A lightweight rolling bearing fault diagnosis method based on improved EfficientNet model

Compared to traditional fault diagnosis methods that rely on manual analysis and cannot fully extract the rich information within signals, deep learning models can achieve more ideal recognition results. However, these models often suffer from large parameter sizes and high computational costs. This paper proposes a method combining the gramian angular field (GAF) encoding technique with an improved EfficientNet-B0 model for bearing fault diagnosis. Firstly, the one-dimensional bearing signal is encoded into a two-dimensional time-series image using the GAF method. Secondly, the two-dimensional image is input into the EfficientNet-B0 model, which incorporates the CBAM attention mechanism, for automatic feature extraction and classification. Finally, in the simulation experiments, bearing datasets from Case Western Reserve University and Paderborn University in Germany were used. The diagnostic method based on the GAF and EfficientNet-B0-CBAM model achieved recognition accuracies of 99.90% and 98.04% for bearing faults, respectively. It can be concluded that the proposed method maintains the lightweight characteristics of the model while achieving higher recognition accuracy and better generalization capability.