Health monitoring of long-span bridges using deep learning driven by sensor measured and numerical response data

Both the vibration and quasi-static load responses of cable-stayed bridges affect their long-term behaviours (eg in fatigue) and so their structural integrity. The associated modal behaviours and (owing to their statically indeterminate nature) the static response are strongly influenced by the spatial stiffness profiles of the bridges. Translation into loads and response of data from a comprehensive network of multi-sensors, shows huge potential to drive a deep learning (DL) approach which can identify these spatial stiffness profiles, and so can reveal any spatial stiffness perturbations arising from any damage states. The role of sensor-verified FE analysis is discussed in providing a means to assess likely damage states for training the DL approach to enable the early defect detection. A significant impact of data quality and sample size on the DL method is discussed in the paper. This paper compares generation of data sets, establishment of learning frameworks, and performance of each DL application. A review of existing literature in the wider field of SHM is also provided, to strengthen the case for this novel approach.