A Method for Domain Generalization in Rolling Bearing Fault Diagnosis Based on Mamba and Causal Prototype Decoupling Constraints
To address the insufficient generalization capability of rolling bearing fault diagnosis models under complex operating conditions such as variable rotational speed, variable load, and variable radial force, this paper proposes a domain generalization fault diagnosis method based on Mamba feature extraction and causal generalization loss. First, the raw vibration signals are standardized and segmented using a sliding window strategy. Then, the selective state space model Mamba is employed to model long-range temporal dependencies and local dynamic variations in fault impact signals. Subsequently, from the perspective of causal invariance, generalization constraints are constructed by treating stable representations related to fault categories as causal features, while regarding amplitude variations, noise disturbances, and speed fluctuations induced by operating-condition changes as non-causal factors. A causal generalization loss consisting of class-conditional causal prototype consistency loss and feature decoupling regularization is designed to enhance the model’s adaptability to unseen operating conditions. Experiments are conducted on the Paderborn University (PU) and JNU datasets, where four operating conditions are regarded as four domains to construct cross-condition diagnosis tasks under a leave-one-condition-out evaluation protocol. The proposed method is compared with support vector machine (SVM), one-dimensional convolutional neural network (1D-CNN), one-dimensional residual network (ResNet1D), temporal convolutional network (TCN), correlation alignment (CORAL), domain-adversarial neural network (DANN), maximum mean discrepancy (MMD), and vanilla Mamba. Experimental results show that the proposed method achieves an average accuracy of 93.56% on the PU dataset and 96.28% on the JNU dataset, outperforming all comparison methods and further demonstrating its effectiveness.