基于LSTM-DNN(长短期记忆-深度神经网络)融合模型的土压平衡盾构土仓压力预测方法  

Earth Chamber Pressure Prediction Method for Earth Pressure Balance Shield Based on LSTM-DNN Fusion Model

作　　者：王伯芝 黄永亮 陈文明 丁爽 刘浩 刘学增[5] 彭子晖[6] 吴炜枫 王嘉烨 WANG Bozhi;HUANG Yongliang;CHEN Wenming;DING Shuang;LIU Hao;LIU Xue-zeng;PENG Zihui;WU Weifeng;WANG Jiaye(School of Qilu Transportation,Shandong University,250002,Jinan,China;Jinan Rail Transit Group,Co.,Ltd.,250014,Jinan,China;Shanghai Tongyan Civil Engineering Technology Co.,Ltd.,200092,Shanghai,China;Shanghai Underground Infrastructure Safety Testing and Maintenance Equipment Engineering Technology Research Center,200092,Shanghai,China;Civil Engineering College,Tongji University,200092,Shanghai,China;Shanghai Tunnel Engineering&Rail Transit Design and Research Institute,200235,Shanghai,China)

机构地区：[1]山东大学齐鲁交通学院,济南250002 [2]济南轨道交通集团有限公司,济南250014 [3]上海同岩土木工程科技股份有限公司,上海200092 [4]上海地下基础设施安全检测与养护装备工程技术研究中心,上海200092 [5]同济大学土木工程学院,上海200092 [6]上海市隧道工程轨道交通设计研究院,上海200235

出　　处：《城市轨道交通研究》2024年第12期39-45,共7页Urban Mass Transit

基　　金：山东省重点研发计划项目(重大科技创新工程)(2019JZZY010428);山东省自然科学基金项目(ZR2020QE256,ZR2020QE243);上海市“科技创新行动计划”启明星项目(扬帆专项)(22YF1446100);上海市青年科技英才扬帆计划项目(23YF1440500)。

摘　　要：[目的]土仓压力是土压平衡盾构施工安全评估的关键参数,准确预测土仓压力有助于施工技术人员及时采取管控措施,进而保障地铁隧道的建设安全性。因此,有必要对土压平衡盾构土仓压力预测方法进行研究。[方法]提出一种多分支的LSTM(长短期记忆)-DNN(深度神经网络)融合模型。LSTM分支通过回溯历史数据提取其时序演变特征,DNN分支提取掘进状态特征,将两者组合后通过全连接层进行融合,实现对土仓压力的预测。依托济南轨道交通1号线实际盾构隧道数据对模型进行验证,并与LSTM模型、DNN模型进行了对比分析。[结果及结论]基于LSTM-DNN融合算法建立的土仓压力预测模型可以高效收敛,且所提模型在训练集和验证集上的预测效果良好。在后续的100步测试中,由LSTM-DNN融合模型得出的土仓压力预测值较好地反映了真实值的变化趋势,其平均偏差为7.65 kPa,相对误差为6.09%,预测精度较高。[Objective]Earth chamber pressure is a key parameter for EPB(earth pressure balance)shield construction assessment.Accurate prediction of earth chamber pressure helps construction technicians take timely control measures to ensure subway tunnel construction safety.Therefore,it is necessary to study the earth chamber pressure prediction method of EPB shield.[Method]A multi-branch LSTM(long and short term memory)-DNN(deep neural network)fusion model is proposed.LSTM branch extracts its time series evolution characteristics by backtracking historical data,while DNN branch extracts excavation state characteristics.The two branches are combined and then integrated through a fully connected layer to realize the prediction of earth chamber pressure.This multi-branch model is verified based on the actual shield tunnel data of Jinan Rail Transit Line 1,and compared with LSTM and DNN models respectively.[Result&Conclusion]The prediction model of earth chamber pressure based on LSTM-DNN fusion algorithm can converge efficiently,and has good prediction effects on the training set and the verification set.In the subsequent 100-step test,the predicted value of earth chamber pressure obtained by the LSTM-DNN fusion model better reflects the change trend of the actual value,with an average deviation of 7.65 kPa and a relative error of 6.09%,indicating a higher prediction accuracy.

关 键 词：城市轨道交通 土仓压力预测 长短期记忆 深度神经网络 

分 类 号：U455.43[建筑科学—桥梁与隧道工程]