Large chamber surrounding rock stress characteristic dynamic prediction method and device

Abstract

The invention provides a large chamber surrounding rock stress characteristic dynamic prediction method and device, and relates to the technical field of underground engineering surrounding rock stability prediction. The method comprises the following steps: establishing a loss function of a dynamic multi-physics field coupled neural network framework by combining a data fitting item residual item, a time sequence continuity regularization item and a large chamber surrounding rock mixed physical constraint loss function constructed based on a layered time sequence parameter, a multi-source physics field parameter and a geological parameter; layered time sequence characteristics of layered excavation of the surrounding rock of the large chamber are extracted, and sampling data are dynamically determined; then, the weights of different physical fields in the large-chamber surrounding rock mixed physical constraint loss function are optimized; and according to the optimized loss function, obtaining a trained dynamic multi-physics field coupled neural network framework, and then generating a dynamic prediction result of the stress characteristics of the surrounding rock of the large chamber. According to the method, the problems of multi-field coupling equation embedding, time sequence dynamic feature capturing and high-risk region data optimization can be effectively solved.

Description

Background Art

[0002] In underground projects such as water conservancy and hydropower, and mining, the stability of the surrounding rock of large chambers is directly related to project safety. During the process of layered excavation, the surrounding rock will experience complex elastoplastic deformation, seepage pressure changes, and damage accumulation. Its mechanical properties exhibit multi-physical field coupling effects and time-series dynamic evolution characteristics. Although traditional numerical simulation methods such as the finite element method can partially describe the mechanical behavior of the surrounding rock, they have the following limitations: First, existing methods simplify the coupling process of elastoplasticity, seepage, and damage fields. For example, the dynamic influence of damage on seepage channels is ignored, or a step-by-step solution strategy leads to error accumulation, making it difficult to truly reflect the interaction of multiple fields. Second, the deformation and damage of the surrounding rock during the layered excavation process have strong time-series dependence, while traditional methods rely on static parameters or segmented assumptions and cannot capture the dynamic correlation between excavation stages. Third, machine learning models based on pure data-driven lack physical law constraints, and their prediction results are prone to deviate from actual physical laws when the training data is insufficient or the working conditions are complex. Moreover, existing physics-informed neural network frameworks are mostly designed for a single physical field and do not effectively integrate multi-field coupling equations and time-series characteristics. Finally, traditional sampling strategies are difficult to optimize data for high-risk areas, resulting in low model training efficiency and insufficient prediction accuracy in key areas. At the same time, the changing geological conditions and the dynamic process of layered excavation require the model to have the ability to adaptively adjust weights, while existing methods usually use fixed weights and are difficult to adapt to complex working conditions. In recent years, some studies have tried to introduce physics-informed neural networks into the field of geotechnical engineering, but their applications are still limited to single physical fields or static working conditions, and they have not effectively solved key problems such as the neural network embedding of multi-field coupling control equations, the dynamic integration of layered excavation time-series data and physical constraints, and adaptive sampling and model optimization for high-risk areas. Summary of the Invention

[0003] To overcome the problems existing in the related art, the present invention provides a training method and device for a neural network framework with dynamic multi-physical field coupling.

[0004] According to the first aspect of the embodiments of the present invention, a training method for a neural network framework with dynamic multi-physical field coupling is provided. The method includes: Obtain the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; Based on the layered time-series parameters, multi-source physical field parameters, and geological parameters, construct a hybrid physical constraint loss function for the surrounding rock of the large chamber; Combine the hybrid physical constraint loss function of the surrounding rock of the large chamber, the residual term of the data fitting term, and the time-series continuity regularization term to establish a loss function for the neural network framework with dynamic multi-physical field coupling; Extract the layered time-series characteristics of the surrounding rock deformation, stress, seepage, and damage from the real-time data of the layered excavation of the surrounding rock of the large chamber; Based on the hierarchical temporal features, a dual-threshold dynamic sampling strategy is used to dynamically determine the sampling data; According to the sampling data, an adaptive weight adjustment algorithm is used to optimize the weights of different physical fields in the mixed physical constraint loss function of the surrounding rock of the large chamber, and an optimized loss function is obtained; According to the optimized loss function, a trained neural network framework for dynamic multi-physical field coupling is obtained.

[0005] In some exemplary embodiments of the present invention, based on the foregoing solution, the hierarchical temporal parameters include: the current excavation layer number , the height of the -th excavation layer , the span of the -th excavation layer , the continuous excavation time of the -th excavation layer and the total number of excavated layers ; The multi-source physical field parameters include: the displacement field of the layer before the -th excavation layer , the stress field of the layer before the -th excavation layer , the seepage pressure field of the layer before the -th excavation layer and the damage factor of the layer before the -th excavation layer ; The geological parameters include: permeability coefficient , damage threshold, critical seepage pressure gradient .

[0006] In some exemplary embodiments of the present invention, based on the foregoing solution, the mixed physical constraint loss function of the surrounding rock of the large chamber includes: Based on the elastoplastic constitutive relationship, a residual term of the stress constraint equation is established; Based on Darcy's law and the damage-seepage coupling relationship, a constraint residual term formed by the seepage field is constructed; Based on the damage evolution theory and the excavation time sequence, a residual term of the constraint equation is constructed; The residual term of the stress constraint equation, the constraint residual term formed by the seepage field, and the residual term of the constraint equation are fitted to obtain the mixed physical constraint loss function of the surrounding rock of the large chamber.

[0007] In some exemplary embodiments of the present invention, based on the foregoing solution, by combining the mixed physical constraint loss function of the surrounding rock of the large chamber, the residual term of the data fitting term, and the temporal continuity regularization term, the loss function of the neural network framework for dynamic multi-physical field coupling includes: Based on the predicted displacement values, measured displacement values, predicted damage values, and measured damage values of each layer in all the excavated layers, calculate the residual terms of the data fitting terms for all the excavated layers; Based on the continuous excavation time of each layer in all the excavated layers and the derivative result of the change rate of the predicted displacement values, calculate the temporal continuity regularization terms for all the excavated layers; Use the first weight coefficient to balance the loss function of the hybrid physical constraints of the surrounding rock of the large chamber to obtain the first balance result; Use the second weight coefficient to balance the temporal continuity regularization terms to obtain the second balance result; Perform a summation calculation on the first balance result, the second balance result, and the residual terms of the data fitting terms to obtain the loss function of the neural network framework for dynamic multi-physical field coupling.

[0008] In some exemplary embodiments of the present invention, based on the foregoing solution, the hierarchical temporal features include strain energy density and damage factor. Based on the hierarchical temporal features, using a dual-threshold dynamic sampling strategy, the dynamically determined sampling data includes: Calculate the probability of the regional sampling points according to the strain energy density and the damage factor; Determine the sampling priority based on the probability of the regional sampling points; Dynamically adjust the sampling points based on the sampling priority; Collect the sampling data of the sampling points.

[0009] In some exemplary embodiments of the present invention, based on the foregoing solution, calculating the probability of the regional sampling points according to the strain energy density and the damage factor includes: Calculate the exponential function value of the strain energy density of the th sampling point and the first preset sensitivity coefficient; Calculate the Gaussian error function value of the damage factor of the th sampling point and the second preset sensitivity coefficient; Perform a summation calculation on the product of the exponential function values and the Gaussian error function values of all sampling points to obtain a summation value; According to the product of the exponential function value and the Gaussian error function value of the th sampling point and the summation value, calculate the probability of the th sampling point.

[0010] According to the second aspect of the embodiments of the present invention, there is provided a method for dynamically predicting the mechanical properties of the surrounding rock of a large chamber, including: Obtain the hierarchical temporal parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; Construct a loss function for the hybrid physical constraints of the surrounding rock of the large chamber based on the stratified timing parameters, multi-source physical field parameters, and geological parameters; Combine the loss function of the hybrid physical constraints of the surrounding rock of the large chamber, the residual term of the data fitting term, and the timing continuity regularization term to establish a loss function for the neural network framework of dynamic multi-physical field coupling; Extract the stratified timing characteristics of the surrounding rock deformation, stress, seepage, and damage from the real-time data of the stratified excavation of the surrounding rock of the large chamber; Based on the stratified timing characteristics, use a double-threshold dynamic sampling strategy to dynamically determine the sampling data; According to the sampling data, use an adaptive weight adjustment algorithm to optimize the weights of different physical fields in the loss function of the hybrid physical constraints of the surrounding rock of the large chamber to obtain an optimized loss function; According to the optimized loss function, obtain a trained neural network framework for dynamic multi-physical field coupling.

[0011] Input the stratified timing parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber, and use the trained neural network framework for dynamic multi-physical field coupling to generate a dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber. The dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

[0012] In some exemplary embodiments of the present invention, based on the foregoing solution, the neural network framework for dynamic multi-physical field coupling includes: An input module, a residual module, a time series module, a feature fusion module, and an output module; Among them, the input module is used to input the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters. The residual module is built based on the residual network. Based on the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters, a loss function of a dynamic multi-physical field coupling neural network framework established by using the constraint equation residual term formed based on the elastoplastic constitutive model, the constraint residual term formed based on the seepage field of Darcy's law, and the constraint equation residual term formed based on damage evolution is used to generate the differences among the displacement field, the stress field, the seepage field, and the damage factor. The time series module is used to generate the preliminary prediction laws of the displacement field, the stress field, the seepage field, and the damage factor with the development of time according to the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters. The feature fusion module introduces a data fitting term residual term and a time series continuity regularization term, and is used to superimpose and fuse the preliminary prediction laws of the displacement field, the stress field, the seepage field, and the damage factor with the development of time output by the time series module and the differences among the displacement field, the stress field, the seepage field, and the damage factor calculated by the residual module to obtain a fusion result. The output module is used to generate a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor based on the fusion result.

[0013] According to the third aspect of the embodiments of the present invention, there is provided a device for dynamically predicting the mechanical characteristics of the surrounding rock of a large chamber, including: A data acquisition module, configured to acquire the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters of the surrounding rock of the large chamber; A first function construction module, configured to construct a mixed physical constraint loss function of the surrounding rock of the large chamber based on the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters; A second function construction module, configured to establish a loss function of a dynamic multi-physical field coupling neural network framework by combining the mixed physical constraint loss function of the surrounding rock of the large chamber, a data fitting term residual term, and a time series continuity regularization term; A feature extraction module, configured to extract the hierarchical time series features of the surrounding rock deformation, stress, seepage, and damage in the real-time data of the layered excavation of the surrounding rock of the large chamber; A sampling data determination module, configured to dynamically determine sampling data based on the hierarchical time series features by using a dual-threshold dynamic sampling strategy; A function optimization module, configured to optimize the weights of different physical fields in the mixed physical constraint loss function of the surrounding rock of the large chamber by using an adaptive weight adjustment algorithm according to the sampling data to obtain an optimized loss function; A network determination module, configured to obtain a trained dynamic multi-physical field coupling neural network framework according to the optimized loss function; A prediction module, configured to input the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber, and generate a dynamic prediction result of the mechanical characteristics of the surrounding rock of the large chamber by using the trained neural network framework with dynamic multi-physical field coupling. The dynamic prediction result of the mechanical characteristics of the surrounding rock of the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

[0014] According to a fourth aspect of the embodiments of the present invention, there is provided an electronic device, including: a processor; and a memory, where computer-readable instructions are stored on the memory, and when the computer-readable instructions are executed by the processor, the dynamic prediction method for the mechanical characteristics of the surrounding rock of the large chamber in the second aspect is implemented.

[0015] According to a fifth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the dynamic prediction method for the mechanical characteristics of the surrounding rock of the large chamber in the second aspect is implemented.

[0016] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: By constructing a hybrid physical constraint loss function that integrates elastoplasticity, seepage, and damage fields, and combining time-series continuity regularization and dynamic sampling strategies, the present invention can effectively solve the problems of embedding multi-field coupling equations, capturing time-series dynamic features, and optimizing data in high-risk areas, and has the advantages of improving the prediction accuracy of the mechanical characteristics of the surrounding rock under complex geological conditions and the generalization ability of the model.

[0017] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The accompanying drawings herein are incorporated into the specification and constitute a part of the present invention, showing embodiments consistent with the present invention, and are used together with the specification to explain the principles of the present invention.

[0019] Figure 1 A schematic diagram of a system architecture showing an exemplary application environment of a training method and device for a neural network framework with dynamic multi-physical field coupling to which the embodiments of the present invention can be applied; Figure 2 A schematic flowchart showing the training method of a neural network framework with dynamic multi-physical field coupling according to some embodiments of the present invention; Figure 3 A schematic flowchart showing the dynamic prediction method for the mechanical characteristics of the surrounding rock of a large chamber according to some embodiments of the present invention; Figure 4 A schematic structural diagram showing a neural network framework with dynamic multi-physical field coupling according to some embodiments of the present invention; Figure 5 Schematically shows a schematic diagram of a device for dynamically predicting the stress characteristics of the surrounding rock of a large chamber according to some embodiments of the present invention; Figure 6 Schematically shows a schematic diagram of the structure of a computer system of an electronic device according to some embodiments of the present invention; Figure 7 Schematically shows a schematic diagram of a computer-readable storage medium according to some embodiments of the present invention. Detailed implementation manners

[0020] Here, exemplary embodiments will be described in detail, and examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0021] The terms used in the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The singular forms "a", "the", and "said" used in the present invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0023] Figure 1 Shows a schematic diagram of the system architecture of an exemplary application environment of a method and device for dynamically predicting the stress characteristics of the surrounding rock of a large chamber to which the embodiments of the present invention can be applied.

[0024] As Figure 1As shown, the system architecture 100 may include one or more of terminal devices such as a desktop computer 101, a portable computer 102, a smart phone 103, etc., a network 104, and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices and the server 105. The network 104 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc. The terminal device may be various electronic devices with data processing functions, and a display screen is provided on the electronic device, and the display screen is used to display the dynamic prediction process and / or results of the stress characteristics of the surrounding rock of the large chamber to the user, including but not limited to the above-mentioned desktop computer, portable computer, smart phone, etc. It should be understood that Figure 1 the numbers of the terminal devices, the network, and the server in

[0025] The dynamic prediction method for the stress characteristics of the surrounding rock of the large chamber provided by the embodiments of the present invention can generally be executed by the terminal device. Correspondingly, the dynamic prediction device for the stress characteristics of the surrounding rock of the large chamber is generally set in the terminal device. However, those skilled in the art can easily understand that the dynamic prediction method for the stress characteristics of the surrounding rock of the large chamber provided by the embodiments of the present invention can also be executed by the server 105. Correspondingly, the dynamic prediction device for the stress characteristics of the surrounding rock of the large chamber can also be set in the server 105. No special limitation is made in this exemplary embodiment.

[0026] In addition, it should be understood that the dynamic prediction method for the stress characteristics of the surrounding rock of the large chamber in the embodiments of the present invention can be configured as a software module. In some implementation scenarios, the dynamic prediction solution for the stress characteristics of the surrounding rock of the large chamber of the present invention can be deployed independently to realize the dynamic prediction of the stress characteristics of the surrounding rock of the large chamber in different regions. In other implementation scenarios, the dynamic prediction solution for the stress characteristics of the surrounding rock of the large chamber of the present invention can be deployed in other software as a functional module of the software, such as being deployed in the analysis software of the surrounding rock of the large chamber. The present invention does not make special restrictions on the application manner of the dynamic prediction method for the stress characteristics of the surrounding rock of the large chamber.

[0027] In the traditional analysis of the stability of surrounding rock in underground engineering, the finite element method often adopts a step-by-step solution strategy to deal with the coupled problem of elastoplastic deformation and seepage field. For example, the damage evolution equation is simplified as a static parameter input, resulting in the neglect of the feedback effect of permeability change on the seepage channel. In terms of time-series dynamic modeling, existing numerical simulations rely on a fixed time step for layered excavation calculation and cannot effectively capture the dynamic correlation between the accumulation of surrounding rock damage and the excavation process. When using a pure data-driven neural network model, the strain energy density data distribution obtained based on the uniform sampling strategy is unbalanced, and the sample density is insufficient in the high-gradient stress area, resulting in a physical law conflict between the predicted result of the damage factor and the elastoplastic constitutive equation.

[0028] For example, during the layered excavation of the underground powerhouse of a hydropower station, the finite element model set the permeability coefficient to a fixed value without considering the The damage factor increases after layer excavation, resulting in the permeability change coefficient The actual situation of increasing flow rate will cause the prediction error of seepage pressure to exceed the benchmark value. When excavating a new layer, the previous stress field data is directly inherited, failing to reflect the impact of residual stress generated by the first five layers on the expansion of the plastic zone in the current layer. When using conventional convolutional neural networks to predict surrounding rock deformation, the sampling coverage of high strain energy density areas (over 1.5 kJ / m³) is insufficient, causing the predicted stress concentration in key areas to deviate from the true physical laws.

[0029] If these issues are not addressed, errors in multi-field coupled modeling will cause support structure design parameters to deviate from actual operating conditions, potentially leading to progressive instability of the surrounding rock mass. The lack of temporal dynamic characteristics will reduce the reliability of excavation process optimization decisions and increase the risk of sudden water inrush accidents. The disconnect between data sampling strategies and physical constraints will exacerbate model prediction inaccuracies in complex geological conditions, especially in areas with weak interlayers or developed joints. Insufficient model generalization will directly affect the scientific nature of engineering safety assessment conclusions.

[0030] When faced with the above problems, the present invention first considers how to effectively embed the elastic-plastic constitutive equation, Darcy's law and damage evolution equation into the neural network training framework to solve the defect of insufficient multi-field coupling modeling. Although the traditional step-by-step solution strategy can simplify the calculation process, it leads to the neglect of the dynamic feedback of damage on the seepage channel. For example, the permeability variation coefficient is not dynamically adjusted with the damage factor. To this end, the present invention attempts to introduce stress field residual terms, seepage field residual terms and damage evolution residual terms into the loss function at the same time to form a mixed physical constraint. However, if the weights of different physical fields are fixedly allocated, it is difficult to adapt to the dynamic changes of the main controlling physical field during the layered excavation process. For example, the first five layers of excavation are dominated by elastic-plastic deformation, and the weight of the seepage field needs to be increased due to damage accumulation in the last three layers of excavation.

[0031] To address the lack of dynamic time series characteristics, this paper explores the addition of a time series continuity regularization term in addition to the data fitting term. Traditional static parameter inputs cannot reflect the impact of the excavation process on residual stresses. For example, the expansion of the plastic zone in the previous excavation layer will change the stress distribution in the current layer. By designing a regularization term based on the rate of change of displacement, the neural network is forced to maintain physical continuity between the prediction results of adjacent excavation layers, avoiding the dynamic correlation caused by the simplified time step.

[0032] Regarding the disconnection between data-driven and physical constraints, the present invention analyzes the defect of insufficient coverage of the uniform sampling strategy in high-risk areas. For example, in areas with high strain energy density, the predicted stress deviates from the true law due to sparse samples. A dual-threshold dynamic sampling strategy is attempted to calculate the regional priority according to the strain energy density and damage factor, and dynamically adjust the distribution of sampling points. At the same time, combined with the adaptive weight adjustment algorithm, the weight coefficients of the physical field constraint terms are optimized according to the real-time sampling data during the training process. For example, when the detected seepage pressure gradient exceeds the critical value, the weight of the seepage field residual term is automatically increased.

[0033] In this regard, as Figure 2 shown, Figure 2 FIG. is a flowchart of a training method for a neural network framework of dynamic multi-physical field coupling according to an exemplary embodiment of the present invention, including the following steps: S210: Obtain the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; S220: Based on the layered time-series parameters, multi-source physical field parameters, and geological parameters, construct a hybrid physical constraint loss function for the surrounding rock of the large chamber; S230: Combine the hybrid physical constraint loss function of the surrounding rock of the large chamber, the data fitting term residual term, and the time-series continuity regularization term to establish a loss function for the neural network framework of dynamic multi-physical field coupling; S240: Extract the layered time-series characteristics of the surrounding rock deformation, stress, seepage, and damage from the real-time data of the layered excavation of the surrounding rock of the large chamber; S250: Based on the layered time-series characteristics, use the dual-threshold dynamic sampling strategy to dynamically determine the sampling data; S260: According to the sampling data, use the adaptive weight adjustment algorithm to optimize the weights of different physical fields in the hybrid physical constraint loss function of the surrounding rock of the large chamber to obtain an optimized loss function; S270: According to the optimized loss function, obtain a trained neural network framework for dynamic multi-physical field coupling.

[0034] Next, the embodiments of the present invention will be described in detail.

[0035] In S210, the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber are obtained.

[0036] The layered time-series parameters refer to the geometric and time parameters that describe the time-varying characteristics of the surrounding rock during the layered excavation process, which can be realized by using the current excavation layer number, layer height, span, continuous excavation time, and total layer number, and are used to capture the time-series dependence of the surrounding rock deformation and damage during the excavation process. It includes: the current excavation layer number , the layer height of the th excavation layer , the Span of the excavation layer , the Continuous excavation time of the excavation layer and the total number of excavated layers .

[0037] Multi-source physical field parameters refer to physical quantities that include displacement field, stress field, seepage pressure field, and damage factor, which can be realized by relevant variables of elastoplastic constitutive equations, Darcy's law, and damage evolution equations, and are used to establish physical constraint conditions for multi-field coupling. It includes: the Displacement field of the layer before the excavation layer , the Stress field of the layer before the excavation layer , the Seepage pressure field of the layer before the excavation layer and the Damage factor of the layer before the excavation layer .

[0038] Geological parameters include: permeability coefficient , damage threshold, critical seepage pressure gradient .

[0039] Current excavation layer number Can be set from 1 to 5 layers. For example, when excavating the third layer = 3; Layer height Can be 3 meters to 5 meters, span Can be 8 meters to 12 meters, continuous excavation time Can be 24 hours to 72 hours 。 Displacement field of the previous layer Can include displacement amounts in the X / Y / Z axes, that is , for example = [0.12m, 0.08m, -0.05m] represents the displacement state after the excavation of the second layer, stress field of the previous layer Can record the principal stress components in tensor form, for example = [15MPa, 10MPa, 8MPa]. Seepage pressure field Can be decomposed into pore water pressure components in three directions. For example, before the excavation of the third layer = 2.3kPa, = 1.8kPa, = 4.1 kPa. Damage factor The value range can be from 0 to 1. For example = 0.35 represents the damage degree after the excavation of the second layer. Permeability coefficient in geological parameters Can be cm / s to cm / s, the critical seepage pressure gradient can be from 0.5 kPa / m to 2.0 kPa / m.

[0040] Specifically, the current excavation layer number and the total number of excavated layers constitute the time series benchmark. For example, when = 5 and = 3, it indicates that the current is in the construction stage of the third layer in the five-layer excavation plan. The layer height and the span jointly define the geometric dimensions of the excavation space. For example, = 4.2 m and = 10.5 m combine to form the spatial characteristics of the third-layer excavation section. The continuous excavation time and the physical field parameters of the previous layer form a time-correlation constraint. For example, when = 48 hours, the model calculates the displacement increment within 48 hours of excavation of the current layer through the displacement field of the previous layer. The stress field of the previous layer and the damage factor of the current layer establish a cross-layer mechanical correlation. For example, when = 12 MPa and = 0.28, the model calculates the stress redistribution caused by the excavation of the third layer based on the elastoplastic constitutive equation. The seepage pressure field of the previous layer and the permeability coefficient form the iteration conditions of the seepage field. For example, = 2.1 kPa after the excavation of the second layer and = cm / s jointly constrain the calculation of the seepage velocity of the third layer. The damage threshold and the critical seepage pressure gradient jointly control the damage evolution process. For example, when = 1.2 kPa / m, the model automatically triggers the seepage-damage coupling calculation module. The combined application of these parameters enables the model to deduce the multi-field coupling effect of the current layer through the physical field state of the previous layer, avoid the error transmission caused by step-by-step solution, and at the same time ensure the applicability of the physical constraint equation under complex geological conditions by defining the rock mass property boundary with geological parameters.

[0041] Thus, the present invention can achieve the precise definition of the parameters of the dynamic multi-physical field coupling model. The hierarchical time-series parameters provide a spatio-temporal dynamic benchmark for layered excavation for the model, enabling the model to accurately track the time series and spatial scale changes during the excavation stage. The multi-source physical field parameters introduce the physical field states of adjacent excavation layers, forcing the model to learn the recursive relationship of multi-field coupling and avoiding the error accumulation in step-by-step solutions in traditional methods. The geological parameters are directly related to the seepage characteristics and damage evolution law of rock masses, providing necessary geological boundary conditions for the residual terms of the seepage field and damage field in the physical constraint equations, and ensuring the generalization ability of the model under complex geological conditions. This parameter definition method solves the key problems caused by the ambiguity of the parameters of the dynamic multi-physical field coupling model, improves the accuracy of the model in capturing the time-series dynamic characteristics of layered excavation during the dynamic multi-physical field coupling process, and realizes the reliable embedding of the multi-field coupling equations.

[0042] The design of traditional loss functions only uses a single physical field constraint equation or overly simplifies the multi-field coupling equation, resulting in the model being unable to accurately capture the dynamic response law of the surrounding rock under multi-field interaction. In response to this, in S220 of the present invention, a hybrid physical constraint loss function for the surrounding rock of large chambers is constructed based on the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters.

[0043] The hybrid physical constraint loss function refers to the loss term that fuses the residuals of the control equations of the stress field, seepage field, and damage field, which can be realized through the weighted combination of the elastoplastic constitutive residual, Darcy's law residual, and damage evolution residual, and is used to embed the multi-field coupling physical law into the neural network training process.

[0044] The elastoplastic deformation constraint term is constructed based on the improved elastoplastic constitutive equation, and its expression is the sum of the squares of the residuals between the predicted stress and the theoretically calculated stress. A damage factor is introduced as the material stiffness reduction coefficient in the theoretical stress calculation. For example, the value range of the elastic modulus can be from 10 GPa to 50 GPa, and the value range of the damage factor can be from 0 to 0.9, and the specific value depends on the type of surrounding rock. The seepage field constraint term is realized by modifying Darcy's law equation, and an adjustment term linearly related to the damage factor is added to the expression of the permeability coefficient. The value range of the permeability change coefficient can be from 0.1 to 5.0, which is used to characterize the amplification effect of different damage degrees on the permeability. The damage evolution constraint term adopts a cumulative form of time-series residual design, and the correlation between damage accumulation and construction procedures is established through the product term of the excavation time and the theoretical damage increment . The unit of the excavation time can be hours or days, specifically depending on the project progress requirements. The weight coefficient adopts an adaptive adjustment mechanism. For example, it can be set in the initial stage = 5:3:2, and the ratio is dynamically adjusted according to the convergence rate of the residuals of each physical field during the training process.

[0045] Specifically, during the neural network training process, the elastoplastic deformation constraint term forces the model to learn the material weakening effect caused by damage by embedding the damage factor into the stress calculation process. For example, when the damage factor reaches 0.5, the material stiffness will decay to 50% of the initial value, and at this time, the calculation of the elastic strain tensor needs to be adjusted accordingly. The seepage field constraint term establishes a quantitative relationship between damage and permeability through the (1 + ) term. When the damage factor increases by 0.1 each time, the permeability coefficient can be amplified by 1% to 50%, and the specific amplification factor is determined by the permeability change coefficient . The damage evolution constraint term captures the temporal cumulative effect of damage during the layered construction process by accumulating the damage increment residuals of each excavation stage. For example, extending the excavation time of each layer by 10% will result in an 8% - 15% increase in the damage increment. The weight coefficient adopts a dynamic adjustment algorithm based on the residual change rate. When the residual decline rate of a certain physical field is lower than the threshold, its corresponding weight coefficient can be increased by 10% - 30% to ensure the balance of multi-field coupling constraints. Through the cooperation with the double-threshold dynamic sampling strategy, this combined loss function preferentially collects sample data in high-damage areas, improving the prediction accuracy of the model in key areas by 15% - 25%.

[0046] As an implementation method, a mixed physical constraint loss function for the surrounding rock of large chambers can be designed including: Among them, represents the residual term of the constraint equation formed based on the elastoplastic constitutive model and , represents the total number of samples, represents the th sample, represents the predicted stress of the th sample, represents the damage factor of the th sample, represents the elastic modulus of the th sample, represents the elastic strain tensor of the th sample, represents the additional stress caused by plastic deformation of the th sample, represents the residual term of the constraint formed by the seepage field based on Darcy's law and , represents the permeability change coefficient of the th sample, Denote the predicted pore water pressure tensor of the th sample, Denote the permeability coefficient tensor of the th sample, Denote the coordinate of the th sample in the direction, Denote the coordinate of the th sample in the direction, Denote the seepage velocity of the th sample in the direction, Denote the derivative with respect to , Denote the residual term of the constraint equation formed based on damage evolution and , Denote the layer number of the excavation layer where the th sample is located, Denote the layer number of the current excavation layer, Is the predicted loss factor of the th sample after the excavation of the th layer ends, Denote the excavation layer and = 1, 2…; Is the predicted loss factor of the th sample after the excavation of the th layer ends; Is the continuous excavation time of the th layer, Is the theoretical loss increment per unit time of the th layer of excavation, , And Are the weights of , And respectively.

[0047] That is to say, The calculation method of the term is as follows: First, traverse the samples, and for each sample , calculate the Euclidean distance square between the predicted stress and the stress after considering the influence of the damage factor . Then sum the distance squares of all samples and divide by the total number of samples to obtain the value of

[0048] The calculation method of the term is: Traverse the , calculate the square of the Euclidean norm. Among them is the permeability change coefficient, is the damage factor, is the permeability coefficient tensor, is the pore water pressure. Then sum up the calculation results of all samples and divide by the total number of samples , to obtain value.

[0049] The calculation method of the item is: traverse samples, for each sample , from the excavation layer where the sample is located to the current excavation layer , calculate the difference in the damage factor between every two layers and the theoretical damage increment and take the absolute value of the difference. Then sum up all the inter-layer differences, and sum up the calculation results of all samples and divide by the total number of samples , to obtain value.

[0050] Finally, multiply , and by the corresponding weight coefficients , and , and add the three together to obtain the final hybrid physical constraint loss function of the surrounding rock of the large chamber .

[0051] Thus, the present invention constructs a hybrid physical constraint loss function that integrates the coupling effects of multiple physical fields. This loss function can effectively balance the coupling relationship among elastoplastic deformation, dynamic changes in the seepage field, and damage evolution. Specifically, the item can directly associate stress calculation with material damage by introducing the damage factor ; the item can capture the dynamic effect of damage on the permeability coefficient through (1 + ); the item adopts a residual design in an accumulative form, which can reflect the chronological cumulative effect of damage evolution during the layered excavation process. Thus, this loss function can accurately capture the dynamic response law of the surrounding rock under the interaction of multiple fields, and overcome the problem of excessive simplification of the multi-field coupling equation in traditional methods. Further, by introducing weight coefficients , and , enabling the present invention to allow dynamic adjustment of the constraint intensity according to the importance of each physical field residual during the training process, avoiding the weakening of the constraint effect of certain physical fields.

[0052] Since relying solely on the physical constraint loss function in the dynamic multi-physical field coupling scenario may cause the model to overly rely on theoretical equations and ignore the fitting of measured data, and at the same time, the dynamic characteristics of the layered excavation time sequence are not explicitly modeled, resulting in jumps or discontinuities in the prediction results in the time dimension. Therefore, in S230, a loss function for the neural network framework of dynamic multi-physical field coupling is established by combining the mixed physical constraint loss function of the large chamber surrounding rock, the residual term of the data fitting item, and the time sequence continuity regularization term.

[0053] The residual term of the data fitting item ensures the consistency between the model output and the real observed data by comparing the mean square error of the predicted displacement and the measured displacement layer by layer, and the mean square error of the predicted damage value and the measured damage value. The time sequence continuity regularization term enforces the model to learn the continuous evolution law of the displacement field by constraining the matching between the displacement change rate of adjacent excavation layers and the time interval. The weight coefficients and are set as dynamically adjustable parameters, and the contribution degrees of the physical constraint term and the time continuity term are optimized through an adaptive weight adjustment algorithm. In areas with higher strain energy density, the value of can be appropriately increased to strengthen the physical constraint effect, and in areas sensitive to displacement mutations, the value of can be increased to enhance the time sequence continuity constraint.

[0054] Specifically, the calculation of the residual term of the data fitting item involves comparing the norms of the predicted displacement and the measured displacement of each excavation layer. For example, when the measured displacement data of a certain excavation layer is missing, the system automatically reduces the weight proportion of the data item of this layer. The time sequence continuity regularization term establishes the correlation between the displacement field change rate and the actual excavation time in the form of mathematical derivatives. For example, when the time interval between adjacent excavation layers is 2 days, the model automatically calculates the reasonable gradient range of the displacement change during this period. The weight coefficients and are dynamically adjusted using an exponential decay strategy, set to 0.8 and 0.5 at the beginning of training to strengthen data fitting, and gradually adjusted to 0.3 and 0.7 as the number of iterations increases to enhance physical regularity. Through this composite loss function structure, the model can not only follow the constitutive equation of elastoplastic mechanics, but also maintain consistency with the measured data, while ensuring the continuous evolution characteristics of the prediction results on the time axis.

[0055] As an implementation method, it is possible to design a loss function for the neural network framework of dynamic multi-physical field coupling by combining the mixed physical constraint loss function of the large chamber surrounding rock, the residual term of the data fitting item, and the time sequence continuity regularization term including: Among them, represents the residual term of the data fitting item and , represents the total number of excavated layers, represents the predicted displacement value of the layer of excavation, represents the measured displacement value of the layer of excavation, predicted damage value of the layer of excavation, represents the measured damage value of the layer of excavation, represents the temporal continuity regularization term and is the continuous excavation time of the layer of excavation, is the continuous excavation time of the layer of excavation, represents the time variable symbol of the continuous excavation time of the layer of excavation, that is, represents the derivative of displacement, the rate of change of displacement; represents the loss function of the mixed physical constraints of the surrounding rock of the large chamber, and are respectively the

[0056] weight coefficients of

[0057] Through the above technical solutions, the present invention can achieve multi-dimensional fusion of data-driven, physical constraints and temporal dynamics, and solve the collaborative optimization problem of model generalization and physical regularity in multi-field coupling prediction. The residual term of the data fitting item directly constrains the deviation between the model output and the engineering measured data, avoiding overfitting caused by pure physical constraints. The temporal continuity regularization term forces the model to capture the continuous gradual change characteristics of the surrounding rock deformation during the excavation process, and can eliminate the prediction jumps of the traditional static model in the time dimension. The introduction of the weight coefficients enables the model to adaptively adjust the weight ratio of the physical constraint term and the time continuity term according to the importance differences of data quality and physical laws under different working conditions. This multi-objective optimization method enables the neural network framework to not only follow basic theories such as elastoplastic mechanics and seepage mechanics, but also fit the actual monitoring data, while ensuring the physical rationality of the prediction results on the time axis, improving the accuracy and reliability of the dynamic prediction of the mechanical properties of the surrounding rock of the large chamber.

[0058] The present invention does not specifically limit the method for extracting hierarchical temporal features. In some embodiments, a hierarchical temporal convolution module or a spatio-temporal attention mechanism can be designed to extract the hierarchical temporal features of surrounding rock deformation, stress, seepage, and damage from the real-time data of the layered excavation of large cavern surrounding rock.

[0059] In the embodiments of the present invention, by designing a basic framework based on a bidirectional LSTM layer or a temporal convolution layer, and arranging the sampled time points in an orderly manner as a time series module to extract the hierarchical temporal features of surrounding rock deformation, stress, seepage, and damage from the real-time data of the layered excavation of large cavern surrounding rock.

[0060] In S250, based on the hierarchical temporal features, a dual-threshold dynamic sampling strategy is used to dynamically determine the sampling data.

[0061] The dual-threshold dynamic sampling strategy refers to a method for calculating the sampling priority of regions based on strain energy density and damage factor, which can be achieved by combining an exponential function and an error function to calculate the probability distribution, and is used to preferentially collect data in high-risk regions to improve the prediction accuracy of key regions.

[0062] The traditional uniform sampling strategy cannot effectively capture the key regions of high strain energy density and high damage factor, resulting in insufficient prediction accuracy for high-risk regions during model training. At the same time, the static sampling mechanism is difficult to adapt to the dynamic evolution of the physical field of surrounding rock during the layered excavation process, causing the dual problems of data redundancy and loss of key features.

[0063] In response to this, in the embodiments of the present invention, the designed hierarchical temporal features include strain energy density and damage factor. On this basis, based on the hierarchical temporal features, the dynamic determination of sampling data using the dual-threshold dynamic sampling strategy includes: Calculating the regional sampling priority according to the strain energy density and the damage factor; Dynamically adjusting the sampling points based on the regional sampling priority; Collecting the sampling data of the sampling points.

[0064] The strain energy density becomes a key index for identifying high-energy aggregation regions by quantifying the energy accumulation degree during the elastoplastic deformation process of the surrounding rock. For example, the measured strain energy density in the edge region of the excavation layer can reach the order of 3.5 - 5.2 MJ / m³. The damage factor is calculated using a continuous damage mechanics model to reflect the degree of crack propagation inside the rock mass. When the damage factor exceeds 0.65, it is determined as a high-risk region. The calculation of the regional sampling priority combines an exponential function and an error function, and the specific expression is: Wherein, represents the probability of the th sampling point, represents the first preset sensitivity coefficient, represents the second preset sensitivity coefficient, represents the strain energy density of the th sampling point, represents the damage factor of the th sampling point, represents the total number of sampling points, represents the th sampling point, represents the exponential function,

[0065]

[0066] Specifically, during the layered excavation process, after each layer of excavation is completed, the distribution data of the strain energy density field and the damage factor field of the current layer are obtained synchronously. Through the dual-threshold dynamic sampling strategy, the priority of all grid cells is sorted first. The area where the strain energy density is higher than 3.2 MJ / m³ and the damage factor exceeds 0.6 is automatically marked as the first-level sampling area. Subsequently, according to the priority calculation result, the sampling point density in the first-level sampling area is increased to 2-3 times that of the normal area, while the sampling point spacing in the low-priority area is expanded to 1.5 times the original spacing. During the sampling data acquisition process, by real-time monitoring the evolution characteristics of the physical field, when it is detected that the growth rate of the strain energy density in a certain area after two adjacent layers of excavation exceeds 25%, the encrypted sampling mechanism in this area is automatically triggered. The collected sampling data is processed by the adaptive weight adjustment algorithm and used to optimize the weight allocation of each physical field in the hybrid physical constraint loss function. Through this dynamic sampling mechanism, the data acquisition efficiency in the excavation layer transition area is increased by more than 40%, and the stress prediction error in the key area is reduced to within 5.8%.The exponential function is configured to perform non-linear amplification on the high strain energy density area. The value range of the first preset sensitivity coefficient can be adjusted between 0.5 and 3.0. For example, in areas with complex geological conditions, can be set to 2.5 to strengthen the strain energy monitoring. The error function is configured to perform probabilistic processing on the damage factor. The value range of the second preset sensitivity coefficient can be adjusted between 0.3 and 1.2. For example, in areas with significant seepage, can be set to 1.0 to focus on damage assessment. The normalization process of the denominator part ensures that the total priority of each area is 1, forming a dynamic weight allocation mechanism.

[0067] When the strain energy density exceeds the critical value of 1.5 MJ / m³, the non-linear amplification effect of the exponential function can increase the priority of the corresponding area by more than 3 times, effectively capturing the energy concentration area. The error function will Mapped to the interval from 0 to 1. When reaching the damage threshold of 0.8, the output value of the error function can reach above 0.9, significantly distinguishing the damage degree. By dynamically adjusting and the combined parameters, the sampling priority allocation can be adapted to the requirements of different engineering stages. For example, at the initial stage of excavation, set : = 2:1, focusing on energy monitoring. At the later stage of excavation, adjust it to 1:2, focusing on damage assessment. The dynamic adjustment of sampling points is based on the priority parameters calculated in real time, and the areas with a probability value higher than 0.15 are preferentially selected for intensive sampling, while the areas with a probability value lower than 0.05 will reduce the sampling frequency or even be eliminated. This dynamic allocation mechanism can increase the sampling density of key areas by 2 - 3 times compared with uniform sampling, while reducing the redundant sampling ratio by more than 40%.

[0068] Based on this, the present invention can not only ensure that the spatio - temporal distribution of the neural network training data set matches the dynamic process of multi - field coupling of surrounding rock, improving the prediction sensitivity and training efficiency of the model for key dangerous areas. It can also effectively overcome the defects of data redundancy and insufficient feature capture of traditional uniform sampling under complex working conditions, capture the key areas of high strain energy density and high damage factors, and improve the prediction accuracy of the model for high - risk areas.

[0069] In S260, according to the sampling data, use the adaptive weight adjustment algorithm to optimize the weights of different physical fields in the hybrid physical constraint loss function of the large chamber surrounding rock, and obtain the optimized loss function.

[0070] The adaptive weight adjustment algorithm refers to a mechanism that dynamically optimizes the weight coefficients of different physical fields according to real - time sampling data. It can be realized by using an adjustment strategy that correlates gradient back - propagation error and residual contribution degree, and is used to balance the weight distribution of the multi - field coupling equation in the loss function.

[0071] The present invention does not limit the specific content of the adaptive adjustment algorithm. For example, in some embodiments, the adaptive adjustment algorithm can be a particle swarm optimization algorithm, a genetic algorithm, a simulated annealing algorithm, etc.

[0072] After that, use the sampling data as the input data of the adaptive stiffness adjustment algorithm to optimize and adjust the hybrid physical constraint loss function of the large chamber surrounding rock, and obtain the optimized loss function.

[0073] In S270, according to the optimized loss function, obtain the trained neural network framework of dynamic multi - field coupling.

[0074] That is to say, the optimized loss function is used as the loss function of the neural network framework for dynamic multi-physical field coupling, and then the trained neural network framework for dynamic multi-physical field coupling is obtained.

[0075] According to the second aspect of the embodiments of the present invention, there is also provided a method for dynamically predicting the mechanical properties of surrounding rock in a large chamber, including: S310: Obtain the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock in the large chamber; S320: Based on the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters, construct a hybrid physical constraint loss function for the surrounding rock in the large chamber; S330: Combine the hybrid physical constraint loss function for the surrounding rock in the large chamber, the data fitting term residual term, and the time-series continuity regularization term to establish a loss function for the neural network framework for dynamic multi-physical field coupling; S340: Extract the hierarchical time-series characteristics of the surrounding rock deformation, stress, seepage, and damage from the real-time data of the layered excavation of the surrounding rock in the large chamber; S350: Based on the hierarchical time-series characteristics, use a double-threshold dynamic sampling strategy to dynamically determine the sampling data; S360: According to the sampling data, use an adaptive weight adjustment algorithm to optimize the weights of different physical fields in the hybrid physical constraint loss function for the surrounding rock in the large chamber to obtain an optimized loss function; S370: According to the optimized loss function, obtain a trained neural network framework for dynamic multi-physical field coupling; S380: Input the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock in the large chamber, and use the trained neural network framework for dynamic multi-physical field coupling to generate a dynamic prediction result of the mechanical properties of the surrounding rock in the large chamber. The dynamic prediction result of the mechanical properties of the surrounding rock in the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

[0076] In S310, the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock in the large chamber are obtained.

[0077] This part has been described in detail in S210, and the present invention will not repeat it here.

[0078] In S320, based on the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters, a hybrid physical constraint loss function for the surrounding rock in the large chamber is constructed.

[0079] This part has been described in detail in S220, and the present invention will not repeat it here.

[0080] In S330, a loss function of a neural network framework for dynamic multi - physical field coupling is established by combining the loss function of the mixed physical constraints of the surrounding rock of the large chamber, the residual term of the data fitting term, and the temporal continuity regularization term.

[0081] This part has been described in detail in S230, and the present invention will not elaborate here.

[0082] In S340, the hierarchical temporal features of the surrounding rock deformation, stress, seepage, and damage are extracted from the real - time data of the layered excavation of the surrounding rock of the large chamber.

[0083] This part has been described in detail in S240, and the present invention will not elaborate here.

[0084] In S350, based on the hierarchical temporal features, a double - threshold dynamic sampling strategy is used to dynamically determine the sampling data.

[0085] This part has been described in detail in S250, and the present invention will not elaborate here.

[0086] In S360, according to the sampling data, an adaptive weight adjustment algorithm is used to optimize the weights of different physical fields in the loss function of the mixed physical constraints of the surrounding rock of the large chamber, and an optimized loss function is obtained.

[0087] This part has been described in detail in S260, and the present invention will not elaborate here.

[0088] In S370, according to the optimized loss function, a trained neural network framework for dynamic multi - physical field coupling is obtained.

[0089] This part has been described in detail in S270, and the present invention will not elaborate here.

[0090] S380: Input the hierarchical temporal parameters, multi - source physical field parameters, and geological parameters of the surrounding rock of the large chamber, and use the trained neural network framework for dynamic multi - physical field coupling to generate a dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber. The dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

[0091] In some embodiments, as shown in Figure 4 the neural network framework for dynamic multi - physical field coupling includes: an input module, a residual module, a time - series module, a feature fusion module, and an output module; Among them, the input module is used to input the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters. The residual module is built based on the residual network. Based on the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters, using the loss function of the dynamic multi-physical field coupling neural network framework established by the constraint equation residual term formed based on the elastoplastic constitutive model, the constraint residual term formed based on the seepage field of Darcy's law, and the constraint equation residual term formed based on the damage evolution, generate the differences between the displacement field, the stress field, the seepage field, and the damage factor. The time series module is used to generate the preliminary prediction rules of the displacement field, the stress field, the seepage field, and the damage factor with the development of time according to the hierarchical time series parameters, the multi-source physical field parameters, and the geological parameters. The feature fusion module introduces the data fitting term residual term and the time series continuity regularization term, and is used to superimpose and fuse the preliminary prediction rules of the displacement field, the stress field, the seepage field, and the damage factor with the development of time output by the time series module and the differences between the displacement field, the stress field, the seepage field, and the damage factor calculated by the residual module to obtain the fusion result. The output module is used to generate the predicted displacement field, the predicted stress field, the predicted seepage field, and the predicted damage factor based on the fusion result.

[0092] The skip connection structure of the residual module can be configured to have 3 residual blocks, each residual block contains two convolutional layers and the ReLU activation function, and the gradient vanishing problem is alleviated through cross-layer connection. The time step of the time series module corresponds to the number of excavation layers. For example, when processing the excavation data of the th layer, the time series module can store the historical states of the previous layers. At the same time, the prediction of the th layer is carried out by combining the time series prediction method. The splicing dimension of the feature fusion module can be dynamically adjusted according to the size of the input feature map. For example, on a 256×256 feature map, 64 channels of static features and 64 channels of dynamic features are spliced into 128 channels along the channel dimension.

[0093] Specifically, the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber enter the neural network framework of dynamic multi-physical field coupling through the input module. The residual module and the time-series module call relevant data for calculation respectively: through the constraint equations formed based on elastoplastic constitutive relations, the constraints formed based on the seepage field of Darcy's law, and the constraint equations formed based on damage evolution introduced by the residual module, the residual terms of relevant equations are calculated, specifically the displacement field, stress field, seepage field, and damage factor input and the displacement field, stress field, seepage field, and damage factor calculated through physical constraint equations. The residual values of these different types form multiple residual blocks in the residual network; through the time-series prediction method built by the time-series module, such as introducing a bidirectional LSTM layer or a temporal convolutional layer, preliminary predicted values of the displacement field, stress field, seepage field, and damage factor over time are obtained. Through the feature fusion module, the data fitting term residual term and the temporal continuity regularization term are introduced, and the displacement field, stress field, seepage field, and damage factor values preliminarily predicted by the time-series module are superimposed and fused with the residuals calculated by the residual module to form a complete neural network of dynamic multi-physical field coupling, and further the displacement field, stress field, seepage field, and damage factor can be predicted and output.

[0094] Thus, the present invention can effectively separate the encoding processes of static field data and dynamic time-series features, avoiding prediction deviations caused by feature confusion; the residual network structure retains shallow geometric information through skip connections, improving the stability of deep feature extraction; the time-series module adopts a bidirectional cyclic structure to capture the forward and backward dependencies during the excavation process, enhancing the modeling accuracy of the damage accumulation effect; the feature fusion module realizes spatio-temporal feature complementarity through channel splicing, strengthening the joint representation of the multi-physical field coupling relationship. This framework optimizes the interaction mechanism between multi-field coupling equations and time-series dynamic features through modular design, significantly improving the prediction robustness of the layered excavation process under complex geological conditions.

[0095] According to the third aspect of the embodiments of the present invention, there is also provided a device for dynamically predicting the mechanical properties of the surrounding rock of a large chamber, as shown in Figure 5 and including: A data acquisition module 510, configured to acquire the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; A first function construction module 520, configured to construct a mixed physical constraint loss function for the surrounding rock of the large chamber based on the layered time-series parameters, multi-source physical field parameters, and geological parameters; A second function construction module 530, configured to establish a loss function for the neural network framework of dynamic multi-physical field coupling by combining the mixed physical constraint loss function for the surrounding rock of the large chamber, the data fitting term residual term, and the temporal continuity regularization term; A feature extraction module 540 is configured to extract the hierarchical time-series features of surrounding rock deformation, stress, seepage, and damage from the real-time data of the layered excavation of the surrounding rock of the large chamber. A sampling data determination module 550 is configured to dynamically determine sampling data based on the hierarchical time-series features by using a dual-threshold dynamic sampling strategy. A function optimization module 560 is configured to optimize the weights of different physical fields in the hybrid physical constraint loss function of the surrounding rock of the large chamber according to the sampling data by using an adaptive weight adjustment algorithm, and obtain an optimized loss function. A network determination module 570 is configured to obtain a trained neural network framework with dynamic multi-physical field coupling according to the optimized loss function. A prediction module 580 is configured to input the hierarchical time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber, and use the trained neural network framework with dynamic multi-physical field coupling to generate a dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber. The dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

[0096] It should be noted that although several modules of the dynamic prediction device for the mechanical properties of the surrounding rock of the large chamber are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present invention, the features and functions of two or more of the above-described modules can be embodied in one module or unit. Conversely, the features and functions of one module described above can be further divided into being embodied by multiple modules or sub-modules.

[0097] In addition, in the exemplary embodiment of the present invention, an electronic device capable of implementing the above-described dynamic prediction method for the mechanical properties of the surrounding rock of the large chamber is also provided.

[0098] Those skilled in the art can understand that various aspects of the present invention can be implemented as a system, a method, or a program product. Therefore, various aspects of the present invention can be specifically implemented in the following forms, namely: a complete hardware embodiment, a complete software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects, which can be collectively referred to as "circuitry", "module", or "system" here.

[0099] Next, refer to Figure 6 to describe the electronic device 600 according to this embodiment of the present invention. Figure 6 The illustrated electronic device 600 is merely an example and should not impose any limitation on the functions and usage scope of the embodiments of the present invention.

[0100] As Figure 6As shown, the electronic device 600 is presented in the form of a general computing device. The components of the electronic device 600 may include, but are not limited to: at least one of the above-mentioned processing units 610, at least one of the above-mentioned storage units 620, a bus 630 connecting different system components (including the storage unit 620 and the processing unit 610), and a display unit 640.

[0101] Among them, the storage unit stores program code, and the program code can be executed by the processing unit 610, so that the processing unit 610 executes the steps according to various exemplary embodiments of the present invention described in the "Exemplary Method" section of the present invention above. For example, the processing unit 610 can execute as Figure 3 shown in S310: Obtain the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; S320: Based on the layered time-series parameters, multi-source physical field parameters, and geological parameters, construct a hybrid physical constraint loss function for the surrounding rock of the large chamber; S330: Combine the hybrid physical constraint loss function of the surrounding rock of the large chamber, the residual term of the data fitting term, and the time-series continuity regularization term to establish a loss function for the neural network framework of dynamic multi-physical field coupling; S340: Extract the layered time-series characteristics of the surrounding rock deformation, stress, seepage, and damage in the real-time data of the layered excavation of the surrounding rock of the large chamber; S350: Based on the layered time-series characteristics, use a double-threshold dynamic sampling strategy to dynamically determine the sampling data; S360: According to the sampling data, use an adaptive weight adjustment algorithm to optimize the weights of different physical fields in the hybrid physical constraint loss function of the surrounding rock of the large chamber to obtain an optimized loss function; S370: According to the optimized loss function, obtain a trained neural network framework for dynamic multi-physical field coupling; S380: Input the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber, and use the trained neural network framework for dynamic multi-physical field coupling to generate a dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber. The dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

[0102] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access storage unit (RAM) 621 and / or a cache storage unit 622, and may further include a read-only storage unit (ROM) 623.

[0103] The storage unit 620 may further include a program / utilities 624 having a set (at least one) of program modules 625. Such program modules 625 include, but are not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include the implementation of a network environment.

[0104] The bus 630 can represent one or more of several types of bus structures, including a memory unit bus or a memory unit controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of the various bus structures.

[0105] The electronic device 600 can also communicate with one or more external devices 670 (such as a keyboard, a pointing device, a Bluetooth device, etc.), can also communicate with one or more devices that enable a user to interact with the electronic device 600, and / or can communicate with any device that enables the electronic device 600 to communicate with one or more other computing devices (such as a router, a modem, etc.). Such communication can be carried out through the input / output (I / O) interface 650. Moreover, the electronic device 600 can also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through the network adapter 660. As shown in the figure, the network adapter 660 communicates with other modules of the electronic device 600 through the bus 630. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in combination with the electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0106] Through the description of the above embodiments, those skilled in the art can easily understand that the exemplary embodiments described herein can be implemented by software, or can be implemented by the way of software in combination with necessary hardware. Therefore, the technical solution according to the embodiments of the present invention can be embodied in the form of a software product, and the software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiments of the present invention.

[0107] In an exemplary embodiment of the present invention, a computer-readable storage medium is also provided, on which a program product capable of implementing the above method of the present invention is stored. In some possible embodiments, various aspects of the present invention can also be implemented in the form of a program product, which includes program code. When the program product runs on a terminal device, the program code is used to enable the terminal device to execute the steps according to various exemplary embodiments of the present invention described in the "Exemplary Method" section above of the present invention.

[0108] Reference Figure 7As shown, a program product 700 for implementing the above-mentioned dynamic prediction method for the stress characteristics of surrounding rocks in large chambers according to an embodiment of the present invention is described. It can be a portable compact disc read-only memory (CD-ROM) and includes program code, and can run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited to this. In the present invention, a readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or component.

[0109] The program product can adopt any combination of one or more readable storage media. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0110] The program code for performing the operations of the present invention can be written in any combination of one or more programming languages. The programming languages include object-oriented programming languages - such as Java, C++, etc., and also include conventional procedural programming languages - such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, executed as an independent software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, by using an Internet service provider to connect through the Internet).

[0111] In addition, the above-mentioned drawings are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present invention, rather than for limiting purposes. It is easy to understand that the processes shown in the above-mentioned drawings do not indicate or limit the chronological order of these processes. Additionally, it is also easy to understand that these processes can be executed, for example, synchronously or asynchronously in multiple modules.

[0112] Through the description of the above embodiments, those skilled in the art can easily understand that the exemplary embodiments described herein can be implemented by software or by a combination of software and necessary hardware. Therefore, the technical solution according to the embodiments of the present invention can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (such as a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present invention.

[0113] After considering the specification and practicing the invention disclosed herein, those skilled in the art will readily conceive of other embodiments of the present invention. The present invention is intended to cover any variations, uses, or adaptations of the present invention, which follow the general principles of the present invention and include common general knowledge or conventional technical means in the technical field not disclosed by the present invention. The specification and embodiments are only regarded as exemplary, and the true scope and spirit of the present invention are pointed out by the claims.

[0114] It should be understood that the present invention is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is only limited by the appended claims.

Claims

1. A training method for a neural network framework with dynamic multi-physical field coupling, characterized in that, Including: Obtain the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; Based on the layered time-series parameters, multi-source physical field parameters, and geological parameters, construct a hybrid physical constraint loss function for the surrounding rock of the large chamber; Combined with the hybrid physical constraint loss function of the surrounding rock of the large chamber, the residual term of the data fitting item, and the time-series continuity regularization term, establish a loss function for the neural network framework of dynamic multi-physical field coupling; Extract the layered time-series characteristics of the surrounding rock deformation, stress, seepage, and damage from the real-time data of the layered excavation of the surrounding rock of the large chamber; Based on the layered time-series characteristics, use a double-threshold dynamic sampling strategy to dynamically determine the sampling data; According to the sampling data, use an adaptive weight adjustment algorithm to optimize the weights of different physical fields in the hybrid physical constraint loss function of the surrounding rock of the large chamber to obtain an optimized loss function; According to the optimized loss function, obtain a trained neural network framework of dynamic multi-physical field coupling.

2. The training method of the neural network framework of dynamic multi-physical field coupling according to claim 1, characterized in that The layered timing parameters include: the current excavation layer number , the height of the excavation layer of the th layer, the span of the excavation layer of the th layer, the continuous excavation time of the excavation layer of the th layer, and the total number of excavated layers ; The multi-source physical field parameters include: the displacement field of the layer immediately preceding the layer being excavated, the stress field of the layer immediately preceding the layer being excavated, the seepage pressure field of the layer immediately preceding the layer being excavated, and the damage factor of the layer immediately preceding the layer being excavated. ; The geological parameters include: permeability coefficient , damage threshold, critical seepage pressure gradient .

3. The training method of the dynamic multi-physical-field coupled neural network framework according to claim 1, characterized in that The hybrid physical constraint loss function of the surrounding rock of the large chamber includes: Based on the elastoplastic constitutive relationship, establish a residual term of the stress constraint equation; Based on Darcy's law and the damage-seepage coupling relationship, construct a constraint residual term formed by the seepage field; Based on the damage evolution theory and the excavation time series, construct a constraint equation residual term; Fit the residual term of the stress constraint equation, the constraint residual term formed by the seepage field, and the constraint equation residual term to obtain the hybrid physical constraint loss function of the surrounding rock of the large chamber.

4. The training method of the dynamic multi-physics field coupled neural network framework according to claim 1, characterized in that Combined with the hybrid physical constraint loss function of the surrounding rock of the large chamber, the residual term of the data fitting item, and the time-series continuity regularization term, establishing a loss function for the neural network framework of dynamic multi-physical field coupling includes: Based on the predicted displacement values, measured displacement values, predicted damage values, and measured damage values of each layer in all the excavated layers, calculate the residual term of the data fitting item for all the excavated layers; Based on the continuous excavation time of each layer in all the excavated layers and the derivative result of the change rate of the predicted displacement value, calculate the time-series continuity regularization term for all the excavated layers; Use the first weight coefficient to balance the hybrid physical constraint loss function of the surrounding rock of the large chamber to obtain a first balance result; Use the second weight coefficient to balance the time-series continuity regularization term to obtain a second balance result; Sum and calculate the first balance result, the second balance result, and the residual term of the data fitting item to obtain a loss function for the neural network framework of dynamic multi-physical field coupling.

5. The training method of the dynamic multi - physical - field coupled neural network framework according to claim 1, characterized in that, The layered time-series characteristics include strain energy density and damage factor. Based on the layered time-series characteristics, Using the double-threshold dynamic sampling strategy to dynamically determine the sampling data includes: According to the strain energy density and the damage factor, calculate the probability of the regional sampling points; Based on the probability of the regional sampling points, determine the sampling priority; Dynamically adjust the sampling points based on the sampling priority; Collect the sampling data of the sampling points.

6. The training method of the neural network framework for dynamic multi-physics field coupling according to claim 5, characterized in that According to the strain energy density and the damage factor, calculating the probability of the regional sampling points includes: Calculate the exponential function value of the strain energy density and the first preset sensitivity coefficient at the sampling point; Calculate the Gaussian error function value of the damage factor and the second preset sensitivity coefficient at the th sampling point; Sum and calculate the product of the exponential function values and the Gaussian error function values of all the sampling points to obtain a sum value; According to the product of the exponential function value and the Gaussian error function value of the th sampling point and the sum value, calculate the probability of the th sampling point.

7. A dynamic prediction method for the stress characteristics of surrounding rock in a large chamber, characterized in that, including: Obtaining the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; Based on the layered time-series parameters, multi-source physical field parameters, and geological parameters, constructing a hybrid physical constraint loss function for the surrounding rock of the large chamber; Combining the hybrid physical constraint loss function of the surrounding rock of the large chamber, the residual term of the data fitting term, and the time-series continuity regularization term to establish a loss function for the neural network framework of dynamic multi-physical field coupling; Extracting the layered time-series characteristics of the surrounding rock deformation, stress, seepage, and damage in the real-time data of the layered excavation of the surrounding rock of the large chamber; Based on the layered time-series characteristics, using a dual-threshold dynamic sampling strategy to dynamically determine the sampling data; According to the sampling data, using an adaptive weight adjustment algorithm to optimize the weights of different physical fields in the hybrid physical constraint loss function of the surrounding rock of the large chamber to obtain an optimized loss function; According to the optimized loss function, obtaining a trained neural network framework of dynamic multi-physical field coupling; Inputting the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber, and using the trained neural network framework of dynamic multi-physical field coupling to generate a dynamic prediction result of the mechanical characteristics of the surrounding rock of the large chamber, where the dynamic prediction result of the mechanical characteristics of the surrounding rock of the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

8. The dynamic prediction method for the mechanical properties of surrounding rock in large chambers according to claim 7, characterized in that The neural network framework of dynamic multi-physical field coupling includes: An input module, a residual module, a time-series module, a feature fusion module, and an output module; Among them, the input module is used to input the layered time-series parameters, the multi-source physical field parameters, and the geological parameters. The residual module is built based on a residual network. Based on the layered time-series parameters, the multi-source physical field parameters, and the geological parameters, using the loss function of the neural network framework of dynamic multi-physical field coupling established by the residual term of the constraint equation formed based on the elastoplastic constitutive model, the constraint residual term formed by the seepage field based on Darcy's law, and the residual term of the constraint equation formed by damage evolution to generate the differences between the displacement field, the stress field, the seepage field, and the damage factor. The time-series module is used to generate the preliminary prediction rules of the displacement field, the stress field, the seepage field, and the damage factor with the development of time according to the layered time-series parameters, the multi-source physical field parameters, and the geological parameters. The feature fusion module introduces the residual term of the data fitting term and the time-series continuity regularization term, and is used to superimpose and fuse the preliminary prediction rules of the displacement field, the stress field, the seepage field, and the damage factor with the development of time output by the time-series module and the differences between the displacement field, the stress field, the seepage field, and the damage factor calculated by the residual module to obtain a fusion result. The output module is used to generate a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor based on the fusion result.

9. A dynamic prediction device for the stress characteristics of surrounding rock in a large chamber, characterized in that, including: A data acquisition module for obtaining the layered time-series parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber; A first function construction module for constructing a hybrid physical constraint loss function for the surrounding rock of the large chamber based on the layered time-series parameters, multi-source physical field parameters, and geological parameters; A second function construction module, configured to establish a loss function for a neural network framework with dynamic multi-physical field coupling by combining the loss function of the mixed physical constraints of the surrounding rock of the large chamber, the residual term of the data fitting term, and the temporal continuity regularization term; A feature extraction module, configured to extract the hierarchical temporal features of the surrounding rock deformation, stress, seepage, and damage in the real-time data of the layered excavation of the surrounding rock of the large chamber; A sampling data determination module, configured to dynamically determine sampling data based on the hierarchical temporal features by using a dual-threshold dynamic sampling strategy; A function optimization module, configured to optimize the weights of different physical fields in the loss function of the mixed physical constraints of the surrounding rock of the large chamber by using an adaptive weight adjustment algorithm according to the sampling data, and obtain an optimized loss function; A network determination module, configured to obtain a trained neural network framework with dynamic multi-physical field coupling according to the optimized loss function; A prediction module, configured to input the hierarchical temporal parameters, multi-source physical field parameters, and geological parameters of the surrounding rock of the large chamber, and use the trained neural network framework with dynamic multi-physical field coupling to generate a dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber, where the dynamic prediction result of the mechanical properties of the surrounding rock of the large chamber includes a predicted displacement field, a predicted stress field, a predicted seepage field, and a predicted damage factor.

10. An electronic device, characterized in that, Comprising: A processor; And A memory, on which computer-readable instructions are stored, and when the computer-readable instructions are executed by the processor, the dynamic prediction method for the mechanical properties of the surrounding rock of the large chamber as described in any one of claims 7-8 is implemented.

Images