Machine learning driven 3D point cloud displacement reconstruction and structural safety evaluation of retaining structures

Abstract

Image-based 3D reconstruction algorithms make it possible to obtain 3D point clouds of retaining structures across different time periods. These point clouds can be used to efficiently derive the full-field displacement of the structures. However, due to environmental influences and limitations in algorithmic accuracy, the generated point clouds may contain a high density of outliers within the full-field displacement data.

To address the issues of outliers and inaccuracies in displacement measurements, this study employs the Improved-PatchmatchNet algorithm, the Improved-ICP algorithm, and the M3C2 algorithm to obtain full-field displacement. Subsequently, machine learning methods are applied to reconstruct point cloud displacement and correct the full-field displacement. Through algorithm improvement and a series of experiments, Support Vector Regression (SVR) is considered the most suitable method for reconstructing displacement data from 3D point clouds. In addition, extensive experiments were conducted to investigate the impact of different kernel functions and parameter settings on displacement accuracy after data reconstruction. The proposed method was validated using a PBGM based evaluation framework, and the results show that the error was less than 1 cm, with an average error of less than 3.27 %. In the field tests, long-term displacement monitoring conducted over 29 months successfully captured structural displacements caused by tunnel shield construction and seismic events. Finally, the safety and reliability of post-earthquake retaining structures were assessed using engineering experience method and PDEM method, revealing that the safety level of the structures is classified as level 1, with a reliability greater than 0.9663, consistent with post-earthquake survey results.