Feature Reduction for Computationally Efficient Damage State Classification Using Binary Tree Support Vector Machines

This paper proposes a computationally efficient methodology for classifying damage in structural hotspots. Data collected from a sensor instrumented lug joint subjected to fatigue loading was preprocessed using a linear discriminant analysis (LDA) to extract features that are relevant for classification and reduce the dimensionality of the data. The data is then reduced in the feature space by analyzing the structure of the mapped clusters and removing the data points that do not affect the construction of interclass separating hyperplanes. The reduced data set is used to train a support vector machines (SVM) based classifier and the results of the classification problem are compared to those when the entire data set is used for training. To further improve the efficiency of the classification scheme, the SVM classifiers are arranged in a binary tree format to reduce the number of comparisons that are necessary. The experimental results show that the data reduction does not reduce the ability of the classifier to distinguish between classes while providing a nearly fourfold decrease in the amount of training data processed.