Robust tool wear prediction under novel operating conditions via physics-guided unsupervised domain adaptation

Abstract

Data-driven tool wear prediction enables efficient tool utilization and prevents unexpected machining downtime, thereby improving quality and productivity. However, existing tool condition monitoring (TCM) approaches face critical limitations in real-world practice due to frequent changes in operating conditions that cause data distribution shifts (i.e., domain shifts), which make trained models obsolete. In addition, obtaining labeled data for fine-tuning under novel operating conditions is impractical because tool wear measurements require considerable costs and domain expertise. Given the aforementioned difficulties, this work proposes a deep unsupervised domain adaptation (DA) method with three novel techniques for robust tool wear prediction under novel operating conditions. First, inverse Gramian subspace matching (IGSM) is proposed to reduce distribution discrepancy between two different domains by aligning subspaces of inverse Gramians without using labeled samples. Second, a pseudo-label-based pairwise regularization (PLPR) is developed to transfer informative knowledge regarding tool wear progression to novel operating conditions. Third, a physics-guided adjustment (PGA) is applied during inference to calibrate physics-compliant predictions on unseen tool wear ranges. Comprehensive experimental results using two real-world milling datasets under various domain shift scenarios demonstrate the proposed method's efficacy in tool wear prediction under novel operating conditions. In particular, the proposed method consistently outperforms existing baselines and state-of-the-art approaches under an unsupervised DA setup, exhibiting its practical effectiveness.