Neural network-based magnetic anomaly response rapid and accurate prediction method
Through the dimensionality reduction and optimization method of the POD-RBF neural network, the problem of high-precision prediction of the magnetic anomaly response characteristics during the projectile penetration process was solved, and fast and accurate magnetic detection and positioning was achieved.
Patent Information
- Application Number
- CN202510506252.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2025-09-16
AI Technical Summary
When existing technologies rely on finite element simulation to analyze magnetic anomaly response characteristics during projectile penetration, computing resources are consumed in large quantities, the research cycle is long, and it is difficult to achieve high-precision magnetic detection and positioning.
A method based on POD-RBF neural network is adopted to process sample data through orthogonal decomposition and dimensionality reduction. The RBF neural network is optimized by combining the PSO algorithm and Sub-NN neural network to construct the mapping relationship between input parameters and modal coefficients, so as to achieve fast and accurate prediction of magnetic anomaly response signals.
High-precision prediction of magnetic anomaly response signals is achieved, computing resource consumption and scientific research cycle are reduced, and prediction efficiency is improved.
Smart Images
Figure CN120653975A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to signal processing technology, and in particular to a method for quickly and accurately predicting magnetic anomaly responses based on a neural network. Background Art
[0002] Magnetic detection technology is a target detection technology characterized by high positioning accuracy, low cost, rapid response, strong discrimination capabilities, and high reliability. Research has found that during projectile penetration, the steel bars in the target plate will disturb the magnetic field around the fuze, making it possible to apply magnetic detection technology to the field of penetration fuzes. However, currently, the main reliance is on finite element simulation to analyze the magnetic anomaly response characteristics of the fuze penetration process. However, the pursuit of high-precision simulation results often requires a large amount of computing resources, and the formulation and verification of control strategies require a large number of magnetic anomaly signal response characteristics under different working conditions as a basis, which poses a huge challenge to both computing resources and scientific research cycles. Summary of the Invention
[0003] The object of the present invention is to provide a method for quickly and accurately predicting magnetic anomaly responses based on a neural network, comprising: Step S100, extracting sample data from simulation parameters, performing dimensionality reduction processing on the sample data using the original orthogonal decomposition method, and intercepting the first K-order POD basis; Step S200, constructing an RBF neural network and optimizing the RBF neural network using a PSO algorithm and a Sub-NN neural network; Step S300: Input simulation parameters to the RBF neural network, and output the predicted modal coefficient matrix after nonlinear feature mapping of the hidden layer; Step S400 , combining the predicted modal coefficients of the RBF neural network and the first K-order POD basis to restore the output prediction signal.
[0004] Furthermore, step S100 specifically includes: Step S101: construct a sample matrix using simulation data , establish the sample covariance matrix ; Step S102, using the covariance matrix Get basis function modes ; Step S103, by calculating the k The high energy principal components in the simulation data are extracted by the cumulative energy contribution rate of the basis function modes ; Step S104: base function mode Arrange in descending order according to the cumulative energy contribution rate, and get the top k POD basis vectors .
[0005] Furthermore, in step S200, the RBF neural network is trained, and the training process includes: Step S201: Input n The simulation parameters of each group are fed into the RBF neural network including the input layer. Each simulation parameter group includes the steel bar diameter. , relative magnetic permeability of the penetrating object , relative magnetic permeability of steel bars ; Step S202: The hidden layer uses the Gaussian basis function as an activation function to complete feature mapping; Step S203: Output the predicted modal coefficient matrix A .
[0006] Furthermore, step S200 uses the PSO algorithm to optimize the neural network, specifically including: Assume that the hidden layer has m hidden layer nodes, the center vector of each node D dimensional, the output layer has K The output node is D In the dimensional optimization space, a particle swarm with a population size of N is constructed, and each particle represents a set of parameters of the RBF neural network, including the hidden layer center vector and output weight; Record the optimal historical position of each particle as the individual extreme value , the best extreme value among all individual extreme values of particles in the group is the global extreme value ; Iterate to the t The second time, i Each particle updates its velocity and position according to these two extreme values; if the termination condition is met, the algorithm ends, otherwise, it continues to iterate.
[0007] Further, the i The update of particle velocity is Location updated to in, is the dimension index, represents the inertia weight, and represents the acceleration constant, and is a random number between [0,1].
[0008] Furthermore, in step S200, the Sub-NN sub-neural network is a convolutional neural network or a recurrent neural network.
[0009] Furthermore, in step S400, a mapping relationship between input parameters and modal coefficients is constructed to reconstruct the prediction signal in, is the predicted modal coefficient matrix A elements.
[0010] The invention discloses a method for quickly and accurately predicting a magnetic anomaly response signal during a penetration process based on an improved POD-RBF neural network, which has a high-precision prediction effect and solves the problems of consuming computing resources and a long scientific research cycle.
[0011] The present invention will be further described below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 It is an overall schematic diagram of the model of the present invention.
[0013] Figure 2 It is the RBF neural network architecture diagram of the present invention.
[0014] Figure 3 It is the optimal expansion coefficient diagram of the three-axis sample data in the example. DETAILED DESCRIPTION
[0015] Combine Figure 1 A method for quickly and accurately predicting magnetic anomaly responses based on a neural network, wherein the neural network is based on an improved POD-RBF neural network, comprises the following steps: Step S100 , extracting sample data from simulation parameters, performing dimension reduction processing on the sample data using the Proven Orthogonal Decomposition (POD) method, and intercepting the first k-order POD basis; Step S200, constructing an RBF neural network and optimizing the RBF neural network using a PSO algorithm and a Sub-NN neural network; Step S300: Input simulation parameters to the RBF neural network, and output the predicted modal coefficient matrix after nonlinear feature mapping of the hidden layer; Step S400 , combining the predicted modal coefficients of the RBF neural network and the first k-order POD basis to restore the output prediction signal.
[0016] The specific process of step S100 is: Step S101: construct a sample matrix using simulation data , establish the sample covariance matrix ,Right now ,in is the number of simulation data; Step S102, by solving the formula Basis function mode , and eigenvalues The total energy of the entire system can be represented as the accumulation of all eigenvalues, that is, ; Step S103, by calculating the k The high energy principal components in the simulation data are extracted by the cumulative energy contribution rate of the basis function modes ,Right now ; Step S104: base function mode Arrange in descending order according to the cumulative energy contribution rate, and get the top k POD basis vectors , .
[0017] Combine Figure 2 In step S200, the RBF neural network includes an input layer, a hidden layer with nonlinear neurons, and a linear output layer. The training process of the RBF neural network includes: Step S201: Input n Group simulation parameters, each group of simulation parameters is , are the three input parameters of the simulation conditions, among which d g is the diameter of the steel bar, is the relative magnetic permeability of the penetrating object, is the relative magnetic permeability of steel bars; Step S202: The hidden layer uses the Gaussian function as the activation function to complete the feature mapping and output , the Gaussian basis function is , c j It is j The center vector of hidden layer neurons (nodes), is the width parameter; Step S203: Output the predicted modal coefficient matrix A , to determine the structure of the RBF neural network.
[0018] The RBF neural network in step S200 is optimized using the PSO algorithm and the Sub-Neural Network (Sub-NN) sub-neural network. PSO is used to globally optimize the center vector and output weights of the hidden layer of the RBF neural network. This algorithm has the advantages of fast global convergence and no need to consider the total number of particles. It alleviates the problem of poor generalization ability and susceptibility to local optimality of the RBF neural network. The nonlinearity in the hidden layer of the nonlinear neuron refers to the nonlinearity of the activation function, which is a Gaussian basis function. The optimization in step S200 includes the following steps: Step S211, assuming that the hidden layer has m hidden layer nodes, the center vector of each node D dimensional, the output layer has K The output node is D In the dimensional optimization space, a particle swarm with a population size of N is constructed. Each particle represents a set of parameters of the RBF neural network, including the hidden layer center vector and output weight. The center vector in the particle position determines the position of the hidden layer node of the RBF neural network. These center vectors are the peak positions of the basis functions and affect the network's response to the input data. Step S212: record the optimal historical position of each particle as the individual extreme value , the best extreme value among all individual extreme values of particles in the group is recorded as the global extreme value ; Step S213, iterate to the t The second time, i Each particle updates its velocity according to the formula based on these two extreme values: , and update its position according to the formula ,in is the dimension index, represents the inertia weight, and represents the acceleration constant, and is a random number between [0,1]; Step S214: If the termination condition is met, the algorithm ends; otherwise, the iteration continues.
[0019] The algorithm termination condition in step S214 may be one of the following conditions: the maximum number of iterations, the fitness reaching a preset threshold, the change of the global optimal solution being less than a preset threshold, and the running time.
[0020] Using a sub-NN (sub-neural network) to optimize the width parameters of the hidden layer of an RBF neural network can effectively improve the predictive ability of the RBF neural network. The sub-NN can be a convolutional neural network (CNN) or a recurrent neural network (RNN). The training process of the sub-NN includes: selecting an appropriate loss function to measure the difference between the width parameters output by the sub-NN and the optimal width parameters. Common loss functions include mean squared error (MSE); using optimization algorithms such as Adam or RMSprop to train the sub-NN, which can accelerate the training process and improve convergence speed; using 64 sets of simulation parameters of the RBF network to train the sub-NN, and evaluating the performance of the sub-NN through methods such as cross-validation.
[0021] In step S400, the mapping relationship between input parameters and modal coefficients is constructed to reconstruct the prediction signal U ( x ) The training samples used in this example are derived from finite element simulations. COMSOL software is used to simulate the triaxial magnetic anomaly signal at the fuze when penetrating a multi-layer target. The target plate is a 5-layer 2m*2m square reinforced concrete target plate. The first layer of the target plate is 300mm thick, and the subsequent target plates are 180mm thick. The target spacing is 3m. The number of x-axis and y-axis steel bars in the target plate is set to 13, and the spacing between adjacent steel bars is 150mm. Add x The direction size is -3.45uT, y The direction size is 32.78uT, z The background magnetic field is -37.71uT. In the sample simulation parameters, the diameter of the steel bar d is set from 10mm to 16mm with a step of 2mm, and the relative magnetic permeability of the projectile is The value ranges from 100 to 400 with a step length of 100. The relative magnetic permeability of steel bars The values range from 600 to 900 with a step length of 100, and there are a total of 64 simulation conditions for each parameter cycle combination.
[0022] The cumulative energy contribution rate of the first few modes of the three-axis sample data obtained by calculation and simulation is almost 100%. In order to balance accuracy and calculation cost, x The axis signal selects the first 4 modes, y The axis signal selects the first 4 modes, z When the first two modes of the shaft signal are selected, the cumulative energy contribution rate reaches 99.99%. A PSO algorithm is used to globally optimize the hidden layer center vector and output weights of the RBF neural network, and the optimized parameters are assigned to the neural network. Furthermore, a sub-neural network is established, with the expansion coefficient of the RBF neural network set as a variable. The corresponding relationship between the simulation parameters and the optimal expansion coefficient is trained, thereby achieving an adaptive output of the optimal expansion coefficient based on the input simulation parameters.
[0023] The expansion coefficient range is set to [0.1, 1.4], with an interval of 0.001, for a total of 1300 possible values. The expansion coefficient value is traversed for each sample and assigned to the RBF network training, for a total of 1300*64 training times. The optimal expansion coefficient of the three axes corresponding to each sample is obtained based on the root mean square error (RMSE), as shown in the following example: Figure 3 shown.
[0024] The corresponding relationship between the simulation parameters and the optimal expansion coefficients of the three axes is input into the sub-neural network. The training algorithm of the sub-neural network adopts the Levenberg-Marquardt (LM) algorithm. The maximum number of training times is set to 1000 and the target error is 1e-3. The mapping relationship between the simulation parameters and their corresponding optimal expansion coefficients is established.
[0025] In order to verify the prediction effect of the model in this embodiment, different numbers of simulation parameters were modified based on the sample simulation parameters and then input into the prediction model. The new parameters were then input into the finite element simulation software to compare the output results of the two. Among them, the modified simulation parameters were random values. The specific input parameters are shown in Table 1. The values in Table 1 are, from left to right, the steel bar diameter, the relative magnetic permeability of the projectile, and the relative magnetic permeability of the steel bar.
[0026] Table 1 Input parameter values The model evaluation indicators are mean absolute percentage error (MAPE) and root mean square error (RMSE). The smaller the two values are, the closer the prediction results are to the actual situation. Table 2, Table 3, and Table 4 are the results of the model evaluation under the input parameters. x axis, y axis, z Axis prediction signal performance evaluation index.
[0027] Table 2 X-axis prediction signal performance evaluation indicators under all input parameters Table 3 Y-axis prediction signal performance evaluation indicators under all input parameters Table 4 Performance evaluation index of Z-axis prediction signal under all input parameters As can be seen from the table, under random parameter input, x The maximum mean absolute percentage error of the axis prediction signal is 4.0777%, and the average is 2.3297%; the maximum root mean square error is 0.0521, and the average is 0.02562. y The maximum mean absolute percentage error of the axis prediction signal is 5.4068%, and the average is 1.977%; the maximum root mean square error is 0.1438, and the average is 0.0506. z The maximum mean absolute percentage error of the axis prediction signal is 2.3965%, and the average is 0.957%; the maximum root mean square error is 0.297, and the average is 0.1056.
[0028] Judging from the results, no matter whether one parameter, two parameters or all parameters are modified on the basis of the sample simulation parameters, the prediction error of the model is basically in the same range, the prediction consistency is high, and the results verify the robustness of the model.
Claims
1. A method for rapid and accurate prediction of magnetic anomaly response based on neural network, characterized in that: include: Step S100, extracting sample data from simulation parameters, performing dimensionality reduction processing on the sample data using the original orthogonal decomposition method, and intercepting the first K-order POD basis; Step S200, constructing an RBF neural network and optimizing the RBF neural network using a PSO algorithm and a Sub-NN neural network; Step S300: Input simulation parameters to the RBF neural network, and output the predicted modal coefficient matrix after nonlinear feature mapping of the hidden layer; Step S400 , combining the predicted modal coefficients of the RBF neural network and the first K-order POD basis to restore the output prediction signal.
2. The method according to claim 1, characterized in that Step S100 specifically includes: Step S101: construct a sample matrix using simulation data , establish the sample covariance matrix ; Step S102, using the covariance matrix Get basis function modes ; Step S103, by calculating the k The high energy principal components in the simulation data are extracted by the cumulative energy contribution rate of the basis function modes ; Step S104: base function mode Arrange in descending order according to the cumulative energy contribution rate, and get the top k POD basis vectors .
3. The method according to claim 1, characterized in that In step S200, the RBF neural network is trained. The training process includes: Step S201: Input n The simulation parameters of each group are fed into the RBF neural network including the input layer. Each simulation parameter group includes the steel bar diameter. d g , relative magnetic permeability of the penetrating object , relative magnetic permeability of steel bars ; Step S202: The hidden layer uses the Gaussian basis function as an activation function to complete feature mapping; Step S203: Output the predicted modal coefficient matrix A .
4. The method according to claim 2, characterized in that Step S200 uses the PSO algorithm to optimize the neural network, specifically including: Assume that the hidden layer has m hidden layer nodes, the center vector of each node D dimensional, the output layer has K The output node is D In the dimensional optimization space, a particle swarm with a population size of N is constructed, and each particle represents a set of parameters of the RBF neural network, including the hidden layer center vector and output weight; Record the optimal historical position of each particle as the individual extreme value , the best extreme value among all individual extreme values of particles in the group is the global extreme value ; Iterate to the t The second time, i Each particle updates its velocity and position according to these two extreme values; if the termination condition is met, the algorithm ends, otherwise, it continues to iterate.
5. The method according to claim 4, characterized in that No. i The update of particle velocity is Location updated to in, d is the dimension index, represents the inertia weight, and represents the acceleration constant, and is a random number between [0,1].
6. The method according to claim 4 or 5, characterized in that In step S200, the Sub-NN neural network is a convolutional neural network or a recurrent neural network.
7. The method according to claim 4 or 5, characterized in that In step S400, a mapping relationship between input parameters and modal coefficients is constructed to reconstruct the prediction signal. U ( x ) in, a i is the predicted modal coefficient matrix A elements.