Gas risk self-adaptive inspection method and system based on digital twin driven extraction drilling group

Abstract

The application discloses a kind of based on digital twin driving's extraction borehole group gas risk self-adapting inspection method and system, it constructs four-in-one technical system of multimodal perception, space-time collaborative optimization, gas path self-consistent control and semantic digital twin: based on space-time graph attention network realizes the accurate quantitative evaluation of borehole group gas emission risk;Based on deep reinforcement learning realizes different time-space multiplexing inspection, so that inspection resources dynamically tilt to high-risk areas;Adopt three-level tree-shaped gas path topology combined with adaptive pressure balance and pulse type self-cleaning technology, realize large-scale borehole group without cross-contamination time multiplexing by single detection unit;Build semantic digital twin, realize full-element real-time mapping of physical space and virtual space and closed-loop management of abnormal events from perception to disposal.The present application can be flexibly adapted to different scale extraction borehole group, without relying on specific sensing equipment, significantly improve the detection frequency of high-risk borehole and gas abnormal response speed.

Description

Technical Field

[0001] This invention relates to the field of intelligent prevention and control technology for coal mine gas disasters, specifically to an adaptive inspection method and system for gas risk of extraction borehole groups based on digital twin-driven technology. Background Technology

[0002] Gas drainage is a core technical measure for preventing major accidents such as coal mine gas explosions, coal and gas outbursts, and spontaneous combustion in goafs. As the terminal sensing unit of the gas drainage system, the real-time monitoring of parameters such as gas concentration, drainage negative pressure, gas flow rate, temperature, and associated gas composition in drainage boreholes is crucial for evaluating drainage effectiveness, identifying early signs of gas disasters, and optimizing drainage processes.

[0003] Currently, the detection of coal mine drainage borehole parameters both domestically and internationally mainly relies on the following three modes: (1) Manual inspection mode. Gas inspectors carry portable optical gas detectors or chromatographic analyzers to conduct regular borehole inspections and manually record data. This mode has inherent defects such as high labor intensity, low efficiency, strong data lag (usually on an hourly or shift basis), and large human reading errors. More seriously, inspectors need to work in harsh environments with high gas, high dust, and strong noise for a long time, facing multiple occupational safety risks such as poisoning, suffocation, gas explosion, and roof collapse, which is difficult to meet the needs of modern coal mine intelligent construction.

[0004] (2) Fixed sensor monitoring mode. Single-point monitoring devices such as methane sensors and flow sensors are installed at each borehole and transmitted to the ground monitoring center via cable or wireless means. Although this mode can achieve real-time acquisition of parameters, it has problems such as high equipment cost (comprehensive cost of thousands to tens of thousands of yuan per borehole), complex downhole wiring, huge maintenance workload, and frequent calibration due to sensor drift. More importantly, this mode adopts a rigid "one borehole, one sensor" architecture, and the inspection mode is rigid. It cannot dynamically adjust the detection frequency according to the borehole risk level, resulting in insufficient monitoring of high-risk boreholes and excessive monitoring of low-risk boreholes, with extremely low resource utilization.

[0005] (3) Simple polling detection mode. Some mines have tried to achieve multiple holes sharing a set of sensors by switching solenoid valves, but it lacks risk self-adaptation capability, only mechanically polls according to a fixed period, and has not solved the problem of cross-contamination of gas path, so the detection accuracy is difficult to guarantee; at the same time, this mode lacks data visualization and abnormal closed-loop handling capabilities, and the response to abnormal events is passive and lagging, and it is impossible to achieve full-process traceability and intelligent analysis.

[0006] With the deepening of intelligent construction in coal mines, the traditional model mentioned above can no longer meet the inherent safety requirements of "safety with fewer people, safety without people". There is an urgent need for an intelligent inspection method and system that can dynamically and adaptively allocate inspection resources based on borehole risks, achieve high-precision detection without cross-contamination, and have full life-cycle visualized management and closed-loop handling capabilities for anomalies. Summary of the Invention

[0007] This invention provides a digital twin-driven adaptive inspection method and system for gas risk in a group of extraction boreholes. The purpose is to overcome the shortcomings of existing technologies, such as rigid inspection modes, single perception dimensions, inefficient allocation of spatiotemporal resources, serious cross-contamination of gas paths, delayed response to anomalies, and lack of full life-cycle visualization and control.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An adaptive inspection method for gas risk in a group of extraction boreholes based on digital twins is characterized by the following steps: Step S1, Multi-scale Risk Dynamic Assessment Step: Construct a dynamic risk assessment model for the drilling borehole group based on a spatiotemporal graph attention network. Collect multimodal operating parameters and geological environment parameters of each borehole in real time. Through hierarchical processing of spatiotemporal embedding layer, graph attention layer, temporal convolutional layer and fully connected output layer, calculate the gas outburst risk level of each borehole. The input of the graph attention layer includes the topological adjacency matrix of the borehole group. The adjacency matrix is ​​constructed based on the spatial distance of the boreholes, the correlation of geological strata and the co-occurrence frequency of historical anomalies. Step S2, Spatiotemporal Cooperative Inspection Planning Step: Based on the risk level, a deep reinforcement learning algorithm is used to dynamically optimize the inspection time slice allocation strategy and generate a spatiotemporal reuse inspection plan. The state space of the deep reinforcement learning algorithm is the current risk distribution and gas path load status of the borehole group, the action space is the inspection time slice allocation and gas path switching sequence of each borehole, and the reward function comprehensively considers risk coverage, detection frequency balance and gas path switching loss. Step S3, Gas Path Intelligent Scheduling and Self-Cleaning Detection Step: According to the inspection plan, the downhole edge intelligent control unit schedules multiple gas path switching units, adopts a three-level tree topology structure of "main path-branch path-bypass path", and uses the same group of multi-parameter detection units to complete the gas parameter detection of all extraction boreholes in a time-division multiplexing manner. The multi-path gas switching unit connects only one branch to the main path at any given time, and performs adaptive pressure balance control during the switching process of each branch. After the single-hole test is completed, a pulsed gas path self-cleaning process is performed: the connection between the branch and the extraction borehole is closed, the connection between the branch and the clean gas source is opened, and clean air is introduced into the branch in a pulsed airflow manner for 3-5 seconds. The pulse frequency and duration are dynamically adjusted based on the feedback of the residual gas concentration in the branch until the residual gas concentration in the pipeline is lower than the preset threshold. Step S4, Semantic Digital Twin Mapping and Visualization Step: The detection data is mapped to the ground semantic digital twin in real time through the Industrial Internet of Things. The digital twin uses multi-level detail modeling technology to construct a three-dimensional geometric model of the extraction borehole group, and binds borehole attributes, operating status and historical event information based on semantic annotation technology to realize three-dimensional visualization of borehole status, full life cycle data traceability and cross-scale status inference. Step S5, Abnormal Event Closed-Loop Handling Steps: When the detection data exceeds the preset multi-level warning threshold, the abnormal borehole encrypted inspection mechanism and emergency response plan are automatically triggered, the abnormal borehole is dynamically upgraded to an extremely high risk level, the low-risk borehole inspection task is suspended, and the abnormal borehole is continuously detected with priority.

[0009] A gas inspection system for a group of extraction boreholes implementing the method adopts a four-level collaborative architecture of "cloud-edge-device-field", specifically including: The ground-based digital twin control unit, deployed in the ground dispatch center, consists of a high-performance industrial server, a GPU-accelerated workstation, a time-series database, a relational database, and a 3D visualization engine. It is used to construct a semantic digital twin of the extraction borehole group and perform global risk assessment, inspection plan generation, data visualization, abnormal event management, report generation, and multi-system collaborative scheduling. The underground edge intelligent control unit is fixedly installed in the extraction roadway chamber. It adopts an intrinsically safe explosion-proof design for mining and integrates a multi-core ARM processor, an NPU neural network acceleration unit, a large-capacity solid-state storage array, and a dual-frequency wireless communication module. It is used to receive ground commands and control gas path switching and data acquisition, perform local risk assessment model inference, edge optimization of inspection plans, data preprocessing and feature extraction, and trigger local audible, visual and voice alarms. The solid-state storage array supports offline caching of ≥90 days of detection data and automatically synchronizes based on a breakpoint resume mechanism after network recovery. The multi-path gas switching unit is electrically connected to the downhole edge intelligent control unit via a CAN bus and an intrinsically safe I / O interface, and adopts a modular pluggable design. It consists of a pilot-operated explosion-proof electromagnetic switching valve group, a precision pressure regulating valve group, a pulse-type clean gas source interface, a pressure balance bypass, and stainless steel / PTFE composite pipelines, and adopts a three-level tree topology structure of "main path-branch path-bypass path". The number of branches can be flexibly expanded according to the number of extraction boreholes, and a single system can support the inspection of 10-1000 extraction boreholes. The multi-parameter detection unit is connected to the main path of the multi-path gas switching unit and integrates a laser methane sensor, an infrared carbon monoxide / carbon dioxide sensor, an electrochemical oxygen sensor, a differential pressure sensor, a thermal mass flow sensor, and a digital temperature and humidity sensor. The laser methane sensor has a detection accuracy of ≤±0.5%FS and a response time of ≤10s. Each sensor adopts a modular rail design, supporting hot-swappable replacement and automatic calibration. The on-site human-machine interaction unit communicates with the underground edge intelligent control unit and adopts a hybrid display scheme of intrinsically safe electronic ink screen and LED array. It is suspended on the pedestrian side of the roadway and is used to display the number, parameters, inspection time, risk level color mark and alarm information of the currently detected borehole in real time. It also supports touch query of the status of nearby boreholes. The industrial communication unit connects the ground digital twin control unit and the downhole edge intelligent control unit respectively, and supports multiple communication methods such as gigabit fiber optic, industrial Ethernet, Wi-Fi 6 and 5G NR; the communication protocol stack includes MQTT, OPC UA and Modbus TCP, and supports TSN time-sensitive networking to meet the deterministic transmission of control commands; The intrinsically safe power supply unit adopts a mining explosion-proof and intrinsically safe DC regulated power supply, providing ≥8 channels of DC18V intrinsically safe output; it has a built-in lithium iron phosphate battery pack with a power outage duration of ≥8 hours.

[0010] The present invention has the following advantages: (1) Order-of-magnitude improvement in inspection efficiency and response speed. Through spatiotemporal collaborative optimization based on deep reinforcement learning, the detection frequency of high-risk boreholes can be increased by more than 10 times (from once per shift in traditional manual inspection to once every 5-10 minutes), and the response time for gas anomalies can be compressed from the traditional hour level to the minute level (as fast as the second level), which greatly improves the early warning capability for gas disasters.

[0011] (2) Extreme optimization of resource utilization and significant cost reduction. Without increasing hardware costs, the time-division multiplexing of up to 1,000 boreholes by a single multi-parameter detection unit reduces sensor and wiring costs by more than 80% compared to the traditional single-point sensor solution. Through differentiated inspection strategy, invalid detection of low-risk boreholes is avoided, and limited detection resources are accurately deployed to high-risk areas.

[0012] (3) A leapfrog improvement in detection accuracy and reliability. The pulsed gas path self-cleaning technology combined with adaptive pressure balance control completely eliminates cross-contamination and pressure fluctuation interference between different borehole gases, improving detection accuracy by more than 20%; the adoption of the laser TDLAS methane sensor fundamentally solves the problems of poisoning, drift and humidity interference of traditional catalytic combustion sensors.

[0013] (4) Upgrading of intelligent and inherently safe management and control models. Semantic digital twin technology has enabled the drilling group to leap from "two-dimensional reports" to "three-dimensional visualization", from "single-point monitoring" to "full life cycle traceability", and from "passive alarm" to "active simulation"; the closed-loop handling of abnormal events has reduced the cost of manual intervention by more than 70%; and unmanned underground inspections have completely eliminated the occupational safety risks of personnel entering high-risk areas.

[0014] (5) Excellent performance in system elasticity and scale adaptability. The modular gas path switching unit supports elastic expansion of 10-1000 boreholes, the "cloud-edge-end-field" four-level architecture supports smooth upgrade from a single working face to the entire mine, and the multi-system collaborative scheduling algorithm realizes global optimal control of large-scale borehole groups, which is suitable for coal mine scenarios with different gas levels, different geological conditions and different production scales. Attached Figure Description

[0015] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0016] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0017] Figure 1 This is a flowchart of a method for adaptive inspection of gas risk in a group of extraction boreholes based on digital twin-driven technology, according to the present invention. Figure 2 This is a complete architectural diagram of the ST-GAT model of the present invention, showing the complete data flow from the input layer to the risk output layer; Figure 3 This is a schematic diagram illustrating the PPO-based inspection time slice allocation optimization principle of the present invention; Figure 4 This is a schematic diagram of the three-stage gas path topology of the present invention, namely "main path-branch path-bypass", and the pulse self-cleaning principle. Figure 5 This is a schematic diagram of the multi-layered architecture and visualization rendering of the semantic extraction borehole group digital twin of the present invention; Figure 6This is a flowchart illustrating the closed-loop handling process of abnormal events in this invention, encompassing "perception-diagnosis-decision-execution-verification". Figure 7 This is a diagram of the "cloud-edge-device-field" four-level collaborative system architecture of this invention; Figure 8 This is a schematic diagram of the dynamic load balancing and regional anomaly association scheduling of multi-system collaborative inspection in this invention. Detailed Implementation

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

[0019] like Figure 1 As shown, an adaptive inspection method for gas risk in a group of extraction boreholes based on digital twins includes the following steps: Step S1: Multi-scale dynamic risk assessment. Construct a dynamic risk assessment model for the extraction borehole group based on the spatiotemporal graph attention network (ST-GAT). Collect multimodal operating parameters and geological environment parameters of each borehole in real time. Through hierarchical processing of spatiotemporal embedding layer, graph attention layer, temporal convolution layer and fully connected output layer, calculate the gas outburst risk level of each borehole.

[0020] Step S1 Input: Input [N, T, F], where N is the total number of extraction boreholes (i.e., the number of boreholes connected to the system), T is the time step (i.e., the number of time series data points processed by the model in one step), and F is the parameter dimension.

[0021] Multimodal operating parameters (8 dimensions): gas concentration sequence, rate of change, negative pressure, flow rate, temperature, humidity, CO, CO2; Geological environmental parameters (6 dimensions): borehole depth, dip angle, coal seam gas content, coal seam thickness, geological structure complexity (fault density), and surrounding rock permeability coefficient.

[0022] like Figure 2 As shown, the logic of step S1 includes: Step S11, Spatiotemporal Embedding Layer: The 14-dimensional raw data (8-dimensional multimodal operating parameters and 6-dimensional geological environment parameters) are mapped to a high-dimensional space through linear transformation, and time-location encoding is superimposed to output a high-dimensional structured vector [N, T, ... The time sequence is marked using positional encoding (telling the model which comes first, minute 1 or minute 720); a fully connected linear layer maps the 14-dimensional original features to a high-dimensional space (taking 128 dimensions as an example below, the high-dimensional structured vector [N, T, ... ]).

[0023] Step S12, Graph Attention Layer: Receives the output from step S11 [N, T, ... Construct an adjacency matrix A (an N×N association table). For each borehole node, use multi-head attention to aggregate the features of associated nodes in adjacency matrix A, outputting [N, T, ... (At this point, the characteristics of each node have been integrated with "the transmission of neighbor anomalies under the influence of its own state and geological structure"). Specifically, firstly, the borehole group topological adjacency matrix A∈ is constructed. (N×N), where N is the number of boreholes; matrix elements It consists of three weighted parts: drilling i and j The calculations include the spatial distance attenuation function (using a Gaussian kernel function), geological stratigraphic correlation (based on the co-occurrence ratio of borehole trajectories in the coal seam), and historical anomaly co-occurrence frequency (statistically counting the number of times two boreholes simultaneously or successively experienced anomalies in the past 30 days). In this step, geological structural complexity (fault density) directly participates in the calculation; the higher the fault density, the lower the weight of the spatial distance attenuation function of adjacent boreholes (the stronger the correlation). At the same time, the surrounding rock permeability coefficient is incorporated into the calculation of the attention vector as a static feature of the node.

[0024] Step S13, Temporal Convolutional Layer: Continuing from the output [N, T, 128] in Step S12, a Causal Dilated CNN is used to capture the temporal dependencies of the drilling parameters. The convolutional kernels are slid along the time axis (dilated convolutions expand the receptive field) to capture long-term evolutionary trends over the past 72 hours (e.g., "continuously climbing over the past 3 hours"), outputting spatiotemporal fusion features [N, 1, 128]. This technique's temporal convolutional layer uses a Causal Dilated CNN to capture the temporal dependencies of the parameters. By stacking multiple layers of dilated convolutions, it captures long-term temporal patterns with an exponentially growing receptive field while maintaining causality (without revealing future information).

[0025] Step S14, Fully Connected Output Layer: The fully connected layer maps to 4 dimensions, outputting extremely high / high / medium / low risk probabilities. A fully connected classifier (FC layer) is connected to map the 128-dimensional features to 4 dimensions (extremely high / high / medium / low), and then the Softmax function is used to convert them into probabilities (the sum of the four probabilities equals 1), outputting the final risk label [N, 4] (each borehole obtains the probability of each of the four risk levels, and the highest probability is taken as the level). In this technique, the fully connected output layer inputs the spatiotemporal embedding vector into the fully connected network, and the Softmax classifier outputs four risk levels: extremely high, high, medium, and low, along with the confidence probability of each level. The risk level is updated every 5 minutes. Update triggering conditions include: timed triggering, triggering when any borehole parameter mutation exceeds 20%, and triggering triggering when an anomaly propagates in a neighboring borehole (when an anomaly occurs in the neighborhood of a borehole, that borehole immediately triggers an evaluation update).

[0026] In step S1, step S11 prepares a "unified dialogue language." Since the original multimodal operation involved geological environmental parameters with different units, step S11 transforms them all into standard-dimensional (128-dimensional) feature vectors for subsequent graph attention calculations. Step S12 overlays surrounding environmental information on this basis, ensuring that the gas value detected by each borehole at the current moment is no longer isolated but carries the common characteristics of boreholes near the fault and boreholes related to the stratigraphy. Step S13 condenses the 72-hour long sequence with spatial information from step S12 into a 128-dimensional comprehensive trend at the current instant. Without the spatial aggregation in step S12, the historical information seen in step S13 would be inaccurate; without the temporal condensation in step S13, step S12 would only see the current neighbor status, missing the abnormal precursors of neighbors in the past few hours. In summary, step S11 provides standardized data, step S12 incorporates the influence of spatial neighbors, step S13 condenses historical trends, and step S14 integrates the results of these three steps to provide the final risk decision.

[0027] Step S2: Spatiotemporal Coordinated Inspection Planning. Based on the risk level, a deep reinforcement learning algorithm is used to dynamically optimize the inspection time slice allocation strategy and generate a spatiotemporally reused inspection plan. The state space of the deep reinforcement learning algorithm is the current risk distribution and gas path load status of the borehole group, and the action space is the inspection time slice allocation and gas path switching sequence of each borehole. The reward function comprehensively considers risk coverage, detection frequency balance, and gas path switching loss, so that high-risk boreholes receive high-frequency short-cycle inspection resources, and low-risk boreholes receive low-frequency long-cycle inspection resources, thereby achieving differentiated inspection resource allocation.

[0028] Step S2 Input: The risk probability matrix [N, 4] output by step S1, where N is the total number of boreholes and each borehole corresponds to 4 probability values ​​(extremely high / high / medium / low).

[0029] Real-time system operating status (from downhole edge control unit and sensors): time interval (minutes) between the last inspection of each borehole, ranging from 0 to 60 minutes; current gas path connectivity status (which borehole number is currently being inspected, or is it in an idle state); cumulative runtime of multi-parameter detection unit (for life management, unit: hours); downhole ambient temperature and humidity (from environmental sensors).

[0030] like Figure 3 As shown, the logic of step S2 includes: Step S21: State Encoding (Feature Concatenation and Normalization) Risk level coding: Extract the level with the highest probability from the [N, 4] probability matrix and convert it into a 4-dimensional One-hot code (Extremely High = [1,0,0,0], High = [0,1,0,0], Medium = [0,0,1,0], Low = [0,0,0,1]). This is equivalent to "hard-determining" the probability distribution as a level label.

[0031] Interval normalization: Divide the time interval between each borehole inspection and the last inspection (original value 0~60 minutes) by the upper limit of the time interval range (60 minutes), normalizing it to the 0~1 range. The purpose is to eliminate the difference in units of measurement, allowing the neural network to treat inputs of different magnitudes equally.

[0032] Other state normalizations: sensor lifespan (cumulative operating time divided by design lifespan) and ambient temperature / humidity are both normalized to the 0~1 range using Min-Max.

[0033] Vector concatenation: Concatenate all the following values ​​in a fixed order into a one-dimensional vector s[t]: Positions 1 through N: Risk One-hot code for each borehole (4 × N values ​​in total); Positions 4N+1 to 5N: Normalized interval duration of each borehole (N values ​​in total). The 5N+1th position: Current gas path conduction status (represented by the normalized value of the borehole number); The 5N+2nd position: sensor normalized lifetime; The 5N+3rd position: Normalized ambient temperature; The 5N+4th position: Normalized ambient humidity; Step S21 outputs s[t], the dimension of s[t] is [4N + N + 4] = [5N + 4], for example, when N=30, the dimension is 154.

[0034] Step S22: PPO Policy Network Inference (Action Generation) Network forward propagation: First, the output s[t] from step S21 is input into an LSTM layer (128 hidden units). The LSTM retains the hidden state h[t-1] from the previous time step to handle temporal dependencies, because the inspection decision cannot only consider the current instant but also the parts inspected in the previous round. The LSTM outputs a 128-dimensional hidden vector, which then passes through three fully connected layers (dimensions of 256, 128, and 64 respectively), each followed by a ReLU activation function. The last layer splits into two heads: The strategy head (Actor) outputs the "logits" of the dominant logarithm for each available action in the action space. The action space consists of three combinations: ① which borehole to drill (N options); ② detection duration (7 discrete values: 5s / 10s / 15s / 30s / 60s / 120s / 300s); ③ switching path (2 options: direct switch 0, or switch 1 via bypass pre-balancing). The total number of actions is N × 7 × 2 combinations.

[0035] Critic: Outputs the value estimate V(s[t]) of the current state (a scalar representing the total reward expected from this state).

[0036] Action sampling: Based on the probability distribution output by the policy head, preliminary actions are obtained through random sampling (rather than directly taking the maximum probability). The advantage of sampling is that it keeps the process exploratory and avoids the strategy from failing to evolve by selecting the same action every time.

[0037] Step S22 Output: Initial Action It contains three fields: target borehole number, detection duration level, and switch path type; The state value estimate V(s[t]) is used only for updating the network in step S24.

[0038] Step S23: Hard allocation rule clipping (motion correction) Step S23 Input: Initial action of step S22 In addition, there are fixed, hard safety rules preset by the user or the mining company. These rules include: Rule 1: Extremely high-risk boreholes (determined by S1, probability > 0.8) must be allocated the entire total inspection cycle T to achieve continuous and uninterrupted inspection.

[0039] Rule 2: Each high-risk borehole (Nh in total) shall be allocated at least T / Nh of time, and the inspection cycle shall be T / Nh.

[0040] Rule 3: Each medium-risk borehole (Nm in total) shall be allocated at least T / Nm of time.

[0041] Rule 4: All low-risk boreholes share the remaining time, but each low-risk borehole receives at least one inspection opportunity within a total cycle T.

[0042] Rule 5: The total inspection cycle T shall be set by the mine according to the gas level, ranging from 1 to 60 minutes, and shall not exceed the limit.

[0043] Step S23 logic: Action legality check: judgment Is the selected borehole of extremely high risk? If so, but its allocated duration is less than T, then it should be forcibly adjusted. = [High-risk hole ID, T, direct switching], i.e., continuous and uninterrupted detection.

[0044] Priority queue jumping logic: If If a low-risk aperture is selected, but there are still extremely high-risk apertures in the current queue that have not been continuously detected, then a forced overwrite will be performed. This is the next extremely high-risk hole.

[0045] Duration limit truncation: If If the allocated duration exceeds the maximum allowable value for that risk level (e.g., the maximum duration for a medium-risk orifice is T / Nm), then it is truncated to T / Nm.

[0046] Path switching correction: If If "direct switching" is selected, but the differential pressure sensor shows that the differential pressure between the current branch and the target branch exceeds 500Pa (from real-time data from S3), then the switch will be forcibly corrected to "switching via bypass pre-balancing" to prevent damage to the sensor.

[0047] Step S23 Output: Compliance Action It includes three compliance fields: target borehole number, compliance inspection duration, and compliance switch path; This is the final executable instruction, passed to step S3 (gas path execution). Simultaneously, The state s[t+1] (updated after execution) is passed to step S24 for calculating the reward.

[0048] Step S24: Reward Calculation and Strategy Update (Self-Evolution) Step S24 Input: Compliance action from step S23 (Decision-making before execution), execution Then, the next state s[t+1] (including the updated inspection interval, new risk level, etc.) is obtained from downhole feedback. Execution. Physical costs incurred during the process: number of gas path switching operations and switching time (from S3 feedback). Old strategy parameters from step S22. And the old value estimate V(s[t]).

[0049] Step S24 logic: Step S241: Calculate the immediate reward R (this is a key signal for PPO learning): As a risk coverage reward, if Selecting an extremely high-risk hole awards +10 points; a high-risk hole awards +5 points; a medium-risk hole awards +1 point; and a low-risk hole awards +0.1 points. If the same extremely high-risk hole is inspected consecutively, an additional +2 points are awarded each time (continuous monitoring is encouraged).

[0050] To ensure a balanced reward for detection frequency, all boreholes are scanned. If a borehole is found to have been inspected for more than 30 minutes since the last inspection (i.e., twice the interval T), a negative penalty of -1 × (interval / T) is imposed to prevent some boreholes from going undetected for a long time.

[0051] As a reward for detection efficiency, the more boreholes detected per unit time (1 minute) and the more stable the parameters (without abnormal fluctuations), the positive reward of +0.5 will be given.

[0052] As a penalty for gas path switching losses, -0.1 is deducted for each gas path switching; if the switching path is "bypass pre-balancing" (which takes longer), an additional -0.05 is deducted.

[0053] Total reward: R = α· + β· + γ· + δ· , where α, β, γ, and δ are weighting coefficients.

[0054] Step S242: Calculate the dominance function (Measure how much better the current movement is than the average movement): Time difference error , where γ=0.99 is the discount factor.

[0055] Calculated using generalized advantage estimation (GAE) Here, λ = 0.95. The larger the value, the better the action is than the average level.

[0056] Step S243: PPO-Clip policy update (correcting network parameter θ) Calculate probability ratio This measures the probability deviation between the new and old strategies for that action. (Pruning the objective function) Where ε = 0.2. The purpose of pruning is to prevent training from collapsing due to drastic policy changes. Gradient ascent is used to maximize... Update the policy network parameters θ; at the same time, update the value network parameters to make V(s[t]) closer to the actual cumulative reward.

[0057] Step S24 output: Updated policy network parameters θ (new version), used for inference in the next round of step S22; reward R and advantage for the current round. Recorded in the log (for offline analysis); the update of θ in this step achieves self-evolution, and in the next decision, the network will tend to choose actions that have historically yielded high rewards.

[0058] In step S2, the risk probability output from S1 is first concatenated with the real-time downhole status (inspection interval, gas path status, sensor lifespan, etc.) to form a standard state vector, which serves as the sole input to the subsequent deep network. This state vector is then used by the PPO policy network to generate preliminary inspection actions that balance risk coverage and resource efficiency. However, the network output may violate safety limits, so it must pass through a hard rule pruning layer to force extremely high-risk boreholes to be upgraded to continuous detection and to cut off the time exceeding the limit. Only after ensuring that all actions are absolutely compliant can they be sent to step S3 for execution. After execution, the system calculates the reward value based on the actual effect (whether high-risk boreholes are covered, whether boreholes are starved, and the size of switching losses), and updates the policy network weights in reverse through the PPO-Clip mechanism, so that the network is more inclined to high-reward behaviors in the next decision. This closed loop ensures that S2 can both maintain the safety red line through hard rules and continuously evolve through reinforcement learning, realizing adaptive inspection planning with continuous optimization of the policy model based on the running data.

[0059] Step S3: Intelligent Gas Path Scheduling and Self-Cleaning Detection Step S3 Input: Compliance action command output in step S23 (Including target borehole ID, compliance inspection duration, and compliance switching path); and real-time feedback from physical sensors, including the pipe pressure value of the currently connected branch. (Read by the main pressure sensor); Pipe pressure value of the target branch (the branch corresponding to the borehole to be switched). (Read by the pressure sensor at the end of the branch).

[0060] like Figure 4 As shown, the logic of step S3 includes: Step S31, Target Switching and Differential Pressure Sensing First, perform pressure difference calculation, calculating the absolute pressure difference ΔP between the current branch and the target branch = | - Then, a feasibility assessment is performed, comparing the calculated ΔP with a preset safety threshold (e.g., 500 Pa). Simultaneously, the switching path type in the S2 instruction ("direct switching" or "switching via bypass pre-balancing") is checked, and a switching execution strategy is generated. If ΔP exceeds the preset safety threshold, regardless of the compliant action instruction... What is it that necessitates the mandatory triggering of the pre-balancing process?

[0061] Step S31 output: differential pressure value ΔP and differential pressure status indicator ("Pre-balancing required" / "Direct switching possible").

[0062] Step S32, Adaptive pressure pre-balancing The differential pressure value ΔP and differential pressure status flag output in step S31, and the switching path type (direct / bypass) in step S23.

[0063] Step S32 logic: First, determine if the bypass is started. If ΔP > 500 Pa or the switching path is "bypass pre-balancing switch", then immediately start the bypass protection circuit. Then, slowly stabilize the pressure and open the bypass precision pressure regulating valve to slowly pressurize or depressurize the target branch at a rate of v = (ΔP - 500) / t_bal (lasting 2~5 seconds) until the real-time differential pressure drops to ΔP ≤ 500 Pa. Finally, quickly close the valve. After the differential pressure reaches the target, immediately send a command to the main line pilot-operated explosion-proof solenoid valve to complete the physical switching of the main gas line from the current branch to the target branch within 50ms.

[0064] Step S32 outputs: Main route switching completion confirmation signal ("Switched to target branch"). If pre-balancing fails (e.g., differential pressure fails to decrease for an extended period), an error is reported to the upper level and the current command is terminated.

[0065] Step S33: Fuzzy PID Flow Steady-State Control Step S33 Input: The main path switching completion confirmation signal output in step S32, as well as the physical quantity setpoint and feedback, and the detection flow rate setpoint calculated based on the current borehole extraction flow rate. (Typically, this is taken as 10% to 20% of the borehole's extraction flow rate); real-time flow sensor feedback value. .

[0066] Step S33 logic: First, perform deviation calculation, calculate the flow deviation e = - The rate of change is ec = de / dt; then adaptive fuzzy PID control is performed, inputting e and ec into the fuzzy controller, dynamically adjusting the PID parameters (ΔKp, ΔKi, ΔKd) based on the built-in fuzzy rule base (constructed based on branch pipe diameter, length, and historical pressure loss data), and finally outputting the control quantity u(t) to adjust the branch valve opening; finally, steady-state locking is judged, and continuous adjustment is performed until the real-time flow stabilizes at... If the flow rate remains within ±5% of the specified range for at least 2 seconds, it is considered a "traffic lockout".

[0067] Step S33 outputs a high-reliability measured data packet. After the flow rate is locked, the multi-parameter detection unit (laser methane sensor, infrared CO sensor, differential pressure sensor, etc.) reads all parameters of the borehole, including the actual gas concentration, CO concentration, negative pressure, flow rate, temperature, and humidity. This data packet, along with the "Detection Complete" flag, is synchronously and in parallel transmitted to steps S4 (digital twin visualization) and S5 (anomaly monitoring); the "Detection Complete" flag is then transmitted to sub-step S34.

[0068] Step S34: Pulse self-cleaning closed loop (cleaning to prevent cross-contamination) Step S34 Input: The "Detection Complete" flag output in step S33, and the residual methane concentration fed back in real time by the gas sensor at the end of the branch. and residual carbon monoxide concentration .

[0069] Step S34 logic: Upon receiving the detection completion signal, immediately close the valve connecting this branch to the extraction borehole, and simultaneously open the valve connecting this branch to the clean gas source (dry compressed air or nitrogen). Then, perform pulse oscillation purging, injecting clean gas flow into the branch in a high-frequency pulse pattern of 5-20 Hz for a base duration of 3-5 seconds, using the airflow oscillation to peel off residual gas adhering to the inner wall of the pipeline. Finally, perform closed-loop compliance judgment, and read the residual concentration at the end sensor after purging. If... >0.01% or If the concentration is >1 ppm, automatic cyclic compensation will be performed: the purging time will be extended by 2 seconds, and the pulse frequency will be increased by 5 Hz (upper limit 20 Hz), and the test will be repeated until the residual concentration meets the standard. The exhaust gas generated during the cleaning process will be discharged into the mine's main extraction pipeline through a dedicated sewage pipeline to avoid secondary pollution in the roadway.

[0070] Step S34 outputs a "Pipeline Cleanliness" compliance flag, which is sent back to the inspection scheduling module in step S2 as a "release" signal to allow the execution of the next S2 instruction. If the cleaning timeout (e.g., total duration exceeding 30 seconds) still fails to meet the standard, an equipment maintenance alarm is triggered.

[0071] In this technology, step S3 is a hard real-time execution chain that prioritizes physical safety and interlocks each other. Its internal logic is to form a detection process that interlocks the differential pressure confirmation, flow stabilization, and gas path cleanliness in sequence. Its core workflow is as follows: First, it receives the inspection instruction ("check borehole number X") issued in step S2, but never acts blindly—it must first determine whether the gas path switching conditions meet the safety threshold through differential pressure sensing (step S31). If the switching impact is too large, the bypass buffer is immediately activated (step S32) to completely eliminate the risk of physical damage to the sensor and pipeline caused by pressure fluctuations. After successfully switching to the target branch, it immediately enters the precision flow stabilization stage (step S33), and locks the airflow within a narrow window of ±5% through adaptive fuzzy PID to ensure that the gas concentration reading is not disturbed by flow rate fluctuations. Only then is the effective parameter of the borehole officially collected and distributed to the visualization and monitoring module. After the reading is completed, the gas path self-cleaning is immediately forced (step S34), and the residual gas on the pipe wall is stripped off by high-frequency pulse gas oscillation and the residual concentration is detected in a closed loop until the pipeline is "cleaned" before a release signal is issued to notify S2 that the next instruction can be issued. The four sub-steps form a unidirectional physical closed loop of "sensing → buffering → detailed reading → cleaning". If the previous step does not send a confirmation signal, the next step can never be started. This hard interlocking mechanism fundamentally eliminates the two core risks of pressure shock damaging the sensor and data distortion caused by cross-contamination of gas between holes.

[0072] Step S4: Semantic Digital Twin Mapping and Visualization Step S4 Input: High-reliability measured data package output from step S33 (including borehole ID, gas concentration, CO concentration, negative pressure, flow rate, temperature, humidity, and risk level); real-time observation distance of the 3D scene virtual camera (determined by the dispatcher's interactive operation); and static mine CAD / BIM base model (roadway, mining area outline) and borehole inclination trajectory database (azimuth, dip angle, and 3D spatial coordinates of each borehole).

[0073] like Figure 5 As shown, the logic for step S4 is as follows: Step S41: First, perform risk color mapping. Based on the risk level (extremely high / high / medium / low) in the data packet, set the rendering color of the corresponding drilling node to red (255,0,0), orange (255,165,0), yellow (255,255,0), and green (0,255,0) respectively.

[0074] Then, dynamic LOD (Level of Detail) switching is performed, automatically adjusting the geometric precision based on the distance between the virtual camera and the borehole model: when the distance is >50 meters, only planar projection points are displayed (LOD0, saving computing power); when the distance is between 10 and 50 meters, a 3D trajectory pipeline is displayed (LOD1); when the distance is <10 meters, a detailed borehole inner wall structure and screen pipe model are loaded (LOD2, supporting close-up detail viewing). The switching process is smooth and flicker-free, ensuring an overall rendering frame rate of no less than 30fps.

[0075] Finally, coordinate binding is performed, binding the numerical fields in the measured data packet to the three-dimensional spatial coordinate nodes of the corresponding borehole ID, realizing the physical mapping of "one borehole, one data".

[0076] Step S41 output: A dynamically updated set of 3D geometric nodes (each borehole is colored according to risk level, and the model accuracy adapts to the viewpoint), which is passed to sub-step S44.

[0077] Step S42: Mounting the semantic knowledge graph (attribute binding) Step S42 Input: The colored 3D geometric node output in step S41 (with borehole ID as the primary key), as well as static and historical data in the ground relational database (PostgreSQL) and historical map database (Neo4j), including basic attributes (hole coordinates, borehole diameter, borehole date, designed extraction volume), historical events (anomaly records, maintenance records, sensor replacement records), and relationships (mining area to which the borehole belongs, responsible work team, geological structure fault zone number, and list of neighboring boreholes in space).

[0078] Step S42 logic: First, RDF semantic tags are generated, using the borehole ID as a unique identifier. The four types of information (basic attributes, operational status, historical events, and relationships) are encapsulated into Resource Description Framework (RDF) triples (subject-predicate-object). Then, node mounting is performed. Through the Digital Thread interface, the RDF tags are mounted to the corresponding 3D geometric node in S4-①. After mounting, the 3D model is no longer just a "static geometric model" but becomes an "interactive data carrier." Finally, dynamic field binding is performed, binding the dynamic parameters (current gas value, negative pressure, etc.) from the real-time data packet in step S3 to the "operational status" subfield of the semantic tags, supporting second-level refresh.

[0079] Step S42 output: 3D nodes carrying complete RDF semantic tags (clicking on any drill hole will display a pop-up window showing all static and dynamic attributes, historical files and related information), passed to sub-step S44.

[0080] Step S43: CFD Anomaly Evolution Simulation (Disaster Trend Prediction) Step S43 Input: Triggering condition, which is automatically activated only when an orange or higher warning (gas ≥ 1.0%) is detected in a borehole in step S5 or S3.

[0081] Input the real-time outflow rate of the abnormal hole (unit: m³ / min) from step S33; the permeability coefficient of the surrounding rock from the static geological database; the boundary conditions of the airflow field in the mining area (wind speed, wind direction, roadway cross-section); and the three-dimensional spatial coordinates of the abnormal hole output from step S41.

[0082] Step S43 logic: First, Kriging geostatistical interpolation is used. Based on the current outflow rate of the anomalous borehole and the synchronization parameters of neighboring boreholes, a semi-variogram is used to fit the spatial variation structure, providing an optimal unbiased estimate of the gas concentration in the unsampled area surrounding the anomalous borehole. Then, lightweight CFD is performed, using the interpolation results as initial boundary conditions. Combined with the surrounding rock permeability coefficient and airflow field data, a simplified Navier-Stokes equation is solved to extrapolate the dynamic evolution path and diffusion envelope of the gas concentration isosurface over the next 10-30 minutes. Finally, a three-dimensional surface is generated, and the extrapolated results (discrete concentration data points) are reconstructed into a renderable three-dimensional isosurface mesh and a semi-transparent influence area envelope.

[0083] Step S43 logic output: dynamic gas concentration isosurface grid, diffusion influence range envelope, and suggested personnel evacuation direction indicators.

[0084] Step S44: Visual rendering and compositing (screen presentation) Step S44 Input: Dynamic 3D geometric nodes (including risk color marks) from step S41.

[0085] Semantic knowledge graph from step S42 (for interactive query).

[0086] Dynamically derived isosurfaces and envelopes from step S43 (existing only in anomalies).

[0087] Interactive commands from the dispatcher (rotate, zoom, click to query, cut profile).

[0088] Step S44 logic: First, scene graph synthesis is performed, mixing the transparency and hierarchical order of geometric nodes, semantic tags, and CFD-derived surfaces in a unified scene graph (ensuring the semi-transparent display of the derivation envelope without obscuring the borehole body). Next, lighting and annotation rendering are performed, using the WebGL pipeline for lighting calculations, automatically generating parameter heatmaps (such as gas concentration cloud maps overlaid on the tunnel floor map), and rendering the borehole profile's parameter curves in real time. Finally, interactive responses are implemented, receiving mouse / touch input; clicking on a borehole immediately extracts and displays its full lifecycle data timeline from the semantic tags in a pop-up window.

[0089] Step S44 output: A smooth 3D interactive scene with ≥30fps on the large screen of the ground dispatch center, supporting roaming, drill-down and historical status playback, for dispatchers to make decisions and for emergency command in step S5.

[0090] Step S41 converts the "abstract numbers" (gas concentration, risk level) from step S3 into "visible geometric shapes" (3D boreholes are colored in red, orange, yellow, and green and become thicker and thinner with the viewing angle), solving the spatial perception problem of "where is the data?" Next, step S42 attaches four-dimensional semantic tags (basic attributes, historical archives, fault associations, and responsibility attribution) to the geometric shapes, upgrading each borehole from a "graphic" to a "knowledge node," allowing users to trace its entire lifecycle with a click, completely eliminating the need for two-dimensional reports. On this basis, the system maintains the flexibility of "standby in peacetime and activation in wartime"—normally only the first two steps are run, but once an anomaly is detected (triggered by S5), step S43 is immediately activated, using the surrounding rock permeability coefficient and outflow rate to start a lightweight CFD simulation engine, predicting the gas diffusion path and threat range 10-30 minutes in advance, gaining a golden window for emergency evacuation. Finally, step S44, as the final assembly workshop, synthesizes the geometric shapes, semantic tags, and simulation surfaces in real time into a smooth 3D battlefield map presented on the dispatch screen. The four interconnected sub-steps—① providing material on "what to display," ② answering the questions of "what and why," ③ predicting the trend of "what will happen next," and ④ integrating all information to allow the dispatcher to "understand the overall situation at a glance"—together form a complete bridge from physical perception to cognitive decision-making.

[0091] Step S5: Closed-loop handling of abnormal events Step S5 Input: The high-reliability measured data package from step S33 (real-time gas concentration CH4, carbon monoxide CO concentration, temperature, negative pressure change rate), and the preset four-level early warning threshold library (fixed safety red line), including the following: Blue (note): CH4 ≥ 0.5%, or concentration change rate ≥ 0.1% / min; Yellow (warning) indicates that CH4 ≥ 0.8%, or abnormal negative pressure fluctuations exceed ± 20%; Orange (hazardous), meaning CH4 ≥ 1.0% or CO ≥ 24 ppm; Red (urgent) indicates that CH4 ≥ 1.5%, or temperature ≥ 40℃, or abnormal ethylene / acetylene (a marker of spontaneous combustion in coal).

[0092] like Figure 6 As shown, the logic for step S5 is as follows: Step S51: Multi-level threshold triggering and resource preemption First, real-time stream comparison is performed, comparing each data packet pushed in step S3 against the four threshold conditions mentioned above in parallel. Then, the warning level is determined, and the highest level among the triggering conditions is taken as the current warning level (for example, if both yellow and orange are triggered simultaneously, it is set to orange). Then, preemptive resource scheduling is performed. Once an orange or red warning is determined, an "interruption command" is immediately generated and forcibly sent to step S2 (inspection planning module) and step S3 (gas path execution module) through the high-speed industrial communication link. This forcibly suspends all low-risk borehole inspection plans currently being executed in step S2, and forcibly interrupts non-abnormal borehole detection actions that are being performed or to be executed in step S3. The risk level of the abnormal borehole is forcibly raised to "extremely high" within the system, and all detection resources of step S3 are locked to the abnormal borehole. The sampling frequency is rapidly increased from the standard 1 time / minute to 1 time / 10 seconds. At the same time, the downhole human-machine interaction unit (electronic ink screen + LED array) is triggered to display alarm color marks and flashing warnings, and the local audio-visual and voice alarms are activated (≥100dB).

[0093] Step S51 outputs: warning level signal (blue / yellow / orange / red) and abnormal borehole ID, and sends it to the forced "interruption command" (containing abnormal borehole ID and encrypted detection parameters) in steps S2 / S3; it is then passed to sub-step S52 for fault diagnosis.

[0094] Step S52, Knowledge Graph Diagnosis Step S52 Input: Warning level and abnormal borehole ID from step S51, and detailed parameter snapshots (complete change curves of gas, negative pressure, CO, and temperature) for 5 minutes before and after the abnormal moment.

[0095] The input also includes a ground-based knowledge graph database (Neo4j) and a rule engine (Drools), which pre-store expert experience rules in the coal mining field, such as: Rule 1: IF (sudden drop in negative pressure > 20%) AND (gas concentration surge > 30%) THEN Diagnosis = "Borehole collapse or screen blockage", Recommendation = "Increase extraction negative pressure and check the wellhead; Rule 2: IF (CO concentration continues to rise >24 ppm) AND (temperature rises >35℃) THEN Diagnosis = "Precursor to spontaneous combustion of coal", Recommendation = "Immediately inject nitrogen or inject inhibitor"; Rule 3: IF (abnormal drilling in adjacent fault zone) AND (drastic fluctuations in gas outburst rate) THEN Diagnosis = "Mining-induced fault activation", Recommendation = "Reduce mining speed and increase monitoring".

[0096] Step S52 logic: First, feature extraction is performed, extracting key criteria (such as "negative pressure drop range", "gas rise slope", "whether CO exceeds the standard") from the parameter snapshot; then, rule matching is performed, inputting the extracted feature combination into the Drools rule engine, and performing forward chain matching with expert rules in the knowledge graph to find the fault mode with the highest confidence; finally, treatment suggestions are generated, and based on the matching results, a structured diagnostic conclusion (human-readable text) and specific actionable treatment suggestions (such as adjusting valve opening and injection medium type) are output.

[0097] Step S52 outputs: a structured diagnostic report, which includes the abnormal borehole ID, the cause of the failure (e.g., "#12 borehole collapse due to fault zone"), handling recommendations (e.g., "increase negative pressure to 25 kPa"), and safety risk warnings, and is then passed to sub-step S53 to generate a work order.

[0098] Step S53: Electronic Work Order Generation and Precise Dispatch Step S53 Input: The structured diagnostic report from step S52, the three-dimensional coordinates of all underground workers provided in real time by the mine personnel positioning system (UWB / Wi-Fi fusion positioning), and the responsibility matrix database (RACI model), which records the responsible team and the list of certified personnel with the necessary qualifications for handling each borehole.

[0099] Step S53 logic: First, the work order is encapsulated, integrating the anomaly location (drill hole number + 3D coordinates), parameter snapshot screenshot, diagnostic conclusion, handling suggestions, safety precautions, and expected handling time into a standardized electronic work order. Then, intelligent dispatch (distance priority) is implemented, filtering from the responsibility matrix to select a list of all personnel qualified to handle the anomaly; next, the personnel positioning system is invoked to calculate the spatial distance from each person in the list to the anomaly drill hole, selecting the closest responsible person as the first dispatch target. A three-level progressive push is used, sending the following information sequentially to the target responsible person's explosion-proof mobile terminal (smart mining lamp or explosion-proof mobile phone): Level 1 (Instant): APP / software push notification (with vibration); Level 2 (no confirmation within 30 seconds): SMS push (including ticket summary link); Level 3 (No confirmation within 60 seconds): Automatically dial a voice call and read out the work order details.

[0100] Step S53 output: The dispatched electronic work order (status marked as "pending handling") is confirmed by the responsible person (feedback to the digital twin). After the responsible person completes the on-site handling, the photos, videos and text records of the handling measures uploaded through the mobile terminal trigger sub-step S54 for re-inspection and verification.

[0101] Step S54: Re-inspection and verification, and closed-loop cancellation / upgrade Step S54 Input: "Handling Completed" confirmation signal from the responsible person's mobile terminal and on-site handling record (photo / video).

[0102] Re-inspection data package: After receiving the "handling completed" signal, the system automatically calls step S3 to perform three independent inspections on the original abnormal borehole (with a 2-minute interval), and generates a complete set of parameters for each inspection.

[0103] The original threshold for triggering the warning for this borehole (e.g., 1.0% for an orange warning).

[0104] Step S54 logic: Take the maximum value from three consecutive re-inspection data and compare it with 80% of the original warning threshold. The results are as follows: Safety line = warning threshold × 0.8 (for example, if the orange warning threshold is 1.0%, then the safety line is 0.8%).

[0105] A. Determine if the data from three consecutive re-inspections are all below the safety threshold (i.e., all three are <0.8%), then the abnormal event is determined to have been effectively handled. A timestamped cancellation record is automatically generated, and the handling process (abnormal snapshot, diagnostic conclusion, handling photos, re-inspection curve) is completely archived to the historical database. The "Restored to Normal" status is written back to the historical event database of step S4 (updating the borehole file); simultaneously, an "abnormal weight attenuation coefficient" is sent to step S1 to reduce the historical anomaly co-occurrence frequency of this borehole in the adjacency matrix (preventing it from being marked as a "key concern").

[0106] B. Deemed Unrecovered (Upgraded): If any one of the three consecutive re-inspection data points is greater than or equal to the safety line, the handling is deemed incomplete or ineffective. The warning level will be automatically upgraded by one level (e.g., orange → red). The system will automatically generate a new, higher-level work order, skipping the original responsible person and reassigning it to a higher-level responsible person (e.g., the mine's chief engineer on duty), and simultaneously triggering the mine-level emergency response plan linkage (e.g., power outage and evacuation).

[0107] Step S54 outputs: cancellation record (restored) or upgrade instruction (not restored). All data is uniformly written back to step S4 (update twin status) and step S1 (correct risk weight), completely closing the current abnormal event management loop.

[0108] In step S5, step S51 continuously monitors each frame of data from step S3. Once the orange or red threshold is reached, a forced interrupt command is immediately issued to steps S2 and S3 with the highest priority, forcibly allocating all system resources (detection frequency, gas path) to the abnormal borehole, ensuring that the hidden danger is locked by "prioritizing the detection of abnormal boreholes." Next, step S52 quickly matches the abnormal parameter combinations (how much the negative pressure dropped, how fast the gas rose) with fault rules in the knowledge base, providing clear diagnoses and treatment plans for borehole collapse, spontaneous combustion, etc. Once the diagnosis is made, step S53 immediately encapsulates it into a structured electronic work order, combined with UWB... The personnel positioning system accurately calculates the closest qualified personnel to the anomaly point and ensures that work orders are delivered through a three-tiered approach via APP, SMS, and voice, issuing commands to individuals specifying "who should repair and what to repair." After the responsible person completes the task and confirms it via mobile terminal, step S54 performs a mandatory automatic re-check three times. Only when the data is consistently below the safety line (80% of the threshold) for three consecutive times is the work order allowed to be closed and archived. The recovery status is then written back to the historical database in step S4 and the adjacency weight matrix in step S1, allowing the entire system to synchronously perceive that the hole has returned to normal. If the re-check fails, the task is deemed invalid, the warning level is automatically escalated, and a new work order is dispatched to the superior.

[0109] like Figure 7 As shown, a gas inspection system for a group of extraction boreholes that implements the above method is presented. This system adopts a four-level collaborative architecture of "cloud-edge-device-field," specifically including: The ground-based digital twin control unit, deployed in the ground dispatch center, consists of a high-performance industrial server, a GPU-accelerated workstation, a time-series database, a relational database, and a 3D visualization engine. It is used to construct a semantic digital twin of the extraction borehole group and perform global risk assessment, inspection plan generation, data visualization, abnormal event management, report generation, and multi-system collaborative scheduling. The underground edge intelligent control unit is fixedly installed in the extraction roadway chamber. It adopts an intrinsically safe explosion-proof design for mining and integrates a multi-core ARM processor, an NPU neural network acceleration unit, a large-capacity solid-state storage array, and a dual-frequency wireless communication module. It is used to receive ground commands and control gas path switching and data acquisition, perform local risk assessment model inference, edge optimization of inspection plans, data preprocessing and feature extraction, and trigger local audible, visual and voice alarms. The solid-state storage array supports offline caching of ≥90 days of detection data and automatically synchronizes based on a breakpoint resume mechanism after network recovery. The multi-path gas switching unit is electrically connected to the downhole edge intelligent control unit via a CAN bus and an intrinsically safe I / O interface, and adopts a modular pluggable design. It consists of a pilot-operated explosion-proof electromagnetic switching valve group, a precision pressure regulating valve group, a pulse-type clean gas source interface, a pressure balance bypass, and stainless steel / PTFE composite pipelines, and adopts a three-level tree topology structure of "main path-branch path-bypass path". The number of branches can be flexibly expanded according to the number of extraction boreholes, and a single system can support the inspection of 10-1000 extraction boreholes. The multi-parameter detection unit is connected to the main path of the multi-path gas switching unit and integrates a laser methane sensor, an infrared carbon monoxide / carbon dioxide sensor, an electrochemical oxygen sensor, a differential pressure sensor, a thermal mass flow sensor, and a digital temperature and humidity sensor. The laser methane sensor has a detection accuracy of ≤±0.5%FS and a response time of ≤10s. Each sensor adopts a modular rail design, supporting hot-swappable replacement and automatic calibration. The on-site human-machine interaction unit communicates with the underground edge intelligent control unit and adopts a hybrid display scheme of intrinsically safe electronic ink screen and LED array. It is suspended on the pedestrian side of the roadway and is used to display the number, parameters, inspection time, risk level color mark and alarm information of the currently detected borehole in real time. It also supports touch query of the status of nearby boreholes. The industrial communication unit connects the ground digital twin control unit and the downhole edge intelligent control unit respectively, and supports multiple communication methods such as gigabit fiber optic, industrial Ethernet, Wi-Fi 6 and 5G NR; the communication protocol stack includes MQTT, OPC UA and Modbus TCP, and supports TSN time-sensitive networking to meet the deterministic transmission of control commands; The intrinsically safe power supply unit adopts a mining explosion-proof and intrinsically safe DC regulated power supply, providing ≥8 channels of DC18V intrinsically safe output; it has a built-in lithium iron phosphate battery pack with a power outage duration of ≥8 hours.

[0110] The downhole edge intelligent control unit also integrates an edge computing framework, which supports the deployment of TensorFlowLite and ONNX Runtime models. It can deploy the spatiotemporal graph attention network model on the edge after INT8 quantization compression, so as to realize local real-time inference of risk assessment with an inference latency of ≤500ms. Even when the network is interrupted, it can still run the risk assessment and inspection scheduling functions independently.

[0111] The modular pluggable design of the multi-path gas switching unit is as follows: Each branch corresponds to a standardized gas path switching module, which integrates a solenoid switching valve, a one-way check valve, a miniature filter, and a quick-connect interface; Each module is connected via a unified backplane bus, and the backplane has built-in pressure balancing air passages and clean air source distribution air passages. When expanding the system, you only need to plug in the new module on the backplane and update the configuration file; there is no need to redesign the piping layout. The backplate is made of 316L stainless steel, and the internal air passages are mirror polished and passivated, with a dead volume of ≤5mL.

[0112] like Figure 8 As shown, when the number of extraction boreholes exceeds the processing capacity of a single system, multiple downhole edge intelligent control units and multi-path gas switching units can be deployed, and uniformly managed by the same ground digital twin control unit through a distributed collaborative scheduling algorithm to achieve multi-system collaborative inspection; the distributed collaborative scheduling algorithm includes: The global risk assessment module comprehensively analyzes the operating parameters and geological data of all boreholes to generate a global risk distribution heat map. The dynamic load balancing module, when multiple high-risk holes appear in the area managed by a certain system, allocates some low-risk holes in that area to nearby idle systems based on the Hungarian algorithm or auction algorithm; The regional anomaly correlation module identifies regional gas outburst events based on spatial clustering algorithms when multiple adjacent boreholes simultaneously exhibit anomalies. This automatically triggers the mine-level emergency response plan and generates a regional coordinated response plan.

[0113] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the systems, devices, apparatuses, modules or units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0114] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such quantities can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0115] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0116] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0117] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0118] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0119] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0120] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A digital twin driving-based gas risk adaptive inspection method for drainage borehole groups, characterized in that, Includes the following steps: Step S1, Multi-scale Risk Dynamic Assessment Step: Construct a dynamic risk assessment model for the drilling borehole group based on a spatiotemporal graph attention network. Collect multimodal operating parameters and geological environment parameters of each borehole in real time. Through hierarchical processing of spatiotemporal embedding layer, graph attention layer, temporal convolutional layer and fully connected output layer, calculate the gas outburst risk level of each borehole. The input of the graph attention layer includes the topological adjacency matrix of the borehole group. The adjacency matrix is ​​constructed based on the spatial distance of the boreholes, the correlation of geological strata and the co-occurrence frequency of historical anomalies. Step S2, Spatiotemporal Cooperative Inspection Planning Step: Based on the risk level, a deep reinforcement learning algorithm is used to dynamically optimize the inspection time slice allocation strategy and generate a spatiotemporal reuse inspection plan. The state space of the deep reinforcement learning algorithm is the current risk distribution and gas path load status of the borehole group, the action space is the inspection time slice allocation and gas path switching sequence of each borehole, and the reward function comprehensively considers risk coverage, detection frequency balance and gas path switching loss. Step S3, Gas Path Intelligent Scheduling and Self-Cleaning Detection Step: According to the inspection plan, the downhole edge intelligent control unit schedules multiple gas path switching units, adopts a three-level tree topology structure of "main path-branch path-bypass path", and uses the same group of multi-parameter detection units to complete the gas parameter detection of all extraction boreholes in a time-division multiplexing manner. The multi-path gas switching unit connects only one branch to the main path at any given time, and performs adaptive pressure balance control during the switching process of each branch. The specific process of the adaptive pressure balance control is as follows: before switching to the next branch, the pressure difference between the current branch and the target branch is collected by the pressure sensor. If the absolute value of the pressure difference exceeds the preset threshold, the bypass pressure regulating valve is opened first for pre-balancing. The main branch switching is then performed after the pressure difference drops to the threshold range. After the switch is completed, the branch flow is adjusted by an adaptive fuzzy PID algorithm to stabilize the detected flow within the set value ±5%. The fuzzy rule base of the fuzzy PID is constructed based on the branch length, pipe diameter and historical pressure loss data. After the single-hole test is completed, a pulsed gas path self-cleaning process is performed: the connection between the branch and the extraction borehole is closed, the connection between the branch and the clean gas source is opened, and clean air is introduced into the branch in a pulsed airflow manner for 3-5 seconds. The pulse frequency and duration are dynamically adjusted based on the feedback of the residual gas concentration in the branch until the residual gas concentration in the pipeline is lower than the preset threshold. Step S4, Semantic Digital Twin Mapping and Visualization Step: The detection data is mapped to the ground semantic digital twin in real time through the Industrial Internet of Things. The digital twin uses multi-level detail modeling technology to construct a three-dimensional geometric model of the extraction borehole group, and binds borehole attributes, operating status and historical event information based on semantic annotation technology to realize three-dimensional visualization of borehole status, full life cycle data traceability and cross-scale status inference. Step S5, Abnormal Event Closed-Loop Handling Steps: When the detection data exceeds the preset multi-level warning threshold, the abnormal borehole encrypted inspection mechanism and emergency response plan are automatically triggered, the abnormal borehole is dynamically upgraded to an extremely high risk level, the low-risk borehole inspection task is suspended, and the abnormal borehole is continuously detected with priority.

2. The method of claim 1, wherein, In step S1, the spatiotemporal graph attention network includes: The input layer is used to receive the borehole multimodal parameter vector and the borehole group topological adjacency matrix; The spatiotemporal embedding layer uses location encoding and geological attribute encoding to perform high-dimensional mapping of input features; The graph attention layer contains multiple graph attention networks. Each layer calculates the spatial association weights between borehole nodes through a multi-head attention mechanism. The spatial association weights are jointly determined by the borehole spatial distance decay function, geological stratum similarity, and historical anomaly co-occurrence frequency. Temporal convolutional layers employ causal dilated convolutional networks to capture the temporal dependence of borehole parameters; The fully connected output layer uses a Softmax classifier to output four risk levels: extremely high, high, medium, and low, and outputs the confidence probability for each level.

3. The method of claim 1, wherein, In step S2, the deep reinforcement learning algorithm employs a proximal policy optimization algorithm, the specific implementation of which includes: Both the policy network and the value network adopt a hybrid architecture of long short-term memory network and fully connected layer; The state vector includes: the current risk level code of each borehole, the time interval since the last inspection, the current gas path connectivity status, the remaining life of the multi-parameter detection unit, and the downhole ambient temperature and humidity; The action vector includes: the borehole number to be detected at the next moment, the detection duration allocated to each borehole, and the gas path switching path; The reward function is defined as: R = α· + β· + γ· + δ· , wherein is a risk coverage reward, is a detection frequency balance reward, is a detection efficiency reward, is a gas path switching loss penalty, and α, β, γ, and δ are weight coefficients. The strategy gradient method is iteratively optimized until the strategy converges, and an optimal inspection time slice allocation scheme is generated.

4. The adaptive inspection method for gas risk in a group of extraction boreholes based on digital twin driving as described in claim 1, characterized in that, In step S2, the generation rule for the spatiotemporal multiplexing inspection plan is to set a total inspection cycle T, where T∈[1,60] minutes, and dynamically allocate time slices according to the borehole risk level: Extremely high-risk boreholes: Allocate the entire total inspection cycle time to achieve continuous and uninterrupted inspection, and when there are multiple extremely high-risk boreholes, allocate the continuous inspection period for each borehole based on a multi-objective optimization algorithm; High-risk drilling: Each drilling hole is allocated a T / The inspection time slots are T / ,in The number of high-risk boreholes; Medium-risk drilling: T / The inspection time slots are T / ,in Number of boreholes classified as medium risk; Low-risk boreholes: All low-risk boreholes share the remaining inspection time. Inspection opportunities are allocated based on a hybrid strategy that combines polling scheduling and random sampling. Each borehole gets at least one inspection opportunity within a total inspection cycle. The total inspection cycle T is adaptively adjusted based on the mine's gas level, mining progress, and historical anomaly frequency.

5. The adaptive inspection method for gas risk in a group of extraction boreholes based on digital twin driving as described in claim 1, characterized in that, In step S3, the three-level tree topology structure of "main road-branch road-bypass road" includes: The main path connects to a multi-parameter detection unit, and the inner wall is electrochemically polished, with a surface roughness Ra≤0.4μm; The branch line is connected to the main line via an electromagnetic switching valve group. Each branch line corresponds to a extraction borehole, and the inner diameter is adaptively configured according to the extraction flow rate of the borehole. The bypass, connected in parallel to both ends of the main line, has a built-in precision pressure regulating valve and buffer chamber to maintain the pressure stability of the main line during branch switching and prevent pressure fluctuations from impacting the multi-parameter detection unit. The electromagnetic switching valve assembly has valve position status feedback and fault self-diagnosis functions.

6. The adaptive inspection method for gas risk in a group of extraction boreholes based on digital twin driving as described in claim 1, characterized in that, In step S3, the pulsed air path self-cleaning process further includes: during the clean air introduction stage, injecting clean air source in the form of high-frequency pressure waves, wherein the frequency of the high-frequency pressure waves is 5-20Hz, and enhancing the stripping effect of residual gas on the inner wall of the pipeline through airflow oscillation. After cleaning is completed, the residual gas concentration is detected by the gas sensor at the end of the branch. If the residual concentration is higher than the preset threshold, the cleaning time is extended and the pulse frequency is increased until the cleaning standard is met.

7. The adaptive inspection method for gas risk in a group of extraction boreholes based on digital twin driving as described in claim 1, characterized in that, In step S4, the process of constructing the semantic digital twin includes: The geometric modeling layer constructs a three-dimensional geometric model of the extraction borehole group based on mine CAD drawings, borehole survey data and geological exploration data, and uses LOD technology to dynamically switch the model accuracy according to the observation distance. The semantic annotation layer binds semantic tags to each borehole node. These semantic tags include basic attribute classes, running status classes, historical event classes, and association relationship classes. The data mapping layer receives real-time downhole monitoring data via the OPC UA protocol and the MQTT IoT protocol, and maps physical space parameters to corresponding nodes in the virtual space. The visualization rendering layer uses a lightweight rendering engine and supports 3D scene roaming, borehole profile display, parameter heatmap overlay, and historical status playback.

8. The adaptive inspection method for gas risk in a group of extraction boreholes based on digital twin driving as described in claim 1, characterized in that, In step S4, the digital twin also integrates an anomaly evolution prediction module. The anomaly evolution prediction module is based on computational fluid dynamics and gas geostatistics methods. It predicts the gas outburst trend and diffusion range based on the real-time parameters of the abnormal borehole, and dynamically displays the gas concentration isosurface and the affected area in a three-dimensional scene, providing visualization support for emergency decision-making.

9. The adaptive inspection method for gas risk in a group of extraction boreholes based on digital twin driving as described in claim 1, characterized in that, In step S5, the multi-level early warning thresholds include: Blue alert: Gas concentration reaches 0.5% or the rate of concentration change exceeds 0.1% / min; Yellow alert: Gas concentration reaches 0.8% or negative pressure fluctuates abnormally by more than ±20%; Orange alert: Gas concentration reaches 1.0% or carbon monoxide concentration exceeds 24 ppm; Red alert: Gas concentration reaches 1.5% or temperature exceeds 40℃ or ethylene / acetylene anomalies are detected; Different levels of early warning correspond to different encrypted inspection strategies and emergency response levels. The orange and red early warnings automatically trigger the linkage of the mine-level emergency plan.

10. The adaptive inspection method for gas risk in a group of extraction boreholes based on digital twin driving according to claim 1, characterized in that, In step S5, the generation and execution process of the closed-loop processing task chain includes: The anomaly diagnosis submodule, based on knowledge graphs and rule engines, automatically matches fault causes and handling suggestions according to the type and range of anomaly parameters. The task generation submodule generates a structured electronic work order that includes the anomaly location, parameter snapshot, diagnostic conclusion, handling suggestions, and estimated handling time. The task push submodule, based on the responsibility matrix and personnel positioning system, pushes work orders to the mobile terminal of the nearest person with the necessary qualifications to handle them. The process recording submodule collects on-site photos, videos, and records of handling measures from personnel via mobile terminals and transmits them back to the digital twin for archiving. The effect verification submodule automatically re-inspects the borehole after the treatment is completed. If the parameters return to normal, the borehole is automatically deactivated. If the parameters still do not return to normal, the warning level is upgraded and a new work order is reassigned.

11. A gas inspection system for a group of extraction boreholes, characterized in that, The method described in any one of claims 1-10 employs a four-level collaborative architecture of "cloud-edge-device-field," comprising: The ground-based digital twin control unit, deployed in the ground dispatch center, consists of a high-performance industrial server, a GPU-accelerated workstation, a time-series database, a relational database, and a 3D visualization engine. It is used to construct a semantic digital twin of the extraction borehole group and perform global risk assessment, inspection plan generation, data visualization, abnormal event management, report generation, and multi-system collaborative scheduling. The underground edge intelligent control unit is fixedly installed in the extraction roadway chamber. It adopts an intrinsically safe explosion-proof design for mining and integrates a multi-core ARM processor, an NPU neural network acceleration unit, a large-capacity solid-state storage array, and a dual-frequency wireless communication module. It is used to receive ground commands and control gas path switching and data acquisition, perform local risk assessment model inference, edge optimization of inspection plans, data preprocessing and feature extraction, and trigger local audible, visual and voice alarms. The solid-state storage array supports offline caching of ≥90 days of detection data and automatically synchronizes based on a breakpoint resume mechanism after network recovery. The multi-path gas switching unit is electrically connected to the downhole edge intelligent control unit via a CAN bus and an intrinsically safe I / O interface, and adopts a modular pluggable design. It consists of a pilot-operated explosion-proof electromagnetic switching valve group, a precision pressure regulating valve group, a pulse-type clean gas source interface, a pressure balance bypass, and stainless steel / PTFE composite pipelines, and adopts a three-level tree topology structure of "main path-branch path-bypass". The number of branches can be flexibly expanded according to the number of extraction boreholes, and a single system can support the inspection of 10-1000 extraction boreholes. The downhole edge intelligent control unit and the multi-path gas switching unit are configured to perform the adaptive pressure balance control as described in claim 1, specifically including: before switching to the next branch, collecting the pressure difference between the current branch and the target branch through a pressure sensor; if the absolute value of the pressure difference exceeds a preset threshold, opening the bypass pressure regulating valve for pre-balancing; after the pressure difference drops to within the threshold range, performing the main path switching; after the switching is completed, adjusting the branch flow rate with an adaptive fuzzy PID algorithm to stabilize the detected flow rate within the range of ±5% of the set value; The multi-parameter detection unit is connected to the main path of the multi-path gas switching unit and integrates a laser methane sensor, an infrared carbon monoxide / carbon dioxide sensor, an electrochemical oxygen sensor, a differential pressure sensor, a thermal mass flow sensor, and a digital temperature and humidity sensor. The laser methane sensor has a detection accuracy of ≤±0.5%FS and a response time of ≤10s. Each sensor adopts a modular rail design, supporting hot-swappable replacement and automatic calibration. The on-site human-machine interaction unit communicates with the underground edge intelligent control unit and adopts a hybrid display scheme of intrinsically safe electronic ink screen and LED array. It is suspended on the pedestrian side of the roadway and is used to display the number, parameters, inspection time, risk level color mark and alarm information of the currently detected borehole in real time. It also supports touch query of the status of nearby boreholes. The industrial communication unit connects the ground digital twin control unit and the downhole edge intelligent control unit respectively, and supports multiple communication methods such as gigabit fiber optic, industrial Ethernet, Wi-Fi 6 and 5G NR; the communication protocol stack includes MQTT, OPC UA and Modbus TCP, and supports TSN time-sensitive networking to meet the deterministic transmission of control commands; The intrinsically safe power supply unit adopts a mining explosion-proof and intrinsically safe DC regulated power supply, providing ≥8 channels of DC18V intrinsically safe output; it has a built-in lithium iron phosphate battery pack with a power outage duration of ≥8 hours.

12. The gas inspection system for a group of extraction boreholes according to claim 11, characterized in that, The downhole edge intelligent control unit also integrates an edge computing framework, which supports TensorFlow Lite and ONNX Runtime model deployment. It can deploy the spatiotemporal graph attention network model described in claim 2 on the edge side after INT8 quantization compression, realize local real-time inference of risk assessment, with inference latency ≤500ms, and can still independently run risk assessment and inspection scheduling functions when the network is interrupted.

13. The gas inspection system for a group of extraction boreholes according to claim 11, characterized in that, The modular pluggable design of the multi-path gas switching unit is as follows: Each branch corresponds to a standardized gas path switching module, which integrates a solenoid switching valve, a one-way check valve, a miniature filter, and a quick-connect interface; Each module is connected via a unified backplane bus, and the backplane has built-in pressure balancing air passages and clean air source distribution air passages. When expanding the system, you only need to plug in the new module on the backplane and update the configuration file; there is no need to redesign the piping layout. The backplate is made of 316L stainless steel, and the internal air passages are mirror polished and passivated, with a dead volume of ≤5mL.

14. The gas inspection system for a group of extraction boreholes according to claim 11, characterized in that, When the number of extraction boreholes exceeds the processing capacity of a single system, multiple downhole edge intelligent control units and multi-path gas switching units are deployed, and uniformly managed by the same ground digital twin control unit through a distributed collaborative scheduling algorithm to achieve multi-system collaborative inspection; the distributed collaborative scheduling algorithm includes: The global risk assessment module comprehensively analyzes the operating parameters and geological data of all boreholes to generate a global risk distribution heat map. The dynamic load balancing module, when multiple high-risk holes appear in the area managed by a certain system, allocates some low-risk holes in that area to nearby idle systems based on the Hungarian algorithm or auction algorithm; The regional anomaly correlation module identifies regional gas outburst events based on spatial clustering algorithms when multiple adjacent boreholes simultaneously exhibit anomalies. This automatically triggers the mine-level emergency response plan and generates a regional coordinated response plan.

Images