Data Preparation to Facilitate Machine Learning Using Pegasus III Data

The local helicity injectors Pegasus III uses for LHI startup can experience cathode spots, which damage the electrode surfaces and degrade injector lifetime. Prediction of cathode spot onset using machine learning (ML) coupled with fast injector shut-down will extend the lifetime of the injectors and reduce plasma impurity content. ML is most successful when large, well-documented and curated datasets are easily accessible for training. To facilitate the incorporation of Pegasus III data into ML, python-based tools have been developed to load and process raw data that is then placed into an MDS+ data structure with appropriate metadata information. The initial data set includes magnetic diagnostics, fast optical time series diagnostics, including an AXUV diode array & optical cathode spot detector, as well as coil currents and power supply diagnostics. The initial ML task will be supervised learning to develop a model to predict cathode spot onset. A cathode-spot start time signal will be developed with defined criteria for classification. Additionally, the curated data set will be uploaded to an OPEN/FAIR database to allow collaboration with outside institutions that would otherwise be unable to access Pegasus III data.