Deep coal bed gas multi-parameter real-time monitoring and safety early warning method and system

Abstract

The invention provides a deep coal bed gas multi-parameter real-time monitoring and safety early warning method, which comprises the following steps of: performing acquisition quality evaluation and anomaly screening on original multi-source monitoring data to obtain screened monitoring data, and performing time-space reference unification processing to obtain a standardized data set; constructing a physical rule constraint condition to perform feature reconstruction on the standardized data set to obtain an enhanced feature set; performing space-time fusion on the enhanced feature set to obtain a fusion early warning result; and constructing a dynamic early warning model to perform real-time early warning according to the fusion early warning result. According to the method, the multi-source monitoring data is subjected to quality evaluation, anomaly screening and time-space reference unification processing, feature reconstruction is completed in combination with physical rule constraints, an early warning result is obtained through time-space fusion, and a dynamic early warning model is constructed, so that the parameter processing efficiency and the acquisition precision are improved, the early warning accuracy is improved, and the risk of false alarm and missing alarm is reduced; and important guarantee is provided for mining safety of deep coal bed gas.

Description

Technical Field

[0001] This invention relates to the field of coalbed methane extraction technology, and in particular to a method and system for real-time monitoring and safety early warning of multiple parameters in deep coalbed methane. Background Technology

[0002] With the continued growth in global demand for clean energy and the near saturation of shallow coalbed methane resources, the development and utilization of deep coalbed methane, as a high-quality unconventional natural gas resource, is of significant strategic importance for ensuring energy supply security and optimizing the energy structure. During deep coalbed methane extraction, monitoring and safety early warning are crucial for ensuring operational safety and improving extraction efficiency. This requires real-time sensing and analysis of key parameters such as coal seam pressure, temperature, gas concentration, and microseismic signals to identify risks such as gas outbursts and coal-rock dynamic disasters in advance.

[0003] In recent years, deep coalbed methane development technology has gradually moved towards large-scale and intelligent development. However, constrained by deep geological conditions, its extraction environment exhibits significant characteristics such as high temperature, high pressure, high corrosivity, and strong reservoir heterogeneity, placing stringent requirements on the environmental adaptability and parameter coordination of monitoring technologies. While existing technologies have developed single or combined monitoring methods based on sensors, distributed optical fibers, and microseismic monitoring, these traditional monitoring models are increasingly unable to meet the demands for real-time, accurate, and integrated monitoring as extraction depth and development complexity increase. Specifically, this manifests in two ways: firstly, the dynamic changes in deep reservoir parameters are drastic, requiring high-frequency, multi-dimensional parameter acquisition to capture early warning signs of disasters; secondly, engineering operations such as fracturing and drainage adjustments during extraction significantly interfere with monitoring data, necessitating more efficient data processing and risk identification capabilities to avoid misjudgments caused by distorted monitoring data. Therefore, designing a multi-parameter real-time monitoring and safety early warning method and system for deep coalbed methane is essential. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for real-time monitoring and safety early warning of multiple parameters in deep coalbed methane. By screening for anomalies and performing fusion constraint processing on monitoring data, combined with a dynamic early warning model, the safety and development efficiency of deep coalbed methane extraction can be improved.

[0005] To achieve the above objectives, the present invention provides the following solution: A method for real-time monitoring and safety early warning of multiple parameters in deep coalbed methane includes the following steps: The original multi-source monitoring data collected during the deep coalbed methane extraction process were assessed for quality and anomaly screening were performed to obtain screened monitoring data. The selected monitoring data were processed to unify the spatiotemporal reference, resulting in a standardized dataset. We construct physical rules and constraints for the mining process, and reconstruct the features of the standardized dataset based on these physical rules and constraints to obtain an enhanced feature set. The enhanced feature set is subjected to spatiotemporal fusion processing to obtain fusion early warning results; A dynamic early warning model is constructed based on the fusion early warning results to provide real-time early warnings.

[0006] Optionally, the raw multi-source monitoring data during the deep coalbed methane extraction process are subjected to acquisition quality assessment and anomaly screening to obtain screened monitoring data, including: A quality assessment threshold table is constructed based on the missing rate, signal-to-noise ratio, and sensor data bias of the original multi-source monitoring data. Anomaly-labeled datasets were obtained by identifying outliers in the original multi-source monitoring data using the isolated forest algorithm. The anomaly-labeled dataset is imputed by interpolation and locally weighted regression to obtain the repaired dataset. The repair dataset is filtered by using a quality assessment threshold table to obtain the filtered monitoring data.

[0007] Optionally, the selected monitoring data undergoes spatiotemporal benchmark unification processing to obtain a standardized dataset, including: A timestamp is added to the filtered monitoring data by a local clock calibration unit to perform time base synchronization and obtain time-stamped data. Using the arrival time of P-waves in microseismic monitoring data as the reference time axis, dynamic time warping calculations are performed on the time-stamped data to obtain the time offset. Time offset correction is performed on the time-stamped data based on the time offset to obtain a time-aligned dataset; A three-dimensional coordinate system for coal seams is constructed based on well logging data in the time-aligned dataset, and spatial coordinates are assigned and three-dimensional coordinates are mapped to the time-aligned dataset to obtain spatially labeled data. Static and dynamic features of spatially labeled data are extracted to standardize the spatially labeled data and obtain a standardized dataset.

[0008] Optionally, static and dynamic features of the spatially labeled data are extracted to standardize the spatially labeled data, resulting in a standardized dataset, including: The mean and extreme values ​​of spatially labeled data are calculated using neighborhood statistics, and the coefficient of variation is calculated based on the mean and extreme values ​​to filter the spatially labeled data and obtain static features. By extracting the temporal trend features of spatially labeled data through a sliding window and calculating mutual information, dynamic features are obtained. The static features are subjected to min-max standardization, and the dynamic features are subjected to median standardization to obtain a standardized dataset.

[0009] Optionally, physical rule constraints in the mining process are constructed, and the standardized dataset is reconstructed based on these physical rule constraints to obtain an enhanced feature set, including: Physical rules and constraints are constructed based on the Langmuir equation and the relationship between permeability and stress during the mining process. Based on the physical equations, the data in the standardized dataset that do not conform to the physical rules and constraints are corrected and calculated to obtain the initial feature set; Simulated samples are generated from the scarce samples in the initial feature set. The distribution difference between the simulated samples and the real samples is calculated. The data augmentation of the initial feature set is performed based on the distribution difference to obtain the expanded feature set. The augmented feature set is reconstructed based on the learning model trained on historical data to obtain the enhanced feature set.

[0010] Optionally, the learning model includes: an input layer, a feature fusion layer, an attention filtering layer, and a feature reconstruction output layer; the input layer maps the expanded feature set to the [-1,1] interval to obtain input features; the feature fusion layer is a two-branch fully connected network, the first branch consists of two fully connected sub-layers with 2N neurons and the activation function of the first sub-layer is LeakyReLU, the second sub-layer has N neurons and the activation function of the second sub-layer is ELU, and the second branch is a fully connected sub-layer with N neurons; the feature fusion layer fuses the input features to obtain initial fused features; the attention filtering layer uses a dual attention filtering mechanism to weight and filter the initial fused features to obtain attention-filtered features; the feature reconstruction output layer transforms and reconstructs the attention-filtered features through a fully connected layer to obtain an enhanced feature set.

[0011] Optionally, the enhanced feature set is subjected to spatiotemporal fusion processing to obtain fusion warning results, including: The enhanced feature set is spatially network-matched with the three-dimensional geological model of the coal seam to obtain a grid dataset. Extract multidimensional features from the grid dataset and construct a multidimensional feature matrix; the multidimensional features include: basic features, trend features, event features, and derived features; Calculate the contribution value of each feature in the multidimensional feature matrix to obtain the feature importance weight matrix; Attention weights are obtained by updating the feature importance weight matrix using a dual-channel attention network. The multidimensional feature matrix is ​​weighted and fused using attention weights to obtain a fused feature vector; The risk output is obtained by using a specialized early warning model to analyze the fused feature vectors and then infer the results. The inference results are corrected and synthesized by using the conflict coefficient to obtain a fused early warning result.

[0012] Optionally, a dynamic early warning model is constructed based on the fused early warning results for real-time early warning, including: Using a three-dimensional spatial grid as nodes, a coal seam monitoring node association diagram is constructed based on the fused early warning results; The risk probability distribution of the nodes is obtained by calculating the risk probability of the coal seam monitoring node association graph using a graph convolutional neural network. Security warnings are issued by comparing the node risk probability distribution with a preset warning threshold.

[0013] A multi-parameter real-time monitoring and safety early warning system for deep coalbed methane includes: The data filtering module is used to assess the acquisition quality and screen for anomalies in the raw multi-source monitoring data during the deep coalbed methane extraction process, and to obtain filtered monitoring data. The standardization module is used to unify the spatiotemporal benchmarks of the screened monitoring data to obtain a standardized dataset. The feature enhancement module is used to construct physical rule constraints in the mining process and reconstruct features of the standardized dataset based on the physical rule constraints to obtain an enhanced feature set. The risk prediction module is used to perform spatiotemporal fusion processing on the enhanced feature set to obtain fusion early warning results; The security early warning module is used to build a dynamic early warning model based on the fusion early warning results, so as to generate risk assessment results and provide real-time early warnings.

[0014] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The method for real-time monitoring and safety early warning of multiple parameters in deep coalbed methane provided by the present invention includes: assessing the acquisition quality and screening for anomalies of the original multi-source monitoring data during the deep coalbed methane extraction process to obtain screened monitoring data; performing spatiotemporal benchmark unification processing on the screened monitoring data to obtain a standardized dataset; constructing physical rule constraints during the extraction process, and reconstructing features of the standardized dataset according to the physical rule constraints to obtain an enhanced feature set; performing spatiotemporal fusion processing on the enhanced feature set to obtain a fused early warning result; and constructing a dynamic early warning model based on the fused early warning result for real-time early warning. This method improves parameter processing efficiency and acquisition accuracy by assessing the quality and screening for anomalies of the original multi-source monitoring data during deep coalbed methane extraction, performing spatiotemporal benchmark unification processing, combining physical rule constraints to complete feature reconstruction, obtaining early warning results through spatiotemporal fusion, and constructing a dynamic early warning model. It also improves early warning accuracy and reduces the risk of false alarms and missed alarms, providing important protection for the safety of deep coalbed methane extraction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the method for real-time monitoring and safety early warning of multiple parameters in deep coalbed methane according to the present invention; Figure 2 This is a schematic diagram of the structure of the deep coalbed methane multi-parameter real-time monitoring and safety early warning system of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] like Figure 1 As shown, this invention provides a method for real-time monitoring and safety early warning of multiple parameters in deep coalbed methane, comprising the following steps: Step 100: Conduct quality assessment and anomaly screening of the original multi-source monitoring data during the deep coalbed methane extraction process to obtain screened monitoring data; Step 200: Perform spatiotemporal benchmark unification processing on the screened monitoring data to obtain a standardized dataset; Step 300: Construct physical rule constraints in the mining process, and reconstruct the features of the standardized dataset based on the physical rule constraints to obtain an enhanced feature set; Step 400: Perform spatiotemporal fusion processing on the enhanced feature set to obtain the fusion warning result; Step 500: Construct a dynamic early warning model based on the fusion early warning results for real-time early warning.

[0020] Specifically, step 100 first constructs a quality assessment threshold table for the three core indicators of the original multi-source monitoring data. The missing rate is the ratio of the number of missing data entries to the total number of data entries; the signal-to-noise ratio is the ratio of the effective signal power to the noise power; and the sensor data deviation is the difference between the actual measured value and the standard reference value. By statistically analyzing the distribution characteristics of the mean, standard deviation, and quartiles of these three indicators in a large amount of historical valid monitoring data, reasonable threshold ranges for each indicator are determined, thus forming the quality assessment threshold table. Then, the isolated forest algorithm is used. Multiple sample subsets are randomly extracted from the original multi-source monitoring data. Features are randomly selected for each subset and randomly segmented to construct multiple isolated trees. The average path length of each data point in all isolated trees is then calculated to obtain an anomaly score. An anomaly score closer to 1 indicates a higher probability of the data being an outlier, while a score closer to 0 indicates a more normal value. Anomalies are marked with anomaly scores to form an anomaly-labeled dataset. Next, for missing or outliers in the anomaly-labeled dataset, linear interpolation is used for filling, with the expression: ; in, These are the missing values ​​to be filled. The sampling time corresponding to the missing value. and The valid sampling times adjacent to the missing value are: and for and The corresponding valid monitoring values. Then, the imputed data is corrected using local weighted regression to obtain the repaired dataset. The formula for calculating the fitted value during the correction process is: ; in, Let x be the regression fit value. For the first i The weights of each sample are calculated using the Gaussian kernel function. Let be the monitoring value of the i-th sample. Finally, the missing rate, signal-to-noise ratio, and sensor data bias of the repaired dataset are compared with the corresponding thresholds in the quality assessment threshold table. Data that meet all the threshold requirements are selected as the filtered monitoring data.

[0021] It should be noted that step 100 initially constructs a multi-dimensional quality assessment threshold table by integrating three key indicators: missing rate, signal-to-noise ratio, and sensor data bias. This enables a comprehensive and accurate assessment of the quality of multi-source monitoring data, avoiding the limitations of a single indicator. The isolated forest algorithm is used for outlier identification, eliminating the need for a pre-defined data distribution model. It accurately captures nonlinear anomaly data in the complex dynamic environment of deep coalbed methane mining, exhibiting stronger data adaptability and improving the accuracy of anomaly identification. Then, interpolation imputation is combined with local weighted regression, ensuring both data integrity and restoring the true data trend. This solves the technical problem that a single repair method cannot simultaneously address both aspects, significantly improving data quality and the accuracy and reliability of the filtered data. It reduces early warning bias caused by data distortion, improves data processing accuracy and efficiency, and lays a core foundation for the accuracy of safety early warnings in deep coalbed methane mining.

[0022] Specifically, step 200 first adds millisecond-accurate timestamps to each of the screened monitoring data using a local clock calibration unit integrating a high-precision crystal oscillator or GPS timing module, generating time-stamped data. Then, using the moment when the P-wave first arrives at the sensor in the microseismic monitoring data as the reference time axis, dynamic time warping is performed between the time-stamped data's time series and the reference time axis; that is, the alignment path with the minimum cumulative distance between the two series is found. The formula for calculating the cumulative distance is: ; in, The monitoring value at the i-th time point of the time-stamped data. Interpolated monitoring value corresponding to the j-th P-wave time on the reference axis The Euclidean distance is used to determine the time offset, and the difference between each corresponding time point in the path is the time offset. Then, each original timestamp of the time-marked data is corrected based on the time offset to obtain the corrected timestamp, thus forming a time-aligned dataset. Next, based on the well logging data in the time-aligned dataset, a three-dimensional coordinate system for the coal seam is constructed with the wellhead center point as the origin O(0,0,0), setting the horizontal direction east as the X-axis, the horizontal direction north as the Y-axis, and the vertical direction downward as the Z-axis. The X and Y coordinates are then calculated based on the horizontal distance L and azimuth θ of the monitoring point using the formulas: X = L × cosθ, Y = L × sinθ. After determining the Z coordinate based on the well logging depth, a unique three-dimensional coordinate is assigned to each data point in the time-aligned dataset, completing the three-dimensional coordinate mapping and obtaining spatially marked data. Finally, the static and dynamic features of the spatially marked data are extracted and standardized to obtain a standardized dataset.

[0023] Furthermore, static features are obtained using neighborhood statistics. First, a spatial neighborhood with a radius of 50 meters centered on each monitoring point is defined. The mean and extreme values ​​of the data within the neighborhood are calculated. Then, the ratio of the standard deviation to the mean of the data within the neighborhood is used as the coefficient of variation, and features with a coefficient of variation < 0.3 are selected as static features. Dynamic features are extracted using a sliding window. First, the time-series trend slope of the data within the window is calculated using the following formula: Next, calculate the mutual information of the time series of different monitoring parameters to obtain the dynamic characteristics. The calculation formula is as follows: ; in, Let X be the joint probability distribution of variables X and Y. p ( x )and p ( y Let X and Y be the marginal probability distributions, respectively. Then, the static features are subjected to min-max standardization, and the dynamic features are subjected to median standardization to obtain the standardized dataset.

[0024] It should be noted that, at the time synchronization level, step 200 combines the basic synchronization of the local clock calibration unit with the geological correlation benchmark of the P-wave arrival time of the microseismic monitoring data, and introduces a dynamic time warping algorithm to calculate the time offset. This can adaptively match the timing differences of different monitoring parameters caused by acquisition frequency, transmission delay, and sensor clock drift, significantly improving time alignment accuracy and avoiding the loss of correlation of multi-source data due to inconsistent time benchmarks. At the spatial mapping level, the simplification of traditional two-dimensional plane coordinates or coarse depth annotation is abandoned. A coal seam-specific three-dimensional coordinate system is constructed based on well logging data. The vertical coordinates are determined by calculating the plane coordinates and well logging depth through horizontal distance and azimuth angle. This achieves precise binding between monitoring data and coal seam geological space, solving the problem that traditional spatial representation cannot reflect the differences in the vertical parameter distribution of coal seams, and providing accurate data support for subsequent spatial risk positioning. In terms of feature extraction, a differentiated processing strategy is adopted for static and dynamic features. Static features are preserved by using neighborhood statistics combined with coefficient of variation screening to retain spatially stable and representative parameter features. Dynamic features are extracted by sliding window and mutual information to accurately capture the temporal change trend of parameters and the correlation between parameters. At the same time, targeted minimum-maximum standardization and median standardization are adopted to avoid the drawbacks of traditional uniform standardization methods that compress the extreme values ​​of static features and mask the trends of dynamic features, thus preserving the physical meaning and discriminative power of features to the maximum extent.

[0025] Specifically, step 300 first constructs constraints based on the core physical mechanism of deep coalbed methane extraction. For the adsorption characteristics of coalbed methane and the changes in reservoir permeability, the Langmuir equation and the permeability-stress correlation equation are used to construct physical rule constraints, with the following expressions: ; in, This represents the amount of coalbed methane adsorption. Where P is the maximum adsorption capacity of the coal seam, and P is the coal seam pressure. Let k be the Langmuir pressure, and k be the current penetration rate. Initial penetration rate, The permeability stress sensitivity coefficient, For the current effective stress, The initial effective stress is then used. Next, based on this physical equation, correction calculations are performed on data in the standardized dataset that do not meet the constraints. If the relationship between adsorption and pressure in a certain data point deviates from the Langmuir equation by more than 5%, or the relationship between permeability and effective stress does not follow an exponential law, then the outlier parameter values ​​in that data are recalculated using the corresponding physical equation, resulting in the initial feature set. Then, based on the scarce samples in the initial feature set—in some embodiments, permeability samples under high effective stress and adsorption samples under high pressure—a generative adversarial network is used to generate simulated samples. The distribution difference between the simulated samples and the real samples is then calculated using KL divergence, with the following formula: ; in, This represents the probability distribution of truly scarce samples in the initial feature set. The probability distribution of the generated simulated samples, The smaller the value, the closer the two distributions are. Then, the generation parameters of the simulated samples are adjusted based on the calculated KL divergence until... If the value is less than a preset threshold of 0.05, the adjusted simulated sample is added to the initial feature set to obtain an expanded feature set. Finally, the expanded feature set is reconstructed using a learning model trained with historical monitoring data covering coal seam parameter data under different mining stages and geological conditions to obtain an enhanced feature set.

[0026] Furthermore, the learning model includes: an input layer, a feature fusion layer, an attention filtering layer, and a feature reconstruction output layer. The input layer first maps each feature value in the expanded feature set to the interval [-1, 1], and the mapping formula is: ; Where x is the original value of a feature in the expanded feature set. and These are the minimum and maximum values ​​of the feature, respectively. The input features are mapped. The feature fusion layer is a two-branch fully connected network. The first branch contains two fully connected sub-layers with 2N neurons (N being the dimension of the input features) and uses the LeakyReLU activation function. The second sub-layer contains N neurons and uses the ELU activation function. The second branch contains one fully connected sub-layer with N neurons. The outputs of the two branches are concatenated according to the feature dimension to obtain the initial fused features. The attention filtering layer calculates the weights of each feature in the initial fused features through a dual attention filtering mechanism. The weight calculation formula is: w=sigmoid(W×F+b), where W is the attention weight matrix, F is the initial fused feature, b is the bias term, and the sigmoid function is used to map the weight values ​​to the [0,1] interval. The initial fused features are then multiplied by the corresponding weights to obtain the attention-filtered features. Then, the feature reconstruction output layer transforms and reconstructs the attention-filtered features through a fully connected layer, outputting an enhanced feature set with dimensions matching the initial feature set and higher data quality.

[0027] It should be noted that step 300 transforms the adsorption characteristics of coalbed methane and the reservoir permeability stress response in deep coalbed methane extraction into quantifiable constraints. The intrinsic relationship between adsorption amount and pressure is bound through the Langmuir equation, and the variation law of permeability with effective stress is defined through the permeability-stress correlation equation. This overcomes the limitations of existing technologies that rely heavily on purely data-driven models and ignore physical rationality, leading to features deviating from actual extraction scenarios. This ensures that subsequent data correction and feature reconstruction always closely match the true physical state of deep reservoirs. By performing targeted correction on outliers in the standardized dataset, rather than simply removing outliers or simple interpolation as in traditional methods, data integrity is preserved while forcing data to conform to physical laws, avoiding interference from pseudo-features in subsequent early warnings. Furthermore, the KL divergence quantification method is introduced to quantify the distribution differences between simulated and real samples. By dynamically adjusting the generation parameters, a high degree of consistency between the distribution of simulated and real samples is ensured, solving the problem of sample distribution shift and decreased generalization ability in existing algorithms. This approach is adaptable to scenarios with scarce extreme working conditions in deep coalbed methane extraction, significantly improving the working condition coverage of the feature set. Finally, the collaborative mechanism formed by the dual-branch fully connected fusion layer and the dual attention filtering layer simultaneously captures the fine dimensions and global correlations of features. The dual attention mechanism can accurately identify features with core value for early warning and assign them high weights, so that the final output enhanced feature set has physical rationality, sample representativeness and key information prominence. This provides highly reliable and highly discriminative input for subsequent spatiotemporal fusion and dynamic early warning models, improving the accuracy and robustness of early warning.

[0028] Specifically, step 400 first performs spatial network matching. Based on a three-dimensional geological model of the coal seam, constructed from borehole and well logging data and including the coal seam's layered structure, fracture development zones, and lithological distribution, the model is divided into spatial grid cells with a cube size of 10m×10m×5m. The three-dimensional coordinates of each data point in the enhanced feature set are matched with the coordinate range of the grid cells, assigning each data point to its corresponding grid cell. Simultaneously, the enhanced features of all data points within each grid cell are integrated to obtain a grid dataset with the grid as the basic unit. Next, basic features, trend features, event features, and derived features are extracted from the grid dataset. In some embodiments, the basic features are the mean and variance of coal seam pressure, temperature, and gas concentration within each grid cell; the trend features are the slope and cumulative change of each basic parameter calculated using a 1-hour sliding window and a 30-minute step size; the event features are the frequency, maximum energy value, and focal depth distribution range of microseismic events within one hour within the grid cell; and the derived features are the ratio of gas concentration to coal seam pressure and the product of the temperature change rate and microseismic energy. If the number of grid cells is M and the total feature dimension is K, then each grid cell corresponds to a 1×K feature vector. Arranging the feature vectors of all grid cells row-wise forms an M×K multidimensional feature matrix. Then, the importance of features in the multidimensional feature matrix is ​​quantized using Gini coefficients, expressed as: ; in, This represents the decrease in the Gini coefficient of the feature, i.e., the contribution value of the feature. The Gini coefficient of the parent node of the decision tree. The number of samples in the child nodes of the decision tree. The number of samples in the parent node. The Gini coefficients of the decision tree child nodes are calculated by averaging all K features. The feature importance weights are then arranged according to their feature dimensions to form a 1×K dimensional feature importance weight matrix. Next, a channel attention layer is used to first perform global average pooling on the feature importance weight matrix to obtain a 1×1 channel description vector. This vector is then processed sequentially through a fully connected layer with K / 4 neurons and a fully connected layer with K neurons to obtain the 1×K dimensional channel attention weights, expressed as: ; in, and These are the weight matrices for the two fully connected layers. This represents the bias term corresponding to a fully connected layer with K neurons. For the Sigmoid function, (·) represents the global average pooling operation. The feature importance weight matrix is ​​used. The multidimensional feature matrix is ​​processed by global max pooling and global average pooling in the spatial dimension through the spatial attention channel, resulting in two M×1 vectors, which are concatenated into an M×2 matrix. This matrix is ​​then passed through a 3×3 convolutional layer with a stride of 1 and padding of 1 to obtain M×1 dimensional spatial attention weights. The channel attention weights are then element-wise multiplied with the feature importance weight matrix, and then element-wise broadcast multiplied with the spatial attention weights to obtain updated M×K dimensional attention weights. Next, the multidimensional feature matrix is ​​element-wise multiplied with the updated attention weights to obtain a weighted feature matrix. Finally, L2 normalization is performed on the feature vector of each grid cell in this matrix to obtain a fused feature vector.

[0029] Furthermore, two trained XGBoost models capable of predicting both gas outburst and coal-rock dynamic disaster risks are employed. In some embodiments, the training data consists of historical monitoring data and corresponding risk event labels. The fused feature vectors are input into the two models, which respectively output the gas outburst risk probability and the coal-rock dynamic disaster risk probability of the grid cells, thus forming the inference results. Finally, the conflict coefficient is calculated using DS evidence theory, expressed as: ; Here, A and B are different risk propositions, and A∩B = non-empty set. For the basic probability assignment of proposition A in the gas outburst model, This represents the basic probability assignment for proposition B in the coal and rock dynamic disaster model. If... A value less than 0.5 indicates low conflict. The total risk probability is calculated directly using the following formula: ; in, for The sum of probabilities of all propositions except A in the model. for The sum of probabilities of all propositions except A in the model. If If the value is ≥0.5, it indicates high conflict. Therefore, a weighting factor is introduced to correct the inference result, expressed as follows: ; ; in, This is a weighting factor, which is 0.6 in some embodiments. This represents the average probability of all grid cells for proposition A. Finally, the comprehensive risk probability and corresponding risk level of each grid cell are obtained, which is the fusion warning result.

[0030] It should be noted that step 400 involves performing cube grid matching between the enhanced feature set and a three-dimensional geological model of the coal seam, which includes the coal seam's layered structure, fracture development zones, and lithological distribution. This ensures that each data point is accurately assigned to its corresponding physical grid cell, providing real physical spatial support for risk localization. The extracted four types of features form a multi-dimensional feature matrix, enabling a more comprehensive capture of multi-dimensional correlation information of risk precursors and avoiding the problem of missing key early warning signals due to a single feature. Furthermore, the Gini coefficient reduction of the random forest algorithm quantifies the contribution value of each feature, ensuring a precise match between the feature importance weight matrix and the feature's early warning value. Simultaneously, the dual-channel attention network strengthens the information of key grid cells, enabling the targeted extraction of the most critical information for early warning. Finally, the trained specialized model and DS evidence theory calculations improve the accuracy of risk identification, achieving deep fusion of multi-source information in the spatiotemporal dimensions. This results in fused early warning results that possess spatial accuracy, feature comprehensiveness, and inference reliability, significantly improving the accuracy and robustness of early warning for deep coalbed methane mining risks.

[0031] Specifically, step 500 uses the cubic spatial grid divided in step 400 as a basis, setting each grid unit as an independent coal seam monitoring node. Based on the fused early warning results, the comprehensive risk probability and corresponding risk level of each node are obtained. Simultaneously, based on the fracture distribution, lithological connectivity, and historical risk propagation paths in the three-dimensional geological model of the coal seam, the correlation between nodes is constructed. If two adjacent grid units have fracture connectivity or both are permeable coal seams, it is determined that there is a correlation edge between the two nodes. All monitoring nodes and correlation edges are integrated to form a coal seam monitoring node correlation graph. Next, based on a graph convolutional neural network, the comprehensive risk probability, risk level quantification value of each node, and the basic geological parameters of the grid unit are used as node feature vectors to obtain the feature matrix and adjacency matrix of the entire correlation graph. After normalizing the adjacency matrix, it is input into the convolutional layer of the graph convolutional neural network along with the feature matrix. After extracting local correlation features through 2-3 convolutional layers, the results are input into a fully connected layer to obtain the updated risk probability value of each node, thus obtaining the node risk probability distribution. Finally, the final risk probability of each node in the node risk probability distribution is compared with a preset early warning threshold. In some embodiments, the warning thresholds are divided into a low-risk threshold (0.3), a medium-risk threshold (0.6), and a high-risk threshold (0.8). If the final risk probability is <0.3, the node is determined to be in a safe state and no warning is triggered; if 0.3 ≤ final risk probability <0.6, it is determined to be in a low-risk state and a yellow warning is triggered; if 0.6 ≤ final risk probability <0.8, it is determined to be in a medium-risk state and an orange warning is triggered; if the final risk probability is ≥0.8, it is determined to be in a high-risk state and a red warning is triggered.

[0032] like Figure 2As shown, the present invention also provides a real-time monitoring and safety early warning system for multiple parameters of deep coalbed methane, comprising: The data filtering module is used to assess the acquisition quality and screen for anomalies in the raw multi-source monitoring data during the deep coalbed methane extraction process, and to obtain filtered monitoring data. The standardization module is used to unify the spatiotemporal benchmarks of the screened monitoring data to obtain a standardized dataset. The feature enhancement module is used to construct physical rule constraints in the mining process and reconstruct features of the standardized dataset based on the physical rule constraints to obtain an enhanced feature set. The risk prediction module is used to perform spatiotemporal fusion processing on the enhanced feature set to obtain fusion early warning results; The security early warning module is used to build a dynamic early warning model based on the fusion early warning results, so as to generate risk assessment results and provide real-time early warnings.

[0033] The beneficial effects of this invention are as follows: 1) Through a closed-loop process of multi-dimensional quality assessment threshold table, isolated forest anomaly identification, interpolation filling and local weighted regression repair, distorted data was eliminated and corrected, solving the problem of environmental interference and transmission error affecting monitoring data in deep coalbed methane mining, and improving the quality and reliability of monitoring data. 2) Dynamic time warping based on the arrival time of microseismic P-waves and a three-dimensional coordinate system for coal seams constructed based on well logging data, combined with the differentiated standardization of static and dynamic characteristics, achieves precise spatiotemporal unification of multi-source data, solving the problems of chaotic spatiotemporal benchmarks and poor feature adaptability in traditional data, and enabling multi-source parameters to have spatiotemporal correlation and comparability. 3) Data reconstruction based on correlation equations and KL divergence control and a dual-branch attention learning model makes up for the shortage of samples in extreme working conditions, highlights the core early warning features, avoids interference from false features, and strengthens the physical rationality and representativeness of features. 4) By using three-dimensional grid space matching, four-dimensional feature matrix and dual-channel attention fusion, and combining special early warning models with DS evidence theory for conflict correction, the system has achieved accurate identification of multiple risks such as gas outbursts and coal and rock dynamic disasters, improved the accuracy and comprehensiveness of risk early warning, and significantly reduced the probability of false alarms and missed alarms. 5) A monitoring correlation graph is constructed using a three-dimensional grid as nodes. The spatial correlation risk of nodes is mined through graph convolutional neural networks. Combined with three-level early warning thresholds, a graded response is achieved. This can accurately locate high-risk areas and dynamically track the risk propagation trend, realizing dynamic real-time early warning and risk location. 6) By identifying risks in advance and accurately defining the warning scope, ineffective shutdowns and blind inspections have been reduced. While ensuring the safety of workers, mining costs have been reduced and development efficiency has been improved, providing technical support for the efficient utilization of deep coalbed methane.

[0034] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0035] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for real-time monitoring and safety early warning of multiple parameters in deep coalbed methane, characterized in that, Includes the following steps: The original multi-source monitoring data collected during the deep coalbed methane extraction process were assessed for quality and anomaly screening were performed to obtain screened monitoring data. The selected monitoring data is subjected to spatiotemporal benchmark unification processing to obtain a standardized dataset; Construct physical rules and constraints for the mining process, and reconstruct the features of the standardized dataset based on the physical rules and constraints to obtain an enhanced feature set; The enhanced feature set is subjected to spatiotemporal fusion processing to obtain a fusion warning result; A dynamic early warning model is constructed based on the fused early warning results to provide real-time early warnings.

2. The method for real-time monitoring and safety early warning of multiple parameters of deep coalbed methane according to claim 1, characterized in that, The raw multi-source monitoring data from the deep coalbed methane extraction process were assessed for acquisition quality and anomaly screening were performed to obtain screened monitoring data, including: A quality assessment threshold table is constructed based on the missing rate, signal-to-noise ratio, and sensor data bias of the original multi-source monitoring data. The isolated forest algorithm is used to identify outliers in the original multi-source monitoring data to obtain an anomaly-labeled dataset. The abnormal marker dataset is interpolated, padded, and subjected to local weighted regression to obtain the repaired dataset; The repair dataset is filtered using the quality assessment threshold table to obtain the filtered monitoring data.

3. The method for real-time monitoring and safety early warning of multiple parameters of deep coalbed methane according to claim 1, characterized in that, The selected monitoring data is subjected to spatiotemporal benchmark unification processing to obtain a standardized dataset, including: A timestamp is added to the filtered monitoring data by a local clock calibration unit to perform time base synchronization and obtain time-stamped data. Using the arrival time of the P-wave in the microseismic monitoring data as the reference time axis, dynamic time warping calculation is performed on the time-marked data to obtain the time offset. The time-marked data is corrected for time deviation based on the time offset to obtain a time-aligned dataset. A three-dimensional coordinate system for coal seams is constructed based on the well logging data in the time-aligned dataset, and spatial coordinates are allocated and three-dimensional coordinates are mapped to the time-aligned dataset to obtain spatially labeled data. The static and dynamic features of the spatially labeled data are extracted to standardize the spatially labeled data, resulting in the standardized dataset.

4. The method for real-time monitoring and safety early warning of multiple parameters of deep coalbed methane according to claim 3, characterized in that, Extracting static and dynamic features from the spatially labeled data to standardize it, resulting in the standardized dataset, includes: The mean and extreme values ​​of the spatially labeled data are calculated using the neighborhood statistics method, and the coefficient of variation is calculated based on the mean and extreme values ​​to filter the spatially labeled data and obtain the static features. The dynamic features are obtained by extracting the temporal trend features of the spatially labeled data through a sliding window and calculating mutual information. The static features are subjected to min-max standardization, and the dynamic features are subjected to median standardization to obtain the standardized dataset.

5. The method for real-time monitoring and safety early warning of multiple parameters of deep coalbed methane according to claim 1, characterized in that, Physical rule constraints are constructed for the mining process, and the standardized dataset is reconstructed based on these physical rule constraints to obtain an enhanced feature set, including: The physical rule constraints are constructed based on the Langmuir equation and the relationship between permeability and stress during the mining process. Based on the physical equations, the data in the standardized dataset that do not conform to the physical rule constraints are corrected and calculated to obtain the initial feature set; Simulated samples are generated based on the scarce samples in the initial feature set. The distribution difference between the simulated samples and the real samples is calculated. Data augmentation is performed on the initial feature set based on the distribution difference to obtain an expanded feature set. The augmented feature set is reconstructed based on the learning model trained on historical data to obtain the enhanced feature set.

6. The method for real-time monitoring and safety early warning of multiple parameters of deep coalbed methane according to claim 5, characterized in that, The learning model includes an input layer, a feature fusion layer, an attention filtering layer, and a feature reconstruction output layer. The input layer maps the expanded feature set to the [-1, 1] interval to obtain input features. The feature fusion layer is a two-branch fully connected network. The first branch consists of two fully connected sub-layers with 2N neurons and a LeakyReLU activation function. The second sub-layer has N neurons and an ELU activation function. The second branch is a fully connected sub-layer with N neurons. The feature fusion layer fuses the input features to obtain initial fused features. The attention filtering layer uses a dual attention filtering mechanism to weight and filter the initial fused features to obtain attention-filtered features. The feature reconstruction output layer transforms and reconstructs the attention-filtered features through a fully connected layer to obtain the enhanced feature set.

7. The method for real-time monitoring and safety early warning of multiple parameters of deep coalbed methane according to claim 1, characterized in that, The enhanced feature set is subjected to spatiotemporal fusion processing to obtain a fusion warning result, including: The enhanced feature set is spatially network-matched with the three-dimensional geological model of the coal seam to obtain a grid dataset. Extract multidimensional features from the grid dataset and construct a multidimensional feature matrix; the multidimensional features include: basic features, trend features, event features, and derived features; Calculate the contribution value of each feature in the multidimensional feature matrix to obtain the feature importance weight matrix; The feature importance weight matrix is ​​updated using a dual-channel attention network to obtain the attention weights. The multidimensional feature matrix is ​​weighted and fused using the attention weights to obtain a fused feature vector; The fused feature vector is used to output risk using a specialized early warning model to obtain the inference results; The inference results are corrected and synthesized using a conflict coefficient to obtain the fusion warning result.

8. The method for real-time monitoring and safety early warning of multiple parameters of deep coalbed methane according to claim 1, characterized in that, A dynamic early warning model is constructed based on the fused early warning results to provide real-time early warning, including: Using a three-dimensional spatial grid as nodes, a coal seam monitoring node association diagram is constructed based on the fused early warning results; The risk probability distribution of the coal seam monitoring node association graph is obtained by calculating the risk probability of the nodes through a graph convolutional neural network. Security warnings are issued by comparing the node risk probability distribution with a preset warning threshold.

9. A real-time monitoring and safety early warning system for multi-parameter deep coalbed methane, characterized in that, include: The data filtering module is used to assess the acquisition quality and screen for anomalies in the raw multi-source monitoring data during the deep coalbed methane extraction process, and to obtain filtered monitoring data. The standardization module is used to perform spatiotemporal benchmark unification processing on the screened monitoring data to obtain a standardized dataset; The feature enhancement module is used to construct physical rule constraints in the mining process and reconstruct the features of the standardized dataset according to the physical rule constraints to obtain an enhanced feature set. The risk prediction module is used to perform spatiotemporal fusion processing on the enhanced feature set to obtain a fusion early warning result; The security early warning module is used to construct a dynamic early warning model based on the fused early warning results, so as to generate risk assessment results and provide real-time early warnings.

Images