基于结构动力特性的结构损伤深度置信网络分层识别研究  

Research on hierarchical identification of structural damage based on dynamic characteristics by deep belief networks

作　　者：常亮亮 姜文恺 杨汉青 孙星 何伟 CHANG Liangliang;JIANG Wenkai;YANG Hanqing;SUN Xing;HE Wei(Xuchang Construction Investment Co.,Ltd.,Xuchang 461000,China;China Railway Fourth Survey and Design Institute Group Co.,Ltd.,Wuhan 430063,China;School of Civil Engineering and Transportation,North China University of Water Resources and Electric Power,Zhengzhou 450045,China;China Railway 16th Bureau Group Co.,Ltd.,Beijing 100018,China)

机构地区：[1]许昌市建设投资有限责任公司,河南许昌461000 [2]中铁第四勘察设计院集团有限公司,湖北武汉430063 [3]华北水利水电大学土木与交通学院,河南郑州450045 [4]中铁十六局集团有限公司,北京100018

出　　处：《地震工程与工程振动》2024年第2期61-71,共11页Earthquake Engineering and Engineering Dynamics

基　　金：中铁十六局集团科技研发项目(K2020-7B);中铁第四勘察设计院集团科技研究开发项目(2020K161)。

摘　　要：为了高效准确地识别结构损伤,将机器学习和智能算法相结合,提出一种基于结构动力特性的结构损伤深度置信网络分层识别方法,分层依次识别损伤位置与损伤程度。为识别损伤位置,利用结构前3阶竖向振动频率和单节点3阶模态位移建立六元向量,以此六元向量作为输入参数,通过深度置信网络识别损伤位置;为识别损伤程度,分别采用前3阶竖向振动固有频率和模态位移或6节点模态曲率差为参数输入深度置信网络识别损伤程度,并以简支梁为模型进行验证。结果表明:识别损伤位置时,即使噪声程度达到10%,仍可准确识别损伤位置;识别损伤程度时,基于6节点模态曲率差的深度置信网络抗噪性强,在15%噪声水平下对损伤程度预测最大相对误差不超过5.08%,均方差为0.4878。与BP神经网络相比,无噪声时,BP神经网络的预测能力优于深度置信网络;相同噪声水平下,深度置信网络的预测能力明显优于BP神经网络,体现了基于结构动力特性的结构损伤深度置信网络分层识别方法鲁棒性强,识别结果精度高。To identify structural damage efficiently and accurately,a hierarchical identification method of structural damage based on structural dynamic characteristics and deep belief network is proposed by combining machine learning with intelligent algorithm,and the damage position and degree are identified in turn.In order to identify the damage location,a 6-element vector is established by using the first three vertical vibration frequencies of the structure and the third modal displacement of a single node,and the damage location is identified by using the 6-element vector as input parameters.To identify the damage degree,the first three natural frequencies and modal displacements of vertical vibration or the 6 nodes modal curvature differences are used as parameters to input the depth confidence network to identify the damage degree.A simple-supported beam is taken as a model to verify it.It is shown that the damage position recognition accuracy can reach 100%even if the noise level reaches 10%.When identifying the damage degree,the deep belief network based on six-node modal curvature difference has strong noise resistance.The maximum relative error of damage degree prediction is less than 5.08%and the mean square error is 0.4878 under the noise of 15%.Compared with BP neural network,the prediction ability of BP neural network is better than that of deep belief network when there is no noise.Under the same noise level,the prediction ability of depth belief network is obviously better than that of BP neural network,which shows that the hierarchical identification method of structural damage based on structural dynamic characteristics and deep belief networks has strong robustness and high accuracy of identification results.

关 键 词：深度置信网络 损伤识别 抗噪性 BP神经网络 

分 类 号：TU311[建筑科学—结构工程] TU18[自动化与计算机技术—控制理论与控制工程] TP183[自动化与计算机技术—控制科学与工程]