An application of SVM, RBF and MLP with ARD on bushings

This paper examines classification models using three classes of artificial neural networks (ANN). The first ANN uses support vector machine activation functions. The second uses multiple-layered perceptron (MLP) activation functions with automatic relevance detection (ARD), and the third uses radial basis activation functions (RBF). In this work the decision is taken to remove or leave a bushing in service based on analysis of bushing parameters using RBF, SVM and MLP. The work finds that the RBF converges to a solution faster than both SVM and MLP. The MLP is the best tool of the three for analyzing large amounts of non-parametric non-linear data. MLP is the most accurate of the three networks. ARD reveals that methane was the most common cause for action on bushings tested using DCA during the two years evaluation period