| The United States is suffering from an aging civil infrastructure crisis; among the solutions are research and development of sensor-based monitoring and abnormality detection technologies. This thesis contributes to this area by investigating an abnormality detection strategy that relies on a pattern classification technique known as two-class Support Vector Machine (SVM). Input for the classifier is vibration data from sensors on a structure, and the output is a classification of this data into one of two classes. For example, class-1 could indicate that a structure is healthy, and class-2 could indicate a specific type of abnormality. Multi-class classification, i.e. the ability to classify more than just two abnormalities, is achieved by creating SVM strategies comprised of a network of two-class SVMs. Each SVM is constructed using a training algorithm that requires example vibration data from both classes that the SVM is intended to classify between. This is a problem for real-world applications because example vibration data rarely exists for abnormal conditions. To solve the issue, abnormalities are simulated using finite element (FE) models. Contributions of this thesis include: a step by step guide for how SVM-based abnormality detection strategies are developed; an application of SVM to a lab structure for detecting the existence, location, and severity of abnormalities; a comparison that assesses how SVM accuracy is affected by a lab structure having more or less sensors available; an application of SVM to a real-world structure; and lastly, a by-product of the other contributions, verification that FE models can be relied upon to simulate vibration data for a structure's healthy and abnormal conditions. An additional, non-technical, contribution is found at the conclusion of the thesis where broader impacts of SVM-based abnormality detection are discussed within the context of public policy. |
