Research on the dimension reduction of the partial discharge feature space for oil-paper insulated power transformer

Abstract Oil-paper insulation condition directly affects transformer reliability. This study evaluates how dimensionality-reduction performance varies with feature type and diagnostic task using laboratory-simulated gas-gap, column-to-plate, pin-to-plate, and inter-turn partial discharges. Statistical (67-dimensional), waveform (13-dimensional), dissolved-gas-analysis (17-dimensional), and combined (97-dimensional) feature spaces were reduced using KPCA, CCMDR, and RST. RST produced the greatest separability improvement for statistical, waveform, and combined features, whereas CCMDR performed best for DGA. With BPNN, combined features with CCMDR achieved the highest defect-type accuracy (93.02%), and combined features with RST achieved the highest severity accuracy (91.86%). Matched RBF-SVM and Random Forest experiments preserved the ordering of all 16 feature-space/reduction configurations (Spearman rho = 1.000 for both tasks). In sequential type-severity recognition, combined CCMDR-RST achieved the highest joint accuracy (90.70%), while waveform RST-RST had a lower normalized batch-processing time (0.228 versus 0.923 s per test sample); both were Pareto-efficient. The results provide feature- and task-specific selection guidance among the evaluated methods. Because the evidence is limited to laboratory single-source defects, field validation is required before deployment.