Learning Entirely Unknown Classes in Time-Series Data Using Convolutional Neural Networks for Insulation Status Assessment of Partial Discharges in Power Cable Joints

In this study, a method for supervised learning comprising a convolutional neural network (CNN) and sliding-window threshold is proposed for data with completely unknown categories. The governing principle is that the sliding-window threshold has the effect of local smoothing on the continuous prediction of CNN, and the data label can be updated after a new change point is generated based on simple rules. The partial discharge (PD) data obtained from the accelerated deterioration tests conducted on nine underground cable straight joints were used for verification. Subsequently, two different initial change points were set at 30% and 70%. The change points of the two cases were found to converge relatively closely. Furthermore, the unsupervised learning method was employed to determine a set of change points, which were similar to the results of the proposed method. Therefore, the rationality of the proposed method was verified using several methods.