You can get smaller: A lightweight self-activation convolution unit modified by transformer for fault diagnosis

The fault diagnosis methods based on convolutional neural network (CNN) have achieved many excellent results. However, owing to the deployment cost, numerous CNNs with large parameters are difficult to be directly applied to industrial practice. Therefore, this work aims to use lower parameters (order of magnitude is thousand) to complete the task of bearing fault diagnosis on the premise that the model has high-accuracy. To achieve this goal, a convolution unit modified by transformer was proposed, who is based upon the self -activation function, which makes the transformer and CNN organically integrated into a whole. Then, based on this unit, a series of novel lightweight diagnosis frameworks were proposed, named SANet. Finally, it was demonstrated that the proposed SANet can complete the high-accuracy diagnosis task with less than three thousand parameters and has strong robustness to noise (Average accuracy in various noise environments: 84.55%), and that SANet can achieve satisfactory results when there are few training samples (The number of samples of each category is 3 x 4), through four research cases. To sum up, based on this novel unit, we provide a series of lightweight frameworks with high-accuracy, strong robustness, and low sample demand, which is expected to promote the process of fault diagnosis technology from theoretical research to industrial practice.

Automated summary

This paper proposes a convolutional unit modified by transformer, which integrates transformer and CNN into a whole, and based on this unit, a series of novel lightweight diagnosis frameworks named SANet are proposed. It is demonstrated that SANet can complete the high-accuracy diagnosis task with less than three thousand parameters and has strong robustness to noise, as well as satisfactory results with few training samples.