Journal
ENGINEERING STRUCTURES
Volume 234, Issue -, Pages -Publisher
ELSEVIER SCI LTD
DOI: 10.1016/j.engstruct.2021.111970
Keywords
Deep learning; Generative adversarial network; Response reconstruction; Earthquake response; Seismic loading; Structural health monitoring
Categories
Ask authors/readers for more resources
This paper introduces a Segment based Conditional Generative Adversarial Network (SegGAN) for structural dynamic response reconstruction, showing superior accuracy and noise immunity over DenseNet and traditional CNN through numerical studies.
Reconstruction of lost responses under external loads, e.g. ambient and seismic loading conditions, is important for structural health monitoring to evaluate the safety of structures. This paper proposes a Segment based Conditional Generative Adversarial Network (SegGAN), which is a powerful deep learning model for solving pixel-to-pixel tasks, to conduct structural dynamic response reconstruction. The proposed network consists of a bottlenecked generator and a segment based discriminator. Generator features skip and dense connections to improve feature extraction, and segment based discriminator uses conditioned input to facilitate generator to learn detailed and robust features. Numerical studies on a steel frame structure model are conducted to evaluate the accuracy and noise immunity of using SegGAN for structural response reconstruction. The feasibility of using the reconstructed response for damage detection is also investigated. Numerical study on a nonlinear building model is performed to validate the capability and accuracy of using the proposed approach in nonlinear response features extraction and response reconstruction. Experimental studies on a laboratory steel frame structure are conducted to investigate the effectiveness and robustness of the proposed approach with limited training data from noisy and operational conditions, and a few sample data from earthquake loads for testing. The reconstruction results of using the proposed approach are compared with those generated by a densely connected network (DenseNet) with the same configuration as the generator and a traditional convolutional neural network (CNN). Responses from available and unavailable sensors are selected as input and output of these three networks, respectively. SegGAN outperforms the other two networks and produces outstanding reconstruction results in both time and frequency domains.
Authors
I am an author on this paper
Click your name to claim this paper and add it to your profile.
Reviews
Recommended
No Data Available