Multiscale similarity ensemble framework for remaining useful life prediction

Accurate prediction of remaining useful life (RUL) is crucially important to perform prognostics and health management. A new similarity-based autoencoder multiscale ensemble (Similarity-based AE MSEN) methodology is proposed in this paper to improve RUL prediction accuracy and characterize RUL prediction uncertainty by considering the differences in equipment degradation rates, monitoring data lengths and fault modes. Firstly, multiscale sliding window sets are designed to divide health index (HI) curves generated by autoencoder into multiscale HI segments, which are used to measure the similarities between training and testing units. Then, multiscale prediction results obtained from similar training units are fused by kernel density estimation to fit a RUL distribution and then provide uncertainty for RUL prediction. The proposed multiscale ensemble strategy can overcome accuracy limitation caused by a fixed time scale and enhance generalization ability. Analysis of e xperimental datasets shows that the proposed multiscale method achieves state-of -the-art prediction performance.