Use of Hidden Markov Models for Partial Discharge Pattern Classification

The importance of partial discharge (PD) measurements for diagnosis of defects in insulation systems is well known. The image patterns obtained in these measurements contain features whose analysis leads to identification of the PD cause. These features are the phase position and amplitudes of PD pulses appearing on the image pattern (usually displayed on elliptic time base on conventional detectors). There is a close similarit y between PD signals and speech. Both are time-varying and similar in behavior. Hidden Markov models (HMM) have been very successful in modeling and recognizing speech. Hence, an attempt was made to employ them to classify PD image patterns. Basis for selection of model and training parameters and the obtained results are discussed. It is shown that successful recognition of PD image patterns using HMM is possible. The ability of HMM to classify some actual PD image patterns has also been ascertained