Physically verifiable intelligent assessment and restoration decision-making for cultural heritage using multimodal knowledge graphs
TL;DR
This approach effectively mitigates prediction biases inherent in purely data-driven models, reducing the root mean square error for long-term crack expansion prediction to 0.15 mm and constructing a semantically and physically coupled ontology knowledge graph.
Abstract
Addressing challenges in cultural heritage damage assessment, namely difficulties in multi-source data fusion, lack of physical mechanism support, and the inability to quantitatively validate restoration decisions, this paper proposes a physically verifiable intelligent assessment and decision-making method driven by multimodal knowledge graphs. This approach constructs a semantically and physically coupled ontology knowledge graph. By incorporating Physical Information Neural Networks (PINNs) and cross-modal reasoning architectures, it integrates physical equation constraints into data-driven models. Taking the Zhou Dynasty bronze vessel ‘Wusi Weiding’ as a case study, this approach effectively mitigates prediction biases inherent in purely data-driven models, reducing the root mean square error (RMSE) for long-term crack expansion prediction to 0.15 mm. Concurrently, the intelligent repair decisions generated based on the upper-confidence bound risk index (UCB-Risk) demonstrate high robustness, significantly outperforming conventional methods in both structural safety factor (enhanced to 3.5) and compliance with cultural heritage preservation principles (achieving 98%).