Can a deep learning based IoT fault diagnosis system identify more than one fault at a time?

Abstract

The experiments in this study propose a fault diagnosis method to incorporate in an internet-of-things (IoT) system for the condition monitoring of high-voltage generating stations. The approach is based on feature extraction with signal processing methods and a deep learning model to tackle fault classification in measured signals that contain one or more faults simultaneously. The proposed system implements feature extraction through the short-time Fourier transform (STFT) of 1-D electro-magnetic interference (EMI) fault signals obtained from online high-voltage (HV) assets. The produced feature maps are then used in parallel with label word embeddings to train and test a deep learning model consisting of, a graph convolutional network (GCN), implemented to learn inter-dependant fault label relationships from label co-occurrence matrices and label word embeddings, and a convolutional neural network (CNN) to extract relevant features from STFT data representations. The proposed system tackles the under-addressed EMI multi-label HV fault diagnosis problem and produces strong results in label classification even when implemented on a heavily imbalanced data set, to the author’s knowledge the system provides an unprecedented level of performance that is industrially acceptable in fault diagnosis and can be successfully implemented on a real-world IoT-based condition monitoring system. In addition, in theory the proposed system is scalable for the prediction of a higher quantity of fault labels present in data instances.