Complex engineered systems often face rapidly evolving risks and inefficient coordination among operational roles. To address these challenges, we propose a safety decision support framework based on a heterogeneous graph neural network. Multi-source sensor time series and role semantics are unified in a heterogeneous graph, which explicitly represents physical interactions, responsibility links, and coordination constraints. A two-level attention mechanism is introduced to fuse multi relational information, and a dual task design jointly supports multi step sensor forecasting and role activation propensity estimation, enabling risk awareness and response planning in a single inference workflow. The approach is validated on a steel continuous casting case study with 10 sensors and 6 roles. Results show low one step forecasting errors, with channel wise RMSE and MAE on the order of 10⁻². Multi step forecasting remains stable and trend consistent up to a 6 step horizon, providing a practical buffer for proactive handling of abnormal events. The predicted role activation propensities are temporally aligned with key process events, supporting timely coordination. Disturbance analysis further demonstrates responsive behavior under abnormal conditions and provides interpretable evidence for decision support. Overall, the proposed framework offers a general modeling paradigm for operator involved monitoring, early warning, and collaborative response, and it is applicable to be extended to civil and infrastructure assets with sensing and monitoring capabilities.
Zhang et al. (Sun,) studied this question.