Remaining Useful Life Prediction Based On Multi Representation Domain Adaptation

With the development of industrial intelligence,the prediction of Remaining Useful Life(RUL)is gradually combined with artificial intelligence technology.Deep learning is widely used in RUL prediction because it does not require prior knowledge and has strong nonlinear fitting ability.However,the current life prediction methods have some shortcomings.First,RUL prediction methods based on deep learning mostly assume that training data and test data have similar distribution.Due to the existence of various operating conditions,fault modes and noise,there are not enough data with similar distribution in the training process,thus reducing the performance of RUL prediction.Second,most of the existing prediction methods based on deep learning are point prediction,which is included in RUL prediction.However,practical engineering applications not only have high requirements on the prediction accuracy of the model,but also expect the model to provide uncertainty of the prediction results,such as confidence interval.The research of confidence interval of RUL prediction is more important for reliability-based maintenance strategy.This thesis presents two distinct network frames,taking these two issues as the primary research topic.In this thesis,a data-driven long and short term memory network and multi-representation domain adaptive(LSTM-MRAN)method are proposed for data prediction under varying working conditions and different distribution conditions.The sliding time window is used to standardize and process the degraded sensor data,and LSTM is used to extract features from the degraded data and mine the time series information between the data.Then,the multi-representation domain adaptive and conditional maximum mean difference loss function is used to reduce the distribution difference between the data.The effectiveness of this method is verified by CMAPSS data set.Compared with non-domain adaptive method and other transfer learning methods,this method provides more reliable RUL prediction results under different working conditions and failure mode data sets.Aiming at the uncertainty of RUL prediction,an interval prediction model based on long and short term memory and lower upper bound estimation(LSTM-LUBE)is proposed for RUL prediction.Firstly,the model uses convolutional self-coding network to encode the multi-dimensional variables collected by multi-sensors into one-dimensional features,which can well characterize the degradation trend and serve as the basic degradation features predicted by RUL.Then,the features were input into the prediction framework composed of LSTM and LUBE for RUL interval prediction.LSTM was used to replace the neural network structure in LUBE,which effectively solved the defect that traditional LUBE network could not analyze the internal time dependence of time series,and the improved LUBE was more suitable for interval prediction analysis of degraded data.In order to verify the validity of the proposed model,the turbofan engine dataset CMAPSS is used for experiments and compared with other interval prediction methods.The results show that the LSTM-LUBE network model further improves the accuracy of RUL interval prediction.