Research on risk decision-making generation method for water conservancy project based on multimodal knowledge graph...

Area of Science:

• Civil Engineering
• Computer Science
• Environmental Science

Background:

• Traditional knowledge graphs for water conservancy project risks are limited by data modalities and information extraction accuracy.
• Effective risk management in water conservancy requires integrating diverse data sources for comprehensive analysis.

Purpose of the Study:

• To propose a multimodal water conservancy project risk knowledge graph.
• To develop a synergistic strategy using multimodal large language models for risk decision-making generation.
• To enhance the accuracy and reliability of risk assessment in water conservancy projects.

Main Methods:

• Developed a multimodal knowledge graph integrating visual and textual data.
• Improved DenseNet model with self-attention and coordinate attention for image classification.
• Utilized BERT-BiLSTM-CRF architecture for extracting textual risk entities.
• Implemented a multi-agent agentic retrieval-augmented generation framework for decision-making.

Main Results:

• The enhanced DenseNet model showed improved precision and recall in image recognition.
• The multimodal knowledge graph and generation framework achieved strong performance on BERTScore and ROUGE-L metrics.
• Demonstrated enhanced reliability and interpretability in risk decision-making outputs.

Conclusions:

• The proposed multimodal knowledge graph and generation strategy effectively address limitations of traditional approaches.
• This work offers a novel perspective for advancing risk management in water conservancy projects through multimodal data integration.
• The findings support more accurate and reliable risk decision-making in complex water conservancy systems.