【作者】 秦飞巍；

【导师】 高曙明；

【作者基本信息】 浙江大学 ， 计算机科学与技术， 2014， 博士

【摘要】 来自不同学科、不同地域的工程师被临时组织在一起,协同完成项目工作已成为当代制造企业产品开发的常见模式。不同领域的工程师们使用不同的CAD系统进行产品设计,导致企业数据库中保存了多种异构CAD产品模型。工程师在进行新产品的开发时,如果能有效地找出便于进行设计重用的现有CAD模型,可以缩短开发周期,降低开发成本,提高开发效率。然而,在异构CAD模型库中搜索出真正满足工程师需要的CAD模型非常困难。为了解决这一问题,本文研究基于语义的异构三维CAD模型检索方法,主要工作包括：(1)提出基于语义的异构CAD模型检索方法该方法基于本体对异构CAD模型进行统一描述,采用本体映射生成异构CAD模型的统一语义描述符;使用深度学习技术对CAD模型库中的模型进行自动分类,形成层次式的索引结构,并对输入查询模型采用先分类,再检索的查询方式,以压缩搜索空间,改善检索性能；利用改进的VSM方法、启发式的图匹配方法对CAD模型进行多模式的相似性度量,以满足工程师多样化的检索需求。(2)构建层次式的CAD特征本体首先分析并总结领域本体的评价准则,并以此为规范构建CAD特征本体。构建的CAD特征本体使用分层表示的结构,由B-rep本体层、公共特征本体层、系统特征本体层组成。公共特征本体层分为公共特征本体核心版和公共特征本体扩展版。公共特征本体核心版参考STEP国际标准进行构建,以便具有领域专业性和权威性,可以满足CAD领域的多种应用需求;公共特征本体扩展版在核心版的基础上进行了扩展和修改,旨在能够有效支持针对异构CAD模型的语义检索。(3)提出基于本体的异构CAD模型统一语义描述生成方法基于CAD特征本体,利用本体映射技术和语义推理技术,生成能够统一描述异构CAD模型的语义描述符,用于CAD模型的相似性度量。在进行从系统特征本体到公共特征本体的映射时,考虑CAD领域特征概念的独特内外部结构,选择并构造多个映射器,从多方面对特征概念进行相似性评价,以提高本体映射的准确率以及识别出CAD领域特征概念之间的复杂对应关系的能力。为了保证描述符语义的准确性和内容的丰富性,通过将领域知识形式化为SWRL规则并以其为基础进行语义推理,实现对统一语义描述符的完善。(4)提出基于深度学习技术的CAD模型分类和索引建立方法首先选择并提取CAD模型中有区分意义的特征,并通过预处理将其转换为高维输入向量；然后通过模拟工程师人工进行三维零件分类的思考过程,构建深层神经网络并将其作为CAD模型分类器；最后适当地选取并组合多种训练策略对深层神经网络进行训练,以提高其泛化性能。该方法将深度学习用于CAD模型自动分类,并以其为基础建立异构CAD模型库的索引结构,能够提高模型分类的准确性和模型检索的效率。(5)提出多模式的CAD模型语义相似性度量方法将VSM (Vector Space Model)引入异构三维CAD模型语义检索,利用改进的VSM方法对CAD模型进行基于向量的相似性度量,以支持快速并具有高查全率的CAD模型语义检索；利用启发式的图匹配方法对异构CAD模型进行基于特征关系图的相似性度量,以支持更准确的CAD模型语义检索。多模式的CAD模型语义相似性度量方法有利于满足工程师对模型检索在效率、准确性、查全性等方面的的不同需求。基于以上研究成果,实现了一个多模式的异构三维CAD模型语义检索原型系统OB-HCMR,通过检索实例验证了本文方法的有效性。更多还原

【Abstract】 It has become a common way in modern manufacturing enterprises to organize multi-disciplinary and geographically distributed engineers working together for product development temporally. These engineers use a variety of CAD systems to design products, thus a large number of heterogeneous CAD models are archived in enterprise databases. When developing new products, if proper CAD models can be found for design reuse, cost reduction and efficiency improvement can be achieved. However, it is difficult to search CAD models which meet engineers’ requirements from heterogeneous CAD model database. This dissertation focuses on issues about semantic based retrieval for heterogeneous3D CAD models, with its main contents being as follows:(1) A semantic based retrieval approach for heterogeneous CAD models is proposed.The proposed approach uses ontology mapping techniques to generate the uniform representation as semantic descriptors for heterogeneous CAD models; and automatically classifies CAD models to form hierarchical index structure with the aid of deep learning techniques. CAD model retrieval driven by classification is achieved, therefore the searching space is narrowed down and the retrieval performance is improved. In addition, improved VSM method and heuristic graph matching method is used to measure similarities between CAD models in a multi-mode way. Therefore various kinds of retrieval requirements from engineers are met.(2) A layered CAD feature ontology is constructed.This paper analyzes and concludes evaluation criteria of domain ontology, then constructs a CAD feature ontology for meeting the evaluation criteria. A layered representation structure is used in the constructed CAD feature ontology, which consists of B-rep layer, common feature ontology (CFO) layer and application feature ontology (AFO) layer. The CFO layer is divided into core version and extended version. The core version of CFO is constructed by referencing international standard-STEP part Ⅲ, in order to be specialized and authorized, and satisfy a variety of application requirements in CAD domain; the extended version of CFO is extended and modified specifically for heterogeneous CAD model retrieval.(3) An ontology based uniform description generation method for heterogeneous CAD models is proposed.Based on CAD feature ontology, the proposed method uses ontology mapping techniques and semantic reasoning techniques to generate the uniform semantic descriptors of heterogeneous CAD models for similarity measurement. When conducting ontology mapping from AFO to CFO, considering the specific internal and external structure of feature concepts, the proposed method chooses and combines multiple mappers, and evaluates similarities from multiple facets. Therefore the accuracy of ontology mapping is improved and the complex correspondences between feature concepts are identified. In addition, for guaranteeing the correctness and abundance of the semantic descriptors, the proposed method carries out semantic reasoning by formalizing domain knowledge into SWRL rules. Therefore contents of the uniform descriptions for heterogeneous CAD models are richer and more correct.(4) A classification and indexing method for CAD model based on deep learning technique is proposed.The proposed method first selects and extracts distinctive features from CAD models, then preprocesses them as high-dimensional input vectors for category recognition. Furthermore, by analogy with the thinking process of engineers, a deep neural network classifier for3D CAD models is constructed with the aid of deep learning techniques. To get an optimal solution, multiple strategies are appropriately chosen and applied in the training phase, which makes the constructed classifier achieve better performance. The proposed method applies deep learning techniques to automatically classify CAD models. Based on classification, a hierarchical index structure in CAD model database is formed. Therefore the classification accuracies and retrieval efficiencies of CAD models are improved.(5) A multi-mode similarity measurement method for3D CAD model is proposed. The proposed method uses improved VSM technique to measure similarities between CAD models in a vector-based way, inclining to recall and the efficiency; also uses heuristic graph match technique to measure similarities in a graph-based way, inclining to precision. The proposed method can satisfy engineers with different kinds of retrieval requirements.Based on the above studies, a prototype system named OB-HCMR is developed. The experimental results are provided to validate the main ideas of the proposed approaches.更多还原