Ensemble Machine Learning Based Adaptive Arc Fault Detection for DC Distribution Systems

The detection of dc arc fault is a challenging task due to the low fault current caused by the high fault impedance, the random nature of arc discharge, and its dependence on current level. The electrical system in real applications creates an even more challenging environment with a large number of electronic loads and versatile operating conditions. This paper presents a Machine Learning (ML) based algorithm for arc fault detection and an experimental testbed for validation. The ML algorithm is trained with experimental arc fault data and an adaptive normalization procedure is proposed to reduce mistriggers. Moreover, a function is designed to ensure detection accuracy with various types of loads. The proposed detection algorithm is implemented on Udoo X86 Ultra microcontroller board and verified with real-time detection tests. The chosen ML algorithm resulted in a high accuracy performance within a relatively low delay time compared to conventional detection methods.