A mineral resource overburden risk assessment method and system

Abstract

The present application provides a kind of mineral resources overburden risk assessment method and system, comprising: obtaining multi-source heterogeneous geological data, the robust neural network training method of game theory gradient control is used to the depth fusion of the multi-source heterogeneous geological data, and the fusion data set is obtained;With the pattern recognition and correlation analysis ability of large model, combined with manifold perception regularization technique, high-dimensional geological data is deeply mined and reasoned to identify geological structure characteristics;A dynamic safe depth evaluation model is established using a fractional spike differential equation neural network with efficient adjoint parameter training, which integrates geological structure risk, mining disturbance effect and engineering safety threshold to calculate the safe mining depth interval and risk probability under different confidence levels;An evaluation report generation module is constructed, which includes core risk point analysis, key technology demonstration and forward-looking risk warning, to generate an intelligent overburden risk assessment report. The present application can improve the accuracy and reliability of overburden risk assessment.

Description

Technical Field

[0001] This invention relates to the field of mineral resource development, specifically to a method and system for assessing the risk of mineral resource overlay, used to evaluate the degree of mutual influence between surface construction projects and underground mineral resources, determine the safe mining depth and risk level, and provide a scientific basis for mineral resource protection and engineering safety. Background Technology

[0002] Mineral resource overlay risk assessment is a crucial technical field for ensuring the safety of surface 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 national requirements for mineral resource protection and construction project safety, the scientific rigor and accuracy of overlay risk assessment technology are receiving growing attention.

[0003] Traditional methods for assessing the risk of mining overburden mainly rely on three-dimensional geological modeling techniques and empirical formulas. A 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 uses a geomechanical model to assess the impact of mining on surface structures.

[0004] The most mainstream approach in current technology is to construct a three-dimensional geological model using geostatistical methods such as Kriging interpolation, then calculate surface subsidence using probability integral methods or influence function methods, and finally determine the safe mining depth based on standards and specifications. This method establishes a continuous geological interface by spatially interpolating discrete borehole data, and calculates surface deformation using empirical formulas based on rock mechanics parameters and mining geometry parameters.

[0005] However, existing technologies have two key problems: First, the depth of geological risk identification is insufficient. Traditional modeling methods are unable to fully explore the complex correlations between multi-source data and cannot deeply reveal the impact 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 dynamics of mining disturbances, resulting in limited accuracy and reliability of the assessment conclusions.

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

[0007] The purpose of this invention is to provide a method and system for assessing the risk of mineral resource overburden. By deeply integrating multi-source heterogeneous geological data, intelligently identifying complex geological structural features, and dynamically evaluating safe mining depth, this method improves the accuracy and reliability of overburden risk assessment and solves 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 objectives, the present invention provides a method for assessing the risk of mineral resource overlay, comprising the following steps:

[0009] Acquire multi-source heterogeneous geological data, including geological exploration data, borehole core data, geophysical exploration data, historical assessment cases, and mining data from surrounding mines. Use a robust neural network training method with game theory gradient control to deeply fuse the multi-source heterogeneous geological data to obtain a fused dataset.

[0010] Based on the fused dataset, the pattern recognition and association analysis capabilities of the large model are utilized, combined with manifold perception regularization technology, to perform in-depth mining and reasoning on high-dimensional geological data, and to identify geological structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures.

[0011] Based on the geological structural features, a dynamic safe mining depth evaluation model is established using a fractional peak differential equation neural network with efficient training of adjoint parameters. The model 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.

[0012] Based on the safe mining depth range and risk probability, an assessment report generation module is constructed that includes core risk point analysis, key technology demonstration and forward-looking risk warning, and automatically generates an intelligent overburden risk assessment report.

[0013] Furthermore, the robust neural network training method employing game theory gradient control is used to deeply fuse the multi-source heterogeneous geological data to obtain a fused dataset, including:

[0014] Using the multi-source heterogeneous geological data as game participants, the Nash equilibrium theory is adopted, and the optimal weight allocation of each data source is achieved through iterative solution to determine the contribution weight of each data source to the final fusion result and obtain the initial weight allocation.

[0015] Based on the initial weight allocation, a robust neural network architecture including residual connections and batch normalization is constructed. Gaussian noise and data augmentation strategies are introduced to improve the network's anti-interference ability. The optimal contribution weight of each data source in the Nash equilibrium state is calculated to obtain the optimal weight allocation state.

[0016] 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 a neural network to obtain the fused dataset.

[0017] Furthermore, the construction of a robust neural network architecture including residual connections and batch normalization includes:

[0018] Based on the initial weight allocation, preprocessed multi-source heterogeneous geological data is used 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 stable training of the deep network.

[0019] Gaussian noise perturbation and data augmentation strategies, including random rotation, scaling and translation transformations, are introduced into the input layer of the neural network to enhance the network's robustness to data perturbation.

[0020] Furthermore, the method of combining manifold sensing regularization technology to perform in-depth mining and reasoning on high-dimensional geological data, identifying geological structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures, includes:

[0021] Based on the geological feature points in the fused dataset, a k-nearest neighbor graph is constructed and the k nearest neighbors of each feature point are calculated. An adjacency matrix between feature points is established to obtain the manifold structure representation of the data.

[0022] Based on the manifold structure representation of the data, the local density and curvature information of feature points in the manifold space are calculated, and the manifold Laplacian matrix is ​​constructed as a regularization term to constrain the data to maintain its intrinsic geometric structure during the training of large models, thereby obtaining regularization constraint parameters.

[0023] The regularization constraint parameters are added to the loss function of the large model based on the Transformer architecture. The model parameters are optimized by multi-head self-attention mechanism and backpropagation algorithm to identify the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures, thereby obtaining the geological structural features.

[0024] Furthermore, the identification of the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures includes:

[0025] The lithological variation sequence characteristics of borehole core data and resistivity and magnetic anomaly information of geophysical exploration data in the fused dataset are analyzed. The sequence pattern matching algorithm is used to automatically identify the burial depth and attitude parameters of the main stratigraphic interfaces and obtain stratigraphic interface distribution information.

[0026] By comprehensively analyzing the degree of core fragmentation in boreholes, low resistivity anomalies in geophysical exploration, and structural records in historical geological data, a support vector machine classification algorithm is used to predict the strike, dip, and influence range of fault fracture zones, thereby obtaining information on the distribution of fault fracture zones.

[0027] A graph neural network is used to model the joint and fracture system as a graph structure, with fracture intersections as nodes and fracture connections as edges. The spatial distribution and connectivity characteristics of the fracture network are learned through a graph convolutional neural network to obtain the development law of joints and fractures.

[0028] By integrating the stratigraphic interface distribution information, the fault fracture zone distribution information, and the joint and fissure development patterns, a regional three-dimensional geological structure model is constructed to obtain the geological structure characteristics.

[0029] Furthermore, the method of establishing a dynamic safe mining depth evaluation model using a fractional peak differential equation neural network with efficient adjoint parameter training includes:

[0030] Based on the geological structural characteristics, the memory effect and long-range correlation of geological materials are described by fractional differential equations, and a peak neural network is designed using a leakage integral distribution model to establish a dynamic safe mining depth evaluation model.

[0031] Based on the dynamic safe mining depth evaluation model, a multi-constraint optimization model is established. The constraints include that the maximum surface subsidence, maximum tilt, and maximum curvature deformation do not exceed the safety threshold. The particle swarm optimization algorithm is used to solve the constrained optimization problem to obtain the safe mining depth range and risk probability.

[0032] Furthermore, the fractional peak differential equation neural network trained with efficient adjoint parameters includes:

[0033] For the state equation in forward propagation Construct the corresponding adjoint equations , 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;

[0034] By solving the adjoint equation, the gradient of the loss function with respect to the network parameters can be efficiently calculated, enabling rapid training and parameter optimization of the neural network.

[0035] Furthermore, the module for generating an assessment report, which includes core risk point analysis, key technology demonstration, and forward-looking risk warning, includes:

[0036] Based on the safe mining depth range and risk probability, cluster analysis is used to divide the assessment area into risk areas of different levels, including high-risk areas, medium-risk areas and low-risk areas. Geological parameters, calculation model parameters and assessment results of each risk area are extracted to form a risk source file.

[0037] Based on the risk source file, natural language generation technology is used, combined with an expert knowledge base and an engineering case library, to match corresponding analysis methods and argumentation logic for each risk type, and automatically generate the intelligent overburden risk assessment report.

[0038] This invention also provides a mineral resource overlay risk assessment system, comprising:

[0039] The multi-source data fusion module 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.

[0040] The geological structure identification module 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 structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures.

[0041] The safe mining depth calculation module 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.

[0042] The report generation module 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.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention employs a robust neural network training method based on game theory gradient control to deeply fuse multi-source heterogeneous geological data, achieving optimal weight allocation and high-quality fusion of different types of geological data, thereby improving data utilization efficiency and fusion accuracy.

[0045] 2. This invention combines manifold perception regularization technology to perform in-depth mining and reasoning on high-dimensional geological data, which can accurately identify complex geological structural features such as stratigraphic interfaces, lithological distribution, fault fracture zones, joints and fissures, thereby improving the accuracy of geological risk identification.

[0046] 3. This invention uses a fractional peak differential equation neural network with efficient adjoint parameter training to establish a dynamic safe mining depth evaluation model, which can comprehensively consider geological structural risks, mining disturbance effects and engineering safety thresholds, thereby improving the accuracy and reliability of safe mining depth assessment.

[0047] 4. This invention constructs an assessment report generation module that includes core risk point analysis, key technology demonstration, and forward-looking risk warning, realizing the intelligent generation of overburden risk assessment reports and improving the professionalism and application value of the assessment results. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 A schematic diagram of the geographic geological information system provided by this invention;

[0050] Figure 2 Flowchart of the mineral resource overlay risk assessment method provided by the present invention;

[0051] Figure 3 The structural diagram of the mineral resource overburden risk assessment system provided by the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0053] like Figure 1 As shown, the Geographic Geological Information System (GGIS) revolves around building a digital platform that integrates geospatial and geological data. Its aim is to address issues such as fragmented data, inefficient analysis, and delayed decision-making in traditional geological work, serving the scientific and precise management of resource exploration, environmental monitoring, engineering construction, disaster prevention, and other fields. The system's full name is "Geographic Geological Information System for a Certain Region." Its necessity lies in achieving standardized integration and management of geographic and geological data, improving the efficiency and accuracy of geological analysis and research, and providing a scientific basis for resource exploration and development. (The references are not mentioned here.)

[0054] In terms of system overview, the core objective is to break down data barriers through "spatial integration, intelligent analysis, and visual representation," providing precise data support and decision-making tools for multiple fields, and maximizing the efficient utilization of geological information and its social value. Functional objectives include resource overview, resource distribution, and project-mineral resource collaboration, while non-functional objectives require a response time of ≤2 seconds and an annual availability of ≥99.9%. The system scope includes functions such as resource overview, regional tectonics, latest information, resource distribution, mineral deposit distribution, project distribution, and project-mineral deposit collaboration. The hardware environment supports Windows 10 and above, and the software environment uses Linux CentOS 7 on the server side, MySQL 8.0 as the database, Redis 6.0 as the middleware, and Chrome 90+ or ​​Edge 90+ as the browser.

[0055] In terms of functional requirements, the system is divided into seven modules based on business logic: Resource Overview, Regional Structure, Latest News, Resource Distribution, Mineral Deposit Distribution, Project Distribution, and Project-Mineral Deposit Collaboration. Each module has a clearly defined function: Resource Overview displays the overall resource distribution across the province; Regional Structure presents geological structure maps of various regions within the province; Latest News publishes relevant national policies and the latest information; Resource Distribution provides statistics on the distribution of mineral resources and the proportion of different mineral types in various cities and prefectures; Mineral Deposit Distribution displays the distribution of national and provincial mineral deposits and delineates mining rights and industrial rights areas; Project Distribution displays the distribution of wind farms, photovoltaic projects, and power transmission projects across the province; and Project-Mineral Deposit Collaboration presents a comprehensive location map of both to aid in comprehensive project analysis.

[0056] Non-functional requirements encompass multiple aspects. In terms of performance, a single user's resource distribution page loading time should be ≤1 second, supporting 10,000 users browsing simultaneously. Regarding security, user passwords should be stored using MD5 salting encryption, with no explicit access control, requiring protection against SQL injection and XSS attacks. For reliability, the system should automatically switch to a backup database without data loss in the event of a database failure, automatically backing up data daily at midnight, with backup files retained for 30 days. In terms of usability, the interface should be simple and easy to operate, with provided operation guide documentation. For scalability, a microservice architecture should be adopted, allowing newly added modules to be deployed independently to support future functional expansion.

[0057] The system architecture is divided into overall architecture and technical architecture. The overall architecture follows a hierarchical structure of front-end layer - API gateway layer - service layer - data layer, specifically: client layer (Web) → API gateway (responsible for routing and authentication) → business service layer (covering modules such as resource overview and resource distribution) → data layer (including MySQL, Redis, and file storage). The core technology stack of the technical architecture is clearly defined. The front-end uses the Vue3 + Vite + Cesium + AMAP + ECharts framework, and the back-end uses the Java language and Spring Boot framework.

[0058] The user interface design follows the principles of simplicity, ease of use, and consistent style. The test plan outlines the testing strategy, including unit testing, integration testing, and performance testing. The test environment must meet relevant requirements, and key test metrics include a functional test pass rate of ≥95% and performance testing meeting response time requirements.

[0059] The implementation and deployment plan clearly defines the system deployment steps as server environment configuration → database initialization → application deployment → data migration. The go-live process adopts a canary release → full release → monitoring and feedback model. Regarding maintenance and support, daily maintenance includes log cleanup and database backup checks. The fault handling procedure stipulates that when a 500 error occurs, the application logs should be checked first, followed by the database connection.

[0060] like Figure 2 As shown, the present invention provides a method for assessing the risk of mineral resource overlay, comprising the following steps:

[0061] Step S1: Obtain multi-source heterogeneous geological data, including geological exploration data, borehole core data, geophysical exploration data, historical assessment cases, and mining data from surrounding mines. Use a robust neural network training method with game theory gradient control to deeply fuse the multi-source heterogeneous geological data to obtain a fused dataset.

[0062] In this step, a comprehensive collection of multi-source heterogeneous geological data for the assessment area is first undertaken. Geological exploration data includes geological maps, drill columnar sections, and geological profiles, which record the spatial distribution, lithological characteristics, and basic structural features of the regional strata. Drill core data is the most direct source of subsurface geological information, including core photographs, lithological descriptions, records of structural features, and physical and mechanical parameter test results. Geophysical exploration data mainly includes electrical, magnetic, and gravity exploration data, which can provide information on the distribution of physical properties of subsurface geological bodies and help identify geological structures such as faults and fracture zones that are not easily discovered directly through drilling. Historical assessment cases and mining data from surrounding mines provide regional geological background and mining impact experience, including historical assessment reports, mining subsidence monitoring data, and records of disaster events.

[0063] After data collection, various data types undergo preprocessing to convert data of different formats and resolutions into a unified digital standard format. For geological maps, vectorization techniques are used to extract geological boundaries; for borehole data, a three-dimensional spatial database is constructed; and for geophysical data, coordinate transformation and data noise reduction are performed. The purpose of preprocessing is to eliminate data noise, unify data formats, and lay the foundation for subsequent deep fusion.

[0064] After data preprocessing, a robust neural network training method based on game theory gradient control is used to deeply fuse multi-source heterogeneous geological data. The core idea of ​​this method is to treat each data source as a player in a game, determine the optimal contribution weight of each data source using game theory principles, and then use a robust neural network to achieve deep data fusion. Compared to traditional data fusion methods, this method can adaptively adjust the weights of each data source, maximizing the retention of useful information, suppressing noise and redundancy, and generating a high-quality fused dataset.

[0065] Step S2: Based on the fused dataset, utilize the pattern recognition and association analysis capabilities of the large model, combined with manifold perception regularization technology, to perform in-depth mining and reasoning on the high-dimensional geological data, and identify geological structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, joints and fissures.

[0066] In this step, a large model is used to perform in-depth mining and inference on high-dimensional geological data based on the fused dataset. The large model refers to a deep learning model based on the Transformer architecture, which is pre-trained on a large corpus of geological literature, exploration reports, and geological maps, possessing powerful geological knowledge understanding and inference capabilities. The fused dataset is input into the large model, and its pattern recognition and association analysis capabilities are used to mine the implicit geological structural features within the data.

[0067] To improve the accuracy of geological structural feature recognition, a manifold-aware regularization technique was incorporated. This technique, based on manifold learning theory, assumes that high-dimensional geological data is distributed on a low-dimensional manifold. By constructing the geometric structure of the data, it constrains the model learning process to maintain the inherent geometric properties of the data. Specifically, a k-nearest neighbor graph of the data points is constructed, and the Laplacian matrix of the manifold is calculated and added as a regularization term to the model's loss function. This method effectively solves the structural distortion problem that may occur in traditional deep learning when processing high-dimensional geological data, thus improving the reliability of feature recognition.

[0068] Through in-depth analysis and reasoning using a large-scale model, the geological structural features of the assessment area were successfully identified, including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures. Stratigraphic interfaces reflect the superposition of strata from different geological periods, lithological distribution describes the spatial distribution patterns of various rock types, and fault fracture zones and joints and fissures are key structural elements affecting rock mass stability. These geological structural features provide fundamental information for subsequent safe mining depth assessments.

[0069] Step S3: Based on the geological structural features, a dynamic safe mining depth evaluation model is established using a fractional peak differential equation neural network with efficient training of adjoint parameters. The model 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.

[0070] In this step, a dynamic safe mining depth evaluation model was established based on the identified geological structural features. This model employs a fractional peak differential equation neural network, which can effectively simulate the propagation process of mining disturbances 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 peak neural network effectively handles the highly temporal geological evolution process by simulating the firing mechanism of biological neurons.

[0071] To improve the training efficiency of the model, an efficient adjoint parameter training algorithm was adopted. This algorithm avoids the gradient vanishing problem in long-term series caused by traditional backpropagation by constructing an adjoint equation, and can efficiently calculate the gradient of the loss function with respect to the network parameters, accelerating the model's convergence process. This algorithm is particularly suitable for handling geomechanical simulation problems involving long-term dependencies, significantly improving the model's training efficiency and prediction accuracy.

[0072] In the evaluation of safe mining depth, geological structural risks, mining disturbance effects, and engineering safety thresholds were comprehensively considered. Geological structural risks mainly stem from unfavorable geological conditions such as faults, fracture zones, and weak interlayers; mining disturbance effects include surface subsidence, horizontal deformation, and tilting deformation; and engineering safety thresholds were determined based on relevant national standards and the actual bearing capacity of the structures. A multi-constraint optimization model was established, and the optimal mining scheme was solved using a particle swarm optimization algorithm. The model ultimately outputs the safe mining depth ranges and corresponding risk probabilities at different confidence levels, providing a scientific basis for decision-making.

[0073] Step S4: Based on the safe mining depth range and risk probability, construct an assessment report generation module that includes core risk point analysis, key technology demonstration and forward-looking risk warning, and automatically generate an intelligent overburden risk assessment report.

[0074] In this step, an intelligent overburden risk assessment report is automatically generated based on the safe mining depth range and risk probability. The report generation module adopts a modular architecture, including a risk identification module, a technical analysis module, a chart generation module, and a text generation module. The risk identification module divides the assessment area into high-risk, medium-risk, and low-risk areas through cluster analysis and extracts key features of each area to form a risk source file. The technical analysis module, based on an expert knowledge base and an engineering case library, matches corresponding analysis methods and argumentation logic for each risk type, generating technical analysis paragraphs.

[0075] The chart generation module automatically creates various professional maps, including geological profiles, risk distribution maps, and safe mining depth contour maps. These maps use standard cartographic specifications and color schemes to intuitively display the assessment results. The text generation module uses natural language generation technology to write fluent and professional report text based on predefined report templates and specific assessment results. The entire report includes three key parts: core risk point analysis, key technology demonstration, and forward-looking risk warning, comprehensively presenting the results and recommendations of the landslide risk assessment.

[0076] The resulting intelligent overburden risk assessment report not only meets professional standards but also has practical value, providing a scientific basis for engineering decisions and ensuring the coordinated development of surface construction projects and underground mineral resource development.

[0077] In a specific embodiment, step S1 employs a robust neural network training method with game theory gradient control to deeply fuse multi-source heterogeneous geological data, including:

[0078] Step S1.1: Using the multi-source heterogeneous geological data as game participants, the Nash equilibrium theory is adopted to iteratively solve the problem so that each data source reaches the optimal weight allocation state, determine the contribution weight of each data source to the final fusion result, and obtain the initial weight allocation.

[0079] In this step, various types of geological data are treated as participants in a game, each aiming to maximize their contribution to the fusion outcome. Nash equilibrium theory provides a method for finding a balance point amidst competition among participants, ensuring that no participant has an incentive to unilaterally change their strategy. In the context of data fusion, this means finding a weighting scheme that achieves an optimal balance of contributions from each data source.

[0080] In practice, a contribution function is first defined for each data source, which measures the information gain of the data source to the fusion result. The contribution function is typically constructed based on factors such as data quality, reliability, and information content. For example, for borehole data, the 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 anomaly features.

[0081] The Nash equilibrium point is found through an iterative solution method. In each iteration, the weights of other data sources are fixed, while the weights of the current data source are optimized to maximize its contribution. After multiple iterations, an equilibrium state is considered to have been reached when the weight changes of all data sources are less than a preset threshold. This method avoids the subjectivity of weight allocation in traditional data fusion and achieves adaptive optimization of data source weights.

[0082] Ultimately, initial weight assignments were obtained for each data source. These weights reflect the relative importance of the information contribution from different data sources, providing an important reference for subsequent neural network fusion. In practical applications, the initial weight assignments may be adjusted according to changes in regional geological conditions and data quality to ensure the optimality of the fusion results.

[0083] Step S1.2: Based on the initial weight allocation, construct a robust neural network architecture that includes residual connections and batch normalization, introduce Gaussian noise and data augmentation strategies to improve the network's anti-interference ability, calculate the optimal contribution weight of each data source in the Nash equilibrium state, and obtain the optimal weight allocation state.

[0084] In this step, a robust neural network architecture was constructed for data fusion based on the initial weight allocation. This neural network adopts a deep learning architecture, containing multiple hidden layers, with residual connections and batch normalization modules added between each layer. Residual connections effectively alleviate the gradient vanishing problem in deep network training by establishing direct channels between shallow and deep features; batch normalization stabilizes the training process and accelerates network convergence by standardizing the output of each layer.

[0085] 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.

[0086] 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.

[0087] 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.

[0088] 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.

[0089] 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.

[0090] 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.

[0091] 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.

[0092] After multiple rounds of training, the network parameters gradually converged, and the fusion performance reached a stable state. At this point, a complete forward propagation was performed on the data from the entire evaluation area to generate the final fused dataset. This fused dataset integrates the advantages of various data sources, possessing higher completeness, consistency, and reliability, providing a high-quality data foundation for subsequent geological structural feature identification.

[0093] In the multi-source heterogeneous geological data acquisition stage, the geological exploration data is first standardized by converting borehole columnar sections and geological profiles of different formats into a unified digital format, establishing a structured database containing key information such as lithology codes, stratigraphic thickness, and occurrence elements. For borehole core data, image recognition technology is used to automatically extract texture features, color distribution, and fracture development from core photographs, and combined with core description records to generate multi-dimensional feature vectors. The processing of geophysical exploration data includes apparent resistivity data from electrical exploration, magnetic anomaly data from magnetic exploration, and gravity anomaly data from gravity exploration. Wavelet transform and Fourier analysis are used to remove noise and extract effective signal features.

[0094] In constructing a robust neural network architecture that includes residual connections and batch normalization, step S1.2 further includes:

[0095] Step S1.2.1: Based on the initial weight allocation, obtain the preprocessed multi-source heterogeneous geological data as the input layer of the neural network. Set up a residual connection module in the neural network and use batch normalization technology to standardize the output of each layer to ensure stable training of the deep network.

[0096] In this step, preprocessed multi-source heterogeneous geological data is first acquired as input to the neural network. Preprocessing includes data cleaning, normalization, missing value handling, and feature extraction to ensure the quality and consistency of the input data. For geological exploration data, key geological elements are extracted to form feature vectors; for borehole core data, multidimensional features containing physical, chemical, and mechanical parameters are constructed; and for geophysical exploration data, various anomaly features and spatial distribution features are extracted.

[0097] Based on this preprocessed data, a deep neural network architecture was constructed. This network contains multiple hidden layers, each consisting of fully connected or convolutional layers, capable of handling different types of geological data. To address the vanishing gradient problem in deep network training, a residual connection module was incorporated into the network. Residual connections, by adding "skip connections," establish a direct pathway between shallow and deep features, allowing gradients to flow smoothly from the output layer to the shallow layers, effectively solving the training difficulties of deep networks.

[0098] 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 calculates the mean and variance of a batch of data, then normalizes the data to zero mean and unit variance, and finally restores the data's expressive power through learnable scaling and offset parameters. This normalization process has three main advantages: first, it accelerates network convergence, allowing for larger learning rates; second, it reduces overfitting and improves the model's generalization ability; and third, it reduces the network's sensitivity to initialization, improving training stability.

[0099] By combining residual connections and batch normalization, stable training of deep neural networks was achieved, laying the technical foundation for subsequent data fusion. This network architecture can effectively learn complex patterns and correlations in multi-source heterogeneous geological data, improving the fusion effect.

[0100] Step S1.2.2: Gaussian noise perturbation and data augmentation strategies are introduced into the input layer of the neural network, including random rotation, scaling and translation transformations, to enhance the network's robustness to data perturbation.

[0101] In this step, several techniques are employed to enhance the robustness of the neural network to data perturbations. First, Gaussian noise perturbation is introduced into the network's input layer. Gaussian noise refers to random noise that follows a normal distribution with a mean of 0 and a variance equal to a preset parameter value. In each training iteration, different amounts of Gaussian noise are added to the original input data, creating slightly perturbed training samples. This perturbation simulates random errors that may exist in actual data acquisition, such as measurement errors, transmission errors, or environmental interference. By continuously exposing itself to these perturbations during training, the network gradually learns to focus on stable and reliable features in the data, ignoring the influence of random noise, thereby improving its generalization ability to unseen data.

[0102] In addition to Gaussian noise, various data augmentation strategies were employed to enrich the diversity of training samples. Random rotation refers to rotating the geological data at random angles, which helps the network learn rotation-invariant feature representations, particularly suitable for the directional analysis of geological structures. Random scaling involves enlarging or reducing the data by different proportions, enabling the network to adapt to geological features at different scales. Random translation involves randomly shifting the data in space, enhancing the network's adaptability to changes in feature location.

[0103] These data augmentation strategies effectively expand the quantity and diversity of training samples, enabling the network to access a wider range of data variations and thus build more robust feature representations. In practical applications, these augmentation strategies are often randomly combined to generate richer and more diverse training samples. By combining Gaussian noise perturbation and data augmentation strategies, neural networks can effectively resist various data perturbations, maintain stable performance in complex and ever-changing geological environments, and provide reliable data support for subsequent geological structural feature identification.

[0104] This design ensures stable training of deep networks while introducing Gaussian noise and data augmentation strategies to improve the network's resilience to input perturbations. In robust neural networks, residual connection structures help solve the gradient vanishing problem during deep network training, while batch normalization accelerates network convergence and improves generalization ability by standardizing the output of each layer.

[0105] In step S2, manifold sensing regularization technology is used to perform in-depth mining and reasoning on high-dimensional geological data to identify geological structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures, including:

[0106] Step S2.1: Based on the geological feature points in the fused dataset, construct a k-nearest neighbor graph and calculate the k nearest neighbors of each feature point, establish an adjacency matrix between feature points, and obtain the manifold structure representation of the data.

[0107] In this step, geological feature points are first extracted from the fused dataset. Geological feature points are representative points in geological space, including borehole locations, geophysical exploration points, and geological boundary feature points. These feature points are distributed in a multidimensional feature space, and each point is represented by a set of feature vectors. These features include spatial coordinates, lithological properties, physical parameters, and chemical composition. The purpose of extracting these feature points is to establish the spatial topology of the geological data, laying the foundation for subsequent manifold learning.

[0108] Based on the extracted geological feature points, a k-nearest neighbor (kNN) graph is constructed. A kNN graph is a graphical model representing the local structure of high-dimensional data, where each node represents a geological feature point, and the edges between nodes represent their proximity relationships. The process of constructing a kNN graph first calculates the distance matrix between feature points. Commonly used 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.

[0109] Based on the distance matrix, k nearest neighbors are selected for each feature point. The parameter k is an important hyperparameter, typically set according to the data size and distribution characteristics. Smaller k values ​​better preserve local structure, while larger k values ​​help capture global characteristics. In geological applications, k values ​​are usually set between 5 and 15, which captures local geological variations while reducing the influence of outliers.

[0110] After selecting the nearest neighbors, an adjacency matrix is ​​established between feature points. The adjacency matrix is ​​an n×n matrix (n is the number of feature points), where each element A[i,j] represents the connection between feature points i and j. In a simple k-nearest 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 assign different weights to edges based on the distance or similarity between points, allowing the adjacency matrix to more accurately reflect the strength of relationships between geological feature points.

[0111] By constructing a k-nearest neighbor graph and an adjacency matrix, a manifold structural representation of the data was obtained. A manifold is a topological space that is locally similar to Euclidean space. In machine learning, high-dimensional data is often distributed on low-dimensional manifolds. The manifold structural representation captures the essential geometric structure of geological data in a 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 typically exhibit strong spatial continuity and local correlations, which are highly consistent with the geometric properties of manifolds.

[0112] Step S2.2: Based on the manifold structure representation of the data, calculate the local density and curvature information of feature points in the manifold space, construct the manifold Laplacian matrix as a regularization term, constrain the data to maintain its intrinsic geometric structure during large model training, and obtain the regularization constraint parameters.

[0113] In this step, based on the established manifold structure representation, the geometric characteristics of feature points in the manifold space are further analyzed. Local density refers to the density of data point distribution around a feature point, reflecting the sufficiency of geological feature sampling or the frequency of change in a certain area. Local density is estimated by calculating the ratio of the number of points to the volume within the k-nearest neighbors of each feature point. High-density areas usually correspond to areas with drastic geological changes or dense sampling, and require more attention in the model.

[0114] Curvature information describes the degree of bending of a manifold at feature points and is an important feature of manifold geometry. When calculating curvature, a tangent plane is first constructed at each feature point, and then the degree to which the data point deviates from this tangent plane is measured. Regions with high curvature typically correspond to inflection points or anomalous areas in geological structures, such as faults, fold axes, or lithological abrupt change zones. These regions have special significance in geological interpretation and need to be preserved during model training.

[0115] Based on 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 a data manifold. It is a variation of the adjacency matrix and can express the similarity relationships between data points and their distribution characteristics on the manifold. For a weighted k-nearest neighbor graph, the Laplacian matrix L is calculated as L = D - W, where W is the weighted adjacency matrix and D is the degree matrix (a diagonal matrix where the diagonal elements are the sum of the weights of corresponding rows or columns).

[0116] The Laplace matrix of a manifold possesses many excellent mathematical properties, especially its eigenvectors, which can serve as basis functions on the manifold to represent functions on the manifold. This property makes the Laplace matrix a core tool in manifold learning. In geological applications, the Laplace matrix can capture the continuous variation patterns and local structural features of geological bodies in space, making it invaluable for identifying complex geological structures.

[0117] The manifold Laplacian matrix is ​​used as a regularization term in the training process of large models. Regularization is a technique to prevent overfitting by adding additional constraints to the loss function to limit model complexity or favor specific solutions. The core idea of ​​manifold regularization is to preserve the inherent geometric structure of the data, allowing model predictions to smoothly change along the manifold. For geological data, this means that similar geological conditions should produce similar model predictions, conforming to the principle of continuity in geology.

[0118] Regularization constraint parameters are hyperparameters that control the strength of regularization and determine the weight of manifold structure constraints in the overall loss function. Optimal regularization constraint parameters are determined using methods such as cross-validation or Bayesian optimization to achieve a balance between model generalization ability and manifold structure preservation. Optimizing these parameters is an iterative process that requires fine-tuning based on domain knowledge and experimental results to obtain optimal model performance.

[0119] Step S2.3: Add the regularization constraint parameters to the loss function of the large model based on the Transformer architecture, optimize the model parameters through multi-head self-attention mechanism and backpropagation algorithm, identify the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures, and obtain the geological structural features.

[0120] In this step, the regularization constraint parameters are first integrated into the loss function of the large model based on the Transformer architecture. The Transformer is a neural network architecture based on a self-attention mechanism, initially applied in natural language processing and later extended to various sequence data processing tasks. In geological applications, the Transformer can effectively process spatial sequence data, capture long-distance dependencies between geological feature points, and is suitable for handling complex geological structure identification problems.

[0121] The loss function of a large model typically consists of two parts: a basic task loss, which measures the difference between the model's predictions and the true labels; and a regularization loss, which constrains the model's complexity or specific properties. In this system, the basic task loss uses cross-entropy loss or mean squared error to measure the model's accuracy in identifying geological structural features. The regularization loss is constructed based on the manifold Laplacian matrix, in the form 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 vary smoothly along the manifold, preserving the intrinsic geometric structure of the geological data.

[0122] Multi-head self-attention is a core component of the Transformer architecture, allowing the model to simultaneously focus on different parts of the input sequence and different representational subspaces. In geological applications, multi-head self-attention enables the model to consider multiple geological factors and multi-scale features simultaneously. For example, some attention heads may focus on local lithological variations, while others may focus on large-scale tectonic features. This multi-faceted feature extraction capability allows the model to more comprehensively understand complex geological conditions.

[0123] Self-attention mechanisms construct global dependencies by calculating the correlation between each element in a sequence and all other elements. Specifically, for each element in the input sequence, the model calculates an attention weight distribution, indicating which other elements that element should pay attention to. This mechanism is particularly suitable for capturing spatial correlations and long-distance dependencies in geological data, such as the correspondence between strata on both sides of a fault or long-distance tectonic connections.

[0124] Model training employs the backpropagation algorithm, which calculates the gradient of the loss function with respect to the model parameters and updates the parameters in the reverse direction of the gradient, progressively optimizing model performance. Backpropagation is a standard optimization method in deep learning, capable of efficiently calculating gradients in complex networks. During training, the system utilizes techniques such as learning rate scheduling and gradient pruning to ensure training stability and convergence speed.

[0125] Through multiple rounds of training, the model gradually learns to identify different types of geological structural features. Stratigraphic interfaces are the boundaries between strata of different geological periods or different lithologies; the model identifies the interface location by learning the characteristic differences and spatial continuity on both sides of the interface. Lithological distribution reflects the spatial distribution patterns of different rock types; the model infers the complete three-dimensional lithological distribution by integrating borehole data and geophysical data. Fault fracture zones are fractured and deformed zones formed by fracturing in rock masses, possessing unique physical and geometric characteristics; the model locates faults by identifying features such as resistivity anomalies, magnetic anomalies, and core fragmentation. Joints and fractures are tiny fractures in rock masses, usually occurring in groups; the model describes the joint and fracture system by analyzing the density, orientation, and connectivity of these fractures.

[0126] Ultimately, the geological structural features, including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures, were output. These features were presented in the form of a three-dimensional spatial model, intuitively describing the geological structure of the assessment area and providing a solid geological foundation for subsequent safe mining depth evaluation.

[0127] This manifold-aware regularization technique is implemented by constructing a k-nearest neighbor graph of geological features, calculating the local density and curvature information of feature points in the manifold space, and constraining the model to maintain the inherent geometric structure of the data during the learning process. In manifold learning theory, the k-nearest neighbor graph can effectively capture the local structure of high-dimensional data, and express the similarity relationship between feature points through the adjacency matrix, laying the foundation for subsequent manifold analysis.

[0128] In step S2.3, identifying the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures further includes:

[0129] Step S2.3.1: Analyze the lithological variation sequence characteristics of the borehole core data and the resistivity and magnetic anomaly information of the geophysical exploration data in the fused dataset, and use the sequence pattern matching algorithm to automatically identify the burial depth and attitude parameters of the main stratigraphic interfaces to obtain stratigraphic interface distribution information.

[0130] In this step, the borehole core data in the fused dataset is analyzed first. Borehole cores are direct samples from underground geology, providing first-hand information on stratigraphic lithology, structure, and physicochemical properties. The system focuses on the lithological variation sequence characteristics in the core data, which reflect the vertical variation patterns of the strata. Lithological variations typically manifest as abrupt changes in lithological type, color, structure, texture, and physical properties; these change points often correspond to stratigraphic interfaces.

[0131] 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.

[0132] 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:

[0133]

[0134] 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 feature 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.

[0135] 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:

[0136]

[0137] 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. This represents the standard deviation of the physical parameters in the lower half of the window. For example, when the mean density of the upper half of the window is 2.5 g / cm³ with a standard deviation of 0.1 g / cm³, and the mean density of the lower half is 2.7 g / cm³ with a standard deviation of 0.15 g / cm³, the standardized variability calculated using the above formula is approximately 1.11. This standardization process eliminates the influence of different parameter dimensions, making the variability comparable.

[0138] In calculating the dissimilarity based on statistical features, the Kullback-Leibler divergence method is used to measure the difference in probability distribution between the upper and lower parts of the window. Its calculation formula is as follows:

[0139]

[0140] in, Represents the distribution divergence based on statistical characteristics. Indicates the number of categories of data values. This indicates that the data in the upper half of the window takes the value of The probability, This indicates that the data in the lower half of the window takes values ​​of The probability of. In practical applications, when Exceeding the preset threshold of 2.0, or Exceeding the preset threshold of 1.0, or When the value exceeds a preset threshold of 0.5, the location will be marked as a potential stratigraphic interface by the system. This multi-indicator comprehensive judgment method can effectively identify stratigraphic interfaces with different origins and manifestations, improving the accuracy and reliability of stratigraphic interface identification.

[0141] Simultaneously, resistivity and magnetic anomalies in geophysical exploration data were analyzed. Resistivity anomalies reflect changes in the electrical properties of subsurface rock masses. Different lithological strata typically exhibit different resistivity characteristics due to differences in mineral composition, porosity, and water content. Significant changes in resistivity values ​​often indicate the presence of stratigraphic interfaces. Similarly, magnetic anomalies reflect spatial variations in the magnetism of rock masses. Abrupt changes in magnetic induction intensity are usually related to lithological changes or geological structures and serve as important auxiliary information for identifying stratigraphic interfaces.

[0142] To integrate borehole and geophysical data, a sequence pattern matching algorithm is employed to automatically identify major stratigraphic interfaces. Sequence pattern matching is a technique for finding specific patterns in sequence data. In geological applications, it can identify characteristic patterns related to stratigraphic interfaces in borehole and geophysical data. Specifically, the system first extracts characteristic patterns from known stratigraphic interfaces as templates, and then searches the entire dataset for sequence fragments similar to these templates, thereby identifying potential stratigraphic interfaces.

[0143] In sequence pattern matching, the Dynamic Time Warping (DTW) algorithm is used to handle different sequence lengths and rates of change. DTW is a method for measuring the similarity between two time series, which can automatically adjust the time axis to find the optimal alignment. This feature enables DTW to handle irregular variations and scale differences commonly found in geological sequences, improving the accuracy and robustness of stratigraphic interface identification.

[0144] By matching sequences of patterns, the burial depths of major stratigraphic interfaces were determined, i.e., the depth locations of the stratigraphic interfaces within the boreholes. For multiple borehole datasets, the system connects discrete interface points into a continuous three-dimensional interface using spatial interpolation methods. Commonly used interpolation methods include kriging, inverse distance weighted interpolation, and spline interpolation; the choice of method depends on the data distribution characteristics and geological background knowledge.

[0145] In addition to burial depth, the attitude parameters of the stratigraphic interfaces were determined, including strike and dip. Strike is the direction of the intersection of the stratigraphic interface and the horizontal plane, usually expressed as the angle with true north; dip is the angle between the stratigraphic interface and the horizontal plane, reflecting the degree of inclination of the strata. These attitude parameters were calculated using the three-point method or least-squares plane fitting method, providing important parameters for subsequent three-dimensional geological modeling.

[0146] Ultimately, complete information on the distribution of stratigraphic interfaces was obtained, including the spatial location, geometric morphology, and attitude characteristics of the main stratigraphic interfaces. This information was presented in the form of a three-dimensional surface model, which intuitively described the stratigraphic structure of the assessment area and provided a basic geological framework for the evaluation of safe mining depth.

[0147] Step S2.3.2: By comprehensively analyzing the degree of core fragmentation in the borehole, the low resistivity anomaly in geophysical exploration, and the structural records in historical geological data, the support vector machine classification algorithm is used to predict the strike, dip, and influence range of the fault fracture zone, and to obtain the distribution information of the fault fracture zone.

[0148] In this step, the fault fracture zone was comprehensively identified and its features extracted. The fault fracture zone is a fractured and deformed zone formed during fault movement, where the rock is sheared, broken, and even pulverized, resulting in a series of characteristic rock types and structural phenomena. As an important geological structure, the fault fracture zone not only affects groundwater flow and rock mass stability but is also a potential trigger for geological hazards, making it crucial for assessing overburden risk.

[0149] First, the degree of core fragmentation in the borehole is analyzed. Core fragmentation is direct evidence of fault fracture zones and is typically described by parameters such as core integrity indices (e.g., RQD value), fracture zone thickness, mineral composition, and structural characteristics. The system analyzes core photographs using image recognition technology to automatically identify fracture zone features, such as the distribution of core fragment size, fracture texture, and secondary mineral infill. Simultaneously, direct observations from the drilling log, such as changes in drilling rate, borehole deviation, and drill string wear, are integrated; these are all auxiliary indicators for determining fault fracture zones.

[0150] Low resistivity anomalies in geophysical exploration are another important characteristic of fault fracture zones. Due to high rock fragmentation, increased porosity, and frequent groundwater activity, fault fracture zones typically exhibit areas with significantly lower resistivity than the surrounding intact rock mass. Analyzing electrical resistivity exploration data helps identify areas of low resistivity anomalies, and combining this with their spatial morphology and extension characteristics determines whether they are indeed fault fracture zones. Furthermore, fault fracture zones may also manifest as magnetic anomalies in magnetic exploration, especially when there are significant differences in magnetic properties between the rock masses on either side of the fault or when magnetic minerals are present within the fault zone.

[0151] Tectonic records in historical geological data are important references for fault identification. These records include geological maps, structural analysis maps, historical drilling reports, and regional tectonic research results, providing information on the regional tectonic background and the distribution of known faults. Fault-related information is extracted from these data using text mining and image recognition techniques, serving as prior knowledge and verification criteria for fault prediction.

[0152] Based on the aforementioned multi-source information, a Support Vector Machine (SVM) classification algorithm is employed to predict the distribution and characteristics of faults. SVM is a powerful supervised learning method that classifies data by constructing an optimal separating hyperplane. In fault identification applications, feature vectors are first extracted from known fault locations as training samples. These feature vectors contain multi-dimensional information such as core fracturing indices, resistivity anomalies, magnetic anomalies, and geological background parameters. Then, these training samples are used to train the SVM model, which learns the decision boundary to distinguish between faulted and non-faulted regions.

[0153] The core of SVM lies in the choice of kernel function, which determines the transformation of the feature space and the complexity of the decision boundary. For tomography tasks, the radial basis function (RBF) kernel typically performs best because it can handle nonlinear relationships between features. Kernel function parameters (such as the γ parameter of the RBF kernel) and penalty coefficient C are determined through grid search and cross-validation methods to balance the model's fitting and generalization abilities.

[0154] Using a trained SVM model, the strike, dip, and extent of the fault's fracture zone were predicted. Strike is the direction of the intersection of the fault plane and the horizontal plane; dip is the direction of the fault plane's inclination, perpendicular to the strike; and extent describes the width and spatial extension of the fault's fracture zone. These parameters were determined by analyzing the fault-related relationships between the strata on either side of the fault, the geometric distribution of the fracture zone in space, and the extension characteristics of geophysical anomalies.

[0155] Ultimately, complete information on the distribution of fault fracture zones was obtained, including the spatial location, geometric parameters, and extent of influence of the faults. This information was presented in the form of a three-dimensional model, clearly describing the fault structure of the assessment area and providing important geological risk basis for subsequent safe mining depth assessment.

[0156] Step S2.3.3: The joint and fracture system is modeled as a graph structure using a graph neural network, with fracture intersections as nodes and fracture connections as edges. The spatial distribution and connectivity characteristics of the fracture network are learned through a graph convolutional neural network to obtain the development law of joints and fractures.

[0157] In this step, the joint and fracture system was analyzed and modeled in depth. Joints and fractures are fracture surfaces in rock masses that do not involve significant displacement; they typically occur in groups, forming a network distribution. Compared to faults, joints and fractures are smaller in scale but numerous, significantly impacting the overall strength and stability of the rock mass. In overburden risk assessment, accurately understanding the development patterns of joints and fractures is crucial for predicting rock mass deformation and stability.

[0158] First, fracture information is extracted from multi-source data. Borehole cores are a crucial source of fracture information; the system analyzes core photographs using image recognition technology to identify the location, orientation, dip angle, aperture, and infill characteristics of fractures. High-precision logging data from geophysical exploration, such as sonic logging, resistivity imaging, and borehole wall television imaging, also provide direct evidence of fractures. Furthermore, fracture information revealed by surface outcrop observations and tunnel engineering is integrated to construct a more comprehensive fracture database.

[0159] Based on the extracted fracture information, a novel graph neural network technique is employed to model the joint fracture system as a graph structure. Graph structures, mathematical models describing the relationships between nodes and edges, are well-suited for expressing the topological characteristics of joint fracture systems. In this model, fracture intersections are represented as nodes, and fracture connections are represented as edges. This representation method intuitively reflects the connectivity and spatial organization of the joint fracture system, providing an ideal data model for subsequent analysis.

[0160] Each node (fracture intersection point) contains attribute information such as location coordinates, number of intersecting fractures, and intersection angle; each edge (fracture connection) contains characteristic parameters such as length, direction, aperture, infill type, and permeability. Through this attribute-rich graph structure, the geometric and physical characteristics of the joint fracture system can be comprehensively captured, providing data support for in-depth analysis.

[0161] Graph neural networks (GNNs) are a class of deep learning models specifically designed for processing graph-structured data. They can learn representations of nodes and edges and infer patterns on the graph. This study employs Graph Convolutional Neural Networks (GCNs), a mainstream GNN architecture, to learn and analyze a fracture network. The core operation of GCN is graph convolution, which updates node representations by aggregating information from node neighbors, thereby capturing local structural features of the graph. By stacking multiple layers of graph convolution, the model can progressively acquire a wider range of structural information, ultimately learning the global features of the jointed fracture system.

[0162] In practical implementation, the system defines specialized graph convolution operations that take into account the unique physical properties of the fractures. For example, edge weights consider not only topological connectivity but also the fracture aperture and permeability, enabling the model to better simulate the propagation of fluids and stresses within the fracture network. Furthermore, the system introduces an attention mechanism to dynamically adjust the importance of different neighboring nodes, more accurately capturing the complex patterns of fracture interactions.

[0163] Through training with a graph convolutional neural network, the spatial distribution patterns and connectivity characteristics of the fracture network were learned. Spatial distribution patterns include fracture directionality, density distribution, and scale distribution. The system identified the orientation of major fracture groups, typically expressed as strike / dip-dip form, such as "NE 45° / SEA -70°". Density distribution describes the spatial variation in fracture density, usually expressed as linear density (total fracture length per unit area) or volumetric density (total fracture area per unit volume). Scale distribution reflects the statistical regularities of parameters such as fracture length and aperture, typically conforming to a power law or log-normal distribution.

[0164] Connectivity is another key attribute of jointed fracture systems, directly affecting the permeability and mechanical behavior of rock masses. This system quantifies the connectivity of the fracture network by analyzing connectivity indices of the graph structure, such as mean degree, clustering coefficient, path length, and number of connected components. Furthermore, it identifies critical nodes and edges in the fracture network; these structures play a decisive role in maintaining network connectivity, and their disruption can lead to significant changes in the overall system performance.

[0165] Ultimately, a complete understanding of joint and fracture development patterns was obtained, including the spatial distribution characteristics, geometric statistical parameters, and connectivity properties of the fractures. These patterns were presented in the form of statistical models and three-dimensional visualization models, providing a scientific basis for understanding the rock mass structure of the assessment area and predicting its mechanical behavior, and serving as an important input for safe mining depth evaluation.

[0166] Step S2.3.4: Integrate the distribution information of the stratigraphic interfaces, the distribution information of the fault fracture zones, and the development law of the joints and fissures to construct a three-dimensional geological structure model of the region and obtain the geological structure characteristics.

[0167] In this step, the stratigraphic interface distribution information, fault fracture zone distribution information, and joint and fracture development patterns identified in the previous steps are integrated to construct a complete three-dimensional geological structure model of the region. The three-dimensional geological structure model is a digital representation of the geometry and spatial relationships of underground geological bodies, capable of intuitively presenting complex geological structures and serving as a crucial foundation for overburden risk assessment.

[0168] This digital representation first employs a three-dimensional Cartesian coordinate system to locate the spatial positions of all geological elements. The horizontal plane uses a planar coordinate system with the east-facing X-axis and the north-facing Y-axis, while the vertical direction uses the downward-positive Z-axis to represent depth. Each geological element is precisely described by a set of spatial coordinate points. The spatial position of any point on a geological interface can be represented as:

[0169]

[0170] in, This represents the spatial location vector of a point on a geological interface. This represents the eastward coordinates of the point. This represents the north coordinates of the point. This represents the depth coordinates of the point. For example, a geological interface can be represented as a set of three-dimensional coordinate points containing thousands or even tens of thousands of such points, which completely depict the distribution of the geological interface in three-dimensional space.

[0171] In terms of the digital representation of geometric morphology, the system employs triangular mesh technology to construct continuous surfaces of stratigraphic interfaces and fault planes. Each triangular patch is fully defined by the spatial coordinates of its three vertices and the connections between the vertices. A triangular patch can be represented as:

[0172] = ( , , )

[0173] in, Represents a triangular facet. , , These represent the spatial position vectors of the three vertices of the triangle. For example, a complex and undulating geological interface may contain five thousand triangular facets. By piecing these triangular facets together according to topological relationships, a complete three-dimensional surface model of the geological interface is formed. This triangular mesh representation can not only accurately describe complex surface morphologies, but also facilitates computer storage, display, and analysis.

[0174] For the digital representation of geological body attribute information, the system associates a multi-dimensional attribute vector with each spatial unit, which can be represented as:

[0175]

[0176] Where A represents the attribute vector at a certain spatial location. Indicates the number of dimensions of the attribute. to These represent different geological attribute parameters, including lithology type code, rock density, elastic modulus, Poisson's ratio, compressive strength, etc. Specifically, in a three-dimensional geological model using voxel representation, the entire evaluation area is divided into regular three-dimensional grid cells, each called a voxel. For example, a voxel cell located at a specific location might record detailed information such as that the rock is sandstone (lithology code 2), its density is 2.65 g / cm³, and its elastic modulus is 15 GPa.

[0177] The spatial topological relationships between geological elements are digitally described using relational data tables. This standardized data format facilitates long-term data storage and exchange, and provides a solid data foundation for subsequent computational analysis and visualization. The entire digital representation system achieves a precise, complete, and computable description of complex three-dimensional geological structures, providing a reliable geological model basis for dynamic safe mining depth evaluation.

[0178] The integration process begins by processing stratigraphic interface data. Stratigraphic interfaces form the basic framework of a 3D geological model, defining the spatial boundaries of different lithological units. Using the main stratigraphic interfaces identified in the previous steps as constraint surfaces, continuous 3D surfaces are generated through spatial interpolation methods. Commonly used interpolation methods include Triangular Irregular Network (TIN) interpolation, Kriging interpolation, and Discrete Smoothing Interpolation (DSI). Among these, the DSI method is particularly suitable for constructing complex geological surfaces because it can simultaneously consider multiple constraints, such as interface points, attitude measurements, and fault truncation, generating smooth and geologically sound surfaces.

[0179] Next, information on fault fracture zones is integrated. Faults are crucial structural elements in geological models, controlling the faulting relationships and spatial distribution of strata. Identified faults are represented as three-dimensional surfaces, and a geometric model is constructed based on their strike, dip, and spatial extent. Importantly, not only are the fault surfaces themselves simulated, but the width and physical property variations of the fault fracture zones are also represented, which is essential for accurately assessing rock mass strength and deformation characteristics.

[0180] During model construction, the intersection relationship between faults and stratigraphic interfaces was addressed. When a fault cuts through strata, it causes faulting and displacement, forming typical fault geological structures. By prioritizing faults over stratigraphic interfaces, the truncation effect of faults on strata was achieved. For complex fault systems, the system also needs to handle the intersection and cutting relationships between faults to ensure the rationality of the model construction.

[0181] The integration of joint and fracture systems is an innovative aspect, as traditional 3D geological models often neglect structural elements at this scale. Based on the fracture development patterns obtained in previous steps, the system uses a stochastic simulation method to generate a fracture network in 3D space that conforms to statistical characteristics. Specifically, the system first defines the directional, density, and scale distribution parameters of fracture groups, and then randomly generates fracture surfaces in 3D space, ensuring that their overall statistical characteristics match the observational data. This method can reasonably infer the fracture distribution across the entire region based on limited observational data.

[0182] During the integration process, logical consistency among geological structures at different scales was ensured. For example, large faults are often accompanied by increased fracture density; certain stratigraphic interfaces may be dominant surfaces for fracture development; and fractures may exhibit different geometric characteristics when traversing different lithologies. By defining the rules governing the interrelationships between these structural elements, the system ensures that the final model conforms to geological principles and field observations.

[0183] In the technical implementation of constructing the three-dimensional geological structure model, a hybrid representation method of voxels and surfaces was adopted. Voxel representation discretizes the three-dimensional space into a regular mesh, with each mesh cell assigned a geological attribute value, which is suitable for expressing continuously changing physical property distributions; surface representation uses triangular facet meshes to describe geological interfaces and fault planes, which is suitable for expressing the boundaries and structural surfaces of geological bodies. This hybrid representation method balances accuracy and efficiency and can comprehensively express complex geological structural features.

[0184] Ultimately, a complete three-dimensional geological structural model of the region was obtained, encompassing core geological structural features such as stratigraphic interfaces, lithological distribution, fault fracture zones, and joint and fracture systems. This model, presented as a three-dimensional digital model, can be observed and dissected from any angle, intuitively demonstrating the spatial relationships and internal structure of the underground geological formations. As the basis for overburden risk assessment, this model provides a geological structural framework for the assessment area, offering crucial input for subsequent safe mining depth calculations and risk assessments.

[0185] This series of steps enables the accurate identification of complex geological structures. In the stratigraphic interface identification process, the sequence pattern matching algorithm automatically identifies the spatial distribution of major stratigraphic boundaries by analyzing lithological variation sequences and geophysical anomaly characteristics in borehole data. The support vector machine classification algorithm is used for fault fracture zone identification, predicting the geometric parameters and influence range of faults through comprehensive analysis of multi-source data features. For joint and fracture systems, graph neural network technology effectively learns the spatial distribution and connectivity characteristics of the fracture network by modeling the joint and fracture system as a graph structure.

[0186] Step S3 employs a fractional peak differential equation neural network with efficient adjoint parameter training to establish a dynamic safe mining depth evaluation model, including:

[0187] Step S3.1: Based on the geological structural characteristics, the memory effect and long-range correlation of geological materials are described by fractional differential equations, and a peak neural network is designed using a leakage integral distribution model to establish a dynamic safe mining depth evaluation model.

[0188] In this step, a dynamic safe mining depth evaluation model was established based on the previously identified geological structural features. 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 disturbances, thereby achieving a more reliable safe mining depth evaluation.

[0189] Fractional differential equations are a mathematical tool that extends the orders of differentiation and integration to non-integer values, offering unique advantages in describing conditions with memory effects and long-range dependencies. Traditional integer differential equations assume that future states depend only on the current state (Markov property), neglecting the cumulative effects of historical states. However, geological materials often exhibit a significant memory effect, meaning that the current deformation of a material depends not only on the current stress state but also on historical loading paths. This memory effect is particularly pronounced in rock and soil materials with significant rheological properties, such as soft rocks, salt rocks, and clay rocks.

[0190] Fractional differential equations, by introducing the fractional derivative operator, can naturally express this memory effect. The fractional derivative is defined as the fractional integral of the integer derivative, or equivalently, the fractional derivative of the integer integral. This mathematical tool allows for the consideration of the influence of all past moments on the current state, with weights that decay with increasing time intervals but never become zero, thus accurately capturing the long-term memory characteristics of geological materials. In safe mining depth assessments, this means that models can consider the cumulative effects of mining activities and more accurately predict the long-term evolution of surface deformation.

[0191] The choice of fractional order is a key parameter in the model, determining the strength of the memory effect. The fractional order α typically ranges from 0 to 2; when α=1, it degenerates into the classical first derivative, and when α=2, it degenerates into the second derivative. For geological materials, the choice of α is based on the material's mechanical properties and field test data. Generally, loose sedimentary rock layers have smaller α values ​​(0.5-0.7), exhibiting a stronger memory effect; hard crystalline rock layers have larger α values ​​(0.8-1.0), closer to the classical elastic model; while strata with significant rheological properties may have α values ​​exceeding 1.0. The most suitable α value is determined through inversion analysis of measured data to improve the model's accuracy.

[0192] When implementing fractional differential equations, numerical approximation methods are employed, such as the Grünwald-Letnikov definition or finite difference schemes, to discretize the continuous fractional derivatives into a computer-implementable form. These methods typically involve a weighted sum of historical states, with weights calculated according to specific recursive relationships, to achieve numerical simulation of the memory effect.

[0193] Spike neural networks are another innovation, mimicking the pulse firing mechanism of biological neurons, making them particularly suitable for handling highly temporal dynamics. Unlike traditional neural networks that use continuous activation functions, neurons in spike neural networks transmit information through discrete pulses (spikes). The neuron's membrane potential evolves over time, firing a pulse and resetting the membrane potential when a threshold is exceeded. This mechanism allows the network to naturally handle temporal information and dynamic processes.

[0194] Spike neurons are represented using the leaky integral firing (LIF) model, a widely used simplified neuron model. In the LIF model, the neuron's membrane potential V(t) evolves according to the following differential equation:

[0195]

[0196] Where τ is the membrane time constant, V rest V(t) is the resting potential, R is the membrane resistance, and I(t) is the input current. When V(t) reaches the threshold V... th At that time, the neuron fires a pulse and resets to V. rest This dynamic characteristic enables spike neural networks to effectively capture the time-dependent and dynamic response features of geology.

[0197] A dynamic safe mining depth evaluation model was constructed by combining fractional differential equations and peak neural networks. The model first converts geological structural features into model inputs, including key parameters such as lithology distribution, fault location, and fracture density. Then, fractional differential equations describe the dynamic response of geological materials under mining disturbances, considering the material memory effect and long-range correlation. Finally, peak neural networks are used to simulate and predict the dynamic evolution, outputting surface deformation and risk probability under different mining depth conditions.

[0198] A dynamic safe mining depth evaluation model was constructed by combining fractional differential equations and peak neural networks. The model first converts geological structural features into model inputs, including key parameters such as lithology distribution, fault location, and fracture density. Then, fractional differential equations describe the dynamic response of geological materials under mining disturbances, considering the material memory effect and long-range correlation. Finally, peak neural networks are used to simulate and predict the dynamic evolution, outputting surface deformation and risk probability under different mining depth conditions.

[0199] The specific calculation of surface deformation is achieved by establishing a fractional-order differential dynamic equation, which accurately describes the deformation process of geological materials over time under mining disturbance. The equation contains several key parameters, among which the fractional-order parameter characterizes the memory properties of the geological material; its value is typically between 0.5 and 1.5, determined based on the rheological properties of the rock mass. The mining area and mining depth parameters directly reflect the scale of mining activities, while the geological structural parameter vector comprehensively characterizes geological conditions such as fault distribution and fracture development. The stress field distribution function describes the stress state of the underground rock mass at different locations and times.

[0200] The spike neural network solves the aforementioned fractional differential equation by simulating the pulse firing mechanism of biological neurons, thereby calculating the subsidence at any location on the Earth's surface at any time. Each output neuron in the network corresponds to a spatial location on the Earth's surface, and its output value is the predicted subsidence at that location. The calculation of the neurons involves multiple layers of weight parameters and threshold parameters, which have been optimized and determined through the previous network training process. The network input includes geological parameters, mining parameters, and time information. After nonlinear transformation in the hidden layers and weighted summation in the output layers, the predicted value of surface subsidence is finally obtained.

[0201] After obtaining the spatial distribution of surface subsidence, the system further calculates two important deformation indices: surface tilt and curvature. Surface tilt is obtained by calculating the gradient of subsidence in the horizontal direction. Specifically, this is done by calculating the partial derivatives of subsidence with respect to east and north coordinates, and then taking the square root of the sum of the squares of the two partial derivatives. The result reflects the steepness of the surface slope. Surface curvature is obtained by calculating the second partial derivatives of subsidence. The second partial derivatives with respect to east and north coordinates are calculated separately, and the larger of the two absolute values ​​is taken as the curvature value at that location. This index reflects the degree of surface bending deformation.

[0202] The risk probability calculation is based on the exceedance probability analysis method of safety thresholds. The system first defines three key safety thresholds: the maximum allowable settlement threshold, the maximum allowable tilt threshold, and the maximum allowable curvature threshold. These thresholds are determined based on the type and structural characteristics of surface structures. Then, for specific mining depth conditions, the system uses Monte Carlo simulation to perform extensive random sampling calculations. In each simulation, considering the uncertainties of geological and model parameters, a set of parameter values ​​is randomly generated, and the surface deformation under this parameter combination is calculated using the aforementioned fractional differential equations and spike neural networks. By counting how many times the maximum surface settlement, maximum tilt, or maximum curvature exceeded the allowable threshold in the total number of simulations, and dividing this number of exceedances by the total number of simulations, the risk probability under that mining depth condition is obtained.

[0203] For example, when evaluating a mining depth of 300 meters, the system performed 10,000 Monte Carlo simulations. In these simulations, 1,500 instances showed surface deformation exceeding the safety threshold. Therefore, the risk probability at this mining depth is 1,500 divided by 10,000, or 0.15 or 15%. This risk probability value quantitatively characterizes the likelihood of surface building safety risks under current geological conditions and mining plans, providing decision-makers with a scientific basis for risk assessment. By systematically calculating the risk probability at different mining depths, a curve showing the change in risk probability with mining depth can be plotted, thus providing a quantitative basis for determining safe mining depth ranges.

[0204] This innovative model architecture enables a comprehensive consideration of geological structural risks, mining disturbance effects, and their temporal evolution characteristics, resulting in a more accurate assessment of safe mining depth. Compared to traditional empirical formulas or integer-order differential equation models, this model better captures the nonlinear dynamic characteristics and long-term evolution trends of geology, improving the accuracy and reliability of predictions.

[0205] Step S3.2: Based on the dynamic safe mining depth evaluation model, establish a multi-constraint optimization model. The constraints include the maximum surface subsidence, maximum tilt, and maximum curvature deformation not exceeding the safety threshold. Use the particle swarm optimization algorithm to solve the constrained optimization problem and obtain the safe mining depth range and risk probability.

[0206] In this step, a multi-constraint optimization model was established based on the dynamic safe mining depth evaluation model. By solving this optimization problem, the safe mining depth range and corresponding risk probabilities under different confidence levels were determined. This process transforms geological safety requirements into mathematical constraints, and uses optimization algorithms to find the optimal mining scheme that meets safety standards, providing a scientific basis for decision-making.

[0207] First, it needs to be clarified that the risk probability mentioned here is essentially the same physical quantity as the risk probability calculated above, both representing the possibility of excessive surface deformation under specific mining depth conditions.

[0208] In S3.2, the quantitative determination of the safe mining depth range under different confidence levels involves several key steps. First, an optimization objective function is established, which aims to maximize the amount of exploitable mineral resources. The calculation of resource quantity involves parameters such as mining area, mining depth, average ore layer density, and loss rate. An optimization algorithm is then used to find the mining depth scheme that yields the maximum resource quantity, representing the optimal economic benefit under safety constraints.

[0209] The constraints of the optimization model are the core of the entire solution process, mainly including two categories: risk probability constraints and physical feasibility constraints. Risk probability constraints require that, at the selected mining depth, the calculated risk probability cannot exceed a pre-set upper limit of permissible risk probability. This upper limit is directly related to the confidence level; the higher the confidence level, the higher the required level of safety assurance, and the lower the corresponding upper limit of permissible risk probability. Specifically, when a 95% confidence level is required, it means there is a 95% certainty that mining is safe, and the corresponding upper limit of permissible risk probability is set at 5%; when a 90% confidence level is required, the upper limit is set at 10%; and when an 80% confidence level is required, the upper limit is set at 20%. Physical feasibility constraints include conditions such as the mining depth being within a technically feasible range, and the surface subsidence, tilt, and curvature meeting corresponding safety standards.

[0210] The calculation of the safe mining depth range is achieved by traversing the possible mining depth range and combining it with 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 in steps of a certain size. Using the aforementioned Monte Carlo simulation method, 10,000 random simulations are performed on each candidate depth, the number of times the limit is exceeded is counted, and the risk probability is calculated. The calculated risk probability is compared with the upper limit of the allowable risk probability to identify all depth values ​​that satisfy the risk probability constraint; these depth values ​​constitute the preliminary feasible region.

[0211] Based on the defined feasible region, the lower and upper limits of the safe mining depth range are further determined. The lower limit is defined as the minimum mining depth that satisfies the risk probability constraint. Physically, when the mining depth is less than this value, the risk of surface deformation is greater due to the shallow burial depth, and the calculated risk probability will exceed the allowable upper limit, failing to meet safety requirements. The lower limit is determined by finding the minimum depth within the feasible region where the risk probability is exactly equal to or slightly less than the allowable upper limit. The upper limit is defined as the optimal mining depth that maximizes resource quantity under the conditions of satisfying both the risk probability constraint and all physical constraints. The upper limit is determined by searching for the optimal solution within the feasible region using a particle swarm optimization algorithm.

[0212] For any mining depth within the interval, the corresponding risk probability can be calculated using the aforementioned method, and this risk probability value will not exceed the set upper limit of the allowable risk probability. Taking a 95% confidence level as an example, the upper limit of the allowable risk probability is 5%. The system iterates through the depth range of 200 meters to 600 meters, calculating 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%, failing to meet safety requirements; when the mining depth reaches 380 meters, the risk probability drops to approximately 5%, satisfying the constraint conditions. Therefore, the lower limit of the interval is determined to be 380 meters. Continuing to search deeper, the risk probability at all depth points within the range of 380 meters to 600 meters does not exceed 5%, satisfying safety requirements. Combining geological condition analysis and mining technology feasibility, while considering the goal of maximizing resource volume, the optimized algorithm determines the optimal depth to be 420 meters; therefore, the upper limit of the interval is set at 420 meters. Ultimately, it was concluded that under a 95% confidence level requirement, the safe mining depth range is 380 meters to 420 meters, and the risk probability at any depth within this range meets the safety requirement of not exceeding 5%.

[0213] Through the aforementioned systematic quantitative analysis process, a strict correspondence was established between confidence levels, permissible risk probabilities, and safe mining depth ranges. Different confidence levels correspond to different upper limits of permissible risk probabilities, thereby determining safe mining depth ranges of varying widths and providing a scientific basis for mineral resource development decisions. This method ensures both the scientific rigor and reliability of the assessment results and provides decision-makers with flexibility, allowing them to select appropriate confidence levels based on factors such as the importance of the project and the sensitivity of structures, achieving an optimal balance between safety and economic benefits.

[0214] The core of multi-constraint optimization models is defining appropriate objective functions and constraints. In safe mining depth evaluation, the objective function is usually set to maximize exploitable resources or economic benefits, while the constraints are based on the safety requirements of surface structures and facilities. Three main types of constraints were considered: maximum surface subsidence, maximum tilt, and maximum curvature deformation.

[0215] Maximum surface settlement refers to the maximum vertical displacement of the ground surface caused by mining activities, and it is directly related to the overall settlement of ground buildings. The safety threshold varies depending on the type and importance of the building. Generally, the maximum settlement of ordinary civil buildings is usually no more than 200-300 mm; important public buildings may be limited to 100-150 mm; while for particularly sensitive buildings or facilities, such as precision instrument factories or high-speed rail lines, the limit may be as low as 10-30 mm. Appropriate settlement limits have been set based on the actual conditions of buildings within the assessment area and relevant regulations.

[0216] Maximum tilt is another key indicator of ground surface deformation, referring to the maximum slope of the ground settlement curve, reflecting the degree of differential settlement a building may experience. Tilting deformation can cause buildings to tilt, affecting their functionality and, in severe cases, threatening structural safety. The tilt limit for general civil buildings is typically 3-6 mm / m; for high-rise buildings or frame structures, the limit may be stricter, such as 2-3 mm / m; while for special facilities, such as long-span bridges or precision equipment workshops, the limit may be as low as 1 mm / m or less. Appropriate tilt limits are determined based on the building's structural type and functional requirements.

[0217] Maximum curvature deformation is an indicator reflecting the degree of surface curvature, defined as the second derivative of the surface settlement curve, and is directly related to the curvature deformation of buildings. Curvature deformation can cause tensile or compressive stresses in buildings, potentially leading to wall cracking or structural damage. For brick-concrete structures, the curvature deformation limit is typically 0.2-0.3 mm / m. 2 For reinforced concrete frame structures, the limit may be 0.3-0.5 mm / m. 2 For special structures or facilities, the limits are determined based on their resistance to deformation. Corresponding curvature deformation limits are set based on the structural characteristics and material properties of the building.

[0218] These constraints collectively constitute 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, and solving such problems requires efficient numerical algorithms.

[0219] The Particle Swarm Optimization (PSO) algorithm is used to solve the constrained optimization problem described above. 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 among particles. In this algorithm, each particle represents a possible mining scheme, including parameters such as mining depth, mining range, and mining order. Particles move in the search space, and their position updates are guided by their individual optimal positions and the global optimal positions.

[0220] The fitness function of a particle is designed to maximize resource extraction or economic benefits while satisfying all constraints. To handle these constraints, a penalty function method is employed, transforming the degree of constraint violation into a fitness penalty term, guiding the search process towards the feasible solution region. This method can effectively handle complex nonlinear constraints and find globally optimal or near-global optimal solutions.

[0221] In practical applications, the particle swarm size is typically set to 50-100 particles, and the number of iterations is 500-1000 to ensure sufficient exploration capability and convergence of the algorithm. Adaptive inertia weights and shrinkage coefficients are also employed to balance the algorithm's global search and local refinement capabilities, improving both solution efficiency and solution quality.

[0222] Considering the uncertainties in geological parameters and model predictions, this approach not only outputs a single optimal solution but also generates safe mining depth ranges at different confidence levels. This is achieved through Monte Carlo simulation, which involves randomly sampling geological and model parameters to generate numerous simulation scenarios. An optimization algorithm is then run in each scenario to obtain a series of optimal mining depths. Statistical analysis of these results determines safe mining depth ranges at different confidence levels, such as 95%, 90%, and 80%.

[0223] In addition, the risk probability associated with each mining depth was calculated, i.e., the likelihood that mining at that depth would result in exceeding a safety threshold. This information is crucial for risk management and decision-making, enabling decision-makers to select appropriate mining schemes based on their own risk preferences.

[0224] Finally, a complete assessment of safe mining depth was output, including safe mining depth ranges at different confidence levels and the corresponding risk probability distributions. These results are presented in the form of numerical reports and visualization charts, intuitively demonstrating the relationship between mining depth and safety risk, and providing a scientific basis for assessing the risk of mineral resource overlay.

[0225] The introduction of fractional differential equations can better describe the memory effect and long-range correlation of geological materials. Their order is typically between 0.5 and 1.5, and optimal parameter values ​​are determined through field experiments and inversion analysis. The design of the spike neural network employs a leakage integral firing model. When the membrane potential exceeds a threshold, the neuron fires a pulse and resets the membrane potential. This mechanism can effectively handle highly time-series geological evolution processes.

[0226] In the process of using a fractional peak differential equation neural network trained with efficient adjoint parameters, step S3.1 further includes:

[0227] Step S3.1.1: For the state equations in forward propagation Construct the corresponding adjoint equations , 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.

[0228] In this step, an innovative solution, the adjoint method, is introduced to address a key challenge in neural network training—the vanishing gradient problem. Traditional backpropagation algorithms, when processing long-term sequences or deep networks, are prone to exponential gradient decay due to the need for long-chain derivatives, ultimately preventing effective updates to network parameters. This problem is particularly prominent in geodynamic simulations, as long-term geological evolution processes typically require simulating long-term sequences.

[0229] First, we need to understand the state equation in forward propagation. The state equation is a differential equation describing how the state changes over time, which can be represented in a neural network as follows: In this context, x represents the state variable, which can be understood as the node activation value or hidden state in a neural network; θ represents the network parameters, including learnable parameters such as weights and biases; t represents the time variable; and f is a nonlinear function describing 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 a geological body.

[0230] Traditional backpropagation algorithms calculate the gradient of the loss function L with respect to the network parameters θ using the chain rule. For long-term sequences, this means the gradient needs to be propagated backward from the output layer to the initial state at each time step, forming a computational chain of hundreds or even thousands of steps. 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 gradient vanishing; if the eigenvalues ​​are greater than 1, it can lead to gradient explosion. Both of these situations severely impact the training performance of neural networks, especially when simulating the long-term memory effect in geology.

[0231] The adjoint method provides an alternative strategy to avoid long chain derivatives. Originating from optimal control theory and variational methods, the adjoint method is widely used in data assimilation and parameter estimation problems in fields such as meteorology and oceanography. Its core idea is to construct an adjoint to directly calculate the gradient of the objective function with respect to the parameters, without the need for stepwise backpropagation.

[0232] In this paper, the adjoint method first defines the adjoint variable λ, which has the same dimension as the original state variable x, but with a different physical meaning. Mathematically, the adjoint variable λ can be understood as a sensitivity or importance index of the state variable x's influence on the loss function. The dynamic evolution of the adjoint variable is described by the adjoint equation: .

[0233] In the adjoint equation It is the Jacobian matrix of the original state equation with respect to the state variables, and its transpose This reflects the mutual influence between state changes; This represents the gradient of the loss function with respect to the state variables, reflecting the direct impact of the state variables on the final evaluation metric. The negative sign indicates that the adjoint evolves backward from the termination time to the initial time, which is a key characteristic of the adjoint method.

[0234] Constructing the adjoint equation involves complex mathematical derivations. First, the discrete-time neural network is dynamically represented as a continuous-time differential equation, establishing the state equation. Then, based on the variational principle, the impact of state changes on the final loss function is derived, resulting in the definition and dynamic equation of the adjoint variable λ. This derivation process ensures the mathematical equivalence between the adjoint method and traditional backpropagation, while avoiding the computational complexity of long chain derivatives.

[0235] It is worth noting that the adjoint equation, though seemingly simple in form, contains a wealth of information. (Jacobi matrix) It captures the internal dynamic structure and interdependencies; gradient terms This ensures a direct correlation between the adjoint and the optimization objective; the negative sign and transpose operations reflect the inverse nature of the adjoint, enabling it to efficiently transmit gradient information.

[0236] In practical calculations, higher-order numerical methods are used to solve the adjoint equations. Since the adjoint equations are linear (although their coefficients are nonlinear), they are generally easier to solve than the original nonlinear state equations. Depending on the characteristics of the problem, appropriate numerical methods are selected, such as the implicit Runge-Kutta method or exponential integration methods, to ensure the stability and accuracy of the numerical solution.

[0237] A key advantage of the adjoint method is its computational complexity, which is independent of the time step. Regardless of the length of the simulated time series, the adjoint method requires only one forward propagation and one backward solution to obtain the accurate gradient of the loss function with respect to the parameters. This characteristic enables efficient handling of long-term geological evolution processes, capturing memory effects and long-term correlations, and providing a solid computational foundation for dynamic safe mining depth assessment.

[0238] By constructing the adjoint equation, the gradient vanishing problem in traditional backpropagation over long time series was successfully avoided, laying the foundation for subsequent 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.

[0239] Step S3.1.2: By solving the adjoint equation, the gradient of the loss function with respect to the network parameters is efficiently calculated, thereby enabling rapid training and parameter optimization of the neural network.

[0240] 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-like differentiation process in traditional backpropagation, thereby achieving more stable and efficient network training.

[0241] First, the adjoint equation needs to be solved to obtain the value of the adjoint variable λ. Adjoint equation It is a linear ordinary differential equation, but it needs to be solved in reverse from the terminal time to the initial time. This solution process starts from the final time point, where λ(T) is initialized to zero or a specific value (depending on the form of the terminal condition), and then gradually evolves back to the initial time point.

[0242] Solving the adjoint equation requires specialized numerical integration methods. Since the adjoint equation evolves in reverse, conventional forward integration methods are not applicable. Numerical methods with inverse adaptive step sizes, such as the inverse Runge-Kutta method or the inverse Adams method, are employed to ensure the stability and accuracy of the numerical solution. These methods can automatically adjust the integration step size, improving computational efficiency while maintaining accuracy.

[0243] To further improve solution efficiency, a checkpoint mechanism for the adjoint variables was implemented. During forward propagation, state information at some key time points is selectively stored as checkpoints for backpropagation of the adjoint equation. This way, during backpropagation, intermediate states only need to be recalculated from the most recent checkpoint, rather than starting from scratch, significantly reducing computational and storage overhead. The checkpoints are set using a logarithmic interval strategy, which is particularly effective for long time series.

[0244] Once the adjoint variable λ is obtained, the gradient of the loss function L with respect to the network parameters θ can be efficiently calculated. According to the theory of the adjoint method, this gradient can be expressed as... This integral expression reflects the cumulative effect of the parameter θ on the loss function L by influencing the dynamics f. In the discrete implementation, integration is transformed into summation, calculating the local contribution at each time step and then accumulating them to obtain the total gradient.

[0245] It is worth noting that the gradient calculation formula described above directly provides global gradient information, avoiding the gradual gradient propagation process in traditional backpropagation. This allows gradient information to be accurately propagated from the termination time to the initial time, without being limited by the eigenvalues ​​of the Jacobian matrix in intermediate steps, effectively solving the problems of gradient vanishing and gradient exploding.

[0246] In practical applications, the combination of automatic differentiation and parallel computing optimization further improves the efficiency and accuracy of gradient calculation. Automatic differentiation automatically generates gradient calculation code through computational graph analysis, avoiding the complexity and potential errors of manually deriving the Jacobian matrix. Parallel computing leverages the multi-core or GPU acceleration capabilities of modern hardware to simultaneously calculate gradients for multiple time steps or multiple parameters, significantly improving computational speed.

[0247] After obtaining 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 employed. These optimizers can automatically adjust the learning rate of each parameter based on historical gradient information, accelerating convergence while maintaining training stability.

[0248] To further improve training performance, batch processing and learning rate scheduling strategies were implemented. The batch processing strategy divides the training data into small batches, using one batch of data to calculate the gradient for each update, increasing the randomness and generalization ability of the training. The learning rate scheduling strategy automatically adjusts the global learning rate according to the training progress, typically using a larger learning rate in the early stages of training to quickly explore the parameter space, and then using a smaller learning rate in the later stages to fine-tune the parameters, achieving better convergence.

[0249] Regularization techniques were also introduced to prevent overfitting. Regularization methods included L1 / L2 regularization (which limits parameter size by adding a parameter norm term to the loss function), Dropout (randomly shutting down some neurons to enhance model robustness), and early stopping (monitoring validation set performance and stopping training before overfitting begins). These techniques ensured the model's generalization ability on unseen data, which is crucial for the reliability of geological simulations.

[0250] To evaluate training effectiveness and parameter optimization progress, comprehensive evaluation metrics and visualization tools were designed. Evaluation metrics include mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). 2 Various methods, such as loss function analysis, are used to measure the accuracy of model predictions from different perspectives. Visualization tools, on the other hand, display the real-time trend of the loss function, the distribution of key parameters, and the comparison between predicted results and true values, helping researchers to intuitively understand the training process and model performance.

[0251] By calculating gradients and optimizing network parameters using the adjoint method, rapid training of neural networks was achieved. Compared to traditional backpropagation, the adjoint method demonstrates significant advantages in handling long-term data series, increasing training speed by several to tens of times while avoiding the vanishing gradient problem and ensuring effective parameter updates. This efficient training method enables the handling of more complex geological models and longer time series, laying the foundation for accurately simulating the long-term dynamic behavior of geology.

[0252] Ultimately, the optimized neural network accurately captures the memory effect and long-range correlation of geological materials, enabling precise prediction of surface deformation. The network parameters, after thorough training, encode the complex dynamic behavior of geology, allowing the model to reliably assess risk probabilities under different mining depths based on geological structural features and mining parameters, providing a scientific basis for safe mining depth decisions.

[0253] This neural network training technique based on the adjoint method is not only applicable to current safe mining depth assessment tasks, but can also be extended to other geological engineering fields, such as groundwater simulation, oil and gas reservoir characterization, and seismic risk assessment, and has broad application prospects and important scientific value.

[0254] Step S4 involves building an assessment report generation module that includes core risk point analysis, key technology demonstration, and forward-looking risk warning, including:

[0255] Step S4.1: Based on the safe mining depth range and risk probability, cluster analysis is used to divide the assessment area into risk areas of different levels, including high-risk areas, medium-risk areas and low-risk areas. Geological parameters, calculation model parameters and assessment results of each risk area are extracted to form a risk source file.

[0256] In this step, based on the previously calculated safe mining depth range and risk probability, the entire assessment area was divided into risk zones and classified. Risk zoning is a crucial step in the overburden risk assessment, aiming to transform spatially continuous risk indicators into discrete risk level areas to facilitate management decisions and risk control. This process involves multiple technical aspects, including data processing, cluster analysis, and risk profile establishment.

[0257] The safe mining depth range refers to the depth 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 uncertainties of geological conditions and mining technology. For example, the safe mining depth range for a certain area might be 350 to 400 meters, meaning that mining depths greater than 400 meters are considered safe, depths less than 350 meters are considered high-risk, and the area between 350 and 400 meters is considered a transitional zone. The calculation of the safe mining depth range is based on the dynamic safe mining depth evaluation model established in the previous steps, comprehensively considering factors such as geological structural characteristics, rock mass mechanical properties, and the sensitivity of surface structures.

[0258] Risk probability is a key indicator for quantitatively characterizing the risk of land cover damage. It refers to the probability that surface structures may experience 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, transferring the uncertainty of model parameters to the final result and forming a probabilistic expression. For example, the risk probability of a certain area at a mining depth of 300 meters might be 0.15, meaning that under this condition, there is a 15% chance of excessive deformation. Risk probability provides a quantitative basis for risk zoning, making risk assessment results more objective and reliable.

[0259] Cluster analysis is an unsupervised machine learning method used to divide data points into groups (clusters) with similar characteristics. In landslide risk assessment, cluster analysis, based on multidimensional risk features, divides spatial locations into high, medium, and low risk level regions. The risk features used include: upper and lower limits of safe mining depth ranges, distribution of risk probability at different mining depths, geological structural complexity indicators, rock mass strength parameters, and sensitivity scores of surface structures. These features form a high-dimensional feature space, with each spatial location point corresponding to a point in the feature space. The clustering algorithm then identifies natural groupings within this feature space.

[0260] The K-means clustering algorithm is mainly used for risk partitioning. K-means is a classic distance-based clustering method that divides data points into a predetermined number of clusters through iterative optimization, resulting in high similarity among points within the same cluster and low similarity between points in different clusters. The algorithm process includes: (1) initializing K cluster centers (K=3, corresponding to high, medium, and low risk levels); (2) assigning each data point to the nearest cluster center; (3) recalculating the center position of each cluster; (4) repeating steps 2 and 3 until the cluster center positions are basically stable or the maximum number of iterations is reached. The K-means algorithm is simple and efficient, suitable for processing large-scale spatial data, but it is sensitive to the selection of initial cluster centers.

[0261] To improve the stability and reliability of clustering results, the K-means++ initialization method was adopted. Unlike the traditional K-means method which randomly selects initial cluster centers, K-means++ uses weighted probability selection to maximize the distance between initial cluster centers, reducing the possibility of the algorithm getting trapped in local optima. The specific steps are: (1) Randomly select a data point as the first cluster center; (2) Calculate the squared distance from each data point to the nearest selected cluster center; (3) Set the selection probability based on the squared distance, with points farther away having a higher 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.

[0262] To further ensure the robustness of the clustering results, a multi-run optimization strategy was implemented. Specifically, the K-means++ algorithm was initialized with different random seeds and run multiple times (typically 10 to 20 times). The result with the minimum total sum of squared distances (i.e., the sum of the squared distances from points within a cluster to the cluster center) was then selected as the final clustering scheme. This strategy effectively reduces the uncertainty introduced by random initialization and improves the repeatability and reliability of the clustering results.

[0263] In certain special cases, K-means may not be suitable for handling non-spherical distributions or data with uneven density. To address this, the density-based spatial clustering algorithm DBSCAN (Density-Based Spatial Clustering of Applications with Noise) has been implemented as an alternative method. DBSCAN defines clusters based on density connectivity, can identify clusters of arbitrary shapes, and can automatically identify noise points. When geological conditions are particularly complex and the risk distribution exhibits an unconventional pattern, the algorithm automatically switches to DBSCAN to ensure the rationality and applicability of the risk zoning.

[0264] After clustering, the assessment area is divided into three levels: high-risk, medium-risk, and low-risk areas. High-risk areas typically have shallow safe mining depths, high risk probabilities, complex geological structures, or sensitive surface structures; medium-risk areas have moderate risk characteristics; and low-risk areas have deep safe mining depths, low risk probabilities, simple geological structures, or insensitive surface structures. The risk zoning results are visualized using GIS (Geographic Information System) to create an intuitive risk zoning map, providing spatial reference for decision-making.

[0265] For each risk area identified, relevant parameters are automatically extracted to form a risk source file. The risk source file is a description of the characteristics of the risk area, including the following key information: (1) Spatial range and coordinates, describing the geographical location and boundaries of the risk area; (2) Geological parameters, including lithological composition, stratigraphic structure, tectonic features (such as faults and folds), hydrogeological conditions, etc.; (3) Engineering parameters, including mining methods, mining depth, mining sequence, support methods, 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) Assessment results, including safe mining depth range, risk probability distribution, predicted surface deformation and its spatiotemporal distribution; (6) Risk level determination basis, explaining the reasons why the area is classified into a specific risk level; (7) Recommended prevention and control measures, specific risk management recommendations for the risk area.

[0266] Risk source files are stored in structured data formats such as JSON (JavaScript Object Notation) or XML (eXtensible Markup Language) to facilitate subsequent querying, analysis, and knowledge mining. For high-risk areas, more detailed files are generated, containing more parameters and deeper analysis; for low-risk areas, a relatively simplified file format is used to reduce redundant information. This differentiated approach improves efficiency while ensuring the integrity of key risk information.

[0267] The risk source file not only records the static risk status but also includes dynamic risk evolution trends. Based on the dynamic safe mining depth evaluation model developed in the preceding steps, the risk status at different points in time is predicted, forming a dynamic description of risk changes over time. This is of great significance for assessing the cumulative effects of long-term mining activities and formulating phased risk management strategies. For example, an area may be low-risk in the early stages of mining, but as mining activities continue, it may gradually transform into a medium- or high-risk area, requiring corresponding adjustments to risk management measures.

[0268] Once the risk source files are created, they are integrated into the overburden risk assessment knowledge base to achieve knowledge accumulation and sharing. The knowledge base is organized using a semantic network structure, establishing relationships between risk factors, risk manifestations, risk assessment methods, and risk control measures, supporting complex knowledge retrieval and reasoning. With the accumulation of projects, the knowledge base is continuously expanded and optimized, gradually forming a comprehensive risk knowledge system covering different geological conditions, mining methods, and building types, providing valuable experience and reference for future overburden risk assessments.

[0269] Step S4.2: Based on the risk source file, natural language generation technology is used, combined with an expert knowledge base and an engineering case library, to match the corresponding analysis methods and argumentation logic for each risk type, and automatically generate the intelligent overburden risk assessment report.

[0270] In this step, based on the previously established risk source profile and combined with professional knowledge and experience cases, an intelligent overburden risk assessment report is automatically generated. This process involves multiple technical aspects such as natural language generation, knowledge retrieval, logical reasoning, and report formatting, ultimately resulting in an assessment report that meets professional standards and has practical value.

[0271] 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 transforms the structured information in risk source files into fluent and professional text descriptions. The basic NLG process includes: content planning (determining the information to be expressed), document planning (determining the organizational structure of the information), sentence planning (designing sentence structure and vocabulary selection), and language implementation (generating the final text). The NLG approach, combining templates and rules, ensures both the professionalism and standardization of the text while providing sufficient flexibility to adapt to different risk scenarios.

[0272] The expert knowledge base is a crucial foundation for report generation. It contains theoretical knowledge, technical specifications, expert experience, and best practices in the field of overburden risk assessment, stored in a structured format. The knowledge base covers conceptual definitions (such as "overburden," "safe mining depth," and "risk probability"), theoretical models (such as "fractional-order continuum mechanics" and "peak neural networks"), assessment methods (such as "numerical simulation" and "statistical analysis"), judgment criteria (such as deformation limits and risk thresholds), and decision-making rules (such as management measures under different risk levels). The knowledge is organized using an ontological approach, establishing a semantic relationship network between concepts to support complex knowledge reasoning and retrieval.

[0273] The engineering case study database is a collection and organization of historical engineering projects, containing a large number of actual cases of risk assessment for landslides. Each case study records information such as project background, geological conditions, assessment methods, assessment results, implementation measures, and feedback on effects. Case studies are categorized by multiple dimensions, including geological type (e.g., "fault-affected areas," "weak interlayer areas"), engineering type (e.g., "coal mining," "metal mining"), and risk type (e.g., "geological structural risk," "mining technology risk"), facilitating retrieval and application. The value of these engineering case studies lies in providing accumulated practical experience, offering valuable lessons and references for current projects.

[0274] Risk types refer to the different natures and causes of overburden risks, reflecting the essential characteristics and impact mechanisms of risks. Common risk types include: (1) geological structure risks, which are related to geological structures such as faults and folds, such as fault transmission risks and stress concentration risks; (2) rock mass characteristic risks, which are related to rock mass strength and deformation characteristics, such as weak interlayer risks and high-pressure water risks; (3) mining technology risks, which are related to mining methods and sequences, such as over-extraction risks and insufficient safety coal pillar risks; (4) surface building risks, which are related to the type and sensitivity of surface buildings, such as ancient building risks and high-rise building risks; (5) cumulative effect risks, which are related to the cumulative impact of long-term mining activities, such as the risk of superimposed impacts from multiple mining operations. Different risk types have different manifestations and impact mechanisms, requiring the use of targeted analysis methods and evaluation standards.

[0275] For each risk type, corresponding analytical methods and argumentation logic were matched. Analytical methods refer to the technical approaches and tools used to assess risk, including qualitative analysis (such as expert judgment, case comparison, and influencing factor analysis) and quantitative analysis (such as numerical simulation, statistical analysis, and reliability analysis). Argumentation logic refers to the reasoning framework and argumentation structure in the analytical process, ensuring the coherence of the analysis and the reliability of the conclusions. Through semantic matching algorithms, the risk characteristics of the current project are compared with the content in the expert knowledge base and engineering case library to find the most similar knowledge points and cases, and to extract applicable analytical methods and argumentation logic.

[0276] The matching process is based on semantic similarity calculation and involves natural language processing techniques. First, key information in the risk source files is converted into semantic vectors. Then, the cosine similarity with entries in the knowledge base and case library is calculated to identify the most relevant content. For example, for risks in fault-affected areas, theoretical knowledge and analytical methods related to faults are retrieved from the knowledge base, while engineering cases with similar geological conditions are searched in the case library to form a comprehensive analytical framework. To improve matching accuracy, weighted similarity calculation is used, assigning higher weights to key features (such as fault type, fault displacement, and fracture zone width) to ensure the relevance and applicability of the matching results.

[0277] 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.

[0278] 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.

[0279] 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.

[0280] 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.

[0281] 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.

[0282] 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.

[0283] like Figure 3 As shown, the present invention also provides a mineral resource overburden risk assessment system, comprising:

[0284] 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.

[0285] 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.

[0286] 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.

[0287] 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.

[0288] 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.

[0289] 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.

[0290] 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.

[0291] 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.

[0292] 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. 2By 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.

[0293] 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.

[0294] 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 for assessing the risk of mineral resource overlay, characterized in that, Includes the following steps: Acquire multi-source heterogeneous geological data, including geological exploration data, borehole core data, geophysical exploration data, historical assessment cases, and mining data from surrounding mines. Deeply fuse this multi-source heterogeneous geological data using a robust neural network training method with game theory gradient control to obtain a fused dataset. The process includes: treating the multi-source heterogeneous geological data as game participants, employing Nash equilibrium theory, and iteratively solving to achieve the optimal weight allocation state for each data source, determining the contribution weight of each data source to the final fusion result, and obtaining an initial weight allocation; based on the initial weight allocation, constructing a robust neural network architecture including residual connections and batch normalization, introducing Gaussian noise and data augmentation strategies to improve the network's anti-interference ability, calculating the optimal contribution weight of each data source under the Nash equilibrium state, and obtaining the optimal weight allocation state; weighting the data sources of the multi-source heterogeneous geological data according to the optimal weight allocation state, and completing the fusion training through forward and backward propagation of the neural network to obtain the fused dataset. Based on the fused dataset, the pattern recognition and association analysis capabilities of the large model are utilized, combined with manifold perception regularization technology, to perform in-depth mining and reasoning on high-dimensional geological data, and to identify geological structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures. Based on the geological structural features, a dynamic safe mining depth evaluation model is established using a fractional-order peak differential equation neural network with efficient training of adjoint parameters. The model 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. Based on the safe mining depth range and risk probability, an assessment report generation module is constructed that includes core risk point analysis, key technology demonstration and forward-looking risk warning, and automatically generates an intelligent overburden risk assessment report.

2. The method according to claim 1, characterized in that, The construction of a robust neural network architecture, including residual connections and batch normalization, includes: Based on the initial weight allocation, multi-source heterogeneous geological data is used 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 stable training of the deep network. Gaussian noise perturbation and data augmentation strategies, including random rotation, scaling and translation transformations, are introduced into the input layer of the neural network to enhance the network's robustness to data perturbation.

3. The method according to claim 1, characterized in that, The method combines manifold sensing regularization technology to perform in-depth mining and reasoning on high-dimensional geological data, identifying geological structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures, including: Based on the geological feature points in the fused dataset, a k-nearest neighbor graph is constructed and the k nearest neighbors of each feature point are calculated. An adjacency matrix between feature points is established to obtain the manifold structure representation of the data. Based on the manifold structure representation of the data, the local density and curvature information of feature points in the manifold space are calculated, and the manifold Laplacian matrix is ​​constructed as a regularization term to constrain the data to maintain its intrinsic geometric structure during the training of large models, thereby obtaining regularization constraint parameters. The regularization constraint parameters are added to the loss function of the large model based on the Transformer architecture. The model parameters are optimized by multi-head self-attention mechanism and backpropagation algorithm to identify the spatial distribution of stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures, thereby obtaining the geological structural features.

4. The method according to claim 1, characterized in that, The method of establishing a dynamic safe mining depth evaluation model using a fractional-order peak differential equation neural network with efficient adjoint parameter training includes: Based on the geological structural characteristics, the memory effect and long-range correlation of geological materials are described by fractional differential equations, and a peak neural network is designed using a leakage integral distribution model to establish a dynamic safe mining depth evaluation model. Based on the dynamic safe mining depth evaluation model, a multi-constraint optimization model is established. The constraints include that the maximum surface subsidence, maximum tilt, and maximum curvature deformation do not exceed the safety threshold. The particle swarm optimization algorithm is used to solve the constrained optimization problem to obtain the safe mining depth range and risk probability.

5. The method according to claim 1, characterized in that, The module for generating assessment reports, which includes core risk point analysis, key technology demonstration, and forward-looking risk warning, includes: Based on the safe mining depth range and risk probability, cluster analysis is used to divide the assessment area into risk areas of different levels, including high-risk areas, medium-risk areas and low-risk areas. Geological parameters, calculation model parameters and assessment results of each risk area are extracted to form a risk source file. Based on the risk source file, natural language generation technology is used, combined with an expert knowledge base and an engineering case library, to match corresponding analysis methods and argumentation logic for each risk type, and automatically generate the intelligent overburden risk assessment report.

6. A mineral resource overlay risk assessment system, characterized in that, include: The multi-source data fusion module 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 from surrounding mines. 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 fused dataset. The method includes: treating the multi-source heterogeneous geological data as game participants, using Nash equilibrium theory, iteratively solving to achieve the optimal weight allocation state for each data source, determining the contribution weight of each data source to the final fusion result, and obtaining an initial weight allocation; based on the initial weight allocation, constructing a robust neural network architecture including residual connections and batch normalization, introducing Gaussian noise and data augmentation strategies to improve the network's anti-interference ability, calculating the optimal contribution weight of each data source under the Nash equilibrium state, and obtaining the optimal weight allocation state; weighting the data sources of the multi-source heterogeneous geological data according to the optimal weight allocation state, and completing the fusion training through forward and backward propagation of the neural network to obtain the fused dataset. The geological structure identification module 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 structural features including stratigraphic interfaces, lithological distribution, fault fracture zones, and joints and fissures. The safe mining depth calculation module is used to establish a dynamic safe mining depth evaluation model using a fractional-order 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 is used to automatically generate an intelligent overburden risk assessment report that includes core risk point analysis, key technology demonstration, and forward-looking risk warning, based on the safe mining depth range and risk probability.

Images