Performance Evaluation of AI-based Algorithms for Condition Assessment of Power Components

This study compares the performance of different artificial intelligence (AI) based algorithms/ classifiers used for partial discharge (PD) classification during insulation diagnostics in power components. During PD measurements, a considerable amount of data is collected, and processing such a huge amount of data is time-consuming and expensive. The useful PD signals can be extracted from the measurements, and AI-based algorithms can be used to process those signals for classification and diagnostics purposes. In this work, the data is collected from three different PD sources, namely, corona, internal, and surface in the high voltage laboratory. Each measurement consists of the PD activity captured in the form of power frequency cycles. The single PD pulses are extracted from the measured signals using the segmentation method. For features extraction, at first discrete wavelet transform (DWT) technique is applied on single pulses, and then statistical parameters (mean, standard deviation, skewness, and kurtosis) are applied to the extracted features. To classify different PD sources, two different classifiers, support vector machine (SVM) and k-nearest neighbors (KNN) with their types, are applied to extracted features. The performance of each classifier is evaluated using the accuracy performance indicator by varying the amount of input PD data from each PD source. The developed understanding will enable researchers/asset managers to extract the required amount of data from the field measurements.