Automated Structural Damage Detection Using One-Class Machine Learning

Data driven SHM methodologies take raw signals obtained from sensor networks, and process them to obtain features representative of the condition of the structure. New measurements are then compared with baselines to detect damage. Because damage-sensitive features also exhibit variation due to environmental and operational changes, these comparisons are not always straightforward and an automated, probabilistic approach is necessary, particularly for largescale sensor networks. In this paper an automated novelty detection methodology based on one-class support vector machines (OCSVM) is proposed and tested on an instrumented experimental steel frame structure. OCSVMs are an advanced machine learning method which can classify new data points based only on data from one class. This enables training of a classifier for damage detection based only on information from a baseline structure. OCSVMs can suffer from over-fitting, a problem which is usually ameliorated by cross-validation. In the absence of any data from the damaged state cross-validation is not possible. In this paper the over-fitting problem is combated by the use of three different recently proposed parameter selection heuristics. These strategies are tested for various damage scenarios of the laboratory structure and the results compared.