基于孔压静力触探试验测试数据的原位固结系数物理信息神经网络反演方法  

A physics-informed neural networks inversion method for in-situ consolidation coefficient based on piezocone penetration test pore pressure data

作　　者：李林[1,2] 左林龙 胡涛涛 宋博恺 LI Lin;ZUO Lin-long;HU Tao-tao;SONG Bo-kai(School of Highway,Chang’an University,Xi’an,Shaanxi 710064,China;Xi’an Key Laboratory of Geotechnical Engineering for Green and Intelligent Transport,Chang’an University,Xi’an,Shaanxi 710061,China)

机构地区：[1]长安大学公路学院,陕西西安710064 [2]长安大学西安市绿色智慧交通岩土工程重点实验室,陕西西安710061

出　　处：《岩土力学》2024年第10期2889-2899,共11页Rock and Soil Mechanics

基　　金：国家自然科学基金(No.52108297);中国博士后基金面上项目(No.2021M692742);中国博士后基金特别资助项目(No.2023T160560);陕西省秦创原引用高层次创新创业人才项目(No.QCYRCXM-2022-29);中央高校基本科研业务费(No.300102212301,No.300102214303)。

摘　　要：固结系数是软基沉降计算和稳定性分析的关键参数,现有固结系数原位测试方法存在耗时长且精度低的缺点。根据孔压静力触探试验(piezoconepenetrationtest,简称CPTU)贯入机制与锥肩超孔隙水压力消散模式,采用圆孔扩张理论和轴对称固结方程描述CPTU锥肩超孔隙水压力的形成、发展和消散过程,利用神经网络自动微分功能将轴对称固结方程嵌入深度神经网络,通过物理方程损失函数、边界条件损失函数和初始条件损失函数形成神经网络的物理信息约束,同时将CPTU孔压测试数据作为数据驱动项,以最小化超孔隙水压力损失函数为优化目标,建立了CPTU孔压测试数据反演场地原位固结系数的物理信息神经网络(physics-informed neural networks,简称PINNs)模型。通过已有离心模型试验数据反演验证了PINNs模型反演场地原位固结系数的有效性,并利用CPTU孔压测试数据分析了PINNs模型反演原位固结系数的鲁棒性。结果表明:提出的PINNs模型能够有效利用CPTU孔压测试数据快速准确地反演场地原位固结系数;由于模型融入了物理机制约束,所需训练数据量少,且对有噪声的孔压测试数据具有较强的鲁棒性和泛化性能,为准确、快速可靠测试场地原位固结系数提供了有效途径。The consolidation coefficient is a crucial parameter for settlement calculation and stability analysis of soft foundations.Existing in-situ testing methods for the consolidation coefficient have the disadvantages of time-consuming and low accuracy.Based on the penetration mechanism of piezocone penetration test(CPTU)and the dissipation pattern of excess pore water pressure at the cone shoulder,the formation,development,and dissipation processes of excess pore water pressure at the CPTU cone shoulder are described using the theory of circular cavity expansion and the axisymmetric consolidation equation.By incorporating the automatic differentiation capability of neural networks,the axisymmetric consolidation equation is embedded into a deep neural network.The physical information constraints of the neural network are formed through the loss functions of physical equations,boundary conditions,and initial conditions.At the same time,the CPTU pore pressure test data serve as a data-driven term.Consequently,with the minimization of the excess pore water pressure loss function as the optimization goal,a physics-informed neural networks(PINNs)model is established for inversely analyzing the in-situ consolidation coefficient using CPTU pore pressure test data.The effectiveness of the PINNs model in inversely analyzing in-situ consolidation coefficient is verified through example analysis and inversion validation using existing centrifuge test data.The robustness of the PINNs model is also analyzed using CPTU pore pressure test data.The results indicate that the proposed PINNs model can effectively use CPTU pore pressure test data to rapidly and accurately invert the site in-situ consolidation coefficient.Due to the integration of physical mechanism constraints,the model requires only a small amount of training data and exhibits strong robustness and generalization performance against noisy pore pressure test data,providing an effective approach for accurate,rapid,and reliable testing of the in-situ consolidation coeffici

关 键 词：原位固结系数 静力触探 孔压测试数据 固结方程 物理信息神经网络 参数反演 

分 类 号：TU447[建筑科学—岩土工程]