Enhanced Meta-Transfer Learning for Few-Shot Fault Diagnosis of Bearings with Variable Conditions

Abstract

The transfer learning method performs better than conventional deep learning when dealing with the few-shot diagnosis situation where obtaining the true bearing defect signal is challenging. In order to leverage transfer learning to overcome the few-shot challenge of variable-condition bearing failure diagnosis, we propose the few-shot fault diagnosis approach based on enhanced meta-transfer learning. First, the network parameters are optimized based on a meta-learner. Second, a meta-learning-based transfer network model is constructed, combined with domain-adaptive methods to obtain a meta-learner with strong generalization ability. Meanwhile, the channel attention module is applied to the feature layer to strengthen the model’s feature expression ability. The proposed method Take advantage of the limited fault feature on small-sample data, while avoiding overfitting and improving the generalization ability. The performance of the proposed approach is verified on the fault data from the low-speed dynamic balance test bench. The consequences indicate that the diagnosis approach based on meta-transfer learning can accurately classify the bearing failures under variable conditions. Contrasted to other approaches, the proposed approaches possess better accuracy and generalization capability.