Enhanced regularized prototypical network for few-shot fault diagnosis of electric submersible pump

Meta-learning has emerged as a promising paradigm for few-shot fault diagnosis (FSFD) in electric submersible pump systems. However, existing approaches often struggle to effectively model the complex spatiotemporal coupling among multivariate time series. Moreover, they usually exhibit limited robustness to uncertain environmental disturbances, which may further degrade diagnostic performance To overcome these limitations, we propose an enhanced regularized prototypical network. Specifically, a crossformer-enhanced embedding module is developed by repurposing the crossformer encoder as a spatiotemporal feature extractor, enabling explicit modeling of inter-variable dependencies in faulty data. Furthermore, a low-rank informed regularization strategy is proposed to suppress redundant variations in the embedding space while preserving fault-related discriminative structures, thereby enhancing representation stability and generalization under interference. Extensive experiments on real-world electric submersible pump fault datasets provided by China National Offshore Oil Corporation demonstrate that the proposed method consistently outperforms state-of-the-art meta-learning methods across various few-shot scenarios, validating its effectiveness for FSFD under limited data conditions.