Dynamic Pattern Recognition Approach for Partial Discharge Signature Analysis Using Hidden Markov Models: a Critique Based on Experimental Investigations

摘要

Minor defects in insulation system of power apparatus such as cracks, inclusions, protrusions etc are unavoidable leading to partial discharges (PD). Since recognition of PD sources is essential for diagnosis and as multiple sources are encountered during real-time measurements, an array of techniques such as Neural Networks (NN), Wavelet Transformation etc have been attempted by researchers for classifying single and partially overlapped PD sources with moderate success. Since PD is a complex non-Markovian process displaying statistical variations among correlated patterns, Hidden Markov Model (HMM) serves as a viable tool in recognizing spatio-temporal and dynamically varying signatures. Hence a non-stationary continuous density HMM (CDHMM) which expresses the hidden state transition probabilities as time-dependant estimates with multivariate Gaussian densities is implemented. A novel hybrid non-stationary CDHMM-Probabilistic Neural Network recognition system is developed which utilizes the complementary advantages of HMM for identification of pattern sequences and NNfor discrimination. Exhaustive studies are carried out on benchmark models and industrial objects to determine the efficacy of the hybrid system in comparison with its stationary counterpart. The focus of the research is on analyzing the role of the optimal state transition sequence of the non-stationary version in capturing the dynamic characteristic of PD patterns.