Research On Deep Learning Based Health Assessment And Remaining Useful Life Prediction Of Rolling Bearing

Fault prediction is one of the main tasks of prognostic and health management(PHM).Its purpose is to predict the remaining useful life of machinery based on the historical and continuous degradation trends observed from condition monitoring data.With the industrial production entering the era of big data,every component of PHM system has been branded with big data.With the rapid development of deep learning in academia and industry,it has significantly improved the recognition accuracy in many traditional recognition tasks and demonstrated its strong ability to deal with complex recognition tasks.As the key component of supporting transmission shaft in rotating machinery,rolling bearing is widely used in mechanical,electrical,petrochemical,metallurgical and some military industries.The performance,life and reliability of bearings play a decisive role in the safe and stable operation of the whole rotating machinery system.Therefore,the research on fault prediction of rolling bearing driven by big data is of great significance.Taking the rolling bearing as the research object,this paper makes full use of the advantages of deep learning and its strong ability in feature extraction,combined with prediction and health management theory,carries out health assessment and remaining useful life prediction(RUL)for the rolling bearing.The purpose is to avoid accidents to the greatest extent,and to provide a new perspective for the main maintenance of enterprises.The innovation and main contents of this paper are as follows:(1)In order to describe the dynamic process of rolling bearing performance degradation accurately,considering the advantage of strong feature extraction ability of deep learning,a novel method combining deep auto-encoder(DAE)with minimum quantization error(MQE)is proposed in this chapter,to evaluate the whole life health of rolling bearings.The original feature is compressed and extracted by the DAE model,and the compressed feature is sorted according to the trend.Then the feature with large trend is selected to construct the health indicator(HI)by using the MQE method.The evaluation criteria based on a single metric is often biased,so a weighted fused evaluation criterion based is presented.The validity of the proposed method is verified by IEEE PHM 2012 data.(2)Aiming at the problem that most of the existing HI construction methods are based on the feature extraction and feature fusion technology of bearing life-cycle vibration signal,these methods need a lot of prior knowledge and the existing HI construction methods based on deep learning are using fixed learning rate to train network,and the network training efficiency is low,a deep convolutional neural network with exponential decayed learning rate(EDCNN)is described to evaluate the health of rolling bearings.The original vibration signals are input to the proposed network.The features extracted from the deep convolutional neural network(DCNN)are input to the deep neural network(DNN)to construct the HI.Exponential decayed learning rate is proposed to train the neural network efficiently.Aiming at the problem of trend burr in HI,a method of detection and correction of trend burr based on threshold limit is presented.Finally,the effectiveness of the proposed method is verified.(3)In order to solve the problem that the phase degradation of bearing is not considered and the failure threshold is difficult to determine in the existing methods of RUL prediction,a HI construction method of rolling bearings based on deep convolutional neural network with polynomial decaying learning rate(PDCNN)considering phase degradation is given in this chapter.The original vibration signal is input into the deep convolution neural network to construct the HI,and the polynomial decayed learning rate is used to train the neural network efficiently.Considering the phase degradation in the HI construction is more in line with the actual degradation process of bearing.The value of the constructed HI is in the interval [0,1],which solves the problem that the failure threshold is difficult to determine in RUL prediction.(4)Aiming at the problem that the uncertainty of RUL prediction using deep learning methods is difficult to quantify,a deep bidirectional gated recurrent unit(DBi GRU)approach of RUL prediction based on bootstrap method is presented in this chapter.DBi GRU network is constructed by stacking bidirectional gated recurrent unit(GRU)network and deep regression network.DBi GRU model can make full use of the knowledge of past and future degradation state of mechanical equipment and transfer between layers effectively,thereby improving the accuracy of RUL prediction.Bootstrap method can obtain the confidence interval(CI)of the RUL,and then express the uncertainty of the deep learning model effectively.Obtaining the CI representing the uncertainty in the deep learning model provide guidance value for the engineering and actual manufacturing.Finally,the effectiveness of the proposed method is verified.(5)On the basis of the previous chapters,the accelerated life test of rolling bearing is set up.On the premise of not changing the contact fatigue failure mechanism of rolling bearing,the accelerated life test is adopted to shorten the experiment time.The methods proposed in the previous chapters are applied to the accelerated life test experiments of bearings,and the effectiveness of the proposed methods is further verified.