Akademik Veri Yönetim Sistemi
IEEE Transactions on Industry Applications, 2026 (SCI-Expanded, Scopus)
Electric Arc Furnaces (EAFs) introduce stochastic current transients that degrade electrode life and power quality. Detecting such events is challenging when high-frequency components are attenuated by the anti-aliasing filters of standard monitoring devices. This paper proposes a fast, deterministic, and training-free transient detector operating on EAF currents sampled at low rates compared to transients. The method forms half-cycle RMS features, extracts a residual via a light Savitzky–Golay trend estimate, and applies numerical differentiation followed by a sliding-window Z-score test with hysteresis logic to robustly label transients. Performance is evaluated on both synthetic signals and field measurements, and is benchmarked against a robust LSTM autoencoder (LSTM-AE) reconstruction-based detector enhanced with robust normalization, training cleaning, and adaptive thresholding. In the evaluated cold-start setting, the proposed approach significantly outperforms the deep learning baseline. While the LSTM-AE suffers from severe event fragmentation and degrades to an F1-score of 0.13 on actual EAF measurements, the proposed method maintains a robust F1-score of 0.68. Furthermore, it operates with substantially lower computational overhead and no training requirement, making it highly suitable for low-latency industrial monitoring in EAF installations.