Aero-Engine Performance Prediction Based On Deep Learning

As the power source of aircraft,the accurate grasp of its health status is of great significance to ensure the safe operation of aircraft and reduce the maintenance cost.Performance prediction is a key part of PHM technology.By predicting performance parameters,it can help pilots and maintenance personnel formulate appropriate flight and maintenance strategies to ensure safety and economy.The aero-engine has a complex structure and rich condition monitoring data,which provides a data foundation and application environment for the implementation of deep learning technology.Therefore,this paper uses deep learning technology to study the performance trend and remaining life prediction of aeroengine based on PHM theory.Firstly,the aero-engine condition monitoring technology is introduced,the EGT parameter in the condition monitoring are analyzed,and the definitions of DEGT parameter and EGTM parameter and their relationship with the aero-engine failure state and the declining state are clarified,and the effects of DEGT and EGTM on the health status of aero-engines were explained,and then DEGT and EGTM were determined as the research objects.Secondly,considering the effect of data quality on the prediction results,the data preprocessing technology is studied,and a Grubbs outlier identification method is proposed to remove outliers in the data.The CEEMD method is used for trend decomposition of the data,and the noise is separated by analyzing the waveform of the autocorrelation function,which provides a data basis for subsequent modeling.Finally,the principles of deep learning technologies CNN,AE,LSTM,and their applicable problems are introduced,and two network models CNN-LSTM and AE-LSTM are proposed.Aiming at the problem of aero-engine performance trend prediction,the research directions of point-to-point prediction,single-step prediction,and multi-step rolling prediction are proposed,and the above two network models are used for prediction.Aiming at the problems of threshold comparison and linear labeling to determine the remaining life,an adaptive interval prediction method based on clustering and a remaining life labeling method based on the decay process are proposed and verified by the CNN-LSTM network model.The comparison with other methods shows that the method proposed in this paper is more accurate in performance prediction and remaining life prediction results.