Identification of corona discharges based on wavelet scalogram images with deep convolutional neural networks

This paper presents a new wavelet-based approach to automate the identification of positive and negative DC corona discharge current pulses. Two electrode systems with variable gap spacings formed corona discharges, and two commercial sensors (a high-frequency current transformer (HFCT) and a shunt resistor) captured transient corona discharge currents. The proposed method employs a continuous wavelet transform to generate time-frequency representations of corona discharge pulse currents, called scalogram images. The effects of sampling interval, data acquisition time, data shifting, and external noise components in the signals on the scalogram images were examined. The well-known pre-trained convolutional neural network models, AlexNet, MobileNet, and ShuffleNet, were tailored, and their ensemble structure was generated to discriminate scalogram images of discharge pulses. A framework was constructed to increase the generalization ability of the study. The results demonstrate that the scalogram images are robust candidates for corona discharge identification.