Detection of Transformer Inter-Turn Faults using Continuous Wavelet Transform and Convolutional Neural Network

Abstract

This paper presents a technique for detecting and locating transformer inter-turn faults. The
technique uses transformer secondary voltages and currents as inputs. It employs continuous
wavelet transform (CWT) for input data processing and a trained convolutional neural network
(CNN) as a decision tool. The processing of input data using the CWT results in six scalogram
images. The six scalogram images are normalized and concatenated into a single image. The
concatenated image is then fed into a trained CNN which indicates the occurrence or otherwise
of an inter-turn fault. When an inter-turn fault is detected, the magnitude/severity of the fault
(defined by the percentage turns affected) as well as the affected phase is determined. The
technique was tested using simulations carried out on a 630kVA, 10.5kV/0.4kV, three-phase
transformer which was modelled using MATLAB-Simulink. Test results show that the technique
accurately detects and locates inter-turn faults. Furthermore, it can be integrated into existing
numerical relays, for enhanced protection of transformers.