Deep Learning Framework and Performance Evaluation for Point-Cloud Time-Series Change Detection in Long-Term Monitoring of Ancient Bridges

Ancient bridges exhibit progressive deformation and localized spalling due to long-term material weathering, humidity–temperature cycles, foundation settlement, and traffic loading, making accurate structural health monitoring essential for heritage preservation. In advanced electromagnetic and intelligent sensing systems, high-precision geometric monitoring complements radar- and wave-based inspection technologies for comprehensive infrastructure assessment. Conventional geometric-differencing methods for point-cloud change detection are susceptible to accumulated registration errors, noise occlusion, and non-uniform point density during long-term monitoring, limiting reliable identification of millimeter-level deformations. This study establishes a deep learning framework for multi-temporal point-cloud change modeling to satisfy the requirements of life-cycle structural monitoring. The temporal characteristics of structural changes and the propagation mechanisms of registration errors are analyzed, and mathematical formulations for cloud-to-cloud (C2C) and cloud-to-plane (C2P) change metrics are presented. The proposed framework provides a theoretical foundation for deep learning-based change detection and performance evaluation, while offering technical support for multimodal structural sensing and intelligent monitoring applications in engineering infrastructures.