A Lightweight Fault Diagnosis Model of Rolling Bearing Based on Gramian Angular Field and EfficientNet-B0

The deep learning model can fully extract the rich information in the signal and provide a better recognition effect as compared to traditional fault diagnosis approaches, which rely on manual analysis. However, the deep learning model has the problem that too many parameters lead to high training cost and low generalization ability. Therefore, this paper proposes a lighter fault diagnosis model for rolling bearings using Gramian Angular Field (GAF) and EfficientNet-B0. Firstly, Gramian Angular Field is used to encode one-dimensional vibration signal into a two-dimensional temporal image, the two-dimensional image is then loaded into the selected EfficientNet-B0 for training in automatic feature extraction and classification recognition before a test set is used to confirm the model’s recognition accuracy. The results show that the recognition rate of bearing faults of the lightweight fault diagnosis model proposed in this paper based on Gramian Angle field and EfficientNet-B0 reaches 99.27%, and the number of parameters of the EfficientNet-B0 is about 1/5 of that of the ResNet-50. Compared with common fault diagnosis methods, this model has a lighter weight convolutional network model, better generalization and higher recognition rate.