Shallow coal seam strong mine pressure dynamic early warning and cooperative control method and system based on multi-source data fusion

Abstract

The invention discloses a shallow coal seam strong mine pressure dynamic early warning and cooperative control method and system based on multi-source data fusion, and relates to the technical field of intelligent control. Coal seam multi-source sensing data is collected, a strong mine pressure feature vector is constructed, and the strong mine pressure feature vector is input into a mixed model of a deep recurrent neural network and a graph neural network; outputting an energy release index and a risk probability value, judging a coal seam strong mine pressure risk level, determining a fracturing position and fracturing parameters of directional hydraulic fracturing by adopting a self-adaptive grid division algorithm when a red early warning signal is triggered, and determining a fracturing parameter of directional hydraulic fracturing by adopting a self-adaptive grid division algorithm when an orange early warning signal is triggered. The spraying area and thickness distribution of energy absorption material spraying are determined through a reinforcement learning model, when a yellow early warning signal is triggered, a supporting resistance adjusting scheme of a hydraulic support is determined through a gradient descent algorithm, risk grading control is achieved, and the accuracy and practicability of coal seam early warning are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, and in particular to a method and system for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion. Background Art

[0002] Shallow coal seams face a unique geomechanical environment during mining. Due to the thin overburden thickness, stress transfer between the surface and the goaf is more direct, resulting in a significant increase in the frequency and intensity of severe mine pressure events. Shallow coal seams are unique in that their overburden is thin, their stress distribution is uneven, and they are significantly affected by surface activity, making it difficult for traditional mine pressure warning methods to accurately capture their dynamic characteristics. Furthermore, shallow coal seam areas often have complex geological structures, with widespread distribution of geological anomalies such as faults and folds, further increasing the difficulty of predicting severe mine pressure events.

[0003] During shallow coal mining, severe mine pressure events such as rock bursts and large-scale roof collapses frequently occur. These accidents not only threaten the lives of miners but also cause severe economic losses, including equipment damage and production interruptions. Traditional static early warning methods are unable to adapt to the complex and changing geological environment of shallow coal seams. Therefore, a dynamic, accurate, and multi-level early warning and coordinated control system is urgently needed to effectively prevent and control severe mine pressure events.

[0004] Existing control measures are often passive and simplistic, lacking differentiated strategies for different risk levels. When risk signals are detected, a blanket shutdown or evacuation of personnel is often adopted, impacting production efficiency.

[0005] To this end, the present invention proposes a method and system for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion. Summary of the Invention

[0006] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a method and system for dynamic early warning and coordinated control of high mine pressure in shallow coal seams based on multi-source data fusion, which implements risk classification management and improves the accuracy and practicality of coal seam early warning.

[0007] To achieve the above objectives, a method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion is proposed, which includes the following steps:

[0008] Step 1: Collect multi-source sensing data of the coal seam, and construct a strong mine pressure feature vector for the multi-source sensing data of the coal seam through wavelet transform and nonlinear dimensionality reduction algorithm;

[0009] The coal seam multi-source sensing data includes stress and strain data collected by distributed optical fiber sensors, energy release data collected by microseismic arrays, surface subsidence data collected by surface subsidence radars, and three-dimensional geological modeling data;

[0010] The strong rock pressure characteristic vector includes three key indicators: stress gradient distribution, energy release rate distribution and crack density distribution, which are stored in the form of a multi-dimensional array;

[0011] Step 2: Input the high-pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network, and output an energy release index and a risk probability value;

[0012] The hybrid model captures the changing patterns of temporal features through a deep recurrent neural network and extracts spatial correlation features using a graph neural network.

[0013] The energy release index represents the ratio of energy accumulation to release within the coal seam, and the risk probability value represents the possibility of a strong mine pressure event. Both are scalar data.

[0014] Step 3: Determine the risk level of strong mine pressure in the coal seam based on the energy release index and the risk probability value. When the energy release index exceeds the first threshold and the risk probability value exceeds the second threshold, a red warning signal is triggered and the process proceeds to step 4. When the energy release index exceeds the third threshold and the risk probability value exceeds the fourth threshold, an orange warning signal is triggered and the process proceeds to step 5. When the energy release index exceeds the fifth threshold and the risk probability value exceeds the sixth threshold, a yellow warning signal is triggered and the process proceeds to step 6. Otherwise, return to step 1 and continue monitoring.

[0015] Step 4: When a red warning signal is triggered, an adaptive grid division algorithm is used to determine the fracturing position and fracturing parameters of directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector;

[0016] The fracturing position and fracturing parameters include fracturing point coordinates, fracturing aperture, fracturing spacing and fracturing depth;

[0017] Step 5: When an orange warning signal is triggered, a reinforcement learning model is used to determine the spraying area and thickness distribution of the energy-absorbing material based on the crack density distribution in the strong rock pressure feature vector;

[0018] The spraying area and thickness distribution include the spraying area boundary coordinates and the spraying thickness at each point;

[0019] Step 6: When a yellow warning signal is triggered, a gradient descent algorithm is used to determine a support resistance adjustment scheme for the hydraulic support based on the energy release rate distribution in the strong rock pressure characteristic vector;

[0020] The support resistance adjustment scheme includes the resistance adjustment amount of each support position;

[0021] The method of collecting coal seam multi-source sensing data includes the following steps:

[0022] The stress and strain data are collected in real time based on the Brillouin scattering principle through an optical fiber network arranged around the coal seam.

[0023] A network of seismic detectors placed around the coal seam collects data on energy release from tiny fractures within the coal seam.

[0024] Synthetic aperture radar interferometry technology is used to obtain surface subsidence data caused by coal mining through satellite or ground radar systems.

[0025] The three-dimensional geological modeling data is obtained by constructing a three-dimensional geological structure model of the coal seam and its surrounding rock through drilling, geophysical exploration and historical mining data combined with geostatistical methods.

[0026] The stress and strain data, energy release data, surface settlement data and three-dimensional geological modeling data are matched with timestamps and spatial position coordinates of the coal seam to form the coal seam multi-source sensing data.

[0027] The method of constructing a strong mine pressure feature vector from the coal seam multi-source sensing data by wavelet transform and nonlinear dimensionality reduction algorithm includes the following steps:

[0028] Step 11: Perform wavelet transform preprocessing on the coal seam multi-source sensing data to extract the multi-scale feature matrix;

[0029] Step 12: Perform nonlinear dimensionality reduction on the multi-scale feature matrix after wavelet transformation to generate a low-dimensional feature space, reduce the data dimension and retain key information;

[0030] Step 13: Based on the low-dimensional feature space after dimensionality reduction, three key indicators are constructed: stress gradient distribution, energy release rate distribution, and crack density distribution;

[0031] Step 14: combining the three key indicators into a strong rock pressure feature vector;

[0032] Inputting the strong mine pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network to output an energy release index and a risk probability value includes the following steps:

[0033] Step 21: Construct a deep recurrent neural network to process the time series data of the strong mine pressure feature vector, generate a time series feature vector, and capture the change pattern in the time dimension;

[0034] Step 22: Construct a graph neural network to process the spatial data of the strong mine pressure feature vector, generate a spatial feature vector, and extract the correlation features in the spatial dimension;

[0035] Step 23: Use the attention mechanism to fuse the output features of the deep recurrent neural network and the graph neural network to generate a fused feature vector;

[0036] Step 24: Construct a prediction unit to generate an energy release index and a risk probability value based on the fusion features;

[0037] Step 25: Use the historically collected coal seam multi-source sensing data and the records of severe mine pressure events to train the hybrid model and output the energy release index and risk probability value;

[0038] The method of determining the fracturing position and fracturing parameters of directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector by using an adaptive grid division algorithm to form a directional hydraulic fracturing control instruction set includes the following steps:

[0039] Step 41: Based on the stress gradient distribution in the strong mine pressure characteristic vector, a three-dimensional stress field model of the coal seam is constructed, and stress concentration areas are identified;

[0040] Step 42: Based on the changing characteristics of the stress gradient, an adaptive meshing algorithm is applied to the identified stress concentration area to determine the mesh density distribution;

[0041] Step 43: Based on the grid density distribution of the adaptive grid division, the fracturing position is determined using the energy release maximization criterion;

[0042] Step 44: Calculate the fracturing parameters based on the determined fracturing location, combined with the stress field characteristics and the pre-collected surrounding rock parameters;

[0043] The method of using a reinforcement learning model to determine the spraying area and thickness distribution of the energy absorbing material spraying comprises the following steps:

[0044] Step 51: converting the crack density distribution data in the strong rock pressure characteristic vector into a three-dimensional space grid representation to construct an environmental state space;

[0045] Step 52: Define the reinforcement learning model as an action space in a two-dimensional discrete-continuous hybrid space, including two dimensions: spray area selection and thickness distribution;

[0046] Step 53: Based on the environment state space and action space, construct a reinforcement learning model based on a deep Q network, wherein the reinforcement learning model includes three core components: a state encoder, an action generator, and a value evaluator;

[0047] Step 54: Define the reward function for the reinforcement learning model to complete the construction of the reinforcement learning model;

[0048] Step 55: Use the deep Q-learning algorithm to train the reinforcement learning model through experience replay and target network technology;

[0049] Step 56: Based on the trained reinforcement learning model, generate a spraying area and thickness distribution plan for the energy-absorbing material spraying;

[0050] The method of determining the support resistance adjustment scheme of the hydraulic support using the gradient descent algorithm based on the energy release rate distribution in the strong rock pressure characteristic vector is as follows:

[0051] Step 61: Convert the energy release rate distribution into a support resistance optimization objective function, and establish a mapping relationship between the energy release rate and the support resistance;

[0052] Step 62: Based on the latest collected energy release rate distribution, initialize the support resistance adjustment amount, set the learning rate and iteration termination condition of the gradient descent algorithm;

[0053] Step 63: Execute the iterative optimization process of the gradient descent algorithm, calculate the gradient of the support resistance optimization objective function with respect to the support resistance of each support, and update the support resistance adjustment amount.

[0054] A dynamic early warning and collaborative control system for high-pressure shallow coal seams based on multi-source data fusion is proposed, which includes a high-pressure vector collection module, a risk index generation module, a risk assessment module, a red early warning module, an orange early warning module, and a yellow early warning module. Each module is electrically connected.

[0055] The strong mine pressure vector collection module collects multi-source perception data of the coal seam, constructs a strong mine pressure feature vector for the multi-source perception data of the coal seam through wavelet transform and nonlinear dimensionality reduction algorithm, and sends the strong mine pressure feature vector to the risk index generation module, red warning module, orange warning module and yellow warning module;

[0056] A risk index generation module inputs the high-pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network, outputs an energy release index and a risk probability value, and sends the energy release index and risk probability value to a risk assessment module;

[0057] The risk assessment module determines the risk level of strong mine pressure in the coal seam according to the energy release index and the risk probability value. When the energy release index exceeds the first threshold and the risk probability value exceeds the second threshold, the module switches to a red warning module. When the energy release index exceeds the third threshold and the risk probability value exceeds the fourth threshold, the module triggers an orange warning signal and switches to an orange warning module. When the energy release index exceeds the fifth threshold and the risk probability value exceeds the sixth threshold, the module triggers a yellow warning signal and switches to a yellow warning module.

[0058] a red warning module, which, when a red warning signal is triggered, uses an adaptive grid division algorithm to determine a fracturing position and fracturing parameters for directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector;

[0059] an orange warning module, which, when an orange warning signal is triggered, uses a reinforcement learning model to determine a spraying area and thickness distribution of an energy-absorbing material based on a crack density distribution in the strong rock pressure feature vector;

[0060] The yellow warning module, when the yellow warning signal is triggered, uses a gradient descent algorithm to determine a support resistance adjustment plan for the hydraulic support based on the energy release rate distribution in the strong mine pressure characteristic vector.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] The present invention collects multi-source raw data of coal seams through multi-source data collection equipment, and then converts the multi-dimensional raw data into strong mine pressure feature vectors containing stress gradient distribution, energy release rate distribution and crack density distribution through wavelet transform and nonlinear dimensionality reduction algorithm. Then, a hybrid model is constructed by combining deep recurrent neural network and graph neural network to realize the dual extraction of time series characteristics and spatial correlation characteristics, while capturing the time evolution law and spatial distribution characteristics before the occurrence of strong mine pressure events. Finally, a multi-level early warning mechanism based on energy release index and risk probability value is established, and the risk level is finely divided into three levels: red, orange and yellow. Corresponding control strategies are designed for different levels, realizing risk hierarchical management and control, avoiding the "one-size-fits-all" extensive management method of traditional early warning systems, and improving the accuracy and practicality of early warning. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of a method and system for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion in Example 1 of the present invention;

[0064] Figure 2 This is a module connection diagram of the shallow coal seam strong mine pressure dynamic warning and collaborative control system based on multi-source data fusion in Example 2 of the present invention. DETAILED DESCRIPTION

[0065] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0066] Example 1

[0067] like Figure 1 As shown in FIG, a method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion includes the following steps:

[0068] Step 1: Collect multi-source sensing data of the coal seam, and construct a strong mine pressure feature vector for the multi-source sensing data of the coal seam through wavelet transform and nonlinear dimensionality reduction algorithm;

[0069] The coal seam multi-source sensing data includes stress and strain data collected by distributed optical fiber sensors, energy release data collected by microseismic arrays, surface subsidence data collected by surface subsidence radars, and three-dimensional geological modeling data;

[0070] The strong rock pressure characteristic vector includes three key indicators: stress gradient distribution, energy release rate distribution and crack density distribution, which are stored in the form of a multi-dimensional array;

[0071] Step 2: Input the high-pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network, and output an energy release index and a risk probability value;

[0072] The hybrid model captures the changing patterns of temporal features through a deep recurrent neural network and extracts spatial correlation features using a graph neural network.

[0073] The energy release index represents the ratio of energy accumulation to release within the coal seam, and the risk probability value represents the possibility of a strong mine pressure event. Both are scalar data.

[0074] Step 3: Determine the risk level of strong mine pressure in the coal seam based on the energy release index and the risk probability value. When the energy release index exceeds the first threshold and the risk probability value exceeds the second threshold, a red warning signal is triggered and the process proceeds to step 4. When the energy release index exceeds the third threshold and the risk probability value exceeds the fourth threshold, an orange warning signal is triggered and the process proceeds to step 5. When the energy release index exceeds the fifth threshold and the risk probability value exceeds the sixth threshold, a yellow warning signal is triggered and the process proceeds to step 6. Otherwise, return to step 1 and continue monitoring.

[0075] Step 4: When a red warning signal is triggered, an adaptive grid division algorithm is used to determine the fracturing position and fracturing parameters of directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector;

[0076] The fracturing position and fracturing parameters include fracturing point coordinates, fracturing aperture, fracturing spacing and fracturing depth;

[0077] Step 5: When an orange warning signal is triggered, a reinforcement learning model is used to determine the spraying area and thickness distribution of the energy-absorbing material based on the crack density distribution in the strong rock pressure feature vector;

[0078] The spraying area and thickness distribution include the spraying area boundary coordinates and the spraying thickness at each point;

[0079] Step 6: When a yellow warning signal is triggered, a gradient descent algorithm is used to determine a support resistance adjustment scheme for the hydraulic support based on the energy release rate distribution in the strong rock pressure characteristic vector;

[0080] The support resistance adjustment scheme includes the resistance adjustment amount of each support position;

[0081] In an embodiment of the present invention, the collecting of coal seam multi-source sensing data includes the following steps:

[0082] The stress and strain data are collected in real time based on the Brillouin scattering principle through an optical fiber network arranged around the coal seam.

[0083] Specifically, the optical fiber network consists of a laser emitting unit, an optical fiber sensing network and a signal analysis unit. The laser emitting unit generates narrow-linewidth laser pulses of a specific wavelength with appropriate pulse width, repetition frequency and peak power. These narrow-linewidth laser pulses are injected into the single-mode optical fiber buried around the coal seam through a ring optical fiber coupler. When the optical fiber is subjected to external stress or strain, the phonons in the optical fiber interact with the incident photons to generate Brillouin scattered light. The frequency shift of the Brillouin scattered light is linearly related to the strain on the optical fiber and has a specific frequency shift coefficient. The signal analysis unit uses coherent detection technology to receive and analyze the scattered light signal, and extracts the spectral characteristics through fast Fourier transform to achieve accurate measurement of the strain and stress at each spatial point of the optical fiber.

[0084] In this embodiment, the optical fiber network is arranged in a grid pattern across the roof, floor, and sides of the coal seam, forming a high-density three-dimensional monitoring network. The collected stresses and strains at each spatial point are stored in a time series format, containing three dimensions: timestamp, spatial coordinates, and strain value, forming a stress-strain distribution tensor as stress-strain data.

[0085] A network of seismic detectors placed around the coal seam collects data on energy release from tiny fractures within the coal seam.

[0086] Specifically, the seismic detector network consists of a seismic detector array, a signal acquisition unit, and a data processing unit. The seismic detector array uses three-component accelerometers with appropriate sensitivity, frequency response range, and dynamic range. Seismic detectors are deployed at multiple energy monitoring points around the coal seam according to pre-set detection locations and transmit the monitored seismic signals to the signal acquisition unit. The signal acquisition unit has an appropriate sampling rate and resolution, and features real-time digital filtering and signal enhancement. It automatically detects microseismic events from seismic signals using a waveform recognition algorithm, extracts the arrival times, amplitudes, and spectral characteristics of P and S waves, and transmits these arrival times, amplitudes, and spectral characteristics to the data processing unit. The data processing unit uses an improved double-difference location algorithm to determine the microseismic source location with a certain level of positioning accuracy. Simultaneously, the data processing unit calculates the moment tensor and energy release value for each microseismic event, achieving a certain level of energy calculation accuracy. The microseismic monitoring system regularly processes the collected data and outputs parameters such as the spatial coordinates, occurrence time, energy level, and focal mechanism of the microseismic event, forming an energy release distribution tensor as energy release data.

[0087] Synthetic aperture radar interferometry technology is used to obtain surface subsidence data and settlement data caused by coal mining through satellite or ground radar systems.

[0088] Specifically, a synthetic aperture radar with a specific band and appropriate central frequency, bandwidth and spatial resolution is used to regularly acquire radar images of the coal seam, and the surface deformation at adjacent times is calculated through differential interferometry processing technology.

[0089] Specifically, the landmark deformation processing process includes conventional steps such as image registration, interferogram generation, phase unwrapping, atmospheric phase correction, and terrain effect elimination. The surface deformation monitoring range covers the area above and around the coal mining face, forming a high-density surface settlement monitoring network. Ultimately, the three-dimensional coordinates and cumulative settlement of each surface monitoring point are output, and derived parameters such as settlement rate, settlement gradient, and settlement curvature are calculated to form a surface settlement distribution tensor as the surface.

[0090] The three-dimensional geological modeling data is obtained by constructing a three-dimensional geological structure model of the coal seam and its surrounding rock through drilling, geophysical exploration and historical mining data combined with geostatistical methods.

[0091] Specifically, the three-dimensional geological modeling process first collects drilling data, geophysical profiles and historical mining data to establish a unified spatial coordinate system and data standards. The drilling data includes information such as lithology, coal seam thickness, roof and floor elevations, and physical and mechanical parameters; the geophysical profiles include three-dimensional seismic exploration and electrical exploration results; and the historical mining data include the distribution of goafs and the impact range of mining with historical records. The Kriging interpolation method and multi-point geostatistics method are used to reconstruct the coal seam morphology and surrounding rock distribution in three-dimensional space. The three-dimensional geological structure model contains multiple attribute fields such as coal seam thickness, inclination, fault distribution, surrounding rock lithology, and physical and mechanical parameters to form a three-dimensional geological structure tensor that composes the three-dimensional geological modeling data.

[0092] The stress and strain data, energy release data, surface settlement data and three-dimensional geological modeling data are matched with timestamps and spatial position coordinates of the coal seam to form the coal seam multi-source sensing data.

[0093] Furthermore, the method of constructing a strong mine pressure feature vector from the coal seam multi-source sensing data by wavelet transform and nonlinear dimensionality reduction algorithm includes the following steps:

[0094] Step 11: Perform wavelet transform preprocessing on the coal seam multi-source sensing data to extract the multi-scale feature matrix;

[0095] Specifically, the wavelet transform preprocessing uses a discrete wavelet transform method to perform multi-scale decomposition on stress-strain data, energy release data, surface settlement data, and 3D geological modeling data. Specifically, for each type of data, a wavelet basis function suitable for its characteristics is selected. For example, in this embodiment, the stress-strain data uses the db4 wavelet basis, the energy release data uses the sym8 wavelet basis, the surface settlement data uses the coif3 wavelet basis, and the 3D geological modeling data uses the bior3.5 wavelet basis.

[0096] During the wavelet decomposition process, each type of data is decomposed into five layers to generate multi-scale wavelet coefficients. For example, in this embodiment, for stress-strain data, the approximate and detailed coefficients of layers 3 to 5 are extracted; for energy release data, the approximate and detailed coefficients of layers 2 to 4 are extracted; for surface settlement data, the approximate and detailed coefficients of layers 3 to 5 are extracted; and for 3D geological modeling data, the approximate and detailed coefficients of layers 2 to 4 are extracted.

[0097] After wavelet transformation, the statistical characteristics of the wavelet coefficients of each layer are calculated, including five statistical quantities such as mean, standard deviation, energy, entropy and kurtosis, to form a preliminary feature matrix. For stress and strain data, a 60-dimensional feature vector is generated; for energy release data, a 45-dimensional feature vector is generated; for surface settlement data, a 60-dimensional feature vector is generated; for 3D geological modeling data, a 45-dimensional feature vector is generated. The feature vectors generated after wavelet transformation of various types of data constitute a multi-scale feature matrix.

[0098] Step 12: Perform nonlinear dimensionality reduction on the multi-scale feature matrix after wavelet transformation to generate a low-dimensional feature space, reduce the data dimension and retain key information;

[0099] Specifically, the nonlinear dimensionality reduction process adopts a hybrid dimensionality reduction method that combines the t-distributed stochastic neighbor embedding algorithm (t-SNE) with an autoencoder.

[0100] In the autoencoder stage, a five-layer neural network structure was constructed, consisting of an input layer, three hidden layers, and an output layer. The input layer had 210 nodes, matching the dimensions of the multi-scale feature matrix. The first hidden layer had 128 nodes and used the ReLU activation function. The second hidden layer (encoding layer) had 64 nodes and used the tanh activation function. The third hidden layer had 128 nodes and used the ReLU activation function. The output layer had 210 nodes, matching the input layer. The autoencoder was trained by minimizing the reconstruction error, extracting a 64-dimensional feature representation of the encoding layer.

[0101] In the t-SNE stage, the output of the autoencoder's encoding layer is used as input, the perplexity parameter is set to 30, the learning rate is set to 200, and the number of iterations is set to 1000, further reducing the 64-dimensional features to a low-dimensional feature space of 27 dimensions. It can be understood that the t-SNE algorithm effectively captures the nonlinear structure of the data by maintaining local similarities between data points, making it particularly suitable for processing the complex nonlinear relationships in multi-source coal seam perception data.

[0102] Step 13: Based on the low-dimensional feature space after dimensionality reduction, three key indicators are constructed: stress gradient distribution, energy release rate distribution, and crack density distribution;

[0103] Specifically, in the 27-dimensional low-dimensional feature space, three types of key indicators are constructed through feature mapping and reconstruction algorithms.

[0104] For the stress gradient distribution, the first nine dimensions of the low-dimensional feature space are selected and combined with the wavelet coefficients of the stress and strain data to generate a three-dimensional spatial grid representation through a tensor reconstruction algorithm. Each grid point contains information about the magnitude and direction of the stress gradient, forming a four-dimensional tensor of 9 × M × N × P, where M, N, and P represent the number of grid cells in the x, y, and z directions of space, respectively.

[0105] For the energy release rate distribution, we select the middle nine dimensions of the low-dimensional feature space and, combined with the wavelet coefficients of the energy release data, use a density estimation algorithm, such as kernel density estimation, to generate the spatial distribution of the energy release rate. The energy release rate represents the amount of energy released per unit volume of coal per unit time, forming a four-dimensional tensor of 9 × M × N × P, which describes the spatiotemporal distribution of energy release.

[0106] For the distribution of crack density, the last 9 dimensions in the low-dimensional feature space are selected, and the wavelet coefficients of the surface settlement data and the three-dimensional geological modeling data are combined to generate the spatial distribution of crack density through the crack identification algorithm. The crack density represents the number and scale of cracks in a unit volume, forming a four-dimensional tensor of 9×M×N×P, which describes the degree of damage to the internal structure of the coal seam. The execution process of the crack identification algorithm is as follows: using the Canny or Sobel operator to perform edge detection on the wavelet coefficients after wavelet transformation to identify possible crack boundaries; segmenting the features of the crack boundary based on the local adaptive threshold to extract potential cracks; applying morphological operations to remove noise and connect broken crack lines; reconstructing the two-dimensional crack information of the crack line into a three-dimensional crack network through interpolation and topological relationships;

[0107] Step 14: combining the three key indicators into a strong rock pressure feature vector;

[0108] Specifically, the strong mine pressure characteristic vector adopts a multi-dimensional array structure, which includes four dimensions: indicator type, characteristic component, spatial position and time series.

[0109] In the indicator type dimension, there are three types, corresponding to stress gradient distribution, energy release rate distribution and crack density distribution respectively; in the characteristic component dimension, each indicator contains 9 components, describing characteristic information in different aspects; in the spatial position dimension, a three-dimensional grid of M×N×P is used to represent the coal seam space; in the time series dimension, it contains historical data of the last T time points.

[0110] The resulting strong coal pressure feature vector is a six-dimensional tensor of 3 × 9 × M × N × P × T. M, N, and P are the number of grid cells in the three spatial directions, typically set to M = 20, N = 20, and P = 10, covering the entire monitored coal seam area. T is the length of the time series, typically set to 24, corresponding to the last 24 hours of monitoring data. Each element in the strong coal pressure feature vector is a normalized dimensionless value, ranging from [-1 to 1].

[0111] Furthermore, inputting the strong mine pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network to output an energy release index and a risk probability value includes the following steps:

[0112] Step 21: Construct a deep recurrent neural network to process the time series data of the strong mine pressure feature vector, generate a time series feature vector, and capture the change pattern in the time dimension;

[0113] Specifically, the deep recurrent neural network uses a bidirectional long short-term memory (Bi-LSTM) architecture, consisting of three stacked layers of LSTM units. Each LSTM unit comprises four key components: a forget gate, an input gate, an output gate, and a memory unit. The forget gate controls the degree of retention of historical information, the input gate controls the degree of acceptance of current input information, the output gate controls the degree of information output, and the memory unit stores long-term dependency information.

[0114] In the deep recurrent neural network (DRN) architecture, the number of LSTM units in the first layer is set to 128, the second layer to 64, and the third layer to 32, extracting higher-level temporal features layer by layer. A dropout layer with a dropout rate of 0.2 is added between each LSTM unit layer to prevent overfitting. The DRN input is a chronological sequence of high-pressure feature vectors, for example, data on stress gradient distribution, energy release rate distribution, and fracture density distribution for the past 30 time points.

[0115] Step 22: Construct a graph neural network to process the spatial data of the strong mine pressure feature vector, generate a spatial feature vector, and extract the correlation features in the spatial dimension;

[0116] Specifically, the graph neural network uses a graph convolutional network structure to capture the correlation between the spatial locations of coal seams. First, an undirected weighted graph is constructed based on the spatial locations of the coal seams, where nodes represent monitoring points and edges represent the spatial relationships between nodes. Edge weights are calculated based on the Euclidean distance between nodes.

[0117] The graph convolutional network contains three layers of graph convolutional layers. The number of output channels of the first layer is 64, the second layer is 32, and the third layer is 16. It extracts higher-order spatial correlation features layer by layer and finally outputs a spatial feature vector.

[0118] Step 23: Use the attention mechanism to fuse the output features of the deep recurrent neural network and the graph neural network to generate a fused feature vector;

[0119] Specifically, the temporal feature vectors output by the deep recurrent neural network and the spatial feature vectors output by the graph neural network are first concatenated to form a preliminary fused feature. The fused feature is then weighted using a multi-head self-attention mechanism, with 8 attention heads and 16 dimensions per attention head.

[0120] The output of the multi-head attention is processed by residual connection and layer normalization to obtain the final fused feature vector.

[0121] Step 24: Construct a prediction unit to generate an energy release index and a risk probability value based on the fusion features;

[0122] Specifically, the prediction unit consists of two parallel fully connected sub-networks, which are used to predict the energy release index and the risk probability value respectively.

[0123] In this embodiment, the energy release index prediction subnetwork consists of two fully connected layers, with 64 hidden layer neurons and 1 output layer neuron. It uses a linear activation function. This subnetwork outputs a scalar value representing the ratio of energy accumulation to energy release within the coal seam.

[0124] The risk probability prediction subnetwork also consists of two fully connected layers, with 64 hidden layer neurons and 1 output layer neuron. It uses a sigmoid activation function to constrain the output value between 0 and 1, representing the probability of a severe mine pressure event. Both subnetworks share the fused feature vector as input but have independent weight parameters to adapt to the characteristics of their respective tasks.

[0125] Step 25: Use the historically collected coal seam multi-source sensing data and the records of severe mine pressure events to train the hybrid model and output the energy release index and risk probability value;

[0126] Specifically, the hybrid model is trained end-to-end, using historically collected multi-source coal seam sensing data and records of severe mine pressure events as the training set. The hybrid model's loss function consists of two components: a mean square error loss for the energy release index and a binary cross entropy loss for the risk probability value. The total loss function is a weighted sum of the two, with an adaptive 1:1 weight ratio.

[0127] The hybrid model was optimized using the Adam optimizer, with an initial learning rate of 0.001 and a cosine annealing strategy to dynamically adjust the learning rate. Training was performed using an early stopping mechanism.

[0128] It should be further noted that, in this embodiment, the first threshold is the high-risk warning line of the energy release index, which is usually set to 0.85, indicating that the energy accumulation inside the coal seam has reached an extremely high level, the release ratio is low, and there is a high risk of sudden large-scale energy release;

[0129] The first threshold is a high-risk warning line for the risk probability value, usually set at 0.75. Exceeding this threshold indicates that the possibility of a severe mine pressure event is extremely high, and the highest level of prevention and control measures must be taken immediately; thus, the first and second thresholds can be used to trigger the conditions for a red alert;

[0130] Similarly, the third threshold and the fourth threshold are used to represent the medium-risk warning line of the energy release index and the medium-risk warning line of the risk probability value, respectively, and are used to trigger the next level orange warning signal;

[0131] The fifth and sixth thresholds are used to represent the low-risk warning line of the energy release index and the low-risk warning line of the risk probability value, respectively, thereby triggering a yellow warning signal of the next level;

[0132] Understandably, for high-risk warnings, the most forceful intervention is required, using directional hydraulic fracturing to directly release the energy accumulated within the coal seam;

[0133] For medium-risk warnings, passive protection measures are needed to enhance the coal seam's ability to resist impact through energy-absorbing materials;

[0134] For low-risk warnings, passive protection measures need to be adopted, auxiliary support measures are adopted, and the support capacity is improved by adjusting the resistance of the hydraulic support;

[0135] Therefore, the method of determining the fracturing position and fracturing parameters of directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector by using an adaptive grid division algorithm to form a directional hydraulic fracturing control instruction set includes the following steps:

[0136] Step 41: Based on the stress gradient distribution in the strong mine pressure characteristic vector, a three-dimensional stress field model of the coal seam is constructed, and stress concentration areas are identified;

[0137] Specifically, the stress gradient distribution data is first converted into a three-dimensional tensor representation, where each tensor element contains the spatial position coordinates (x, y, z) and the corresponding stress gradient vector The modulus of the stress gradient vector represents the rate of stress change, and the direction represents the direction of the fastest stress growth.

[0138] In the process of constructing the three-dimensional stress field model, the trilinear interpolation algorithm is used to perform spatial interpolation on the discrete stress gradient data to generate a continuous stress gradient field. In this stress gradient field, for any point P(x, y, z) in space, its stress gradient value is obtained by weighted average calculation of the eight nearest sampling points around it, and the weight coefficient is inversely proportional to the distance.

[0139] Stress concentration areas are identified using a local extremum detection algorithm. By calculating the spatial second-order derivative of the stress gradient, the algorithm identifies regions where the stress gradient amplitude exceeds a preset threshold. In a specific implementation of the present invention, a three-dimensional Laplace operator is applied to the stress gradient field. When the Laplace operator value exceeds a preset threshold, the region is marked as a stress concentration area. Stress concentration areas are represented as three-dimensional connected domains, each of which contains a set of spatial coordinates and a corresponding stress gradient value.

[0140] Step 42: Based on the changing characteristics of the stress gradient, an adaptive meshing algorithm is applied to the identified stress concentration area to determine the mesh density distribution;

[0141] Specifically, the adaptive meshing algorithm is implemented based on an octree structure. First, the entire coal seam is divided into a uniform initial grid of 1m x 1m x 1m. This grid is then recursively subdivided based on the rate of change of the stress gradient. Regions with greater rates of change in the stress gradient are given finer meshes.

[0142] The criterion for mesh subdivision is to calculate the difference between the maximum and minimum stress gradients within the current mesh cell. If this difference exceeds a preset threshold, τ1, the mesh cell is divided into eight sub-meshes. This process is repeated recursively until one of the following termination conditions is met: the stress gradient change within the mesh cell is less than a threshold, τ2; or the mesh size reaches a preset minimum value (10 cm × 10 cm × 10 cm); or the recursive depth reaches a preset maximum value (6 layers).

[0143] Therefore, the adaptive grid finally generated has the following characteristics: the grid density is high in the area where the stress gradient changes sharply, and the grid density is low in the area where the stress gradient changes gently, thereby improving the calculation efficiency while ensuring the calculation accuracy.

[0144] Step 43: Based on the grid density distribution of the adaptive grid division, the fracturing position is determined using the energy release maximization criterion;

[0145] Specifically, an energy release potential assessment model was constructed based on the adaptive grid division. For each grid cell, an energy release potential index was calculated. This energy release potential comprehensively considers the following factors: stress gradient value, the angle between the main stress direction and the coal seam strike, and the distance from existing fractures.

[0146] Therefore, the calculation formula of the energy release potential index is:

[0147]

[0148] in, is the stress gradient value; θ is the angle between the main stress direction and the coal seam trend; d is the distance to the nearest fracture; w1, w2, and w3 are weight coefficients, which are determined by optimizing historical data; and d0 is the characteristic distance parameter.

[0149] Based on the calculated energy release potential index, a non-dominated sorting genetic algorithm (NSGA-II) was used for multi-objective optimization. The objective functions included maximizing the total energy release, minimizing the number of fracturing points, and maximizing the spatial uniformity between fracturing points. The fracturing locations were then determined by solving for a Pareto-optimal set of solutions. These locations consisted of three-dimensional (x, y, z) coordinates and corresponding energy release potential indices.

[0150] Step 44: Calculate the fracturing parameters based on the determined fracturing location, combined with the stress field characteristics and the pre-collected surrounding rock parameters;

[0151] Specifically, for each determined fracturing location, the optimal fracturing parameters are calculated based on elastic mechanics theory and fracture propagation models. Fracturing parameters include fracturing aperture, fracturing spacing, fracturing depth, and fracturing pressure.

[0152] Specifically, the calculation of the fracturing aperture is based on the modified Hubbert formula, taking into account the stress field distribution and rock properties:

[0153]

[0154] Where D is the fracturing aperture; E is the elastic modulus of rock; σt is the tensile strength of rock; is the stress gradient value; α and β are empirical coefficients, which are calibrated by field data.

[0155] The calculation of fracture spacing is based on stress interference theory, which ensures that the stress fields between adjacent fracture points enhance rather than cancel each other out:

[0156]

[0157] Where L is the fracturing spacing; h is the coal seam thickness; σh is the minimum horizontal principal stress; γ and δ are empirical coefficients.

[0158] The calculation of fracturing depth takes into account the maximum energy release and the stability of the surrounding rock:

[0159]

[0160] Among them, H is the fracturing depth; h is the coal seam thickness; K is the stress intensity factor; E is the rock elastic modulus; σv is the vertical component of stress; and ε is the preset safety factor.

[0161] The calculation of fracturing pressure is based on the modified fracturing mechanics model:

[0162]

[0163] Wherein, P is the fracturing pressure and λ is the correction coefficient.

[0164] It can be understood that the rock elastic modulus and rock tensile strength are parameters of the surrounding rock parameters; after obtaining the fracturing position and fracturing parameters, the corresponding fracturing parameters can be implemented at the fracturing position to complete the release of the energy accumulated in the coal seam;

[0165] Furthermore, the method of using a reinforcement learning model to determine the spraying area and thickness distribution of the energy absorbing material spraying includes the following steps:

[0166] Step 51: converting the crack density distribution data in the strong rock pressure characteristic vector into a three-dimensional space grid representation to construct an environmental state space;

[0167] Specifically, the fracture density distribution data was first discretized into a regular three-dimensional grid structure. The grid cell size was set to 0.5m × 0.5m × 0.5m, ensuring spatial resolution while controlling computational complexity. Each grid cell contains three key attributes: fracture density value, fracture direction distribution, and fracture connectivity. The fracture density value represents the number or total length of fractures per unit volume; the fracture direction distribution represents the main direction of the fractures in the form of direction cosines; and fracture connectivity describes the degree of connection between adjacent fractures.

[0168] In this embodiment of the present invention, the environmental state space is represented by a multi-channel three-dimensional tensor with the dimensions [D, H, W, C], where D, H, and W correspond to the depth, height, and width of the coal seam space, respectively, and C is the number of characteristic channels, including five channels: fracture density, principal direction cosines, and connectivity. Each element in the environmental state space represents the fracture characteristics of the corresponding spatial location and serves as input to the reinforcement learning model.

[0169] Step 52: Define the reinforcement learning model as an action space in a two-dimensional discrete-continuous hybrid space, including two dimensions: spray area selection and thickness distribution;

[0170] Specifically, in this embodiment, the action space is designed as a two-dimensional discrete-continuous hybrid space.

[0171] The first dimension is a discrete action, representing the selection of a spraying area. The coal seam space is divided into several candidate spraying areas, each corresponding to a discrete action. In this example, based on the coal seam size and spraying equipment characteristics, the coal seam space is divided into 20 candidate spraying areas, each of which can be set to 5m×5m in size.

[0172] The second dimension is the continuous action, representing the spray thickness distribution strategy. The spray thickness range is set to 5mm to 30mm and is dynamically adjusted based on changes in crack density. The continuous action is represented by a parameterized function, using a thickness distribution model based on radial basis functions to express the spatial variation pattern. Specifically, the thickness distribution function is defined as the weighted sum of multiple radial basis functions, each of which controls the thickness variation in a local area, with the weight coefficients learned by the reinforcement learning model.

[0173] Step 53: Based on the environment state space and action space, construct a reinforcement learning model based on a deep Q network, wherein the reinforcement learning model includes three core components: a state encoder, an action generator, and a value evaluator;

[0174] Specifically, the deep Q-network employs a two-stream network architecture, handling discrete action selection and continuous thickness allocation, respectively. The state encoder, taking the environment state as input, consists of a three-dimensional convolutional network consisting of four consecutive three-dimensional convolutional blocks. Each convolutional block contains a 3×3×3 three-dimensional convolutional layer, a batch normalization layer, and a LeakyReLU activation function. The first convolutional block has 32 output channels, and the number of channels doubles with each subsequent block. The final layer uses global average pooling as the state representation.

[0175] The action generator consists of two parts: a region selection network and a thickness allocation network. The region selection network uses a fully connected structure, consisting of two hidden layers, each with 128 neurons. The output layer has an output layer with the same number of neurons as the candidate regions, representing the Q value of each region to select the spraying area. The thickness allocation network also uses a fully connected structure, but the output layer contains the action selected from the action space, which controls the spraying thickness.

[0176] The value estimator integrates state representations and action information to predict the long-term reward of the current state-action pair. It consists of a three-layer fully connected network. Its input is a concatenation of a vector of the environment state and the action representation. Its output is a scalar Q-value, which represents the expected effect of the current decision.

[0177] Step 54: Define the reward function for the reinforcement learning model to complete the construction of the reinforcement learning model;

[0178] In an embodiment of the present invention, the reward function weighs three aspects: energy absorption effect, material consumption, and construction difficulty;

[0179] Specifically, the reward function is designed as a multi-objective weighted form, which includes the following three components:

[0180] Energy Absorption Performance Award R1 evaluates the coverage and energy absorption capacity of the spray solution in the crack area. This award is calculated as a weighted integral of the crack density and the spray thickness. Areas with high crack density and thicker energy absorption materials receive a higher award.

[0181] The material consumption penalty R2 is proportional to the total amount of spray material used, encouraging the algorithm to conserve material while ensuring energy absorption. It is calculated as the spatial integral of the spray thickness multiplied by the unit cost and serves as a negative reward.

[0182] The construction difficulty penalty R3 evaluates the feasibility of the spraying plan, including two indicators: thickness gradient and area continuity. Excessive thickness gradient increases construction difficulty, and discontinuous spraying areas increase equipment movement costs.

[0183] The comprehensive reward function is expressed as a weighted sum of the energy absorption effect reward, material consumption penalty, and construction difficulty penalty. The weights of each factor are dynamically adjusted based on the actual project requirements. It should be noted that the above reward function is only an example for illustrative purposes. In actual implementation, various rewards or penalties can be added, deleted, or modified based on actual needs. This is not a limitation here, but any modification to the reward function falls within the scope of protection of this invention.

[0184] Step 55: Use the deep Q-learning algorithm to train the reinforcement learning model through experience replay and target network technology;

[0185] Specifically, the training process of the deep Q learning algorithm includes the following steps:

[0186] First, initialize the experience replay buffer. Generally, the buffer capacity is set to 10,000 transition samples. Each transition sample contains five elements: the current state, the action performed, the reward obtained, the next state, and the termination flag.

[0187] Secondly, initialize the parameters of the main network and the target network. The main network is used for action selection and value evaluation, and the target network is used to calculate the target Q value to improve training stability. The target network parameters are initialized to be the same as those of the main network, but the update frequency is lower, and it is updated every 100 training steps. It should be noted that in the target network technology, the structures of the main network and the target network are exactly the same, and both contain the three core components of the state encoder, action generator, and value evaluator in the above-mentioned reinforcement learning model. The difference is that the main network is used for real-time decision-making, that is, selecting the optimal action under the current state, while the target network is used to calculate the target Q value. The reason for this setting is that if only one network is used to generate actions and calculate target values, it will lead to the problem of "chasing moving targets", making it difficult for training to converge;

[0188] During the training of the reinforcement learning model, an ε-greedy strategy is used to balance exploration and exploitation. The initial ε value is set to 0.9, indicating a 90% probability of randomly selecting an action for exploration and a 10% probability of selecting the action with the highest Q value. The ε value decays linearly with training, eventually dropping to 0.1, increasing the exploitation of high-value actions.

[0189] For each training step, batches of data are randomly sampled from the buffer, with a batch size of 64. The target Q-value is calculated: for non-terminal states, the target Q-value is the immediate reward plus a discount factor multiplied by the maximum Q-value predicted by the target network for the next state; for terminal states, the target Q-value is simply the immediate reward. The discount factor is set to 0.95 to balance short-term and long-term rewards.

[0190] The mean squared error loss function is used to calculate the difference between the Q values ​​predicted by the primary network and the Q values ​​output by the target network. The primary network parameters are updated using the Adam optimizer with a learning rate of 0.001. Training continues until the loss function converges or the preset maximum number of training steps is reached.

[0191] Step 56: Based on the trained reinforcement learning model, generate a spraying area and thickness distribution plan for the energy-absorbing material spraying;

[0192] Specifically, the latest coal seam fracture density distribution is first converted into a vector representation of the environmental state and input into a trained reinforcement learning model. The reinforcement learning model's region selection network outputs a Q value for each candidate region, and the candidate with the highest Q value is selected as the spraying area. Simultaneously, the thickness distribution network generates the parameters of a parameterized thickness distribution function to determine the thickness value at each point within the spraying area.

[0193] It can be understood that after obtaining the spraying area and thickness value, the corresponding spraying thickness can be implemented in the spraying area to enhance the ability of the coal seam to resist impact through the energy absorbing material;

[0194] Furthermore, the method of determining the support resistance adjustment scheme of the hydraulic support using the gradient descent algorithm based on the energy release rate distribution in the strong rock pressure characteristic vector is as follows:

[0195] Step 61: Convert the energy release rate distribution into a support resistance optimization objective function, and establish a mapping relationship between the energy release rate and the support resistance;

[0196] Specifically, the support resistance optimization objective function is defined as a weighted combination of the spatial uniformity of the energy release rate distribution and the overall energy release level. The spatial uniformity of the energy release rate distribution represents the spatial variance of the energy release rate, reflecting the uniformity of energy release; the overall energy release level represents the difference between the overall energy release level and the safety threshold. The goal of support resistance adjustment is to make the energy release rate distribution more uniform by adjusting the support resistance at each support position, while keeping the overall energy release level within a safe range.

[0197] To establish the mapping relationship, the support resistance influence matrix is ​​first constructed. This matrix describes the degree to which changes in resistance at each support location affect the energy release rate in the surrounding area. This matrix can be calculated based on historical monitoring data and a mechanical model, taking into account inter-support interactions and spatial attenuation characteristics. This matrix can also be used to predict the effect of any support resistance adjustment on the energy release rate distribution.

[0198] Step 62: Based on the latest collected energy release rate distribution, initialize the support resistance adjustment amount, set the learning rate and iteration termination condition of the gradient descent algorithm;

[0199] Specifically, the support resistance adjustment amount is initialized to a zero vector, indicating that the initial adjustment amount of each support position is zero. The learning rate setting adopts an adaptive strategy, with the initial value set to 0.05, and is dynamically adjusted as the rate of descent of the objective function changes during the iteration process. When the rate of descent of the objective function is large, the learning rate is increased to accelerate convergence; when the rate of descent of the objective function decreases, the learning rate is reduced to avoid oscillation. The iterative termination condition is set as follows: the improvement of the objective function for three consecutive iterations is less than the preset threshold (usually 0.001), or the maximum number of iterations (usually 100 times) is reached.

[0200] Step 63: Execute the iterative optimization process of the gradient descent algorithm, calculate the gradient of the support resistance optimization objective function with respect to the support resistance of each support, and update the support resistance adjustment amount;

[0201] Specifically, in each iteration, the adjusted energy release rate distribution is first predicted based on the current support resistance adjustment using the mapping relationship established in step 61. The function value of the support resistance optimization objective function is then calculated, and the gradient of the support resistance optimization objective function with respect to each support resistance is solved. The gradient calculation uses a numerical differentiation method to perform a small perturbation on the support resistance at each support position and observe the rate of change of the objective function to obtain a gradient vector.

[0202] It should be noted that the gradient vector represents the direction and magnitude of the objective function's change at each support location. Areas with concentrated energy release rates typically have larger gradient values, indicating that support resistance adjustments in these areas have a more significant impact on the optimization objective. Based on the calculated gradient vector, the support resistance adjustment is updated using the gradient descent principle: the new adjustment is equal to the current adjustment minus the product of the learning rate and the gradient.

[0203] It is understandable that after obtaining the support resistance adjustment amount, the support resistance of the hydraulic support can be adjusted, thereby achieving the function of improving the support capacity by adjusting the resistance of the hydraulic support.

[0204] Example 2

[0205] like Figure 2 As shown in the figure, the shallow coal seam strong mine pressure dynamic warning and collaborative control system based on multi-source data fusion includes a strong mine pressure vector collection module, a risk index generation module, a risk assessment module, a red warning module, an orange warning module and a yellow warning module; wherein each module is electrically connected;

[0206] The strong mine pressure vector collection module collects multi-source perception data of the coal seam, constructs a strong mine pressure feature vector for the multi-source perception data of the coal seam through wavelet transform and nonlinear dimensionality reduction algorithm, and sends the strong mine pressure feature vector to the risk index generation module, red warning module, orange warning module and yellow warning module;

[0207] A risk index generation module inputs the high-pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network, outputs an energy release index and a risk probability value, and sends the energy release index and risk probability value to a risk assessment module;

[0208] The risk assessment module determines the risk level of strong mine pressure in the coal seam according to the energy release index and the risk probability value. When the energy release index exceeds the first threshold and the risk probability value exceeds the second threshold, the module switches to a red warning module. When the energy release index exceeds the third threshold and the risk probability value exceeds the fourth threshold, the module triggers an orange warning signal and switches to an orange warning module. When the energy release index exceeds the fifth threshold and the risk probability value exceeds the sixth threshold, the module triggers a yellow warning signal and switches to a yellow warning module.

[0209] a red warning module, which, when a red warning signal is triggered, uses an adaptive grid division algorithm to determine a fracturing position and fracturing parameters for directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector;

[0210] an orange warning module, which, when an orange warning signal is triggered, uses a reinforcement learning model to determine a spraying area and thickness distribution of an energy-absorbing material based on a crack density distribution in the strong rock pressure feature vector;

[0211] The yellow warning module, when the yellow warning signal is triggered, uses a gradient descent algorithm to determine a support resistance adjustment plan for the hydraulic support based on the energy release rate distribution in the strong mine pressure characteristic vector.

[0212] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion, characterized by: The following steps are involved: Step 1: Collect multi-source sensing data of the coal seam, and construct a strong mine pressure feature vector for the multi-source sensing data of the coal seam through wavelet transform and nonlinear dimensionality reduction algorithm; Step 2: Input the high-pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network, and output an energy release index and a risk probability value; Step 3: Determine the risk level of strong mine pressure in the coal seam based on the energy release index and the risk probability value. When the energy release index exceeds the first threshold and the risk probability value exceeds the second threshold, a red warning signal is triggered and the process proceeds to step 4. When the energy release index exceeds the third threshold and the risk probability value exceeds the fourth threshold, an orange warning signal is triggered and the process proceeds to step 5. When the energy release index exceeds the fifth threshold and the risk probability value exceeds the sixth threshold, a yellow warning signal is triggered and the process proceeds to step 6. Otherwise, return to step 1 and continue monitoring. Step 4: When a red warning signal is triggered, an adaptive grid division algorithm is used to determine the fracturing position and fracturing parameters of directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector; Step 5: When an orange warning signal is triggered, a reinforcement learning model is used to determine the spraying area and thickness distribution of the energy-absorbing material based on the crack density distribution in the strong rock pressure feature vector; Step 6: When the yellow warning signal is triggered, a gradient descent algorithm is used to determine a support resistance adjustment scheme for the hydraulic support based on the energy release rate distribution in the strong mine pressure characteristic vector.

2. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 1 is characterized in that: The coal seam multi-source sensing data includes stress and strain data collected by distributed optical fiber sensors, energy release data collected by microseismic arrays, surface subsidence data collected by surface subsidence radars, and three-dimensional geological modeling data.

3. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 1 is characterized in that: The strong rock pressure characteristic vector includes three key indicators: stress gradient distribution, energy release rate distribution and crack density distribution.

4. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 3 is characterized in that: The method of collecting coal seam multi-source sensing data includes the following steps: The stress and strain data are collected in real time based on the Brillouin scattering principle through an optical fiber network arranged around the coal seam. A network of seismic detectors placed around the coal seam collects data on energy release from tiny fractures within the coal seam. Synthetic aperture radar interferometry technology is used to obtain surface subsidence data and settlement data caused by coal mining through satellite or ground radar systems. The three-dimensional geological modeling data is obtained by constructing a three-dimensional geological structure model of the coal seam and its surrounding rock through drilling, geophysical exploration and historical mining data combined with geostatistical methods. The stress-strain data, energy release data, surface settlement data and three-dimensional geological modeling data, after timestamp matching and matching of the spatial position coordinates of the coal seam, together constitute the coal seam multi-source sensing data.

5. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 4 is characterized in that: The method of constructing a strong mine pressure feature vector from the coal seam multi-source sensing data by wavelet transform and nonlinear dimensionality reduction algorithm includes the following steps: Step 11: Perform wavelet transform preprocessing on the coal seam multi-source sensing data to extract the multi-scale feature matrix; Step 12: Perform nonlinear dimensionality reduction on the multi-scale feature matrix after wavelet transformation to generate a low-dimensional feature space, reduce the data dimension and retain key information; Step 13: Based on the low-dimensional feature space after dimensionality reduction, three key indicators are constructed: stress gradient distribution, energy release rate distribution, and crack density distribution; Step 14: Combine the three types of key indicators into a strong mine pressure feature vector.

6. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 5 is characterized in that: Inputting the strong mine pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network to output an energy release index and a risk probability value includes the following steps: Step 21: Construct a deep recurrent neural network to process the time series data of the strong mine pressure feature vector, generate a time series feature vector, and capture the change pattern in the time dimension; Step 22: Construct a graph neural network to process the spatial data of the strong mine pressure feature vector, generate a spatial feature vector, and extract the correlation features in the spatial dimension; Step 23: Use the attention mechanism to fuse the output features of the deep recurrent neural network and the graph neural network to generate a fused feature vector; Step 24: Construct a prediction unit to generate an energy release index and a risk probability value based on the fusion features; Step 25: Use the historically collected coal seam multi-source sensing data and the records of severe mine pressure events to train the hybrid model and output the energy release index and risk probability value.

7. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 6 is characterized in that: The method of determining the fracturing position and fracturing parameters of directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector by using an adaptive grid division algorithm to form a directional hydraulic fracturing control instruction set includes the following steps: Step 41: Based on the stress gradient distribution in the strong mine pressure characteristic vector, a three-dimensional stress field model of the coal seam is constructed, and stress concentration areas are identified; Step 42: Based on the changing characteristics of the stress gradient, an adaptive meshing algorithm is applied to the identified stress concentration area to determine the mesh density distribution; Step 43: Based on the grid density distribution of the adaptive grid division, the fracturing position is determined using the energy release maximization criterion; Step 44: Calculate the fracturing parameters based on the determined fracturing position, combined with the stress field characteristics and the pre-collected surrounding rock parameters.

8. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 7 is characterized in that: The method of using a reinforcement learning model to determine the spraying area and thickness distribution of the energy absorbing material spraying comprises the following steps: Step 51: converting the crack density distribution data in the strong rock pressure characteristic vector into a three-dimensional space grid representation to construct an environmental state space; Step 52: Define the reinforcement learning model as an action space in a two-dimensional discrete-continuous hybrid space, including two dimensions: spray area selection and thickness distribution; Step 53: Based on the environment state space and action space, construct a reinforcement learning model based on a deep Q network, wherein the reinforcement learning model includes three core components: a state encoder, an action generator, and a value evaluator; Step 54: Define the reward function for the reinforcement learning model to complete the construction of the reinforcement learning model; Step 55: Use the deep Q-learning algorithm to train the reinforcement learning model through experience replay and target network technology; Step 56: Based on the trained reinforcement learning model, generate a spraying area and thickness distribution plan for the energy-absorbing material spraying.

9. The method for dynamic early warning and coordinated control of strong mine pressure in shallow coal seams based on multi-source data fusion according to claim 8, characterized in that: The method of determining the support resistance adjustment scheme of the hydraulic support using the gradient descent algorithm based on the energy release rate distribution in the strong rock pressure characteristic vector is as follows: Step 61: Convert the energy release rate distribution into a support resistance optimization objective function, and establish a mapping relationship between the energy release rate and the support resistance; Step 62: Based on the latest collected energy release rate distribution, initialize the support resistance adjustment amount, set the learning rate and iteration termination condition of the gradient descent algorithm; Step 63: Execute the iterative optimization process of the gradient descent algorithm, calculate the gradient of the support resistance optimization objective function with respect to the support resistance of each support, and update the support resistance adjustment amount.

10. A dynamic early warning and collaborative control system for strong mine pressure in shallow coal seams based on multi-source data fusion, which is used to implement the method for dynamic early warning and collaborative control of strong mine pressure in shallow coal seams based on multi-source data fusion as described in any one of claims 1 to 9, characterized in that: It includes a strong mine pressure vector collection module, a risk index generation module, a risk assessment module, a red warning module, an orange warning module, and a yellow warning module; wherein each module is electrically connected; The strong mine pressure vector collection module collects multi-source perception data of the coal seam, constructs a strong mine pressure feature vector for the multi-source perception data of the coal seam through wavelet transform and nonlinear dimensionality reduction algorithm, and sends the strong mine pressure feature vector to the risk index generation module, red warning module, orange warning module and yellow warning module; A risk index generation module inputs the high-pressure feature vector into a hybrid model of a deep recurrent neural network and a graph neural network, outputs an energy release index and a risk probability value, and sends the energy release index and risk probability value to a risk assessment module; The risk assessment module determines the risk level of strong mine pressure in the coal seam according to the energy release index and the risk probability value. When the energy release index exceeds the first threshold and the risk probability value exceeds the second threshold, the module switches to a red warning module. When the energy release index exceeds the third threshold and the risk probability value exceeds the fourth threshold, the module triggers an orange warning signal and switches to an orange warning module. When the energy release index exceeds the fifth threshold and the risk probability value exceeds the sixth threshold, the module triggers a yellow warning signal and switches to a yellow warning module. a red warning module, which, when a red warning signal is triggered, uses an adaptive grid division algorithm to determine a fracturing position and fracturing parameters for directional hydraulic fracturing based on the stress gradient distribution in the strong rock pressure characteristic vector; an orange warning module, which, when an orange warning signal is triggered, uses a reinforcement learning model to determine a spraying area and thickness distribution of an energy-absorbing material based on a crack density distribution in the strong rock pressure feature vector; The yellow warning module, when the yellow warning signal is triggered, uses a gradient descent algorithm to determine a support resistance adjustment plan for the hydraulic support based on the energy release rate distribution in the strong mine pressure characteristic vector.

Images