Wear characterization of the cutting tool in milling processes using shape and texture descriptors

Resumo

This thesis proposes several methods to evaluate the wear level of tools in milling processes. Using shape descriptors based on contour, moments and orientations, and also texture ones, the wear regions of the cutting inserts were characterised and using different classifiers the wear was modelled. The proposed models were evaluated on classic shape datasets as well as in new insert datasets created specifically for this problem. The motivation behind our work was that optimizing tool replacement operations may produce a significant improvement in manufacturing efficiency and competitiveness. It is known that changing tools at the right moment is essential,not only because of the cost of cutting tools themselves, but also for the indirect costs due to the unproductive time needed to carry out the tool replacement. This thesis addresses the task of tool wear monitoring in edge profile milling processes using computer vision and machine learning techniques allowing to replace the tools at the optimum moment and thus saving direct and indirect costs. We have created three insert datasets to assess our solutions in similar conditions to the real environment: Insert edges dataset, Insert region dataset, and Insert High-Resolution Dataset. The general shape can be extract from the boundary pixels of an image. For this reason, our first proposal for shape description was a contour-based approach. We proposed RCPDH and CPDH36R, both based on CPDH (Contour points distribution histogram). Next, we decided to describe the images using both moments and orientations. We introduced two new methods named aZIBO and B-ORCHIZ, both of them based on ZMEG approach. They are composed for global and local shape descriptors. The combination of local and global descriptors allows us to have an idea of the general form and to access the local details of the form. In this way, we take advantage of both methods of description. B-ORCHIZ proved to be the best moment-orientation descriptor obtaining an accuracy up to 91.92% in the shape description problem. Results were promising not only for generic shape datasets but also in real applications obtaining results up to 87.02% in terms of accuracy in the tool wear monitoring problem. As B-ORCHIZ descriptor proved to be useful in the tool wear monitoring field, we explored the combination of it with our shape descriptor proposal, ShapeFeat. ShapeFeat takes into account ten different features extracted from the binary region of each image. Combining ShapeFeat with B-ORCHIZ and evaluating three types of fusion methods, we obtained results which improves the performance significantly, yielding a 91.44% of hit rate in the complete dataset evaluation using two classes and a 82.90% using three classes. Our final proposal consisted of a new online, low cost and fast approach to determine the state and categorize the wear of cutting tools used in edge profile milling processes. The method proposed is based on dividing the region of the cutting edge in patches. We presented five alternatives to carry out this division and we described each patch with different texture methods. The experiments were carried out using a Support Vector Machine (SVM) with intersection kernel. Results shown that our proposal ”Small Edge Division” (SED) outperforms the other region division proposals achieving an Fscore value of 0.9032.