Use of descriptors based on moments from digital images for tool wear monitoring

The widely used criteria to determine the need for tool replacement are too much conservative. The consequence is that tools are only used for a small fraction of their possible useful life. The economical influence of tool replacement costs over total production costs demands better criteria according to current technology. The use of different moments to describe tool wear images and to classify the tool condition in wear classes has been studied. The moments used as descriptors in this paper show different behavior with regard to wear identification, concluding that Hu and Legendre descriptors provide the best performance. These descriptors have been classified using a finite mixture mclust model considering three wear classes (low, medium and high). The achieved results from the clustering have been checked by means of discriminant analysis techniques, linear and quadratic, using the Fowlkes–Mallow index as quality factor. The projection of image data by means of linear discriminant analysis provides useful wear maps for tool monitoring. These wear maps show us the wear classes and the frontiers among them, in such a way that wear evolution for current tool can be mapped. The quadratic discriminant analysis allows us to assign to the current tool, a probability of belonging to a wear class. This probability is used as a new wear criterion in substitution of the current conservative criteria, making possible to reduce tool replacement costs.