Adaptive Detection Method for Arc Faults in Low Voltage Power Supply Systems

With increasingly emerging AC devices, arc fault characteristics would become more uncertain and random. Therefore, it is urged to investigate an adaptive arc fault detection method. Firstly, an experimental platform is constructed to acquire diverse arc fault signals according to IEC 62606 and GB/T31143. A feature-driven intelligent detection network is well-trained based on limited arc fault data. However, it was found that this established network could not detect other unknown arc fault situations. Next, these unknown situations can be accurately picked out by the local maximum mean discrepancy (LMMD) method. The LMMD method is also applied to build the loss function of the detection network. Then the deep subdomain adaptation network (DSAN) is used to minimize the loss value so that the modified detection network could also become effective for unknown situations. Finally, it is verified that the detection accuracy of the proposed method improves by an average of 21% for unknown series and parallel arc faults, regardless of the chosen training conditions.