A Comparative Study of Different Techniques for UHF Signal Classification of Partial Discharges

In this article, a comparison of different machine learning techniques for the classification of UHF partial discharge signals is presented, namely: Artificial Neural Network, Support Vector Machine, and Fuzzy Logic. The database used consisted of signals obtained in the laboratory through a printed monopole antenna with a bandwidth between 319 MHz and 1.6 GHz, which captured signals from three test objects: oil cell, potential transformer, and ceramic insulator. After creating the database, various features were extracted from the signals, including statistical parameters of the signal in the time domain and its normalized autocorrelation function, as well as characteristics of the signals in the frequency domain. Subsequently, the features that best characterized the signals were selected using the Relief algorithm and then applied to the classification algorithms. From the obtained results, it was noted that the Support Vector Machine achieved the highest accuracy in classifying the defects, reaching 98.57%, followed by Artificial Neural Networks and Fuzzy Logic, with 97% and 83.33%, respectively.