Data processing methods and preliminary testing of machine learning algorithms for Pegasus-III

Pegasus-III is a spherical tokamak investigating non-solenoidal plasma initiation techniques. Machine learning (ML) methods are being developed for audio fault detection, predictive control systems, and cathode spot detection on Pegasus-III. Reinforcement learning based radial position control algorithms will be developed with collaborators, attempting both model based and non-model based methods. Local helicity injectors on Pegasus-III may be damaged by external arcing, known as "cathode spots." To minimize damage, ML methods can be used to shut off injectors at the onset of cathode spots. A dataset consisting of 50 discharges has been curated for preliminary testing of cathode spot identification using ML algorithms. Data sets include injected voltage and current traces, and a boolean value indicating the onset of a cathode spot. The data are used to train an XGboost model that yielded an average accuracy of 65%, with accuracy expected to improve as more data points are added to the set. The data is also being used with binary classification models. Additionally, a convolutional neural network algorithm is being developed to use fast camera imaging of the injectors to identify cathode spots.