Real-time anomaly detection for disruption prediction: the JET case

This article shows the development of a new kind of real-time disruption predictor that is based on detecting anomalies in the data flow. The new predictor neither depends on data from past discharges nor is based on signal amplitude thresholds. In JET, using only the locked mode signal, the new predictor shows results comparable to the JET APODIS predictor but without the need of a training process with past data. The predictor has been tested with JET discharges in the range 82460- 87918. This range corresponds to all ITER-like Wall experimental campaigns (2011 - 2014). The discharge dataset consists of 1738 non-disruptive discharges and all unintentional disruptions (566 disruptive shots). The results show 8.98% of false alarms, 10.60% of missed alarms, 3.18% of tardy detections, 83.57% of valid alarms, 2.65% of premature alarms and average anticipation time of 389 ms. These rates are compared in the article with the results of the JET APODIS predictor and the JET disruption predictor based on crossing a threshold of the locked mode signal amplitude.