电压特征能量驱动的电气火灾安全感知方法——仿真数据集与分析代码

Series arc faults in low-voltage AC wiring, caused by loose connections, insulation aging, or conductor breaks, produce fault currents that remain below the tripping threshold of conventional overcurrent protection devices. The arc burns unattended. This work addresses the detection gap with a method that relies solely on the voltage waveform at a single monitoring point, avoiding the load-dependence that limits current-based approaches. Continuous wavelet transform analysis isolates the 3-40 kHz band as the frequency region where arc-induced transients concentrate. Loads are classified by AC-side high-frequency impedance into two categories. For switch-mode power supply (SMPS) loads, a full-cycle RMS energy criterion (Rule A) exploits the rectifier dead-zone to achieve high-sensitivity discrimination. For non-SMPS loads, resistive heaters and induction motors among them, a phase-window energy mapping criterion (Rule B) locates the reignition signature within 15-45 degrees after current zero-crossing. A Mayr arc model augmented with a controlled breakdown-reignition module is implemented in MATLAB/Simulink and exercised across four representative load types: SMPS power supply (300 W), resistive heater (1500 W), induction motor (750 W), and VFD-driven load (1100 W). For the first three types, arc-state band energy reaches 64-82 times the normal-state baseline. An adaptive three-step countermeasure is developed for the variable-frequency drive (VFD) scenario, where the 8 kHz PWM carrier overlaps the sensing band and suppresses the energy ratio to 0.9. The cascade, adaptive carrier detection, a bank of five IIR notch filters (Q = 30, -3 dB bandwidth approx. 267 Hz per filter), and multi-subband energy verification, restores the ratio to 37.7. Independent testing on the Wu (2024) public arc-fault dataset (100 kS/s sampling, UL 1699B compliant) covers ten load configurations, including four mixed-load cases: energy ratios span 2.1 to 79.6, all exceeding the 1.5x detection threshold. End-to-end sensing latency is under 200 ms.