Real-time and accurate plasma boundary reconstruction is critical for plasma control in tokamaks. Reconstructing plasma boundary through visible light diagnostics provides a promising approach to plasma shape control during steady-state discharges. In this study, a two-stage plasma boundary detection framework (YOLO-GRAY) has been developed on the EAST. The framework first utilizes You Only Look Once (YOLO) model to rapidly localize the optical boundary emission region and then calculates the precise boundary position using grayscale edge detection within that region. This two-stage approach enables more precise and robust detection of the plasma boundary position. However, due to modifications in the visible light diagnostics, the accuracy of the YOLO model pre-trained on J-Port dataset decreased to 0.493 on images with a new field of view (FOV). By employing transfer learning and fine-tuning the pre-trained YOLO model on merely 50 new FOV images, the model’s detection accuracy was improved to 0.963. Furthermore, the algorithm maintains robust performance during long-pulse operations. In 450 s long-pulse experiment on EAST, the Last Closed Flux Surface displayed accumulated deviations of 0.5 cm and 1.5 cm in GAP and GAP, respectively, whereas the optical boundary positions remained essentially stable. Building upon this framework, we developed a real-time optical boundary reconstruction system based on heterogeneous computing. Hybrid CPU–GPU scheduling ensures that reconstruction for a single frame is completed within 1.6 ms. With this system, multi-point plasma shape control using optical boundaries was demonstrated for the first time on EAST, validating the feasibility of visible-light-based shape control. Overall, the YOLO-GRAY algorithm demonstrates stable, precise reconstruction of optical plasma boundaries, and visible-light-based shape control holds great promise for future fusion devices such as ITER.
Real-time optical plasma boundary reconstruction for plasma position control at the TCV Tokamak