Identification of MHD modes on EAST using a deep learning framework

Abstract

The improvement of plasma parameters is severely limited by magnetohydrodynamic (MHD) instabilities. The identification of MHD modes is crucial for the study and control of MHD instabilities. In this study, an MHD mode identifier is developed based on a temporal convolutional network and long short-term memory (LSTM) network. The identifier is trained and tested on a small dataset containing 33 shots. Firstly, the temporal convolutional network encodes 27 diagnostic signals and then decodes them using LSTM network with different parameters to obtain the MHD modes and their frequency and intensity. The identifier exhibits an accuracy of approximately 98.38% on the test set and can accurately calculate the frequency and intensity of the MHD modes. To further examine the performance of the identifier, seven shots outside the dataset are used for shot-by-shot testing. The identifier can accurately identify the time period of tearing modes, and the identification accuracies of 2/1 and 3/2 tearing modes are 92.7% and 100%, respectively. The identification accuracy of the fishbone mode is slightly worse, only 82.1%. This is because the fishbone mode occurs intermittently. The frequent switching between the fishbone mode and no MHD behavior affects the identification of the fishbone mode. Overall, through the training of the small datasets, the identifier exhibits a good identification performance for the MHD modes. The proposed data-driven identifier can serve as a reference for establishing a large MHD mode database of EAST as well as a real-time MHD identification and control algorithm.