Research On Fault Diagnosis Of Oil-immersed Transformer Based On Reconstructed Clustering Analysis

The Dissolved Gas Analysis(DGA) method can monitor the internal state of the running transformer without power failure,and is effective for discovering the latent faults and fault development trends inside the transformer.Therefore,it has been widely used in engineering practice.The International Electrotechnical Commission(IEC)has recommended a variety of DGA analysis methods,including the characteristic gas method,the ratio method,the David triangle method,etc.,but these methods have too many fault boundaries,and the fault recognition accuracy rate is usually not sound.Scholars from various countries have improved them by using artificial neural networks,genetic algorithms and other intelligent algorithms,and some effects have been achieved.However,these methods are still limited to the analysis method of the ratio method.It is difficult to fundamentally make up for the problem that the fault boundary is too absolute in the DGA method,so it is limited to a certain extent in practical application and promotion.Fuzzy C-means Method(FCM)classifies things according to similarity,and uses membership degree to describe the category of samples.It can effectively avoid the defects of excessive classification,and has a good application prospect in DGA data analysis.Based on the fuzzy clustering method,this paper studies the fault analysis method of large oil-immersed transformers based on DGA data.The specific research contents and achievements are as follows.(1)Based on DGA data,the initial value sensitivity problem of fuzzy clustering algorithm is analyzed and studied.In this paper,the traditional fuzzy clustering algorithm is used to classify the DGA data,and the results show the traditional FCM algorithm is not suitable for valid transformer fault classification using DGA data,as it could not return a valid pattern containing the six fault types.Different clustering results were obtained for different iteration starting points of the same dataset,and the classification result is too sensitive to the starting point of the iteration.A Gaussian distribution point set is generated on the two-dimensional plane to test the distribution characteristics of the sample.The test shows that it is beneficial to the fuzzy clustering algorithm to classify the samples when the sample exhibits cluster characteristics in the cluster space and the distribution between the clusters is relatively balanced.When the sample distribution is scattered and there is no obvious interval between the data of each category,the objective function of the algorithm has many local extremum points,and the iterative process is sensitive to the initial value at this time.The DGA data of various typical faults of the transformer is not compact in spatial distribution,and there is no obvious cluster feature.There is overlap between the data,and there is no clear boundary between the various fault type data.Therefore,the classification results are not ideal when using traditional FCM to analyze DGA data.(2)Reconstruct the fuzzy clustering algorithm so that the membership function varying monotonically with sample differences.The traditional FCM algorithm uses the power of the reciprocal of the Euclidean distance to describe the similarity between the samples.The resulting membership function has many local extremum points,which is not a monotonic function of the sample difference.Therefore,when using this method to diagnose transformer faults,the failure pattern of the transformer cannot be effectively identified.In this paper,based on the basic requirement of "the larger the sample difference is,the smaller the membership degree is",after a rigorous mathematical derivation,the membership degree calculation function is reconstructed,so that it is monotonic to the sample difference,and there is no local extreme point.In this way,the division of the cluster space by the algorithm is optimized.Moreover,with Zangwill's theorem,the convergence of the reconstructed clustering algorithm is analyzed.The analysis shows that starting from an arbitrary starting point,the reconstructed clustering algorithm proposed in this paper can converge to the extreme point of the objective function,which can realize the partitioning of the sample set.Combined with the artificially generated two-dimensional Gaussian distribution point set and the Iris data set in the UCI(University of California Irvine,UCI)database,the proposed reconstruction clustering algorithm is tested.The test results show that the reconstructed algorithm can classify the sample set,and compared with the traditional algorithm,the proposed algorithm considerably improves the convergence speed of FCM clustering and outperforms the traditional one.The reconstructed algorithm is used to analyze the DGA dataset,as expected,data samples corresponding to the same fault data were categorized into the same subclass,and different fault data samples were classified into different subclasses.(3)A multi-track parallel clustering algorithm based on chaotic sequence is constructed.Although the reconstructed clustering algorithm eliminates the local extremum of the membership function,the objective function of the clustering algorithm is a typical non-convex function,and the iterative optimization process belongs to the local search method based on gradient calculation.So the situation where the optimal process falls into the local extreme point still exists.In this paper,an iterative method for parallel computation of multiple trajectories is constructed.In the iterative process,information can be shared and utilized between each trajectory.The position of each point is determined not only by the gradient information of the previous point of the trajectory,but also affected by the trajectory history information and the remaining trajectory history information;thereby the trend of the trajectory can be adjusted.Moreover,in the iterative process,chaotic sequences are introduced to fully search for potential feasible solutions in the cluster space,and the optimization ability of the algorithm is improved.The algorithm is tested with DGA data,and the test shows that with the increase of the number of parallel trajectories and the increase of chaotic detection frequency,the global optimization ability of the method is effectively enhanced.(4)DGA fault diagnosis analysis based on reconstructed clustering algorithm and its application.Based on a large number of(two thousand) DGA data,the reconstructed clustering method is used to obtain the chromatographic data of six typical faults such as low temperature overheating,medium temperature overheating,high temperature overheating,partial discharge,spark discharge and arc discharge,and they are used as its reference standard sequence for fault diagnosis.For the DGA data to be diagnosed,the membership degree between it and the above six reference standard sequences is calculated in turn,and the degree of belonging to various typical faults is characterized by membership degree,thereby the fault diagnosis is implemented.The method of this paper abandons the idea of the ratio method and avoids the defects such as the faulty boundary data of traditional DGA analysis methods such as David triangle method and three-ratio method.The test of a large number of field fault cases(300 fault transformer examples)shows that the proposed method can describe the severity of the fault in more detail,especially for the lack of code data,the composite fault and the fault in the intermediate state,and the fault diagnosis correct rate is high,which is an effective fault diagnosis method.