Zone-resolved optical wear assessment with band-position referencing for condition-aware metal band sawing

Abstract In band sawing, monitoring signals are commonly averaged over the whole band and therefore cannot localise damaged teeth. This study presents an in-process optical pipeline that images moving teeth, segments the visible wear region with a compact grayscale U-Net, condenses each mask into three lateral cutting-edge zones and assigns detections to a weld-seam-referenced band coordinate. Four findings generalise beyond the present setup. The choice of input modality matters more than network size: an otherwise identical colour model collapses under cutting fluid and reflections, because the colour channels encode the appearance of the process rather than the wear; classical thresholding is not a viable alternative either. Wear extent and its lateral distribution behave differently and should be kept as separate descriptors, since the summed wear land width grew on every tooth while concentration and left–right imbalance stayed tooth-specific. The uncertainty of the position assignment is measurable rather than assumed, and because it spans several teeth, the tooth group is the smallest unit such a map can reliably resolve. Finally, values from single frames scatter too strongly for an isolated verdict, which makes consolidation over repeated revolutions part of the method rather than an optional refinement. Segmentation reached a mean Dice similarity coefficient of 0.817 on moving-band images at about 3 ms per image, giving spatially resolved, uncertainty-aware wear information for targeted inspection.