Research On Unsupervised Feature Extraction Of Transformer Surface Vibration Signal

Transformer as one of the important equipment in power system,the operating status of the power system has a very important impact on the safety and economic benefits.The characteristic extraction of the vibration signal is the basis and premise of monitoring the transformer status by the vibration method.In this paper,we study the feature extraction of transformer surface vibration signals in operation,and use the wavelet packet analysis technology to represent the band-energy characteristics of transformer surface vibration signals.According to the mutual information and feature selection criteria UmRMR(unsupervised maximum correlation minimum redundancy)To sort and select the features of the vibration signal on the transformer surface,to realize the feature extraction of the vibration signal on the surface of the transformer and to provide technical support for the analysis of the operating status of the transformer by the vibration method.According to the mechanism of transformer vibration and the need of fault diagnosis,a set of convenient vibration signal acquisition system was selected to set the relevant acquisition parameters.The spectrum analysis of several different types of transformer vibration signals showed that different types of transformers According to the law of frequency and energy distribution of the transformer vibration signal in actual operation,the wavelet packet transform is used to decompose the vibration energy of the vibration signal.A wavelet packet analysis based on the surface vibration signal of the transformer-Energy representation to complete the characterization of the transformer vibration signal.According to the vibration method,there is a lack of fault samples,unknown fault types,and a large amount of redundant and irrelevant data in the transformer on-line monitoring data.Through mutual information to measure the correlation and redundancy between features and to determine the dependence between features,unsupervised UmRAR criterion is used as the basis to evaluate the importance of features.Unsupervised feature selection method of transformer surface vibration signal based on mutual information to realize the importance ordering and selection of features.The experimental data of the standard dataset are validated to verify the feasibility and effectiveness of this method.After extracting the characteristics of the measured vibration signal of the transformer,the experimental data set is designed and constructed,and the clustering and classification simulation are completed on the experimental data set.The results show that the feature selection method of this paper can correctly distinguish the different measurements of the transformer on the basis of effectively reducing the data dimension Compared with the traditional method of arranging in the order of the features,the point surface vibration signals can select the important features and improve the performance of the learning algorithm.