摘要
针对大跨度桥梁等工程结构在紊流场作用下的抖振响应预测问题,以薄平板为例,将数值模拟的薄平板抖振响应时程结果作为训练与测试数据,选用风场时程数据作为输入,并将薄平板的横向位移、竖向位移以及扭转角响应时程数据作为输出,分别采用带外部输入的非线性自回归(NARX)、长短期记忆(LSTM)、卷积长短期记忆(Conv LSTM)、注意力机制长短期记忆(LSTM-AM)神经网络模型预测薄平板的抖振响应。进一步地,将迁移学习(TL)方法与上述神经网络模型相结合,提出基于Davenport准定常抖振理论获取大量源任务数据的方法。通过筛选出的可用源任务数据,训练上述神经网络模型并经共享权重、微调参数后完成对薄平板目标任务数据的预测,并最终构建了TL-Conv LSTM-AM组合模型来预测薄平板抖振响应的思路。研究结果表明:在薄平板抖振响应预测中,LSTM模型的预测精度要高于NARX模型;引入卷积计算和注意力机制均有利于时序数据的预测,因此Conv LSTM和LSTM-AM模型的抖振响应预测精度相比单一的LSTM模型的预测精度要高;当上述神经网络模型结合迁移学习方法后能有效提升抖振响应的预测精度,但在局部区域仍有一定偏差;提出的TL-Conv LSTM-AM组合模型在高紊流情况下3个方向的抖振响应预测精度R2均超过了0.99,低紊流情况下预测精度R2也超过了0.98,其中扭转角的预测精度R2也接近于0.99。因此,所提出的TL-Conv LSTM-AM组合模型在薄平板抖振响应预测中的精度较高,具有较强的泛化性。研究成果可为大跨度桥梁等工程结构的抖振响应时程预测提供新的思路。
To predict the buffeting response of long-span bridges and other engineering structures under the action of a turbulent wind field,a thin plate was considered in this study.The time-history results of the buffeting response of the thin plate obtained via numerical simulation were adopted as the training and testing data.The time-history data of the wind field were then used as the inputs,and the time-history data of the lateral displacement,vertical displacement,and torsional angle of the thin plate were used as the outputs.Nonlinear autoregressive with external input(NARX),long short-term memory(LSTM),convolutional LSTM(Conv LSTM),and LSTM-attention mechanism(LSTM-AM)neural network models were adopted to predict the buffeting response of the thin plate.Moreover,the transfer learning(TL)method was combined with the neural network models mentioned above,and a method based on the Davenport quasi-steady buffeting theory was proposed to obtain large amounts of source task data.Using the selected available source task data,the above neural network models were trained,and the prediction for the target task data of the thin plate was completed using shared weights and parameter fine-tuning methods.Finally,a combined TL-Conv LSTM-AM model was constructed to predict the buffeting response of the thin plate.The results show that the prediction accuracy of the LSTM model is higher than that of the NARX model in predicting the buffeting response of the thin plate.Because the introduction of the convolution calculation and attention mechanism are conducive to the prediction of time-series data,the Conv LSTM and LSTM-AM models have higher prediction accuracies than the single LSTM model.When the above neural network models are combined with the TL method,the prediction accuracy of the buffeting response can be effectively improved.However,there are still some deviations in some local regions.For the proposed combined model of TL-Conv LSTM-AM,values of R2 are all larger than 0.99 in the buffering response of the three direction
作者
胡朋
丁艳
韩艳
张非
程渭
HU Peng;DING Yan;HAN Yan;ZHANG Fei;CHENG Wei(School of Civil Engineering,Changsha University of Science&Technology,Changsha 410114,Hunan,China)
出处
《中国公路学报》
EI
CAS
CSCD
北大核心
2023年第8期96-111,共16页
China Journal of Highway and Transport
基金
国家自然科学基金项目(52178451,51878080,52178452)
湖南省自然科学基金项目(2020JJ3035,2021RC4031)
长沙理工大学研究生科研创新项目(CXCLY2022037)。
关键词
桥梁工程
抖振响应预测
迁移学习
薄平板
长短期记忆网络
bridge engineering
buffeting response prediction
transfer learning
thin plate
long short-term memory network
