A Class-Aware Supervised Contrastive Graph Network for CNC machining condition monitoring with class-imbalanced data

Machining state monitoring plays a critical role in ensuring quality, productivity, and reliability in CNC manufacturing. However, real-world industrial scenarios are often characterized by severe class imbalance and limited annotated data, which significantly challenge the robustness and generalization of existing learning-based approaches. To address this issue, we propose a Class-Aware Supervised Contrastive Graph Network (CSCGN), a unified framework that rethinks imbalanced industrial signal learning from a representation-centric perspective. Instead of treating signal segments independently, the proposed approach explicitly models the intrinsic structural relationships of one-dimensional sensor signals through a dynamic graph mechanism, while incorporating class-aware contrastive learning to enhance discriminative representation under imbalance. By jointly integrating multi-view signal representations, physics-informed topological modeling, and imbalance-aware optimization, the proposed framework effectively captures both structural dependencies and minority-class characteristics in noisy industrial environments. Extensive experiments on representative CNC machining tasks, including chatter detection and tool wear monitoring, demonstrate that CSCGN consistently outperforms state-of-the-art methods across various imbalance settings. The results highlight its strong capability in minority-class recognition while maintaining overall robustness, indicating its potential for reliable deployment in real-world industrial systems.