An intelligent prediction and parameter optimization method for rock mass mechanical properties based on digital twinning

Abstract

The application provides a kind of rock mass mechanical property intelligent prediction and parameter optimization method based on digital twinning, belongs to the cross technical field of geotechnical engineering and intelligent computing, method includes: S1, obtain data;S2, scale mapping and unified expression;S3, mode division;S4, extract feature data;S5, feature coding and fusion;S6, construct mechanical behavior prediction and parameter optimization model;S7, output prediction result and / or optimization result;S8, adaptive closed loop optimization.The application realizes the fast and accurate prediction of rock contact surface mechanical response, and the intelligent inversion and optimization of key mechanical parameters of the contact surface by proposing the intelligent analysis scheme of experiment-numerical simulation fusion, combined with cross-scale data modeling and physical priori constraint, and finally comprehensively improves the calculation efficiency, prediction accuracy and engineering applicability of rock contact surface mechanical analysis.

Description

Technical Field

[0001] This invention relates to the field of geotechnical engineering and intelligent computing, and in particular to a method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins. Background Technology

[0002] The mechanical behavior of contact surfaces such as rock joint surfaces, fracture surfaces, and structural surfaces is a core fundamental issue in the stability analysis and disaster prevention of rock mass engineering. Their shear strength, deformation characteristics, and parameter evolution directly determine the safety and reliability of rock mass engineering structures such as slopes, underground caverns, and tunnels. Currently, mainstream research methods in the industry face significant bottlenecks.

[0003] Indoor mechanical testing methods, such as direct shear tests and cyclic shear tests, can realistically reproduce the mechanical response characteristics of rock contact surfaces. However, due to strict limitations in sample preparation, loading conditions, and equipment capabilities, they are difficult to cover multi-parameter and multi-condition scenarios in complex engineering projects. The testing costs are high, and they cannot form a data expression form with unified physical meaning under multi-scale and multi-condition conditions, which seriously limits the application of data-driven methods in engineering prediction and parameter inversion.

[0004] Numerical simulation methods, by constructing a model of the rock mass and its contact relationship, can flexibly simulate the shear slip and deformation processes at the rock contact surface. However, the calculation results are highly sensitive to key parameters such as contact stiffness, friction coefficient, and cohesion. These parameters generally rely on empirical values ​​or repeated manual calculations to determine, lacking an intelligent and unified mapping and inversion mechanism between experimental data and numerical simulation results. This results in extremely low parameter inversion efficiency and makes it difficult to maintain the consistency and stability of parameter values ​​under different working conditions.

[0005] In existing technologies, experimental data and numerical simulation results are mostly used independently. A complete technical solution has not yet been formed that can express and fuse the two types of data in multiple modes under unified physical constraints, and simultaneously realize intelligent prediction of mechanical behavior and adaptive optimization of parameters. This cannot meet the core needs of refined analysis and efficient calculation in complex rock mass engineering. Summary of the Invention

[0006] This invention provides an intelligent prediction and parameter optimization method for rock mass mechanical properties based on digital twins. The core objective is to overcome the fundamental shortcomings of existing rock contact surface mechanical analysis techniques: key mechanical parameters rely on manual experience for value selection, numerical simulation computation is inefficient, experimental data and numerical simulation data are difficult to effectively integrate within a unified physical framework, and mechanical information at different scales is difficult to collaboratively model. By proposing an intelligent analysis scheme that integrates experimental and numerical simulations, combines cross-scale data modeling with physical prior constraints, it achieves rapid and accurate prediction of the mechanical response of rock contact surfaces, as well as intelligent inversion and optimization of key mechanical parameters of the contact surfaces. Ultimately, this comprehensively improves the computational efficiency, prediction accuracy, and engineering applicability of rock contact surface mechanical analysis.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins, comprising:

[0009] S1. Obtain experimental data and numerical simulation data of the rock contact surface, as well as the corresponding matching material parameters, working condition parameters and original contact surface image data;

[0010] S2. Perform a three-level cross-scale partitioning of all data obtained in S1, from micro to meso to macro. Based on homogenization theory, construct a scale transformation matrix to complete the scale mapping and unified expression of data features, and obtain a scale-unified cross-scale dataset.

[0011] S3. The multi-scale dataset is divided into four modalities according to its physical representation, resulting in four types of modal data: parameters, curves, images, and time series.

[0012] S4. The image modal data of S3 is processed by a multi-scale image feature extraction model to extract cross-scale image modal feature data;

[0013] S5. The other three types of modal data and cross-scale image modal feature data in S3 are preprocessed, encoded and fused using a multimodal feature processing model to obtain the experimental-numerical simulation cross-scale fused feature representation.

[0014] S6. Construct a mechanical behavior prediction and parameter optimization model. Use the cross-scale fusion feature representation of experiment-numerical simulation as input, experimental data and numerical simulation data as training labels, introduce rock mechanics physical constraints to complete model training, and obtain the trained prediction model.

[0015] S7. Input the parameters of the working condition to be analyzed into the prediction model, and output the mechanical response prediction results and / or mechanical parameter optimization results of the corresponding rock contact surface;

[0016] S8. Compare the verification data corresponding to the working condition to be analyzed with the mechanical response prediction results or mechanical parameter optimization results. Based on the comparison results, adaptively update the corresponding models of S4, S5, and S6 to form an adaptive closed-loop optimization process that combines experiment, numerical simulation, and artificial intelligence.

[0017] In this specification, the parameter modal data mentioned in S3 includes at least one of normal stress, shear displacement, contact stiffness, and friction coefficient at different scales; the curve modal data is used to characterize at least one of shear stress-displacement relationship curves, normal displacement evolution curves, and cyclic loading response curves at different scales or different loading levels; the image modal data includes at least one of rock contact surface morphology images, scanned contour maps, and contact state distribution maps at micro, meso, or macro scales; and the time series modal data is used to characterize the evolution of the mechanical response of the rock contact surface over time under different scales or cross-scale coupled working conditions.

[0018] In this specification, the multi-scale image feature extraction model described in S4 adopts a multi-scale convolutional neural network structure to extract features in parallel from rock contact surface images with different spatial resolutions and scale levels. It forms a unified cross-scale image feature representation through scale alignment or feature aggregation. At the same time, it combines the feature description method of local texture statistical analysis to extract geometric morphology and surface roughness features related to the mechanical behavior of rock contact surfaces.

[0019] In this specification, the multimodal feature processing model described in S5 introduces an attention weighting mechanism based on prior mechanical and physical constraints of rock contact surfaces during feature encoding and fusion. The attention weights are initialized based on at least one of the mechanical prior information, such as the evolution law of contact stiffness, shear-normal coupling relationship, and friction constitutive properties, and are adaptively corrected based on a data-driven approach during model training.

[0020] In this specification, the rock mechanics physical constraints described in S6 include at least one of mechanical equilibrium constraints, boundary condition constraints, and material constitutive consistency constraints. These physical constraints participate in the construction of the model loss function or the updating of attention weights to guide the model training process using prior physical information. The specific steps of the model training include:

[0021] Consistency processing and feature reconstruction are performed on the experimental data and numerical simulation data corresponding to the cross-scale fusion feature representation of experiment-numerical simulation to form a feature vector for model input.

[0022] The feature vectors are divided into training datasets and validation datasets according to a preset ratio;

[0023] The training and validation datasets are scaled to eliminate the impact of differences in the dimensions of different physical quantities on model training.

[0024] The processed training dataset is input into the mechanical behavior prediction and parameter optimization model, and the network weight parameters are iteratively updated through the backpropagation algorithm under the condition of satisfying the preset physical constraints.

[0025] When the prediction error of the model on the validation dataset meets the preset convergence condition, the trained prediction model is output.

[0026] In this specification, the mechanical response prediction results described in S7 include at least one of the rock contact surface shear strength, normal deformation, contact stiffness evolution, and friction characteristics; the mechanical parameter optimization results are output in at least one of the following forms: optimal parameter value, parameter candidate interval, and parameter probability distribution.

[0027] In this specification, the verification data corresponding to the working condition to be analyzed mentioned in S8 includes experimental results, numerical simulation results, or engineering monitoring results under that working condition. After comparison, the scene to be analyzed is first classified based on at least one of the working condition parameters, loading path, and contact surface state change characteristics. Then, according to the scene classification results, the comparison relationship between the prediction error and the preset error threshold, the corresponding model update strategy is adaptively selected. The preset error threshold is dynamically adjusted based on at least one of the scene type, historical prediction error distribution, and engineering safety level.

[0028] In this specification, the model update strategy includes at least one of incremental training, transfer learning, and model retraining; the model update process includes at least one of adaptive adjustment of model parameters, reconfiguration of model structure, and redistribution of feature weights, wherein the feature weight redistribution process is simultaneously affected by both mechanical and physical prior constraints and data-driven optimization results.

[0029] In this specification, based on the prediction model trained by S6, and combined with the 3D scanning data and geological survey data from the engineering site, a digital twin of the rock contact surface is constructed. The digital twin is used to map the geometry, mechanical state and evolution process of the real rock contact surface. Sensor data from the engineering site are collected in real time and input into the digital twin to realize the real-time update of the digital twin's state.

[0030] In this specification, based on the real-time status of the digital twin, the mechanical response prediction results and mechanical parameter optimization results output by S7, and combined with the preset engineering safety threshold, at least one of the following is output to guide engineering decision-making: support structure adjustment scheme, loading scheme optimization suggestions, and risk warning information.

[0031] In summary, the present invention has at least the following beneficial effects:

[0032] It has achieved cross-scale and multimodal deep fusion of experimental data and numerical simulation data under a unified physical framework, breaking the industry bottleneck of long-term independent application of the two types of data, constructing a complete multi-source information collaborative modeling system, and significantly improving the accuracy and stability of predicting the mechanical behavior of rock contact surfaces.

[0033] By embedding the entire process of rock mechanics physical prior constraints into the feature fusion, training optimization and updating of the artificial intelligence model, the "black box" problem of pure data-driven models is solved, ensuring the physical rationality and engineering reliability of the model output results. At the same time, it avoids the high computational cost caused by repeated parameter calculations in traditional numerical simulations, and significantly improves the computational efficiency of mechanical analysis.

[0034] It realizes intelligent inversion and multi-objective optimization of key mechanical parameters of rock contact surface, greatly reduces the dependence of parameter values ​​on human experience, solves the pain points of poor consistency and stability of parameter values ​​in traditional methods, and improves the objectivity, reliability and engineering applicability of parameter determination.

[0035] An adaptive closed-loop optimization mechanism integrating experiment, numerical simulation, and artificial intelligence was constructed, which supports continuous iteration and adaptive correction of the model under different working conditions and engineering scenarios. It has strong scenario adaptability and generalization performance, and can be widely applied to the refined analysis of various complex rock mass engineering projects, with good engineering promotion value.

[0036] It can be extended to construct digital twins of rock contact surfaces, realize real-time mapping and dynamic control of the mechanical state of rock mass in engineering sites, and provide direct technical support for the adjustment of engineering support schemes, construction optimization and risk warning, thus expanding the engineering application boundaries of the method. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of an intelligent prediction and parameter optimization method for rock mechanical properties based on digital twins, which is involved in this invention. Detailed Implementation

[0039] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0040] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0041] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0042] like Figure 1 As shown in the figure, this embodiment provides a method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins, including:

[0043] S1. Obtain experimental data and numerical simulation data of the rock contact surface, as well as the corresponding matching material parameters, working condition parameters and original contact surface image data;

[0044] S2. Perform a three-level cross-scale partitioning of all data obtained in S1, from micro to meso to macro. Based on homogenization theory, construct a scale transformation matrix to complete the scale mapping and unified expression of data features, and obtain a scale-unified cross-scale dataset.

[0045] S3. The multi-scale dataset is divided into four modalities according to its physical representation, resulting in four types of modal data: parameters, curves, images, and time series.

[0046] S4. The image modal data of S3 is processed by a multi-scale image feature extraction model to extract cross-scale image modal feature data;

[0047] S5. The other three types of modal data and cross-scale image modal feature data in S3 are preprocessed, encoded and fused using a multimodal feature processing model to obtain the experimental-numerical simulation cross-scale fused feature representation.

[0048] S6. Construct a mechanical behavior prediction and parameter optimization model. Use the cross-scale fusion feature representation of experiment-numerical simulation as input, experimental data and numerical simulation data as training labels, introduce rock mechanics physical constraints to complete model training, and obtain the trained prediction model.

[0049] S7. Input the parameters of the working condition to be analyzed into the prediction model, and output the mechanical response prediction results and / or mechanical parameter optimization results of the corresponding rock contact surface;

[0050] S8. Compare the verification data corresponding to the working condition to be analyzed with the mechanical response prediction results or mechanical parameter optimization results. Based on the comparison results, adaptively update the corresponding models of S4, S5, and S6 to form an adaptive closed-loop optimization process that combines experiment, numerical simulation, and artificial intelligence.

[0051] In some embodiments, the parametric modal data in S3 includes at least one of normal stress, shear displacement, contact stiffness, and friction coefficient at different scales; the curve modal data is used to characterize at least one of shear stress-displacement relationship curves, normal displacement evolution curves, and cyclic loading response curves at different scales or different loading levels; the image modal data includes at least one of rock contact surface morphology images, scanned contour maps, and contact state distribution maps at micro, meso, or macro scales; and the time series modal data is used to characterize the evolution of the mechanical response of the rock contact surface over time under different scales or cross-scale coupled conditions.

[0052] In some embodiments, the multi-scale image feature extraction model in S4 adopts a multi-scale convolutional neural network structure to extract features in parallel from rock contact surface images with different spatial resolutions and scale levels. It forms a unified cross-scale image feature representation through scale alignment or feature aggregation. At the same time, it combines the feature description method of local texture statistical analysis to extract geometric morphology and surface roughness features related to the mechanical behavior of rock contact surfaces.

[0053] In some embodiments, the multimodal feature processing model described in S5 introduces an attention weighting mechanism based on prior mechanical and physical constraints of rock contact surfaces during feature encoding and fusion. The attention weights are initialized based on at least one of the mechanical prior information, such as the evolution law of contact stiffness, shear-normal coupling relationship, and friction constitutive properties, and are adaptively corrected based on a data-driven approach during model training.

[0054] In some embodiments, the rock mechanics physical constraints in S6 include at least one of mechanical equilibrium constraints, boundary condition constraints, and material constitutive consistency constraints. These physical constraints participate in the construction of the model loss function or the update of attention weights to guide the model training process using prior physical information. The specific steps of the model training include:

[0055] Consistency processing and feature reconstruction are performed on the experimental data and numerical simulation data corresponding to the cross-scale fusion feature representation of experimental-numerical simulation to form a feature vector for model input.

[0056] The feature vectors are divided into training datasets and validation datasets according to a preset ratio;

[0057] The training and validation datasets are scaled to eliminate the impact of differences in the dimensions of different physical quantities on model training.

[0058] The processed training dataset is input into the mechanical behavior prediction and parameter optimization model, and the network weight parameters are iteratively updated through the backpropagation algorithm under the condition of satisfying the preset physical constraints.

[0059] When the prediction error of the model on the validation dataset meets the preset convergence condition, the trained prediction model is output.

[0060] In some embodiments, the mechanical response prediction results in S7 include at least one of the rock contact surface shear strength, normal deformation, contact stiffness evolution, and friction characteristics; the mechanical parameter optimization results are output in the form of at least one of the parameter optimal value, parameter candidate interval, and parameter probability distribution.

[0061] In some embodiments, the verification data corresponding to the working condition to be analyzed in S8 includes experimental results, numerical simulation results, or engineering monitoring results under that working condition. After comparison, the scene to be analyzed is first classified based on at least one of the working condition parameters, loading path, and contact surface state change characteristics. Then, according to the scene classification results, the comparison relationship between the prediction error and the preset error threshold, the corresponding model update strategy is adaptively selected. The preset error threshold is dynamically adjusted according to at least one of the scene type, historical prediction error distribution, and engineering safety level.

[0062] In some embodiments, the model update strategy includes at least one of incremental training, transfer learning, and model retraining; the model update process includes at least one of adaptive adjustment of model parameters, reconfiguration of model structure, and redistribution of feature weights, wherein the feature weight redistribution process is simultaneously influenced by both mechanical and physical prior constraints and data-driven optimization results.

[0063] In some embodiments, based on the prediction model trained by S6, and combined with the three-dimensional scanning data and geological survey data of the engineering site, a digital twin of the rock contact surface is constructed. The digital twin is used to map the geometry, mechanical state and evolution process of the real rock contact surface. Sensor data from the engineering site are collected in real time and input into the digital twin to realize the real-time update of the digital twin's state.

[0064] In some embodiments, based on the real-time status of the digital twin, the mechanical response prediction results and mechanical parameter optimization results output by S7, and combined with a preset engineering safety threshold, at least one of the following is output to guide engineering decision-making: support structure adjustment scheme, loading scheme optimization suggestion, and risk warning information.

[0065] This invention constructs a complete technical system of "cross-scale multimodal data fusion - physical constraint AI modeling - adaptive closed-loop optimization", and the core implementation process is as follows:

[0066] Multi-source data acquisition and system construction: Acquire experimental data and numerical simulation data of rock contact surfaces, as well as corresponding material parameters, working condition parameters, and contact surface morphology image data; complete cross-scale division of data according to three levels of micro, meso, and macro scales; realize unified mapping and expression of data at different scales based on homogenization theory and scale transformation matrix; at the same time, according to the physical expression form, divide the data into four modes: parameters, curves, images, and time series, and construct a complete cross-scale multimodal data system.

[0067] Multi-scale and multi-modal feature extraction and fusion: For image modal data, a multi-scale image feature extraction model is constructed to extract core image features of contact surface geometry, roughness, and spatial distribution at different scales; for four modal data, a multi-modal feature processing model is constructed to complete data preprocessing and independent feature encoding, and an attention weighting mechanism based on rock mechanics physical prior constraints is introduced to achieve deep fusion of features at different scales and modalities, forming a cross-scale fused feature representation of experimental and numerical simulation.

[0068] Construction of Mechanical Behavior Prediction and Parameter Optimization Model: Using cross-scale fusion features as input, a mechanical behavior prediction and parameter optimization model is constructed, which includes regression prediction and intelligent optimization modules. During the model training and optimization process, physical constraints such as mechanical equilibrium, boundary conditions, and material constitutive consistency are embedded to guide the model to learn the nonlinear mapping relationship between working parameters, contact surface morphology and mechanical response, thereby completing the model training.

[0069] Predictive optimization output and closed-loop update: Input the parameters of the working condition to be analyzed into the trained model, and output the predicted mechanical behavior of rock contact surface shear strength, normal deformation, contact stiffness evolution, friction characteristics, etc., as well as the optimal values, candidate intervals or probability distributions of key mechanical parameters. At the same time, the predicted results are compared with newly added experimental, numerical simulation or engineering monitoring data, and the model update strategy is adaptively selected based on the prediction error and working condition change characteristics, forming an adaptive closed-loop optimization process of experiment-numerical simulation-artificial intelligence collaboration.

[0070] Furthermore, this method can construct a digital twin of the rock contact surface based on the prediction model and engineering site data, thereby realizing real-time mapping, dynamic control, and risk warning of the rock mass condition at the engineering site.

[0071] The technical concept of this invention is as follows:

[0072] The experiment data of the rock contact surface, the numerical simulation data obtained by numerical simulation method, and the material parameters and working condition parameters corresponding to the experimental data and numerical simulation data are obtained. The experimental data includes mechanical response data of the rock contact surface obtained by indoor test or field test, and the numerical simulation data includes mechanical response data of the rock contact surface obtained by discontinuous deformation analysis method. At the same time, the original image data characterizing the morphology or contact state of the rock contact surface are obtained. The original image data includes contact surface scan images, morphology contour maps or contact state distribution maps at different scales.

[0073] The experimental data, numerical simulation data, and raw image data are divided into cross-scale categories according to micro, meso, and macro scales to construct a three-level cross-scale data system of micro-meso-macro.

[0074] Based on the physical representation of data at different scales, the data is modally divided to form a multimodal data system that includes at least parametric modal data, curve modal data, image modal data, and time series modal data.

[0075] For the aforementioned image modal data, a multi-scale image feature extraction model is constructed. Through image recognition and feature extraction processing, image modal feature data characterizing the geometric morphology, surface roughness, and spatial distribution of the rock contact surface at different scales are extracted.

[0076] The parametric modal data, curve modal data, time series modal data, and the multi-scale image modal feature data are respectively cleaned, outlier data is removed, features are extracted and normalized, and features of different modes and scales are encoded and fused under the unified physical quantity definition and scale mapping conditions.

[0077] In the feature fusion process, an attention weighting mechanism based on prior mechanical and physical constraints of rock contact surfaces is introduced. The attention weights are initialized based on prior mechanical information and adaptively corrected during model training using a data-driven approach. This constructs a cross-scale fusion dataset or fusion feature representation of experimental and numerical simulation, which is then used for subsequent prediction of rock contact surface mechanical behavior and parameter optimization modeling.

[0078] A mechanical behavior prediction and parameter optimization model is constructed using the experimental-numerical simulation cross-scale fusion dataset or fusion feature representation as input. During the training process, one or more of the following constraints are introduced: mechanical equilibrium constraints, boundary condition constraints, or material constitutive consistency constraints. This allows the model to learn the mapping relationship between working condition parameters, contact surface morphology features, and rock contact surface mechanical behavior under the premise of satisfying the mechanical and physical rationality of the rock contact surface, thus obtaining a trained rock contact surface mechanical behavior prediction model.

[0079] Input the parameters of the working condition to be analyzed into the trained rock contact surface mechanical behavior prediction model, and output the mechanical response prediction results and / or the optimization results of the contact surface mechanical parameters of the corresponding rock contact surface. The mechanical response prediction results include one or more of shear strength, normal deformation, contact stiffness evolution or friction characteristics.

[0080] The experimental results, numerical simulation results, or engineering monitoring results under the working condition to be analyzed are obtained and compared with the prediction results. When the prediction error exceeds the preset threshold, the image feature extraction model, multimodal feature processing model, or mechanical behavior prediction and parameter optimization model are adaptively selected for incremental training, transfer learning, or retraining according to the characteristics of working condition changes or scene type, forming an adaptive closed-loop optimization process of experiment-numerical simulation-artificial intelligence collaboration.

[0081] This invention constructs a fusion system of cross-scale experimental data and numerical simulation data based on discontinuous deformation analysis methods, and achieves collaborative modeling of multimodal and multi-scale information under a unified physical framework, thereby improving the accuracy and stability of predicting the mechanical behavior of rock contact surfaces.

[0082] This invention introduces an artificial intelligence model with embedded mechanical and physical prior constraints to achieve rapid prediction of the mechanical response of rock contact surfaces, avoiding the high computational cost caused by repeated trial calculations and parameter adjustments in traditional numerical simulation methods, and significantly improving computational efficiency.

[0083] This invention enables intelligent inversion and optimization of key mechanical parameters of rock contact surfaces under mechanical and physical constraints, reducing reliance on human experience in parameter selection and improving the objectivity, reliability, and engineering applicability of parameter determination.

[0084] This invention supports continuous updating and adaptive correction of the model under different working conditions and scenarios. It is suitable for efficient analysis and prediction of the mechanical behavior of rock contact surfaces under complex rock mass engineering conditions, and has good scalability and engineering promotion value.

[0085] This method is based on experimental data, numerical simulation data and contact surface image data of rock contact surfaces at the micro, meso, and macro scales. By constructing a three-level cross-scale data system of micro-meso-macro, and introducing homogenization theory and scale transformation matrix, it realizes the mapping and expression of features at different scales in a unified feature space.

[0086] Based on this, combined with multimodal artificial intelligence modeling technology, parametric modal data, curve modal data, time series modal data and image modal data are jointly modeled to establish a nonlinear mapping relationship between working condition parameters, contact surface morphology features and rock contact surface mechanical response. This enables the prediction of rock contact surface mechanical behavior and intelligent inversion and optimization of key mechanical parameters, forming an adaptive closed-loop optimization process that combines experiment, numerical simulation and artificial intelligence.

[0087] Based on the manifestation of rock contact surface mechanical behavior at different spatial scales, experimental data and numerical simulation data are divided into microscale data, mesoscale data and macroscale data.

[0088] Microscale data are used to characterize particle contact properties, micro-roughness structure, and local geometric undulations.

[0089] Microscale data are used to characterize crack distribution features, contact element combination characteristics, and local structural surface response behavior;

[0090] Macroscale data are used to characterize the mechanical response behavior of the overall structural surface under external loads.

[0091] To achieve consistent representation of features at different scales, a scale transformation matrix based on homogenization theory and numerical equivalence principle is introduced to map and transform feature parameters at different scales, making them comparable in a unified feature space, thereby eliminating the feature misalignment problem caused by cross-scale feature differences.

[0092] The method and process are briefly described as follows:

[0093] The experimental data of the rock contact surface, the numerical simulation data obtained by the numerical simulation method, and the original image data of the contact surface are obtained, and the material parameters and working condition parameters corresponding to the data are obtained simultaneously.

[0094] The experimental data, numerical simulation data, and image data are divided into micro-scale, meso-scale, and macro-scale to construct a three-level cross-scale original data system.

[0095] Based on homogenization theory and scale transformation matrix, scale mapping and unified expression are performed on data features at different scales, realizing the mapping of micro, meso and macro scale features in a unified feature space, and eliminating the feature misalignment problem caused by cross-scale feature differences.

[0096] For image modal data at different scales, a multi-scale image feature extraction model is constructed. Through image recognition and feature extraction processing, image modal feature data that characterize the geometric morphology, surface roughness and spatial distribution of rock contact surfaces at micro, meso and macro scales are extracted.

[0097] Data cleaning, outlier removal, feature extraction, and normalization are performed on the parametric modal data, curve modal data, time series modal data, and the image modal feature data, respectively.

[0098] Under the conditions of unified physical quantity definition and scale mapping, features of different modes and scales are encoded and fused to construct experimental-numerical simulation cross-scale fused datasets or fused feature representations.

[0099] Construct a mechanical behavior prediction and parameter optimization model with the experimental-numerical simulation cross-scale fusion dataset or fusion feature representation as input, and introduce one or more of the following: mechanical equilibrium constraints, boundary condition constraints, or material constitutive consistency constraints.

[0100] The model is trained using the fused dataset to obtain a trained prediction model of the mechanical behavior of rock contact surfaces.

[0101] Input the parameters of the working condition to be analyzed into the trained rock contact surface mechanical behavior prediction model, and output the corresponding rock contact surface mechanical response prediction results and / or the optimization results of key mechanical parameters of the contact surface.

[0102] Obtain the experimental results, numerical simulation results, or engineering monitoring results under the working condition to be analyzed, and compare them with the model prediction results;

[0103] When the prediction error exceeds the preset threshold, the current scene is judged based on the working condition parameters, loading path or contact surface state change characteristics, and the model update strategy is adaptively selected to update or correct the image feature extraction model, multimodal feature processing model or mechanical behavior prediction and parameter optimization model, forming an adaptive closed-loop optimization process of experiment-numerical simulation-artificial intelligence collaboration.

[0104] The detailed implementation examples of image modal feature extraction are as follows:

[0105] In this embodiment, the image modal feature extraction step is used to convert the original image modal data of the rock contact surface into image modal feature data with clear physical meaning, so as to participate in subsequent multimodal feature fusion and mechanical behavior prediction modeling.

[0106] Image modal data sources and preprocessing:

[0107] The image modal data includes rock contact surface morphology images, scanned contour maps, or contact state distribution maps, which can be obtained by three-dimensional laser scanning, structured light scanning, digital image correlation methods, or microscopic imaging methods.

[0108] Before feature extraction, the image modal data is preprocessed. The preprocessing steps include image denoising, grayscale normalization, size resampling, and background region removal to reduce the impact of imaging noise and differences in acquisition conditions on the feature extraction results.

[0109] Image feature extraction model construction:

[0110] For the preprocessed image modal data, an image feature extraction model is constructed. The image feature extraction model adopts a multi-scale convolutional neural network structure to extract features in parallel from rock contact surface images with different spatial resolutions and scale levels. A unified cross-scale image feature representation is formed through scale alignment or feature aggregation. The model is used to extract multi-scale geometric and roughness features related to the mechanical behavior of rock contact surfaces, providing a priori feature basis for subsequent physical constraint modeling and attention weight allocation.

[0111] The feature extraction structure based on the convolutional neural network includes at least one set of convolutional layers, nonlinear activation layers, and pooling layers, which are used to automatically learn the spatial structural features in the rock contact surface morphology image.

[0112] The texture feature extraction structure based on local statistical analysis is used to extract statistical feature parameters that reflect the micro-roughness and local undulation features of the contact surface.

[0113] Convolutional feature extraction process:

[0114] In one embodiment, the input image modality data The convolution feature mapping can be represented as follows:

[0115] ;

[0116] in, Indicates the first Each convolutional kernel is located at... The characteristic value output at the location; Indicates the kernel number. ,in This represents the total number of convolutional kernels; Indicates the first The coordinates of each convolution kernel within the kernel Weight parameters at the location; Indicates the first The bias parameters corresponding to each convolutional kernel; and Represents the local coordinate index inside the convolution kernel; This indicates the size of the convolution kernel in the horizontal direction; This indicates the size of the convolution kernel in the vertical direction; Represents a nonlinear activation function; and This indicates the pixel positions covered by the convolution kernel as it slides across the input image. The number of convolution kernels can be varied. and kernel size It can extract the morphological and structural features of rock contact surfaces at different scales.

[0117] Through multi-layer convolution operations, macroscopic morphological features and local structural features in the rock contact surface morphology image are extracted step by step.

[0118] Local texture statistical feature extraction:

[0119] In another embodiment, to enhance the physical interpretability of image features, local texture statistical analysis is performed on the convolution feature extraction results or the original image modal data to extract statistical parameters characterizing the roughness and spatial distribution features of the contact surface.

[0120] The local texture statistical features include, but are not limited to, local height mean, standard deviation, slope distribution parameters, and gray-level co-occurrence matrix features. These local texture statistical features are used to characterize the actual contact area variation trend and local stress concentration potential of the rock contact surface during mechanical loading. Its local roughness characteristic parameters can be expressed as:

[0121] ;

[0122] This represents the root mean square roughness characteristic value of the local region; This represents the total number of pixels participating in the statistical calculation within the local area; This variable represents the index of a pixel within a local region, and its value range is... ; Indicates the first The height or grayscale value corresponding to each pixel; This represents the arithmetic mean of the height or grayscale values ​​of all pixels within the local area.

[0123] The depth features extracted by the convolutional neural network are concatenated or weighted with local texture statistical features to form an image modality feature vector characterizing the geometric morphology, surface roughness, and spatial distribution characteristics of the rock contact surface. ;

[0124] The image modal feature vectors, as intermediate feature variables with clear physical meanings, participate together with parametric modal data, curve modal data, and time series modal data in the subsequent multimodal feature encoding, fusion, and construction of mechanical behavior prediction and parameter optimization models.

[0125] Detailed implementation examples of multimodal feature encoding and fusion are as follows:

[0126] In this embodiment, the multimodal feature encoding and fusion step is used to convert parameter modal data, curve modal data, time series modal data and image modal feature data from different sources and with different physical representations into a unified feature representation, and to realize collaborative modeling of multimodal information, providing a stable and reliable model input for subsequent prediction of rock contact surface mechanical behavior and parameter optimization.

[0127] In one embodiment, the input data involved in multimodal feature encoding and fusion includes:

[0128] Parametric modal data are used to characterize the working parameters and material parameters of the rock contact surface;

[0129] Curve modal data are used to characterize the changes in the mechanical response of rock contact surfaces during loading or shearing.

[0130] Time-series modal data are used to characterize the evolution of the mechanical response of rock contact surfaces over time under cyclic loading or multi-stage loading conditions;

[0131] Image modal feature data is used to characterize the geometry, surface roughness, and spatial distribution characteristics of rock contact surfaces.

[0132] The aforementioned multimodal features have different physical dimensions, data structures, and time scales, and therefore require unified encoding and fusion processing.

[0133] For the physical properties of different modal features, corresponding feature encoding sub-models are constructed to encode each modal feature independently.

[0134] Parametric modal feature encoding: Normalize and perform feature mapping on parametric modal data to map it to a high-dimensional feature space, forming parametric modal feature vectors. ;

[0135] Curve modal feature encoding: The curve modal data is discretized into a unified sequence of sampling points, and the curve morphological features are extracted through function modeling or sequence coding models to form a curve modal feature vector. ;

[0136] Time-series modal feature encoding: Construct a time-series feature encoding model for time-series modal data to extract time-series feature vectors that reflect the evolution of mechanical response over time. ;

[0137] Image modal feature encoding: The image modal feature vector output from the image modal feature step is subjected to dimensionality compression and feature mapping to form an image modal feature vector with the same scale as other modal features. ;

[0138] In one embodiment, to eliminate the influence of differences in physical dimensions and numerical ranges between different modal features, the modal feature vectors are subjected to uniform scaling, and their normalized form can be expressed as: ;

[0139] in, Represents the eigenvector of the t-th mode; Indicates the number after standardization. Modal feature vector. t represents the modal category index. , where T is the total number of modes; Let represent the mean of the feature vector of the t-th modality on the training dataset; Indicates the first The standard deviation of the modality feature vector on the training dataset.

[0140] Meanwhile, mechanical consistency constraints are introduced during the feature encoding process to ensure that the encoded features satisfy the basic mechanical equilibrium relationship and material constitutive constraints.

[0141] In one embodiment, a multimodal feature fusion model is used to fuse the feature vectors of each modality. The fusion method includes one or more of feature concatenation, weighted fusion, or attention-weighted fusion. The attention weights are set during the initialization phase based on prior mechanical and physical information such as the evolution law of contact stiffness of the rock contact surface, shear-normal coupling relationship, or friction constitutive properties, and are adaptively corrected during model training through a data-driven approach. The attention-weighted fusion method assigns weights according to the contribution of different modal features to the prediction target, and its fused feature representation is as follows: ;

[0142] in: This represents the fused multimodal feature vector; This represents the feature vector corresponding to the s-th modality; s represents the modality category index, s=1,2,…,S; S represents the total number of modalities participating in the fusion. This represents the attention weight coefficient corresponding to the s-th modality feature.

[0143] The attention weight coefficients satisfy the normalization constraint: ;

[0144] During model training, the weight coefficients are adaptively adjusted by optimizing the objective function, so that the fused feature vector can more effectively represent the contribution of different modal data to the task of predicting the mechanical behavior of rock contact surfaces.

[0145] The weight coefficients are obtained by adaptive learning during the training process of the fusion model.

[0146] The multimodal features obtained by the fusion are represented As a unified input feature, it is fed into the subsequent rock contact surface mechanical behavior prediction and parameter optimization model to establish the nonlinear mapping relationship between working condition parameters, contact surface morphology features and rock contact surface mechanical response.

[0147] Through the above-mentioned multimodal feature encoding and fusion processing, the experimental data, numerical simulation data and image information are utilized in a coordinated manner, thereby improving the model's ability to express and predict the mechanical behavior of rock contact surfaces under complex working conditions.

[0148] A detailed implementation example of the mechanical behavior prediction and parameter optimization model is as follows:

[0149] In this embodiment, the mechanical behavior prediction and parameter optimization model is used to establish a mapping relationship between working condition parameters, contact surface morphology features and rock contact surface mechanical response based on multimodal fusion features, so as to realize the prediction of rock contact surface mechanical behavior and intelligent inversion and optimization of key mechanical parameters.

[0150] Model input and output definitions:

[0151] In one embodiment, the mechanical behavior prediction and parameter optimization model is represented by the output multimodal fusion features. Input for the model.

[0152] The output of the model includes:

[0153] The predicted results of the mechanical behavior of the rock contact surface include shear strength, normal deformation, contact stiffness evolution and friction characteristics;

[0154] The optimization results of key mechanical parameters of rock contact surfaces include optimal parameter values, candidate parameter intervals, or parameter probability distributions.

[0155] Predictive model structure construction:

[0156] In one embodiment, a mechanical behavior prediction model based on a deep neural network is constructed. The model includes an input layer, several hidden layers, and an output layer. Feature transformation between the layers is achieved through nonlinear mapping.

[0157] The forward propagation process of the model can be represented as: ;

[0158] in, This indicates the predicted results of the mechanical behavior of the rock contact surface; This represents the input feature vector after multimodal feature fusion; Represents the mapping function of a deep neural network; This represents the set of model parameters.

[0159] This model enables a nonlinear mapping between multimodal fusion features and mechanical response.

[0160] Physical consistency constraints are introduced:

[0161] In one embodiment, a comprehensive loss function, including a data consistency term and a mechanical consistency term, is used during model training. The comprehensive loss function of the model can be expressed as:

[0162] ;

[0163] in: Total training loss; Indicates the model prediction results With corresponding experimental data or numerical simulation results The error between;

[0164] The term represents the mechanical consistency constraint error term, where R(·) represents the residual operator composed of mechanical equilibrium equations, contact conditions, or material constitutive relations, and β represents the mechanical parameters of the contact surface to be identified. is the weighting coefficient (which can be dynamically adjusted according to the accuracy requirements and physical constraint priorities of the engineering scenario; the exemplary value range is 0.1 to 10), and N is the number of samples. For sample index;

[0165] Multi-objective optimization mathematical model:

[0166] In the parameter optimization phase, the parameter inversion problem is constructed as a multi-objective optimization problem, the mathematical expression of which is:

[0167] :

[0168] in: This represents the perturbation variable. Ω denotes the feasible region of the parameters; Indicates in and The model predicts the mechanical response of the rock contact surface under the given conditions; This represents the experimentally measured mechanical response data of the rock contact surface; Indicates to The expectation operator; Indicates to The gradient operator; This represents the L2 norm. For the goal of data consistency; For robustness objectives; For parameter sensitivity targets;

[0169] In one embodiment, the mechanical consistency condition is used as a feasible region constraint in the parameter optimization process, and its mathematical expression is:

[0170] ;

[0171] in This is the allowable threshold for mechanical residuals; This indicates the range of parameter values ​​determined by engineering experience or experiments.

[0172] The aforementioned multi-objective optimization problem is solved using a multi-objective intelligent optimization algorithm based on non-dominated sorting. In one specific implementation, an improved non-dominated sorting genetic algorithm III is used to search the parameter space to obtain a set of Pareto optimal solutions that satisfy the multi-objective constraints.

[0173] ;

[0174] Where P represents the Pareto optimal solution set; Let represent the o-th candidate parameter solution; O represents the number of solutions contained in the Pareto optimal solution set. A parameter solution is called a non-dominated solution and constitutes the Pareto optimal solution set P when no other parameter solution is simultaneously superior to it on all optimization objectives, and it is not inferior to other solutions on at least one objective function.

[0175] In some embodiments, when the engineering scenario has higher requirements for physical rationality (such as safety-sensitive projects like high slopes and deep-buried tunnels), it can be The value is set to 5 to 10 to strengthen the guiding role of mechanical consistency constraints in model training and prioritize ensuring that the prediction results conform to the basic laws of rock mechanics.

[0176] When engineering scenarios require higher data fitting accuracy (such as indoor test condition verification, small-scale rock mass analysis), it is possible to... The value is set to 0.1 to 2 to prioritize the model's ability to fit experimental and numerical simulation data and improve prediction accuracy under local conditions.

[0177] For conventional rock mass engineering scenarios, it can be The initial values ​​are set to 1 to 3, and are adaptively fine-tuned during training based on the error feedback from the validation set, balancing data fitting and physical compliance.

[0178] In some embodiments, the multi-scale convolutional neural network used in the multi-scale image feature extraction model can adopt the following typical structure to adapt to the feature extraction requirements of rock contact surface morphology images:

[0179] 1. Input layer: Receives rock contact surface morphology images (including micro / meso / macro scale scan images) with a resolution of 512×512 pixels, with 1 input channel (grayscale image) or 3 input channels (RGB color image);

[0180] 2. Convolution-Pooling Module Groups: Set up 4 groups of convolution-pooling modules in sequence, with the following configuration:

[0181] Group 1: Convolutional layer (64 3×3 convolutional kernels, stride 1, padding 1) → ReLU activation → 2×2 max pooling;

[0182] Group 2: Convolutional layer (128 3×3 convolutional kernels, stride 1, padding 1) → ReLU activation → 2×2 max pooling;

[0183] Group 3: Convolutional layer (256 3×3 convolutional kernels, stride 1, padding 1) → ReLU activation → 2×2 max pooling;

[0184] Group 4: Convolutional layer (512 3×3 convolutional kernels, stride 1, padding 1) → ReLU activation → 2×2 max pooling;

[0185] 3. Global Average Pooling Layer: Global average pooling is performed on the feature maps output by the last set of convolutional layers to output a 512-dimensional cross-scale image feature vector;

[0186] 4. Feature Fusion Interface: This interface concatenates the feature vector with other modal features and inputs it into the multimodal feature processing model.

[0187] In some embodiments, the improved non-dominated sorting genetic algorithm III addresses the multi-objective optimization problem of mechanical parameters of rock contact surfaces, and makes the following core improvements based on the standard NSGA-III:

[0188] 1. Constraint Initialization: During the population initialization phase, a hard constraint on the mechanically feasible region is introduced, generating only initial individuals that satisfy the basic laws of rock mechanics (e.g., the friction coefficient is limited to a range of 0.2~0.8, and the contact stiffness is limited to a range of...). ), to avoid invalid iterations and physically infeasible solutions;

[0189] 2. Physical constraint penalty mechanism: In the crossover and mutation operation, a penalty function is applied to the generated offspring individuals that violate mechanical consistency constraints (such as shear strength exceeding the rock ultimate strength, and the relationship between normal deformation and stress violating constitutive laws), thereby reducing their fitness value and guiding the algorithm to converge toward the physical feasible region;

[0190] 3. Engineering-oriented reference point setting: In the reference point setting stage, combined with engineering experience and parameter sensitivity analysis, targets closely related to engineering safety (such as shear strength and contact stiffness) are selected as reference directions to improve the engineering applicability of the optimization results;

[0191] 4. Adaptive Termination Condition: Set the termination condition to "the distribution difference of non-dominated solutions in three consecutive generations is less than..." Meanwhile, the maximum number of iterations is limited to 500 generations to balance optimization accuracy and computational efficiency.

[0192] In some embodiments, the engineering application of the method of the present invention is demonstrated by taking the analysis of the mechanical behavior of the rock mass structure of a high slope project of a mountain highway as an example:

[0193] 1. Operating conditions and data foundation:

[0194] Project background: The slope is 68m high, the rock mass is moderately weathered granite, the structural plane dips at 35°, and it is a typical bedding slope;

[0195] Operating parameters: normal stress range 2~8MPa, shear displacement range 0~20mm, loading method is graded cyclic loading;

[0196] Data sources: Indoor direct shear test data (300 sets), 3DEC numerical simulation data (900 sets), and three-dimensional laser scanning morphology images (1200 images) were obtained, forming a total of 1200 sets of cross-scale multimodal samples, which were divided into training set (960 sets) and validation set (240 sets) in an 8:2 ratio.

[0197] 2. Model training configuration:

[0198] Cross-scale processing: Based on homogenization theory, a scale transformation matrix is ​​constructed to achieve a unified representation of micro-micro-macro three-level data;

[0199] Feature extraction: The aforementioned exemplary multi-scale convolutional neural network is used to extract image modal features, and the multi-modal feature fusion is completed by combining a mechanical prior attention mechanism;

[0200] Training parameters: in the comprehensive loss function Mechanical equilibrium constraints and constitutive consistency constraints are introduced, and the Adam optimizer (initial learning rate) is used. Training is stopped when the prediction error on the validation set is less than 0.5% for 10 consecutive rounds.

[0201] 3. Comparison of prediction results with on-site monitoring:

[0202] Mechanical response prediction: Input the working condition parameters such as normal stress and shear displacement monitored on site into the model, and output the predicted results of shear strength and normal deformation of the structural surface; compared with the monitoring data of stress sensor and displacement gauge on site, the prediction error of shear strength is <3% and the prediction error of normal deformation is <2%, which meets the engineering accuracy requirements;

[0203] Parameter optimization: The model outputs the optimal contact stiffness of this structural surface as follows: The optimal friction coefficient was 0.58, with a deviation of less than 5% from the field inversion results, verifying the reliability of parameter optimization.

[0204] Engineering value: Based on the prediction results and optimization parameters, suggestions for adjusting the slope support structure are proposed, which effectively reduces the risk of slope slippage and proves the engineering applicability of the method of this invention.

[0205] In some embodiments, taking the mechanical behavior analysis of the contact surface between the surrounding rock and the support of a deep-buried tunnel as an example, the generalizability of the method of the present invention is further verified:

[0206] 1. Operating conditions and data foundation:

[0207] Project background: The tunnel is 800m deep, the surrounding rock is a thin layer of limestone, the contact angle between the surrounding rock and the support is 15°, and the normal stress range is 5~12MPa;

[0208] Data sources: 400 sets of on-site surrounding rock deformation monitoring data, 800 sets of UDEC numerical simulation data, and 1200 structured light scan contact surface images were obtained to form 1200 sets of samples. The ratio of training set to validation set remained at 8:2.

[0209] 2. Model Training and Prediction:

[0210] Training configuration: In the comprehensive loss function (Strengthening physical constraints to adapt to deep-buried high-stress scenarios), the improved NSGA-Ⅲ algorithm is used to optimize parameters such as the contact stiffness and cohesion between the surrounding rock and the support.

[0211] Prediction results: The model output showed an error of <4% in the shear strength prediction of the surrounding rock-support contact surface and <2.5% in the normal deformation prediction. The deviation between the optimized contact stiffness parameters and the monitoring results after on-site grouting reinforcement was <6%, providing a reliable basis for the design of tunnel support parameters.

[0212] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0213] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0214] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0215] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0216] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0217] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0218] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0219] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0220] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0221] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins, characterized in that, include: S1. Obtain experimental and numerical simulation data of the rock contact surface, as well as the corresponding matching material parameters, working condition parameters, and original contact surface image data; S2. Perform a three-level cross-scale partitioning of all data obtained in S1, from micro to meso to macro. Based on homogenization theory, construct a scale transformation matrix to complete the scale mapping and unified expression of data features, and obtain a scale-unified cross-scale dataset. S3. Divide the multi-scale dataset into modes according to its physical representation to obtain parametric mode data, curve mode data, image mode data and time series mode data; S4. Process image modal data through a multi-scale image feature extraction model to extract cross-scale image modal feature data; S5. The parametric modal data, curve modal data, time series modal data, and cross-scale image modal feature data are preprocessed, encoded, and fused using a multimodal feature processing model to obtain an experimental-numerical simulation cross-scale fused feature representation. In the feature encoding and fusion process, the multimodal feature processing model introduces an attention weighting mechanism based on prior physical constraints of rock contact surface mechanics. The attention weights are initialized based on at least one of the mechanical prior information in contact stiffness evolution law, shear-normal coupling relationship, and friction constitutive properties, and are adaptively corrected based on a data-driven approach during model training. S6. Construct a mechanical behavior prediction and parameter optimization model. Use the cross-scale fusion feature representation of experiment-numerical simulation as input, experimental data and numerical simulation data as training labels, introduce rock mechanics physical constraints to complete model training, and obtain the trained prediction model. S7. Input the parameters of the working condition to be analyzed into the prediction model, and output the mechanical response prediction results and / or mechanical parameter optimization results of the corresponding rock contact surface; S8. Compare the verification data corresponding to the working condition to be analyzed with the mechanical response prediction results or mechanical parameter optimization results, and adaptively update the corresponding models of S4, S5 and S6 based on the comparison results to form an adaptive closed-loop optimization process of experiment-numerical simulation-artificial intelligence collaboration. Based on the prediction model trained by S6, and combined with the 3D scanning data and geological survey data from the engineering site, a digital twin of the rock contact surface is constructed. The digital twin is used to map the geometry, mechanical state and evolution process of the real rock contact surface. Sensor data from the engineering site are collected in real time and input into the digital twin to realize the real-time update of the digital twin's state. Based on the real-time status of the digital twin, the mechanical response prediction results and / or mechanical parameter optimization results output by S7, and combined with the preset engineering safety threshold, at least one of the following is output to guide engineering decision-making: support structure adjustment scheme, loading scheme optimization suggestions, and risk warning information.

2. The method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins according to claim 1, characterized in that, The parameter modal data in S3 includes at least one of normal stress, shear displacement, contact stiffness, and friction coefficient at different scales; the curve modal data is used to characterize at least one of shear stress-displacement relationship curves, normal displacement evolution curves, and cyclic loading response curves at different scales or different loading levels; the image modal data includes at least one of rock contact surface morphology images, scan profile maps, and contact state distribution maps at micro, meso, or macro scales; the time series modal data is used to characterize the evolution of the mechanical response of the rock contact surface over time under different scales or cross-scale coupled conditions.

3. The method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins according to claim 1, characterized in that, The multi-scale image feature extraction model described in S4 adopts a multi-scale convolutional neural network structure to extract features in parallel from rock contact surface images with different spatial resolutions and scale levels. It forms a unified cross-scale image feature representation through scale alignment or feature aggregation. At the same time, it combines the feature description method of local texture statistical analysis to extract geometric morphology and surface roughness features related to the mechanical behavior of rock contact surfaces.

4. The method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins according to claim 1, characterized in that, The rock mechanics physical constraints described in S6 include at least one of mechanical equilibrium constraints, boundary condition constraints, and material constitutive consistency constraints. These physical constraints participate in the construction of the model loss function or the update of attention weights to guide the model training process using prior physical information. The specific steps of the model training include: Consistency processing and feature reconstruction are performed on the experimental data and numerical simulation data corresponding to the cross-scale fusion feature representation of experimental-numerical simulation to form a feature vector for model input. The feature vectors are divided into training datasets and validation datasets according to a preset ratio; The training and validation datasets are scaled to eliminate the impact of differences in the dimensions of different physical quantities on model training. The processed training dataset is input into the mechanical behavior prediction and parameter optimization model, and the network weight parameters are iteratively updated through the backpropagation algorithm under the condition of satisfying the preset physical constraints. When the prediction error of the model on the validation dataset meets the preset convergence condition, the trained prediction model is output.

5. The method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins according to claim 1, characterized in that, The mechanical response prediction results mentioned in S7 include at least one of the rock contact surface shear strength, normal deformation, contact stiffness evolution, and friction characteristics; the mechanical parameter optimization results are output in the form of at least one of the parameter optimal value, parameter candidate interval, and parameter probability distribution.

6. The method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins according to claim 1, characterized in that, The verification data corresponding to the working condition to be analyzed in S8 includes experimental results, numerical simulation results, or engineering monitoring results under that working condition. After comparison, the scene to be analyzed is first classified based on at least one of the working condition parameters, loading path, and contact surface state change characteristics. Then, according to the scene classification results, the comparison relationship between the prediction error and the preset error threshold, the corresponding model update strategy is adaptively selected. The preset error threshold is dynamically adjusted according to at least one of the scene type, historical prediction error distribution, and engineering safety level.

7. The method for intelligent prediction and parameter optimization of rock mass mechanical properties based on digital twins according to claim 6, characterized in that, The model update strategy includes at least one of incremental training, transfer learning, and model retraining; the model update process includes at least one of adaptive adjustment of model parameters, reconfiguration of model structure, and redistribution of feature weights, wherein the feature weight redistribution process is simultaneously influenced by both mechanical and physical prior constraints and data-driven optimization results.

Images