A Hybrid Signal Processing and Deep Learning Framework for Accurate Transformer Fault Identification

Accurate discrimination among magnetizing inrush currents and internal fault currents is still a major problem in power transformer protection. Conventional time domain analysis gives little or no insight into transient characteristics and stimulates advanced signal processing techniques. This paper presents a new hybrid approach by combining the Wavelet Transform and the Curvelet Transform to exploit feature extraction. The Wavelet Transform is effective in picking up localized time frequency information while the Curvelet Transform provides optimal sparse representation of directional discontinuities. Classification is further enhanced for accuracy by submitting the features to advanced deep learning methods, i.e. Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks. The resultant Wavelet-Curvelet-DL method is found to be more robust, computationally efficient, and of higher accuracy in distinguishing magnetizing inrush from internal fault currents. Simulation results have also proved that the methodology allows for fast and reliable fault detection and this is a step forward in the protection schemes required in modern smart grids.