摘要
随着城市中大型建筑的快速增长,其结构安全健康已经成为社会普遍关注的重点。大型建筑物的结构位移预测是实现建筑物安全的重要技术保障。然而,现有的结构位移预测方法,由于没有考虑到各个位移点之间的关系,从而导致预测的精准度不高。因此为了实现准确的建筑位移预测,提出一种融合图卷积网络和蒸馏Transformer的建筑位移预测模型。该模型能够提取各位移点的非欧几何空间数据特征,并通过蒸馏的自注意力机制提取权重较大的特征,从而有效地减少模型计算量、参数和内存消耗。与其他模型相比,该模型在竖向位移预测方面表现出良好的性能,当位移趋势发生变化时,预测误差小于0.4 mm,在预测步长为9时MSE为1.104、MAE为0.787,较其他模型的MSE和MAE分别提高了4%和3%。
With the rapid growth of large buildings in cities,their structural safety and fitness have become the focus of national and social attention.Structural displacement prediction of large buildings is an important technical guarantee for realizing building safety.However,the existing structural displacement prediction methods do not take into account the relationship between displacement points,which leads to poor accuracy of prediction.Therefore,in order to realize accurate building displacement prediction,a building displacement prediction model is proposed by fusing graph convolutional network and distillation Transformer.The model is able to extract the non-Euclidean geometric spatial data features of each displacement point and the features with larger weights through the self-attention mechanism of distillation,which effectively reduces the model computation,parameters and memory consumption.Compared with other models,this model shows good performance in vertical displacement prediction,with a prediction error of less than 0.4 mm when the displacement trend changes,and MSE of 1.104 and MAE of 0.787 at a prediction step of 9,which is 4%and 3%higher than the MSE and MAE of other models,respectively.
作者
吴宏杰
田闯闯
陶然
傅启明
马洁明
崔志明
WU Hongjie;TIAN Chuangchuang;TAO Ran;FU Qiming;MA Jieming;CUI Zhiming(School of Electronic&Information Engineering,SUST,Suzhou 215009,China;School of Intelligent Engineering,Xi’an Jiaotong-Liverpool University,Suzhou 215123,China)
出处
《苏州科技大学学报（自然科学版）》
CAS
2024年第4期128-138,共11页
Journal of Suzhou University of Science and Technology（Natural Science Edition）
基金
国家自然科学基金项目(62073231,62176175)
国家重点研发计划项目(2020YFC2006602)
苏州大学江苏省计算机信息处理技术重点实验室开放课题(KJS2166)
苏州大学江苏省大数据智能工程实验室开放课题(SDGC2157)。
关键词
蒸馏自注意力机制
图卷积神经网络
建筑竖向位移
distillation self-attention mechanism
graph convolutional neural network
building vertical displacement
