Deep learning and image processing for the automated analysis of thermal events on the first wall and divertor of fusion reactors

A multi-stage process that detects, tracks and classifies thermal events automatically using thermal imaging of the inside of fusion reactors is presented. The process relies on the Cascade R-CNN algorithm for the detection and classification and on the SORT algorithm for the tracking. The process is trained using a dataset of 325 thermal events distributed in seven classes, manually annotated from 20 infrared movies of the inside of the WEST tokamak. This dataset is created using user-friendly annotation tools, based on simple thresholding. The performance of the process is evaluated using modified indicators that emphasize the importance of the detection of the hottest zones of the hot spots. The modified mean average precision on a test dataset establishes at 27%.