Remaining Useful Life Prediction Of Rolling Bearing Based On Time-frequency Analysis And Deep Learning

As an important part of machinery and equipment,to ensure the safety of industrial production,accurate prediction of the Remaining Useful Life of rolling bearings(RUL)can prevent the failure of bearings leading to mechanical equipment failure.During the operation of mechanical equipment,the vibration data collected are often non-smooth and non-linear because the bearing vibration signals are very susceptible to background noise during collection and transmission,so this paper utilizes different timefrequency analysis methods to extract the time-frequency degradation characteristics in the original bearing vibration data.In predictive modeling of degradation trends of bearings,traditional physical or mathematical models cannot create a practically oriented unified model,therefore,in this paper,deep learning networks are utilized in predictive modeling of pictorial degradation data,with the following main research components.(1)An image-based TET degradation data feature construction with a bearing RUL prediction network based on Convolutional Neural Network(CNN)and Bi-directional Long Short-Term Memory(BILSTM)is proposed.Firstly,for the non-smooth and nonlinear characteristics of the original vibration data of the bearings,Transient-Extracting Transform(TET)is used to obtain the time-frequency characteristics of the bearings and transform them into time-frequency maps.At the same time,to address the confounding problem of the time-frequency map,data rearrangement,bilinear interpolation and channel splicing are used to reduce the confounding phenomenon based on the analysis of the energy distribution and variation in the time-frequency map;secondly,a CNN is built to automatically extract the spatially correlated degradation features of the degradation data;then,in order to further obtain the temporal relationship between the degradation features,BILSTM is used to further calculate the degradation features to predict the Health Indicator(HI)The prediction results in the bearing dataset PHM2012 and XJTU-SY show that the RUL prediction model proposed in this paper can effectively predict the remaining service life of bearings.(2)A bearing Health Stage(HS)segmentation model based on deep stacked AutoEncode(AE)with KMeans algorithm and a residual deep convolutional RUL prediction network based on attention mechanism are proposed.First,an attention-based residual depth convolutional network is constructed using Continuous Wavelet Transform(CWT)to extract the time-frequency features of bearings,and an attention-based residual depth convolutional network is constructed with Res Net18 as the backbone for the underutilization of degradation information of bearing time-frequency features in both spatial and channel dimensions by neural networks,and second,for the RUL prediction process using a large amount of bearing health stage data will cause the prediction results to deviate from the degradation trend,resulting in the degradation of prediction accuracy,this paper uses deep AE and KMeans algorithm to construct an unsupervised model for bearing HS division,which divides the entire operating cycle of the bearing into four operating stages;finally,the trained prediction network is used to perform HI prediction on the unhealthy data of the test set,and through data smoothing and optimal regression fitting HI to achieve RUL prediction.Experimental simulations on the PHM2012 dataset show that the bearing HS classification and prediction network proposed in this paper can not only effectively classify the bearing healthy and unhealthy operation data,but also effectively improve the RUL prediction accuracy.