Deep learning-based process for the automatic detection, tracking, and classification of thermal events on the in-vessel components of fusion reactors

This paper presents an automated process that detects, tracks, and classifies thermal events using infrared movies of the inside of the vessel for machine protection. This process relies on a Region-Based Convolutional Neural Network, a deep learning model, for the detection of thermal events in infrared images. This model is trained using a dataset of thermal events, obtained by manually annotating thermal events in movies from the WEST tokamak, which is equipped with 12 infrared cameras that provide information about the surface temperature of the in-vessel components. The labels characterizing the thermal events are chosen in a custom-designed ontology, which is being developed. This automated process can correctly detect, track and classify most of the regular thermal events appearing in the infrared movies of WEST, in a manner compatible with the real-time plasma operation. These events can be used for post-pulse analysis by deriving from them metadata, displayed in a dedicated dashboard, that help human operators understand better and more quickly what happened during a pulse.