RUL prediction of milling tool under variable conditions based on multi-dimensionally screened features and physically modified attention network

Predicting the remaining useful life (RUL) of milling tools is crucial for maximizing tool utilization and reducing machining costs. However, accurately predicting RUL faces three main issues: large data variations across variable conditions, potential conflicts with physical laws, and estimating prediction uncertainty. To address these limitations, this paper proposes multi-dimensionally screened features and a physically modified attention network to predict the milling tool RUL under variable conditions. A multi-dimensionally screened features method is designed to obtain minimal variation across variable conditions by evaluating correlation and similarity. The screened features are integrated with physical constraints to construct a unified health indicator. Finally, the physics-based model and Wiener process jointly guide the attention network training, narrowing the parameter search space and enhancing RUL prediction accuracy. Compared to data-driven and hybrid-driven models, the proposed model achieves optimal prediction performance and quantifies prediction uncertainty under variable conditions, with the lowest MSE reaching 0.68 and MAPE within 10%, showing an average improvement of 35.08% over hybrid-driven methods.