Online un-supervised tearing mode detection with sequentially discounting algorithms for JT-60SA

A non-empirical tearing mode detector model based on statistical anomaly detection with sequentially discounting algorithms has been developed, and its prediction performance has been demonstrated with experiment data in the JT-60SA initial operation phase. Detection and prediction of the occurrence of disruptions is an important issue to protect the tokamak device from the damage caused by disruptions, and empirical models using machine learning techniques are being developed. A non-empirical model based on anomaly detection is one possible solution to an inherent problem of uncertainty as to whether such empirical models can be extrapolated to future tokamaks. The developed detector model has shown a good prediction performance, and its processing time was smaller than the control cycle time of the JT-60SA without any advanced devices such as GPUs. The result suggests that the statistical anomaly detection with sequentially discounting algorithms is a good choice to be implemented in the plasma controller not only in the JT-60SA but also in the future devices such as ITER and DEMO.