基于双层深度置信网络的梁桥结构损伤识别方法研究  

Damage identification method of the beam bridge structures based on a double-layer deep belief network

作　　者：闫嵩 彭华春[2] 杨汉青 何伟[3] YAN Song;PENG Huachun;YANG Hanqing;HE Wei(China Railway 16th Bureau Group Co.,Ltd.,Beijing 100018,China;China Railway Siyuan Survey and Design Group Co.,Ltd.,Wuhan 430063,Hubei,China;North China University of Water Resources and Electric Power,Zhengzhou 450045,Henan,China)

机构地区：[1]中铁十六局集团有限公司,北京100018 [2]中铁第四勘察设计院集团有限公司,湖北武汉430063 [3]华北水利水电大学,河南郑州450045

出　　处：《地震工程学报》2024年第1期66-73,104,共9页China Earthquake Engineering Journal

基　　金：河南省科技攻关计划项目(182102310890);河南省高等学校重点科研项目计划项目(19A560014);中铁十六局集团科技研发项目(K2020-7B);铁四院科技研究开发项目(2020K161)。

摘　　要：为高效准确识别桥梁结构损伤,将深度学习与结构动力特性相结合,提出基于双层深度置信网络的桥梁结构损伤识别方法。首先取结构前3阶竖向振动频率和跨中节点前3阶竖向振动模态位移为参数,将其共同作为首层深度置信网络(DBN)的输入数据对结构的损伤位置进行识别;然后以1阶竖向振动的模态位移差作为参数,基于二层DBN对结构损伤程度进行预测;最后以郑许市域铁路桥梁为例进行验证。计算结果显示,当不考虑误差时,基于双层深度置信网络的结构损伤方法进行识别且结果精确;当噪声程度不超过10%时,定位识别结果准确率达100%;当噪声程度不超过15%时,定量识别结果最大绝对误差限不超过1.15%,识别结果准确;与传统的BP神经网络方法相比,本方法识别精度更高,抗噪性更强。To accurately and efficiently identify structural damage in bridges,we propose a method based on a double-layer deep belief network(DBN).This approach combines deep learning with structural dynamic characteristics of structural engineering.First,the initial three vertical vibration frequencies of the structure,along with the first three vertical vibration modal displacements of midspan nodes,are taken as parameters.These parameters serve as the input data for the first-layer DBN to identify the damage location of the structure.Following this,the differences in the modal displacement of the first-order vertical vibration are taken as parameters.These are then used in the second-layer DBN to predict the extent of the structure damage.As a case study,we applied this method to the Zhengzhou-Xuchang suburban railway bridge.The calculation results show that when the error is not considered,the results of the structural damage identification method based on the double-layer DBN are precise.When the noise level does not exceed 10%,the accuracy of the location identification results is 100%.Even when the noise level does not exceed 15%,the maximum absolute error of quantitative identification results is not larger than-1.15%.Compared with the traditional BP neural network method,the proposed method demonstrates higher recognition accuracy and a stronger capability to resist noise.

关 键 词：DBN 损伤识别 抗噪性 固有频率 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]