Mineral resource overburden risk assessment method and system

Abstract

The invention provides a mineral resource overburden risk assessment method and system, and the method comprises the steps: obtaining multi-source heterogeneous geological data, carrying out the deep fusion of the multi-source heterogeneous geological data through employing a game theory gradient control robust neural network training method, and obtaining a fusion data set; performing deep mining and reasoning on high-dimensional geological data by utilizing the pattern recognition and correlation analysis capability of a large model and combining a manifold perception regularization technology, and recognizing geological structure features; establishing a dynamic safety mining depth evaluation model by using a fractional peak differential equation neural network with efficient accompanying parameter training, and calculating safety mining depth intervals and risk probabilities under different confidence degrees by integrating geological structure risks, mining disturbance effects and engineering safety thresholds; and constructing an assessment report generation module including core risk point analysis, key technology demonstration and prospective risk early warning, and generating an intelligent overburden risk assessment report. According to the invention, the accuracy and reliability of the pressing and covering risk assessment can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of mineral resources development, in particular to a mineral resource overburden risk assessment method and system for assessing the mutual influence degree between ground construction projects and underground mineral resources, determining the safe mining depth and risk level, and providing a scientific basis for mineral resource protection and engineering safety. BACKGROUND

[0002] Mineral resource overburden risk assessment is an important technical field to ensure the safety of ground construction projects and the rational development of underground mineral resources, involving multiple disciplines such as geological engineering, mining engineering, and geotechnical engineering. With the increasing requirements of the state for mineral resource protection and construction project safety, the scientificity and accuracy of overburden risk assessment technology are increasingly valued.

[0003] Traditional overburden risk assessment methods mainly rely on three-dimensional geological modeling technology and empirical formula calculation. The typical assessment process includes geological data collection, three-dimensional geological model construction, safe mining depth calculation, and risk level classification. Another common method is finite element analysis based on numerical simulation, which establishes a geomechanics model to assess the influence degree of mining on surface buildings.

[0004] The most mainstream method in the prior art is to construct a three-dimensional geological model using geostatistical methods such as Kriging interpolation, and then calculate the surface subsidence by combining the probability integral method or the influence function method, and finally determine the safe mining depth according to the specification standard. This method establishes a continuous geological interface by spatially interpolating discrete borehole data, and calculates the surface deformation based on rock mass mechanical parameters and mining geometric parameters using empirical formulas.

[0005] However, there are two key problems in the prior art: first, the depth of geological risk identification is insufficient, and traditional modeling methods cannot fully explore the complex correlation between multi-source data, and cannot deeply reveal the influence mechanism of geological structures such as faults and joints on mining stability; second, the demonstration of safe mining depth relies too much on empirical formulas, and lacks comprehensive consideration of the complexity of geological conditions and the dynamic nature of mining disturbance, resulting in limited accuracy and reliability of the assessment conclusion.

[0006] Therefore, there is an urgent need for an intelligent overburden risk assessment method and system that can deeply integrate multi-source heterogeneous geological data, accurately identify complex geological structure features, and dynamically evaluate safe mining depth. SUMMARY

[0007] The purpose of the present application is to provide a mineral resource overburden risk assessment method and system that improves the accuracy and reliability of overburden risk assessment by deeply integrating multi-source heterogeneous geological data, intelligently identifying complex geological structure features, and dynamically evaluating safe mining depth, solving the technical problems of insufficient depth of geological risk identification and limited accuracy of safe mining depth demonstration in traditional assessment methods.

[0008] To achieve the above object, the application provides a mineral resource overburden risk assessment method, comprising the following steps: Obtain multi-source heterogeneous geological data, including geological exploration data, drilling core data, geophysical exploration data, historical assessment cases and surrounding mine exploitation data, and use a robust neural network training method based on gradient control of game theory to deeply fuse the multi-source heterogeneous geological data to obtain a fused data set; Based on the fused data set, use the pattern recognition and correlation analysis capability of a large model, combined with manifold perception regularization technology, to deeply mine and reason high-dimensional geological data to identify geological structure features including stratigraphic interface, lithology distribution, fault fracture zone and joint fissure; Based on the geological structure features, use a fractional spike differential equation neural network with efficient adjoint parameter training to establish a dynamic safe mining depth evaluation model, comprehensively consider geological structure risk, mining disturbance effect and engineering safety threshold, and calculate safe mining depth intervals and risk probabilities under different confidence levels; According to the safe mining depth intervals and risk probabilities, build an evaluation report generation module including core risk point analysis, key technology demonstration and forward-looking risk warning to automatically generate an intelligent overburden risk assessment report.

[0009] Further, the robust neural network training method based on gradient control of game theory is used to deeply fuse the multi-source heterogeneous geological data to obtain a fused data set, comprising: The multi-source heterogeneous geological data is taken as a game participant, Nash equilibrium theory is used, and each data source is iteratively solved to reach an optimal weight distribution state to determine the contribution weight of each data source to the final fusion result, and an initial weight distribution is obtained; Based on the initial weight distribution, a robust neural network architecture containing residual connection and batch normalization is built, Gaussian noise and data enhancement strategy are introduced to improve network anti-interference ability, the optimal contribution weight of each data source under Nash equilibrium state is calculated, and an optimal weight distribution state is obtained; Each data source of the multi-source heterogeneous geological data is weighted and linearly combined according to the optimal weight distribution state, and fusion training is completed through neural network forward propagation and back propagation to obtain the fused data set.

[0010] Further, the robust neural network architecture containing residual connection and batch normalization is built, comprising: Based on the initial weight distribution, obtain the preprocessed multi-source heterogeneous geological data as the input layer of the neural network, set a residual connection module in the neural network, use batch normalization technology to standardize the output of each layer to ensure stable training of the deep network; For the input layer of the neural network, Gaussian noise disturbance and data enhancement strategies are introduced, including random rotation, scaling and translation transformation, to enhance the robustness of the network to data disturbance.

[0011] Further, the combination of manifold-aware regularization techniques deeply mines and reasons high-dimensional geological data, identifies geological structure features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joint fissures, including: Based on the geological feature points in the fusion data set, a k-nearest neighbor graph is constructed and the k nearest neighbors of each feature point are calculated to establish an adjacency relationship matrix between feature points and obtain a manifold structure representation of the data; Based on the manifold structure representation of the data, the local density and curvature information of the feature points in the manifold space are calculated, a manifold Laplacian matrix is constructed as a regularization term to constrain the inherent geometric structure of the data during large model training, and a regularization constraint parameter is obtained; The regularization constraint parameter is added to the loss function of the large model based on the Transformer architecture, and the model parameters are optimized through the multi-head self-attention mechanism and the back propagation algorithm to identify the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joint fissures, and obtain the geological structure features.

[0012] Further, the identification of the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joint fissures includes: Analyze the lithological variation sequence characteristics of the drilling core data and the resistivity anomaly and magnetic anomaly information of the geophysical exploration data in the fusion data set, and use a sequence pattern matching algorithm to automatically identify the depth and occurrence parameters of the main stratigraphic interface to obtain stratigraphic interface distribution information; Comprehensively analyze the core fracture degree in the borehole, the low resistance anomaly in geophysical exploration, and the structure record in historical geological data, and use a support vector machine classification algorithm to predict the strike, dip, and fracture zone influence range of the fault to obtain fault fracture zone distribution information; Model the joint fissure system as a graph structure using a graph neural network, with the intersection of fissures as nodes and the connection of fissures as edges, learn the spatial distribution law and connectivity characteristics of the fissure network through a graph convolutional neural network, and obtain the joint fissure development law; Integrate the stratigraphic interface distribution information, the fault fracture zone distribution information, and the joint fissure development law to construct a regional three-dimensional geological structure model and obtain the geological structure features.

[0013] Further, the establishment of the dynamic safe drilling depth evaluation model using the fractional spike differential equation neural network with efficient adjoint parameter training includes: Based on the geological structure characteristics, the memory effect and long-range correlation of the geological material are described by a fractional differential equation, a spike neural network is designed by using a leaky integral firing model, and a dynamic safe mining depth evaluation model is established; Based on the dynamic safe mining depth evaluation model, a multi-constraint optimization model is established, the constraint conditions include that the maximum surface subsidence, the maximum inclination and the maximum curvature deformation do not exceed the safety threshold, and a particle swarm optimization algorithm is used to solve the constraint optimization problem to obtain the safe mining depth interval and the risk probability.

[0014] Further, the fractional spike differential equation neural network with efficient adjoint parameter training comprises: For the state equation in the forward propagation , the corresponding adjoint equation is constructed , wherein lambda is an adjoint variable, L is a loss function, and theta is a network parameter, avoiding the gradient vanishing problem of traditional back propagation on a long time sequence; By solving the adjoint equation, the gradient of the loss function with respect to the network parameter is efficiently calculated, realizing the rapid training and parameter optimization of the neural network.

[0015] Further, the evaluation report generation module comprising core risk point analysis, key technology demonstration and forward-looking risk early warning comprises: Based on the safe mining depth interval and the risk probability, clustering analysis is used to divide the evaluation area into different risk areas, including high-risk areas, medium-risk areas and low-risk areas, and the geological parameters, calculation model parameters and evaluation results of each risk area are extracted to form a risk source archive; According to the risk source archive, a natural language generation technology is used to match the corresponding analysis method and demonstration logic for each risk type in combination with an expert knowledge base and an engineering case base, and the intelligentized pressure overlay risk evaluation report is automatically generated.

[0016] The application also provides a mineral resource pressure overlay risk evaluation system, comprising: A multi-source data fusion module is used to obtain multi-source heterogeneous geological data, including geological exploration data, drilling core data, geophysical exploration data, historical evaluation cases and surrounding mine exploitation data, and a robust neural network training method based on game theory gradient control is used to deeply fuse the multi-source heterogeneous geological data to obtain a fusion data set; A geological structure identification module is used to identify the geological structure characteristics including stratigraphic interface, lithology distribution, fault fracture zone and joint fissure based on the fusion data set, using the pattern recognition and correlation analysis capabilities of a large model in combination with manifold perception regularization technology to deeply mine and reason the high-dimensional geological data; A safe mining depth calculation module is used to establish a dynamic safe mining depth evaluation model by using a fractional spike differential equation neural network with high-efficiency accompanying parameter training, to calculate the safe mining depth interval and risk probability under different confidence levels by comprehensively considering the geological structure risk, mining disturbance effect and engineering safety threshold. A report generation module is used to construct an evaluation report generation module containing core risk point analysis, key technology demonstration and forward-looking risk warning according to the safe mining depth interval and risk probability, to automatically generate an intelligent pressure overlay risk evaluation report.

[0017] The present application has the following advantages: 1. The present application uses a robust neural network training method based on game theory gradient control to deeply fuse multi-source heterogeneous geological data, realizes optimal weight distribution and high-quality fusion of different types of geological data, and improves data utilization efficiency and fusion accuracy. 2. The present application combines manifold perception regularization technology to deeply mine and reason high-dimensional geological data, which can accurately identify complex geological structure features such as stratigraphic interface, lithology distribution, fault fracture zone and joint fissure, and improves the accuracy of geological risk identification. 3. The present application uses a fractional spike differential equation neural network with high-efficiency accompanying parameter training to establish a dynamic safe mining depth evaluation model, which can comprehensively consider geological structure risk, mining disturbance effect and engineering safety threshold, and improve the accuracy and reliability of safe mining depth evaluation. 4. The present application constructs an evaluation report generation module containing core risk point analysis, key technology demonstration and forward-looking risk warning, realizes the intelligent generation of pressure overlay risk evaluation report, and improves the professionalism and application value of the evaluation result. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0019] Figure 1 A schematic diagram of a geographic and geological information system provided by the present application; Figure 2 A flowchart of a mineral resource pressure overlay risk evaluation method provided by the present application; Figure 3 A structure diagram of a mineral resource pressure overlay risk evaluation system provided by the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with specific examples and drawings.

[0021] As Figure 1 shown, the geographic and geological information system is developed around the construction of a digital platform integrating geographic and geological professional data, aiming to solve the problems of scattered data, low efficiency of analysis, lagging decision-making and other problems in traditional geological work, and serve the scientific and precise management of resource exploration, environmental monitoring, engineering construction, disaster prevention and control and other fields. The system is called a certain regional geographic and geological information system, and its construction necessity lies in the standardized integration and management of geographic and geological data, the improvement of the efficiency and accuracy of geological analysis and research, and the provision of scientific decision-making basis for resource exploration and development. The reference material is not mentioned. In terms of system overview, the core goal breaks down data barriers through "spatial integration, intelligent analysis and visual expression", providing precise data support and decision-making tools for multiple fields, and realizing efficient utilization of geological information and maximization of social value. Among them, the functional goals include resource overview, resource distribution, project and mineral land coordination and other functions, and the non-functional goals require a response time of ≤2 seconds and an annual availability of ≥99.9%. The system scope includes resource overview, regional structure, latest information, resource distribution, mineral land distribution, project distribution, project and mineral land coordination and other functions, and the running hardware environment supports Windows 10 and above system, the software environment uses Linux CentOS 7 on the server side, the database is MySQL 8.0, the middleware is Redis6.0, and the browser needs Chrome 90+ or Edge 90+. In terms of functional requirements, according to business logic, it is divided into seven modules of resource overview, regional structure, latest information, resource distribution, mineral land distribution, project distribution and project and mineral land coordination. Each module has clear functions, resource overview displays the resource distribution overview of the province, regional structure presents the geological structure map of each region in the province, latest information publishes relevant national policies and latest information, resource distribution statistics the mineral resource distribution and mineral species proportion of each city, mineral land distribution displays the national and provincial mineral land distribution and divides the mining right and industrial right area, project distribution displays the wind power plant, photovoltaic and power transmission project distribution in the province, and project and mineral land coordination presents the comprehensive location area map of the two to assist project comprehensive analysis. Non-functional requirements include various requirements in performance, single-user query resource distribution page loading ≤ 1 second, supporting 10,000 users online at the same time; in security, user password uses MD5 salted encryption storage, no explicit permission control, needs to prevent SQL injection and XSS attacks; in reliability, the system automatically switches to the standby database when the database is down and the data is not lost, automatically backs up the data at 0 o'clock every morning, and the backup file is kept for 30 days; in ease of use, the interface is simple and easy to operate, and operation guide documents are provided; in scalability, microservice architecture is adopted, and new modules can be independently deployed to support future function expansion. The system architecture is divided into overall architecture and technical architecture. The overall architecture follows the hierarchical structure of front-end layer-API gateway layer-service layer-data layer, specifically client layer (Web)→API gateway (responsible for routing, authentication)→business service layer (covering resource overview, resource distribution modules)→data layer (including MySQL, Redis, file storage); The core technology stack of the technical architecture is clear, the front-end uses vue3+vite+cesium+amap+echart framework, and the back-end uses Java language and Spring Boot framework.

[0022] The user interface design follows the principles of simplicity and ease of use, and unified style. The test plan outline determines the test strategy of unit testing, integration testing, and performance testing. The test environment needs to meet the relevant requirements, and the key test indicators are the function test pass rate ≥ 95% and the performance test meets the response time requirements. The implementation and deployment plan clearly defines the system deployment steps as server environment configuration→database initialization→application deployment→data migration, and the online process adopts the mode of gray release→full release→monitoring feedback. In terms of maintenance and support, daily maintenance includes log cleaning, database backup inspection, and fault handling process specifies that when 500 errors occur, first check the application log, then check the database connection.

[0023] As shown in Figure 2 The mineral resource overburden risk assessment method provided by the present application comprises the following steps: Step S1: Obtain multi-source heterogeneous geological data, including geological exploration data, drilling core data, geophysical exploration data, historical assessment cases, and surrounding mine exploitation data, and use a robust neural network training method based on game theory gradient control to deeply fuse the multi-source heterogeneous geological data to obtain a fused data set.

[0024] In this step, the multi-source heterogeneous geological data of the evaluation area is comprehensively collected. Geological exploration data includes geological maps, drilling columnar sections, geological profiles, etc. These data record the spatial distribution of regional strata, lithological characteristics and basic structural characteristics. Drilling core data is the most direct source of underground geological information, including core photos, lithological description, structural feature records and physical and mechanical parameter test results. Geophysical exploration data mainly includes electrical, magnetic and gravity exploration data. These data can provide the distribution of physical properties of underground geological bodies, which helps to identify faults, fracture zones and other geological structures that are not easily discovered by drilling. Historical evaluation cases and surrounding mine exploitation data provide regional geological background and exploitation experience, including historical evaluation reports, exploitation subsidence monitoring data and disaster event records.

[0025] After data collection is complete, pre-process each type of data, converting data of different formats and resolutions into a unified digital standard format. For geological maps, use vectorization technology to extract geological boundaries; for drilling data, build a three-dimensional spatial database; for geophysical data, perform coordinate conversion and data noise reduction. The purpose of pre-processing is to eliminate data noise and unify data formats, laying the foundation for subsequent deep fusion.

[0026] After data pre-processing is complete, use a robust neural network training method based on game theory gradient control to perform deep fusion of multi-source heterogeneous geological data. The core idea of this method is to consider each data source as a game participant, determine the optimal contribution weight of each data source based on game theory principles, and then use a robust neural network to achieve deep fusion of the data. Compared with traditional data fusion methods, this method can adaptively adjust the weights of each data source, maximize the retention of useful information, suppress noise and redundancy, and generate high-quality fused data sets.

[0027] Step S2: Based on the fused data set, use the pattern recognition and association analysis capabilities of the large model, combined with manifold-aware regularization techniques, to perform deep mining and reasoning on high-dimensional geological data, and identify geological structural features including stratigraphic interfaces, lithological distribution, fault fracture zones and joint fissures.

[0028] In this step, based on the fused data set, the large model is used to perform deep mining and reasoning on high-dimensional geological data. The large model refers to a deep learning model based on the Transformer architecture, which has been pre-trained on a large corpus of geological literature, exploration reports and geological maps. It has strong geological knowledge understanding and reasoning capabilities. By inputting the fused data set into the large model, its pattern recognition and association analysis capabilities are used to mine the implicit geological structural features in the data.

[0029] To improve the accuracy of geological structure feature recognition, manifold-aware regularization techniques are combined. This technique is based on manifold learning theory, assuming that high-dimensional geological data is distributed on a low-dimensional manifold. By constructing the geometric structure of the data, the model learning process is constrained to maintain the intrinsic geometric properties of the data. In specific implementation, the k-nearest neighbor graph of data points is constructed, and the manifold Laplacian matrix is calculated, which is added to the loss function of the model as a regularization term. This method effectively solves the structure distortion problem that may occur when traditional deep learning is used to process high-dimensional geological data, improving the reliability of feature recognition.

[0030] Through deep mining and reasoning of large models, the geological structure features of the evaluation area are successfully identified, including stratigraphic interfaces, lithological distribution, fault fracture zones, and joint fissures. Stratigraphic interfaces reflect the superimposed relationship of rock layers in different geological periods, lithological distribution describes the spatial distribution of various rocks, and fault fracture zones and joint fissures are key structural elements affecting rock mass stability. These geological structure features provide basic information for subsequent safety mining depth evaluation.

[0031] Step S3: Based on the geological structure features, a dynamic safety mining depth evaluation model is established using a fractional spike differential equation neural network with efficient adjoint parameter training. The model integrates geological structure risk, mining disturbance effect, and engineering safety threshold to calculate the safety mining depth interval and risk probability under different confidence levels.

[0032] In this step, a dynamic safety mining depth evaluation model is established based on the identified geological structure features. This model uses a fractional spike differential equation neural network, which can effectively simulate the propagation process of mining disturbance under complex geological conditions. The introduction of fractional differential equations allows the model to better describe the memory effect and long-range correlation of geological materials, while the spike neural network effectively handles the strong temporal nature of geological evolution processes by simulating the discharge mechanism of biological neurons.

[0033] To improve the training efficiency of the model, an efficient adjoint parameter training algorithm is used. This algorithm avoids the gradient vanishing problem of traditional backpropagation on long time series by constructing an adjoint equation, which can efficiently calculate the gradient of the loss function with respect to the network parameters and accelerate the convergence process of the model. This algorithm is particularly suitable for handling geological mechanics simulation problems involving long-term dependencies, significantly improving the training efficiency and prediction accuracy of the model.

[0034] In the safety mining depth evaluation process, the geological structure risk, mining disturbance effect and engineering safety threshold are comprehensively considered. The geological structure risk mainly comes from unfavorable geological conditions such as faults, fracture zones and weak interlayers; the mining disturbance effect includes surface subsidence, horizontal deformation and inclination deformation; and the engineering safety threshold is determined based on relevant national standards and the actual bearing capacity of the building. A multi-constraint optimization model is established, and the optimal mining scheme is solved by a particle swarm optimization algorithm, and finally the safe mining depth interval and the corresponding risk probability under different confidence levels are output, providing a scientific basis for decision-making.

[0035] Step S4: Based on the safe mining depth interval and risk probability, an evaluation report generation module including core risk point analysis, key technology demonstration and forward-looking risk warning is constructed, and an intelligent overburden pressure risk evaluation report is automatically generated.

[0036] In this step, based on the safe mining depth interval and risk probability, an intelligent overburden pressure risk evaluation report is automatically generated. The report generation module adopts a modular architecture, including a risk identification module, a technology analysis module, a chart generation module and a text generation module. The risk identification module divides the evaluation area into high-risk, medium-risk and low-risk areas through cluster analysis, and extracts the key features of each area to form a risk source archive. The technology analysis module matches the corresponding analysis method and demonstration logic for each risk type based on the expert knowledge base and the engineering case base, and generates a technology analysis paragraph.

[0037] The chart generation module automatically creates various professional maps, including geological profile, risk distribution, safe mining depth contour map, etc. These maps use standard mapping specifications and color schemes to visually display the evaluation results. The text generation module uses natural language generation technology to write smooth and professional report text based on predefined report templates and specific evaluation results. The entire report contains three key parts: core risk point analysis, key technology demonstration and forward-looking risk warning, which comprehensively presents the results and recommendations of the overburden pressure risk evaluation.

[0038] The finally generated intelligent overburden pressure risk evaluation report not only meets professional standards, but also has practical value, and can provide a scientific basis for engineering decision-making to ensure the coordinated development of ground construction projects and underground mineral resources development.

[0039] In specific embodiments, the robust neural network training method using game theory gradient control is used to deeply fuse multi-source heterogeneous geological data in step S1, including: Step S1.1: The multi-source heterogeneous geological data is taken as a game participant, and the Nash equilibrium theory is used to iteratively solve the optimal weight distribution of each data source to determine the contribution weight of each data source to the final fusion result, and obtain the initial weight distribution.

[0040] In this step, various types of geological data are considered as participants in a game, and each participant's goal is to maximize their contribution to the fusion result. Nash equilibrium theory provides a method to find a balance point in the competition between participants, so that each participant has no incentive to change their strategy unilaterally. In the context of data fusion, this means finding a weight distribution scheme that optimizes the contribution of each data source to the optimal balance.

[0041] In implementation, first, a contribution function is defined for each data source, which measures the information gain of the data source to the fusion result. The contribution function is usually constructed based on factors such as data quality, reliability, and information quantity. For example, for drilling data, its contribution may be related to data density, coverage, and measurement accuracy; for geophysical data, the contribution may be related to signal quality and the significance of abnormal features.

[0042] Nash equilibrium is found by an iterative solution method. In each iteration, the weights of other data sources are fixed, and the weight of the current data source is optimized to maximize its contribution. After multiple iterations, when the weight change of all data sources is less than a preset threshold, it is considered that the equilibrium state is reached. This method avoids the subjectivity of weight distribution in traditional data fusion and realizes the adaptive optimization of data source weights.

[0043] Finally, the initial weight distribution of each data source is obtained. These weights reflect the relative importance of different data sources in information contribution, providing an important reference for subsequent neural network fusion. In practical applications, the initial weight distribution may be adjusted as the regional geological conditions and data quality change, ensuring the optimality of the fusion result.

[0044] Step S1.2: Based on the initial weight distribution, a robust neural network architecture is constructed, which includes residual connections and batch normalization. Gaussian noise and data augmentation strategies are introduced to improve network anti-interference ability, and the optimal contribution weight of each data source in the Nash equilibrium state is calculated to obtain the optimal weight distribution state.

[0045] In this step, based on the initial weight distribution, a robust neural network architecture is constructed for data fusion. This neural network adopts a deep learning architecture, which includes multiple hidden layers with residual connections and batch normalization modules between each layer. Residual connections establish a direct channel between shallow and deep features, effectively alleviating the gradient vanishing problem in deep network training; batch normalization standardizes the output of each layer, stabilizing the training process and accelerating the convergence of the network.

[0046] To improve the network's robustness against interference, Gaussian noise and data augmentation strategies were introduced. Gaussian noise was added to the network's input or intermediate layers to simulate random perturbations that might occur during actual data acquisition, enabling the network to learn more robust feature representations. Data augmentation strategies included random rotations, scaling, and translation transformations, which generated more training samples, enhancing the network's generalization ability and allowing it to adapt to data input under different conditions.

[0047] During training, the initial weight allocation is used as prior knowledge to guide the network's learning. The network's loss function consists of two parts: first, the data reconstruction error, which measures the fidelity of the fused result to the original data; and second, the Nash equilibrium constraint, which ensures that the contributions of each data source conform to the game equilibrium state. By alternately optimizing these two objectives, the network parameters and data source weights are continuously adjusted until the optimal weight allocation state under Nash equilibrium is finally achieved.

[0048] This neural network-based weight optimization method is more flexible and adaptable than direct mathematical solutions, capable of handling more complex nonlinear relationships and uncertainties, and yielding more accurate weight allocation results. The final optimal weight allocation state will be used in the subsequent data fusion process to ensure the quality and reliability of the fusion results.

[0049] Step S1.3: The data sources of the multi-source heterogeneous geological data are weighted linearly combined according to the optimal weight allocation state, and the fusion training is completed through forward and backward propagation of the neural network to obtain the fused dataset.

[0050] In this step, multi-source heterogeneous geological data are deeply fused based on the optimal weight allocation. First, the feature vectors of each data source are weighted linearly according to the optimal weights to form an initial fused representation. This weighting method ensures that different data sources participate in the fusion process reasonably according to their contribution, avoiding the negative impact of low-quality data sources on the final result.

[0051] Next, the initial fused representation is fed into the trained neural network for forward propagation. During forward propagation, the data undergoes multiple nonlinear transformations, and the network extracts more abstract feature representations layer by layer, ultimately generating high-level fused features. These deep features can capture the complex relationships between different data sources, achieving deep integration of information.

[0052] The neural network is trained using the backpropagation algorithm. This algorithm calculates the gradient of the loss function with respect to the network parameters and updates the parameters in the reverse direction of the gradient, progressively optimizing the network performance. The loss function typically includes multiple aspects such as reconstruction error, feature consistency, and spatial continuity, comprehensively evaluating the quality of the fusion result. During training, a batch processing approach is used, randomly selecting a batch of samples for training each time to improve training efficiency and generalization ability.

[0053] After multiple rounds of training, the network parameters gradually converge, and the fusion performance reaches a stable state. At this time, a complete forward propagation is performed on the data of the entire evaluation area to generate the final fusion data set. This fusion data set integrates the advantages of each data source, has higher integrity, consistency and reliability, and provides a high-quality data foundation for subsequent geological structure feature recognition.

[0054] In the multi-source heterogeneous geological data acquisition stage, first, the geological exploration data is standardized to convert different formats of drilling columnar section and geological profile into a unified digital format, and a structured database containing key information such as lithology code, stratum thickness, and occurrence elements is established. For drilling core data, image recognition technology is used to automatically extract texture features, color distribution and fracture development degree from core photos, and combine with core description records to generate multi-dimensional feature vectors. The processing of geophysical exploration data includes apparent resistivity data of electrical prospecting, magnetic anomaly data of magnetic prospecting and gravity anomaly data of gravity prospecting. Wavelet transform and Fourier analysis are used to remove noise and extract effective signal features.

[0055] In the process of building a robust neural network architecture containing residual connection and batch normalization, step S1.2 further comprises: Step S1.2.1: Based on the initial weight distribution, the preprocessed multi-source heterogeneous geological data is obtained as the input layer of the neural network, a residual connection module is set in the neural network, and batch normalization technology is used to standardize the output of each layer to ensure the stable training of the deep network.

[0056] In this step, first, the preprocessed multi-source heterogeneous geological data is obtained as the input of the neural network. Preprocessing includes data cleaning, normalization, missing value processing and feature extraction, etc. to ensure the quality and consistency of the input data. For geological exploration data, key geological elements are extracted to form feature vectors; for drilling core data, multi-dimensional features containing physical, chemical and mechanical parameters are constructed; for geophysical exploration data, various anomaly features and spatial distribution features are extracted.

[0057] Based on these preprocessed data, a deep neural network architecture is constructed. This network contains multiple hidden layers, each composed of fully connected layers or convolutional layers, capable of processing different types of geological data. To solve the problem of gradient vanishing in deep network training, a residual connection module is set in the network. Residual connection establishes a direct channel between shallow features and deep features by adding "skip connection", so that the gradient can be transmitted smoothly from the output layer to the shallow layer, effectively solving the problem of deep network training difficulty.

[0058] Another key technique is batch normalization, which adds a batch normalization layer after each hidden layer to standardize the layer output. Specifically, batch normalization first computes the mean and variance of a batch of data, then standardizes the data to have zero mean and unit variance, and finally restores the data's expressive power through learnable scaling and shifting parameters. This standardization has three main advantages: first, it speeds up network convergence and allows for larger learning rates; second, it reduces overfitting and improves model generalization; and third, it reduces network sensitivity to initialization and improves training stability.

[0059] Through the combination of residual connections and batch normalization, stable training of deep neural networks is achieved, laying a technical foundation for subsequent data fusion. This network architecture can effectively learn complex patterns and relationships in multi-source heterogeneous geological data, improving fusion results.

[0060] Step S1.2.2: For the input layer of the neural network, introduce Gaussian noise perturbation and data augmentation strategies, including random rotation, scaling and translation transformation, to enhance the network's robustness to data perturbations.

[0061] In this step, multiple techniques are used to improve the robustness of the neural network to data perturbations. First, Gaussian noise is introduced at the input layer of the network. Gaussian noise refers to random noise that follows a normal distribution, with a mean of 0 and a variance of a pre-set parameter value. In each training iteration, different Gaussian noise is added to the original input data to form slightly perturbed training samples. This perturbation simulates random errors that may exist in the actual data acquisition process, such as measurement errors, transmission errors or environmental disturbances. By constantly exposing the network to these perturbations during training, the network gradually learns to focus on stable and reliable features in the data, ignoring the effects of random noise, thereby improving its generalization ability to unseen data.

[0062] In addition to Gaussian noise, multiple data augmentation strategies are used to enrich the diversity of training samples. Random rotation refers to rotating the geological data at a random angle, which helps the network learn rotation-invariant feature representations, especially suitable for directional analysis of geological structures. Random scaling scales up or down the data by different proportions, allowing the network to adapt to geological features of different scales. Random translation shifts the data in space randomly, enhancing the network's ability to adapt to changes in feature location.

[0063] These data augmentation strategies effectively expand the number and diversity of training samples, allowing the network to encounter a wider range of data variations and establish a more robust feature representation. In practical applications, these augmentation strategies are often randomly combined to generate a more diverse set of training samples. By combining Gaussian noise perturbation and data augmentation strategies, the neural network can effectively resist various data perturbations and maintain stable performance in complex and variable geological environments, providing reliable data support for subsequent geological structure feature recognition.

[0064] This design ensures stable training of the deep network while introducing Gaussian noise and data augmentation strategies to improve the network's resistance to input perturbations. In robust neural networks, the residual connection structure helps to solve the gradient vanishing problem in deep network training, and the batch normalization technique accelerates network convergence and improves generalization ability by standardizing the output of each layer.

[0065] In step S2, manifold-aware regularization techniques are combined to deeply mine and reason high-dimensional geological data, identifying geological structure features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joint fissures, including: Step S2.1: Based on the geological feature points in the fusion data set, construct a k-nearest neighbor graph and calculate the k nearest neighbors of each feature point, establish the adjacency relationship matrix between feature points, and obtain the manifold structure representation of the data.

[0066] In this step, first extract the geological feature points from the fusion data set. Geological feature points refer to representative points in the geological space, including drilling locations, geophysical survey points, and geological boundary feature points. These feature points are distributed in a multi-dimensional feature space, and each point is represented by a set of feature vectors, including spatial coordinates, lithological properties, physical parameters, and chemical compositions. The purpose of extracting these feature points is to establish the spatial topology of geological data and lay the foundation for subsequent manifold learning.

[0067] Based on the extracted geological feature points, construct a k-nearest neighbor graph. The k-nearest neighbor graph is a graph model that represents the local structure of high-dimensional data, where each node represents a geological feature point, and the edges between nodes represent their proximity. The process of constructing a k-nearest neighbor graph first calculates the distance matrix between feature points. Common distance metrics include Euclidean distance, Mahalanobis distance, or cosine similarity. For geological data, Euclidean distance is suitable for numerical features such as physical parameters, while cosine similarity is more suitable for representing the similarity of categorical features such as lithology.

[0068] On the basis of the distance matrix, k nearest neighbors are selected for each feature point. The parameter k is an important hyperparameter, usually set according to the data size and distribution characteristics. A smaller k value can better preserve the local structure, while a larger k value can help capture global characteristics. In geological applications, k is usually set between 5 and 15, which can capture both local geological variation characteristics and reduce the influence of outliers.

[0069] After selecting the nearest neighbors, an adjacency matrix between feature points is established. The adjacency matrix is an n x n matrix (n is the number of feature points), and the matrix element A[i,j] represents the connection between feature points i and j. In a simple k-neighbor graph, if point j is one of the k nearest neighbors of point i, then A[i,j]=1, otherwise A[i,j]=0. More complex methods will give different weights to the edges according to the distance or similarity between points, so that the adjacency matrix can more accurately reflect the relationship strength between geological feature points.

[0070] By constructing the k-neighbor graph and the adjacency matrix, the manifold structure representation of the data is obtained. Manifold refers to a topological space that is locally similar to Euclidean space. In machine learning, high-dimensional data often distributes on a low-dimensional manifold. The manifold structure representation captures the essential geometric structure of geological data in high-dimensional feature space, providing a foundation for subsequent manifold learning and feature extraction. This representation method is particularly suitable for processing complex geological data, because geological features often exhibit strong spatial continuity and local correlation, which is highly consistent with the geometric properties of manifolds.

[0071] Step S2.2: Based on the manifold structure representation of the data, calculate the local density and curvature information of the feature points in the manifold space, construct the manifold Laplacian matrix as a regularization term to constrain the geometric structure of the data during the training process of the large model, and obtain the regularization constraint parameter.

[0072] In this step, based on the established manifold structure representation, the geometric characteristics of the feature points in the manifold space are further analyzed. Local density refers to the density of data points around a feature point, reflecting the sampling sufficiency or variation frequency of geological features in a certain area. The local density is estimated by calculating the ratio of the number of points within the k-neighborhood of each feature point to the volume. High-density areas usually correspond to areas with intense geological changes or dense sampling, which need to be given more attention in the model.

[0073] Curvature information describes the degree of curvature of the manifold at a feature point, which is an important feature of manifold geometry. When calculating the curvature, a tangent plane is first constructed at each feature point, and then the degree of deviation of data points from the tangent plane is measured. Areas with high curvature usually correspond to turning points or abnormal areas of geological structures, such as faults, fold axes, or lithology mutation zones. These areas have special significance in geological interpretation and need to be preserved in model training.

[0074] Based on the local density and curvature information, a manifold Laplacian matrix is constructed. The manifold Laplacian matrix is an important tool for describing the geometric structure of the data manifold. It is a transformation of the adjacency matrix, which can express the similarity relationship between data points and their distribution characteristics on the manifold. For the weighted k-neighbor graph, the calculation formula of the Laplacian matrix L is L = D - W, where W is the weighted adjacency matrix, and D is the degree matrix (diagonal matrix, diagonal elements are the sum of weights of corresponding rows or columns).

[0075] The manifold Laplacian matrix has many excellent mathematical properties, especially its eigenvectors can be used as basis functions on the manifold to represent functions on the manifold. This property makes the Laplacian matrix a core tool in manifold learning. In geological applications, the Laplacian matrix can capture the continuous variation patterns and local structural characteristics of geological bodies in space, and has important value for identifying complex geological structures.

[0076] The manifold Laplacian matrix is added as a regularization term to the training process of the large model. Regularization is a technique to prevent overfitting of the model. By adding an additional constraint term to the loss function, the model complexity or preference for a specific solution is limited. The core idea of manifold regularization is to maintain the intrinsic geometric structure of the data, so that the model prediction changes smoothly on the manifold. For geological data, this means that similar geological conditions should produce similar model predictions, consistent with the continuity principle in geology.

[0077] The regularization constraint parameter is a hyperparameter that controls the strength of regularization, determining the weight of the manifold structure constraint in the overall loss function. Through cross-validation or Bayesian optimization, the optimal regularization constraint parameter is determined to balance the model generalization ability and the preservation of geological structure. The optimization of these parameters is an iterative process that requires fine-tuning based on domain knowledge and experimental results to achieve the best model performance.

[0078] Step S2.3: Add the regularization constraint parameter to the loss function of the large model based on the Transformer architecture, optimize the model parameters through the multi-head self-attention mechanism and backpropagation algorithm, identify the spatial distribution of stratigraphic interfaces, lithology distribution, fault fracture zones, and joint fissures, and obtain the geological structure characteristics.

[0079] In this step, the regularization constraint parameter is integrated into the loss function of the large model based on the Transformer architecture. Transformer is a neural network architecture based on self-attention mechanism, originally applied to natural language processing, and later extended to various sequence data processing tasks. In geological applications, Transformer can effectively process spatial sequence data, capture long-range dependencies between geological feature points, and is suitable for processing complex geological structure recognition problems.

[0080] The loss function of the large model usually contains two parts: one is the basic task loss, which measures the difference between the model prediction and the true label; the other is the regularization loss, which constrains the model complexity or specific properties. In this system, the basic task loss uses cross-entropy loss or mean square error to measure the recognition accuracy of the model for geological structure features. The regularization loss is constructed based on the manifold Laplacian matrix, and the form is λ'·f T ·L·f, where f is the model output, L is the Laplacian matrix, and λ' is the regularization constraint parameter. This regularization term encourages the model output to change smoothly on the manifold, preserving the intrinsic geometric structure of geological data.

[0081] Multi-head self-attention mechanism is the core component of Transformer architecture, allowing the model to simultaneously focus on different parts of the input sequence and different representation subspaces. In geological applications, multi-head self-attention allows the model to consider multiple geological factors and multi-scale features simultaneously, for example, some attention heads may focus on local lithology changes, while others may focus on large-scale structural features. This multi-angle feature extraction capability enables the model to better understand complex geological conditions.

[0082] Self-attention mechanism builds global dependencies by computing the correlation between each element in the sequence and all other elements. Specifically, for each element in the input sequence, the model computes an attention weight distribution indicating which other elements it should focus on. This mechanism is particularly suitable for capturing spatial correlations and long-range dependencies in geological data, such as the correspondence of strata on both sides of a fault or long-distance structural connections.

[0083] Model training uses the backpropagation algorithm to compute the gradient of the loss function with respect to the model parameters and update the parameters in the opposite direction of the gradient to gradually optimize model performance. Backpropagation is a standard optimization method in deep learning, which can efficiently compute gradients in complex networks. During training, the system uses learning rate scheduling, gradient clipping, and other techniques to ensure the stability and convergence speed of training.

[0084] Through multiple rounds of training, the model gradually learns to recognize different types of geological structural features. The stratigraphic interface is the boundary between different geological periods or different lithological strata, and the model identifies the interface position by learning the feature differences and spatial continuity on both sides of the interface. Lithological distribution reflects the distribution of different rock types in space, and the model integrates drilling data and geophysical data to infer the complete three-dimensional lithological distribution. Fault fracture zone is a fractured deformation zone formed by rock mass due to faulting, with unique physical and geometric characteristics. The model locates faults by identifying resistivity anomalies, magnetic anomalies, and core fractures. Joint fissure is a small fracture in rock mass, usually appearing in groups, and the model describes the joint fissure system by analyzing the density, direction, and connectivity of the fissures.

[0085] Finally, the geological structural features including stratigraphic interface, lithological distribution, fault fracture zone, and joint fissure are output. These features are presented in the form of a three-dimensional spatial model, which intuitively describes the geological structure of the evaluation area and provides a solid geological foundation for subsequent safety evaluation of deep exploration.

[0086] The implementation of this manifold-aware regularization technique constrains the model learning process to maintain the intrinsic geometric structure of the data by constructing a k-nearest neighbor graph of geological features and calculating the local density and curvature information of feature points in the manifold space. In manifold learning theory, the k-nearest neighbor graph can effectively capture the local structure of high-dimensional data, expressing the similarity relationship between feature points through an adjacency relationship matrix, laying the foundation for subsequent manifold analysis.

[0087] In step S2.3, the spatial distribution of stratigraphic interface, lithological distribution, fault fracture zone, and joint fissure is identified, further comprising: Step S2.3.1: Analyze the lithological variation sequence features of the drilling core data and the resistivity anomaly and magnetic anomaly information of the geophysical exploration data in the fusion data set, automatically identify the depth and occurrence parameters of the main stratigraphic boundary using a sequence pattern matching algorithm, and obtain the stratigraphic interface distribution information.

[0088] In this step, the drilling core data in the fusion data set is first analyzed. Drilling core is a direct sampling of underground geology, providing first-hand information on stratigraphic lithology, structure, and physical and chemical properties. The system focuses on the lithological variation sequence features in the core data, which reflect the variation of strata in the vertical direction. Lithological variation usually manifests as sudden changes in lithology type, color, structure, tectonic, and physical properties, and these change points often correspond to stratigraphic interface positions.

[0089] A sliding window technique is used to detect lithological change points. Specifically, a fixed-size window is set on the borehole columnar section and gradually slid along the borehole depth direction, calculating the difference in lithological characteristics between the upper and lower parts of the window. When the difference exceeds a preset threshold, the location is marked as a potential stratigraphic interface. The calculation of the difference can be based on changes in lithological coding, abrupt changes in physical parameters, or differences in statistical characteristics; the specific method depends on the data type and quality.

[0090] Specifically, the methods for calculating the degree of difference include three types: difference calculation based on lithology coding, difference calculation based on physical parameters, and difference calculation based on statistical characteristics. In the difference calculation based on lithology coding, the lithological feature vectors of the upper and lower halves of the window are first extracted. Then, the Euclidean distance between the two feature vectors is calculated as the difference index, and its calculation formula is as follows:

[0091] in, Indicates the degree of difference based on lithological coding. The dimension of the feature vector. Indicates the upper half of the window The values ​​of each characteristic component, Indicates the lower half of the window. The numerical value of each feature component. For example, when the upper half of the window is sandstone, its lithology code is 2, and when the lower half is shale, its lithology code is 5. If only the lithology code is considered as a single feature (i.e., If the feature vector contains multiple dimensions, then the sum of the squares of the differences in each dimension needs to be taken to obtain the comprehensive difference index.

[0092] In the calculation of differences based on physical parameters, for continuous physical parameters such as density and resistivity, the standardized difference method is used for quantification, and the calculation formula is as follows:

[0093] in, Indicates the standardized variability based on physical parameters. This represents the average of the physical parameters in the upper half of the window. This represents the mean of the physical parameters in the lower half of the window. This represents the standard deviation of the physical parameters in the upper half of the window. The standard deviation of the physical parameter in the lower half of the window. For example, when the mean of the density in the upper half of the window is 2.5 g / cm3, the standard deviation is 0.1 g / cm3, the mean of the density in the lower half is 2.7 g / cm3, and the standard deviation is 0.15 g / cm3, the normalized difference is about 1.11 calculated by the above formula. This standardization can eliminate the influence of different parameter dimensions, making the difference degree comparable.

[0094] In the difference degree calculation based on statistical characteristics, the KL divergence (Kullback-Leibler divergence) method is used to measure the probability distribution difference of the upper and lower parts of the window, and the calculation formula is:

[0095] wherein, represents the distribution divergence based on statistical characteristics, represents the number of categories of data values, represents the probability of the data value in the upper half of the window being , and represents the probability of the data value in the lower half of the window being . In practical applications, when exceeds the preset threshold of 2.0, or exceeds the preset threshold of 1.0, or exceeds the preset threshold of 0.5, the position will be marked as a potential stratigraphic interface by the system. Through this multi-index comprehensive judgment method, different causes and forms of stratigraphic interfaces can be effectively identified, improving the accuracy and reliability of stratigraphic interface identification.

[0096] At the same time, analyze the resistivity anomaly and magnetic anomaly information in the geophysical exploration data. Resistivity anomaly reflects the change of underground rock mass electrical property, different lithology strata usually show different resistivity characteristics due to the difference of mineral composition, porosity and water content. When the resistivity value changes significantly, it often indicates the existence of stratigraphic interface. Similarly, magnetic anomaly reflects the spatial variation of rock mass magnetism, the sudden change of magnetic induction intensity is usually related to lithology change or geological structure, which is an important auxiliary information for identifying stratigraphic interface.

[0097] In order to integrate drilling and geophysical data, sequence pattern matching algorithm is used to automatically identify main stratigraphic interfaces. Sequence pattern matching is a technology to find specific patterns in sequence data, which can identify characteristic patterns related to stratigraphic interfaces in drilling and geophysical data in geological applications. In specific implementation, the system first extracts characteristic patterns from known stratigraphic interfaces as templates, and then searches for similar sequence fragments in the entire data set to identify potential stratigraphic interfaces.

[0098] In the sequence pattern matching process, the Dynamic Time Warping (DTW) algorithm is used to handle the problem of different sequence lengths and varying rates. DTW is a method to measure the similarity of two time series, which can automatically adjust the time axis to find the best alignment. This feature allows DTW to handle the irregular changes and scale differences commonly found in geological sequences, improving the accuracy and robustness of stratigraphic interface recognition.

[0099] Through sequence pattern matching, the depth of the main stratigraphic interface, i.e., the depth position of the stratigraphic interface in the borehole, is determined. For multiple borehole data, the system connects the discrete interface points into a continuous three-dimensional interface through spatial interpolation methods. Common interpolation methods include Kriging, inverse distance weighting, and spline interpolation, and the choice of method depends on the data distribution characteristics and geological background knowledge.

[0100] In addition to the depth, the occurrence parameters of the stratigraphic interface, including the strike and dip angle, are also determined. The strike is the direction of the intersection line of the stratigraphic interface and the horizontal plane, usually represented by the angle with the north direction; the dip angle is the angle between the stratigraphic interface and the horizontal plane, reflecting the degree of stratigraphic inclination. These occurrence parameters are calculated through the three-point method or least squares plane fitting method, providing important parameters for subsequent three-dimensional geological modeling.

[0101] Finally, complete stratigraphic interface distribution information is obtained, including the spatial position, geometric shape, and occurrence characteristics of the main stratigraphic interface. These information is presented in the form of a three-dimensional curved surface model, which intuitively describes the stratigraphic structure of the evaluation area and provides a basic geological framework for safe mining evaluation.

[0102] Step S2.3.2: Comprehensive analysis of the degree of core breakage in boreholes, low resistance anomalies in geophysical exploration, and tectonic records in historical geological data, using support vector machine classification algorithm to predict the strike, dip, and influence range of the fault fracture zone, obtaining the distribution information of the fault fracture zone.

[0103] In this step, the fault fracture zone is comprehensively identified and characterized. The fault fracture zone is a broken deformation zone formed during fault movement, in which rocks are sheared, broken, and even pulverized, forming a series of characteristic rock types and tectonic phenomena. As an important geological structure, the fault fracture zone not only affects groundwater flow and rock mass stability, but also is a potential cause of geological disasters, which has important significance for the evaluation of overburden risk.

[0104] First, analyze the degree of core fragmentation in the borehole. Core fragmentation is a direct evidence of fault fracture zones, usually described by parameters such as core integrity indicators (e.g., RQD values), fracture zone thickness, fracture zone mineral composition, and fracture zone structural characteristics. The system analyzes core photos using image recognition technology to automatically identify fracture zone characteristics such as core fragment size distribution, fragmentation texture, and secondary mineral filling. At the same time, integrate direct observation results from drilling records, such as drilling rate changes, borehole deflection, and drill tool wear, which are auxiliary indicators for judging fault fracture zones.

[0105] Low-resistance anomalies in geophysical exploration are another important feature of fault fracture zones. Fault fracture zones usually exhibit significantly lower resistivity than surrounding intact rock bodies due to high rock fragmentation, increased porosity, and frequent groundwater activity. Analyze electrical prospecting data to identify low-resistance anomaly areas and determine whether they are fault fracture zones based on their spatial form and extension characteristics. In addition, fault fracture zones may also exhibit magnetic anomalies in magnetic prospecting, especially when there is a significant difference in rock magnetism on both sides of the fault or when there is magnetic mineral filling in the fault zone.

[0106] Tectonic records in historical geological data are important references for fault identification. These records include geological maps, tectonic analysis maps, historical drilling reports, and regional tectonic research results, providing regional tectonic background and known fault distribution information. Extract fault-related information from these data using text mining and image recognition techniques as prior knowledge and verification for fault prediction.

[0107] Based on the above multi-source information, support vector machine (SVM) classification algorithm is used to predict the distribution and characteristics of faults. Support vector machine is a powerful supervised learning method that classifies data by constructing an optimal separating hyperplane. In fault identification applications, first extract feature vectors from known faults as training samples, which include core fragmentation indicators, resistivity anomaly characteristics, magnetic anomaly characteristics, and geological background parameters. Then, use these training samples to train the SVM model to learn the decision boundary that distinguishes between fault and non-fault areas.

[0108] The core of SVM is the selection of kernel functions, which determines the transformation of feature space and the complexity of decision boundary. For fault identification tasks, the radial basis function (RBF) kernel usually performs best because it can handle non-linear relationships between features. Use grid search and cross-validation methods to determine kernel function parameters (such as RBF kernel γ parameter) and penalty coefficient C, balancing the fitting ability and generalization ability of the model.

[0109] The strike, dip, and influence range of the fault are predicted by the trained SVM model. The strike is the direction of the intersection between the fault plane and the horizontal plane; the dip is the direction of the fault plane inclination, perpendicular to the strike; the influence range describes the width and spatial extension of the fault fracture zone. These parameters are determined by analyzing the faulting relationship between the two sides of the stratum, the geometric distribution of the fracture zone in space, and the extension characteristics of the geophysical anomaly.

[0110] Finally, the complete fault fracture zone distribution information is obtained, including the spatial position, geometric parameters, and influence range of the fault. These information is presented in the form of a three-dimensional model, which clearly describes the fault structure of the evaluation area and provides important geological risk basis for subsequent safety mining evaluation.

[0111] Step S2.3.3: Model the joint fracture system as a graph structure using graph neural networks, with fracture intersection points as nodes and fracture connections as edges, learn the spatial distribution and connectivity characteristics of the fracture network through graph convolutional neural networks, and obtain the development law of the joint fracture.

[0112] In this step, the joint fracture system is analyzed and modeled in depth. Joint fractures are discontinuities in rock mass that do not involve significant displacement, usually occurring in groups and forming a network-like distribution. Compared with faults, joint fractures are smaller in size but numerous in quantity, and have a significant impact on the overall strength and stability of the rock mass. In the pressure overlay risk assessment, accurately grasping the development law of joint fractures is crucial for predicting rock mass deformation and stability.

[0113] First, extract the fracture information from multiple sources. Drilling cores are an important source of fracture information, and the system analyzes core photos through image recognition technology to identify the location, direction, dip angle, opening, and filling characteristics of fractures. High-precision logging data in geophysical exploration, such as sonic logging, resistivity imaging, and borehole television imaging, also provide direct evidence of fractures. In addition, fracture information from surface outcrop observations and tunnel engineering is integrated to build a more complete fracture database.

[0114] Based on the extracted fracture information, graph neural network technology is innovatively used to model the joint fracture system as a graph structure. Graph structure is a mathematical model that describes the relationship between nodes and edges, and is very suitable for expressing the topological characteristics of the joint fracture system. In this model, the intersection points of fractures are represented as nodes of the graph, and the connections between fractures are represented as edges of the graph. This representation method can intuitively reflect the connectivity and spatial organization structure of the joint fracture system, providing an ideal data model for subsequent analysis.

[0115] Each node (intersection of fissures) contains attribute information such as position coordinates, number of intersecting fissures, and intersection angle; each edge (connection of fissures) contains characteristic parameters such as length, direction, opening, filling type, and permeability. Through this attribute-rich graph structure, the geometric and physical characteristics of the joint fissure system can be fully captured, providing data support for in-depth analysis.

[0116] Graph neural networks are a class of deep learning models specifically designed to handle graph-structured data, capable of learning node and edge representations and reasoning about patterns on graphs. The graph convolutional neural network (GCN) is a mainstream graph neural network architecture used to learn and analyze fissure networks. The core operation of GCN is graph convolution, which updates node representations by aggregating information from neighboring nodes, thereby capturing local structural features of the graph. By stacking multiple layers of graph convolution, the model can gradually acquire larger-scale structural information, ultimately learning the global features of the joint fissure system.

[0117] In practical implementation, the system defines specialized graph convolution operations that take into account the unique physical properties of fissures. For example, the weight of an edge not only considers topological connections but also incorporates the opening and permeability of the fissure, allowing the model to better simulate the propagation of fluids and stresses in the fissure network. In addition, the system introduces an attention mechanism to dynamically adjust the importance of different neighbor nodes, more accurately capturing the complex patterns of fissure interaction.

[0118] Through the training of the graph convolutional neural network, the spatial distribution patterns and connectivity features of the fissure network are learned. Spatial distribution patterns include the directionality, density distribution, and scale distribution of fissures. The system identifies the directions of the main fissure groups, typically represented in the form of strike / dip-angle, such as "NE 45° / SE-70°". The density distribution describes the variation in the density of fissures in space, usually represented by line density (total length of fissures per unit area) or volume density (total area of fissures per unit volume). The scale distribution reflects the statistical patterns of fissure length, opening, and other parameters, typically following a power-law or log-normal distribution.

[0119] Connectivity features are another key attribute of the joint fissure system, directly affecting the permeability and mechanical behavior of the rock mass. The system quantifies the connectivity of the fissure network by analyzing connectivity indicators of the graph structure, such as average degree, clustering coefficient, path length, and number of connected components. In addition, the system identifies key nodes and edges in the fissure network, which play a decisive role in maintaining network connectivity, and their destruction can lead to significant changes in the performance of the entire system.

[0120] Finally, the complete joint fracture development law is obtained, including the spatial distribution characteristics, geometric statistical parameters and connectivity characteristics of the fractures. These laws are presented in the form of statistical models and three-dimensional visualization models, which provide a scientific basis for understanding and evaluating the rock mass structure of the region and predicting its mechanical behavior, and are important inputs for safety mining evaluation.

[0121] Step S2.3.4: integrating the stratum interface distribution information, the fault fracture zone distribution information and the joint fracture development law to construct a regional three-dimensional geological structure model and obtain the geological structure characteristics.

[0122] In this step, the stratum interface distribution information, fault fracture zone distribution information and joint fracture development law identified in the previous steps are integrated to construct a complete regional three-dimensional geological structure model. The three-dimensional geological structure model is a digital expression of the geometry and spatial relationship of the underground geological body, which can intuitively present complex geological structures and is an important basis for pressure overlay risk assessment.

[0123] This digital expression first uses a three-dimensional Cartesian coordinate system to locate the spatial positions of all geological elements, in which the horizontal plane uses an eastward X-axis and a northward Y-axis coordinate system, and the vertical direction uses a downward Z-axis to represent the depth coordinate. Each geological element is accurately described by a set of spatial coordinate points, and the spatial position of any point on the geological interface can be represented as:

[0124] wherein, represents the spatial position vector of a point on the geological interface, represents the eastward coordinate of the point, represents the northward coordinate of the point, represents the depth coordinate of the point. For example, a stratum interface can be represented as a point set containing thousands or even tens of thousands of such three-dimensional coordinate points, which completely describe the distribution of the stratum interface in three-dimensional space.

[0125] In terms of digital expression of geometry, the system uses triangular mesh technology to construct continuous surfaces of stratum interfaces and fault surfaces. Each triangular patch is completely defined by the spatial coordinates of its three vertices and the connection relationship between the vertices, and the triangular patch can be represented as: = ( , , ) wherein, represents a triangular patch, , , respectively. For example, a complexly undulating stratigraphic interface can contain five thousand triangular facets, which are pieced together according to topological relationships to form a complete three-dimensional curved surface model of the stratigraphic interface. This triangular mesh expression not only can accurately describe the complex curved surface morphology, but also is convenient for computer storage, display and analysis.

[0126] For digital expression of the attribute information of the geological body, the system associates a multi-dimensional attribute vector with each spatial unit, which can be expressed as:

[0127] where A represents an attribute vector at a spatial position, represents the number of dimensions of the attribute, respectively represent different geological attribute parameters, including lithology type code, rock density, elastic modulus, Poisson's ratio, compressive strength, etc. Specifically, in a voxelized three-dimensional geological model, the entire evaluation area is divided into regular three-dimensional grid units, each of which is called a voxel. For example, a voxel unit at a particular location can record that it is sandstone (lithology code 2), density is 2.65 grams per cubic centimeter, elastic modulus is 15 gigapascals, etc.

[0128] The spatial topological relationships between geological elements are digitally described by a relationship data table. By using this standardized data format, not only is it convenient for long-term data preservation and exchange and sharing, but also it provides a good data basis for subsequent calculation analysis and visualization display. The entire digital expression system realizes accurate, complete and computable description of complex three-dimensional geological structures, and provides a reliable geological model basis for dynamic safety mining evaluation.

[0129] The integration process first processes the stratigraphic interface data. The stratigraphic interface is the basic framework of the three-dimensional geological model, which defines the spatial boundaries of different lithological units. The main stratigraphic interfaces identified in the previous step are used as constraint surfaces, and a spatial interpolation method is used to generate a continuous three-dimensional surface. Common interpolation methods include triangular irregular network (TIN) interpolation, Kriging interpolation and discrete smoothing interpolation (DSI), etc. Among them, the DSI method is particularly suitable for constructing complex geological surfaces, because it can consider multiple constraint conditions such as interface points, occurrence measurements, fault cutting, etc., to generate smooth and geologically reasonable surfaces.

[0130] ​Next, the fault fracture zone information is integrated. Faults are important structural elements in the geological model, controlling the faulting relationship and spatial distribution of strata. The identified faults are represented as three-dimensional surfaces, and a geometric model is constructed based on their strike, dip, and spatial extension. Importantly, not only the fault surfaces themselves are simulated, but also the width and physical property changes of the fault fracture zone, which are crucial for accurately assessing rock mass strength and deformation characteristics.

[0131] During the model construction process, the intersection relationship between faults and strata interfaces is handled. When a fault cuts through a stratum, it can cause the stratum to be faulted and displaced, forming typical faulted geological structures. By setting the priority of the fault higher than that of the strata interface, the faulting effect of the fault on the stratum is achieved. For complex fault systems, the system also needs to handle the intersection and cutting relationship between faults to ensure the rationality of the model structure.

[0132] The integration of the joint fissure system is an innovative point, as traditional three-dimensional geological models often ignore this scale of structural elements. Based on the fissure development rules obtained in the previous steps, the system uses a random simulation method to generate a fissure network in three-dimensional space that conforms to statistical characteristics. Specifically, the system first defines the direction distribution, density distribution, and scale distribution parameters of the fissure group, and then randomly generates fissure surfaces in three-dimensional space so that their overall statistical characteristics conform to the observed data. This method can reasonably infer the fissure distribution in the entire region based on limited observed data.

[0133] During the integration process, the logical consistency between different scale geological structures is ensured. For example, large faults are usually accompanied by an increase in fissure density; certain strata interfaces may be the preferred faces for fissure development; and fissures may exhibit different geometric characteristics when they pass through different lithologies. The system ensures that the final model conforms to geological principles and field observation facts by defining the mutual relationship rules between these structural elements.

[0134] In terms of technical implementation of building a three-dimensional geological structure model, a hybrid representation method of voxels (Voxel) and surfaces (Surface) is used. Voxel representation discretizes three-dimensional space into regular grids, and each grid cell is assigned a geological attribute value, suitable for expressing continuous changing physical property distribution; surface representation uses a triangular mesh to describe geological interfaces and fault surfaces, suitable for expressing the boundaries of geological bodies and structural surfaces. This hybrid representation method balances precision and efficiency, and can fully express the characteristics of complex geological structures.

[0135] Finally, a complete regional 3D geological structure model is obtained, including key geological structure characteristics such as stratigraphic interface, lithology distribution, fault fracture zone, and joint fissure system. This model is presented in the form of a 3D digital model, which can be observed and sectioned from any angle, intuitively displaying the spatial relationship and internal structure of underground geological structures. As the basis for overburden risk assessment, this model provides an assessment of the regional geological structure framework, providing key inputs for subsequent safe drilling depth calculation and risk assessment.

[0136] This series of steps realizes the accurate identification of complex geological structures. In the stratigraphic interface identification process, the sequence pattern matching algorithm can automatically identify the spatial distribution of the main stratigraphic boundary by analyzing the lithology change sequence and geophysical anomaly characteristics in the drilling data. The support vector machine classification algorithm is used for fault fracture zone identification, which predicts the geometric parameters and influence range of faults by comprehensively analyzing the characteristics of multi-source data. For the joint fissure system, the graph neural network technology effectively learns the spatial distribution and connectivity characteristics of the fissure network by modeling the joint fissure system as a graph structure.

[0137] In step S3, a dynamic safe drilling depth evaluation model is established using a fractional spike differential equation neural network with efficient adjoint parameter training, including: Step S3.1: Based on the geological structure characteristics, the memory effect and long-range correlation of geological materials are described by fractional differential equations, and a spike neural network is designed using a leaky integral firing model to establish a dynamic safe drilling depth evaluation model.

[0138] In this step, a dynamic safe drilling depth evaluation model is established based on the previously identified geological structure characteristics. The innovation of this model lies in the introduction of fractional differential equations and spike neural networks, which can more accurately describe the dynamic response of geological materials under mining disturbance, thereby achieving more reliable safe drilling depth evaluation.

[0139] Fractional differential equations are a mathematical tool that extends the order of differentiation and integration to non-integer values, which has unique advantages in describing materials with memory effects and long-range correlations. Traditional integer-order differential equations assume that the future state only depends on the current state (Markov property), ignoring the cumulative effect of historical states. However, geological materials often exhibit significant memory effects, meaning that the current deformation of the material depends not only on the current stress state but also on the historical loading path. This memory effect is particularly pronounced in rheological materials such as soft rock, salt rock, and clay rock.

[0140] Fractional differential equations naturally express this memory effect by introducing fractional derivative operators. Fractional derivative is defined as fractional integration of integer order derivative, or equivalently, fractional differentiation of integer order integration. This mathematical tool enables to consider the influence of all past time instants on the current state, with weights decaying with increasing time interval but never vanishing, thus accurately capturing the long-term memory property of geological materials. In safety assessment of deep mining, this means that the model can consider the cumulative effect of mining activities and more accurately predict the long-term evolution of surface deformations.

[0141] The choice of the fractional order is a key parameter in the model, which determines the strength of the memory effect. The fractional order a usually ranges from 0 to 2, with a = 1 degenerating to the classical first-order derivative and a = 2 degenerating to the second-order derivative. For geological materials, the choice of a is based on the mechanical properties of the materials and field test data. Generally, loose sedimentary rock layers have smaller a values (0.5-0.7), showing stronger memory effects; hard crystalline rock layers have larger a values (0.8-1.0), being closer to the classical elastic model; and strata with obvious rheological properties may have a values exceeding 1.0. The most suitable a value is determined by inverse analysis of the measured data to improve the accuracy of the model.

[0142] In implementing fractional differential equations, numerical approximation methods such as Grünwald-Letnikov definition or finite difference scheme are used to discretize the continuous fractional derivative into a computer-implementable form. These methods usually involve weighted sums of historical states, with weights calculated according to a specific recurrence relation to numerically simulate the memory effect.

[0143] Spiking neural networks are another innovation, which simulate the pulse firing mechanism of biological neurons and are particularly suitable for handling strong time-series dynamics. Unlike traditional neural networks that use continuous activation functions, neurons in spiking neural networks transmit information through discrete spikes (spikes). The membrane potential of a neuron evolves over time, and when it exceeds a threshold, it fires a spike and resets the membrane potential. This mechanism enables the network to naturally handle time information and dynamic processes.

[0144] Spiking neurons are implemented using the leaky integrate-and-fire (LIF) model, which is a widely used simplified neuron model. In the LIF model, the membrane potential V(t) of a neuron evolves according to the following differential equation:

[0145] where τ is the membrane time constant, V rest is the resting potential, R is the membrane resistance, and I(t) is the input current. When V(t) reaches the threshold V th , the neuron fires a spike and resets to V restThis dynamic characteristic enables the spiking neural network to effectively capture the time-dependent and dynamic response features of geology.

[0146] In combination with the fractional differential equation and the spiking neural network, a dynamic safe mining depth evaluation model is constructed. First, the geological structure features are converted into model inputs, including key parameters such as lithology distribution, fault location, and fracture density. Then, the fractional differential equation is used to describe the dynamic response of geological materials under mining disturbance, considering the memory effect and long-range correlation of the materials. Finally, the spiking neural network is used to simulate and predict the dynamic evolution, outputting the surface deformation and risk probability under different mining depths.

[0147] In combination with the fractional differential equation and the spiking neural network, a dynamic safe mining depth evaluation model is constructed. First, the geological structure features are converted into model inputs, including key parameters such as lithology distribution, fault location, and fracture density. Then, the fractional differential equation is used to describe the dynamic response of geological materials under mining disturbance, considering the memory effect and long-range correlation of the materials. Finally, the spiking neural network is used to simulate and predict the dynamic evolution, outputting the surface deformation and risk probability under different mining depths.

[0148] The specific surface deformation calculation is realized by establishing a fractional differential dynamic equation, which can accurately describe the deformation process of geological materials evolving over time under mining disturbance. The equation contains multiple key parameters, including the fractional order parameter, which is used to characterize the memory characteristics of geological materials, and its value is usually between 0.5 and 1.5, determined according to the rheological properties of the rock mass. The mining area parameter and the mining depth parameter directly reflect the scale of mining activities, and the geological structure parameter vector comprehensively represents the fault distribution, fracture development degree, and other geological conditions. The stress field distribution function describes the stress state of the underground rock mass at different positions and times.

[0149] The spiking neural network solves the above fractional differential equation by simulating the pulse firing mechanism of biological neurons, thereby calculating the settlement at any position on the surface at any time. Each output neuron in the network corresponds to a spatial position on the surface, and its output value is the predicted settlement at that position. The calculation of neurons involves multiple layers of weight parameters and threshold parameters, which have been optimized and determined through the network training process. The input of the network includes geological parameters, mining parameters, and time information, which are transformed nonlinearly by the hidden layer and summed by the output layer, finally obtaining the predicted value of the surface settlement.

[0150] After obtaining the spatial distribution of ground subsidence, the system further calculates two important deformation indicators, ground inclination and curvature. Ground inclination is obtained by calculating the gradient of subsidence in the horizontal direction. Specifically, the partial derivatives of subsidence with respect to east and north coordinates are calculated, and then the square root of the sum of the squares of the two partial derivatives is taken to obtain the result, which reflects the steepness of the ground slope. Ground curvature is obtained by calculating the second-order partial derivatives of subsidence. The second-order partial derivatives with respect to east and north coordinates are calculated, and the absolute value of the larger one is taken as the curvature value at that location, which reflects the degree of ground bending deformation.

[0151] The calculation of risk probability is based on the exceeding probability analysis method of safety threshold. The system first defines three key safety thresholds, namely the maximum allowed subsidence threshold, the maximum allowed inclination threshold, and the maximum allowed curvature threshold, which are determined according to the type and structural characteristics of the ground buildings. Then, for a specific mining depth condition, the system uses the Monte Carlo simulation method to perform a large number of random sampling calculations. In each simulation, considering the uncertainty of geological parameters and model parameters, a set of parameter values is randomly generated, and the ground deformation under this parameter combination is calculated using the fractional differential equation and the spike neural network described earlier. By counting how many times the ground maximum subsidence exceeds the allowed threshold, or the maximum inclination exceeds the allowed threshold, or the maximum curvature exceeds the allowed threshold in the total number of simulations, and dividing this number by the total number of simulations, the risk probability under the current mining depth condition is obtained.

[0152] For example, when evaluating a mining depth of 300 meters, the system performs 10,000 Monte Carlo simulation calculations, and in these 10,000 simulations, it is found that there are 1,500 times when the ground deformation exceeds the safety threshold. The risk probability of this mining depth is 1,500 divided by 10,000, which is 0.15 or 15%. This risk probability value quantitatively represents the possibility of occurrence of ground building safety risk under the current geological conditions and mining scheme, providing a scientific risk assessment basis for decision-makers. By systematically calculating the risk probability of different mining depths, a curve of risk probability changing with mining depth can be drawn, providing a quantitative basis for determining the safe mining depth interval.

[0153] This innovative model architecture enables comprehensive consideration of geological structure risk, mining disturbance effect, and its time evolution characteristics, achieving more accurate safety mining depth evaluation. Compared with traditional empirical formulas or integer-order differential equation models, this model can better capture the nonlinear dynamic characteristics and long-term evolution trend of geology, improving the accuracy and reliability of prediction.

[0154] Step S3.2: Based on the dynamic safe mining depth evaluation model, a multi-constrained optimization model is established, the constraint conditions include that the maximum surface subsidence, maximum inclination, and maximum curvature deformation do not exceed the safety threshold, and a particle swarm optimization algorithm is used to solve the constrained optimization problem to obtain the safe mining depth interval and risk probability.

[0155] In this step, based on the dynamic safe mining depth evaluation model, a multi-constrained optimization model is established, and by solving the optimization problem, the safe mining depth interval and the corresponding risk probability under different confidence levels are determined. This process converts geological safety requirements into mathematical constraint conditions, and finds the optimal mining scheme that meets the safety standards through optimization algorithm, providing a scientific basis for decision-making.

[0156] First of all, it needs to be clear that the risk probability mentioned here is exactly the same physical quantity as the risk probability calculated before, both representing the possibility of exceeding the surface deformation under the condition of a certain mining depth.

[0157] In the process of S3.2, the quantitative determination process of the safe mining depth interval under different confidence levels contains several key steps. First, the optimization objective function is established, which is set to maximize the amount of mineable mineral resources. The calculation of resource quantity involves parameters such as mining area, mining depth, average density of the ore layer, and loss rate. Through optimization algorithm, the mining depth scheme that can obtain the maximum resource quantity is found, which represents the optimal economic benefit under the safety constraint conditions.

[0158] The constraint conditions of the optimization model are the core of the entire solving process, mainly including two categories: risk probability constraint and physical feasibility constraint. The risk probability constraint requires that the calculated risk probability under the selected mining depth cannot exceed the pre-set upper limit of the allowable risk probability. The setting of this allowable risk probability upper limit is directly related to the confidence level. The higher the confidence level, the higher the safety guarantee required, and the lower the allowable risk probability upper limit. Specifically, when requiring 95% confidence, it means that there is 95% confidence that the mining is safe, and the corresponding allowable risk probability upper limit is set to 5%; when requiring 90% confidence, the allowable risk probability upper limit is set to 10%; when requiring 80% confidence, the allowable risk probability upper limit is set to 20%. The physical feasibility constraint includes conditions such as that the mining depth must be within the technically feasible range, and the surface subsidence, inclination, and curvature must meet the corresponding safety standards.

[0159] The calculation of the safe mining depth interval is achieved by traversing the possible mining depth range and combining an optimization algorithm. The system first sets a large depth search range, for example, from 200 meters to 600 meters, and then calculates the risk probability corresponding to each depth value one by one at a certain step size. Through the Monte Carlo simulation method described above, 10000 random simulations are performed for each candidate depth, the number of exceedances is counted, and the risk probability is calculated. The calculated risk probability is compared with the upper limit of the allowable risk probability, and all depth values that meet the risk probability constraint are identified. These depth values constitute the preliminary feasible region.

[0160] On the basis of the determined feasible region, the lower limit and the upper limit of the safe mining depth interval are further solved. The lower limit of the interval is defined as the minimum mining depth that meets the risk probability constraint. Its physical meaning is that when the mining depth is less than this value, due to the shallow depth, the risk of ground surface deformation is greater, and the calculated risk probability will exceed the allowable upper limit, not meeting the safety requirements. The lower limit is determined by finding the minimum depth value in the feasible region whose risk probability is equal to or slightly less than the allowable upper limit. The upper limit of the interval is defined as the optimal mining depth that can maximize the resource quantity under the conditions of meeting the risk probability constraint and all physical constraint conditions. The upper limit is determined by searching for the optimal solution in the feasible region through the particle swarm optimization algorithm.

[0161] For any mining depth within the interval, its corresponding risk probability can be calculated according to the method described above. This risk probability value must not exceed the set allowable risk probability upper limit. Taking the case of 95% confidence as an example for quantitative description, the allowable risk probability upper limit is 5% at this time. The system traverses the depth range from 200 meters to 600 meters, and calculates the risk probability every 10 meters. The calculation results show that when the mining depth is less than 380 meters, the risk probability exceeds 5%, not meeting the safety requirements; when the mining depth reaches 380 meters, the risk probability just drops to about 5%, meeting the constraint condition, so the lower limit of the interval is determined as 380 meters. Continue to search deeper, the risk probability of all depth points in the range of 380 meters to 600 meters does not exceed 5%, meeting the safety requirements. Combined with the analysis of geological conditions and the feasibility of mining technology, and considering the goal of maximizing resource quantity, the optimization algorithm determines the optimal depth as 420 meters, so the upper limit of the interval is set as 420 meters. Finally, under the requirement of 95% confidence, the safe mining depth interval is determined as 380 meters to 420 meters, and the risk probability of any depth within this interval meets the safety requirement of not exceeding 5%.

[0162] Through the above systematic quantitative analysis process, a strict correspondence between the confidence level, the allowable risk probability and the safe mining depth interval is established. Different confidence requirements correspond to different upper limits of allowable risk probability, and then determine the safe mining depth interval of different width, which provides a scientific basis for mineral resource development decision. This method not only guarantees the scientificity and reliability of the evaluation results, but also provides a flexible selection space for decision makers, which can select the appropriate confidence level according to the importance of the project and the sensitivity of the building, and achieve the optimal balance between safety and economic benefits.

[0163] The core of the multi-constraint optimization model is to define appropriate objective function and constraint conditions. In the safety mining depth evaluation, the objective function is usually set as maximizing the exploitable resource quantity or economic benefit, while the constraint conditions are based on the safety requirements of surface buildings and facilities. Three main constraint conditions are considered: maximum surface subsidence, maximum inclination and maximum curvature deformation.

[0164] The maximum surface subsidence refers to the maximum vertical displacement of the ground surface caused by mining activities, which is directly related to the overall subsidence of the ground buildings. According to the type and importance of different buildings, the safety threshold is different. Generally speaking, the maximum subsidence of ordinary civil buildings is usually not more than 200-300mm; important public buildings may be limited within 100-150mm; and for particularly sensitive buildings or facilities, such as precision instrument workshops or high-speed rail lines, the limit value may be as low as 10-30mm. According to the actual situation of buildings in the evaluation area and relevant specifications, appropriate subsidence limit values are set.

[0165] The maximum inclination is another key indicator of surface deformation, which refers to the maximum slope of the surface subsidence curve, reflecting the degree of differential settlement that the building may experience. Inclination deformation will cause the building to tilt, affecting its use function, and in severe cases, it may threaten the structural safety. The inclination limit value of general civil buildings is usually 3-6mm / m; for high-rise buildings or frame structures, the limit value may be more stringent, such as 2-3mm / m; and for special facilities, such as large-span bridges or precision equipment workshops, the limit value may be as low as 1mm / m or less. According to the structural type and functional requirements of the building, appropriate inclination limit values are determined.

[0166] The maximum curvature deformation is an indicator reflecting the degree of surface bending, defined as the second derivative of the surface subsidence curve, which is directly related to the bending deformation of the building. Curvature deformation will cause the building to produce tensile or compressive stress, which may cause wall cracking or structural damage. For brick-concrete structure buildings, the curvature deformation limit value is usually 0.2-0.3mm / m 2 ; for reinforced concrete frame structures, the limit value may be 0.3-0.5mm / m 2between 0.5% and 1.5%; while for special structures or facilities, the limit needs to be determined according to their deformation resistance. Based on the structural characteristics and material properties of the building, corresponding curvature deformation limits are set.

[0167] These constraints collectively form the boundary conditions for safe mining. The goal is to determine the maximum possible mining depth while satisfying all safety constraints. Mathematically, this forms an optimization problem with multiple nonlinear constraints, which requires efficient numerical algorithms to solve.

[0168] Particle Swarm Optimization (PSO) algorithm is used to solve the above constrained optimization problem. PSO is a global optimization method based on swarm intelligence, simulating the foraging behavior of bird flocks, and finding the optimal solution through information sharing between particles. In this algorithm, each particle represents a possible mining scheme, including mining depth, mining range, and mining sequence, etc. The particle moves in the search space, and its position update is guided by the individual optimal position and the global optimal position.

[0169] The fitness function of the particle is designed to maximize the amount of mined resources or economic benefits while satisfying all constraints. To handle constraints, the penalty function method is used to convert the degree of constraint violation into a penalty term of fitness, guiding the search process to converge to the feasible solution region. This method can effectively handle complex nonlinear constraints and find the global optimal or near-global optimal solution.

[0170] In practical applications, the size of the particle swarm is usually set to 50-100 particles, and the number of iterations is 500-1000, to ensure that the algorithm has sufficient exploration ability and convergence. Adaptive inertia weight and contraction coefficient techniques are also used to balance the global search and local refinement capabilities of the algorithm, improving the solution efficiency and solution quality.

[0171] Considering the uncertainty of geological parameters and model predictions, not only a single optimal solution is output, but also a range of safe mining depths under different confidence levels is generated. This is achieved through the Monte Carlo simulation method, which randomly samples geological parameters and model parameters to generate a large number of simulation scenarios, then runs the optimization algorithm in each scenario to obtain a series of optimal mining depths. Through statistical analysis of these results, the safe mining depth intervals under different confidence levels such as 95%, 90%, 80%, etc. are determined.

[0172] In addition, the risk probability associated with each mining depth is also calculated, i.e. the probability of exceeding the safety threshold when mining at that depth. This information is crucial for risk management and decision-making, allowing decision-makers to choose appropriate mining schemes according to their risk preferences.

[0173] Finally, the complete safe mining depth assessment results are output, including the safe mining depth intervals under different confidence levels and the corresponding risk probability distributions. These results are presented in the form of numerical reports and visual charts, providing a visual display of the relationship between mining depth and safety risk, and providing a scientific basis for mineral resource overburden risk assessment.

[0174] The introduction of fractional differential equations can better describe the memory effect and long-range correlation of geological materials, with an order usually between 0.5 and 1.5, and the most suitable parameter value is determined through field tests and inversion analysis. The design of the spiking neural network uses a leaky integral firing model, in which the neuron fires a pulse and resets the membrane potential when the membrane potential exceeds the threshold. This mechanism can effectively handle the strong temporal nature of geological evolution processes.

[0175] In the process of using fractional spiking differential equation neural networks with efficient adjoint parameter training, step S3.1 further includes: Step S3.1.1: For the state equation in forward propagation , construct the corresponding adjoint equation , where λ is the adjoint variable, L is the loss function, and θ is the network parameter, avoiding the gradient vanishing problem of traditional backpropagation on long time series.

[0176] In this step, an innovative solution, the adjoint method, is introduced to address the key challenge of gradient vanishing in neural network training. Traditional backpropagation algorithms, when dealing with long time series or deep networks, can easily lead to exponential decay of gradients due to the need for long chain differentiation of gradients, ultimately making it difficult for network parameters to be effectively updated. This problem is particularly prominent in geodynamic modeling, as long-term geological evolution processes often require the simulation of long time series.

[0177] First, we need to understand the state equation in forward propagation. The state equation is a differential equation that describes the change of state over time, which in neural networks can be represented as . Where x represents the state variable, which can be understood as the node activation value or hidden state in neural networks; θ represents the network parameters, including weights and biases, etc. learnable parameters; t represents the time variable; f is a nonlinear function that describes the rate of change of the state, usually determined by the network structure and activation function. In geodynamic modeling, the state variable x can represent the deformation, stress state or other key physical quantities of the geological body.

[0178] Traditional backpropagation algorithms compute the gradient of the loss function L with respect to the network parameters θ using the chain rule. For long time series, this means that the gradient needs to be backpropagated from the output layer through each time step to the initial state, forming a chain of computations that can be hundreds or even thousands of steps long. During this process, if the eigenvalues of the Jacobian matrix at each step are less than 1, the gradient decays exponentially with increasing time steps, leading to vanishing gradients; if the eigenvalues are greater than 1, it can lead to exploding gradients. Both cases can severely impact the training effectiveness of neural networks, especially when simulating long-term memory effects in geology.

[0179] The adjoint method provides an alternative strategy that avoids long chain derivatives. The adjoint method is derived from optimal control theory and variational methods and is widely used in data assimilation and parameter estimation problems in meteorology, oceanography, and other fields. The core idea is to construct an adjoint that directly computes the gradient of the objective function with respect to the parameters without step-by-step backpropagation.

[0180] In this context, the adjoint method first defines an adjoint variable λ, which has the same dimension as the original state variable x but has a different physical meaning. From a mathematical point of view, the adjoint variable λ can be understood as the sensitivity or importance index of the state variable x on the loss function. The dynamic evolution of the adjoint variable is described by the adjoint equation: .

[0181] In the adjoint equation, is the Jacobian matrix of the original state equation with respect to the state variable x, and its transpose reflects the mutual influence between state changes; represents the gradient of the loss function with respect to the state variable, reflecting the direct influence of the state variable on the final evaluation index. The negative sign indicates that the adjoint evolves from the termination time to the initial time, which is a key feature of the adjoint method.

[0182] The process of constructing the adjoint equation involves complex mathematical derivation. First, the discrete time step neural network dynamics is represented as a continuous time differential equation, establishing the state equation . Then, based on the variational principle, the influence of state changes on the final loss function is derived, obtaining the definition and dynamic equation of the adjoint variable λ. This derivation process ensures the mathematical equivalence of the adjoint method and the traditional backpropagation, but avoids the computational complexity of long chain derivatives.

[0183] It is worth noting that the adjoint equation looks simple in form, but contains rich information. The Jacobian matrix captures the internal dynamic structure and interdependence; the gradient term ensures the direct connection between the adjoint and the optimization objective; the negative sign and the transpose operation reflect the reverse nature of the adjoint, allowing it to efficiently transmit gradient information.

[0184] In practical computation, high-order numerical methods are employed to solve the adjoint equation. Since the adjoint equation is linear (although its coefficients are nonlinear), it is generally easier to solve compared to the original nonlinear state equations. Depending on the problem characteristics, appropriate numerical methods such as implicit Runge-Kutta methods or exponential integration methods are chosen to ensure the stability and accuracy of the numerical solution.

[0185] One key advantage of the adjoint method is that its computational complexity is independent of the time step size. Regardless of the length of the simulated time series, the adjoint method only requires one forward propagation and one backward solution to obtain the exact gradient of the loss function with respect to the parameters. This property enables efficient handling of long-term geological evolution processes, capturing memory effects and long-term correlations, providing a solid computational foundation for dynamic safe production depth evaluation.

[0186] By constructing the adjoint equation, the gradient vanishing problem of traditional backpropagation over long time series is successfully avoided, laying a foundation for the next step of parameter optimization. This method not only improves computational efficiency but also enhances the model's ability to capture long-term dependencies, which is of great significance for accurately simulating the dynamic response of geology.

[0187] Step S3.1.2: Efficiently calculate the gradient of the loss function with respect to the network parameters by solving the adjoint equation, achieving fast training and parameter optimization of the neural network.

[0188] In this step, based on the previously constructed adjoint equation, efficient optimization of neural network parameters is achieved. The core value of the adjoint method lies in directly providing the gradient information of the loss function with respect to the network parameters, avoiding the chain rule differentiation process in traditional backpropagation, thereby realizing more stable and efficient network training.

[0189] First, the adjoint equation needs to be solved to obtain the value of the adjoint variable λ. The adjoint equation is a linear ordinary differential equation, but needs to be solved from the terminal time to the initial time. This solving process starts from the final time point, λ(T) is initialized to zero or a specific value (depending on the form of the terminal condition), and then evolves step by step towards the initial time point.

[0190] Solving the adjoint equation requires special numerical integration methods. Since the adjoint equation is evolved in reverse, conventional forward integration methods are not applicable. Reverse self-adaptive step numerical methods such as reverse Runge-Kutta methods or reverse Adams methods are adopted to ensure the stability and accuracy of the numerical solution. These methods can automatically adjust the integration step size, ensuring computational accuracy while improving computational efficiency.

[0191] To further improve the solving efficiency, a checkpoint mechanism for the adjoint variables is implemented. During the forward propagation, the state information at some key time points is selectively stored as checkpoints for the backward solving of the adjoint equations. In this way, during the backward solving, only the intermediate states need to be recomputed from the nearest checkpoint, rather than from the beginning, greatly reducing the computational and storage overhead. The placement of checkpoints adopts a logarithmic interval strategy, which is particularly effective in cases with long time series.

[0192] Once the adjoint variables λ are obtained, the gradient of the loss function L with respect to the network parameters θ can be efficiently computed. According to the theory of the adjoint method, this gradient can be expressed as This integral expression reflects the cumulative effect of the parameters θ on the loss function L by affecting the dynamics f. In discrete implementations, the integral is transformed into a summation, computing the local contribution at each time step and then accumulating to obtain the total gradient.

[0193] It is worth noting that the above gradient computation formula directly provides global gradient information, avoiding the process of step-by-step gradient transmission in traditional backpropagation. This enables the gradient information to be accurately transmitted from the termination time to the initial time, without being limited by the eigenvalues of the intermediate step Jacobian matrices, effectively solving the problems of gradient vanishing and gradient explosion.

[0194] In practical applications, combined with automatic differentiation technology and parallel computing optimization, the efficiency and accuracy of gradient computation are further improved. Automatic differentiation technology generates gradient computation code automatically through computation graph analysis, avoiding the complexity and possible errors of manually deriving the Jacobian matrix. Parallel computing utilizes the multi-core or GPU acceleration capabilities of modern hardware to simultaneously compute the gradients of multiple time steps or multiple parameters, significantly improving the computation speed.

[0195] After obtaining the gradient information, advanced optimization algorithms are used to update the network parameters. Considering the complexity and non-convexity of the geological model, adaptive learning rate optimizers such as Adam (Adaptive Moment Estimation) or RMSProp (Root Mean Square Propagation) are used. These optimizers can automatically adjust the learning rate of each parameter based on the historical information of the gradient, maintaining the stability of the training while accelerating the convergence.

[0196] To further improve the training effect, batch processing and learning rate scheduling strategies are implemented. The batch processing strategy divides the training data into small batches, and each update uses the gradient computed from one batch of data, increasing the randomness and generalization ability of the training. Learning rate scheduling automatically adjusts the global learning rate according to the training progress, usually using a larger learning rate in the early stage of training to quickly explore the parameter space, and using a smaller learning rate in the later stage to fine-tune the parameters, achieving better convergence effect.

[0197] Regularization techniques are also introduced to prevent overfitting of the model. Regularization methods include L1 / L2 regularization (limiting parameter size by adding parameter norm terms in the loss function), Dropout (randomly turning off some neurons to enhance the robustness of the model), and Early Stopping (monitoring the performance of the validation set and stopping training before overfitting begins). These techniques ensure the generalization ability of the model on unseen data, which is crucial for the reliability of geological simulation.

[0198] To evaluate the training effect and parameter optimization progress, comprehensive evaluation indicators and visualization tools are designed. Evaluation indicators include Mean Squared Error (MSE), Mean Absolute Error (MAE), and Determination Coefficient (R 2 ) and other indicators, which measure the accuracy of model prediction from different angles. Visualization tools display the trend of loss function changes, key parameter distribution, and the comparison between predicted results and true values in real time, helping researchers intuitively understand the training process and model performance.

[0199] The gradient is calculated by the adjoint method and the network parameters are optimized, realizing the fast training of neural network. Compared with traditional back propagation, the adjoint method shows significant advantages in processing long time series, with training speed improved by several times to several tens of times, while avoiding the problem of gradient vanishing, ensuring the effective update of parameters. This efficient training method enables the processing of more complex geological models and longer time series, laying a foundation for accurately simulating the long-term dynamic behavior of geology.

[0200] Finally, the optimized neural network can accurately capture the memory effect and long-range correlation of geological materials, realizing accurate prediction of surface deformation. The network parameters are fully trained and encode the complex dynamic behavior of geology, enabling the model to reliably assess the risk probability under different mining depth conditions based on geological structure characteristics and mining parameters, providing a scientific basis for safe mining depth decision-making.

[0201] This neural network training technology based on the adjoint method not only applies to the current safe mining depth evaluation task, but also can be widely applied to other geological engineering fields such as groundwater simulation, oil and gas reservoir characterization, and earthquake risk assessment, with broad application prospects and important scientific value.

[0202] In step S4, an evaluation report generation module is constructed, including core risk point analysis, key technology demonstration, and forward-looking risk warning, which includes: Step S4.1: Based on the safe mining depth interval and risk probability, the evaluation area is divided into different risk areas by clustering analysis, including high-risk areas, medium-risk areas, and low-risk areas. Extract the geological parameters, calculation model parameters, and evaluation results of each risk area to form a risk source archive.

[0203] In this step, the entire evaluation area is risk-zoned and ranked based on the previously calculated safe mining depth intervals and risk probabilities. Risk zoning is a critical component of the pressure-overburden risk assessment, aiming to transform the spatially continuous risk indicators into discrete risk level regions for ease of management decision-making and risk control. This process involves multiple technical steps such as data processing, clustering analysis, and risk profile establishment.

[0204] The safe mining depth interval refers to the range of depths within which mineral resources can be safely mined under a specific risk threshold. It typically consists of a lower limit and an upper limit, reflecting the uncertainty of geological conditions and mining technology. For example, a safe mining depth interval for a certain area might be 350 meters to 400 meters, indicating that depths greater than 400 meters are considered safe, while depths less than 350 meters are at higher risk, and depths between 350 meters and 400 meters are in a transitional state. The calculation of the safe mining depth interval is based on the dynamic safe mining depth evaluation model established in the previous step, taking into account factors such as geological structure characteristics, rock mass mechanical properties, and the sensitivity of surface buildings.

[0205] Risk probability is a key indicator for quantifying pressure-overburden risk, representing the probability of surface buildings experiencing excessive deformation or damage at a specific mining depth. The calculation of risk probability is based on methods such as Monte Carlo simulation or probability distribution fitting, which transfer the uncertainty of model parameters to the final results, forming a probabilistic expression. For example, the risk probability for a certain area at a mining depth of 300 meters might be 0.15, indicating a 15% chance of excessive deformation under that condition. Risk probability provides a quantitative basis for risk zoning, making the risk assessment results more objective and reliable.

[0206] Clustering analysis is an unsupervised machine learning method used to divide data points into groups (clusters) with similar characteristics. In pressure-overburden risk assessment, clustering analysis divides spatial location points into high, medium, and low risk level regions based on multi-dimensional risk features. The risk features used include: the upper and lower limits of the safe mining depth interval, the distribution of risk probability at different mining depths, geological structure complexity indicators, rock mass strength parameters, and surface building sensitivity scores. These features form a high-dimensional feature space, with each spatial location point corresponding to a point in the feature space, and the clustering algorithm identifies natural groupings in the feature space.

[0207] The K-means clustering algorithm is mainly used for risk zoning. K-means is a classic distance-based clustering method that divides data points into a predetermined number of clusters through iterative optimization, so that the similarity of points in the same cluster is high, and the similarity of points between different clusters is low. The algorithm process includes: (1) initialize K cluster centers (K=3 in this case, corresponding to high, medium and low risk levels); (2) assign each data point to the nearest cluster center; (3) recalculate the center position of each cluster; (4) repeat steps 2 and 3 until the cluster center position is basically stable or the maximum number of iterations is reached. K-means algorithm is simple and efficient, suitable for processing large-scale spatial data, but sensitive to the selection of initial cluster centers.

[0208] To improve the stability and reliability of the clustering results, the K-means++ initialization method is used. Unlike traditional K-means that randomly selects initial cluster centers, K-means++ uses weighted probability selection to make the distance between initial cluster centers as far apart as possible, reducing the possibility of the algorithm falling into local optimum. The specific steps are: (1) randomly select a data point as the first cluster center; (2) calculate the square of the distance between each data point and the nearest selected cluster center; (3) set the selection probability based on the square of the distance, the farther the distance, the greater the probability of being selected as the next cluster center; (4) repeat steps 2 and 3 until K cluster centers are selected. This initialization method significantly improves the stability and clustering quality of K-means.

[0209] To further ensure the robustness of the clustering results, the optimal strategy of multiple runs is implemented. Specifically, the K-means++ algorithm is initialized with different random seeds and run multiple times (usually 10 to 20 times), and then the result with the smallest total distance square sum (i.e. the sum of the square of the distance between points in a cluster and the cluster center) is selected as the final clustering scheme. This strategy effectively reduces the uncertainty brought by random initialization, improves the repeatability and reliability of the clustering results.

[0210] In some special cases, K-means may not be suitable for handling non-spherical distribution or uneven density data. For this case, the density clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is also implemented as an alternative method. DBSCAN defines clusters based on density connectivity, can identify clusters of arbitrary shape, and can automatically identify noise points. When the geological conditions are particularly complex and the risk distribution presents an irregular shape, the DBSCAN algorithm will automatically switch to ensure the rationality and applicability of risk zoning.

[0211] After clustering, the evaluation area is divided into three levels: high-risk areas, medium-risk areas, and low-risk areas. High-risk areas usually have shallow safe mining depths, high risk probabilities, complex geological structures, or sensitive surface buildings; medium-risk areas have medium-level risk characteristics; low-risk areas have deep safe mining depths, low risk probabilities, simple geological structures, or insensitive surface buildings. The risk zoning results are visualized through GIS (Geographic Information) to form an intuitive risk zoning map, providing a spatial reference for decision-making.

[0212] For each risk area divided, relevant parameters are automatically extracted to form a risk source profile. The risk source profile is a description of the characteristics of the risk area, containing the following key information: (1) spatial range and coordinates, describing the geographical location and boundaries of the risk area; (2) geological parameters, including lithology composition, stratigraphic structure, tectonic characteristics (such as faults, folds), hydrogeological conditions, etc.; (3) engineering parameters, including mining methods, mining depth, mining sequence, support method, etc.; (4) calculation model parameters, including boundary conditions, material parameters (such as elastic modulus, Poisson's ratio, fractional order parameters), initial conditions, numerical method settings, etc.; (5) evaluation results, including safe mining depth interval, risk probability distribution, predicted surface deformation and its spatio-temporal distribution; (6) risk level determination basis, explaining the reason why the area is divided into a specific risk level; (7) recommended prevention and control measures, specific risk management recommendations for the risk area.

[0213] The risk source profile is stored in a structured data format, such as JSON (JavaScript Object Notation) or XML (eXtensible Markup Language), facilitating subsequent queries, analysis, and knowledge mining. For high-risk areas, more detailed profiles are generated, containing more parameters and more in-depth analysis; for low-risk areas, a relatively simplified profile format is used to reduce redundant information. This differentiated processing improves efficiency while ensuring the integrity of key risk information.

[0214] The risk source profile not only records the static risk state, but also contains the dynamic risk evolution trend. Based on the dynamic safe mining depth evaluation model in the previous steps, the risk state at different time points is predicted to form a dynamic description of the risk over time. This is of great significance for assessing the cumulative effects of long-term mining activities and developing phased risk management strategies. For example, a certain area may be low-risk in the early stages of mining, but as mining activities continue, it may gradually change to medium-risk or high-risk areas, requiring corresponding adjustments to risk management measures.

[0215] After the risk source profile is formed, it is integrated into the overburden risk assessment knowledge base to achieve knowledge accumulation and sharing. The knowledge base uses a semantic network structure to organize and establish the correlation between risk factors, risk manifestations, risk assessment methods, and risk control measures, supporting complex knowledge retrieval and reasoning. As projects accumulate, the knowledge base is continuously expanded and optimized, gradually forming a comprehensive risk knowledge system covering different geological conditions, different mining methods, and different building types, providing valuable experience for future overburden risk assessment.

[0216] Step S4.2: According to the risk source profile, using natural language generation technology, combining expert knowledge base and engineering case base, matching the corresponding analysis method and argumentation logic for each risk type, automatically generating the intelligent overburden risk assessment report.

[0217] In this step, based on the previously established risk source profile, combined with professional knowledge and experience cases, an intelligent overburden risk assessment report is automatically generated. This process involves multiple technical links such as natural language generation, knowledge retrieval, logical reasoning, and report layout, ultimately forming an assessment report that meets professional standards and has practical value.

[0218] Natural Language Generation (NLG) is an important branch of artificial intelligence, focusing on converting structured data into human-readable natural language text. In the generation of overburden risk assessment reports, NLG technology converts structured information in the risk source profile into fluent and professional text descriptions. The basic process of NLG includes content planning (determining the information to be expressed), document planning (determining the organization structure of information), sentence planning (designing sentence structure and word selection), and language implementation (generating the final text). Using a combination of templates and rules in NLG methods ensures the professionalism and standardization of the text, while providing sufficient flexibility to adapt to different risk situations.

[0219] The expert knowledge base is an important knowledge base for generating reports. The expert knowledge base contains theoretical knowledge, technical specifications, expert experience, and best practices in the field of overburden risk assessment, stored in a structured form. The knowledge base content covers concept definitions (such as "overburden", "safe mining depth", "risk probability", etc.), theoretical models (such as "fractional order continuum mechanics", "spiking neural network", etc.), assessment methods (such as "numerical simulation", "statistical analysis", etc.), judgment criteria (such as deformation limit value, risk threshold value, etc.), and decision rules (such as management measures under different risk levels). The organization of knowledge uses the ontology method to establish a semantic relationship network between concepts, supporting complex knowledge reasoning and retrieval.

[0220] The engineering case library is a collection and organization of historical engineering projects, containing a large number of actual overburden risk assessment cases. Each case records project background, geological conditions, assessment methods, assessment results, implementation measures, and effect feedback information. Cases are classified by multiple dimensions such as geological type (e.g. "fault impact area", "soft interlayer area", etc.), engineering type (e.g. "coal mining", "metal mining", etc.), risk type (e.g. "geological structure risk", "mining technology risk", etc.), etc. for easy retrieval and application. The value of engineering cases lies in the accumulation of practical experience, which can provide reference for current projects.

[0221] Risk type refers to different nature and cause of overburden risk, reflecting the essential characteristics and influence mechanism of risk. Common risk types include: (1) Geological structure risk, related to geological structures such as faults and folds, such as fault conduction risk, stress concentration risk, etc.; (2) Rock mass property risk, related to rock mass strength and deformation characteristics, such as soft interlayer risk, high pressure water risk, etc.; (3) Mining technology risk, related to mining methods and sequence, such as over-mining risk, insufficient safety coal pillar risk, etc.; (4) Surface building risk, related to surface building type and sensitivity, such as ancient building risk, high-rise building risk, etc.; (5) Cumulative effect risk, related to the cumulative impact of long-term mining activities, such as multiple mining influence superposition risk, etc. Different risk types have different forms of expression and influence mechanisms, and require targeted analysis methods and evaluation standards.

[0222] For each risk type, a corresponding analysis method and argumentation logic are matched. Analysis method refers to the technical path and tool for assessing risk, including qualitative analysis (such as expert judgment, case comparison, influence factor analysis, etc.) and quantitative analysis (such as numerical simulation, statistical analysis, reliability analysis, etc.). Argumentation logic refers to the reasoning framework and argumentation structure in the analysis process, ensuring the coherence of the analysis process and the reliability of the conclusion. Through semantic matching algorithm, the risk characteristics of the current project are compared with the content in the expert knowledge base and the engineering case library, the most similar knowledge points and cases are found, and the applicable analysis methods and argumentation logic are extracted.

[0223] The matching process is based on semantic similarity calculation, involving natural language processing technology. First, the key information in the risk source archive is converted into a semantic vector, then the cosine similarity with the items in the knowledge base and case library is calculated, and the most relevant content is identified. For example, for the risk of fault-affected area, the theoretical knowledge and analysis methods related to fault in the knowledge base are retrieved, and engineering cases under similar geological conditions in the case library are searched, forming a comprehensive analysis framework. To improve the matching accuracy, weighted similarity calculation is used, giving higher weight to key features such as fault type, fault throw, and fracture zone width, to ensure the relevance and applicability of the matching results.

[0224] Based on matching analysis methods and logical reasoning, technical analysis paragraphs are automatically generated. These paragraphs are presented in natural language and include descriptions of risk characteristics, analysis of risk causes, assessment of risk impact, and recommendations for risk control. Paragraph generation combines template filling and dynamic text generation, ensuring both the completeness and professionalism of the content while providing sufficient flexibility to adapt to different risk scenarios. A rich built-in language template library covers common expression structures and terminology usage for various risk analyses, ensuring the standardization and readability of the generated text.

[0225] Chart generation is a crucial part of report production. Based on risk source files and analysis results, various professional maps are automatically created, such as geological profiles, risk distribution maps, safe mining depth contour maps, and surface subsidence prediction maps. The generated maps adhere to professional cartographic standards and color schemes, ensuring their scientific accuracy and intuitiveness. Interactive chart functionality is also supported, allowing users to view the 3D geological model and risk distribution from different angles through zooming, panning, and rotating, enhancing the report's information delivery. Furthermore, a text-image association feature is implemented; when a user views a specific chart, relevant text descriptions are automatically highlighted, improving report readability and user experience.

[0226] In terms of report structure, a standardized template framework was adopted to ensure the completeness and logic of the content. A typical overburden risk assessment report includes the following main chapters: (1) Project Overview, introducing the project background, assessment purpose and scope; (2) Geological Condition Analysis, describing the regional geological structure, stratigraphic distribution and hydrogeological conditions; (3) Assessment Methodology Introduction, explaining the technical route, model methods and data sources used; (4) Safe Mining Depth Calculation, showing the calculation process and results of the safe mining depth range; (5) Risk Zoning and Classification, showing the risk zoning map and risk characteristics of each area; (6) Risk Analysis, conducting technical analysis for different risk types; (7) Countermeasures and Recommendations, proposing risk management and control measures; (8) Conclusion, summarizing the main findings and recommendations of the assessment. This structural framework not only conforms to industry standards, but also facilitates decision-makers to quickly obtain key information.

[0227] To ensure report quality, a multi-layered quality control mechanism was implemented. First, a data consistency check was conducted to ensure that the data cited in the report matched the calculation results. Second, a content completeness check was performed to ensure that all necessary content was included in the report. Finally, a professional reasonableness check was conducted, simulating a professional review process to assess the reasonableness of the report's key conclusions and recommendations. These quality control measures ensured the accuracy and reliability of the generated report.

[0228] Ultimately, a complete intelligent overburden risk assessment report was produced. This report not only complies with industry standards and technical specifications but also possesses targeted and practical value, providing a scientific basis for engineering decisions. The report's risk identification, analysis, and evaluation results are clear and explicit, and the recommended risk control measures are specific and feasible, guiding subsequent engineering implementation and management decisions to ensure the coordinated development of surface construction projects and underground mineral resource development.

[0229] The generation of the intelligent overburden risk assessment report marks the successful completion of the entire assessment process. Compared with traditional manually compiled reports, intelligent reports have advantages such as high generation efficiency (reducing the time from several days to several hours), comprehensive content, in-depth analysis, and professional expression, greatly improving the efficiency and technical level of overburden risk assessment. More importantly, intelligent reports can be quickly updated based on new data and circumstances, achieving dynamic assessment and continuous optimization, providing strong technical support for risk management in complex and changing environments.

[0230] like Figure 3 As shown, the present invention also provides a mineral resource overburden risk assessment system, comprising: The multi-source data fusion module 501 is used to acquire multi-source heterogeneous geological data, including geological exploration data, borehole core data, geophysical exploration data, historical assessment cases and mining data of surrounding mines. The multi-source heterogeneous geological data is deeply fused using a robust neural network training method with game theory gradient control to obtain a fused dataset. The geological structure identification module 502 is used to perform in-depth mining and reasoning on high-dimensional geological data based on the fused dataset, utilizing the pattern recognition and correlation analysis capabilities of the large model, combined with manifold perception regularization technology, to identify geological structure features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures. The safe mining depth calculation module 503 is used to establish a dynamic safe mining depth evaluation model using a fractional peak differential equation neural network with efficient adjoint parameter training. It integrates geological structural risks, mining disturbance effects and engineering safety thresholds to calculate the safe mining depth range and risk probability under different confidence levels. The report generation module 504 is used to construct an assessment report generation module that includes core risk point analysis, key technology demonstration and forward-looking risk warning based on the safe mining depth range and risk probability, and automatically generate an intelligent overburden risk assessment report.

[0231] The various modules of this system work collaboratively to form a complete mineral resource overburden risk assessment process. The multi-source data fusion module is responsible for data acquisition and preprocessing, and achieves deep fusion of heterogeneous data through a robust neural network with game theory gradient control; the geological structure identification module identifies complex geological structure features based on the fused dataset and combined with manifold perception regularization technology; the safe mining depth calculation module uses a fractional peak differential equation neural network to establish a dynamic safe mining depth evaluation model; and the report generation module is responsible for the automatic generation of intelligent overburden risk assessment reports.

[0232] This invention uses a wind farm construction project as an example to detail the implementation process of mineral resource overburden risk assessment. The wind farm is located above a coal mine area, requiring an assessment of the mutual impact between wind farm construction and underground coal seam mining to determine the safe mining depth and risk level.

[0233] First, the multi-source data fusion module collected geological exploration data, borehole core data, geophysical exploration data, historical assessment cases, and mining data from surrounding mines in the region. After preprocessing these heterogeneous data, a robust neural network training method with game theory gradient control was used for deep fusion. Under Nash equilibrium, the system automatically determined the optimal contribution weights of each data source, with borehole core data having a weight of 0.35, geophysical exploration data 0.25, geological exploration data 0.20, historical assessment cases 0.15, and surrounding mine mining data 0.05. Through forward and backward propagation of the robust neural network, a high-quality fused dataset was finally obtained.

[0234] Then, the geological structure identification module, based on the fused dataset and combined with manifold-aware regularization technology, performs in-depth mining and inference on high-dimensional geological data. The system first constructs a k-nearest neighbor graph (k=5), establishes an adjacency matrix between feature points, and obtains the manifold structure representation of the data. Based on the manifold structure representation, the local density and curvature information of feature points in the manifold space are calculated, and a manifold Laplacian matrix is ​​constructed as a regularization term to constrain the data's intrinsic geometric structure during large model training. Through a large model with a Transformer architecture and a multi-head self-attention mechanism, the system successfully identified the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fractures in the region. The identification results show that there are two main faults in the region, trending NE45° and dipping 75°, with fracture zones 2-5 meters wide; two sets of main joints are developed in the shale strata above the coal seams, trending NE and NW respectively, with a spacing of 0.5-2 meters and moderate connectivity.

[0235] Next, based on the identified geological structural features, the safe mining depth calculation module established a dynamic safe mining depth evaluation model using a fractional peak differential equation neural network with efficient adjoint parameter training. This model uses fractional differential equations (order 0.8) to describe the memory effect of rock mass deformation, and the peak neural network designed through a leakage integral distribution model effectively captures the dynamic response characteristics during the mining process. The system established a multi-constraint optimization model, with constraints including a maximum surface settlement not exceeding 10 mm, a maximum inclination not exceeding 2 mm / m, and a maximum curvature deformation not exceeding 0.2 mm / m. 2 By solving the constrained optimization problem using the particle swarm optimization algorithm, the safe mining depth ranges under different confidence levels were obtained: 380-420 meters at 95% confidence level, 350-450 meters at 90% confidence level, and 320-480 meters at 80% confidence level.

[0236] Finally, the report generation module automatically generated an intelligent overburden risk assessment report based on the safe mining depth range and risk probability. The system used cluster analysis to divide the assessment area into high-risk areas (near fault fracture zones, accounting for approximately 15% of the total area), medium-risk areas (joint-developed areas, accounting for approximately 35% of the total area), and low-risk areas (structurally stable areas, accounting for approximately 50% of the total area). For each risk area, the system extracted geological parameters, calculation model parameters, and assessment results to form a risk source profile. Based on the risk source profile, the system used natural language generation technology, combined with an expert knowledge base and engineering case library, to match corresponding analysis methods and argumentation logic for each risk type, ultimately generating a complete assessment report including core risk point analysis, key technology argumentation, and forward-looking risk warnings. The report concluded by recommending that wind farm tower foundation locations avoid high-risk areas, reinforcement measures be taken in medium-risk areas, and coal seam mining depth be controlled below 400 meters to ensure the safety of wind farm construction and coal mining.

[0237] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method of assessing the risk of mineral resource overburden, characterised in that, The method comprises the following steps: obtaining multi-source heterogeneous geological data including geological exploration data, drilling core data, geophysical exploration data, historical evaluation cases and surrounding mine exploitation data, using a robust neural network training method based on gradient control of game theory to deeply fuse the multi-source heterogeneous geological data to obtain a fused data set; based on the fused data set, using the pattern recognition and correlation analysis capability of a large model, combining manifold perception regularization technology to deeply mine and reason high-dimensional geological data, and identifying geological structure features including stratigraphic interface, lithological distribution, fault fracture zone and joint fissure; based on the geological structure features, using a fractional spike differential equation neural network with efficient adjoint parameter training to establish a dynamic safe mining depth evaluation model, comprehensively considering geological structure risk, mining disturbance effect and engineering safety threshold, and calculating safe mining depth intervals and risk probabilities under different confidence levels; according to the safe mining depth intervals and risk probabilities, constructing an evaluation report generation module including core risk point analysis, key technology demonstration and forward-looking risk warning, and automatically generating an intelligent overburden risk evaluation report.

2. The method of claim 1, wherein, The robust neural network training method based on gradient control of game theory is used to deeply fuse the multi-source heterogeneous geological data to obtain a fused data set, which comprises: the multi-source heterogeneous geological data is taken as a game participant, Nash equilibrium theory is used, and each data source is iteratively solved to reach an optimal weight distribution state to determine the contribution weight of each data source to the final fusion result, and an initial weight distribution is obtained; based on the initial weight distribution, a robust neural network architecture including residual connection and batch normalization is constructed, Gaussian noise and data enhancement strategy are introduced to improve the network anti-interference ability, the optimal contribution weight of each data source under the Nash equilibrium state is calculated, and an optimal weight distribution state is obtained; each data source of the multi-source heterogeneous geological data is weighted and linearly combined according to the optimal weight distribution state, and fusion training is completed through neural network forward propagation and back propagation to obtain the fused data set.

3. The method of claim 2, wherein, The robust neural network architecture including residual connection and batch normalization comprises: based on the initial weight distribution, the preprocessed multi-source heterogeneous geological data is taken as the input layer of the neural network, a residual connection module is set in the neural network, batch normalization technology is used to standardize the output of each layer to ensure stable training of the deep network; Gaussian noise disturbance and data enhancement strategy are introduced to the input layer of the neural network, including random rotation, scaling and translation transformation, to enhance the robustness of the network to data disturbance.

4. The method of claim 1, wherein, The manifold perception regularization technology is combined to deeply mine and reason high-dimensional geological data to identify geological structure features including stratigraphic interface, lithological distribution, fault fracture zone and joint fissure, which comprises: based on the geological feature points in the fused data set, a k-nearest neighbor graph is constructed and the k nearest neighbors of each feature point are calculated to establish an adjacency relationship matrix between feature points and obtain the manifold structure representation of the data; Based on the manifold structure representation of the data, the local density and curvature information of the feature points in the manifold space are calculated, a manifold Laplacian matrix is constructed as a regularization term to constrain the intrinsic geometric structure of the data in the large model training process, and a regularization constraint parameter is obtained; The regularization constraint parameter is added to the loss function of the large model based on the Transformer architecture, and the model parameters are optimized through the multi-head self-attention mechanism and the back propagation algorithm to identify the spatial distribution of the stratigraphic interface, lithology distribution, fault fracture zone and joint fissure, and obtain the geological structure characteristics.

5. The method of claim 4, wherein, The identification of the spatial distribution of the stratigraphic interface, lithology distribution, fault fracture zone and joint fissure comprises: Analyzing the lithology change sequence characteristics of the drilling core data and the resistivity anomaly and magnetic anomaly information of the geophysical exploration data in the fusion data set, using a sequence pattern matching algorithm to automatically identify the depth and occurrence parameters of the main stratigraphic interface, and obtaining stratigraphic interface distribution information; Comprehensively analyzing the core crushing degree in the borehole, the low resistance anomaly in the geophysical exploration and the structure record in the historical geological data, using a support vector machine classification algorithm to predict the trend, tendency and fracture zone influence range of the fault, and obtaining fault fracture zone distribution information; Using a graph neural network to model the joint fissure system as a graph structure, taking the intersection point of the fissure as a node and the connection of the fissure as an edge, learning the spatial distribution rule and connectivity characteristics of the fissure network through a graph convolutional neural network, and obtaining the joint fissure development rule; Integrating the stratigraphic interface distribution information, the fault fracture zone distribution information and the joint fissure development rule, a regional three-dimensional geological structure model is constructed, and the geological structure characteristics are obtained.

6. The method of claim 1, wherein, The establishment of the dynamic safe mining depth evaluation model using the fractional spike differential equation neural network with efficient adjoint parameter training comprises: Based on the geological structure characteristics, the memory effect and long-range correlation of geological materials are described by a fractional differential equation, a spike neural network is designed using a leaky integral firing model, and a dynamic safe mining depth evaluation model is established; Based on the dynamic safe mining depth evaluation model, a multi-constraint optimization model is established, the constraint conditions include that the maximum subsidence, maximum inclination and maximum curvature deformation on the ground surface do not exceed the safety threshold, and a particle swarm optimization algorithm is used to solve the constraint optimization problem to obtain the safe mining depth interval and risk probability.

7. The method of claim 6, wherein, The use of the fractional spike differential equation neural network with efficient adjoint parameter training comprises: For the state equation in the forward propagation , the corresponding adjoint equation is constructed , where λ is the adjoint variable, L is the loss function, and θ is the network parameter, avoiding the gradient vanishing problem of traditional back propagation on long time series; By solving the adjoint equation, the gradient of the loss function with respect to the network parameters is efficiently calculated, and the fast training and parameter optimization of the neural network are realized.

8. The method of claim 1, wherein, The construction of the evaluation report generation module containing core risk point analysis, key technology demonstration and forward-looking risk warning comprises: Based on the safe mining depth interval and risk probability, a clustering analysis is used to divide the evaluation area into different levels of risk areas, including high-risk areas, medium-risk areas and low-risk areas, geological parameters, calculation model parameters and evaluation results of each risk area are extracted, and a risk source archive is formed; According to the risk source archive, a natural language generation technique is adopted, combined with an expert knowledge base and an engineering case base to match corresponding analysis methods and demonstration logic for each risk type, and an intelligentized overburden pressure risk assessment report is automatically generated.

9. A mineral resource pressure cover risk assessment system, characterized by, Comprise: A multi-source data fusion module is configured to obtain multi-source heterogeneous geological data, including geological exploration data, drilling core data, geophysical exploration data, historical assessment cases, and surrounding mine exploitation data, and a robust neural network training method based on gradient control of game theory is used to deeply fuse the multi-source heterogeneous geological data to obtain a fused data set; A geological structure identification module is configured to use the pattern recognition and correlation analysis capabilities of a large model to deeply mine and reason high-dimensional geological data based on the fused data set, combined with manifold perception regularization technology, to identify geological structure characteristics including stratigraphic interfaces, lithological distribution, fault fracture zones, and joint fissures; A safe mining depth calculation module is configured to use a fractional spike differential equation neural network with high-efficiency companion parameter training to establish a dynamic safe mining depth evaluation model, and to calculate safe mining depth intervals and risk probabilities under different confidence levels by comprehensively considering geological structure risks, mining disturbance effects, and engineering safety thresholds; A report generation module is configured to construct an evaluation report generation module containing core risk point analysis, key technology demonstration, and forward-looking risk warning based on the safe mining depth intervals and risk probabilities, and to automatically generate an intelligentized overburden pressure risk assessment report.

Images