Predictive Analysis of Transformer Faults Through Vibration Signatures and One-Dimensional Convolutional Neural Networks

Vibration analysis is one of the advanced, non-destructive, and cost-efficient techniques for monitoring the operational state and structural integrity of transformers. This method has been investigated as one of the innovative approaches for transformer condition monitoring and fault prognosis in the last few decades. On the other hand, deep learning approaches have gained significant popularity in recent years for detecting faults in transformers within various scenarios. In this paper, we propose and examine the utility of one-dimensional CNN (1D-CNN) to construct a predictive model of transformer fault using vibration data emulated in the laboratory. The developed CNN model is capable of accurately predicting voltage fluctuations in transformers, including overvoltage and undervoltage, as well as detecting turn-to-turn short circuit failures at an early stage. In particular, the model that was developed for predicting transformer excitation voltage showed impressive results, with a relative absolute error (RAE) and a root relative squared error (RRSE) of 1.14% and 3.38%, respectively. Similarly, the model built to predict inter-turn short circuit faults showed a remarkable performance with an RAE of 1.03% and an RRSE of 2.91%.