DC series arc diagnosis based on deep-learning algorithm with frequency-domain characteristics

Abstract

A series arc produces both light and heat, which results in a risk of fire. Hence, its prompt detection is very important. The series arc has different frequency characteristics depending on the operating conditions of the inverter and the line impedance. As a result, it is difficult to detect with conventional methods that use specific frequency bands. In this study, series arcs generated by three-phase pulse-width-modulated and three-phase model-predictive-control inverter loads, operating under various conditions, were detected using a deep-learning algorithm. To detect a series arc fault, the frequency characteristics were analyzed using all the frequency bands of a fast Fourier transform (FFT) result. To use all the frequency bands, a deep neural network (DNN), which is a deep-learning algorithm, was used. DNNs can learn quickly from large amounts of data and model complex nonlinear relationships. The DNN model proposed in this paper was designed to use all the frequency bands, and three cases were tested by changing the current magnitude and switching frequency. The arc current detection accuracy was calculated for three test cases and the obtained results were analyzed in detail based on the percentages of true/false and mean absolute error metrics. The analysis confirmed that the arc current was accurately detected, which validates the suitability of the DNN for series arc detection.