Narrow Band Frequency Response Analysis of Power Transformers with Deep Learning

Abstract

Frequency response analysis (FRA) is a standard technique for monitoring the integrity of the mechanical structure of power transformer windings. To date, however, there remains no suitable method for online testing using this technique. One of the main issues that persists is that any hardware designed to measure the frequencies in the range of interest would filter out frequency bands used for assessment by humans. The growth of pattern recognition capabilities in deep learning networks, however, now offers the possibility of detecting different types of faults in a narrow frequency band, which is simply not possible for human experts. This paper explores the ability of a selection of typical networks to classify common faults within different bands. The results show that networks are able to identify faults in bands where humans are unable to find them, which has implications for signal processing and electronics design in developing a system for online monitoring.