高斯过程机器学习在临河地铁车站基坑变形预测中的应用

Application of Gaussian process machine learning in deformation prediction of foundation pit of subway station near river

为了更准确地对临河基坑开挖过程中出现的变形进行预测和风险预警,以天津地铁7号线王兰庄站出入口基坑临河施工为例,提出了一种基于改良高斯过程机器学习的临河基坑变形预测方法。首先,将施工过程收集的变形位移监测数据作为学习样本;其次,利用改良高斯过程回归算法对学习样本进行训练,并通过自适应不同的训练函数,获取较为合理的基坑变形代理模型;最后,通过代理模型预测基坑变形位移,并识别基坑风险,同时与传统高斯预测结果、有限元计算结果和实测值进行比较。结果表明:1)改良高斯过程回归方法较常规高斯过程的模拟效果更好,在样本点充足的条件下,可准确地根据基坑变形数据学习并预测后续基坑开挖的变形位移;2)相较于工程应用较多的有限元计算方法,采用改良高斯回归过程机器学习方法,可根据现场的实际监测数据对变形进行动态地学习、分析与预测。研究成果更贴近现场实际情况,具有较强的推广价值,可为基坑变形预测提供参考。

In order to more accurately predict the deformation and risk warning during the excavation of riverside foundation pit in underground engineering,taking the riverside excavation of Wanglanzhuang station of Tianjin Metro Line 7 as an example,a method for deformation prediction of riverside excavation based on improved Gaussian process machine learning was proposed.Firstly,the deformation and displacement monitoring data collected during construction were taken as learning samples.Secondly,the improved Gaussion process regression algorithm was used to train the learning samples,and a more reasonable deformation proxy model was obtained by adapting different training functions.Finally,the proxy model was used to predict the deformation and displacement of foundation pit and identify the risk of foundation pit.And the results were compared with the traditional Gaussian prediction results,the finite element calculation results and the measured values.The results show that:1)The improved Gaussian regression process method has better simulation effect than the conventional Gaussian process.Under the condition of sufficient sample points,it can accurately learn and predict the deformation and displacement of the subsequent foundation pit excavation according to the deformation data;2)Compared with the finite element calculation method widely used in engineering,the modified Gaussian regression process machine learning method can dynamically learn,analyze and predict the deformation according to the actual monitoring data on site.The results is more appropriate to the actual situation on site and has strong promotion value,and can provide reference for the prediction of foundation pit deformation.