Research On A Comprehensive Experiment Platform And Algorithms For Partial Discharge Pattern Recognition

As a significant representation for insulation defects, partialdischarges have long been an important part of insulation detection.The methods applied to partial discharge detections have differentprinciples and realizations, which directly leads to the difficulty incomparison of results. With an experiment platform built supportingparallel analysis of multiple detection signals, systematical research on therelevance and difference between detection methods and physicalquantities is then possible. Due to the difference in insulation structuresand mechanisms, partial discharges are diverse. The harm of differentpartial discharges are varied. It is thus meaningful to recognize partialdischarge types correctly for accurate assessments on insulationdegradation and reasonable strategy of maintenance.In this dissertation, the concept of a comprehensive experimentplatform is put forward. It is designed and built, supporting the ERAmethod, the high-frequency current transformer technology, the ultra-highfrequency technology and the acoustic emission technologies. All-roundresearch on physical phenomena generated by partial discharges is possible.The partial-discharge-causing insulation defect types are summarized,multiple typical defect models are designed and made.Partial discharge pattern recognition methods are systematicallyresearched. Basing on the Hilbert transform, the novel algorithm forfeature extraction from oscillating signals is presented; basing on thecomparative agglomeration algorithm, novel cluster features of phase-resolved partial discharges are presented. A dimensionality reductionmethod based on linear discriminant analysis solved by a modified particleswarm optimization is presented. Classification based on the minimumerror-rate Bayesian decision using kernel density estimation is realized.The algorithm is verified and analyzed with experiment data. The resultshows the algorithm presented in this dissertation provides goodperformance, which is better than that of a back-propagation neuralnetwork.