Real-time inversion method for TBM extrusion load probability based on agent model
By using a proxy model-based approach, the problem of real-time monitoring of compression loads during TBM tunneling was solved, achieving efficient and accurate load estimation and early warning. This overcomes the computational time and uncertainty issues of traditional methods, and meets the real-time and risk quantification requirements of TBM construction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SINOHYDRO BUREAU 14 CO LTD
- Filing Date
- 2026-02-12
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies make it difficult to achieve accurate and real-time monitoring of compression loads during TBM tunneling, especially in deeply buried soft and fractured strata or high-stress fault fracture zones, leading to construction delays and economic losses. Furthermore, traditional methods suffer from time-consuming calculations and unpredictable and uncertain solutions.
By employing a surrogate model-based approach, a mapping relationship between the compressive load and the TBM structural response is constructed by defining the prior distribution of the compressive load vector. Combining sensor data and likelihood functions, a Bayesian framework is used for probabilistic inversion to achieve real-time estimation and early warning of the compressive load.
It achieves millisecond-level response to compression loads, breaking through the lag of traditional high-fidelity numerical simulations, accurately inverting non-uniform loads, quantifying the uncertainty of inversion results, and meeting the real-time and risk assessment requirements of TBM tunneling.
Smart Images
Figure CN122287295A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of tunnel engineering and rock mechanics, specifically to a real-time inversion method for the probability of TBM compression load based on a surrogate model. Background Technology
[0002] Full-face tunnel boring machines (TBMs) have become the preferred equipment for tunnel construction due to their high efficiency and safety. However, when traversing deeply buried, weak, and fractured strata or high-stress fault fracture zones, they are highly susceptible to large deformations of the surrounding rock and shield crushing accidents. If the TBM gets stuck, it will cause serious delays and economic losses. Therefore, accurately and in real-time monitoring of the magnitude and distribution of the crushing load acting on the shield is crucial for guiding the adjustment of TBM tunneling parameters and preventing TBM jamming accidents.
[0003] Currently, methods for obtaining shield compression loads are mainly divided into two categories: theoretical calculation methods and field monitoring inversion methods. Theoretical calculation methods are mainly based on classical formulas such as the Fenner-Pacher curve and the Terzaghi loosened earth pressure theory. However, these theories typically assume that the rock mass is an isotropic continuous medium and that the stress on the shield is regularly distributed (e.g., uniform or elliptical). In actual engineering, fractured surrounding rock exhibits significant discontinuous, nonlinear, and anisotropic characteristics, and the load distribution often presents a highly irregular shape, leading to large deviations between theoretical calculation results and actual conditions, making it difficult to guide construction.
[0004] Numerical inversion based on on-site monitoring: This method acquires monitoring data by installing strain gauges or displacement sensors inside the shield and then performs inverse analysis using numerical simulation. It is currently a highly accurate method. However, existing technologies suffer from two main bottlenecks: (1) In order to accurately simulate the large compression deformation of fractured rock mass and the contact behavior of rock mass-shield, a high-precision three-dimensional numerical model is usually required. Such high-fidelity numerical simulations involve huge computational loads, and a single calculation often takes several hours or even days. Inversion based on swarm intelligence algorithms or gradient descent usually requires calling thousands of numerical models, resulting in extremely long inversion time and serious lag, which cannot meet the real-time requirements of TBM tunneling.
[0005] (2) Most existing inversion methods are deterministic inversions, which use optimization algorithms to find a set of "mathematically optimal" load parameters to make the calculated values closest to the monitored values. However, such methods face two major challenges: First, they ignore random uncertainty: the geological parameters of deep rock masses have great spatial variability, and sensor monitoring data inevitably contain noise. Deterministic inversion ignores these factors, only providing a single numerical result, and cannot provide the confidence interval or probability distribution of the inversion results, making it difficult to quantify engineering risks. Second, there is a serious problem of ill-posed solutions: due to the limited number of monitoring points on the shield, and the continuous and complex spatial distribution of external compression loads, the inversion process faces the problem of insufficient information. This leads to the non-uniqueness of the inversion solution, that is, there may be many completely different load distribution patterns that can produce the same sensor response. Deterministic inversion is very prone to getting trapped in local optima or converging to a mathematically valid but physically unreasonable "pseudo-true value", thus misleading construction decisions.
[0006] Therefore, there is an urgent need to develop a real-time inversion method for TBM extrusion load probabilities that can retain the computational accuracy of high-fidelity numerical simulation, significantly reduce computation time, effectively overcome the ill-posedness of the solution caused by insufficient monitoring information, and fully quantify the uncertainty of the inversion results. Summary of the Invention
[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing a real-time inversion method for the probability of TBM compression load based on a surrogate model.
[0008] The objective of this invention is achieved through the following technical solution: Step S1: For the TBM shield tunneling environment, define the compression load vector to be inverted, and determine the prior distribution of each component of the load vector; Step S2: Establish a load-TBM structure model of the TBM shield, use an active learning strategy to iteratively select sample points for simulation calculation, and build and train a surrogate model to describe the mapping relationship between the crushing load and the TBM shield structure response. Step S3: Obtain the measured shield response data at the TBM tunneling site, and construct a likelihood function that characterizes the degree of matching between the observed data and the predicted data by combining the observation error of the sensor and the prediction error of the surrogate model. Step S4: Based on the likelihood function and prior distribution, solve for the posterior probability density distribution of the compressive load to obtain the optimal estimate and confidence interval of the compressive load; Step S5: Set the allowable crush load threshold for the TBM shield. When the failure probability is higher than the preset warning line, a graded early warning signal will be automatically triggered.
[0009] Furthermore, the compression load vector mentioned in step S1 is described using a discretized model based on the load mesh; specifically: The outer surface of the TBM shield is divided into several independent load-bearing grids in the circumferential and axial directions. The compression load vector is composed of the pressure value components at the center nodes of each grid; or a spatial interpolation function is used to map the pressure values of a finite number of control nodes into a continuous pressure distribution surface.
[0010] Furthermore, when defining the compression load vector, the intrinsic orthogonal decomposition method is used to reduce the dimensionality of the high-dimensional load parameters; specifically: Construct a snapshot matrix containing multiple typical load distribution patterns, calculate its eigenvalues and eigenvectors, and select the top... k Using the principal eigenvectors as basis functions, the compression load vector to be inverted is represented as follows: k A linear combination of basis functions, thereby transforming the inversion objective into a linear combination of basis functions. k Inversion of weighted coefficients.
[0011] Further, in step S1, the specific steps for determining the prior distribution of the load vector are as follows: Model Construction: A refined numerical model of the surrounding rock-TBM interaction system is established. The numerical model adopts any one of the following methods: continuum mechanics, discontinuous mechanics, or continuous-discontinuous coupling method. Stochastic simulation: Obtain the statistical characteristics of the physical and mechanical parameters of the surrounding rock, and perform multiple forward calculations on the refined numerical model using Monte Carlo simulation or stochastic finite element method; Prior generation: The TBM shield contact force data obtained by statistical forward calculation is fitted to obtain the probability density function of the compression load, which is used as the prior distribution of the compression load vector in the subsequent probability inversion.
[0012] Furthermore, based on geological survey data or design parameters, the a priori range of values for the load vector is determined, specifically as follows: A refined numerical model of the surrounding rock-TBM system was established to obtain the probability distribution of TBM compression load, which was used as the prior distribution of compression load.
[0013] Furthermore, the active learning strategy in step S2 specifically includes: First, a small number of initial sample points are generated using Latin hypercube sampling, and high-fidelity numerical calculations are performed to establish an initial surrogate model. Then, an iterative loop of prediction-evaluation-addition (increasing sample points) is entered. The potential value of each candidate point in the parameter space is evaluated using the acquisition function, and the point that contributes the most to improving the model accuracy is dynamically selected as a new sample point for supplementary calculation. The training set is continuously updated and the model is retrained until the prediction accuracy or convergence of the surrogate model meets the preset termination condition.
[0014] Furthermore, in step S2, the surrogate model is constructed using any one or a combination of Gaussian process regression, support vector regression, radial basis function neural network, or neural network.
[0015] Furthermore, in step S3, the shield response data includes strain data, deformation displacement data, and segment contact pressure data collected by sensors arranged on the inner surface of the shield.
[0016] Furthermore, step S4 is specifically executed as follows: Within the Bayesian framework, the Markov Chain-Monte Carlo (MCMC) algorithm is used to sample the posterior distribution of load parameters. During each iteration of MCMC sampling, the surrogate model is invoked to replace the high-fidelity numerical simulation to calculate the likelihood function value, thereby achieving real-time inversion.
[0017] Furthermore, the inversion process is carried out continuously following the TBM tunneling process, employing a sequential Bayesian update strategy; specifically: ... t The posterior probability distribution of the compressive load obtained by time-inversion is used as... t The prior probability distribution inverted at time +1 is used to continuously revise the estimate of the compression characteristics of the strata ahead using historical tunneling data.
[0018] The beneficial effects of this invention are: This invention utilizes a surrogate model with millisecond-level response to replace time-consuming high-fidelity numerical simulations, compressing the probabilistic inversion cycle to minutes or even seconds. This effectively solves the lag problem of traditional high-fidelity numerical inversion and meets the real-time sensing requirements for compression loads at TBM tunneling sites. Simultaneously, it employs a grid-based discretization parameterization method, breaking through the traditional theoretical assumptions regarding regular load shapes. This enables accurate inversion of complex distortion loads during TBM tunneling, such as non-uniformity, local bias, and fault displacement. Furthermore, based on a Bayesian probabilistic framework, this invention effectively alleviates the ill-posedness of solutions caused by insufficient monitoring information, while overcoming the limitation of traditional deterministic inversion in assessing risk, thus quantifying the uncertainty of the inversion results. Combined with a sequential Bayesian update strategy, it fully utilizes historical tunneling data, enabling the inversion system to have continuous correction capabilities and improving prediction stability in long-distance tunneling. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.
[0020] Figure 1 The flowchart below shows a real-time inversion method for the probability of TBM extrusion load based on a surrogate model according to the present invention. Figure 2 This is a structural diagram of a real-time inversion device for TBM compression load probability based on a surrogate model according to the present invention. Detailed Implementation
[0021] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0022] like Figure 1 As shown, this invention provides a real-time inversion method for the probability of TBM compression load based on a surrogate model. The method includes the following steps: Step S1: For the TBM tunneling environment, define the compression load vector x acting on the TBM shield to be inverted, and determine the a priori value range of each component of the load vector based on geological survey data or design parameters. The specific steps are as follows: S11: To accurately describe the non-uniform, locally biased loads that may occur in deep strata, the outer surface of the TBM shield is divided into grids in the circumferential and axial directions. The original physical parameters to be inverted are defined as the pressure values p at each load grid node. i The compression load vector is then expressed as: In the formula, N Represents the number of nodes in the load mesh. For the first N The pressure vector of each load grid node.
[0023] S12: Based on the engineering geological survey report, conduct refined numerical model calculations to determine the physical upper and lower limits of the pressure vector at each load grid node, and assume that it follows a uniform distribution as a priori distribution. p prior (x).
[0024] S13: If there are too many grid nodes (more than 20), direct inversion will lead to the "curse of dimensionality". Intrinsic orthogonal decomposition can be used to reduce dimensionality. The specific implementation steps are as follows: S131: Generation based on analytical solution n p The sample load field matrix X, covering various working conditions (e.g., uniform, biased, fault displacement), is expressed as: In the formula, Indicates the first n p Group sample load vector; S132: Perform singular value decomposition on the sample load field matrix X, and extract the initial... k The principal eigenvectors of the maximally singular values are Φ1, Φ2, ..., Φ3. k As a basis function; S133: Define the compression load vector x to be inverted as this k A linear combination of principal eigenvectors, i.e., defining the weighted coefficient vector w = [ w 1 , w 2 , ..., w k ] T This transforms the direct inversion of the compression load vector x into solving for the weighted coefficient vector w.
[0025] Step S2: Establish a high-fidelity numerical simulation model of the TBM shield (hereinafter referred to as the "high-fidelity model"). Employ an active learning strategy to iteratively select sample points for simulation calculations, and construct and train a surrogate model M(x) to describe the mapping relationship between the compressive load and the structural response of the TBM shield. S21: Gaussian process regression is selected as the surrogate model. When the amount of data is extremely large, deep neural networks can also be selected. S22: Generate a small number (e.g., 20-50) of initial points in the parameter space using Latin hypercube sampling, call the high-fidelity model to calculate the corresponding shield response, and train the initial proxy model; S23: Construct an adaptive acquisition function based on the LOLA-Voronoi criterion: S231: Within the preset parameter value range, generate a large-scale candidate sample point set using random sampling, Latin hypercube sampling, or grid sampling methods. Χ ; S232: Calculate the Local Linear Approximation (LOLA) score: Determine candidate sample points x using a nearest neighbor search algorithm. i ∈ Χ The neighborhood point set {x j y j A local linear prediction model is constructed using the least squares method, and the local Jacobian matrix J at that point is calculated. i Define the LOLA score: S233: Perform Voronoi cell partitioning on the current sample set and calculate x for each sample point. i The volume or equivalent diameter of the corresponding cell cavityV (x i ).
[0026] S234: Define the acquisition function H (x i The sum of the two indicators mentioned above is: in, W 1 and W 2 represents the weighting coefficient. The candidate point x that maximizes the acquisition function is searched from the candidate sample point set. new This will be used as the input parameter for the next high-fidelity numerical model.
[0027] S24: Repeat S23, and stop iterating when the model's prediction variance is less than a preset threshold.
[0028] Step S3: Obtain measured shield response data y at the TBM tunneling site obs Combining the observation error of the sensor and the prediction error of the surrogate model M(x), a likelihood function is constructed to characterize the degree of matching between the observed data and the predicted data. The prior distribution of the load vector x is then updated based on the likelihood function. The specific steps are as follows: S31: The error between the predicted data and the observed data can be assumed to be: In the formula, χ meas It is the measurement error of the sensors at each observation point, χ surr It is a surrogate model error; S32: Optionally, assuming that the measurement errors follow a normal distribution with a mean of 0 and are independent of each other, construct the total covariance moment Σ=Σ meas +Σ surr Under the premise that the compression load vector x to be inverted is: Here, Σ meas Σ represents the random measurement bias between observed data and actual physical quantities; surr The covariance matrix of the prediction error of the surrogate model; m Represents the dimension of the observation vector.
[0029] S33: Update the prior distribution of the compression load vector x to be inverted, and obtain the posterior distribution of x: Optionally, for the convenience of numerical computation, the posterior distribution of x is rewritten in logarithmic form: In the formula, the integrating factor cSince the logarithmic transformation does not change the location of the extreme values or the probability ratios of the posterior distribution, it is not explicitly solved for.
[0030] Step S4: Based on the likelihood function and prior distribution, solve for the posterior probability density distribution of the compressive load within the Bayesian statistical framework to obtain the optimal estimate and confidence interval of the compressive load. The specific steps are as follows: S41: Optionally, the Markov chain Monte Carlo method (MCMC) is used to sample from the posterior distribution. During the sampling process, the likelihood function needs to be calculated at each iteration. At this time, the trained surrogate model M(x) is directly called instead of the computationally expensive "high-fidelity numerical model" to achieve millisecond-level response. S42: Optionally, considering the characteristics of continuous tunneling by the TBM, a recursive update strategy is adopted, namely: the first... t The posterior probability distribution obtained at the end of the inversion As the first t The prior distribution of the +1 inversion, i.e. Step S5: Based on the sampling results, output the inversion results and issue engineering warnings. The specific steps are as follows: S51: Remove the samples preheated by the MCMC sampling chain, draw the frequency distribution histogram of each component of the extrusion load vector x to be inverted based on the remaining samples, and calculate its mean as the optimal estimate. S52: Set the threshold F for the compressive load on the shield structure. lim Traverse the sample chain generated by MCMC sampling and count the number of samples whose predicted response exceeds the threshold. S fail Compared with the total number of samples The ratio of these two values is the current probability of failure. p f : In the formula, I (⋅) is the characteristic function, x j These are sample points generated by MCMC sampling. When the failure probability exceeds a preset warning threshold, a tiered early warning signal is automatically triggered.
[0031] Corresponding to the aforementioned embodiment of a real-time inversion method for TBM compression load probability based on a surrogate model, the present invention also provides an embodiment of a real-time inversion device for TBM compression load probability based on a surrogate model.
[0032] See Figure 2The present invention provides a real-time inversion device for TBM extrusion load probability based on a surrogate model, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement a real-time inversion method for TBM extrusion load probability based on a surrogate model as described in the above embodiment.
[0033] The embodiment of the TBM extrusion load probability real-time inversion device based on a surrogate model provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 2 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is the TBM extrusion load probability real-time inversion device based on a surrogate model provided by the present invention. (Except for...) Figure 2 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.
[0034] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.
[0035] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0036] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a real-time inversion method for TBM extrusion load probability based on a surrogate model as described in the above embodiments.
[0037] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.
[0038] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the aforementioned real-time inversion method for TBM extrusion load probability based on a proxy model.
[0039] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.
Claims
1. A method for real-time inversion of TBM extrusion load probability based on a proxy model, characterized in that, Includes the following steps: Step S1: For the TBM shield tunneling environment, define the compression load vector to be inverted, and determine the prior distribution of each component of the load vector; Step S2: Establish a load-TBM structure model of the TBM shield, use an active learning strategy to iteratively select sample points for simulation calculation, and build and train a surrogate model to describe the mapping relationship between the crushing load and the TBM shield structure response. Step S3: Obtain the measured shield response data at the TBM tunneling site, and construct a likelihood function that characterizes the degree of matching between the observed data and the predicted data by combining the observation error of the sensor and the prediction error of the surrogate model. Step S4: Based on the likelihood function and prior distribution, solve for the posterior probability density distribution of the compressive load to obtain the optimal estimate and confidence interval of the compressive load; Step S5: Set the allowable crush load threshold for the TBM shield. When the failure probability is higher than the preset warning line, a graded early warning signal will be automatically triggered.
2. The TBM extrusion load probabilistic inversion method based on the agent model according to claim 1, characterized in that, The compression load vector mentioned in step S1 is described using a discretized model based on the load mesh; specifically: The outer surface of the TBM shield is divided into several independent load-bearing grids in the circumferential and axial directions. The compression load vector is composed of the pressure value components at the center nodes of each grid; or a spatial interpolation function is used to map the pressure values of a finite number of control nodes into a continuous pressure distribution surface.
3. The TBM extrusion load probabilistic inversion method based on the agent model according to claim 1, characterized in that, When defining the compression load vector, the intrinsic orthogonal decomposition method is used to reduce the dimensionality of the high-dimensional load parameters; specifically: A snapshot matrix containing several typical load distribution patterns is constructed, and its eigenvalues and eigenvectors are calculated. The first k few principal eigenvectors are selected as the basis functions, and the inversion target is expressed as a linear combination of these k basis functions, thus transforming the inversion target into the inversion of several weighted coefficients. k 4. The TBM extrusion load probabilistic inversion method based on the agent model of claim 1, wherein, In step S1, the specific steps for determining the prior distribution of the load vector are as follows: Model Construction: A refined numerical model of the surrounding rock-TBM interaction system is established. The numerical model adopts any one of the following methods: continuum mechanics, discontinuous mechanics, or continuous-discontinuous coupling method. Stochastic simulation: Obtain the statistical characteristics of the physical and mechanical parameters of the surrounding rock, and perform multiple forward calculations on the refined numerical model using Monte Carlo simulation or stochastic finite element method; Prior generation: The TBM shield contact force data obtained by statistical forward calculation is fitted to obtain the probability density function of the compression load, which is used as the prior distribution of the compression load vector in the subsequent probability inversion.
5. The TBM extrusion load probabilistic inversion method based on the agent model according to claim 1, characterized in that, Based on geological survey data or design parameters, the a priori value range of each component of the load vector is determined, specifically as follows: A refined numerical model of the surrounding rock-TBM system was established to obtain the probability distribution of TBM compression load, which was used as the prior distribution of compression load.
6. The TBM extrusion load probabilistic inversion method based on the agent model according to claim 1, characterized in that, The active learning strategy in step S2 specifically includes: First, a small number of initial sample points are generated using Latin hypercube sampling, and high-fidelity numerical calculations are performed to establish an initial surrogate model. Then, an iterative loop of prediction-evaluation-adding sample points is entered. The potential value of each candidate point in the parameter space is evaluated using the acquisition function, and the point that contributes the most to improving the model accuracy is dynamically selected as a new sample point for supplementary calculation. The training set is continuously updated and the model is retrained until the prediction accuracy or convergence of the surrogate model meets the preset termination condition.
7. The TBM extrusion load probabilistic inversion method based on the agent model according to claim 1, characterized in that, In step S2, the surrogate model is constructed using any one or a combination of Gaussian process regression, support vector regression, radial basis function neural network, or neural network.
8. The TBM extrusion load probabilistic inversion method based on the agent model according to claim 1, characterized in that, In step S3, the shield response data includes strain data, deformation displacement data, and segment contact pressure data collected by sensors arranged on the inner surface of the shield.
9. The TBM extrusion load probabilistic inversion method based on a proxy model according to claim 6, characterized in that, The specific execution method of step S4 is as follows: Within the Bayesian framework, the Markov Chain-Monte Carlo (MCMC) algorithm is used to sample the posterior distribution of load parameters. During each iteration of MCMC sampling, the surrogate model is invoked to replace the high-fidelity numerical simulation to calculate the likelihood function value, thereby achieving real-time inversion.
10. The TBM extrusion load probabilistic inversion method based on a surrogate model according to claim 1, characterized in that, The inversion process is performed continuously following the TBM tunneling process, employing a sequential Bayesian update strategy; specifically: ... t The posterior probability distribution of the compressive load obtained by time-inversion is used as... t The prior probability distribution inverted at time +1 is used to continuously revise the estimate of the compression characteristics of the strata ahead using historical tunneling data.