Fault diagnosis of rolling bearings in high-noise environments based on a variational gated hierarchical memory network

Abstract Rolling bearing fault diagnosis under strong noise environments has remained a major challenge in industrial applications. Traditional signal processing methods rely on manual feature engineering, making feature extraction difficult in complex noise conditions. While existing deep learning approaches can automatically learn features, they exhibit insufficient robustness under extreme noise conditions, particularly with limited capability for handling non-Gaussian and composite noise. To address this technical bottleneck, this paper proposes an intelligent fault diagnosis method based on variational gated hierarchical memory network. The method integrates three core modules: variational probabilistic temporal feature extraction, probability-driven adaptive gating, and dual-level memory enhancement, achieving end-to-end noise-robust fault diagnosis. First, a variational gated temporal convolutional network is designed to map deterministic features to probability distribution parameters through a variational encoder, enabling natural quantification of feature uncertainty. Second, a probability-driven adaptive gating mechanism is introduced to dynamically adjust gating weights based on the uncertainty of variational features, realizing noise-aware adaptive feature selection. Finally, a dual-level adaptive memory enhancement module hierarchical adaptive variational memory bank is constructed, employing category-specific and globally-shared dual memory architecture to provide stable memory support and feature enhancement under complex noise environments. Experimental results demonstrate that the proposed method significantly outperforms existing approaches on two public benchmark datasets, maintaining diagnostic accuracy above 99.5% even under extreme noise conditions (−30 dB SNR), with strong robustness against white noise, pink noise, power frequency interference, and mixed noise. This research provides a breakthrough technical solution for bearing fault diagnosis under strong noise backgrounds, offering significant theoretical value and broad engineering application prospects.