The autonomous Prognostics and Health Management (PHM) of complex combustion systems faces fundamental challenges: reliance on domain expertise for manual setup, scarcity of empirical data for rare accident scenarios, and prohibitive computational costs of high-fidelity simulations. This study proposes a novel framework that integrates Large Language Models (LLMs), autonomous Multiphase Flow with Interphase eXchanges (MFiX) simulation, Reduced-Order Models (ROMs), and machine learning. The LLM serves as a generative world model that synthesizes accident scenarios from semantic resources and translates abstract fault descriptions into structured simulation parameters, enabling a fully automated workflow. To address the computational burden of extensive data generation, Operator Inference (OpInf) constructs a data-driven ROM. A machine learning classifier trained on the augmented dataset then performs real-time fault diagnosis. By bridging qualitative semantic reasoning and quantitative physical dynamics, the proposed framework automates end-to-end scenario generation and diagnosis while ensuring robustness against rare, unforeseen hazards. This integration represents a meaningful advancement in digital twin technologies, establishing a foundation for truly autonomous and proactive disaster management systems.