Computer Vision-Based Quantification of Wear on CNN-Generated Masks of Gear Cutting Tools

Gears are usually soft machined using manufacturing processes involving a geometrically defined cutting edge. The aim is to balance workpiece quality and manufacturing costs by controlling tool wear. Although optical systems can capture wear with high precision, the subsequent quantification process often relies on subjective human interpretation of the images, introducing uncertainty. In this paper, a post-processing algorithm for segmentation masks generated by a convolutional neural network to quantify wear on gear cutting tools is presented. For the first time, an algorithm has been developed for three common gear soft machining processes: gear hobbing, bevel gear cutting, and gear skiving. Two approaches using different clustering, filtering, and thresholding techniques were compared to identify a suitable post-processing algorithm based on their ability to accurately detect wear contours. To assess the algorithm’s performance manual and algorithm generated measurements are compared. For gear hobbing the absolute mean difference between manually measured and algorithm generated maximum wear width is ΔVB max = 26.6 µm with a standard deviation of σ = 16.26 µm, for gear skiving ΔVB max = 37.06 µm with a standard deviation of σ = 47.83 µm and for bevel gear cutting ΔVB max = 41.75 µm with a standard deviation of σ = 37.49 µm.