Rock burst early warning method and system based on data-mechanism dual drive

Abstract

The invention discloses a data-mechanism dual-drive-based rock burst early warning method and system, and the method comprises the following steps: deploying a multi-modal sensor network to collect coal and rock stratum data, building a rock burst disaster precursor information sample database, providing a rock burst disaster multi-modal data precursor feature recognition algorithm, and carrying out the recognition of rock burst disaster multi-modal data precursor features. Mining the relevance between the multi-modal data and disaster-causing key risk indexes, and establishing a rock burst disaster multi-modal data prediction model; establishing a three-dimensional geological geometric model, fusing a multi-field coupling dynamics constitutive model and a catastrophe criterion, constructing a PINN physical information neural network prediction model of the rock burst disaster, and obtaining a time-space evolution rule of an energy field of a target area; providing a loss function coupling calculation method of a multi-modal data driving sample error and a physical driving control equation residual error, dynamic data and mechanism prediction result weight, comprehensively calculating a risk score, and accurately judging a top disaster danger level.

Description

Technical Field

[0001] The present invention relates to the technical field of impact dynamic disaster monitoring and early warning, and particularly to a rock burst early warning method and system based on dual drive of data and mechanism. Background Art

[0002] Rock burst is a dynamic disaster that seriously threatens the safe production of mines, manifested as the phenomenon that coal and rock masses suddenly and violently break and release energy under high stress, which can lead to roadway deformation, support damage and even casualties. The existing rock burst early warning methods mainly include the following categories: (1) Early warning methods based on a single parameter: such as microseismic monitoring method, drill cuttings method, electromagnetic radiation method, etc. These methods have their own advantages and disadvantages, but there are limitations in early warning relying only on a single parameter, which is prone to false alarms or missed alarms. For example, microseismic monitoring can capture the fracture activities of coal and rock strata, but it is difficult to distinguish harmful and harmless fractures; the drill cuttings method is intuitive but has a limited monitoring range; the electromagnetic radiation method has high sensitivity but is easily affected by external interference. (2) Data-driven methods based on statistical models: By establishing statistical regression models or machine learning models, analyze the correlation between the characteristics in historical data and the occurrence of rock bursts. These methods rely on a large amount of high-quality historical data, have high requirements for the distribution characteristics of the data, and lack the consideration of physical mechanisms, with limited generalization ability. (3) Mechanism-driven methods based on mechanical theories: By establishing geomechanical models, simulate the stress distribution and energy evolution of coal and rock masses during the mining process. These methods have clear physical meanings, but due to the complex and changeable geological conditions, it is difficult to accurately determine the model parameters, resulting in limited prediction accuracy. Most of the existing early warning methods choose one of the two ideas of "data-driven" and "mechanism-driven", lacking a technical solution that effectively integrates the advantages of both. The data-driven method captures feature correlations but lacks physical constraints, and the mechanism-driven method has a theoretical basis but is difficult to adapt to complex geological conditions. Therefore, there is an urgent need for a new type of early warning method that can combine the advantages of data-driven and mechanism-driven to improve the accuracy and reliability of rock burst early warning.

[0003] Through the retrieval of the prior art, it is found that the Chinese patent publication number is CN116227921.A, the publication date is June 6, 2023, and the patent name is: A warning method based on dual driving of data and physical models. This patent obtains the first original data, establishes a simulation model for risk warning based on the first original data, obtains a risk simulation result according to the simulation model, establishes a physical model for risk warning, the physical model includes indicators and weights, obtains the weight value of the physical model according to the indicators, obtains the second original data, obtains the indicator value of the physical model according to the second original data, obtains a risk warning result according to the indicator value and the weight value. When the error between the risk simulation result and the risk warning result is greater than or equal to the preset error threshold, iteratively update the indicator value and the weight value. When the error is less than the preset error threshold, stop iteratively updating the indicator value and the weight value, and output the risk warning result. The deficiencies of this method are as follows: Although a warning method based on dual driving of data and physical models is proposed, the relationship between the data and the physical model is not constructed, and the long short-term memory neural network method and the physics-informed neural network method are not used to process the results, resulting in poor warning accuracy and reliability. Summary of the Invention

[0004] In response to the problems and requirements raised above, this solution proposes a rockburst warning method and system based on data-mechanism dual driving, which can achieve the above technical objectives and bring many other technical effects due to the following technical features.

[0005] An object of the present invention is to propose a rockburst warning method based on data-mechanism dual driving, including the following steps:

[0006] S10: Arrange multi-modal sensors in the coal seam to collect physical parameters of coal and rock strata, microseismic data, electromagnetic radiation data, and stress monitoring data related to rockburst in real time;

[0007] S20: Clean the multi-modal data, reduce the random noise in the sample data, and perform standardization processing on the multi-modal data to construct a rockburst precursor information sample database;

[0008] S30: Use continuous wavelet transform to extract the time-frequency domain features and spatio-temporal correlation features of the signal from the preprocessed multi-modal data, and through the multi-modal neural network model of the long short-term memory neural network, explore the correlation between the multi-modal data and the key risk indicators of disaster-causing;

[0009] S40: Based on the on-site borehole exploration, in-situ stress testing, and mine pressure observation data, construct a three-dimensional geological and physical model of the target coal seam, determine the initial conditions and boundary constraints of the model, and combine with the dynamic critical criterion of the catastrophe criterion to construct a physics-informed neural network (PINN) for rockburst disasters, simulate the evolution law of the energy field distribution during the mining process, and evaluate the rockburst risk;

[0010] S50: Construct a coupling calculation method for the loss function of the multi-modal data-driven sample error and the physical-driven control equation residual, dynamically adjust the weights of data and mechanism, comprehensively calculate the risk score, establish a multi-level early warning response, automatically adjust the threshold of the hazard coefficient according to real-time data, and optimize the early warning response.

[0011] In addition, according to the rockburst early warning method based on data-mechanism dual drive of the present invention, the following technical features may also be included:

[0012] In an example of the present invention, in the step S30, continuous wavelet transform is used to extract the time-frequency domain features and spatio-temporal correlation features of the signal, and the correlation between multi-modal data and key disaster-causing risk indicators is mined, including the following steps:

[0013] S31: Use continuous wavelet transform to perform time-frequency domain analysis on the time series data, and extract the energy distribution, peak frequency, and time domain statistics of the multi-modal data signal;

[0014] S32: Use the regression analysis method to quantify the relationship between the data feature f i and the risk indicator r j : The specific expression is:

[0015]

[0016] where β0 is the intercept, β i is the regression coefficient, and ∈ is the error term;

[0017] S33: Use the Pearson correlation coefficient r to calculate the correlation between the multi-modal features and the risk indicators, and select the features with |r|>0.6 as the LSTM input. The expression of the correlation coefficient r is as follows:

[0018]

[0019] where x i is the feature value, y i is the risk indicator value, and represent the mean value, and n is the number of samples.

[0020] In an example of the present invention, the step S40 specifically includes the following steps:

[0021] S41: Obtain the thickness and dip angle of the rock formation through borehole exploration, construct a three-dimensional geomechanical model using finite element software, obtain the initial stress field through in-situ stress testing, apply the overburden load according to the burial depth h at the top of the model, and use fixed displacement constraints at the bottom and around;

[0022] S42: The three-dimensional geomechanical model adopts the Mohr-Coulomb criterion, and the input parameters include rock mass density, elastic modulus, Poisson's ratio, internal friction angle, and compressive strength. Through mine pressure monitoring, obtain the stress changes during the mining process, and use this as the input condition to simulate the variation laws of the stress field and energy field during the mining process;

[0023] S43: Based on the extracted multi-modal data features, optimize the parameters and adjust the boundary conditions of the physical model;

[0024] S44: Calculate the critical energy threshold using rock mechanics theory,

[0025] Calculate the actual cumulative energy E through microseismic event or stress sensor data measured ;

[0026] Judge whether the difference between the actual energy E faced by the target area measured and the critical energy E critical is greater than the tolerable safety margin δ. If it holds, trigger an early warning, and the calculation is as follows:

[0027] ΔE = E measured - E critical > δ

[0028] where the safety margin δ is obtained based on the analysis of historical disaster data;

[0029] S45: Construct a physics-informed neural network, embed the mechanical control equations into the network, and by inputting the spatio-temporal coordinates (x1, x2, x3, t), output the relevant physical quantities (u1, u2, u3, σ 11 , σ 22 , σ 33 , σ 12 , σ 13 , σ 23 ).

[0030] In an example of the present invention, in the step S43, the parameter optimization and boundary condition adjustment of the physical model specifically include the following steps:

[0031] Optimize using the Bayesian optimization framework to find a set of parameters θ to make the output M(θ) of the physical model as close as possible to the observed data D obs , to minimize the difference between the model prediction and the measured data, and the objective function is constructed as follows:

[0032] minθ ||M(θ) - D obs || 2 + R(θ)

[0033] where M(θ) is the output of the physical model under parameter θ, D obs is the observed data, and R(θ) is the regularization term;

[0034] The boundary condition update adopts a sliding time window strategy. In each new time window, based on the stress field and energy field distributions inverted from the monitoring data, the load conditions and displacement constraints on the model boundary are updated.

[0035] In an example of the present invention, in the step S45, constructing the physics-informed neural network further includes: establishing a minimized loss function, and the specific steps are as follows:

[0036] Data term:

[0037] In the formula, N d is the number of data points; u k (X m , t m ) is the k-th predicted physical quantity, is the k-th observed value, X m is the spatial coordinate (m); t m is the time (s); k is the variable index;

[0038] Physical term:

[0039] In the formula, N p is the number of collocation points; is the divergence of the stress tensor (MPa / m); i, j are the direction indices (1, 2, 3 represent x, y, z);

[0040] Boundary term:

[0041] In the formula, N b is the number of boundary points; u k (X p , t p ) is the k-th predicted boundary value; is the boundary condition value; X p is the boundary point coordinate; t p is the time;

[0042] Total loss function of PINN: L PINN = λ1L data + λ2L physics + λ3L boundary

[0043] Where λ is the weight coefficient.

[0044] In an example of the present invention, in step S50, the expression for constructing the coupling calculation method of the loss function of the multi-modal data-driven sample error and the physical-driven control equation residual is:

[0045] L total =α·L LSTM +β·L PINN

[0046]

[0047] β = 1 - α

[0048] Where α and β are dynamically adjusted by the gradient norm; is the gradient of the LSTM loss with respect to the parameters; is the gradient of the PINN loss with respect to the parameters.

[0049] In an example of the present invention, in step S50, dynamically adjusting the weights of data and mechanism includes: fusing the data and mechanism results through a dynamic weighting strategy and calculating the risk score, and its expression is:

[0050] RiskScore = w data P LSTM +w physics I EnergyThreshold

[0051]

[0052] w physics = 1 - w data

[0053] In the formula, the sensitivity coefficient k,; data uncertainty mechanism model error P LSTM is the probability (0 - 1) predicted by the LSTM model; I Energythreshold = 1 or 0.

[0054] In an example of the present invention, in step S50, comprehensively calculating the risk score, establishing a multi-level early warning response, automatically adjusting the threshold of the hazard coefficient according to real-time data, and optimizing the early warning response, specifically including the following steps:

[0055] S51: Risk level I: 0.3 ≤ RiskScore < 0.6, and the response measure is to increase the monitoring frequency;

[0056] S52: Risk level II: 0.6 ≤ RiskScore < 0.9, and the response measures are to limit the number of operating personnel and reduce the mining speed;

[0057] S53: Risk Level III: RiskScore ≥ 0.9, and the response measures are to immediately stop the operation, evacuate personnel, and activate the emergency plan;

[0058] S54: Evaluate the system performance by recording the false alarm rate and missed alarm rate indicators, and combine with the actual situation of the mine to evaluate and adjust the warning threshold every quarter.

[0059] In an example of the present invention, the method for adjusting the warning threshold is as follows:

[0060] Calculation of the system comprehensive performance index:

[0061]

[0062] In the formula, w1, w2, and w3 are weight coefficients; FPR is the false alarm rate; TPR is the true positive rate; T lead is the warning lead time; T ref is the reference time;

[0063] When the false alarm rate > 10%, the warning thresholds at all levels are increased by 0.05;

[0064] When a missed alarm event occurs, the warning threshold of the corresponding level is decreased by 0.1;

[0065] According to the seasonal characteristics of rock burst disasters, the threshold is appropriately decreased by 0.05 - 0.1 during the high-incidence season.

[0066] Another object of the present invention is to propose a rock burst warning system based on data-mechanism dual-drive, which is characterized by including:

[0067] A data acquisition module configured to arrange multi-modal sensors in the coal seam to collect in real time the physical parameters of coal and rock strata, microseismic data, electromagnetic radiation data, and stress monitoring data related to rock burst;

[0068] A data processing module configured to clean the multi-modal data, reduce the random noise in the sample data, and perform standardization processing on the multi-modal data to construct a rock burst precursor information sample database;

[0069] A feature extraction module configured to extract the time-frequency domain features and spatio-temporal correlation features of the signal from the preprocessed multi-modal data by using continuous wavelet transform, and mine the correlation between the multi-modal data and the key disaster risk indicators through a multi-modal neural network model of a long short-term memory neural network;

[0070] The three-dimensional simulation module is configured to construct a three-dimensional geological and physical model of the target coal seam based on on-site borehole exploration, in-situ stress testing, and mine pressure observation data, determine the initial conditions and boundary constraints of the model, and combine with the dynamic critical criterion of the disaster-causing criterion to construct a physics-informed neural network (PINN) for rockburst disasters, simulate the evolution law of the energy field distribution during the mining process, and evaluate the rockburst risk;

[0071] The risk and early warning module is configured to construct a coupling calculation method for the loss function of multi-modal data-driven sample error and physics-driven control equation residual, dynamically adjust the weights of data and mechanism, comprehensively calculate the risk score, establish a multi-level early warning response, automatically adjust the threshold of the hazard coefficient according to real-time data, and optimize the early warning response.

[0072] The present invention has the following beneficial effects compared with the prior art:

[0073] The present invention adopts a data-mechanism dual-driven early warning mode, which organically combines the statistical relevance of data-driven and the physical constraints of mechanism-driven. It not only makes full use of the statistical laws in the monitoring data but also ensures that the prediction results conform to the physical mechanism of rockburst occurrence, significantly improving the early warning accuracy;

[0074] The present invention adopts a multi-modal sensor network to comprehensively capture various precursor signals before the occurrence of rockburst. By cross-modal feature fusion, it improves the expression ability of features and reduces the one-sidedness that may be brought by a single data source;

[0075] The present invention introduces a physics-informed neural network (PINN), embeds the control equation constraints into the deep learning framework, realizes the accurate simulation of the spatio-temporal evolution of the energy field, and enhances the physical interpretability of the model;

[0076] The present invention proposes a dynamic weight allocation strategy based on data uncertainty and physical error, increasing the weight of the physical model when the data quality is poor and increasing the weight of the data model when the physical model parameters are inaccurate, making the system have an adaptive ability;

[0077] The present invention establishes a multi-level early warning threshold automatic adjustment mechanism, dynamically optimizes the early warning standard in combination with indicators such as false alarm rate and missed alarm rate, and further improves the reliability and applicability of the early warning system.

[0078] In the following, the optimal embodiments of implementing the present invention will be described in more detail with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly introduced below. Among them, the accompanying drawings are only used to show some embodiments of the present invention and do not limit all embodiments of the present invention thereto.

[0080] Figure 1 It is a flowchart of a rockburst early warning method based on data - mechanism dual - drive according to an embodiment of the present invention;

[0081] Figure 2 It is a schematic diagram of the deployment of a multi - modal sensor network according to an embodiment of the present invention;

[0082] Figure 3 It is a framework diagram of a long short - term memory neural network model (LSTM) according to an embodiment of the present invention;

[0083] Figure 4 It is a data - driven rockburst early warning process according to an embodiment of the present invention;

[0084] Figure 5 It is a structure diagram of a PINN physics - informed neural network according to an embodiment of the present invention;

[0085] Figure 6 It is a data - mechanism dual - drive coupling early warning model diagram according to an embodiment of the present invention. Detailed implementation manners

[0086] In order to make the objectives, technical solutions, and advantages of the technical solutions of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the specific embodiments of the present invention. The same reference numerals in the drawings represent the same components. It should be noted that the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0087] Unless otherwise defined, the technical terms or scientific terms used herein shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention pertains. The "first", "second", and similar terms used in the specification and claims of this patent application for the present invention do not denote any order, quantity, or importance, but are only used to distinguish different components. Similarly, terms such as "a" or "one" do not necessarily denote a quantity limitation. The terms "including" or "comprising" and the like mean that the elements or items appearing before the term cover the elements or items listed after the term and their equivalents, without excluding other elements or items. The terms "connected" or "coupled" and the like are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms "upper", "lower", "left", "right", etc. are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0088] A rock burst early warning method based on data-mechanism dual-driving according to the first aspect of the present invention, as Figure 1 shown, includes the following steps:

[0089] S10: Arrange multi-modal sensors in the coal seam to collect physical parameters of coal and rock strata, microseismic data, electromagnetic radiation data, and stress monitoring data related to rock bursts in real time;

[0090] S20: Clean the multi-modal data, reduce the random noise in the sample data, and perform standardization processing on the multi-modal data to construct a precursor information sample database for rock bursts and capture the dangerous precursor signals of rock bursts;

[0091] S30: Use continuous wavelet transform to extract the time-frequency domain features and spatio-temporal correlation features of the signal from the preprocessed multi-modal data, and through the multi-modal neural network model of the long short-term memory neural network, explore the correlation between the multi-modal data and the key risk indicators of disaster-causing;

[0092] S40: Based on on-site borehole exploration, in-situ stress testing, and mine pressure observation data, construct a three-dimensional geological physical model of the target coal seam, determine the initial conditions and boundary constraints of the model, and combine the dynamic critical criterion of the disaster-causing criterion to construct a PINN physical information neural network for rock burst disasters, simulate the evolution law of the energy field distribution during the mining process, and evaluate the rock burst risk;

[0093] S50: Construct a coupling calculation method for the loss function of the multi-modal data-driven sample error and the physical-driven control equation residual, dynamically adjust the weights of data and mechanism, comprehensively calculate the risk score, establish a multi-level early warning response, automatically adjust the threshold of the danger coefficient according to real-time data, and optimize the early warning response.

[0094] This early warning method adopts a data-mechanism dual-driving early warning mode, which organically combines the statistical correlation of data driving with the physical constraint of mechanism driving. It not only makes full use of the statistical laws in the monitoring data but also ensures that the prediction results conform to the physical mechanism of rock burst occurrence, significantly improving the early warning accuracy;

[0095] This early warning method adopts a multi-modal sensor network to comprehensively capture various precursor signals before the occurrence of rock bursts, improves the expression ability of features through cross-modal feature fusion, and reduces the one-sidedness that may be brought by a single data source;

[0096] This early warning method introduces a PINN physical information neural network, embeds the control equation constraints into the deep learning framework, realizes the accurate simulation of the spatio-temporal evolution of the energy field, and enhances the physical interpretability of the model;

[0097] This early warning method proposes a dynamic weight allocation strategy based on data uncertainty and physical errors, increasing the weight of the physical model when the data quality is poor and increasing the weight of the data model when the physical model parameters are inaccurate, enabling the system to have an adaptive ability;

[0098] This early warning method establishes a multi-level early warning threshold automatic adjustment mechanism, dynamically optimizing the early warning criteria by combining indicators such as the false alarm rate and the missed alarm rate, further improving the reliability and applicability of the early warning system.

[0099] In an example of the present invention, as Figure 2 shown, the step S10 specifically includes:

[0100] S11: Deploy a multi-modal sensor network inside the target mine, such as evenly distributing microseismic sensors along the mining and excavation faces and main roadways, with a spacing of 50 - 100 meters; densely arrange electromagnetic radiation and stress sensors in high-risk areas (such as geological fault zones and the edges of goafs), with a spacing of 20 - 50 meters. Real-time collect the physical parameters of coal and rock strata related to rock bursts. Sensor types and parameters:

[0101] (1) Microseismic sensor: Monitor the low-frequency vibration signals generated by coal and rock fractures, with a frequency range of 10 Hz - 1 kHz, a sampling rate of 1 kHz, and a sensitivity of 0.1 mV / g.

[0102] (2) Electromagnetic radiation sensor: Detect the electromagnetic signals induced by stress changes, with a frequency range of 1 kHz - 10 MHz, a sampling rate of 10 MHz, and a range of 0 - 100 μV.

[0103] (3) Stress sensor: Measure the stress state of coal and rock strata, with a range of 0 - 50 MPa, a sampling rate of 1 Hz, and an accuracy of ±0.1 MPa.

[0104] S12: Use geophysical exploration, drilling and other methods to obtain information on the occurrence state and structural characteristics of coal and rock strata in the target research area, collect regional coal and rock samples, and obtain coal and rock physical and mechanical parameters such as density (unit: kg / m 3 ), elastic modulus (unit: GPa), Poisson's ratio, internal friction angle (unit: °), compressive strength (unit: MPa), etc. through laboratory tests.

[0105] S13: The data is stored in the form of a time series, and each record contains a timestamp, a sensor ID, a signal type, and an original value. The data acquisition system uses high-precision clock synchronization to ensure the time consistency of multi-modal data.

[0106] In an example of the present invention, the step S20 specifically includes the following steps:

[0107] S21: Data cleaning, removing outliers (such as stress > 50 MPa) beyond the sensor range and missing values with signal interruptions, and filling the missing data using linear interpolation; the formula is as follows:

[0108]

[0109] Where t1 and t2 are the times of the valid data points adjacent to the missing value, x(t1) and x(t2) are the corresponding values, and t is the time point for which interpolation is required (t1 < t < t2).

[0110] S22: Using the wavelet denoising method, selecting the Daubechies (db4) wavelet, with the decomposition level being 3 - 8 layers, to remove environmental noise. The formula for the denoising threshold λ is as follows:

[0111]

[0112] Where σ is the noise standard deviation (estimated through the stationary section of the signal), and N is the total number of signal sampling points (unitless);

[0113] Applying soft threshold processing to the wavelet coefficients:

[0114]

[0115] Where ω is the original wavelet coefficient, is the wavelet coefficient after denoising, and sign(ω) is the sign function, returning the positive or negative sign of ω (+1 or -1).

[0116] S23: Standardize each type of data sample x to map the data to the [0, 1] interval;

[0117]

[0118] Where μ is the mean, σ is the standard deviation, and x' is the standardized data; this step eliminates the dimension difference for subsequent model analysis.

[0119] S24: Integrate the preprocessed data into a structured database, with the data storage format being HDF5, supporting efficient query and large - scale processing. The database includes historical disaster data (recording the multi - modal data corresponding to the occurred rock burst events, as positive samples) and normal mining background data (recording the monitoring data when no disasters occur, as negative samples).

[0120] In an example of the present invention, in the step S30, as Figure 4 shown, use continuous wavelet transform to extract the time - frequency domain features and spatio - temporal correlation features of the signal from the preprocessed multi - modal data, including the following steps:

[0121] S31: Use continuous wavelet transform (CWT) to perform time-frequency domain analysis on time series data (such as microseismic waveform signals) to extract the energy distribution, peak frequency, and time-domain statistics of multimodal data signals;

[0122]

[0123] Where x(t) is the input signal; ψ(t) is the selected wavelet function, such as Morlet wavelet; ψ * represents complex conjugate; a is the scale parameter; b is the translation parameter; t is the time;

[0124] Output the CWT coefficient matrix and extract the features as follows:

[0125] (1) Energy distribution: Energy(a) = ∑b|W(a,b)| 2

[0126] (2) Main frequency: Freq max =argmax a Energy(a)

[0127] (3) Time domain statistics: root mean square (RMS), peak, kurtosis, and skewness.

[0128] S32: Using regression analysis method, quantify data features f i and risk indicator r j (such as stress concentration and energy release): The specific expression is:

[0129]

[0130] Among them, β0 is the intercept, β i is the regression coefficient, ∈ is the error term;

[0131] S33: Use the Pearson correlation coefficient r to calculate the correlation between multimodal features and risk indicators (such as energy accumulation), and select features with |r|>0.6 as LSTM input. The expression of the correlation coefficient r is as follows:

[0132]

[0133] Among them, x i is the eigenvalue, y i is the risk indicator value, and represents the mean, and n is the number of samples.

[0134] In one example of the present invention, in step S30, as Figure 3 As shown in Figure 2, the multimodal neural network model of the long short-term memory neural network includes:

[0135] Construct an LSTM network structure. The input layer receives feature vectors; the hidden layer contains multiple LSTM units (usually set to 64 - 128 nodes); the output layer uses the sigmoid activation function to output the risk probability P LSTM , ranging from 0 to 1;

[0136] The core calculations of the LSTM unit include:

[0137] Forget gate: f t = σ(W f ·|h t-1 ,x t | + b f )

[0138] Input gate: i t = σ(W i ·|h t-1 ,x t | + b i )

[0139] Candidate state:

[0140] State update:

[0141] Output gate: o t = σ(W o ·|h t-1 ,x t | + b o )

[0142] Hidden state: h t = o t ·tanh(C t )

[0143] where σ is the sigmoid activation function, W and b are the weights and biases respectively, h t-1 is the hidden state at the previous moment, and x t is the current input.

[0144] Use historical data (including positive and negative samples) for supervised learning to minimize the loss function as follows:

[0145]

[0146] where y i is the true risk value (0 or 1, dimensionless), is the predicted value (range [0,1], dimensionless), and N is the number of samples.

[0147] The training parameters are set as follows: Adam optimizer is selected; the learning rate is 0.001; the batch size is 32; the number of training epochs is 200.

[0148] In an example of the present invention, step S40 specifically includes the following steps:

[0149] S41: Obtain parameters such as rock formation thickness and dip angle through borehole exploration, and use FLAC 3D finite element software to construct a three-dimensional geomechanical model. The model uses tetrahedral elements or hexahedral elements, with a mesh scale of 1m×1m×1m and a total number of elements of about 105. Obtain the initial stress field through in-situ stress testing. The overlying load (σ z =ρgh, g = 9.8m / s 2 ) is applied to the top of the model according to the buried depth h, and fixed displacement constraints are used at the bottom and around;

[0150] S42: The three-dimensional geomechanical model adopts the Mohr-Coulomb criterion, and the input parameters include rock mass density, elastic modulus, Poisson's ratio, internal friction angle, compressive strength, etc. Through mine pressure monitoring, obtain the stress changes during the mining process, and use this as the input condition to simulate the variation laws of the stress field and energy field during the mining process;

[0151] S43: Based on the extracted multi-modal data features, optimize the parameters of the physical model and adjust the boundary conditions;

[0152] S44: Calculate the critical energy threshold using rock mechanics theory as follows:

[0153]

[0154] where σ c is the critical stress of the rock mass (MPa), determined by on-site testing or empirical formula; E is the elastic modulus (GPa), measured through rock mechanics experiments; V rock is the volume of the dangerous area, m 3 ;

[0155] Calculate the actual cumulative energy E measured through microseismic events or stress sensor data as follows:

[0156]

[0157] where e(t) is the energy release rate per unit time;

[0158] Judge whether the difference between the actual energy E measured in the target area and the critical energy E critical is greater than the tolerable safety margin δ. If it holds, trigger an early warning, and the calculation is as follows:

[0159] ΔE = E measured - E critical > δ

[0160] Among them, the safety margin v is obtained based on the analysis of historical catastrophe data;

[0161] S45: Construct a physics-informed neural network (PINN), embed the mechanical control equations into the network, and by inputting the spatio-temporal coordinates (x1, x2, x3, t), output the relevant physical quantities (u1, u2, u3, σ 11 , σ 22 , σ 33 , σ 12 , σ 13 , σ 23 ).

[0162] In an example of the present invention, in the step S43, the parameter optimization and boundary condition adjustment of the physical model specifically include the following steps:

[0163] Optimize using the Bayesian optimization framework to find a set of parameters θ such that the output M(θ) of the physical model is as close as possible to the observed data D obs , in order to minimize the difference between the model prediction and the measured data, and the objective function is constructed as follows:

[0164] min θ ||M(θ) - D obs || 2 + R(θ)

[0165] Among them, M(θ) is the output of the physical model under the parameters θ, D obs is the observed data, and R(θ) is the regularization term;

[0166] The boundary condition update adopts a sliding time window strategy. In each new time window, based on the stress field and energy field distributions inversed from the monitoring data, the load conditions and displacement constraints on the model boundary are updated. The boundary condition update frequency is determined according to the change rate of the monitoring data. When the change is significant, the update frequency is increased, and when the change is gentle, the update frequency is decreased, which not only ensures the real-time performance of the model but also controls the calculation cost.

[0167] In an example of the present invention, as Figure 5 shown, in the step S45, constructing the physics-informed neural network further includes: establishing a loss function to be minimized, and the specific steps are as follows:

[0168] Data term:

[0169] In the formula, N d is the number of data points; u k (X m , tm ) is the k-th predicted physical quantity, is the k-th observed value, X m is the spatial coordinate (m); t m is the time (s); k is the variable index (1 - 9, including displacement and stress components);

[0170] Physical term:

[0171] In the formula, N p is the number of collocation points; is the divergence of the stress tensor (MPa / m); i, j are the direction indices (1, 2, 3 represent x, y, z);

[0172] Boundary term:

[0173] In the formula, N b is the number of boundary points; u k (X p , t p ) is the k-th predicted boundary value; is the boundary condition value (such as the fixed displacement is 0 m); X p is the boundary point coordinate; t p is the time;

[0174] Total loss function of PINN: L PINN = λ1L data + λ2L physics + λ3L boundary

[0175] In the formula, λ is the weight coefficient (such as 1, 1, 0.1).

[0176] In an example of the present invention, in the step S50, the expression for constructing the coupling calculation method of the loss function of the multi-modal data-driven sample error and the physical-driven control equation residual is:

[0177] L total = α·L LSTM + β·L PINN

[0178]

[0179] β = 1 - α

[0180] where α and β are dynamically adjusted by the gradient norm; is the gradient of the LSTM loss with respect to the parameters; is the gradient of the PINN loss with respect to the parameters.

[0181] Iteratively optimize the neural network model based on the obtained loss function, and the training parameters are set as follows: Adam optimizer, learning rate 0.001, batch size 128, and number of training epochs 5000.

[0182] In an example of the present invention, in the step S50, as Figure 6 shown, dynamically adjusting the weights of data and mechanism includes: fusing the data and mechanism results through a dynamic weighting strategy, and calculating the risk score, the expression of which is:

[0183] RiskScore = w data P LSTM + w physics I EnergyThreshold

[0184]

[0185] w physics = 1 - w data

[0186] In the formula, the sensitivity coefficient k is determined through cross-validation to ensure moderate sensitivity of weight adjustment, and usually takes 1; the data uncertainty is calculated from the standard deviation and mean based on the last 100 data points; the mechanism model error is calculated based on 50 historical calibration points; P LSTM is the probability (0 - 1) predicted by the LSTM model; I EnergyThreshold = 1 (when ΔE > δ) or 0.

[0187] In an example of the present invention, in the step S50, comprehensively calculate the risk score, establish a multi-level early warning response, automatically adjust the threshold of the hazard coefficient according to real-time data, and optimize the early warning response, which specifically includes the following steps:

[0188] S51: Yellow early warning (risk level I): 0.3 ≤ RiskScore < 0.6, and the response measure is to increase the monitoring frequency;

[0189] S52: Orange early warning (risk level II): 0.6 ≤ RiskScore < 0.9, and the response measure is to limit the number of operating personnel and reduce the mining speed;

[0190] S53: Red early warning (risk level III): RiskScore ≥ 0.9, and the response measure is to immediately stop the operation, evacuate the personnel, and start the emergency plan;

[0191] S54: Evaluate the system performance by recording indicators such as the false alarm rate (target < 5%) and the missed alarm rate (target 0%), and evaluate and adjust the early warning threshold every quarter in combination with the actual situation of the mine.

[0192] In an example of the present invention, the method for adjusting the warning threshold is as follows:

[0193] Calculation of the system comprehensive performance index:

[0194]

[0195] In the formula, w1, w2, w3 are weight coefficients (unitless, such as 1 / 3); FPR is the false positive rate (range [0, 1]); TPR is the true positive rate (range [0, 1]); T lead is the warning lead time (unit: second); T ref is the reference time (such as the average lead time of disasters, unit: second);

[0196] When the false positive rate > 10%, the warning thresholds at all levels are increased by 0.05;

[0197] When a missed alarm event occurs, the warning threshold of the corresponding level is decreased by 0.1;

[0198] According to the seasonal characteristics of rock burst disasters, the threshold is appropriately decreased by 0.05 - 0.1 in the high-incidence season (such as the rainy season).

[0199] According to a rock burst warning system based on data - mechanism dual - drive in the second aspect of the present invention, it includes:

[0200] A data acquisition module, configured to arrange multimodal sensors in the coal seam to collect physical parameters of coal and rock strata, microseismic data, electromagnetic radiation data, and stress monitoring data related to rock burst in real time;

[0201] A data processing module, configured to clean the multimodal data, reduce random noise in the sample data, perform standardized processing on the multimodal data, construct a sample database of rock burst precursor information, and capture dangerous precursor signals of rock burst;

[0202] A feature extraction module, configured to extract time - frequency domain features and spatio - temporal correlation features of signals from the pre - processed multimodal data by using continuous wavelet transform, and mine the correlation between multimodal data and key disaster - causing risk indicators through a multimodal neural network model of a long short - term memory neural network;

[0203] A three - dimensional simulation module, configured to construct a three - dimensional geological physical model of the target coal seam based on on - site borehole exploration, in - situ stress testing, and mine pressure observation data, determine the initial conditions and boundary constraints of the model, combine with the dynamic critical criterion of the disaster - causing criterion, construct a PINN physics - informed neural network for rock burst disasters, simulate the evolution law of the energy field distribution during the mining process, and evaluate the rock burst risk;

[0204] The risk and early warning module is configured to construct a coupled calculation method for the loss function of multi-modal data-driven sample error and physically-driven control equation residuals, dynamically adjust the weights of data and mechanism, comprehensively calculate the risk score, establish a multi-level early warning response, automatically adjust the threshold of the risk coefficient according to real-time data, and optimize the early warning response.

[0205] Specifically, the early warning system further includes: a data acquisition subsystem, a data analysis and mechanism modeling subsystem, an intelligent early warning decision-making subsystem, and an intelligent early warning decision-making subsystem;

[0206] (1) Data acquisition subsystem:

[0207] Multi-modal sensor network: including microseismic sensors, electromagnetic radiation sensors, and stress sensors;

[0208] Data acquisition equipment: sensor signal conditioning module, high-speed A / D conversion module, data cache module;

[0209] Communication network: industrial Ethernet backbone network (transmission rate ≥ 1 Gbps), fiber optic communication system;

[0210] Data preprocessing unit: edge computing device, realizing preliminary data filtering and compression.

[0211] (2) Data analysis and mechanism modeling subsystem:

[0212] Data processing server: configured with a high-performance CPU (≥ 16 cores) and a GPU acceleration card (≥ 16 GB video memory);

[0213] Data storage device: combination of high-speed SSD (≥ 2 TB) and large-capacity storage array (≥ 20 TB);

[0214] Model training module: realizing the training and optimization of LSTM network and PINN network;

[0215] Parameter inversion module: updating physical model parameters in real time according to monitoring data;

[0216] Visualization engine: 3D modeling and visualization of risk distribution.

[0217] (3) Intelligent early warning decision-making subsystem:

[0218] Risk assessment module: calculating the comprehensive risk score in real time;

[0219] Early warning grading module: determining the early warning level according to the risk score;

[0220] Response decision module: automatically generating response suggestions according to the early warning level;

[0221] Alarm trigger module: controlling the audible and visual alarm device.

[0222] (4) Man - machine interaction and application subsystem:

[0223] Display terminals: The large - screen display system in the surface control center and explosion - proof tablets underground;

[0224] Acoustic - optical alarm: Acoustic - optical alarm devices arranged underground, supporting remote triggering;

[0225] Mobile application: The mobile APP for managers, supporting real - time viewing of monitoring data and early warning information;

[0226] Data interface: The data exchange interface with the mine safety production management system.

[0227] The specific layout and connection methods are as follows. As Figure 2 shown, inside the coal body, along at least one side of the coal mining direction, electromagnetic radiation sensors, stress sensors, and microseismic sensors are arranged. Among them, the electromagnetic radiation sensors and stress sensors are alternately arranged at intervals in turn. For example, they are alternately arranged every 20 meters, and the microseismic sensors are alternately arranged in turn. For example, they are alternately arranged every 50 meters; the electromagnetic radiation sensors, stress sensors, and microseismic sensors are connected to the signal acquisition instrument through signal connection cables. The signal acquisition instrument is connected to the multi - source information system. The signal acquisition instrument is powered by a mine - use power supply. If electromagnetic radiation sensors, stress sensors, and microseismic sensors are arranged on both sides of the coal mining direction, a time synchronizer needs to be set between the two signal acquisition instruments; the signal acquisition instrument controls the electromagnetic radiation sensors, stress sensors, and microseismic sensors to collect electromagnetic radiation data, microseismic data, and stress monitoring data respectively, and transmits the above data to the multi - source information system for processing.

[0228] This early warning system adopts a data - mechanism dual - drive early warning mode, which organically combines the statistical relevance of data - driven and the physical constraints of mechanism - driven. It not only makes full use of the statistical laws in the monitoring data but also ensures that the prediction results conform to the physical mechanism of rock burst occurrence, significantly improving the early warning accuracy;

[0229] This early warning system adopts a multi - modal sensor network to comprehensively capture various precursor signals before the occurrence of rock burst. By cross - modal feature fusion, it improves the expression ability of features and reduces the one - sidedness that may be brought by a single data source;

[0230] This early warning system introduces the PINN (Physics - informed Neural Network), embeds the control equation constraints into the deep learning framework, realizes the accurate simulation of the spatio - temporal evolution of the energy field, and enhances the physical interpretability of the model;

[0231] The early warning system proposes a dynamic weight allocation strategy based on data uncertainty and physical errors, increasing the weight of the physical model when the data quality is poor and increasing the weight of the data model when the physical model parameters are inaccurate, enabling the system to have an adaptive ability;

[0232] The early warning system establishes a multi-level early warning threshold automatic adjustment mechanism, dynamically optimizing the early warning standard by combining indicators such as the false alarm rate and the missed alarm rate, further improving the reliability and applicability of the early warning system.

[0233] In the foregoing, the exemplary implementation manners of the rock burst early warning method and system based on data-mechanism dual drive proposed by the present invention have been described in detail with reference to the preferred embodiments. However, those skilled in the art can understand that, without departing from the concept of the present invention, various modifications and variations can be made to the above specific embodiments, and various combinations can be made to the various technical features and structures proposed by the present invention, without exceeding the protection scope of the present invention. The protection scope of the present invention is determined by the appended claims.

Claims

1. A rockburst warning method based on data-mechanism dual drive, characterized in that It includes the following steps: S10: Arrange multimodal sensors in the coal seam to collect physical parameters of coal and rock strata, microseismic data, electromagnetic radiation data, and stress monitoring data related to rock burst in real time; S20: Clean the multimodal data, reduce the random noise in the sample data, and perform standardization processing on the multimodal data to construct a sample database of precursor information for rock burst; S30: Use continuous wavelet transform to extract the time-frequency domain features and spatio-temporal correlation features of the signal from the preprocessed multimodal data. Through the multimodal neural network model of the long short-term memory neural network, explore the correlation between the multimodal data and the key risk indicators of disaster-causing; S40: Based on on-site borehole exploration, in-situ stress testing, and mine pressure observation data, construct a three-dimensional geological and physical model of the target coal seam, determine the initial conditions and boundary constraints of the model, and combine the dynamic critical criterion of the disaster-causing criterion to construct a PINN physics-informed neural network for rock burst disasters, simulate the evolution law of the energy field distribution during the mining process, and evaluate the rock burst risk; S50: Construct a coupling calculation method for the loss function of multimodal data-driven sample error and physically-driven control equation residual, dynamically adjust the weights of data and mechanism, comprehensively calculate the risk score, establish a multi-level early warning response, automatically adjust the threshold of the hazard coefficient according to real-time data, and optimize the early warning response.

2. The rock burst early warning method based on data-mechanism dual drive according to claim 1, characterized in that In the step S30, using continuous wavelet transform to extract the time-frequency domain features and spatio-temporal correlation features of the signal from the preprocessed multimodal data includes the following steps: S31: Use continuous wavelet transform to perform time-frequency domain analysis on the time series data, and extract the energy distribution, peak frequency, and time domain statistics of the multimodal data signal; S32: Use the regression analysis method to quantify the data feature f i and the risk indicator r j The relationship is: The specific expression is: where β0 is the intercept, β i is the regression coefficient, and ∈ is the error term; S33: Use the Pearson correlation coefficient r to calculate the correlation between the multimodal features and the risk indicators, and select the features with |r|>0.6 as the input of the LSTM. The expression of the correlation coefficient r is as follows: Among them, x i is the eigenvalue, y i is the risk index value, and represent the mean value, and n is the number of samples.

3. The rock burst early warning method based on data-mechanism dual drive according to claim 1, characterized in that The step S40 specifically includes the following steps: S41: Obtain the rock layer thickness and dip angle through borehole exploration, use finite element software to construct a three-dimensional geomechanical model, obtain the initial stress field through in-situ stress testing, apply the overlying load according to the buried depth h at the top of the model, and use fixed displacement constraints at the bottom and around; S42: The three-dimensional geomechanical model adopts the Mohr-Coulomb criterion, and the input parameters include rock mass density, elastic modulus, Poisson's ratio, internal friction angle, and compressive strength. Through mine pressure monitoring, obtain the stress change during the mining process, and use this as the input condition to simulate the change laws of the stress field and energy field during the mining process; S43: Based on the extracted multimodal data features, optimize the parameters of the physical model and adjust the boundary conditions; S44: Calculate the critical energy threshold using rock mechanics theory, and calculate the actual cumulative energy E through microseismic event or stress sensor data measured , and judge the actual energy E measured faced by the target area critical whether the difference from the critical energy E is greater than the tolerable safety margin δ. If it holds, trigger an alarm, and the calculation is as follows: ΔE = E measured - E critical > δ Among them, the safety margin δ is obtained based on the analysis of historical disaster data; S45: Construct a physical information neural network, embed the mechanical control equations into the network, and output relevant physical quantities (u1, u2, u3, σ 11 , σ 22 , σ 33 , σ 12 , σ 13 , σ 23 ) by inputting the spatio-temporal coordinates (x1, x2, x3, t).

4. The rock burst early warning method based on data-mechanism dual drive according to claim 3, characterized in that In the step S43, the parameter optimization and boundary condition adjustment of the physical model specifically include the following steps: Optimized using the Bayesian optimization framework to find a set of parameters θ such that the output M(θ) of the physical model is as close as possible to the observed data D obs , to minimize the difference between the model prediction and the measured data, and the objective function is constructed as follows: min θ ||M(θ) - D obs || 2 + R(θ) where M(θ) is the output of the physical model under parameter θ, D obs is the observed data, and R(θ) is the regularization term; The boundary condition update adopts a sliding time window strategy. In each new time window, based on the stress field and energy field distributions inversed from the monitoring data, the load conditions and displacement constraints on the model boundary are updated.

5. The data-mechanism dual-driven rock burst early warning method according to claim 3, characterized in that In the step S45, constructing the physics-informed neural network further includes: establishing a minimized loss function, and the specific steps are as follows: Data item: Where N d is the number of data points; u k (X m , t m ) is the k-th predicted physical quantity, is the k-th observed value, X m is the spatial coordinate (m); t m is the time (s); k is the variable index; Physical item: where N p is the number of collocation points; is the divergence of the stress tensor (MPa / m); i, j are the direction indices (1, 2, 3 represent x, y, z); Boundary term: Where N b is the number of boundary points; u k (X p , t p ) is the k-th predicted boundary value; is the boundary condition value; X p is the boundary point coordinate; t p is the time; PINN total loss function: L PINN = λ1L data + λ2L physics + λ3L boundary In the formula, λ is the weight coefficient.

6. The data-mechanism dual-driven rock burst early warning method according to claim 1, characterized in that In the step S50, the expression of the coupling calculation method of the loss function for constructing the multi-modal data-driven sample error and the physical-driven control equation residual is: L total = α·L LSTM + β·L PINN β＝1-α where α and β are dynamically adjusted by the gradient norm; is the gradient of the LSTM loss with respect to the parameters; is the gradient of the PINN loss with respect to the parameters.

7. The data-mechanism dual-driven rock burst early warning method according to claim 1, characterized in that In the step S50, dynamically adjusting the weights of data and mechanism includes: fusing the data and mechanism results through a dynamic weighting strategy, and calculating the risk score, and its expression is: RiskScore = w data P LSTM + w physics I EnergyThreshold w physics = 1 - w data where the sensitivity coefficient is k; data uncertainty mechanism model error P LSTM is the probability (0 - 1) predicted by the LSTM model; I EnergyThreshold = 1 or 0.

8. The data-mechanism dual-driven rock burst early warning method according to claim 1, characterized in that In the step S50, comprehensively calculating the risk score, establishing a multi-level early warning response, automatically adjusting the danger coefficient threshold according to the real-time data, and optimizing the early warning response, specifically including the following steps: S51: Risk level I: 0.3 ≤ RiskScore < 0.6, and the response measure is to increase the monitoring frequency; S52: Risk level II: 0.6 ≤ RiskScore < 0.9, and the response measure is to limit the number of operating personnel and reduce the mining speed; S53: Risk level III: RiskScore ≥ 0.9, and the response measure is to immediately stop the operation, evacuate the personnel, and start the emergency plan; S54: By recording the false alarm rate and missed alarm rate indicators, evaluating the system performance, and combining the actual situation of the mine, evaluating and adjusting the early warning threshold every quarter.

9. The data-mechanism dual-driven rock burst early warning method according to claim 8, characterized in that The adjustment method of the early warning threshold is as follows: Calculation of the system comprehensive performance index: where w1, w2, and w3 are weight coefficients; FPR is the false positive rate; TPR is the true positive rate; T lead is the early warning lead time; T ref is the reference time; When the false alarm rate > 10%, the early warning thresholds at all levels are increased by 0.05; After a missed alarm event occurs, the early warning threshold of the corresponding level is reduced by 0.1; According to the seasonal characteristics of the rock burst disaster, the threshold is appropriately reduced by 0.05 - 0.1 in the high-incidence season.

10. A rock burst early warning system based on data-mechanism dual drive, characterized in that Including: A data acquisition module configured to arrange multi-modal sensors in the coal seam to collect in real time the physical parameters of coal and rock strata, microseismic data, electromagnetic radiation data, and stress monitoring data related to rock burst; A data processing module configured to clean the multi-modal data, reduce the random noise in the sample data, and perform standardization processing on the multi-modal data to construct a rock burst precursor information sample database; A feature extraction module, configured to extract the time-frequency domain features and spatio-temporal correlation features of signals from the preprocessed multi-modal data by using continuous wavelet transform, and to mine the correlation between the multi-modal data and the key disaster-causing risk indicators through a multi-modal neural network model of a long short-term memory neural network; A three-dimensional simulation module, configured to construct a three-dimensional geological and physical model of the target coal seam based on on-site borehole exploration, in-situ stress testing and mine pressure observation data, determine the initial conditions and boundary constraints of the model, and combine with the dynamic critical criterion of the disaster-causing criterion to construct a physics-informed neural network (PINN) of rockburst disasters, simulate the evolution law of the energy field distribution during the mining process, and evaluate the rockburst risk; A risk and early warning module, configured to construct a coupling calculation method for the loss function of the multi-modal data-driven sample error and the physical-driven control equation residual, dynamically adjust the weights of the data and the mechanism, comprehensively calculate the risk score, establish a multi-level early warning response, automatically adjust the threshold of the hazard coefficient according to the real-time data, and optimize the early warning response.

Images