Research On EMUs Bearing Condition Assessment And Remaining Useful Life Prediction Method Based On Deep Learning

The research on Health Condition(HC)assessment and Remaining Useful Life(RUL)prediction algorithms can provide strong support for the operation and maintenance management of bearings in Electric Multiple Units(EMUs).With the rapid development of sensor technology,the data-driven Deep Learning(DL)method has gradually become an effective research method in the field of Prognostics and Health Management(PHM).However,in data-poor scenarios,the performance of supervised DL-based methods usually plummets or even fails,such as in scenarios with complex operating conditions and non-life-cycle vibration monitoring data.The Transfer Learning(TL)method constructs the mapping relationship between source domain data and target domain data through semi-supervised or unsupervised modeling,so as to achieve the same or similar predictive tasks on the target domain with the source domain,which largely solves the problems caused by insufficient data or lack of labels.This degree thesis analyzes the advantages and disadvantages of bearings HC assessment and RUL prediction methods based on DL,focusing on the problems that the degradation data come from before bearings are not run-to-failure,the degradation data is scarce,and the labels are missing.For the degradation prediction(HC assessment and RUL prediction)of three different scenarios of mixed multiconditions,cross-condition,and cross-equipment,combined with the corresponding difficulties of research tasks,semi-supervised or unsupervised data-driven DL and Deep Transfer Learning(DTL)models are designed.Finally,four research works are realized: the bearings RUL prediction algorithm of the hybrid multi-conditions,the bearings RUL assessment algorithm across the operation conditions,and the bearings RUL prediction algorithm across the operation conditions and across the equipments.The specific work and contributions are as follows:(1)RUL prediction algorithm of the hybrid multi-conditions: Aiming at the problem with sparse target operating condition data,an auto-encoder based on feature reconstruction and feature dimensionality reduction is designed.By fusing and re-reconstructing the bearings vibration data of mixed multi-conditions,the distribution difference between different operation conditions is reduced.In addition,based on the reconstructed degradation representation,a non-fixed-length gated recurrent unit method is introduced to capture the degradation law in the time dimension.Then,the model trained under the mixed multi-conditions data is applied to RUL prediction under the target condition.Finally,the proposed algorithm for predicting the RUL of the hybrid multi-conditions bearings is verified on the bearings dataset of the French FEMTO-ST Institute,the bearings dataset of Xi’an Jiaotong University,and the bearings test bench of the high-speed railway traction motor bearings,which has outstanding performance in RUL prediction.(2)HC assessment algorithm of the Cross-Condition: Aiming at the situation that the bearings vibration data under the target operation condition are not full-life(Only the vibration data in the initial healthy stage and early degradation stage),based on the feature extraction of the traditional convolutional neural network,a feature mining algorithm based on graph convolutional network is designed,which builds a deep representation on the associations between different degradation features.In order to better adapt the source domain operation condition data to the target operation condition data,in addition to the loss function of HC assessment in the source domain,three additional loss constraint terms are added.Then,the above four loss items are superimposed to form the total loss value in the model training process.The model is trained through iterative learning to optimize the parameters.Finally,the real-time Health Indicator(HI)of the bearings in the source domain and target domain is generalized to a generalized degradation path(From HI equals 1 in the initial healthy stage to that HI equals to 0 when the occurrence of failure).The proposed method is verified and analyzed on the bearings data set of the French FEMTO-ST Institute and the high-speed railway traction motor bearings test bench.In order to further verify the generalization ability of the proposed method,the analysis is also verified on the NASA turbofan engine data set.Extensive experimental results demonstrate the effectiveness and generalizability of the proposed cross-condition HC assessment method.(3)RUL prediction algorithm of cross-condition/cross-equipment : Aiming at the problem that the supervised RUL prediction method fails when the degradation label data of bearings are missing in the target operating condition,an unsupervised cross-condition and cross-equipment bearings RUL prediction method is proposed.According to the large distribution differences in the process of cross-condition and cross-equipment RUL prediction,an unsupervised adversarial domain adaptation model constructed by three components: feature extractor,regression predictor,and domain discriminator is designed.The proposed method was originally designed for the migration gap problem of bearings RUL prediction across equipments,during the experimental verification process,the migration prediction scenario across operation conditions on the same equipment was also considered.In the experimental verification part,a large number of sufficient experiments show that the proposed method can achieve significant performance in the RUL prediction of bearings under cross-condition and cross-equipment.