Dynamic monitoring of tool wear in turning using a video-based artificial intelligence approach

It has been demonstrated many times in the literature that images of turned surfaces carry traces of wear and can be used for image-based wear monitoring; however, most studies rely on static photographs, and acquiring comparable frames during actual cutting requires high-speed imaging hardware. This entails cost and integration challenges. In this study, the monitoring of tool wear from ordinary video frames is targeted. Since standard video frames are blurrier and contain less detail than photographs, designing surface-specific features is important. Accordingly, two features that capture wear-induced changes in the surface waviness profile (Sine ERMS, Gauss ERMS) were developed and, together with the mean Fourier amplitude, formed the FSG feature set. Features extracted from video frames were evaluated with a linear SVM, and for Video–FSG the classification accuracy and F1 score were, respectively, Acc≈0.65 and F1≈0.31. Applying the same approach to photographs yielded Acc≈0.71 and F1≈0.51 for Photo–FSG. Although photographs provide higher monitorability, video frames also offered meaningful discriminative power. For comparison, three commonly used, general-purpose GLCM features (contrast, homogeneity, entropy) were tested on both video frames and static photographs; F1≈0.24 for Video–GLCM and F1≈0.04 for Photo–GLCM demonstrated that the proposed features outperform GLCM in both modalities. The findings provide quantitative evidence that wear cues can be monitored from ordinary video frames with suitable features.