A damage-integrated image-point cloud registration network for characterizing macro-meso hysteretic features of semi-rigid precast concrete beam-column joints under strong earthquakes

Semi-rigid precast concrete (PC) connections can offer good design adaptability to provide desired strength, stiffness, and energy dissipation. When subjected to earthquake-excited cyclic loading, they display prominent non-integral global deformation, complex local multi-scale surface features, and temporal multi-valued attributes. To overcome the challenge of establishing the mapping relation between surface features and internal damage states in semi-rigid PC joints, we propose HysD-I2PC, a novel registration network. This system integrates three key components: an image-point cloud registration module, a cross-dimensional fusion mechanism, and a damage prediction module, enabling end-to-end pose estimation and temporal damage prediction. A comprehensive dataset consisting of 2016 samples from four tested semi-rigid PC joints and 92 simulated ones, as well as the public KITTI, is used for the comparative analysis with DeepI2P, CorrI2P, VP2P, and CoFil2P. The outcome of this study highlights the rationality of the architecture of HysD-I2PC and its superior performance in describing the prominent macro-meso hysteresis features and predicting temporal damage of the joints under reversed cyclic loading. HysD-I2PC possesses ResNet-Transformer and Point Transformer in the image and point cloud branches, respectively, a cross-dimensional feature fusion mechanism and a damage-driven temporal prediction branch. It is proved to be optimal from the ablation experiments with a loss function reflecting comprehensively the contributions from feature similarity, differential pose, and hysteretic damage. For the generated dataset, the errors of relative rotation, relative translation, and damage prediction of the proposed registration network are only 1.07 ± 0.39°, 1.13 ± 0.72 mm, and 6.81 ± 5.40 %, respectively.