TRANSFORMER FAULT PROGNOSIS USING VIBRATION SIGNALS AND FORECASTING TECHNIQUES

Abstract

Vibration signature analysis is considered as an advanced and economical methods to evaluate transformer operating condition and mechanical integrity. Transformer condition monitoring and fault prognosis have been investigated and discussed from the second decade of this century, while the modern and innovative approaches such Artificial Intelligence (AI), are very quickly under development in different applications and they have employed recently in this field. In this thesis, we first discuss the advantages and disadvantages of the conventional techniques along with mathematical/practical approaches that are capable to monitor transformer working condition effectively. Afterwards, analytical approach to model the transformer vibration is conducted and vibrational model of transformer is provided. Then, for fault prediction deep neural networks, namely convolutional neural network (CNN) architectures, were employed. Different experimental works to emulate various crucial faults in transformer operational condition is conducted, and recorded vibrational data will be analyzed using CNN models. The experimental results are also compared with mathematical modeling by validating the recorded data approach in this study. In this regard, two case studies were examined in the experimental laboratory. Firstly, transformer voltage excitation test was conducted to collect transformer vibration data using experimental transformer with different loads. Secondly, the transformer vibration waveforms were recorded by emulating transformer inter-turn short circuit using variable resistor. The focus of the thesis is observing the possibility of applying deep neural networks, in particular, CNNs, for time-series vibration signals to predict the transformer excitation and turn-to-turn short circuit fault. Therefore, 1D-CNN architecture was constructed by selecting the best predictive model from a prespecified space of hyperparameters. The constructed CNN model for transformer excitation voltage exhibited a remarkable performance with RRSE of 4.49% and RAE of 2.49%. At the same time, the model constructed for the inter-turn short circuit fault classification achieved a remarkable accuracy of 99.86%. Finally, the achieved results were compared with previous studies and discussed in detail.