Spatial variability soft rock tunnel deformation prediction method and system based on deep learning

Abstract

The invention provides a spatial variability soft rock tunnel deformation prediction method based on deep learning, and belongs to the technical field of tunnel deformation prediction, and the method comprises the steps: constructing a surrounding rock parameter database, and determining the optimal edge distribution of different surrounding rock parameters through employing an akaike information criterion; constructing a reference value initial model, and generating a related standard uniformly distributed random field; performing equal probability transformation on the correlation standard uniformly distributed random field to generate a cross-correlation non-Gaussian distributed random field; encoding the data matrix of the non-Gaussian distribution random field into a parameter field image; and performing soft rock tunnel deformation prediction on the parameter field image by adopting a trained deformation prediction model. According to the method, parameter independent hypothesis limitation is broken through, the multi-parameter coupling space variability is accurately represented, and the prediction precision is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of tunnel deformation prediction, and in particular relates to a method and system for predicting deformation of soft rock tunnels with spatial variability based on deep learning. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the comprehensive advancement of the construction of a comprehensive three-dimensional transportation network, tunnel projects are moving towards longer and deeper tunnels. This poses significant challenges to the construction safety of soft rock tunnels in complex geological conditions. In weak surrounding rock formations, the parameters of the rock exhibit significant spatial variability due to factors such as geological structure, geostress distribution, and groundwater. This heterogeneity can easily lead to large deformation hazards in tunnels, seriously threatening construction safety and project progress.

[0004] Traditional deformation prediction methods, often based on empirical formulas or simplified mechanical models, struggle to accurately characterize the spatial heterogeneity and multi-field coupling of rock mass parameters, leading to significant deviations between predicted results and measured data. While current prediction techniques based on numerical simulation can partially reflect the mechanical behavior of surrounding rock, they face bottlenecks such as high parameter sensitivity and low computational efficiency when dealing with complex geological conditions, and they struggle to effectively integrate multi-source monitoring data during construction. With the widespread adoption of IoT monitoring technology, the accumulation of massive amounts of geological exploration data, real-time monitoring information, and historical engineering cases has opened up new possibilities for establishing data-driven intelligent prediction models.

[0005] Currently, deformation prediction for soft rock tunnel excavation under complex geological conditions faces multiple technical bottlenecks of uncertainty, mainly including the following problems: 1. Inadequate characterization of the spatial randomness of geological parameters in soft rock tunnels. Numerous experimental studies in geotechnical engineering have shown that most rock and soil parameters have non-Gaussian distributions. Existing prediction methods, often based on empirical and homogenized assumptions, struggle to effectively characterize the spatially random distribution characteristics of the surrounding rock's physical and mechanical parameters. Particularly in transitional zones within tectonic fracture zones, traditional deterministic prediction models are unable to accurately quantify the anisotropy of parameter spatial distributions and their differential impact on the deformation sensitivity of the surrounding rock.

[0006] 2. Quantifying the spatial variability of coupled multi-rock parameters is difficult. In addition to spatial autocorrelation, numerous experimental results indicate that there is a certain degree of cross-correlation among multi-rock parameters of tunnels. Key parameters related to soft rock tunnel deformation (such as elastic modulus, cohesion, cohesion, and internal friction angle) exhibit nonlinear coupling relationships, making traditional independent random field models unable to characterize the cross-correlation structure between these parameters.

[0007] 3. Insufficient value mining of multi-source monitoring data. Current methods have not yet established an effective mechanism for integrating multi-source heterogeneous monitoring data (geosurface radar, convergence monitoring, and microseismic signals) with prediction models. This lacks the ability to assimilate and dynamically update data in real time, limiting the timeliness and accuracy of deformation prediction during construction.

[0008] 4. The nonlinear dynamic coupling mechanism between surrounding rock parameters and deformation response is still unclear. The discrete representation of high-dimensional random field parameters leads to the curse of dimensionality. When the mesh is refined to the order of n × 10^3, the dimensionality of the parameter space increases dramatically, making it impossible to construct deformation function using traditional response surface methods. Summary of the Invention

[0009] To overcome the shortcomings of the aforementioned existing technologies, this paper proposes a deep learning-based method and system for predicting deformation in soft rock tunnels based on spatial variability. By constructing a deep learning model that incorporates geological spatial variation, this innovatively combines random field theory with deep neural networks to effectively characterize the inherent relationship between the spatial distribution characteristics of surrounding rock parameters and deformation responses in soft rock tunnels. This method not only enables accurate prediction of tunnel deformation under complex geological conditions but also allows real-time updating of model parameters to adapt to dynamic construction environments, providing key technical support for intelligent tunnel construction and risk prevention and control.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, a method for predicting deformation of soft rock tunnels with spatial variability based on deep learning is disclosed, comprising: A surrounding rock parameter database is constructed and the optimal edge distribution of different surrounding rock parameters is determined using the Akaike Information Criterion; the surrounding rock parameters include elastic modulus, Poisson's ratio, cohesion, and internal friction angle; constructing a benchmark numerical initial model based on the surrounding rock parameters, extracting the center coordinates of the grid cells of the initial model, discretizing the center coordinates of the grid cells by random fields and calculating the mutual correlation coefficients between the surrounding rock parameters to generate a relevant standard uniform distribution random field; Based on the optimal marginal distribution, the related standard uniform distribution random field is converted into an equal probability random field to generate a cross-correlated non-Gaussian distribution random field; the data matrix of the cross-correlated non-Gaussian distribution random field is encoded into a parameter field image; The trained deformation prediction model is used to predict the deformation of the soft rock tunnel based on the parameter field image.

[0011] In the second aspect, a spatially variable soft rock tunnel deformation prediction system based on deep learning is disclosed, comprising: A data acquisition module is configured to: construct a surrounding rock parameter database and determine the optimal marginal distribution of different surrounding rock parameters using the Akaike information criterion; A first processing module is configured to: construct a benchmark numerical initial model based on the surrounding rock parameters, extract the center coordinates of the grid cells of the initial model, discretize the center coordinates of the grid cells by random fields and calculate the mutual correlation coefficients between the surrounding rock parameters to generate a related standard uniform distribution random field; The second processing module is configured to: perform an equal probability transformation on the relevant standard uniform distribution random field based on the optimal marginal distribution to generate a cross-correlated non-Gaussian distribution random field; encode the data matrix of the cross-correlated non-Gaussian distribution random field into a parameter field image; The prediction module is configured to: use a trained deformation prediction model to predict the deformation of the soft rock tunnel on the parameter field image.

[0012] In a third aspect, an electronic device is disclosed, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are run by the processor, the steps of the above-mentioned deep learning-based spatial variability soft rock tunnel deformation prediction method are completed.

[0013] In a fourth aspect, a computer-readable storage medium is disclosed for storing computer instructions. When the computer instructions are executed by a processor, the steps of the above-mentioned deep learning-based spatial variability soft rock tunnel deformation prediction method are completed.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This method focuses on the four key constitutive parameters of soft rock tunnel surrounding rock—elastic modulus, Poisson's ratio, cohesion, and internal friction angle. The multivariate surrounding rock parameter selection is comprehensive and easily accessible, covering the core mechanical properties of stiffness, deformation, and strength. Multi-parameter randomization allows the random field model to better reflect the spatial variation of soft rock and enhance parameter interaction. The integration of multi-source data, including field surveys (such as boreholes and geological radar) and laboratory tests (triaxial tests and direct shear tests), significantly enhances the robustness of the parameter statistical distribution and reduces the possibility of model overfitting or bias amplification due to data sparsity. Furthermore, adaptive optimal distribution selection and bandwidth optimization are performed on the multivariate surrounding rock parameters, weakening the strong assumptions of traditional parameterized distributions. This data-driven approach allows for the precise characterization of complex distributions, such as asymmetric and multimodal soft rock parameters. This significantly improves the generalization capability and marginal distribution fitting reliability of the random field model, better characterizing the random distribution characteristics of surrounding rock parameters in space.

[0015] The present invention adopts the KL expansion method combined with the parameter cross-correlation to construct the covariance matrix, which accurately characterizes the cross-correlation characteristics between parameters and avoids the distortion of geomechanical response caused by traditional independent random field generation.

[0016] The present invention adaptively couples the stress release coefficient method with the random field model, combines the hybrid drive mechanism of Python scripts and Command commands, and accurately quantifies the asymmetric influence of the spatial variability of soft rock on the relaxation effect of the tunnel surrounding rock based on the dynamic parameter feedback of full-section progressive excavation. At the same time, it uses an automated iterative process to achieve efficient calculation of large-scale random field models and synchronous capture of settlement data, realizing data communication and dynamic updates; at the same time, it combines the end-to-end architecture of the neural network to directly map the geological parameters to the temporal evolution relationship of the surrounding rock deformation, thereby improving the prediction accuracy and reasoning efficiency.

[0017] The present invention strictly maintains the mutual correlation and spatial variability between parameters through the joint driving of equal probability transformation and cross-correlation covariance matrix, and realizes lossless reconstruction of the joint probability structure of multivariate surrounding rock parameters based on the mathematical support of the optimal marginal distribution function. It not only ensures the statistical consistency between the marginal distribution of each parameter and the measured data, but also can accurately characterize the nonlinear coupling characteristics of multiple parameters in space through the coordinated control of the covariance matrix.

[0018] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 This is a flowchart of the method for predicting deformation of soft rock tunnels with spatial variability based on deep learning described in Example 1 of the present invention.

[0021] Figure 2 This is a first-time realization diagram of the cross-correlation random field of multivariate surrounding rock parameters of a soft rock tunnel described in Example 1 of the present invention. DETAILED DESCRIPTION

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0023] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0024] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0025] Example 1 In one or more embodiments, a method for predicting deformation of soft rock tunnels with spatial variability based on deep learning is disclosed, such as Figure 1 As shown, the following steps are included: Step S1: construct a surrounding rock parameter database and use the Akaike information criterion to determine the optimal marginal distribution of different surrounding rock parameters.

[0026] Step S1-1, before building the surrounding rock parameter database, also includes building a parameterized random field model, selecting four key constitutive model parameters of the soft rock tunnel surrounding rock (elastic modulus E, Poisson's ratio ν, cohesion c, internal friction angle φ) as random field input variables, and establishing a non-Gaussian random field model that considers the multi-parameter cross-correlation covariance structure; The random field model parameters mainly include the spatial autocorrelation and spatial cross-correlation of key indicators such as elastic modulus E, Poisson's ratio ν, cohesion c, and internal friction angle φ. They can study the excavation deformation mechanism of soft rock tunnels under spatial variability conditions and more clearly reflect the deformation characteristics of soft rock tunnels.

[0027] This example focuses on the four key constitutive parameters of soft rock tunnel surrounding rock (E, ν, c, φ). The multivariate surrounding rock parameter selection is comprehensive and easily accessible, covering core mechanical properties such as stiffness, deformation, and strength, balancing the completeness of the parameter system with engineering operability. The selected parameters can be directly obtained from conventional geological survey reports or through simple laboratory tests, significantly reducing data acquisition costs. These selected parameters overcome the limitations of traditional methods that only consider the random distribution or simple linear correlation of a single parameter, making the random field model more consistent with the actual spatial variation of soft rock and addressing the problem of prediction bias caused by ignoring parameter interactions.

[0028] Step S1-2: Construct a surrounding rock parameter database. Collect no less than 50 sets of surrounding rock sample data including elastic modulus E, Poisson's ratio ν, cohesion c, and internal friction angle φ through field surveys and indoor tests. Build a multi-source heterogeneous database of soft rock mechanical parameters, extract the spatial characteristics richly represented by the four parameters, and perform normalization preprocessing.

[0029] Corresponding representative multivariate surrounding rock parameters are collected in the target soft rock tunnel to determine the spatial variability of multivariate surrounding rock parameters such as E, ν, c, and φ, and to establish a sample database that can reflect the actual spatial distribution characteristics of the studied tunnel.

[0030] This implementation integrates multi-source data, including field surveys (e.g., boreholes, geological radar) and laboratory tests (e.g., triaxial tests, direct shear tests). Using spatial feature extraction techniques, it explores the spatial correlations of parameters, overcoming the limitations of traditional methods that rely on a single data type. This allows the database to more comprehensively characterize the mechanical behavior of soft rock. Compared to traditional methods with small sample sizes or single data sources, this significantly enhances the robustness of parameter statistical distributions and reduces the potential for model overfitting or bias amplification due to data sparsity. Standardized processes (e.g., outlier removal and dimensional normalization) are used to preprocess the raw data, improving data quality.

[0031] Step S1-3, optimal marginal distribution probability density function selection: non-parametric kernel density estimation methods such as Gaussian kernel, uniform kernel, triangular kernel and Ipanechkov kernel are used, combined with the Akaike Information Criterion (AIC) to optimize the marginal distribution and determine the optimal marginal distribution and bandwidth parameters of parameters such as E, ν, c, and φ.

[0032] Specifically, the maximum likelihood estimation method is used to calculate the optimal marginal distribution and bandwidth parameters of the multivariate surrounding rock parameters such as E, ν, c, and φ in the database. The kernel density function is shown in Formula 1: (1) Where n is the sample size, h is the bandwidth, K(·) is the kernel function, The data obtained for the i-th measurement.

[0033] As shown in Table 1, it mainly includes non-parametric probability kernel functions with Gaussian kernel, uniform kernel, triangular kernel and Ipanechkov kernel functions.

[0034]

[0035] Note: I (•) is the indicator function. When |u|≤1, I (|u|≤1)=1, otherwise I (|u|≤1)=0.

[0036] Step S1-4, using the AIC criterion to quantitatively evaluate the fitting ability of the kernel density estimation with different kernel functions embedded; For each group of surrounding rock parameters, the AIC index values ​​of the four kernel functions can be calculated according to Formula 2, and the kernel density function corresponding to the minimum AIC value is selected as the optimal marginal distribution of the group of surrounding rock parameters; (2) Where, N is the number of samples, u i is the cumulative probability value of the i-th sample data, f is the probability density value; d is the bandwidth parameter of the kernel density function; kis the number of parameters in the kernel density function. In this embodiment, since there is only one bandwidth parameter, the number of parameters for each surrounding rock parameter is k =1.

[0037] Among them, the solution formula for the bandwidth parameter of the optimal marginal distribution is as follows: (3) Where σ is the standard deviation of the sample data.

[0038] In the marginal distribution modeling stage, a non-parametric kernel density estimation method (integrating Gaussian kernel, uniform kernel, triangular kernel and Ipanechkov kernel) is adopted, and the AIC criterion is combined to adaptively select the optimal distribution and adjust the bandwidth of multivariate surrounding rock parameters such as E, ν, c, φ. This weakens the strong assumption restrictions of traditional parameterized distributions (such as normal and lognormal), and accurately characterizes the asymmetric, multi-peak and other complex distribution forms of soft rock parameters through data-driven, significantly improving the generalization ability of the random field model and the reliability of marginal distribution fitting, and better characterizing the random distribution characteristics of surrounding rock parameters in space.

[0039] Step S2: construct a benchmark numerical initial model based on the surrounding rock parameters, extract the center coordinates of the grid cells of the initial model, discretize the center coordinates of the grid cells into random fields, calculate the correlation coefficients between the surrounding rock parameters, and generate a related standard uniform distribution random field.

[0040] Step S2-1, establish the benchmark numerical initial model: based on the surrounding rock parameters, in FLAC 3D The platform establishes an isotropic homogeneous soft rock tunnel finite difference model, uses the Mohr-Coulomb criterion to assign values ​​to the model, and obtains the initial model; Specifically, based on the geometric parameters (section shape, burial depth) of actual soft rock tunnels and the measured mechanical parameters of surrounding rocks, the FLAC 3D The platform constructs a homogenized finite-difference model, employing the Mohr-Coulomb elastoplastic constitutive model to define the mechanical behavior of the surrounding rock. Structured meshing enables refined modeling of the stratum-tunnel system. Fixed constraints are applied to the bottom of the model, while free-field boundaries are used laterally to simulate infinite domain effects. The top is set as a free deformation surface. An explicit dynamic relaxation algorithm is used to iteratively solve the initial geostress field. Displacement back-analysis and verification of the model are performed using field monitoring data to ensure consistency between the numerical model and actual engineering practice, providing a reliable benchmark for subsequent random field analysis.

[0041] Based on the measured soft rock parameters, an isotropic homogeneous model was established on the FLAC3D platform, and the Mohr-Coulomb criterion was used to assign values. A high-precision initial model was constructed using the finite difference method. This method not only retains the core characteristics of the soft rock mechanical response (such as elastic-plastic deformation and shear failure mechanism), but also removes spatial variation interference through the homogenization assumption. This provides a high-reliability benchmark for the subsequent cross-correlation random field and uncertainty quantification of multivariate surrounding rock parameters, and significantly reduces the initial value sensitivity error of numerical simulations under complex geological conditions.

[0042] Step S2-2, extracting mesh topology data: calling FLAC 3D The built-in Python API exports the center coordinates of the initial tunnel model grid cells in step S2-1 and generates a structured txt text data file containing the cell topology relationship; Specifically, in FLAC 3D In the Python API, za.pos() is called to traverse the tunnel model grid cells, extract the cell number and center coordinates, and generate a coordinate data file named "Coord" in txt format. The built-in Python core code for reading and exporting the grid cell center coordinates is as follows: Coord = za.pos()[za.in_group("tunnel")]#Extract coordinates np.savetxt('Coord.txt', Coord, fmt = '%.18e', delimiter=' ')#Save coordinates By calling FLAC 3D The built-in Python API enables efficient and automated extraction of grid cell center coordinates and topological relationships, generating coordinate data files in TXT format for the tunnel model. This breaks through the inefficient modes of traditional manual export, complex FISH language export, or non-standardized data conversion, accurately preserves the model's geometric and spatial correlation features, and provides a highly compatible input interface for subsequent parameter random field mapping and Monte Carlo simulation, significantly reducing manual operation errors and data format adaptation costs.

[0043] Step S2-3, discretizing the random field and calculating the correlation coefficients between surrounding rock parameters to generate a related standard uniform distribution random field; The Karhunen-Loeve expansion method (KL expansion method) is used in MATLAB to discretize the center coordinates of the soft rock tunnel grid cells obtained in step S2-2 into a Gaussian autocorrelation random field. A covariance matrix that considers the mutual correlation of parameters is constructed to generate N groups of standard uniform random field samples that meet the spatial correlation structure. Specifically, based on the exported tunnel model coordinates Coord.txt file, the KL expansion method shown in Equation 4 is used to discretize the random field to obtain the standard normal distribution random field: (4) Where, is a standard normal distribution random field; is the jth eigenvalue of the autocorrelation function; is the jth characteristic function of the autocorrelation function; is a random vector matrix; i is the random field number; j Truncate ordinal numbers for random field variables; M It is a cutoff term and can be determined when the expected energy ratio factor is 0.95.

[0044] Furthermore, the Pearson correlation coefficient is used to determine the mutual correlation coefficient R of parameters such as E, ν, c, and φ based on the multi-source heterogeneous database of soft rock mechanical parameters, and the Cholesky decomposition is performed to obtain the lower triangular matrix L , and multiplying with an independent standard normal random vector to obtain ξL , we can further get the relevant standard normal distribution random field: (5) Where, is the relevant standard normal distribution random field; L is the matrix of the parameter cross-correlation coefficient decomposition.

[0045] Furthermore, according to Equation 6, the relevant standard normal distribution random field H C Convert to the relevant standard uniform distribution random field H U : (6) Where, is the relevant standard uniformly distributed random field; is the standard normal cumulative distribution function.

[0046] This embodiment uses the KL expansion method combined with parameter cross-correlation to construct a covariance matrix, and realizes low-dimensional orthogonal expression of high-dimensional random fields through eigenvalue decomposition. This breaks through the limitations of traditional methods on the assumption of independent parameters and significantly enhances the coupled modeling capabilities of spatial variation characteristics of different soft rock parameters. At the same time, the covariance matrix accurately characterizes the cross-correlation characteristics between parameters, avoiding the distortion of geomechanical responses caused by traditional independent random field generation. This makes the random field discretization process of heterogeneous soft rock tunnels more consistent with the complex characteristics of multi-parameter co-evolution in actual engineering, providing rigorous mathematical support and an efficient computational framework for subsequent coupled spatial variability analysis.

[0047] Step S3: Based on the optimal marginal distribution, the relevant standard uniform distribution random field is converted into an equal probability random field to generate a cross-correlation non-Gaussian distribution random field. The cross-correlation non-Gaussian distribution random field data is output in the form of a matrix to form a multivariate surrounding rock parameter cross-correlation random field data matrix.

[0048] Specifically, based on the optimal marginal distribution and bandwidth parameters of the parameters E, ν, c, φ obtained in step S1, the equal probability mapping from the standard uniform field to the physical parameter field is realized in MATLAB through equal probability transformation, and the multivariate surrounding rock parameter cross-correlation random field data matrix is ​​output; the multivariate surrounding rock parameter cross-correlation random field RF can be obtained through equal probability transformation of the kernel density function: (7) (8) (9) (10) Where, 、 、 and They represent the inverse functions of the marginal distribution kernel density function with the optimal parameters E, ν, c and φ obtained in step S1 respectively.

[0049] like Figure 2 The figure shows a one-time realization diagram of the cross-correlation random field of multivariate surrounding rock parameters of soft rock tunnels. This embodiment breaks through the simplified assumption of traditional single variable independent mapping by jointly driving the equal probability transformation and the cross-correlation covariance matrix. In the process of converting the standard uniform field to the physical parameter field, the cross-correlation and spatial variability between parameters are strictly maintained. Based on the mathematical support of the optimal marginal distribution function, the lossless reconstruction of the joint probability structure of multivariate surrounding rock parameters is achieved, which not only ensures the statistical consistency of the marginal distribution of each parameter with the measured data, but also accurately depicts the nonlinear coupling characteristics of multiple parameters in space through the coordinated control of the covariance matrix, significantly improving the characterization ability of the random geological model of heterogeneous soft rock tunnels for the complex geomechanical behavior of actual engineering.

[0050] Step S4: encoding the multivariate surrounding rock parameter cross-correlation random field data matrix into a parameter field image based on the non-Gaussian distribution random field.

[0051] Based on the cross-correlation random fields of multivariate surrounding rock parameters such as E, ν, c, and φ generated by the KL expansion method, the channel separation-fusion technology is used to map different physical fields (E field, c field, and φ field) to the three primary color channels of the RGB color space. An H×W×D×3-dimensional image tensor data structure is constructed. The piecewise linear transformation is used to achieve conformal mapping from the non-uniform parameter value range to the [0,1] interval. Finally, the parameter field image composed of the E field, c field, and φ field is formed as the input of the prediction model.

[0052] By adopting the parameter normalization and 3D RGB encoding fusion method, the multi-parameter cross-correlated random field data of soft rock tunnels are converted into a structured 3D image tensor, and the data association between the parameter field and the vault settlement is established. Through the RGB multi-channel color gamut mapping technology, the distributed visualization expression of the parameter space variation characteristics and the retention of statistical correlation are simultaneously achieved, effectively decoupling the multi-parameter features. The resulting 3D RGB-settlement standardized multidimensional dataset provides highly compatible and strongly correlated multimodal data support for the intelligent prediction of tunnel deformation based on deep learning.

[0053] Step S5: Use the trained deformation prediction model to predict the deformation of the soft rock tunnel based on the parameter field image.

[0054] A multi-scale convolutional neural network architecture was designed, using the ReLU activation function and batch normalization layer, and hyperparameters were optimized through grid search to establish an end-to-end prediction model of "parameter field image-sedimentation displacement"; Specifically, the multi-scale convolutional neural network includes an input layer, a multi-scale feature extraction module, a residual attention module and an output module connected in sequence. The input layer receives the parameter field image input; the multi-scale feature extraction module includes a parallel three-branch structure and a fusion unit. The first branch is a macro branch, including a 7×7 convolution layer, the second branch is a meso branch, including a 3×3 convolution layer, and the third branch is a micro branch including 1×1 and 5×5 depth-separable convolution layers. Each branch includes conventional layers such as normalization layers and activation layers. The fusion unit splices features through cross-branch channels; the residual attention module includes three levels of connected convolution blocks. Each convolution block includes an identity map and a residual branch. The identity map passes the input of the convolution block to the addition layer. Node, the specific structure of the residual branch is convolution layer-normalization layer-activation function layer-SE (Squeeze-and-Excitation) module. The SE module obtains weighted features through global average pooling → fully connected layer → Sigmoid activation → channel weight multiplication, dynamically weighting key parameter channels, and the output of each convolution block is the element-by-element addition of the weighted features and the features of the identity mapping; the output features of the residual attention module are output through the output module to output the settlement displacement deformation prediction value. The output module includes a global pooling layer, a fully connected layer, and an output layer.

[0055] A multi-scale feature fusion mechanism is constructed based on a convolutional neural network. Multi-scale convolutional layers extract local and global features through parallel convolution paths with different kernel sizes. Hierarchical feature splicing is used to fuse shallow high-resolution features with deep semantic features. The hierarchical feature extraction module captures the spatial correlation and physical property coupling of parameter fields such as E, ν, c, and φ, highlighting the spatial synergistic variation patterns between parameters. Residual connections and cross-layer feature splicing are used to enhance gradient propagation efficiency, and a channel attention mechanism is combined to dynamically weight key geological parameter features.

[0056] It should be noted that the model training process also includes using historical data based on the non-Gaussian distribution random field obtained in step S3 to perform tunnel excavation simulation and record tunnel settlement deformation data for output data of model training.

[0057] This example is based on FLAC 3D The built-in Python interface performs cell traversal, coordinate matching, and parameter batch assignment. The matrix data calculated by the random field is assigned to each grid cell of the numerical model in the flac software according to the coordinate sequence number, and the random field data is batch mapped to FLAC. 3D Grid units are used to digitally reconstruct the spatial variability of the heterogeneity of multiple surrounding rock parameters such as E, ν, c, and φ in soft rock tunnels, completing the random field modeling process before tunnel excavation; Through the self-developed Python automation interface, random field data and FLAC 3D The efficient and seamless integration avoids the inefficiency and subjectivity of traditional manual unit-by-unit value assignment; based on the batch mapping algorithm, the random field data of multivariate surrounding rock parameters are accurately embedded in the numerical model grid cells while maintaining the statistical characteristics of parameter mutual correlation and spatial variability, ensuring the accuracy of digital reconstruction of complex geomechanical behaviors; through the automated process of the interface, the efficiency of multi-scale and multi-parameter coupled modeling of heterogeneous soft rock tunnels is significantly improved, providing technical support for subsequent high-fidelity numerical simulation and excavation deformation analysis.

[0058] A stress release method was used to simulate progressive excavation of the entire tunnel cross-section. The iterative calculation process was controlled by the built-in Python language and Command commands, and the vault settlement displacement values ​​after each random field model excavation stabilization were recorded. The specific implementation steps were as follows: the total number of release steps was determined based on the stress release rate, where the relaxation factor decayed step by step according to a linear law. During the numerical simulation, the maximum unbalanced force at each node on the tunnel inner wall was monitored in real time, and the dynamic stress release mechanism was simulated by dynamically applying reverse compensation node forces. When the cumulative stress release reached 50%, the initial support structure was activated, forming a surrounding rock-support collaborative bearing system. Subsequently, iterative calculations were performed to achieve a dynamic balance between the residual stress release and the support resistance, and the vault settlement displacement values ​​of each random field model were saved as deformation data.

[0059] Through the adaptive coupling of the stress release coefficient method and the random field model, combined with a hybrid drive mechanism of Python scripts and Command commands, and based on the dynamic parameter feedback of full-section progressive excavation, the asymmetric influence of the spatial variability of soft rock on the relaxation effect of the tunnel surrounding rock is accurately quantified. At the same time, an automated iterative process is used to achieve efficient calculation of large-scale random field models and synchronous capture of settlement data, significantly improving the statistical reliability and engineering applicability of the arch settlement values ​​under complex geological conditions, providing a high-precision data base for the probabilistic assessment of tunnel construction risks.

[0060] Based on the cross-correlated random fields of multivariate surrounding rock parameters such as E, ν, c, and φ generated by the KL expansion method, the different physical fields (E field, c field, and φ field) are mapped to the three primary color channels of the RGB color space using the channel separation-fusion technique. An H×W×D×3-dimensional image tensor data structure is constructed. A piecewise linear transformation is used to achieve conformal mapping from the non-uniform parameter value range to the [0,1] interval. Finally, a data set is formed with the parameter field image composed of the E field, c field, and φ field as input and the tunnel vault settlement value as output.

[0061] This dataset improves the generalization ability of the model through data enhancement strategies such as random rotation and mirror flipping. The settlement output value corresponding to the rotated image remains unchanged, providing data support for subsequent deep learning-driven prediction of surrounding rock deformation of soft rock tunnels.

[0062] In this embodiment, the training process of the deformation prediction model is as follows: historical data is processed through the above-mentioned process to obtain training samples, including deformation data and parameter field images, and a neural network is trained to obtain a trained deformation prediction model.

[0063] During the network training process, adaptive learning rate scheduling and hybrid regularization strategies (including spatial dropout and weight constraints) are introduced at the network optimization level, and a composite loss function is used to simultaneously optimize the time domain accuracy and curve morphology similarity of the displacement sequence.

[0064] Among them, the expression of the composite loss function is: (11) Where, is the time domain accuracy loss, is the curve shape loss, is the regularization term; Taking the time domain accuracy loss as the main guide, the mean absolute error between the predicted value and the measured value at all time steps is calculated: (12) Where, is the measured value, is the predicted value, is the number of training samples; The curve shape loss is an auxiliary term, and the standard DTW algorithm is used to calculate the shape difference between the predicted value and the measured value curves: (13) Where, is the curve of the measured value, is the curve of the predicted value; For fixed weight regularization, standard L2 regularization is used to prevent overfitting: (14) Where, is the weight. In this embodiment, the model architecture design fully considers the characteristics of engineering data, and realizes nonlinear modeling from the spatial variability of geological parameters to the temporal evolution of surrounding rock deformation through end-to-end mapping, ultimately forming an intelligent prediction system with both efficient reasoning capabilities and strong generalization. On the validation set, it demonstrates computational efficiency and prediction accuracy that are superior to traditional numerical simulation methods.

[0065] By integrating the spatial correlation and physical coupling of parameter fields such as E, ν, c, and φ through convolutional fusion, and combining residual connections with a channel-wise attention mechanism, the model optimizes gradient propagation paths and geological feature weights, respectively. Layered feature concatenation and hybrid regularization (spatial dropout + weight constraint) are employed to simultaneously control displacement accuracy and morphological similarity using a composite loss function, enhancing model robustness. Its end-to-end architecture directly maps geological parameters to the temporal evolution of surrounding rock deformation, improving prediction accuracy and inference efficiency. It also supports cross-project migration, providing a highly effective solution for tunnel excavation settlement deformation early warning under complex conditions.

[0066] Furthermore, we will deploy an intelligent prediction system, develop a transfer learning framework, and embed the trained CNN model into FLAC. 3D The calculation process can realize real-time intelligent prediction of tunnel deformation under new random field conditions.

[0067] Integrate the trained CNN model into FLAC based on the transfer learning framework 3D The platform's built-in Python language enables rapid prediction of excavation deformation in soft rock tunnels using a multivariate cross-correlated random field of geological parameters. A multi-dimensional visual interactive interface is developed, and a closed-loop analysis system combining numerical simulation and machine learning is established, improving computational efficiency while ensuring prediction accuracy.

[0068] This embodiment breaks down the barriers between multiple software programs, such as MATLAB, Python, and FLAC3D. Based on the transfer learning framework, a fully trained CNN model is embedded in Python built into FLAC3D. Only a small amount of new working condition data is required to quickly adapt to different geological conditions, enabling real-time and efficient prediction of tunnel deformation. This significantly reduces the cost of repeated model training and ultimately forms a closed-loop decision support system from simulation to early warning, providing intelligent guarantees for improving tunnel construction safety and efficiency.

[0069] Example 2 In one or more embodiments, a deep learning-based spatially variable soft rock tunnel deformation prediction system is disclosed, specifically comprising: A data acquisition module is configured to: construct a surrounding rock parameter database and determine the optimal marginal distribution of different surrounding rock parameters using the Akaike information criterion; A first processing module is configured to: construct a benchmark numerical initial model based on the surrounding rock parameters, extract the center coordinates of the grid cells of the initial model, discretize the center coordinates of the grid cells by random fields and calculate the mutual correlation coefficients between the surrounding rock parameters to generate a related standard uniform distribution random field; The second processing module is configured to: perform an equal probability transformation on the relevant standard uniform distribution random field based on the optimal marginal distribution to generate a cross-correlated non-Gaussian distribution random field; encode the multivariate surrounding rock parameter cross-correlated random field data matrix into a parameter field image; The prediction module is configured to: use a trained deformation prediction model to predict the deformation of the soft rock tunnel on the parameter field image.

[0070] Example 3 This embodiment provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the steps of the above-mentioned deep learning-based spatial variability soft rock tunnel deformation prediction method are completed.

[0071] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the above-mentioned deep learning-based spatial variability soft rock tunnel deformation prediction method are completed.

[0072] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0073] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0074] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide the functions for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0075] The descriptions of the various embodiments in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0076] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A spatially variable soft rock tunnel deformation prediction method based on deep learning, characterized by: include: A surrounding rock parameter database is constructed and the optimal edge distribution of different surrounding rock parameters is determined using the Akaike Information Criterion; the surrounding rock parameters include elastic modulus, Poisson's ratio, cohesion, and internal friction angle; constructing a benchmark numerical initial model based on the surrounding rock parameters, extracting the center coordinates of the grid cells of the initial model, discretizing the center coordinates of the grid cells by random fields and calculating the mutual correlation coefficients between the surrounding rock parameters to generate a relevant standard uniform distribution random field; Based on the optimal marginal distribution, the related standard uniform distribution random field is converted into an equal probability random field to generate a cross-correlated non-Gaussian distribution random field; the data matrix of the cross-correlated non-Gaussian distribution random field is encoded into a parameter field image; The trained deformation prediction model is used to predict the deformation of the soft rock tunnel based on the parameter field image.

2. The spatially variable soft rock tunnel deformation prediction method based on deep learning according to claim 1, characterized in that: The maximum likelihood estimation method is used to calculate the optimal marginal distribution of surrounding rock parameters, and the non-parametric probability kernel density function with Gaussian kernel, uniform kernel, triangular kernel and Ipanechkov kernel functions is calculated: Where n is the sample size, h is the bandwidth, K(·) is the kernel function, is the i-th data; The Akaike information criterion is used to quantitatively evaluate the fitting ability of the kernel density estimation of the kernel function, and the kernel density function corresponding to the minimum AIC value is selected as the optimal marginal distribution of the surrounding rock parameters: Where, N is the number of samples, u i is the cumulative probability value of the i-th sample data, f is the probability density value; d is the bandwidth parameter of the kernel density function; k is the number of parameters in the kernel density function.

3. The spatially variable soft rock tunnel deformation prediction method based on deep learning according to claim 1, characterized in that: The center coordinates of the grid cells are randomly discretized using the KL expansion method to obtain a standard normal distribution random field. The mutual correlation coefficient of the surrounding rock parameters is determined based on the surrounding rock parameter database. The mutual correlation coefficient is decomposed by Cholesky to obtain a lower triangular matrix. The lower triangular matrix is ​​multiplied and coupled with an independent standard normal random vector to obtain a correlated standard normal distribution random field: Where, is the relevant standard normal distribution random field; L is the matrix of parameter cross-correlation coefficient decomposition; M is the truncated item; is the jth eigenvalue of the autocorrelation function; is the jth characteristic function of the autocorrelation function; is a random vector matrix; i is the random field number; j Truncate ordinal numbers for random field variables; Based on the relevant standard normal distribution random field, the relevant standard uniform distribution random field is obtained as follows: Where, is the relevant standard uniformly distributed random field; is the standard normal cumulative distribution function.

4. The spatially variable soft rock tunnel deformation prediction method based on deep learning according to claim 1, characterized in that: The cross-correlated non-Gaussian distribution random field can be obtained by equal probability transformation of the kernel density function, including: Where, 、 、 and The inverse functions of the kernel density function of the marginal distribution representing the optimal parameters of elastic modulus, Poisson's ratio, cohesion and internal friction angle, respectively.

5. The spatially variable soft rock tunnel deformation prediction method based on deep learning according to claim 1, characterized in that: The data matrix of the cross-correlated non-Gaussian distribution random field is encoded into a parameter field image, specifically: Channel separation-fusion technology is used to map the physical fields of different parameters to the three primary color channels of the RGB color space, construct a three-dimensional image tensor data structure, and use piecewise linear transformation to achieve conformal mapping of the non-uniform parameter value range to the [0,1] interval, forming a parameter field image consisting of the elastic modulus field, cohesion field, and internal friction angle field.

6. The spatially variable soft rock tunnel deformation prediction method based on deep learning according to claim 1, characterized in that: The deformation prediction model training also includes performing tunnel simulation excavation based on a non-Gaussian distribution random field to record tunnel settlement deformation data for output data of model training; The total number of release steps is determined based on the stress release rate, where the relaxation factor decays step by step according to a linear law; During the numerical simulation, the maximum unbalanced force at each node on the tunnel inner wall is monitored in real time, and the dynamic stress release mechanism is simulated by dynamically applying reverse compensation node forces. When the cumulative stress release reaches 50%, the initial support structure is activated to form a surrounding rock-support collaborative bearing system. Afterwards, the dynamic balance between residual stress release and support resistance is achieved through iterative calculation, and the arch settlement displacement value of each random field model is saved as deformation data.

7. The spatially variable soft rock tunnel deformation prediction method based on deep learning according to claim 1, characterized in that: The multi-scale convolutional neural network includes an input layer, a multi-scale feature extraction module, a residual attention module and an output module connected in sequence; the input layer receives a parameter field image input; the multi-scale feature extraction module includes a parallel three-branch structure and a fusion unit, the first branch is a macro branch, including a 7×7 convolution layer, the second branch is a meso branch, including a 3×3 convolution layer, and the third branch is a micro branch including a 1×1 and a 5×5 depth-separable convolution layer; the residual attention module includes three levels of connected convolution blocks, each convolution block includes an identity map and a residual branch, the identity map passes the input of the convolution block to the addition node, the specific structure of the residual branch is a convolution layer-normalization layer-activation function layer-SE module, the output of each convolution block is the output feature of the residual branch and the feature of the identity map, element-by-element addition, and the output feature of the residual attention module is output through the output module to output the settlement displacement deformation prediction value.

8. A spatially variable soft rock tunnel deformation prediction system based on deep learning, characterized by: include: A data acquisition module is configured to: construct a surrounding rock parameter database and determine the optimal marginal distribution of different surrounding rock parameters using the Akaike information criterion; A first processing module is configured to: construct a benchmark numerical initial model based on the surrounding rock parameters, extract the center coordinates of the grid cells of the initial model, discretize the center coordinates of the grid cells by random fields and calculate the mutual correlation coefficients between the surrounding rock parameters to generate a related standard uniform distribution random field; The second processing module is configured to: perform an equal probability transformation on the relevant standard uniform distribution random field based on the optimal marginal distribution to generate a cross-correlated non-Gaussian distribution random field; encode the data matrix of the cross-correlated non-Gaussian distribution random field into a parameter field image; The prediction module is configured to: use a trained deformation prediction model to predict the deformation of the soft rock tunnel on the parameter field image.

9. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor. When the computer instructions are executed by the processor, the method for predicting deformation of a soft rock tunnel with spatial variability based on deep learning as described in any one of claims 1 to 7 is completed.

10. A computer-readable storage medium, characterized in that Used to store computer instructions, which, when executed by a processor, complete the spatial variability soft rock tunnel deformation prediction method based on deep learning as described in any one of claims 1 to 7.

Images