Aero Engine Remaining Useful Life Prediction Based On Multi-variate Data Fusion

The aero engine,with its reliability and stability being guarantees for safe flights,is the core component of an aircraft power system.The degradation phenomenon of aero engines proves to be inevitable as it usually works in high-temperature and high-pressure conditions.In order to avoid security risks and contain maintenance cost as well,it makes great sense to build effective systems in prognostics and heath management.Benefiting from the rapid development of machine learning in recent years,researches about predicting remaining useful life of aero engines with data mining techniques have been in the ascendant.In this thesis,monitored signals of multiple performance parameters during degradation are taken as original data.By combining classic signal processing methods with data-driven algorithms,two methodologies,called ES-XGBoost and ES-GRU-Attn(x)respectively,are proposed.The effectiveness of the different parts in the methodologies are discussed with experiment results.In comparison with other mainstream methodologies about aircraft engine remaining useful life predicion,the advantages of the proposed ones are illustrated.The research contents and results of this thesis mainly include as follows:1.Research of denoising methods on monitered signals.A modified version of exponential smoothing and a singular value decomposition based method are adopted in order to improve data quality.Experiments in later chapters illustrate that exponential smoothing helps to improve predictive ability.2.Research of setting manners of remaining useful life labels.In order to improve the widely adopted piece-wise linear manner,a health indicator curve based setting strategy is attempted.3.Research of predicting remaining useful life of aero engines utilizing XGBoost.In consideration that ensemble models based on regression trees are capable of dealing with high-dimensional feature vectors,XGBoost is adopted to build the predictive model,combined with the denoising methods and label setting strategies.Experiments show that under piece-wise linear labels,the proposed ES-XGBoost methodology results in excellent performance.4.Research of predicting remaining useful life of aero engines utilizing GRU and Attention mechanism.In consideration that classic machine learning methods process data without exploiting its sequential features,GRU based predictive models are proposed.To deal with long-time dependencies loss problem in recurrent neural networks,Attention mechanism is introduced,which is performed in a modified way to adapt to the feature vectors in this thesis.Experiments show that under piece-wise linear labels,the proposed ES-GRU-Attn(x)methodology results in competitive performance.