Jixie chuandong (Jul 2023)
Fault Diagnosis of Rolling Bearings Based on Two-step Transfer Learning and EfficientNetV2
Abstract
A rolling bearing fault diagnosis model based on two-step transfer learning and EfficientNetV2 (TSTE) is proposed for the real fault diagnosis environment in engineering, where the scarcity of available data leads to the low accuracy of the intelligent diagnosis model in bearing health status diagnosis. Firstly, the model is trained on the full life time bearing data set and then the model shallow weights are freezed to train it on the multi-condition bearing data set for the first transfer learning. Secondly, by constructing a class-imbalance dataset, the impact of scarcity of available data on fault diagnosis performance is studied in actual fault environments is studied. Then, the synthetic minority oversampling technique (SMOTE) oversampling method and the edited nearest neighbors (ENN) under the sampling method are proposed to expand the fault data, reconstructing the class-imbalanced dataset into a class-balanced dataset. Finally, the model is trained on the class-balanced dataset, freezing the model's bottom weights and training the model deeper for a second transfer learning, enabling the model to take control of the failure characteristics of the balanced dataset. The experiments are evaluated by a variety of metrics, while comparing it with other methods, and using the Grad-CAM method for feature visualization. The results show that the proposed method is able to transfer the fault diagnosis knowledge accumulated by the model in a laboratory environment to actual engineering equipment. It is suitable for the diagnosis of rolling bearing faults in situations where test data is scarce.
