Using a Pattern Recognition-Based Technique to Assess the Condition of Silicone Rubber Outdoor Insulators

Abstract

Several transmission and distribution companies worldwide have started to replace their existing outdoor ceramic insulators with silicone rubber insulators. The use of silicone rubber insulators in outdoor insulators was first introduced in the market almost 30 years ago. Various studies have looked at the characteristics of this material under contaminated conditions. Despite the numerous advantages of silicone rubber insulators, they still suffer from several disadvantages, especially with the lack of sufficient and reliable tests from the manufacturers. The main disadvantage of silicone rubber insulators over ceramic ones is ageing. Ageing of silicone rubber insulators can occur due to arcing, partial discharge, weather conditions, and other factors. When silicone rubber insulators age with time, they may lose their hydrophobicity. Based on the water-filming resistance of these insulators, the hydrophobicity of their surface is manually classified into seven classes. The aim of this thesis is to develop an automatic system to classify and assess the condition of silicone rubber insulators using image processing and pattern recognition techniques. Accordingly, a database of images that represent the seven classes of surface hydrophobicity of silicone rubber will be created. In this thesis, several image processing techniques have been used to extract textural and statistical features. These methods include discrete cosine transformation, wavelet transformation, Radon transformation, contourlet transformation, and using a gray-level co-occurrence matrix. Stepwise regression was used as a dimensionality reduction technique and method of feature selection. Various classifiers were examined to evaluate the extracted features. The examined classifiers included linear, polynomial, k-nearest neighbor, and neural networks. A database of 358 gray-level 481x481 sized images was prepared to represent the seven hydrophobicity classes. An excellent recognition rate of 96.5% was achieved using fused features selected by a stepwise regression and classified by a neural network classifier. The 3.5% misclassified images were mainly due to confusions between adjacent classes that exhibited high levels of visual similarity. The system proposed by this thesis can be used to help utilities assess their silicone rubber insulators automatically and effectively.