Challenges of randomness in tool wear with small samples: A physics-informed cross-domain monitoring method

Abstract

Cutting tool wear monitoring is crucial for enabling predictive maintenance in machining processes. However, uncertainties in tool degradation during small-batch personalized machining present significant challenges to achieving accurate monitoring. This study addresses the randomness of tool wear through a dual-level approach: data and model. At the data level, an empirical tool wear model is developed based on nonlinear wear mechanisms, which is integrated with a domain-discriminative generative adversarial network to construct a target domain tool wear data generation framework. At the model level, a feature extractor tailored for transfer learning is designed using nonlinear relationships inherent in tool wear mechanisms, complemented by fine-tuning the classifier with the generated target domain tool life cycle data to handle domain shifts caused by randomness. The proposed method is validated using both public datasets and workshop experiments under both fixed and variable cutting conditions. Compared with baseline models, ablation models, and several state-of-the-art data generation and transfer learning models, the proposed approach demonstrates superior adaptability and robustness in handling the randomness in tool wear, even with highly imbalanced and small datasets. The results confirm the effectiveness of the proposed method in tool wear monitoring for small-batch personalized machining processes.