Temporal Modeling of Point-cloud Evolution for Predictive Structural Assessments

Recent advances have enabled the use of modern remote sensing technologies such as 3D laser scanning and photogrammetry for the long-term health monitoring of structural systems. These new and emerging vision-based technologies have provided an opportunity to create high-resolution 3D point clouds that capture the insitu geometry of structures in a contactless and nondestructive manner, and offer a fundamentally different source of information than conventional sensor systems. However, the complex nature of point cloud data means that there is a need for new methods of analyzing and leveraging these unstructured data sets. Previous work by the authors has shown that it is possible to quantify geometric and colorimetric structural changes in point clouds, such as those caused by plastic deformations or the spread of corrosion, by quantifying spatial and color differentials on a pointwise basis. In this paper, the authors extend these efforts to longer-term life-cycle modeling by characterizing these differentials as stochastic processes. Point cloud differentials, both geometric and colorimetric, are parameterized via convex hull computations and then fit to vector autoregressive time-series models. These parameterized models can then be used to forecast future condition states of a structure. The results of experimental validation of this process on both synthetic and laboratory specimens is presented as well. The presented experiments illustrate the robustness of the autoregressive approach to varying process models, as well as to measurement noise. The flexibility of the approach with respect to differences in structural materials and flaw topologies is studied as well. The results indicate that the accuracy of this overall approach to point cloud analytics is dependent on the quality of the generated point cloud data, as expected. However, it is also dependent on an understanding of stochastic life-cycle dynamics, most notably if the underlying stochastic process strongly deviates from gaussian assumptions of normality, and highlights avenues for future work.