Orthogonal least square center selection technique - A robust scheme for multiple source Partial Discharge pattern recognition using Radial Basis Probabilistic Neural Network

Partial Discharge (PD) pattern recognition has emerged as a subject of vital interest for the diagnosis of complex insulation system of power equipment to personnel handling power system utilities and researchers alike, since the phenomenon serves inherently as an excellent non-intrusive testing technique. Recently, the focus of researchers has shifted to the recognition of defects in insulation due to multiple PD sources, as it is often encountered during real-time PD measurements. A survey of research literature indicates clearly that the recognition of fully overlapped PD patterns is yet an unresolved issue and that techniques such as Mixed Weibull Function, Neural Network (NN), Wavelet Transformation, etc. have been attempted with only reasonable success. Since most digital PD online acquisition systems record data for a stipulated and considerable duration as mandated by international standards, the database is large. This poses substantial complexity in classification during the training phase of the NNs. These difficulties may be attributed to ill-conditioned data, non-Markovian nature of discharges, curse of dimensionality of the data, etc. Since training methods based on random selection of centers from a large training set of fixed size are found to be relatively insensitive and detrimental to classification in many cases, a Forward Orthogonal Least Square algorithm (FOLS) is utilized in order to reduce the number of hidden layer neurons and obtain a parsimonious yet optimal set of centers. This algorithm, in addition, obviates the need for a separate clustering method making the procedure inherently viable for on-line PD recognition. This research work proposes a novel approach of utilizing Radial Basis Probabilistic Neural Network (RBPNN) with FOLS center selection algorithm for classification of multiple PD sources. Exhaustive analysis is carried out to ascertain the efficacy of classification of the proposed RBPNN-FOLS algorithm to cater to large training data set. A detailed comparison of the performance of the proposed scheme with that of the standard version of Probabilistic Neural Network (PNN) and Heteroscedastic PNN (HRPNN) that was taken up for study by the authors in their previous work indicates firstly the effectiveness of FOLS algorithm in obtaining parsimonious centers, points out secondly the capability of the Radial Basis Probabilistic Neural Network (RBPNN) model to integrate the advantages of the Radial Basis Function Neural Network (RBFNN) and PNN in classifying multiple PD sources and finally throws light on the exceptional capability of the FOLS-RBPNN in discriminating the sources of PD due to varying applied voltages also.