Interpretable root-cause analysis of construction defects via knowledge-enhanced causal reasoning with large language models

Accurate root-cause identification of construction defects is a core challenge in ensuring engineering quality and structural safety; however, existing methods are fundamentally limited by insufficient domain knowledge, weak causal reasoning, and lack of interpretability in their analytical outputs. This paper proposes the knowledge-enhanced causal reasoning large language model framework (KE-CausalLLM), which deeply integrates a structured engineering knowledge graph, a counterfactual inference mechanism driven by a structural causal model, and a retrieval-augmented generation (RAG) large language model (LLM) to achieve interpretable root-cause analysis of construction defects. The framework comprises three core modules: a domain-specific engineering knowledge graph construction module that incorporates engineering standards, historical cases, and expert knowledge, containing 3867 entity nodes and 9214 relational edges; a counterfactual inference-based causal reasoning module that automatically identifies the minimal sufficient root-cause set via the do-operator; and a knowledge graph-augmented LLM reasoning engine that generates structured analytical reports-including root-cause identification, causal chain evidence, standard clause references, and remediation recommendations-through RAG-based dynamic subgraph retrieval and chain-of-thought prompting. The experimental dataset was independently collected by our research team over 18 months, spanning four provinces and 12 active construction projects, comprising 4152 annotated samples covering six defect categories: cracks, leakage, hollowing, rebar corrosion, settlement, and honeycombing. Experimental results demonstrate that KE-CausalLLM significantly outperforms seven baseline methods in accuracy (91.8%), F1 score (90.9%), and AUC (0.957). Ablation studies validate the independent contribution of each module; cross-project-type generalization experiments confirm the framework’s robustness across diverse engineering scenarios; and human evaluation of interpretability yields a weighted composite score of 4.29 out of 5. This study provides a new technical paradigm and theoretical foundation for intelligent quality management in construction engineering.