Bearing fault diagnosis is essential for rotating machinery safety but remains challenging due to complex time–frequency patterns, limited samples, and imbalanced fault types. Existing models struggle to exploit multi-scale features and coordinate expert predictions effectively. To address this, we propose a hierarchical multi-scale expert fusion network that integrates multiscale convolutional feature extraction with a dual-path Mixture-of-Experts decision mechanism. Its core innovation is a masked symmetric Kullback–Leibler divergence regularization at the decision level, which promotes expert predictions and mitigates homogenization under small-sample and imbalanced conditions. The effectiveness of the proposed method is verified on multiple benchmark datasets, where it achieves 99.36% accuracy on the SEU dataset.

Rotating machinery,Feature learning,Intelligent fault diagnosis,Multi-scale feature extraction