A nonlinear dynamic ensemble remaining useful life prediction method considering multi-source data uncertainty

Abstract

Multi-performance indicator fusion methods have been widely applied for prognostics and health management, but the monitoring indicators are significantly influenced by the internal and external operational environment of the measuring instrument, which creates greater uncertainty to the prediction results. To address this issue, a nonlinear dynamic ensemble remaining useful life (RUL) prediction framework considering multi-source data uncertainty is proposed in this paper. Firstly, a multi-performance indicator fusion method considering data uncertainty is proposed. This method explicates the multi-indicator data as various proxies of the degradation state of equipment by establishing a multivariate implicit nonlinear state function and a multivariate measurement function, and constructs an optimal fusion strategy by designing a new objective function related to the performance of the indicator. Besides, a multi-model nonlinear dynamic ensemble method is proposed to compensate for the inadequacy of a single model to accurately characterize the degradation trajectory by integrating the prediction results of different degradation models in real-time. Finally, a Likal-Recursive algorithm is developed to address the challenge of estimating latent variables in multivariate state-space models without relying on initial parameter assumptions. The superior performance and effectiveness of the proposed framework are validated using the C-MAPSS dataset and multi-sensor datasets of rolling bearings.