Partial discharge classification in power electronics applications using machine learning

A study of machine learning (ML) methods for classification of data from partial discharges (PDs) is described. A novel set of features are suggested and tested using an extensive set of machine learning based algorithms. The aim is to classify PDs occurring within insulation systems and power electronics devices (PED). Due to the increased use of pulse width modulation waveform (PWM) in PEDs, an increased insulation degradation has been observed due to a more intense PD exposure. This study aims to develop suitable tools to detect types of defects to facilitate diagnostics as well as to improve isolation system design. To evaluate the performance of ML based classification, several algorithms have been developed to detect and classify PDs from different kind of material defects with the aim to address the reason behind the appearance of partial discharges. Experiments with different PD source locations and volume of the defect and voltage rise time were investigated on an artificial cavity test object. Relevant signal features found important are for example the maximum magnitude, duration, the distance from polarity shift, the time distance between PDs and the absolute value of the area of the detected PD waveform. It has been observed that forming such PD features based on their time occurrence results in an accurate and generalized solution. With these features the best results were achieved with the deep learning LSTM architecture reaching a test accuracy of 98.3%. For industry applications, feature engineering is useful to reduce amount of data necessary to be analyzed by the neural network or ML algorithm