Metal mine dynamic disaster intelligent prevention and control method and system based on environment reconstruction and disaster energy utilization

By combining multi-source sensing data with intelligent algorithms, the full life-cycle perception and modeling of dynamic disasters in deep metal mines has been realized, enabling intelligent early warning and risk zone delineation. This has solved the problems of lagging prevention and control and insufficient energy utilization in existing technologies, and achieved systematic prevention and control and energy utilization.

CN123072176APending Publication Date: 2026-11-10CENT SOUTH UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611404538.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-10
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies for the prevention and control of dynamic disasters in deep metal mines suffer from problems such as lagging monitoring and early warning, passive prevention and control measures, serious information silos, and insufficient utilization of disaster energy, resulting in unsystematic prevention and control and high energy consumption.

Method used

A multi-source high-dimensional disaster-prone feature matrix is ​​constructed using multi-source sensing data. Deep belief networks, convolutional neural networks, and long short-term memory networks are used to identify and warn of disaster-prone factors. Combined with multi-objective genetic algorithms, prevention and control schemes are generated to achieve intelligent prevention and control and disaster energy utilization throughout the entire life cycle.

Benefits of technology

It has achieved full-domain perception and modeling of disaster-prone environments, multi-scale intelligent early warning and precise three-dimensional risk zone delineation, proactive prevention and early warning response to metal mine disasters, and effective utilization of the disaster energy of deep high-energy-storage rock masses.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072176A_ABST
    Figure CN123072176A_ABST
Patent Text Reader

Abstract

The application provides a metal mine dynamic disaster intelligent prevention and control method and system of environment reconstruction and disaster energy utilization, global and full life cycle perception and modeling of disaster-pregnant environment and disaster process; multi-scale intelligent early warning and three-dimensional risk area accurate division are realized through autonomous identification of disaster-pregnant factors, and then scientific decision and scheme automatic issuing are realized based on the three-dimensional strategy of "mining optimization-environment reconstruction-disaster energy utilization", active prevention and control and early warning response of metal mine disasters are realized, linkage of monitoring and control means is realized, and effective utilization of deep high-energy rock mass disaster energy is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an intelligent prevention and control method and system for dynamic disasters in metal mines that integrates environmental reconstruction and disaster energy utilization. Background Technology

[0002] As global metal mining continues to extend into deeper and higher-stress areas, the mining environment is becoming increasingly complex. Deep surrounding rock faces multiple challenges, including significantly increased triaxial stress levels in the original rock, altered rock properties due to rising deep geothermal temperatures, and the coupling effect of high-pressure groundwater systems. Furthermore, frequent blasting and tunneling during mining activities cause continuous redistribution of the stress field, creating an extremely complex dynamic response environment. Under the combined effects of these multiple factors, the mechanical behavior of deep rock masses exhibits high nonlinearity, leading to frequent dynamic disasters such as rockbursts, rock bursts, large-scale roof collapses, and extensive spalling. These disasters are characterized by their suddenness, high concealment, and high destructiveness, seriously threatening safe mine production.

[0003] Currently, the main shortcomings of dynamic disaster prevention and control measures in deep metal mines are as follows: First, monitoring and early warning methods are mostly reactive. While existing microseismic, stress, and deformation monitoring systems can provide basic data, early warning thresholds are usually set based on field experience, lacking the ability to adaptively adjust to different geological conditions and mining stages. Alarms are often issued only after obvious disaster signs, leading to delayed prevention and control. Second, prevention and control measures lack systematicity and specificity. Common measures such as anchor bolt (cable) support, grouting reinforcement, and large-diameter pressure relief hole construction are mostly implemented passively after a disaster or in the event of a potential disaster. Some active pressure relief or pre-splitting blasting schemes are based on experience-based design and fail to achieve dynamic linkage and optimization with real-time monitoring data and the dynamic disaster's gestation state. Third, information silos are severe, failing to form a closed-loop management system. Microseismic monitoring, geostress monitoring, geological modeling, and production scheduling are often managed by different departments, with inconsistent data standards, making it difficult to achieve an integrated intelligent closed loop of "monitoring-early warning-decision-control." Fourth, the potential utilization value of disaster energy is overlooked. The large amount of elastic potential energy stored in deep, hard, and brittle rock masses is the energy source for rockbursts. Traditional energy management concepts mainly focus on "peak shaving and energy reduction" and preventing concentrated energy release, neglecting the possibility of guiding this energy to participate in rock breaking operations under controllable conditions. This not only wastes energy but also means that breaking deep, hard rocks still requires a large amount of mechanical or fossil fuel energy, increasing mining energy consumption and costs. Summary of the Invention

[0004] This application proposes an intelligent prevention and control method and system for dynamic disasters in metal mines that integrates environmental reconstruction and disaster energy utilization, which can solve one of the problems existing in the background technology.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] Firstly, a smart prevention and control method for dynamic disasters in metal mines, which combines environmental reconstruction and disaster energy utilization, is provided, including:

[0007] Obtain multi-source sensing data;

[0008] Based on the multi-source sensing data, a multi-source high-dimensional disaster-pregnancy feature matrix is ​​constructed that integrates macroscopic spatial topology and long- and short-term time dependencies.

[0009] Using a pre-trained autonomous disaster-prone factor identification and multi-scale intelligent early warning model, the multi-source high-dimensional disaster-prone feature matrix is ​​processed to identify the disaster-prone pattern classification, dynamic risk probability, and dynamic risk early warning level of a spatial region. The spatial region is then divided into risk zones to obtain the three-dimensional spatial range of potential disaster source areas. The autonomous disaster-prone factor identification and multi-scale intelligent early warning model includes a cascaded structure of: a Deep Belief Network (DBN), a 3D Convolutional Neural Network (3D-CNN), a Long Short-Term Memory (LSTM) network, and a classification head. The Deep Belief Network is used to extract and represent the multi-source high-dimensional disaster-prone feature matrix on a single grid. The underlying potential correlation features of the internal physical damage evolution state are obtained. The three-dimensional convolutional neural network is used to perform deep decoupling and fusion of spatiotemporal features on the three-dimensional feature tensor containing spatial coordinates corresponding to the underlying potential correlation features to obtain a global spatial feature sequence characterizing the spatial damage evolution law of the rock mass. The long short-term memory network is used to capture the nonlinear temporal evolution law of disaster-inducing factors on the time axis of the global spatial feature sequence to obtain a spatiotemporal fusion feature vector. The classification head is used to obtain the disaster-inducing pattern classification, dynamic risk probability results and dynamic risk warning level of the potential disaster source area based on the spatiotemporal fusion feature vector.

[0010] Using a four-dimensional digital twin as the physical simulation environment, and employing a multi-objective genetic algorithm, under the condition of satisfying safety and mechanical constraints, a comprehensive risk, efficiency and disaster energy conversion rate are considered to generate a prevention and control optimization scheme for the three-dimensional spatial range of potential disaster source areas, which includes mining optimization, environmental reconstruction and disaster energy utilization.

[0011] In one possible design approach of the first aspect, the multi-source high-dimensional pregnancy feature matrix is:

[0012] In the formula, For the current target time step t, a high-dimensional feature tensor that integrates macroscopic spatial topology and long- and short-term time dependencies; A global three-dimensional spatial mesh set based on a four-dimensional digital twin; For indexing historical continuous time series within a sliding time window; For a specific moment And the absolute coordinates are The feature column vector of a single grid node; The specific time step of the sliding time window.

[0013] In one possible design of the first aspect, the deep belief network is composed of several multi-layer Restricted Boltzmann Machines (RBMs) stacked together. The deep belief network learns the joint probability distribution among variables by minimizing the system energy function. The multi-layer RBMs include cascaded visible layers and several hidden layers. The energy function of a single-layer RBM... for:

[0014] In the formula, and These are the visible layer and hidden layer state vectors, respectively. and These represent the total number of nodes in the visible layer and the hidden layer, respectively. and The first The visible node and the first The state values ​​of the hidden nodes; and These are the corresponding bias terms; This represents the connection weights between the visible and hidden layers.

[0015] In one possible design approach of the first aspect, the three-dimensional convolutional neural network utilizes multiple three-dimensional convolutional kernels to perform sliding window operations in stress and energy release fields to autonomously identify various spatial disaster-causing anomaly structures. The forward propagation formula of the three-dimensional convolutional layer of the three-dimensional convolutional neural network is:

[0016] In the formula, This is the channel index of the input feature map of the previous layer. For the first The output of the convolutional layer is the first A three-dimensional feature map; This is the 3D feature map input from the previous layer; This represents the corresponding 3D convolution kernel weight matrix; * indicates 3D convolution operation; For bias terms; It is a non-linear activation function;

[0017] After the final 3D feature map is downsampled by the pooling layer, the 3D convolutional neural network outputs the global spatial feature sequence.

[0018] In one possible design approach of the first aspect, the cell state update and hidden layer output formulas of the Long Short-Term Memory (LSTM) network core are as follows:

[0019] In the formula, The spatial feature sequence input at the current time; and These are the hidden layer states of the previous and current time steps, respectively. and These are historical and updated cell states, used for long-term memory of pregnancy and disaster information; , , These are the activation vectors for the forget gate, input gate, and output gate, respectively. ; It is the Sigmoid activation function. Represents the product of matrix elements.

[0020] In one possible design approach of the first aspect, the spatial region is divided into risk zones to obtain the three-dimensional spatial extent of the potential disaster source area, specifically including:

[0021] If at a high-risk grid point Spatial search radius The number of grid points contained within is greater than or equal to the set minimum density threshold. Then that point The density constraint formula for a node identified as a cluster core is:

[0022] In the formula, For the initial set of high-risk discrete points, For connected points within the neighborhood, This represents the total number of grid elements within the set;

[0023] Starting from the core point, perform a union operation on all grid cells with achievable density to form clusters. Furthermore, the actual physical three-dimensional envelope occupied by this cluster in the four-dimensional digital twin base is defined as a potential disaster source area. Potential disaster source area The formula for generating spatial boundaries is:

[0024] In the formula, The center of mass is The three-dimensional volume of a single independent mesh;

[0025] Automatic calculation of disaster source area The minimum bounding box geometric span in a three-dimensional coordinate system;

[0026] The following scales are used to determine the risk level based on the geometric span of the smallest bounding box: if the high-risk area clustered out has a very small span, it is judged as a stope level; if the high-risk area spreads out in a continuous manner on the horizontal plane, spanning several stopes, but does not penetrate vertically, it is judged as a mid-level level; if the high-risk area penetrates deep strata in the vertical direction, it is judged as a mining area level.

[0027] Based on the scale determination results and the exclusive thresholds that match the scale determination results, a graded early warning signal is automatically generated;

[0028] By performing density clustering and connectivity analysis on high-risk spatial grid cells that exceed safety thresholds, the three-dimensional spatial range of potential disaster source areas is dynamically delineated, and their absolute coordinate boundaries, disaster-causing energy volume, and spatial topological relationship with current mining projects are clarified.

[0029] In one possible design approach of the first aspect, the safety mechanical constraints specifically include: surrounding rock strength and yield criterion constraints, allowable deformation and support coordination constraints, and dynamic load safety constraints for induced work; the multi-objective optimization function in the multi-objective genetic algorithm includes: a first objective function to minimize the residual disaster-causing energy density in the disaster source area, a second objective function to minimize the comprehensive cost of engineering treatment under the premise of ensuring production capacity, and a third objective function to force the elastic potential energy of a specific high-energy-storage hard rock layer to be converted into effective energy release for fracturing hard rock or pressure relief work to the greatest extent possible;

[0030] The multi-objective genetic algorithm includes:

[0031] Transforming physical control strategies into multidimensional gene chromosome sequences ;

[0032] Introducing Boolean Discrete Control Vectors A global integrated prevention and control scheme vector is generated from the multidimensional gene chromosome sequence;

[0033] Under the boundary constraints of the four-dimensional digital twin base, a three-dimensional multi-objective optimization function model based on the global integrated prevention and control scheme vector is constructed;

[0034] A non-dominated sorting multi-objective genetic algorithm is employed to deduce the objective function in parallel. The algorithm iteratively generates a Pareto optimal solution set covering a multi-scale solution space through crossover and mutation operations, and by introducing a non-dominated sorting mechanism based on crowding distance. Subsequently, digital twin feature parameters of the current working condition are extracted in real time, and a conditional state machine is used to autonomously match the optimal combination vector best suited to the current environment from the Pareto front. The selected global optimal parameter sequence Automatically compile and convert these instructions into industrial control commands; then send these commands to the downhole intelligent terminal for execution.

[0035] In one possible design approach of the first aspect, the multi-source sensing data includes: energy index EI, fracture scale evolution value, spectral characteristics, microseismic spatial concentration, stress concentration factor, surrounding rock deformation rate, and geological fault geometric sensitivity weight.

[0036] Secondly, a smart prevention and control system for dynamic disasters in metal mines, which integrates environmental reconstruction and disaster energy utilization, is provided. The system includes one or more functional units or modules for performing the method described in any of the possible implementations of the first aspect.

[0037] Thirdly, an electronic device is provided, comprising: a processor, and a memory coupled to the processor, the memory for storing a computer program; the processor for executing the computer program stored in the memory such that the electronic device performs the method as described in any possible implementation of the first aspect.

[0038] Beneficial effects:

[0039] Based on the above technical solutions, the system achieves full-domain and full-lifecycle perception and modeling of the disaster-prone environment and disaster process; it realizes multi-scale intelligent early warning and precise three-dimensional risk zone delineation by autonomously identifying disaster-prone factors; and it then makes scientific decisions and automatically issues solutions based on the three-dimensional strategy of "mining optimization - environmental reconstruction - disaster energy utilization". This enables proactive prevention and early warning response to metal mine disasters, linkage between monitoring and governance methods, and effective utilization of disaster energy in deep high-energy-storage rock masses. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the overall process of an embodiment of this application.

[0042] Figure 2 This is a logic diagram of the autonomous identification and intelligent early warning algorithm for pregnancy-related factors in an embodiment of this application.

[0043] Figure 3 This is a topology diagram of the intelligent sensing network for deep rock dynamic disasters according to an embodiment of this application.

[0044] Figure 4 This is a schematic diagram of the prevention and control strategy decision logic tree in an embodiment of this application.

[0045] Figure 5 This is a closed-loop control principle diagram of the disaster-sensitive environment feedback control system according to an embodiment of this application. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] It should be noted that although functional modules are divided in the device schematic diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification and the above-mentioned figures are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0048] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0049] The purpose of this invention is to provide an intelligent prevention and control method and system for dynamic disasters in metal mines based on environmental reconstruction and disaster energy utilization. Addressing the problems of passive prevention and control methods, delayed early warning response, fragmented monitoring and management approaches, and insufficient energy utilization in deep, high-energy-storage rock masses in existing technologies, this invention aims to achieve comprehensive, full-lifecycle perception and modeling of the disaster-prone environment and disaster processes; to achieve multi-scale intelligent early warning and precise three-dimensional risk zone delineation through autonomous identification of disaster-prone factors; and to conduct scientific decision-making and automatic scheme distribution based on a three-dimensional strategy of "mining optimization—environmental reconstruction—disaster energy utilization," while utilizing the feedback regulation mechanism of disaster-sensitive environments and quantitative assessment methods to achieve a closed-loop optimized proactive prevention and control goal.

[0050] like Figure 1 As shown, according to one aspect of this embodiment, an intelligent prevention and control method for dynamic disasters in metal mines is provided, which includes the following consecutive steps:

[0051] Step S1: Holographic perception and dynamic construction of multi-source heterogeneous data;

[0052] First, a high-precision sensing network is constructed in the deep mining area of ​​the metal mine and its overlying surrounding rock. This network integrates microseismic sensor arrays, ground stress monitoring stations, acoustic emission probes, surrounding rock displacement gauges, and pore water pressure gauges. Various monitoring data are transmitted to the ground data center in real time through underground industrial Ethernet or fiber optic communication systems.

[0053] The data acquisition covers the three-dimensional stress field and its rate of change around the roadway and mining area, surrounding rock convergence and displacement deformation data, production data from blasting and tunneling operations, and acoustic emission and microseismic waveform signals excited by microfractures in the rock mass. For the dynamic response characteristics, the system solves for the four-dimensional coordinates of the seismic source in the non-uniform rock mass based on ray tracing and residual minimization objective functions. Its positioning calculation formula is:

[0054] In the formula, The objective function for the travel time residuals in microseismic source localization. To minimize the objective operator, Let i be the initial initiation time of the earthquake source to be determined, and let i be the number of microseismic sensors that were effectively triggered. For the first The initial triggering time of the micro-vibration sensor, where N is the effective number of triggering sensors. For the spatial variable wave velocity field updated in real time in the four-dimensional digital twin base, Spatial coordinates of the earthquake source Known fixed physical coordinates of the i-th microseismic sensor Nonlinear ray propagation path between them These are the cumulative arc length coordinates along the propagation path curve. The arc length of the first type of line integral is defined as the infinitesimal element along the propagation path.

[0055] Among them, the spatial variable wave velocity field Three-dimensional absolute coordinates based on downhole microseismic sensor network The system generates microseismic waveform signals acquired in real time by various sensors: Based on a three-dimensional geological exploration model and prior parameters from indoor ultrasonic rock testing, the system assigns initial reference wave velocities to the rock layer grids within the digital twin base, and determines the arrival time of the first wave picked up by the microseismic sensor network to trigger the initial motion. Substituting into the ray tracing equation, the actual reference wave velocity of each ray path is calculated and calibrated. Based on this absolute calibration, the system continuously collects waveform travel time residual data captured by each microseismic sensor in passive rock mass micro-fracture events. Using a joint iterative reconstruction algorithm, the slowness of the global grid is dynamically inverted and corrected at high frequency. This allows the wave velocity anomaly evolution induced by stress concentration or damage fracture in local rock mass to be mapped in real time and with high fidelity into a non-uniform three-dimensional wave velocity field.

[0056] Simultaneously, the microseismic radiation energy was calculated using the time-domain integral formula. :

[0057] In the formula, The density of the rock in the fracture zone, This represents the propagation speed of elastic waves in rock mass. Let r be the time series of particle vibration velocity received by the sensor and the instrument response, where r is the propagation distance from the source to the sensor. t represents the duration of the earthquake source.

[0058] Among them, the time series of particle vibration velocity received by the sensor after the instrument response. Essentially, it refers to the oscillation velocity of surrounding rock particles captured by physical sensors in the time domain, which is a function of time; while the aforementioned spatially variable wave velocity field This refers to the propagation velocity of elastic waves in a non-homogeneous rock mass, which is a function of spatial coordinates. The system is first based on a spatially variable wave velocity field. By performing nonlinear spatiotemporal inversion, the coordinates of the seismic source are precisely located, thereby deriving the three-dimensional spatial path distance from the seismic source to a specific sensor. Subsequently, the system will record the distance. As a geometric diffusion compensation coefficient, it is introduced into the equation of particle vibration velocity. The time-domain square integral equation is thus constructed, allowing for a high-fidelity calculation of the absolute radiative kinetic energy released by the microsource. .

[0059] Secondly, the aforementioned multi-source heterogeneous data undergoes preprocessing and spatial mapping. Wavelet transform is used to filter out environmental noise. For the spatially unevenly distributed discrete monitoring data of geostress and pore pressure, ordinary kriging is employed for feature scale unification and spatial interpolation completion. For any point to be estimated in the grid... Attribute estimates The calculation is performed using an unbiased estimator:

[0060] In the formula, n is the total number of known measuring points. Let m be the spatial location of the known measuring point. To ensure an unbiased estimate and minimize the variance of the estimate for the spatial location of the u-th known measurement point, and The weighting coefficients for the corresponding measurement points are obtained by solving the Kriging equations, which consist of two core equations:

[0061] In the formula, The semi-variogram values ​​between known measuring points characterize the autocorrelation of the spatial properties of the rock mass; The semivariogram values ​​between the known points and the points to be estimated; These are Lagrange multipliers. By strictly adhering to the mathematical constraints of minimizing the estimation variance and normalizing the weights, the artificial smoothing effect and statistical bias caused by traditional spatial interpolation are effectively eliminated, ensuring that the inverted data field has extremely high mathematical and statistical fidelity.

[0062] For high-frequency, non-stationary microvibrations and acoustic emission waveforms, the system uses wavelet transform for threshold denoising. For the clean waveform, the system does not directly use the amplitude sequence, but instead performs parallel dynamic inversion of the radiated energy E, energy index EI, and b-value for calculation.

[0063] The system performs time-domain-frequency domain integration on the effective microseismic signals to calculate the radiation energy E and seismic moment M of a single rupture event. To eliminate the masking effect of the absolute magnitude on the local stress concentration state, the system extracts the energy index EI. Its calculation formula is as follows:

[0064] In the formula, c and d are local geological empirical constants fitted based on historical microseismic data of the target mine. When EI>1, it indicates that the actual released energy in the grid area is higher than the average energy under the same magnitude, suggesting a high-stress locked state.

[0065] Simultaneously, the system extracts the magnitude sequence of all microseismic events within a set sliding time window (e.g., the past 24 hours), and dynamically solves for the b-value characterizing the microcrack propagation scale using the maximum likelihood estimation method (MLE) based on the Gutenberg-Richter relation:

[0066] In the formula, e is the natural constant. The average magnitude of the microseismic events within the time window. This represents the minimum complete magnitude under the current sensor network monitoring capabilities. The continuous decrease in the b-value is directly mapped as a precursory feature of micro-cracks within the rock mass leading to macroscopic, large-scale fracture penetration.

[0067] For discrete deformation readings of surrounding rock obtained by lidar or multi-point displacement gauges, since mechanical vibrations in the field can easily introduce burr noise, the system first uses a moving average filter to filter the original displacement sequence. Perform time-domain smoothing preprocessing:

[0068] In the formula, To be at the current central time node The smooth net displacement value output after filtering; The original displacement sequence is affected by downhole transient environmental noise; k is the center discrete time index of the current data node to be estimated; The preset single-sided sliding time window step is used to define the range of reference data before and after the center point; 2 +1 represents the total number of valid sampling points within the sliding time window; j represents the relative time offset index within the time window, with a value range of [- , By introducing a symmetrical time window observation matrix, transient displacement data jumps caused by heavy equipment operation or rockfall impacts were effectively suppressed.

[0069] After obtaining the smooth displacement curve, the system uses the first-order forward difference formula to calculate the surrounding rock deformation rate of the spatial grid node at a specific time. :

[0070] In the formula, Due to time difference.

[0071] Deformation rate By eliminating the static baseline data accumulated from historical slow creep in the absolute displacement, the nonlinear accelerated rheological characteristics of the rock mass under recent mining disturbances are highlighted.

[0072] For discrete production logs such as those documented for blasting and tunneling, the system preprocesses them from text modality into a four-dimensional spatiotemporal constrained Boolean matrix, and extracts the three-dimensional centroid coordinates of each operation. At the moment of detonation Generate constraint functions :

[0073] In the formula, The effective physical sweep radius of the dynamic load disturbance caused by blasting. This represents the hysteresis time window for stress redistribution. When the Mask value of a grid node is 1, the system will assign additional dynamic disturbance triggering weights to that grid during subsequent disaster-prone mode classification.

[0074] Finally, combining the mine geological model (including strata, fault structures, lithology, etc.) with the current mining space topology and process parameters, a spatial base is generated using mesh generation technology to construct a four-dimensional model of the deep rock mass, encompassing "static geology, dynamic stress, and dynamic response." Digital twin model: The model is driven by the time dimension t. It uses an ensemble Kalman filter (EnKF) data assimilation algorithm to access and fuse monitoring data in real time. The standard spatiotemporal features generated after preprocessing such as wavelet denoising, Kriging interpolation, moving average, and dimensionless transformation are used as the discrete observation vectors in the field. Injected into the prior full-field stress state vector predicted by numerical simulation. In the process, dynamic inversion yields posterior analytical values. The Kalman filter data assimilation equation is as follows:

[0075] Wherein, the Kalman gain matrix The assimilation weight of the system for the measured data is determined, and its calculation formula is as follows:

[0076] In the formula, H is the observation operator matrix. The background error covariance matrix is ​​calculated from the set. This represents the sensor measurement error covariance matrix. Based on this complete data assimilation process, the digital twin model can dynamically update the stress redistribution and surrounding rock stability evolution state with high fidelity, providing a rigorous physical boundary constraint and quantitative basis for subsequent autonomous identification of disaster-causing factors and multi-scale intelligent early warning.

[0077] Step S2: Autonomous identification of pregnancy-related factors and multi-scale intelligent early warning;

[0078] First, extract the absolute spatiotemporal coordinates from the multi-source monitoring data. The energy index EI, b-value evolution, spectral characteristics, and microseismic spatial concentration obtained from the pure waveform transformation, along with the stress concentration factor and smoothed net displacement value obtained from the full-field posterior stress state vector solution after dynamic injection of the prior model using ensemble Kalman filtering (EnKF), are also analyzed. It is mapped to the surrounding rock deformation rate through the first-order time-domain difference. By incorporating spatiotemporal physical characteristics and geological fault structural elements, a sliding time window of specific step length L is introduced to sequentially stack spatial features within consecutive time steps, thereby constructing a multi-source high-dimensional disaster-prone feature matrix sample. Its expression is:

[0079] In the formula, The input sample is a high-dimensional feature tensor that integrates macroscopic spatial topology and long- and short-term time dependencies at the current target time step t. A global three-dimensional spatial mesh set based on a four-dimensional digital twin; For indexing historical continuous time series within a sliding time window; For a specific moment And the absolute coordinates are The feature column vector of a single grid node, where EI is the energy index and b is the fracture-scale evolution value. For spectral characteristics, For microseismic spatial concentration, The stress concentration factor is... The deformation rate of the surrounding rock. This represents the quantified geometric sensitivity weights for geological faults.

[0080] To eliminate dimensional interference between heterogeneous data and extract underlying latent correlation features, a high-dimensional feature matrix is ​​input into a deep belief network (DBN) for unsupervised pre-training. For example... Figure 3 As shown, a DBN is composed of multiple layers of Restricted Boltzmann Machines (RBMs) stacked together. It learns the joint probability distribution among variables by minimizing the system energy function. The energy function of a single layer of RBM is... The calculation formula is:

[0081] In the formula, and These are the visible layer and hidden layer state vectors, respectively. and These represent the total number of nodes in the visible layer and the hidden layer, respectively. and The first The visible node and the first The state values ​​of the hidden nodes; and These are the corresponding bias terms; The connection weights between the visible and hidden layers are defined. Through layer-by-layer feature dimensionality reduction and weight initialization using DBN, a highly robust disaster-causing latent feature vector is output, effectively avoiding gradient vanishing and local optima traps in deep network training. The final output is a low-dimensional, high-order one-dimensional feature vector representing the physical damage evolution state within the single grid.

[0082] Secondly, such as Figure 2 As shown, to address the non-uniformity of the disaster-prone environment in deep rock masses in three-dimensional space, deep decoupling and fusion of spatiotemporal features are performed. Based on the topological mesh index of a four-dimensional digital twin base, and strictly following the corresponding real physical spatial coordinates, the latent feature vectors output by the DBN are restored to include spatial coordinates. The three-dimensional feature tensor is obtained and input into a three-dimensional convolutional neural network (3D-CNN) to extract spatial disaster-causing features. Multiple three-dimensional convolutional kernels are used to perform sliding window operations in stress and energy release fields to autonomously identify spatial disaster-causing anomalies such as high-stress-concentration gradient bands and microseismic energy-dense kernels. The forward propagation formula for the three-dimensional convolutional layer is:

[0083] In the formula, This is the channel index of the input feature map of the previous layer. For the first The output of the convolutional layer is the first A three-dimensional feature map; This is the 3D feature map input from the previous layer; This represents the corresponding 3D convolution kernel weight matrix; * indicates 3D convolution operation; For bias terms; The activation function is nonlinear. After downsampling through pooling layers, the CNN outputs a global spatial feature sequence characterizing the spatial damage evolution of the rock mass.

[0084] Finally, to capture the nonlinear temporal evolution of disaster-causing factors along the time axis t, the spatial feature sequence output by the CNN is connected to a Long Short-Term Memory (LSTM) network. LSTM, through its internal gating mechanisms (forget gate, input gate, output gate), filters and retains long-term historical disaster-causing features, eliminates redundant interference, and achieves accurate prediction of the temporal evolution trend of dynamic disasters. The core cell state update and hidden layer output formulas of LSTM are as follows:

[0085] In the formula, The spatial feature sequence input at the current time; and These are the hidden states of the previous time step and the current time step, respectively; and These are historical and updated cell states, used for long-term memory of pregnancy and disaster information; , , These are the activation vectors for the forget gate, input gate, and output gate, respectively. and These are the weight matrix and bias term of the forget gate, respectively; and These are the weight matrix and bias term of the candidate cell state calculation unit, respectively; It is the Sigmoid activation function. Represents the product of matrix elements.

[0086] The spatiotemporal fusion feature vector output by the LSTM is then fed into a fully connected layer and a Softmax classifier to calculate the dynamic risk probability of dynamic disasters such as rockbursts occurring at the working face within a specific future time period. It simultaneously outputs the dominant disaster-generating mode labels (such as tectonic activation type, high stress concentration type, or high energy accumulation type), and uses the DBSCAN spatial topology clustering algorithm to connect and reconstruct them into three-dimensional potential disaster source areas. This involves risk zoning, and its core mathematical model is based on spatial... - The density threshold determination for the neighborhood, the key determination and boundary generation formulas are as follows:

[0087] If a certain high-risk grid point Spatial search radius The number of grid points contained within is greater than or equal to the set minimum density threshold. If the value is true, then the point is determined to be a cluster core point, and its density constraint formula is:

[0088] In the formula, For the initial set of high-risk discrete points, For connected points within the neighborhood, This represents the total number of grid elements within the set.

[0089] Secondly, starting from the core point, the system performs a union operation on all grid cells with achievable density to form clusters. Furthermore, the actual physical three-dimensional envelope occupied by this cluster in the four-dimensional digital twin base is defined as a potential disaster source area. The formula for generating its spatial boundary is:

[0090] In the formula, The center of mass is The three-dimensional volume of a single, independent grid. The system is based on the disaster source area. Automatically calculate the minimum bounding box geometric span of the region in a three-dimensional coordinate system. If the high-risk areas identified by clustering have a very small span ( Standard span of the mining area ,and Segmented Standard Height The system classifies it as a mining area level; if the high-risk area spreads horizontally across several mining areas but does not extend vertically (…), it is classified as a mining area level. Mid-section standard height The system classifies it as a mid-level zone; if the high-risk area penetrates deep strata vertically ( Mid-section standard height The system judged it to be at the mining area level.

[0091] Ultimately, based on the judgment results at this scale and their matching specific thresholds, the system automatically generates red, orange, yellow, and blue graded early warning signals, which are then visualized and rendered in a four-dimensional digital twin model. Simultaneously, by performing density clustering and connectivity analysis on high-risk spatial grid cells exceeding safety thresholds, the system dynamically delineates the three-dimensional spatial range of potential disaster source areas, clarifying their absolute coordinate boundaries, disaster-causing energy volume, and spatial topological relationship with current mining operations. This provides precise spatial targets and quantitative basis for the scientific decision-making of the "mining-reconstruction-utilization" prevention and control plan in subsequent step S3.

[0092] Step S3: Make scientific decisions based on the prevention and control plan of "mining-reconstruction-utilization";

[0093] For the designated potential disaster source areas The central control system uses a four-dimensional digital twin as its physical simulation environment and employs a multi-objective genetic algorithm (MOGA) to autonomously generate an optimal prevention and control scheme that includes strategy A (mining optimization), strategy B (environmental reconstruction), and strategy C (disaster energy utilization) logic and their combinations, while meeting safety and mechanical constraints and comprehensively balancing production efficiency, governance costs, and risk levels.

[0094] In this multi-objective optimization process, the dynamic risk warning level output in step S2 is used as a penalty term to adaptively adjust the safety constraint boundary and strategy weight: under the blue and yellow warning states, the system constraint threshold is relatively loose, and the MOGA algorithm focuses on optimizing production efficiency, mainly generating strategy A (mining optimization) which is mainly based on adjusting the advance speed and optimizing the cyclic advance; while when the warning level jumps to the orange or even red extremely high-risk state, the algorithm will forcibly tighten the safety constraint tolerance threshold, veto the mining efficiency, and instead forcibly generate strategy B (environmental reconstruction) which is mainly based on local strong pressure relief and targeted grouting, and even a comprehensive heavy transformation plan that combines emergency evacuation and global disaster energy utilization.

[0095] Meanwhile, the specific disaster-prone mode classification determined in step S2 serves as a heuristic rule, directly narrowing the initial solution space of the algorithm and achieving targeted strategy matching: if the mode is determined to be "structural activation type", the algorithm prioritizes the combination of leaving fault isolation columns (strategy A) and cross-fault high prestressed locking (strategy B); for the "high stress concentration type" mode, the algorithm focuses on generating large-diameter boreholes for pressure relief or pre-fracturing blasting (strategy B) to cut off the stress transmission path; for the "high energy accumulation type" mode, the algorithm not only outputs strategy B, which weakens the rock mass stiffness through hydraulic fracturing, but also actively matches strategy C (disaster energy utilization), which involves in-situ extraction of deep gas or water resources, thereby achieving the optimal solution for disaster prevention, mitigation and resource utilization while meeting subsequent safety and mechanical constraints.

[0096] The aforementioned safety mechanical constraints mainly include four aspects: stress, deformation, energy, and dynamic load.

[0097] (1) Constraints on surrounding rock strength and yield criterion: To ensure that the optimized mining and environmental reconstruction scheme will not induce large-scale rock mass shear or tensile failure, the redistribution stress tensor of each grid cell in the disaster source area and its adjacent safety pillar is constrained. The macroscopic strength envelope constraint g1 of the rock mass must be satisfied:

[0098] In the formula, This represents the maximum principal stress equivalent that the base evolves after the strategy is implemented. The uniaxial compressive strength of deep, intact rock. An environmental reduction factor is used to account for the development of rock mass joints and fissures and the intensity of engineering disturbance.

[0099] (2) Allowable deformation and support coordination constraints: To prevent large deformation or tearing of the support system in roadways or mining areas, the calculated relative convergence displacement of the roadway wall and the absolute displacement within the anchoring layer shall not exceed the limit safety deformation g2 set for deep engineering.

[0100] In the formula, To implement the strategy at a specific time step Inner The norm of the spatial displacement vector of each key monitoring node To match the ultimate allowable tensile elongation or shear slip of the current support material (such as high-strength yield anchors / cables).

[0101] (3) Dynamic radiation energy anti-rockburst constraint: Focusing on the prevention of rockburst disasters, it is required that after the adjustment of the scheme (such as pressure relief blasting or grouting reconstruction), the maximum residual elastic strain energy density inside the rock mass in the target area must be strictly lower than the critical energy index threshold g3 that triggers rockburst.

[0102] In the formula, The residual elastic energy density of the mesh after regulation is calculated based on physical environment deduction. The critical elastic energy reserve limit for target rock strata to undergo dynamic failure was determined and entered into the system through indoor rock mechanics tests.

[0103] (4) Safety constraints for induced blasting: For induced blasting operations introduced in Strategy C, when the dynamic energy shock wave generated by the blasting radiates outward, the vibration velocity of the particles reaching the boundaries of adjacent mining faces, important electromechanical equipment, or permanent structures must be less than the safety regulation threshold g4.

[0104] In the formula, Based on the parameters of small charge in induced blasting Spatial attenuation distance The calculated peak vibration velocity of the particle, The maximum safe vibration velocity that underground intelligent equipment or structures can withstand.

[0105] In practice, the system first transforms the three physical control strategies into computer-optimizable multidimensional gene chromosome sequences. Among them, the mining optimization decision variables This includes blasting timing interval parameters to stagger and reduce peak dynamic load energy, mining advance direction spatial vectors to avoid high-stress zones, and stope structure parameters to optimize stress distribution; environmental reconfiguration decision variables. This includes parameters for large-diameter pressure-relief boreholes, hydraulic fracturing top-cutting pressure, and the strength of grouting reinforcement materials. The aim is to physically alter the rock mass structure, reconstructing a high-energy-storage state into a stress-controlled, disaster-insensitive environment; disaster energy utilization decision variables. For specific high-energy-storage hard rock strata, we define the critical threshold for induced energy release and the parameters for induced blasting charge to guide the elastic potential energy of the rock mass into rock-breaking power.

[0106] To enable flexible individual or coordinated applications for specific operating conditions, the system introduces Boolean discrete control vectors. T, where This leads to the generation of a global integrated prevention and control scheme vector. :

[0107] Under the boundary constraints of the four-dimensional digital twin platform, the system constructs a three-dimensional multi-objective optimization function model based on the aforementioned variables. The first objective function... Aiming to minimize risk, the strategy aims to minimize the disaster source area by optimizing the evolving stress field after the implementation of the coupling strategy. Residual disaster-causing energy density within:

[0108] In the formula, For native high energy storage density, , , These are the energy densities lost through mining optimization and transfer, environmental reconstruction and dissipation, and disaster energy utilization and conversion.

[0109] Second objective function To maximize overall efficiency while minimizing the overall cost of engineering and management processes, while ensuring production capacity:

[0110] In the formula, This is the production efficiency conversion factor. For the volume of ore mined, , , For the cost of implementing each strategy independently, This refers to the cost of system collaboration.

[0111] Third objective function To maximize the energy conversion rate, the elastic potential energy of specific high-energy-storage hard rock layers must be converted to the maximum extent into effective energy release for fracturing hard rock or decompression work:

[0112] In the formula, To induce a successful transformation of effective rock-breaking dynamics, To accumulate elastic energy in the disaster-stricken area The parameters for the small charge quantity used in inducing blasting.

[0113] After completing the model construction, the central control system employs a non-dominated sorting multi-objective genetic algorithm to deduce the aforementioned objective function in parallel. The algorithm iteratively generates a Pareto optimal solution set covering a multi-scale solution space through crossover and mutation operations, and by introducing a non-dominated sorting mechanism based on crowding distance. Subsequently, the system extracts digital twin feature parameters of the current operating condition in real time (such as disaster probability, distance to fault approach, elastic energy accumulation gradient, etc.), and uses a conditional state machine to autonomously match the optimal combination vector that best fits the current environment from the Pareto front. Ultimately, the system will select the globally optimal parameter sequence. Automatically compile and translate into industrial control commands for automated drilling, blasting, and mining equipment; that is, adaptively match and output propulsion axial force that matches the current rock type. Torque In addition to impact energy parameters, variable frequency flexible drilling with constant power or constant torque is achieved, combining blasting timing intervals with small charge parameters. The system automatically compiles precise millisecond-level delay detonation current network commands into micro-delay detonation control terminals (such as electronic detonator encoders). These commands are then sent to downhole intelligent terminals, thereby achieving closed-loop physical execution of the underlying collaborative process of "mining-reconstruction-utilization" and turning harm into benefit.

[0114] Step S4: Feedback regulation and effect evaluation of disaster-sensitive environments;

[0115] After the decision-making plan is issued, the downhole intelligent control equipment automatically executes and implements monitoring while drilling (MWD) and feedback correction operations. During implementation, the intelligent drilling and grouting robots collect multi-source spatiotemporal mechanical parameters in real time, including torque T, drilling pressure W, rotational speed N, and drilling speed V, while performing environmental reconstruction or depressurization drilling operations. To accurately quantify the rock mass mechanical state in front of the drill bit, the system calculates the specific energy (SE) while drilling based on the aforementioned multi-source sensing data to characterize the density and stress concentration level of the rock mass in real time.

[0116] In the formula, A represents the cross-sectional area of ​​the drill bit. If the system detects a sudden increase in the drilling specific energy SE, or if the fluid diffusion pressure in the grouting pipeline exhibits an abnormal nonlinear decay, it immediately determines that the bottom-level execution terminal has encountered a high-stress locking zone or a dominant water-conducting fracture channel. At this time, the equipment's central control unit will trigger an adaptive PID control algorithm to dynamically correct the drilling posture parameters or adjust the grouting ratio and flow rate, achieving bottom-level spatiotemporal feedback control between the intelligent equipment and the disaster-sensitive environment.

[0117] Subsequently, for the target area after induced blasting or environmental reconstruction, the system continuously captures subsequent microseismic response parameters based on a four-dimensional digital twin base, and introduces a risk reduction effect assessment index. The effect of elastic potential energy release and transfer is quantitatively determined:

[0118] In the formula, and These represent the cumulative microseismic radiation kinetic energy released in the target area within a set time window before and after the implementation of the prevention and control plan. Combining the frequency decay characteristics of microseismic events after induced blasting with the amplitude of the spatial stress gradient decrease, the system objectively assesses whether the high-risk rock mass has been successfully reconstructed into a safe state with dispersed stress and controllable energy.

[0119] After the operation is completed and a rigorous observation period has passed, the system forcibly re-executes steps S1 and S2 for a holographic re-evaluation. If the multi-scale comprehensive risk probability P(t) drops below the safety threshold and the risk reduction index... If the criteria are met, the collaborative prevention and control measures for that round are deemed effective, the system archives the parameters, and proceeds to the next mining cycle. Conversely, if the assessment fails to meet the criteria, it indicates a deviation between the rock mechanics parameters or early warning network weights in the digital twin base and the actual environment. The system then retrieves the abnormal drilling parameters and microseismic travel-time residuals from the execution process, dynamically inverts and updates the prior geomechanical parameters using Bayesian inference formulas.

[0120] In the formula, The set of rock mass characteristic parameters to be corrected (including the cohesion of deep jointed rock masses) internal friction angle Environmental reduction factor elastic modulus Poisson's ratio and the locally redistributed residual principal stress tensor Critical storage limit of rock mass and elastic energy limit ), Let D be the prior probability distribution before the update, and D be the newly acquired real feedback data in the mutual feedback regulation. Let be the likelihood function. The updated posterior probability. The data assimilation module, which is re-injected into S1, drives the MOGA algorithm of the central control system to regenerate the revised next-generation control scheme, thereby building a highly reliable closed-loop intelligent prevention and control system of "monitoring-early warning-decision-execution-feedback-re-early warning" at the underlying logic level.

[0121] To make the technical solution of this embodiment clearer, the system architecture, workflow, algorithm logic, and intelligent control principle of this embodiment will be further explained in conjunction with the accompanying drawings. As shown in the drawings, this embodiment, in practical applications, firstly... Figure 1 The diagram illustrates the construction of a high-precision multi-source sensing network underground. This network comprises a microseismic sensor array, a ground stress monitoring station, acoustic emission probes, surrounding rock displacement gauges, and related auxiliary sensors, with roadways, mining areas, and high-stress sensitive zones as key deployment areas. Various real-time monitoring data are transmitted to the ground data center via an underground industrial ring network, where preliminary noise reduction, clock synchronization, and feature extraction are performed at edge computing nodes. The data acquisition includes the three-dimensional stress field and its rate of change around roadways and mining areas, surrounding rock convergence and displacement deformation data, blasting and tunneling production data, and acoustic emission and microseismic waveform signals excited by rock microfractures. For dynamic response characteristics, the system solves for the four-dimensional coordinates of the seismic source in non-uniform rock masses based on ray tracing and residual minimization objective functions. Its positioning calculation formula is:

[0122] Simultaneously, the microseismic radiation energy was calculated using the time-domain integral formula. :

[0123] The aforementioned multi-source heterogeneous data is preprocessed and spatially mapped, and a standard kriging algorithm is used for feature scale unification and spatial interpolation completion. For any point to be estimated in the grid... Attribute estimates The calculation is performed using an unbiased estimator:

[0124] Subsequently, the ground computing platform integrates the above data with the mine geological model, structural model, and current mining parameters, constructing a digital twin engine such as... Figure 1 The four-dimensional deep rock mass dynamic response model marked in the process enables real-time simulation of rock mass stress redistribution, damage propagation, and energy accumulation zone evolution.

[0125] For high-frequency, non-stationary microvibrations and acoustic emission waveforms, the system uses wavelet transform for threshold denoising. For the clean waveform, the system does not directly use the amplitude sequence, but instead performs parallel dynamic inversion of the radiated energy E, energy index EI, and b-value for calculation.

[0126] The system performs time-domain-frequency domain integration on the effective microseismic signals to calculate the radiation energy E and seismic moment M of a single rupture event. To eliminate the masking effect of the absolute magnitude on the local stress concentration state, the system extracts the energy index EI. Its calculation formula is as follows:

[0127] In the formula, c and d are local geological empirical constants fitted based on historical microseismic data of the target mine. When EI>1, it indicates that the actual released energy in the grid area is higher than the average energy under the same magnitude, suggesting a high-stress locked state.

[0128] Simultaneously, the system extracts the magnitude sequence of all microseismic events within a set sliding time window (e.g., the past 24 hours), and dynamically solves for the b-value characterizing the microcrack propagation scale using the maximum likelihood estimation method (MLE) based on the Gutenberg-Richter relation:

[0129] In the formula, e is the natural constant. The average magnitude of the microseismic events within the time window. This represents the minimum complete magnitude under the current sensor network monitoring capabilities. The continuous decrease in the b-value is directly mapped as a precursory feature of micro-cracks within the rock mass leading to macroscopic, large-scale fracture penetration.

[0130] For discrete deformation readings of surrounding rock obtained by lidar or multi-point displacement gauges, since mechanical vibrations in the field can easily introduce burr noise, the system first uses a moving average filter to filter the original displacement sequence. Perform time-domain smoothing preprocessing:

[0131] In the formula, To be at the current central time node The smooth net displacement value output after filtering; The original displacement sequence is affected by downhole transient environmental noise; k is the center discrete time index of the current data node to be estimated; The preset single-sided sliding time window step is used to define the range of reference data before and after the center point; 2 +1 represents the total number of valid sampling points within the sliding time window; j represents the relative time offset index within the time window, with a value range of [- , By introducing a symmetrical time window observation matrix, transient displacement data jumps caused by heavy equipment operation or rockfall impacts were effectively suppressed.

[0132] After obtaining the smooth displacement curve, the system uses the first-order forward difference formula to calculate the surrounding rock deformation rate of the spatial grid node at a specific time. :

[0133] Deformation rate By eliminating the static baseline data accumulated from historical slow creep in the absolute displacement, the nonlinear accelerated rheological characteristics of the rock mass under recent mining disturbances are highlighted.

[0134] For discrete production logs such as those documented for blasting and tunneling, the system preprocesses them from text modality into a four-dimensional spatiotemporal constrained Boolean matrix, and extracts the three-dimensional centroid coordinates of each operation. At the moment of detonation Generate constraint functions:

[0135] In the formula, The effective physical sweep radius of the dynamic load disturbance caused by blasting. This represents the hysteresis time window for stress redistribution. When the Mask value of a grid node is 1, the system will assign additional dynamic disturbance triggering weights to that grid during subsequent disaster-prone mode classification.

[0136] The model is driven by the time dimension t, and uses an ensemble Kalman filter (EnKF) data assimilation algorithm to access and fuse monitoring data in real time, transforming the discrete observation vectors in the field into a unified whole. Injected into the prior full-field stress state vector predicted by numerical simulation In the process, dynamic inversion yields posterior analytical values. :

[0137] Wherein, the Kalman gain matrix The assimilation weight of the system for the measured data is determined, and its calculation formula is as follows:

[0138] As the mining face advances, if... Figure 2As shown, the intensity, frequency, and spatial clustering of microseismic events show an abnormal increase, the b-value decreases significantly, and stress monitoring shows a continuous accumulation trend, at which point the system automatically enters the disaster pre-prone identification stage. The algorithm module first performs time-series-spatial field feature fusion on key indicators such as microseismic energy index, acoustic emission ring count, stress concentration factor, ground stress gradient, and surrounding rock deformation acceleration, and inputs these features into a DBN / CNN-LSTM deep neural network structure. The high-dimensional feature matrix is ​​first input into the deep belief network (DBN) for unsupervised pre-training, such as... Figure 3 As shown, a DBN is composed of multiple layers of Restricted Boltzmann Machines (RBMs) stacked together. It learns the joint probability distribution among variables by minimizing the system energy function. The energy function of a single layer of RBM is... The calculation formula is:

[0139] The output low-dimensional, high-order one-dimensional feature vectors are used in a three-dimensional convolutional neural network (3D-CNN) to extract spatial patterns. Multiple three-dimensional convolutional kernels are used to perform sliding window operations in stress and energy release fields to autonomously identify spatially hazardous anomalies such as high-stress-concentrated gradient bands and microseismic energy-dense kernels. The forward propagation formula for its three-dimensional convolutional layers is:

[0140] After pooling layer downsampling, the CNN outputs a global spatial feature sequence that characterizes the spatial damage evolution of the rock mass.

[0141] In the LSTM part, a temporal evolution relationship is established. The core cell state update and hidden layer output formulas of the LSTM are as follows:

[0142] The spatiotemporal fusion feature vector output by the LSTM is then fed into a fully connected layer and a Softmax classifier to calculate the dynamic risk probability of dynamic disasters such as rockbursts occurring at the working face within a specific future time period. This allows the system to output the current rock mass's disaster-prone type (e.g., tectonic activation type, high-energy accumulation type) and the probability of rock mass instability within a certain future time window. The DBSCAN spatial topology clustering algorithm is then used to connect and reconstruct these areas into a three-dimensional potential disaster source region. Risk zoning:

[0143] The system is based on the disaster source area Automatically calculate the minimum bounding box geometric span of the region in a three-dimensional coordinate system. If the high-risk areas identified by clustering have a very small span ( Standard span of the mining area ,and Segmented Standard Height The system classifies it as a mining area level; if the high-risk area spreads horizontally across several mining areas but does not extend vertically (…), it is classified as a mining area level. Mid-section standard height The system classifies it as a mid-level zone; if the high-risk area penetrates deep strata vertically ( Mid-section standard height The system classifies this as a mining area. Based on this classification, the system automatically generates... Figure 2 The red, orange, yellow, and blue graded early warning signals are shown and visualized in a four-dimensional digital twin model.

[0144] Targeting the designated potential disaster source areas The central control system uses a four-dimensional digital twin platform as its physical simulation environment. Utilizing a multi-objective genetic algorithm (MOGA), it autonomously generates an optimal prevention and control scheme that integrates strategy A (mining optimization), strategy B (environmental reconstruction), and strategy C (disaster energy utilization) logic and their combinations, while satisfying safety and mechanical constraints and comprehensively balancing production efficiency, governance costs, and risk levels. The safety and mechanical constraints mainly include four aspects: stress, deformation, energy, and dynamic load.

[0145] (1) Constraints on surrounding rock strength and yield criterion: To ensure that the optimized mining and environmental reconstruction scheme will not induce large-scale rock mass shear or tensile failure, the redistribution stress tensor of each grid cell in the disaster source area and its adjacent safety pillar is constrained. The macroscopic strength envelope constraint g1 of the rock mass must be satisfied:

[0146] (2) Allowable deformation and support coordination constraints: To prevent large deformation or tearing of the support system in roadways or mining areas, the calculated relative convergence displacement of the roadway wall and the absolute displacement within the anchoring layer shall not exceed the limit safety deformation g2 set for deep engineering.

[0147] (3) Dynamic radiation energy anti-rockburst constraint: Focusing on the prevention of rockburst disasters, it is required that after the adjustment of the scheme (such as pressure relief blasting or grouting reconstruction), the maximum residual elastic strain energy density inside the rock mass in the target area must be strictly lower than the critical energy index threshold g3 that triggers rockburst.

[0148] (4) Safety constraints for induced blasting: For induced blasting operations introduced in Strategy C, when the dynamic energy shock wave generated by the blasting radiates outward, the vibration velocity of the particles reaching the boundaries of adjacent mining faces, important electromechanical equipment, or permanent structures must be less than the safety regulation threshold g4.

[0149] When the warning level reaches orange or red, the system will automatically enter a state similar to... Figure 4 The strategy decision-making stage is shown in the diagram. The decision tree first assesses whether the rock mass in the disaster-prone area possesses high energy storage characteristics, whether it is located in a necessary mining passage, and whether the energy release path is singular. If the risk can be reduced simply by changing mining parameters, an optimized mining strategy is selected. If the rock mass has strong structural integrity and high stress concentration, an environmental reconstruction strategy is selected, including large-diameter pressure relief drilling, hydraulic fracturing and top cutting, and grouting reinforcement. If the disaster-prone area is hard and brittle rock with huge energy storage, to avoid uncontrollable rockbursts, the system further calculates the induced energy release threshold and guides the system into a disaster energy utilization strategy, releasing the elastic potential energy stored in the rock mass in a controllable manner through methods such as small-charge induced blasting. This is combined with the global comprehensive prevention and control scheme vector. Under the boundary constraints of the four-dimensional digital twin platform, the system constructs a three-dimensional multi-objective optimization function model based on the aforementioned variables. The first objective function... Aiming to minimize risk, the strategy aims to minimize the disaster source area by optimizing the evolving stress field after the implementation of the coupling strategy. Residual disaster-causing energy density within:

[0150] Second objective function To maximize overall efficiency while minimizing the overall cost of engineering and management processes, while ensuring production capacity:

[0151] Third objective function To maximize the energy conversion rate, the elastic potential energy of specific high-energy-storage hard rock layers must be converted to the maximum extent into effective energy release for fracturing hard rock or decompression work:

[0152] according to Figure 3 The system's decision-making logic involves selecting the globally optimal parameter sequence. Automatically compile and translate into industrial control commands for automated drilling, blasting, and mining equipment, including adaptive matching and outputting propulsion axial forces that match the current rock type. Torque In addition to impact energy parameters, variable frequency flexible drilling with constant power or constant torque is achieved, combining blasting timing intervals with small charge parameters. Automatically compiled into precise millisecond-level delay initiation current network commands for micro-delay initiation control terminals (such as electronic detonator encoders).

[0153] During the execution phase, such as Figure 4 As shown, the intelligent drilling rig, grouting robot, and blasting control unit all maintain two-way communication with the ground central control system. Taking the intelligent drilling rig as an example, it collects multi-source spatiotemporal mechanical parameters in real time, including torque T, drilling pressure W, rotational speed N, and drilling speed V, and calculates the specific energy SE while drilling.

[0154] When the drilling rig reached Figure 4 When drilling through areas with natural fractures, induced fractures, or high-stress rock masses, the drilling specific energy (SE) may experience a sudden, high-gradient increase, or the fluid diffusion pressure in the grouting pipeline may exhibit an abnormal, nonlinear decay. The control system automatically adjusts drilling parameters accordingly, thereby avoiding safety issues such as stuck drill bits and borehole collapse, and improving the drilling success rate in complex geological environments. After drilling is completed, the system... Figure 3 The plan calls upon a digital initiation system to execute an induced blast. The shockwave generated by the blast will trigger... Figure 4 The microseismic response shown indicates that the monitoring network immediately captures the mainshock and aftershock events, and assesses the risk reduction effect through an index. The effect of elastic potential energy release and transfer is quantitatively determined:

[0155] In the formula, and These represent the cumulative microseismic radiation kinetic energy released in the target area within a set time window before and after the implementation of the prevention and control plan. Combining the frequency decay characteristics of microseismic events after induced blasting with the amplitude of the spatial stress gradient decrease, the system objectively assesses whether the high-risk rock mass has been successfully reconstructed into a safe state with dispersed stress and controllable energy.

[0156] If the multi-scale comprehensive risk probability P(t) drops below the safety threshold and the risk reduction index... If the criteria are met, the collaborative prevention and control measures are deemed effective, and the system enters the environmental reconstruction phase. This involves a smart drilling rig performing large-diameter pressure-relief drilling or a grouting robot reinforcing the potentially fractured zone. Conversely, if the assessment fails to meet the criteria, it indicates a deviation between the rock mechanics parameters or early warning network weights in the digital twin base and the actual environment. The system then retrieves abnormal drilling parameters and microseismic travel-time residuals from the execution process, dynamically inverts and updates the prior geomechanical parameters using Bayesian inference formulas.

[0157] Updated posterior probability The data assimilation module, re-injected into S1, drives the MOGA algorithm of the central control system to regenerate the corrected next-generation control scheme, thus achieving... Figure 4 The diagram shows a fully closed-loop intelligent control system consisting of "monitoring—early warning—decision-execution—feedback—re-early warning".

[0158] As can be seen from the above description in conjunction with the accompanying drawings, this embodiment can achieve high spatial accuracy and high temporal resolution in the identification of potential hazards in deep and complex geological environments. Through innovative means such as active induction, environmental reconstruction, and utilization of hazard energy, it realizes the transformation of dynamic hazard prevention and control from passive avoidance to active intervention, transforming the dangerous energy of high-energy-storage rock masses into usable rock-breaking power, improving mine safety and mining efficiency, while avoiding energy waste, and has significant technical advantages.

[0159] This application also provides an intelligent prevention and control system for dynamic disasters in metal mines, including one or more functional units or modules, for performing the methods described in any of the above embodiments.

[0160] This application also provides an electronic device, including: a processor, and a memory coupled to the processor, the memory being used to store a computer program; the processor being used to execute the computer program stored in the memory, so that the electronic device performs the method as described in any of the above embodiments.

[0161] Electronic devices can be computing devices such as desktop computers, laptops, handheld computers, and cloud servers. These electronic devices may include, but are not limited to, processors and memory.

[0162] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the device via various interfaces and lines.

[0163] The memory can be used to store the computer program, and the processor implements various functions of the electronic device by running or executing the computer program stored in the memory and calling the data stored in the memory.

[0164] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0165] This application also provides a storage medium, which is a computer-readable storage medium. The computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0166] This application also provides a computer program product, including: a computer program or instructions that, when the computer program or instructions are run on a computer, cause the computer to perform any of the above possible implementation methods.

[0167] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. A smart prevention and control method for dynamic disasters in metal mines based on environmental reconstruction and disaster energy utilization, characterized in that, include: Obtain multi-source sensing data; Based on the multi-source sensing data, a multi-source high-dimensional disaster-pregnancy feature matrix is ​​constructed that integrates macroscopic spatial topology and long- and short-term time dependencies. Using a pre-trained autonomous disaster-prone factor identification and multi-scale intelligent early warning model, the multi-source high-dimensional disaster-prone feature matrix is ​​processed to identify the disaster-prone pattern classification, dynamic risk probability, and dynamic risk early warning level of a spatial region. The spatial region is then divided into risk zones to obtain the three-dimensional spatial range of potential disaster source areas. The autonomous disaster-prone factor identification and multi-scale intelligent early warning model includes a cascaded structure of: a Deep Belief Network (DBN), a 3D Convolutional Neural Network (3D-CNN), a Long Short-Term Memory (LSTM) network, and a classification head. The Deep Belief Network is used to extract and represent the multi-source high-dimensional disaster-prone feature matrix on a single grid. The underlying potential correlation features of the internal physical damage evolution state are obtained. The three-dimensional convolutional neural network is used to perform deep decoupling and fusion of spatiotemporal features on the three-dimensional feature tensor containing spatial coordinates corresponding to the underlying potential correlation features to obtain a global spatial feature sequence characterizing the spatial damage evolution law of the rock mass. The long short-term memory network is used to capture the nonlinear temporal evolution law of disaster-inducing factors on the time axis of the global spatial feature sequence to obtain a spatiotemporal fusion feature vector. The classification head is used to obtain the disaster-inducing pattern classification, dynamic risk probability results and dynamic risk warning level of the potential disaster source area based on the spatiotemporal fusion feature vector. Using a four-dimensional digital twin as the physical simulation environment, and employing a multi-objective genetic algorithm, under the condition of satisfying safety and mechanical constraints, a comprehensive risk, efficiency and disaster energy conversion rate are considered to generate a prevention and control optimization scheme for the three-dimensional spatial range of potential disaster source areas, which includes mining optimization, environmental reconstruction and disaster energy utilization.

2. The method as described in claim 1, characterized in that, The multi-source high-dimensional pregnancy feature matrix is ​​as follows: In the formula, For the current target time step t, a high-dimensional feature tensor that integrates macroscopic spatial topology and long- and short-term time dependencies; A global three-dimensional spatial mesh set based on a four-dimensional digital twin; For indexing historical continuous time series within a sliding time window; For a specific moment And the absolute coordinates are The feature column vector of a single grid node; The specific time step of the sliding time window.

3. The method as described in claim 1, characterized in that, The deep belief network is composed of several multi-layer Restricted Boltzmann Machines (RBMs) stacked together. The deep belief network learns the joint probability distribution among variables by minimizing the system energy function. Each RBM includes a cascaded visible layer and several hidden layers. The energy function of a single-layer RBM... for: In the formula, and These are the visible layer and hidden layer state vectors, respectively. and These represent the total number of nodes in the visible layer and the hidden layer, respectively. and The first The visible node and the first The state values ​​of the hidden nodes; and These are the corresponding bias terms; This represents the connection weights between the visible and hidden layers.

4. The method as described in claim 1, characterized in that, The three-dimensional convolutional neural network utilizes multiple three-dimensional convolutional kernels to perform sliding window operations in stress and energy release fields, autonomously identifying various spatial disaster-causing anomaly structures. The forward propagation formula of the three-dimensional convolutional layer of the three-dimensional convolutional neural network is as follows: In the formula, This is the channel index of the input feature map of the previous layer. For the first The output of the convolutional layer is the first A three-dimensional feature map; This is the 3D feature map input from the previous layer; This represents the corresponding 3D convolution kernel weight matrix; * indicates 3D convolution operation; For bias terms; It is a non-linear activation function; After the final 3D feature map is downsampled by the pooling layer, the 3D convolutional neural network outputs the global spatial feature sequence.

5. The method as described in claim 1, characterized in that, The formulas for cell state updates and hidden layer outputs in the core of the Long Short-Term Memory (LSTM) network are as follows: In the formula, The spatial feature sequence input at the current time; and These are the hidden layer states of the previous and current time steps, respectively. and These are historical and updated cell states, used for long-term memory of pregnancy and disaster information; , , These are the activation vectors for the forget gate, input gate, and output gate, respectively. ; It is the Sigmoid activation function. Represents the product of matrix elements.

6. The method as described in claim 1, characterized in that, The spatial region is divided into risk zones to obtain the three-dimensional spatial extent of potential disaster source areas, specifically including: If at a high-risk grid point Spatial search radius The number of grid points contained within is greater than or equal to the set minimum density threshold. Then that point The density constraint formula for a node identified as a cluster core is: In the formula, For the initial set of high-risk discrete points, For connected points within the neighborhood, This represents the total number of grid elements within the set; Starting from the core point, perform a union operation on all grid cells with achievable density to form clusters. Furthermore, the actual physical three-dimensional envelope occupied by this cluster in the four-dimensional digital twin base is defined as a potential disaster source area. Potential disaster source area The formula for generating spatial boundaries is: In the formula, The center of mass is The three-dimensional volume of a single independent mesh; Automatic calculation of disaster source area The minimum bounding box geometric span in a three-dimensional coordinate system; The following scales are used to determine the risk level based on the geometric span of the smallest bounding box: if the high-risk area clustered out has a very small span, it is judged as a stope level; if the high-risk area spreads out in a continuous manner on the horizontal plane, spanning several stopes, but does not penetrate vertically, it is judged as a mid-level level; if the high-risk area penetrates deep strata in the vertical direction, it is judged as a mining area level. Based on the scale determination results and the exclusive thresholds that match the scale determination results, a graded early warning signal is automatically generated; By performing density clustering and connectivity analysis on high-risk spatial grid cells that exceed safety thresholds, the three-dimensional spatial range of potential disaster source areas is dynamically delineated, and their absolute coordinate boundaries, disaster-causing energy volume, and spatial topological relationship with current mining projects are clarified.

7. The method as described in claim 1, characterized in that, The specific safety mechanical constraints include: surrounding rock strength and yield criterion constraints, allowable deformation and support coordination constraints, and dynamic load safety constraints for induced work; the multi-objective optimization functions in the multi-objective genetic algorithm include: a first objective function to minimize the residual disaster-causing energy density in the disaster source area, a second objective function to minimize the comprehensive cost of engineering treatment under the premise of ensuring production capacity, and a third objective function to force the elastic potential energy of a specific high-energy-storage hard rock layer to be converted into effective energy release for fracturing hard rock or pressure relief work. The multi-objective genetic algorithm includes: Transforming physical control strategies into multidimensional gene chromosome sequences ; Introducing Boolean Discrete Control Vectors A global integrated prevention and control scheme vector is generated from the multidimensional gene chromosome sequence; Under the boundary constraints of the four-dimensional digital twin base, a three-dimensional multi-objective optimization function model based on the global integrated prevention and control scheme vector is constructed; A non-dominated sorting multi-objective genetic algorithm is employed to deduce the objective function in parallel. The algorithm iteratively generates a Pareto optimal solution set covering a multi-scale solution space through crossover and mutation operations, and by introducing a non-dominated sorting mechanism based on crowding distance. Subsequently, digital twin feature parameters of the current working condition are extracted in real time, and a conditional state machine is used to autonomously match the optimal combination vector best suited to the current environment from the Pareto front. The selected global optimal parameter sequence Automatically compile and convert these instructions into industrial control commands; then send these commands to the downhole intelligent terminal for execution.

8. The method as described in claim 1, characterized in that, The multi-source sensing data includes: energy index (EI), fracture scale evolution value, spectral characteristics, microseismic spatial concentration, stress concentration factor, surrounding rock deformation rate, and geological fault geometric sensitivity weight.

9. A smart prevention and control system for dynamic disasters in metal mines, characterized by environmental reconstruction and disaster energy utilization, wherein... It includes one or more functional units or modules for performing the method of any one of claims 1-8.

10. An electronic device, characterized in that, The electronic device includes: a processor, and a memory coupled to the processor. The memory is used to store computer programs; The processor is configured to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-8.