Quiet-State-Anchored Simulation-to-Reality Adaptation for Arcing High-Impedance Fault Detection

High-impedance fault detectors are commonly trained in simulation, yet their transfer to physical arcing remains poorly quantified. Using a report-calibrated source generator and 983 physical vegetation-conduction tests from the Vegetation High-Impedance Fault (VeHIF) data set, we found that a compact one-dimensional convolutional neural network (1-D CNN) retained strong target ranking (area under the receiver operating characteristic curve (AUROC) 0.985), but its source-domain threshold reduced F1 to 0.628: decision calibration dominated the transfer gap. We therefore set the operating threshold as a score quantile of amplitude-silent unlabeled target windows and refine the model by two rounds of conservative self-training. Across ten seeds, the detector achieved F1 0.950±0.024 and AUROC 0.990±0.002, which was 32.2 F1 points above source-threshold transfer (paired Wilcoxon p=0.002); five further recording-wise splits reproduced the comparison. With a 3-of-5 persistence rule, 94.8% of fault tests raised an alarm (median decision-complete delay 0.6 s) and no background or non-ignition recording did; the pooled pre-fault window rate of 10.3% (11.8–16.2% on further splits) came from sub-threshold conduction in about 9% of tests. Surrogate tests showed that the detector responds to arc waveform shape but not current presence. On an independent 60 Hz covered-conductor campaign, a zero-shot check retained partial ranking (AUROC 0.89) but required re-anchoring to its quiet state. Quiet-state anchoring is thus a practical label-free calibration for channels dominated by the fault-loop current; superposed load current and arcing loads remain outside its scope.