Automatic Recognition And Classification Of Plasma Patterns In Electron Cyclotron Emission Imaging Based On Machine Learning Methods

The Electron Cyclotron Imaging(ECEI)diagnostic system can provide two-dimensional electron temperature fluctuation imaging on the cross section of Tokamak.By analyzing the ECEI data,it can be found that Tokamak has various plasma events(Such as sawtooth mode,tearing mode,balloon mode,etc.),the discovery of these patterns is of great significance for understanding the physical mechanism of high temperature plasma and improving plasma confinement.Due to the low efficiency of manual analysis of ECEI data,a large amount of ECEI data has not been fully exploited.In this paper,the ECEI data is analyzed by machine learning methods.The plasma patterns are automatically identified and classified by unsupervised learning and supervised learning.For plasma patterns recognition,this paper uses automatic spectral clustering method to analyze the ECEI data,so as to automatically determine whether there is a plasma pattern and the time and space location of the pattern in each shot.There are many kinds of plasma patterns in Tokamak.Different patterns have different characteristics.Therefore,it is difficult to establish a universal recognition model using the traditional feature-based plasma pattern criterion.In this paper,a new criterion of plasma pattern based on similarity is proposed.According to the criterion,a plasma pattern recognition model based on automatic spectral clustering algorithm is established.The model can automatically determine the shot number,time and space location of the plasma patterns.Since criterion of the plasma pattern is independent of the characteristics of each plasma pattern,the model can not only identify the plasma pattern that has been discovered,but also explore new plasma patterns.For plasma pattern classification,this paper uses convolutional neural network to construct a plasma pattern classifier.The pattern classifier can be used to further analyze the output of the pattern recognition model to determine the specific type of the pattern.In general,the main difficulty in pattern classification is the lack of labeled data and feature extraction.However,the result of the plasma pattern recognition model can provide a large amount of labeled data for the training convolutional neural network.At the same time,the convolutional neural network can automatically extract the characteristics of the time series signal,thus avoiding the difficulty of feature extraction.Finally,the paper analyzes the ECEI data of 5 shots to verify the correctness of the model.In this paper,the pattern recognition model discovers sawtooth patterns and a possible new plasma pattern in the training dataset.Then we use the output of the pattern recognition model as the input of convolutional neural networks and obtained a sawtooth classifier.It is the first time that the machine learning methods are applied to the pattern recognition and classification of ECEI data of EAST Tokamak.This method can quickly analyze ECEI data,which is of great significance for people to understand plasma physics and optimize tokamak control.