| Abstract |
| Taiwan, located in
the Pacific Ring of Fire, experiences frequent earthquakes that pose
significant risks to urban buildings. Conventional post-earthquake building
damage assessment primarily relies on manual field inspections, which are
time-consuming, costly, hazardous, and often subject to human judgment. To
address these limitations and the scarcity of real-world damage data, this
study proposes an automated post-earthquake building damage classification
framework that integrates physics-based collapse simulation with deep
learning. A simulationdriven data generation workflow is developed using the
open-source 3D software Blender coupled with the bullet constraints builder
(BCB) physics engine. Based on the discrete element method (DEM), large-scale
synthetic point cloud datasets with explicit physical attributes are
generated and manually labeled in accordance with established domestic and
international standards for post-disaster assessment. For automated damage
classification, the PointNet deep learning architecture is adopted, and
K-fold cross-validation is applied to ensure robust model training and
evaluation. Experimental results show that the proposed model achieves an
average classification accuracy exceeding 95.00% on validation datasets
structurally consistent with the training data. Moreover, an accuracy of
86.67% is maintained on an independent test dataset, indicating promising
generalization capability and knowledge transfer potential. The results
demonstrate the feasibility of combining synthetic collapse simulation data
with deep learning for the automated assessment of post-earthquake building
damage. The proposed framework offers a scalable, costeffective, and
automation-oriented solution that supports rapid post-disaster decision-making
and enhances the digitalization of structural damage assessment workflows. |
|
