Feature Pruning for Partial Discharge Classification using IndFeat and ReliefF Algorithm

In this work, feature pruning was implemented to a Partial Discharge (PD) classification system to reduce the size of input data. Reducing the input data feature size will ease the computational burden and reduce the data bandwidth required, which is important for real-time fault classification application. Different insulation defects will produce different PD patterns; hence PD pattern recognition is of interest to researchers worldwide because it enables the precognition of any incipient faults. PD data used in this work were obtained from 5 cross-linked polyethylene (XLPE) cables with artificial cable joint defects. Statistical features were calculated from the measured phased-resolved PD (PRPD) data. Two feature pruning algorithms, Independent Significant Feature Test (IndFeat) and ReliefF were used to decrease the dimension of input data for the classifier training process. Support Vector Machine (SVM) and Artificial Neural Network (ANN) were used as the classifier. Results show that when using the pruned data compared to unpruned data, the PD classifier was able to achieve similar performance even in the presence of noise contamination. The feature pruning algorithms also performed better compared to dimensionality reduction method such as principal component analysis (PCA).