Physics-informed BiLSTM with attention mechanism for tool wear prediction

Abstract

A physics–data fusion model for tool wear prediction, termed PI-ATBiLSTM, was proposed in the present study to address the limitations of conventional purely data-driven approaches, which do not readily incorporate physical knowledge of the cutting process and depend heavily on large quantities of labelled data. A bidirectional long short-term memory network was integrated with an attention mechanism to capture temporal dependencies and critical wear-related features from the monitoring data obtained from nine identical cutting tools. In addition, a physics-consistent loss function was introduced to embed constraints governing wear evolution into the training process, thereby improving the physical consistency of the predicted wear trajectories. The experimental results showed that, compared with the conventional data-driven model, PI-ATBiLSTM reduced the average root mean square error, mean absolute error and mean absolute percentage error by 60.92%, 60.83% and 62.50%, respectively. Robust predictive performance was maintained under different cutting conditions. The model was further validated on the publicly available PHM2010 dataset, achieving average root mean square error, mean absolute error and mean absolute percentage error values of 1.88 μm, 1.42 μm and 0.014, respectively. Overall, the proposed framework provides an accurate and robust physics–data fusion approach to tool wear prediction for intelligent manufacturing applications.