Discrimination of the Structures in Nuclear Facility by Deep Learning Based on 3D Point Cloud Data

Abstract

This paper describes a method for discriminating structures in a nuclear facility based on deep learning using three-dimensional (3D) point cloud data. To promote safe and secure decommissioning, estimating and assuming the conditions of a nuclear facility based on measured sensor data are important. Especially, data on dose rate in a workspace are useful to plan a decommissioning task, and the shape and material of structures in the workspace are required for radiation dose simulation. Shape data can be obtained using equipment such as a 3D laser scanner. However, obtaining the material data of objects is difficult. Therefore, we consider that major material can be estimated from the category of structures in a nuclear facility. In this study, we propose a structure discrimination method based on 3D semantic segmentation using 3D point cloud data comprising labeled data points by referring to the structural category labels of 3D computer-aided design data of an existing nuclear facility. We evaluated the discrimination performance of the proposed method through hold-out validation.