MMR-EENet: a multiscale convolutional and dual-path MoE framework for intelligent fault diagnosis
TL;DR
A macro–meso–residual expert ensemble network (MMR-EENet) that combines multiscale convolutional feature extraction with a dual-path mixture-of-expert decision mechanism, enhancing decision robustness and feature diversity is proposed.
Abstract
Accurate and robust bearing fault diagnosis is critical for the safe operation of rotating machinery, yet remains challenging due to complex time–frequency patterns, limited training samples, and imbalanced fault types. Existing convolution- and attention-based models partially address these issues, but their ability to exploit multi-scale features and coordinate multiple expert predictions is still limited. To address this limitation, this paper proposes a macro–meso–residual expert ensemble network (MMR-EENet) that combines multiscale convolutional feature extraction with a dual-path mixture-of-expert decision mechanism. Specifically, the network extracts representations at multiple receptive-field scales, capturing both global structural patterns and local discriminative details. These features are then integrated via a mixture-of-experts mechanism, enhancing decision robustness and feature diversity. Furthermore, a masked symmetric Kullback–Leibler divergence regularization is introduced at the decision level to encourage complementary predictions between experts while mitigating homogenization under small-sample conditions. Extensive experiments on multiple datasets, including BUCEA and XJTUSpurgear, demonstrate that MMR-EENet outperforms existing baselines. On the SEU dataset, it achieves up to 98.96% accuracy, 2%–3% higher than the best competing model.