Research On The Exhaust Gas Temperature Prediction Method Of Aviation Piston Engine

Accurate monitoring and prediction of aeroengine performance is the core content of aeroengine health management system.The exhaust gas temperature of aeroengine is a key index for airlines to maintain their engines.The variation trend of exhaust gas temperature is an objective reflection of engine performance degradation and service life during flight.Therefore,predicting the trend of exhaust gas temperature is helpful for making maintenance plans,predicting the remaining life of aircraft engines,and reducing maintenance costs to reduce aircraft engine failures.The prediction method based on physical model requires a complete engine component feature map and the complex modeling process limits the computational performance and applicability.It is an effective method to predict exhaust gas temperature by data mining of historical flight parameters collected by airborne sensors.As a powerful nonlinear model,artificial neural network has been widely used in engine state parameter prediction.In this thesis,the prediction method of exhaust gas temperature is studied in depth from the following aspects:(1)Firstly,the prediction methods of aeroengine exhaust gas temperature at home and abroad are reviewed and summarized,and the classical time series prediction models are briefly summarized,including their prediction principles,advantages and limitations.The evaluation indexes are also summarized,including RMSE,MAPE and MAE,and their mathematical descriptions are given to lay a foundation for further research.(2)Data sources and datasets are described.The types of missing values and the methods of filling in missing values are introduced.Based on these methods,the filling results of partial measured data are given,and multiple average interpolation technique is proposed to fill in missing values.The types and detection methods of outliers are summarized,and the standardized transformation of data is introduced.This thesis introduces the feature selection method and its implementation process,summarizes their advantages and disadvantages,and carries out feature selection based on correlation analysis.(3)The prediction problem is described in formulaic and the difference between univariate time series and multivariate time series data is discussed.In addition,the spatial feature extraction network and temporal feature extraction network are introduced,and the differences and connections among RNN,LSTM and GRU networks are summarized.On this basis,a spatio-temporal feature fusion network STFFNet is proposed,and the network architecture is given.(4)This thesis introduces the overview of ensemble learning,which can be divided into boosting,bagging and stacking method.The three methods are introduced in detail.Furthermore,an EGT prediction scheme based on ensemble learning is proposed,which is described in detail,and the flow of the optimal weight solution method is given.