CNN disruption predictor at JET: Early versus late data fusion approach

This work focuses on the development of a data driven model, based on Convolutional Neural Networks (CNNs), for the real-time detection of disruptive events at JET. The predictor exploits the ability of CNNs in learning relevant spatiotemporal information straight from 1D plasma profiles, avoiding hand-engineered feature extraction procedures. In this paper, the radiation profiles from both the bolometer horizontal and vertical cameras have been considered amongst the predictor inputs, with the aim of discriminating between the core radiation due to impurity accumulations and the outboard radiation phenomena. Moreover, an innovative predictor architecture is proposed, where two separate CNNs are trained to focus on events with different timescales, that is, the destabilization of radiation, electron density and temperature profiles, and the mode-locking and current profile variations. The outputs of the two CNNs are combined with a logic OR function to provide the disruption alarm trigger. The advantages of this data fusion approach impact on the predictor performance, with a very limited number of false alarms (only 1 in the considered test set), and on the model output interpretability as the two different branches are triggered by different types of events.