Interpretable tool wear monitoring: Architecture with large-scale CNN and adaptive EMD

Abstract

In manufacturing, tool wear monitoring (TWM) is crucial for ensuring product quality and processing efficiency. Numerous data-driven models based on deep learning have been developed to enhance the accuracy of TWM. However, most models necessitate traditional signal preprocessing and heavily depend on expert knowledge for parameter settings. To address these issues, this study proposes a novel TWM architecture, LCNN-EMD, which combines large-scale convolutional neural networks (LCNN) with adaptive empirical mode decomposition (EMD). By analyzing cutting process mechanisms, it is found that sensor signals can be decoupled into low-frequency noise from the machine's inherent frequencies, mid-frequency features from the tool's chatter, and high-frequency noise from the sensor's quantization. The interpretability of the LCNN-EMD architecture is rooted in its sophisticated capability to analyze and suppress specific frequency features from the input signals. The LCNN adeptly captures low- and mid-frequency features while effectively suppressing high-frequency noise. Concurrently, adaptive EMD dynamically mitigates low-frequency noise, ensuring the preservation of critical mid-frequency features. This dual mechanism not only enhances the accuracy and robustness of LCNN-EMD but also provides it with high interpretability, addressing the limitations of traditional models that depend on extensive expert knowledge and static signal preprocessing. Finally, we conduct tool life experiments and compare the LCNN-EMD with mainstream models for a comprehensive evaluation. Experimental results indicate that LCNN-EMD consistently outperforms the comparative models in terms of accuracy and stability.