Research On Error Traceability Detection Method Of Industrial Three-Dimensional Drawing Files Based On Machine Learning

NX is one of the commonly used software for drawing three-dimensional graphics in industrial design,its main functions include industrial design,product design,simulation,product optimization,ect,and a very important step is to detect for the completed industrial drawing files.The existing detection tool of the NX platform is the Check-Mate,which uses some authentication rules to complete the detection,thesis will use machine learning to conduct error detection of the three-dimensional drawing files to make the detection function more intelligent.But before that,the three-dimensional drawing file needs to be processed first,so the main work content of thesis can be summarized from the two directions of data and model:(1)In terms of data processing: Through the analysis of the original file,it is found that the information contained in the file can be represented by nodes,and there is a connection between nodes,therefore,thesis digitizes the three-dimensional drawing file into graph structure data.The operation of the entire data set is based on the xml format file,by learning the language features and syntax requirements of the xml format file,thesis extracts the relationship between nodes and the information contained in node,so as to construct the graph structure data and its feature attributes as the input of the data set of thesis,that is the input of the model.(2)In terms of models and experiments: The detection objective of thesis is to detect the error nodes in the three-dimensional drawing file,so the nodes are classified,and the wrong nodes are regarded as negative samples,conversely,as positive samples,therefore,the experimental task is regarded as the classification task in machine learning in thesis.In thesis,the multi-layer linear graph convolutional network,the attention-based graph convolutional network,and the sampling and aggregation-based graph convolutional network are used to test the error node detection task.Aiming at the idea of multi-layer linear convolution of the first model,the model training is carried out from the selection of network layers,and the networks of 2,3,4 and 5 layers are selected respectively for the experiment,it is considered that the effect of the model is relatively better when the number of network layers is 4.For the second model,a network with 3 multi-head attention is selected for training.Aiming at the aggregation function method proposed in the third model,experiments are carried out from four ways of mean aggregation,pool aggregation,LSTM aggregation and gcn aggregation,after comprehensive analysis,it is considered that the gcn aggregation function model is relatively better.After comparative analysis,different networks have their own advantages and disadvantages.To sum up,thesis constructs a data set that suitable for graph convolutional network,and trains on three different network models to verify the feasibility of graph convolutional network in industrial three-dimensional drawing file detection.