Bearing Multi-sensor Fusion Fault Diagnosis Based on an Adaptive ResGAT Network

Abstract

The rolling bearing condition monitoring signal under strong noise interference is characterized by non-stationary multi-component signals, and the fault information contained in a single sensor signal is limited, which cannot fully characterize the operating state of the equipment. This study proposed a multi-sensor fusion fault diagnosis method based on the adaptive residual graph attention convolutional neural network (ResGAT), which uses multiple sensor monitoring signals to accurately identify the rolling bearing fault information under different working conditions. Firstly, the vibration signals collected by multiple sensors were decomposed into wavelet coefficient matrices by variational mode decomposition (VMD) and wavelet packet decomposition (WPD), and the graph structure data containing multi-sensor network information was constructed based on the radius graph strategy. Secondly, based on the short-circuit characteristics of the residual network, an adaptive ResGAT was designed, which used the output and residual of the network to deeply mine the redundant fault information of multi-sensor fusion data. Finally, the proposed ResGAT model was applied to rolling bearing fault diagnosis datasets under three different working conditions: constant speed, variable speed, and composite fault. The research results show that compared with existing methods, the proposed method has higher classification accuracy and robustness under three working conditions.

Keywords

• Multi-sensor fusion
• Graph neural network
• Rolling bearing
• Fault diagnose
• Wavelet packet decomposition