Internal reconnection events (IREs) are rapid magnetohydrodynamic phenomena that play an important role in the confinement and stability of spherical tokamak plasmas. Reliable identification of IREs in experimental data is challenging due to short discharge durations, ambiguous event boundaries, and the limited availability of labeled data. In this study, we propose an unsupervised, event-level IRE detection framework based on anomaly detection techniques and apply it to experimental data from the VEST spherical tokamak. The proposed framework combines a two-stage detection strategy using plasma current and Hα emission signals with sliding-window segmentation and event-level evaluation, enabling physically meaningful IRE identification without labeled training data. Three unsupervised models—K-Nearest Neighbors (KNN), One-Class Support Vector Machine (OCSVM), and an autoencoder (AE)—are evaluated within a unified framework. All models achieve stable detection performance, with precision exceeding 80% and recall above 70% under a precision-oriented operating point. To enhance detection robustness, a KNN-based cleaning procedure is introduced during training to remove noise-driven, locally isolated windows, significantly reducing spurious detections while preserving physically meaningful IRE signatures. Event-level analysis indicates that missed detections under this operating regime predominantly correspond to weak events with limited impact on global plasma behavior. The proposed framework is fully unsupervised, computationally efficient, and readily extensible to other spherical tokamak devices, providing a flexible foundation for incorporating additional diagnostics, such as Mirnov coil signals, toward precursor-aware detection and future predictive modeling of IRE activity.
Ok et al. (Sun,) studied this question.