Plasma array microwave protection method and system based on time sequence neural network

Abstract

The invention discloses a plasma array microwave protection method and system based on a time sequence neural network, and relates to the field of microwave protection. The method comprises the following steps: constructing a neural network model and setting initialization parameters; the method comprises the following steps: reading a numerical sequence of interaction of high-power microwaves and a plasma array, performing maximum and minimum normalization processing on each data feature in the numerical sequence to obtain an input sequence and an output target sequence, making a data set, and dividing the data set into a training set, a verification set and a test set; training the model by using the training set, carrying out back propagation on the loss function, and carrying out parameter optimization by using an optimizer to obtain a trained multi-path feature information fusion time sequence neural network model; performing forward propagation on input data in the verification set by using the model to obtain a prediction error of the current training model; and performing forward propagation on input data in a test set by using the model to obtain a prediction result. According to the method, high-power microwave inversion, future numerical response prediction and missing value inference can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of microwave protection, and in particular to a plasma array microwave protection method and system based on a time sequence neural network. Background Art

[0002] Microwave technology has made significant progress in recent years and has been widely used in a variety of key areas, such as microwave imaging, radar detection, wireless communications, and satellite navigation. However, electromagnetic interference (EMI) has become increasingly serious, potentially damaging to human health and the environment and disrupting the normal operation of communication systems. Consequently, with the continued development of microwave technology, research on electromagnetic shielding and high-power microwave protection has become crucial. Therefore, developing efficient electromagnetic shielding materials and protective measures to ensure human health, ecological safety, and the stability of communication systems has become a critical issue in science and technology. Researchers have made significant progress in the field of electromagnetic shielding. Although traditional materials such as rectangular housings, metal mesh, and modern glass can provide electromagnetic shielding to a certain extent, they are insufficient in protecting against high-power microwaves. To address this challenge, researchers have developed a variety of innovative shielding technologies over the past few decades to protect against the effects of high-power microwaves, including frequency selective surfaces (FSS), energy selective surfaces (ESS), and broadband adaptive waveguides. However, these technologies are generally limited to specific frequency ranges and lack the necessary dynamic adjustability, limiting their adaptability and flexibility in practical applications. While some degree of frequency adjustment capability has been achieved by introducing tunable components, this has been shown to adversely affect other components and increase the complexity of the manufacturing process.

[0003] Therefore, the development of electromagnetic shielding technology with wide bandwidth, high efficiency and good dynamic adjustability is of great significance for improving the electromagnetic shielding effect and meeting the needs of modern technology. In contrast, plasma arrays are a good candidate. When high-power microwaves hit the columnar array with a specific electric field strength, incident angle and frequency to excite plasma and produce an electromagnetic shielding effect, the columnar plasma array will immediately prevent the high-power microwaves from penetrating free space. However, not all high-power microwaves can excite gas discharge to form plasma and thus produce electron avalanches. The protection effect often depends on the electric field strength, incident frequency, incident angle of the high-power microwave and the initial electron density of the plasma. Future research needs to further explore the application of plasma arrays in electromagnetic shielding, and how to optimize these parameters to improve the protection effect. Therefore, for those skilled in the art, how to use columnar plasma arrays for high-power microwave protection more efficiently and intelligently is an urgent problem to be solved. Summary of the Invention

[0004] The purpose of the present invention is to provide a plasma array microwave protection method and system based on a time series neural network to solve the problems raised in the background technology. The method uses the numerical response of the plasma under high-power microwave irradiation to achieve high-precision deduction results, which can realize high-power microwave inversion, predict future numerical responses and infer missing values.

[0005] To achieve the above objectives, the present invention provides the following solutions: On the one hand, a plasma array microwave protection method based on a time-series neural network is provided, and the specific steps include the following:

[0006] Constructing a neural network model and setting initialization parameters of the neural network model;

[0007] The interaction process between high-power microwaves and plasma is simulated based on Maxwell's curl equation, electron density drift-diffusion equation, electron energy drift-diffusion equation, heavy matter transfer equation and Poisson's equation;

[0008] Reading a numerical sequence of the interaction between the high-power microwave and the plasma array, performing maximum and minimum normalization processing on each data feature in the numerical sequence, and obtaining an input sequence and an output target sequence;

[0009] Convert the input sequence and the output target sequence into tensors to create a dataset, and divide the dataset into a training set, a validation set, and a test set;

[0010] Using the training set to train the neural network model, backpropagating the loss function, and using an optimizer to optimize the parameters to obtain a trained multi-path feature information fusion time series neural network model;

[0011] Performing forward propagation on the input data in the validation set using the multi-path feature information fusion time series neural network model to obtain a prediction error of the current training model;

[0012] The multi-path feature information fusion temporal neural network model is used to forward propagate the input data in the test set to obtain the prediction results.

[0013] Preferably, a position encoding network layer is added to the neural network model to add position information to the native data.

[0014] Preferably, the neural network model also includes a variable embedding layer network and a sequence embedding layer network, which perform data segmentation on the input sequence from the perspective of multiple variables and the perspective of long time series respectively. On this basis, the multi-head attention mechanism is applied to realize feature extraction of the input data, and the attention score of the preprocessed sequence is calculated from different perspectives to capture the data dependence of the time series after passing through the embedding layer. The multi-head attention mechanism is applied in the decoder part to realize multi-path information fusion technology.

[0015] Preferably, the stochastic gradient descent algorithm is selected as the optimizer of the neural network model and the MSE function is selected as the loss function of the model.

[0016] Preferably, during a training process of a neural network, the encoder completes one forward propagation and the decoder completes multiple forward propagations. During the training process, the input sequence received by the decoder during the first forward propagation is an all-zero sequence.

[0017] Preferably, the initialization parameters include the spatial dimension of the neural network model, the number of heads of the multi-head attention mechanism, the batch size of the data, the number of variables in the encoder input sequence, the sequence length of the encoder input sequence, the number of variables in the decoder input sequence, the sequence length of the decoder input sequence, the number of repetitions of the encoder, and the number of repetitions of the decoder.

[0018] Preferably, the maximum and minimum values of the input data are normalized, and the calculation formula is:

[0019]

[0020] Among them, γ and β are the parameters that need to be learned in the neural network, μ and σ 2 They represent the mean and variance of the sequence respectively, x is the input data, and ε is a non-zero number used to ensure numerical stability and prevent division by zero.

[0021] Preferably, in the neural network model, the feedforward neural network uses one-dimensional convolution to extract feature point information from the data. The one-dimensional convolution captures local features through local connections, and the weights of the same convolution kernel are shared across the entire input data.

[0022] On the other hand, a plasma array microwave protection system based on a time series neural network is provided, comprising an initialization module, a deduction simulation module, a data acquisition module, a preprocessing module, a training module, a verification module, and a prediction module; wherein,

[0023] The initialization module is used to construct a neural network model and set initialization parameters of the neural network model;

[0024] The deduction simulation module is used to simulate the interaction process between high-power microwaves and plasma based on Maxwell's curl equation, electron density drift-diffusion equation, electron energy drift-diffusion equation, heavy matter transfer equation and Poisson's equation;

[0025] The data acquisition module is used to read the numerical sequence of the interaction between the high-power microwave and the plasma array, perform maximum and minimum value normalization processing on each data feature in the numerical sequence, and obtain an input sequence and an output target sequence;

[0026] The preprocessing module is used to convert the input sequence and the output target sequence into tensors to produce a dataset, and divide the dataset into a training set, a validation set, and a test set;

[0027] The training module is used to train the neural network model using the training set, perform backpropagation on the loss function, and use an optimizer to optimize the parameters to obtain a trained multi-path feature information fusion time series neural network model;

[0028] The verification module is used to use the multi-path feature information fusion time series neural network model to perform forward propagation on the input data in the verification set to obtain the prediction error of the current training model;

[0029] The prediction module is used to use the multi-path feature information fusion temporal neural network model to perform forward propagation on the input data in the test set to obtain a prediction result.

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0031] (1) The time series neural network model based on multipath feature information fusion shows superior performance compared to the existing iTransformer in high-power microwave protection problems, requires fewer training parameters, achieves extremely high accuracy and stronger stability, and has excellent nonlinear representation learning and time series deduction capabilities;

[0032] (2) A variable embedding layer network and a sequence embedding layer network are proposed to segment the input sequence from the perspective of multiple variables and long time series. On this basis, a multi-head attention mechanism is applied to extract the features of the input data, and the attention score of the preprocessed sequence is calculated from two different perspectives to capture the data dependency of the time series after passing through the embedding layer. On this basis, the multi-head attention mechanism is applied in the decoder to realize multi-path information fusion technology;

[0033] (3) The multi-path feature information fusion time series neural network model constructed by the present invention has high deduction accuracy and strong robustness, and can be widely applied to various time series tasks, such as time series inversion, time series prediction and time series inference. It can realize real-time and intelligent enabling in high-power microwave protection, realize high-power microwave inversion (infer the field strength, frequency and incident angle of high-power microwave), predict future numerical responses and infer missing values;

[0034] (4) Aiming at the mathematical modeling of a cylindrical tube array filled with argon plasma under high-power microwave irradiation and the realization of high-power microwave efficient protection, the present invention segments and enriches the input sequence data from the perspective of multiple variables and long time series, and further proposes a multi-path feature information fusion time series neural network model algorithm (Multi-path Feature Transformer, MFT) based on the encoding-decoding architecture. It can not only predict the numerical response of the plasma at a future moment and infer the missing value at any position in the plasma response sequence, but also further realize high-power microwave inversion based on the incomplete plasma response sequence. The present invention uses artificial intelligence to enable high-power microwave protection, which is a brand-new innovation. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0036] Figure 1 is a flow chart of the method of the present invention;

[0037] Figure 2 This is a diagram of the algorithm architecture of the present invention;

[0038] Figure 3 This is a comparison chart of the training loss, validation loss, and loss of the iTransformer model in Model 1 of the embodiment of the present invention;

[0039] Figure 4 This is a comparison chart of the test accuracy of Model 1 in the embodiment of the present invention and the accuracy of the iTransformer model;

[0040] Figure 5 This is the first result diagram of the present invention predicting the electron density of 5 unknown future data based on 15 historical data;

[0041] Figure 6 This is a second result diagram of the present invention predicting the electron density of 5 unknown future data based on 15 historical data;

[0042] Figure 7 This is the third result diagram of the present invention predicting the electron density of 5 unknown future data based on 15 historical data;

[0043] Figure 8 This is the fourth result diagram of the present invention predicting the electron density of 5 unknown future data based on 15 historical data;

[0044] Figure 9 This is a comparison chart of the training loss, validation loss, and iTransformer model loss in model 2 of the embodiment of the present invention;

[0045] Figure 10 This is a comparison chart of the test accuracy of Model 2 in the embodiment of the present invention and the accuracy of the iTransformer model;

[0046] Figure 11 This is a comparison chart of the training loss, validation loss, and loss of the iTransformer model in the third embodiment of the present invention;

[0047] Figure 12 This is a comparison chart of the accuracy of the model 3 in the embodiment of the present invention and the iTransformer model;

[0048] Figure 13 This is the first result diagram of the present invention based on the plasma density sequence with a sequence length of 10 to infer the missing values of any two positions;

[0049] Figure 14 This is the second result diagram of the present invention based on the plasma density sequence with a sequence length of 10 to infer the missing values of any two positions;

[0050] Figure 15 This is the third result diagram of the present invention based on the plasma density sequence with a sequence length of 10 to infer the missing values of any two positions;

[0051] Figure 16 This is the fourth result diagram of the present invention based on the plasma density sequence with a sequence length of 10 to infer the missing values of any two positions. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] The purpose of this invention is to provide a plasma array microwave protection method based on a time sequence neural network, such as Figure 1 、 2 As shown, the specific steps include:

[0054] S1. Build a neural network model and set the initialization parameters of the neural network model;

[0055] S2. Simulate the interaction between high-power microwaves and plasma based on Maxwell's curl equation, electron density drift-diffusion equation, electron energy drift-diffusion equation, heavy matter transfer equation, and Poisson's equation;

[0056] S3, reading the numerical sequence of the interaction between the high-power microwave and the plasma array, performing maximum and minimum normalization processing on each data feature in the numerical sequence, and obtaining an input sequence and an output target sequence;

[0057] S4. Convert the input sequence and output target sequence into tensors to create a dataset, and divide it into training set, validation set and test set;

[0058] S5. Use the training set to train the neural network model, perform backpropagation on the loss function, and use the optimizer to optimize the parameters to obtain a trained multi-path feature information fusion time series neural network model;

[0059] S6. Use the multi-path feature information fusion time series neural network model to forward propagate the input data in the validation set to obtain the prediction error of the current training model;

[0060] S7. Use the multi-path feature information fusion time series neural network model to forward propagate the input data in the test set to obtain the prediction results.

[0061] Furthermore, in step S1, the initialization parameters include the spatial dimension of the neural network model, the number of heads of the multi-head attention mechanism, the batch size of the data, the number of variables of the encoder input sequence, the sequence length of the encoder input sequence, the number of variables of the decoder input sequence, the sequence length of the decoder input sequence, the number of repetitions of the encoder, and the number of repetitions of the decoder.

[0062] Furthermore, in step S2, a wave equation in the plasma was established to analyze the propagation process of high-power microwaves in the plasma; secondly, an electron density / energy drift-diffusion equation was established to study the changes in electron density / electron energy inside the plasma under the action of high-power microwaves; then, the Poisson equation was added to study the changes in the electron potential inside the plasma; finally, a heavy matter transfer equation was established to study the effects of the incident electromagnetic waves on other charged particles inside the plasma.

[0063] Maxwell's curl equation:

[0064]

[0065] Among them, j, ω, E, H, μ, σ, and ε represent the imaginary unit, angular frequency, Hamiltonian operator, electric field, magnetic field, magnetic permeability, conductivity, and dielectric constant, respectively.

[0066] Electron density drift diffusion equation:

[0067]

[0068] Among them, n e is the electron density; μ e is the electron mobility; D e is the electron diffusion rate; R e is the electron source term; m i is the molar mass fraction of the colliding particles in the i-th reaction; k i is the reaction rate of the ith reaction; N t is the total particle number density in the plasma.

[0069] Electron energy drift diffusion equation:

[0070]

[0071] Among them, Δε i is the energy loss of the ith reaction of the interaction between high-power microwave and plasma; ε is the energy; σ k is the plasma collision cross section; F is the electron energy distribution function.

[0072] Heavy mass transfer equation

[0073]

[0074] Where ρ is the gas density; x k is the mole fraction of the kth particle; u is the average fluid velocity; j k is the diffusion flux of the kth particle; R k is the rate of change of the kth particle.

[0075] Poisson's equation

[0076]

[0077] Where V is the gas density; ε0 is the dielectric constant of vacuum; ε r is the relative dielectric constant of plasma; q is the electron charge.

[0078] Furthermore, in step S3, the calculation formula for normalizing the maximum and minimum values of each data feature in the numerical sequence is:

[0079]

[0080] Among them, x min Represents the minimum value in the entire sequence, x max Represents the maximum value in the entire sequence.

[0081] It also includes the maximum and minimum normalization of the input data. The calculation formula is:

[0082]

[0083] Among them, γ and β are the parameters that need to be learned in the neural network, μ and σ 2 They represent the mean and variance of the sequence respectively, x is the input data, and ε is a non-zero number used to ensure numerical stability and prevent division by zero.

[0084] Furthermore, a positional encoding network layer is added to the neural network model to add position information to the native data:

[0085]

[0086] Among them, PE represents the position encoding value, POS represents the variable position, which ranges from 0 to N-1, and d represents the current dimension, which ranges from 0 to D / 2-1.

[0087] The present invention adopts the Encoder-Decoder architecture to complete the time series task. The Encoder part adopts a combination of Patching Embedding (sequence embedding layer network), Variate Embedding (variable embedding layer network) and Multi-head Attention (multi-head attention mechanism). The Decoder part adopts the Teaching Multivariate Attention (teaching multivariate attention) mechanism to fuse multi-path feature information. Residual connections are used between network layers to transmit information.

[0088] The neural network model also includes a variable embedding layer network (Variate Embedding) and a sequence embedding layer network (Patch Embedding), which segment the input sequence from the perspective of multiple variables and long time series respectively. On this basis, the multi-head attention mechanism (Multi-Head Attention) is applied to extract the features of the input data, calculate the attention score of the preprocessed sequence from different perspectives, capture the data dependence of the time series after passing through the embedding layer, and apply the multi-head attention mechanism in the decoder part to realize multi-path information fusion technology.

[0089] The embedding layer network of Variate Embedding enriches the features of the input data from the perspective of multivariate long time series:

[0090]

[0091] Where L represents the length of the input time series, M represents the number of blocks after segmentation, S represents the length of the non-overlapping area between two consecutive blocks, and Q represents the length of each sub-block.

[0092] The embedding layer network of Patch Embedding enriches the features of the input data from the perspective of multivariate long time series:

[0093]

[0094] Where L represents the length of the input time series, W represents the number of blocks after segmentation, J represents the length of the non-overlapping area between two consecutive blocks, and K represents the length of each sub-block.

[0095] Multi-HeadAttention mechanism:

[0096]

[0097] Multi-head Q,K,V =Concat(head1,...,head h )W O ;

[0098] Among them, d k =D / n_head, n_head represents the number of heads of the multi-head attention mechanism, W o 、W i Q 、W i K 、W i V are the network parameters that the neural network needs to train.

[0099] In this embodiment, the stochastic gradient descent algorithm is selected as the optimizer of the neural network model, and the MSE function is selected as the loss function of the model.

[0100] This invention proposes a forward teaching process called the Teaching Process. Specifically, during one network training phase, the decoder is trained N times and the encoder is trained once. This approach is often used in validation or testing when there is no external real-world target (reference information) input. This model not only learns the correct knowledge but also incorporates its own thinking, significantly improving its robustness, accuracy, and generalization capabilities.

[0101] The teaching process is used for the decoder and consists of two steps. For training, in the first step, the decoder takes a multivariate sequence of all zeros as input and uses the MultivariateAttention mechanism to capture the global dependencies of the input sequence. In the second step, the decoder randomly chooses to use either the MultivariateAttention mechanism or the MaskedMultivariateAttention mechanism based on the TeachingRatio, acting on the true target sequence or the previous output, respectively. In this step, the decoder performs multiple rounds of representation learning, each input being derived from the previous output, fully extracting global and local dependencies. For validation and testing, in the first step, the decoder takes a multivariate sequence of all zeros as input and uses the MultivariateAttention mechanism to capture the global dependencies of the input sequence. In the second step, the decoder randomly chooses to use the MultivariateAttention mechanism or MaskedMultivariateAttention based on the Teaching Ratio. Each input information comes from the previous output result, fully extracting the global and local dependencies of the information. This process is independent of the actual target sequence and is consistent with the real situation where the target result is unknown before prediction.

[0102] The feed-forward network uses one-dimensional convolution to extract feature point information from the data. One-dimensional convolution captures local features through local connections. The weights of the same convolution kernel are shared across the entire input data, improving parameter utilization. The calculation method for parameter changes is:

[0103] n out =(n in +2×n padding -n filter )+1;

[0104] Among them, n in 、n out 、n padding 、n filter They represent the number of input variables, the number of output variables, the amount of padding, and the size of the convolution kernel respectively.

[0105] The accuracy formula is used to measure the similarity between the predicted value and the true value. i With p iWhen they are equal, Accuracy is equal to 1.

[0106]

[0107] Among them, t i 、p i represent the target value and the predicted value respectively.

[0108] The following three time series were designed to verify the results and conduct experimental comparison with the iTransformer model:

[0109] Model 1: Use multi-path feature information fusion time series neural network model to realize the numerical prediction of plasma electron density at future time.

[0110] The spatial size of the model is set to 1024, the number of heads of the multi-head attention mechanism is 32, the batch size is 512, the number of variables in the encoder input sequence is 4, the length of the encoder input sequence is 15, the number of variables in the decoder input sequence is 1, the length of the decoder input sequence is 5, the number of encoder repetitions is 3, and the number of decoder repetitions is 3.

[0111] During the training phase, the response results calculated by numerical simulation are clipped without distortion using a rectangular window function to ensure that the information is not affected by noise. Known historical data with a sequence length of 15 is used as the model input data, that is, the encoder input data. The sequence values corresponding to the historical data for the next five moments are used as the model's target data. The TeachingRatio determines whether the decoder input data is the model's target data or a sequence of zero values. The network model obtains optimized model parameters after continuous iterative training.

[0112] During the validation phase, the validation set is used to verify whether the network model's prediction accuracy after each iteration is overfitting or underfitting. Similarly, known historical data with a sequence length of 15 is used as the model input data, that is, the encoder input data. The trained model naturally infers the predicted values for the next five moments. Figure 3 The training loss and validation loss of the present invention are shown, as well as the loss comparison with the iTransformer model.

[0113] During the testing phase, the test set is used to test the prediction accuracy of the network model after training. Similarly, known historical data with a sequence length of 15 is used as the model input data, that is, the input data of the encoder. The trained model naturally infers the predicted values for the next five moments. Figure 4 The prediction accuracy of the present invention on the test set is shown, as well as the accuracy comparison with the iTransformer model. Figure 5 、 Figure 6 、 Figure 7 、 Figure 8 The prediction results of the model are shown in Table 1, which are compared with the prediction results of the iTransformer model.

[0114]

[0115] Table 1

[0116] Model 2: High-power microwave inversion is achieved using a multi-path feature information fusion time series neural network model.

[0117] The spatial size of the model is set to 1024, the number of heads of the multi-head attention mechanism is 32, the batch size is 512, the number of variables in the encoder input sequence is 4, the length of the encoder input sequence is 30, the number of variables in the decoder input sequence is 4, the length of the decoder input sequence is 1, the number of encoder repetitions is 3, and the number of decoder repetitions is 3.

[0118] During the training phase, the response results calculated by numerical simulation are clipped without distortion using a rectangular window function to ensure that the information is not affected by noise. Feature response data with 4 variables and a sequence length of 30 is used as the model input, i.e., the encoder input data. HPM feature parameters with 4 variables and a sequence length of 1 are used as the model target data. The teaching ratio determines whether the decoder input data is the model target data or a zero-value sequence. The network model is trained iteratively to obtain optimized model parameters.

[0119] During the validation phase, the validation set was used to verify whether the network model's prediction accuracy after each iteration was overfitting or underfitting. Similarly, feature response data with 4 variables and a sequence length of 30 was used as the model input, i.e., the encoder input. The trained model then naturally inferred the HPM feature parameters. Figure 9 The training and validation losses of the present invention are shown, along with a comparison with the iTransformer model's loss. During the testing phase, a test set was used to test the inversion accuracy of the network model after training. Similarly, feature response data with 4 variables and a sequence length of 30 was used as the model input, i.e., the encoder input. The trained model naturally inferred the HPM feature parameters. Figure 10 The prediction accuracy of the present invention on the test set is shown in Table 2, as well as the accuracy comparison with the iTransformer model. The comparison analysis of the predicted HPM parameters and the true HPM parameters is shown in Table 2, and Table 3 reflects the comparative analysis of the present invention in achieving HPM inversion.

[0120] category Electric field (V / m) Frequency (GHz) Incident angle (rad) True value 14000 9 0 The present invention 13993.2490 8.9990 0.0011

[0121] Table 2

[0122]

[0123] Table 3

[0124] Model 3: Use the multi-path feature information fusion time series neural network model to realize the inference of missing values in the plasma electron density series.

[0125] The spatial size of the model is set to 2048, the number of heads of the multi-head attention mechanism is 32, the batch size is 512, the number of variables in the encoder input sequence is 1, the length of the encoder input sequence is 10, the number of variables in the decoder input sequence is 1, the length of the decoder input sequence is 2, the number of encoder repetitions is 3, and the number of decoder repetitions is 3.

[0126] During the training phase, the response results calculated by numerical simulation are clipped without distortion using a rectangular window function to ensure that the information is not affected by noise. Feature response data with a variable number of 1 and a sequence length of 10 (including two randomly missing feature responses) is used as the model input data, i.e., the encoder input data. Missing values with a variable number of 1 and a sequence length of 2 are used as the model's target data. The teaching ratio determines whether the decoder input data is the model's target data or a sequence of zero values. The network model obtains optimized model parameters after continuous iterative training.

[0127] During the validation phase, the validation set was used to verify whether the network model's prediction accuracy after each iteration was overfitting or underfitting. Similarly, feature response data with a variable count of 1 and a sequence length of 10 (including two randomly missing feature responses) was used as the model input data, i.e., the encoder input data. The trained model naturally inferred the two randomly missing feature values based on the known eight feature responses. Figure 11 The training loss and validation loss of the present invention are shown, as well as the loss comparison with the iTransformer model.

[0128] During the testing phase, the test set was used to test the inversion accuracy of the network model after training. Similarly, feature response data with a variable count of 1 and a sequence length of 10 (including two randomly missing feature responses) was used as the model input data, i.e., the encoder input data. The trained model naturally inferred the two randomly missing feature values based on the known eight feature responses. Figure 12 The prediction accuracy of the present invention on the test set is shown, as well as the accuracy comparison with the iTransformer model. Figure 13 、 Figure 14 、 Figure 15 、 Figure 16 The deduction results of the model are shown and compared with the deduction results of the iTransformer model.

[0129] Table 4 analyzes the performance of missing values in the inferred plasma electron density response.

[0130]

[0131] Table 4

[0132] On the other hand, a plasma array microwave protection system based on a time series neural network is provided, comprising an initialization module, a deduction simulation module, a data acquisition module, a preprocessing module, a training module, a verification module, and a prediction module; wherein,

[0133] An initialization module, used to construct a neural network model and set initialization parameters of the neural network model;

[0134] The deduction simulation module is used to simulate the interaction process between high-power microwaves and plasma based on Maxwell's curl equation, electron density drift-diffusion equation, electron energy drift-diffusion equation, heavy matter transfer equation and Poisson's equation;

[0135] A data acquisition module is used to read the numerical sequence of the interaction between the high-power microwave and the plasma array, perform maximum and minimum normalization processing on each data feature in the numerical sequence, and obtain an input sequence and an output target sequence;

[0136] A preprocessing module, configured to convert the input sequence and the output target sequence into tensors to produce a dataset, and divide the dataset into a training set and a test set;

[0137] The training module is used to train the neural network model using the training set, backpropagate the loss function, and use the optimizer to optimize the parameters to obtain the trained multi-path feature information fusion time series neural network model;

[0138] The verification module is used to use the multi-path feature information fusion time series neural network model to perform forward propagation on the input data in the verification set to obtain the prediction results;

[0139] The prediction module is used to use the multi-path feature information fusion time series neural network model to perform forward propagation on the input data in the test set to obtain the prediction results.

[0140] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A plasma array microwave protection method based on a time series neural network, characterized in that: The specific steps include the following: Constructing a neural network model and setting initialization parameters of the neural network model; The interaction process between high-power microwaves and plasma is simulated based on Maxwell's curl equation, electron density drift-diffusion equation, electron energy drift-diffusion equation, heavy matter transfer equation and Poisson's equation; Reading a numerical sequence of the interaction between the high-power microwave and the plasma array, performing maximum and minimum normalization processing on each data feature in the numerical sequence, and obtaining an input sequence and an output target sequence; Convert the input sequence and the output target sequence into tensors to create a dataset, and divide the dataset into a training set, a validation set, and a test set; Using the training set to train the neural network model, backpropagating the loss function, and using an optimizer to optimize the parameters to obtain a trained multi-path feature information fusion time series neural network model; Performing forward propagation on the input data in the validation set using the multi-path feature information fusion time series neural network model to obtain a prediction error of the current training model; The multi-path feature information fusion temporal neural network model is used to forward propagate the input data in the test set to obtain the prediction results.

2. The plasma array microwave protection method based on a time series neural network according to claim 1, characterized in that: In the neural network model, a position encoding network layer is added to add position information to the native data.

3. The plasma array microwave protection method based on a time series neural network according to claim 1, characterized in that: The neural network model also includes a variable embedding layer network and a sequence embedding layer network, which segment the input sequence from the perspective of multiple variables and long time series respectively. On this basis, the multi-head attention mechanism is applied to realize feature extraction of the input data, calculate the attention score of the preprocessing sequence from different perspectives, capture the data dependence degree of the time series after passing through the embedding layer, and apply the multi-head attention mechanism in the decoder part to realize multi-path information fusion technology.

4. The plasma array microwave protection method based on a time series neural network according to claim 1, characterized in that: The stochastic gradient descent algorithm was selected as the optimizer of the neural network model and the MSE function was selected as the loss function of the model.

5. The plasma array microwave protection method based on a time series neural network according to claim 3, characterized in that: In the process of training a neural network, the encoder completes one forward propagation and the decoder completes multiple forward propagations. During the training process, the input sequence received by the decoder during the first forward propagation is an all-zero sequence.

6. The plasma array microwave protection method based on a time series neural network according to claim 5, characterized in that: The initialization parameters include the spatial dimension of the neural network model, the number of heads of the multi-head attention mechanism, the batch size of the data, the number of variables in the encoder input sequence, the sequence length of the encoder input sequence, the number of variables in the decoder input sequence, the sequence length of the decoder input sequence, the number of repetitions of the encoder, and the number of repetitions of the decoder.

7. The plasma array microwave protection method based on a time series neural network according to claim 1, characterized in that: It also includes the maximum and minimum normalization of the input data. The calculation formula is: Among them, γ and β are the parameters that need to be learned in the neural network, μ and σ 2 They represent the mean and variance of the sequence respectively, x is the input data, and ε is a non-zero number used to ensure numerical stability and prevent division by zero.

8. The plasma array microwave protection method based on a time series neural network according to claim 1, characterized in that: In the neural network model, the feedforward neural network uses one-dimensional convolution to extract feature point information from the data. The one-dimensional convolution captures local features through local connections, and the weights of the same convolution kernel are shared across the entire input data.

9. A plasma array microwave protection system based on a time-series neural network, characterized in that: It includes initialization module, deduction simulation module, data acquisition module, preprocessing module, training module, verification module and prediction module; among them, The initialization module is used to construct a neural network model and set initialization parameters of the neural network model; The deduction simulation module is used to simulate the interaction process between high-power microwaves and plasma based on Maxwell's curl equation, electron density drift-diffusion equation, electron energy drift-diffusion equation, heavy matter transfer equation and Poisson's equation; The data acquisition module is used to read the numerical sequence of the interaction between the high-power microwave and the plasma array, perform maximum and minimum value normalization processing on each data feature in the numerical sequence, and obtain an input sequence and an output target sequence; The preprocessing module is used to convert the input sequence and the output target sequence into tensors to produce a dataset, and divide the dataset into a training set, a validation set, and a test set; The training module is used to train the neural network model using the training set, perform backpropagation on the loss function, and use an optimizer to optimize the parameters to obtain a trained multi-path feature information fusion time series neural network model; The verification module is used to use the multi-path feature information fusion time series neural network model to perform forward propagation on the input data in the verification set to obtain the prediction error of the current training model; The prediction module is used to use the multi-path feature information fusion temporal neural network model to perform forward propagation on the input data in the test set to obtain a prediction result.

Images