Plasma Disruption Prediction Based On Deep Learning And Anomaly Detection

Disruption is an unstable phenomenon in the discharge of the Tokamak devices,which will do great harm to the tokamaks.To solve this problem,researchers have developed some plasma disruption mitigation measures,which can greatly reduce or avoid the damage caused by disruption.However,these methods require a certain amount of action time,so it is necessary to predict the disruption before it occurs.At present,many disruption predictors have been designed.From the early method of setting a threshold for a single diagnostic signal to the neural network disruption predictor based on machine learning,the accuracy of prediction has been greatly improved and the false alarm rate is also greatly reduced.However,the neural network method requires a lot of data for training which is very difficult to obtain in the future tokamak devices,especially the disruptive discharges.In this paper,the distuption prediction methods with only a small amount of disruptive data is studied.Firstly,a disruption database was developed to deal with the data processing problem in Tokamaks.The disruption database includes diagnostic data storage module,data automatic processing module,and disruption database service module.They automate the acquisition,processing and analysis of the data required in disruption predictors more efficiently.Then,a time series prediction model of diagnostic signals based on convolutional neural network and recurrent neural network is designed.Through the prediction residual of the model,the anomaly will be detected to further detect disruption.This model only needs data from non-disruptive shots for training,and the model trained on the J-TEXT can achieve a success rate of 83% with a false alarm rate of 18%.Finally,the data from J-TEXT and HL-2A are used to study the transference of models between different Tokamaks.In this part,the data of two devices are used to train and evaluate the same model respectively.The model trained on J-TEXT is frozen in part of the network layer,then using a small amount of HL-2A data for transfer learning training.Both methods have achieved positive results in offline experiment.The results in this paper prove the feasibility of developing a disruption prediction system with a small amount of disruption data based on anomaly detection method or transfer learning method.It provides a solution for the disruption prediction problem of future large Tokamaks.