Remaining useful life prediction of mechanical equipment based on attention kernel dynamic time warping

Remaining useful life (RUL) prediction method based on similarity measurement has the advantages of strong interpretability and a simple implementation process. However, existing similarity measure methods are limited by linear distance measurement, which fundamentally restricts their ability to characterize nonlinear degradation dynamics. Furthermore, the equal weighting of all time steps neglects the higher predictive value of pre-failure data, thereby reducing the prediction accuracy. To overcome this problem, a RUL prediction method based on attention kernel dynamic time warping (AKDTW) is proposed. The significance of this study lies in the fact that by constructing the AKDTW distance, the health indicator is mapped to a high-dimensional Hilbert space, thereby effectively capturing the nonlinear characteristics of equipment degradation. Furthermore, by designing attention weights to act on the kernel distance, the nonlinear similarity contribution of critical time steps is amplified, further focusing on degradation information in the pre-failure stage. The verification results on the C-MAPSS engine dataset show that the RUL prediction results of this method are better than traditional linear distance measurement methods and some deep learning network methods, indicating that the proposed method can provide reliable technical support for mechanical equipment health management.