A multi-layer neural network approach to identification of mechanical damage in power transformer windings

Power transformers are among the most critical of assets for electric utilities, in the financial impact that their failure can bring. Asset Managers need to be able to determine the right time for replacement, refurbishment or relocation of these devices, with an increasing degree of confidence, in order to minimize the total cost of operation over the equipments’ life. This has brought a change from scheduled maintenance to condition based monitoring (CBM), where the state of the transformer is continuously monitored to evaluate its working condition. A key method of transformer CBM, which effectively detects mechanical damage to the structure of the transformer windings, is Frequency Response Analysis (FRA). FRA relies on comparison of electrical admittance signatures to determine if the winding has become deformed. One of the major problems it still faces is the interpretation of differences in the signatures. To date, experts are needed to analyse graphs, drawing from experience in order to produce educated guesses as to what the differences in admittance functions denote. However, in the recent past, there has been some headway in programming computer based solutions for the problem of interpretation. The use of Artificial Neural Networks (ANNs) has perhaps been the most promising in this respect. ANNs perform in the same way that human experts do, drawing upon experience to map a change in shape of a signature to a physical change in the winding system. However, one of the major drawbacks of these methods is the large training data-sets required for the neural network to learn. The work reported in this thesis seeks to address this problem by generating training datasets from analytical models of the transformer. Due to the large number of simulations that need to be performed a customized solution method was developed to speed up computations. A combination of back propagation and radial basis function networks were then used to classify the type, location and severity of winding movement. The results showed that the neural network approach was not only accurate but tolerant to high noise levels.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat