GNN-based deep reinforcement learning for MBD product model recommendation

Abstract

ABSTRACT Digital twin is more and more widely used, and the delivery demand of digital twin is more and more prominent at the same time of product physical delivery. Research on the digital twin product model recommendation method is of great significance for the rapid construction and reuse of digital twins. The methods currently in use, however, principally concentrate on geometric reuse and pay little attention to functional or knowledge reuse. In this paper, a graph neural network (GNN)-based deep reinforcement learning (DRL) for product model recommendation is presented. First, an MBD (model-based definition)-based semantic feature attribute adjacency graph (MSFAAG) is introduced to structured MBD model as the carrier of the digital twin product model. The MSFAAG is then embedded into continuous vector spaces using a GNN to obtain the categorization of these MBD models. Finally, DRL is used to adaptively identify more important semantic features, including manufacturing semantics and functional semantics, to obtain more detailed model classification results. The experiment effectively improves the reuse efficiency of the non-geometric aspects of the digital twin product and MBD model. Compared with other traditional recommendation algorithms, the algorithm proposed in this paper has higher accuracy and can well meet the design requirements of users.