Multi-view dynamic manifold reconstruction with adaptive cross-attention fusion enables noise-robust bearing fault diagnosis

Abstract

To address the issues that bearing fault features are easily submerged under strong noise and that manifold topologies are prone to collapse, this paper proposes a Multi-view Dynamic Manifold Reconstruction and Adaptive Cross-Attention Fusion Network (MDM-CA Net). First, based on the theory of phase space reconstruction, dynamic K-nearest neighbor graphs are constructed in both the time domain and the frequency domain. A dual-channel graph convolutional network is employed to mine the intrinsic geometric structure of the signal, while complementary information from multiple views is utilized to filter out spurious connections induced by noise, thereby isolating noise outliers at the representation level. Second, a noise-aware temporal enhancement module centered on a bidirectional gated recurrent unit and integrated with a global attention mechanism is developed to adaptively suppress background noise interference. Finally, a Transformer-style cross-attention strategy is introduced, where temporal features serve as queries to retrieve critical patterns from spatial topologies, enabling deep semantic alignment and nonlinear interaction between heterogeneous spatiotemporal features. Experimental results on three public datasets, CWRU, SEU, and PU, demonstrate that MDM-CA Net achieves optimal comprehensive performance under both same-domain and cross-domain operating conditions. Under the extreme condition with a signal-to-noise ratio as low as − 4 dB, it still maintains diagnostic accuracies of 94.2%, 93.5%, and 95.8%, respectively. Ablation studies confirm the synergistic enhancement effect among the modules: the multi-view mechanism drives high precision, while cross-attention drives high recall, and their collaboration achieves an optimal balance.