Automatic Recognition Of Machining Features Based On Deep Learning For Machining Surface Point Cloud Data

The computer-aided technologies such as CAD/CAPP/CAM promote the intelligentization of mechanical product design and manufacture.In order to realize the intercommunication between design information in CAD and manufacturing information needed by CAPP,and realize computer integrated manufacturing system(CIMS),machining feature recognition is one of the key technologies.The existing machining feature recognition algorithms based on rule and symbol reasoning mainly have the problems of low recognition efficiency and difficulty in recognizing intersecting features.In recent years,deep learning technology has made great achievements in the field of computer vision.It provides a new way for automatic recognition of machining features.In this paper,the definition and classification of machining features are analyzed,the defects of traditional machining feature recognition methods and the methods based on convolution neural network are studied,and a machining feature recognition method based on deep learning for machining surface point cloud data is proposed.The main work of this paper is embodied in the following aspects:(1)Analyze the definition and classification of features,and the main methods and applications of recognizing machining features based on machine learning,focusing on the convolutional neural network and the recognition algorithm based on point cloud.(2)Aiming at the problem of lack of training neural network for a large number of point cloud data sets,a method of obtaining machining feature surface sets from CAD model and sampling them into point clouds is given.At the same time,a random point cloud destruction program is developed to simulate the boundary disappearance of feature cross-training,so as to establish a cross feature point cloud data set.(3)Based on PointNet point cloud recognition framework,a convolution neural network for automatic recognition of machining features is constructed.The control variable method is used to optimize the overall structure of neural network,so that it contains as few training parameters as possible on the premise of ensuring recognition.(4)The convolutional neural network is implemented on Tensorflow2.1platform,and the training and experimental scheme is designed.The recognition accuracy of 24 types of machining features is tested,which proves that the recognition of this method is still robust even when the boundary disappears due to feature intersection.An example of machining feature recognition of a part model is given.The innovations of this paper are as follows: 1.The combination of machining feature surface and point cloud data is used to represent machining features,which can not only ensure the topology and geometric structure of machining features,but also directly apply deep learning to identify them,thus avoiding the problem of difficult identification due to the difficulty of representing the topological structure of cross features in rule-based methods.2.Aiming at the problem that deep learning lacks a large amount of training data,a uniform sampling program of point cloud and a destruction program of machining feature surface are developed.3.The neural network structure is optimized on the basis of PointNet.The method is simple and efficient,insensitive to noise and defective point cloud data,and still has good robustness to machining surface damage caused by feature intersection.