Automated indoor construction inspection platform based on source-level non-redundant point cloud

Data redundancy and high processing costs hinder the practical application of point cloud-based construction quality inspection. This paper proposes an automated platform for indoor quality inspection utilizing machine learning to segment point cloud data at the source level with non-redundancy (SNPCD, Source-level Non-redundant Point Cloud Data). Multiscale geometric features (MGF) were developed to enhance the representation of SNPCD. A comparative study on the SNPCD dataset demonstrated that gradient boosting decision trees outperformed random forests. On this basis, the metaheuristic algorithm, Newton-Raphson-Based Optimizer (NRBO), was applied to identify optimal hyperparameters. The NRBO-MGF-CatBoost approach exhibited clear advantages over the traditional Optuna method. Furthermore, a data augmentation method tailored for SNPCD was proposed. Experimental results indicate that moderate up-sampling data augmentation can effectively improve model performance. A real-world case study validates the feasibility and accuracy of the platform proposed in this paper.