Data-driven technique for disruption prediction in GOLEM tokamak using stacked ensembles with active learning

Abstract

In a tokamak, disruption is defined as losing control over a confined plasma resulting in sudden extinction of the plasma current. Machine learning offers potent solutions to classify plasma discharges into disruptive and non-disruptive classes. Evolving experimental programs reduce the performance of machine learning models, and also, the need for labeling the huge volume of data incurs more labor cost and time. This paper proposes a data-driven based machine learning technique that employs an active learning approach for labeling and classification of plasma discharges. The designed model uses 117 normally terminated shots and 70 disruptive shots with 14 labeled diagnostic signals. The stacking classifier is built over three base learners: logistic regression, reduced error pruning tree, and categorial boost algorithm, and the logistic regression technique is used at the meta-learner. An active learning approach is proposed for labeling the unlabeled dataset using a modified uncertainty sampling technique with minimal queries. The proposed model queries the unlabeled data to an oracle based on a selection strategy with uncertainty sampling using entropy metrics. The new labeled data and the class probabilities of the base classifiers are channeled to the final predictor for classifying the plasma discharge. The proposed model achieves an accuracy of 98.75% in classifying the disruptive vs non-disruptive discharges, with a minimally trained dataset, and also, it is free from aging of predictors.