Fault Diagnosis Method for Rolling Bearing under Variable Working Conditions Using Improved Residual Neural Network

Aiming at the bad effect of fault diagnosis of rolling bearing due to complex and changeable working environments, ambient noise influence and insufficient valid sample data, an improved residual neural network method for fault diagnosis is proposed under variable working conditions. The acquired time domain signals of rolling bearing are taken as the inputs, and according to the strong time-varying characteristic of time domain signals of rolling bearing, an improved data pooling layer based on the Inception module is constructed. To extract the feature information effectively, following the Inception module idea, the data pooling layer is constructed by three small 3×3 stacked convolutional layers in series and by adding residual connection. A kind of residual block with a skipping connecting line is designed by adding a skipping connecting line, which can enhance the learning efficiency of characteristic information. Because the dilated convolution can expand the receptive field, the normal convolution in the residual block with a skipping connecting line is replaced by a dilated block, so a dilated and residual block with skipping connecting line is designed. The neural network is designed by the two kinds of residual blocks in end-to-end connection. Compared with SVM+EMD+Hilbert envelope spectrum, BPNN+EMD+Hilbert envelope spectrum and ResNet, the results show that the average accuracy of the proposed method in the variable noise experiment is 97.34%, the accuracy in the variable load experiment is 88.83%-96.76%, which are higher than the other methods, and the average accuracy is higher than ResNet method with a lower mean variance of 0.000 6 in the variable working condition experiment, so the noise resistance and generalization ability of the proposed method are verified. © 2020, Editorial Office of Journal of Xi'an Jiaotong University. All right reserved.