A lightweight neural network with strong robustness for bearing fault diagnosis

Abstract

• SIRCNN uses depthwise separable convolution realize lightweight model design. • SIRCNN has good performance in terms of diagnosis speed, model size and accuracy. • SIRCNN has high denoising ability under different noise environments. Traditional methods of rolling bearing fault diagnosis generally have the following disadvantages: low accuracy of fault severity identification, the need for artificial feature extraction, poor noise resistance and high requirements for diagnostic equipment. To overcome these disadvantages, an intelligent bearing fault diagnosis method based on Stacked Inverted Residual Convolution Neural Network (SIRCNN) is proposed. Compared with machine learning and classical convolutional neural networks, SIRCNN has a smaller model size, faster diagnosis speed and extraordinary robustness. The lightweight of the model is achieved through the application of depthwise separable convolution. Moreover, using the inverted residual structure ensures the accuracy of the model in noisy environments. The experimental results show that the fault diagnosis of rolling bearing based on SIRCNN can effectively identify the type and severity of bearing fault under different noise environments, improve the diagnostic efficiency and reduce the performance requirements for the diagnostic equipment.