Weighted majority rule ensemble classifier for sensor fault classification for plasma position control in Tokamaks

Abstract

Tokamaks are the most promising devices which employ magnetic confinement of Deuterium and Tritium plasma to achieve nuclear fusion for power generation. Magnetic flux sensors are employed to measure the position of the plasma column inside the device. Faults occurring in these sensors may cause the failure of control of the plasma position, thereby terminating the fusion process. In this paper, we present a comparison of different Machine Learning (ML) algorithms to detect and classify the sensors’ faults. We applied four different ML algorithms; namely, Logistic Regression based Multi-layer Perceptron Classifier (MLP), Support Vector Classifier (SVC), K-Nearest Neighbor (KNN), and Decision Tree (DT) as the individual ML classifiers for the task of sensor fault classification. We propose an ensemble classifier (ECF) using the classifiers, as mentioned earlier. We provide a comparative assessment of the classification accuracy of different classifiers, namely, MLP, SVC, KNN, DT, and ECF. The ECF utilized a weighted majority voting method to combine the decision of individual faults classifiers. Overall, robustness to the miss-classification of new data by the individual sensor fault classifiers by applying the ECF is reported.