Real-time sensor fault detection in Tokamak using different machine learning algorithms

The "Tokamak" is a device that facilitates nuclear fusion between Deuterium and Tritium. Multiple arrays of magnetic sensors are used to detect the plasma position inside a Tokamak. This paper presents the application of different machine learning (ML) based fault detection techniques for the identification and classification of faults that happen in typical magnetic position sensors. The performances of these machine learning based fault detection algorithms are evaluated for two scenarios as follows. Firstly, during the "Self-Test" mode, i.e., before the start-up of the plasma discharge, with known current waveforms in the external coils. Secondly, by using the simulated plasma discharge waveforms. Their performances are compared in terms of computational complexities and latency in view of deciding the best fault detection algorithm. The machine learning techniques are implemented in real-time on the Xilinx Kintex-7 and Xilinx Zync-7 series FPGA. From the obtained comparison results, it is observed that out of the six machine learning approaches, namely Mull-Layer Perceptron (MLP) neural network, K-Nearest Neighbor (KNN), Support Vector Classifier (SVC), Gaussian Naive Bayes (GNB), Decision Tree (DT) and Random Forest classifier (RF) employed for Tokamak sensor fault detection, the Random Forest Classifier based approach was found to be the best in terms of speed and accuracy.