A method for drilling and exploring in advance of concealed geological disasters in a coal mining working face

Abstract

The application discloses a kind of coal mine mining working face advanced concealed geological disaster while drilling exploration methods, belong to geological disaster identification technical field.The method is set between drill bit and drill pipe Detection drill pipe, through built-in sensing mechanism while drilling collection drilling thrust, rotation torque, three-dimensional vibration, gas concentration, drill depth and other multi-source parameters;Multi-source parameters are denoised, time domain and frequency domain feature extraction, form multidimensional feature vector with drill depth as index;Based on multidimensional feature vector, geological anomaly information such as fault fracture zone, fissure development zone at different depth positions along the drill hole is intelligently identified;Gas enrichment area category, relative gas pressure gradient and gas content trend are identified using multi-task learning model;High-pressure dense gas reservoir, structural fissure seepage channel and soft structure coal outburst risk and other gas-geological types are determined by fusing geological and gas information.The application can improve the real-time identification capability of advanced concealed disaster in mining working face.

Description

Technical Field

[0001] This invention relates to the field of coal mine geological hazard identification and drilling exploration technology, and in particular to a drilling exploration method for advanced identification of hidden geological hazards in coal mine working faces, gas distribution inversion, and gas-geological coupling risk assessment. Background Technology

[0002] Currently, the main methods for detecting hidden geological hazards in coal mine working faces are divided into two categories: drilling and geophysical exploration. However, traditional conventional drilling technology has obvious limitations: conventional drilling uses ordinary drill rods in combination, and the drill rods do not have real-time monitoring capabilities. Post-drilling analysis is required after the borehole construction is completed, through methods such as core sampling and gas desorption, resulting in data acquisition delays. At the same time, it is easily affected by drilling vibrations and equipment noise, leading to low identification of geological anomalies. Geophysical exploration is easily affected by factors such as differences in the electrical properties of the strata, tectonic interference, and mining disturbances, resulting in multiple interpretations of the detection results and insufficient resolution for identifying small-scale hidden geological hazards (such as small faults and localized gas-rich areas).

[0003] In recent years, exploration while drilling (WD) technology has gradually become a research hotspot in coal mine geological hazard investigation. Some technologies attempt to acquire WD data by mounting sensors on the drill pipe, but existing WD technologies still have many shortcomings: First, most technologies only collect drilling parameters or gas parameters individually, lacking a collaborative monitoring mechanism for geological and gas parameters, making it impossible to achieve coupled analysis of geological structures and gas hazards, and difficult to assess the risk of geological-gas synergistic disasters; Second, WD monitoring data has multi-source heterogeneous characteristics, with data collected by different sensors showing deviations in time and space. Existing technologies lack effective data depth alignment methods and have not carried out targeted noise reduction processing for the noise characteristics of WD data, resulting in low accuracy in extracting time-frequency domain feature parameters, affecting the accuracy of subsequent hazard identification; Third, the hazard assessment stage of existing WD technologies relies heavily on manual experience, with low levels of intelligence, making it difficult to quickly and accurately identify the range of gas-rich areas, gas pressure and content distribution along the borehole, as well as the type and boundaries of geological anomalies. Summary of the Invention

[0004] This invention aims to solve the above problems and provides a method for proactively detecting hidden geological hazards in coal mine working faces during drilling. The technical solution adopted is as follows: S1. A detection drill rod is connected between the drill bit and the drill rod, and a monitoring and sensing mechanism is installed inside the detection drill rod; S2. The drill bit drives the probe drill rod to perform geological drilling in front of the mining working face. During the drilling process, the monitoring and sensing mechanism collects the drilling monitoring parameters in real time and stores and uploads the drilling monitoring parameters to the host computer according to the sampling timestamp. The drilling monitoring parameters can be divided into the following categories according to the data type: in-hole drilling monitoring data, out-of-hole gas dynamic data, drilling space data and construction process data. S3. Based on the borehole depth, coordinate the different types of monitoring data while drilling, perform noise reduction on the monitoring parameters while drilling, divide the data analysis window along the borehole depth direction, calculate the time domain characteristics, frequency domain characteristics, gas dynamic characteristics and borehole trajectory characteristics of the monitoring data while drilling in each analysis window, and form a multi-dimensional feature vector indexed by the borehole depth. S4. The multidimensional feature vectors are fused and analyzed to identify the geological anomaly types and corresponding confidence levels of each analysis window along the borehole; S5. Based on the mechanical response characteristics and gas dynamic characteristics, a mechanical-gas full feature vector is established, and a multi-task learning model is used to identify the gas enrichment zone category, relative gas pressure gradient and gas content trend. The multi-task learning model is obtained by training with historical borehole data. S6. Input the geological anomaly identification results obtained in step S4 and the gas distribution identification results obtained in step S5 into the mechanical-gas fusion response rule base to determine the gas-geological type of each analysis window.

[0005] Based on the above technical solution, the in-hole drilling monitoring data includes: drilling thrust F, rotational torque T, triaxial vibration and drilling speed; the external gas dynamic data includes gas concentration C, concentration increment ΔC, and concentration change rate; the borehole space data includes: borehole depth D and trajectory inclination angle θ; and the construction process data includes sampling timestamp, drill rod length and rod replacement time.

[0006] Based on the above scheme, in step S2, the host computer establishes a drilling monitoring parameter database, which includes multiple borehole records. Each borehole record includes a borehole number, a borehole time sequence, a borehole depth sequence, a borehole trajectory sequence, a drilling thrust sequence, a rotational torque sequence, a three-dimensional vibration sequence, and a gas concentration sequence. A corresponding sampling timestamp and depth identifier are set for each sampling point.

[0007] Preferably, step S3 includes: S3-1. Establish a depth function D(t) based on the correspondence between borehole depth and sampling time, and interpolate and resample the drilling thrust, rotational torque, three-dimensional vibration and borehole trajectory data according to the preset depth interval Δd; S3-2. The aligned drilling monitoring data is denoised. The denoising method can be one of outlier removal, median replacement, moving average, wavelet threshold denoising and empirical mode class decomposition denoising. S3-3. Along the drilling depth direction, an analysis window is defined as a nominal proportion of the length of a single drill rod. The length of the analysis window is 1 / 4 to 1 / 2 of the length of a single drill rod, and adjacent analysis windows can overlap by 10% to 50% of their window length. S3-4. Perform time-domain and frequency-domain characteristic calculations on different drilling parameters within the analysis window to form a multidimensional feature vector for the analysis window. The multidimensional feature vector includes the mean drilling thrust, the standard deviation of drilling thrust, the mean rotational torque, the standard deviation of rotational torque, the three-dimensional vibration dominant frequency, the energy ratio of the three-dimensional vibration sensitive frequency band, the vibration spectrum entropy, the mean gas concentration, the gas concentration increment, the gas concentration change rate, the trajectory dip angle, and the trajectory azimuth angle.

[0008] Based on the above scheme, step S4 includes: S4-1. Construct a database of geological anomaly response patterns, including typical combination and change patterns of multidimensional feature vectors of drilling monitoring parameters in terms of time series and numerical values ​​under various typical geological anomaly conditions; S4-2. Set the reference values ​​for drilling thrust, rotational torque, and three-dimensional vibration signal characteristics when drilling in normal coal seams, as well as the deviation range relative to the reference values. Within the deviation range, define multiple different levels of numerical ranges, and determine the level of each parameter based on the parameter values ​​in the multi-dimensional feature vector of the drilling monitoring parameters. S4-3. The identification framework is defined as Θ={normal coal seam, fault fracture zone, fracture development zone, collapse column or cavity, lithological interface or weak interlayer}. Drilling thrust, rotational torque, three-dimensional vibration and borehole trajectory are used as evidence sources. The DS evidence theory evaluation method is used to fuse evidence. The geological anomaly type with the highest basic probability after fusion is taken as the identification result, and the corresponding probability is used as the confidence level.

[0009] Based on the above scheme, the DS evidence theory evaluation method includes: S4-3-1. Define the recognition frame and focal element: Suppose that all possible evaluation results for a certain problem can form a complete set Θ={ T 1, T 2, , T 5} When all elements in a set Θ satisfy the condition of mutual exclusion, the set Θ is called a recognition frame; all subsets of the recognition frame Θ... A The set formed is called the power set of the recognition frame Θ, denoted as 2. Θ Then the power set within the comprehensive evaluation framework for geological anomaly types is: (4-1) S4-3-2. Constructing the basic probability assignment function: For a recognition frame Θ, the quality function of any subset A is its power set 2. Θ The mapping to [0, 1] can be represented as m:2 Θ →[0, 1], where m is the quality function of Θ on the recognition frame, and the quality function must satisfy the following conditions: (4-2) In the formula: m(A) represents the basic probability distribution of a subset, which indicates the degree of support for the evaluation result; Introducing the normal membership function as the quality function, its calculation formula can be expressed as equation (4-3). (4-3) In the formula: x is the characteristic value of the evaluation index. Identify the center of the interval for each evaluation indicator. and These are the upper and lower limits of each recognition interval; when x = At this point, the membership degree is at its maximum value of 1, indicating a state of complete trust. S4-3-3. Trust Function: definition A The trust function of Θ is Bel :2 Θ →Mapping to [0, 1]; Bel ( A ) indicates evidence against a subset A The overall level of trust constitutes a subset. A The lower bound of the probability distribution; the trust function satisfies the following conditions: (4-4) S4-3-4. Likelihood function: definition A The likelihood function of Θ is Pl :2 Θ The mapping of →[0, 1] represents the evaluation result. A The degree of suspicion Pl ( A ) constitute a subset A The upper bound of the probability distribution; the likelihood function satisfies the following conditions: (4-5) By calculating a certain evaluation result in the identification framework Θ A Trust function Be1 (A ) and likelihood function Pl ( A ), thus deriving the trust interval [ Be1 ( A ), Pl ( A ], which indicates the evaluation results A The degree of confirmation; S4-3-5. Synthesis Rules: Assumption m 1, m 2, ..., m N Corresponding to A 1, A 2, ..., A N The mass function, then N The formula for synthesizing the evidence information is shown in equation (4-6). (4-6) Where: ⊕ represents an orthogonal sum, K K is the conflict coefficient, reflecting the degree of conflict between two pieces of evidence. The value of K is positively correlated with the degree of conflict. The condition for two evidence indicators to be fused using Dempster's combination rule is that K < 1. ； According to the fusion rules, the trust function is first obtained. Bel ( A Then, based on the principle of maximum membership, the geological anomaly type is determined by the maximum trust function value.

[0010] The beneficial effects of this invention are as follows: A probe drill rod is installed behind the drill bit so that during the synchronous drilling process of the probe drill rod and the drill bit, the monitoring and sensing mechanism can realize real-time synchronous acquisition of drilling data, avoiding the lag in data acquisition; A complete workflow for depth alignment, noise reduction, and time-frequency domain feature parameter calculation was designed. This effectively eliminates the deviation of data from different sensors in the time and space dimensions, reduces the impact of interference factors such as drilling vibration and equipment noise on the data, significantly improves the extraction accuracy of time-frequency domain feature parameters, and thus improves the accuracy of geological identification results. By conducting coordinated monitoring and assessment of mechanical properties and gas parameters, the limitations of monitoring and assessing single-type parameters can be avoided. This not only improves the comprehensiveness of disaster identification but also enables coupled analysis of the synergistic disaster-causing effects of mechanical and gas parameters. Attached Figure Description

[0011] Figure 1 : Flowchart of the method of this invention. Detailed Implementation

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments: In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0013] In the description of this invention, it should be understood that the terms "center," "length," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," and "inner," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0014] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0015] like Figure 1 As shown, a method for proactively investigating hidden geological hazards in coal mine working faces during drilling includes the following steps: S1. A detection drill rod is connected between the drill bit and the drill rod, and a monitoring and sensing mechanism is installed inside the detection drill rod; S2. The drill bit drives the probe drill rod to perform geological drilling in front of the mining working face. During the drilling process, the monitoring and sensing mechanism collects the drilling monitoring parameters in real time, and stores and uploads the drilling monitoring parameters to the host computer according to the sampling timestamp. The drilling monitoring parameters can be divided into the following categories according to data type: in-hole drilling monitoring data, out-of-hole gas dynamic data, drilling space data, and construction process data; "outside the hole" in the out-of-hole gas dynamic data refers to the hole opening during the drilling process, such as the position of the windward side of the hole opening; S3. Based on the borehole depth, coordinate the different types of monitoring data while drilling, perform noise reduction on the monitoring parameters while drilling, divide the data analysis window along the borehole depth direction, calculate the time domain characteristics, frequency domain characteristics, gas dynamic characteristics and borehole trajectory characteristics of the monitoring data while drilling in each analysis window, and form a multi-dimensional feature vector indexed by the borehole depth. S4. The multidimensional feature vectors are fused and analyzed to identify the geological anomaly types and corresponding confidence levels of each analysis window along the borehole; S5. Based on the mechanical response characteristics and gas dynamic characteristics, a mechanical-gas full feature vector is established, and a multi-task learning model is used to identify the gas enrichment zone category, relative gas pressure gradient and gas content trend. The multi-task learning model is obtained by training with historical borehole data. S6. Input the geological anomaly identification results obtained in step S4 and the gas distribution identification results obtained in step S5 into the mechanical-gas fusion response rule base to determine the gas-geological type of each analysis window.

[0016] Furthermore, the drilling monitoring parameters in step S2, categorized by data type, main data content, and function, are shown in the table below: Table 1. Data Categories, Contents, and Functions of Monitoring While Drilling (MSD) Parameters

[0017] In step S2, after receiving the drilling monitoring parameters, the host computer establishes a drilling monitoring parameter database. The drilling monitoring parameter database includes multiple borehole records. Each borehole record includes a borehole number, a borehole time sequence, a borehole depth sequence, a borehole trajectory sequence, a drilling thrust sequence, a rotational torque sequence, a three-dimensional vibration sequence, and a gas concentration sequence. A corresponding sampling timestamp and depth identifier are set for each sampling point.

[0018] Step S3 includes: S3-1. Establish a depth function D(t) based on the correspondence between borehole depth and sampling time, and interpolate and resample the drilling thrust, rotational torque, three-dimensional vibration and borehole trajectory data according to the preset depth interval Δd; S3-2. The aligned drilling monitoring data is denoised. The denoising method can be one of outlier removal, median replacement, moving average, wavelet threshold denoising and empirical mode class decomposition denoising, in order to filter out high-frequency noise caused by equipment vibration and operation fluctuations, and retain effective signals and trends that can reflect advanced and hidden geological information. S3-3. Along the drilling depth direction, an analysis window is defined as a nominal proportion of the length of a single drill rod. The length of the analysis window is 1 / 4 to 1 / 2 of the length of a single drill rod, and adjacent analysis windows can overlap by 10% to 50% of their window length. S3-4. Perform time-domain and frequency-domain characteristic calculations on different drilling parameters within the analysis window to form a multidimensional feature vector for the analysis window. The multidimensional feature vector includes the mean drilling thrust, the standard deviation of drilling thrust, the mean rotational torque, the standard deviation of rotational torque, the three-dimensional vibration dominant frequency, the energy ratio of the three-dimensional vibration sensitive frequency band, the vibration spectrum entropy, the mean gas concentration, the gas concentration increment, the gas concentration change rate, the trajectory dip angle, and the trajectory azimuth angle.

[0019] Step S4 includes: S4-1. Construct a geological anomaly response pattern database, including typical combination and numerical variation patterns of multidimensional feature vectors of drilling monitoring parameters under various typical geological anomaly conditions, which may include the following table: Table 2. Comparison of characteristics of monitoring parameters for various typical geological anomalies during drilling.

[0020] S4-2. Set the benchmark values ​​for drilling thrust, rotational torque, and three-dimensional vibration signal characteristics when drilling in normal coal seams, as well as the deviation range relative to the benchmark values. Within the deviation range, define multiple distinct levels for each numerical range. Determine the level of each parameter based on the parameter values ​​in the multi-dimensional feature vector of the drilling monitoring parameters. An example of the classification is as follows: Table 3. Example of Classification of Monitoring While Drilling Parameters

[0021] S4-3. The identification framework is defined as Θ={normal coal seam, fault fracture zone, fracture development zone, collapse column or cavity, lithological interface or weak interlayer}. Drilling thrust, rotational torque, three-dimensional vibration and borehole trajectory are used as evidence sources. The DS evidence theory evaluation method is used to perform evidence fusion. The geological anomaly type with the highest basic probability after fusion is taken as the identification result, and the corresponding probability is used as the confidence level.

[0022] The DS evidence theory evaluation method includes: S4-3-1. Define the recognition frame and focal element: Suppose that all possible evaluation results for a certain problem can form a complete set Θ={ T 1, T 2, , T 5} When all elements in a set Θ satisfy the condition of mutual exclusion, the set Θ is called a recognition frame; all subsets of the recognition frame Θ... A The set formed is called the power set of the recognition frame Θ, denoted as 2. Θ Then the power set within the comprehensive evaluation framework for geological anomaly types is: (4-1) S4-3-2. Constructing the basic probability assignment function: For a recognition frame Θ, the quality function of any subset A is its power set 2. Θ The mapping to [0, 1] can be represented as m:2 Θ →[0, 1], where m is the quality function of Θ on the recognition frame, and the quality function must satisfy the following conditions: (4-2) In the formula: m(A) represents the basic probability distribution of a subset, which indicates the degree of support for the evaluation result; The determination of the quality function directly determines the accuracy of the evaluation results. In order to reduce the influence of subjective factors on the probability distribution of evaluation indicators, this paper introduces the normal membership function as the quality function, and its calculation formula can be expressed as equation (4-3). (4-3) In the formula: x is the characteristic value of the evaluation index. Identify the center of the interval for each evaluation indicator. and These are the upper and lower limits of each recognition interval; when x = At this point, the membership degree is at its maximum value of 1, indicating a state of complete trust. S4-3-3. Trust Function: definition A The trust function of Θ is Bel :2 Θ The mapping of →[0, 1] reflects the influence of evaluation indicators on evaluation results. A The level of trust; Bel ( A ) indicates evidence against a subset A The overall level of trust constitutes a subset. A The lower bound of the probability distribution; the trust function satisfies the following conditions: (4-4) S4-3-4. Likelihood function: definition A The likelihood function of Θ is Pl :2 Θ The mapping from →[0, 1] can more comprehensively describe the degree of trust that evaluation indicators place in the evaluation results; it represents the level of confidence in the evaluation results. A The degree of suspicion Pl ( A ) constitute a subset A The upper bound of the probability distribution; the likelihood function satisfies the following conditions: (4-5) By calculating a certain evaluation result in the identification framework Θ A Trust function Be1 ( A ) and likelihood function Pl ( A ), thus deriving the trust interval [ Be1 ( A ), Pl ( A ], which indicates the evaluation results A The degree of confirmation; S4-3-5. Synthesis Rules: The DS evidence theory possesses a strong ability to fuse multi-source information, stemming from its core rules of synthesis. Assumptions m 1, m 2, ..., m N Corresponding to A 1, A 2, ..., A N The mass function, then N The formula for synthesizing the evidence information is shown in equation (4-6). (4-6) Where: ⊕ represents an orthogonal sum, K K is the conflict coefficient, reflecting the degree of conflict between two pieces of evidence. The value of K is positively correlated with the degree of conflict. The condition for two evidence indicators to be fused using Dempster's combination rule is that K < 1. ； According to the fusion rules, the trust function is first obtained. Bel ( A Then, based on the principle of maximum membership, the geological anomaly type is determined by the maximum trust function value.

[0023] Step S5 includes: S5-1. Establish the mechanical-gas full feature vector for each analysis window, including coal and rock mass strength characteristics, coal and rock mass integrity, and gas dynamic characteristics. The coal and rock mass strength characteristics include the average drilling thrust, average rotational torque, and normalized comprehensive strength of the analysis window. The coal and rock mass integrity includes the standard deviation of rotational torque, drilling fluctuation coefficient, and spectral entropy or energy ratio of three-dimensional vibration in the predetermined sensitive frequency band within the analysis window. The gas dynamic characteristics include gas concentration, gas concentration increment, and gas concentration change rate. S5-2. Define the identification targets of gas distribution characteristics, including the classification of gas-rich areas, the identification results of relative gas pressure gradient, and the identification results of gas content trend; Specifically, the gas enrichment area is classified as a discrete label. Based on the pre-set gas concentration range, non-enriched areas, general enriched areas, and significantly enriched areas are defined, and each depth point or depth window is assigned to the corresponding category. This classification is used to guide the target area positioning of gas drainage boreholes.

[0024] The relative gas gradient identification result is a continuous value, representing the relative pressure change trend along the depth direction. It also identifies abrupt gas pressure change points along the depth direction, where the absolute value of the gradient change exceeds a predetermined value. These abrupt gas pressure change points are typically critical points where gas is suddenly released, inducing blowouts or outbursts, and are crucial for guiding drilling safety and extraction design.

[0025] The gas content trend identification result is a continuous value, representing a relative gas content change curve along the borehole depth direction. By using a small number of calibration points, the trend curve can be calibrated to approximate absolute content values. Calibration points can be laboratory measurements, etc. Gas content trend identification is used to assess the overall gas distribution in coal seams and optimize regional gas control schemes.

[0026] The parameter change trends and parameter value ranges in the above modes are all preset, and the setting principles can be domain knowledge and / or historical cases.

[0027] S5-3. Build a multi-task learning framework model, take the full feature vector of mechanics-gas as the model input, and take the identification target of gas distribution characteristics as the model output; the model uses gradient boosting decision tree and other models under the multi-task learning framework to handle the above classification and regression tasks at the same time, and uses historical data with labels (enrichment area category, pressure gradient, content trend value) for training.

[0028] The method for building the multi-task learning framework model includes: S5-3-1. Input Layer Design: The constructed mechanics-gas full feature vector is used as input; in order to eliminate the dimensional influence of multiple physical quantities and improve the convergence speed of the model, the input data is standardized, the mean and standard deviation of each dimension feature are calculated, and the processed feature vector is input into the network. (5-1) (5-2) In the formula, x i For the first i The original feature value of the first feature, obtained during the feature extraction stage. i The raw input data of the dimension, such as the average drilling thrust or gas concentration collected by the system; For the first i The standardized feature values ​​are the data that are finally input into the neural network after processing. After being processed by the normalization formula, the data is usually transformed into a standard normal distribution with a mean of 0 and a standard deviation of 1. μ i and σ i The first i The mean and standard deviation of the dimensional feature, i.e., the 3rd feature. i The arithmetic mean and dispersion of all raw data in the entire sample set across all dimensions; S5-3-2. Shared Hidden Layer Design: A multi-layer fully connected network is used to extract the shared features at the bottom layer. This layer receives standardized features and aims to explore the deep nonlinear coupling relationship between coal and rock strength, fragility and gas dynamics. Through activation function mapping, it provides shared features with strong generalization ability for downstream multi-tasks.

[0029] (5-3) In the formula, h l For the first l The output feature vector of a layer is the result of linear transformation and nonlinear activation of the current hidden layer, representing a more abstract and higher-level feature representation extracted by the network at this layer; ReLU is the modified linear unit activation function, whose mathematical definition is... f ( x )=max(0, x Its core function is to introduce nonlinear fitting ability into the model. Without the activation function, the multilayer neural network degenerates into a simple linear regression and will be unable to capture complex physical laws. W l For the first lThe weight matrix of a layer is a core parameter that the neural network continuously updates during training, determining the importance of each feature passed from the previous layer in the current layer. h l-1 For the first l The output feature vector of layer -1, for the -1st layer l In terms of layers, it refers to input data. If the current layer is the first hidden layer (i.e., ...) l =1), then h 0 This is the initial mechanical-gas full eigenvector after standardization. x '; b l For the first l The bias vector of a layer is a parameter that the model needs to learn. It is used to translate the feature space as a whole, increasing the flexibility of the network to fit complex data distributions. S5-3-3. Task-Specific Output Layer Design: The shared features are accessed through three independent branch networks; the gas enrichment area classification task uses a normalized exponential function to output the category probability; the relative pressure gradient and content trend identification task is a regression task, which uses a linear layer to output the predicted value, thereby achieving joint prediction of gas distribution characteristics; (5-4) (5-5) In the formula, p The output category probability vector represents the probability that the model predicts the coal and rock mass within the current analysis window belongs to a specific gas enrichment zone category; Softmax is the normalized exponential activation function, whose core function is to transform the original output of the linear layer into a standard probability distribution form, so that the sum of the probabilities of all predicted categories is strictly 1, thereby clearly determining the most likely gas enrichment state. W 1 represents the specific weight matrix for the classification task branch, which are parameters that the model learns independently specifically for the gas classification task during joint training. h L For the first L Shared feature vectors of layers L Represents the last layer of the shared hidden layer. h L It is a high-order feature that the model deeply extracts and fuses from the underlying drilling parameters and initial gas dynamic characteristics; b 1 is a specific bias vector for the classification task branch, used to fine-tune the classification decision boundary; y k For the first k The continuous prediction output value of a regression task, in this application, is... k When = 2, it represents the specific value of the relative gas pressure gradient predicted by the model.k When =3, it represents the predicted trend value of gas content; W k For the first k A specific weight matrix for each regression task branch, specifically designed to fit the independent parameter matrix for predicting continuous physical quantities; b k For the first k A specific bias vector for each branch of the regression task. k This serves as a task index identifier, used to distinguish which specific regression prediction task branch is currently being calculated; S5-3-4. Model Training and Loss Function Design: A multi-task joint loss function is constructed, which includes classification cross-entropy loss and regression mean squared error loss, and hyperparameters are introduced to balance the weights of each task. An adaptive optimization algorithm is used to update the parameters based on error backpropagation until the total loss function converges, thus completing model training. (5-6) In the formula, L Let be the total loss function, and let be the final target value that the entire multi-task learning model attempts to minimize in each training iteration. The model updates the weight matrix based on the total loss function using the backpropagation algorithm. W l and bias vector b l ; L ce Cross-entropy loss is used for classification tasks in gas-rich areas to measure the class probability of the model output. p Differences between labels and actual gas-rich areas; L mse1 The mean squared error loss of the first regression task in the relative gas pressure gradient identification task is used to measure the continuous numerical values ​​predicted by the model. y 2. The squared difference between the actual pressure gradient value and the actual pressure gradient value; L mse2 The mean squared error loss of the second regression task in the gas content trend identification task is used to measure the model's predicted values. y The differences λ1, λ2, and λ3 between the model and the true trend values ​​are task weight hyperparameters, which are used to balance the influence of different tasks on the backbone network and ensure that the model can achieve optimal performance on the three gas prediction tasks.

[0030] Step S6 includes: S6-1. Define the characteristic parameters for gas-geological type identification, including the dynamic characteristics of gas escape and the mechanical response characteristics; The dynamic characteristics of gas emission include concentration baseline, emission intensity, emission sensitivity, and emission morphology index; the concentration baseline is the stable background concentration during drilling intervals; emission intensity refers to the absolute emission amount relative to the concentration baseline; emission sensitivity refers to the rate of change of gas concentration with borehole depth to indicate the urgency of gas emission; the emission morphology index is used to describe the shape of the gas concentration peak to reflect the uniformity of the gas source. The mechanical response characteristics include normalized comprehensive mechanical strength (SI), normalized fragmentation characteristic (FI), and drilling efficiency index. The normalized comprehensive mechanical strength (SI) is a weighted average of normalized drilling thrust and rotational torque, used to quantify the overall strength of the encountered coal and rock mass. The normalized fragmentation characteristic (FI) is calculated from the fluctuation variance of rotational torque and the energy of a specific frequency band of three-dimensional vibration, used to quantify the degree of fragmentation or fracture development of the rock mass. The drilling efficiency index includes characteristic indicators such as drilling thrust / rotational torque, drilling thrust / three-dimensional vibration, and three-dimensional vibration / rotational torque, reflecting information about the encountered coal and rock mass through changes in these indicators. S6-2. Establish a mechanical-gas integrated response rule base, which lists typical gas geological types, including high-pressure tight gas reservoir model, tectonic fracture seepage channel model, and weak tectonic coal outburst risk model: High-pressure tight gas reservoir model: This geological type is characterized by intact rock mass but high-pressure gas content. The quantitative identification condition for this type is the increase in gas emission concentration. ΔC The increase reached ΔC ≥30% C 0, C 0 is the predetermined standard value; normalized comprehensive mechanical strength SI ≥0.65; Normalized fragmentation characteristic FI ≤0.3; the drilling fluctuation coefficient K remains within 3 consecutive analysis windows. K ≤0.4; Structural fracture seepage channel model: This geological type is characterized by numerous fractures in the coal and rock mass, providing channels for gas migration. The quantitative identification condition for this type is a 10% increase in gas emission concentration. C 0≤ ΔC <30% C 0 indicates a moderate upward trend; the normalized comprehensive mechanical strength (SI) exhibits discrete fluctuations with an overall value fluctuation range of 0.3 < SI <0.65; Normalized fragmentation characteristic 0.3< FI ≤0.75; the drilling fluctuation coefficient K remained at 0.4 for three consecutive analysis windows. K Within the range of ≤0.6; Soft geological structure coal outburst risk model: This geological type is characterized by extremely fragmented and weak coal bodies, exhibiting rheological or creep instability, and is highly susceptible to inducing gas dynamic disasters. The quantitative identification condition for this type is: gas emission concentration increment. ΔC > C 0, meaning any degree of abnormal increase; normalized comprehensive mechanical strength. SI ≤0.3; Normalized fragmentation characteristic FI ≥0.75; the drilling fluctuation coefficient K remains within 3 consecutive analysis windows. K ≥0.6.

[0031] The quantitative conditions for each of the above models are statistically summarized in Table 4 below: Table 4. Quantitative conditions for evaluating different gas-geological types

[0032] It also includes step S7, S7. Mark the geological anomaly types, gas distribution characteristics, and gas-geological types at each depth location, and arrange the above markings along the borehole trajectory for each depth location. The marking format can combine text and charts to visualize the detection and identification results.

[0033] The present invention has been described above by way of example, but the present invention is not limited to the specific embodiments described above. Any modifications or variations made based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method for proactively investigating hidden geological hazards in coal mine working faces during drilling, characterized in that, Includes the following steps: S1. A detection drill rod is connected between the drill bit and the drill rod, and a monitoring and sensing mechanism is installed inside the detection drill rod; S2. The drill bit drives the probe drill rod to perform geological drilling in front of the mining working face. During the drilling process, the monitoring and sensing mechanism collects the drilling monitoring parameters in real time and stores and uploads the drilling monitoring parameters to the host computer according to the sampling timestamp. The drilling monitoring parameters can be divided into the following categories according to the data type: in-hole drilling monitoring data, out-of-hole gas dynamic data, drilling space data and construction process data. S3. Based on the borehole depth, coordinate the different types of monitoring data while drilling, perform noise reduction on the monitoring parameters while drilling, divide the data analysis window along the borehole depth direction, calculate the time domain characteristics, frequency domain characteristics, gas dynamic characteristics and borehole trajectory characteristics of the monitoring data while drilling in each analysis window, and form a multi-dimensional feature vector indexed by the borehole depth. S4. The multidimensional feature vectors are fused and analyzed to identify the geological anomaly types and corresponding confidence levels of each analysis window along the borehole; S5. Based on the mechanical response characteristics and gas dynamic characteristics, a mechanical-gas full feature vector is established, and a multi-task learning model is used to identify the gas enrichment zone category, relative gas pressure gradient and gas content trend. The multi-task learning model is obtained by training with historical borehole data. S6. Input the geological anomaly identification results obtained in step S4 and the gas distribution identification results obtained in step S5 into the mechanical-gas fusion response rule base to determine the gas-geological type of each analysis window.

2. The method for advanced exploration of hidden geological hazards in coal mine working faces according to claim 1, characterized in that: The in-hole monitoring data includes: drilling thrust F, rotational torque T, triaxial vibration and drilling speed; the external gas dynamic data includes gas concentration C, concentration increment ΔC and concentration change rate; the borehole spatial data includes: borehole depth D and trajectory inclination angle θ; and the construction process data includes sampling timestamp, drill rod length and rod replacement time.

3. A method for advanced drilling exploration of concealed geological hazards in coal mine working faces according to claim 1 or 2, characterized in that, In step S2, the host computer establishes a drilling monitoring parameter database, which includes multiple borehole records. Each borehole record includes a borehole number, a borehole time sequence, a borehole depth sequence, a borehole trajectory sequence, a drilling thrust sequence, a rotational torque sequence, a three-dimensional vibration sequence, and a gas concentration sequence. A corresponding sampling timestamp and depth identifier are set for each sampling point.

4. The method for advanced exploration of concealed geological hazards in coal mine working faces according to claim 2, characterized in that, Step S3 includes: S3-1. Establish a depth function D(t) based on the correspondence between borehole depth and sampling time, and interpolate and resample the drilling thrust, rotational torque, three-dimensional vibration and borehole trajectory data according to the preset depth interval Δd; S3-2. The aligned drilling monitoring data is denoised. The denoising method can be one of outlier removal, median replacement, moving average, wavelet threshold denoising and empirical mode class decomposition denoising. S3-3. Along the drilling depth direction, an analysis window is defined as a nominal proportion of the length of a single drill rod. The length of the analysis window is 1 / 4 to 1 / 2 of the length of a single drill rod, and adjacent analysis windows can overlap by 10% to 50% of their window length. S3-4. Perform time-domain and frequency-domain characteristic calculations on different drilling parameters within the analysis window to form a multidimensional feature vector for the analysis window. The multidimensional feature vector includes the mean drilling thrust, the standard deviation of drilling thrust, the mean rotational torque, the standard deviation of rotational torque, the three-dimensional vibration dominant frequency, the energy ratio of the three-dimensional vibration sensitive frequency band, the vibration spectrum entropy, the mean gas concentration, the gas concentration increment, the gas concentration change rate, the trajectory dip angle, and the trajectory azimuth angle.

5. The method for advanced drilling exploration of concealed geological hazards in coal mine working faces according to claim 4, characterized in that, Step S4 includes: S4-1. Construct a database of geological anomaly response patterns, including typical combination and change patterns of multidimensional feature vectors of drilling monitoring parameters in terms of time series and numerical values ​​under various typical geological anomaly conditions; S4-2. Set the reference values ​​for drilling thrust, rotational torque, and three-dimensional vibration signal characteristics when drilling in normal coal seams, as well as the deviation range relative to the reference values. Within the deviation range, define multiple different levels of numerical ranges, and determine the level of each parameter based on the parameter values ​​in the multi-dimensional feature vector of the drilling monitoring parameters. S4-3. The identification framework is defined as Θ={normal coal seam, fault fracture zone, fracture development zone, collapse column or cavity, lithological interface or weak interlayer}. Drilling thrust, rotational torque, three-dimensional vibration and borehole trajectory are used as evidence sources. The DS evidence theory evaluation method is used to fuse evidence. The geological anomaly type with the highest basic probability after fusion is taken as the identification result, and the corresponding probability is used as the confidence level.

6. The method for advanced drilling exploration of concealed geological hazards in coal mine working faces according to claim 5, characterized in that, The DS evidence theory evaluation method includes: S4-3-1. Define the recognition frame and focal element: Suppose that all possible evaluation results for a certain problem can form a complete set Θ={ T 1, T 2, , T 5} When all elements in a set Θ satisfy the condition of mutual exclusion, the set Θ is called a recognition frame; all subsets of the recognition frame Θ... A The set formed is called the power set of the recognition frame Θ, denoted as 2. Θ Then the power set within the comprehensive evaluation framework for geological anomaly types is: （4-1） S4-3-2. Constructing the basic probability assignment function: For a recognition frame Θ, the quality function of any subset A is its power set 2. Θ The mapping to [0, 1] can be represented as m:2 Θ →[0, 1], where m is the quality function of Θ on the recognition frame, and the quality function must satisfy the following conditions: （4-2） In the formula: m(A) represents the basic probability distribution of a subset, which indicates the degree of support for the evaluation result; Introducing the normal membership function as the quality function, its calculation formula can be expressed as equation (4-3). （4-3） In the formula: x is the characteristic value of the evaluation index. Identify the center of the interval for each evaluation indicator. and These are the upper and lower limits of each recognition interval; when x = At this point, the membership degree is at its maximum value of 1, indicating a state of complete trust. S4-3-3. Trust Function: definition A The trust function of Θ is Bel :2 Θ →Mapping to [0, 1]; Bel ( A ) indicates evidence against a subset A The overall level of trust constitutes a subset. A The lower bound of the probability distribution; the trust function satisfies the following conditions: （4-4） S4-3-4. Likelihood function: definition A The likelihood function of Θ is Pl :2 Θ The mapping →[0, 1] represents the evaluation result. A The degree of suspicion Pl ( A ) constitute a subset A The upper bound of the probability distribution; the likelihood function satisfies the following conditions: （4-5） By calculating a certain evaluation result in the identification framework Θ A Trust function Be1 ( A ) and likelihood function Pl ( A ), thus deriving the trust interval [ Be1 ( A ), Pl ( A ], which indicates the evaluation results A The degree of confirmation; S4-3-5. Synthesis Rules: Assumption m 1, m 2, ..., m N Corresponding to A 1, A 2, ..., A N The mass function, then N The formula for synthesizing the evidence information is shown in equation (4-6). （4-6） Where: ⊕ represents an orthogonal sum, K K is the conflict coefficient, reflecting the degree of conflict between two pieces of evidence. The value of K is positively correlated with the degree of conflict. The condition for two evidence indicators to be fused using Dempster's combination rule is that K < 1. ； According to the fusion rules, the trust function is first obtained. Bel ( A Then, based on the principle of maximum membership, the geological anomaly type is determined by the maximum trust function value.

7. A method for advanced drilling exploration of concealed geological hazards in coal mine working faces according to claim 1 or 2, characterized in that, Step S5 includes: S5-1. Establish the mechanical-gas full feature vector for each analysis window, including coal and rock mass strength characteristics, coal and rock mass integrity, and gas dynamic characteristics. The coal and rock mass strength characteristics include the average drilling thrust, average rotational torque, and normalized comprehensive strength of the analysis window. The coal and rock mass integrity includes the standard deviation of rotational torque, drilling fluctuation coefficient, and spectral entropy or energy ratio of three-dimensional vibration in the predetermined sensitive frequency band within the analysis window. The gas dynamic characteristics include gas concentration, gas concentration increment, and gas concentration change rate. S5-2. Define the identification targets of gas distribution characteristics, including gas enrichment zone classification, relative gas pressure gradient and gas content trend; S5-3. Construct a multi-task learning framework model, using the mechanics-gas full feature vector as the model input and the gas distribution feature identification target as the model output. The multi-task learning model includes a shared feature extraction layer, a gas enrichment area classification output branch, a relative gas pressure gradient regression output branch, and a gas content trend regression output branch. The model training adopts a joint loss function that is a weighted sum of classification cross-entropy loss and two regression mean square error losses.

8. The method for advanced exploration of concealed geological hazards in coal mine working faces according to claim 7, characterized in that, The method for building the multi-task learning model includes: S5-3-1. Input Layer Design: The constructed mechanics-gas full feature vector is used as input; the mean and standard deviation of each dimension feature are calculated to obtain the processed feature vector, which is then input into the network. （5-1） （5-2） In the formula, x i For the first i The original feature value of the first feature, obtained during the feature extraction stage. i The original input data for the dimension; For the first i The standardized feature values ​​of each feature; μ i and σ i The first i The mean and standard deviation of the dimensional feature, i.e., the 3rd feature. i The arithmetic mean and dispersion of all raw data in the entire sample set across all dimensions; S5-3-2. Shared Hidden Layer Design: Extracting shared features from the lower layers using a multi-layer fully connected network （5-3） In the formula, h l For the first l The output feature vector of a layer is the result of linear transformation and nonlinear activation of the current hidden layer, representing a more abstract and higher-level feature representation extracted by the network at this layer; ReLU is the modified linear unit activation function, whose mathematical definition is... f ( x )=max(0, x ); W l For the first l Layer weight matrix; h l-1 For the first l -1 layer output feature vector; b l For the first l Layer bias vector; S5-3-3. Task-Specific Output Layer Design: The shared features are accessed through three independent branch networks; the gas enrichment area classification task uses a normalized exponential function to output the category probability; the relative pressure gradient and content trend identification task is a regression task, which uses a linear layer to output the predicted value, thereby achieving joint prediction of gas distribution characteristics; （5-4） （5-5） In the formula, p The output category probability vector represents the probability that the model predicts the coal and rock mass within the current analysis window belongs to a specific gas-rich area category; Softmax is the normalized exponential activation function, used to transform the original output of the linear layer into a standard probability distribution form, so that the sum of the probabilities of all predicted categories is 1. W 1 represents the specific weight matrix for the classification task branch; h L For the first L Shared feature vectors of layers L This is the last layer of the shared hidden layer; b 1 represents the specific bias vector for the classification task branch; y k For the first k The continuous prediction output value of a regression task, when k When =2, y k The specific value of the relative gas pressure gradient predicted by the model, when k When =3, y k This represents the predicted trend value of gas content. W k For the first k A specific weight matrix for each regression task branch; b k For the first k A specific bias vector for each branch of the regression task. k This serves as a task index identifier, used to distinguish which specific regression prediction task branch is currently being calculated; S5-3-4. Model Training and Loss Function Design: A multi-task joint loss function is constructed, which includes classification cross-entropy loss and regression mean squared error loss, and hyperparameters are introduced to balance the weights of each task. An adaptive optimization algorithm is used to update the parameters based on error backpropagation until the total loss function converges, thus completing model training. （5-6） In the formula, L Let be the total loss function, and let be the final target value that the entire multi-task learning model attempts to minimize in each training iteration. The model updates the weight matrix based on the total loss function using the backpropagation algorithm. W l and bias vector b l ; L ce Cross-entropy loss is used for classification tasks in gas-rich areas to measure the class probability of the model output. p Differences between labels and actual gas-rich areas; L mse1 The mean squared error loss of the first regression task in the relative gas pressure gradient identification task is used to measure the continuous numerical values ​​predicted by the model. y 2. The squared difference between the actual pressure gradient value and the actual pressure gradient value; L mse2 The mean squared error loss of the second regression task in the gas content trend identification task is used to measure the model's predicted values. y The differences λ1, λ2, and λ3 between the actual trend values ​​and the actual trend values ​​are the task weight hyperparameters, which are used to balance the influence of different tasks on the backbone network.

9. A method for advanced drilling exploration of concealed geological hazards in coal mine working faces according to claim 1 or 2, characterized in that, Step S6 includes: S6-1. Define the characteristic parameters for gas-geological type identification, including the dynamic characteristics of gas escape and the mechanical response characteristics; S6-2. Establish a mechanical-gas integrated response rule base, which lists typical gas geological types, including high-pressure tight gas reservoir model, tectonic fracture seepage channel model, and weak tectonic coal outburst risk model: High-pressure tight gas reservoir model: gas emission concentration increment ΔC The increase reached ΔC ≥30% C 0, C 0 is the predetermined standard value; normalized comprehensive mechanical strength SI ≥0.65; Normalized fragmentation characteristic FI ≤0.3; the drilling fluctuation coefficient K remains within 3 consecutive analysis windows. K ≤0.4; Construction of fracture seepage channels: Gas emission concentration increment is 10%. C 0≤ ΔC <30% C 0; The normalized comprehensive mechanical strength SI exhibits discrete fluctuations, with the overall value fluctuating within a range of 0.3 < SI <0.65; Normalized fragmentation characteristic 0.3< FI ≤0.75; the drilling fluctuation coefficient K remained at 0.4 for three consecutive analysis windows. K Within the range of ≤0.6; Coal outburst risk model in weak geological structures: Incremental gas emission concentration ΔC > C 0; Normalized comprehensive mechanical strength SI ≤0.3; Normalized fragmentation characteristic FI ≥0.75; the drilling fluctuation coefficient K remains within 3 consecutive analysis windows. K ≥0.6.

Images