Non-contact vision-based response reconstruction and reinforcement learning guided evolutionary algorithm for substructural condition assessment

Abstract

Structural health monitoring of large-span bridges and high-rise buildings is crucial for ensuring safety and serviceability. However, accurately capturing motion in these structures using consumer-grade cameras is challenging due to their limited Field of Vision (FOV). To address this issue, in this study, a novel output-only substructural condition assessment framework based on the reinforcement learning-guided evolutionary algorithm and vision-based displacement response reconstruction technique is proposed. On the one hand, displacement responses of the target substructure are extracted from the vibration video using subpixel template matching algorithm with camera pose correction, which is suitable for integration with substructure strategy to detect elemental damage. A vision-based substructural displacement response reconstruction technique is developed based on transmissibility matrix and Tikhonov regularization. The measured and reconstructed displacements are utilized to established the objective function. On the other hand, to solve the optimization-based damage identification problem, a new reinforcement learning guided evolutionary algorithm, named sparse Q-learning guided evolutionary algorithm (SQEA), is proposed. In the proposed SQEA, sparse initial population is produced by reducing the dimension of unknown parameters to be identified. Six different search strategies, including DE/rand/1, DE/rand/2, DE/best/1, DE/best/2, Jaya mutation, perturbation with the Cauchy mutation, are used to construct a search strategy pool. A representative reinforced learning algorithm, Q-Learning algorithm is introduced to adaptively select the most suitable search strategy. Experimental tests on a steel frame structure and a three-span beam structure are performed to validate the accuracy, efficiency, and robustness of the proposed approach. Results demonstrate that the damage locations and extents can be accurately identified without the measurement of input forces, expanding the application of low-cost vision-based displacement measurement in substructural condition assessment. Furthermore, the performance of the improved L-curve method over traditional L-curve and Bayesian inference regularization, the superiority of the proposed SQEA over other state-of-the-art intelligent algorithms are investigated.