Semi-supervised vibration-based structural health monitoring via deep graph learning and contrastive learning

Civil structures are vital and expensive assets that are regularly inspected and monitored, resulting in a large volume of measured data. Thus, labeling structural health monitoring-related data is a tedious, time-consuming, and tricky process. In order to alleviate the dependence on labeled data, this study investigates a semi-supervised structural damage detection approach, named semi-SDD, for evaluating structures’ health status based on vibration data from multiple sensors mounted across the structure. First, a deep graph neural network is designed to combine spatial information of sensor locations with time-varying vibration data into latent representations. Next, the latent representation is empowered via contrastive learning before going through a multiple-layer perceptron layer to identify the structure's state. The applicability and performance of the proposed framework are consistently validated through three examples, including both numerically generated data and experimentally measured data (from the literature). Furthermore, additional comparison, parametric and robustness studies are carried out to gain helpful insight into the proposed method's performance.