Research On Transfer Prediction Method Of Rolling Bearings Remaining Useful Life Based On Deep Feature Representation

Recently,prediction remaining useful life(RUL)by collecting,analyzing and modeling the degradation data of machinery has become a hot research topic in the field of Prognostics and Health Management(PHM).However,traditional RUL prediction methods usually ignore the distribution differences and temporal dependencies among the degradation data,which will result in the reduction of prediction accuracy and stability.Aiming at RUL prediction problems of rolling bearings,this paper adopts two perspectives of temporal characteristics and data distribution and introduces deep learning techniques to effectively exploit feature representation information of degradation sequence data.And two kinds of RUL prediction methods based on temporal information representation and transfer learning are constructed respectively.The main work and contributions are as follows:(1)Aiming at the problem that the representative ability of bearings degradation feature is not enough,a new auto-encoder ELM-based feature extraction method is proposed in this paper to verify the validity of deep learning techniques by means of fault diagnosis.First,we run a Fast Fourier Transform(FFT)algorithm on raw data to obtain the frequency spectrum.Second,we build a multi-layer extreme learning machine auto-encoder and train it with frequency data to obtain the deep feature and construct a fault diagnosis model.Experiments are conducted on bearing data sets of CWRU and the results show good representative ability of deep feature compared with the existing fault diagnosis methods and common features.(2)For the problem of ignoring temporal dependency in degradation modeling,a new RUL prediction approach based on deep feature representation and long short-term memory(LSTM)neural network is proposed.First,deep features with good representational ability can be obtained from the convolutional neural network(CNN)by inputting Hilbert-Huang marginal spectrum of the vibration signal.Second,a new singular value decomposition(SVD)correlation coefficient based health state assessment method is proposed to divide the degradation process into different health states.Finally,the features of fast-degradation process and corresponding RUL values are fed into an LSTM neural network to construct RUL prediction model,and online bearings' RUL can be predicted.Experiments are conducted on IEEE PHM Challenge 2012 bearing data sets.The results show that the RUL prediction accuracy and stability can be improved by combining the good representation performance of deep features and temporal dependence.(3)Aiming at the problem of ignoring data distribution difference in degradation modeling,a new RUL prediction method based on deep feature and transfer learning is proposed.First,the Hilbert-Huang Transform marginal spectrum of raw vibration signal is calculated as input data and then contractive denoising auto-encoder(CDAE)is introduced to extract deep features.Second,by using the obtained deep features and Pearson's correlation coefficient,a new health condition assessment method is proposed.Finally,using the deep features adapted by transfer component analysis(TCA)and their RUL values,an RUL prediction model is trained by means of least-square support vector machine during fast degradation state.Results on the PHM Challenging 2012 dataset show that the remaining life prediction method proposed in this paper can effectively improve the prediction accuracy of the regression model by reducing the data distribution difference between different domains.As a conclusion,this dissertation provides a new solution for RUL prediction of rolling bearings from perspectives of deep learning and transfer learning.So we claim our work has significant theoretical and practical engineering value.