A dynamic parameterized design method and system for a coal mine roadway support scheme

Abstract

The application discloses a kind of coal mine roadway supporting scheme dynamic parameterization design method and system, and design method includes the following steps: obtaining the multi-source heterogeneous data of coal mine lane, constructs the knowledge graph of coal mine roadway supporting field, forms original characteristic parameter set;Regression and classification double-path prediction framework is constructed, and the original characteristic parameter is carried out depth feature extraction, associated learning and candidate scheme generation, and the initial scheme of supporting parameter is output;By physical information neural network, the endogenous constraint is applied to the initial scheme of supporting parameter, while combining the rules of supporting specification information, the double verification is carried out to supporting scheme, and through the automatic adjustment of the parameter that does not satisfy constraint condition is carried out to the antagonistic correction mechanism;The supporting scheme is updated dynamically, and the parameter deviation caused by the discontinuity of geological exploration or environmental disturbance is corrected, and the adaptive optimization and iterative update of supporting scheme are realized by reinforcement learning strategy and incremental closed loop mechanism, and the final supporting scheme is obtained.

Description

Technical Field

[0001] This invention relates to intelligent design and safety management technology for coal mine engineering, and in particular to a dynamic parameterized design method and system for coal mine roadway support schemes. Background Technology

[0002] Coal mine roadways are a crucial component of the mine production system, and their stability directly impacts mine safety and production efficiency. However, accidents are closely related to roadway surrounding rock instability and support failure. Current coal mine roadway support design primarily relies on the engineering experience and manual calculations of technical personnel. The design process is highly dependent on the support knowledge and engineering experience of designers. A large amount of data resources scattered across exploration reports, design documents, regulations, and historical cases are not systematically utilized. This makes it difficult to achieve precise and personalized matching of support schemes under complex geological conditions, easily leading to problems such as parameter redundancy, conservative design, or localized failures, thus failing to meet the demands of modern coal mine safety and efficient construction.

[0003] With the development of information technology, big data, and artificial intelligence, intelligentization has provided new technological pathways for coal mine engineering. Existing research includes Wu et al. proposing a coal mine safety early warning model based on neural networks and fuzzy hierarchical analysis, and constructing a mobile intelligent mining platform to achieve comprehensive online monitoring of the mine system; Xia et al. proposed a "field-area-network" model and supporting software system to improve the intelligence level and monitoring and evaluation capabilities of coal mine gas extraction operations. Therefore, artificial intelligence technology has significant potential in coal mine safety management and production optimization. However, the above achievements mainly focus on the intelligentization of safety early warning, production scheduling, and single business links. Research on the intelligent generation and dynamic design of coal mine roadway support schemes is still in its initial stage. Existing related research largely remains at the level of empirical formulas or local parameter prediction. The support theory system is still immature, the generated parameters are difficult to form complete and executable support schemes, and the dynamic changes and randomness of the surrounding rock conditions during excavation are generally ignored, making it impossible to effectively learn from and inherit the wealth of engineering experience contained in historical support cases.

[0004] Furthermore, current regulations and design documents related to coal mine roadway support suffer from scattered parameter sources, inconsistent expression standards, and significant missing parameters. The lack of a systematic correlation mechanism between different data directly impacts the reliability and reusability of support schemes. Mature technical systems are still lacking for knowledge representation, correlation modeling, and inference to complete missing parameters for large-scale, multi-source, heterogeneous engineering data. Simultaneously, existing methods generally lack the ability to deeply integrate data-driven results with mechanical mechanisms and engineering specifications, making it difficult to balance physical rationality and engineering compliance in generated schemes, and even more difficult to dynamically adjust based on real-time monitoring information during construction.

[0005] In recent years, information technology and artificial intelligence have increasingly penetrated the engineering field, promoting the intelligent development of mines. However, existing research is mostly focused on local scenarios such as safety early warning, production scheduling, or single-parameter prediction, and has not yet formed a systematic intelligent design method for the entire process of coal mine roadway support schemes. The number of existing support-related studies is limited, and the theoretical system is still immature. Most are based on empirical formulas or simple regression models, and the generated parameters often exhibit discrete and fragmented characteristics, making it difficult to construct a complete and feasible roadway support scheme. At the same time, existing methods ignore the dynamic changes and random disturbances of roadway surrounding rock conditions during excavation, lacking the ability to model the "evolution during excavation" process. Once generated, support parameters tend to be static, making it difficult to adapt to the time-varying characteristics of surrounding rock mechanical behavior under complex geological conditions. Furthermore, a large amount of engineering experience data contained in historical roadway support cases has long been stored in heterogeneous forms such as text and tables, lacking a unified semantic expression and association mechanism. Existing technologies struggle to systematically mine, reason, and transfer this data, resulting in support design still heavily relying on human experience and making it difficult to achieve cross-scenario reuse and intelligent decision-making.

[0006] Therefore, there is an urgent need for an intelligent design method for the entire process of coal mine roadway support, capable of unified perception and fusion of multi-source heterogeneous engineering data, constructing a domain knowledge expression system, and realizing automatic reasoning and completion of support parameters. On this basis, combined with data-driven models and mechanical and regulatory constraints, a physically reasonable and engineering-compliant support scheme can be generated. Furthermore, it can be dynamically updated and adaptively evolved based on real-time monitoring data during roadway excavation, thereby realizing the transformation of coal mine roadway support schemes from static experience-based design to dynamic parameterized intelligent design, and improving the safety, reliability, and standardization of support design under complex geological conditions. Summary of the Invention

[0007] This solution addresses the problems and needs raised above by proposing a dynamic parameterized design method and system for coal mine roadway support schemes. Due to the adoption of the following technical features, it can achieve the above-mentioned technical objectives and bring about several other technical benefits.

[0008] One objective of this invention is to provide a dynamic parameterized design method for coal mine roadway support schemes, comprising the following steps: S10: Obtain multi-source heterogeneous data of coal mine roadways, perform multi-modal data extraction on the multi-source heterogeneous data, then perform deep fusion on the extracted data to construct a knowledge graph in the field of coal mine roadway support, and infer and complete the missing or incomplete engineering parameters to form an original feature parameter set. S20: Construct a dual-path prediction framework for regression and classification based on conditional diffusion model and heterogeneous graph transformation network, perform deep feature extraction, association learning and candidate scheme generation on the original feature parameters, and output an initial scheme of support parameters that meets the given geological conditions and engineering constraints. S30: Apply intrinsic constraints of surrounding rock mechanics and support structure stress to the initial support parameter scheme through a physical information neural network, and perform dual verification of the support scheme by combining the rules of support specification information. Automatically adjust parameters that do not meet the constraints through an adversarial correction mechanism. S40: Model the evolution of spatiotemporal effects during tunnel excavation, use real-time monitoring data during excavation as feedback increments to dynamically update the support scheme, correct parameter deviations caused by discontinuities in geological exploration or environmental disturbances, and achieve adaptive optimization and iterative updates of the support scheme through reinforcement learning strategies and incremental closed-loop mechanisms to obtain the final support scheme and realize dynamic parameterized design of support parameters.

[0009] Furthermore, the dynamic parameterized design method and system for coal mine roadway support schemes according to the present invention may also have the following technical features: In one example of the present invention, step S10, which involves multimodal data extraction from the multi-source heterogeneous data, specifically includes: For text data, the CasRel model is used for entity recognition and relation extraction. For table data, the BERT and pointer network models are used to extract structured parameters. For image data, the CNN-Transformer hybrid vision model is used to extract rock layer thickness, porosity and geological features.

[0010] In one example of the present invention, step S10 involves reasoning to complete missing or incomplete engineering parameters, specifically including: The knowledge graph completion mechanism, based on graph embedding models or rule-based reasoning models, performs correlation reasoning on key parameters such as surrounding rock lithology, roof and floor conditions, burial depth, stress environment, tunnel cross-sectional dimensions, and historical support methods. It also verifies these parameters using mechanical formulas and industry standards to complete missing engineering parameter information.

[0011] In one example of the present invention, step S20 specifically includes the following steps: S21: The conditional diffusion model generates diverse candidate solutions for support parameters under given geological conditions and engineering constraints: A forward diffusion process is constructed, vectorizing the surrounding rock geological conditions and engineering constraint information to form conditional vectors. The conditional vector denoising process is constrained by geological features and design specifications when generating support parameters. During model training, a loss function is constructed for optimization and iterative denoising is performed. The trained denoising network is used to correct the current state, gradually approximating the distribution of support parameters that conforms to the geological conditions. Multiple sets of candidate support parameter vectors are generated through multiple independent sampling processes to obtain a parameter solution set covering different design possibilities. The generated candidate solutions are then subjected to engineering constraint screening and correction to obtain a set of candidate support parameters that meet engineering feasibility. S22: The heterogeneous graph transformation network characterizes the complex relationships between different types of nodes, realizing cross-modal correlation modeling between lithological characteristics, structural parameters, and support methods. The heterogeneous graph transformation network constructs a graph structure containing multiple types of nodes and multiple relational edges to model cross-modal correlations between lithological characteristics, structural parameters, and support methods. It performs unified vectorization representation of the original data of different types of nodes, normalizes continuous variables, and transforms discrete variables into low-dimensional dense vectors through embedding mapping, thus forming a unified node feature representation. Based on this, relation-specific feature transformation matrices are introduced for different types of relationships, and information transfer and feature updates between nodes are realized through a neighborhood aggregation mechanism. Through multi-layer graph transformation operations, local and global structural information are continuously fused to obtain a high-dimensional semantic representation that simultaneously includes surrounding rock conditions, parameter configurations, and support strategies, which is then used as input to the subsequent regression and classification dual-path prediction framework. S23: In the dual-path prediction framework of regression and classification, the regression path is used to predict continuous parameters such as anchor length, spacing, anchor cable preload, and shotcrete thickness; the classification path is used to determine the support type, support level, and combination method, thereby realizing the overall generation of the support scheme. The regression path and the classification path are trained collaboratively through a multi-task joint learning mechanism. Their joint loss function is a weighted combination of regression loss and classification loss, which enables continuous parameter prediction and discrete decision discrimination to be optimized collaboratively in a unified feature space. Through the result fusion mechanism, the support type and level determined by the classification path are used as constraints to match and verify the continuous parameters output by the regression path, thereby generating a complete support scheme that meets engineering specifications and mechanical requirements.

[0012] In one example of the present invention, in step S22, the node representation update process satisfies the following formula: In the formula, For nodes In the The feature vector of the layer; It is a non-linear activation function; For relational types, A collection of relation types; For neighboring nodes; For the first A learnable weight matrix is ​​set separately for self-looping relationships within the layer; This is the normalization factor.

[0013] In one example of the present invention, step S30 specifically includes the following steps: S31: The geological parameters of the surrounding rock, the original or candidate support parameters, and the boundary condition information are uniformly encoded to form an input vector; then, the input data is nonlinearly transformed by a feature mapping module composed of a multi-layer fully connected network to obtain an implicit feature representation characterizing the mechanical response of the surrounding rock-support system. S32: The surrounding rock deformation control equation, the support structure force balance equation and boundary conditions are embedded into the network in the form of physical constraints. By constructing a physical residual function to participate in model training, the network output can simultaneously meet the data fitting requirements and the basic laws of engineering mechanics. The physical residual function is a weighted combination of the residuals of the surrounding rock balance equation, the constitutive relation residuals and the boundary condition residuals, and together with the data residual terms, it constitutes the total loss function. S33: The response output module calculates key response indicators such as the surrounding rock displacement field, stress field distribution, plastic zone range, and internal forces of the support structure. Based on the above results, the rationality of the input support parameters is evaluated. When the output results are detected to not meet mechanical constraints or exceed preset deformation and safety thresholds, a parameter correction mechanism is triggered to adjust key parameters such as anchor length, spacing, anchor cable preload, and shotcrete thickness. The corrected parameters are then re-input into the network for forward calculation and physical consistency verification until the constraints are met or the maximum number of iterations is reached. Simultaneously, during the inference phase, the numerical simulation results of the finite difference software FLAC3D are combined to compare and verify the output of the physical information neural network, and the current support parameters are locally optimized based on the deviation.

[0014] In one example of the present invention, in step S32, the expressions for the embedded surrounding rock deformation equation, the support structure force balance equation, and the boundary conditions are as follows: The surrounding rock under axisymmetric plane strain conditions satisfies the following static equilibrium equations: =0 In the formula, Radial stress, For tangential stress, Radial coordinates; Under small deformation conditions, the deformation of the surrounding rock satisfies the strain-displacement relationship equation: In the formula, Radial displacement; The constitutive relation of the surrounding rock satisfies the generalized Hooke's law in the elastic stage, and the following equations under axisymmetric plane strain conditions: )] In the formula, For elastic modulus, Poisson's ratio, For axial stress; During the plastic stage, the strength of the surrounding rock satisfies the Mohr-Coulomb yield criterion: In the formula, For cohesion, It is the internal friction angle; The axial force on the anchor bolts in the support structure satisfies a one-dimensional equilibrium equation: =0 In the formula, For the axial force of the anchor bolt, The load is applied to the surrounding rock. The shotcrete layer satisfies the circumferential force equilibrium relationship in the case of thin walls: In the formula, To support the reaction force, The radius of the alleyway, This refers to the thickness of the spray layer; Boundary conditions include internal boundary conditions at the tunnel wall and far-field stress boundary conditions: The internal boundary conditions at the tunnel wall are: The far-field stress boundary conditions are: In the formula, This refers to the stress in the original rock.

[0015] In one example of the present invention, in step S32, the expression for the total loss function is: in: For the total loss function, For data residuals, The residuals of the surrounding rock equilibrium equations, For constitutive relation residuals, For boundary condition residuals, , , , These are the corresponding weighting coefficients.

[0016] In one example of the present invention, step S40 specifically includes the following steps: S41: During the tunnel excavation process, the surrounding rock response is first dynamically predicted based on the time series modeling method. The construction process is discretized into a time series. Using the excavation step sequence as an index, the monitoring data of surrounding rock displacement, stress, convergence deformation, and support structure stress are constructed as multivariate time series inputs. Combined with geological conditions and initial support parameters, the evolution law of the surrounding rock response over time is modeled through the time series prediction model, thereby obtaining the prediction results of surrounding rock deformation and stability indicators at several future moments. The time series prediction model continuously updates the latest monitoring data through a sliding time window, so that the model parameters are continuously corrected with the construction process, and the real-time tracking and trend prediction of the surrounding rock status are completed. S42: After obtaining the predicted results of the surrounding rock response, the predicted surrounding rock state and the current support parameters are combined as the system state input reinforcement learning decision model. The state characterizes the stability and stress characteristics of the surrounding rock. The actions are to adjust the length, spacing, preload of anchor cables and thickness of shotcrete. The reward function comprehensively considers the surrounding rock deformation control effect, safety factor and support cost. Thus, the optimal parameter adjustment scheme is generated through reinforcement learning strategy, realizing the transformation of support design from static scheme to dynamic decision. S43: During the execution phase, the support parameter adjustment scheme output by reinforcement learning is applied to the current design, and the response of the surrounding rock is compared and analyzed in conjunction with real-time monitoring data. The deviation between the prediction results and the actual monitoring results is used as a feedback signal. On the one hand, it is used to update the parameters of the time series model to improve the subsequent prediction accuracy. On the other hand, it is used for iterative optimization of the reinforcement learning strategy. The decision performance is continuously improved through strategy updates or value function corrections. S44: The updated support parameters are re-input into the physical information neural network and numerical simulation module for mechanical consistency verification and stability analysis. When the test results do not meet the constraints, the reinforcement learning model is triggered to make a decision again.

[0017] Another objective of this invention is to provide a dynamic parameterized design system for coal mine roadway support schemes, comprising: The intelligent perception module is configured to acquire multi-source heterogeneous data from coal mine roadways, extract multi-modal data from the multi-source heterogeneous data, then perform deep fusion of the extracted data to construct a knowledge graph in the field of coal mine roadway support, and infer and complete missing or incomplete engineering parameters to form an original feature parameter set. The design intelligent module is configured to construct a regression and classification dual-path prediction framework based on a conditional diffusion model and a heterogeneous graph transformation network. It performs deep feature extraction, association learning, and candidate scheme generation on the original feature parameters and outputs an initial scheme of support parameters that meets the given geological conditions and engineering constraints. The physical verification intelligent module is configured to apply intrinsic constraints of surrounding rock mechanics and support structure stress to the initial support parameter scheme through a physical information neural network, and at the same time perform dual verification of the support scheme based on support specification information, and automatically adjust parameters that do not meet the constraints through an adversarial correction mechanism. The evolutionary intelligence module is configured to model the spatiotemporal effects during tunnel excavation, use real-time monitoring data during excavation as feedback increments to dynamically update the support scheme, correct parameter deviations caused by discontinuities in geological exploration or environmental disturbances, and achieve adaptive optimization and iterative updates of the support scheme through reinforcement learning strategies and incremental closed-loop mechanisms to obtain the final support scheme and realize the dynamic parameterization design of support parameters.

[0018] In one example of the present invention, the physical verification intelligent module includes: The physical information neural network unit is configured to embed engineering mechanics laws into the model training and reasoning process, using the embedded surrounding rock deformation equation, support structure force balance equation and boundary conditions as constraints, and to identify and correct parameter combinations that violate mechanical laws; at the same time, it is combined with the finite difference software FLAC3D for simulation to perform overall support system force and stability analysis. The rule verification unit is configured to perform discrete rule verification on key parameters such as anchor bolt length, spacing, anchor cable preload, shotcrete thickness, and support form matching relationship based on support specification information and national or industry coal mine roadway support design specifications. Parameters that do not comply with specifications or have insufficient safety factors are marked, corrected, or rolled back.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention revolves around the entire process of coal mine roadway support schemes, encompassing "data perception, scheme generation, mechanical verification, and dynamic control." It achieves unified perception and knowledge-based representation of multi-source heterogeneous engineering data, overcoming the problems of data fragmentation, parameter missingness, and difficulty in inheriting experience in existing technologies. Based on this, it introduces a conditional diffusion model and a heterogeneous graph transformation network to establish a dual-path prediction framework of regression and classification, overcoming the shortcomings of traditional models that can only output local parameters and are difficult to form complete schemes. Simultaneously, through a physical information neural network and a standard rule verification module, it embeds the surrounding rock mechanics mechanism and engineering specifications into the reasoning process, compensating for the lack of physical constraints and insufficient engineering feasibility in existing data-driven methods. Furthermore, by combining real-time monitoring data during excavation, an evolutionary intelligent agent performs online correction and adaptive updates of support parameters, solving the deficiency of existing support schemes in reflecting dynamic changes during roadway construction. This achieves the transformation of coal mine roadway support schemes from static experience-based design to dynamic parameterized intelligent design.

[0020] This invention presents a dynamic parameterized design method for coal mine roadway support schemes. This method enables full-process management of coal mine roadway support schemes, from multi-source heterogeneous data perception, knowledge graph completion, intelligent generation, and dual verification of physical and standard data, to dynamic control during construction. It can automatically generate support parameter schemes that conform to geological conditions and engineering specifications, and dynamically update and optimize them based on real-time monitoring data during excavation. This improves the accuracy, reliability, and construction safety of support design under complex geological conditions, and timely prevents safety risks such as surrounding rock instability and support structure failure. The generated roadway support schemes can be applied to various support-related scenarios, such as coal mine construction and tunnel construction. This invention reduces reliance on manual experience, significantly improving engineering efficiency, construction safety, and the operability of support schemes.

[0021] The preferred embodiments of the invention will be described in more detail below with reference to the accompanying drawings, so as to facilitate an understanding of the features and advantages of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. The drawings are merely illustrative of some embodiments of the present invention and are not intended to limit the scope of the present invention to all embodiments.

[0023] Figure 1 A flowchart of a dynamic parameterized design method for coal mine roadway support schemes according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a dynamic parameterized design system for coal mine roadway support schemes according to an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0025] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “comprising” or “including” mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as “connected” or “linked” are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as “upper,” “lower,” “left,” and “right” are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0026] According to a first aspect of the present invention, a dynamic parameterized design method for coal mine roadway support schemes is provided, such as... Figure 1 As shown, it includes the following steps: S10: Acquire multi-source heterogeneous data of coal mine roadways, extract multi-modal data from the multi-source heterogeneous data, and then perform deep fusion on the extracted data to construct a knowledge graph in the field of coal mine roadway support. Using a relational graph neural network model with a knowledge graph completion mechanism, and combining mechanical formulas and industry standards, infer and complete missing or incomplete engineering parameters to form a standardized, continuous, and physically reasonable set of original feature parameters. The multi-source heterogeneous data includes at least geological exploration data, historical support case data, engineering design documents, national standards and specifications, and real-time monitoring data during excavation. Specifically, this step aims to collect, extract, fuse, and preliminarily verify multi-source heterogeneous data related to coal mine roadways, providing a continuous, complete, and reasonable set of feature parameters for support scheme generation. This includes multi-modal data extraction, knowledge graph construction, and domain data completion. First, multi-source heterogeneous data extraction is performed: Text data: A joint named entity recognition model based on relation extraction is used to perform entity recognition and relation extraction on unstructured text such as exploration reports and design documents, extracting information such as lithology, surrounding rock grade, anchor bolt specifications, and support type. Tabular data: BERT and pointer network models are used to parse semi-structured tabular data, extracting key fields such as roof and floor names, rock names, and thicknesses. Image data: Convolutional neural networks combined with visual Transformers are used to extract spatial features and structural parameters from borehole columnar diagrams, support layout diagrams, and construction photos. Structured monitoring data: Real-time monitoring data from pressure gauges, displacement gauges, stress sensors, etc., are uniformly encoded and standardized. Then, the knowledge graph is constructed, and data completion is performed based on the knowledge graph. A triplet knowledge graph integrating two structures, "entity-relationship-entity" and "entity-attribute-attribute value," is built. Node and relation encoding is performed using a graph embedding model and a rule-based reasoning model. For missing parameters or data in detection blind spots, inference completion is performed using a relational graph convolutional network combined with graph reasoning techniques, along with mechanical formulas such as the Mohr-Coulomb strength criterion and industry standards (such as support safety factors). The completed parameters are then weighted using a self-attention mechanism to eliminate redundancy and output a high-dimensional feature tensor, ensuring data continuity, physical rationality, and regulatory compliance. The expression for the Mohr-Coulomb strength criterion is: In the formula, For shear stress, For cohesion, For effective stress, It is the internal friction angle.

[0027] S20: Construct a dual-path prediction framework for regression and classification based on conditional diffusion model and heterogeneous graph transformation network, perform deep feature extraction, association learning and candidate scheme generation on the original feature parameters, and output an initial scheme of support parameters that meets the given geological conditions and engineering constraints. Specifically, this step aims to generate accurate parameters and intelligently map the support scheme into constructable 2D and 3D drawings. This includes implementing dual-path prediction (regression and classification), candidate scheme generation, and intelligent drawing generation. For parameterized prediction of coal mine roadway support schemes, a dual-path system of regression and classification is established. Regression prediction uses a conditional diffusion model to generate continuous parameters, such as roof and floor names, roadway support methods, bolt types, cable types, and mesh types. Classification prediction uses a heterogeneous graph transformation network to analyze roadway topology and node characteristics (e.g., faults, soft rock, water-soaked areas) to determine the support level, support form, and combination method. After prediction, the conditional diffusion model iteratively generates multiple sets of candidate parameter combinations through forward noise addition and backward noise reduction. Pareto optimization is then performed based on safety factors, construction costs, and efficiency to output the optimal support parameter matrix. After the scheme is generated, the coal mine roadway support scheme is visualized and generated, and the support parameters are automatically mapped to the two-dimensional layout drawing and the three-dimensional support model. By using the linkage with the AutoCAD API, the construction drawings can be quickly exported, detailed drawings can be generated and modified drawings can be updated. The drawings can be updated in real time after the design is adjusted, which can assist in construction and approval.

[0028] S30: The initial support parameter scheme is subjected to intrinsic constraints of surrounding rock mechanics and support structure stress through a physical information neural network. At the same time, the support scheme is double-verified from both the physical mechanism and industry standard levels by combining the rules of support specification information. The parameters that do not meet the constraints are automatically adjusted through an adversarial correction mechanism. The physical verification agent further uses the finite difference software FLAC3D to simulate and perform stress analysis, deformation prediction and stability assessment on the combined support scheme of anchor bolts, supports and shotcrete, etc., and outputs a physically reasonable, engineering compliant and safe and reliable support scheme. Specifically, this step aims to ensure that the generated support scheme is safe and reliable at both the physical and mechanical levels and industry standards. Its main functions include: constructing a physical information neural network, performing support simulation, and verifying against standards. The physical information neural network embeds the surrounding rock deformation equation, the force balance equation of the support structure, and boundary conditions. Based on this, the scheme undergoes mechanical constraint verification, such as stress divergence balance testing and displacement-strain consistency testing. Parameter combinations that violate mechanical laws generate high loss values, which are automatically corrected through backpropagation. Simultaneously, the coal mine roadway support scheme is simulated using the finite difference software FLAC3D to evaluate the mechanical response of the support system under different surrounding rock conditions, and to check the safety factor, anchor bolt stress, and plastic ring coverage to ensure compliance with design specifications. Finally, a database of national and industry support standards is built into the physical verification agent to perform discrete rule verification on anchor bolt length, spacing, preload, shotcrete thickness, and support form matching. Parameters that do not comply with standards or have insufficient safety factors are automatically marked, corrected, or rolled back.

[0029] S40: The spatiotemporal effects during tunnel excavation are modeled. Real-time monitoring data during excavation is used as feedback increments to dynamically update the support scheme, correcting parameter deviations caused by discontinuities in geological exploration or environmental disturbances. Adaptive optimization and iterative updates of the support scheme are achieved through reinforcement learning strategies and incremental closed-loop mechanisms to obtain the final support scheme, realizing dynamic parameterized design of support parameters. The evolutionary agent can accumulate learning from historical monitoring data, improving the model's generalization ability and prediction accuracy under complex geological conditions.

[0030] This step aims to achieve adaptive adjustment of the support scheme through monitoring data during excavation, enhancing the robustness and safety of the scheme during actual excavation. This part mainly includes: spatiotemporal response modeling, setting up a real-time incremental correction mechanism, and closed-loop optimization and long-term evolution. First, a spatiotemporal graph convolutional network is used to extract the displacement, stress, and deformation rate characteristics of each measuring point in the roadway over time, predicting the trend of surrounding rock deformation during construction. Second, a real-time incremental correction mechanism is set up to compare the actual monitored values ​​with the model's predicted values ​​and calculate the deviation increment; if the deviation exceeds a critical threshold (e.g., the roof delamination requirement should meet the following formula), the design agent is triggered to iterate and generate parameters again, adjusting the anchor density, length, or support type in real time. Then, through continuous training on historical case libraries and excavation data, the self-optimization strategy of dynamically updating support parameters is achieved, enabling the system to automatically recommend the optimal support scheme based on the construction stage, surrounding rock conditions, and construction speed. Finally, closed-loop optimization and long-term evolution are performed. Each dynamic correction data is stored in a case library for online training of the sensing and design agents, realizing the transformation of the support scheme from static design to dynamic, parameterized, and intelligent closed-loop evolution. Among these, the roof delamination requirement should satisfy the following formula: In the formula: The extraction height is in mm. Allowable delamination value, in mm.

[0031] This design methodology revolves around the entire process of coal mine roadway support schemes, encompassing "data perception, scheme generation, mechanical verification, and dynamic control." It achieves unified perception and knowledge-based representation of multi-source heterogeneous engineering data, overcoming the problems of data fragmentation, parameter missingness, and difficulty in inheriting experience in existing technologies. Building upon this, it introduces a conditional diffusion model and a heterogeneous graph transformation network to establish a dual-path prediction framework of regression and classification, overcoming the shortcomings of traditional models that can only output local parameters and are unable to form complete schemes. Simultaneously, through a physical information neural network and a standard rule verification module, it embeds the surrounding rock mechanics mechanism and engineering specifications into the reasoning process, compensating for the lack of physical constraints and insufficient engineering feasibility in existing data-driven methods. Furthermore, by combining real-time monitoring data during excavation, an evolutionary intelligent agent performs online correction and adaptive updates of support parameters, addressing the deficiency of existing support schemes in reflecting dynamic changes during roadway construction. This achieves a transformation of coal mine roadway support schemes from static experience-based design to dynamic, parameterized, and intelligent design.

[0032] This design method enables full-process management of coal mine roadway support schemes, from multi-source heterogeneous data perception, knowledge graph completion, intelligent generation, and dual verification of physical and standard data, to dynamic control during construction. It can automatically generate support parameter schemes that conform to geological conditions and engineering specifications, and dynamically update and optimize them based on real-time monitoring data during excavation. This improves the accuracy, reliability, and construction safety of support design under complex geological conditions, and timely prevents safety risks such as surrounding rock instability and support structure failure. The generated roadway support schemes can be applied to various support-related scenarios, such as coal mine construction and tunnel construction. This invention reduces reliance on manual experience and significantly improves engineering efficiency, construction safety, and the operability of support schemes.

[0033] In one example of the present invention, step S10, which involves multimodal data extraction from the multi-source heterogeneous data, specifically includes: For text data, the CasRel model is used for entity recognition and relation extraction. For table data, the BERT and pointer network models are used to extract structured parameters. For image data, the CNN-Transformer hybrid vision model is used to extract rock layer thickness, porosity and geological features.

[0034] In one example of the present invention, step S10 involves reasoning to complete missing or incomplete engineering parameters, specifically including: The knowledge graph completion mechanism, based on graph embedding models or rule-based reasoning models, performs correlation reasoning on key parameters such as surrounding rock lithology, roof and floor conditions, burial depth, stress environment, tunnel cross-sectional dimensions, and historical support methods. It also verifies these parameters using mechanical formulas and industry standards to complete missing engineering parameter information.

[0035] In one example of the present invention, step S20 specifically includes the following steps: S21: The conditional diffusion model generates diverse candidate solutions for support parameters under given geological conditions and engineering constraints. It standardizes the support parameter vectors in historical engineering data and constructs a forward diffusion process, gradually superimposing Gaussian noise onto the original support parameters to obtain noise samples at different time steps. Next, it vectorizes the surrounding rock geological conditions and engineering constraint information to form conditional vectors, which are then input into a denoising network in the form of conditional embedding. These vectors are fused with the feature representations of the noise samples through a conditional modulation mechanism, ensuring that the denoising process is constrained by geological features and design specifications when generating support parameters. Then, during the model training phase, it calculates the error between the denoising network's predictions and the actual noise, and uses this error to construct a loss function for optimization, learning the mapping relationship for recovering support parameters under different noise levels and given conditions. In the inference phase, an initial noise vector is randomly sampled from a standard Gaussian distribution. Under the constraint of a given condition vector, denoising is iterated step by step according to a preset time step. In each step, the trained denoising network is used to correct the current state, making it gradually approach the distribution of support parameters that conforms to the geological conditions. Furthermore, multiple sets of candidate support parameter vectors are generated through multiple independent sampling processes. Each set of candidate solutions corresponds to a potential support design scheme, thereby obtaining a parameter solution set covering different design possibilities. Finally, the generated candidate solutions are subjected to engineering constraint screening and correction processing, including verifying the value range of anchor bolt length, spacing, anchor cable preload, and shotcrete thickness, checking parameter matching relationships, and conducting a preliminary safety factor assessment. Candidate solutions that do not meet the constraints are eliminated or adjusted, thereby obtaining a set of candidate support parameters that meet engineering feasibility. S22: Heterogeneous graph transformation networks characterize the complex relationships between different types of nodes, enabling cross-modal correlation modeling between lithological characteristics, structural parameters, and support methods. By constructing a graph structure containing multiple types of nodes and multiple relational edges, heterogeneous graph transformation networks model cross-modal correlations between lithological characteristics, structural parameters, and support methods. First, surrounding rock mechanical parameters, support structure design parameters, and support form and grade information are abstracted into different types of nodes. Then, corresponding edge connections are established based on their engineering semantic relationships, including the influence of lithological characteristics on structural parameters, the constraint relationship of support methods on parameter configuration, and the relationships between different parameters. The coupling relationship is further explored. The original data of different types of nodes are then uniformly vectorized, with continuous variables normalized and discrete variables transformed into low-dimensional dense vectors through embedding mapping, thus forming a unified node feature representation. Based on this, relationship-specific feature transformation matrices are introduced for different types of relationships, and information transfer and feature updates between nodes are achieved through a neighborhood aggregation mechanism. Furthermore, through multi-layer graph transformation operations, local and global structural information are continuously fused to obtain a high-dimensional semantic representation that simultaneously includes surrounding rock conditions, parameter configurations, and support strategies. This high-dimensional semantic representation is then used as input to the subsequent regression and classification dual-path prediction framework. S23: In the dual-path prediction framework of regression and classification, the regression path is used to predict continuous parameters such as anchor bolt length, spacing, anchor cable preload, and shotcrete thickness; the classification path is used to determine the support type, support level, and combination method, thereby achieving the overall generation of the support scheme. In this dual-path prediction framework, the regression and classification paths are modeled in parallel based on the high-dimensional semantic features output by the shared feature extraction module. The regression path is used to achieve precise prediction of continuous support parameters, while the classification path is used to achieve discrete decision-making for the support scheme. Specifically, the regression path includes a feature mapping submodule, a parameter constraint submodule, and a continuous output submodule. First, the shared features are input into a multi-layer fully connected network for nonlinear mapping. Then, the network output is limited in range and physically constrained through a parameter constraint mechanism to ensure that parameters such as anchor bolt length, spacing, anchor cable preload, and shotcrete thickness meet engineering experience and specification requirements. Finally, a continuous parameter vector is generated through a linear output layer, and the prediction results are optimized using a mean squared error loss function during training to obtain high-precision support parameter prediction results. Meanwhile, the classification path includes a feature discrimination submodule. The system comprises a main module, a multi-task classification submodule, and a probability output submodule. Shared features are input into the classification network for discriminative mapping. Based on the decision structure of the support scheme, it is divided into support form discrimination units, support level discrimination units, and support combination method discrimination units. Each unit outputs the probability distribution of its corresponding category, which is processed through a normalization function to obtain the final classification result. Cross-entropy loss is used for optimization during training. Furthermore, the regression path and classification path are collaboratively trained through a multi-task joint learning mechanism. Their joint loss function is a weighted combination of regression loss and classification loss, enabling continuous parameter prediction and discrete decision discrimination to be collaboratively optimized in a unified feature space. A result fusion mechanism uses the support form and level determined by the classification path as constraints to match and verify the continuous parameters output by the regression path, thereby generating a complete support scheme that meets engineering specifications and mechanical requirements.

[0036] In the multi-task joint learning framework, the joint loss function of the regression path and the classification path can be expressed as: in: For the joint loss function, To recover the losses from the mission, To classify task losses, , These are the weighting coefficients.

[0037] In one example of the present invention, step S20 further includes: automatically mapping the generated support parameters to the two-dimensional layout drawing and the three-dimensional support model, and realizing the automatic generation and real-time updating of construction drawings, detailed drawings and modified drawings through the AutoCAD API linkage, so as to realize the visualization output of the support scheme and construction guidance.

[0038] In one example of the present invention, in step S22, the node representation update process satisfies the following formula: In the formula, For nodes In the The feature vector of the layer; It is a non-linear activation function; For relational types, A collection of relation types; For neighboring nodes; For the first A learnable weight matrix is ​​set separately for self-looping relationships within the layer; This is the normalization factor.

[0039] In one example of the present invention, step S30 specifically includes the following steps: S31: The geological parameters of the surrounding rock, the original or candidate support parameters, and the boundary condition information are uniformly encoded to form an input vector; then, the input data is nonlinearly transformed by a feature mapping module composed of a multi-layer fully connected network to obtain an implicit feature representation characterizing the mechanical response of the surrounding rock-support system. S32: The surrounding rock deformation control equation, the support structure force balance equation and boundary conditions are embedded into the network in the form of physical constraints. By constructing a physical residual function to participate in model training, the network output can simultaneously meet the data fitting requirements and the basic laws of engineering mechanics. The physical residual function is a weighted combination of the residuals of the surrounding rock balance equation, the constitutive relation residuals and the boundary condition residuals, and together with the data residual terms, it constitutes the total loss function. S33: The response output module calculates key response indicators such as the surrounding rock displacement field, stress field distribution, plastic zone range, and internal forces of the support structure. Based on the above results, the rationality of the input support parameters is evaluated. When the output results are detected to not meet mechanical constraints or exceed preset deformation and safety thresholds, a parameter correction mechanism is triggered to adjust key parameters such as anchor length, spacing, anchor cable preload, and shotcrete thickness. The corrected parameters are then re-input into the network for forward calculation and physical consistency verification until the constraints are met or the maximum number of iterations is reached. Simultaneously, during the inference phase, the numerical simulation results of the finite difference software FLAC3D are combined to compare and verify the output of the physical information neural network, and the current support parameters are locally optimized based on the deviation.

[0040] The overall structure of the physical information neural network includes encoded input, feature mapping, physical constraint embedding, and response output. The physical information neural network uses the embedded surrounding rock deformation equation, the support structure force balance equation, and boundary conditions as constraints to embed engineering mechanics laws into the model training and reasoning process, and identifies and corrects parameter combinations that violate mechanical laws. At the same time, it combines the finite difference software FLAC3D for simulation to perform stress and stability analysis of the overall support system. The rule verification module based on support specification information uses national or industry coal mine roadway support design specifications as a basis to perform discrete rule verification on key parameters such as anchor bolt length, spacing, anchor cable preload, shotcrete thickness, and support form matching relationship, and marks, corrects, or rolls back parameters that do not conform to the specifications or have insufficient safety factors.

[0041] In one example of the present invention, in step S32, the expressions for the embedded surrounding rock deformation equation, the support structure force balance equation, and the boundary conditions are as follows: The surrounding rock under axisymmetric plane strain conditions satisfies the following static equilibrium equations: =0 In the formula, Radial stress, For tangential stress, Radial coordinates; Under small deformation conditions, the deformation of the surrounding rock satisfies the strain-displacement relationship equation: In the formula, Radial displacement; The constitutive relation of the surrounding rock satisfies the generalized Hooke's law in the elastic stage, and the following equations under axisymmetric plane strain conditions: )] In the formula, For elastic modulus, Poisson's ratio, For axial stress; During the plastic stage, the strength of the surrounding rock satisfies the Mohr-Coulomb yield criterion: In the formula, For cohesion, It is the internal friction angle; The axial force on the anchor bolts in the support structure satisfies a one-dimensional equilibrium equation: =0 In the formula, For the axial force of the anchor bolt, The load is applied to the surrounding rock. The shotcrete layer satisfies the circumferential force equilibrium relationship in the case of thin walls: In the formula, To support the reaction force, The radius of the alleyway, This refers to the thickness of the spray layer; Boundary conditions include internal boundary conditions at the tunnel wall and far-field stress boundary conditions: The internal boundary conditions at the tunnel wall are: The far-field stress boundary conditions are: In the formula, This refers to the stress in the original rock.

[0042] In one example of the present invention, step S40 specifically includes the following steps: S41: During the tunnel excavation process, the surrounding rock response is first dynamically predicted based on the time series modeling method. The construction process is discretized into a time series. Using the excavation step sequence as an index, the monitoring data of surrounding rock displacement, stress, convergence deformation, and support structure stress are constructed as multivariate time series inputs. Combined with geological conditions and initial support parameters, the evolution law of the surrounding rock response over time is modeled through the time series prediction model, thereby obtaining the prediction results of surrounding rock deformation and stability indicators at several future moments. The time series prediction model continuously updates the latest monitoring data through a sliding time window, so that the model parameters are continuously corrected with the construction process, and the real-time tracking and trend prediction of the surrounding rock status are completed. S42: After obtaining the predicted results of the surrounding rock response, the predicted surrounding rock state and the current support parameters are combined as the system state input reinforcement learning decision model. The state characterizes the stability and stress characteristics of the surrounding rock. The actions are to adjust the length, spacing, preload of anchor cables and thickness of shotcrete. The reward function comprehensively considers the surrounding rock deformation control effect, safety factor and support cost. Thus, the optimal parameter adjustment scheme is generated through reinforcement learning strategy, realizing the transformation of support design from static scheme to dynamic decision. S43: During the execution phase, the support parameter adjustment scheme output by reinforcement learning is applied to the current design, and the response of the surrounding rock is compared and analyzed in conjunction with real-time monitoring data. The deviation between the prediction results and the actual monitoring results is used as a feedback signal. On the one hand, it is used to update the parameters of the time series model to improve the subsequent prediction accuracy. On the other hand, it is used for iterative optimization of the reinforcement learning strategy. The decision performance is continuously improved through strategy updates or value function corrections. S44: The updated support parameters are re-input into the physical information neural network and numerical simulation module for mechanical consistency verification and stability analysis. When the test results do not meet the constraints, the reinforcement learning model is triggered to make a decision again.

[0043] Based on time series modeling methods or reinforcement learning strategies, the response of the surrounding rock and the support effect during the tunnel excavation process are continuously tracked, and the model parameters are updated online according to real-time monitoring data to achieve adaptive evolution of the support scheme. At the same time, incremental learning is carried out using a historical case library to achieve self-optimization of the perception, design and physical verification agent.

[0044] According to a second aspect of the present invention, a dynamic parameterized design system for coal mine roadway support schemes, such as... Figure 2 As shown, it includes: The intelligent perception module is configured to acquire multi-source heterogeneous data from coal mine roadways, extract multimodal data from the multi-source heterogeneous data, and then perform deep fusion on the extracted data to construct a knowledge graph in the field of coal mine roadway support. A relational graph neural network model with a knowledge graph completion mechanism is then used, combined with mechanical formulas and industry standards, to infer and complete missing or incomplete engineering parameters, forming a standardized, continuous, and physically reasonable set of original feature parameters. The multi-source heterogeneous data includes at least geological exploration data, historical support case data, engineering design documents, national standards and specifications, and real-time monitoring data during excavation. The design intelligent module is configured to construct a regression and classification dual-path prediction framework based on a conditional diffusion model and a heterogeneous graph transformation network. It performs deep feature extraction, association learning, and candidate scheme generation on the original feature parameters and outputs an initial scheme of support parameters that meets the given geological conditions and engineering constraints. The physical verification intelligent module is configured to apply intrinsic constraints of surrounding rock mechanics and support structure stress to the initial support parameter scheme through a physical information neural network. At the same time, it combines a rule verification module based on support specification information to perform dual verification of the support scheme from both physical mechanism and industry specification levels. It also automatically adjusts parameters that do not meet the constraints through an adversarial correction mechanism. The physical verification intelligent agent further uses the finite difference software FLAC3D to perform simulation, conduct stress analysis, deformation prediction and stability assessment of the combined support scheme of anchor bolts, supports and shotcrete, etc., and output a physically reasonable, engineering compliant and safe and reliable support scheme. The evolutionary intelligence module is configured to model the spatiotemporal evolution of effects during tunnel excavation. It uses real-time monitoring data as feedback increments to dynamically update the support scheme, correcting parameter deviations caused by discontinuities in geological exploration or environmental disturbances. Through reinforcement learning strategies and an incremental closed-loop mechanism, it achieves adaptive optimization and iterative updates of the support scheme, obtaining the final support scheme and realizing dynamic parameterized design of support parameters. The evolutionary intelligence agent can accumulate and learn from historical monitoring data, improving the model's generalization ability and prediction accuracy under complex geological conditions.

[0045] This design system revolves around the entire process of coal mine roadway support schemes, encompassing "data perception, scheme generation, mechanical verification, and dynamic control." It achieves unified perception and knowledge-based representation of multi-source heterogeneous engineering data, overcoming the problems of data fragmentation, parameter missingness, and difficulty in inheriting experience in existing technologies. Based on this, it introduces a conditional diffusion model and a heterogeneous graph transformation network to establish a dual-path prediction framework of regression and classification, overcoming the shortcomings of traditional models that can only output local parameters and are unable to form complete schemes. Simultaneously, through a physical information neural network and a standard rule verification module, it embeds the surrounding rock mechanics mechanism and engineering specifications into the reasoning process, compensating for the lack of physical constraints and insufficient engineering feasibility in existing data-driven methods. Furthermore, by combining real-time monitoring data during excavation, an evolutionary intelligent agent performs online correction and adaptive updates of support parameters, addressing the deficiency of existing support schemes in reflecting dynamic changes during roadway construction. This achieves the transformation of coal mine roadway support schemes from static experience-based design to dynamic, parameterized, and intelligent design.

[0046] This design system enables full-process management of coal mine roadway support schemes, from multi-source heterogeneous data perception, knowledge graph completion, intelligent generation, and dual verification of physical and standard data, to dynamic control during construction. It can automatically generate support parameter schemes that conform to geological conditions and engineering specifications, and dynamically update and optimize them based on real-time monitoring data during excavation. This improves the accuracy, reliability, and construction safety of support design under complex geological conditions, and timely prevents safety risks such as surrounding rock instability and support structure failure. The generated roadway support schemes can be applied to various support-related scenarios, such as coal mine construction and tunnel construction. This invention reduces reliance on manual experience and significantly improves engineering efficiency, construction safety, and the operability of support schemes.

[0047] In one example of the present invention, the physical verification intelligent module includes: The physical information neural network unit is configured to embed engineering mechanics laws into the model training and reasoning process, using the embedded surrounding rock deformation equation, support structure force balance equation and boundary conditions as constraints, and to identify and correct parameter combinations that violate mechanical laws; at the same time, it is combined with the finite difference software FLAC3D for simulation to perform overall support system force and stability analysis. The rule verification unit is configured to perform discrete rule verification on key parameters such as anchor bolt length, spacing, anchor cable preload, shotcrete thickness, and support form matching relationship based on support specification information and national or industry coal mine roadway support design specifications. Parameters that do not comply with the specifications or have insufficient safety factors are marked, corrected, or rolled back.

[0048] It should be noted that the dynamic parameterized design system for coal mine roadway support scheme of the present invention can also perform any of the processing described in the previously described dynamic parameterized design method for coal mine roadway support scheme, and the specific details are not repeated here.

[0049] The foregoing description, with reference to preferred embodiments, details an exemplary implementation of the dynamic parameterized design method and system for coal mine roadway support proposed by the present invention. However, those skilled in the art will understand that various modifications and alterations can be made to the above specific embodiments without departing from the concept of the present invention, and various combinations can be made to the various technical features and structures proposed by the present invention without exceeding the protection scope of the present invention, which is determined by the appended claims.

Claims

1. A dynamic parameterized design method for coal mine roadway support schemes, characterized in that, Includes the following steps: S10: Obtain multi-source heterogeneous data of coal mine roadways, perform multi-modal data extraction on the multi-source heterogeneous data, then perform deep fusion on the extracted data to construct a knowledge graph in the field of coal mine roadway support, and infer and complete the missing or incomplete engineering parameters to form an original feature parameter set. S20: Construct a dual-path prediction framework for regression and classification based on conditional diffusion model and heterogeneous graph transformation network, perform deep feature extraction, association learning and candidate scheme generation on the original feature parameters, and output an initial scheme of support parameters that meets the given geological conditions and engineering constraints. S30: Apply intrinsic constraints of surrounding rock mechanics and support structure stress to the initial support parameter scheme through a physical information neural network, and perform dual verification of the support scheme by combining the rules of support specification information. Automatically adjust parameters that do not meet the constraints through an adversarial correction mechanism. S40: Model the spatiotemporal effect evolution during the tunnel excavation process, use real-time monitoring data during excavation as feedback increments to dynamically update the support scheme, correct parameter deviations caused by discontinuities in geological exploration or environmental disturbances, and achieve adaptive optimization and iterative updates of the support scheme through reinforcement learning strategies and incremental closed-loop mechanisms to obtain the final support scheme and realize the dynamic parameterization design of support parameters.

2. The dynamic parameterized design method for coal mine roadway support schemes according to claim 1, characterized in that, In step S10, the multimodal data extraction of the multi-source heterogeneous data specifically includes: For text data, the CasRel model is used for entity recognition and relation extraction. For table data, the BERT and pointer network models are used to extract structured parameters. For image data, the CNN-Transformer hybrid vision model is used to extract rock layer thickness, porosity and geological features.

3. The dynamic parameterized design method for coal mine roadway support schemes according to claim 1, characterized in that, In step S10, missing or incomplete engineering parameters are inferred and completed, specifically including: The knowledge graph completion mechanism, based on graph embedding models or rule-based reasoning models, performs correlation reasoning on key parameters such as surrounding rock lithology, roof and floor conditions, burial depth, stress environment, tunnel cross-sectional dimensions, and historical support methods. It also verifies these parameters using mechanical formulas and industry standards to complete missing engineering parameter information.

4. The dynamic parameterized design method for coal mine roadway support schemes according to claim 1, characterized in that, Step S20 specifically includes the following steps: S21: The conditional diffusion model generates diverse candidate solutions for support parameters under given geological conditions and engineering constraints: A forward diffusion process is constructed, vectorizing the surrounding rock geological conditions and engineering constraint information to form conditional vectors. The conditional vector denoising process is constrained by geological features and design specifications when generating support parameters. During model training, a loss function is constructed for optimization and iterative denoising is performed. The trained denoising network is used to correct the current state, gradually approximating the distribution of support parameters that conforms to the geological conditions. Multiple sets of candidate support parameter vectors are generated through multiple independent sampling processes to obtain a parameter solution set covering different design possibilities. The generated candidate solutions are then subjected to engineering constraint screening and correction to obtain a set of candidate support parameters that meet engineering feasibility. S22: The heterogeneous graph transformation network characterizes the complex relationships between different types of nodes, realizing cross-modal correlation modeling between lithological characteristics, structural parameters, and support methods. The heterogeneous graph transformation network constructs a graph structure containing multiple types of nodes and multiple relational edges to model cross-modal correlations between lithological characteristics, structural parameters, and support methods. It performs unified vectorization representation of the original data of different types of nodes, normalizes continuous variables, and transforms discrete variables into low-dimensional dense vectors through embedding mapping, thus forming a unified node feature representation. Based on this, relation-specific feature transformation matrices are introduced for different types of relationships, and information transfer and feature updates between nodes are realized through a neighborhood aggregation mechanism. Through multi-layer graph transformation operations, local and global structural information are continuously fused to obtain a high-dimensional semantic representation that simultaneously includes surrounding rock conditions, parameter configurations, and support strategies, which is then used as input to the subsequent regression and classification dual-path prediction framework. S23: In the dual-path prediction framework of regression and classification, the regression path is used to predict continuous parameters such as anchor length, spacing, anchor cable preload, and shotcrete thickness; the classification path is used to determine the support type, support level, and combination method, thereby realizing the overall generation of the support scheme. The regression path and the classification path are trained collaboratively through a multi-task joint learning mechanism. Their joint loss function is a weighted combination of regression loss and classification loss, which enables continuous parameter prediction and discrete decision discrimination to be optimized collaboratively in a unified feature space. Through the result fusion mechanism, the support type and level determined by the classification path are used as constraints to match and verify the continuous parameters output by the regression path, thereby generating a complete support scheme that meets engineering specifications and mechanical requirements.

5. The dynamic parameterized design method for coal mine roadway support schemes according to claim 4, characterized in that, In step S22, the node representation update process satisfies the following formula: In the formula, For nodes In the The feature vector of the layer; It is a non-linear activation function; For relation types, A collection of relation types; For neighboring nodes; For the first A learnable weight matrix is ​​set separately for self-looping relationships within the layer; This is the normalization factor.

6. The dynamic parameterized design method for coal mine roadway support schemes according to claim 1, characterized in that, Step S30 specifically includes the following steps: S31: The geological parameters of the surrounding rock, the original or candidate support parameters, and the boundary condition information are uniformly encoded to form an input vector; then, the input data is nonlinearly transformed by a feature mapping module composed of a multi-layer fully connected network to obtain an implicit feature representation characterizing the mechanical response of the surrounding rock-support system. S32: The surrounding rock deformation control equation, the support structure force balance equation and boundary conditions are embedded into the network in the form of physical constraints. By constructing a physical residual function to participate in model training, the network output can simultaneously meet the data fitting requirements and the basic laws of engineering mechanics. The physical residual function is a weighted combination of the residuals of the surrounding rock balance equation, the constitutive relation residuals and the boundary condition residuals, and together with the data residual terms, it constitutes the total loss function. S33: The response output module calculates key response indicators such as the surrounding rock displacement field, stress field distribution, plastic zone range, and internal forces of the support structure. Based on the above results, the rationality of the input support parameters is evaluated. When the output results are detected to not meet mechanical constraints or exceed preset deformation and safety thresholds, a parameter correction mechanism is triggered to adjust key parameters such as anchor length, spacing, anchor cable preload, and shotcrete thickness. The corrected parameters are then re-input into the network for forward calculation and physical consistency verification until the constraints are met or the maximum number of iterations is reached. Simultaneously, during the inference phase, the numerical simulation results of the finite difference software FLAC3D are combined to compare and verify the output of the physical information neural network, and the current support parameters are locally optimized based on the deviation.

7. The dynamic parameterized design method for coal mine roadway support schemes according to claim 6, characterized in that, In step S32, the expressions for the embedded surrounding rock deformation equation, the force balance equation of the support structure, and the boundary conditions are as follows: The surrounding rock under axisymmetric plane strain conditions satisfies the following static equilibrium equations: =0 In the formula, Radial stress, For tangential stress, Radial coordinates; Under small deformation conditions, the deformation of the surrounding rock satisfies the strain-displacement relationship equation: In the formula, Radial displacement; The constitutive relation of the surrounding rock satisfies the generalized Hooke's law in the elastic stage, and the following equations under axisymmetric plane strain conditions: )] In the formula, For elastic modulus, Poisson's ratio, For axial stress; During the plastic stage, the strength of the surrounding rock satisfies the Mohr-Coulomb yield criterion: In the formula, For cohesion, It is the internal friction angle; The axial force on the anchor bolts in the support structure satisfies a one-dimensional equilibrium equation: =0 In the formula, For the axial force of the anchor bolt, The load is applied to the surrounding rock. The shotcrete layer satisfies the circumferential force equilibrium relationship in the case of thin walls: In the formula, To support the reaction force, The radius of the alleyway, This refers to the thickness of the spray layer; Boundary conditions include internal boundary conditions at the tunnel wall and far-field stress boundary conditions: The internal boundary conditions at the tunnel wall are: The far-field stress boundary conditions are: In the formula, This refers to the stress in the original rock.

8. The dynamic parameterized design method for coal mine roadway support schemes according to claim 6, characterized in that, In step S32, the expression for the total loss function is: in: For the total loss function, For data residuals, The residuals of the surrounding rock equilibrium equations, For constitutive relation residuals, For boundary condition residuals, , , , These are the corresponding weighting coefficients.

9. The dynamic parameterized design method for coal mine roadway support schemes according to claim 1, characterized in that, Step S40 specifically includes the following steps: S41: During the tunnel excavation process, the surrounding rock response is first dynamically predicted based on the time series modeling method. The construction process is discretized into a time series. Using the excavation step sequence as an index, the monitoring data of surrounding rock displacement, stress, convergence deformation, and support structure stress are constructed as multivariate time series inputs. Combined with geological conditions and initial support parameters, the evolution law of the surrounding rock response over time is modeled through the time series prediction model, thereby obtaining the prediction results of surrounding rock deformation and stability indicators at several future moments. The time series prediction model continuously updates the latest monitoring data through a sliding time window, so that the model parameters are continuously corrected with the construction process, and the real-time tracking and trend prediction of the surrounding rock status are completed. S42: After obtaining the predicted results of the surrounding rock response, the predicted surrounding rock state and the current support parameters are combined as the system state input reinforcement learning decision model. The state characterizes the stability and stress characteristics of the surrounding rock. The actions are to adjust the length, spacing, preload of anchor cables and thickness of shotcrete. The reward function comprehensively considers the surrounding rock deformation control effect, safety factor and support cost. Thus, the optimal parameter adjustment scheme is generated through reinforcement learning strategy, realizing the transformation of support design from static scheme to dynamic decision. S43: During the execution phase, the support parameter adjustment scheme output by reinforcement learning is applied to the current design, and the response of the surrounding rock is compared and analyzed in conjunction with real-time monitoring data. The deviation between the prediction results and the actual monitoring results is used as a feedback signal. On the one hand, it is used to update the parameters of the time series model to improve the subsequent prediction accuracy. On the other hand, it is used for iterative optimization of the reinforcement learning strategy. The decision performance is continuously improved through strategy updates or value function corrections. S44: The updated support parameters are re-input into the physical information neural network and numerical simulation module for mechanical consistency verification and stability analysis. When the test results do not meet the constraints, the reinforcement learning model is triggered to make a decision again.

10. A dynamic parametric design system for coal mine roadway support schemes, characterized in that, include: The intelligent perception module is configured to acquire multi-source heterogeneous data from coal mine roadways, extract multi-modal data from the multi-source heterogeneous data, then perform deep fusion of the extracted data to construct a knowledge graph in the field of coal mine roadway support, and infer and complete missing or incomplete engineering parameters to form an original feature parameter set. The design intelligent module is configured to construct a regression and classification dual-path prediction framework based on a conditional diffusion model and a heterogeneous graph transformation network. It performs deep feature extraction, association learning, and candidate scheme generation on the original feature parameters and outputs an initial scheme of support parameters that meets the given geological conditions and engineering constraints. The physical verification intelligent module is configured to apply intrinsic constraints of surrounding rock mechanics and support structure stress to the initial support parameter scheme through a physical information neural network, and at the same time perform dual verification of the support scheme based on support specification information, and automatically adjust parameters that do not meet the constraints through an adversarial correction mechanism. The evolutionary intelligence module is configured to model the spatiotemporal effects during tunnel excavation, use real-time monitoring data during excavation as feedback increments to dynamically update the support scheme, correct parameter deviations caused by discontinuities in geological exploration or environmental disturbances, and achieve adaptive optimization and iterative updates of the support scheme through reinforcement learning strategies and incremental closed-loop mechanisms to obtain the final support scheme and realize the dynamic parameterization design of support parameters.

Images