Work

Physically Interpretable Series Arc Fault Detection via Hilbert Transform-Derived Half-Cycle Energy Ratios and LightGBM

Year: 2026
Type: article
Abstract: Series arc faults in low-voltage AC distribution systems pose a significant fire hazard but are difficult to detect because the fault current may remain below the operating thresholds of conventional circuit breakers and residual current devices. This paper proposes a physics-guided and lightweight detection framework that combines Hilbert-domain transient analysis, Half-Cycle Instantaneous Energy Ratio (HIER) modeling, and an edge-constrained LightGBM classifier. A test platform compliant with GB/T 31143 was constructed to acquire high-sampling-rate current signals from eight representative resistive, inductive, and nonlinear household loads under normal and series-arcing conditions. By treating each AC half-cycle as the fundamental modeling unit, three complementary HIER descriptors were developed to quantify localized energy bursts, out-of-band spectral dispersion, and frequency–amplitude instability. These descriptors were combined with analytic-signal envelope, instantaneous-frequency, and Teager–Kaiser energy features and aggregated over a 100 ms sliding window to form a compact 13-dimensional feature vector. On the reference test set, the proposed model achieved an accuracy of 98.52%, an arc-fault recall of 99.26%, an F1-score of 98.47%, and a ROC-AUC of 99.85%. Ablation experiments confirmed that the HIER descriptors provide complementary discriminative information, while SHAP analysis showed that the classifier decisions are consistent with the transient-energy and non-stationary characteristics of arc discharge. The complete model occupied approximately 0.53 MB, and tree-prefix pruning reduced its size to 194.2 KB while maintaining an F1-score above 0.98. These results demonstrate a favorable balance among detection performance, physical interpretability, and deployment efficiency for resource-constrained arc-fault detection devices.
Source: Energies
Language: English
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7236315927
FWCI:
Cites: 20
Cited by:
Related to: 10