Quality‐Based Tool Remaining Useful Life Prediction Under Time‐Varying Processing Parameters by Windowed Sparse‐Attention Transformer

ABSTRACT Accurate remaining useful life (RUL) prediction for manufacturing tools under time‐varying processing parameters (TVPPs) is essential for optimizing production schedules. However, during the planning phase, conventional time‐series models often suffer from limited training data and a lack of quantitative physical interpretability. To address these challenges, this study proposes a hybrid physics‐based and data‐driven structure that enables quantitative physical information in tool degradation prediction. First, physics‐based models derived from fundamental processing and degradation mechanisms are employed to generate foundational data. A novel Windowed Sparse‐attention Transformer (WSAT) model is then introduced to forecast tool degradation trends under TVPPs. By explicitly embedding degradation mechanisms and aligning its architecture with degradation features, the WSAT model achieves reliable predictions even with limited training datasets acquired under time‐constant parameters. Crucially, quantitative physical insights are extracted by mapping the data‐driven predictions under TVPPs back into the physical domain. Finally, unlike conventional approaches relying on a scalar degradation metric, the proposed framework determines RUL by explicitly analyzing the impact of degradation on processing reliability and product quality. This approach presents accurate and physically rational results while ensuring rapid responsiveness to varying quality requirements. Comparative studies on a rolling process case demonstrate that the proposed framework achieves superior prediction accuracy under limited data conditions.