Research On Data-driven Prediction Methods For Remaining Useful Life Of Aero-engine

Aero-engine Prognostics and Health Management(PHM)is an advanced technology under development in recent years,which is an important means to promote the reform of maintenance support and improve the safety,reliability and affordability of aircraft engine.Prognostics,as the core technology of PHM and also the most difficult and challenging integrated technology,is aimed at predicting the failure and Remaining Useful Life(RUL)of components or the system,further providing support for the operation planning and maintenance decision.Data-driven RUL prediction mainly concentrates on learning and describing the system degradation modes,and predicting the fault or damage propagation path and performance degradation trends according to condition monitoring data using artificial intelligence and case-based reasoning approaches etc.so as to perform RUL assessment.Data-driven approaches have strong practicability since there are no needs of understanding the details of equipment and the complex failure mechanism deeply.On the basis of studying PHM concepts and architecture,comparing and analyzing the advantages and shortcomings of various RUL prediction approaches,this paper concentrates on the research of data-driven RUL prediction method,aiming at providing technical reserves for the development of aero-engine PHM system.The main work and innovations are as follows:(1)RUL prediction-oriented health assessment for aero engine.The health assessment process for complex equipment with multiple operation regimes is studied,including sensor selection,operation regimes partition and identification,health assessment model construction,etc.Then a sensor selection method based on information entropy theory is proposed with the aim to choose the proper and available condition monitoring sensors for RUL prediction.Besides,Kernel Principle Component Analysis(KPCA)and quadratic programming are adopted to perform data fusion for the selected sensors,thus composite Health Index(HI)that can represent the degradation of the system is constructed.The HI is evaluated based on aero engine run-to-failure simulation data,which shows that HI is more suitable for the modeling and analysis of data-driven RUL prediction,as HI can better reflect the system performance degradation trends and the failure HI values of all samples are less diverse.(2)RUL prediction method based on degradation trajectory similarity.The principle and framework for Similarity-Based Prognosis(SBP)method is presented.Then the major procedures,such as degradation extraction,similarity evaluation,RUL estimation aggregation are illustrated.To tackle the problem of the existing method that the RUL prediction may produce a relatively larger error and uncertainty when the system is far from failure,a new method which combines Degradation Trajectory Extrapolation and Similarity Based Prognosis(TE-SBP)is proposed.Firstly the degradation trajectory is extrapolated using sate space model for system performance degradation,and then RUL prediction is achieved through SBP.The effectiveness and advantage of TE-SBP method is validated with a case study.(3)RUL prediction method using relevance vector machine.A Wavelet Relevance Vector Machine(WRVM)is conceived via the combination of wavelet theory and RVM.Moreover,the parameters of WRVM are optimized by Particle Swarm Optimization(PSO)algorithm and the modified fitness function is utilized to avoid the neglect of algorithm structural risks.On the basis of this,inspired by pattern recognition,a novel direct RUL prediction method based on WRVM is proposed to overcome the shortcomings of multi-steps forward prediction method and the direct mapping from condition monitoring data to RUL.The proposed method is assessed with a case study,which demonstrates that the RUL prediction method based on WRVM is superior to existing frequently used RUL prediction strategy as it can not only get rid of the difficulties in failure threshold selection and error accumulation in multi-steps forward prediction,but also describe the characteristics of degradation modes better than the direct mapping between the monitoring data and RUL.(4)RUL prediction method based on stochastic process.In order to express the stochasticity in engine degradation process,a RUL prediction method based on stochastic process is employed.A general stochastic process-based RUL prediction frame is developed,and then the realization of the important procedures,such as model priori hypothesis,parameters update via Bayesian rule,model hyper-parameters estimation using Expectation-Maximization(EM)algorithm and the update of RUL distribution,are presented taking the exponential-like model with description of stochastic dynamics as an example.With the aim to reduce the influence on prediction accuracy caused by the random errors of stochastic process,an improved Particle Filter based on Electromagnetism-like Mechanism(EMPF)is used to estimate actual system health state.Finally the effectiveness of the proposed method is verified through a case study.(5)The fusion of various RUL prediction methods.In order to improve the robustness of RUL prediction and take advantages of different methods,the fusion of various approaches is conducted.Firstly the fusing framework,process and common fusing methods are presented.On the basis of this,a Kalman filter-based fusing method is focused on and applied for fusion.The results of case study show that the fusion of various RUL prediction methods prevails over any single prediction method for higher accuracy and better robustness.In addition,the fusing method based on Kalman filter can achieve better fusion performance compared to the information entropy-based fusing method and simple weighted average method.