Detection of Arc Discharge through Sound Event Detection Techniques and Public Datasets

Abstract

Arc discharge in switchboard is a potential fire hazard and caused by short-circuits in power supply cords with damaged electrical insulation. Arc detection based on electric current and resistance can be destructed while arc discharge occurs due to circuit contact, and humidity condition of switchboard affects the detection performance of optical signal-based arc detection. This paper proposes a deep learning-based arc discharge detection that uses sound event detection (SED) method. The neural network was trained on public datasets to exclude the processfor obtainingsound dataset of arc discharge. Filter-augmentation techniques were applied to account for the acoustical differences between the public dataset and target environments. The network uses cepstrum in addition to STFT as network input feature for effective extraction of harmonic spectral patterns of arc discharge. The results show that the proposed neural network achieved an F1 score of 0.83 and accuracy of 0.89 when tested on a test-bed with arc discharge generator based on UL1699.