A Dual-Task Self-Supervised Framework for Fault Diagnosis of Rotating Machinery under Limited Annotations

Reliable fault diagnosis of rotating machinery is critical for maintaining safety and efficiency in complex industrial environments. However, conventional deep learning methods are limited by their dependence on large-scale labeled datasets, which are expensive and impractical to acquire in real-world applications. To address this limitation, we propose a dual-task self-supervised framework, CGLR-Net, which jointly models global and local features to learn robust and discriminative representations under limited annotations. The framework integrates three components: a Global Alignment Module to enforce cross-condition consistency, a Local Discrimination Module to enhance fine-grained separability, and a Channel Attention Module to emphasize critical diagnostic channels while suppressing redundancy. Experiments on the Paderborn University bearing dataset show that CGLR-Net consistently outperforms state-of-the-art baselines, achieving more than 91 % accuracy with only ten labeled samples per class. Ablation and visualization analyses further confirm the complementarity of global and local tasks, as well as the effectiveness of channel attention in enhancing feature separability and robustness. This study provides a practical solution for small-sample intelligent fault diagnosis and establishes a methodological foundation for applying self-supervised learning in real-world industrial systems.