第七届全国防灾减灾工程学术会议

101 / 2025-05-12 10:41:10

融合DFNN与MLLM的桥梁多层次风险识别研究

bridge health monitoring; multi-level risk identification; deep feedforward neural network; multimodal large language model; disaster prevention and mitigation engineering

Abstract: To address the shortcomings of traditional bridge risk identification methods in multimodal data fusion and cross-modal semantic understanding, a bridge multi-level risk identification method integrating Deep Feedforward Neural Network (DFNN) and Multimodal Large Language Model (MLLM) is proposed. Based on finite element simulation and field monitoring ata, a multimodal risk scenario library is constructed, incorporating structural responses, damage images, and inspection texts. The DFNN model is designed to extract sensing features, while the MLLM enables cross-modal semantic reasoning, with feature fusion enhancing the overall identification capability. Results show that the proposed method achieves an accuracy of 94.2% for risk type classification, 92.8% for location identification, and 91.7% for severity prediction, demonstrating good generalization and robustness in both simulation and experimental bridge tests. This study provides a novel approach and technical support for intelligent bridge monitoring and disaster prevention and mitigation, with promising engineering application value.