A Bayesian non-parametric clustering approach for semi-supervised Structural Health Monitoring

A key challenge in Structural Health Monitoring (SHM) is the lack of availability of data from a full range of changing operational and damage conditions, with which to train an identification/classification algorithm. This paper presents a framework based on Bayesian non-parametric clustering, in particular Dirichlet Process (DP) mixture models, for performing SHM tasks in a semi-supervised manner, including an online feature extraction method. Previously, methods applied for SHM of structures in operation, such as bridges, have required at least a year's worth of data before any inferences on performance or structural condition can be made. The method introduced here avoids the need for training data to be collected before inference can begin and increases in robustness as more data are added online. The method is demonstrated on two datasets; one from a laboratory test, the other from a full scale test on civil infrastructure. Results show very good classification accuracy and the ability to incorporate information online (e.g. regarding environmental changes).