Research On Remaining Useful Life Prediction Method Of Bearings Based On Improved BiLSTM

Rolling bearing as a basic component plays an important role in rotating machinery,which is widely used in a variety of rotating equipment.Bearing failure is one of the key factors leading to machine shutdown.Therefore,in the industrial field,the remaining useful life prediction of rolling bearing is very important.With the rapid development of artificial intelligence and Internet of Things technology,data processing technology and sensor technology have been significantly enhanced.Human beings have entered a new era of industrial big data.The data-driven residual life prediction method of bearings has made great progress in the application research of industrial big data environment.The research work of this paper mainly focuses on two main steps of data-driven residual life prediction(the construction of health indicators and the construction of life prediction model).The traditional data-driven bearing health index construction method still needs some prior knowledge,such as feature index selection,health index construction,failure threshold setting and so on.The prediction results rely heavily on expert experience.In order to solve this problem,a health index construction model based on Deep Residual ShrinkageNetwork(DRSN)is proposed based on deep learning method.The attention mechanism and soft threshold structure of DRSN can effectively eliminate the influence of noise-related characteristics,so as to construct the health index of bearings.The performance of the proposed bearing health index construction method is verified by a large number of comparative experiments using the PHM2012 challenge data set.The method for constructing the bearing health index is analyzed and compared with the residual ShrinkageNetwork(ResNet)and the artificially extracted feature mean.Among them,the performance of the DRSN-based bearing health index construction method is better.Data-driven method is divided into statistical method and deep learning method.In this paper,two kinds of rolling bearing life prediction methods are proposed for these two fields.One method is to use Holt double exponential model to establish prediction model to predict the remaining life of the bearing at the moment.Another method is to use Bidirectional Long-Short-Term MemoryNetwork(BiLSTM)in deep learning to build bearing remaining life prediction model.The two methods are :(1)Ordinary double exponential prediction model needs to adjust four hyperparameters,resulting in low prediction accuracy.In view of the above problems,a prediction method of residual life of rolling bearings based on Holt double exponential model is proposed.Holt double exponential model can be used to predict time series data with trend,and the sparrow search algorithm is used to update the two parameters that need to be adjusted.The results show that this method can effectively reduce the prediction error of bearing residual life.Compared with the double exponential prediction model,the accuracy of this method is improved by 8.4%.(2)The remaining life prediction model of bearing is constructed by BiLSTM network.Aiming at the problem that the number of neurons and the learning rate of BiLSTM hidden layer are not easy to set,the sparrow search algorithm is used to optimize the parameters.The input of the bearing residual prediction model is the health index constructed by DRSN.After smoothing the health index extracted by DRSN,the normalized life is taken as the label and input into the optimized BiLSTM prediction model to complete the prediction of the bearing residual life.The experimental results show that the error of the optimized BiLSTM bearing residual life prediction model is the smallest.The root mean square errors of the three bearing residual life prediction models based on the optimized BiLSTM,BiLSTM and Long-Short-Term MemoryNetwork(LSTM)are 1.41%,2.71% and 5.64%,respectively,which verify the effectiveness of the method.