Autonomous Post-earthquake Damage Estimation Augmented by Graphic-Based Digital Twin and Deep Learning
Advisor: Professor Billie F. Spencer
Abstract
This dissertation presents an autonomous framework for post-earthquake structural damage estimation, addressing challenges in data scarcity, subjective damage interpretation, computational demands, and limited access to damaged regions. Four key research tasks were undertaken: (1) synthetic damage simulation for data augmentation, (2) development of a bi-directional graphics-based digital twin framework for damage state estimation, (3) efficient structural response simulation with a data/physics-driven approach, and (4) development of a structural condition assessment framework enhanced by photo-realistic synthetic environment.
Task 1 developed a synthetic damage simulation method to augment limited dataset of crack detection. By augmenting a cracked bridge girder dataset, models trained on the augmented dataset outperformed those using real-world data alone, validating the effectiveness and realism of synthetic damage simulation. Task 2 extended the synthetic damage simulation approach to create highly realistic synthetic damage representations derived from finite element model (FEM) results. The development of the Bi-GBDT framework, connecting observed damage patterns with structural behavior (e.g., deformation, stress) for forward and backward predictions. Validation using a reinforced concrete shear wall experiment demonstrated its accuracy in both prediction directions. Task 3 combined physics- and data-driven methods for accurate and scalable simulations, addressing the computational challenges of scaling the simulations in Bi-GDBT framework to large and complex structures. The framework incorporated advanced neural networks (ASRNN) with conditional augmentation and attention mechanisms, achieving a balance between accuracy and efficiency in a case study of a three-story frame/shear-wall building. Task 4 developed a comprehensive and scalable framework for autonomous damage inspection of a three-story reinforced concrete shear wall building, which covers autonomous UAV-based data acquisition, deep learning-based damage segmentation, and the damage assessment framework proposed in previous tasks. By integrating realistic damage simulations, efficient structural response modeling, and bidirectional analysis of damage patterns and structural conditions, this task addressed the challenges of automatically assessing complex, large-scale structures in post-earthquake scenarios.