| Classification techniques are widely used in condition monitoring applications. To design a condition monitoring system for the endurance test of automotive light assemblies, classification techniques using the support vector machine with Gaussian radial basis function kernel (Gaussian RBF SVM) and the corresponding parameter estimation are presented in this thesis. The features are exacted from vibration signals. They consist of time domain parameters and frequency band energy distribution calculated using wavelet packet transform. Particularly, our goal is to develop a classifier with high accuracy and good generalization ability. Based on the one-on-one strategy, a multiplex parameter estimation method is implemented to achieve minimum training error with minimum support vector count bound on leave-one-out error for each SVM. A comparative study shows that the proposed classifier outperforms other common classifiers, such as the Bayesian and neural network based classifiers in classification accuracy and generalization ability. |
