Work

Cross‐device and cross‐modal arc detection: transfer learning between EAST visible and WEST infrared diagnostics

Year: 2026

Type: article

Abstract: This work presents an engineering-oriented data processing and model training pipeline for Experimental Advanced Superconducting Tokamak (EAST) visible and (W) Environment in Steady-state Tokamak (WEST) infrared data, enabling a systematic study of cross-device and cross-modal electric arc detection. A lightweight convolutional neural network is adopted as a cross-device arc detection model, and joint training and transfer learning experiments are conducted on EAST visible (VIS) and WEST infrared (IR) data. Single‐modality baselines are first trained on EAST and WEST, then evaluated zero‐shot on the opposite device, followed by a comparison of frozen‐backbone and fine‐tuned transfer for VIS → IR and IR → VIS. The results show that models trained from scratch on each diagnostic channel already achieve high precision and recall for arc detection, whereas direct zero-shot cross-modal transfer drives the arc recall on the opposite domain almost to zero. In contrast, starting from the single-modality baselines and modestly fine tuning higher convolutional layers and the classification head on the target domain can significantly restore-cross device, cross-modal detection performance, with positive class precision and recall approaching or even surpassing the single-domain baselines. Class activation map analysis further indicates that, under fine-tuned and joint training configurations, the models consistently focus on the antenna, the arc, and its surrounding scattering region, supporting physically meaningful explanations and suggesting that cross-device, cross-modal training can capture arc related structures that generalize across devices while providing a lightweight, scalable route to reusing imaging-based diagnostic knowledge with reduced annotation and retuning effort. Given that machine protection is safety-critical and that imaging diagnostics are never identical across devices, the transferability demonstrated in this work is directly relevant to equipping next-generation tokamaks such as ITER and BEST with a reliable arc-detection capability before a large native training set becomes available. On a newly commissioned device, arc events are rare, sporadic and safety-relevant, so deferring detection until a sufficiently large native training corpus has accumulated would leave the machine unprotected during the early operational phase, when limited operating experience makes the associated risks most difficult to characterize; cross-machine and cross-modal transfer instead enables a new device to inherit a usable detection capability from the outset. Accordingly, the present work is intended not merely as a benchmark of cross-domain accuracy, but as evidence that scalable, transferable arc monitoring can constitute a machine-protection layer of future fusion reactors.

Source: Nuclear Fusion

DOI: 10.1088/1741-4326/aeb036