Precursor warning of NTM/TM trigger regime in EAST based on supervised learning with rolling-aware hard-negative training

Neoclassical tearing modes (NTM) and tearing modes (TM) pose significant risks to long-pulse operation in the Experimental Advanced Superconducting Tokamak. Machine learning approaches promise fast precursor recognition, but fixed-window classifiers can produce deceptively high validation scores that do not translate into practical online alarm performance. In this work, we develop a reproducible, multi-diagnostic framework with rolling-aware training using hard negative examples from entire non-disruptive discharges, teaching the model what safe operation actually looks like. Using 55 discharges (35 with NTM/TM onset) and 826 hand-crafted features, combining generic channel statistics with physics-informed cross-diagnostic coupling measures, a LightGBM classifier achieved grouped 5-fold cross-validation ROC-AUC of 0.9817 (near-perfect), but fails in rolling simulation with 1.0 false alarms per negative shot. After rolling-aware training, the false-alarm rate drops to 0.10 per negative shot at 0.857 sensitivity (catching 85.7% of true events), with a median warning lead time of 1445 ms—about 1.5 s, valuable for mitigation. Generalisation is confirmed by out-of-fold AUC of 0.858 ± 0.0635 and temporal-block holdout AUC of 0.7986. SHAP interpretation based on the 1 kHz and 10 kHz sampling rates using a rolling window shows that low-frequency features of Mirnov probe signals dominate the top importance ranks, while ECE band power and channel-wise statistical features provide complementary information. These results establish a realistic, interpretable baseline for EAST NTM/TM warning and demonstrate that rolling-aware hard-negative training is essential to bridge the gap between offline accuracy and online reliability in fusion field.