Abstract：Rolling bearings are installed in various machine tools and other production machinery, and are prone to faults and failures, requiring continuous monitoring to ensure their safe and reliable operation. In this paper, a multiple parallel graphs convolutional neural residual network (MPGCN-Resnet) is designed to complete the fault diagnosis of rolling bearings. This method consists of four parts. Among them, the time-frequency graph acquisition part based on Cmor wavelet can complete refined processing of reconstruction and disassembly in various fault vibration signals. The feature acquisition part based on a multi-parallel network can improve generalization and accelerate convergence. The feature learning part under the residual structure graph neural network can utilize the residual structure to complete feature learning and can realize the in-depth exploration of fault characteristics of rolling bearings. The GAP-Softmax fault classification part can complete the effective diagnosis of rolling bearing faults. The CWRU bearing data set is selected to complete the comparison and analysis experiments of this method and the IHDSVM-Alexnet and MSATM methods in terms of variable working conditions, variable noise, accuracy, and loss values. The results show that the average fault diagnosis accuracy of this method for rolling bearings can reach 96.4%, which is higher than 91% in a -6 dB noise environment and greater than 90% when the load suddenly changes by 3×0.75KW. This method has higher fault diagnosis accuracy for rolling bearings in various variable working conditions and variable noise environments than the other two methods, and can alleviate the problems of increased parameters and excessive calculation.
2024, Vol. 45 
