Feature Extraction Techniques from Contaminated-Insulator Leakage Current Data for Deep Learning Applications in Condition Monitoring of High Voltage Insulators

This paper presents advanced feature extraction techniques for analyzing leakage current data from contaminated high voltage insulators, aimed at enhancing deep learning applications in condition monitoring. We extract and evaluate time-domain, frequency-domain, and wavelet domain features from leakage current signals. Using Random Forest and permutation importance, we identify key features such as temperature, relative humidity, peak-to-peak values, wavelet energy, and spectral centroid. Correlation and scatter matrix visualizations illustrate the relationships between features and their ability to distinguish pollution levels. A careful selection of features facilitates efficient training of deep learning algorithms to generalize across various contamination levels and environmental conditions.