Online Tool Condition Monitoring Using Unreliable Pseudo-Labels

Abstract

Tool condition monitoring in high-speed cutting machining is essential to ensure the machining surface accuracy requirements, improve the tool utilization and extend the machine tool life. However, it is challenging to screen and process the data of each stage of feed-path. Moreover, how to utilize the massive unlabeled data of different machining parameters in the actual machining process is an open problem. To address these challenges, this paper proposes the TCM-U2PL model, comprising a teacher model and a student model, which can adaptively extract the data of cutting stages with tool condition features and improve model performance using unlabeled data. First, the teacher model consists of two independent classifiers in a multi-branch classification model, which can adaptively extract and classify the tool condition features in the cutting stage and can label part of the unlabeled data as positive samples and negative samples. Then, the student model identifies the tool condition with high accuracy by minimizing the marginal distribution discrepancy and maximizing the conditional distribution alignment. The model was validated on the tool condition dataset, and TCM-U2PL achieved a classification accuracy of 85.7%, significantly outperforming CNN, DA-DBN, and NSVDD models.