Prediction Of Residual Life Of Rolling Bearing Based On Deep Learning

The remaining life prediction of rolling bearing is an important problem in the health management of rotating machinery system.The traditional rolling bearing life prediction algorithm based on signal analysis and processing combined with artificial feature design and feature processing has high requirements for expert experience,long time-consuming and poor universality.The deep neural network model can automatically extract and process the features from the original time-domain vibration signal of rolling bearing without expert experience and manual design.Combining CUSUM strain point detection algorithm with the feature self extraction ability of deep neural network can improve the accuracy of residual life prediction,and realize the residual life prediction of rolling bearing through end-to-end structure.The main research contents of this paper are as follows:(1)In order to improve the end-to-end residual life prediction of rolling bearing signal and improve the accuracy of residual life prediction,the third chapter of this paper proposes PBiGRU(Pooled Bidirectional Gate Recurrent Unit)net and PBiGRU net construction algorithm based on the characteristics of deep cycle network technology,which is good at processing time series data,and based on the characteristics of one-dimensional time series vibration signal of rolling bearing.PBiGRU network introduces the pool layer into Bi GRU network,so that the network has the ability of data dimensionality reduction and avoids the generation of gradient explosion and gradient disappearance.The specific training process of the model is to segment the data processed by CUSUM strain point detection algorithm,train the model,and then output it as the remaining life prediction data after Savitzky-Golay smoothing.The model has good prediction accuracy.(2)Aiming at the problem of more noise in the vibration signal data of rolling bearing collected in the industrial field,the fourth chapter of this paper combines the idea of stacking integrated learning and the powerful signal noise reduction ability of DRSN(Deep Residual Shrinkage Network),takes multiple DRSN networks with different convolution kernel scales as the basic learner and fully connected neural network as the meta learner,MSDRSN(Multiscale Stacking Deep Residual Shrink Network)model is proposed.The model not only improves the accuracy of DRSN,but also retains the noise reduction ability of DRSN.Experiments show that the accuracy of residual life prediction of the model in the environment of noisy data set is higher than that of multi-scale convolutional neural network,and higher than that of base learning DRSN,which proves the effectiveness of the proposed model.