A U-net-based intelligent approach for belt morphology quantification and wear monitoring

Abstract

A novel method based on image signals is proposed to achieve the fast morphology quantification of single abrasive grain and precise wear monitoring of pyramid abrasive belt in a robotic grinding system. The Regions of Interest (ROI) of multiple abrasive grains are precisely and adaptively segmented based on the U-Net deep learning algorithm, and the Intersection over Union (IoU) of predictions reaches above 0.9. Identification of effective abrasive grains is achieved by the intensity distribution, region boundary, and nearest neighbor searching for centroids. In addition, the morphological parameters of effective abrasive grains are calculated to quantify the multiple worn forms such as continuity, non-uniform, and adhesion. The calculated dimensional parameters (area ratio and perimeter ratio) intuitively characterize the evolution of belt wear in the time domain, and the calculated shape parameters (aspect ratio and extent) greatly illustrate the effect of force and adhesion on abrasive grain wear. A comprehensive experimental study of the grinding performance shows that the material removal rate is mainly affected by the macroscopic morphological wear of the compact grains, and the machined surface quality depends more on the interaction between small abrasive agglomerates and workpiece. Furthermore, the thresholds of the mean area ratio and perimeter ratio with high independence for variable process parameters can be used as the judging criteria for determining the belt service life.