The Research On Fault Feature Extraction And Diagnosis Method Of Power Transformer

The safe and stable operation of power transformer is of great significance to the entire power system.The application of fault diagnosis technology is an important means to ensure the safety of power transformers.Dissolved Gas Analysis(DGA)can provide an important basis for transformer fault diagnosis.The gas contents support online monitoring and analysis of it helps to detect early latent faults to effectively prevent serious faults.It is a key concern of power transformer fault diagnosis to study how to extract the feature quantities sensitive to the operating state of the equipment from the original measurable characteristic gas data,and to study the fault classification based on this.In this paper,the problem of power transformer fault state identification is studied from two aspects: feature extraction and fault classification.The main research contents are as follows:(1)Due to the correlation and redundancy of the original characteristic data of the power transformer,the subsequent fault state recognition performance will be affected,thus Kernel Neighbor Preserving Embedding(KNPE),an nonlinear manifold learning method is introduced on the purpose of solving this problem.The following three improved models are proposed,and the example analysis verifies the effectiveness of these models.?The KNPE algorithm ignores the category information and the relationship between samples when evaluating the similarity of neighbors.Aim at this problem,a new similarity measuring method——Supervised Kernel Shared Nearest Neighbor(SKSNN)is proposed and is applied to KNPE to form a A new Kernel Neighbor Preserving Embedding Based on Supervised Kernel Shared Nearest Neighbor(SKSNN-KNPE)feature extraction algorithm.SKSNN-KNPE makes full use of the sample category information and the closeness between the samples to construct the local neighbor graph,which is more conducive to mining the low-dimensional sensitive features in the original data set to effectively expose the low-dimensional manifold structure of the original data set.?Aiming at the problem that the KNPE algorithm does not grasp the global distribution characteristics of the sample data,an orthogonal global and locality preserving embedding(OGLPE)nonlinear manifold learning method is proposed.The method takes into account the global distribution variance maintenance characteristics and the local nonlinear manifold structure retention characteristics,and introduces orthogonalization conditions to eliminate the redundant information between the embedded vectors.Compared with the KNPE algorithm,it is more beneficial to extract the effective information in the original data set.?It is difficult for the unsupervised kernel attribute reduction algorithm to determine the kernel parameters.To solve this problem,an adaptive kernel parameter optimization learning framework for multi-classification tasks is constructed,which is combined with KNPE algorithm to form a Supervised Adaptive Kernel Neighbor Preserving Embedding(SAKNPE)feature extraction algorithm.The algorithm makes full use of the sample class information to establish a data-sensitive adaptive kernel function,and with the aim of maximizing intraclass similarity and minimizing the similarity between classes,the weight vector representing the category information in the adaptive kernel function is calculated to realize the adaptive optimization learning of kernel functions to prevent adverse effects of improper selection of kernel parameters on the final result.SAKNPE has significant advantages over unsupervised KNPE algorithms when dealing with multi-class learning problems such as power transformer operating state identification.(2)Considering the shortcomings of k-nearest neighbor(kNN)classification algorithm,a cloud adaptive fuzzy k-nearest neighbor(CAFkNN)classification method is proposed.The classification ability and anti-noise ability of CAFkNN method are verified on the UCI data sets.Then the application of CAFkNN to power transformer fault state identification shows that it has strong generalization ability and high robustness,and can maintain excellent diagnostic results even in the face of small sample problem or unbalanced data set.(3)In order to effectively integrate the feature information in DGA data and further improve the accuracy and reliability of the diagnosis results,a comprehensive power transformer diagnosis model with information fusion realized in feature layer and decision layer is established.In the feature layer,feature space fusion is implemented by weighting,which improves the recognition accuracy of single diagnostic models.In the decision-making layer,based on DS evidence theory,the results of each single diagnosis model are effectively integrated to overcome the one-sidedness and uncertainty of the conclusion of a single diagnosis mode.The case study shows that the comprehensive diagnosis model can effectively improve the accuracy and reliability of the diagnosis results compared with the diagnosis model based on single feature space and single classifier.