Hierarchical Trustworthiness Assessment and Operational State Estimation of Multi-Agent Systems in Smart Manufacturing

Abstract

Smart manufacturing increasingly relies on multi-agent systems (MASs) to coordinate distributed sensing, control, quality analysis, maintenance, and resource allocation. As agent autonomy increases, system risks can no longer be captured by a single accuracy metric; local misjudgment, coordination mismatch, fault propagation, privilege misuse, and poor traceability must be considered together. This paper develops a hierarchical trustworthiness assessment framework with six dimensions: functional reliability, interaction credibility, robustness and resilience, operational safety, cyber and data security, and transparency and accountability. For operational evidence extraction, a three-branch Fusion ST-GAT model combines current-state features, temporal dynamics, and process-topology-aware graph representations. Experiments are conducted on the Tennessee Eastman Process benchmark using a chronological training/validation split before sliding-window generation, three random seeds, and Platt scaling calibration. The Fusion ST-GAT achieves an AUC of 0.9898±0.0031 and an F1-score of 0.9806±0.0032, whereas the MLP baseline reaches 0.9935±0.0015 and 0.9857±0.0020, indicating that higher model complexity does not automatically produce better overall performance on this benchmark. Under combined Gaussian noise and 10% element-wise missingness, the Fusion ST-GAT retains an F1-score of $0.9776 \pm 0.0030$. Agent-occlusion analysis for IDV(14) shows the strongest sensitivity to reactor-node inputs. The results support operational-state estimation, but they do not replace governance evidence such as access control, audit records, or accountability allocation.