A rolling bearing fault diagnosis method based on a new data fusion mechanism and improved CNN

Abstract

The development of modern industry has accelerated the need for intelligent fault diagnosis. Nowadays, most bearing fault diagnosis methods only use the information of one sensor, and the diagnostic knowledge contained in single-sensor data is often insufficient, which leads to insufficient diagnostic accuracy under complex working conditions. In addition, although convolutional neural network (CNN) has been widely used in fault diagnosis, the network structures used are still relatively traditional, and the ability of feature extraction is relatively poor. To solve the problems, firstly, this paper innovatively uses coordinate attention (CA) to more fully mine fusion information after concatenate (Cat) operation and proposes a new data fusion mechanism, Cat-CA. Then an improved Residual Block is proposed, and a novel improved CNN is built by stacking this Block. Finally, the Cat-CA-ICNN is built by combining Cat-CA and improved CNN, and its effectiveness and superiority are verified using two datasets.