Research On The Prediction Methods Of Remaining Useful Life Of Industrial Equipment Based On Deep Learning

Complex and high-precision equipment malfunctioning due to degradation and failure can cause significant losses.Remaining useful life of industrial equipment is an important part of prognostics and health management.The remaining useful life is predicted based on the operating status data of the equipment,which is used to formulate a maintenance plan,which can make it operate safely and reliably,saving a lot of manpower and financial costs.The remaining useful life prediction has important application significance in predictive maintenance in the fields of aerospace,wind power and high-speed trains.The application of deep learning methods in the field of remaining useful life prediction has made some progress,but there is still a lot of room for optimization.In this dissertation,aiming at the problems of the remaining useful life prediction of the recurrent neural network has limitations,the equipment operating data distribution is different under different working conditions,and there are unlabeled data,the following work is mainly done:(1)In order to solve the problem that the recurrent neural network cannot solve the longterm dependence and treat the data of each dimension of the input multidimensional time series equally,this dissertation adds a feature attention mechanism on the basis of the Long Short Term Memory,and builds a feature-based selection of remaining useful life prediction model(FA-LSTM)fused with attention mechanism.FA-LSTM uses a three-layer fully connected layer to implement a feature attention mechanism,which can give different weights to different input sensor data in multi-dimensional sensor monitoring data during model training,so that the model has the ability to select features.It was verified on the aviation turbofan engine data set,and compared with FA-LSTM,LSTM,convolutional neural network,gradient lifting decision tree and other models.Experimental results show that the FA-LSTM model can effectively improve the prediction effect of remaining useful life,which is 3.76 less than the average root mean square error of the LSTM model.Finally,the effect of different time window sizes on the model performance in FA-LSTM is verified experimentally.(2)Aiming at the problems of different probability distribution of sensor monitoring data and the existence of a large amount of unlabeled data in industrial equipment running under different working conditions,this dissertation introduces the domain adaptation method of transfer learning in remaining useful life prediction.Combining the LSTM model and the Wasserstein distance on the basis of the twin network,the remaining life prediction is achieved by means of adversarial training,and a remaining useful life prediction model(DA-LSTM)based on domain adaption and twin network is proposed.The DA-LSTM model uses labeled source domain data and unlabeled target domain data to jointly train the network,narrows the distribution distance of the source domain data features and the target domain data features at the feature extraction layer,and is used for remaining useful life prediction in the target domain data.Verification was performed among four subsets with different data distributions in the aviation turbofan engine data set.The experimental results show that the DA-LSTM model can reduce the error of remaining useful life prediction in the target domain.The mean of root mean square error of the DA-LSTM model is 1.87 less than that of the long-short-term memory model trained using only source domain data.In summary,this dissertation has made some explorations on the application of deep learning in remaining useful life prediction and transfer learning in remaining useful life prediction,and proposed two models which achieved good results of remaining useful life,which has certain reference significance.