COMPARATIVE STUDY OF HMM AND BPNN IN DETECTING CORONA DISCHARGE ON 20 KV CUBICLE BASED ON VOLTAGE AND SOUND

Abstract

Corona discharge is a common disturbance in 20 kV cubicles and may lead to insulation degradation, power losses, and potential equipment failure. Early detection is therefore essential to support preventive maintenance and enhance distribution system reliability. This study presents a comparative analysis of the Hidden Markov Model (HMM) and Backpropagation Neural Network (BPNN) for identifying corona discharge based on acoustic signals and voltage variations. Acoustic data were recorded using a needle–rod configuration and processed through Linear Predictive Coding (LPC) to obtain cepstral features. The classification results show that HMM provides high accuracy, achieving 100% in both training and testing for noise-based clustering and 84.44% in voltage-based testing. Meanwhile, BPNN demonstrates stable performance with training and testing accuracies of 95.93% and 82.35% for voltage-based clustering, and 95.69% and 100% for noise-based clustering, respectively. Overall, the findings indicate that although HMM performs more consistently for noise-based classification, BPNN provides competitive accuracy and better adaptability for early detection of corona discharge in 20 kV cubicles. These results support the development of intelligent monitoring systems for improving insulation condition assessment and reducing potential failures in medium-voltage distribution networks.