Semi-supervised multi-source meta-domain generalization method for tool wear state prediction under varying cutting conditions

Abstract

Accurate tool wear state prediction during machining is essential for lowering production costs and ensuring quality. Conventional deep learning-based methods perform excellently under constant cutting conditions with sufficient labeled data. During training, domain-adaptation-based transfer learning methods require exposure to partial target domain data to accurately predict under varying cutting conditions. However, the performance of these two methods drops considerably under unseen cutting conditions. Therefore, we propose a semi-supervised multi-source meta-domain generalization (SSM2DG) method for label scarcity training scenarios, which can generalize the deep models to predict tool wear state under unseen cutting conditions. SSM2DG adopts an episodic training strategy based on meta-learning to simulate cutting condition shifts. Then, one basic model based on multiple attention mechanisms is presented for enhancing temporal and cross-channel feature extraction. Next, the entropy-based semi-supervised learning method is incorporated into the episodic training procedure to enrich intra-domain representations using unlabeled data. Besides, class relationship alignment loss based on the distributional sliced-Wasserstein distance and memory centroid-based loss are designed to improve generalization ability. Finally, milling experiments under variable working conditions and multiple ablation studies verify the effectiveness and superiority of the proposed method.