Investigation of AI-Based Detection Methods for AC and DC Series Arc Faults in Three-Phase DC/AC Inverters

Abstract

Series arc faults occurring at the AC output and DC input sides of a three-phase pulse-width modulation inverter can degrade system reliability and increase the risk of electrical accidents. Because the effects of an arc on the output-current waveform vary depending on the fault location and operating condition, the diagnostic characteristics of AC and DC series arcs should be evaluated separately. In this study, the performance of several artificial-intelligence models was compared using a single A-phase output-current signal measured under various inverter operating conditions. The measured current was used as a time-domain sequence and was also transformed into a frequency-domain magnitude spectrum using the fast Fourier transform. The experimental results showed that the CNN-Transformer achieved the highest diagnostic accuracy for both AC and DC arc faults with both input representations. In particular, the frequency-domain input provided high precision and recall for both AC and DC arc conditions. These results demonstrate that both AC output-side and DC input-side series arcs can be effectively diagnosed from the output-current characteristics, while the diagnostic performance depends on the fault location, input representation, and model structure.