基于L2-DFNN的桥梁多层次风险识别研究

打开文本图片集

中图分类号：U445.7 文献标志码：A 文章编号：1000-0844（2025）06—1357—12

DOI：10.20000/j.1000-0844.20240914001

Abstract： A three-span continuous girder bridge was studied to establish a multilevel risk scenario library and address the insuficient identification capability of traditional risk assessment methods under complex and unknown bridge risk scenarios. Feature learning was performed on different risk types， locations，and severity levels of bridges using an improved deep feedforward neural network （DFNN）. The trained network was used to predict and analyze typical risk scenarios， achieving multilevel risk identification for bridges. Results demonstrate that L2 regularization significantly enhances the generalization capability of DFNN， yielding training accuracies of 91.6% for classification， 74.9% for localization，and 93.7% for quantitative identification. In the finite element simulations of typical risk identification tasks， the prediction accuracies reach 80% ， （20 100% ，and 97.41% ，respectively. The proposed L2-DFNN effectively enables multilevel bridge risk identification， providing critical support for bridge health monitoring.

Keywords： multilevel risk identification； deep feedforward neural network （DFNN）； L2 regularization; finite element simulation; bridge health monitoring

0 引言

桥梁在长期服役过程中，受到外界荷载、自然灾害和材料老化等因素的影响，内部的损伤会不断累积，逐步降低桥梁承载能力，导致桥梁风险事故频发、突发。（剩余15840字）