Transformer Fault Prognosis Using Deep Recurrent Neural Network Over Vibration Signals

Abstract

Vibration analysis is considered as a cost-efficient and nondestructive technique to monitor the transformer operating conditions and evaluate the transformer mechanical integrity. This method enables transformer fault prognosis before insulation catastrophic failure. In this work, the possibility of using deep neural networks in capturing the hidden patterns of vibration time series to predict the transformer underexcitation and overexcitation and interturn fault progress prediction in early stages is examined. This focus is warranted because deep learning techniques lend themselves to integrating the feature extraction into the predictive model construction stage. In this regard, deep recurrent neural network (RNN) architecture, including unidirectional and bidirectional gated recurrent units (GRUs) and long short-term memory (LSTM) models, are adopted. The constructed RNN for predicting excitation voltage exhibits a remarkable performance with a relative absolute error (RAE) of 0.56%. Predicting the interturn fault proved to be a more challenging problem and the constructed RNN for this purpose showed an RAE of 17.58%. The source code for implementing all constructed models is available at https://github.com/azollanvari/FaultPrognosisDL .