Interpretable deep learning approach for tool wear monitoring in high-speed milling

Tool wear monitoring (TWM) is critical in modern high-speed milling, and an effective TWM system will improve machining precision, increase tool life and reduce production costs. As a novel data-driven approach with strong learning capability, deep learning has been introduced and studied for manufacturing process monitoring, but it is rarely applied as an independent method in practice for TWM due to the poor interpretability of the monitoring results. In this study, a multi-scale pyramid attention network (MPAN) is proposed. MPAN can not only accurately monitor tool wear based on sensory signals, but also introduce the interpretability from both the aspect of network structure design and feature extraction. With the prior knowledge of signal periodicity is introduced into the structure design, the extracted multi-scale features can cover almost all the characteristic periods. In addition, the periodicity of interest can be studied based on the attention distribution. The effectiveness and feasibility of this method are verified on high-speed milling experiments. This is the first attempt to interpret deep-learning based approach for TWM.