Automated Health Status Assessment for a Combustion System via LLM-Generated Failure Scenarios

Automated health status assessment of complex combustion systems—such as fluidised bed combustion in power plant applications—faces three persistent challenges: dependence on domain expertise, scarcity of empirical data on rare accident scenarios, and the prohibitive high-fidelity simulation costs. We propose a framework that integrates domain-grounded large language models, high-fidelity simulations, data-driven reduced-order models, and machine learning for computationally efficient, simulation-based health status assessment. The large language model synthesises failure scenarios and translates failure descriptions into simulation parameters. We construct a reduced-order model via operator inference to reduce the computational costs by 99.74% over the same 0–15 s time horizon and train an extreme gradient boosting classifier to effectively categorise fault signatures and differentiate between faulty and nominal operational states. All reduced-order model trajectories remained numerically stable up to 30 s, and the classifier achieved an accuracy, precision, recall, and F1-score of 1.0 in both clean on-grid and independent off-grid validation subjected to mild noise conditions. The framework automates the workflow from scenario generation to simulation-based health status assessment, supporting more efficient exploration of rare fault conditions in digital twin applications. These results demonstrate feasibility within the reduced, sensitivity-guided operating envelope considered in this study.