随着城市地下空间的大规模建设，盾构工程逐渐朝着“大断面、深覆土、长距离”趋势发展，这使得盾构施工环境更加复杂，盾构施工性能与操作参数间关系的复杂性也随之增加。同时，各施工性能目标间（例如：盾构施工的安全与效率）存在一定的关联性与冲突性，这样盾构操作参数的设定和施工性能提升面临极大挑战。为提升盾构施工综合性能，本文选取地表沉降、刀具磨损、掘进能耗效率与比能四个指标为优化控制目标，其中掘进能耗效率与比能通过能耗计算模型构建其优化目标函数，对于与操作参数关系复杂的地表沉降、刀具磨损进一步运用数值仿真、机器学习、机理分析等相关技术分析复杂映射关系以支持其优化目标函数建立，在此基础上对所选四个目标的优化求解框架进行研究，以实现盾构操作参数的智能决策。主要包括：

（1）针对地表沉降与盾构操作参数的复杂映射关系，建立基于有限元的地表沉降代理模型以快速求解地表沉降。首先在 Plaxis 3D 中建立盾构分步开挖模型以仿真计算沉降，其得到的仿真沉降符合沉降监测规律；然后利用拉丁超立方抽样获取输入参数组合并仿真对应沉降输出，构建有限元-沉降数据集；基于该数据集训练 SAGRNN（Simulated Annealing-General Regression Neural Network）算法构建代理模型，代理模型评估沉降的精度高（R2=0.89），且计算效率（单环求解时间约 1 s）远高于限元仿真，对沉降优化目标函数的建立有较好支持性。
（2）针对刀具磨损与盾构操作参数的复杂映射关系，建立基于机理-数据混合驱动的刀具磨损预测模型以准确预测刀具磨损。首先基于 Archard 模型中粘着磨损与磨粒磨损推导预测刀具磨损量的显式公式；然后对公式中难确定的磨损系数采用神经网络进行预测，且将由相关文献确定的磨损系数阈值作为约束添加到神经网络的损失函数中进行训练；最终得到的刀具磨损预测值准确度高、泛化性能好，测试集中仍能保持较好预测性能（R2=0.8879），且较少出现预测值大幅偏离实际值，对刀具磨损优化目标函数的建立有较好支持性。
（3）针对具有多个优化目标的盾构施工操作参数优化问题，建立基于目标重要性的多目标优化框架以支持盾构操作参数的选取。首先利用语言判断矩阵对目标重要性进行群体决策评估，计算目标重要性排序向量；然后基于目标重要性对优化问题进行分层序列求解，得到了第一层为地表沉降优化，第二层为刀具磨损、掘进能耗效率、掘进比能优化的分层优化模型；基于多目标遗传算法，从上到下求解得到优化问题的 Pareto 解集，在 Pareto 解集中基于目标重要性使用加权和法选取最优解（即最优操作参数）。
本文提出的盾构操作参数优化框架可辅助盾构操作参数设置，有助于提高盾构施工安全、成本、效率等性能。同时本文的研究方法也为盾构施工性能-操作参数关系分析，盾构操作参数智能设定及无人盾构技术建立了理论基础。

关键词：盾构操作参数；多目标优化；代理模型；机理-数据混合驱动；分层序列法

With the development of large-scale underground space construction in cities, shield tunneling engineering is gradually developing towards the trends of "larger cross-section, deeper burial depth and longer distance", which increase the complexity of the engineering environment and the difficulties in tunnel construction. There are several correlations and conflicts among different tunneling performance objectives (e.g., construction safety and efficiency), resulting challenges in Tunnel Boring Machine (TBM )operational parameters setting and tunneling performance improvement.
In order to assist in the decision-making of TBM operational parameters and optimize the comprehensive performance of shield tunneling construction, four key performance indexes (i.e., ground settlement, cutter wear, excavation energy efficiency, and specific energy) are selected as optimization objectives. For excavation energy efficiency and specific energy, their calculation model are developed based on energy consumption. For the tunneling performance with complex relations to operational parameters (i.e., ground settlement and cutter wear), numerical simulation, machine learning, mechanism analysis are engaged to determine the complex mapping relationships, which support the construction of objective function in optimization problems. On such basis, the optimization framework of the selected objectives are studied. The main content as follow:
(1) For the complex mapping relationship between ground settlement and shield TBM operational parameters, a machine learning surrogate model for tunneling-induced settlement based on the finite element model is proposed. Plaxis 3D (a finite element software) is first used to establish shield tunneling step excavation model for settlement simulation, and simulation settlement fits the monitoring settlement development rules well; then, Latin hypercube sampling is used to obtain input parameter combinations and corresponding settlement outputs are obtained by simulation, to construct a finite elementsettlement dataset; based on such dataset, a SA-GRNN (Simulated Annealing-General Regression Neural Network) model is trained as a surrogate model for the finite element model, the surrogate model has high accuracy (R2=0.89) in evaluating settlement and its computational efficiency (computing time for a ring is about 1s) is much higher than that of finite element simulation, which provides good support for establishing the optimization objective function of settlement.
(2) For the complex mapping relation between TBM cutter wear and TBM operational parameters, a TBM cutter wear prediction model based on a mechanism-data hybrid driven approach is proposed. The average cutter wear prediction calculation formula is firstly deduced from the Archard’s Adhesive wear model and Abrasive wear model; for the wear coefficient that is difficult to determine in the formula, a neural network is used to predict based on historical data, and the wear coefficients threshold determined according to the related literature is added to the loss function of the neural network for training; the hybrid driven prediction model shows high accuracy and good generalization performance, the R2 of testing set can still reach 0.8879, and the serious error’s probability of the predicted value is much smaller than other prediction model, which provides good support for establishing the optimization objective function of cutter wear.
(3) For the solution of the TBM operational parameters optimization problems with multiple objectives, a multi-objective optimization framework based on objectives’ importance is constructed. At first, the importance of the objectives is evaluated by the language judgment matrix, and the objectives’ importance ranking vector is calculated; then, based on the importance of objectives, the optimization problem is solved by Stratified Sequencing Method, and the first stratum is for ground settlement optimization, the second stratum is for cutter wear, excavation energy efficiency, and specific energy optimization; at last, the Pareto solution set of the optimization problem is solved from the first stratum to the last stratum based on multi-objective genetic algorithms, and the optimal solution, i.e., the optimal operational parameters, is selected from the Pareto solution set based on the Weight Sum Method.
The proposed TBM operational parameters optimizing framework can assist in setting TBM parameters and contribute to improving performance in terms of construction safety, cost, and efficiency. Furthermore, the employed research methodology establishes a theoretical foundation for analyzing the relationship between shield tunneling performance and operation parameters, as well as for unmanned TBM.