Neural network based PD source classification using a combined topology of unsupervised and supervised learning algorithm

Partial discharge distribution patterns derived from EEC270 defined parameters are investigated. It has been found that normative references can be employed not only to monitor the discharge activity level but also to identify different type of discharge sources. This is because the distribution pattern contains information about the origin of discharge source. PD distribution of phase resolved (PRPD) and pulse height resolved (PHPD) patterns are analyzed with the aid of a computer-based PD detector. The parameters of various distribution patterns are extracted to form a feature vector or fingerprint. There is an advantage of using the combined topology of unsupervised and supervised learning neural networks to train feature vectors of large size. The results in classifying PD sources on a few samples have confirmed the usefulness of this new approach.