This thesis will leverage Large Language Models (LLMs) to automate the construction of the Knowledge Graph (ontology, schema, relationships) from structured and unstructured sources (e.g., sensor data, technical manuals, regulations, logs, etc.) and enable advanced KG utilization (e.g., semantic querying, reasoning, scenario planning, energy optimization, etc.). Work to be done:
- Domain Review & LLM Strategy
- Survey existing maritime ontologies and standards relevant to ship data
- Review available data sources within the TwinShip project
- Assess applicability of LLMs for semantic extraction, ontology design, and knowledge enrichment
- Compare general-purpose and domain-specific LLMs (e.g., Llamarine)
- LLM-Assisted Ontology & KG Construction
- Collect representative structured (sensor data, logs) and unstructured (manuals, regulations) datasets
- Use LLMs to extract entities, relationships, and attributes
- Propose ontology elements automatically with LLM support
- Refine and validate ontology proposals with domain experts
- Integrate validated elements into the knowledge graph
- Semantic Enrichment Pipeline
- Design pipelines for harmonization and integration across heterogeneous datasets
- Connect pipeline outputs to LLM-assisted extraction processes
- Implement data quality assessment (detect inconsistencies, missing values, outliers)
- Develop anomaly detection methods for faulty sensor readings or unusual operational data
- Apply data recovery techniques (imputation, interpolation, contextual correction)
- Ensure enriched data is robust and consistent for KG applications
- Use-Cases Enabled by LLM-Augmented KG
- Develop semantic queries for energy-related analysis (fuel performance, energy prediction, compliance checking)
- Apply KG reasoning for anomaly detection (abnormal energy use, equipment malfunctions, regulatory deviations)
- Link anomaly detection to recovery strategies (corrective measures, operational adjustments)
- Demonstrate scenario-based decision-making in voyage planning or system configuration
- Explore integration of LSDTs (LLM-Augmented Semantic Digital Twins)
- LLM-Driven Reasoning & Insight Generation
- Generate natural language explanations from KG queries o Provide recommendations for energy optimization strategies
- Deliver anomaly alerts with contextual reasoning and suggested interventions
- Support interactive DSS queries in natural language
- Validation & Metrics
- Test the KG construction pipeline and enrichment process on TwinShip pilot vessel datasets
- Measure accuracy of entity and relation extraction
- Benchmark time efficiency compared to manual KG development
- Evaluate effectiveness of anomaly detection and recovery mechanisms
- Assess improvements in semantic query expressiveness and DSS support
- Analyze resilience of the KG against noisy, incomplete, or inconsistent input data