Series arc fault detection via arcing physical characteristics-guided feature extraction: a dual-signal multi-timescale method

• Reveals why arc faults from poor contacts differ from insulation aging. • Proposes a dual-signal multi-timescale feature extraction framework. • Designs a two-stage decision logic for series arc fault detection. • Achieves 99.2 % detection accuracy for series arc faults in complex scenarios. • Enables deployment on resource-limited edge devices with low computational cost. Series arc faults (SAFs) are a primary cause of electrical fires in building distribution systems, requiring reliable detection to mitigate fire risks and ensure operational safety. Existing methods, though effective under simplified scenarios, suffer from false alarms and missed detections in complex scenarios involving diverse load types, circuit topologies, arc-generating modes, and arc-like conditions. To address the issues, this paper proposes a SAF detection method that integrates a dual-signal multi-timescale feature extraction framework, an arc fault decision tree, and arc fault decision criteria. Firstly, an experimental platform simulating SAFs in building distribution systems is established, and the arcing characteristics are analyzed through key electrical parameters—arc voltage, current, and zero-sequence current coupling signal—across four stages: arc-igniting, arcing, arc-extinguishing, and zero-current under different arc-generating modes. Secondly, based on these characteristics, with a particular focus on the current and zero-sequence current coupling signal, a comprehensive arc fault feature set is constructed by extracting features from both single-cycle and multi-cycle perspectives. Thirdly, the arc fault decision tree is trained using the constructed feature set, and the arc fault decision criteria are designed based on the persistence characteristics of arc faults. Finally, the superiority of the proposed method is demonstrated through offline tests, online verification on a self-developed prototype, anti-interference experiments, and a comparative analysis. Experimental results show that the proposed method achieves a detection accuracy of 99.20 % under test conditions encompassing thirteen loads, four circuit topologies, and two arc-generating modes, outperforming five representative benchmark methods under identical test conditions. Furthermore, during 360 online tests covering both training conditions and previously unseen real-world operating conditions, the proposed method satisfied standard detection time requirements in 354 tests while maintaining zero false alarms under arc-like conditions such as load switching.