Vi-DP: Low-Latency Video-Based Disruption Prediction with Multi-FOV Fusion in EAST

Plasma disruption poses a significant safety challenge for future tokamaks and fusion reactors. Prior work on disruption prediction relies primarily on multiple scalar zero-dimensional (0D) signals and one-dimensional (1D) profiles as input of the predictor. Instead, this study investigates the feasibility of exclusively utilizing visible light video signals, consisting of sequential two-dimensional (2D) images, to achieve low-latency disruption prediction in EAST for the first time. To address the high data throughput of video input, striking a balance between real-time performance and accuracy, a lightweight deep neural network with combined spatial and temporal feature extraction structure is proposed for feature learning and sequence modeling. The datasets come from the wide-angle viewing systems (WAVS) distributed across EAST ports C, F, and K. The model is evaluated on both single-port and multi-port fusion datasets at different sampling rates, with all configurations demonstrating robust prediction performance. In particular, with multiple field-of-view (FOV) fusion dataset, the model achieves a true positive rate (TPR) of 92.8% and a false positive rate (FPR) of 5.8%, demonstrating performance on par with state-of-the-art non-video-based disruption prediction models. Notably, the model takes merely 6.9 ms for inference, whereas its average warning time is 458 ms, which opens up the possibility for future online disruption warning and active control based on real-time video stream input. These initial results confirm the inherent predictive value of visible video for disruption warning and lay an important foundation for future endeavors focused on online disruption prediction that incorporates visual information. Furthermore, for future large-scale cross-device disruption warning databases, it could be advantageous to incorporate 2D visual signals in conjunction with conventional 0D and 1D signals to provide more comprehensive and robust features for disruption prediction.