Real-time capable MHD mode tracking using deep learning algorithm and high-speed video cameras on HBT-EP

We present recent developments on deep-learning-based MHD mode tracking using the newly upgraded high-speed videography system on HBT-EP. This algorithm uses the convolutional neural network architecture to process video frames recorded during plasma discharges by one or multiple cameras and predict the n=1 mode amplitude and phase over time. Training and testing datasets were assembled from multiple run campaigns conducted on HBT-EP, during which two Phantom S710 cameras operating at 250 kfps, 128×64 pixels resolution were set to measure Dα emissions on the plasma’s poloidal cross-sections at different toroidal positions. The target mode information was obtained through least-squares fitting using the in-vessel magnetic sensors. The resulting model is able to track the n=1 external kink mode consistently over the testing shots, with the amplitude and phase errors bounded by ±2 G (over 0-20 G amplitude range) and ±20 degrees during the mode-active period of a discharge. This indicates a significant improvement over the previous SVD-based method [1]. Variations of the model’s architecture as well as the inference latency and robustness of different models are also investigated.

[1] Angelini, et al., Plasma Phys Contr Fusion, 57, 045008 (2015).