Reducing label dependence in vibration-based drill-bit condition monitoring with masked feature pretraining

Reliable tool condition monitoring (TCM) plays a critical role in precision machining, where progressive wear can lead to dimensional inaccuracies, degraded surface finish, and unplanned downtime. Despite advances in data-driven diagnostics, most machine-learning solutions remain constrained by their reliance on extensive labelled datasets, which poses a major barrier to industrial adoption. To address this limitation, this work introduces a Self-Supervised Masked-Feature Pretraining (SSL-MFP) framework that learns latent vibration representations by reconstructing partially masked time-frequency features, thereby eliminating the need for class labels during the initial learning stage. The pretrained encoder is subsequently fine-tuned using only a small subset of the labelled dataset for downstream drill-wear classification, markedly reducing annotation demands. The framework is evaluated on a fused vibration-feature dataset and benchmarked against established supervised baselines spanning machine-learning and deep-learning architectures. Results indicate that the proposed approach achieves classification accuracy comparable to that of fully supervised models while utilizing significantly fewer labelled samples, demonstrating effective generalization under limited annotation conditions. Furthermore, the learned feature manifold exhibits distinct class separability, evidencing the representational strength of the self-supervised encoder. Overall, the SSL-MFP paradigm provides a data-efficient foundation for TCM, enabling industrial deployment where labelling costs and adaptation are critical challenges.