A cutting tool wear data-physics fusion-driven monitoring method

Abstract

Tool wear monitoring (TWM) serves as a bridge between process information and tool condition. However, existing monitoring methods encounter limitations regarding long-term monitoring performance, and a single feedforward fusion approach often fails to consistently deliver high-precision predictions and reliable maintenance s1ort. In this paper, a TWM method based on Transformer-GRU and particle filter (TG-PF) driven by data-physics is proposed. Specifically, deep-level features are extracted from multi-source signals, and feature fusion is performed to accurately capture the tool's state changes. Based on the Transformer-GRU model, the mapping relationship between the historical series and the future tool wear is established, with the wear value obtained from the Transformer-GRU model serving as the observed value. The parameters of the physical model are updated during each iteration of the particle filter to achieve the fused prediction result. The fusion method incorporates a closed-loop dynamic error correction mechanism, mitigating long-term error accumulation and reducing prediction uncertainty. The TWM experiments conducted on thin-walled and rectangular block parts demonstrate that the prediction error of the fusion method is reduced by 19.19 % and 43.36 %, respectively, when compared to the single data-driven method, and the TG-PF method outperforms 11 advanced TWM methods. The fusion method leverages historical data and incorporates a dynamic updating mechanism, serving as a crucial approach to enhancing accuracy and mitigating uncertainty in long-term TWM.