Deep learning for fault detection in transformers using vibration data

The purpose of this paper is to evaluate the virtue of deep neural networks in detecting incipient failures of transformers, in particular windings looseness, via vibration data analysis. The transformer vibration technique is a non-invasive method to monitor winding looseness. It is based on the analysis of vibration spectra measured by sensors located on the transformer tank. In this paper, we rely on measurements that have been made in a dedicated lab under two different conditions: in presence or in absence of the clamping pressure on the windings. The analysis of data, oriented to fault detection, is performed by feedforward neural networks which, by experimental results, proved effective for a reliable prediction. Special emphasis is given to the robustness of the prediction to sensor misplacement and various techniques are carried out to evaluate and to enforce generalization to out-of-sample-data for the obtained classifier.