Development of SAM-Augmented U-Net Model With Transfer Learning for Multiple Tool Wear Detection

This article presents a novel tool wear detection system that accurately identifies wear on cutting tools in manufacturing processes by combining the U-Net semantic segmentation network with the segment anything model (SAM). Traditional deep learning approaches for tool wear detection require extensive manual data annotation to prepare training data, which is both inefficient and labor intensive. The proposed method addresses these limitations by leveraging the SAM model to generate high-quality pseudo-labels for new tool data. This is achieved by comparing the segmentation masks produced by SAM with the outputs of a pretrained model and selecting the masks with the highest intersection over union (IoU) as the most realistic pseudo-labels. This self-supervised approach enables the acquisition of training data that closely resembles real annotations, without the need for manual labeling. The system employs a stitching method to create a panoramic image of the cutting tool, which is then segmented using the U-Net framework to accurately identify the cutting edges. Transfer learning techniques are applied to fine-tune the model, ensuring optimal performance across both the source and target domains. Experimental results validate the proposed approach’s effectiveness, achieving a remarkable 35% improvement in mean IoU (mIoU) on new tool types, while maintaining excellent performance on the original tools. The self-training strategy proves to be an effective solution for overcoming the limitations of traditional tool wear detection methods, especially when faced with new tool geometries. By accurately identifying cutting-edge positions and detecting tool wear, the proposed system represents a promising approach for online monitoring of tool wear in manufacturing processes, with significant potential to ensure product quality and reduce production costs.