Prediction Of Remaining Useful Life Of Aero-Engine Based On Deep Learning

As the most core component of aircraft,aero-engine plays a crucial role in the overall flight safety.Due to the complexity of aero-engine structure and working environment,performance degradation problem often occurs,which brings potential safety problems.Airlines need to spend a lot of money overhauling them to ensure flight safety.The ac-curate prediction of the remaining useful life of the engine can help to make reasonable maintenance strategy,timely diagnosis and maintenance,and greatly save maintenance cost while ensuring safety.At present,many researchers have studied this problem,and the research on the remaining useful life prediction of aero-engine based on deep learning has received wide attention.However,there are still many problems,including insuffi-cient extraction of temporal features,lack of more effective feature fusion methods,lack of research on the relative information of sensors,lack of thinking on the generalization ability of the model,lack of an online prediction platform.This thesis studies the above problems and puts forward some feasible solutions.The main work of this thesis is as follows:1.A prediction model of aero-engine remaining useful life based on multi-feature fusion is proposed.Firstly,aiming at the interference phenomenon in temporal feature extraction,a dual encoder structure with double step input is proposed to deal with short-term temporal features more efficiently while preserving long-term dependencies.Then,aiming at the problem of lack of effective feature fusion,a temporal feature attention mechanism and a sensor feature attention mechanism are designed to focus on useful in-formation from the time dimension and the sensor dimension,respectively.Finally,in order to solve the problem of missing correlation between sensors,the permutation en-tropy value which can reflect the complexity and changing trend of signals is calculated,and the permutation entropy encoder layer is designed to incorporate the relative informa-tion between sensors.Through the experiments,the effectiveness of the proposed method is proved by comparing with other advanced methods.2.A method for predicting the remaining useful life of aero-engine based on uncer-tainty modeling is proposed.Considering the complexity of engine working environment,aiming at the problem of insufficient generalization ability of the current model,an uncer-tainty modeling module is added on the basis of the proposed model based on multi-feature fusion.During model training,Gaussian distribution is established for feature statistics related to important data distribution by uncertainty estimation,and the uncertainty of po-tential domain offset in test data is simulated by the feature statistics generated by random sampling in the distribution.By incorporating uncertainty into model training,the model’s ability to cope with uncertain scenarios is improved and its robustness is enhanced3.An online system for aero-engine remaining useful life prediction is designed and implemented.The system integrates the proposed prediction model,realizes the core functions related to prediction,including model storage,online training,life prediction,prediction record,and other auxiliary functions,such as data visualization,log audit,user management,and meets the basic requirements of the prediction system in real application scenarios,where safety is guaranteed.The prediction model based on multi-feature fusion and the prediction method based on uncertainty modeling proposed in this thesis have achieved good prediction results on two publicly available aero-engine simulation datasets,proving the effectiveness of the methods.The aero-engine remaining useful life prediction system designed and imple-mented in this thesis is helpful to promote the deployment of the models.