Digital image processing with deep learning for automated cutting tool wear detection

Abstract

Tool wear is a cost driver in the metal cutting industry. Besides costs for the cutting tools themselves, further costs appear - equipment downtime for tool changes, reworking of damaged surfaces, scrap parts or damages to the machine tool itself in the worst case. Consequently, tools need to be exchanged on a regular basis or at a defined tool wear state. In order to detect and monitor the tool wear state different approaches are possible. In this publication, a deep learning approach for image processing is investigated in order to quantify the tool wear state. In a first step, a Convolutional Neural Networks (CNN) is trained for cutting tool type classification. This works well with an accuracy of 95.6% on the test dataset. Finally, a Fully Convolutional Network (FCN) for semantic segmentation is trained on individual tool type datasets (ball end mill, end mill, drills and inserts) and a mixed dataset to detect worn areas on the microscopic tool images. The accuracy metric for this kind of task, Intersect over Union (IoU), is around 0.7 for all networks on the test dataset. This paper contributes to the perspective of a fully automated cutting tool wear analysis method using machine tool integrated microscopes in the scientific and industrial environment.