Wavelet‐guided dual‐graph transformer for low‐frequency oscillation detection and root cause localization in industrial processes

Abstract Industrial process data often contain intermittent, low frequency oscillations that are weak, slowly drifting in frequency, and propagate across multiple coupled control loops, conditions under which fixed thresholds, pairwise test and uncalibrated scores perform poorly. To address these challenges, a wavelet‐guided dual‐graph spatiotemporal transformer framework is proposed for plant‐wide oscillation detection and root‐cause localization with calibrated confidence. First, a complex Morlet continuous wavelet transform is applied to identify oscillatory events and to define two analysis windows per event: an early window capturing onset dynamics and a peak window centred on maximal oscillatory energy. Second, for each window, Granger‐causality and band‐limited coherence graphs are constructed, yielding four directed graphs that encode causal and coherence structure over the specific oscillation interval. Third, a dual‐graph message‐passing model fuses these graphs at each time step, and a temporal transformer aggregates contextual information to produce per‐sensor logits. Finally, a sigmoid output head provides root‐cause probabilities for each channel, and post‐hoc calibration converts raw scores into reliable probabilities suitable for operational thresholding. On real hot strip mill data, the proposed framework outperforms strongest granger baselines for root cause Localisation, achieving top‐1 accuracy of 0.90 and a top‐3 accuracy of 0.95, while the continuous wavelet transform (CWT)‐ridge detector improves oscillations event detection. Reliability analysis and a row‐normalized calibration matrix yields ECE = 0.142 and Brier = 0.049, isotonic calibration slightly pulls mid‐range probabilities towards the centre and makes high confidence predictions a bit stronger without changing the order of predictions. Early/peak Granger‐causality and coherence heatmaps provide transparent, auditable evidence, yielding a practical end‐to‐end deployable solution for plant‐wide low‐frequency oscillation detection and root‐cause localization.