A Novel Time-Varying Failure Risk Assessment Framework for Marine Diesel Engines Integrating Large Language Models and Bayesian Networks

Fault risks in marine diesel engines (MDEs) propagate across coupled subsystems and evolve with component degradation, but existing methods rarely integrate accident narratives, causal structure, and time-varying reliability. This study develops a novel framework that integrates large language models (LLMs), rough–fuzzy DEMATEL, interpretive structural modeling (ISM), and Bayesian networks (BNs) with service-time-dependent priors for time-varying failure analysis. First, the risk-influencing factors (RIFs) are extracted from accident and maintenance records using LLMs, text embeddings, semantic clustering, and expert consolidation. Rough–fuzzy DEMATEL and ISM are used to identify causal relationships and the hierarchical structure. The RIFs, bottom-level components, and target failure are then mapped into a multilayer BN parameterized using Noisy-OR relationships and Weibull-derived time-varying priors. In a case study of MDE hard starting, 338 cause descriptions from 35 records yielded 12 RIFs, with semantic coverage above 93% across four evaluation models. When hard starting was observed, the posterior probability of mechanical failure of the fuel injection system reached 60.92%, compared with 36.91% for governor and mechanical actuation system failure. Over 0–10,000 h of cumulative service, the model-inferred probability of hard starting during a single starting attempt increased from 38.09% to 75.23% under the specified model parameterization. The framework supports causal interpretation, troubleshooting prioritization, and service-time-dependent maintenance prioritization.