Discriminative Analysis of HVDC Discharges Over Composite Insulators by Feature Selection Combined With SVM

HV insulation systems under ac or dc are subject to partial discharge (PD) phenomena, which happen randomly and can lead to progressive damage of the dielectric material and/or the outage of the line in which they occur. In order to automate the predictive maintenance, these discharges must be monitored in real time to recognize their energetic nature: low-energy (LE) ones that extinguish by themselves or contrarily high-energy (HE) ones, which can lead to flashovers that imply line outage. One has then to focus on HE discharges since they constitute the major risk for the insulation systems and the network stability. For this aim, an experimental investigation was performed for HVdc voltages applied to composite insulators, from levels corresponding to LE PDs to flashover occurrence. Thereby, low noise nanosecond scale measurements are taken simultaneously for voltages, currents, and radiated magnetic fields by both types of discharge where a qualitative database was obtained. This article presents the use of efficient and fast feature selection software that uses the Fisher score as an evaluation criterion. The features are selected and then classified from the most to the least discriminant. The number of the most successive discriminant features that correspond to the best HE discharge recognition ratio is then searched. The obtained results from the achieved algorithm, combined with support vector machine (SVM), allow an important reduction of feature number from 25 000 to only 37 and an improvement of 4.9% of HE discharge recognition when compared to the results obtained before feature selection. The approach is suitable for real-time diagnosis of insulator chains close to substations and power plants.