Research On Data-driven Aero-engine RUL Prediction Model

Prognostics health management(PHM)of aero-engine is an important means to improve engine safety and reduce engine maintenance costs.It is an important strategy for intelligent engine maintenance in recent years.The prediction of the remaining useful life(RUL)of aero-engine is one of the important issues in PHM-related research.Through accurate RUL prediction,a corresponding maintenance plan can be formulated to avoid sudden engine failures and unpredictable losses.Based on the analysis of the typical data-driven RUL prediction methods proposed in recent years,this paper aims at the problem of traditional neural networks that are difficult to capture multi-scale information and the problem of low RUL prediction accuracy of aero-engine under multiple operating conditions.Two RUL prediction models are proposed: MSTCN-FS and LSTM-ATT.The main work contents are as follows:(1)In the RUL prediction work,it is difficult for traditional neural networks to capture information of different scales,and the extracted features cannot simultaneously reflect the degradation trend of a single component and the entire engine system.In response to this problem,this paper proposes a multi-scale temporal convolutional network fusion model(MSTCN-FS).The model first separately trains multiple time series convolutional networks with the same network structure and different sliding windows to extract the time-series features of the sensors,and at the same time captures the local information and global information of different scales in the sequence to obtain different potential features;At the same time,considering the importance of manual features,the extracted manual features and multi-scale information are merged to obtain a complete feature set,and finally the aircraft engine RUL prediction is realized in the regression layer.(2)Aero-engine has different degradation trends under multiple operating conditions,which will affect the engine’s RUL,but many studies have not considered this situation,resulting in low engine RUL prediction accuracy.In response to this problem,this paper proposes an RUL prediction model(LSTM-ATT)based on an improved LSTM network.The model first uses the LSTM network to learn representative sequence features from the original sensor data;Then,uses the attention mechanism to learn the importance of sensor features and time steps,and assigns greater weights to more important features and time steps;Finally,feature fusion of automatically extracted features and multi-condition parameters,to achieve accurate prediction of RUL.The two models proposed in this article were tested on two data subsets of the C-MAPSS dataset,respectively,and compared with other advanced RUL prediction models.Compared with traditional machine learning models such as gradient boosting decision tree,random forest,and so on,and deep learning models such as bidirectional long short-term memory networks,deep convolutional neural networks,and deep convolutional generative adversarial networks,the two models proposed in this article have the best prediction performance under different evaluation indicators.Experiments show that the two models proposed in this paper are more effective in predicting the RUL and provide more reliable support for the efficient maintenance strategy of the aero-engine.