Gas extraction intelligent hole sealing control method

Abstract

The application provides a gas extraction intelligent hole sealing control method, and belongs to the technical field of gas extraction control based on computer data processing; first, a multi-modal data set of four kinds of heterogeneous data is constructed, the heterogeneous data is preprocessed, and data features are extracted; the heterogeneous data features are input into a feature enhancement module to obtain heterogeneous data enhancement features, then a hole sealing monitoring point graph structure is constructed, edge weight weighting calculation is performed on the enhancement features related to each monitoring point to obtain a hole sealing area weighted graph feature; the weighted graph feature is input into a hole sealing defect evolution recognition model to obtain a hole sealing defect state prediction value and a corresponding gas escape early warning level; finally, a multi-objective optimization problem is established and solved to obtain a grouting parameter Pareto optimal solution under the gas escape early warning level, and the grouting equipment is controlled according to the optimal solution. The application realizes intelligent recognition and prediction of the space-time evolution process of the hole sealing state, and realizes adaptive adjustment of the hole sealing parameters.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of gas extraction control based on computer data processing, and particularly relates to a gas extraction intelligent hole sealing control method. BACKGROUND

[0002] Gas extraction is a key link in the safety production of coal mines, and its core is to extract gas in the coal seam through drilling, thereby reducing the gas pressure in the mine and preventing the accumulation of gas from causing explosions or the urgent need for hidden hole sealing failure problems in the mining process, which plays an irreplaceable role in ensuring the safety production of coal mines.

[0003] With the in-depth promotion of intelligent mine construction, the traditional method of relying on manual experience to judge the sealing quality has been difficult to meet the engineering needs of high precision and real-time. The existing technology usually uses a single sensor to monitor the grouting pressure or gas concentration, which lacks dynamic perception ability in multiple dimensions and the whole process of hole sealing state, making it difficult to find subtle defects in time and unable to predict the gas escape path in advance. By fusing multi-source sensor data and deep learning algorithm, a hole sealing state space evolution model is constructed, which can break through the limitations of traditional methods. This intelligent hole sealing quality monitoring and early warning technology not only improves the reliability of the gas extraction system, but also provides prospective decision-making basis for underground disaster prevention and control.

[0004] The existing intelligent hole sealing method for gas extraction is as follows:

[0005] A Chinese invention patent with application number CN202410057770.3 discloses an intelligent hole sealing method and device for underground gas extraction in coal mines, which relates to the technical field of gas extraction construction. The method includes: after installing a single-channel extraction hole under the constraint of hole sealing length, issuing a grouting instruction; obtaining the annulus characteristic information of the double-wall annulus in the single-channel extraction hole and the slurry characteristic information of the predetermined slurry to form target grouting factor characteristic information; analyzing the target grouting factor characteristic information through an intelligent grouting decision model and a gas extraction database to obtain a target grouting decision; controlling the hole sealing robot to inject the predetermined slurry into the double-wall annulus according to the target grouting decision and issuing an extraction preparation instruction; placing the predetermined screen pipe into the center channel of the single-channel extraction hole to complete the hole sealing of the coal mine extraction hole. This solves the problem of low intelligent degree and poor sealing precision in the existing technology of underground gas extraction in coal mines, which affects the effect and safety of coal mine gas extraction;

[0006] The Chinese invention patent with the application number CN202111468824.8 discloses a reusable, automatic pressure supplementing gas extraction drill hole sealing system, which includes a liquid storage tank, a sealing grouting pump assembly, a bag type sealing device, a negative pressure recovery pump assembly, and a PLC automatic control system. The bag type sealing device is installed outside the gas extraction pipe and extends into the drill hole. The inlet of the sealing grouting pump assembly is connected with the outlet of the liquid storage tank. The bag type sealing device is provided with a grouting pipe and a recovery pipe inside. The outlet of the sealing grouting pump assembly is connected with the left port of the grouting pipe. The inlet of the negative pressure recovery pump assembly is connected with the left port of the recovery pipe. The outlet of the negative pressure recovery pump assembly is connected with the inlet at the top of the liquid storage tank through a filter. The PLC automatic control system includes a PLC controller, an electromagnetic valve, a pressure gauge, and a pressure sensor. The system realizes repeated grouting and pressure supplementing, and the grouting liquid and the sealing device can be reused. The problems of low gas extraction concentration, rapid attenuation, and poor extraction effect are solved. The drilling efficiency is improved, and the system realizes intelligent and automatic drilling sealing.

[0007] However, the above method has the following limitations:

[0008] 1) Insufficient intelligence. The existing technology mainly uses simple control logic and preset parameters based on PLC for sealing operation, lacking advanced intelligent analysis algorithms. The system cannot perform deep learning and autonomous decision-making on the collected multi-dimensional data, making it difficult to adapt to complex and variable geological conditions and dynamically changing sealing environments in the well;

[0009] 2) Insufficient multi-physical field coupling analysis. The existing technology does not consider the coupling effects of mechanics, thermodynamics, acoustics, and other fields, and the evaluation dimension of sealing integrity is single. It is difficult to comprehensively and accurately reflect the real state of sealing by relying on individual parameters alone, resulting in deviations between quality evaluation results and actual working conditions;

[0010] 3) Inadequate data utilization. A large amount of monitoring data is only used for simple display and recording, without deep mining and analysis. The system lacks modeling and prediction capabilities for sealing evolution rules, and the value of data has not been fully exploited. It cannot provide effective data support for preventive maintenance and optimization decisions;

[0011] 4) Low system integration. Each monitoring subsystem usually operates independently, lacking a collaborative working mechanism. The system fails to achieve a closed-loop control process, resulting in low overall efficiency. The automation level is limited, and manual intervention is still required in key links, affecting response speed and processing efficiency;

[0012] 5) Poor dynamic adaptability. The existing method mainly uses a static control mode and cannot respond to dynamic changes in the sealing process in real time. It lacks the ability to identify progressive defects such as micro-crack expansion and gas leakage path formation. SUMMARY

[0013] In view of the above problems, the application designs an intelligent sealing control method based on hole sealing evolution identification and gas escape early warning, multi-source sensor networks are deployed to collect real-time mechanical, gas, acoustic and thermal multi-modal signals of the sealing section, and a deep learning model is constructed to realize intelligent identification and prediction of the spatio-temporal evolution process of the sealing state. The system dynamically generates the optimal control strategy through a multi-objective game optimization algorithm, and drives the intelligent execution device to realize adaptive adjustment of the sealing parameters.

[0014] The application provides a gas extraction intelligent sealing control method, including the following steps:

[0015] S1, collecting and acquiring four types of heterogeneous data of grouting pressure, pore gas concentration, acoustic emission and infrared thermal imaging in the gas extraction process;

[0016] The four types of heterogeneous data are preprocessed, and the data features of the four types of heterogeneous data are extracted, to obtain a grouting pressure time domain feature set, a pore gas concentration gradient feature, a vibration signal frequency domain feature set and a sealing area heat distribution feature set;

[0017] S2, inputting the four types of heterogeneous data features obtained in S1 into a sealing state spatio-temporal evolution feature enhancement module to obtain four types of heterogeneous data enhanced features; then, based on a sealing state spatio-temporal evolution multi-modal feature correlation network, a sealing monitoring point graph structure is constructed, the edge weight of the four types of heterogeneous data enhanced features involved in each monitoring point is calculated, and a sealing area weighted graph feature is obtained;

[0018] S3, inputting the sealing area weighted graph feature into the sealing defect evolution identification model which is constructed and trained to obtain a sealing defect state prediction value and a corresponding gas escape early warning level;

[0019] S4, taking the sealing strength, material cost and extraction stability as the optimization objectives, taking the grouting parameters as the design variables to establish a multi-objective optimization problem, using a non-dominated sorting genetic algorithm to solve the problem to obtain the Pareto optimal solution of the grouting parameters under the gas escape early warning level output in S3, and controlling the grouting equipment according to the optimal solution.

[0020] Preferably, in S1, four types of heterogeneous sensors are deployed at multiple monitoring points of the sealing section of the gas extraction borehole, including grouting pressure sensors, pore gas concentration sensors, acoustic emission sensors, and infrared thermal imagers; the grouting pressure sensors monitor the grouting pressure fluctuation in real time during the grouting process to obtain grouting pressure data, which can capture the characteristics of incomplete grouting or local blockage; the pore gas concentration sensors are used to collect pore gas concentration data at the interface between the sealing material and the coal body, and the gas concentration is used to identify micro-leakage and gas escape rate; the acoustic emission sensors are used to collect vibration signal data during the stress process of the sealing material, and the vibration signal reflects the interface peeling and internal crack evolution; the infrared thermal imager is used to obtain the thermal distribution image of the sealing area, and the abnormal points in the thermal distribution image reflect the existence of gas leakage in the area.

[0021] Preferably, in S1, the data features of the four types of heterogeneous data are extracted, specifically:

[0022] The pressure peak value, rising rate, and steady duration of the grouting pressure data are calculated, and the three time domain features and the grouting pressure data form a grouting pressure time domain feature set; the gradient of the pore gas concentration data at each sampling time is calculated to obtain the pore gas concentration gradient feature; the amplitude, energy, and main frequency of the vibration signal data are calculated, and the three time domain features and the vibration signal data form a vibration signal frequency domain feature set; the mean value of each temperature point in the thermal distribution image of the sealing area is calculated, and the thermal distribution image of the sealing area and the mean temperature of the sealing area form a sealing area thermal distribution feature set.

[0023] Preferably, the sealing state spatiotemporal evolution feature enhancement module includes a grouting pressure time domain feature enhancement network, which includes a Kalman filter layer, a long short-term memory network layer, a mask mask layer, and a feature concatenation layer; the grouting pressure data in the grouting pressure time domain feature set is input into the Kalman filter layer to remove Gaussian white noise and then input into the long short-term memory network layer to model the time dependence of grouting pressure peak values at different times, obtaining grouting pressure time dependence features; the grouting pressure time dependence features, grouting pressure peak value, grouting pressure rising rate, and grouting pressure steady duration are input into the mask mask layer for feature length alignment and then input into the feature concatenation layer to obtain grouting pressure enhanced features.

[0024] Preferably, the hole sealing state space-time evolution feature enhancement module comprises a pore gas concentration gradient feature enhancement network, the pore gas concentration gradient feature enhancement network comprising a first one-dimensional convolution layer, a radial basis activation function layer, a second one-dimensional convolution layer, a mask layer and a feature splicing layer; first, the pore gas concentration gradient features are input into the one-dimensional convolution layer to obtain local pore gas concentration gradient features, then the local pore gas concentration gradient features are input into the radial basis activation function layer for Gaussian nonlinear transformation to capture pore gas concentration gradient distribution features; thirdly, the local pore gas concentration gradient features are input into the second one-dimensional convolution layer to obtain advanced pore gas concentration gradient features; fourthly, the pore gas concentration gradient distribution features and the advanced pore gas concentration gradient features are input into the mask layer for feature length alignment and then input into the feature splicing layer to obtain pore gas concentration gradient enhanced features.

[0025] Preferably, the hole sealing state space-time evolution feature enhancement module comprises a vibration signal frequency domain feature enhancement network, the vibration signal frequency domain feature enhancement network comprising a short-time Fourier transform layer, a wavelet envelope decomposition layer, a mask layer and a feature splicing layer; first, the vibration signal data in the vibration signal frequency domain feature set are input into the short-time Fourier transform layer for short-time Fourier transform of the vibration signal, the short-time Fourier transform layer performing localized analysis of the vibration signal in the time and frequency domains to obtain vibration signal frequency domain data; the vibration signal frequency domain data are input into the wavelet envelope decomposition layer for modal decomposition to obtain a feature set containing vibration signal features of multiple frequency subbands, and the features that are significantly outlying in the set are deleted to eliminate outlying abnormal data noise contained in the vibration signal data, thereby obtaining a denoising feature set; the denoising feature set, the amplitude, the energy and the main frequency in the vibration signal frequency domain feature set are simultaneously input into the mask layer for feature length alignment and then input into the feature splicing layer to obtain vibration signal frequency domain enhanced features.

[0026] Preferably, the hole sealing state space-time evolution feature enhancement module comprises a hole sealing region heat distribution feature enhancement network, the hole sealing region heat distribution feature enhancement network comprising a pre-trained VGG16 network layer, a full connection layer, a mask layer and a feature splicing layer; first, the hole sealing region heat distribution images in the hole sealing region heat distribution feature set are input into the pre-trained VGG16 network layer, the first 10 layers of the pre-trained VGG16 network layer being frozen to improve the generalization ability in the hole sealing region heat distribution image feature extraction task, and the features of the hole sealing region heat distribution images are gradually extracted through convolution layers and pooling layers to obtain hole sealing region heat distribution spatial texture features, the spatial texture features including temperature gradient features and abnormal region contour features; the advanced features of the hole sealing region heat distribution images are input into the full connection layer to be converted into a one-dimensional hole sealing region heat distribution feature vector; the hole sealing region heat distribution feature vector and the hole sealing region temperature mean value are simultaneously input into the mask layer for feature length alignment and then input into the feature splicing layer to obtain hole sealing region heat distribution enhanced features.

[0027] Preferably, the multi-modal feature correlation network of the temporal and spatial evolution of the sealing state in S2 comprises a graph relationship modeling layer, an edge weight calculation layer, and a feature weighting layer; first, the relative coordinates of each monitoring point in the sealing area are input into the graph relationship modeling layer to obtain a graph structure with each monitoring point in the sealing area as a node and the Euclidean distance between each monitoring point as an edge, and the node features are the grouting pressure enhancement feature, the pore gas concentration gradient enhancement feature, the vibration signal frequency domain enhancement feature, and the sealing area heat distribution enhancement feature; the edges between the nodes in the graph structure are input into the edge weight calculation layer to obtain the edge weights between the nodes, and the edge weights reflect the physical association between the nodes, which are determined according to the distance and physical action strength between the nodes; the closer the distance and the stronger the physical action, the greater the edge weight;

[0028] Finally, the node features and the edge weights in the graph structure are input into the feature weighting layer to perform point multiplication operation of the node features and the corresponding edge weights of the nodes to obtain the weighted graph features of the sealing area.

[0029] Preferably, the sealing defect evolution identification model in S3 comprises a Bayesian algorithm layer and a softmax activation function layer; the weighted graph features of the sealing area obtained are input into the Bayesian algorithm layer to obtain the posterior probability distribution of the sealing area defect state, and the posterior probability distribution of the sealing area defect state is input into the softmax activation function layer to obtain the predicted value of the sealing defect state; if the sealing defect state is predicted to be micro-leakage, the gas escape warning level is I; if the sealing defect state is predicted to be interface peeling, the gas escape warning level is II; if the sealing defect state is predicted to be crack expansion, the gas escape warning level is III; and if the sealing defect state is predicted to be a through channel, the gas escape warning level is IV.

[0030] Preferably, the specific process of S4 comprises:

[0031] S41, defining three optimization objectives of maximizing the sealing strength, minimizing the material cost, and minimizing the gas extraction flow fluctuation, wherein the sealing strength is measured by the compressive strength and bonding strength of the sealing material, reflecting the mechanical stability, and the gas extraction flow fluctuation minimization reflects the extraction stability;

[0032] S42, defining the sealing state regulation decision variable; the decision variable is composed of the grouting material type, the grouting speed, and the reinforcement position, wherein the grouting material type affects the sealing strength of the sealing material and the total cost of the sealing material, the grouting speed affects the sealing strength of the sealing material and the gas extraction flow fluctuation, and the reinforcement position affects the sealing strength of the sealing material and the gas extraction flow fluctuation;

[0033] S43 uses a non-dominated sorting genetic algorithm to solve the optimization problem and obtains the Pareto optimal solution for the control of the sealing state with three optimization objectives: maximizing the balance sealing strength, minimizing the material cost, and minimizing the fluctuation of the extraction flow rate.

[0034] S44 controls the grouting pump and segmented valves by adjusting the grouting material type, grouting speed and reinforcement position corresponding to the Pareto optimal solution of the sealing state, thereby achieving control of the sealing state; wherein the grouting pump performs grouting according to the grouting material type and grouting speed, and the segmented valves control the grouting area according to the reinforcement position.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) Improve the comprehensiveness and accuracy of pore sealing quality monitoring: By deploying a multi-source sensor network, comprehensive monitoring of the mechanical state, gas concentration distribution, internal material damage, and temperature field changes during the pore sealing process is achieved. The multimodal sensing system of this invention can capture multi-dimensional information such as grouting pressure fluctuations, gas micro-leakage, and material crack propagation in real time. Through spatiotemporal evolution characteristic analysis, it can accurately identify defects such as interface peeling and local blockage, significantly improving the accuracy of pore sealing quality assessment;

[0037] (2) Enhancing the intelligence level of gas escape early warning: By constructing a dynamic identification model for the evolution of sealing defects, intelligent prediction and graded early warning of gas escape paths are realized. This invention can predict potential gas escape paths in advance by analyzing the spatiotemporal correlation of gas concentration gradient changes, acoustic emission event distribution, and thermal imaging anomaly areas, and output graded early warnings according to the severity of defects;

[0038] (3) Improve adaptability under complex geological conditions: Introduce a multi-objective game optimization algorithm to enable the system to adapt to different coal seam geological conditions and engineering requirements. This invention establishes a multi-objective optimization model for sealing strength, material cost, and extraction stability, which can dynamically adjust grouting parameters and improve the system's adaptability under different geological conditions. Attached Figure Description

[0039] Figure 1 This is a flowchart illustrating the overall technical route of the present invention.

[0040] Figure 2 This is a structural diagram of the module for enhancing the spatiotemporal evolution characteristics of the sealed state.

[0041] Figure 3 The flowchart shows the process of using a non-dominated sorting genetic algorithm to solve the multi-objective optimization problem of grouting parameters.

[0042] Figure 4 This is a comparison chart of multimodal fusion performance in the embodiments.

[0043] Figure 5The image shows a heatmap of defect identification performance in this embodiment. Detailed Implementation

[0044] This invention proposes an intelligent sealing control method for gas extraction, the overall process of which is as follows: Figure 1 As shown:

[0045] S1. Construction of a multimodal dataset for the spatiotemporal evolution of grout sealing status: First, addressing the issue of single data in traditional grout sealing monitoring, this invention constructs a multimodal dataset based on grouting dynamics and gas seepage theory, containing four types of heterogeneous data: grouting pressure, pore gas concentration, acoustic emission, and infrared thermal imaging. Expert annotation categorizes the grout sealing status into normal and various defect states, and labels the defect evolution stages. Second, preprocessing of the four types of heterogeneous data achieves alignment in the time dimension. Finally, data features of the four types of heterogeneous data are extracted to obtain a set of temporal features for grouting pressure, a set of gradient features for pore gas concentration, a set of frequency features for vibration signals, and a set of thermal distribution features in the grout sealing area.

[0046] S2, the module for enhancing the spatiotemporal evolution features of the sealing state; input the four types of heterogeneous data features obtained in S1 into this module, first enhance the features of the four types of heterogeneous data features to obtain the four types of heterogeneous data enhanced features, then construct the sealing monitoring point map structure, and perform edge weight weighting calculation on the four types of heterogeneous data enhanced features involved in each monitoring point to obtain the weighted map features of the sealing area.

[0047] S3, Sealing Defect Evolution Identification Model; Input the weighted map features of the sealing area into the sealing defect evolution identification model to obtain the predicted value of the sealing defect state and its corresponding gas escape warning level;

[0048] S4 is an adaptive control module for sealing status. It establishes a multi-objective optimization problem with sealing strength, material cost, and extraction stability as optimization objectives and grouting parameters as design variables. It uses a non-dominated sorting genetic algorithm to solve the problem and obtain the Pareto optimal solution of the grouting parameters under the gas escape warning level output by S3. The grouting equipment is then controlled based on the optimal solution.

[0049] The specific implementation process of the present invention will be described in detail below with reference to specific embodiments.

[0050] I. Construction and Processing of Multimodal Dataset of Spatiotemporal Evolution of Sealing State

[0051] Multi-source sensor deployment and data acquisition: Four types of heterogeneous sensors are deployed at multiple monitoring points in the sealing section of the gas drainage borehole, including grouting pressure sensors, pore gas concentration sensors, acoustic emission sensors, and infrared thermal imagers. The grouting pressure sensors monitor grouting pressure fluctuations in real time during the grouting process, acquiring grouting pressure data that can capture characteristics of incomplete grouting or localized blockages. The pore gas concentration sensors collect pore gas concentration data at the interface between the sealing material and the coal seam; gas concentration can be used to identify micro-leakage and gas escape rates. The acoustic emission sensors are used for… Vibration signal data of the sealing material during stress collection can reflect interface peeling and internal crack evolution; an infrared thermal imager is used to acquire thermal distribution images of the sealing area, and abnormal points in the thermal distribution images reflect gas leakage in the area; experts label the data samples based on the real-time status of the sealing and historical test results to ensure the accuracy and consistency of the labeling; the labeling includes the normal state of the sealing and various defect states, where the defect states are specifically micro-leakage, interface peeling, crack propagation and penetrating channels, and the corresponding gas escape degree ranges from slight to severe.

[0052] Preprocessing of multimodal data on spatiotemporal evolution of grouting state: Since the sampling frequencies of different modal sensors are inconsistent, this invention uses the Kriging interpolation method to interpolate the temperature data at each pixel location in the grouting pressure data, pore gas concentration data, vibration signal data and thermal distribution image to achieve alignment in the time dimension.

[0053] Multimodal feature extraction of the spatiotemporal evolution of the sealing state: Preprocessed grouting pressure data, pore gas concentration data, vibration signal data, and thermal distribution image of the sealing area are used. Three temporal features of the grouting pressure data are calculated: peak pressure, rate of rise, and steady-state duration. These three temporal features, along with the grouting pressure data, form a grouting pressure temporal feature set. The gradient of the pore gas concentration data at each sampling time is calculated to obtain the pore gas concentration gradient feature. Three frequency domain features of the vibration signal data are calculated: amplitude, energy, and dominant frequency. These three temporal features, along with the vibration signal data, form a vibration signal frequency domain feature set. The mean value of each temperature point in the thermal distribution image of the sealing area is calculated. The thermal distribution image of the sealing area and the mean temperature of the sealing area form a thermal distribution feature set of the sealing area.

[0054] The formula for calculating the characteristics of pore gas concentration gradient is as follows:

[0055] ;

[0056] in Characteristics of pore gas concentration gradient. This indicates the change in pore gas concentration. It represents the change over time.

[0057] II. Design of the Spatiotemporal Evolution Feature Enhancement Module for Sealing State

[0058] The module for enhancing the spatiotemporal evolution characteristics of the sealing state includes a grouting pressure temporal characteristic enhancement network, a pore gas concentration gradient characteristic enhancement network, a vibration signal frequency domain characteristic enhancement network, and a sealing area thermal distribution characteristic enhancement network.

[0059] 1. Grouting Pressure Temporal Feature Enhancement Network: The grouting pressure temporal feature enhancement network includes a Kalman filter layer, a long short-term memory network layer, a mask layer, and a feature splicing layer. Grouting pressure data from the grouting pressure temporal feature set is input into the Kalman filter layer for Gaussian white noise removal, and then input into the long short-term memory network layer to model the time dependency of grouting pressure peak values ​​at different times, obtaining the grouting pressure time dependency feature. The grouting pressure time dependency feature, grouting pressure peak value, grouting pressure rise rate, and grouting pressure steady-state duration are input into the mask layer for feature length alignment, and then input into the feature splicing layer to obtain the grouting pressure enhancement feature.

[0060] The specific formula for removing Gaussian white noise from grouting pressure data using the Kalman filter layer is shown below:

[0061] ;

[0062] in express Prior state estimation of grouting pressure at time t, Represents the state transition matrix. express Post-hoc state estimation of grouting pressure at time t. Represents the control input matrix. express The grouting pressure control input at any given time; the prior estimate of grouting pressure is corrected by combining the measured grouting pressure data, using the following formula:

[0063] ;

[0064] in surface Denoising grouting pressure data at any given time. Indicates Kalman gain, express The measured value of grouting pressure at any given time. A matrix representing the measured values ​​of grouting pressure data;

[0065] 2. Pore Gas Concentration Gradient Feature Enhancement Network: The pore gas concentration gradient feature enhancement network consists of a first one-dimensional convolutional layer, a radial basis activation function layer, a second one-dimensional convolutional layer, a mask layer, and a feature concatenation layer. First, the pore gas concentration gradient features are input into the one-dimensional convolutional layer to obtain local pore gas concentration gradient features. Second, the local pore gas concentration gradient features are input into the radial basis activation function layer for Gaussian nonlinear transformation to capture the pore gas concentration gradient distribution features. Third, the local pore gas concentration gradient features are input into the second one-dimensional convolutional layer to obtain high-level pore gas concentration gradient features. Finally, the pore gas concentration gradient distribution features and high-level pore gas concentration gradient features are input into the mask layer for feature length alignment and then input into the feature concatenation layer to obtain the enhanced pore gas concentration gradient features.

[0066] 3. Vibration Signal Frequency Domain Feature Enhancement Network: The vibration signal frequency domain feature enhancement network includes a short-time Fourier transform layer, a wavelet envelope decomposition layer, a mask layer, and a feature concatenation layer. First, the vibration signal data from the vibration signal frequency domain feature set is input into the short-time Fourier transform layer to perform a short-time Fourier transform on the vibration signal. The short-time Fourier transform performs localized analysis of the vibration signal in the time and frequency domains, obtaining the vibration signal frequency domain data. The vibration signal frequency domain data is then input into the wavelet envelope decomposition layer for modal decomposition, obtaining a feature set containing vibration signal features from multiple frequency sub-bands. Significant outlier features in this set are removed to eliminate outlier noise in the vibration signal data, resulting in a denoised feature set. Finally, the amplitude, energy, and dominant frequency from the denoised feature set and the vibration signal frequency domain feature set are simultaneously input into the mask layer for feature length alignment, and then input into the feature concatenation layer to obtain the enhanced vibration signal frequency domain features.

[0067] 4. Enhancing Network for Thermal Distribution Features in Sealed Areas: This network consists of a pre-trained VGG16 network layer, a fully connected layer, a mask layer, and a feature concatenation layer. First, the thermal distribution image of the sealed area from the feature set is input into the pre-trained VGG16 network layer. The first 10 layers of the pre-trained VGG16 network have their parameters frozen to improve generalization ability in the task of extracting features from the thermal distribution image of the sealed area. Features of the thermal distribution image of the sealed area are extracted step by step through convolutional layers and pooling layers to obtain spatial texture features of the thermal distribution of the sealed area. These spatial texture features include temperature gradient features and abnormal region contour features. The high-level features of the thermal distribution image of the sealed area are input into the fully connected layer to transform them into a one-dimensional thermal distribution feature vector of the sealed area. The thermal distribution feature vector of the sealed area and the mean temperature of the sealed area are simultaneously input into the mask layer for feature length alignment before being input into the feature concatenation layer to obtain the enhanced features of the thermal distribution of the sealed area.

[0068] 5. Spatiotemporal Evolution Multimodal Feature Association Network for Sealing State: To characterize the feature correlations of different monitoring points in the sealing area, capture the potential interactive influence relationships between the features of the monitoring points, and improve the accuracy of model prediction, this application designs a spatiotemporal evolution multimodal feature association network for sealing state, including a graph relationship modeling layer, an edge weight calculation layer, and a feature weighting layer. First, the relative coordinates of each monitoring point in the sealing area are input into the graph relationship modeling layer to obtain a graph structure with each monitoring point in the sealing area as a node and the Euclidean distance between each monitoring point as an edge. The node features are the obtained grouting pressure enhancement feature, pore gas concentration gradient enhancement feature, vibration signal frequency domain enhancement feature, and sealing area heat distribution enhancement feature. The edges between each node in the graph structure are input into the edge weight calculation layer to obtain the edge weights between each node. The edge weights reflect the physical correlation between nodes and are determined according to the distance between nodes and the intensity of physical interaction. The closer the distance and the stronger the physical interaction, the larger the edge weight. The calculation formula is as follows:

[0069] ;

[0070] in Representing nodes in a graph structure and nodes Edge weights between them Indicates the weighting coefficient. Represents a node and nodes The Euclidean distance between them Indicates the distance attenuation coefficient;

[0071] Finally, the features of each node and the weights of the edges in the graph structure are input into the feature weighting layer, and the dot product operation between the features of each node and the corresponding edge weights is performed to obtain the weighted graph features of the sealing area.

[0072] III. Design of Dynamic Identification Model for the Evolution of Sealing Defects

[0073] This invention designs a dynamic identification model for the evolution of sealing defects, including a Bayesian algorithm layer and a softmax activation function layer, to achieve accurate identification of the evolution process of sealing defects, as detailed below:

[0074] The obtained weighted map features of the sealing area are input into the Bayesian algorithm layer to obtain the posterior probability distribution of the defect state of the sealing area. The formula for calculating the posterior probability is as follows:

[0075] ;

[0076] in Indicates the weighted map features in the hole sealing region. Down The posterior probability, Indicating defect state D The likelihood probability, express The prior probability, express Probability of evidence; It is one of four types of defects: microleakage, interface peeling, crack propagation, and through-channel; the posterior probabilities of the four types of defects constitute the posterior probability distribution of the defect state in the sealing area.

[0077] The posterior probability distribution of the defect state in the sealing area is input into the softmax activation function layer to obtain the predicted value of the sealing defect state. If the sealing defect state is predicted as micro-leakage, the gas escape warning level is Level I; if the sealing defect state is predicted as interface peeling, the gas escape warning level is Level II; if the sealing defect state is predicted as crack propagation, the gas escape warning level is Level III; if the sealing defect state is predicted as a through channel, the gas escape warning level is Level IV.

[0078] IV. Adaptive Control of Sealing Status

[0079] Multi-objective optimization model construction: Three optimization objectives are defined: maximizing sealing strength, minimizing material cost, and minimizing extraction flow rate fluctuation. Sealing strength is measured by the compressive and adhesive strength of the sealing material, reflecting mechanical stability. Minimizing extraction flow rate fluctuation reflects extraction stability. The specific formula is as follows:

[0080] ;

[0081] in Indicates the sealing strength of the sealing material. Indicates the compressive strength of the sealing material. Indicates the bonding strength of the sealing material. , These represent the weighting coefficients for the compressive strength and bond strength of the sealing material, respectively. These two coefficients are determined by the gas escape warning level output by S3-2. If the gas escape warning level is Level I, then... If the gas escape warning level is Level II, then If the gas escape warning level is Level III, then 7. If the gas escape warning level is Level IV, then ; This represents the total cost of the sealing material. Indicates the first The unit cost of this type of sealing material Indicates the first The amount of sealing material used Indicates the type of sealing material; This indicates the fluctuation in gas extraction flow rate. Indicates the length of time. express Gas extraction flow rate at any given time This represents the average gas extraction flow rate;

[0082] Define the decision variables for controlling the sealing state: The decision variables consist of the type of grouting material, the grouting speed, and the reinforcement location. Among them, the type of grouting material affects the sealing strength and the total cost of the sealing material, the grouting speed affects the sealing strength and the fluctuation of the gas extraction flow rate, and the reinforcement location affects the sealing strength and the fluctuation of the gas extraction flow rate.

[0083] A Pareto optimal solution for the control of the sealing state is obtained by using a non-dominated sorting genetic algorithm to solve the defined optimization problem, which aims to maximize the equilibrium sealing strength, minimize the material cost, and minimize the fluctuation of the extraction flow rate.

[0084] By controlling the grouting pump and segmented valves according to the Pareto optimal solution for the sealing state, the grouting material type, grouting speed, and reinforcement location are controlled to achieve control of the sealing state and prevent further deterioration of sealing defects. The grouting pump performs grouting according to the grouting material type and grouting speed, and the segmented valves control the grouting area according to the reinforcement location.

[0085] V. Experimental Verification and Analysis

[0086] This experiment, conducted on a test platform built at a coal mine underground gas drainage borehole, involved 30 days of continuous testing, covering typical working conditions under different coal seam geological conditions (soft and medium-hard seams), different borehole depths (50-150m), and different drainage stages. Data acquisition was completed simultaneously using grouting pressure sensors, pore gas concentration sensors, acoustic emission sensors, and infrared thermal imaging devices. Baseline data was also obtained through manual borehole inspection and gas flow metering. The dataset contains 300 complete borehole cycle data sets, divided into training, validation, and test sets in a 6:2:2 ratio.

[0087] 1. Sealing defect identification performance

[0088] The verification was performed from three dimensions: recognition accuracy, early warning lead time, and robustness to modal loss, as shown in Table 1.

[0089] Evaluation indicators:

[0090] Defect identification accuracy: The proportion of correctly identified defect samples out of the total samples;

[0091] Early Warning Time (EAT): The time difference between a system warning and the actual occurrence of a defect;

[0092] Modal Missing Adaptation Rate (MAR): The rate at which recognition accuracy is maintained when a single modality is missing;

[0093] Table 1. Experimental results of sealing defect identification performance.

[0094]

[0095] 2. Multimodal fusion performance

[0096] The multimodal feature fusion module underwent specific verification, as shown in Table 2:

[0097] Evaluation indicators:

[0098] Feature Association Strength (FCS): The mutual information value between multimodal features;

[0099] Spatiotemporal alignment error (STA): the deviation between timestamps and spatial coordinates of multi-source data;

[0100] Fusion computation time (T): The time required to fuse a single set of data;

[0101] Table 2 Experimental results of multimodal fusion performance

[0102]

[0103] 3. Optimization effect of regulation strategy

[0104] The verification results for the multi-objective optimization and control module are shown in Table 3:

[0105] Evaluation indicators:

[0106] Sealing efficiency (SE): the percentage reduction in gas leakage rate;

[0107] Material cost savings rate (CR): Cost comparison with traditional solutions;

[0108] Sampling flow stability (FS): Standard deviation of flow fluctuation;

[0109] Table 3. Experimental Results of Optimization of Regulation Strategies

[0110]

[0111] Through Tables 1, 2, and 3, and Figure 4 , Figure 5 Experimental results show that the present invention maintains high accuracy under different working conditions, and the early warning time meets the needs of on-site handling, especially under complex conditions such as soft coal seams. The multimodal fusion module significantly improves the feature correlation strength through graph structure modeling, while reducing spatiotemporal alignment errors. The control strategy effectively saves material costs and improves extraction stability while ensuring sealing efficiency, and complies with coal mine safety regulations.

[0112] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0113] While the specific embodiments of the present invention have been described above, they are not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for intelligent sealing control of gas extraction, characterized in that, Includes the following steps: S1 collects four types of heterogeneous data during the gas extraction process: grouting pressure, pore gas concentration, acoustic emission, and infrared thermal imaging. The four types of heterogeneous data were preprocessed and their data features were extracted to obtain the time-domain feature set of grouting pressure, the pore gas concentration gradient feature set, the frequency-domain feature set of vibration signal, and the thermal distribution feature set of the sealing area. S2, input the four types of heterogeneous data features obtained in S1 into the spatiotemporal evolution feature enhancement module of the sealing state to obtain four types of heterogeneous data enhancement features; Then, based on the spatiotemporal evolution multimodal feature association network of the sealing state, a sealing monitoring point graph structure is constructed. The edge weights of the four types of heterogeneous data enhancement features involved in each monitoring point are calculated to obtain the weighted graph features of the sealing area. The module for enhancing the spatiotemporal evolution characteristics of the sealing state includes a grouting pressure time-domain feature enhancement network, a pore gas concentration gradient feature enhancement network, a vibration signal frequency-domain feature enhancement network, and a sealing area heat distribution feature enhancement network. The spatiotemporal evolution multimodal feature association network of the sealing state includes a graph relationship modeling layer, an edge weight calculation layer, and a feature weighting layer. First, the relative coordinates of each monitoring point in the sealing area are input into the graph relationship modeling layer to obtain a graph structure with each monitoring point in the sealing area as a node and the Euclidean distance between each monitoring point as an edge. The node features are grouting pressure enhancement features, pore gas concentration gradient enhancement features, vibration signal frequency domain enhancement features, and sealing area heat distribution enhancement features. The edge weights between nodes in the graph structure are obtained by inputting the edge weights into the edge weights calculation layer. The edge weights reflect the physical relationships between nodes and are determined based on the distance between nodes and the strength of the physical interaction. The closer the distance and the stronger the physical interaction, the greater the edge weight. Finally, the features of each node in the graph structure are input into the feature weighting layer, and the dot product operation between the node features and the corresponding edge weights is performed to obtain the weighted graph features of the sealing area. S3, input the weighted map features of the sealing area into the constructed and trained sealing defect evolution recognition model to obtain the predicted value of the sealing defect state and its corresponding gas escape warning level; S4 establishes a multi-objective optimization problem with sealing strength, material cost, and extraction stability as optimization objectives and grouting parameters as design variables. The problem is solved using a non-dominated sorting genetic algorithm to obtain the Pareto optimal solution of the grouting parameters under the gas escape warning level output by S3. The grouting equipment is then controlled based on the optimal solution.

2. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: In S1, four types of heterogeneous sensors are deployed at multiple monitoring points in the sealing section of the gas drainage borehole, including a grouting pressure sensor, a pore gas concentration sensor, an acoustic emission sensor, and an infrared thermal imager. The grouting pressure sensor monitors the grouting pressure fluctuations in real time during the grouting process and collects grouting pressure data, which can capture characteristics of incomplete grouting or local blockage. The pore gas concentration sensor is used to collect pore gas concentration data at the interface between the sealing material and the coal body, and the gas concentration is used to identify micro-leakage and gas escape rate. The acoustic emission sensor is used to collect vibration signal data of the sealing material during the stress process, and the vibration signal reflects interface peeling and internal crack evolution. The infrared thermal imager is used to obtain thermal distribution images of the sealing area, and abnormal points in the thermal distribution images reflect the presence of gas leakage in the area.

3. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: The data features extracted from the four types of heterogeneous data in S1 are as follows: Three time-domain features of grouting pressure data are calculated: peak pressure, rate of rise, and steady-state duration. These three time-domain features and the grouting pressure data together form a set of grouting pressure time-domain features. The gradient of pore gas concentration data at each sampling time is calculated to obtain the pore gas concentration gradient features. The amplitude, energy, and dominant frequency of vibration signal data are calculated. These three time-domain features, together with the vibration signal data, form a set of frequency domain features of the vibration signal. Calculate the mean value of each temperature point in the thermal distribution image of the sealing area, and combine the thermal distribution image of the sealing area and the mean temperature of the sealing area to form a set of thermal distribution features of the sealing area.

4. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: The spatiotemporal evolution feature enhancement module for the sealing state includes a grouting pressure time-domain feature enhancement network, which includes a Kalman filter layer, a long short-term memory network layer, a mask layer, and a feature splicing layer. The grouting pressure data in the grouting pressure time-domain feature set is input into the Kalman filter layer for Gaussian white noise removal, and then input into the long short-term memory network layer to model the time dependence of the grouting pressure peak at different times, thereby obtaining the grouting pressure time dependence feature. The grouting pressure time-dependent features, grouting pressure peak value, grouting pressure rise rate, and grouting pressure steady-state duration are input into the mask layer, aligned with the feature length, and then input into the feature splicing layer to obtain the grouting pressure enhancement features.

5. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: The spatiotemporal evolution feature enhancement module for the sealing state includes a pore gas concentration gradient feature enhancement network, which comprises a first one-dimensional convolutional layer, a radial basis activation function layer, a second one-dimensional convolutional layer, a mask layer, and a feature splicing layer. First, the pore gas concentration gradient features are input into the one-dimensional convolutional layer to obtain local pore gas concentration gradient features. Second, the local pore gas concentration gradient features are input into the radial basis activation function layer for Gaussian nonlinear transformation to capture the pore gas concentration gradient distribution features. Third, the local pore gas concentration gradient features are input into the second one-dimensional convolutional layer to obtain higher-level pore gas concentration gradient features. Finally, the pore gas concentration gradient distribution features and the higher-level pore gas concentration gradient features are input into the mask layer for feature length alignment and then input into the feature splicing layer to obtain the enhanced pore gas concentration gradient features.

6. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: The spatiotemporal evolution feature enhancement module for the sealing state includes a vibration signal frequency domain feature enhancement network, which comprises a short-time Fourier transform layer, a wavelet envelope decomposition layer, a mask layer, and a feature splicing layer. First, the vibration signal data from the vibration signal frequency domain feature set is input into the short-time Fourier transform layer to perform a short-time Fourier transform on the vibration signal. The short-time Fourier transform performs localized analysis of the vibration signal in the time and frequency domains to obtain the vibration signal frequency domain data. The vibration signal frequency domain data is then input into the wavelet envelope decomposition layer for modal decomposition to obtain a feature set containing vibration signal features from multiple frequency sub-bands. Significantly outlier features in this set are removed to eliminate outlier and abnormal data noise contained in the vibration signal data, resulting in a denoised feature set. Finally, the amplitude, energy, and dominant frequency from the denoised feature set and the vibration signal frequency domain feature set are simultaneously input into the mask layer for feature length alignment before being input into the feature splicing layer to obtain the enhanced vibration signal frequency domain features.

7. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: The module for enhancing the spatiotemporal evolution features of the sealed area includes a network for enhancing the thermal distribution features of the sealed area. This network comprises a pre-trained VGG16 network layer, a fully connected layer, a mask layer, and a feature concatenation layer. First, the thermal distribution image of the sealed area from the feature set is input into the pre-trained VGG16 network layer. The first 10 layers of the pre-trained VGG16 network have their parameters frozen to improve generalization ability in the task of extracting features from the thermal distribution image of the sealed area. Features of the thermal distribution image of the sealed area are extracted step by step through convolutional layers and pooling layers to obtain spatial texture features of the thermal distribution of the sealed area. These spatial texture features include temperature gradient features and abnormal region contour features. The high-level features of the thermal distribution image of the sealed area are input into the fully connected layer to transform them into a one-dimensional thermal distribution feature vector of the sealed area. The thermal distribution feature vector and the average temperature of the sealed area are simultaneously input into the mask layer for feature length alignment before being input into the feature concatenation layer to obtain the enhanced features of the thermal distribution of the sealed area.

8. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: The sealing defect evolution identification model in S3 includes a Bayesian algorithm layer and a softmax activation function layer; the obtained weighted map features of the sealing area are input into the Bayesian algorithm layer to obtain the posterior probability distribution of the sealing area defect state, and the posterior probability distribution of the sealing area defect state is input into the softmax activation function layer to obtain the predicted value of the sealing defect state. If the sealing defect is predicted to be micro-leakage, the gas escape warning level is Level I; if the sealing defect is predicted to be interface peeling, the gas escape warning level is Level II; if the sealing defect is predicted to be crack propagation, the gas escape warning level is Level III; if the sealing defect is predicted to be a through channel, the gas escape warning level is Level IV.

9. The intelligent sealing control method for gas extraction as described in claim 1, characterized in that: The specific process of S4 includes: S41 defines three optimization objectives: maximizing sealing strength, minimizing material cost, and minimizing drainage flow fluctuation. Sealing strength is measured by the compressive strength and bonding strength of the sealing material, reflecting mechanical stability, while minimizing drainage flow fluctuation reflects drainage stability. S42, define the decision variables for controlling the sealing state; the decision variables consist of the type of grouting material, the grouting speed, and the reinforcement location. Among them, the type of grouting material affects the sealing strength and the total cost of the sealing material, the grouting speed affects the sealing strength and the fluctuation of the gas extraction flow rate, and the reinforcement location affects the sealing strength and the fluctuation of the gas extraction flow rate. S43 uses a non-dominated sorting genetic algorithm to solve the optimization problem and obtains the Pareto optimal solution for the control of the sealing state with three optimization objectives: maximizing the balance sealing strength, minimizing the material cost, and minimizing the fluctuation of the extraction flow rate. S44 controls the grouting pump and segmented valves by adjusting the grouting material type, grouting speed and reinforcement position corresponding to the Pareto optimal solution of the sealing state, thereby achieving control of the sealing state; wherein the grouting pump performs grouting according to the grouting material type and grouting speed, and the segmented valves control the grouting area according to the reinforcement position.

Images