An intelligent fault diagnosis of rotating machinery in class-incremental scenarios

Diagnostic models usually presume the presence of all fault types in the provided dataset. Due to changing work circumstances and longer service times, new faults will undoubtedly arise. However, when previous fault data is unavailable, only retraining with new fault types might result in catastrophic forgetting problems, rendering the diagnostic model unreliable. To tackle this challenge, an intelligent fault diagnosis of rotating machinery based on class-incremental learning (CIL) framework has been proposed, namely the Zipper Network (ZipNet). The ZipNet-based CIL framework first creates an invertible mapping module and then utilizes dual invertible mapping modules updated alternatively to replay known fault types and learn new ones. Furthermore, an orthogonal constraint module is included to enhance the ability to accurately identify all forms of faults. Comprehensive tests are carried out on two fault datasets to confirm the efficiency and capability of the ZipNet-based CIL framework.