Naive Bayes Multi-Label Classification Approach for High-Voltage Condition Monitoring

This paper addresses for the first time the multi-label classification of High-Voltage (HV) discharges captured using the Electromagnetic Interference (EMI) method for HV machines. The approach involves feature extraction from EMI time signals, emitted during the discharge events, by means of 1D-Local Binary Pattern (LBP) and 1D-Histogram of Oriented Gradients (HOG) techniques. Their combination provides a feature vector that is implemented in a naive Bayes classifier designed to identify the labels of two or more discharge sources contained within a single signal. The performance of this novel approach is measured using various metrics including average precision, accuracy, specificity, hamming loss etc. Results demonstrate a successful performance that is in line with similar application to other fields such as biology and image processing. This first attempt of multi-label classification of EMI discharge sources opens a new research topic in HV condition monitoring.