Tool Wear Prediction Based on Adaptive Feature-Temporal Weighted Method and Long Short-term Memory Model

Tool wear prediction constitutes a critical enabler for intelligent manufacturing systems, providing scientific decision support for proactive maintenance scheduling. Tool wear prediction is a critical challenge in manufacturing, as accurate predictions can optimize production processes, reduce downtime, and lower maintenance costs. However, existing methods face significant limitations: traditional machine learning approaches rely heavily on complex feature engineering, requiring extensive domain expertise, which limits their generalizability and scalability. Additionally, conventional Long Short-term Memory (LSTM) models, while effective for structured data on tool machining, struggle to adequately assign importance to features, leading to suboptimal prediction accuracy. These gaps highlight the need for a more robust and efficient approach to tool wear prediction. This study develops an adaptive LSTM framework with dual attention mechanisms to automate tool wear prediction in smart manufacturing. The method processes vibration signals through four stages—signal division, multi-domain feature extraction, adaptive feature-temporal weighting, and wear quantification. The proposed method is experimentally validated, and the results demonstrate superior performance with lower errors than conventional LSTM models, achieving a reduction of 12.27%, 34.31%, and 40.21% in RMSE on C1, C4, and C6, respectively, and 20.48%, 37.71%, and 39.12% in MAE. The developed model helps producers determine the precise timing of tool changes, reducing the risk of downtime due to tool failure and thus increasing productivity.