Prototype Space Boundary Alignment Network for Bearing Continuous Fault Diagnosis Under Class-Incremental Scenarios

In recent years, the fault diagnosis method based on continuous learning has been widely used in the field of health monitoring of key components of rotating machinery. Most existing continuous learning fault diagnosis methods are deployed in an environment of relatively balanced datasets. However, the fault data generated by mechanical equipment under actual working conditions often have a long-tailed distribution, and the lack of tail fault data worsens the catastrophic forgetting problem of traditional continuous learning methods. A new prototype space boundary alignment network (PSBAN) for bearing continuous fault diagnosis under class incremental scenarios is proposed to solve the above problems. First, in the first stage, a new fusion loss function is designed to make the extracted features more robust. Then, in the second stage, the cluster mean algorithm is used to construct the prototype space of the class. At the same time, a total variance method is proposed to accurately estimate the central trend of the class and thus prevent the overdependence on the tail class from making the variance estimation inaccurate. Finally, the boundary of the prototype space is aligned by using the total variance. Experiments validate that the proposed PSBAN has higher accuracy than other methods do and effectively alleviates the catastrophic forgetting problem under the long-tailed distribution.