Purpose <p>Fault diagnosis of rotating machinery often suffers from severe class imbalance because fault samples are scarce under practical operating conditions. Existing data augmentation methods still face challenges in simultaneously preserving sample quality, diversity, and structural consistency. This study aims to develop an effective diffusion-based generation framework for imbalanced fault diagnosis.</p> Methods <p>A local–global residual enhanced diffusion model (LG-ResDiff) is proposed. The denoising network is improved by introducing a Local Information Perception Feed-Forward Network to strengthen local feature extraction and global dependency modeling. A residual-enhanced bottleneck structure is further designed to stabilize feature propagation during reverse diffusion. Conditional time–label embeddings are incorporated to guide class-aware generation of minority-class samples. The generated samples are then combined with original training data to construct balanced datasets for fault classification.</p> Results <p>Experiments conducted on bearing and gear datasets demonstrate that the proposed method can generate high-quality time-frequency representations with strong structural fidelity and distribution consistency. Compared to various generative methods, LG-ResDiff achieves superior generation quality and consistently improves fault diagnosis accuracy under various imbalance ratios. Under fully balanced conditions, the proposed method achieves a diagnostic accuracy of 99.5% on the bearing dataset and 99.3% on the gear dataset.</p> Conclusion <p>The proposed LG-ResDiff provides an effective solution for minority-class sample augmentation in rotating machinery fault diagnosis. By jointly enhancing local detail representation and global structural modeling, it improves the fidelity, diversity, and stability of generated samples and offers reliable support for intelligent fault diagnosis under imbalanced data conditions.</p>