Mine rock burst monitoring method, prevention and control method, storage medium and electronic equipment

Abstract

The application discloses a kind of mine rock burst monitoring method, prevention and control method, storage medium and electronic equipment, comprising: the deformation and failure control equation of coal rock mass and energy balance equation are converted into penalty term of loss function, construct multi-scale hierarchical coupling PINN model;With the Euclidean distance between sensor, geological structure correlation and observation signal correlation as the multi-dimensional attribute of graph edge, construct space heterogenous graph, generate virtual node in the monitoring blind area of not laying sensor, the feature of each virtual node is spacially interpolated and spliced with space-time coordinates, form encoding input feature, and input to multi-scale hierarchical coupling PINN model, output the prediction result of stress field and energy field;By variational inference, introduce probability distribution to network weight, so that multi-scale hierarchical coupling PINN model outputs probability distribution prediction result containing confidence interval, and generates cognitive uncertainty graph.The application can monitor and prevent and control mine rock burst in real time, improve safety.

Description

Technical Field

[0001] This invention relates to the field of coal mine safety mining technology, and in particular to a method for monitoring, controlling, storing, and operating a mine rockburst. Background Technology

[0002] In coal mining engineering, rockbursts, as a typical dynamic disaster, are closely related to the evolution of stress and energy fields, exhibiting strong nonlinearity, time-varying characteristics, and multi-scale abrupt changes. In high-stress areas and other regions, they are prone to a dangerous pattern of "slow incubation, rapid onset, and severe damage," seriously threatening construction safety.

[0003] Existing methods for monitoring rockbursts mainly include: 1) Single-index threshold alarm methods, which set fixed thresholds for monitoring indicators such as the number of microseismic events, stress values, and energy release rates. An alarm is triggered when any indicator exceeds the set threshold. This type of method is a post-event or pre-disaster type of early warning, which can only reflect the dangerous state that has occurred at the current moment and cannot predict or anticipate the impact hazard. 2) Pure data-driven artificial intelligence (AI) methods, which use existing models and historical monitoring data as training samples to build an early warning model. This type of method has defects such as overfitting and lack of interpretability, resulting in low confidence in the early warning. 3) General-purpose physical information neural network (PINN) methods, which add partial differential equation (PDE) constraint terms (i.e., physical loss) to the loss function of the neural network so that the output of the neural network simultaneously satisfies data fitting and physical laws. However, the general-purpose PINN model is a single-scale architecture, which cannot cope with multi-scale features and has gradient imbalance problems. Meanwhile, existing rockburst monitoring methods are disconnected from prevention and control methods. Prevention and control designs mainly rely on manual experience, making it impossible to conduct advance assessments and easily leading to over-prevention or under-prevention. Furthermore, there is a lack of quantitative prediction of prevention and control effectiveness and a closed-loop feedback optimization mechanism. Therefore, there is an urgent need to provide a method for monitoring and controlling rockbursts that can systematically integrate physical mechanism constraints, multi-scale modeling, spatial geological information perception, uncertainty quantification, and closed-loop adaptive optimization. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing rockburst monitoring and control methods that cannot systematically integrate physical mechanism constraints, multi-scale modeling, spatial geological information perception, uncertainty quantification, and closed-loop adaptive optimization, and to provide a mine rockburst monitoring method, control method, storage medium, and electronic equipment.

[0005] The technical solution of the present invention provides a method for monitoring rockburst in mines, comprising: The deformation and failure control equations and energy balance equations of coal and rock mass are transformed into penalty terms of the loss function. A multi-scale hierarchical coupled physical information neural network model is constructed, consisting of three hierarchical nested networks including micro-scale subnetworks, meso-scale subnetworks and macro-scale subnetworks, to perform cross-scale joint inference. A spatial heterogeneous graph is constructed using Euclidean distance between sensors, geological structure correlation, and observation signal correlation as multidimensional attributes of the graph edges. Virtual nodes are generated in the monitoring blind area where no sensors are deployed. The features of each virtual node are spatially interpolated and then spliced ​​with spatiotemporal coordinates to form coded input features that fuse spatial topological information. The coded input features are then input into the multi-scale hierarchical coupled physical information neural network model to output the predicted results of stress field and energy field. By introducing a probability distribution into the network weights of the trained multi-scale hierarchical coupled physical information neural network model through variational inference, the multi-scale hierarchical coupled physical information neural network model outputs a probability distribution prediction result containing confidence intervals, and generates a cognitive uncertainty map. The cognitive uncertainty map represents the prediction confidence distribution of the multi-scale hierarchical coupled physical information neural network model in various spatial regions of the working face and mining area.

[0006] Furthermore, the process of transforming the deformation and failure control equations and energy balance equations of coal and rock mass into penalty terms of a loss function, and constructing a multi-scale hierarchical coupled physical information neural network model comprising three nested networks—micro-scale, meso-scale, and macro-scale—for cross-scale joint extrapolation, previously included: Real-time access to monitoring data on microseismic activity, ground acoustics, borehole stress gauges, anchor bolt force gauges, support working resistance, and mining stress; Spatiotemporal alignment and fusion are performed using Kalman filtering, Bayesian fusion, and graph neural network-assisted spatiotemporal interpolation algorithms to form a four-dimensional monitoring dataset; The multi-scale hierarchical coupled physical information neural network model is driven to perform dynamic inversion with preset spatial and temporal resolutions.

[0007] Furthermore, the training of the three nested networks employs an alternating-joint hybrid training strategy anchored by mesoscale subnetworks, including: In each training cycle, the upstream and downstream subnets are frozen sequentially and trained separately for the mesoscale subnet, the macroscale subnet, and the microscale subnet, and then the freeze is lifted for joint fine-tuning of all parameters.

[0008] Furthermore, during the training process of the multi-scale hierarchical coupled physical information neural network model, the meta-network drives the adaptive physical constraint weight adjustment. The meta-network takes the current residual magnitude, spatial gradient distribution and historical change trend of each loss term as input, and generates local adaptive loss weights for each spatial region in real time. By dynamically adjusting the relative contributions of data loss terms and physical loss terms, a balance is maintained between monitoring data fitting and physical equation constraints.

[0009] Furthermore, the meta-network-driven adaptive physical constraint weight adjustment includes: If the difference between the current physical residual of the partial differential equation in a spatial region and the historical physical residual in the previous time exceeds a preset residual threshold, the meta-network increases the loss weight of the partial differential equation in that region to a preset multiple of the benchmark value. If the confidence score of the monitoring data in one of the spatial regions is lower than the preset quality score threshold, the meta-network reduces the data loss weight of that spatial region.

[0010] Furthermore, the meta-network includes a region encoder and a weight generator. The region encoder compresses the input feature vector of each spatial region into a region embedding vector, and concatenates it with the spatial coordinate encoding of the region to form a region representation vector. The weight generator takes the region representation vector as input and outputs the local adaptive loss weights for that region. The meta-network shares parameters across all spatial regions.

[0011] The technical solution of the present invention also provides a method for preventing and controlling rockbursts in mines, comprising: Obtain the stress field prediction results, energy field prediction results, and cognitive uncertainty map of the mine working face. The cognitive uncertainty map represents the prediction confidence distribution of the stress field and energy field prediction results in various spatial regions of the working face and mining area. Based on the cognitive uncertainty diagram, combined with the stress field prediction results and the energy field prediction results, a hierarchical response decision is output. The stress field prediction results and the energy field prediction results are input into the prevention and control scheme generation module to generate a targeted prevention and control scheme. The targeted prevention and control scheme is then input into the multi-scale hierarchical coupled physical information neural network model for digital twin advanced simulation. The digital twin advanced simulation results of the prevention and control scheme are output, including the predicted mean and confidence interval of the stress field and energy field after the scheme is implemented. After implementing the prevention and control plan, the evaluation result of the prevention and control effect is used as the feedback signal of the multi-scale hierarchical coupled physical information neural network model, triggering the online Bayesian update of the multi-scale hierarchical coupled physical information neural network model, so that the parameters of the multi-scale hierarchical coupled physical information neural network model co-evolve with the prevention and control plan.

[0012] Furthermore, the output hierarchical response decision also includes: Obtain the parameter prior distribution from the preset prior distribution knowledge base that is most similar to the geological conditions of the new mining area, and use a preset number of effective rockburst events as training samples to correct the parameters.

[0013] The present invention also provides a storage medium that stores computer instructions. When the computer executes the computer instructions, it is used to perform all the steps of the mine rockburst monitoring method or the mine rockburst prevention and control method as described above.

[0014] The present invention also provides an electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the mine rockburst monitoring method or the mine rockburst prevention and control method as described above.

[0015] The above technical solution has the following beneficial effects: 1. By transforming the deformation and failure control equations and energy balance equations of coal and rock masses into penalty terms of loss functions, a multi-scale hierarchical coupled physical information neural network model is constructed, consisting of three nested networks: micro-scale, meso-scale, and macro-scale subnets. This model enables cross-scale joint extrapolation, achieving physical information transfer between adjacent scales and resulting in high prediction accuracy. A spatial heterogeneous graph is constructed using Euclidean distance between sensors, geological structural correlation, and correlation of observation signals as multi-dimensional attributes of the graph edges. Virtual nodes are generated in monitoring blind areas where no sensors are deployed. The features of each virtual node are spatially interpolated and then concatenated with spatiotemporal coordinates to form coded input features that integrate spatial topological information. These coded input features are then input into the multi-scale analysis... The layered coupled physical information neural network model outputs predicted stress and energy fields, enabling automatic learning of the spatial topology and geological control laws of the sensor network. It generates geocentric virtual nodes in monitoring blind zones, providing spatially continuous high-dimensional feature maps for PINN, thus solving the problem of blind zone monitoring caused by the sparse and non-uniform distribution of downhole sensors. By introducing probability distributions into the network weights of the trained multi-scale layered coupled physical information neural network model through variational inference, the model outputs probability distribution predictions with confidence intervals and generates a cognitive uncertainty map. This allows for direct correlation between the cognitive uncertainty map and hierarchical response decisions, reducing the risk of false alarms and missed alarms.

[0016] 2. By acquiring the stress field prediction results, energy field prediction results, and cognitive uncertainty map of the mine working face, and based on the cognitive uncertainty map, combined with the stress field prediction results and energy field prediction results, a graded response decision is output. The stress field prediction results and energy field prediction results are input into the prevention and control scheme generation module to generate a targeted prevention and control scheme. The targeted prevention and control scheme is then input into a multi-scale hierarchical coupled physical information neural network model for digital twin advanced simulation, outputting the digital twin advanced simulation results of the prevention and control scheme. After the prevention and control scheme is executed, the evaluation results of the prevention and control effect are used as feedback signals to the multi-scale hierarchical coupled physical information neural network model, triggering the online Bayesian update of the multi-scale hierarchical coupled physical information neural network model. This enables the parameters of the multi-scale hierarchical coupled physical information neural network model to co-evolve with the prevention and control scheme, realizing real-time feedback of the prevention and control effect to drive the co-evolution of model parameters and prevention and control strategies. This solves the problem of rapid performance decay of traditional static models as mining progresses, improves the accuracy and confidence of subsequent predictions, and forms a closed-loop adaptive control system with "monitoring as simulation, early warning as prevention and control, and effect as feedback" throughout the entire process, fundamentally solving the problem of accurate early warning and intelligent prevention and control of rockburst. Attached Figure Description

[0017] The disclosure of this invention will become more readily understood by referring to the accompanying drawings. It should be understood that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings: Figure 1 A flowchart illustrating a method for monitoring rockburst in a mine, provided in an embodiment of the present invention; Figure 2 A flowchart illustrating a method for preventing and controlling rockbursts in mines, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device for monitoring or preventing rockbursts in mines, provided as an embodiment of the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0019] It is readily understood that, based on the technical solution of this invention, various structural and implementation methods can be interchanged by those skilled in the art without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of the invention.

[0020] The directional terms such as up, down, left, right, front, back, front, back, top, and bottom mentioned or possibly used in this specification are defined relative to the structures shown in the accompanying drawings. They are relative concepts and may therefore vary depending on their location and usage. Therefore, these or other directional terms should not be interpreted as restrictive.

[0021] like Figure 1 The diagram shown is a flowchart of a mine rockburst monitoring method according to an embodiment of the present invention, including: Step S101: Transform the deformation and failure control equations and energy balance equations of coal and rock mass into penalty terms of the loss function, and construct a multi-scale hierarchical coupled physical information neural network model with three levels of nested networks including micro-scale subnetworks, meso-scale subnetworks and macro-scale subnetworks for cross-scale joint inference. Step S102: Construct a spatial heterogeneous graph using the Euclidean distance between sensors, geological structure correlation, and observation signal correlation as multidimensional attributes of the graph edges. Generate virtual nodes in the monitoring blind area where no sensors are deployed. After spatial interpolation, the features of each virtual node are spliced ​​with the spatiotemporal coordinates to form coded input features that fuse spatial topological information. Input the coded input features into the multi-scale hierarchical coupled physical information neural network model to output the predicted results of stress field and energy field. Step S103: Introduce a probability distribution into the network weights of the trained multi-scale hierarchical coupled physical information neural network model through variational inference, so that the multi-scale hierarchical coupled physical information neural network model outputs a probability distribution prediction result containing confidence intervals, and generates a cognitive uncertainty map, which represents the prediction confidence distribution of the multi-scale hierarchical coupled physical information neural network model in various spatial regions of the working face and mining area.

[0022] Specifically, in step S101, the control equation for deformation and failure of coal and rock mass is transformed into a penalty term of the loss function, and a multi-scale hierarchical coupled physical information neural network model containing three nested sub-networks is constructed. The three nested sub-networks include a micro-scale sub-network, a meso-scale sub-network, and a macro-scale sub-network.

[0023] Specifically, the microscale subnet has a temporal resolution of milliseconds and a spatial resolution of centimeters, and embeds the mesoscale damage constitutive equation as a physical constraint. The mesoscale subnet has a temporal resolution of minutes and a spatial resolution of meters, and embeds the elastoplastic constitutive equation and energy balance equation as physical constraints. The macroscale subnet has a temporal resolution of hours to days and a spatial resolution of tens of meters, and embeds the mining-induced stress propagation equation and roof collapse criterion as physical constraints. Adjacent scale subnets are connected through a cross-scale consistency loss function to achieve cross-scale joint simulation. The cross-scale consistency loss function is:

[0024] in, For the downscale projection operator, This is the upscale projection operator.

[0025] The three scale subnets share the same physical parameter space, which includes elastic modulus, Poisson's ratio, and cohesion. The microscale subnet captures the rock particle-level damage evolution via an upward projection operator. The stress redistribution results of the mesoscale subnet are passed to the mesoscale subnet as local material weakening information, and then passed to the mesoscale subnet via the downprojection operator. Feedback is fed back to the microscale subnet to verify the rationality of the damage propagation path, while the upward projection operator is used. Uploaded to the macro-scale subnet to drive the prediction of hazardous area evolution at the mining area scale.

[0026] The training of the three nested networks employs an alternating-joint hybrid training strategy anchored by mesoscale subnetworks, including: In each training cycle, the upstream and downstream subnets are frozen sequentially for training the mesoscale subnet, macroscale subnet, and microscale subnet separately, and finally the freeze is lifted for joint fine-tuning of all parameters.

[0027] In this embodiment, cross-scale joint extrapolation is performed at the micro, meso, and macro scales to establish an extrapolation link of "microscopic damage - mesoscopic stress redistribution - macroscopic catastrophe release" from the physical mechanism level. This ensures that the prediction results of each scale subnet in the overlapping spatial domain remain physically consistent, solving the technical problem that the accuracy of prediction results in cross-scale regions caused by the existing single-scale PINN model is significantly reduced.

[0028] In step S102, each downhole microseismic geophone, stress gauge, and borehole strain gauge is used as a graphical node. Three types of graph edges are constructed: Euclidean distance edges (based on the actual 3D spatial distance between two sensors, with connections established only when the distance is less than a preset radius), geologically related edges (based on whether the two sensors are located in the same coal seam, cross the same fault zone, and have different dip angles as 3D edge attribute vectors), and signal-related edges (based on the Pearson correlation coefficient of the two sensor observation signals within a preset time window, with connections established only when the absolute value of the correlation coefficient is greater than a preset threshold). The node feature vector of each graph node includes the sensor's current observation value, historical time series characteristics, and 3D burial depth coordinates.

[0029] Message passing is performed 3 to 5 times on the heterogeneous graph, with each round including aggregation and update. Aggregation refers to each node collecting feature information and edge attributes of all its neighboring nodes (including neighbors connected by different types of edges). Update refers to fusing the aggregated neighbor information with the node's own features using a learnable update function to generate a new node feature vector. The message passing formula is:

[0030] in, For nodes In the The feature vector of the layer, For nodes The neighborhood group, and For learnable weight matrix ( Regarding edge attributes (function) It is a non-linear activation function.

[0031] In the monitoring blind zone where no sensors are deployed (defined as an area where the distance to the nearest real sensor node exceeds a preset radius), the Graph Neural Network (GNN) automatically generates virtual nodes using the Graph Laplacian interpolation method. The feature vector of a virtual node is obtained by weighted averaging of the features of its k nearest real nodes, with the interpolation weights of the real nodes proportional to the reciprocal of the Euclidean distance from the real node to the virtual node. Virtual nodes do not participate in the output calculation of subsequent message passing to avoid the propagation of uncertainty from virtual nodes to real nodes.

[0032] In this embodiment, by modeling the sensor network as a spatial heterogeneous graph containing Euclidean distance edges, geological correlation edges, and signal correlation edges, the GNN automatically learns "which sensors physically belong to the same geological unit" and "which sensors observe the same mechanical process" during message passing. The information of the three edge types is gradually fused in multiple rounds of message passing, ultimately forming the geological perception characteristics of each sensor. This allows the feature interpolation of the virtual nodes in the blind zone to be based not only on spatial distance but also on deep geological and physical correlations, thus obtaining physically reasonable interpolation results even under complex geological conditions such as fault zones and lithological abrupt change zones.

[0033] In step S103, variational inference is used to introduce a Gaussian distribution prior to the network weights. Specifically, for each weight parameter w of the main PINN, its prior distribution is assumed to be:

[0034] in, The pre-defined prior variance is used.

[0035] The variational distribution is constructed as follows:

[0036] in, For the variational parameters to be learned (mean and posterior standard deviation), the variational parameters are optimized by minimizing the Kullback-Leibler divergence. The formula for calculating the Kullback-Leibler divergence is:

[0037] During the inference phase (i.e., the prediction phase after model deployment), each network weight is derived from its variational posterior distribution. The network is sampled M times (M being 10-20, the number of Monte Carlo sampling times). Each sampling yields a set of network outputs. After collecting M sets of outputs, the mean value at each spatial location is calculated. and standard deviation Construct a 95% confidence interval [μ-1.96s, μ+1.96s].

[0038] Output the following four types of information: mean stress field and 95% confidence interval, energy field mean The probability density distribution of the 95% confidence interval, the impact hazard level (e.g., four levels: safe, caution, warning, and danger), and the cognitive uncertainty map (in the form of a spatial heatmap, with color depth representing the prediction standard deviation of each grid node). The relative size, with light-colored areas representing low uncertainty and dark-colored areas representing high uncertainty.

[0039] In this embodiment, variational inference upgrades the network weights from fixed scalar values ​​to probability distributions, ensuring that each prediction by the model carries a confidence interval and cognitive uncertainty information. This transforms the model's prediction of danger into a prediction of danger with 95% confidence, providing a scientific triggering basis for the tiered response decision-making mechanism. The cognitive uncertainty map further enables a precise deployment strategy of prioritizing the addition of sensors where uncertainty exists, optimally allocating limited monitoring resources to areas with the greatest information gaps.

[0040] In one embodiment, step S101 further includes the following prior to: Real-time access to monitoring data on microseismic activity, ground acoustics, borehole stress gauges, anchor bolt force gauges, support working resistance, and mining stress; Spatiotemporal alignment and fusion are performed using Kalman filtering, Bayesian fusion, and graph neural network-assisted spatiotemporal interpolation algorithms to form a four-dimensional monitoring dataset; The multi-scale hierarchical coupled physical information neural network model is driven to perform dynamic inversion with preset spatial and temporal resolutions.

[0041] Specifically, the microseismic system provides the three-dimensional location (x, y, z), released energy, and frequency of microseismic events. The ground acoustic monitoring system provides acoustic emission event parameters (amplitude, energy, frequency). The borehole stress gauge provides borehole axial and radial stress counts. The anchor bolt force gauge provides anchor bolt axial force. The support working resistance sensor provides hydraulic support working resistance. The mining-induced stress monitoring system provides the pressure distribution of the working face's advanced support.

[0042] Data preprocessing was performed on real-time monitoring data of microseismic activity, ground acoustics, borehole stress gauges, anchor bolt dynamometers, support working resistance, and mining stress. First, a mean filter (window size of 5 sampling points) was applied independently to each sensor channel to remove impulse noise. Then, outliers were marked and corrected using a sliding window adaptive threshold (mean within the window ± 3 times the standard deviation within the window) (linear interpolation was used as a substitute). Second, the data for each channel was normalized to the [0,1] interval. Finally, all channels were aligned according to the Global Positioning System (GPS) / Network Time Protocol (NTP) timestamps, and the sampling frequency was unified to 1Hz. Channels with original sampling frequencies higher than 1Hz were downsampled, and channels with original sampling frequencies lower than 1Hz were augmented using linear interpolation.

[0043] Kalman filtering is used to estimate the state and smooth the temporal sequence of the signal. A Bayesian fusion algorithm is employed to perform weighted fusion of redundancy counts from multiple sensors for the same physical quantity. A GNN-assisted spatiotemporal interpolation algorithm is used to make the spatial dimension continuous. Finally, a high-fidelity monitoring dataset in a unified four-dimensional spatial coordinate system (x, y, z, t) is formed, driving a multi-scale hierarchical coupled physical information neural network model with a spatial resolution of 1 meter and a temporal resolution of 1 minute for dynamic inversion and visualization.

[0044] In this embodiment, Kalman filtering, Bayesian fusion, and graph neural network-assisted spatiotemporal interpolation algorithms are used to assimilate multi-source data from microseismic, ground acoustic, borehole stress gauge, anchor bolt force gauge, support working resistance, and mining stress monitoring data in all time and space, forming a unified four-dimensional monitoring dataset. This dataset is then used to dynamically invert the multi-scale hierarchical coupled physical information neural network model, thereby achieving transparency and confidence visualization of the mining environment.

[0045] In one embodiment, during the training process of the multi-scale hierarchical coupled physical information neural network model, the meta-network drives the adaptive physical constraint weight adjustment. The meta-network takes the current residual magnitude, spatial gradient distribution and historical change trend of each loss term as input, and generates local adaptive loss weights for each spatial region in real time. By dynamically adjusting the relative contributions of data loss terms and physical loss terms, a balance is maintained between monitoring data fitting and physical equation constraints.

[0046] Specifically, the lightweight meta-weight network (MWN) employs a two-stage design: a region encoder and a weight generator. The region encoder concatenates the 12-dimensional input feature vectors of each spatial region into an 11-dimensional region representation vector. The weight generator takes this 11-dimensional region representation vector as input and independently outputs four adaptive loss weights for that region. The meta-weight network shares parameters across all spatial regions; that is, the same set of network parameters processes the representation vector of each region sequentially, generating weights for each region one by one. This avoids the mismatch between ultra-high-dimensional input and the capacity of the lightweight network, while preserving the spatial independence of weight decisions for each region.

[0047] When the physical residual of a partial differential equation (PDE) in a certain spatial region increases by more than a preset residual threshold (e.g., 20%) from the previous time step, the meta-network automatically increases the PDE loss weight for that region by a preset multiple of the baseline value (e.g., 1.5 to 3 times, with the increase proportional to the increase rate), thus strengthening the optimizer's satisfaction of physical constraints in that region. When the monitoring data in a certain region suffers quality degradation due to sensor drift or temporary communication interruption (the confidence score output by the data quality detection module is below 0.6), the meta-network reduces the data loss weight for that region (e.g., to 0.3 to 0.5 times the baseline value).

[0048] In this embodiment, the meta-network takes loss residual, spatial gradient and historical trend information as input, and senses the degree of constraint satisfaction in each spatial region in real time and dynamically adjusts the weight allocation. This enables the automatic increase of physical loss weight in regions where physical constraints are not fully satisfied to force mechanical consistency, and the automatic suppression of data loss weight in regions where data quality is reduced due to sensor failure or interference, so as to prevent low-quality data from polluting physical deduction. This ensures that the physical consistency and data consistency of each region are optimally balanced throughout the training process.

[0049] like Figure 2 As shown, a flowchart of a mine rockburst prevention and control method according to an embodiment of the present invention is provided, including: Step S201: Obtain the stress field prediction results, energy field prediction results, and cognitive uncertainty map of the mine working face. The cognitive uncertainty map represents the distribution of the prediction confidence of the stress field and energy field prediction results in various spatial regions of the working face and mining area. Step S202: Based on the cognitive uncertainty diagram, combined with the stress field prediction results and the energy field prediction results, output a graded response decision; Step S203: Input the stress field prediction results and the energy field prediction results into the prevention and control scheme generation module to generate a targeted prevention and control scheme, and input the targeted prevention and control scheme into the multi-scale hierarchical coupled physical information neural network model for digital twin advanced simulation, and output the digital twin advanced simulation results of the prevention and control scheme. The simulation results include the predicted mean and confidence interval of the stress field and energy field after the scheme is implemented. Step S204: After implementing the prevention and control plan, the evaluation result of the prevention and control effect is used as the feedback signal of the multi-scale hierarchical coupled physical information neural network model to trigger the online Bayesian update of the multi-scale hierarchical coupled physical information neural network model, so that the parameters of the multi-scale hierarchical coupled physical information neural network model co-evolve with the prevention and control plan.

[0050] Specifically, the mine rockburst prevention and control method of this embodiment is described using the stress field prediction results, energy field prediction results, and cognitive uncertainty diagram of the mine working face output by the aforementioned mine rockburst monitoring method as input. However, the prevention and control method of the present invention is not limited to this input source, and can also use the stress field prediction results, energy field prediction results, and cognitive uncertainty diagram of the mine working face output by any source as input.

[0051] In step S201, the stress field prediction results, energy field prediction results, and cognitive uncertainty diagram of the mine working face are obtained. The stress field prediction results include the predicted values ​​of each component of the stress tensor in each spatial region of the working face and mining area. The energy field prediction results include the predicted values ​​of the elastic strain energy density in each spatial region. The cognitive uncertainty diagram characterizes the distribution of prediction confidence of the stress field and energy field prediction results in each spatial region of the working face and mining area.

[0052] In step S202, based on the cognitive uncertainty diagram and combined with the stress field prediction results and energy field prediction results, a graded response decision is executed: when the confidence level of the impact hazard level prediction is higher than 95%, the highest level evacuation warning is triggered; when the confidence level is between 80% and 95%, an enhanced monitoring warning is triggered; when the confidence level is lower than 80%, sensor optimization deployment suggestions are output.

[0053] In step S203, the stress field prediction results and energy field prediction results are input into the prevention and control scheme generation module. The prevention and control scheme generation module performs hazard feature identification: scanning the stress concentration factor (the ratio of the local maximum principal stress to the regional average principal stress) in the entire space domain, and marking regions with a factor greater than 2.0 as high stress gradient regions; scanning the elastic strain energy density distribution, and marking regions with a factor greater than 100 KJ / m 3 The area is designated as an area of ​​abnormal energy accumulation; the growth trend of microseismic event density and total released energy over the past 24 hours is analyzed, and areas with growth rates exceeding a preset rate threshold are marked as active damage areas.

[0054] The prevention and control plan generation module includes the following expert rule base: when the suspended area of ​​a high-level, solid roof exceeds 500m². 2 If the total energy of microseismic events shows an upward trend for three consecutive days, it is judged as a risk of roof-type rockburst, and hydraulic fracturing or deep-hole blasting is recommended to relieve pressure. If the width of the remaining coal pillar is less than 20m and the stress concentration factor continues to increase, it is judged as a risk of coal pillar-type rockburst, and large-diameter (e.g., greater than or equal to 100mm) boreholes are recommended to relieve pressure, with the borehole spacing being 5-8 times the borehole diameter.

[0055] The generated targeted control scheme outputs the following parameters in a structured data format: the three-dimensional starting coordinates, azimuth (angle with due north), dip (angle with the horizontal plane), design borehole diameter, design borehole depth, and design spacing of the pressure relief borehole; or the target layer depth, estimated fracture propagation direction, design injection pressure, and design injection flow rate of the hydraulic fracturing.

[0056] Before the targeted prevention and control plan is issued, the parameters of the plan are transformed into input conditions for a digital twin model. These parameters are then converted into boundary conditions or material parameter changes and fed back into a multi-scale, layered, coupled physical information neural network model to simulate the redistribution of stress and energy fields after the plan's implementation. The parameters for pressure relief boreholes are transformed into a cylindrical coordinate system with each borehole axis as the reference point, reducing the elastic modulus of the coal and rock mass using a linear reduction function within a range of 3-5 times the borehole radius. The parameters for hydraulic fracturing are transformed into applying fracture opening displacement boundary conditions along the fracture propagation direction at the target stratum. The displacement is calculated using linear elastic fracture mechanics formulas based on the injection pressure and rock fracture toughness.

[0057] A multi-scale, hierarchically coupled physical information neural network model was used to simulate the stress field redistribution after the implementation of the scheme in three scale subnets. The time span was set to 24 to 72 hours after the scheme implementation. Two key indicators of the simulation results were output: the expected value of stress reduction (the relative reduction of the average maximum principal stress in the high-stress area before and after the scheme implementation) and the expected value of energy release efficiency (the ratio of released elastic strain energy to stored energy during the simulation), as well as the 95% confidence intervals of these two indicators. The control scheme was deemed to meet the safety standards and approved for implementation only when the expected value of stress reduction was greater than 30% and the lower bound of the confidence interval of the expected value of stress reduction was greater than 15%.

[0058] In step S204, after the implementation of the prevention and control plan, the changes in indicators in the area are continuously monitored at a preset time resolution (e.g., 1 minute), including the frequency of microseismic events, the total energy released by microseismic events, and the slope parameter (obtained by least-squares linear fitting of the magnitude-frequency distribution of microseismic events). 72 hours after the implementation of the prevention and control plan, the evaluation indicators of the prevention and control effect are calculated one by one, including stress drop, daily attenuation rate of total microseismic energy, and change in the slope parameter. Specifically, the evaluation is as follows: 1) When the stress drop is greater than 1.3, it is rated as good; when the stress drop is greater than 1.1 and less than or equal to 1.3, it is rated as average; when the stress drop is less than or equal to 1.1, it is rated as insufficient. 2) When the daily attenuation rate of total microseismic energy is greater than 0.5, it is rated as good. 3) When the change in the slope parameter is greater than 0.3, it is rated as good.

[0059] The evaluation of the prevention and control effect is compared with the predicted values ​​in the pre-implementation phase. If the deviation between the measured value and the predicted value of any indicator exceeds a preset deviation threshold (e.g., the pre-implementation predicted a stress reduction of 40%, but the measured value is only 28%, a deviation of -30%), then an online Bayesian update of the multi-scale hierarchical coupled physical information neural network model is triggered, enabling the parameters of the multi-scale hierarchical coupled physical information neural network model to co-evolve with the prevention and control plan. The update goal of the online Bayesian update is to correct the posterior distribution of physical parameters after absorbing the observation data of this construction. The current posterior distribution (before implementation) is used as the prior distribution for the next update (after implementation). The measured data of the prevention and control effect of this construction is used as the new observation data. The joint posterior distribution of physical parameters is updated through the Bayesian formula. The expected value of the updated parameter distribution moves towards the direction of the measured results, while the uncertainty of parameter estimation (posterior variance) gradually decreases with the accumulation of observation data.

[0060] This embodiment enables real-time feedback of prevention and control effects to drive the co-evolution of model parameters and prevention and control strategies, solving the problem of rapid performance degradation of traditional static models as mining progresses. It improves the accuracy and confidence of subsequent predictions, forming a closed-loop adaptive control system with "monitoring as simulation, early warning as prevention and control, and effect as feedback" throughout the entire process, fundamentally solving the problem of accurate early warning and intelligent prevention and control of rockburst.

[0061] In one embodiment, step S202 further includes: Obtain the parameter prior distribution from the preset prior distribution knowledge base that is most similar to the geological conditions of the new mining area, and use a preset number of effective rockburst events as training samples to correct the parameters.

[0062] Specifically, from the accumulated PINN training results of multiple mining areas (such as 3-5 mining areas with different geological conditions), the posterior distributions of the following five key mechanical parameters for each mining area are extracted (the mean and posterior standard deviation are recorded respectively): elastic modulus, Poisson's ratio, cohesion, internal friction angle, and tensile strength. These parameters are archived according to geological type and burial depth range as two-level classification labels, constructing a three-dimensional indexed knowledge base of the prior distributions of coal and rock mechanical parameters.

[0063] When deployed in a new mining area, the corresponding parameter test values ​​and geological condition description text are extracted from the geological exploration report of the target mining area. After performing Term Frequency-Inverse Document Frequency (TF-IDF) encoding on the geological condition description text, the cosine similarity with the description text of each archived entry in the knowledge base is calculated. Combining the deviation of the numerical parameter test values, a weighted comprehensive similarity score is calculated, and the posterior parameter distribution of the entry with the highest similarity score is selected as the Bayesian prior of the PINN for the target mining area.

[0064] The preset number of effective rockburst events is 10 to 20.

[0065] In this embodiment, when deploying a new mining area, the posterior distribution of each physical parameter is extracted from the accumulated PINN training results of the mining area and a prior knowledge base is constructed. The new mining area only needs a very small number of effective impact events to gradually correct the prior distribution to the actual physical state of the target mining area through Bayesian updates, without having to learn the entire high-dimensional input-output mapping relationship from scratch. This shortens the deployment cycle, enables rapid adaptation across mining areas, and improves the cross-mining area versatility of rockburst prevention technology.

[0066] One embodiment of the present invention provides a storage medium for storing computer instructions. When the computer executes the computer instructions, it performs all the steps of the mine rockburst monitoring method or the mine rockburst prevention and control method as described in any of the above method embodiments.

[0067] like Figure 3 As shown, a hardware structure diagram of an electronic device for controlling an ammonia synthesis system according to an embodiment of the present invention is provided, including: At least one processor 301; and, Memory 302 is communicatively connected to at least one processor 301; wherein, The memory 302 stores instructions that can be executed by at least one processor 301, which enables the at least one processor 301 to perform the mine rockburst monitoring method or the mine rockburst prevention and control method as described above in any of the above method embodiments.

[0068] Figure 3 Take processor 301 as an example.

[0069] The electronic device is preferably an electronic control unit (ECU).

[0070] The electronic device may also include an input device 303 and an output device 304.

[0071] The processor 301, memory 302, input device 303 and output device 304 can be connected by a bus or other means. The figure shows an example of connection by bus.

[0072] The memory 302, as a non-volatile storage medium, can be used to obtain non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the mine rockburst monitoring method or mine rockburst prevention and control method in the embodiments of this application. Figures 1-2 The method flow is shown. The processor 301 executes various functional applications and data processing by running non-volatile software programs, instructions, and modules acquired in the memory 302, thereby realizing the mine rockburst monitoring method or mine rockburst prevention and control method in the above embodiments.

[0073] The memory 302 may include an acquisition program area and an acquisition data area. The acquisition program area may acquire the operating system and applications required for at least one function. The acquisition data area may acquire data created based on the use of the mine rockburst monitoring method or the mine rockburst prevention and control method. Furthermore, the memory 302 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 302 may optionally include memory remotely located relative to the processor 301. This remote memory can be connected via a network to the apparatus performing the mine rockburst monitoring method or the mine rockburst prevention and control method. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0074] The input device 303 can receive user clicks and generate signal inputs related to user settings and function control of the mine rockburst monitoring method or mine rockburst prevention method. The output device 304 may include display devices such as a display screen.

[0075] When the one or more modules are accessed in the memory 302 and are run by the one or more processors 301, the mine rockburst monitoring method or the mine rockburst prevention and control method in any of the above method embodiments are executed.

[0076] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0077] The above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the embodiments of the present invention have been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring rockburst in mines, characterized in that, include: The deformation and failure control equations and energy balance equations of coal and rock mass are transformed into penalty terms of the loss function. A multi-scale hierarchical coupled physical information neural network model is constructed, consisting of three hierarchical nested networks including micro-scale subnetworks, meso-scale subnetworks and macro-scale subnetworks, to perform cross-scale joint inference. A spatial heterogeneous graph is constructed using Euclidean distance between sensors, geological structure correlation, and observation signal correlation as multidimensional attributes of the graph edges. Virtual nodes are generated in the monitoring blind area where no sensors are deployed. The features of each virtual node are spatially interpolated and then spliced ​​with spatiotemporal coordinates to form coded input features that fuse spatial topological information. The coded input features are then input into the multi-scale hierarchical coupled physical information neural network model to output the predicted results of stress field and energy field. By introducing a probability distribution into the network weights of the trained multi-scale hierarchical coupled physical information neural network model through variational inference, the multi-scale hierarchical coupled physical information neural network model outputs a probability distribution prediction result containing confidence intervals, and generates a cognitive uncertainty map. The cognitive uncertainty map represents the prediction confidence distribution of the multi-scale hierarchical coupled physical information neural network model in various spatial regions of the working face and mining area.

2. The mine rockburst monitoring method as described in claim 1, characterized in that, The process involves transforming the deformation and failure control equations and energy balance equations of coal and rock mass into penalty terms of a loss function, constructing a multi-scale hierarchical coupled physical information neural network model with three nested networks (micro-scale, meso-scale, and macro-scale) for cross-scale joint inference, and previously included: Real-time access to monitoring data on microseismic activity, ground acoustics, borehole stress gauges, anchor bolt force gauges, support working resistance, and mining stress; Spatiotemporal alignment and fusion are performed using Kalman filtering, Bayesian fusion, and graph neural network-assisted spatiotemporal interpolation algorithms to form a four-dimensional monitoring dataset; The multi-scale hierarchical coupled physical information neural network model is driven to perform dynamic inversion with preset spatial and temporal resolutions.

3. The mine rockburst monitoring method as described in claim 1, characterized in that, The training of the three nested networks employs an alternating-joint hybrid training strategy anchored by mesoscale subnetworks, including: In each training cycle, the upstream and downstream subnets are frozen sequentially and trained separately for the mesoscale subnet, the macroscale subnet, and the microscale subnet, and then the freeze is lifted for joint fine-tuning of all parameters.

4. The method for monitoring rockburst in mines as described in claim 1, characterized in that, During the training process of the multi-scale hierarchical coupled physical information neural network model, the meta-network drives the adaptive physical constraint weight adjustment. The meta-network takes the current residual magnitude, spatial gradient distribution and historical change trend of each loss term as input, and generates local adaptive loss weights for each spatial region in real time. By dynamically adjusting the relative contributions of data loss terms and physical loss terms, a balance is maintained between monitoring data fitting and physical equation constraints.

5. The mine rockburst monitoring method as described in claim 4, characterized in that, The meta-network-driven adaptive physical constraint weight adjustment includes: If the difference between the current physical residual of the partial differential equation in a spatial region and the historical physical residual in the previous time exceeds a preset residual threshold, the meta-network increases the loss weight of the partial differential equation in that region to a preset multiple of the benchmark value. If the confidence score of the monitoring data in one of the spatial regions is lower than the preset quality score threshold, the meta-network reduces the data loss weight of that spatial region.

6. The mine rockburst monitoring method as described in claim 4, characterized in that, The meta-network includes a region encoder and a weight generator. The region encoder compresses the input feature vector of each spatial region into a region embedding vector, and concatenates it with the spatial coordinate encoding of the region to form a region representation vector. The weight generator takes the region representation vector as input and outputs the local adaptive loss weights for that region. The meta-network shares parameters across all spatial regions.

7. A method for preventing and controlling rockbursts in mines, characterized in that, include: Obtain the stress field prediction results, energy field prediction results, and cognitive uncertainty map of the mine working face. The cognitive uncertainty map represents the prediction confidence distribution of the stress field and energy field prediction results in various spatial regions of the working face and mining area. Based on the cognitive uncertainty diagram, combined with the stress field prediction results and the energy field prediction results, a hierarchical response decision is output. The stress field prediction results and the energy field prediction results are input into the prevention and control scheme generation module to generate a targeted prevention and control scheme. The targeted prevention and control scheme is then input into the multi-scale hierarchical coupled physical information neural network model for digital twin advanced simulation. The digital twin advanced simulation results of the prevention and control scheme are output, including the predicted mean and confidence interval of the stress field and energy field after the scheme is implemented. After implementing the prevention and control plan, the evaluation result of the prevention and control effect is used as the feedback signal of the multi-scale hierarchical coupled physical information neural network model, triggering the online Bayesian update of the multi-scale hierarchical coupled physical information neural network model, so that the parameters of the multi-scale hierarchical coupled physical information neural network model co-evolve with the prevention and control plan.

8. The method for preventing and controlling rockbursts in mines as described in claim 7, characterized in that, The output hierarchical response decision previously included: Obtain the parameter prior distribution from the preset prior distribution knowledge base that is most similar to the geological conditions of the new mining area, and use a preset number of effective rockburst events as training samples to correct the parameters.

9. A storage medium, characterized in that, The storage medium stores computer instructions, which, when executed by the computer, are used to perform all the steps of the mine rockburst monitoring method as described in any one of claims 1-6 or the mine rockburst prevention and control method as described in claim 7 or 8.

10. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the mine rockburst monitoring method as described in any one of claims 1-6 or the mine rockburst prevention and control method as described in claim 7 or 8.

Images