基于信号分解与重构的基坑变形预测研究  

Study on Foundation Pit Deformation Prediction Based on Signal Decomposition and Reconstruction

作　　者：李志伟 冼进业 祝敏刚 Li Zhiwei;Xian Jinye;Zhu Mingang(Guangzhou Urban Planning&Design Survey Research Institute,Guangzhou 510060,China;Collaborative Innovation Center for Natural Resources Planning and Marine Technology of Guangzhou,Guangzhou 510060,China;Guangdong Enterprise Key Laboratory for Urban Sensing,Monitoring and Early Warning,Guangzhou 510060,China;Guangzhou Virtual Power Network Technology Co.,Ltd.,Guangzhou 510060,China;China Power Construction Group Urban Planning and Design Institute Co.,Ltd.,Guangzhou 510060,China)

机构地区：[1]广州市城市规划勘测设计研究院有限公司,广东广州510060 [2]广州市资源规划和海洋科技协同创新中心,广东广州510060 [3]广东省城市感知与监测预警企业重点实验室,广东广州510060 [4]广州市虚拟动力网络技术有限公司,广东广州510060 [5]中国电建集团城市规划设计研究院有限公司,广东广州510060

出　　处：《市政技术》2025年第1期144-150,158,共8页Journal of Municipal Technology

基　　金：广东省城市感知与监测预警企业重点实验室基金项目(2020B121202019)。

摘　　要：提出了一种基于信号分解与重构的基坑变形预测方法。首先,运用经验模态分解(EMD)对基坑变形序列数据进行分解,得到数个本征模态函数(IMF),并获取不同频率的特征信号,将变形序列分解成趋势项与波动项。通过门控循环单元(GRU)结合自注意力机制,对波动项的序列数据进行建模,并对比筛选不同时长的输入数据,以确定最佳输入时长,从而提高预测精度。对于趋势项,采用多项式拟合方法进行预测,最终将趋势项与波动项的预测结果相加得到最终预测结果。以广州某基坑的监测数据为实验对象,对所提方法进行了验证。结果表明,当输入步长为5时,模型的MSE仅为0.92,R2为0.980,有效提升了预测的准确性。本研究可以为类似的基坑监测项目提供借鉴。This paper proposes a foundation pit deformation prediction method based on signal decomposition and reconstruction.Firstly,the empirical mode decomposition(EMD)method is used to decompose the foundation pit deformation sequence data into several intrinsic mode functions(IMF),obtain characteristic signals of different frequencies,and decompose the deformation sequence into trend and fluctuation components.The sequence data of the fluctuation component is modeled by the gated recurrent unit(GRU)combined with the self-attention mechanism.Different input data lengths are compared and selected to determine the optimal input duration,thereby prediction accuracy will be improved.For the trend component,a polynomial fitting method is applied in prediction.Finally,the final prediction result is obtained by summing the prediction results of the trend and fluctuation components.The proposed method is validated by monitoring data from a foundation pit in Guangzhou.The results show that when the input step is 5,the mean squared error(MSE)of the model is only 0.92,the coefficient of prediction(R2)is 0.980,which effectively improved the prediction accuracy.This study can provide reference for similar foundation pit monitoring projects.

关 键 词：基坑变形预测 GRU 自注意力机制 经验模态分解 

分 类 号：TU433[建筑科学—岩土工程] TP183[建筑科学—土工工程]