Remaining Useful Life Prediction for Aero-engines based on Improved Dynamic Ensemble Learning

The data collected by various sensors in monitoring the operating status of aero-engines can be used to predict the Remaining Useful Life (RUL) of aero-engines. This dataset has characterisitcs of high dimensions and large scale, which increase the difficulty of accurately predicting RUL. To obtain more accurate prediction results, this paper proposes a prediction model based on dynamic ensemble learning to predict RUL of aero-engines. The model selects the K nearest neighbor samples of one testing sample, dynamically determines the weight of each learner by evaluating the local performance of this learner in the neighbor samples, and constructs a weighted kernel density estimation function based on previously calculated weights to achieve integrated prediction of multiple base learners dynamically. In order to better determine the similarity between the data, an improved adaptive KNN (K-Nearest Neighbor) algorithm is introduced, and the importance of each sensor is introduced into the traditional distance measurement, and the adaptive K value selection is realized through the relationship between the global average density and the local density. In order to reflect the short-term and long-term dependencies between samples in dataset better, neural network LSTM (Long Short-Term Memory) is selected as the base learner of the dynamic ensemble learning model. Finally, the aircraft engine simulation data set C-MAPSS released by NASA is used for simulation verification. The experimental results show that the model proposed in this paper can improve the forecast precision of aero-engines’ RUL.