Imaginary Part). The same approach is applied to the vertical gravity gradient anomaly VGGA. It is organized into a two-dimensional matrix and then 2FFT decomposition is performed to obtain the vertical gravity gradient anomaly frequency-domain real part matrix VGGAFR (Vertical Gravity GradientAnomalies Frequency Domain Real Part) and the vertical gravity gradient anomaly frequency-domain imaginary part matrix VGGAFI (Vertical Gravity Gradient Anomalies Frequency Domain Imaginary Part). Finally, data in the form of (longitude, latitude, GA, VGGA, GAFR, GAFI, VGGAFR, VGGAFI, wavelength band depth WBD) can be obtained. This format of data can be directly used for network model training. Among them, the test data directly input into the network model are GA, VGGA, GAFR, GAFI, VGGAFR, VGGAFI, longitude and latitude are used as position data to assist subsequent distance-based depth error evaluation processing, and the wavelength band depth WBD is used as the output label.

[0197] Here, the physical meaning of using FFT to decompose data is explained: GA and VGGA can directly reflect the terrain depth. The existing inversion algorithms use GA as the input and then calculate the depth data. Referring to Fig. 3(a), it is the GA data, and Fig. 3(b) is the terrain data. The two graphs are very similar. The frequency domain information obtained through FFT decomposition can directly reflect the tortuosity of the terrain (which is widely used in the fields of signal separation and filtering). This is the most important reason for using FFT here. Therefore, the present invention proposes to jointly use the time domain information (GA and VGGA) and the frequency domain information (GAFR, GAFI, VGGAFR, and VGGAFI) for depth prediction (since the frequency domain information image has no obvious features, no examples are given here). Moreover, through a large number of tests, the present invention has found and designed a BP neural network structure with better performance, and the topology is also related to the time domain and frequency domain.

[0198] Regarding the verification data, the present invention does not extract the verification data from the preprocessed data, but uses the data obtained from other channels as the verification data. It is mainly multi-beam depth data (Multi-beam Depth) or seismic depth data (Seismic Depth), which are hereinafter referred to as verification data or calibration data, and the format is in the form of (longitude, latitude, depth). The reason for doing this is that the data obtained from the multi-beam and seismic channels are more accurate, and at the same time, it can ensure a better refined description of the BP neural network model.

[0199] As Figure 10 shown, another model training process, step 1:

[0200] First, decompose the gravity data based on BEMD. Taking the decomposition of the gravity anomaly GA as an example, organize the gravity anomaly GA data into a two-dimensional matrix, and decompose the gravity data based on the BEMD method. Finally, multiple IMFs (Intrinsic Mode Function) and the Res residual component are obtained: where GA represents the gravity anomaly data, IMF i represents the separated modal data, and Res represents the residual data. The core idea of BEMD is to decompose the data GA to be processed into IMFs i in the order of the data composition components, and the IMFs i and IMFs i+1 are in a sequential relationship, and Res is the residual left after the last decomposition. Here, the physical meaning of using BEMD to decompose the mode is explained: GA is composed of the superposition of multiple different signals, and BEMD is to separate different signals. The IMFs of different modes iFeedback on the tortuosity of the terrain at different angles. Therefore, the present invention proposes an idea of predicting the terrain by means of the BEMD-separated modal data and with the aid of a BP neural network. After a large number of tests, it is proved that the ability of the model to depict the terrain in detail is significantly improved.

[0201] Then, preprocess the gravity data and depth data to obtain test data and calibration data. Regarding the training data, the input data is mainly gravity data, including gravity anomaly GA, depth, etc. These data are from geological ship tests or satellite altimetry, and the format is in the form of (longitude, latitude, gravity anomaly GA, depth). Since the neural network training data requires clear input and output labels, the depth here is the label data. As mentioned before, the model obtained by only relying on GA data training is not fine enough, so more dimensional data is needed to supplement.

[0202] In order to obtain more dimensional data, the present invention proposes an idea of modal separation based on BEMD This is one of the core points of the present invention, and there is no similar approach in the existing work. Finally, data in the form of (longitude, latitude, GA, IMF1, IMF2,..., IMF N , Res, depth) can be obtained, and this format of data can be directly used for network model training.

[0203] Regarding the calibration data, the present invention does not extract the calibration data from the preprocessed data, but uses the data obtained from other channels as the calibration data. It is mainly multi-beam depth data or seismic depth data, which are hereinafter referred to as calibration data or verification data, and the format is in the form of (longitude, latitude, depth). The reason for doing this is that the data obtained from multi-beam and seismic channels is more accurate, and at the same time, it can ensure better fine-grained depiction of the BP neural network model.

[0204] As Figure 9 shown, a model training process, step 2: Train a BP neural network based on the test data. Specifically, it can include:

[0205] It should be noted that the present invention does not make theoretical improvements to the BP neural network itself, but proposes unique ideas in aspects such as data processing and model training. Based on the direction of seabed terrain prediction, through a large number of tests and theoretical calculations, the present invention proposes a network topology model with better adaptability. Specifically as follows Figure 4 The BP neural network topology diagram shown.

[0206] As Figure 4, the gravity data is pre - processed through Step 1 to obtain input data such as GA, VGGA, GAFR, GAFI, VGGAFR, and VGGAFI. Distinguishing by time domain and frequency domain, the time - domain data is connected to the input layer (2) (i.e., the first input layer), and the frequency - domain data is connected to the input layer (4) (i.e., the second input layer), and then connected to the corresponding fully - connected layer. The fully - connected layer is mainly used for feature extraction training. The topological key point proposed by the present invention is to separately train and calculate the time domain and frequency domain in the input layer and the hidden layer, and then merge the vectors at the output layer and finally output uniformly (for details, refer to Figure 5 ). The reason for this design: The time - domain data has a strong similarity with the seabed terrain data image, so it can be approximately understood as a linear relationship. While the frequency - domain data mainly reflects the degree of tortuosity of the seabed terrain data, and the images do not have strong similarity, so it can be approximately understood as a non - linear relationship. Through different types of topological designs, implementations, and tests, it is found that the model designed in this way performs better.

[0207] Figure 5 It is the data flow diagram corresponding to the BP neural network. There are 6 inputs including input data GA, VGGA, GAFR, GAFI, VGGAFR, and VGGAFI. The time - domain data and the frequency - domain data are separated both in the input layer and the hidden layer, and unified output is completed through vector splicing at the output layer.

[0208] The topological design proposed by the present invention starts from the linear and non - linear relationships between data. Through theoretical calculations and a large number of tests, the data accuracy and performance of the model output are better.

[0209] As Figure 10 shown, another model training process, Step 2: Train the BP neural network based on test data. It should be noted that the present invention does not make theoretical improvements to the BP neural network itself, but proposes unique ideas in aspects such as data processing and model training. Based on the seabed terrain prediction direction, through a large number of tests and theoretical calculations, the present invention proposes a network topology model with better adaptability. Specifically, as Figure 8 shown in the BP neural network topology diagram.

[0210] As Figure 8, the gravity data is pre - processed through Step 1 to obtain input data in the form of (longitude, latitude, GA, IMF1, IMF2,......, Res, depth). The original data is connected to the input layer (1), and the decomposed modal data is connected to the input layer (2), and then connected to the corresponding fully - connected layer. The fully - connected layer is mainly used for feature extraction training. The topological key point proposed by the present invention is that the original data and the modal data are separately trained and calculated in the input layer and the hidden layer, and the vectors are merged at the output layer, and finally unified output. The reason for such a design: There is a strong similarity between the gravity data GA and the seabed terrain data image, so it can be approximately understood as a linear relationship. The modal data mainly reflects the tortuosity degree of the seabed terrain data, and there is no strong similarity with the image, so it can be approximately understood as a non - linear relationship, and there is an order - dependent relationship between the IMF i and IMF i+1 in the modal data. Through different types of topological design, implementation and testing, it is found that the model designed in this way performs better.

[0211] In summary, the topological design proposed by the present invention starts from the linear and non - linear relationships between data. After theoretical calculation and a large number of tests, the data accuracy and performance of the model output are better.

[0212] In the subsequent steps, the training processes of the two models have the same principle and will not be distinguished and described separately.

[0213] Step 3: Calculate the error based on a single batch of validation data. Specifically, it can include:

[0214] Step 2 is based on all data for training, but the model performance is verified by calculating the batch data error. This is because there is a lot of depth data such as multi - beam and seismic data. If too much data is used for prediction and batch verification is not adopted, the calculation efficiency is too low, resulting in the model training, verification, and output being too slow. Generally, the batch size BatchSize is 512, which can be configured. If it is configured as the number of the maximum depth data, the verification process degenerates into all depth data participating in the verification.

[0215] Different from directly using gravity data to divide the test set and the validation set, the present invention uses depth data obtained from other channels as validation data, which is also one of the core points of the present invention. The format of the validation data is in the form of (longitude, latitude, depth), and the longitude and latitude of the validation data are not necessarily the same as those of the gravity data. For this reason, the present invention uses a depth error evaluation method based on distance.

[0216] Such as Figure 6, GA11 to GA34 represent gravity data, and D1, D2, D3 represent depth data. For the verification data of each (longitude, latitude, depth), according to the longitude and latitude, find the four gravity data points closest to itself in the depth data, and then calculate the distances of the corresponding points. The closer the distance, the greater the influence. It is mainly calculated in three cases.

[0217] Case1: Taking D1 as an example, the gravity data points closest to D1 are GA11, GA12, GA21, and GA22. Use the longitude and latitude to calculate the distances from D1 to these four points, which are l1, l2, l3, and l4 respectively. Then the estimated depth of D1 is:

[0218]

[0219] Among them, Depth 11 represents the wavelength band depth WBD of GA11, and the others are similar. PredictDepth represents the depth predicted based on these four points. If the longitude and latitude of D1 are exactly the same as those of a certain point in the gravity data, then PredictDepth is the wavelength band depth WBD of that point.

[0220] Case2: Taking D2 as an example, the gravity data points closest to D2 are GA22 and GA23. Use the longitude and latitude to calculate the distances from D2 to these two points, which are l5 and l6 respectively. Then the estimated depth of D2 is:

[0221]

[0222] Among them, Depth 22 represents the wavelength band depth WBD of GA22, and the others are similar. PredictDepth represents the depth predicted based on these two points. If the longitude and latitude of D2 are exactly the same as those of a certain point in the gravity data, then PredictDepth is the wavelength band depth WBD of that point.

[0223] Case3: Taking D3 as an example, because the longitude and latitude of D3 are outside the gravity data, D3 does not participate in the calculation. During data preprocessing, data like D3 that exceeds the gravity longitude and latitude range will be filtered out and not participate in the calculation.

[0224] The formula for calculating the distance based on the longitude and latitude of two points is as follows, where lng1 and lat1 represent the longitude and latitude of point 1, lng2 and lat2 represent the longitude and latitude of point 2, asin and sqrt represent the inverse cosine operation and the square root operation respectively, Distance is the finally calculated distance, and 6371 represents the radius of the earth, with the unit of km.

[0225] dlon = lng2 - lng1

[0226] dlat = lat2 - lat1

[0227]

[0228] After traversing and calculating the predicted depths of all verification data one by one, the error depth sequence = predicted depth sequence - true depth sequence. Based on the error depth sequence, RMS (Root Mean Square) is calculated, and this is the error. The present invention adjusts the neural network based on this error, which is one of the core inventive points.

[0229] Step 4: Error backpropagation. Use the error backpropagation algorithm to backpropagate the gradient information of the error loss function. This is a prior art and is directly used here.

[0230] Step 5: Gradient descent to adjust the neural network parameters. Adjust the neural network parameters based on the Adam algorithm. The Adam algorithm, namely Adaptive Moment Estimation, can calculate the adaptive learning rate for each parameter. Compared with other adaptive learning rate algorithms, the Adam algorithm has a faster convergence speed, more effective learning effect, and can correct problems existing in other optimization techniques, such as the disappearance of the learning rate, slow convergence, or large fluctuations in the loss function caused by high-variance parameter updates, etc. This is a prior art and is directly used here.

[0231] Step 6: Complete all batches of verification data. Assume that BatchSize is 512 and the total amount is 5120, then there are a total of 10 batches of data for calculating the error and adjusting the model. If all batches of data are traversed, go to Step 7; otherwise, go to Step 3.

[0232] Step 7: Calculate the overall error based on the verification data. After the batch data training is completed, based on the current model, all depth data participate in calculating the overall error, and the calculation method is the same as that in Step 3.

[0233] Step 8: Meet the minimum error. If the calculated error RMS is less than the minimum MinIterRMS (this can be configured, with a default value of 1), it means that the currently adjusted model is excellent enough and can be directly used. Go to Step 10; otherwise, go to Step 9.

[0234] MinIterRMS has a physical meaning. It represents the average error depth. Assume that the depth unit involved in the calculation is km and MinIterRMS is 1, then it means that the average error of the output depth matrix is required to be between [-1 km, 1 km].

[0235] Step 9: Reach the maximum number of iterations. If the number of iterations exceeds the maximum limit MaxIterCnt, go to Step 10, indicating that the best seabed terrain depth prediction model has been found; otherwise, go to Step 2 to continue training and adjusting the network model.

[0236] Step 10: Obtain the final seabed terrain depth prediction model. When reaching this step, the final depth prediction model has been obtained and can be applied to depth prediction.

[0237] Exemplarily based on Figure 1 the implementation environment shown, another aspect of the embodiments of the present invention provides a seabed terrain prediction method. Taking the application of this model training method in the server 101 as an example for illustration, it can be understood that this model training method can also be applied to the terminal 102.

[0238] Referring to Figure 11 , Figure 11 is a flowchart of the seabed terrain prediction method applied to a server provided by the embodiments of the present invention. The execution subject of this seabed terrain prediction method can be any of the foregoing computer devices. Referring to Figure 11 , this method includes the following steps:

[0239] T100. Obtain the gravity data at each position in the seabed area to be recognized;

[0240] T200. Through fast Fourier transform, perform data dimension expansion on the gravity data, and then organize and obtain the input data;

[0241] T300. Input the input data into the seabed terrain depth prediction model to predict the seabed terrain depth at each position in the seabed area to be recognized; wherein, the seabed terrain depth prediction model is trained by the model training method in the previous steps S100 to S500;

[0242] T400. According to the predicted depth at each position in the seabed area to be recognized, obtain the seabed terrain of the seabed area to be recognized.

[0243] Exemplarily, in some specific embodiments, when the seabed terrain depth prediction model is obtained, the seabed terrain can be directly predicted. As Figure 12 shown, it is mainly divided into two steps:

[0244] Step 1: Preprocess the gravity data to obtain the input data. Based on FFT decomposition of the depth data, the process refers to the previous data preprocessing process, and finally data in the form of (longitude, latitude, GA, VGGA, GAFR, GAFI, VGGAFR, VGGAFI, wavelength band depth WBD) is obtained, and these data are the output data.

[0245] Step 2: Input data into the model to predict the seabed topography. Directly input the input data into the model, and then the predicted depths at different longitudes and latitudes can be obtained, and finally the entire seabed topography can be obtained.

[0246] Refer to Figure 13 , Figure 13 which is a flowchart of another seabed topography prediction method applied to a server provided by an embodiment of the present invention. The execution subject of this seabed topography prediction method can be any of the foregoing computer devices. Refer to Figure 13 , and this method includes the following steps:

[0247] F100. Obtain the gravity data at each position in the seabed area to be recognized;

[0248] F200. Through two-dimensional empirical mode decomposition, perform data dimension augmentation on the gravity data, and then organize and obtain the input data;

[0249] F300. Input the input data into the seabed topography depth prediction model to predict the seabed topography depth at each position in the seabed area to be recognized; wherein, the seabed topography depth prediction model is trained by the model training method in the previous steps U100 to U500;

[0250] F400. Obtain the seabed topography of the seabed area to be recognized according to the predicted depths at each position in the seabed area to be recognized.

[0251] Exemplarily, in some specific embodiments, when the seabed topography depth prediction model is obtained, the seabed topography can be directly predicted. As Figure 14 shown, it is mainly divided into two steps:

[0252] Step 1: Preprocess the gravity data to obtain the input data. Based on BEMD decomposition of the gravity data, organize the gravity anomaly GA data into a two-dimensional matrix, decompose the gravity data based on the BEMD method, and finally obtain multiple IMFs and Res residuals. Finally, data in the form of (longitude, latitude, GA, IMF1, IMF2,..., IMF N , Res, depth) is obtained, and these data are the output data.

[0253] Step 2: Input data into the model to predict the seabed topography. Directly input the input data into the model, and then the predicted depths at different longitudes and latitudes can be obtained, and finally the entire seabed topography can be obtained.

[0254] In summary, the overall process of the present invention for model training and seabed topography prediction based on a BP neural network is as follows:

[0255] 1) Preprocess the gravity data and depth data, and separate information in different dimensions such as time domain and frequency domain through technologies such as FFT (Fast Fourier Transform), and obtain test data and verification data after processing.

[0256] 2) Train a custom BP neural network based on the test data, use a custom error evaluation system, then update the model through batch processing, and obtain the final model through multiple iterations.

[0257] 3) Preprocess the gravity data and then input it into the final model to obtain the seabed terrain data.

[0258] Among them, the overall process of the present invention for model training and seabed terrain prediction by combining BEMD with a neural network is as follows:

[0259] 1) Preprocess the gravity data and depth data, separate information in different modes through BEMD technology, and obtain test data and calibration data after processing.

[0260] 2) Train a custom BP neural network based on the test data, use a custom error evaluation system, then update the model through batch processing, and obtain the final model through multiple iterations.

[0261] 3) Preprocess the gravity data and then input it into the final model to obtain the seabed terrain data.

[0262] With the advent of the big data era and the rise of artificial intelligence, machine learning methods such as decision trees, support vector machines, and artificial neural networks have been introduced into seabed terrain inversion. The present invention uses a BP neural network for seabed terrain prediction. Compared with existing methods, the neural network has a small computational amount. Most importantly, it has a good ability to describe non-linear data relationships.

[0263] Compared with the prior art, the main advantages of the present invention are as follows:

[0264] 1) Data preprocessing technology. Through a large number of tests, it is found that the model fineness obtained by relying solely on gravity data for neural network training is insufficient, and the error of the predicted seabed terrain is large. To solve this problem, the present invention proposes to describe the neural network model from multiple dimensions such as time domain and frequency domain. Experiments prove that the model's ability to finely describe the terrain has been significantly improved.

[0265] 2) BP neural network structure. Through a large number of tests, starting from the linear and non-linear relationships of the data, the present invention proposes a better topological structure of the neural network, which has good potential in the field of seabed terrain prediction, ensuring both the efficiency of model training and the fineness of the model's terrain description.

[0266] 3) Design of neural network loss function. The existing approach is to divide a single data source into test data and validation data, and design the loss function based on the test set. In this application, data from other channels is introduced to design the loss function. Specifically, the present invention introduces multi-beam depth data or seismic depth data, and optimizes the depth model based on these depth data, ensuring the fineness of the terrain depicted by the model.

[0267] Based on the disadvantages of the prior art, the improved solution can at least have the following beneficial effects:

[0268] 1) Small overall computational amount and model reusability. After training a model using a large amount of gravity data, the model can be directly reused without a large computational amount later, and is suitable for predicting the seabed terrain of large areas.

[0269] 2) No need for parameter tuning. The neural network does not have dozens of parameters such as high-pass filtering and low-pass filtering, and belongs to a black-box model without interpretability, so there is no need for parameter tuning.

[0270] 3) More refined terrain depiction. The neural network has good non-linear simulation ability and can finely depict the terrain well.

[0271] On the other hand, as Figure 15 shown, an embodiment of the present invention provides a model training device 1000, including: a first module 1010 for obtaining an initial training data set; obtaining validation data; the initial training data set includes matching gravity data and wavelength band depth data; the validation data includes multi-beam depth data and seismic depth data; a second module 1020 for performing data dimension expansion on the gravity data in the initial training data set through fast Fourier transform, and then sorting to obtain test data; a third module 1030 for using the wavelength band depth data as an output label, inputting the test data into a preset neural network for network training to obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer, and an output layer, and the hidden layer includes multiple fully connected layers; a fourth module 1040 for performing distance-based depth error evaluation processing based on a single batch of validation data in combination with the depth prediction result to obtain a depth error; a fifth module 1050 for performing backpropagation on the depth error and correcting the parameters of the preset neural network in combination with gradient descent until a preset condition is met to obtain a seabed terrain depth prediction model; the preset condition includes one or a combination of using the validation data reaching a preset number of batches, the depth error being less than a preset threshold, and the network training reaching a preset number of iterations.

[0272] It should be noted that as Figure 15As shown, the model training device 1000 provided by the embodiments of the present invention is also applicable to another model training method provided by the embodiments of the present invention. Specifically: The first module 1010 is configured to obtain an initial training data set; obtain verification data; the initial training data set includes matching gravity data and depth data; the verification data includes multi-beam depth data and seismic depth data; the second module 1020 is configured to perform data dimension augmentation on the gravity data in the initial training data set through two-dimensional empirical mode decomposition, and then organize and obtain test data; the third module 1030 is configured to use the depth data as an output label, input the test data into a preset neural network for network training to obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer, and an output layer, and the hidden layer includes multiple fully connected layers; the fourth module 1040 is configured to perform distance-based depth error evaluation processing based on a single batch of verification data in combination with the depth prediction result to obtain a depth error; the fifth module 1050 is configured to perform backpropagation on the depth error and correct the parameters of the preset neural network in combination with gradient descent until a preset condition is met to obtain a seabed terrain depth prediction model; the preset condition includes one or more combinations of the verification data used reaching a preset number of batches, the depth error being less than a preset threshold, and the network training reaching a preset number of iterations.

[0273] In some embodiments, optionally, the model training device may further include at least one of the following: a sixth module, configured to filter the verification data when no gravity data point is obtained at the position corresponding to the verification data; a seventh module, configured to use the predicted depth obtained from the gravity data corresponding to a gravity data point as the estimated depth of the position corresponding to the verification data when the position corresponding to the verification data coincides with a certain gravity data point.

[0274] On the other hand, the embodiments of the present invention provide a seabed terrain prediction device, including: an eighth module, configured to obtain gravity data at each position in the seabed area to be recognized; a ninth module, configured to perform data dimension augmentation on the gravity data through fast Fourier transform, and then organize and obtain input data; a tenth module, configured to input the input data into the seabed terrain depth prediction model to perform seabed terrain depth prediction to obtain the predicted depth at each position in the seabed area to be recognized; wherein, the seabed terrain depth prediction model is trained by the previous model training method; an eleventh module, configured to obtain the seabed terrain of the seabed area to be recognized according to the predicted depth at each position in the seabed area to be recognized.

[0275] It should be noted that the seabed terrain prediction device provided in the embodiments of the present invention is equally applicable to another seabed terrain prediction method provided in the embodiments of the present invention. Specifically: The eighth module is used to obtain the gravity data at each position in the seabed area to be identified; the ninth module is used to perform data dimension augmentation on the gravity data through two-dimensional empirical mode decomposition, and then organize and obtain the input data; the tenth module is used to input the input data into the seabed terrain depth prediction model to predict the seabed terrain depth at each position in the seabed area to be identified, where the seabed terrain depth prediction model is trained by the previous model training method; the eleventh module is used to obtain the seabed terrain of the seabed area to be identified according to the predicted depths at each position in the seabed area to be identified.

[0276] The content of the method embodiments of the present invention is applicable to the device embodiments. The functions specifically implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those of the above methods.

[0277] On the other hand, the embodiments of the present invention further provide an electronic device, which includes at least one processor and at least one memory for storing at least one program, and the processor executes the program to implement any one of the previous model training methods or seabed terrain prediction methods; taking one processor and one memory as an example.

[0278] The processor and the memory can be connected through a bus or other means.

[0279] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the device through a network. Examples of the above networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0280] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0281] The content of the method embodiments of the present invention is applicable to the electronic device embodiments. The functions specifically implemented by the electronic device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those of the above methods.

[0282] Another aspect of the embodiments of the present invention further provides a computer-readable storage medium, which stores a program, and the program is executed by a processor to implement any of the previous model training methods or seabed terrain prediction methods.

[0283] The content of the method embodiments of the present invention is applicable to the embodiments of this computer-readable storage medium. The functions specifically implemented by the embodiments of this computer-readable storage medium are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those of the above methods.

[0284] The embodiments of the present invention also disclose a computer program product or a computer program, which includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to enable the computer device to execute the previous methods.

[0285] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the claims and their equivalents.

[0286] The above has specifically described the preferred embodiments of the present invention, but the present invention is not limited to the embodiments. Those skilled in the art can also make various equivalent deformations or substitutions without violating the spirit of the present invention, and these equivalent deformations or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. A model training method, characterized in that, Including: Obtaining an initial training data set; obtaining validation data; The initial training data set includes matching gravity data and depth data; The validation data includes multi-beam depth data and seismic depth data; Through two-dimensional empirical mode decomposition, the data dimension of the gravity data in the initial training data set is amplified, and then the test data is sorted out; Among them, the process of amplifying the data dimension of the gravity data in the initial training data set through two-dimensional empirical mode decomposition and then sorting out the test data includes: Organize the gravity data into a two-dimensional matrix; Project the two-dimensional matrix into the two-dimensional frequency domain through two-dimensional empirical mode decomposition, and then decompose the two-dimensional matrix into multiple independent modes to obtain a residual component and several intrinsic mode function components; Sort out the test data according to the intrinsic mode function components, the residual component and the gravity data; Using the depth data as the output label, input the test data into a preset neural network for network training to obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer and an output layer, and the hidden layer includes multiple fully connected layers; Among them, the test data includes the gravity data and the residual component and several intrinsic mode function components processed based on the gravity data; the input layer includes a first input layer and a second input layer, and the hidden layer includes a first hidden layer and a second hidden layer; the process of inputting the test data into a preset neural network for network training to obtain a depth prediction result includes: Input the gravity data into the first input layer, and then perform a first multi-layer fully connected process on the result of the first input layer through the first hidden layer to obtain a first feature; Input the intrinsic mode function components and the residual component into the second input layer, and then perform a second multi-layer fully connected process on the result of the second input layer through the second hidden layer to obtain a second feature; Obtain the depth prediction result through the output layer according to the first feature and the second feature; Based on a single batch of the validation data, combine the depth prediction result to perform a distance-based depth error evaluation process to obtain a depth error; Backpropagate the depth error and combine gradient descent to correct the parameters of the preset neural network until a preset condition is met to obtain a seabed terrain depth prediction model; the preset condition includes one or more combinations of the validation data used reaching a preset batch number, the depth error being less than a preset threshold, and the network training reaching a preset number of iterations; Among them, the process of backpropagating the depth error and combining gradient descent to correct the parameters of the preset neural network includes: Using the error backpropagation algorithm, backpropagate the gradient information of the depth error in the preset neural network, and implement gradient descent to correct the parameters of the preset neural network based on the adaptive moment estimation method.

2. The model training method according to claim 1, wherein The process of performing a distance-based depth error evaluation process based on a single batch of the validation data and combining the depth prediction result to obtain a depth error includes: Based on the validation data of a single batch, obtain a plurality of gravity data points through the positions corresponding to the validation data; the gravity data points represent the positions corresponding to the gravity data; and then determine the distance data between the positions corresponding to the validation data and each of the gravity data points; According to the gravity data corresponding to the gravity data points, obtain the corresponding predicted depth from the depth prediction results; According to the distance data and the predicted depths corresponding to each of the gravity data points, obtain the estimated depth of the position corresponding to the validation data; and then combine the validation data of a single batch to obtain an error depth sequence; Based on the error depth sequence, calculate the root mean square value to obtain the depth error.

3. The model training method according to claim 2, wherein The method further includes at least one of the following: When the gravity data points are not obtained through the positions corresponding to the validation data, filter the validation data; When the position corresponding to the validation data coincides with a certain gravity data point, use the predicted depth obtained from the gravity data corresponding to the gravity data point as the estimated depth of the position corresponding to the validation data.

4. A model training device, characterized in that, It includes: A first module for obtaining an initial training data set; obtaining validation data; The initial training data set includes matching gravity data and depth data; The validation data includes multi-beam depth data and seismic depth data; A second module for performing data dimension augmentation on the gravity data in the initial training data set through two-dimensional empirical mode decomposition, and then organizing and obtaining test data; Among them, the performing data dimension augmentation on the gravity data in the initial training data set through two-dimensional empirical mode decomposition, and then organizing and obtaining test data includes: Organize the gravity data into a two-dimensional matrix; Project the two-dimensional matrix into the two-dimensional frequency domain through two-dimensional empirical mode decomposition, and then decompose the two-dimensional matrix into multiple independent modes to obtain a residual component and several intrinsic mode function components; Organize and obtain test data according to the intrinsic mode function components, the residual component and the gravity data; A third module for using the depth data as an output label, inputting the test data into a preset neural network for network training to obtain a depth prediction result; the preset neural network includes an input layer, a hidden layer and an output layer, and the hidden layer includes multiple fully connected layers; Among them, the test data includes the gravity data and the residual component and several intrinsic mode function components processed based on the gravity data; the input layer includes a first input layer and a second input layer, and the hidden layer includes a first hidden layer and a second hidden layer; the inputting the test data into a preset neural network for network training to obtain a depth prediction result includes: Input the gravity data into the first input layer, and then perform a first multi-layer fully connected process on the result of the first input layer through the first hidden layer to obtain a first feature; Input the intrinsic mode function components and the residual component into the second input layer, and then perform a second multi-layer fully connected process on the result of the second input layer through the second hidden layer to obtain a second feature; The output layer obtains a depth prediction result according to the first feature and the second feature; A fourth module, configured to perform distance-based depth error evaluation processing based on the depth prediction result by combining the validation data of a single batch, so as to obtain a depth error; A fifth module, configured to perform backpropagation on the depth error, and combine gradient descent to correct the parameters of the preset neural network until a preset condition is met, so as to obtain a seabed terrain depth prediction model; the preset condition includes one or a combination of more of the following: the validation data used reaches a preset number of batches, the depth error is less than a preset threshold, and the network training reaches a preset number of iterations; Wherein, the performing backpropagation on the depth error and combining gradient descent to correct the parameters of the preset neural network includes: Using the error backpropagation algorithm, performing backpropagation on the gradient information of the depth error in the preset neural network, and implementing gradient descent based on the adaptive moment estimation method to correct the parameters of the preset neural network.

5. A method for predicting submarine topography, characterized in that, Including: Obtaining gravity data at each position in the seabed area to be recognized; Through two-dimensional empirical mode decomposition, performing data dimension augmentation on the gravity data, and then sorting to obtain input data; Inputting the input data into the seabed terrain depth prediction model to perform seabed terrain depth prediction to obtain the predicted depth at each position in the seabed area to be recognized; wherein, the seabed terrain depth prediction model is trained by the model training method according to any one of claims 1 to 3; According to the predicted depth at each position in the seabed area to be recognized, obtaining the seabed terrain of the seabed area to be recognized.

6. An electronic device, characterized in that, Including a processor and a memory; The memory is used to store a program; The processor executes the program to implement the method according to any one of claims 1 to 3 or claim 5.

7. A computer storage medium storing a program executable by a processor, characterized in that The program executable by the processor, when executed by the processor, is used to implement the method according to any one of claims 1 to 3 or claim 5.

Images