Benchmark of automated machine learning with state-of-the-art image segmentation algorithms for tool condition monitoring

In condition monitoring of cutting inserts for machine tools, vision-based solutions enable detailed wear pattern analysis. Besides the main failure modes of flank wear and tool breakage, other defects, such as chipping, grooves, and build-up-edges, can be detected and quantified. However, manual analysis of the images is time consuming and traditional machine vision systems have limited capabilities adapting to changing situations, such as different insert types. As a result, robust machine learning techniques are researched to support the process of classifying images and detecting defects through image segmentation. For the latter, a variety of highly optimized networks exists. Still, these networks require tuning by machine learning experts. In contrast, automated machine learning is a recent trend that greatly reduces these efforts through automated network selection and hyperparameter optimization. In this study, automated machine learning is compared with manually trained segmentation networks on the example of tool condition monitoring. To achieve this, a heterogeneous dataset of over 200 industrial cutting tool images is recorded and evaluated. Comparing the manually trained segmentation networks to the automated machine learning framework, it is determined that the automated machine learning solution is easier to handle, faster to train and achieves better accuracies than other approaches.