A novel and interpretable approach to data augmentation for enhanced prediction of cutting tool remaining useful life

Abstract The prediction of tool remaining useful life (RUL) is crucial for ensuring the quality and reliability of components in high-end equipment. However, in practical applications, tool wear data are typically sparse and exhibits irregular time-series patterns, which pose significant challenges to the development of accurate and reliable predictive models for RUL. To address these challenges, this paper presents a novel and interpretable data augmentation collaborative regression prediction model (DL-SVRs). DL-SVRs is a deep regression framework with a three-layer Support Vector Machine structure, which innovatively transforms the time-series regression problem into an imbalanced multi-class classification problem. Specifically, DL-SVRs employs the interpretable Improved Sample-Characteristic Oversampling Technique (ISCOTE) for data augmentation, using the first and second layers of Support Vector Classifiers, and incorporates an adaptive scaling factor to adjust the amount of the augmented data. Subsequently, DL-SVRs integrates ISCOTE with Support Vector Regressors (SVRs) in the third layer to prediction RUL by monitoring wear values. The effectiveness and superiority of the proposed method are validated through experiments on the PHM2010 milling tool dataset and a self-collected turning tool dataset TTWD. The experimental results demonstrate that, compared to renowned algorithms such as SMOTE and MWMOTE, the proposed DL-SVRs model offers distinct advantages, providing valuable insights for RUL prediction in high-end equipment.