A PPCA based non-parametric modeling and retrieval of PD signal buried in excessive noise

The problem of on-line recognition and retrieval of relatively weak industrial signal such as Partial Discharges (PD), buried in excessive noise has been addressed in this paper. The major bottleneck being the recognition and suppression of stochastic pulsive interference (PI), due to, overlapping broad band frequency spectrum of PI and PD pulses. Therefore, on-line, on-site, PD measurement is hardly possible in conventional frequency based DSP techniques. The observed PD signal is modeled as linear combination of systematic and random components employing probabilistic principal component analysis (PPCA) and pdf of the underlying stochastic process is obtained. The PD/PI pulses are assumed as the mean of the process and modeled instituting non-parametric methods based on smooth FIR filters and a maximum aposteriori probability (MAP) procedure employed therein to estimate the filter coefficients. The classification of the pulses are undertaken, using a simple PCA classifier. The methods proposed by the Authors were found to be effective in, automatic retrieval of PD pulses, completely rejecting PI.