Research On Data-driven Remaining Useful Life Prediction Method Based On Deep Learning And Uncertainty Quantification

With the rapid development of science and technology,the complexity,precision and intelligence of equipment are gradually increasing.Traditional maintenance strategies have been insufficient to meet modern maintenance due to its lag in operation.Prognosis and health management(PHM)technology came into being in this context.As the core of PHM,the remaining useful life(RUL)prediction technology directly determines the accuracy and timeliness of system's failure warning and maintenance decision,it has important research value.Among the remaining useful life prediction methods,data-driven RUL prediction technology has become a hot research topic because it does not rely on the domain knowledge of equipment.Although the existing data-driven RUL prediction methods have achieved good prediction accuracy,it ignore the uncertainty quantification of the prediction model and impossible to obtain confidence intervals.Therefore,based on the data-driven remaining useful life prediction technology,combining deep learning and uncertainty quantification method,and considering the complex operating conditions of equipment,the research contents of the remaining useful life prediction in this thesis are as follows:1.Study the application of deep learning technology in the prediction of remaining useful life,proposing a remaining useful life prediction method based on Autoencoder-TCN network,which combining the timing feature extraction of temporal convolutional network(TCN)and the dimensionality reduction ability of autoencoder(AE).Improving the accuracy of RUL prediction under complex operating conditions by K-means clustering and data enhancement methods.Finally,the experiment was conducted on the public data set of C-MAPSS,which proved that the Autoencoder-TCN network has higher accuracy than other models under complex working conditions.2.In view of the situation that traditional deep learning methods cannot quantify uncertainty,introducing the accidental uncertainty and cognitive uncertainty by application of bayesian neural network.And proposing an uncertainty modeling method based on bayesian long short term memory(BayesLSTM)network,which can quantify the uncertainty and obtain the confidence interval of the RUL prediction result and its RUL prediction results is in the form of probability distribution.Finally,using The PHM08 and C-MAPSS data sets analyzing the RUL prediction results and the changing trend of uncertainty,which proved the effectiveness of the method.3.Aiming at the problem of the large amount of calculation in bayesian neural network,proposing a RUL prediction method based on approximate bayesian inference.By introducing Monte Carlo(MC)dropout,the ordinary neural network can also output the probability distribution of RUL prediction and get confidence interval.Finally,the experiment was conducted on the C-MAPSS data set,which verified the superiority of the method.