基于人工神经网络多模型迁移学习的隧(巷)道机械化掘进装备控制参数自主决策方法

An autonomous decision-making method for mechanized tunneling equipment control parameters based on transfer learning of multiple ANN models

TBM工法是一种应用于隧(巷)道建设的机械化掘进方法。为实现TBM稳态段掘进过程中装备自身对操作参数的优化决策,建立一种基于迁移学习思想的TBM控制参数自主决策模型。首先提出核心优化策略:在充分尊重现场经验的基础上,提高掘进效率、降低掘进能耗,并以稳态段贯入度(PRev)、刀盘转速(RPM)为输出参数完成数学建模;其次构建优化决策模型,模型基本架构采用深度神经网络,包含2个源域子模型和1个目标域主干模型,分别执行控制参数回归、破岩比能预测、控制参数优化决策任务,并通过迁移学习和网络层冻结的方法实现了源域与目标域的统一;而后确定模型关键超参数取值,采用正交试验和贝叶斯优化相结合的方式确定了源域子模型的最优超参数组合,基于层次分析法确定了目标域主干模型目标函数的关键权重;最后,依托吉林引松供水工程总干线输水隧洞四标段TBM施工数据集(4459组有效数据),对所建立模型进行训练和测试。结果表明,TBM稳态段净掘进速率平均提升了15.55%,刀盘破岩比能平均下降了7.13%,并降低了方差,改善了稳态段掘进过程的整体平稳性。研究成果可为长大隧道、矿山巷道建设工程中的TBM智能化控制系统提供技术支撑。

TBM method is a mechanized excavation process applied to roadway and tunnel construction.In order to realize the optimization of operating parameters by automatic execution of TBM in the process of stable tunneling,an autonomous decision-making model of TBM control parameters based on the transfer learning method is established.First,the core optimization strategy is proposed,which aims to improve the tunneling efficiency and reduce the energy consumption on the basis of fully reference to manual experience.Taking the penetration rate per revolution(PRev)and the cutterhead revolutions per minutes(RPM)as output parameters,the mathematical modeling is completed.Secondly,an optimization model is constructed.Deep artificial neural networks is used as the basic architecture,including two auxiliary networks in the source domain and one main network in the target domain,which respectively performs the regression of control parameters,the prediction of rock breaking specific energy,and the optimization of control parameters.The methods of transfer learning and frozen-layers are adopted to achieve the unification of the source and target domain.Third,key hyperparameters are identified.The optimal hyperparameters of auxiliary networks are determined by a combination of orthogonal experiments and Bayesian optimization.The key weights of the objective function of the main network in the target domain are determined based on analytic hierarchy process.Finally,relying on the TBM construction data set(4459 effective data)collected from the transferring water project from Songhua River in Jilin,the established model is trained and tested.The results show that in stable phase,the heading efficiency is increased by 15.55%on average,the energy consumption is decreased by 7.13%on average,and the variance is reduced,which improves the overall stability.The research results can provide technical support for the establishment of TBM intelligent control system.