Disruption predictor based on neural network and anomaly detection on J-TEXT

Abstract

Disruption prediction is essential for the safe operation of a large scale tokamak. Existing disruption predictors based on machine learning techniques have good prediction performance, but all these methods need large training datasets including many disruptions to develop their successful prediction capability. Future machines are unlikely to provide enough disruption samples since these cause excessive machine damage and the prediction models used are difficult to extrapolate to a machines that the predictor was not trained on. In this paper, a disruption predictor based on a deep learning and anomaly detection technique has been developed. It regards the disruption as an anomaly, and can learn on non-disruptive shots only. The model is trained to extract the hidden features of various nondisruptive shots with a convolutional neural network and a long-shot term memory (LSTM) recurrent neural network. It will predict the future trend of selected diagnostics, then using the predicted future trend and the measured signal to calculate an outlier factor to determine if a disruption is coming. It was tested with J-TEXT discharges in flat top phase and can demonstrate comparable performance to current machine learning disruption prediction techniques, without requiring a disruption data set. This could be applied to future tokamaks and reduce the dependency on disruptive experiments.

Review: Machine learning for advancing low-temperature plasma modeling and simulation

“…LSTM networks have been applied in fluid dynamics for the prediction of the temporal dynamics of turbulent flow through channels based on DNS data 156 . They have further been used for the prediction of disruption in magnetic confinement fusion plasmas 157 , 158 …”

“…156 They have further been used for the prediction of disruption in magnetic confinement fusion plasmas. 157,158 Moreover, in a study on a hierarchical approach to multiscale data-driven modeling, LSTM networks have been included for comparison. 159 Another fundamentally related variant of RNNs that was included in this study are echo-state networks (ESN).…”

Disruption prediction with artificial intelligence techniques in tokamak plasmas

“…Approaches relying on a mixture of time/frequency domains, including wavelet decompositions, have also been pursued [17][18][19] . With regard to classifier technologies, real-time compatible predictors have typically been based on artificial neural networks, support vector machines, fuzzy logic, generative topographic mapping and deep learning and have been studied on a broad range of tokamaks, including ADITYA (India) 20 , ASDEX Upgrade (Germany) 21 , DIII-D (United States) [22][23][24] , J-TEXT (China) 25 , NSTX (United States) 26 , ALCATOR C-MOD (United States) 27 , JT-60U (Japan) 28 , EAST (China) [29][30][31] , HL-2A (China) 32 and JET (United Kingdom) [33][34][35] .…”

“…The classifiers employed in the studies on tokamaks mentioned above [20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35] were developed using real-time valid solutions, which guarantee response times within a specified time window. The predictors discussed in the remainder of this work have been tested offline with real-time compatible technologies and using only real-time available signals.…”

Using HPC infrastructures for deep learning applications in fusion research

“…Our goal is to find this by automated means, using deep learning. The problem falls within the scope of anomaly detection [31,32], where a model is trained to tell whether a given sample is anomalous or not, based on how distinct that sample is from the samples that the model has seen during training. In our context, this implies training a model on non-disruptive pulses (the "normal" behavior), and then applying it to disruptive pulses in order to identify the anomalous behavior.…”