Research On Data-driven Remaining Useful Life Prediction Of Turbofan Engine

With the continuous improvement of the level of industrialization,the structure of industrial equipment has become more complex.Aero-engines are exposed to harsh environments of high temperature and high pressure for a long time,and they are prone to performance degradation and even failure.Therefore,Remaining Useful Life(RUL)of aero-engines,as the core of Prognostics and Health Management(PHM),directly affects the decision-making of aero-engine maintenance or failure alarms,which has research significance.Based on the data-driven method,this paper uses the turbofan engine performance degradation data provided by NASA as the benchmark data to carry out the remaining life prediction research of the turbofan engine.The main research contents include:(1)Aiming at the high dimensionality of turbofan engine performance degradation data and the time-series characteristics of the data,a model for predicting the remaining life of the turbofan engine based on a long and short-term memory(LSTM)network is constructed.Based on the C-MAPSS public data set,the prediction performance of the established life prediction model is evaluated through two evaluation indicators: Root Mean Squared Error(RMSE)and Score function.And further based on the experimental results to optimize the network parameters of the remaining life prediction model based on LSTM.(2)Aiming at the problem that many noises in the turbofan engine degradation data affect the prediction accuracy,a neural network improvement model based on Bi LSTM-NN is constructed on the basis of the LSTM-NN prediction model.The model mainly improves the remaining life prediction model based on LSTM-NN from three aspects,including data preprocessing,suppression of overfitting and network training model.In addition,in order to enable the objective function of the improved Bi LSTM-NN model to reflect the actual physical situation where the predicted fault occurrence time point is later than the actual fault occurrence point,there will be serious consequences,this paper sets the network objective function as a Score custom function.Finally,the experimental results show that the prediction accuracy of the improved model is significantly improved.(3)Aiming at the limitation of a single network,a fusion model based on recurrent neural network and convolutional neural network is proposed.First,dilated causal convolution,self-normalized Scaled Exponential Linear Units(SELU)activation function and Depthwise Separable Convolution(DSC)Neural Network are used to construct a CNN-SCNN model.Secondly,the Bi LSTM-NN model and the CNN-SCNN model are merged in two different ways: sequential stacking and dual-channel parallelism.Finally,the experimental results show that the fusion model has higher prediction accuracy than a single prediction model.