Study On Aeroengine Gas Parameters Predction Based On Ensemble Learning

Aeroengine is a high-speed rotating equipment with high complexity and poor operating environment and is relevant to the safety of the aircraft.Its reliability has a significant impact on the safety of the aircraft.The engine failure accounts for a large proportion of the total aircraft failure.The International Air Transport Association(IATA)pointed out that in 2014 the engine maintenance costs accounted for 40% of the total cost of aircraft MRO(Maintenance,Repair and Overhaul),reaching $24.84 billion.Aeroengine performance parameters prediction is of great significance to avoid the occurrence of failure,and is useful for future CBM(Condition Based Maintenance).Consequently,preprocessing method,point prediction method based on ensemble learning and the method for interval forecasting in gas path parameters prediction filed are studied systematically in this paper.This paper also develops some relevant tools.The validity of the methods and the tools has been verified by applying them to the real industrial exhaust gas temperature margin(EGTM)series.Aiming at the problem of the existence of gross error and noise in the original parameters,the preprocessing method for the gas path parameters has been studied in this paper.A preprocessing method based on empirical mode decomposition(EMD)and subfilter weighted FIR median hybrid(SWFMH)is proposed.Firstly,a modified method of extending data which is based on nearest similarity distance is used to solve the end effect of the EMD algorithm.Secondly,the EMD algorithm is used to extract the trend component of the data.Then the trend component is carried out from data.According to the detrended data,the Pauta criterion is used to detect data points with gross errors and then the gross errors are removed.Then EMD algorithm is used to find the noise modes of the data without gross errors,and SWFMH is applied to these noise modes to denoise.Finally,some modes are chosen to reconstruct to get the preprocessed data.Based on this method,a tool with preprocessing capability is developed.The validity of this method and the tool has been verified by applying them to the simulation series and the real industrial EGTM series.Aiming at the problem that single global modeling methods cannot get satisfactory point prediction results when predicting gas path parameters,the point prediction method based on ensemble learning for the gas path parameters has been studied in this paper.In this paper,a combination method called dynamic weighted kernel density estimation(DWKDE)is proposed.Firstly,neighboring samples of the test samples are chosen.Secondly,the weights of the base learners are dynamically determined by evaluating the base learners' local performance in the neighboring samples.Finally,the integrated prediction of the time series is realized by using the weighted kernel density estimation based on the weights.The proposed combination method is insensitive to outliers and deviations from normality.Apply this combination method to Ada Boost.RT and Ada Boost.R2 and experiments that have been conducted on benchmark data and EGTM series verify that the proposed algorithms can achieve higher prediction accuracy than the single neural network and the traditional ensemble learning algorithms.This paper also studies how the number of base learners influences the accuracies of several ensemble learning algorithms.A tool with point forecasting capability is developed.Aiming at the problem that the point prediction results cannot characterize the uncertainties of the parameters and the potential risks of engine operation,the interval forecasting method for the gas path parameters has been studied in this paper.In this paper,a neural network-based adaptive interval forecasting(NNBAIF)model is proposed to predict the fluctuation range of gas path parameters under a certain confidence level.The trend component and historical data are used as the input of neural network.The control parameters are optimized by the harmony search algorithm when constructing the output set.Then according to the parameters,the output set of neural network is constructed.The input set and the output set are used to train the neural network.This model enables engineers to get the prediction intervals(PIs)without point prediction results.Based on the proposed model,a tool with interval forecasting capability is developed.The experimental results show that the PI coverage probability(PICP)and mean PI width(MPIW)obtained by using the model are superior to another method of a literature.This paper can enrich the methodology of time series prediction and the prediction of aeroengine health parameters(including point prediction and interval forecasting).In addition,the methods in this paper can improve the prediction accuracy of gas path parameters,improve the safety of aircraft and reduce operation and maintenance cost.