Intelligent prediction approach of spatial structure response and performance deterioration by integrating spatiotemporal deep learning and digital twins

Abstract

Global spatial structures face potential safety hazards as they age and are subjected to external influences. As traditional response and performance deterioration prediction methods have exhibited limitations when dealing with long time-series data and spatially complex features, this study proposed a novel approach integrating spatiotemporal deep learning (SDL) and digital twins (DT) to address this challenge. An SDL framework driven by component strain data from physical monitoring was proposed to predict the structural responses. Three time-series deep learning frameworks were proposed to predict ambient trends, which were input in a DT model to output the structural responses driven by virtual future ambient data. Then errors of both are fused through a proposed deep neural network to improve the final prediction accuracy. Finally, deterioration prediction was achieved by comparing the deviations between the final predicted performance responses and those calculated in real time from the initial structure’s DT replica, both in the same ambient conditions. The proposed approach, verified through a long-term health monitoring experiment of a scaled cable dome structure, demonstrated high precision, with the maximum mean error of cable strain deterioration prediction only 10.45 με. It can provide a new intelligent solution for the safe operation and maintenance of spatial structures.