Fault Diagnostics of Power Transformers Using Autoencoders and Gated Recurrent Units with Applications for Sensor Failures

With the industrial internet of things and SG technology developing, more and more operation data could be accessible, and condition-based maintenance shows promise for electrical equipment. This paper aims to develop a data-driven fault diagnosis utilizing operation data for high voltage equipment condition monitoring. To understand the asset management of power transformers, an interview is conducted as the expertise input for the study. The paper uses deep learning in an unsupervised way to model normal behaviors and identify underlying operational risks. The autoencoders are used to compress the raw data and extract the key features and the gated recurrent unit to model the dependencies between normal behaviors of power transformers. Finally, the method employs control charts to generate the alarm to indicate the underlying anomalies. The paper uses an online dataset to test the applications for sensor failures. The results show that the method can identify the operational risks before sensor failures.