An Intelligent Framework for Optimal Feature Selection of PD Signals Produced by Electrical Defects in MV Switchgear

This study presents a novel feature-wise non-linear least absolute shrinkage and selection operator (Lasso) based technique for optimal feature selection of high-dimensional features in Partial discharges (PD) data. Laboratory conditions are employed to develop four insulation defects in medium voltage (MV) switchgear, extracting raw PD data. An experimental platform comprising a hybrid PD detection technique is employed to capture 4000 conventional and 4000 unconventional PD signals. Also, 4000 cumulative energy (CE) signals are generated from unconventional PD signals. From this dataset comprising 12000 signals, 35 typical PD features, 490 two-dimensional features, and 777 three-dimensional features are achieved. Finally, a novel feature-wise non-linear Lasso-based feature selection technique is implemented for optimal feature selection of PD data. The effectiveness of the proposed method is evaluated through different classifiers. The proposed technique demonstrates exceptional performance in feature selection for PD data in MV switchgear with the potential for application to other power system components.