Causal graph-based spatial-temporal attention network for RUL prediction of complex systems

作     者：Zheng, Shuwen Liu, Jie Chen, Yunxia Fan, Yu Xu, Dan 

作者机构：Beihang Univ Sch Reliabil & Syst Engn Beijing Peoples R China Beihang Univ Sch Energy & Power Engn Beijing Peoples R China 

出 版 物：《COMPUTERS & INDUSTRIAL ENGINEERING》 (Comput Ind Eng)

年 卷 期：2025年第201卷

核心收录：

学科分类：1201[管理学-管理科学与工程(可授管理学、工学学位)] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Aeronautical Science Foundation of China 

主　　题：Remaining useful life Spatial-temporal features Causality Complex systems Prognostics 

摘      要：With diverse sensors in complex systems, Remaining Useful Life (RUL) prediction involves fusing multivariate time-series data. Incorporating features from spatial-and-temporal dimensions is a trending and challenging topic. Existing state-of-the-art researches typically fuse spatial-temporal features according to the correlation analysis, limiting their ability to capture representative features for RUL prediction. Moreover, the dynamic feature importance in system degradation is often ignored. To address the above limitations, this work proposes a Causal graph-based Spatial-Temporal Attention Network (CaSTAN) for RUL prediction of complex systems. Prior knowledge is first combined with causal discovery algorithms to explore cause-effect relationships among variables. Subsequently, strategies are introduced for adaptively extracting the spatial, temporal and causal features during the degradation process. The extracted representations are then aggregated using a gating mechanism, allowing to selectively fuse features for RUL prediction. Case studies concerning aircraft turbofan engines are conducted, along with extensive comparative and experimental analyses. The results underscore the method s effectiveness in enhancing predictive maintenance of complex systems.