Estimation of Missing Stress Data for Structural Health Monitoring Based on Multitask Learning and Multisource Information

Strain monitoring is of importance for structural health monitoring (SHM). The stress responses obtained by strain sensors play an important role in structural damage detection and health status assessment. However, the strain monitoring sometimes suffers from data loss due to sensor fault, power failure, transmission failure, and so on, which inevitably create barriers to subsequent data analysis and structural condition diagnosis. Although there are some data recovery approaches based on the correlations between stress data from strain sensors, they struggle to perform the expected data reconstruction when simultaneous data loss occurs in all strain sensors. To this end, a novel estimation approach based on multitask learning and deep learning is presented in this paper to address the data loss arising from simultaneous failures of all strain sensors. A data-driven estimation model is developed based on the presented approach for characterizing the potential relationships between the stress responses and wind speed, wind direction, and structural temperature information from a SHM system. The effects of network structure and hyperparameters on the performance of the estimation model are investigated, and the optimal estimation model is obtained by optimizing these parameters. The feasibility and accuracy of the estimation approach are verified by utilizing the monitoring data collected by the SHM system on a 600 m high supertall building. The results show that the proposed approach performs satisfactorily in the stress response estimation, demonstrating that the estimation approach is an effective tool for recovering the lost stress data in SHM systems for supertall buildings. Notably, this paper aims to offer a new way for addressing the data loss caused by simultaneous failures of all strain sensors in SHM and field measurements.