A hierarchical classification scheme for computationally efficient damage classification

Abstract

This article presents a methodology for data mining of sensor signals in a structural health monitoring (SHM) framework for damage classification using a machine-learning-based approach called support vector machines (SVMs). A hierarchical decision tree structure is constructed for damage classification and experiments were conducted on metallic and composite test specimens with surface mounted piezoelectric transducers. Damage was induced in the specimens by fatigue, impact, and tensile loading; in addition, specimens with seeded delaminations were also considered. Data were collected from the surface mounted sensors at different severities of induced damage. A matching pursuit decomposition (MPD) algorithm was used as a feature extraction technique to preprocess the sensor data and extract the input vectors used in classification. Using this binary tree framework, the computational intensity of each successive classifier is reduced and the efficiency of the algorithm as a whole is increased. The results obtained using this classification show that this type of architecture works well for large data sets because a reduced number of comparisons are required. Due to the hierarchical set-up of the classifiers, performance of the classifier as a whole is heavily dependent on the performance of the classifier at higher levels in the classification tree.