Failure Prediction of Metal Oxide Arresters using Artificial Neural Networks

Abstract

This paper investigates the suitability of Artificial Neural Networks (ANN) technique to test the correlation between thermal stress and growth in varistor grains, in order to predict failure of metal oxide (MO) arresters. Input data to the ANN model consisted of 60 points of thermal stress intensity to which low-voltage varistor samples were continuously exposed to, for a constant time-period and voltage stress. Varistor grain response obtained experimentally (targets) were compared to the ANN model output. Results show a strong relationship between the two outputs. The MSE value predicted is as low as 1.157e™27 and the error histogram shows balanced error for the trained data.