Classification of Tool Wear State based on Dual Attention Mechanism Network

To assure the quality of product processing, precise abrasion detections must be performed on the machine's cutting tools. Consequently, the improvement of abrasion detection is crucial for the upkeep of devices in terms of processing capacity and cutting performance. The technique of tool surface abrasion imaging is one of the detection methods. This paper proposes a deep learning and computer vision-based monitoring model for conducting abrasion monitoring over cutting tools, as conventional imaging techniques always require high precision and are criticized for a complicated calculation process and their time-consuming nature resulting from manual calibration. This method is built on the SE-ResNet50 based online abrasion state monitoring model and introduces an enhanced dual-attention mechanism to learn the dependency of pixel characteristics and the inter-correlation between channels, respectively. It is proposed that the Enhance Module Network capture the underlying information on a greater scale. To achieve the self-adaptive perception of the network weights corresponding with distinct abrasion categories, attributes are recovered from the input photos, hence eliminating the complexity and restrictions associated with manual extraction. The established abrasion status categorization method is experimentally validated on a three-axis milling machine with cemented carbide tools. The results indicated that the proposed method can classify tool wear state more accurately from the raw data collected by industrial cameras under the premise of ensuring efficiency. Its recognition accuracy is up to 96.99%, and the generalization ability can obtain good results, which provides a novel concept for tool condition monitoring in actual industrial scene.