A method for predicting tool remaining useful life based on a physics-informed neural network

Abstract This paper proposes a Physics-Informed Neural Network (PINN)-based method for tool remaining useful life (RUL) prediction by incorporating physical wear evolution laws into a data-driven prediction framework. The method improves prediction accuracy and enhances cross-scenario adaptability through consistency with tool wear mechanisms. First, a quantitative mapping model between tool RUL and flank wear is established based on tool wear theory, clarifying the mathematical relationship between the tool degradation state and its RUL. A PINN is then constructed by embedding the wear rate derived from the Archard wear equation as a physical constraint in the loss function. A hybrid loss function integrating data-driven and physics-informed terms is designed, and a Long Short-Term Memory (LSTM)-based time-series prediction model is developed. The method is validated using the PHM2010 dataset, including comparative tests on labeled sub-datasets D 1 , D 4 , and D 6 against multiple baseline methods, as well as generalization tests on unlabeled sub-datasets D 2 , D 3 , and D 5 . The results show that the proposed method achieves a Mean Absolute Error of 4.25 μm and a Root Mean Square Error of 4.83 μm on the test dataset, reducing MAE by 57.33% compared with Support Vector Machine and by 37.13% compared with Convolutional Neural Network. The end-of-life prediction deviation rate under the same working conditions is 2.46%, and the model shows reliable generalization ability on unlabeled data. It should be noted that, in the current implementation, the wear coefficient k in the Archard constraint is fixed to a handbook value and treated as a constant during training; online adaptive estimation of k is left to future work. These results provide a physically interpretable technical approach for intelligent tool health management and maintenance decision-making.