The main purpose of this work is to propose a modern one-dimensional convolutional neural network (1 D CNN) configurations for distinguishing separate PD impulses from different types of PD sources while the parameters of these sources are changed. Three PD sources were built for signal generation: corona discharge, discharge in a void, and surface discharge. The reason for using separate PD impulses for classification is to develop a universal tool with the ability to recognize an insulation defects by analysing very few events in the insulation in a short range of time. Additionally, we found the optimal sample rates for the data acquisition for these network configurations. The necessity of signal filtering was also tested. The following configurations of a neural network were proposed: configuration for classification raw PD impulses; configuration for classification of PD impulses represented by power spectral density, for both filtered and unfiltered variants.

In this work we presented the results of the experiment in which a reference signal was used for determining degradation level of cable insulation. A set of reference signals was constructed using programing language. Onwards, each variant of the constructed signal was subsequently sent to the coaxial cable by aim of a following: digital-analog converter, radio frequency power amplifier with high linearity, and high frequency current injection air core transformer. The signals, passed through the cable, were received with a current transformer and processed using algorithm written in Python language. Processed signals were utilized for constructing data set for modern one-dimensional convolutional neural network. Neural network was defined in Keras, then optimal configuration of the network structure and its parameters were found. Also, visualization by gradient-based localization method was used to interpret the results of classification for constructed classes. Based on the classification accuracy, the most appropriate parameters of reference signal were determined.

No abstract

No abstract

This paper compares the performance of onedimensional and two-dimensional convolutional neural networks in the task of analyzing a reference signal while determining the degradation level of single-core polymerinsulated cable. In this work was designed the set of reference signals and several forms of representing of these signals in the form of one-dimensional and two-dimensional tensors. Then, an experimental determination of the most effective version of the reference signal is carried out in terms of classification accuracy and the most effective form of representation of this signal was found, as well as most efficient type of neural network.