Intelligent Identification Of Assembly Interface Based On Improved Graph Attention Network

With the development of CAD technology,more and more product models are shared to the Internet and applied to the production process of relevant enterprises.Kinematic semantics is the relative motion betwteen two parts in the product model.And it’s an important feature of most products(especially mechanical products),and also a key content of mainstream digital product design,which reflects the working principle of product models.Because the retrieved product models usually do not have reusable kinematic semantics,reconstructing kinematic semantics has become one of the most important topics in the fields of model reuse,model retrieval,feature recognition,pattern recognition and so on.The key of kinematic semantics reconstruction is to identify the assembly interface on the part.However,there are few general methods to automatically and effectively identify the above contents,and some related work usually depends on labor,which is time-consuming and inefficient.To address the above problems,this paper proposes a kinematic semantics reconstruction method based on improved graph attention network.The main works of this paper are as follows:(1)We present a new product model descriptor,attribute adjacency graph with face structure fingerprint(FSFAAG).Compared with the existing attribute adjacency graph that only retains the content attributes,the descriptor proposed in this paper adds the face structure fingerprint to the attribute set when describing the product model.This descriptor can records the local structure information of each model face,and uniformly describe each input part as well as its assembly interfaces(they can have various geometric shapes).(2)A novel dual-level anti-interference-filtering mechanism is proposed and introduced into the neural network architecture.Combining the dual-level antiinterference filtering mechanism with graph attention network,it can solve the problem of sparse distribution and small number of assembly interfaces in the product model.(3)Based on semantic consistency and attribute adjacency graph,this paper proposes an assembly-interface-faces-aggregation method.This method can quickly aggregate the adjacent-assembly-interface faces with the same type,so as to accurately identify each assembly interface of the part model.(4)To make the proposed approach more general,this paper marks and constructs a model library which have wide coverage of part models and can be rapid expanded.(5)A prototype system for reconstructing kinematic semantics is designed and implemented.Through different experiments,we verify the effectiveness of our approach in kinematic semantics reconstruction.The results of network comparison experiments show that the assembly-interface-face-recognition accuracy of our approach can reach about 92%,which is about 2%-5% higher than that of other classical graph neural networks.The ablation experimental results present that the assemblyinterface-face-recognition accuracy of the complete network model is about 3%-14%higher than that after ablation.