Speaker
Description
Real-time and accurate plasma boundary reconstruction is critical for tokamak plasma control. Visible light diagnostics offer a promising solution for plasma shape control during steady-state discharges. In this study, a two-stage plasma boundary detection framework (YOLO-GRAY) was developed on EAST tokamak. The framework utilizes YOLO to rapidly localize the optical emission region, followed by grayscale feature detection to precisely determine the optical boundary position. This approach enables more precise and robust boundary extraction. However, due to modifications in visible light diagnostics, the detection accuracy of the previously developed YOLOv8n-seg-CBAM model dropped to 0.493 for images with a new field of view (FOV). By employing transfer learning with only 50 new FOV images, accuracy improved to 0.963. Furthermore, the algorithm maintains robust performance during long-pulse operations. In a 450 s experiment, the Last Closed Flux Surface(LCFS) exhibited accumulated deviations of 0.5 cm in $Gap_{in}$ and 1.5 cm in $Gap_{out}$, whereas optical boundary positions remained stable. Building upon this, a real-time reconstruction system based on heterogeneous computing was developed. Hybrid CPU-GPU scheduling ensures single-frame reconstruction within 1.6 ms. Utilizing this system, multi-point plasma shape control based on optical boundaries was successfully demonstrated for the first time on EAST, validating the feasibility of optical-based plasma shape control. Overall, this system holds significant promise for plasma control in future fusion devices.
| Minioral | Yes |
|---|---|
| IEEE Member | No |
| Are you a student? | Yes |