An adaptive causal path reasoning model for marine diesel engine fault diagnosis using knowledge graph
Henglong Shen, Hui Cao, Zeren Ai, Saishuai Dai, Longde Wang · Knowledge-Based Systems · 2025
Marine diesel engine diagnostics often suffer from limited interpretability and insufficient integration of expert knowledge. To address these challenges, this paper proposes a novel adaptive reasoning model, Knowledge Graph-based Marine Diesel Engine Fault Reasoning (KG-MDEFR), which simultaneously optimises structural and semantic features. KG-MDEFR takes a query triple as input and incrementally constructs a multi-hop reasoning graph via a layer-wise path generator. It expands to semantically relevant entities while preserving local connectivity, which effectively controlseffectively controlling node expansion. A semantic-aware sampling mechanism scores candidate entities using graph embeddings and selects paths through a parameterised selection strategy combined with Gumbel top-K sampling. A structural masking strategy is also employed, combining edge attention and triple scoring to model the structural features of entity relations. This reduces dependence on fixed graph structures and enhances generalisation in inductive scenarios. Experimental results demonstrate that KG-MDEFR outperforms baseline models in transductive reasoning on diesel engine fault graphs (MRR = 0.9230, Hit@1 = 0.8915) and that it generalises well to unseen entities (MRR = 0.6785, Hit@1 = 0.6354). Case studies further confirm its ability to trace causal reasoning paths and deliver interpretable fault analysis. The model provides a practical and deployable knowledge-graphknowledge reasoning solution for intelligent marine engine maintenance.