Dynamic Balanced Dual Prototypical Domain Generalization for Cross-Machine Fault Diagnosis

Cross-machine domain-generalization-based fault diagnosis (DGFD) methods have shown promising prospects because they assume that the model trained on available source machines can generalize to unseen target machines. However, there are still some problems remaining unsolved. First, most existing methods learned domain-invariant representations from raw vibrational signals directly, and it makes the task challenging because the raw signals contain lots of domain-related information coming from measurement differences (e.g., sampling frequencies and sensor types, rotating speeds, and resonance frequencies). Second, existing cross-machine DGFD methods learn domain-invariant representation by aligning distributions at the class level but failing to suppress domain-specific information at the domain level. Third, the difficulty level of learning to classify samples from different domains varies for a neural network model, leading to the domain imbalance problem, which may damage the model's generalization ability. To solve the measurement issue, a novel signal preprocessing module is first devised to mitigate the negative effects of raw vibrational signals. Then, this article proposes a dynamic balanced dual prototypical network (DBDP-Net). In DBDP-Net, the dual prototype loss based on class and domain prototypes is designed to learn domain-invariant feature representation by reducing distribution discrepancy at class and domain levels. Finally, to reduce the domain imbalance problem, we propose a dynamic weighted strategy to balance the feature learning process of different domains. Extensive experiments are conducted on five bearing fault datasets collected from different machines, and the results show that the proposed DBDP-Net is superior to the state-of-the-art (SOTA) DGFD methods.