Research On Feature Representation Method For 3D CAD Model Clustering

The design of 3D CAD model has always been the most basic link for manufacturing industry to control the accuracy of products,and it is also the basic consideration for whether the products meet the production requirements.In order to reuse a large number of 3D CAD models piled up in the industrial database,effective clustering has become the focus of research in recent years,one of which,it is very important to enhance the expression ability of CAD models.In view of the expression methods of local area of 3D CAD model still have the problem of low discrimination,we proposed a local area expression method of fusing edge attribute information.This method makes statistics of the intersection edge type attribute information between face and face in the model to form feature vector,As another important attribute feature and integrated into the traditional local descriptor to make the local area of the model increased with ability of difference.In view of the fact that the whole expression method of 3D CAD model in the clustering stage,it fails to calculate the similarity according to the different importance degree of each local area in the model,we proposed a new expression method of 3D CAD model based on the weighting of local area.This method gives a large weight to some rare and distinctive local areas,and a small weight to the common local areas when the model is recombination.The experimental results show that the NMI and V-measure values of the four typical clustering algorithms are improved,and the number of correct clustering of the model is increased.In view of the problem of the loss of the key information due to the high complexity of the artificial definition of the feature rules and the tedious process by the traditional model expression method,we proposed a method based on the deep neural network to learn and express the key features of the CAD model.The method uses the convolution neural network and the Autoencoder to learn the representation of the 3D CAD model,and the training results or the middle hidden layer of eigenvector are obtained as the only representation of the model.In terms of the final clustering results,the NMI and V-measure values obtained by this method are improved compared with the baseline method,and the clustering results can meet the requirements of engineering designers.