基于时序InSAR与机器学习的岩溶地区铁路沿线形变监测及预测方法  

Deformation monitoring and prediction method along railway in karst areas based on TS-InSAR and machine learning

作　　者：徐林荣[1] 何元幸子 邓志兴 肖源杰 李永威 陈昀灏 XU Linrong;HE Yuanxingzi;DENG Zhixing;XIAO Yuanjie;LI Yongwei;CHEN Yunhao(School of Civil Engineering,Central South University,Changsha 410075,China;School of Civil Engineering,Hunan University,Changsha 410082,China)

机构地区：[1]中南大学土木工程学院,湖南长沙410075 [2]湖南大学土木工程学院,湖南长沙410082

出　　处：《中南大学学报(自然科学版)》2025年第3期998-1014,共17页Journal of Central South University:Science and Technology

基　　金：铁路基础研究联合基金资助项目(U2268213);国家自然科学基金资助项目(42172322);中国中铁股份有限公司科技研究开发计划项目(2023-重点-09)。

摘　　要：岩溶塌陷灾害对线性工程稳定性和安全性的影响不容忽视,地表形变监测结果能直接反映灾害体的当前状态。为捕捉岩溶地区铁路既有线的沉降靶区并分析其形变特征,提出一种基于时序InSAR与机器学习结合的铁路沿线形变监测及预测方法。以京广铁路广州段为研究对象,首先基于时序InSAR技术获取并分析铁路沿线2019—2023年地表形变,结合现场勘察对铁路沿线形变时空特征、特征点时序形变特征及靶区形变稳定性进行分析。其次,采用小波变换(wavelet transform,WT)算法对3处路基历史塌陷点时序形变进行降噪处理,再基于反向传播(back propagation,BP)、极限学习机(extreme learning machine,ELM)、长短时记忆网络(long short-term memory,LSTM)这3种经典的机器学习模型进行形变的预测分析,并利用最佳机器学习模型对其进行超前预测。研究结果表明:铁路沿线存在A、B两处沉降靶区,2021—2023年,区内沉降量逐年增大且沉降范围不断扩大,其累计沉降量峰值为-118.01 mm;路基历史塌陷处均位于年平均形变速率的漏斗中心或边缘,结合正态分布统计,区域A、B均于2021年达到最不稳定状态,通过预测分析得到WT-LSTM模型的预测效果最佳,在训练集和测试集上的拟合优度分别高于0.98、0.94,基于超前预测结果发现未来形变预测的时间约为1 a。研究成果可为岩溶地区铁路沿线沉降早期预防提供关键技术支撑。The impact of karst collapse disaster on the stability and safety of linear engineering should not be ignored.The surface deformation monitoring can directly reflect the current state of the disaster body.To capture the settlement target areas of the railway in karst areas and analyze its deformation characteristics,a hybrid method was proposed based on TS-InSAR and ML for monitoring and predicting the surface deformation along the railway.The Guangzhou section of the Beijing—Guangzhou Railway was taken as the study area.Firstly,based on the TS-InSAR technology,the surface deformation along the railway from 2019 to 2023 was obtained and analyzed.Then,combined with the site survey results,the spatial and temporal deformation characteristics along the railway,deformation characteristics of the time series in characteristic points and the stability of deformation in the target area were analyzed.Secondly,the wavelet transform(WT)algorithm was used to reduce the noise of the time series deformation of the three historical railway collapse points.Then,back propagation(BP),extreme learning machine(ELM)and long short-term memory(LSTM)were used to predict the time series deformation of the feature points.Finally,the optimal model of three classical ML models was used to carry out over-prediction of time series deformation.The results show that there are two settlement target areas along the railway which are A and B.In 2021—2023,the settlement in the area increases year by year,and the settlement range is expanded.The peak of cumulative settlement is-118.01 mm.The historical collapse sites of the subgrade are all located at the center or edge of the average annual deformation rate funnel for each period.Combining the normal distribution statistics,both A and B are the most unstable state in 2021.The WT-LSTM model has the best prediction effect.The goodness of fit on the train set and the test set are higher than 0.98 and 0.94,respectively.The scale of future deformation prediction is about one year based on the resul

关 键 词：铁路 地表形变 时序InSAR 机器学习 时序预测 

分 类 号：U213.1[交通运输工程—道路与铁道工程]