Diffusion-augmented contrastive learning framework for quantitative diagnosis under limited data conditions

Quantitative diagnosis of bearing faults is essential for ensuring the safe and reliable operation of mechanical transmission systems, especially under complex and variable operating conditions. However, in real-world scenarios, collecting sufficient labeled fault data remains a major challenge, hindering accurate fault classification and severity estimation. While diffusion models have shown strong potential in data generation, existing methods primarily use them to expand training sets without explicitly modeling fault semantics or guiding the learning process. To address these limitations, we propose a novel Diffusion-Augmented Contrastive Learning (DiCL) framework for quantitative bearing fault diagnosis under limited data conditions. First, a fault-controllable denoising diffusion probabilistic model (DDPM) is developed to generate class-conditional synthetic signals across various fault types and severity levels. These synthetic samples are further used to construct compound fault labels that reflect complex or multi-fault conditions. Second, a dual-branch contrastive learning strategy is adopted, where two input samples are jointly processed to form contrastive pairs. This mechanism enables the feature extraction network to learn fault-discriminative representations by reinforcing shared fault characteristics and suppressing irrelevant variations. Third, a cycle-consistency constraint is introduced via a composite loss function to enforce semantic alignment among samples of the same fault class. The proposed DiCL framework is evaluated on two bearing fault datasets: the Paderborn University bearing dataset and a laboratory-scale mechanical transmission system dataset. Experimental results demonstrate that DiCL achieves high-fidelity data generation and superior diagnostic performance. Notably, even with only 5% of the fault training set, DiCL attains over 80% classification accuracy on both benchmark datasets, significantly outperforming state-of-the-art baseline methods.