3D CAD Model Classification Based On Deep Learning Algorithm

3D CAD technology has become the indispensable and auxiliary means for the manufacturing enterprises in the product design and production. The statistics and research analysis show that in the process of new product development:only 20%designs are totally new, while 40% designs are created by reuse of existing product designs directly, and 40% designs are created by partially modifying existing product designs. Therefore, correct classification, retrieval and reusing existing 3D CAD model can lower design time efficiently and cut down design costs for products.However, CAD model is traditionally classified by engineers, which is a waste of time and energy, more importantly, errors occurred frequently in the tedious artificial operation. So it is necessary to develop a kind of CAD model classification method-which is more intelligent and more automatic. Based on excellent classification effect through deep learning in image, natural language and other fields, this subject intends to introduce deep learning technology to 3D CAD model classification and realize CAD model classification without supervision.In order to realize an effective combination of 3D model classification and deep learning algorithms; this article includes these two parts:·Due to the complexity of 3D models; they cannot be used as input of deep learning algorithm directly. Therefore, it is necessary to get a proper descriptor to do the preprocessing for 3D model. Through comparison and research on 3D model representation, this paper chooses the shaped D2 descriptor to express given 3D models. This descriptor calculated sampling functions through the distance between two arbitrary points on 3D surface; which extracts and translates the characteristic of an arbitrary and degenerated 3D model to the comparatively easy shaped probability distribution Meahwhile,D2 descriptor owns translation, rotation, scaling invariance as well as robustness for 3D CAD.·Through comparing the different frames of present deep learning algorithms, understanding the training methods and characteristics of multiple frames, this article chooses an appropriate architecture of Deep Belief Network which is stacked by multiple RBMs(Restricted Boltzmann Machine) to classify 3D model study training. This hierarchical structure can extract the 3D model data from low dimension to high dimension in characteristics, better explained model data.Based on the research above, this article basically achieves deep learning algorithms and 3D CAD model classification combinations; testing for this algorithm by 3D CAD model instances,it verifies the effectiveness of algorithms.