Classification of multiple PD Sources by Signal Features and LSTM Networks

Abstract

Measurement and classification of partial discharges (PD) is an often used tool for the assessment of the insulation of high-voltage power transformers. The patterns emerging from the measurements can be used to identify the type of source which caused the PD. Multiple active PD sources can’t be identified by this method however, if their patterns overlap. Retroactive separation of these overlapping patterns is difficult. A simpler approach is using the individual PD signals. Since the signals occur at different times, they are already separated. This new approach uses single PD impulses as input data. Different features are extracted from the PD signal and these features are used for the classification. For the feature extraction statistical methods, signal form analysis and LSTM neural networks were used. Classification was implemented via the Random Forest algorithm. The focus lied on the exploration and comparison of different feature sets. To make comparison easier, only two classification methods were used. Training data consisted of PD signals from single, artificial PD sources submerged in mineral oil. This new approach can correctly classify PD sources by a single signal, so overlapping PRPD patterns have no influence on this method.