Brownian Distance Covariance-Based Few-Shot Learning Framework Considering Noisy Labels for Fault Diagnosis of Train Transmission Systems

The significance of intelligent fault diagnosis techniques is increasing in maintaining the security and reliability of railway operations. In particular, few-shot learning shows promise since it can address the issue of limited fault samples. However, the existing approaches have the following shortcomings. First, they ignore rich fault information in the joint distributions of multi-dimensional features, limiting the improvement of diagnosis accuracy. Second, they lack specialized mechanisms to alleviate severe degradation in diagnosis accuracy caused by label mistakes in engineering scenarios. To address the above issues, a Brownian distance covariance-based few-shot learning framework of fault diagnosis considering noisy labels is proposed for train transmission systems. In network construction, a novel joint distribution expression (JDE) layer is developed and embedded into the prototypical network, implementing similarity measures based on the joint distribution characteristics. In network training, a new antimislabeling learning strategy is designed for few-shot fault diagnosis tasks, in which a similarity-weighted prototype aggregation mechanism mitigates the negative effects of mislabeled support samples and a loss function with outlier attenuation avoids disruptions from mislabeled query samples. Taking a fault diagnosis study case of motors, gearboxes, and axle boxes as examples, it is demonstrated that the proposed framework can effectively learn and recognize faults, even using limited training datasets containing parts of mislabeled samples. Moreover, the superiority of the proposed framework is demonstrated by comparing it with other state-of-the-art methods.