| In this dissertation, we introduce neural network based algorithms to classify signals with the conjunction of Wavelet Transform and Genetic Algorithm techniques. Specifically, we use these algorithms on wheel bearing fault detection and diagnosis systems. Wavelet Transforms have been widely used for pattern recognition applications. Features extracted from scales (sub-bands) produced by Wavelet Transforms are highly correlated due to the redundancy of the scales (sub-bands). These correlated features may cause the classifiers to converge very slowly and reduce the classification performance. Feature dimension reduction techniques are essential to making wavelets more powerful. In this dissertation, we introduce Genetic Algorithm to reduce the feature dimension in two steps: (1) selecting the subset of sub-bands, and (2) selecting features belonging to these sub-bands. We use both multilayer perceptron (MLP) and support vector machine (SVM) as classifiers. The original SVMs are developed for a two-class problem. To extend the SVMs to our applications, we develop a multi-SVM that is optimized by adjusting the cost-factor of each individual SVM. In addition, we provide some advanced topics that will be helpful for the future research of railroad wheel bearing condition monitoring applications. |
