Discrimination of the structures in nuclear facility by deep learning based on 3D point cloud data

Imabuchi, Takashi  ; Tanifuji, Yuta ; Kawabata, Kuniaki   

This paper describes a method for discrimination of the structures in nuclear power station by deep learning based on 3D point cloud data. In order to promote safe and steady decommissioning work, it is important to estimate and assume the condition in nuclear power station based on the measured sensor data. Especially, the data of the dose rate in the workspace is useful to plan the decommissioning task and, the shape and the material property of the structures in the workspace are required for the dose rate simulation. Shape data can be obtained by such as 3D Scan, however, it is difficult to acquire the material property data of the objects. Therefore, we consider that it is possible that the major material property can be estimated from the category of the structures in nuclear power station. In this paper, we proposed a structure discrimination method by 3D semantic segmentation with 3D point cloud data that consists of labeled points by referring category labels of CAD data of existing nuclear facility. We reported discrimination performance of the proposed method by hold-out validation.