Application of Pretrained Model for Zero Shot Tool Wear Monitoring with High Fidelity

To minimize CNC tool wear related non-compliance, wastage form premature tool replacement is a common practice in high precision low-volume high-mix machining environments. An in-line vision-based manufacturing process monitoring for tool wear is essential to effectively minimize such wastage. To date, a significant amount of work has been reported for off-line visual tool inspection using skillful classical image processing techniques. Lately, application-specific AI algorithms have demonstrated reliable wear detection with adequate image training. To overcome the limitations of classical and application specific AI algorithms, this application-oriented research article focuses on a fast, reliable, industry-ready, and easy-to-use in-line tool wear monitoring system. In this regard, this study investigates the capability of an open-sourced pretrained zero-shot image segmentation model for tool wear monitoring for the first time. Few versions of pre-trained deep learning Segment Anything Model (SAM) is studied for image segmentation. Then, an efficient tool wear monitoring process pipeline is established with great success for erosive tool wear in dry milling. With an automatic processing pipeline with a cycle time of only 0.572 seconds per image, a 3.81% relative error in gradual flank wear is observed. Moreover, different lighting conditions to mimic industrial environments are tested with a coefficient of variance of only 10%. In summary, by leveraging the pretrained SAM model, its zero-shot segmentation capabilities, and ease of implementation and adaptability; this method demonstrates a new approach towards manufacturing process monitoring.