Research On Super-resolution Based Machine Vision Detection Method For Micro-milling Tool Wear State

Detection of tool wear is a key aspect of monitoring the machining process on CNC machine tools.Among the methods for detecting tool wear,the direct method is a hot topic in the research field of process monitoring due to its intuitive and accurate advantages.Currently,the tool images used in the direct method are mostly high-resolution images without noise.However,in actual machining process,the images of the tools captured are typically low-resolution images with noise.If low-resolution images are directly used for detecting tool wear,accuracy may be reduced.To reconstruct low-resolution tool images into high-resolution images and detect the wear state of the tool,a super-resolution convolutional neural network was designed based on the characteristics of the tool images.A loss function was proposed to guide the training process of the network.A tool wear detection method was proposed to extract the wear value of the tool from the tool images that have undergone super-resolution reconstruction.The main research content and innovation points are as follows:(1)Three representative super-resolution convolutional neural networks were selected,and the performance of deep learning methods on tool image super-resolution reconstruction problems was analyzed.First,the network was trained using datasets of different image types,and whether the characteristics of tool images have an impact on the training results of the network was analyzed.Second,the impact of pre-training on the network reconstruction results was analyzed due to the problem of a small dataset of tool images.Finally,two sets of experimental results were analyzed based on objective image quality evaluation indicators and subjective evaluation.(2)A tool image super-resolution convolutional neural network model based on the residual learning mechanism was constructed,taking into account the special nature of tool images and the shortcomings of existing super-resolution convolutional neural networks.This network uses a combination of local residual learning and global residual learning.In addition,a loss function was proposed that includes tool image edge loss information to address the deficiency of commonly used loss functions that can cause loss of edge information in images.The proposed network is trained by using the proposed loss function.The proposed network demonstrates superior performance on both objective and subjective evaluation metrics,as shown by the experimental results.(3)A method for detecting micro milling tool wear based on Hough transform and threshold segmentation was proposed.This method first performs super-resolution reconstruction on the images,then segments the wear area from the tool image,and calculates the wear value of the tool’s flank and radial wear value.Through experiments,it was found that this method can handle tool images at any position and rotation angle,and can accurately extract the wear value of the tool’s flank.