A generalization enhanced spatial-time-global competitive network for remaining useful life prediction

Abstract The prediction of degradation and remaining useful life (RUL) for mechanical equipment has received widespread attention. Due to the large number of mechanical equipment components and varying operating conditions, it is difficult to extract degradation features. This paper proposes a generalized enhanced spatial-time-global (ESTG) competitive network for predicting RUL. The model mainly includes spatial competition (SC) module, temporal competition (TC) module, and global competition (GC) module. The SC module focuses on feature mining of multi-source data, which can sensitively capture key information from different spatial locations; the TC module adaptively adjusts the threshold based on the learning ability of the model itself, constructs threshold-enhanced recognition boundaries, and eliminates the influence of irrelevant noise; the GC module introduces a feature matrix as a scaling factor to adjust key information dynamically. In addition, interpretability analysis was also employed to validate the proposed method’s capability in capturing RUL feature representation trends. Finally, the effectiveness of ESTG competitive network is validated by the XJTU-SY and PHM bearing datasets, along with a rail transit pantograph carbon sliding plate wear dataset. The experimental results show that the proposed method achieved an average 44% reduction in root mean square error.