Study On Oil-immersed Power Transformer Fault Diagnosis Based On Relevance Vector Machine

Fault diagnosis of power transformer based on Dissolved Gas Analysis (DGA) is a sensitive potential failure detection technique, which can be carried out while transformer is running. This method is of great important practical significance, which can promote the realization of condition maintenance from original regular maintenance, and improve the operation and maintenance level. Based on the analysis of the characteristics and shortcomings of the existing diagnosis methods, the relevance vector machine (RVM), which can solve the small-sample, high-dimensional, and non-linear classification problems, was firstly applied to the fault diagnosis of oil-immersed power transformer in this paper. A new way for DGA-based fault diagnosis is explored.A RVM-based fault diagnosis model of oil-immersed power transformer was built by binary tree method. The affection of the feature variables and kernel functions on diagnostic performance was investigated, and the implementation procedure of fault diagnosis was provided in detail. The diagnosis model can provide probabilistic outputs, and is especially suitable for online diagnosis due to high diagnosis speed. It solved the diagnosis problem of lacking of sample data, and its diagnosis performances were validated by case studies.A fault diagnosis method for oil-immersed power transformer based on multiclass RVM was proposed. This diagnosis method can directly implement multi-state identification and provide the probability of each state. It takes the advantage of original RVM which decomposes fault diagnosis into multiple binary classifications, and overcomes the disadvantages of classification overlap, classification failure, multiple-classifier need and error accumulation. The accuracy of this diagnosis method was validated by real-world diagnosis cases.Combination kernel learning method, as well as kernel parameters optimization method was studied. On this basis, a fault diagnosis method of oil-immersed power transformer based on multi-kernel learning RVM was proposed. The method can integrate the feature information reflecting the operating state from different perspectives. DGA-based diagnosis cases verified the effectiveness of the integration method.Cost-sensitive RVM (CS-RVM) based on Bayesian risk theory was proposed and applied to fault diagnosis of oil-immersed power transformer. The CS-RVM based diagnosis method introduced the idea of considering misdiagnosis cost to fault diagnosis, aiming to minimize misdiagnosis cost. It could overcome the problem of bringing no meaningful results for only pursuing high classification accuracy. Experimental results showed that CS-RVM diagnosis method tended to increase the diagnostic accuracy of high misdiagnosis cost category, and diagnosis speed was high enough to meet the engineering requirement.