Artificial intelligence-based seawater intrusion prediction model training method and device

Abstract

The invention relates to the technical field of seawater intrusion prediction, in particular to a seawater intrusion prediction model training method and device based on artificial intelligence, and the method specifically comprises the following steps: collecting seawater intrusion related data to construct a local training data set for a local seawater intrusion prediction feature extraction module; parameters of a local seawater intrusion prediction feature extraction module and a central seawater prediction feature extraction module of each client are interacted based on a federated learning architecture, so that parameter updating of the global seawater intrusion prediction feature extraction module is realized. Meanwhile, the accuracy of a local seawater intrusion prediction feature extraction module is improved by using a dynamic balance optimization algorithm, finally, features captured by the local seawater intrusion prediction feature extraction module are analyzed through a seawater intrusion prediction classifier module, and the seawater intrusion prediction classifier module adopts a high-order partial derivative constrained extreme learning machine. According to the invention, the stability and accuracy of the model in processing the complex marine environment data can be improved through the dynamic balance optimization algorithm and the extreme learning machine.

Description

Technical Field

[0001] The present invention relates to the technical field of seawater intrusion prediction, and in particular to a seawater intrusion prediction model training method and device based on artificial intelligence. Background Art

[0002] Seawater intrusion is a common environmental problem in coastal areas, which has a serious impact on the local ecological environment, agricultural water use and residents' lives. With climate change and the intensification of human activities, the problem of seawater intrusion has become more complicated. Its occurrence is affected by many environmental factors, including water temperature, salinity, pressure, flow rate, geological characteristics and historical intrusion events. Accurately predicting the occurrence and level of seawater intrusion is the key to managing and mitigating its impact.

[0003] The prior art has the following objective disadvantages: (1) In existing seawater intrusion prediction and related tasks, the centralized model in the technology requires all data to be aggregated to a central server, which is prone to data privacy leakage and limits its application in data-sensitive fields; (2) Traditional feature extraction models rely on the gradient descent method during training, which is prone to gradient vanishing, gradient explosion, or falling into the local optimal solution, resulting in insufficient adaptability of seawater intrusion prediction models to complex environmental data; (3) Traditional seawater intrusion prediction classifiers have insufficient performance in processing complex nonlinear feature relationships and are unable to capture the implicit patterns in seawater intrusion data, thus affecting the accuracy of prediction. (4) Traditional optimization algorithms lack robustness to seawater intrusion data noise and outliers, and are unable to dynamically adapt to the importance of different weights, resulting in an inefficient training process and being easily affected by unbalanced data distribution; (5) When faced with the problem of imbalanced seawater intrusion data, traditional loss functions cannot effectively improve the classification performance of minority class samples, resulting in inaccurate predictions of mild, moderate or severe intrusion events.

[0004] Therefore, the present invention proposes a seawater intrusion prediction model training method and device based on artificial intelligence to solve the above problems. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention develops a seawater intrusion prediction model training method and device based on artificial intelligence. The present invention can improve the stability and accuracy of the seawater intrusion prediction model when processing complex marine environment data.

[0006] On the one hand, the technical solution of the present invention to solve the technical problem is a seawater intrusion prediction model training method based on artificial intelligence, which is as follows: Collect seawater intrusion related data, build a local training data set, extract features through the local seawater intrusion prediction feature extraction module in the client, then interact the parameters of the local seawater intrusion prediction feature extraction module of each client with the central seawater prediction feature extraction module, and then the central seawater prediction feature extraction module aggregates the updated parameters of each local seawater intrusion prediction feature extraction module to achieve global seawater intrusion prediction feature extraction module parameter update, and use the dynamic balance optimization algorithm to improve the accuracy of the local seawater intrusion prediction feature extraction module, finally, the features captured by the local seawater intrusion prediction feature extraction module are analyzed by the seawater intrusion prediction classifier module, which adopts the extreme learning machine with high-order partial derivative constraints.

[0007] The client's local training data set is as follows: The local training data set of the client is obtained by collecting and annotating data related to seawater intrusion; The relevant data of seawater intrusion include the geographical distribution, time and intensity information of historical seawater intrusion events, the chemical composition, temperature, salinity, pressure, flow rate, topographic parameters of seawater; The sources of relevant data on seawater intrusion include satellite remote sensing data, marine meteorological station data, hydrological monitoring data, geological and hydrogeological data, and historical event data; The data collection method of seawater intrusion is realized through sensors, remote sensing technology, marine meteorological observation stations, unmanned surface vehicles, and seabed sensors. The data is transmitted, processed and verified during the collection process, and finally stored in a unified structured data format, specifically a structured CSV format; The collected data is manually labeled, and the labeled categories include no invasion, slight invasion, moderate invasion, and severe invasion; No invasion means that under the current environmental conditions, the seawater has not invaded, marked as 0; slight invasion means that under the current environmental conditions, the seawater has invaded slightly, marked as 1; moderate invasion means that under the current environmental conditions, the seawater has invaded slightly or moderately, marked as 1; severe invasion means that under the current environmental conditions, the seawater has invaded slightly or moderately, marked as 1.

[0008] The global seawater intrusion prediction feature extraction module parameter update process is as follows: Each client is trained according to its own local training data set through its own local seawater intrusion prediction feature extraction module. Based on the federated learning architecture, the parameters obtained after the training of each local seawater intrusion prediction feature extraction module interact with the central seawater intrusion prediction feature extraction module. The central seawater intrusion prediction feature extraction module receives the parameters of each local seawater intrusion prediction feature extraction module and sends the received parameters to each local seawater intrusion prediction feature extraction module. The central seawater intrusion prediction feature extraction module summarizes the parameters of all local seawater intrusion prediction feature extraction modules and then sends the summarized parameters to each local seawater intrusion prediction feature extraction module. Each local seawater intrusion prediction feature extraction module updates its parameters according to the received information, and then shares the updated parameters through the central seawater intrusion prediction feature extraction module again to achieve the re-update of the parameters of each local seawater intrusion prediction feature extraction module. After multiple iterations, the global seawater intrusion prediction feature extraction module parameter update is achieved.

[0009] Dynamic balance optimization algorithm: The local seawater intrusion prediction feature extraction module includes a five-layer fully connected neural network. The parameters of the neural network training process are optimized through a dynamic balance optimization algorithm. The dynamic balance optimization algorithm specifically includes an isolated system energy conservation method and a dynamic balance concept. The specific optimization process is as follows; (1) The weights and biases of the neural network in the local seawater intrusion prediction feature extraction module are randomly initialized. The initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix; (2) Set the total energy of the system, which is the process of optimizing the parameters of the neural network in the local seawater intrusion prediction feature extraction module. The total energy of the system is expressed as: , in, represents the total energy of the system, It is set to a random value at initialization. represents the number of parameters in the neural network, express The index of Indicates the assignment to the neural network The energy of the parameters; The energy assigned to the parameters in the neural network is calculated by the sensitivity of the parameters. The calculation formula is as follows: , in, Indicates the neural network The sensitivity of the parameters, Indicates the neural network The sensitivity of the parameters, Indicates the neural network The number of neurons in the layer where the parameter is located, Indicates the number of layers of the neural network; Then according to the total energy of the system The energy assigned to the parameter is calculated by the sensitivity calculation of the parameter, and the calculation formula is as follows: , in, Indicates the assignment to the neural network The energy of the parameters, express Another index of Indicates the neural network The sensitivity of the parameters, Indicates lifting network The sum of the sensitivities of the parameters; The above process is iterated in each layer of the neural network, and the data propagation mode during the iteration is forward propagation; (4) The energy of the parameters of the seawater intrusion feature extraction neural network is dynamically adjusted according to the error feedback. The error feedback is the energy change of the parameters. The calculation formula is as follows: , , in, Indicates the neural network The energy change of each parameter is represents the learning rate parameter of the neural network, represents the symbol of partial derivative, represents the trend function, Indicates the neural network The weight of the parameters, Indicates the neural network The weight of the parameters, Indicates the neural network The weight of each parameter changes; (5) Update the energy of the parameter according to the energy change of the parameter. The calculation formula is as follows: , in, Indicates the neural network The energy after the parameter update; (6) The updated value is calculated using the energy transfer function. The calculation formula is as follows: , , in, represents the conversion rate hyperparameter, represents the adjustment factor considering the impact of historical energy, represents the symbolic function, represents the energy transfer function, represents the number of iterations, express The index of Indicates the neural network The iteration The energy of the parameters; (7) Based on the difference between the output of the neural network, i.e., the final output of the local seawater intrusion prediction feature extraction module, and the target output, the energy allocation of each parameter is recalculated and adjusted. The calculation formula is as follows: , , in, Indicates the neural network The energy of parameter adjustment, represents the adjustment intensity of parameter energy, represents the difference between the output of the neural network and the target output, Indicates the neural network The updated gradient of the parameter energy, A function that represents the relationship between error and gradient; The adjusted energy is calculated as follows: , in, Indicates the neural network The final energy after parameter adjustment; (8) Fine-tune the weights of the parameters in the neural network according to the energy of the parameters. The calculation formula is as follows: , in, represents the fine-tuning coefficient, Indicates the neural network The weights after fine-tuning of parameters; (9) Perform multiple iterations and set iteration conditions until the preset iteration conditions are met and the iteration stops.

[0010] Seawater intrusion prediction classifier module: The seawater intrusion prediction classifier module receives the output from the local seawater intrusion prediction feature extraction module, and classifies the results according to the output of the local seawater intrusion prediction feature extraction module, thereby realizing the prediction of the seawater intrusion level; The seawater intrusion prediction classifier module adopts an extreme learning machine with high-order partial derivative constraints. It captures the complex nonlinear relationship of the seawater intrusion data after feature extraction by adhering to the partial derivative constraints, and optimizes the weight of the output layer in the seawater intrusion prediction classifier module by combining the least squares method with the gradient descent method, thereby improving the accuracy of seawater intrusion classification prediction.

[0011] Extreme Learning Machine: (1) Initialize the parameters of the extreme learning machine and define the output representation of the hidden layer of the extreme learning machine. The calculation formula is as follows: , , in, represents the output matrix of the hidden layer of the extreme learning machine, represents the extreme learning machine Activation function, Represents the input sample matrix of the extreme learning machine, with a size of , represents the number of samples, represents the number of features, represents the weight matrix input to the hidden layer, represents the standard deviation of the weight initialization of the extreme learning machine, represents a normal distribution with a mean of 0 and a variance of 1. represents the bias term of the hidden layer of the extreme learning machine, and its initialization value is zero; (2) In the training process of the extreme learning machine model, a high-order partial derivative constraint strategy is adopted, and the update method of the module parameters is adjusted to enhance the learning ability of the complex nonlinear relationship of the seawater intrusion data after feature extraction, so that the module can capture more complex associations between features. The high-order loss function calculation formula of the extreme learning machine is as follows: , in, represents the high-order loss function of the extreme learning machine, Indicates the input The true labels of samples, Indicates the input The hidden layer output of the extreme learning machine corresponding to the sample, Indicates The hidden layer output of the extreme learning machine corresponding to the sample, Indicates The hidden layer output of the extreme learning machine corresponding to the sample, represents the number of input samples, , and express Three different indexes, represents the tuning parameter of the higher-order partial derivative constraint, represents the tuning parameter of the third-order partial derivative constraint, represents the symbol of partial derivative, represents the second-order partial derivative, represents the third-order partial derivative, represents the extreme learning machine An input sample matrix, represents the extreme learning machine An input sample matrix; (3) The dynamic learning rate adjustment strategy makes the gradient of the extreme learning machine parameter update more flexible. The weight update calculation formula of the extreme learning machine is as follows: , , , , in, represents the weight update amount of the extreme learning machine, represents the learning rate of the extreme learning machine, represents the total loss function of the extreme learning machine, represents the balanced loss function of the extreme learning machine, represents the first regularization coefficient of the extreme learning machine, represents the first elements, represents the extreme learning machine The weight sparsification strength of the elements, represents the number of elements in the hidden layer of the extreme learning machine, represents the extreme learning machine The absolute value of the element weight, represents the influencing factor of the balance loss function, An adjustment factor indicating that the placement frequency is too low, Indicates the frequency of occurrence, express Activation function, Indicates the input The hidden layer output of the extreme learning machine corresponding to the sample, represents the hyperparameter that controls the variation of coefficientized strength, Represents the weight decay exponent for adjusting the sparsification strength; Then the learning rate of the extreme learning machine is adjusted by dynamic adjustment. The calculation formula is as follows: , in, represents the learning rate of the adjusted extreme learning machine, represents the learning rate adjustment factor of the extreme learning machine, Indicates the gradient change of the loss function of the extreme learning machine; (4) The output layer weights of the extreme learning machine are optimized by combining the least squares method with the gradient descent method, so that the prediction results are closer to the true value. The calculation formula is as follows: , in, represents the final weight matrix of the output layer of the extreme learning machine, which is used for the next iteration. represents the second regularization coefficient of the extreme learning machine, represents transpose, represents the identity matrix; (5) Perform multiple iterations and set iteration conditions until the preset iteration conditions are met and the iteration stops.

[0012] On the other hand, the present invention also provides an artificial intelligence-based seawater intrusion prediction model training device, including a seawater intrusion prediction model, and executing an artificial intelligence-based seawater intrusion prediction model training method. The structure of the seawater intrusion prediction model is multiple clients and a server. The clients are each provided with a local seawater intrusion prediction feature extraction module and a seawater intrusion prediction classifier module. The server is provided with a central seawater intrusion prediction feature extraction module. Each client has an independent local training data set. After each local seawater intrusion prediction feature extraction module processes the data in the local training data set, the final classification result is uploaded to the server.

[0013] The effects provided in the summary of the invention are only the effects of the embodiments, rather than all the effects of the invention. The above technical solution has the following advantages or beneficial effects: By exchanging parameters between the local seawater intrusion prediction feature extraction module and the central seawater intrusion prediction feature extraction module, the original data is not shared between clients or with the central server during the training process. Instead, only the seawater intrusion prediction feature extraction model parameters are exchanged for updating, which can achieve data privacy protection and break the limitations of applications in data-sensitive fields. In the local seawater intrusion prediction feature extraction module, the dynamic balance algorithm is used. Drawing on the concepts of energy conservation and dynamic balance of isolated systems in physics, the natural distribution and redistribution mechanism of energy in the system is used to optimize the parameters of the neural network. By simulating the dynamic process of energy transfer in an isolated system, it is not necessary to rely on traditional gradient information to adjust weights and biases, which can enhance the efficiency and robustness of neural network parameter optimization, and avoid the phenomenon of gradient disappearance, gradient explosion or falling into the local optimal solution. At the same time, in the process of training the neural network using seawater intrusion data, the gradient dependence is abandoned, and the optimization instability caused by improper step size selection is avoided. In the process of training the neural network using seawater intrusion data, an adaptive energy adjustment mechanism is used to enable the network to dynamically find the optimal solution, and enhance the robustness of the feature extraction of seawater intrusion data, showing higher adaptability to data noise and outliers. By using the extreme learning machine with high-order partial derivative constraints as the seawater intrusion prediction classifier module, the high-order partial derivative constraints can capture the complex nonlinear relationship of the seawater intrusion data after feature extraction, so that the seawater intrusion prediction classifier module can capture the subtle changes and complex relationships of the seawater intrusion data after feature extraction; and by optimizing the output layer weights by combining the least squares method with the gradient descent method, the accuracy of seawater intrusion prediction classification can be improved; the dynamic learning rate adjustment strategy can make the gradient update of the extreme learning machine parameter update more flexible, avoiding the training process from falling into the local optimum; the weight sparsification mechanism is adopted to make the feature extraction and seawater intrusion prediction classification process more efficient, showing higher adaptability to seawater intrusion data noise and minority class samples; through the calculation of the total loss function of the extreme learning machine, it can be avoided that when the seawater intrusion data category is unbalanced, the classification performance of the minority class samples cannot be effectively improved, resulting in inaccurate prediction of mild, moderate or severe invasion events, alleviating the impact of the imbalance of seawater intrusion data categories on model training, and effectively improving the accuracy of seawater intrusion prediction of minority class samples.

[0014] In summary, the present invention improves the stability of the seawater intrusion prediction model when processing complex marine environment data, overcomes the shortcomings of traditional classifiers in processing nonlinear and weakly changing patterns such as seawater intrusion data, and improves the accuracy of seawater intrusion prediction for minority class samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0016] Figure 1 It is a schematic diagram of the structure of the device of the present invention.

[0017] Figure 2This is a comparison chart of the convergence speed between the dynamic balance optimization algorithm in the neural network of the present invention and the gradient descent method.

[0018] Figure 3 This is a performance comparison chart of the dynamic balance optimization algorithm and the gradient descent method in the neural network of the present invention on noise robustness.

[0019] Figure 4 This is a graph showing the relationship between sparsification intensity and prediction accuracy in the extreme learning machine of the present invention. DETAILED DESCRIPTION

[0020] In order to clearly illustrate the technical features of the present invention, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below.

[0021] Example 1 A seawater intrusion prediction model training method based on artificial intelligence is as follows: Collect seawater intrusion related data, build a local training data set, extract features through the local seawater intrusion prediction feature extraction module in the client, then interact the parameters of the local seawater intrusion prediction feature extraction module of each client with the central seawater prediction feature extraction module, and then the central seawater prediction feature extraction module aggregates the updated parameters of each local seawater intrusion prediction feature extraction module to achieve global seawater intrusion prediction feature extraction module parameter update, and use the dynamic balance optimization algorithm to improve the accuracy of the local seawater intrusion prediction feature extraction module, finally, the features captured by the local seawater intrusion prediction feature extraction module are analyzed by the seawater intrusion prediction classifier module, which adopts the extreme learning machine with high-order partial derivative constraints.

[0022] The client's local training data set is as follows: The local training data set of the client is obtained by collecting and annotating data related to seawater intrusion; The relevant data of seawater intrusion include the geographical distribution, time and intensity information of historical seawater intrusion events, the chemical composition, temperature, salinity, pressure, flow rate, topographic parameters of seawater; The sources of relevant data on seawater intrusion include satellite remote sensing data, marine meteorological station data, hydrological monitoring data, geological and hydrogeological data, and historical event data; Satellite remote sensing data: environmental data such as ocean surface temperature and salinity obtained through satellite monitoring; Marine weather station data: meteorological and hydrological data collected by weather stations located along the coastline and offshore areas, including temperature, wind speed, precipitation, etc.; Hydrological monitoring data: water depth, salinity, pH value, dissolved oxygen and other data obtained by water quality monitoring devices deployed in coastal areas; Geological and hydrogeological data: Provides geological survey data of coastlines and groundwater levels, as well as records of historical seawater intrusion; Historical event data: including historical records of seawater intrusion and geographical location annotations; The data collection method of seawater intrusion is realized through sensors, remote sensing technology, marine meteorological observation stations, unmanned surface vehicles, and seabed sensors. The data is transmitted, processed and verified during the collection process, and finally stored in a unified structured data format, specifically a structured CSV format; The collected data is manually labeled, and the labeled categories include no invasion, slight invasion, moderate invasion, and severe invasion; No invasion means that under the current environmental conditions, the seawater has not invaded, marked as 0; slight invasion means that under the current environmental conditions, the seawater has invaded slightly, marked as 1; moderate invasion means that under the current environmental conditions, the seawater has invaded slightly or moderately, marked as 1; severe invasion means that under the current environmental conditions, the seawater has invaded slightly or moderately, marked as 1.

[0023] The global seawater intrusion prediction feature extraction module parameter update process is as follows: Each client is trained according to its own local training data set through its own local seawater intrusion prediction feature extraction module. Based on the federated learning architecture, the parameters obtained after the training of each local seawater intrusion prediction feature extraction module interact with the central seawater intrusion prediction feature extraction module. The central seawater intrusion prediction feature extraction module receives the parameters of each local seawater intrusion prediction feature extraction module and sends the received parameters to each local seawater intrusion prediction feature extraction module. The central seawater intrusion prediction feature extraction module summarizes the parameters of all local seawater intrusion prediction feature extraction modules and then sends the summarized parameters to each local seawater intrusion prediction feature extraction module. Each local seawater intrusion prediction feature extraction module updates its parameters according to the received information, and then shares the updated parameters through the central seawater intrusion prediction feature extraction module again to achieve the re-update of the parameters of each local seawater intrusion prediction feature extraction module. After multiple iterations, the global seawater intrusion prediction feature extraction module parameter update is achieved.

[0024] Dynamic balance optimization algorithm: The local seawater intrusion prediction feature extraction module includes a five-layer fully connected neural network. The parameters of the neural network training process are optimized through a dynamic balance optimization algorithm. The dynamic balance optimization algorithm specifically includes an isolated system energy conservation method and a dynamic balance concept. The specific optimization process is as follows; (1) The weights and biases of the neural network in the local seawater intrusion prediction feature extraction module are randomly initialized. The initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix; (2) Set the total energy of the system, which is the process of optimizing the parameters of the neural network in the local seawater intrusion prediction feature extraction module. The total energy of the system is expressed as: , in, represents the total energy of the system, It is set to a random value at initialization. represents the number of parameters in the neural network, express The index of Indicates the assignment to the neural network The energy of the parameters; The energy assigned to the parameters in the neural network is calculated by the sensitivity of the parameters. The calculation formula is as follows: , in, Indicates the neural network The sensitivity of the parameters, Indicates the neural network The sensitivity of the parameters, Indicates the neural network The number of neurons in the layer where the parameter is located, Indicates the number of layers of the neural network; Then according to the total energy of the system The energy assigned to the parameter is calculated by the sensitivity calculation of the parameter, and the calculation formula is as follows: , in, Indicates the assignment to the neural network The energy of the parameters, express Another index of Indicates the neural network The sensitivity of the parameters, Indicates lifting network The sum of the sensitivities of the parameters; The above process is iterated in each layer of the neural network, and the data propagation mode during the iteration is forward propagation; (4) The energy of the parameters of the seawater intrusion feature extraction neural network is dynamically adjusted according to the error feedback. The error feedback is the energy change of the parameters. The calculation formula is as follows: , , in, Indicates the neural network The energy change of each parameter is represents the learning rate parameter of the neural network, represents the symbol of partial derivative, represents the trend function, Indicates the neural network The weight of the parameters, Indicates the neural network The weight of the parameters, Indicates the neural network The weight of each parameter changes; (5) Update the energy of the parameter according to the energy change of the parameter. The calculation formula is as follows: , in, Indicates the neural network The energy after the parameter update; (6) The updated value is calculated using the energy transfer function. The calculation formula is as follows: , , in, represents the conversion rate hyperparameter, set , represents the adjustment factor considering the impact of historical energy, represents the symbolic function, represents the energy transfer function, represents the number of iterations, express The index of Indicates the neural network The iteration The energy of the parameters; (7) Based on the difference between the output of the neural network, i.e., the final output of the local seawater intrusion prediction feature extraction module, and the target output, the energy allocation of each parameter is recalculated and adjusted. The calculation formula is as follows: , , in, Indicates the neural network The energy of parameter adjustment, represents the adjustment intensity of parameter energy, represents the difference between the output of the neural network and the target output, Indicates the neural network The updated gradient of the parameter energy, A function that represents the relationship between error and gradient; The adjusted energy is calculated as follows: , in, Indicates the neural network The final energy after parameter adjustment; (8) Fine-tune the weights of the parameters in the neural network according to the energy of the parameters. The calculation formula is as follows: , in, Indicates the fine-tuning coefficient, set , Indicates the neural network The weights after fine-tuning of parameters; (9) Perform multiple iterations and set iteration conditions until the preset iteration conditions are met and the iteration stops.

[0025] Seawater intrusion prediction classifier module: The seawater intrusion prediction classifier module receives the output from the local seawater intrusion prediction feature extraction module, and classifies the results according to the output of the local seawater intrusion prediction feature extraction module, thereby realizing the prediction of the seawater intrusion level; The seawater intrusion prediction classifier module adopts an extreme learning machine with high-order partial derivative constraints. It captures the complex nonlinear relationship of the seawater intrusion data after feature extraction by adhering to the partial derivative constraints, and optimizes the weight of the output layer in the seawater intrusion prediction classifier module by combining the least squares method with the gradient descent method, thereby improving the accuracy of seawater intrusion classification prediction.

[0026] Extreme Learning Machine: (1) Initialize the parameters of the extreme learning machine and define the output representation of the hidden layer of the extreme learning machine. The calculation formula is as follows: , , in, represents the output matrix of the hidden layer of the extreme learning machine, represents the extreme learning machine Activation function, Represents the input sample matrix of the extreme learning machine, with a size of , represents the number of samples, represents the number of features, represents the weight matrix input to the hidden layer, represents the standard deviation of the weight initialization of the extreme learning machine, represents a normal distribution with a mean of 0 and a variance of 1. represents the bias term of the hidden layer of the extreme learning machine, and its initialization value is zero; (2) In the training process of the extreme learning machine model, a high-order partial derivative constraint strategy is adopted, and the update method of the module parameters is adjusted to enhance the learning ability of the complex nonlinear relationship of the seawater intrusion data after feature extraction, so that the module can capture more complex associations between features. The high-order loss function calculation formula of the extreme learning machine is as follows: , in, represents the high-order loss function of the extreme learning machine, Indicates the input The true labels of samples, Indicates the input The hidden layer output of the extreme learning machine corresponding to the sample, Indicates The hidden layer output of the extreme learning machine corresponding to the sample, Indicates The hidden layer output of the extreme learning machine corresponding to the sample, represents the number of input samples, , and express Three different indexes, Represents the adjustment parameter of the high-order partial derivative constraint, set , Represents the adjustment parameter of the third-order partial derivative constraint, set , represents the symbol of partial derivative, represents the second-order partial derivative, represents the third-order partial derivative, represents the extreme learning machine An input sample matrix, represents the extreme learning machine An input sample matrix; (3) The dynamic learning rate adjustment strategy makes the gradient of the extreme learning machine parameter update more flexible. The weight update calculation formula of the extreme learning machine is as follows: , , , , in, represents the weight update amount of the extreme learning machine, represents the learning rate of the extreme learning machine, represents the total loss function of the extreme learning machine, represents the balanced loss function of the extreme learning machine, represents the first regularization coefficient of the extreme learning machine, set , represents the first elements, represents the extreme learning machine The weight sparsification strength of the elements, represents the number of elements in the hidden layer of the extreme learning machine, represents the extreme learning machine The absolute value of the element weight, Represents the influencing factor of the balance loss function, set , Indicates the adjustment factor for low placement frequency, set , Indicates the frequency of occurrence, express Activation function, Indicates the input The hidden layer output of the extreme learning machine corresponding to the sample, Represents the hyperparameter that controls the change in coefficient strength, set , Represents the weight decay exponent for adjusting the sparsification strength, set ; Then the learning rate of the extreme learning machine is adjusted by dynamic adjustment. The calculation formula is as follows: , in, represents the learning rate of the adjusted extreme learning machine, represents the learning rate adjustment factor of the extreme learning machine, Indicates the gradient change of the loss function of the extreme learning machine; (4) The output layer weights of the extreme learning machine are optimized by combining the least squares method with the gradient descent method, so that the prediction results are closer to the true value. The calculation formula is as follows: , in, represents the final weight matrix of the output layer of the extreme learning machine, which is used for the next iteration. represents the second regularization coefficient of the extreme learning machine, set , represents transpose, represents the identity matrix; (5) Perform multiple iterations and set iteration conditions until the preset iteration conditions are met and the iteration stops.

[0027] Example 2 like Figure 1 As shown, a seawater intrusion prediction model training device based on artificial intelligence includes a seawater intrusion prediction model, and executes a seawater intrusion prediction model training method based on artificial intelligence. The structure of the seawater intrusion prediction model is four clients and one server. The clients are each provided with a local seawater intrusion prediction feature extraction module and a seawater intrusion prediction classifier module. The server is provided with a central seawater intrusion prediction feature extraction module. Each client has an independent local training data set. After each local seawater intrusion prediction feature extraction module processes the centralized data of the local training data set, the final classification result is uploaded to the server. The four clients are denoted as client 0, client 1, client 2, and client 3 respectively.

[0028] Example 3 In order to better demonstrate the technical effect of the present invention, the present invention is applied to practical applications to train the seawater intrusion prediction model, thereby improving the accuracy of the prediction. Figures 2 to 4 As shown in the figure, they are comparison diagram and relationship diagram, respectively. Figure 2 It can be seen that the dynamic balance optimization algorithm in the neural network of the present invention is compared with the gradient descent method in terms of convergence speed. For the same number of iterations, the error value of the dynamic balance optimization algorithm is smaller. Figure 3 It can be seen that the performance of the dynamic balance optimization algorithm in the neural network of the present invention is compared with the gradient descent method in terms of noise robustness. Under the same noise level, the accuracy of the dynamic balance optimization algorithm is higher. This proves that the dynamic balance optimization algorithm is used to train the local seawater intrusion feature extraction neural network module, and the parameters are optimized by simulating the dynamic process of energy transfer. The error value is reduced. It can be seen that the method of the present invention can effectively avoid the gradient disappearance and gradient explosion problems, and can improve the stability of the model when processing complex marine environment data; Figure 4 It can be seen that the relationship between the sparsification intensity and the prediction accuracy in the extreme learning machine of the present invention is that the higher the sparsification intensity, the higher the accuracy. Therefore, it can be seen that only the weighted sparsification mechanism can make the feature extraction and seawater intrusion prediction and classification process more efficient, show higher adaptability to seawater intrusion data noise and minority category samples, higher accuracy, and more accurate prediction results.

[0029] In summary, the present invention improves the stability and accuracy of the seawater intrusion prediction model when processing complex marine environment data.

[0030] Although the above describes the specific implementation mode of the invention in conjunction with the drawings, it is not intended to limit the scope of protection of the invention. Based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present invention.

Claims

1. A seawater intrusion prediction model training method based on artificial intelligence, characterized by: Collect seawater intrusion related data, build a local training data set, extract features through the local seawater intrusion prediction feature extraction module in the client, then interact the parameters of the local seawater intrusion prediction feature extraction module of each client with the central seawater prediction feature extraction module of the server, and then the central seawater prediction feature extraction module aggregates the updated parameters of each local seawater intrusion prediction feature extraction module to achieve global seawater intrusion prediction feature extraction module parameter update, and use the dynamic balance optimization algorithm to improve the accuracy of the local seawater intrusion prediction feature extraction module, finally, analyze the features captured by the local seawater intrusion prediction feature extraction module through the seawater intrusion prediction classifier module of the client, and the seawater intrusion prediction classifier module adopts the extreme learning machine with high-order partial derivative constraints.

2. The artificial intelligence-based seawater intrusion prediction model training method according to claim 1 is characterized in that: The client's local training data set is as follows: The local training data set of the client is obtained by collecting and annotating data related to seawater intrusion; The relevant data of seawater intrusion include the geographical distribution, time and intensity information of historical seawater intrusion events, the chemical composition, temperature, salinity, pressure, flow rate, topographic parameters of seawater; The sources of relevant data on seawater intrusion include satellite remote sensing data, marine meteorological station data, hydrological monitoring data, geological and hydrogeological data, and historical event data; The data collection method of seawater intrusion is realized through sensors, remote sensing technology, marine meteorological observation stations, unmanned surface vehicles, and seabed sensors. The data is transmitted, processed and verified during the collection process, and finally stored in a unified structured data format, specifically a structured CSV format; The collected data is manually labeled, and the labeled categories include no invasion, slight invasion, moderate invasion, and severe invasion; No invasion means that under the current environmental conditions, the seawater has not invaded, marked as 0; slight invasion means that under the current environmental conditions, the seawater has invaded slightly, marked as 1; moderate invasion means that under the current environmental conditions, the seawater has invaded slightly or moderately, marked as 1; severe invasion means that under the current environmental conditions, the seawater has invaded slightly or moderately, marked as 1.

3. The artificial intelligence-based seawater intrusion prediction model training method according to claim 2 is characterized in that: The global seawater intrusion prediction feature extraction module parameter update process is as follows: Each client is trained according to its own local training data set through its own local seawater intrusion prediction feature extraction module. Based on the federated learning architecture, the parameters obtained after the training of each local seawater intrusion prediction feature extraction module interact with the central seawater intrusion prediction feature extraction module. The central seawater intrusion prediction feature extraction module receives the parameters of each local seawater intrusion prediction feature extraction module and sends the received parameters to each local seawater intrusion prediction feature extraction module. The central seawater intrusion prediction feature extraction module summarizes the parameters of all local seawater intrusion prediction feature extraction modules and then sends the summarized parameters to each local seawater intrusion prediction feature extraction module. Each local seawater intrusion prediction feature extraction module updates its parameters according to the received information, and then shares the updated parameters through the central seawater intrusion prediction feature extraction module again to achieve the re-update of the parameters of each local seawater intrusion prediction feature extraction module. After multiple iterations, the global seawater intrusion prediction feature extraction module parameter update is achieved.

4. The artificial intelligence-based seawater intrusion prediction model training method according to claim 3 is characterized in that: Dynamic balance optimization algorithm: The local seawater intrusion prediction feature extraction module includes a five-layer fully connected neural network. The parameters of the neural network training process are optimized through a dynamic balance optimization algorithm. The dynamic balance optimization algorithm specifically includes an isolated system energy conservation method and a dynamic balance concept. The specific optimization process is as follows; (1) The weights and biases of the neural network in the local seawater intrusion prediction feature extraction module are randomly initialized. The initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix; (2) Set the total energy of the system, which is the process of optimizing the parameters of the neural network in the local seawater intrusion prediction feature extraction module. The total energy of the system is expressed as: ， in, represents the total energy of the system, It is set to a random value during initialization. represents the number of parameters in the neural network, express The index of Indicates the assignment to the neural network The energy of the parameters; The energy assigned to the parameters in the neural network is calculated by the sensitivity of the parameters. The calculation formula is as follows: ， in, Indicates the neural network The sensitivity of the parameters, Indicates the neural network The sensitivity of the parameters, Indicates the neural network The number of neurons in the layer where the parameter is located, Indicates the number of layers of the neural network; Then according to the total energy of the system The energy assigned to the parameter is calculated by the sensitivity calculation of the parameter, and the calculation formula is as follows: ， in, Indicates the assignment to the neural network The energy of the parameters, express Another index of Indicates the neural network The sensitivity of the parameters, Indicates lifting network The sum of the sensitivities of the parameters; The above process is iterated in each layer of the neural network, and the data propagation mode during the iteration is forward propagation; (4) The energy of the parameters of the seawater intrusion feature extraction neural network is dynamically adjusted according to the error feedback. The error feedback is the energy change of the parameters. The calculation formula is as follows: ， ， in, Indicates the neural network The energy change of each parameter is represents the learning rate parameter of the neural network, represents the symbol of partial derivative, represents the trend function, Indicates the neural network The weight of the parameters, Indicates the neural network The weight of the parameters, Indicates the neural network The weight of each parameter changes; (5) Update the energy of the parameter according to the energy change of the parameter. The calculation formula is as follows: ， in, Indicates the neural network The energy after the parameter update; (6) The updated value is calculated using the energy transfer function. The calculation formula is as follows: ， ， in, represents the conversion rate hyperparameter, represents the adjustment factor considering the impact of historical energy, represents the symbolic function, represents the energy transfer function, represents the number of iterations, express The index of Indicates the neural network The iteration The energy of the parameters; (7) Based on the difference between the output of the neural network, i.e., the final output of the local seawater intrusion prediction feature extraction module, and the target output, the energy allocation of each parameter is recalculated and adjusted. The calculation formula is as follows: ， ， in, Indicates the neural network The energy of parameter adjustment, represents the adjustment intensity of parameter energy, represents the difference between the output of the neural network and the target output, Indicates the neural network The updated gradient of the parameter energy, A function that represents the relationship between error and gradient; The adjusted energy is calculated as follows: ， in, Indicates the neural network The final energy after parameter adjustment; (8) Fine-tune the weights of the parameters in the neural network according to the energy of the parameters. The calculation formula is as follows: ， in, represents the fine-tuning coefficient, Indicates the neural network The weights after fine-tuning of parameters; (9) Perform multiple iterations and set iteration conditions until the preset iteration conditions are met and the iteration stops.

5. The artificial intelligence-based seawater intrusion prediction model training method according to claim 4 is characterized in that: Seawater intrusion prediction classifier module: The seawater intrusion prediction classifier module receives the output from the local seawater intrusion prediction feature extraction module, and classifies the results according to the output of the local seawater intrusion prediction feature extraction module, thereby realizing the prediction of the seawater intrusion level; The seawater intrusion prediction classifier module adopts an extreme learning machine with high-order partial derivative constraints. It captures the complex nonlinear relationship of the seawater intrusion data after feature extraction by adhering to the partial derivative constraints, and optimizes the weight of the output layer in the seawater intrusion prediction classifier module by combining the least squares method with the gradient descent method, thereby improving the accuracy of seawater intrusion classification prediction.

6. The artificial intelligence-based seawater intrusion prediction model training method according to claim 5 is characterized in that: Learning Machine: (1) Initialize the parameters of the extreme learning machine and define the output representation of the hidden layer of the extreme learning machine. The calculation formula is as follows: ， ， in, represents the output matrix of the hidden layer of the extreme learning machine, represents the extreme learning machine Activation function, Represents the input sample matrix of the extreme learning machine, with a size of , represents the number of samples, represents the number of features, represents the weight matrix input to the hidden layer, represents the standard deviation of the weight initialization of the extreme learning machine, represents a normal distribution with a mean of 0 and a variance of 1. represents the bias term of the hidden layer of the extreme learning machine, and its initialization value is zero; (2) In the training process of the extreme learning machine model, a high-order partial derivative constraint strategy is adopted, and the update method of the module parameters is adjusted to enhance the learning ability of the complex nonlinear relationship of the seawater intrusion data after feature extraction, so that the module can capture more complex associations between features. The high-order loss function calculation formula of the extreme learning machine is as follows: ， in, represents the high-order loss function of the extreme learning machine, Indicates the input The true labels of samples, Indicates the input The hidden layer output of the extreme learning machine corresponding to the sample, Indicates The hidden layer output of the extreme learning machine corresponding to the sample, Indicates The hidden layer output of the extreme learning machine corresponding to the sample, represents the number of input samples, , and express Three different indexes, represents the tuning parameter of the higher-order partial derivative constraint, represents the tuning parameter of the third-order partial derivative constraint, represents the symbol of partial derivative, represents the second-order partial derivative, represents the third-order partial derivative, represents the extreme learning machine An input sample matrix, represents the extreme learning machine An input sample matrix; (3) The dynamic learning rate adjustment strategy makes the gradient of the extreme learning machine parameter update more flexible. The weight update calculation formula of the extreme learning machine is as follows: ， ， ， ， in, represents the weight update amount of the extreme learning machine, represents the learning rate of the extreme learning machine, represents the total loss function of the extreme learning machine, represents the balanced loss function of the extreme learning machine, represents the first regularization coefficient of the extreme learning machine, represents the first elements, represents the extreme learning machine The weight sparsification strength of the elements, represents the number of elements in the hidden layer of the extreme learning machine, represents the extreme learning machine The absolute value of the element weight, represents the influencing factor of the balance loss function, An adjustment factor indicating that the placement frequency is too low, Indicates the frequency of occurrence, express Activation function, Indicates the input The hidden layer output of the extreme learning machine corresponding to the sample, represents the hyperparameter that controls the variation of coefficientized strength, Represents the weight decay exponent for adjusting the sparsification strength; Then the learning rate of the extreme learning machine is adjusted by dynamic adjustment. The calculation formula is as follows: ， in, represents the learning rate of the adjusted extreme learning machine, represents the learning rate adjustment factor of the extreme learning machine, Indicates the gradient change of the loss function of the extreme learning machine; (4) The output layer weights of the extreme learning machine are optimized by combining the least squares method with the gradient descent method, so that the prediction results are closer to the true value. The calculation formula is as follows: ， in, represents the final weight matrix of the output layer of the extreme learning machine, which is used for the next iteration. represents the second regularization coefficient of the extreme learning machine, represents transpose, represents the identity matrix; (5) Perform multiple iterations and set iteration conditions until the preset iteration conditions are met and the iteration stops.

7. A seawater intrusion prediction model training device based on artificial intelligence, comprising a seawater intrusion prediction model, and executing a seawater intrusion prediction model training method based on artificial intelligence as claimed in any one of claims 1 to 6, characterized in that: The structure of the seawater intrusion prediction model consists of multiple clients and a server. The clients are equipped with local seawater intrusion prediction feature extraction modules and seawater intrusion prediction classifier modules. The server is equipped with a central seawater intrusion prediction feature extraction module. Each client has an independent local training data set. After each local seawater intrusion prediction feature extraction module processes the data in the local training data set, the final classification result is uploaded to the server.

Images