Studies On Fault Diagnostics For Oil-immersed Power Transformer

Identifying problem at an early stage before catastrophic failure occurs is a great benefit for reliable operation of a power transformer which is one of the most significant equipment in power system. In order to ensure the security and reliability of power system, it is necessary to improve the fault diagnosis technique for power transformer.The fault can be divided into several categories according to different sorting standards such as structure of power transformer, fault characteristic, etc. The fault information of power transformer can be obtained by various detecting techniques such as oil chromatography analysis, routine electrical tests, partial discharge tests, etc. Moreover, it has characteristics such as complementarity, repellency and redundancy. The fault diagnosis of power transformer is to ascertain specific fault type through integrated analysis of fault information on basis of a proper classified mode. In other words, it is a process to build a fault diagnostic model and is the key point of condition-based maintenance (CBM) technology.Dissolved gas-in-oil analysis (DGA) method has been adopted extensively by utilities to detect internal faults without interrupting services. It can determine the fault type through the analysis of the composition of the dissolved gases, rates of generation and specific content ratios. But the actual diagnosis must also consider other information of transformer such as size, volume of oil, type of transformer etc. Artificial neural network (ANN) has been become a promising approach compared with conventional ratio method owing to its self-learning and adaptation capability. Following this approach, incipient fault detection in power transformers using ANN can be reduced to an association process of inputs (patterns of gases concentration) and outputs (fault type). Back-propagation Neural Network (BPNN) is one of most widely used neural network model, but the slow convergence and being susceptible to network expansion restrain its future application. Another problem of BPNN is the determination of the nodes of hidden layer. Probabilistic Neural Network (PNN) has a fast training process to adjust weights for network training without any iterative or recursive computations. The decision boundaries and the training set can be easily modified using new data as it becomes available. The two different diagnosis models were established by Neural Network Tools Box available in MATLAB. The performance of two models was discussed in this paper. It can be concluded from the simulation results that PNN model is more suitable for the on-line monitoring and diagnosis for power transformer due to its characteristics such as simple network structure and being easily implemented for industrial application.Fuzzy neural network is organic integration of neural network and fuzzy logic, it possesses the capability to deal with uncertain information and storage knowledge. As a result, it is an unique approach for pattern recognition problems. A diagnosis model based on TNFIN(Tsukamoto-Type Neural Fuzzy Inference Network) that is a typical network of FNN is proposed in this paper. Firstly the network architecture is decided according to the input-output pattern, then network parameters can be updated according to the hybrid algorithm that integrates gradient decent algorithm and LMS algorithm on the basis of supervised learning and error minimization. The results of verification show that it improves the shortcomings of BPNN such as slow convergence and local minima. It is feasible to apply FNN in the diagnosis of power transformer.On the basis of information fusion, a synthetic diagnosis model for power transformer is proposed based on TNFIN and evidential reasoning. The model is modularized into three parts that are functioned as qualitative analysis, preliminary location and secondary location respectively. Firstly it takes advantage of the virtue of TNFIN such as fast convergence and fixed network structure to determine the quality of a certain fault. Afterward, the specific location is ascertained using evidential reasoning owing to its excellent capability to deal with the uncertain knowledge. Consequently, the diagnosis model can provide some intuitionistic and comprehensive referenced information to support the establishment of inspection strategy. Results of diagnostic example indicate that the presented model can improve the reliability and accuracy of fault diagnosis.