Multi-stage residual life prediction of aero-engine based on real-time clustering and combined prediction model.

• Multi-stage division method is developed using real-time clustering with sliding window. • The weights of three models are dynamically updated by means of AIC-based model selection method. • R2 and Pearson coefficient are introduced to explain the weight setting and model selection results. • The convergence and generalization of presented model are proved. • Proposed combined model has better prediction performance and interpretability. Due to mass uncertain issues affecting health status of aero-engines, their degradation process commonly exhibits multi-stage features. Also, key characteristics underlying degradation process cannot be precisely described by traditional remaining useful life (RUL) prediction methods. Thus, for multi-stage RUL prediction, we develop a novel real-time combined approach that effectively explains the weights of each base model. The degradation process is divided into multiple stages through real-time clustering. Then, a combined prediction model including Wiener process, LSTM network, and XGBoost is introduced, which can optimally select multiple models for prediction according to AIC. The convergence and generalization of proposed model are proved. Besides, we adopt R2 and Pearson correlation to analyze model selection, further explaining the weights of every base model. The effectiveness of proposed model is validated by comparing with state-of-art methods available. RMSE results of three-stage and four-stage decrease by 2.72% and 2.69% respectively, compared with results of other models. Especially, RMSE of the stage with lowest value reduces by about 4.27%. From 90% whole life analysis, RE results are reduced by 2.77% (three stages) and 2.75% (four stages). Experimental results show that proposed method enhances prediction accuracy and improves model interpretability. [ABSTRACT FROM AUTHOR]