Spectral subband centroid energy vectors algorithm and artificial neural networks for acoustic emission pattern classification

Abstract

This work proposes and evaluates a methodology for monitoring and diagnosis of polymeric insulators in operation based on the parameterization of acoustic emissions (AE) created by corona and electrical surface discharges. The parameterization was performed with the use of the spectral subband centroid energy vectors (SSCEV) algorithm, which compresses the frequency spectrum and presents the results of the AE energies in several frequency bands. Thus, it was possible to calculate the dominant acoustic emission frequencies. This parameter was used as reference for an operating point of the insulators and, therefore, it was used to classify them. This classification was correlated to the classification obtained by visual inspection in the laboratory, where the insulators were divided into three distinct classes: clean, polluted and damaged. Aiming to insert an aid to the decision-making, this work still proposes the use of artificial neural networks (ANN) for pattern recognition. In this way, we performed a sensitivity analysis of the parameters that influence the SSCEV and ANN, in order to obtain the values and configurations with higher performance. The use of Levenberg-Marquardt training algorithm has proved to be more suitable, since it showed hit rates and convergence up to 97.66 percent and 70 epochs, respectively.