Hybrid Deep Learning Models for Enhanced Classification of Phase-Resolved Partial Discharge Patterns From High-Voltage Rotating Machine Insulation

Abstract

Partial discharge (PD) monitoring plays a crucial role in identifying insulation defects in high-voltage rotating machinery, where accurate classification is essential for improving the reliability and efficiency of condition-based maintenance (CBM). This work proposes hybrid convolutional neural network (CNN) models to classify phase-resolved PD (PRPD) patterns from six different defects in a rotating machine insulation. Various hybrid models were evaluated by integrating CNN with machine learning (ML) algorithms, which include support vector machines (SVMs), k-nearest neighbors (KNNs), logistic regression (LR), decision trees (DTs), random forests (RFs), and naive Bayes (NB). The results reveal that all proposed hybrid models consistently outperform CNN in terms of computational efficiency, by achieving an average accuracy of 94.87% across all models using two optimizers, ADAM and stochastic gradient descent with momentum (SGDM). Notably, CNN-RF (CNN-RF) and CNN-KNN (CNN-KNN) models achieve the best performance, with an accuracy exceeding 96% with lower computational time compared to CNN, which only achieves 94.44% accuracy. Thus, this work provides valuable insight into enhancing PRPD classification with lower computational cost while increasing the classification accuracy of PRPD patterns from rotating machine insulation.