Discrimination of structures in plant using deep learning models trained by 3D CAD semantics

This paper describes a 3D point cloud segmentation pipeline that contributes to the efficiency of decommissioning works at the Fukushima Daiichi Nuclear Power Station. For decommissioning works, simulations and calculations for preliminary work planning using 3D structural models are crucial from a safety and efficiency viewpoint. However, 3D modeling works typically require high costs. Therefore, we aim to improve the efficiency of 3D modeling by segmenting geometric shape regions into categories in a 3D point cloud state using deep learning. Our pipeline uses 3D computer-aided design semantics to create a training dataset that reduces annotation costs and helps learn human knowledge. Performance evaluation results show that the discriminator can discriminate major structural categories with high accuracy using deep learning models. However, we confirm that even the state-of-the-art model has limitations in discriminating structures containing similar shapes between categories and structures in categories with a small number of training data. In the analysis of evaluation results, we discuss challenges encountered by our pipeline for practical applications.