Dynamic Balance Domain-Adaptive Meta-Learning for Few-Shot Multi-Domain Motor Bearing Fault Diagnosis Under Limited Data

Under varying operating conditions, motor bearings undergo continuous changes, necessitating the development of deep learning models capable of robust fault diagnosis. While meta-learning can enhance generalization in low-data scenarios, it is often susceptible to overfitting. Domain adaptation mitigates this by aligning feature distributions across domains; however, most existing methods primarily focus on global alignment, overlooking intra-class subdomain variations. To address these limitations, we propose a novel Dynamic Balance Domain-Adaptation based Few-shot Diagnosis framework (DBDA-FD), which incorporates both global and subdomain alignment mechanisms along with a dynamic balancing factor that adaptively adjusts their relative contributions during training. Furthermore, the proposed framework implicitly leverages the concept of symmetry in feature distributions. By simultaneously aligning global and subdomain-level representations, DBDA-FD enforces a symmetric structure between source and target domains, which enhances generalization and stability under varying operational conditions. Extensive experiments on the CWRU and PU datasets demonstrate the effectiveness of DBDA-FD, achieving 97.6% and 97.3% accuracy on five-way five-shot and three-way five-shot tasks, respectively. Compared to state-of-the-art baselines such as PMML and ADMTL, our method achieves up to 1.4% improvement in accuracy while also exhibiting enhanced robustness against domain shifts and class imbalance.