Multiscale inverted residual convolutional neural network for intelligent diagnosis of bearings under variable load condition

Abstract

In industrial production, it is particularly important to diagnose the bearing fault in time under variable loads. The intelligent diagnosis method has strong robustness without human intervention, but it needs a lot of raw data. However, large amounts of data storage is relatively difficult and slow transmission speed. Meanwhile, under different loads, the same fault feature has no significant difference in the process of bearing degradation. To address these problems, this article proposes a new multiscale inverted residual convolutional neural network (MIRCNN) method for fault diagnosis of variable load bearing. Firstly, a semi tensor product compressed sensing (CS) method based on parallel orthogonal matching pursuit (POMP) is proposed. The vibration signal is reconstructed with the proposed method to solve the problems of difficult data storage and slow transmission speed. Then, the convolutional neural network (CNN) is designed for high-dimensional signals, so that the one-dimensional signal is converted to three-dimensional image for further training. Finally, the multiscale algorithm is applied to the CNN architecture, and MIRCNN is established by adding inverted residual learning. It can extract the different features between fault signals of variable load bearings, improving the ability to identify faults. Experimental results on two rolling bearing test beds with different bearing types and operating conditions and compared with existing state-of-the-art methods to prove the effectiveness and accuracy of the proposed method.