Research On Bearing Health Index Construction And Remaining Useful Life Prediction Based On Deep Learning

With the development of information and sensor technology,industry has entered the 4.0 era.How to use a variety of sensors to collect massive data with low value density in real time has become an urgent problem to be solved.One of the most valuable research directions is to predict the remaining useful life of mechanical equipment in real time according to the data signal of mechanical equipment,so as to provide reference for condition based maintenance of mechanical equipment.At present,there are some problems in the research methods of health index（HI）construction and life prediction,such as relying on expert knowledge,difficult to analyze complex mechanical systems,and unable to predict mechanical equipment in real time and accurately.When the research method based on deep learning has made great progress in other fields,it is urgent to use advanced models to monitor the health index status of mechanical equipment and effectively predict RUL.After in-depth study of the existing literature,aiming at the problems existing in the traditional methods,combined with the deep learning method,taking the bearing as the research object,this thesis studies the construction of health index and the prediction of remaining useful life.The main research contents are as follows:（1）Aiming at the problems of relying on expert knowledge,poor generalization and redundancy of features in the current construction methods,a construction method of bearing degraded health index based on the combined model of stacked denoised autoencoder（SDAE）and self-organizing map（SOM）is proposed.The original data is denoised by SDAE and multi-dimensional features are extracted adaptively,and then the multi-dimensional features are reduced by SOM.Compared with the HI curve constructed by various methods under different working conditions on PHM2012,it has better performance in the proposed evaluation indicators.（2）In order to overcome the problem of excessive prediction error caused by the lack of adjacent real value in long-term prediction.A bearing life prediction method based on multistep long short term memory network（MS-LSTM）is proposed.The proposed method optimizes the input-output structure of the original LSTM by adding the single-step prediction value to the input time step,and uses the constructed HI value as the prediction training and test data to predict the remaining useful life of the bearing,which speeds up the training speed and greatly reduces the parameter adjustment difficulty.The comparative experiment on PHM2012 data set shows that the optimized model significantly improves the prediction accuracy in long-term prediction.（3）The key equipment health management system is designed and implemented,and the two methods mentioned above are combined and applied in the system.The two methods mentioned above are combined and applied in the system.The health status is monitored by using the hi curve of the fan and motor,and then its life is predicted by using the deployed MS-LSTM model.The above research results show that the whole research method based on deep learning proposed in this thesis has certain advantages over the existing methods in characterizing the health state and life prediction of bearings,and has good universality in other complex mechanical equipment.