Prediction of Discharge Electrode Structures from Electroacoustic Signals Based on Machine Learning

Accurate identification of electrode structures is essential for diagnosing discharges induced by insulation defects in power systems. This study presents a machine learning–driven diagnostic framework that maps electroacoustic patterns to electrode geometries. A versatile discharge platform was developed, supporting three power modes (pulsed, direct current, and alternating current) and eight electrode configurations to simulate typical defect scenarios. Mel-frequency cepstral coefficients (MFCCs) and short-time Fourier transform (STFT) with a Hamming window were employed to extract discharge electroacoustic features sensitive to different electrode structures under various excitation sources. Four classification algorithms—deep neural network (DNN), convolutional neural network (CNN), k-nearest neighbors (KNN), and support vector machine (SVM)—were evaluated in terms of classification accuracy and F1 score. Among them, the DNN achieved the highest predictive performance, enabling precise recognition of the discharge source geometry. This approach replaces traditional physics-based modeling with a data-driven method, facilitating condition assessment and defect localization in high-voltage insulation systems. It significantly enhances real-time monitoring capabilities and demonstrates strong potential for industrial deployment.