基于LSTM神经网络和残余力向量法的结构损伤识别  

Structural Damage Recognition Based on LSTM Neural Network and Residual Force Vector Method

作　　者：宋福春[1] 杨子豪 付聿旻 崔福和 白祥鸽 SONG Fuchun;YANG Zihao;FU Yumin;CUI Fuhe;BAI Xiangge(School of Transportation and Geomatics Engineering,Shenyang Jianzhu University,Shenyang,China,110168)

机构地区：[1]沈阳建筑大学交通与测绘工程学院,辽宁沈阳110168

出　　处：《沈阳建筑大学学报（自然科学版）》2023年第5期872-879,共8页Journal of Shenyang Jianzhu University：Natural Science

基　　金：“十三五”国家重点研发计划重点项目(2018YFC0809600,2018YFC0809606)。

摘　　要：目的 为减少传统残余力向量法的工作量,提高计算效率,提出一种采用LSTM神经网络与残余力向量法相结合方法。方法 以结构损伤后的残余力作为LSTM神经网络的损伤识别指标,建立输入与输出之间模型,同时运用分步损伤识别法,对可能存在损伤的结构进行判断,并通过简支梁模型进行验证。结果 LSTM神经网络对简支梁损伤情况判断较为准确,在样本数为350组的情况下,其分类准确率为97%,训练结果的均方根误差值为0.64,预测结果的最大误差为3.7%;噪声水平在10%及以下时,最大误差为6.8%,噪声水平在15%及以下时仍可对单损伤做出较为准确的判断,最大误差为9.4%,抗噪性较好。结论 所设计的基于残余力LSTM神经网络对结构损伤定位与程度识别效果较好,具有一定可行性。In order to reduce the workload of the traditional residual force vector method and improve the computational efficiency,a combination method using LSTM neural network and residual force vector method is proposed.The residual force of the structure after damage is used as the damage identification index of the LSTM neural network,and a model between input and output is established,while the step-by-step damage identification method is applied to determine the structure that may have damage,which is verified by the simple-supported beam model.The LSTM neural network is more accurate in judging damage of the simple beam,with a classification accuracy of 97%,a root mean square error value of 0.64 for the training results and a maximum error of 3.7%for the prediction results when the sample data is 350 groups;the maximum error is 6.8%when the noise level is 10%and below,and it can still make a single damage when the noise level is 15%and below The maximum error is 9.4%,which is a good noise immunity.The designed residual force-based LSTM neural network is effective in localising and identifying the extent of structural damage,and has a certain degree of feasibility.

关 键 词：结构损伤识别 LSTM神经网络 残余力向量法 损伤评估 

分 类 号：TU311.4[建筑科学—结构工程] TP183[自动化与计算机技术—控制理论与控制工程]