Application of Machine Learning Technique and Sweep Frequency Response Analysis to Classify Stator Winding Insulation Defects

The on-line partial discharge monitoring system has become a significant tool for protecting high voltage motors' stator winding from incipient failure. High voltage motors are typically produced through the Global vacuum pressure impregnation (VPI) process. Therefore, on-line partial discharge (PD) sensors, such as high voltage coupling capacitors, are connected to high voltage feeders in a terminal box. However, installing external PD sensors has additional costs and requires more effort. To overcome these disadvantages, researchers have developed a 6.6 kV Global VPI motor embedded with six capacitive slot coupler (CSC) PD sensors and successfully operated it in a power plant. In this report, a machine learning technique that can classify the defect types of stator windings is introduced. The classification process is based on the pattern recognition of phase-resolved partial discharge spectra obtained from CSC PD sensors. In recent decades, critical turn-to-turn insulation failures have been reported in many power plants. A research project that assesses the degradation condition of turn-to-turn insulation was initiated in 2020 to overcome these failures. The project's first goal was to find any parameter that can represent the degradation depth of turn-to-turn or strand-to-strand insulation. The second is to classify the type of insulation defect in stator windings. After reviewing several machine learning techniques, a multiclass support vector machine (SVM) was selected as the main classification algorithm. In the experiment, 35 stator windings with 4 defect types (slot discharge, turn insulation discharge, turn-to-main insulation discharge, and main insulation discharge) were trained using the SVM algorithm, and 7 test stator windings were selected to classify the defect types. Due to the small number of training samples, the classification accuracy was 71.4% with the radial basis function kernel. The accuracy is expected to improve further with a larger training sample. The application of sweep frequency response analysis is the third goal of the project for monitoring the health condition of the individual stator winding. This technique may contribute to assessing the deteriorated condition of turn insulation by switching surge.