An improved fault diagnosis method of rolling bearings based on LeNet-5

This paper proposed a convolutional neural network fault diagnosis method based on improved LeNet-5, aiming at the incompleteness of rolling bearing fault samples. This method takes the original time-domain vibration signals of rolling bearing containing multiple speeds as the input of the model in the form of two-dimensional grayscale images. The input size is determined according to the signal characteristics, and features are extracted adaptively through convolution operations. This method introduces a batch normalization operation to improve the model and uses the softmax classifier to implement fault classification and recognition. Finally, the t-distribution neighborhood embedding algorithm (t-SNE) is used to objectively demonstrate the feature extraction effect of the method. The rationality and effectiveness of the improved model were verified by the multi-fault experimental analysis of rolling bearings. Experimental results show that by training the rolling bearing fault data at four speeds can learn the common characteristics of bearing fault samples with limited speed, accurate classification of rolling bearing faults can be achieved. And the fault data at other speeds are also valid, which broadens the speed generalization ability of rolling bearing fault diagnosis. The BP neural network (BPNN) and support vector machine (SVM) algorithms were compared with the method proposed in this paper, which proves that the method has good robustness and generalization ability. This work can provide reference and reference for ensuring the reliability of rolling bearings and the safe operation of equipment. © 2021, Editorial Office of Journal of Vibration and Shock. All right reserved.