Bayesian dynamic forecasting of structural strain response using structural health monitoring data

Research on structural health monitoring (SHM) is nowadays evolving from SHM-based diagnosis towards SHM-based prognosis. The structural strain response, as a localized response, has gained growing attention for application to structural condition assessment and prognosis in that continuous strain measurement can offer information about the stress experienced by an in-service structure and is better suited to characterize local deficiency and damage of the structure than global responses. As such, accurate forecasting of the structural strain response in real time is essential for both structural condition diagnosis and prognosis. In this paper, a Bayesian modeling approach embedding model class selection is proposed for dynamic forecasting purpose, which enables the probabilistic forecasting of structural strain response and bears a strong capability of modeling the underlying non-stationary dynamic process. As opposed to the classical time series models, the proposed Bayesian dynamic linear model (BDLM) accommodates both stationary and non-stationary time series data and delineates the time-dependent structural strain response through invoking different hidden components, such as overall trend, seasonal (cyclical), and regressive components. It in turn paves an effective way for incorporating the newly observed time-variant data into the model framework for structural response prediction. By embedding a novel model class selection paradigm into the BDLM, the proposed algorithm enables simultaneous model class selection and probabilistic forecasting of strain responses in a real-time manner. The utility of the proposed approach and its forecasting accuracy are examined by using the real-world monitoring data successively collected from a three-tower cable-stayed bridge.