Development of Novel Convolutional Neural Networks for Detection of Damage to Nuclear Storage Containers

Lee David Taylor

doi:10.7273/000005197

Over the last decade, the use of machine-based monitoring systems has exploded in popularity due to the relatively low cost and improved accuracy over human based monitoring systems. Los Alamos National Laboratories (LANL) is investigating the potential of using machine-based tools to monitor containers which are used to store nuclear materials using multiple modes of detection. In this work, the mode of detection to be examined is a 3D convolutional neural network which will use point-cloud data as an input, captured by depth cameras. Training these networks requires a significant amount of data as examples in order to be considered robust enough for human lives to depend on them. Preexisting labeled data is not always readily available and the process of capturing data and labeling a custom dataset is time consuming and expensive. Point cloud data poses a particularly difficult problem as these are typically massive metric spaces that are not useable the typical network architecture used in 2D networks. In this research, no real data was available for the purposes of training so synthetic data would need to be used instead. Synthetic data is generated using a 3D model of the container made available by LANL. A series of iteratively more complex datasets are developed for use in training and testing in two generations of neural networks which are developed for the purposes of the detection of surface damage using the whole of a point cloud. The second-generation network shows itself capable of accurately prediction of surface damage within a pre-defined ideal distance from the camera. In addition to these networks, two auxiliary networks were created which explore a per-point classification verses per-cloud classification. The first of these auxiliary networks is a multi-label classification network which classifies features of the container and also labels locations of damage. The second network is intended to classify points in a scene as container/no container for use in the process of labeling of real data or extracting the points which constitute the container as a separate point cloud. Ultimately, the viability of synthetic data as a sole training data source is not determined with per-cloud networks in this research; however, the per-point container extraction network showed promising early result with the first real data available. This work will assist future researchers developing convolutional neural networks which use point cloud data, and best practices for using synthetic data to stand-in or augment their training data.

Development of Novel Convolutional Neural Networks for Detection of Damage to Nuclear Storage Containers

Files and links (1)

Abstract

Metrics

Details