Multimodal knowledge graph construction for risk identification in water diversion projects

Abstract

The operation risks of water diversion projects involve numerous influencing factors, complex interrelationships, and heterogeneous data from multiple sources. This study presents a multimodal knowledge graph construction approach for water diversion projects, aiming to comprehend and identify the key risks associated with engineering operation and their propagation patterns. Utilizing term-masked pre-trained language models enhances comprehension of specialized terminology and identifies risk entities within the text. Employing high-order residual convolutional neural networks improves the processing capability for complex graph data, extracting risk information from images. Aggregating multimodal knowledge graphs based on the semantic relationships among entities and conditional probability to determine the coupled features of different risks. Employing complex network theory, analyze node degree and betweenness centrality to identify the diffusion effects and propagation levels of risks. The results indicate that the knowledge extraction accuracy of our method is high (with an average F1 score of 95.85%), enabling the qualitative analysis and quantitative calculation of operational risks in engineering. Relevant studies can effectively enhance the reliability of engineering safety management and reduce the impact of engineering risks on water supply security.