Study on fault diagnosis of ultra-low-speed bearings under variable working conditions based on improved EfficientNet network

Bearing fault diagnosis plays an important part in preventing rotating equipment faults, especially in the field of ultra-low-speed bearing fault diagnosis. Due to their low fault frequency and insignificant fault characteristics, it is difficult to realise the fault diagnosis of ultra-low-speed bearings using traditional methods; therefore, based on acoustic emission (AE) signals, this study proposes an ultra-low-speed bearing recognition model with EfficientNet as the backbone feature extraction network and successfully achieves bearing fault diagnosis under small-sample variable working conditions combined with transfer learning. The coordinate attention (CA) mechanism is introduced into the EfficientNet backbone feature extraction network to improve the ability of the model to extract detailed position information. The AdamW optimisation algorithm is introduced to improve the generalisation ability of the model. Combined with the idea of transfer learning, the data under different working conditions are trained and tested to form a high-performance and lightweight small-sample variable condition bearing recognition model called EfficientNet-CA-AdamW (EfficientNet-CAA). Comparison experiments show that the EfficientNet-CAA model proposed in this study has an accuracy of 99.81% for ultra-low-speed bearing recognition when the training samples are sufficient. Furthermore, the recognition accuracy is smoother and the loss function is significantly lower compared with convolutional neural network (CNN) models such as AlexNet, VGG-16, ResNet-34, ShuffleNet-V2 and EfficientNet-B0. In small-sample variable condition fault recognition, it has more powerful advantages compared with the other models. The recognition accuracy under variable conditions can reach more than 98%, which is significantly higher than that of the other models, and effectively improves the bearing fault recognition accuracy under small-sample variable conditions. In this study, the CA mechanism and the AdamW optimisation algorithm are introduced to lessen the difficulty of extracting detailed features and address the lack of generalisation ability of the EfficientNet model, which provides an idea for the application of the deep learning model to small-sample bearing fault diagnosis under variable working conditions.