A New Hybrid Feature Extraction Method for Partial Discharge Signals Classification

摘要

In this paper, a new hybrid feature extraction method combining adaptive optimal radially Gaussian kernel (AORGK) time-frequency representation with two dimensional nonnegative matrix factorization (2DNMF) is proposed for partial discharge (PD) classification. Firstly, AORGK is applied to obtain the time-frequency matrices of PD ultra-high-frequency (UHF) signals. Then 2DNMF is employed to compress the AORGK amplitude (AORGKA) matrices to extract various feature vectors with different (d_1d_2) combinations, i.e. (5, 5), (5, 10), (10, 5) and (10, 10). Finally, the extracted features are classified by fuzzy k nearest neighbor (FkNN) classifier and back propagation neural network (BPNN). 600 samples sampled from four typical artificial defect models in Laboratory are adopting for testing of the proposed feature extraction algorithm. It is shown that the successful rate by FkNN and BPNN are all higher than 80%, and FkNN has superior classification accuracies than BPNN under four circumstances of (d_1 d_2) combinations. In addition, FkNN achieves the highest classification accuracy 93.73% with (10, 5) combination. The results demonstrate that it is feasible to apply the proposed algorithm to PD signal classification.%W artykule przedstawiono nową hybrydową metodę klasyfikacji wyładowań niezupełnych (ang. Partial Discharge), wykorzystującą algorytm AORGK (ang. Adaptive Optimal Radially-Gaussian Kernel) o nieujemnej, matrycowej faktoryzacji dwuwymiarowej (ang. 2-Dimensional Nonnegative Matrix Factorization). W metodzie wykorzystano także algorytm k najbliższych sąsiadów oparty na teorii zbiorów rozmytych (ang. Fuzzy k Nearest Neighbour Classifier) oraz sieci neuronowe (ang. Back Propagation Neural Network).