Various realization methods of machine-part classification based on deep learning

Parts classification can improve the efficacy of the manufacturing process in a computer-aided process planning system. In this study, we investigate various methodologies to assist with parts classification based on deep learning technologies, including a two-dimensional convolutional neural network (2D-CNN) trained using both picture data and CSV files; and a three-dimensional convolutional neural network (3D-CNN) trained using voxel data. Additionally, two novel methods are proposed: (1) feature recognition for the processing parts based on syntactic patterns, where their feature quantities are computed and saved to comma-separated variable (CSV) files that are subsequently employed to train the 2D-CNN model; and (2) voxelization of parts, wherein the voxel data of the parts is obtained for training the 3D-CNN model. The two methods are compared with a 2D-CNN model trained with the images of parts to classify. Results indicated that the 2D-CNN model trained with CSV data yielded the best performance and highest accuracy, followed by the 3D-CNN model, which was simpler and easier to implement and utilized better learning ability for the parts’ details. The 2D-CNN model trained with picture files evinced the lowest accuracy and a complex training network.