Serviceability-oriented multimodal data integration for prescriptive road-tunnel maintenance digital twins

Road tunnels are indispensable components of transport networks, providing critical connections while operating continuously under demanding environmental, geological, and hydrogeological conditions. Their large dimensions, confined geometry, and uninterrupted operation make comprehensive inspection challenging. Full-face inspection technologies such as light detection and ranging (LiDAR) can rapidly capture high-resolution surface data; however, the resulting digital models are often descriptive rather than explanatory or instructive. This limitation is critical because seemingly minor defects may indicate latent deterioration and, if untreated, can lead to severe safety and operational consequences. With advances in artificial intelligence, particularly large language models (LLMs) and robotic inspection platforms, tunnel inspection is expected to move beyond defect detection towards knowledge-driven diagnosis and decision support. This paper presents a serviceability-oriented Prescriptive Twin framework for road-tunnel maintenance, that is, a digital twin that represents asset conditions and recommends maintenance actions from inspection evidence and codified engineering knowledge. The framework connects inspection data, prescriptive knowledge, decision-making, and verification through a Web Ontology Language (OWL 2) domain ontology populated by a schema-guided LLM pipeline. The ontology integrates tunnel maintenance standards, road asset data requirements, and an extended Construction Operations Building Information Exchange (COBie) schema, while the pipeline converts narrative standards into expert-validated seven-level Failure Mode and Effects Analysis (FMEA) chains. A multimodal layer maps red–green–blue (RGB), RGBD (RGB with depth), thermal, and ground-penetrating radar (GPR) outputs to causal-chain entry points, such as level 4 indicators and level 5 causes. The framework specifies an identity-resolution rule that binds observations from different modalities, and from successive inspection cycles, to a persistent defect instance, and a governance layer that binds every decision threshold to a dated clause of a specific standard edition, so that a prescription can be reconstructed against the edition under which it was issued. Chain traversal maps onto the inspection report and an inter-cycle change report. Demonstrated on two defects in an operating road tunnel, the framework generates traceable prescriptions and shows how multimodal evidence improves intervention accuracy: at the worked example, thermal and GPR evidence escalated the intervention tier and expanded the repair scope by a factor of 3.2 relative to RGB-only inspection, delivering an industry-deployable pathway for standards-compliant maintenance decision support.