Machine learning based sensor fault detection schemes for plasma position control in tokamak

Abstract

The Tokamak is a device that facilitates nuclear fusion via magnetic confinement of Deuterium and tritium plasma. Circular arrays of magnetic flux sensors are employed outside the vacuum vessel to measure the position of the hot plasma. There are many sources of faults that can affect the readings of these magnetic sensors, namely, stuck-at-zero fault, offset faults, and noise faults. Similarly, the sensors can be influenced by other active or passive currents in the neighboring superconducting coils. Thus, it is essential to detect any faults and recover the faulty sensor to be used for plasma position control inside a Tokamak. In this work, the use of Machine Learning (ML) techniques for sensor fault detection is explored to utilize the knowledge of the experts to automate the sensor fault detection tasks. Since sensor responses are highly non-linear and the exact mathematical model of a sensor is unavailable, data-driven methods are preferred. Therefore, this work targets to explore the effectiveness of ML algorithms to classify the sensor faults. A sensor fault database was made from the simulated and experimental sensor measurement data from the ‘Aditya’ Tokamak situated at the Institute of Plasma Research, Gandhinagar, Gujarat. The three most occurring fault types in a Tokamak, namely Stuck-at-Zero fault, Offset fault, and Noise Faults, were simulated and added to the expected/ideal measurements to create the fault signature database. The ML models were trained using this database. The work reported in this thesis describes the development of algorithms for the detection of various types of sensor faults that occur in plasma position sensors in a Tokamak. Firstly various ML algorithms, namely, Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Gaussian Naive Bayes (GNB), Bernoulli’s Naïve Bayes (BNB), Ridge Classifier, Linear Support Vector Classifier (LSVC), Radial Basis Function Support Vector Classifier (RBFSVC), Nearest Centroid (NC), Radius Neighbor (RN), K-Nearest Neighbor (KNN), Multi-Layer Perceptron (MLP) Neural Network, and Decision Tree (DT) are evaluated for the fault classification. The ML algorithms were compared in terms of classification accuracy and execution time. The Decision Tree method was found to generate the highest scores in classifying the sensor faults, with minimum classification time. Furthermore, a weighted majority voting-based Ensemble Classifier (ECF) was designed constituting Four different ML classifiers, namely, Logistic Regression based Multi-layer Perceptron Classifier (MLP), Support Vector Classifier (SVC), K-Nearest Neighbor (KNN), and Decision Tree (DT). Improvement in overall accuracy was observed with the ECF for correctly identifying and classifying the faults. ML algorithms necessitate significant computational resources, i.e., Central Processing Unit (CPU) and Random Access Memory (RAM), during their training period. Therefore, Graphics Processing Units (GPU) acceleration of the ML models was investigated, as they contain a large number of floating-point processors. A Dense Neural Network (DNN) and a stacked ensemble classifier were evaluated for sensor fault classification. Since the Decision tree algorithm was found to provide the best classification accuracy and classification time, it was chosen to be implemented in FPGA for real-time sensor fault classification. A CPU+FPGA implementation method was used to realize the fault classifier on a Digilent Xilinx PYNQ FPGA platform. Finally, a real-time Hardware-in-Loop (HIL) simulation was designed to evaluate the ML models for sensor fault detection and classification tasks. It consisted of an Android application for user input, a Single Board Computer (SBC) to generate the simulated sensor measurement data, an FPGA board with the ML classifier implementation. A Graphical User Interface (GUI) was developed in ‘Python + Tkinter’ to display the sensor status, with an audible alarm and data-logging system. The application of machine learning algorithms for sensor fault detection for plasma position control in Tokamak shows promising results in automated identification and classification of faults. Automatic fault detection will help reduce the downtime of the Tokamak to search and replace the faulty sensor and improve the overall efficiency of the nuclear fusion process for the generation of electricity.