Low Complexity Classification of Power Asset Faults for Real Time IoT-based Diagnostics

Abstract

This paper investigates a new application of Capsule Neural Network (CapsNet), in combination with Constant-Q Transform (CQT), for insulation fault signal detection in High Voltage (HV) power plants. First, a mapping from insulation fault time-series signals to time-frequency images is obtained using the CQT, providing both time and frequency information. Different ways of exploiting the resulting complex-valued CQT are proposed; the CQT magnitude as a 1-channel image and the real-imaginary values of the CQT as a 2-channel image. This paper presents novel work in HV condition monitoring by utilising the CQT and CapsNet methods. Feature extraction and classification, from the produced CQT spectrum, is performed by CapsNet and the Residual Neural Network (ResNet). A performance comparison between both models, shows that CapsNet outperforms the ResNet in terms of classification accuracy with lower computation. The reduced computation and improved classification accuracy proves ideal, for system implementation on an edge embedded device incorporated in an Internet of Things (IoT) arrangement.