Adaptive underwater 3D surface scanning using Bayesian optimisation: A controlled proof-of-concept for suction-caisson applications

Abstract

Suction caisson installation can be affected by soil plug hazards such as plug heave, uplift, and uneven plug geometry. Current internal monitoring methods generally provide measurements at a single location and may not resolve the full surface expression of asymmetric uplift. This paper presents a proof-of-concept adaptive 3D scanning framework for reconstructing prescribed underwater surface geometries representative of soil plug hazard modes. The method combines Bayesian optimisation for sequential scan planning, Gaussian process regression for surface reconstruction, contextual updating with elapsed time as the context variable, and a sliding-window strategy to preserve causal operation as the surface evolves during measurement acquisition. The framework is evaluated in proof-of-concept laboratory experiments designed to represent heave, uplift, and wobbly uplift at average surface speeds of 0.2, 1, and 2 mm/s using both single- and dual-scanner configurations. Across the tested conditions, the system reconstructs evolving surface geometries with centimetre-scale RMSE, captures asymmetric and migrating high-risk regions, and maintains comparable performance when the monitored area is partitioned between two scanners. The results demonstrate the feasibility of information-driven full-surface monitoring under controlled kinematic conditions, while further validation on deformable soil plugs and under offshore deployment constraints remains necessary.