Assessing One-vs-All 1D-CNN Classifiers for Multi-Label Classification of Partial Discharge Waveforms in 3D-Printed Dielectric Samples with Different Void Sizes

Effective insulation degradation diagnosis is essential for monitoring the reliability of any electrical system. One cause of degradation within insulation materials is the occurrence of partial discharge in voids. The severity of the degradation is related to the size of these voids inside the material. Hence, non-invasive classification of the void size could be important for cost-effective maintenance. However, multiple void sizes can exist concurrently within the insulation material which makes the problem a multi-label classification problem. In this paper, the performance of a collection of one-versus-all one-dimensional convolutional neural network (CNN) was investigated to classify different void sizes inside 3D-printed dielectric samples. Training of the CNN classification algorithm was done on single void-size samples and testing was done on single and multiple void-size samples. The CNN took a set of PD time-series waveforms as the input and investigation was carried out to assess the performance of such a system when multi-labeled signals were presented in the testing phase. In addition, the effect of the number of the classified classes on the performance of the proposed system was considered.