使用CNN(卷积神经网络)-LSTM(长短期记忆)联合神经网络预测盾构隧道施工引起的地面沉降

Prediction of Land Subsidence Caused by Shield Tunnel Construction with Joint CNNLSTM Neural Network

[目的]地铁盾构隧道施工会引起周围地面沉降,影响周围环境。传统地面沉降预测方法难以综合考虑沉降影响因素,对此,为提高地面沉降的预测精度,使用CNN(卷积神经网络)-LSTM(长短期记忆)联合神经网络,对盾构隧道施工引起的地面沉降进行预测。[方法]以某地铁施工区间地面沉降监测数据为研究对象,使用CNN对影响参数(压缩模量、黏聚力、内摩擦角、泊松比、土层厚度、隧道埋深和施工参数)与地面沉降监测值进行连接,使用LSTM神经网络对地面沉降进行分析,建立了基于CNN-LSTM联合神经网络的地面沉降预测模型,探讨了同时考虑多个因素对地面沉降预测值的影响。[结果及结论]使用CNN对地面沉降相关的影响参数特征提取效果较好;所建CNN-LSTM模型的准确率比单独使用LSTM模型的准确率提高了3%、比传统BP(反向传播)神经网络模型准确率提高了9%;所建CNN-LSTM模型,对单测点短时间地面沉降预测准确率达到93%,预测值与监测值吻合较好。

[Objective]Metro shield tunnel construction may cause surrounding land subsidence,affecting the surrounding environment.Traditional land subsidence prediction models are difficult to comprehensively consider the influencing factors of land subsidence.Therefore,in order to improve the prediction accuracy of land subsidence,the CNN(convolutional neural network)-LSTM(long-short-term memory)joint neural network is used to predict the land subsidence caused by shield tunnel construction.[Method]With the monitored land subsidence data of a metro section as the research object,CNN is used to connect the influencing parameters(including compressive modulus,cohesion,internal friction angle,Poisson's ratio,soil thickness,tunnel buried depth,and construction parameters)and monitored land subsidence data.The LSTM neural network is used to analyze the land subsidence,and a land subsidence prediction model based on the CNN-LSTM joint neural network is established.Simultaneous consideration of the multiple factor influence on land subsidence prediction is explored.[Result&Conclusion]Using CNN has a good effect on extracting the parameter features related to land subsidence.The prediction accuracy of the established CNN-LSTM model is3%higher than that of the LSTM model alone,and 9%higher than that of the traditional BP(back propagation)neural network model.The prediction accuracy of the established CNNLSTM model reaches 93%when predicting the short time land subsidence in single measuring point,and the predicted value is in good agreement with the monitored value.