An On-line Feature Extraction Method for Transformer Vibration Signals

FFT or wavelet transform method is usually used in transformer on-line vibration signal analysis apparatus. In order to extract features of transformer vibration signals reliably and accurately, a method which is based on FFT and wavelet packet transform was proposed for type identification of transformer vibration signals using k-nearest neighbor (KNN). First of all, the known transformer vibration signals were decomposed by FFT and several characteristic bands which were used to determine the layer number of wavelet packet decomposition were extracted; secondly, the feature vectors contained energy entropy of a selective wavelet packet decomposition for characteristic bands was obtained; Finally, KNN classifiers are utilized for pattern recognition. Five kinds of winding and core state vibration signals under conditions of transformer short-circuit test in the laboratory were extracted and recognized using the above methods. Simulation results demonstrate that the method proposed can extract the features of transformer vibration signals accurately and reliably, providing a good reference for the recognition of partial discharge signals.