Study On Feature Extraction And Pattern Recognition Approaches For Partial Discharge Signal In Power Transformer

Power transformer is the key equipment in the electrical system and it is important to maintain its safety and stable operation. Partial discharge is the important symptom and manifestation of insulation deterioration in power transformer. Effective detection of partial discharge has practical significance in insulation evaluation of power transformer. Deep studies on interference suppression method, feature extraction method and pattern recognition method for partial discharge were conducted based on the principles and properties of partial discharge signal. The main work of this dissertation is as follows.A denoising method for partial discharge signal in power transformer based on translation invariant wavelet footprint was proposed. The denoising principle of wavelet footprint was studied and the method was applied in denoising of partial discharge signal. Cycle spanning method was used for eliminating the translation dependence, so as to restrain the oscillation phenomenon and strengthen denoising effection. Effectiveness of the method was validated by case studies.A novel feature extraction method for partial discharge based on cross-wavelet transform was proposed, aimed at the high sensitivity of the traditional feature extraction method to noise. The analytical characteristics of signal in time and frequency domain were described by cross-wavelet transform. The method was used for processing of the partial discharge signal and the feature parameters were obtained which represented the characteristics of the cross diagram. Finally the correlation coefficient matrix method was applied to correlation analysis between feature parameters. Case analysis demonstrated that the proposed method could effectively avoid the influence of noise.A pattern recognition method for partial discharge based on active learning SVM was proposed. The idea of active learning was applied to "one against one" multi-class SVM classifier. The sampling function based on posterior probability was used for sample selection. Those samples which were valuable to the classifier were selected for training. Case analysis showed that, the proposed method could reduce the number of training samples and improve the learning efficiency on the premise of maintaining high recognition accuracy.The theory of principal component analysis (PCA) method was considered for deep analysis. PCA method was used for processing high-dimensional statistical parameters. Fewer principal component factors were extracted to represent original signal characteristics. The relevance vetor machine (RVM) classifier was applied to pattern recognition with the characteristic parameters before and after dimension reduction. The effectiveness of the proposed method for dimension reduction was validated by case analysis.A pattern recognition method for partial discharge based on the multi-kernel multi-class relevance vetor machine (MMRVM) was proposed. The method integrated information of different discharge sources with different kernel functions. Particle swarm optimization algorithm was utilized for optimal design to obtain the optimal combination parameters. Experimental data of partial discharge indicated that the designed classification model integrated various feature information and could represent partial discharge characteristics comprehensively with higher diagnostic accuracy.