Adversarial-Based Super Feature Reconstruction Meta-Transfer Network for Weak Feature Enhancement and Fault Diagnosis of Harmonic Drive

Deep transfer learning methods have achieved promising progress in the fault diagnosis field with insufficient data. However, these methods still face some challenges in industrial scenarios: First, the fault characteristics are extremely weak in the early fault stage. Second, a significant distribution discrepancy exists between laboratory conditions and industrial scenarios. To address the above problems, an improved weak feature enhancement and fault diagnosis network named adversarial-based superfeature reconstruction meta transfer (ASFRMT) is developed for harmonic drive. First, a superfeature reconstruction strategy is designed to enhance the feature representation of weak faults based on the adversarial with severe fault features, where a feature content loss and a distribution matching loss are provided to further extract the superfeature of weak faults. Second, a fault classification module is built for fault diagnosis based on enhanced weak features. Furthermore, the model-agnostic meta-transfer learning framework is employed from laboratory environment to industrial scenario. In particular, the output layer of the classification module is adjusted and re-trained to be suitable for harmonic drive. The comprehensive experiments are conducted on the cross roller bearing of harmonic drive, which verified the efficiency and robustness of the ASFRMT network compared to other fault diagnosis methods.