The article presents the results concerning the use of clustering methods to identify signals of acoustic emission (AE) generated by partial discharge (PD) in oil-paper insulation. The conducted testing featured qualitative analysis of the following clustering methods: single linkage, complete linkage, average linkage, centroid linkage and Ward linkage. The purpose of the analysis was to search the tested series of AE signal measurements, deriving from three various PD forms, for elements of grouping (clusters), which are most similar to one another and maximally different than in other groups in terms of a specific feature or adopted criteria. Then, the conducted clustering was used as a basis for attempting to assess the effectiveness of identification of particular PD forms that modelled exemplary defects of the power transformer’s oil-paper insulation system. The relevant analyses and simulations were conducted using the Matlab estimation environment and the clustering procedures available in it. The conducted tests featured analyses of the results of the series of measurements of acoustic emissions generated by the basic PD forms, which were obtained in laboratory conditions using spark gap systems that modelled the defects of the power transformer’s oil-paper insulation.
A transformer is an important part of power transmission and transformation equipment. Once a fault occurs, it may cause a large-scale power outage. The safety of the transformer is related to the safe and stable operation of the power system. Aiming at the problem that the diagnosis result of transformer fault diagnosis method is not ideal and the model is unstable, a transformer fault diagnosis model based on improved particle swarm optimization online sequence extreme learning machine (IPSO-OS-ELM) algorithm is proposed. The improved particle swarmoptimization algorithm is applied to the transformer fault diagnosis model based on the OS-ELM, and the problems of randomly selecting parameters in the hidden layer of the OS-ELM and its network output not stable enough, are solved by optimization. Finally, the effectiveness of the improved fault diagnosis model in improving the accuracy is verified by simulation experiments.