Research On The J-TEXT Plasma Electron Temperature Prediction And Automatic Locked Mode Classification Based On Machine Learning

The tokamak plasma information needs to be obtained with the help of various diagnostic systems,but the normal operation of the diagnostic system will encounter severe challenges in its strong electromagnetic radiation environment.For example,the failure of the diagnostic system makes it impossible to detect important information inside the plasma,which will greatly affect the physical research of the plasma.But the diagnostic signal is a multidirectional mapping of plasma information,which has various connections.Therefore,other diagnostic signals can be used to infer the missing diagnostic information and eliminate the impact on physical research.This paper uses machine learning to study the fitting relationship between plasma parameters,and uses other plasma diagnostic information to estimate electron temperature on the J-TEXT,which can make up for the research impact caused by the lack of diagnostic signals for measuring electron temperature.On the basis of this research,this paper also uses the diagnostic signal to carry out the model study of automatic classification,recognition and labeling of phenomena such as locked mode,which is helpful to fully excavate the internal relations of the plasma and further study the physical mechanism of the Magnetohydrodynamic instability(MHD).This paper analyzes the relationship between the plasma diagnostic signal and the electron temperature,and uses the fully-connected Back-Propagation Neural Network(BPNN)and the Generalized Regression Neural Network(GRNN)to build the model.The model uses electron cyclotron radiation measurement or plasma electron temperature detected by X-ray Imaging Crystal Spectrometer(XICS)as the output target,and other plasma diagnostic signals(soft Xray radiation intensity,electron density,plasma current,loop voltage,and toroidal magnetic field strength,etc.)as input parameters.As a result,a neural network model for predicting plasma electron temperature is trained and established,which realize the estimation of the electron temperature and the reconstruction of its profile.In addition,it analyzes the sensitivity of plasma parametersto electron temperature.The BPNN model can predict the electron temperature through the basic parameters of the plasma,and even some electron temperature perturbation signals guided by MHD(such as sawtooth oscillation)can be well predicted,and the average error is within 5%.It trained multiple network models to predict the electron temperature at different radial positions,thereby reconstructing the electron temperature profile.It compares the prediction effects of two kinds of neural networks on the electron temperature.BPNN has a better prediction effect on the perturbation of the electron temperature,and GRNN has a faster prediction speed on the equilibrium temperature.The method verifies that the information in plasma diagnosis is interrelated.It can use neural network to predict between diagnostic signals.Even if the relevant diagnosis fails and therefore cannot provide effective information,machine learning can still be used to achieve the acquisition of key plasma information,providing support for plasma information acquisition.This paper studies the relationship between plasma information and magnetic fluid instability,and uses cluster analysis and Support Vector Machine(SVM)to achieve automatic classification and recognition of MHD in plasma.It uses unsupervised learning K-means and hierarchical clustering model to study the classification and recognition effects of various diagnostic signals on MHD in plasma.It chooses signals such as magnetic perturbation,soft Xray radiation and electron density as input parameters to realize real-time classification and recognition and automatic marking of phenomena such as locked mode,penetration mode,sawtooth,and tearing mode in the plasma,with an accuracy rate of 95 %the above.At the same time,locked mode is one of the factors that cause plasma disrupture,which seriously endangers the safe operation of the fusion device,and is one of the important contents of plasma physics research.This paper uses supervised learning to train the SVM model to identify and classify various locked mode,and the correct rate is over 94%.The method achieves accurate classification,identification and labeling of various MHD in plasma.It is helpful for the indepth study of physical phenomena such as locked mode,penetration mode,and sawtooth in the future.Based on the correlation between plasma internal information and diagnostic signals,this paper uses machine learning to explore the plasma electron temperature prediction and MHD classification research in the J-TEXT,and obtain better predictions and classifications effect.This method will help in-depth physical analysis and provide effective assistance for the fusion research of the J-TEXT.