Tokamak plasma rupture prediction model construction method and application

Abstract

The invention belongs to the technical field of Tokamak plasma rupture prediction, and discloses a Tokamak plasma rupture prediction model construction method and application, and the method comprises the steps: employing a training data set to train a multi-mode convolution self-coding neural network, and obtaining a rupture prediction model; wherein one training sample in the training data set is a slice of different diagnosis signals at the current moment; the multi-modal convolution self-coding neural network comprises a convolution feature extraction module, a splicing module, a feature self-coding module, a segmentation module and a deconvolution feature reconstruction module. According to the method, the constructed prediction model can process data of various modes and automatically perform feature extraction; and the model can be trained by only using non-fracture samples, and also can use fracture precursor samples when fractures occur, so that the model can adapt to a data environment in which the sample proportion on a future device is uncertain.

Description

Technical Field

[0001] The present invention belongs to the technical field of tokamak plasma rupture prediction, and more specifically, relates to a tokamak plasma rupture prediction model construction method and application. Background Art

[0002] The tokamak device uses a magnetic field to confine deuterium and tritium in the form of plasma to carry out nuclear fusion reactions. When the device is running, the plasma is often subject to various interferences, which makes its operating state unstable and causes sudden quenching. This phenomenon is called rupture. The sudden release of energy in the plasma during rupture will damage the device. Future devices (such as ITER) will be larger in scale, and the energy in the plasma will also be greater, so the harm caused by rupture will be greater. The instability of the plasma can be observed before the rupture occurs, which makes it possible to predict the rupture. During the operation of the tokamak device, the state of the plasma is monitored by sensors detecting signals of multiple modes such as electricity, magnetism, and radiation of the plasma. For different types of ruptures, the diagnostic signals with precursors are not necessarily the same, so generally a predictor is built by combining multiple diagnostic signals related to rupture. After the predictor recognizes the precursor, an alarm is issued to activate the mitigation system to reduce damage to the tokamak device.

[0003] The current tokamak plasma rupture theory is not fully understood, and it is difficult to establish a reliable physical-driven rupture predictor. Therefore, data-driven algorithms based on machine learning are currently used to achieve rupture prediction. Such predictors automatically learn information related to plasma rupture from a large amount of historical discharge data on existing devices and use it for rupture prediction, which solves the defects of physical-driven methods. A variety of machine learning algorithms, including neural networks and support vector machines, have been used to develop rupture predictors and deployed on existing devices such as J-TEXT, EAST, DIII-D, and JET. The performance of some predictors can reach a success rate of 90% and a false alarm rate of less than 10%. However, these rupture predictors are mainly based on supervised learning methods and require rich and balanced rupture precursor samples and non-rupture sample training. The cost of obtaining rupture cannons on large-scale devices in the future is extremely high, and there are very few rupture precursor samples available in the database in the early stage of operation of new devices. This extremely unbalanced sample ratio will reduce the performance of the supervised learning model or even make it invalid.

[0004] In order to solve the defects of supervised learning predictors, the research of rupture predictors based on anomaly detection has been spawned. This type of rupture predictor is trained only with easily accessible non-rupture samples, overcoming the sample imbalance problem of supervised learning predictors. However, since the samples for rupture prediction are high-dimensional sample data of different modalities, that is, the format of the input diagnostic signal, such as type, length, number of channels and sampling frequency, is different, it is impossible to directly use such complex high-dimensional data for automatic feature extraction; directly using high-dimensional original samples for training and reasoning requires large amount of calculation and the process is extremely time-consuming; and the use of manual feature extraction methods may lead to the loss of key information in the data. In the prior art, there is a method of splicing diagnostic signals of different modalities together into a large two-dimensional feature map and then performing feature self-encoding to predict rupture. Since all diagnostic signals are spliced ​​together, the correlation between the prediction results and the model input is unclear and lacks interpretability. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a method and application for constructing a tokamak plasma rupture prediction model, the purpose of which is to construct a tokamak plasma rupture prediction model that can automatically extract features from multimodal diagnostic signals and improve the interpretability of the model.

[0006] To achieve the above object, according to a first aspect of the present invention, a method for constructing a tokamak plasma disruption prediction model is provided, comprising:

[0007] A multimodal convolutional autoencoder neural network is trained using a training data set to obtain a rupture prediction model; wherein a training sample in the training data set is a slice of different diagnostic signals at the current moment; the multimodal convolutional autoencoder neural network includes:

[0008] A convolution feature extraction module is used to extract features and reduce the dimension of each diagnostic signal slice in the current training sample to obtain the features of each diagnostic signal slice;

[0009] A splicing module, used for flattening and splicing the features of each diagnostic signal slice to form a one-dimensional vector containing the features of each diagnostic signal slice;

[0010] A feature self-encoding module, used for compressing and reconstructing the one-dimensional vector;

[0011] A segmentation module, used to segment the reconstructed one-dimensional vector;

[0012] A deconvolution feature reconstruction module, used to restore the reconstructed features of each diagnostic signal slice obtained after segmentation to the corresponding reconstructed slice;

[0013] The anomaly detection module is used to calculate the reconstruction error between the restored reconstructed slice and the slice of the corresponding input diagnostic signal to obtain the reconstruction error corresponding to the current training sample; based on the reconstruction error corresponding to the current training sample, determine whether the current training sample is a rupture precursor sample.

[0014] Furthermore, during the training of the multimodal convolutional autoencoder neural network, the loss function used is:

[0015]

[0016] Where n is the number of training samples, y is the label of the training sample, the label of the non-rupture sample is 1, and the label of the rupture precursor sample is 0; RCE i MSE represents the average mean square error between each diagnostic signal slice of the i-th training sample and its reconstructed slice, which is used to characterize the reconstruction error corresponding to the i-th training sample; i,ip , MSE i,bt , …, MSE i,mirnov Respectively represent the mean square error between each diagnostic signal slice and its reconstructed slice in the i-th training sample; ip, bt, …, mirnov is used to represent the input n diag diagnostic signal; α is the weight assigned to the rupture precursor sample, which is a hyperparameter; β is the dynamic balance parameter between non-rupture samples and rupture precursor samples, which is used to stabilize the attention of the rupture prediction model to the rupture precursor samples; n N 、n D are the number of non-rupture samples and rupture precursor samples in the training sample set, respectively.

[0017] Furthermore, the convolution feature extraction module and the deconvolution feature reconstruction module are mirror images of each other.

[0018] Further, the diagnostic signal includes a single-channel diagnostic signal and a multi-channel diagnostic signal;

[0019] When the diagnostic signal is a single-channel diagnostic signal, the convolution feature extraction module uses multi-layer one-dimensional convolution to extract the time characteristics of the diagnostic signal slice, and the deconvolution feature reconstruction module uses multi-layer one-dimensional deconvolution to restore the slice;

[0020] When the diagnostic signal is a multi-channel diagnostic signal, the convolution feature extraction module uses multi-layer two-dimensional convolution to extract the temporal features and spatial features of the diagnostic signal slices, and the deconvolution feature reconstruction module uses multi-layer two-dimensional deconvolution to restore the slices.

[0021] Further, the length of the diagnostic signal slice is the product of the sampling frequency of the current diagnostic signal and the time window selected during slicing;

[0022] The sampling frequency is determined as follows:

[0023] Perform spectrum analysis on the current diagnostic signal to determine the frequency range of spectrum difference between the rupture precursor sample and the non-rupture sample; and set the sampling frequency to twice the frequency range of the spectrum difference.

[0024] Further, judging whether the current training sample is a rupture precursor sample based on the reconstruction error corresponding to the current training sample includes:

[0025] The reconstruction error corresponding to the current training sample is compared with the reconstruction error warning threshold, and the training sample corresponding to the reconstruction error exceeding the reconstruction error warning threshold is the rupture precursor sample.

[0026] According to a second aspect of the present invention, a method for real-time prediction of tokamak plasma disruption is provided, comprising:

[0027] The slice data of each diagnostic signal at the current moment is input into the rupture prediction model, and whether the slice data at the current moment is rupture precursor data is output, so as to realize the real-time prediction of tokamak plasma rupture; wherein, the rupture prediction model is constructed by the tokamak plasma rupture prediction model construction method described in any one of the first aspects.

[0028] According to a third aspect of the present invention, there is provided an electronic device comprising a computer-readable storage medium and a processor;

[0029] The computer-readable storage medium is used to store executable instructions;

[0030] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the tokamak plasma disruption prediction model construction method described in any one of the first aspects, or / and, to execute the tokamak plasma disruption real-time prediction method described in the second aspect.

[0031] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the method for constructing a tokamak plasma disruption prediction model as described in any one of the first aspects, or / and the method for real-time prediction of tokamak plasma disruption as described in the second aspect is implemented.

[0032] According to a fifth aspect of the present invention, a computer program product is provided. When the computer program product is run on a computer, the computer executes the method for constructing a tokamak plasma disruption prediction model as described in any one of the first aspects, or / and executes the tokamak plasma disruption real-time prediction method as described in the second aspect.

[0033] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects:

[0034] (1) In the present invention, by time slicing the multimodal diagnostic signal, each training sample includes slices of different diagnostic signals at the current moment, and the features of each diagnostic signal slice in each training sample are extracted in turn by the convolution feature extraction module, and the features of each diagnostic signal slice are flattened and spliced ​​before being input into the feature self-encoding module for compression and reconstruction. Correspondingly, the compressed and reconstructed one-dimensional vector is segmented, and the reconstructed features are restored to the corresponding reconstructed slices. Based on this, the anomaly detection module can distinguish whether the current input sample is a rupture precursor sample by calculating the reconstruction error between the slices of each reconstructed diagnostic signal and the slices of the corresponding diagnostic signal of the original input, thereby realizing the prediction of the rupture of the tokamak plasma. The method of the present invention realizes automatic feature extraction of complex high-dimensional multimodal diagnostic signals; and the anomaly detection module predicts rupture by comparing the reconstruction error between the input and output corresponding to each diagnostic signal, and the reconstruction error between the overall output and input of the model. In this way, the contribution of each input diagnostic signal to the prediction result can be obtained by comparison. Therefore, the present invention can reflect the correlation between the prediction model input and the prediction result, and improve the interpretability of the prediction model.

[0035] (2) As a preferred embodiment, based on the optimized loss function, the rupture prediction model based on anomaly detection in the present invention can be trained using only non-rupture samples, and can also be trained using rupture precursor samples when rupture occurs, thereby improving the data utilization efficiency of the model. In addition, a balance coefficient is introduced into the loss function so that the model's attention to each type of sample remains unchanged during each retraining, that is, the model remains stable. Therefore, the prediction model in the present invention can adapt to the unpredictable data environment on new devices in the future, efficiently utilize data, and improve rupture prediction performance.

[0036] (3) As a preferred method, each diagnostic signal is downsampled by Shannon sampling theorem to reduce the amount of model calculation and ensure information integrity.

[0037] In summary, the prediction model constructed in the present invention can process data of multiple modalities and automatically extract features; and the model can be trained only with non-rupture samples, and can also use rupture precursor samples when rupture occurs. It is a model that can adapt to data environments with uncertain sample ratios on future devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the multimodal convolutional autoencoder neural network structure in an embodiment of the present invention.

[0039] Figure 2It is a schematic diagram of the structure of the convolution feature extraction module and the deconvolution feature reconstruction module in the embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0041] In the present invention, the terms "first", "second", etc. in the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0042] Example 1

[0043] like Figure 1 and Figure 2 As shown, the method for constructing a tokamak plasma disruption prediction model provided in an embodiment of the present invention mainly includes:

[0044] The multimodal convolutional autoencoder neural network is trained using a training data set to obtain a rupture prediction model; wherein a training sample in the training data set is a slice of different diagnostic signals at the current moment; the multimodal convolutional autoencoder neural network includes:

[0045] A convolution feature extraction module is used to extract features and reduce the dimension of each diagnostic signal slice in the current training sample to obtain the features of each diagnostic signal slice; wherein the number of convolution feature extraction modules is consistent with the number of input diagnostic signals;

[0046] A splicing module, used for flattening and splicing the features of each diagnostic signal slice to form a one-dimensional vector containing the features of each diagnostic signal slice;

[0047] A feature self-encoding module is used to compress and reconstruct the one-dimensional vector to obtain a reconstructed one-dimensional vector;

[0048] A segmentation module, used for segmenting the reconstructed one-dimensional vector to obtain the reconstructed features of each diagnostic signal slice;

[0049] A deconvolution feature reconstruction module, used to restore the reconstructed features of each diagnostic signal slice to the corresponding reconstructed slice;

[0050] The anomaly detection module is used to calculate the similarity between the slice reconstructed by the deconvolution feature reconstruction module and the slice of the corresponding input diagnostic signal, obtain the average similarity between all input diagnostic signals and their corresponding reconstructed signals, and use the average similarity as the reconstruction error; and compare the reconstruction error with the reconstruction error warning threshold. The input sample corresponding to the value exceeding the reconstruction error warning threshold is the rupture precursor sample, thereby realizing the tokamak plasma rupture prediction.

[0051] Specifically, the structure of the multimodal convolutional autoencoder neural network is related to the form of the input samples. Since each diagnostic signal is a high-dimensional signal of different modes, the signal type, number of channels, sampling frequency, data length, etc. are all different. Taking the diagnostic signal on the J-TEXT reactor as an example, the selected diagnosis is divided into two categories: single-channel diagnosis (such as: Ip, Bt, etc.) and multi-channel diagnosis (for example: Mirnov array, ne array, etc.). Single-channel diagnosis has only one time series and only contains the time characteristics of the diagnosis. Multi-channel diagnosis refers to multiple time series (channels) collected by an observation array composed of multiple sensors, including the time and space characteristics of the diagnosis. Each convolution feature extraction module is adapted to the characteristics of its corresponding diagnostic signal.

[0052] Each sample is composed of slices of each diagnosis, and the length of each slice is the product of the sampling rate of the corresponding diagnosis and the time window selected when slicing. The last point of each diagnostic slice represents the same moment, indicating the sampling time of the sample. Different diagnostic signals are diagnostic signals after downsampling the diagnostic signals collected by each sensor in the tokamak device. In an embodiment of the present invention, the diagnostic signals collected by each sensor are downsampled by Shannon sampling theorem to reduce the amount of model calculation and ensure information integrity. The sampling frequency is determined by the following method:

[0053] Perform spectrum analysis on the current diagnostic signal to determine the frequency range of the spectrum difference between the rupture precursor sample and the non-rupture sample; set the sampling frequency to twice the spectrum difference frequency. For example, perform spectrum analysis on the Mirnov signal on J-TEXT and observe that the spectrum of the rupture precursor sample and the non-rupture sample differs significantly within 30kHz, so the sampling rate of the Mirnov signal is set to 60kHz. The sampling rate selection of other diagnostic signals is similar, and the minimum sampling rate is set to 1kHz.

[0054] The selected time windows for different diagnostic signals are not necessarily the same. In the embodiment of the present invention, for diagnostic signals without oscillation characteristics, the time window is set to 10ms; for diagnostic signals with oscillation characteristics, the time window must be at least longer than one period of the main harmonic oscillation of the diagnostic signal.

[0055] In the embodiment of the present invention, the selected diagnostic signals and slices related to rupture are shown in Table 1.

[0056] Table 1 Diagnostic signs and sections related to rupture

[0057]

[0058] As a preference, Figure 2 As shown, in the embodiment of the present invention, since the shapes of the diagnostic slices are not necessarily the same, the structures of the convolution feature extraction module and the deconvolution feature reconstruction module are not necessarily the same. Diagnostic signals can generally be divided into two categories: single-channel diagnosis and multi-channel diagnosis. For slices of single-channel diagnostic signals, the feature extraction module uses multiple layers of one-dimensional convolution to extract time features, while the deconvolution feature reconstruction module uses multiple layers of one-dimensional deconvolution to restore the slices. For multi-channel diagnostic slices, the feature extraction module uses multiple layers of two-dimensional convolution to extract features including time and space, and the deconvolution feature reconstruction module uses multiple layers of two-dimensional deconvolution to restore the slices. For the same diagnostic signal, the structures of the corresponding convolution feature extraction module and deconvolution feature reconstruction module are mirror images of each other.

[0059] Preferably, the feature self-encoding module is a fully connected network with multiple layers connected in sequence. In the embodiment of the present invention, a five-layer fully connected network is used, and the number of nodes in each layer of the fully connected network is 992, 400, 200, 400, and 992, respectively. It can be adjusted according to actual needs.

[0060] As one of the specific implementation methods, during the training process, if traditional anomaly detection is performed, that is, when all training samples are non-fracture samples, the loss function used is:

[0061]

[0062] Where n is the number of training samples, RCE i represents the MSE average between each diagnostic signal slice corresponding to the i-th training sample and its corresponding reconstructed slice, which is used to characterize the reconstruction error (RCE) of the i-th training sample returned by the network; n diag is the number of diagnostic signals selected in the input training sample set; MSE i,ip , MSE i,bt , …, MSE i,mirnov They represent the mean square error (MSE) between each diagnostic signal slice and its corresponding reconstructed slice in the current i-th training sample, and are used to characterize the similarity between the reconstructed diagnostic signal slice and its corresponding input diagnostic signal slice. For example, MSE i,ip It represents the similarity between the diagnostic signal slice reconstructed by the diagnostic signal ip in the i-th training sample at the current moment and the slice input by the diagnostic signal ip at the current moment.

[0063] During the training process, if the training samples are all non-rupture samples, the model aims to minimize the RCE of non-rupture samples. When predicting, the model returns a smaller RCE for non-rupture samples; the RCE returned for rupture precursor samples that have not been learned is relatively large, and the RCE warning threshold is set to distinguish rupture precursor samples. Among them, the RCE warning threshold can be obtained as a hyperparameter through training.

[0064] As a further design of the present invention, the present invention takes into account that in the future, when a new device starts to operate from the first shot, the available data accumulates and increases over time. If the model can only be trained using non-rupture samples, the data utilization rate is low, resulting in great waste. In addition, when the prediction model in the present invention is deployed on a new device using an adaptive learning method from scratch, it is necessary to regularly use the training set to retrain the model. Since the number of ruptures is unpredictable, the ratio of rupture precursor samples and non-rupture samples in the training set will change each time the training is retrained, that is, the model's attention to each type of sample will change, which will cause the trained model to be unstable and affect the rupture prediction performance of the model. Therefore, in an embodiment of the present invention, when training the rupture prediction model, an optimized loss function is used, so that the rupture prediction model based on anomaly detection can be trained using both non-rupture samples and rupture precursor samples when rupture occurs, thereby improving the data utilization efficiency of the model, and being able to adapt to the data environment with uncertain sample ratios on future devices, thereby improving the model prediction performance.

[0065] The optimized loss function is:

[0066]

[0067] Among them, y is the label of the input sample, the label of the non-rupture sample is 1, and the label of the rupture precursor sample is 0. α is the weight assigned to the precursor sample, which is a hyperparameter. β is the dynamic balance parameter of non-rupture samples and rupture precursor samples, which is used to stabilize the model's attention to the precursor samples. The model automatically calculates it before retraining the model. N 、n D are the number of non-rupture samples and rupture precursor samples in the training set. During training, when there are no rupture precursor samples, the model only minimizes the RCE of non-rupture samples; when there are rupture precursor samples, the model minimizes the RCE of precursor samples -1 , that is, to maximize the RCE of the rupture precursor sample.

[0068] Based on the optimized loss function, the rupture prediction model based on anomaly detection in the present invention can be trained using only non-rupture samples, and can also be trained using rupture precursor samples when rupture occurs, thereby improving the data utilization efficiency of the model. In addition, the balance coefficient is introduced into the loss function so that the model's attention to each type of sample remains unchanged during each retraining, that is, the model remains stable. Therefore, the prediction model in the present invention can adapt to the unpredictable data environment on new devices in the future, efficiently utilize data, and improve rupture prediction performance.

[0069] The rupture prediction model after the training is completed can be used for real-time rupture prediction. In real-time rupture prediction, the predictor (trained rupture prediction model) analyzes the discharge samples in chronological order to see whether they are abnormal samples (rupture precursor samples).

[0070] In the present invention, by time slicing the multimodal diagnostic signal, each training sample includes slices of different diagnostic signals at the current moment, and the features of each diagnostic signal slice in each training sample are extracted in turn by the convolution feature extraction module, and the features of each diagnostic signal slice are flattened and spliced, and then input into the feature self-encoding module for compression and reconstruction, and correspondingly, the compressed and reconstructed one-dimensional vector is segmented, and the reconstructed features are restored to the corresponding reconstructed slices. Based on this, the anomaly detection module can distinguish whether the current input sample is a rupture precursor sample by calculating the similarity between the slices of each reconstructed diagnostic signal and the slices of the corresponding diagnostic signal of the original input, so as to realize the prediction of the rupture of the tokamak plasma. The method of the present invention realizes automatic feature extraction of complex high-dimensional multimodal diagnostic signals; and the anomaly detection module predicts rupture by comparing the similarity of the difference between the input and output corresponding to each diagnostic signal, and the similarity between the overall output and input of the model. In this way, the contribution of each input diagnostic signal to the prediction result can be obtained by comparison, which improves the interpretability of the prediction model.

[0071] Example 2

[0072] The embodiment of the present invention provides a real-time prediction method for tokamak plasma rupture, comprising:

[0073] The slice data of different diagnostic signals at the current moment are input into the rupture prediction model to obtain whether the slice data at the current moment is rupture precursor data, so as to realize the real-time prediction of the rupture of the tokamak plasma; wherein the rupture prediction model is constructed by the tokamak plasma rupture prediction model construction method in Example 1. For details, please refer to the description in Example 1, which will not be repeated here.

[0074] Example 3

[0075] An embodiment of the present invention provides an electronic device, including a computer-readable storage medium and a processor;

[0076] Computer-readable storage media for storing executable instructions;

[0077] The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the method for constructing a tokamak plasma disruption prediction model in Example 1, or / and to execute the tokamak plasma disruption real-time prediction method in Example 2. For related technical solutions, please refer to the description in Example 1 and Example 2.

[0078] Example 4

[0079] The embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the program is executed by a processor, the method for constructing a tokamak plasma disruption prediction model as in embodiment 1 is implemented, or / and, the method for real-time prediction of tokamak plasma disruption as in embodiment 2 is implemented. For related technical solutions, see the descriptions in embodiments 1 and 2.

[0080] Example 5

[0081] The embodiment of the present invention provides a computer program product. When the computer program product is run on a computer, the computer executes the method for constructing a tokamak plasma disruption prediction model in embodiment 1, or / and executes the tokamak plasma disruption real-time prediction method in embodiment 2. For related technical solutions, see the description in embodiments 1 and 2.

[0082] The rupture prediction model in the embodiment of the present invention is recorded as an E-CAAD predictor. The E-CAAD predictor is a deep learning model based on a multimodal convolutional autoencoder neural network structure, which compresses and reconstructs the input samples. In order to enable the model to automatically extract features, the structure of the neural network is designed according to the characteristics of each diagnostic signal on the tokamak device. Based on the processed diagnostic signal data, a convolutional feature extraction module and a deconvolution feature reconstruction module are added to the network. The convolutional feature extraction module can reduce the dimension of the samples of each input mode to obtain the diagnostic features corresponding to each mode; the obtained diagnostic features will be imported into the fully connected feature autoencoder module for comprehensive analysis and reconstruction; and the deconvolution feature reconstruction module will restore the reconstructed features to the original signal samples, that is, the reconstructed samples. The final predictor realizes rupture prediction by comparing the reconstruction error (RCE) of the input multimodal original samples and the reconstructed multimodal samples.

[0083] When the model detects anomalies, when it is trained only with non-rupture samples, the model has good reconstruction performance for non-rupture samples and a small RCE; the reconstruction performance for rupture precursor samples that have not been learned is poor and the RCE is large. Rupture prediction is achieved by setting the RCE threshold. Furthermore, by optimizing the loss function of the model, it can be trained using only non-rupture samples, and can also be trained using rupture precursor samples when rupture occurs, thereby improving the data utilization efficiency of the model. In addition, a balance coefficient is introduced into the loss function so that the model's attention to each type of sample remains unchanged during each retraining, that is, the model remains stable. Therefore, the E-CAAD model can adapt to the unpredictable data environment on new devices in the future, efficiently utilize data, and achieve rupture prediction.

[0084] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a tokamak plasma disruption prediction model, characterized in that: include: A multimodal convolutional autoencoder neural network is trained using a training data set to obtain a rupture prediction model; wherein a training sample in the training data set is a slice of different diagnostic signals at the current moment; the multimodal convolutional autoencoder neural network includes: A convolution feature extraction module is used to extract features and reduce the dimension of each diagnostic signal slice in the current training sample to obtain the features of each diagnostic signal slice; A splicing module, used for flattening and splicing the features of each diagnostic signal slice to form a one-dimensional vector containing the features of each diagnostic signal slice; A feature self-encoding module, used for compressing and reconstructing the one-dimensional vector; A segmentation module, used to segment the reconstructed one-dimensional vector; A deconvolution feature reconstruction module, used to restore the reconstructed features of each diagnostic signal slice obtained after segmentation to the corresponding reconstructed slice; The anomaly detection module is used to calculate the reconstruction error between the restored reconstructed slice and the slice of the corresponding input diagnostic signal to obtain the reconstruction error corresponding to the current training sample; based on the reconstruction error corresponding to the current training sample, determine whether the current training sample is a rupture precursor sample.

2. The method for constructing a tokamak plasma disruption prediction model according to claim 1, characterized in that: During the training of the multimodal convolutional autoencoder neural network, the loss function used is: Where n is the number of training samples, y is the label of the training sample, the label of the non-rupture sample is 1, and the label of the rupture precursor sample is 0; RCE i MSE represents the average mean square error between each diagnostic signal slice of the i-th training sample and its reconstructed slice, which is used to characterize the reconstruction error corresponding to the i-th training sample; i,ip , MSE i,bt , …, MSE i,mirnov Respectively represent the mean square error between each diagnostic signal slice and its reconstructed slice in the i-th training sample; ip, bt, …, mirnov is used to represent the input n diag diagnostic signal; α is the weight assigned to the rupture precursor sample, which is a hyperparameter; β is the dynamic balance parameter between the non-rupture sample and the rupture precursor sample, which is used to stabilize the attention of the rupture prediction model to the rupture precursor sample; n N 、n D are the number of non-rupture samples and rupture precursor samples in the training sample set, respectively.

3. The method for constructing a tokamak plasma disruption prediction model according to claim 1 or 2, characterized in that: The convolution feature extraction module and the deconvolution feature reconstruction module are mirror images of each other.

4. The method for constructing a tokamak plasma disruption prediction model according to claim 3, characterized in that: The diagnostic signal includes a single-channel diagnostic signal and a multi-channel diagnostic signal; When the diagnostic signal is a single-channel diagnostic signal, the convolution feature extraction module uses multi-layer one-dimensional convolution to extract the time characteristics of the diagnostic signal slice, and the deconvolution feature reconstruction module uses multi-layer one-dimensional deconvolution to restore the slice; When the diagnostic signal is a multi-channel diagnostic signal, the convolution feature extraction module uses multi-layer two-dimensional convolution to extract the temporal features and spatial features of the diagnostic signal slices, and the deconvolution feature reconstruction module uses multi-layer two-dimensional deconvolution to restore the slices.

5. The method for constructing a tokamak plasma disruption prediction model according to claim 1, characterized in that: The length of the diagnostic signal slice is the product of the sampling frequency of the current diagnostic signal and the time window selected during slicing; The sampling frequency is determined as follows: Perform spectrum analysis on the current diagnostic signal to determine the frequency range of spectrum difference between the rupture precursor sample and the non-rupture sample; and set the sampling frequency to twice the frequency range of the spectrum difference.

6. The method for constructing a tokamak plasma disruption prediction model according to claim 1, characterized in that: Judging whether the current training sample is a rupture precursor sample based on the reconstruction error corresponding to the current training sample includes: The reconstruction error corresponding to the current training sample is compared with the reconstruction error warning threshold, and the training sample corresponding to the reconstruction error exceeding the reconstruction error warning threshold is the rupture precursor sample.

7. A real-time prediction method for tokamak plasma rupture, characterized in that: include: The slice data of each diagnostic signal at the current moment is input into the rupture prediction model, and whether the slice data at the current moment is rupture precursor data is output, so as to realize the real-time prediction of tokamak plasma rupture; wherein, the rupture prediction model is constructed by the tokamak plasma rupture prediction model construction method according to any one of claims 1 to 6.

8. An electronic device, characterized in that: comprising a computer readable storage medium and a processor; The computer-readable storage medium is used to store executable instructions; The processor is used to read the executable instructions stored in the computer-readable storage medium to execute the tokamak plasma disruption prediction model construction method described in any one of claims 1-6, or / and, to execute the tokamak plasma disruption real-time prediction method described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, the method for constructing a tokamak plasma disruption prediction model as described in any one of claims 1 to 6 is implemented, or / and, the method for real-time prediction of tokamak plasma disruption as described in claim 7 is implemented.

10. A computer program product, characterized in that When the computer program product is run on a computer, the computer is enabled to execute the method for constructing a tokamak plasma disruption prediction model according to any one of claims 1 to 6, or / and execute the tokamak plasma disruption real-time prediction method according to claim 7.

Images