第19卷第2期地下空间与工程学报Vol.192023年4月ChineseJournalofUndergroundSpaceandEngineeringApr.2023基于神经网络的盾构滚刀磨损量预测方法探讨丁小彬1,2,谢宇轩1,薛皓文1,黄威然3(1.华南理工大学土木与交通学院,广州510640;2.华南岩土工程研究院,广州510640;3.广州轨道交通建设监理有限公司,广州510010)摘要:盾构施工中采集的参数众多,难以直接反映滚刀磨损量的发展规律。统计31种现有滚刀磨损研究中影响参数的出现频次,以广州地铁18号线番禺广场至南村万博区间2号始发井中间风井右线盾构区间为依托,选择14种输入参数,整理出共2386条样本用于BP神经网络模型开发。分别采用序贯模型优化(SMBO)和遗传算法(GA)进行超参数优化。模型预测值与实测值决定系数达0.832,准确预测了滚刀磨损量。选择预测准确度较高模型进行敏感性分析,验证所选的14种参数对于滚刀磨损量预测的贡献,结果表明,考虑土压、新旧刀、盾构深度对于滚刀磨损量预测有较大贡献。本文成果可为其他地层下的盾构滚刀磨损量预测与分析提供参考。关键词:盾构掘进;滚刀磨损;神经网络;优化算法中图分类号:U455.3文献标识码:A文章编号:1673-0836(2023)02-0560-11InvestigationofQuantitativePredictionofTBMDiscCutterWearbyANNDingXiaobin1,2,XieYuxuan1,XueHaowen1,HuangWeiran3(1.SchoolofCivilEngineeringandTransportation,SouthChinaUniversityofTechnology,Guangzhou510640,P.R.China;2.SouthChinaInstituteofGeotechnicalEngineering,Guangzhou510640,P.R.China;3.GuangzhouMassTransitEngineeringConsultantCo.,Ltd.,Guangzhou510010,P.R.China)Abstract:Thereareplentifulpotentialinfluentialparameterscollectedduringtheconstructionandmanyofthemdonothaveanexplicitcontributiontodisccutterwear.Accordingtotheoccurrencefrequencyof31typesofinfluentialparametersinpreviouscutterwearanalysisstudies,adatabasewith14inputparametersand2386samplesareobtainedfromthetunnelingsectionbetween2#launchshaftandventilationshaftofGuangzhouMetroLine18PanyusquaretoNancunWanboStationtodevelopBPneuralnetworkmodels.ModelsareoptimizedbySMBO(SequentialModel-basedOptimization)andGA(GeneticAlgorithm)andthosewithhigheraccuracyareputintosensitivityanalysistoinvestigatethecontributionoftheselected14infl...