Coal seam drilling intelligent early warning system based on unsupervised gas pressure data learning

Abstract

The invention discloses an unsupervised coal seam drilling intelligent early warning system for learning gas pressure data. The unsupervised coal seam drilling intelligent early warning system comprises a multi-source sensing data acquisition module, an edge preprocessing and feature engineering module, a physical mechanism embedded unsupervised clustering module, a dynamic risk mapping and grading early warning module and a man-machine interaction and decision support module. The multi-source sensing data acquisition module is deployed in a drilling machine operation area and is used for acquiring gas pressure in real time. The invention relates to the technical field of artificial intelligence and mine safety monitoring. According to the unsupervised coal seam drilling intelligent early warning system for learning the gas pressure data, a physical mechanism embedded unsupervised clustering architecture is constructed, so that the dependence on manual annotation data is thoroughly eliminated, and full-automatic, self-adaptive and high-precision early warning from original sensing data to risk level output is realized; the technical scheme has remarkable innovativeness and engineering practicability in the aspects of feature engineering, clustering algorithm, risk mapping, system integration and the like.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and mine safety monitoring technology, specifically to an intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data. Background Technology

[0002] In the coal mine safety production system, gas disaster prevention and control always occupies a core position, directly affecting the life safety of underground workers and the mine's ability to maintain continuous and stable production. Coal seam drilling, as a key preliminary process for gas extraction and geological exploration, involves dynamic changes in gas pressure during the process, which essentially carries multi-dimensional information about the stability of the coal seam structure, the state of gas occurrence, and potential outburst risks. Traditionally, gas pressure monitoring mainly relies on fixed-point, intermittent manual measurement or semi-automated sensing devices. These methods have low data acquisition frequency and delayed response, making it difficult to meet the urgent needs of modern intelligent mines for "real-time risk perception, dynamic assessment, and proactive early warning."

[0003] The patent publication number "CN112926270B" discloses a method for analyzing the coupling relationship of multiple factors of gas and constructing an early warning model, which specifically includes the following steps: Step 1, collecting gas concentration data of the working face, upper corner, and coal seam, as well as hydraulic support pressure data of the working face; Step 2, performing outlier detection on the four-dimensional data using a k-means algorithm based on the optimization of cluster initial value selection; Outliers in gas data can be regarded as outliers in the clustering process, so the extraction of outliers can be based on the clustering method to detect outliers by examining the relationship between the object and the cluster.

[0004] As shown in the above technologies, existing technologies mostly focus on supervised learning frameworks, which rely on historical labeled datasets to train classification or regression models to predict whether the current drilling status is within the danger threshold range. However, coal mine geological conditions are highly heterogeneous and non-repeatable. The evolution of gas pressure corresponding to different working faces, different coal seams, and even different borehole trajectories are significantly different, making it difficult for historical labeled data to cover all potential risk patterns. On the other hand, extreme disaster events such as gas outbursts are characterized by low frequency, suddenness, and non-steady state, making "positive samples" (i.e., real dangerous events) that can be used for training extremely scarce. This leads to sample imbalance during model training, ultimately causing the model's generalization ability to drop sharply when facing unknown or rare risk patterns, and even resulting in misjudgments. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data, which solves the problem of high-precision, low-latency, and strong generalization early warning of potential gas disasters during coal seam drilling.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: A coal seam drilling intelligent early warning system based on unsupervised learning of gas pressure data includes: a multi-source sensor data acquisition module, an edge preprocessing and feature engineering module, a physical mechanism embedded unsupervised clustering module, a dynamic risk mapping and hierarchical early warning module, and a human-computer interaction and decision support module; the multi-source sensor data acquisition module is deployed in the drilling rig operating area to collect three types of time-series data in real time: gas pressure, drilling parameters, and environmental status, and synchronously transmits them to the edge computing node via an industrial Ethernet bus or CAN bus; The edge preprocessing and feature engineering module performs sliding median filtering, cubic spline interpolation and imputation, and zero-mean unit variance standardization on the original data. Based on a 60-second sliding window, it extracts 28-dimensional dynamic features, including pressure mean, standard deviation, skewness, kurtosis, maximum and minimum values, rise and fall rates, spectral energy entropy, wavelet packet decomposition coefficient energy ratio, autocorrelation and cross-correlation function lag coefficients, pressure gradient change rate, and pressure fluctuation frequency. Then, it performs principal component analysis to reduce the dimensionality and retain the top 8 principal components with a cumulative variance contribution rate of ≥95% as the output feature vector. The physical mechanism embedded unsupervised clustering module adopts an improved constrained K-means algorithm. Its initial cluster centers are preset into three categories based on gas dynamics theory: stable low-pressure zone, fluctuating medium-pressure zone, and high-pressure abrupt change zone. During the iteration process, physical consistency indicators are introduced, including pressure gradient consistency index, spectral structure similarity, and cross-correlation stability coefficient with drilling speed sequence. These drive the cluster splitting mechanism or cluster merging mechanism to ensure that the final clusters have high cohesion and inter-class separability in both statistical distribution and physical mechanism dimensions.

[0007] Preferably, the dynamic risk mapping and hierarchical early warning module receives clustering results according to the formula: R=α·P mean + β·dP max / dt + γ·H entropy + δ·N spike A risk score is calculated, where α, β, γ, and δ are all weighting coefficients, and α+β+γ+δ=1. After score normalization, the scores are divided into safety level, attention level, warning level, and danger level according to the threshold. The cluster evolution trajectory is monitored to trigger dynamic upgrade rules. The human-computer interaction and decision support module provides a visual interface on both the downhole explosion-proof display screen and the ground monitoring center. When the risk level reaches the warning level or above, the process parameter adjustment strategy library is automatically called to generate control commands and sent to the drilling rig control system for execution via CAN bus.

[0008] Preferably, the physical mechanism embedded unsupervised clustering module sets three types of physical constraints for cluster centers during the cluster initialization phase: the stable low-pressure zone satisfies a pressure mean <0.3MPa and a standard deviation <0.05MPa; the fluctuating medium-pressure zone satisfies a pressure mean of 0.3–1.2MPa and a standard deviation of 0.05–0.2MPa; and the high-pressure abrupt change zone satisfies a pressure mean >1.2MPa or a pressure rise rate >0.5MPa / min or a spectral energy entropy >1.8.

[0009] Preferably, during the iterative process of the physical mechanism embedded unsupervised clustering module, if the physical consistency index of a cluster is lower than a preset threshold, i.e., the pressure gradient consistency index < 0.7, the spectral structure similarity < 0.6, and the cross-correlation stability coefficient < 0.5, then the cluster splitting mechanism is triggered, and the samples are divided into two groups according to the median pressure rise rate and the sub-cluster centers are recalculated; if the physical consistency of two adjacent clusters meets the standard and the cluster distance is less than 0.3 times the average intra-cluster distance, then the cluster merging mechanism is triggered, and the new cluster center is the weighted average of the original cluster centers.

[0010] Preferably, in the risk scoring function used, P mean dP is the average pressure within the cluster. max / dt is the maximum rate of pressure rise within the cluster, H entropy N represents the mean spectral energy entropy within the cluster. spike This represents the average number of pressure jumps per unit time within the cluster.

[0011] Preferably, the physical mechanism embedded unsupervised clustering module is also configured with a dynamic upgrade trigger: if the risk score of a safety-level cluster continues to rise within 5 consecutive time windows and the magnitude is >0.1, it is automatically upgraded to the attention level; if the score of a warning-level cluster jumps to >0.2 within a single time window, a danger-level alarm is immediately triggered; the risk level classification result synchronously outputs the corresponding cluster identification code C-XXX, where XXX is a three-digit number, and the number is sorted according to the cluster center's position in the feature space.

[0012] Preferably, the human-machine interaction and decision support module displays the current borehole location, six-channel gas pressure curve, three-dimensional projection distribution map of clusters, four-color risk indicator lights, and text warning information in real time at the downhole end; and provides historical risk evolution heat map, multi-bore risk comparison bar chart, and cluster center migration trajectory map at the surface end; when the risk level is determined to be warning level or dangerous level, it automatically matches a scheme from the pre-stored 12 process parameter adjustment strategy library, generates a CAN bus control frame containing target drilling speed, target thrust, pause command or water injection command, and sends it to the drilling rig PLC for execution, with a command response delay of <200ms.

[0013] Preferably, the edge computing nodes cache the feature vectors of the most recent 24 hours and perform incremental K-means fine-tuning every hour. If the distance between a new data point and all existing cluster centers is greater than 1.5 times the average intra-cluster distance, a new cluster is created and the physical consistency check process is initiated. If the check passes, the cluster is included in the risk mapping system; if it fails, it is marked as a cluster to be observed and will not participate in the scoring. The system also has a built-in performance evaluation module that calculates the cluster profile coefficient, the root mean square of intra-cluster distance, and the variance of the risk score daily. If the profile coefficient is less than 0.5 for three consecutive days, the model re-initialization process is triggered, the current clustering model is cleared, and the model is retrained with the data of the most recent 72 hours.

[0014] Beneficial effects This invention provides an intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data. Compared with existing technologies, it has the following advantages: 1. This intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data completely eliminates the dependence on manually labeled data by constructing a physical mechanism embedded unsupervised clustering architecture. It achieves fully automatic, adaptive, and high-precision early warning from raw sensor data to risk level output. Its technical solution has significant innovation and engineering practicality in feature engineering, clustering algorithm, risk mapping, and system integration. It effectively solves the core contradictions of existing technologies, such as weak generalization ability, high false alarm rate, and dependence on expert labeling, and provides reliable technical support for intelligent safety monitoring in coal mines.

[0015] 2. This intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data uses a multi-threaded parallel processing architecture to independently execute the entire process of feature extraction, cluster analysis, and risk mapping for each borehole. The early warning results of each borehole are independent of each other and do not interfere with each other. The ground monitoring center can centrally monitor and prioritize the risk status of all boreholes and supports multi-dimensional filtering and statistics by risk level, borehole location, and work team.

[0016] 3. This intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data outputs the identification code of the current cluster when the system triggers an early warning, which facilitates post-event traceability and pattern analysis. The ground monitoring center database automatically records all cluster identification codes and their corresponding original feature vectors, risk scores, timestamps, and borehole numbers to build a risk pattern knowledge base, providing data support for subsequent process optimization and geological prediction. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart illustrating the workflow of the embedded unsupervised clustering module based on the physical mechanism of this invention. Figure 3This is a logic block diagram of the dynamic risk mapping and hierarchical early warning module of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-3 This intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data provides a technical solution: It includes: a multi-source sensor data acquisition module, an edge preprocessing and feature engineering module, a physical mechanism embedded unsupervised clustering module, a dynamic risk mapping and hierarchical early warning module, and a human-computer interaction and decision support module; the multi-source sensor data acquisition module is deployed in the drilling rig operation area to collect three types of time-series data in real time: gas pressure, drilling parameters and environmental status, and transmits them synchronously to the edge computing node through an industrial Ethernet bus or CAN bus; The multi-source sensor data acquisition module consists of a gas pressure sensor array, a drilling parameter sensing unit, and an environmental condition monitoring unit. The gas pressure sensor array includes six high-precision piezoresistive pressure sensors, which are installed at the drill pipe cavity near the drill bit end, the middle of the drill pipe, the tail end of the drill pipe, the inside of the wellhead sealing device, the inlet of the extraction pipeline, and the wellhead return gas channel, respectively. The range is 0–5 MPa, the accuracy is 0.1%FS, and the sampling frequency is 100Hz. The drilling parameter sensing unit includes a torque sensor, a speed encoder, a thrust sensor, and a drilling speed meter, which measure the drill rig spindle torque, hydraulic motor speed, propulsion cylinder thrust, and drill pipe axial displacement rate, respectively. The environmental condition monitoring unit includes a methane concentration sensor, a temperature sensor, a humidity sensor, and an atmospheric pressure sensor, which are installed within 5 meters of the drilling face. All sensor signals are fed into the edge computing node via a 16-channel synchronous data acquisition card, with a timestamp synchronization error of <1m.

[0020] The edge preprocessing and feature engineering module performs sliding median filtering, cubic spline interpolation imputation, and zero-mean unit variance standardization on the original data. Based on a 60-second sliding window, it extracts 28-dimensional dynamic features, including pressure mean, standard deviation, skewness, kurtosis, maximum and minimum values, rise and fall rates, spectral energy entropy, wavelet packet decomposition coefficient energy ratio, autocorrelation and cross-correlation function lag coefficients, pressure gradient change rate, and pressure fluctuation frequency. Then, it performs principal component analysis to reduce the dimensionality and retain the top 8 principal components with a cumulative variance contribution rate of ≥95% as the output feature vector. The physical mechanism embedded unsupervised clustering module adopts an improved constrained K-means algorithm. Its initial cluster centers are preset into three categories based on gas dynamics theory: stable low-pressure zone, fluctuating medium-pressure zone, and high-pressure abrupt change zone. During the iteration process, physical consistency indicators are introduced, including pressure gradient consistency index, spectral structure similarity, and cross-correlation stability coefficient with drilling speed sequence. These drive the cluster splitting mechanism or cluster merging mechanism to ensure that the final clusters have high cohesion and inter-class separability in both statistical distribution and physical mechanism dimensions.

[0021] The system defaults to using the physically constrained K-means algorithm, but can also switch to DBSCAN density clustering or Gaussian mixture model clustering, which are suitable for constructing complex abrupt change scenarios and thick coal seam multi-peak distribution scenarios, respectively. Users can customize the calculation formulas for new features through configuration files, such as the second derivative of pressure, pressure-drilling speed phase difference, and specific frequency band energy ratio. New features automatically participate in principal component analysis, and whether to retain them is determined based on the variance contribution rate. The risk score weight coefficients α, β, γ, and δ can be finely adjusted online at the ground end, with a range of ±0.1 and a step size of 0.01. Adjustments take effect immediately without restarting. The system interfaces with the mine automation platform through the OPC UA protocol interface to receive external information such as extraction efficiency and ventilation status to assist in decision-making, and uses an AES-256 encryption module to ensure data transmission security.

[0022] The dynamic risk mapping and hierarchical early warning module receives clustering results and processes them according to the formula: R=α·P mean + β·dP max / dt + γ·H entropy + δ·N spike Risk scores are calculated, where α, β, γ, and δ are weighted coefficients, and α+β+γ+δ=1. After score normalization, the scores are divided into safety level, attention level, warning level, and danger level according to thresholds. The cluster evolution trajectory is monitored to trigger dynamic upgrade rules. The human-machine interaction and decision support module provides a visual interface on both the downhole explosion-proof display screen and the ground monitoring center. When the risk level reaches the warning level or above, the system automatically calls the process parameter adjustment strategy library to generate control commands and sends them to the drilling rig control system for execution via the CAN bus. When the system triggers a warning, it simultaneously outputs the identification code of the current cluster for easy traceability and pattern analysis. The ground monitoring center database automatically records all cluster identification codes and their corresponding original feature vectors, risk scores, timestamps, and borehole numbers to build a risk pattern knowledge base, providing data support for subsequent process optimization and geological prediction.

[0023] The physical mechanism embedded unsupervised clustering module sets three types of physical constraints for cluster centers during the cluster initialization phase: stable low-pressure regions satisfying a pressure mean <0.3MPa and a standard deviation <0.05MPa; fluctuating medium-pressure regions satisfying a pressure mean of 0.3–1.2MPa and a standard deviation of 0.05–0.2MPa; and high-pressure abrupt change regions satisfying a pressure mean >1.2MPa or a pressure rise rate >0.5MPa / min or a spectral energy entropy >1.8. During the iteration process, if the physical consistency index of a cluster is lower than the preset threshold, i.e., pressure gradient consistency index <0.7, spectral structure similarity <0.6, and cross-correlation stability coefficient <0.5, a cluster splitting mechanism is triggered, dividing the samples into two groups according to the median pressure rise rate and recalculating the sub-cluster centers; if the physical consistency of two adjacent clusters meets the standard and the cluster distance is less than 0.3 times the average intra-cluster distance, a cluster merging mechanism is triggered, and the new cluster center is the weighted average of the original cluster centers.

[0024] In the risk scoring function used, P mean dP is the average pressure within the cluster. max / dt is the maximum rate of pressure rise within the cluster, H entropy N represents the mean spectral energy entropy within the cluster. spike This represents the average number of pressure spikes per unit time within a cluster. This module is also configured with a dynamic upgrade trigger: if a "safe" cluster experiences a continuous increase in risk score with an amplitude >0.1 within five consecutive time windows, it is automatically upgraded to "attention level"; if a warning-level cluster experiences a score jump >0.2 within a single time window, a danger-level alarm is immediately triggered. The risk level classification results are synchronously output with the corresponding cluster identification code C-XXX, where XXX is a three-digit number ordered according to the cluster center's position in the feature space.

[0025] The human-machine interaction and decision support module displays the current borehole location, six-channel gas pressure curve, three-dimensional projection distribution map of clusters, four-color risk indicator lights, and text warning information in real time at the downhole end; and provides historical risk evolution heat map, multi-bore risk comparison bar chart, and cluster center migration trajectory map at the surface end; when the risk level is determined to be warning level or dangerous level, it automatically matches a scheme from the pre-stored 12 process parameter adjustment strategy library, generates a CAN bus control frame containing target drilling speed, target thrust, pause command or water injection command, and sends it to the drilling rig PLC for execution, with a command response delay of <200ms.

[0026] Edge computing nodes cache the feature vectors of the most recent 24 hours and perform incremental K-means fine-tuning every hour. If the distance between a new data point and all existing cluster centers is greater than 1.5 times the average intra-cluster distance, a new cluster is created and a physical consistency check process is initiated. If the check passes, the cluster is included in the risk mapping system; if it fails, it is marked as a cluster to be observed and will not participate in the scoring. The system also has a built-in performance evaluation module that calculates the cluster profile coefficient, root mean square of intra-cluster distance, and variance of risk score daily. If the profile coefficient is less than 0.5 for 3 consecutive days, the model re-initialization process is triggered, the current clustering model is cleared, and it is retrained with data from the most recent 72 hours.

[0027] The edge computing nodes are configured with dual power inputs and dual eMMC storage cards for hot standby, and the data acquisition cards adopt a dual-channel synchronous architecture. The software adopts a dual-process daemon mechanism, with the main process responsible for real-time clustering and the daemon process monitoring heartbeats and restarting in case of abnormalities. Communication uses dual channels of TCP / IP and CAN bus, with data packets appended with CRC32 checksums and serial numbers. The system supports offline operation mode, and when communication is interrupted, local storage of early warning records lasts for up to 72 hours, which are automatically uploaded after recovery. User access management adopts a three-level RFID authentication system, assigning different functional permissions to operators, engineers, and administrators respectively.

[0028] This invention completely eliminates the reliance on manually labeled data by constructing a physical mechanism-embedded unsupervised clustering architecture, achieving fully automatic, adaptive, and high-precision early warning from raw sensor data to risk level output. Its technical solution has significant innovation and engineering practicality in feature engineering, clustering algorithm, risk mapping, and system integration. It effectively solves the core contradictions of existing technologies, such as weak generalization ability, high false alarm rate, and reliance on expert labeling, and provides reliable technical support for intelligent safety monitoring in coal mines.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A coal seam drilling intelligent early warning system based on unsupervised learning of gas pressure data, characterized in that: include: The system includes a multi-source sensor data acquisition module, an edge preprocessing and feature engineering module, a physical mechanism embedded unsupervised clustering module, a dynamic risk mapping and hierarchical early warning module, and a human-computer interaction and decision support module. The multi-source sensor data acquisition module is deployed in the drilling rig's operating area to collect three types of time-series data in real time: gas pressure, drilling parameters, and environmental status. The data is then synchronously transmitted to the edge computing node via an industrial Ethernet bus or CAN bus. The edge preprocessing and feature engineering module performs sliding median filtering, cubic spline interpolation and imputation, and zero-mean unit variance standardization on the original data. Based on a 60-second sliding window, it extracts 28-dimensional dynamic features, including pressure mean, standard deviation, skewness, kurtosis, maximum and minimum values, rise and fall rates, spectral energy entropy, wavelet packet decomposition coefficient energy ratio, autocorrelation and cross-correlation function lag coefficients, pressure gradient change rate, and pressure fluctuation frequency. Then, it performs principal component analysis to reduce the dimensionality and retain the top 8 principal components with a cumulative variance contribution rate of ≥95% as the output feature vector. The physical mechanism embedded unsupervised clustering module adopts an improved constrained K-means algorithm. Its initial cluster centers are preset into three categories based on gas dynamics theory: stable low-pressure zone, fluctuating medium-pressure zone, and high-pressure abrupt change zone. During the iteration process, physical consistency indicators are introduced, including pressure gradient consistency index, spectral structure similarity, and cross-correlation stability coefficient with drilling speed sequence. These drive the cluster splitting mechanism or cluster merging mechanism to ensure that the final clusters have high cohesion and inter-class separability in both statistical distribution and physical mechanism dimensions.

2. The intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data according to claim 1, characterized in that: The dynamic risk mapping and hierarchical early warning module receives clustering results and calculates them according to the formula: R=α·P mean + β·dP max / dt + γ·H entropy + δ·N spike A risk score is calculated, where α, β, γ, and δ are all weighting coefficients, and α+β+γ+δ=1. After score normalization, the scores are divided into safety level, attention level, warning level, and danger level according to the threshold. The cluster evolution trajectory is monitored to trigger dynamic upgrade rules. The human-computer interaction and decision support module provides a visual interface on both the downhole explosion-proof display screen and the ground monitoring center. When the risk level reaches the warning level or above, the process parameter adjustment strategy library is automatically called to generate control commands and sent to the drilling rig control system for execution via CAN bus.

3. The intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data according to claim 1, characterized in that: The physical mechanism embedded unsupervised clustering module sets three types of physical constraints for cluster centers during the cluster initialization phase: stable low-pressure regions satisfy a pressure mean <0.3MPa and a standard deviation <0.05MPa; fluctuating medium-pressure regions satisfy a pressure mean of 0.3–1.2MPa and a standard deviation of 0.05–0.2MPa; and high-pressure abrupt change regions satisfy a pressure mean >1.2MPa or a pressure rise rate >0.5MPa / min or a spectral energy entropy >1.

8.

4. The intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data according to claim 1, characterized in that: During the iterative process of the embedded unsupervised clustering module of the physical mechanism, if the physical consistency index of a cluster is lower than the preset threshold, i.e., the pressure gradient consistency index < 0.7, the spectral structure similarity < 0.6, and the cross-correlation stability coefficient < 0.5, the cluster splitting mechanism is triggered, and the samples are divided into two groups according to the median pressure rise rate and the sub-cluster centers are recalculated; if the physical consistency of two adjacent clusters meets the standard and the cluster distance is less than 0.3 times the average intra-cluster distance, the cluster merging mechanism is triggered, and the new cluster center is the weighted average of the original cluster centers.

5. The intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data according to claim 2, characterized in that: In the risk scoring function used, P mean dP is the average pressure within the cluster. max / dt is the maximum rate of pressure rise within the cluster, H entropy N represents the mean spectral energy entropy within the cluster. spike This represents the average number of pressure jumps per unit time within the cluster.

6. The intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data according to claim 1, characterized in that: The physical mechanism embeds an unsupervised clustering module and is configured with a dynamic upgrade trigger: if the risk score of a safety-level cluster continues to rise within 5 consecutive time windows and the magnitude is >0.1, it will be automatically upgraded to the attention level; if the score of a warning-level cluster jumps to >0.2 within a single time window, a danger-level alarm will be triggered immediately; the risk level classification result will synchronously output the corresponding cluster identification code C-XXX, where XXX is a three-digit number, and the number is sorted according to the cluster center's position in the feature space.

7. The intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data according to claim 1, characterized in that: The human-computer interaction and decision support module displays the current borehole location, six-channel gas pressure curve, three-dimensional projection distribution map of clusters, four-color risk indicator lights and text warning information in real time at the downhole end; and provides historical risk evolution heat map, multi-bore risk comparison bar chart and cluster center migration trajectory map at the surface end. When the risk level is determined to be warning level or dangerous level, the system automatically matches a scheme from the pre-stored library of 12 process parameter adjustment strategies, generates a CAN bus control frame containing the target drilling speed, target thrust, pause command or water injection command, and sends it to the drilling rig PLC for execution, with a command response delay of <200ms.

8. The intelligent early warning system for coal seam drilling based on unsupervised learning of gas pressure data according to claim 1, characterized in that: Edge computing nodes cache feature vectors from the past 24 hours and perform incremental K-means fine-tuning every hour. If the distance between a new data point and all existing cluster centers is greater than 1.5 times the average intra-cluster distance, a new cluster is created and a physical consistency check process is initiated. If the check passes, the cluster is included in the risk mapping system; otherwise, it is marked as a cluster to be observed and not included in the scoring. The system also has a built-in performance evaluation module that calculates the cluster profile coefficient, root mean square of intra-cluster distance, and variance of risk score daily. If the profile coefficient is less than 0.5 for three consecutive days, a model re-initialization process is triggered, clearing the current clustering model and retraining it with data from the past 72 hours.

Images