ABSTRACT Reliability assessment of complex multi‐state systems (MSSs) is essential for their safe and efficient operation. Survival signature, a powerful tool for reliability analysis, faces significant computational challenges when applied to MSSs due to combinatorial explosion. However, research on efficient computation of survival signature for MSSs remains scarce and challenging. To address this issue, this study proposes a graph neural network (GNN)‐based approach for predicting survival signatures with improved computational efficiency, which integrates both the topological structure and component state information of MSSs. The proposed method utilizes the graph attention network v2 (GATv2) to dynamically aggregate node features through learnable attention weights. Furthermore, it incorporates the jumping knowledge (JK) framework to adaptively integrate multi‐scale features across different network layers, thereby mitigating over‐smoothing and enhancing hierarchical feature extraction. The numerical example and the application example of dual‐axis positioning mechanisms for satellite antennas demonstrate that, compared with traditional methods such as Monte Carlo simulation (MCS) or enumeration, the proposed approach not only ensures high prediction accuracy and significantly reduces the computational cost of survival signature evaluation, but also efficiently performs multi‐state Birnbaum importance analysis. It provides an effective tool for reliability analysis and optimization of MSSs.
Li et al. (Tue,) studied this question.