Search

CN-121980188-A - Direct shear test soil parameter calibration method integrating numerical simulation and machine learning

CN121980188ACN 121980188 ACN121980188 ACN 121980188ACN-121980188-A

Abstract

The invention provides a direct shear test soil body parameter calibration method integrating numerical simulation and machine learning, and relates to the field of geotechnical engineering. The invention adopts the direct shear test soil parameter calibration method integrating the numerical simulation and the machine learning, and carries out systematic deep fusion on the high-fidelity finite element numerical simulation, KERNEL PCA dimension reduction and PINN agent model and Bayes parameter inversion for the first time. Through the gradual damage process from edge to inside and the non-uniform stress strain distribution of the soil body in the high-fidelity finite element numerical simulation accurate reduction direct shear test, the method has the advantages that The dimension reduction method extracts the low-dimension essential characteristics of the high-dimension response curve, and embeds the constitutive model and the intensity criterion And the model is proxied to ensure physical rationality, and finally, parameter estimation is realized through a Bayesian parameter inversion framework. The combination solves the problem that the stress strain of the soil body is uneven in the direct shear test and interferes with parameter calibration, so that the calibration result is more attached to the actual mechanical characteristics of the soil body.

Inventors

  • CUI WENJIE
  • DAI CHENGHAO
  • ZHANG SHUAIJIE
  • WU XIAOTIAN
  • FAN YIXIN

Assignees

  • 北京航空航天大学

Dates

Publication Date
20260505
Application Date
20260128

Claims (10)

  1. 1. The direct shear test soil parameter calibration method integrating numerical simulation and machine learning is characterized by comprising the following steps of: s1, performing an indoor direct shear test, recording a plurality of groups of positive stress and shear stress test results obtained by the test, and drawing a positive stress-shear stress curve I to obtain a preliminary soil body strength parameter; s2, based on the initial soil body strength parameter obtained in the S1 as an initial parameter input value, adopting an elastoplastic constitutive model and combining a fluid-solid coupling finite element numerical simulation method, and carrying out overall process fine simulation on a direct shear test to obtain a positive stress-shear stress curve II based on finite element numerical simulation calculation; S3, determining a soil body intensity parameter based on a finite element according to a positive stress-shear stress curve II obtained in the S2 and calculated based on finite element numerical simulation, comparing the positive stress-shear stress curve II with a positive stress-shear stress curve I obtained in the S1, and outputting the soil body intensity parameter based on the finite element as an actual soil body intensity parameter if the root mean square error of a fitting result is less than 5 percent; S4, establishing a Bayesian parameter inversion based system The machine learning agent model selects 20 groups of independent samples which do not participate in training in a parameter space, calculates the average absolute error of an agent model prediction curve and a high-fidelity simulation curve in S4, carries out migration test on different OCRs, verifies the prediction capability of the agent model in the whole parameter space in S4, and outputs calibrated soil body strength parameters, 95% confidence interval of quantization parameter reliability and the test result after verification Inverse mapping reconstructed positive stress-shear stress curve III; S5, comparing the positive stress-shear stress curve III obtained in the S4 with the positive stress-shear stress curve I obtained in the S1, outputting the calibrated soil body intensity parameter based on the agent model obtained in the S4 as the actual soil body intensity parameter if the root mean square error of the fitting result is less than 5%, and repeating the S4 iteration until the fitting result is within the allowable error range if the fitting result is not within the allowable error range, and outputting the calibrated final soil body intensity parameter.
  2. 2. The method for calibrating soil parameters of direct shear test by combining numerical simulation and machine learning according to claim 1, wherein in the indoor direct shear test of S1, axial stresses with different magnitudes are applied to the soil The shearing box is pushed horizontally, so that the upper shearing box and the lower shearing box are displaced relatively, and the shearing strain of soil body is observed Recording test results of multiple groups of positive stress and shear stress tests until the soil body reaches a damage condition, and drawing a positive stress-shear stress curve I: Obtaining an initial soil body strength parameter internal friction angle according to the slope and intercept of the positive stress-shear stress curve I And cohesive force 。
  3. 3. The direct shear test soil body parameter calibration method integrating numerical simulation and machine learning as claimed in claim 1, wherein the simulation process in S2 specifically performs the following steps: S21, setting a finite element grid and initial conditions; Establishing a two-dimensional high-fidelity finite element numerical simulation model in a plane strain state, dispersing by using a sufficient number of high-order constant parameter units, directly endowing a uniform initial stress field to the whole sample area, and adjusting the consolidation ratio Simulating stress states corresponding to different consolidation degrees; s22, setting boundary conditions, namely completely fixing a base of a shear box to simulate a stable support, applying horizontal displacement to the boundary of an upper shear box, keeping a lower shear box fixed, realizing a shear process through relative displacement, setting the horizontal displacement of a middle node of a contact surface of an upper box body and a lower box body to be 1/2 of the displacement of the upper box body, and simulating a rigid loading top cover at the top of the shear box through a unit layer with extremely high rigidity; S23, simulating a loading process by applying incremental horizontal displacement to the boundary of the upper shear box to simulate the shearing process, setting the shearing rate to be consistent with an actual experiment, and applying a vertical load to the center of the rigid top cover in a form of concentrated force, wherein the loading force is kept constant in the whole process; s24, finite element result calculation.
  4. 4. A direct shear test soil body parameter calibration method integrating numerical simulation and machine learning according to claim 3 is characterized in that in S24, horizontal counter forces of all nodes of an active boundary are accumulated in real time in the calculation process, the sum is total shear force of a sample, shear stress is obtained through the ratio of the applied horizontal total shear force to the initial length of a shear box set in a high-fidelity finite element numerical simulation grid, integral horizontal displacement and vertical displacement of a top cover are recorded, positive stress is obtained through the ratio of the vertical force applied by the top cover to the top area of the grid sample, and a positive stress-shear stress curve II is drawn according to the positive stress-shear stress curve.
  5. 5. The method for calibrating soil parameters for direct shear test by combining numerical simulation and machine learning according to claim 1, wherein S4 is established The proxy model has the following steps: s41, constructing an input parameter space, generating a data set and preprocessing; generating the data set comprises the steps of: The soil body intensity parameter based on finite element in S2 is the internal friction angle ' And cohesion ' As a core regulation parameter, combined consolidation ratio Building a complete input parameter space The positive stress-shear stress curve II which is output by high-fidelity finite element numerical simulation is taken as an output object Wherein Discrete sampling points for the curve covering the whole shearing process; based on the initial soil intensity parameter in S1 as the reference of interval definition, combining engineering experience and soil mechanical property definition Generating 500-2000 groups of input parameter samples by Latin hypercube sampling method, inputting single groups of samples into a high-fidelity finite element numerical simulation model one by one for calculation, outputting a corresponding 500-2000 groups of positive stress-shear stress curves II, and constructing an original data set containing input parameters and corresponding output curves ; The preprocessing process comprises inputting parameters And discrete data of the output positive stress-shear stress curve two Performing feature scale normalization processing to eliminate interference of different parameter dimension differences and numerical range differences on model training convergence speed and prediction accuracy; S42, performing high-dimensional output dimension reduction by adopting kernel principal component analysis KERNEL PCA; Introducing a Gaussian kernel function The method comprises the steps of mapping high-dimensional curve data to a low-dimensional feature space, calculating feature values and feature vectors of a kernel matrix, and calculating the contribution rate according to the accumulated variance Screening the number of optimal principal components by criteria of (2) Outputting the original high-dimensional output Compression into low-dimensional feature vectors Synchronous preservation Mapping matrix and kernel function parameters; S43, constructing PINN agent models and training fused with physical constraints; s44, building a Bayesian inversion framework; S45, performing result iteration optimization and parameter calibration; And S46, performing model generalization and result output.
  6. 6. The method for calibrating soil parameters for direct shear test by combining numerical simulation and machine learning according to claim 5, wherein in S43, PINN agent model comprises input layer, hidden layer, output layer and physical constraint layer; the input layer receives normalized input parameter space ; 4 Layers are arranged on the hidden layer, the number of neurons in each layer is 20, a ReLU activation function is adopted, and nonlinear association between input parameters and low-dimensional features is fitted through a depth network structure; The output layer directly outputs the low-dimensional feature vector And (3) with The feature space after dimension reduction corresponds to the feature space; The physical constraint layer is embedded into the model in the form of differential equation, and specifically comprises the following steps: Molar coulomb intensity criterion The corresponding constraint equation quantifies that the shear stress does not exceed the shear strength , Is shear strength; correcting the Cambridge constitutive model including elastic strain increment equation And the plastic strain delta equation In the elastic strain increment equation Is of Young's modulus, Is poisson's ratio, In order to achieve a shear modulus, the polymer is, Is an effective stress tensor increment, Is the effective stress increment of volume, As a Croneck function, in the incremental equation of plastic strain Is a plastic strain tensor increment, Is a plastic multiplier, M is a critical stress ratio, Is the current stress ratio, p is the average effective stress, Is effective stress tensor, and shear swelling softening association type contains shear swelling constraint And softening effect constraints Wherein 、 As a parameter of the initial intensity of the light, In order to accumulate the plastic strain, 、 Is the softening coefficient; Embedding PINN the above expressions into a loss function of the model; PINN model training was performed using a mixing loss function: Wherein To predict low-dimensional features Low dimensional features corresponding to the second positive stress-shear stress curve Is defined as the mean square error of Wherein ' Is a low-dimensional feature vector; For quantifying the deviation of the predicted outcome from the physical constraint; The physical constraint weight is used for balancing data fitting and physical consistency; Dividing the data set obtained in the step S41 into a training set, a verification set and a test set according to the proportion of 8:1:1, wherein the training process adopts Optimizer, initial learning rate is set as And is introduced into Regularization avoids overfitting; half learning rate and lowest decay to the minimum when the verification set error is continuously 50 rounds of iteration without reduction The error of the verification set is continuously 100 times without reduction or the total iteration number reaches 5000 times, and training is stopped and the final product is stored when any condition is met And (5) a proxy model.
  7. 7. The method for calibrating the soil parameters of the direct shear test by combining numerical simulation and machine learning according to claim 6, wherein the construction of the Bayesian parameter inversion framework in S44 comprises the following steps: S441, based on the initially determined soil body intensity parameter obtained in the S1 and combining the soil body intensity parameter range defined by engineering experience and soil body mechanical property, the inner friction angle of the core inversion parameter is obtained ' And cohesion ' Set non-information prior distribution Excessive interference of priori information on inversion result is avoided, and the consolidation ratio is improved Setting narrow distribution according to the actual measurement range of the test; S442, constructing a likelihood function, taking the positive stress-shear stress curve obtained by the indoor direct shear test in S1 as an inversion target, and performing Low-dimensional features of model output General purpose medicine The inverse mapping reconstruction obtains a positive stress-shear stress curve three marks of machine learning agent model inversion as Defining likelihood functions using a multivariate Gaussian distribution Wherein The positive stress-shear stress curve I obtained in the indoor direct shear test is an observation error covariance matrix, diagonal elements are test measurement error variances, and the test measurement error variances are obtained through result dispersion statistics of 3 parallel direct shear tests.
  8. 8. The method for calibrating the soil parameters of the direct shear test by combining numerical simulation and machine learning according to claim 7, wherein the step S45 specifically comprises the following steps: S451A posterior distribution of MCMC pair by adopting Markov chain Monte Carlo algorithm High-efficiency sampling is carried out, and iteration is continued until a sampling chain meets the following conditions A convergence criterion; S452, carrying out statistical analysis on a posterior sample set obtained by sampling through a Markov Chain Monte Carlo (MCMC) algorithm in S433, and taking the mean value of posterior distribution as a calibrated soil body strength parameter Outputting the calibrated soil parameters And (3) quantifying the uncertainty range of the calibration result.
  9. 9. The method for calibrating soil parameters in direct shear test by integrating numerical simulation and machine learning according to claim 8, wherein in S46, 20 groups of independent parameter samples which do not participate in training are selected, the average absolute error of a positive stress-shear stress curve III generated by calculation PINN and a positive stress-shear stress curve II calculated by finite element are combined with a multi-OCR scene migration test to ensure that PINN model is stable and reliable in a full parameter space under different working conditions, and after extensive verification is passed, calibrated soil parameters are output Confidence interval and channel And (3) inversely mapping the reconstructed positive stress-shear stress curve III.
  10. 10. The method for calibrating the soil parameters of the direct shear test by integrating numerical simulation and machine learning according to claim 8, wherein S5 is characterized in that according to the fitting degree indexes of the corresponding positive stress-shear stress curve III and the positive stress-shear stress curve I extracted in S46, the positive stress-shear stress curve III corresponding to the parameters of the sample with the minimum RMSE is screened out: If the positive stress-shear stress curve III and the positive stress-shear stress curve I Judging that the fit reaches the standard, and obtaining according to a positive stress-shear stress curve III The soil body intensity parameter after the calibration; If the positive stress-shear stress curve III and the positive stress-shear stress curve I Feeding back the current optimal parameters to an input parameter space, redriving PINN a model and finite element simulation iteration, generating a new positive stress-shear stress curve III, and verifying until the new positive stress-shear stress curve III meets the requirements And outputting the following results: first, fitting with the positive stress-shear stress curve A positive stress-shear stress curve III of (2); Fitting standard normal stress-shear stress curve three corresponding standard soil body parameters ; Thirdly, calibrating soil parameters Is a 95% confidence interval; RMSE specific values of the positive stress-shear stress curve three and the positive stress-shear stress curve one.

Description

Direct shear test soil parameter calibration method integrating numerical simulation and machine learning Technical Field The invention relates to the field of geotechnical engineering, in particular to a direct shear test soil body parameter calibration method integrating numerical simulation and machine learning. Background The direct shear test is to measure shear strength index of soil body, which comprises internal friction angleAnd cohesive forceIs the most basic indoor test method. The basic principle is that the soil sample is sheared by the relative displacement of the shearing box, and the normal stress of the soil sample is measured in the fixed stateShear strength under. In the traditional parameter determination method, most of soil samples in the indoor direct shear test are regarded as unit bodies which deform uniformly, and the normal stress-shear stress rule measured by the indoor test is directly adopted as the basis for calibrating the soil body strength parameter. However, in the direct shear test, a rigid shear box is adopted to apply shear deformation, so that soil body is gradually destroyed from edge to inside in the shear process, namely, the stress strain in the soil body is uneven, and further, the nonlinear development of a shear stress-displacement curve is initiated, the strength characteristic of the soil body can not be actually reflected, and the engineering applicability and reliability of the method are restricted. Therefore, the direct shear test soil body parameter calibration method integrating numerical simulation and machine learning is provided to solve the problems. Disclosure of Invention The invention aims to provide a direct shear test soil body parameter calibration method integrating numerical simulation and machine learning, which carries out systematic deep fusion on high-fidelity finite element numerical simulation, KERNEL PCA dimension reduction and PINN agent model and Bayes parameter inversion for the first time. Through the gradual damage process from edge to inside and the non-uniform stress strain distribution of the soil body in the high-fidelity finite element numerical simulation accurate reduction direct shear test, the method has the advantages thatThe dimension reduction method extracts the low-dimension essential characteristics of the high-dimension response curve,Constitutive model and strength criterion embeddingAnd the model is proxied to ensure physical rationality, and probabilistic parameter estimation is finally realized through a Bayesian parameter inversion framework. The combination solves the interference of uneven stress strain of the soil body to parameter calibration in the direct shear test, so that the calibration result is more fit with the actual mechanical property of the soil body, the engineering application range is widened, and the combination is widely applied to large-scale scientific research and practice. In order to achieve the above purpose, the invention provides a direct shear test soil body parameter calibration method integrating numerical simulation and machine learning, which comprises the following steps: s1, performing an indoor direct shear test, recording test results to obtain a plurality of groups of positive stress and shear stress test results, and drawing a positive stress-shear stress curve I to obtain an initial soil body strength parameter; s2, based on the initial soil body strength parameter obtained in the S1 as an initial parameter input value, adopting an elastoplastic constitutive model and combining a fluid-solid coupling finite element numerical simulation method, and carrying out overall process fine simulation on a direct shear test to obtain a positive stress-shear stress curve II based on finite element numerical simulation calculation; S3, determining a soil body intensity parameter based on a finite element according to a positive stress-shear stress curve II obtained in the S2 and calculated based on finite element numerical simulation, comparing the positive stress-shear stress curve II with a positive stress-shear stress curve I obtained in the S1, and outputting the soil body intensity parameter based on the finite element as an actual soil body intensity parameter if the root mean square error of a fitting result is less than 5 percent; S4, establishing a Bayesian parameter inversion based system The machine learning agent model selects 20 groups of independent samples which do not participate in training in a parameter space, calculates the average absolute error of an agent model prediction curve and a high-fidelity simulation curve in S4, carries out migration test on different OCRs, verifies the prediction capability of the agent model in the whole parameter space in S4, and outputs a calibrated soil body strength parameter, a 95% uncertainty interval of quantization parameter reliability and a channel after verificationInverse mapping reconstructed positive stress-shear stre