Plasma parameter diagnosis method based on artificial intelligence
By constructing the residual neural network model and using the deep learning framework, the accuracy problem of electronic density diagnosis of plasma falling edges is solved, the robustness and accuracy of diagnosis is improved, and the gradient disappearance and explosion problems in BP neural network are overcome.
Patent Information
- Application Number
- CN202510023128.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-07
- Publication Date
- 2025-05-09
AI Technical Summary
The prior art is difficult to accurately diagnose the electron density of the falling edge of plasma, and the BP neural network-based method has problems of gradient disappearance and explosion, resulting in unsatisfactory model training effect.
The plasma parameter diagnosis method based on artificial intelligence is adopted, by obtaining measured data, building plasma parameter modeling, calculating S parameters at different frequency points, constructing a residual neural network model, and using the preprocessed data set to train the model to realize the diagnosis of plasma parameters.
Through the deep learning framework, the complex relationship between plasma parameters and S parameters is automatically extracted and learned, and the problem of accurate diagnosis of electron density in falling edges is solved, improving the robustness and accuracy of diagnosis.
Smart Images
Figure CN119962356A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of electromagnetic computing, and in particular relates to a plasma parameter diagnosis method based on artificial intelligence. Background Art
[0002] When an aircraft flies at a supersonic speed at an altitude of 20km-100km above the ground, its front end generates a significant shock wave due to the high-speed impact. At the same time, due to the high-speed relative motion between the aircraft and the surrounding air, strong friction occurs. Combined with the viscous stagnation effect brought by the shock wave and the compression of the air, the heat near the aircraft increases rapidly. This temperature increase causes the gas to ionize, forming an active plasma state and forming a "plasma sheath" around the aircraft. The charged particles in this sheath have a strong effect on electromagnetic waves, such as absorption, reflection and scattering, which may cause communication interruption, which is called the "black barrier problem". For the electromagnetic field, the electron density and collision frequency of this sheath are key factors.
[0003] At present, there are two main methods for plasma parameter diagnosis based on microwave reflection method, one is based on the traditional microwave reflection diagnosis method, and the other is based on the BP neural network plasma parameter diagnosis method. Based on the traditional microwave reflection method, there are mainly plasma parameter diagnosis methods based on broadband reflection coefficient curve curvature analysis, and plasma electron density and collision frequency joint diagnosis methods based on wave impedance invariant points. The other is based on the BP neural network plasma parameter diagnosis method.
[0004] Traditional microwave reflection diagnostic methods rely on the cutoff characteristics of electromagnetic waves in plasma to evaluate electron density. Specifically, different electromagnetic wave frequencies will be totally reflected at different electron density critical interfaces. By measuring the phase changes of electromagnetic waves at these different frequencies, the electron density of the plasma can be determined. However, the limitation of this method is that it can only detect the electron density on the rising edge of the plasma distribution, but cannot accurately diagnose the electron density on the falling edge. Another plasma parameter inversion method based on BP neural network provides a new approach, but because BP neural network is prone to gradient vanishing and explosion problems, its model training effect is not ideal. Summary of the invention
[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides an artificial intelligence-based plasma parameter diagnosis method, which solves the problem that the electron density at the falling edge cannot be accurately diagnosed, and the method has high robustness.
[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is: a plasma parameter diagnosis method based on artificial intelligence, comprising the following steps:
[0007] S1. Obtain measured data and conduct plasma parameter modeling based on the influence of plasma on electromagnetic waves;
[0008] S2. Based on the modeling results of plasma parameters, calculate the S parameters at different frequencies;
[0009] S3, taking the S parameter as a sample of the data set, and taking the distribution parameter of the plasma as a label of the data set, and preprocessing the data set;
[0010] S4, constructing a residual neural network model for plasma parameter diagnosis, and using the preprocessed data set to train the residual neural network model;
[0011] S5. Use the trained residual neural network model to diagnose plasma parameters and complete artificial intelligence-based plasma parameter diagnosis.
[0012] The beneficial effects of the present invention are as follows: compared with the traditional method, the present invention automatically extracts and learns the complex relationship between plasma parameters and S parameters through the framework of deep learning, without relying on complex empirical formulas or models, and solves the problem that the electron density of the falling edge cannot be accurately diagnosed. Therefore, the present invention has higher robustness and can cope with various plasma states and conditions.
[0013] Furthermore, the step S2 comprises the following steps:
[0014] S201, selecting a wide-band antenna, and meshing the model based on the modeling results of the plasma parameters;
[0015] S202. Based on the grid division result, calculate the S parameters at different frequency points.
[0016] The beneficial effect of the above further scheme is that by calculating the broadband reflection coefficient, a foundation is laid for the subsequent diagnosis of cross-scale plasma parameters.
[0017] Furthermore, the data set is preprocessed, which specifically includes:
[0018] Adjust the format of S parameters to correspond to the residual neural network model structure;
[0019] The labels of the data set are normalized using logarithmic transformation to complete the preprocessing of the data set.
[0020] The beneficial effects of the above further scheme are: by adjusting the format of the S parameters, the residual network can be used to extract richer features, and the use of logarithmic processing can reduce the impact of the cross-order magnitude of the plasma parameters.
[0021] Furthermore, the residual neural network model includes:
[0022] A convolution layer, used to extract low-dimensional features of plasma parameters and electromagnetic wave data based on the reflection coefficient of the antenna, wherein the convolution layer includes a 7×7 convolution layer, a batch normalization layer, a ReLU activation function, and a 3×3 maximum pooling layer connected in sequence;
[0023] A residual sequence generation layer, used for extracting high-dimensional features of plasma parameters and electromagnetic wave data based on the reflection coefficient of the antenna by increasing the number of channels, wherein the residual sequence generation layer includes 4 residual sequences, and the 4 residual sequences include 3 residual units, 4 residual units, 6 residual units and 3 residual units respectively;
[0024] Feature fusion layer, used to fuse low-dimensional features and high-dimensional features;
[0025] The decision construction layer is used to integrate the fusion results using the fully connected layer and output the diagnosis results of the plasma parameters.
[0026] The beneficial effect of the above further scheme is that compared with the model based on BP neural network, the residual neural network has significant advantages. First, the residual network effectively solves the common gradient vanishing and gradient exploding problems of deep neural networks through a special "residual connection" structure, allowing the network to be deeper and thus capture more features. Secondly, this structure of the residual network makes it more stable and converges faster during the training process. This means that within the same training time, the residual network can achieve better performance. Finally, due to its excellent characteristics, the method based on the residual network makes the model have better tolerance to noise and outliers, and improves the robustness of parameter diagnosis. Therefore, the present invention combines the advancement of deep learning and the stability of the residual network to make the inversion of plasma parameters more accurate.
[0027] Furthermore, the training and optimization process of the residual neural network model is as follows:
[0028] Divide the preprocessed data set into training set, test set and validation set;
[0029] Forward propagation of a set of training set samples, inputting the training set samples into the residual neural network model under the current weight to extract features, obtaining a set of output prediction values, and calculating the root mean square error of the first set of data between the prediction value and the true value;
[0030] Back-propagating the root mean square error of the first set of data, updating the weights of the current residual neural network model, adjusting the hyperparameters, and continuing to extract features to obtain a set of new predicted values for output, and calculating the root mean square error of the second set of data between the new predicted values and the true values, until the root mean square error of the second set of data is less than a preset threshold, completing the training of the residual neural network model, wherein the training process uses the validation set to verify the trained residual neural network model, and after the training is completed, the residual neural network model is tested using the test set;
[0031] The root mean square error of the second set of data is selected as the loss function of the residual neural network model, and the parameters of the residual neural network model are optimized by the stochastic batch gradient descent method with momentum, and the learning rate of the residual neural network model is dynamically adjusted by using the cosine annealing technology.
[0032] Furthermore, the expression for optimizing the parameters of the residual neural network model is as follows:
[0033]
[0034] v t =β·v t-1 +(1-β)·g t
[0035] θ t =θ t-1 -α·v t
[0036] in, and g t Both represent the loss function J in the residual neural network model parameter θ t-1 The gradient at v t represents the velocity at time step t, β represents the momentum term, and v t-1 represents the velocity at the previous time step, θ t represents the residual neural network model parameters after time step t update, α represents the learning rate, and θ t-1 Represents the residual neural network model parameters after updating at time step t-1. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 1 The figure is a flow chart of the method of the present invention.
[0038] Figure 2 This is the structure diagram of the residual neural network model.
[0039] Figure 3 This is a performance verification diagram of the residual neural network model. DETAILED DESCRIPTION
[0040] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0041] Example
[0042] In order to solve the problems existing in the background technology, the present invention proposes a plasma parameter diagnosis method based on artificial intelligence. Compared with the BP neural network, the residual neural network effectively avoids the gradient vanishing and explosion problems by adopting residual connections, thereby making the model training more stable and effective. Its deep structure helps to capture richer data features. Relying on this method, the plasma parameters can be more accurately predicted through the measured reflection coefficient related information. Figure 1 As shown, the present invention provides a plasma parameter diagnosis method based on artificial intelligence, and its implementation method is as follows:
[0043] S1. Obtain measured data and conduct plasma parameter modeling based on the influence of plasma on electromagnetic waves;
[0044] In this embodiment, plasma parameter modeling is performed based on measured data such as flight experiments or wind tunnel experiments and in combination with the unique influence characteristics of plasma on electromagnetic waves.
[0045] The transmission of electromagnetic waves in plasma is affected by two key parameters: electron density and collision frequency. In order to understand these effects more accurately, the plasma can be layered and the electron density and collision frequency of each layer can be diagnosed, or a common distribution (such as double Gaussian distribution or double exponential electron density distribution) can be used to equivalently simulate the distribution of plasma. The present invention takes double Gaussian distribution as an example, and the formula is shown in (1):
[0046]
[0047] The key parameters involved are the plasma peak density Ne peak , plasma frequency v e , double Gaussian distribution upper and lower edge coefficients a1, a2, peak density coordinate z1, distance d1 between plasma and antenna surface, sheath thickness z t , the electron density n of the plasma at distance d e (d).
[0048] S2. Based on the modeling results of plasma parameters, the S parameters at different frequencies are calculated. The implementation method is as follows:
[0049] S201, selecting a wide-band antenna, building a model in combination with the modeling results of plasma parameters, and performing meshing;
[0050] S202. Based on the grid division result, calculate the S parameters at different frequency points.
[0051] In this embodiment, the antenna-plasma model and the corresponding grid involved can be generated separately through CAD or CAE software.
[0052] In this embodiment, an efficient full-wave electromagnetic simulation algorithm is used or S parameters at different frequencies are calculated. The present invention takes a hybrid finite element-boundary element-multilayer fast multipole algorithm (FE-BI-MLFMA, combined element) efficient full-wave electromagnetic simulation algorithm as an example.
[0053] First, a wide-band antenna is selected. Taking the wide-band horn antenna in the present invention as an example, the operating frequency range of this antenna is 2GHz-40GHz. Based on the plasma parameter data of step S1, combined with the antenna, a corresponding model is established and meshed, and then the efficient full-wave electromagnetic simulation algorithm is used to calculate the S parameters at each frequency point using the Heyuanji calculation example.
[0054] S3, taking the S parameter as a sample of the data set, and taking the distribution parameter of the plasma as a label of the data set, and preprocessing the data set;
[0055] In this embodiment, the data set is preprocessed, which is specifically as follows:
[0056] Adjust the format of S parameters to correspond to the residual neural network model structure;
[0057] The labels of the data set are normalized using logarithmic transformation to complete the preprocessing of the data set.
[0058] In this embodiment, the obtained S parameters are used as samples of the data set, and the distribution parameters of the plasma are used as labels of the data set.
[0059] In this embodiment, the data set is preprocessed as follows:
[0060] The format of the S parameters is adjusted to match the deep learning model structure. For example, if 1024 frequency points are set during simulation, the amplitude and phase of the S parameters of these frequency points will be obtained, and then converted into a 32*32*3 data set.
[0061] Label processing: Since the peak density and collision frequency of plasma are significantly higher than other parameters in magnitude (about 10-15 orders of magnitude difference), logarithmic transformation is used to normalize the data labels to speed up model training and improve prediction efficiency.
[0062] S4. Constructing a residual neural network model for plasma parameter diagnosis, and using the preprocessed data set to train the residual neural network model, wherein the residual neural network model includes:
[0063] A convolution layer, used to extract low-dimensional features of plasma parameters and electromagnetic wave data based on the reflection coefficient of the antenna, wherein the convolution layer includes a 7×7 convolution layer, a batch normalization layer, a ReLU activation function, and a 3×3 maximum pooling layer connected in sequence;
[0064] A residual sequence generation layer, used for extracting high-dimensional features of plasma parameters and electromagnetic wave data based on the reflection coefficient of the antenna by increasing the number of channels, wherein the residual sequence generation layer includes 4 residual sequences, and the 4 residual sequences include 3 residual units, 4 residual units, 6 residual units and 3 residual units respectively;
[0065] Feature fusion layer, used to fuse low-dimensional features and high-dimensional features;
[0066] The decision construction layer is used to integrate the fusion results using the fully connected layer and output the diagnosis results of the plasma parameters.
[0067] In this embodiment, the training and optimization process of the residual neural network model is as follows:
[0068] Divide the preprocessed data set into training set, test set and validation set;
[0069] Forward propagation of a set of training set samples, inputting the training set samples into the residual neural network model under the current weight to extract features, obtaining a set of output prediction values, and calculating the root mean square error of the first set of data between the prediction value and the true value;
[0070] Back-propagating the root mean square error of the first set of data, updating the weights of the current residual neural network model, adjusting the hyperparameters, and continuing to extract features to obtain a set of new predicted values for output, and calculating the root mean square error of the second set of data between the new predicted values and the true values, until the root mean square error of the second set of data is less than a preset threshold, completing the training of the residual neural network model, wherein the training process uses the validation set to verify the trained residual neural network model, and after the training is completed, the residual neural network model is tested using the test set;
[0071] The root mean square error of the second set of data is selected as the loss function of the residual neural network model, and the parameters of the residual neural network model are optimized by the stochastic batch gradient descent method with momentum, and the learning rate of the residual neural network model is dynamically adjusted by using the cosine annealing technology.
[0072] In this embodiment, the present invention divides the data set into a training set, a validation set, and a test set in a ratio of 8:1:1. The training set is used as training data to train the residual neural network model and learn data features; the validation set is used to evaluate the performance of the residual neural network model during the training process. If the performance of the residual neural network model on the validation set no longer improves within several consecutive cycles, the training can be stopped in advance to prevent overfitting; the test set is used to evaluate the final performance of the residual neural network model on unseen data, and can be used as an unbiased estimate of the expected performance of the residual neural network model in practical applications.
[0073] In this embodiment, Figure 2 As shown, the residual neural network model for plasma parameter diagnosis is constructed as follows:
[0074] Convolutional layer construction: It consists of a 7×7 convolutional layer, a batch normalization layer, a ReLU activation function, and a 3×3 maximum pooling layer in sequence to extract low-dimensional features of plasma parameters and electromagnetic wave data.
[0075] Residual sequence generation: There are 4 residual sequences in the entire network model, which are composed of 3, 4, 6 and 3 residual units respectively. Each residual unit is composed of 3 convolutional layers, batch normalization and ReLU activation functions. The first residual unit of the last three residual sequences uses 1×1 convolution, which can increase the number of channels and help extract high-dimensional features of plasma parameters and electromagnetic wave data.
[0076] Feature fusion: After all residual sequences, a global average pooling layer is introduced to fuse features of various dimensions and effectively reduce the feature dimension, thereby improving the execution efficiency of the residual neural network model.
[0077] Decision construction: A fully connected layer is used to efficiently integrate the previously extracted features and output the final diagnosis results of plasma parameters.
[0078] In this embodiment, the residual neural network model is trained and optimized as follows:
[0079] The root mean square error between the true value and the predicted value is selected as the loss function, and its expression is shown in formula (2). The parameters of the network model are optimized by the stochastic batch gradient descent algorithm with momentum to ensure that the final loss function can converge. The expression is shown in formulas (3)-(5). In addition, the cosine annealing technology is used to dynamically adjust the learning rate, combined with the ten-fold cross-validation method to improve the accuracy, robustness and generalization of model prediction. During the training process, the validation set is used to evaluate the model performance. If the performance of the model on the validation set does not improve in several consecutive cycles, the training is stopped in advance to prevent overfitting.
[0080]
[0081] Where n represents the number of data samples, y i represents the true value of the i-th prediction, represents the i-th predicted value.
[0082]
[0083] The update speed of the parameters is:
[0084] v t =β·v t-1 +(1-β)·g t , (4)
[0085] θ t =θ t-1 -α·v t , (5)
[0086] in, and g t Both represent the loss function J in the residual neural network model parameter θ t-1 The gradient at v t represents the velocity at time step t, β represents the momentum term, and v t-1 represents the velocity at the previous time step, θ t represents the residual neural network model parameters after time step t update, α represents the learning rate, and θ t-1 Represents the residual neural network model parameters after updating at time step t-1.
[0087] In this embodiment, after the training is completed, the final performance of the residual neural network model is verified using a test set that was not used in the training, in order to illustrate its performance in practical applications and comprehensively evaluate the generalization of the model. Figure 3 .
[0088] In this embodiment, in addition to using the amplitude and phase of the S parameter, the real part and imaginary part of the S parameter may also be used for analysis.
[0089] S5. Use the trained residual neural network model to diagnose plasma parameters and complete artificial intelligence-based plasma parameter diagnosis.
[0090] In summary, through the above design, the present invention can more accurately predict plasma parameters through the measured reflection coefficient related information.
Claims
1. A plasma parameter diagnosis method based on artificial intelligence, characterized in that: The following steps are involved: S1. Obtain measured data and conduct plasma parameter modeling based on the influence of plasma on electromagnetic waves; S2. Based on the modeling results of plasma parameters, calculate the S parameters at different frequencies; S3, taking the S parameter as a sample of the data set, and taking the distribution parameter of the plasma as a label of the data set, and preprocessing the data set; S4, constructing a residual neural network model for plasma parameter diagnosis, and using the preprocessed data set to train the residual neural network model; S5. Use the trained residual neural network model to diagnose plasma parameters and complete artificial intelligence-based plasma parameter diagnosis.
2. The plasma parameter diagnosis method based on artificial intelligence according to claim 1, characterized in that: The step S2 comprises the following steps: S201, selecting a wide-band antenna, building a model in combination with the modeling results of plasma parameters, and performing meshing; S202. Based on the grid division result, calculate the S parameters at different frequency points.
3. The plasma parameter diagnosis method based on artificial intelligence according to claim 1, characterized in that: The data set is preprocessed, specifically: Adjust the format of S parameters to correspond to the residual neural network model structure; The labels of the data set are normalized using logarithmic transformation to complete the preprocessing of the data set.
4. The plasma parameter diagnosis method based on artificial intelligence according to claim 1, characterized in that: The residual neural network model includes: A convolution layer, used to extract low-dimensional features of plasma parameters and electromagnetic wave data based on the reflection coefficient of the antenna, wherein the convolution layer includes a 7×7 convolution layer, a batch normalization layer, a ReLU activation function, and a 3×3 maximum pooling layer connected in sequence; A residual sequence generation layer, used for extracting high-dimensional features of plasma parameters and electromagnetic wave data based on the reflection coefficient of the antenna by increasing the number of channels, wherein the residual sequence generation layer includes 4 residual sequences, and the 4 residual sequences include 3 residual units, 4 residual units, 6 residual units and 3 residual units respectively; Feature fusion layer, used to fuse low-dimensional features and high-dimensional features; The decision construction layer is used to integrate the fusion results using the fully connected layer and output the diagnosis results of the plasma parameters.
5. The plasma parameter diagnosis method based on artificial intelligence according to claim 1, characterized in that: The training and optimization process of the residual neural network model is as follows: Divide the preprocessed data set into training set, test set and validation set; Forward propagation of a set of training set samples, inputting the training set samples into the residual neural network model under the current weight to extract features, obtaining a set of output prediction values, and calculating the root mean square error of the first set of data between the prediction value and the true value; Back-propagating the root mean square error of the first set of data, updating the weights of the current residual neural network model, adjusting the hyperparameters, and continuing to extract features to obtain a set of new predicted values for output, and calculating the root mean square error of the second set of data between the new predicted values and the true values, until the root mean square error of the second set of data is less than a preset threshold, completing the training of the residual neural network model, wherein the training process uses the validation set to verify the trained residual neural network model, and after the training is completed, the residual neural network model is tested using the test set; The root mean square error of the second set of data is selected as the loss function of the residual neural network model, and the parameters of the residual neural network model are optimized by the stochastic batch gradient descent method with momentum, and the learning rate of the residual neural network model is dynamically adjusted by using the cosine annealing technology.
6. The plasma parameter diagnosis method based on artificial intelligence according to claim 1, characterized in that: The expression for optimizing the parameters of the residual neural network model is as follows: v t =β·v t-1 +(1-β)·g t i t =θ t-1 -a·v t in, and g t Both represent the loss function J in the residual neural network model parameter θ t-1 The gradient at v t represents the velocity at time step t, β represents the momentum term, and v t-1 represents the velocity at the previous time step, θ t represents the residual neural network model parameters after time step t update, α represents the learning rate, and θ t-1 Represents the residual neural network model parameters after updating at time step t-1.