Classification of Partial Discharge Signals Using 1D Convolutional Neural Networks

For decades, partial discharge (PD) measurement has been used as a common tool for assessing the insulation condition of high voltage (HV) systems. Background noise and interference resulting from the measurement environment and other power electronic devices in the setup make PD diagnosis challenging and more difficult. Signal processing tools employed for PD classification usually require a significant effort and expertise to extract semi-automated features from the time domain PD signals. The performance of a PD detection system depends heavily on the quality of these features. With the emergence of new technologies, wherein the interference pulses become more similar to PD pulses, automatic feature extraction has become a necessary prerequisite to have a reliable PD detection system. Therefore, the implementation of techniques based on deep neural networks that enable automated feature extraction and classification is needed. In this paper, a one dimensional convolutional neural network has been designed that takes a set of time series waveforms as the input and is capable of classifying PD sources in the presence of additive Gaussian noise and discrete spectral interference.