Feature Computing and Signal Classifications

Abstract

In this chapter, we describe the feature computing and pattern analysis methods for VAG signal classifications. The purpose of feature selection is to study the feature correlations and then exclude the redundant features before pattern classifications. The reduction of feature dimensions may avoid excessive computation expenses, such that the pattern analysis methods based on the most informative features can also achieve favorable classification results. The signal classification methods include Fisher's linear discriminant analysis, radial basis function network, Vapnik and least-squares support vector machines, Bayesian decision rule, and multiple classifier fusion systems. The text also presents the common diagnostic performance techniques such as cross-validation, confusion matrix, and receiver operating characteristic curves. Finally, we review the state-ofthe-art methods for the VAG signal classifications and compare the results reported in recent literature.