Remaining Life Prediction Method Of Rolling Bearing Based On T-SNE And Deep Convolutional Neural Networks

Rolling bearing is one of the most important basic parts in mechanical equipment,and its health status directly determines the performance,quality and reliability of the equipment and its host products.Through the reliable remaining life prediction of the bearing,the operating status of the equipment and the remaining working time can be grasped in time,which is convenient for the staff to formulate the maintenance plan as soon as possible to improve the efficiency of the equipment and reduce the economic loss.However,the current data-driven life prediction methods are mostly based on obtaining different degradation characteristic information from multiple sensors,and rarely extract different fault information from single sensor data.Therefore,the thesis conducts in-depth research on five parts about noise reduction of vibration signal,feature parameters extraction,feature parameters selection,degradation state assessment and remaining life prediction.The main research contents are as follows:(1)In response to the phenomenon that the vibration data collected by sensors contains a lot of noise and interference signals due to the complex working environment of rotating machinery equipment,this thesis is devoted to propose a wavelet packet threshold noise reduction method based on Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(CEEMDAN).Then use this method to process the actual vibration signal for noise reduction,and extract the characteristic parameters in the time domain,frequency domain,and time-frequency domain from the noise-reduced signal.Finally,an unsupervised feature parameter selection method is used,that is,three assessment indicators are defined based on the nature of the degradation features.The comprehensive index value is obtained by weighting the indicators to screen out the feature parameters that can more accurately describe the degradation trend of bearing.(2)In view of the fact that the degradation characteristic parameters of rolling bearings have obvious stages,and the data samples collected during the failure process are different from the healthy samples.The thesis is devoted to propose a degradation assessment index construction method by combining t-Distributed Stochastic Neighbor Embedding(t-SNE)and correlation analysis.Then obtain the approximate range of the degradation start time according to the degradation assessment index,and then use the characteristics of t-SNE to separate different samples to determine the specific degradation time.Finally,verify the accuracy according to the envelope spectrum of the sample data at the time of degradation.And demonstrate the initial failure sensitivity of this method by comparing with other dimensionality reduction methods.(3)As the vibration data of rolling bearings is a one-dimensional time series,and the extracted feature parameters have the characteristics of long-term dependence,this thesis is devoted to propose a remaining life prediction method which combines Bi-directional Long Short-Term Memory(Bi LSTM)and Deep Convolution Neural Networks(DCNN).Firstly,the remaining life of the bearing is normalized as the label for network training according to the degradation start time determined by t-SNE.Then take advantages of Bi LSTM to obtain context information from time series and DCNN to extract the abstract features,so as to establish the mapping relationship between feature parameters and remaining life.Then the result analysis and error comparison of the remaining life prediction value of the tested bearing are carried out to verify the effectiveness of the deep learning model.