Coal seam drilling gas pre-extraction state evaluation method based on semi-supervised learning

Abstract

The invention discloses a semi-supervised learning-based coal seam drilling gas pre-extraction state evaluation method, which is characterized by comprising the following steps of: 1, performing semi-supervised learning on the coal seam drilling gas pre-extraction state; the method comprises the following steps of: 1, preprocessing collected original data, and converting unstructured data into a numerical form which can be processed by a model; key features are extracted and used for providing high-quality input for a subsequent model; and 2, constructing an intelligent evaluation system fusing subjective experience and data driving, and realizing data acquisition, evaluation and feedback through real-time monitoring and model prediction. The method has the characteristics of high efficiency, high precision, low cost, dynamic optimization and robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of evaluation of the drainage state of pre-drainage boreholes in coal seams, and particularly relates to a method for evaluating the state of pre-drainage of coal seam boreholes based on semi-supervised learning. Background Art

[0002] During the pre-drainage process, manual on-site sampling and detection methods are usually used to evaluate the drainage effect. Although this method is intuitive in operation, its limitation is that it is difficult to accurately grasp the measurement time, and it cannot provide real-time feedback on the pre-drainage effect. For a long time, it may potentially pose a major hidden danger of imbalance between excavation and mining.

[0003] Traditional evaluation methods have the following disadvantages:

[0004] 1. Lack of single borehole evaluation: Existing methods focus on the overall drainage efficiency and ignore the need for fine control of single boreholes;

[0005] 2. Insufficient multi-parameter fusion: Over-reliance on a single index of gas concentration, without comprehensively considering time-series characteristics such as negative pressure, temperature, and historical data;

[0006] 3. Poor real-time and dynamic performance: Traditional models are difficult to adapt to the spatial heterogeneity of gas occurrence and the dynamic changes in the drainage process.

[0007] Supervised learning models (such as KELM, DNN) and unsupervised learning models (such as K-Means) are the mainstream methods for current gas drainage evaluation.

[0008] However, they have the following disadvantages:

[0009] 1. Supervised learning trains the model through labeled data and can achieve classification of the drainage effect, but requires a large number of manually labeled samples, and its generalization ability is limited when the sample size is insufficient;

[0010] 2. Unsupervised learning (such as K-Means) mines potential patterns through data clustering and does not require labeling, but cannot directly give qualitative evaluation results and needs to be interpreted in combination with manual experience, resulting in strong subjectivity in evaluation. Summary of the Invention

[0011] In order to overcome the above technical problems, the purpose of the present invention is to provide a method for evaluating the state of pre-drainage of coal seam boreholes based on semi-supervised learning, which has the characteristics of high efficiency, high precision, dynamic optimization, and robustness.

[0012] The technical solution adopted by the present invention is:

[0013] A method for evaluating the state of pre-drainage of coal seam boreholes based on semi-supervised learning, comprising the following steps;

[0014] Step 1: Preprocess the obtained original data, convert unstructured data (such as categorical variables) into a numerical form that can be processed by the model; extract key features to provide high-quality input for the subsequent model.

[0015] Step 2: Build an intelligent evaluation system that integrates subjective experience and data-driven approach, and through real-time monitoring and model prediction, achieve data collection, evaluation, and feedback.

[0016] The specific content of Step 1 is as follows:

[0017] Step 1.1: Collect data by combining the refined layout control system for gas pre-drainage boreholes in coal seam working faces and manual collection.

[0018] Step 1.2: Preprocess the obtained data.

[0019] The system in Step 1.1 includes an intelligent control system for gas drainage pipelines, a measurement and monitoring system, and a data transmission and processing system.

[0020] The intelligent control system for gas drainage pipelines is used to dynamically adjust the drainage negative pressure, optimize the pipeline layout and valve opening, and balance the extraction flow of each borehole in real time through intelligent control algorithms to ensure the efficient extraction of boreholes in high gas areas, while avoiding air mixing in low gas areas due to excessive negative pressure.

[0021] The measurement and monitoring system is used to collect key parameters of gas extraction in multiple dimensions in real time, including dynamic data such as methane concentration, extraction negative pressure, ambient temperature, and working condition flow, as well as static attributes such as borehole location and sealing quality; through a distributed sensor network and high-precision instruments, achieve refined monitoring of single boreholes and pipeline network systems, provide a real and reliable data source for the evaluation model, and at the same time provide an intuitive visualization interface of the extraction status for on-site engineers.

[0022] The data transmission and processing system is used to efficiently integrate multi-source heterogeneous data, and transmit the collected data to the central server in real time through industrial Ethernet and wireless communication technology, and complete data cleaning, standardization, and feature engineering processing.

[0023] The specific content of Step 1.2 is

[0024] (1) Data conversion based on One-hot encoding:

[0025] When processing non-numerical data obtained during the data collection process, it is required that categorical variables be input in numerical form. One-hot encoding creates independent columns for each category by converting each value of the categorical variable into a binary form, eliminating the ordinal relationship between categorical variables.

[0026] (2) Normalization:

[0027] Adopt the MinMaxScaler normalization method, that is, subtract the minimum value from each data value and then divide by the difference between the maximum value and the minimum value, so as to map the data into the range of [0,1]. The formula for min-max normalization is as follows:

[0028]

[0029] Among them, X is the original data, X max and X min are the maximum and minimum values of the features respectively.

[0030] Specifically in the step (1):

[0031] Mark the categorical variables, and then generate the corresponding binary representation through the one-hot encoder. The number of columns of the one-hot encoding of each categorical variable is equal to the number of its possible values;

[0032] Merge the columns of the binary representation with the original data set to form a new data structure suitable for model input.

[0033] Specifically, the step 2 is as follows:

[0034] Step 2.1: Establish the AHP-FEM evaluation system. Quantify the expert experience through the Analytic Hierarchy Process (AHP), and combine the Fuzzy Evaluation Method (FEM) to construct the classification criteria for the drainage effect (excellent / good / fair / poor), providing a subjective label basis for the model; provide a small number of labeled samples for the subsequent semi-supervised learning model to solve the problem that supervised learning relies on a large number of manual annotations.

[0035] Step 2.2: Construct a Gaussian mixture model to provide a basic clustering tool for semi-supervised learning. The probability soft assignment characteristic of GMM and the fast convergence of K-Means complement each other, laying an algorithmic foundation for the subsequent hybrid model design;

[0036] Step 2.3: K-Means clustering algorithm;

[0037] Step 2.4: The proposed optimized semi-supervised evaluation model; propose the AHP-FEM-SSGMM / SSK-Means model, combine the labels generated by AHP-FEM with GMM / K-Means clustering, and guide the clustering direction through a small number of labeled samples; provide a basic clustering tool for semi-supervised learning. The probability soft assignment characteristic of GMM and the fast convergence of K-Means complement each other, laying an algorithmic foundation for the subsequent hybrid model design;

[0038] Step: 2.5 PCA optimization algorithm. Reduce the dimension of the 8-dimensional original data, reduce the computational complexity and retain the key features, improve the clustering efficiency and stability of GMM / K-Means in high-dimensional data, and ensure the feasibility of the model in actual engineering applications.

[0039] Step 2.1 is specifically as follows:

[0040] (1) Selection of evaluation indicators

[0041] A series of factors are selected as the evaluation indicators for the drainage boreholes. According to the different data sources, the weight calculation methods are mainly divided into three categories: subjective method, objective method, and comprehensive method;

[0042] (2) Weight calculation

[0043] The hierarchical structure model is used to decompose the complex evaluation problem of the coal seam borehole drainage state into a hierarchical system that can be quantitatively analyzed, assign quantitative weights to the evaluation indicators, and provide a marking basis and support for the subsequent fuzzy evaluation system.

[0044] (3) Establishment of the evaluation label system.

[0045] In the hierarchical structure model of step (2), the top layer is the target layer, followed by the criterion layer, sub-criterion layer, and the bottom layer is the specific scheme layer; the elements in each layer form a judgment matrix through pairwise comparison, and the weight is calculated using the eigenvector method. For the indicators in the same layer, pairwise comparison is performed to form a judgment matrix, and industry experts are invited to score the comparison results according to the 9-scale method;

[0046] AHP ensures the rationality of the expert scoring through consistency testing. First, normalize each vector of the judgment matrix:

[0047]

[0048] represents the element in the i-th row and j-th column of the normalized judgment matrix. m ij represents the element in the i-th row and j-th column of the original judgment matrix. n represents the order of the judgment matrix, that is, the number of indicators participating in the comparison. i and j represent the i-th row and j-th column respectively

[0049] Add up the rows of the matrix obtained after normalization:

[0050]

[0051] Calculate the average value of each row of the normalized matrix to obtain a column vector, which is the weight vector;

[0052]

[0053] Calculate the maximum eigenvalue and the corresponding eigenvector of the judgment matrix. The maximum eigenvalue is obtained through the following formula:

[0054]

[0055] Among them, n is the order of the judgment matrix, A is the judgment matrix, w is the normalized weight vector, and (Aw) i is the i-th element of the Aw vector, and w i is the i-th element of the w vector. The eigenvector is the normalized weight vector, which can be obtained by the square root method or the sum method;

[0056] Calculate the consistency index CI of the judgment matrix, and its definition is:

[0057]

[0058] The closer CI is to 0, the better the consistency of the judgment matrix. CI = 0 indicates that the judgment matrix is completely consistent. Then, calculate the consistency ratio CR of the judgment matrix, and its definition is:

[0059]

[0060] RI (random consistency index) is calculated based on the average CI value of randomly generated judgment matrices and is related to the order of the judgment matrix. The smaller the CR value, the better the consistency of the judgment matrix.

[0061] Usually, when CR < 0.1, it is considered that the consistency of the judgment matrix is acceptable; otherwise, the judgment matrix needs to be corrected. Use the analytic hierarchy process to calculate the weights of the evaluation indexes for the drainage effect of coal seam pre-drainage boreholes;

[0062] The specific content of step (3) is as follows:

[0063] In AHP, the relative relationships of evaluation indexes usually show an increasing or decreasing trend;

[0064] Formulas (8) and (9) are the calculation formulas for the descending semi-trapezoidal model and the ascending semi-trapezoidal model respectively. μ(X i ) represents the membership degree value of the i-th evaluation index X i . In the descending semi-trapezoidal model, the larger its value, the smaller the contribution rate to satisfaction. In the ascending semi-trapezoidal model, the opposite is true;

[0065] The mathematical expression of the trapezoidal membership function is:

[0066]

[0067] Among them, X i represents the specific data value of the i-th evaluation index, x imax represents the maximum value of the i-th evaluation index in the dataset, x imin represents the minimum value of the i-th evaluation index in the dataset, and μ(X i ) represents the membership degree value of the i-th evaluation index X i ;

[0068] The satisfactory value for the evaluation of the drainage effect of the pre-drainage boreholes for coal seam gas is expressed as:

[0069]

[0070] The comment set of the satisfactory value is divided into four categories, namely "excellent, good, medium, poor".

[0071] The specific content of step 2.2 is as follows:

[0072] It is assumed that each sample in the dataset is randomly generated from one of K Gaussian distributions, where K is the number of clusters, each Gaussian distribution corresponds to a cluster, and its parameters are the mean μ k , variance Σ k and weight π k , satisfying The GMM is expressed as:

[0073]

[0074] where, N(x|μ k ,Σ k ) is the probability density function of the Gaussian distribution, expressed as:

[0075]

[0076] where, d is the dimension of the sample, ∣Σ k ∣ is the determinant of the variance matrix;

[0077] This posterior probability is expressed as:

[0078]

[0079] where, p(k∣x) is the conditional probability of the sample given the cluster, p(k) is the prior probability of the cluster, p(x) is the marginal probability of the sample, and this posterior probability is also called the latent variable, denoted as γ nk , representing the probability that the nth sample belongs to the kth cluster.

[0080] The calculation steps of GMM are as follows:

[0081] The number of clusters k, the initial parameters μ k , Σ k and π k , N is the number of samples, X = {x1, x2,..., x N} is the dataset, is the parameter set.

[0082] The specific content of step 2.3 is as follows:

[0083] The goal of K-Means is to find a set of cluster centers such that the sum of the distances between each sample and its corresponding cluster center is minimized, i.e., minimizing the following objective function:

[0084]

[0085] where k is the number of clusters, C i is the i-th cluster, μ i is the center of the i-th cluster, ||x - μ i || is the Euclidean distance between the sample x and the cluster center μ i .

[0086] Specifically, step 2.4 is as follows:

[0087] Cluster the mixed samples through GMM and K-Means clusterers, map the information of the labeled samples to the unlabeled samples, generate SSGMM and SSK-Means classifiers, and achieve the classification of the mixed samples. Use the SSGMM classifier to evaluate the quality and quantity of the samples screened by AHP-FEM. Through the automated learning process, obtain the optimal input samples and high-precision classifiers.

[0088] Specifically, step 2.5 is as follows:

[0089] There are a total of 8 evaluation indicators for the pre-drainage data of coal seam boreholes. However, in a high-dimensional space, the distance between data points becomes sparser, which will lead to a decline in the performance of the clustering algorithm. By performing dimensionality reduction processing through the PCA optimization algorithm, redundant features in the data can be removed, improving the ability of the clustering algorithm to grasp the internal structure of the data. In addition, high-dimensional data is usually accompanied by more computational overhead, and dimensionality reduction can significantly reduce the computational complexity and improve the efficiency of the algorithm.

[0090] The dimensionality reduction method is operated as follows.

[0091] (1) Assume that the original data is an n×d matrix X, where n is the number of samples, d is the dimension, and the target dimension is k, satisfying k < d.

[0092] (2) Make the mean and variance of each feature become 0 and 1 respectively, so that the data has the same scale on different features.

[0093]

[0094] where X ij represents the element in the i-th row and j-th column of the data matrix, represents the mean of the j-th column, and σ j represents the standard deviation of the j-th column.

[0095] (3) Calculate the covariance matrix S of the data.

[0096]

[0097] Among them, X T represents the transpose of the data matrix, S is a d×d symmetric matrix, the diagonal elements of the covariance matrix represent the variance of each feature, and the off-diagonal elements represent the covariance between two features.

[0098] (4) Calculate the eigenvalues and eigenvectors of the covariance matrix:

[0099] Sv = λv (17)

[0100] where v is an eigenvector of the covariance matrix and λ is the corresponding eigenvalue.

[0101] (5) Select the eigenvectors corresponding to the top k largest eigenvalues as the new coordinate axes of the data, and form a d×k projection matrix W.

[0102] W = [v1, v2,..., v k (18)

[0103] (6) Transform the data from the original coordinate system to the new coordinate system, that is, multiply the data by the projection matrix to obtain a reduced-dimensional data matrix Y of n×k.

[0104] Y = XW (19)

[0105] Among them, each row of Y represents the representation of a sample in the new coordinate system, and each column of Y represents the representation of a principal component in the new coordinate system. Using PCA for dimensionality reduction can map high-dimensional data to a low-dimensional space for visualization.

[0106] AHP-FEM provides prior knowledge (weights, labels) for the model to guide the clustering direction;

[0107] GMM / K-Means mines feature associations through data-driven methods to make up for the subjective limitations of AHP-FEM;

[0108] PCA optimizes the input quality, enhances the robustness of the clustering algorithm, reduces the computational complexity, and improves the computational efficiency.

[0109] Advantages of the present invention:

[0110] 1. Strengthen the coal seam gas safety management system: significantly improve the scientificity and timeliness of single borehole drainage regulation, and provide an effective technical path for the improvement of coal seam gas drainage efficiency and gas drainage regulation.

[0111] 2. Efficiency improvement: reduce the manual annotation workload by more than 90%, and the processing speed is 4-6 times faster than the traditional method.

[0112] 3. Precision Optimization: In classification tasks, the clustering degree is better. From the examples, the MVCR index reaches 82.64% and the ARI reaches 85.83%, which is a 15 - 20% improvement compared to single clustering algorithms.

[0113] 4. Reduce Labeling Costs: Utilize a small amount of labeled data and a large amount of unlabeled data to reduce the need for manual labeling.

[0114] 5. Improve Model Generalization Ability: By mining the internal structure of the data, optimize the classification boundary and enhance the adaptability to unknown samples.

[0115] 6. Applicable to Complex Scenarios: When the data distribution is irregular or there is noise, it is more robust than pure supervised or unsupervised methods. Description of the Drawings

[0116] Figure 1 It is a schematic diagram of the system of the present invention.

[0117] Figure 2 It is a schematic diagram of the One - hot encoding of the present invention.

[0118] Figure 3 It is a schematic diagram of the architecture of the hybrid optimization model of the present invention.

[0119] Figure 4 It is a visualization diagram of the GMM drilling clustering result.

[0120] Figure 5 It is a visualization diagram of the K - Means drilling clustering result.

[0121] Figure 6 It is a drilling evaluation diagram of the SSGMM algorithm.

[0122] Figure 7 It is a drilling evaluation diagram of the SSK - Means algorithm. Detailed Embodiment

[0123] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0124] A method for evaluating the pre - drainage gas state of coal seam boreholes based on semi - supervised learning includes the following steps;

[0125] 1 Data setting; The purpose and effect of this step are to eliminate noise, outliers, and redundant information in the original data, ensuring data quality; unify the dimensions of different features (such as temperature, pressure, flow rate, etc.) to avoid model bias caused by dimensional differences; convert unstructured data (such as categorical variables) into a numerical form that can be processed by the model (such as One-hot encoding); extract key features (such as historical drainage data) to enhance the model's ability to capture temporal patterns and uneven gas distribution. This is used to provide high-quality input for subsequent models, support refined evaluation, and enhance the model's generalization ability.

[0126] 1.1 Data source

[0127] Data is collected by combining the refined layout control system for gas pre-drainage boreholes in coal seams and manual collection. This system consists of three parts: an intelligent control system for gas drainage pipelines, a measurement and monitoring system, and a data transmission and processing system. It can monitor and control various parameters of gas drainage in real time, selectively extract target pipeline data, and improve the drainage efficiency and safety. The system schematic diagram is as Figure 1 shown.

[0128] The intelligent control system for gas drainage pipelines is used to dynamically adjust the drainage negative pressure, optimize the pipeline layout and valve opening, and balance the drainage flow of each borehole in real time through intelligent control algorithms to ensure the efficient drainage of boreholes in high-gas areas. At the same time, it avoids air mixing in low-gas areas due to excessive negative pressure. This system integrates pressure sensors and electric control valves, which can automatically adjust the drainage parameters according to the difference in coal seam permeability, reduce the risk of pipeline blockage, and improve the uniformity and continuity of gas drainage.

[0129] The measurement and monitoring system is used to collect key parameters of gas drainage in multiple dimensions in real time, including dynamic data such as methane concentration, drainage negative pressure, ambient temperature, and working condition flow rate, as well as static attributes such as borehole location and sealing quality. Through a distributed sensor network and high-precision instruments, it realizes the refined monitoring of single boreholes and pipeline network systems, provides a real and reliable data source for the evaluation model, and at the same time provides an intuitive visualization interface of the drainage status for on-site engineers.

[0130] The data transmission and processing system is used to efficiently integrate multi-source heterogeneous data, and transmit the collected data to the central server in real time through industrial Ethernet and wireless communication technology, and complete data cleaning, standardization, and feature engineering processing. This system supports historical data storage and quick query, provides time-series feature training samples for semi-supervised learning models, and reveals gas drainage laws through data mining technology to provide data support for optimization decisions, ultimately achieving a double improvement in drainage efficiency and safety.

[0131] 1.2 Data preprocessing

[0132] (1) Data conversion based on One-hot encoding

[0133] When dealing with finitely enumerable non-numerical data (non-numerical data obtained during the data acquisition process), the model usually requires categorical variables to be input in numerical form. One-hot encoding creates independent columns for each category by converting each value of the categorical variable into a binary form, eliminating the ordinal relationship between categorical variables. First, the categorical variables are labeled, and then the corresponding binary representations are generated through a one-hot encoder. The number of one-hot encoded columns for each categorical variable is equal to the number of its possible values. Finally, these columns of binary representations are merged with the original dataset to form a new data structure suitable for model input. A schematic diagram of one-hot encoding is shown as Figure 2 shown below.

[0134] (2) Normalization

[0135] This paper adopts the method of MinMaxScaler normalization, that is, subtracting the minimum value from each data (data acquisition) value and then dividing by the difference between the maximum value and the minimum value, so as to map the data into the range of [0,1]. The formula for maximum-minimum normalization is as follows:

[0136]

[0137] where X is the original data, X max and X min are the maximum and minimum values of the feature respectively.

[0138] 2 Model establishment. The purpose and effect of this step are to construct an intelligent evaluation system that integrates subjective experience and data-driven: ① Solve the limitations of traditional evaluation: By combining the Analytic Hierarchy Process (AHP) and the Fuzzy Evaluation Method (FEM), quantify expert experience and reduce subjective judgment bias; ② Break through the annotation dependence of supervised learning: Use semi-supervised learning (SSGMM / SSK-Means) to integrate a small amount of labeled data and a large amount of unlabeled data, reducing the manual annotation cost; ③ Capture the internal laws of data: Through the Gaussian Mixture Model (GMM) and the K-Means clustering algorithm, mine the spatio-temporal correlations of multi-dimensional parameters (such as negative pressure, concentration, historical flow, etc.) to adapt to the characteristics of uneven coal seam gas distribution.

[0139] It is used for ① Supporting refined regulation: Provide a hierarchical evaluation (excellent / good / medium / poor) for the single-hole gas drainage state to guide targeted on-site measures; ② Ensuring safe production: Through real-time monitoring and model prediction, early detect potential hazards such as air leakage and blocked holes, reducing the risk of gas overrun; ③ Promoting intelligent management: The model can be embedded in the gas drainage system to realize the closed-loop of "data acquisition - evaluation - feedback", providing a scientific decision-making tool for coal mine gas control.

[0140] 2.1 Establishment of the AHP-FEM evaluation system

[0141] (1) Selection of evaluation indicators

[0142] After fully considering the types of monitoring data and their inherent time-series characteristics, this study uses a method that combines on-site investigation and theoretical analysis, makes full use of the characteristics of monitoring data, ensures the operability of the selected evaluation indicators, and is consistent with conventional monitoring practices. At the same time, the selection of evaluation indicators is based on the data range that the existing monitoring system can provide to ensure its practicality and effectiveness. Based on this, a series of factors are selected as the evaluation indicators for drainage boreholes, and the detailed list is shown in Table 1. According to the different data sources, the weight calculation methods are mainly divided into three categories: subjective method, objective method, and comprehensive method.

[0143] Table 1 Evaluation indicators for borehole drainage effect

[0144]

[0145] (2) Weight calculation

[0146] AHP decomposes the problem into multiple levels and constructs a hierarchical structure model. The hierarchical structure model is used to decompose the complex evaluation problem of coal seam borehole drainage state into a hierarchical system that can be quantitatively analyzed, quantitatively assign weights to evaluation indicators, and provide a basis and support for the subsequent fuzzy evaluation system.

[0147] In this model, the top layer is the goal layer, followed by the criterion layer, sub-criterion layer, and the bottom layer is the specific scheme layer. The elements in each layer form a judgment matrix through pairwise comparison, and the weight is calculated using the eigenvector method. For the indicators at the same level, pairwise comparison is carried out to form a judgment matrix, and industry experts are invited to score the comparison results according to the 9-scale method. The scoring rules are shown in Table 2.

[0148] Table 2 AHP scoring rules

[0149]

[0150] AHP fully considers the decision maker's subjective perception of the problem and ensures the rationality of expert scoring through consistency testing. First, normalize each vector of the judgment matrix:

[0151]

[0152] represents the element in the i-th row and j-th column of the normalized judgment matrix. m ij represents the element in the i-th row and j-th column of the original judgment matrix. n represents the order of the judgment matrix, that is, the number of indicators participating in the comparison. i and j represent the i-th row and j-th column respectively;

[0153] Add up the rows of the matrix obtained after normalization:

[0154]

[0155] Calculate the average value of each row of the normalization matrix to obtain a column vector, which is the weight vector.

[0156]

[0157] Calculate the maximum eigenvalue and the corresponding eigenvector of the judgment matrix. The maximum eigenvalue can be obtained through the following formula:

[0158]

[0159] where n is the order of the judgment matrix, A is the judgment matrix, w is the normalized weight vector, (Aw) i is the i-th element of the Aw vector, and w i is the i-th element of the w vector. The eigenvector is the normalized weight vector, which can be obtained by the square root method or the sum method.

[0160] Calculate the consistency index CI of the judgment matrix, and its definition is:

[0161]

[0162] The closer CI is to 0, the better the consistency of the judgment matrix. CI = 0 indicates that the judgment matrix is completely consistent. Next, calculate the consistency ratio CR of the judgment matrix, and its definition is:

[0163]

[0164] RI (random consistency index) is calculated based on the average CI value of randomly generated judgment matrices and is related to the order of the judgment matrix. The smaller the CR value, the better the consistency of the judgment matrix. Usually, when CR < 0.1, it is considered that the consistency of the judgment matrix is acceptable; otherwise, the judgment matrix needs to be corrected. Using the analytic hierarchy process to calculate the weights of the evaluation indexes for the drainage effect of coal seam pre-drainage boreholes, the results are shown in Tables 3 and 4. It can be seen from the table that after calculation, the maximum eigenvalue λ max = 8.433, the consistency index CI = 0.061. Looking up the table, the consistency ratio RI = 1.41. Finally, it can be calculated that CR = 0.043 < 0.1, and the consistency of this judgment matrix is good.

[0165] Table 3 Judgment Matrix, Weights and Consistency Test of Evaluation Indexes

[0166]

[0167]

[0168] Table 4 Values of Random Consistency Index RI

[0169]

[0170] (3) Establishment of Evaluation Label System

[0171] In order to more flexibly reflect the fuzzy attribution relationship between evaluation indicators in AHP, a semi-trapezoidal distribution function is introduced when determining weights to reduce the absolute subjectivity of expert judgment. In AHP, the relative relationship of evaluation indicators usually shows an increasing or decreasing trend, and the semi-trapezoidal distribution function can effectively simulate this trend to construct a more reliable and interpretable evaluation system.

[0172] Formulas (8) and (9) are the calculation formulas for the descending semi-trapezoidal model and the ascending semi-trapezoidal model respectively. μ(X i ) represents the membership degree value of the i-th evaluation indicator X i . In the descending semi-trapezoidal model, the larger its value, the smaller the contribution rate to satisfaction. In the ascending semi-trapezoidal model, the opposite is true.

[0173] The mathematical expression of the trapezoidal membership function is as follows:

[0174]

[0175] The evaluation satisfactory value of the coal seam gas pre-drainage borehole drainage effect is expressed as:

[0176]

[0177] The comment set of the satisfactory value is divided into four categories, namely "excellent, good, medium, poor". The specific classification criteria are as follows:

[0178] Excellent: P ∈ (0.75, 1]. The drainage effect of the coal seam pre-drainage borehole is excellent, the pre-drainage efficiency is high, the borehole spacing is reasonable, the sealing state is good, there is no air leakage, the drainage pipeline state is stable, the negative pressure setting and pipeline layout are reasonable, the drainage concentration is high, and no adjustment is required.

[0179] Good: P ∈ (0.5, 0.75]. The drainage effect of the coal seam pre-drainage borehole is good, the pre-drainage efficiency is relatively high, all indicators meet the national requirements, the sealing state is good, the negative pressure setting is reasonable, and the drainage concentration reaches the standard. It is recommended to check whether there is air leakage in the borehole and adjust the drainage negative pressure appropriately according to the site conditions.

[0180] Medium: P ∈ (0.25, 0.5]. The drainage effect of the coal seam pre-drainage borehole is average, the pre-drainage efficiency is low, there is air leakage, the negative pressure setting needs to be adjusted, the drainage concentration is low, the borehole layout spacing is unreasonable, and relevant parameters need to be optimized.

[0181] Difference: P ∈ [0, 0.25], the extraction effect of the pre-drainage boreholes in the coal seam is poor, the pre-drainage efficiency is extremely low, and the extraction concentration is seriously insufficient. It is urgent to comprehensively adjust the boreholes and extraction parameters.

[0182] 2.2 Gaussian Mixture Model

[0183] The Gaussian Mixture Model (GMM) is a probability-based clustering algorithm that assumes that each cluster follows a Gaussian distribution, and the dataset is generated by a linear combination of these Gaussian distributions.

[0184] Table 5 GMM algorithm flow chart

[0185]

[0186]

[0187] Assume that each sample in the dataset is randomly generated by one of K Gaussian distributions, where K is the number of clusters. Each Gaussian distribution corresponds to a cluster, and its parameters are the mean μ k , variance Σ k and weight π k , satisfying Therefore, GMM can be expressed as:

[0188]

[0189] where, N(x|μ k , Σ k ) is the probability density function of the Gaussian distribution, expressed as:

[0190]

[0191] where, d is the dimension of the sample, and |Σ k | is the determinant of the variance matrix.

[0192] An important feature of GMM is that it not only gives the maximum likelihood estimate of which cluster each sample belongs to, but also gives the posterior probability of each sample belonging to each cluster, that is, soft assignment. This posterior probability can be expressed as:

[0193]

[0194] where, p(k|x) is the conditional probability of the sample given the cluster, p(k) is the prior probability of the cluster, and p(x) is the marginal probability of the sample. This posterior probability is also called the latent variable, denoted as γ nk , representing the probability that the nth sample belongs to the kth cluster. The calculation steps of GMM are as follows:

[0195] The number k of clusters, and the initial parameters μ of each cluster k , Σ k and π k , N is the number of samples, X = {x1, x2,..., x N} is the data set, is the parameter set. The logical pseudocode of GMM is organized as shown in Table 5.

[0196] 2.3 K-Means clustering algorithm

[0197] Table 6 K-Means algorithm flow chart

[0198]

[0199] K-Means is a partitioning-based clustering algorithm that assumes that each cluster is composed of a center point and the samples around it, and each sample belongs to only one cluster. The goal of K-Means is to find a set of cluster centers such that the sum of the distances between each sample and its belonging cluster center is minimized, that is, to minimize the following objective function:

[0200]

[0201] where k is the number of clusters, C i is the i-th cluster, μ i is the center of the i-th cluster, ||x - μ i || is the Euclidean distance between the sample x and the cluster center μ i . Table 6 is the K-Means algorithm flow chart.

[0202] 2.4 The proposed optimized semi-supervised evaluation model

[0203] Due to the lack of labeled samples, the traditional GMM and K-Means algorithms lead to unclear training objectives and unintuitive clustering results, and rely on human judgment of the distribution of the feature space. This study proposes a hybrid semi-supervised learning model that combines the analytic hierarchy process and the fuzzy evaluation method to label the features of a small number of samples and mix them with unlabeled samples. The mixed samples are clustered by GMM and K-Means clusterers, and the information of the labeled samples is mapped to the unlabeled samples to generate SSGMM and SSK-Means classifiers to achieve the classification of the mixed samples. The SSGMM classifier is used to evaluate the quality and quantity of the samples screened by AHP-FEM. Through the automated learning process, the optimal input samples and high-precision classifiers are obtained.

[0204] 2.5 PCA optimization algorithm

[0205] There are a total of 8 evaluation indicators for the pre-drainage data of coal seam boreholes. However, in a high-dimensional space, the distance between data points becomes sparser, which will lead to a decline in the performance of clustering algorithms. By performing dimensionality reduction through the PCA optimization algorithm, redundant features in the data can be removed, improving the clustering algorithm's ability to grasp the internal structure of the data. In addition, high-dimensional data is usually accompanied by more computational overhead, and dimensionality reduction can significantly reduce the computational complexity and improve the efficiency of the algorithm. The dimensionality reduction method is operated as follows.

[0206] (1) Assume that the original data is an n×d matrix X, where n is the number of samples, d is the dimension, and the target dimension is k, satisfying k < d.

[0207] (2) Make the mean and variance of each feature become 0 and 1 respectively, so that the data has the same scale on different features.

[0208]

[0209] Among them, X ij represents the element in the i-th row and j-th column of the data matrix, represents the mean of the j-th column, and σ j represents the standard deviation of the j-th column.

[0210] (3) Calculate the covariance matrix S of the data.

[0211]

[0212] Among them, X T represents the transpose of the data matrix. S is a d×d symmetric matrix. The diagonal elements of the covariance matrix represent the variance of each feature, and the off-diagonal elements represent the covariance between two features.

[0213] (4) Calculate the eigenvalues and eigenvectors of the covariance matrix:

[0214] Sv = λv (17)

[0215] Among them, v is an eigenvector of the covariance matrix, and λ is the corresponding eigenvalue.

[0216] (5) Select the eigenvectors corresponding to the top k largest eigenvalues as the new coordinate axes of the data, and form a d×k projection matrix W.

[0217] W = [v1, v2,..., v k (18)

[0218] (6) Transform the data from the original coordinate system to the new coordinate system, that is, multiply the data by the projection matrix to obtain a reduced-dimensional data matrix Y of n×k.

[0219] Y = XW (19)

[0220] Among them, each row of Y represents the representation of a sample in the new coordinate system, and each column of Y represents the representation of a principal component in the new coordinate system. Using PCA dimensionality reduction processing to map high-dimensional data to a low-dimensional space is convenient for visualization.

[0221] Taking the 215 working face of Huangling No. 2 Coal Mine as the research object, data collection is carried out by combining the refined layout control system of gas pre-drainage boreholes in the coal seam working face and manual collection. The research selects the drainage data of boreholes #325 - #425 on March 2 of the 215 working face, and takes the drainage data on March 1 as the background for comparative analysis. The schematic diagram of the system structure is as Figure 1 shown. Using AHP-FEM to perform feature marking on 10 gas drainage samples, mixing them with the remaining 91 unmarked samples, after optimization processing using PCA, clustering analysis is carried out using the SSGMM and SSK-Means algorithms respectively, and the final clustering results are as Figure 4 、 5 、6, and 7 shown.

[0222] The results show that the SSGMM and SSK-Means clustering algorithms divide the borehole drainage data into four clusters, which are represented by different colors and symbols. The clustering result of SSGMM shows a smaller shaded area and better clustering density compared with SSK-Means. Among them, cluster 1 (blue) is the densest, and cluster 2 (pink) is the most dispersed, which may be affected by noise or outliers. The GMM probability-based clustering algorithm can capture different shapes and densities of data points, but has a high computational complexity; K-Means is based on distance, with small computational volume but is not suitable for non-circular or spherical distributions and is easily affected by the initial value. Using AHP and FEM to carry out manual annotation and evaluation on the clustering results based on the GMM and K-Means algorithms, the drainage effect evaluation of different clusters of boreholes is obtained. The evaluation results of the two algorithms are highly consistent, indicating that the clustering method can effectively distinguish the drainage effects of boreholes. After calculation, in terms of the excellent rate of the drainage effect of the test boreholes, the SSGMM algorithm is 59.41%, and the SSK-Means algorithm is 74.26%, and the overall drainage effect is good. The similarity rate analysis for the excellent, good, medium, and poor comment sets shows that the similarity rate of the "excellent" comment set is the lowest (46.67%), and the similarity rate of the "poor" comment set is the highest (100%). This indicates that the two algorithms have high accuracy in distinguishing between boreholes with better and worse drainage effects, but there are differences in distinguishing excellent drainage effects, perhaps due to the different numbers of clustering results.

Claims

1. A method for evaluating the state of pre-extraction gas in coal seam drilling based on semi-supervised learning, characterized in that: The steps include: Step 1: Preprocess the collected raw data and convert the unstructured data into a numerical form that can be processed by the model; extract Key features, used to provide high-quality input for subsequent models; Step 2: Build an intelligent evaluation system that integrates subjective experience and data-driven, and realize data collection, evaluation, and feedback through real-time monitoring and model prediction.

2. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 1, characterized in that: The step 1 is specifically as follows: Step 1.1: Data collection is carried out by combining the refined control system of gas pre-extraction drilling holes in coal seam working faces with manual collection; Step 1.2: Preprocess the acquired data; The system obtained in step 1.1 includes a gas extraction pipeline intelligent management and control system, a measurement and monitoring system, and a data transmission and processing system; The gas extraction pipeline intelligent control system is used to dynamically adjust the extraction negative pressure, optimize the pipeline layout and valve opening, and balance the extraction flow of each borehole in real time through an intelligent control algorithm to ensure efficient extraction of boreholes in high-gas areas, while avoiding air mixing in low-gas areas due to excessive negative pressure; The measurement and monitoring system is used to collect key gas extraction parameters in real time in multiple dimensions, including dynamic data of methane concentration, extraction negative pressure, ambient temperature, and operating flow rate, as well as static properties of drilling position and sealing quality; Through distributed sensor networks and high-precision instruments, refined monitoring of single holes and pipe network systems is achieved, providing a real and reliable data source for the evaluation model, while providing an intuitive drainage status visualization interface for field engineers; The data transmission and processing system is used to efficiently integrate multi-source heterogeneous data, transmit the collected data to the central server in real time through industrial Ethernet and wireless communication technology, and complete data cleaning, standardization and feature engineering processing; The step 1.2 is specifically as follows: (1) Data conversion based on One-hot encoding: When processing non-numerical data obtained during data collection, categorical variables are required to be input in numerical form. One-hot encoding converts each value of the categorical variable into binary form, creates independent columns for each category, and eliminates the order relationship between categorical variables; (2) Normalization: The MinMaxScaler normalization method is used, that is, each data value is subtracted from the minimum value, and then divided by the difference between the maximum and minimum values, so as to map the data to the range of [0,1]. The formula for maximum and minimum normalization is as follows: Among them, X is the original data, X max and X min are the maximum and minimum values ​​of the feature, respectively.

3. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 2 is characterized in that: The step (1) specifically comprises: The categorical variables are labeled and then the corresponding binary representation is generated through a one-hot encoder. The number of one-hot encoding columns for each categorical variable is equal to the number of its possible values. Merge the binary representation of the columns with the original dataset to form a new data structure suitable for model input.

4. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 1, characterized in that: The step 2 is specifically as follows: Step 2.1: Establishment of the AHP-FEM evaluation system, quantifying expert experience through the analytic hierarchy process (AHP), and combining the fuzzy evaluation method (FEM) to construct the classification standard (excellent / good / medium / poor) of extraction effect, providing a subjective label basis for the model; Step 2.2: Construct a Gaussian mixture model to provide a basic clustering tool for semi-supervised learning; Step 2.3: K-Means clustering algorithm; Step 2.4: Propose the AHP-FEM-SSGMM / SSK-Means model, combine the labels generated by AHP-FEM with GMM / K-Means clustering, and guide the clustering direction through a small number of labeled samples; provide a basic clustering tool for semi-supervised learning; Step: 2.5 PCA optimization algorithm is used to reduce the dimensionality of the 8-dimensional original data.

5. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 4 is characterized in that: The step 2.1 is specifically as follows: (1) Selection of evaluation indicators A series of factors are selected as evaluation indicators for extraction boreholes. According to different data sources, the weight calculation methods are mainly divided into three categories: subjective method, objective method and comprehensive method; (2) Weight calculation The hierarchical model is used to decompose the complex coal seam drilling and extraction status assessment problem into a hierarchical system that can be quantified and analyzed, and to assign quantitative weights to the evaluation indicators; (3) Establishment of evaluation label system; In the hierarchical structure model of step (2), the top layer is the target layer, followed by the criterion layer, the sub-criterion layer, and the bottom layer is the specific solution layer; the elements in each layer are compared in pairs to form a judgment matrix, and the weights are calculated using the eigenvector method. For the indicators at the same level, they are compared in pairs to form a judgment matrix, and industry experts are invited to score the comparison results according to the 9-scale method; AHP ensures the rationality of expert scoring through consistency testing. First, each vector of the judgment matrix is ​​normalized: Represents the element in the i-th row and j-th column of the normalized judgment matrix. ij Represents the element in the i-th row and j-th column of the original judgment matrix. n represents the order of the judgment matrix, that is, the number of indicators involved in the comparison. i,j represent the i-th row and j-th column respectively; Add the rows of the normalized matrix together: Calculate the average value of each row of the normalized matrix to obtain a column vector, which is the weight vector; Calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix. The maximum eigenvalue is obtained by the following formula: Where n is the order of the judgment matrix, A is the judgment matrix, and w is the normalized weight vector, (Aw) i is the i-th element of the Aw vector, w i is the i-th element of the w vector, and the eigenvector is the normalized weight vector, which can be obtained by the square root method or the sum method; Calculate the consistency index CI of the judgment matrix, which is defined as: The closer CI is to 0, the better the consistency of the judgment matrix. CI = 0 means that the judgment matrix is ​​completely consistent. Next, the consistency ratio CR of the judgment matrix is ​​calculated, which is defined as: RI is calculated based on the average CI value of the randomly generated judgment matrix and is related to the order of the judgment matrix. The smaller the CR value, the better the consistency of the judgment matrix. When CR<0.1, the consistency of the judgment matrix is ​​considered acceptable, otherwise the judgment matrix needs to be modified; the weight calculation of the extraction effect evaluation index of the coal seam pre-extraction drilling hole is carried out by using the hierarchical analysis method; The step (3) is specifically: In AHP, the relative relationship of evaluation indicators usually shows an increasing or decreasing trend; Formulas (8) and (9) are the calculation formulas for the descending semi-trapezoidal model and the ascending semi-trapezoidal model, respectively. i ) represents the i-th evaluation index X i The membership value of is greater in the descending semi-trapezoidal model, the smaller the contribution rate to satisfaction is, while the opposite is true in the ascending semi-trapezoidal model; The mathematical expression of the trapezoidal membership function is: Among them, X i represents the specific data value of the i-th evaluation index, x imax represents the maximum value of the i-th evaluation index in the data set, x imin represents the minimum value of the i-th evaluation index in the data set, μ(X i ) represents the i-th evaluation index X i The membership value of The satisfactory value of coal seam gas pre-extraction drilling effect evaluation is expressed as:

6. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 4, characterized in that: The step 2.2 is specifically as follows: Assume that each sample in the data set is randomly generated from one of K Gaussian distributions, where K is the number of clusters, each Gaussian distribution corresponds to a cluster, and its parameter is the mean μ k , variance Σ k and weight π k ,satisfy GMM is represented as: Among them, N(x|μ k ,Σ k ) is the probability density function of the Gaussian distribution, expressed as: Where d is the dimension of the sample, |Σ k ∣ is the determinant of the variance matrix; This posterior probability is expressed as: Among them, p(k|x) is the conditional probability of a given cluster, p(k) is the prior probability of the cluster, and p(x) is the marginal probability of the sample. This posterior probability is also called a latent variable, denoted by γ nk , which represents the probability that the nth sample belongs to the kth cluster. The calculation steps of GMM are as follows: The number of clusters k, the initial parameter μ of each cluster k ,Σ k and π k , N is the number of samples, X={x1,x2,...,x N } is the dataset, Is a collection of parameters.

7. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 4, characterized in that: The step 2.3 is specifically as follows: The goal of K-Means is to find a set of cluster centers so that the sum of the distances between each sample and the center of its cluster is minimized, that is, to minimize the following objective function: Where k is the number of clusters, C i is the i-th cluster, μ i is the center of the ith cluster, ||x-μ i || is the sample x and the cluster center μ i The Euclidean distance of .

8. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 4, characterized in that: The step 2.4 is specifically as follows: The mixed samples are clustered by GMM and K-Means clusterers, the information of labeled samples is mapped to unlabeled samples, SSGMM and SSK-Means classifiers are generated to realize the classification of mixed samples, and the quality and quantity of samples screened by AHP-FEM are evaluated by SSGMM classifier. Through the automated learning process, the optimal input samples and high-precision classifiers are obtained.

9. The method for evaluating the state of coal seam drilling pre-extraction gas based on semi-supervised learning according to claim 4, characterized in that: The step 2.5 is specifically as follows: The PCA optimization algorithm is used to perform dimensionality reduction. The dimensionality reduction method is shown in the following steps; (1) Assume that the original data is an n×d matrix X, where n is the number of samples, d is the dimension, and the target dimension is k, satisfying k<d; (2) Change the mean and variance of each feature to 0 and 1 respectively, so that the data has the same scale on different features; Among them, X ij represents the element in the i-th row and j-th column of the data matrix, represents the mean of the jth column, σ j represents the standard deviation of the jth column; (3) Calculate the covariance matrix S of the data; Among them, X T represents the transpose of the data matrix, S is a d×d symmetric matrix, the diagonal elements of the covariance matrix represent the variance of each feature, and the off-diagonal elements represent the covariance between two features; (4) Calculate the eigenvalues ​​and eigenvectors of the covariance matrix: Sv=λv(17) Where v is an eigenvector of the covariance matrix and λ is the corresponding eigenvalue; (5) Select the eigenvectors corresponding to the first k largest eigenvalues ​​as the new coordinate axes of the data to form a d×k projection matrix W; <h2 style=";text-align:left;direction:ltr">W=[v1,v2,...,v<h2 style=";text-align:left;direction:ltr"> k <h2 style=";text-align:left;direction:ltr"> (18) (6) Convert the data from the original coordinate system to the new coordinate system, that is, multiply the data by the projection matrix to obtain an n×k dimensionally reduced data matrix Y; Y=XW (19) Among them, each row of Y represents the representation of a sample in the new coordinate system, and each column of Y represents the representation of a principal component in the new coordinate system. PCA dimensionality reduction processing is used to map high-dimensional data to a low-dimensional space for easy visualization.

Images