1In the product design stage, a large number of 3D CAD models are accumulated. This has led enterprises to pay increasing attention to the effective management and reuse of model resources in order to improve design efficiency and shorten the product life cycle. However, in the process of resource reuse, low efficiency is caused by the complexity of the assembly model structure, the designer's personal knowledge or inexperience, or unclear model function annotation. In addition, the high cost, the strong subjective nature, and the incompleteness of manual annotation affect the accuracy of model labeling, which decreases retrieval efficiency. To solve this problem, the automatic labeling of the model function has been a concern.The automated annotation process uses the labeled models in a library to obtain the semantic information of an unknown model through a similarity comparison. Zhang et al(2001) have proposed the active learning mechanism, which generates the semantic annotation of the model according to the geometric similarity with other models. Similar methods include the support vector machine (Ip and Regli,2005;Wang et al.,2011) and the Bayesian recursive learning mechanism (Barutcuoglu and DeCoro,2006). Leifman et al (2005) used relevant feedback technology to record user intention during the model retrieval process, which alleviates the gap between geometry and semantics of the model to a certain extent. Zhang et al (2010) and Li et al (2013) presented the automatic labeling method, which uses the multiscale feature extraction method based on an attributed adjacency graph (AAG) and the local shape distribution histogram to obtain the shape information of the CAD model and calculate the shape similarity between different models. According to the known semantic classification information of the model in a sample database, a probabilistic annotation framework is constructed to make semantic annotation of the CAD model in order to establish a relationship AbstractIncreasing attention has been paid to the effective management and reuse of CAD model resources. Aiming at the problems of low efficiency of model reuse and poor accuracy in the function labeling process of assembly models, this paper presents an effective probability-based model labeling strategy for complex assemblies. It is a proper method to realize automatic labeling of assembly functional semantics through active learning. Different from part model retrieval, an assembly model is described through graph theory and a bag-of-relationships model. The assembly relationships of assembly model and the information of key functional parts are considered synthetically. Then a two-tiered model retrieval mechanism is constructed to reduce the computation time cost and improve retrieval efficiency. Further, the concept of functional ontology is introduced to establish the normalized expression of the key functional semantics of the assembly model. The functional semantic annotation of the key parts of the CAD assembly model is carried out through a prob...