LDGM: Domain-invariant bearing fault diagnosis via Mamba-based linear complexity modeling and feature perturbation for edge deployment

In large mechanical bearings such as water turbine generators, fault diagnosis is often carried out at the edge end due to the risk of data transmission. The edge-end diagnosis model faces the challenge of cross-domain diagnosis ability posed by diverse operating conditions, but the limited computing power of edge devices is difficult to support the computational overhead of conventional cross-domain models. To address these challenges, a lightweight domain generalization model (LDGM) for edge deployment is proposed. The main innovation lies in the Mamba-MP architecture, which effectively decouples the global pattern and local transient influence features, and maximizes feature extraction while ensuring model lightweight. In addition, in order to enhance the domain robustness, the data augmentation alignment training strategy is introduced. This strategy exploits the decoupling property of LDGM and applies the cooperative perturbation to the global time–frequency information and instantaneous features to enhance the prediction consistency in order to improve the generalization ability under different working conditions. Experimental results on 14 cross-condition transfer tasks show that the average diagnostic accuracy of LDGM under different transfer tasks is more than 99%, which is significantly better than the comparison baseline.