Research On 3D CAD Model Retrieval Method For Design Reuse

Computer-Aided Design(CAD)technology has become an essential aspect in the manufacturing industry.Its excellent features such as visualization and digitalization can help engineers better understand the shape and structure of products in a more intuitive way.This is beneficial for engineers to improve the efficiency of product design and reduce development time.The industry now has access to a vast collection of 3D CAD models,thanks to the rapid development of internet technology.Retrieving suitable models and reusing designs can save costs and improve a company’s competitiveness.Deep learning has made great progress in 3D model retrieval,but 3D CAD models have their own characteristics.Applying 3D model retrieval methods directly to CAD model retrieval can lead to some problems,such as the multi view retrieval method that performs well on 3D models will lose structural information and engineering semantic information..Existing methods also do not make use of the rich design semantic information contained in the parametric design features of 3D CAD models,so they cannot reflect the design intent of engineers very well.Currently,3D CAD model retrieval methods cannot meet the needs of retrieval and design reuse in industrial manufacturing.This dissertation focuses on the above issues for research.The specific research content is as follows:(1)Aiming to address the issue of structural and engineering semantic information loss in mult view methods,we propose a 3D CAD model retrieval framework that combines multi-view and B-rep information.The structural information and engineering semantic information in the B-rep representation are complemented with the appearance information contained in the multi view.The framework is divided into three modules: Dual-branch Feature Extraction Module,Multi-head Attention Enhancement Module,and Multiple Modal Fusion Module.An improved neural network is utilized by the Dual-branch Feature Extraction Module to extract features from both modalities of the CAD model.Since multi view and B-rep representations represent the same objective object,there is a latent relationship between them.Therefore,we propose an Attention Mentoring Enhancement Module to use multi view information to enhance the B-rep information,making the B-rep contain more important features and effectively improving the discriminative ability of B-rep features.In the Multiple Modal Fusion Module,to achieve more effective fusion,a joint loss function that includes a correlation loss function is proposed.The correlation loss function is designed to minimize the distance between different modal representations in the feature space.The proposed method achieved the best retrieval performance on both the dataset constructed from mechanical parts models obtained from a real manufacturing company and the publicly available Fab Wave dataset.(2)Parametric design features contain rich design semantic information and reflect the design intent of engineers.Existing deep learning based 3D CAD model retrieval methods have not fully utilized design features for retrieval,which cannot meet the retrieval and design reuse needs in industrial manufacturing.To address this issue,this paper first generates a graph representation of the 3D CAD model,namely the design feature graph.The nodes in the design feature graph represent the design features of the CAD model.By selecting sketch parameters and design parameters from the design features,a fixed-length vector is obtained as the attribute of the nodes in the design feature graph.This approach fully utilizes the geometric properties and design semantics of the CAD model’s design features.Next,this paper constructs a graph neural network based on the design feature graph,including graph convolution,graph pooling,and residual connection structures.Considering the small size of the design feature graph data,this paper uses an aggregation method that controls the sampling ratio to update nodes and employs an edge contraction mechanism to effectively utilize the dependencies contained in the design feature graph.Meanwhile,this paper fully utilizes the features of different levels through residual connection structures.Finally,experiments was conducted on a 3D CAD model dataset containing parametric design features.According to the experimental results,the proposed method exhibits excellent retrieval performance and speed.It can effectively meet the demand for 3D CAD model retrieval based on design features in the field of industrial manufacturing.