CN-121980653-A - Soft foundation consolidation degree rapid deduction method based on depth operator number intelligent model

CN121980653ACN 121980653 ACN121980653 ACN 121980653ACN-121980653-A

Abstract

The invention relates to a soft foundation consolidation degree rapid deduction method based on a depth operator digital intelligence model, which comprises the steps of obtaining soil layer thickness, a layered structure, consolidation coefficients of each layer and drainage boundary conditions of a soft foundation, carrying out dimensionless treatment on physical depth and physical time, generating a dimensionless consolidation coefficient field sample changing along with the dimensionless depth and dispersing at a plurality of fixed depth positions, inputting dimensionless time-depth coordinates into a main network of a physical information-depth operator network, inputting a consolidation coefficient discrete vector and soil layer distribution parameters into a branch network of the physical information-depth operator network, fusing output of the main network and output of the branch network to obtain an estimated value of dimensionless hyperstatic pore water pressure, and training the physical information-depth operator network based on a physical information loss function to obtain a soft foundation consolidation prediction model. The method can realize rapid and accurate deduction of the consolidation process.

Inventors

GUO JING
SONG BOKAI
ZHANG CHENWEI
PU QI
ZHANG LI
YE FEI
HAN XINGBO
LI LIN

Assignees

中铁十二局集团有限公司
中铁十二局集团第三工程有限公司
长安大学

Dates

Publication Date: 20260505
Application Date: 20260119

Claims (10)

1. A soft foundation consolidation degree rapid deduction method based on a depth operator number intelligent model is characterized by comprising the following steps: Acquiring the soil layer thickness, the layered structure, the consolidation coefficient of each layer and the drainage boundary condition of the soft soil foundation, carrying out dimensionless treatment on the physical depth and the physical time to obtain a dimensionless depth interval and a dimensionless time interval, and forming dimensionless time-depth coordinates; Generating a dimensionless consolidation coefficient field sample which changes along with the dimensionless depth in the dimensionless depth interval, and dispersing the dimensionless consolidation coefficient field sample at a plurality of fixed depth positions to form a consolidation coefficient discrete vector; Inputting the dimensionless time-depth coordinates to a backbone network of a physical information-depth operator network, inputting the consolidation coefficient discrete vector and soil layer distribution parameters to a branch network of the physical information-depth operator network, and fusing the output of the backbone network and the output of the branch network to obtain an estimated value of dimensionless hyperstatic pore water pressure; Training a physical information-depth operator network based on a physical information loss function to obtain a soft soil foundation consolidation prediction model, wherein the physical information loss function is constructed based on a one-dimensional consolidation control equation, an initial condition and a drainage boundary condition, and the soft soil foundation consolidation prediction model is used for predicting an ultra-static pore water pressure field according to a dimensionless consolidation coefficient field, a drainage boundary condition and soil layer distribution parameters of a target project.
2. The method for rapidly deducting soft foundation consolidation degree based on a depth operator number intelligent model according to claim 1, wherein the step of performing dimensionless treatment on physical depth and physical time to obtain a dimensionless depth interval and a dimensionless time interval and forming dimensionless time-depth coordinates comprises the following steps: Adopting the total soil layer thickness and the characteristic consolidation duration to carry out dimensionless treatment on the physical depth and the physical time to obtain a dimensionless depth interval and a dimensionless time interval: Wherein, the Is a depth of a dimensionless shape, In order to be a dimensionless time, the method comprises the steps of, Is the non-dimensional hyperstatic pore water pressure, For the physical depth of the object to be measured, In order for the physical time to be a physical time, To be at physical depth And physical time The physical hyperstatic pore water pressure is lower, Is characterized by the total thickness of the soil layer, For the duration of the characteristic consolidation, Is characterized by hyperstatic pore water pressure.
3. The method for quickly deducting soft foundation consolidation degree based on a depth operator number intelligent model according to claim 1, wherein generating a dimensionless consolidation coefficient field changing with a dimensionless depth in the dimensionless depth interval, and dispersing the dimensionless consolidation coefficient field at a plurality of fixed depth positions to form a consolidation coefficient discrete vector, comprises: selecting in the dimensionless depth interval Discrete depth points Defining a Gaussian random vector : Wherein, the To a dimensionless depth A gaussian random variable at; The Gaussian random vector Covariance matrix of (2) The method comprises the following steps: Wherein, the Is the standard deviation of the two-dimensional image, For the length of the correlation, Performing element-by-element exponential operation; For the covariance matrix Performing Cholesky decomposition and introducing the Gaussian random vector The dimensionless consolidation coefficient field samples at discrete points are defined as: Wherein, the Is Gaussian random vector Is used for the average value vector of (a), In the form of a lower triangular matrix, For random vectors subject to normal distribution of multiple criteria ; Forming a consolidation coefficient discrete vector according to the dimensionless consolidation coefficient field samples on the discrete points 。
4. The method for rapidly deducting soft foundation consolidation degree based on a depth operator number intelligence model according to claim 3, wherein the physical information-depth operator network comprises a backbone network, a branch network and an output layer, wherein, The input of the backbone network is dimensionless time-depth coordinates Outputting the trunk feature vector ; The input of the branched network is a consolidation coefficient discrete vector And soil layer distribution parameters, output as branch feature vectors ; The output layer carries out inner product operation on the main feature vector and the branch feature vector to obtain an estimated value of dimensionless hyperstatic pore water pressure: Wherein, the In consolidation coefficient field for physical information-depth operator network The estimated value of the non-dimensional hyperstatic pore water pressure is calculated, Is vector quantity 、 Is used in the manufacture of a printed circuit board, Is a dimensionless consolidation coefficient field.
5. The soft foundation consolidation rapid deduction method based on the depth operator number intelligent model according to claim 1, wherein the main network and the branch network both adopt a multi-layer feedforward full-connection structure, the multi-layer feedforward full-connection structure comprises an input layer, a plurality of hidden layers and an output layer, and an activation function of each hidden layer adopts a ReLU or a tanh.
6. The soft foundation consolidation rapid deduction method based on the depth operator number intelligent model according to claim 5, wherein the number of hidden layers in the main network is 6, and each layer contains 60 neurons; the number of hidden layers in the branched network is 5 layers, each layer containing 60 neurons.
7. The method for rapidly deducting soft foundation consolidation degree based on the depth operator number intelligent model according to claim 1, wherein the soil layer distribution parameters of the branched network of the physical information-depth operator network comprise permeability coefficients and volume compression coefficients.
8. The soft foundation consolidation rapid deduction method based on the depth operator number intelligent model according to claim 1, wherein the dimensionless form of the one-dimensional consolidation control equation is: Wherein, the Is the non-dimensional hyperstatic pore water pressure, Is a depth of a dimensionless shape, In order to be a dimensionless time, the method comprises the steps of, Is a non-dimensional coefficient, and is a non-dimensional coefficient, , Is characterized by the total thickness of the soil layer, For the duration of the characteristic consolidation, Is a dimensionless consolidation coefficient field sample; The initial conditions are: Wherein, the The water pressure distribution of the ultra-pore along the depth at the initial loading moment; The drainage boundary conditions comprise top single-sided drainage and top double-sided drainage; The boundary conditions of the top single-sided drainage are: the boundary conditions of the top-bottom double-sided drainage are as follows: 。
9. The soft foundation consolidation level rapid deduction method based on the depth operator number intelligent model according to claim 8, wherein the physical information loss function is: Wherein, the As a function of the total loss, As a partial differential equation loss function, As a function of the loss of the initial condition, A loss function for boundary conditions; non-negative weight coefficients for each loss term; the partial differential equation loss function Expressed as: Wherein, the Is the number of partial differential equation collocation points, Respectively the first The partial differential equation matches the dimensionless depth and dimensionless time of the points, In consolidation coefficient field for physical information-depth operator network The estimated value of the non-dimensional hyperstatic pore water pressure is calculated, For the partial derivative of the dimensionless time, For the second partial derivative of the dimensionless depth, Is a dimensionless coefficient; the initial condition loss function Expressed as: In the formula, For the number of the matching points of the initial condition, Is the first The initial conditions match the dimensionless depth of the points, An initial dimensionless hyperstatic pore water pressure at the depth; The boundary condition loss function The expression of (2) is: Wherein, the As a function of the top boundary loss, Is a bottom boundary loss function; Top single-sided drainage boundary conditions Lower, top boundary loss function The method comprises the following steps: Wherein, the Is the number of top boundary collocation points, Is the first Dimensionless time of the top boundary collocation points; Top-bottom double-sided drainage boundary Bottom boundary loss function at the time The method comprises the following steps: Top single-sided drainage boundary conditions Bottom boundary loss function at the time The method comprises the following steps: Wherein, the Is the number of the bottom boundary collocation points, Is the first Dimensionless time of the bottom boundary collocation points.
10. The soft foundation consolidation rapid deduction method based on the depth operator number intelligent model according to claim 1, wherein training the physical information-depth operator network based on the physical information loss function to obtain a soft foundation consolidation prediction model comprises the following steps: Updating parameters of a physical information-depth operator network by adopting an Adam optimizer, wherein the initial learning rate is At decay rate according to inverse time decay strategy Dynamically adjusting the learning rate, wherein the training iteration step number is 10000, the batch size is 32, the sample number ratio of the training set to the test set is 8:2, and the physical information loss function is gradually reduced to obtain a soft soil foundation consolidation prediction model.

Description

Soft foundation consolidation degree rapid deduction method based on depth operator number intelligent model Technical Field The invention belongs to the technical field of geotechnical mechanics and engineering, and particularly relates to a soft foundation consolidation degree rapid deduction method based on a depth operator number intelligence model. Background The soft foundation widely exists in roadbed filling, foundation treatment and structural foundation design, and accurate and efficient prediction of consolidation degree is an important foundation for soft foundation design, construction control and post-construction settlement evaluation. The dissipation process of pore water pressure in natural soft soil is controlled by multiple factors such as drainage boundary conditions, soil layer distribution, consolidation coefficient space variability and the like, and analytical solution is generally only applicable under ideal conditions of simple soil layer and uniform consolidation coefficient. In practical engineering, stratum generally presents layering heterogeneous characteristics, consolidation coefficients obviously change along with depth, drainage boundaries are variously combined, and approximate solution is often needed by a numerical method. However, the traditional numerical analysis has high calculation cost when repeatedly calculating multiple working conditions, cannot meet the real-time evaluation requirement, and is difficult to meet the requirements of soft foundation design and informatization construction on quick evaluation, dynamic adjustment and intelligent decision. The soft foundation consolidation analysis methods commonly used in the current engineering are mainly a Finite Element Method (FEM) and a Finite Difference Method (FDM). The method constructs a grid and a constitutive model by carrying out space-time dispersion on a consolidation control equation, and solves the space-time distribution of pore water pressure and consolidation degree. For a given working condition, the result is reliable, but when the drainage condition, the stacking process or the filling height is frequently adjusted, repeated modeling and recalculation are often needed, the modeling cost and calculation cost are high, when the consolidation coefficient is in strong uneven distribution or represented by a random field, the degree of freedom is sharply increased, the calculation efficiency is further reduced, and the practical requirements of multi-working condition, rapid evaluation and informatization construction are difficult to meet. Along with the development of artificial intelligence, data driving models such as a support vector machine, an ensemble learning model, a convolutional neural network and a cyclic neural network are introduced into soft foundation consolidation prediction, and rapid estimation of consolidation degree or sedimentation is realized through an input-output relation in learning historical data. However, the method is highly dependent on large-scale and high-quality samples, is sensitive to data distribution, does not explicitly consider physical constraints such as a consolidation control equation, a drainage boundary and the like, is easy to have the conditions of insufficient prediction precision and limited generalization capability, and once working conditions exceed the coverage range of training data, the performance of the model is obviously reduced. In order to make up for the defect that the pure data driving method is separated from a physical mechanism, the Physical Information Neural Network (PINNs) can simultaneously restrict data errors and physical residual errors in the training process by embedding the partial differential control equation and the residual errors of initial/boundary conditions thereof into a loss function, so that a certain effect is achieved in the problems of soft soil consolidation prediction, consolidation coefficient inversion and the like. However, under the conditions of multi-parameter working conditions, long-term consolidation and complex boundary conditions, PINNs is easy to have unstable training, slow convergence and even difficult convergence, and the network is usually required to be trained for different working conditions, so that 'one-time training and multi-working condition prediction' are difficult to realize. The depth operator network (DeepONet) is based on an operator general approximation theory, solves a partial differential equation to be regarded as an operator learning problem from an input function space to a solution function space, encodes input functions (such as consolidation coefficient fields) through a branch network, encodes time-space coordinates through a main network, and can rapidly infer a large number of new input functions once operator training is completed, so that multi-working-condition rapid prediction is realized. The physical information DeepONet introduces control equation residua