Coal mine underground gas inspection collaborative decision-making method and device

Abstract

The invention relates to a coal mine underground gas inspection collaborative decision-making method and a coal mine underground gas inspection collaborative decision-making device. Comprising the steps of amphibious inspection robot system initialization, inertial data preprocessing and state prediction, point cloud acquisition and state updating, multi-source data fusion and map updating, spatio-temporal feature node construction, dynamic spatio-temporal diagram modeling and information spreading, emission source probability field generation, three-dimensional semantic understanding, terrain geometric parameter extraction and trafficability quantitative decision making. The gas emission source positioning method has the advantage that the gas emission source positioning method based on dynamic space-time diagram modeling and physical information fusion is provided. A node network with space-time constraint is constructed, a learnable neural network is utilized to calculate node association weights, and a gating modulation mechanism of environmental factors such as wind speed is combined, so that efficient fusion and reasoning of sparse and dynamic gas data are realized. An attention mechanism and kernel density estimation capable of optimizing bandwidth are further introduced, a continuous and explainable gas emission probability field is generated, and the limitation of traditional point type monitoring is broken through.

Description

Technical Field

[0001] This invention belongs to the field of underground coal mine inspection technology, and relates to a collaborative decision-making method and device for underground coal mine gas inspection. Background Technology

[0002] Currently, the penetration rate of intelligent technology in the coal industry is gradually increasing, rising from 15% in 2018 to approximately 50% today. Under this trend, technologies such as the "Mine Eye Intelligent Detection" twin-driven multi-scenario inspection robot are emerging, aiming to solve the inspection challenges in underground environments characterized by high dust, low light, and complex dynamics. However, the general situation in the industry is that while existing robots can move autonomously, they struggle with accurate analysis and decision-making, especially when it comes to multi-source perception fusion and disaster early warning, where the system's intelligence level is severely insufficient. Specifically, in the fields of amphibious inspection and gas detection, existing technical solutions suffer from the following three key bottlenecks: 1. Functional mechanical superposition, isolated sensing data. Currently, most amphibious inspection vehicles combined with gas detectors on the market are essentially just a physical superposition of an amphibious platform for "walking and flying" and an independent gas detection module. This approach cannot achieve deep data-level collaboration, resulting in the inspection process only acquiring discrete, isolated point concentration information, lacking dynamic correlation analysis of key environmental factors such as gas diffusion, temperature, and wind speed within the tunnel space. This isolated sensing mode limits the system's judgment of gas anomalies to the rudimentary stage of threshold alarms, failing to provide a data foundation for accurate disaster source location and situation assessment.

[0003] 2. The core algorithms suffer from insufficient adaptability and poor environmental robustness. Faced with the extremely complex environment of underground coal mines—characterized by the absence of GPS, high dust levels, and strong electromagnetic interference—the stability and accuracy of existing algorithms face severe challenges. Specifically, this manifests in the following ways: ① In terms of localization and mapping algorithms: Traditional multi-sensor fusion SLAM is prone to pose estimation degradation, matching failures, and map drift in long underground corridors and low-texture areas, directly affecting the planning of subsequent inspection paths and the spatial registration accuracy of sensing data. ② In terms of gas sensing and source localization algorithms: Existing research mostly focuses on single-point concentration monitoring. For the core issue of retrieving a continuous three-dimensional gas concentration field from sparse, dynamic sampling point data and accurately locating the emission source (such as gas release from goaf), relevant algorithm research is insufficient, and inaccurate localization is often caused by complex airflow interference such as underground air leakage. The algorithms struggle to effectively couple concentration data with environmental context (wind speed, temperature), resulting in poor model physical interpretability and a high false alarm rate.

[0004] 3. The system lacks an intelligent closed loop and has weak decision-making autonomy. Currently, most inspection systems operate their various functional modules (such as SLAM, gas detection, and visual recognition) relatively independently, failing to form a closed loop of "perception-modeling-decision-action." For example, the path planning module cannot proactively avoid disasters or prioritize inspections based on real-time gas risk heat maps; the terrain analysis module also cannot provide quantifiable traversability parameters for obstacle-crossing decisions. The system heavily relies on remote control and pre-set programs, lacking autonomous response and collaborative decision-making capabilities in post-disaster or unknown complex environments.

[0005] In conclusion, the industry urgently needs an intelligent inspection system that truly achieves deep collaboration between "machine-electronics-computing". This system not only needs to overcome the limitations of amphibious capabilities in its mechanical structure, but more importantly, it needs to solve the fundamental problems of existing solutions—such as "strong inspection but weak detection," delayed early warning, and inaccurate source location—through an algorithm system that can tightly integrate multi-source heterogeneous data, adapt to harsh downhole environments, and possess a complete intelligent closed loop from perception to decision-making. This would enable a leap from "passive detection" to "proactive prediction". Summary of the Invention

[0006] This invention aims to solve the problem of the current lack of intelligent gas inspection systems in coal mines with deep collaboration.

[0007] The present invention solves the above-mentioned technical problems through the following technical means: A collaborative decision-making method for underground gas inspection in coal mines includes the following steps: S1. Amphibious Inspection Robot System Initialization: Complete lidar startup, gas sensor preheating, and algorithm loading; S2. Inertial Data Preprocessing and State Prediction: Raw data acquisition; raw data correction and compensation; state prediction; S3. LiDAR Data Preprocessing and Observation Generation: Point Cloud Data Acquisition; Noise Reduction Filtering and Motion Distortion Correction; State Update; S4. Multi-source data fusion and state update: Calculate Kalman gain; map update; S5. Spatiotemporal Feature Node Construction: Spatiotemporal label node creation; feature vector construction; S6. Dynamic Spatiotemporal Graph Modeling and Information Propagation: Defining the spatiotemporal neighborhood of nodes; Multilayer Perceptron Optimization; Dynamic Adjustment of Gated Cyclic Units; S7. Generation of Efferent Source Probability Field: Calculation of node attention weights; Calculation of spatial probability distribution of efferent sources; Risk assessment; S8. Three-dimensional semantic understanding: point cloud recognition and segmentation; ② Point cloud semantic label assignment; S9. Terrain geometric parameter extraction: Extract target point cloud clusters; calculate unit normal vectors; calculate key parameters; S10. Accessibility Quantification Decision: Comparison of geometric parameters with robot physical capability thresholds; accessibility decision discrimination.

[0008] This invention achieves systematic innovation in the field of intelligent mine inspection. Its core lies in constructing an intelligent decision-making framework that integrates deep fusion of multi-source sensing, adaptive optimization, and semantic understanding. The algorithm proposes a method for locating gas emission sources by combining dynamic spatiotemporal graph modeling with physical information fusion. By constructing a node network with spatiotemporal constraints and utilizing a learnable neural network to calculate node association weights, combined with a gating modulation mechanism for environmental factors such as wind speed, efficient fusion and inference of sparse, dynamic gas data are achieved. Furthermore, by introducing an attention mechanism and kernel density estimation with optimizable bandwidth, a continuous and interpretable gas emission probability field is generated, overcoming the limitations of traditional point-based monitoring.

[0009] Preferably, the specific method of step S2 is as follows: S2a, Raw data acquisition: The six-axis inertial measurement unit (IMU sensor) integrated into the system-controlled robot collects the angular velocity and acceleration in the three-axis directions in real time; S2b, Raw data correction and compensation: ① Temperature compensation, using the formula and formula Corrections are made for angular velocity and acceleration; ② Zero bias correction is performed using the formula Calculate the corrected real-time state; S2c, state prediction: using the formula Calculate separately k The position, velocity, and attitude angle at any given moment; where Indicates the corrected angular velocity, express k angular velocity at time, and Indicates the compensation coefficient, Indicates the corrected acceleration, express k acceleration at any moment They are respectively Position, velocity, and attitude angle at any given moment Indicates the sampling time interval. express Position, velocity, and attitude angle at any given moment This indicates the corrected acceleration. This indicates the corrected angular velocity.

[0010] Preferably, the specific method of step S3 includes: S3a, point cloud data acquisition: the system controls the lidar to rotate and scan to obtain a point cloud composed of three-dimensional points; S3b, noise reduction filtering and motion distortion correction: noise reduction filtering, using the formula and formula Remove outliers caused by dust; correct motion distortion using formulas. After transforming all points, the corrected point cloud is obtained; S3c, State Update: The current frame point cloud is compared with the initial map inside the system, and then the formula is used. Calculate the rotation matrix R and translation vector describing the pose of the point cloud relative to the map in the current frame. In the formula Indicates the points being scanned. Indicates and Adjacent points, Represents each point and its nearest neighbors. The average distance between points The standard deviation is , and the corrected laser point is . The collection time is The end time of this frame scan is ; Represents the rotation matrix. Indicates the first The original coordinates of the laser points Indicates the first The target coordinates of the laser points are given by the translation vector. .

[0011] Preferably, the specific method of step S4 includes: S4a, Kalman gain calculation: using the formula Calculate the observation residuals, and then use the formula Calculate Kalman gain; S4b, map update: using the formula Update the height status using the formula Update the covariance, Through transformation Integrate into the system's initial map to update the map; utilize formulas Adjusting the reliability of observations; where For the first The predicted state value at time 10:00. Indicates the first The observation residual at time, express k The state at any given moment, h For the observation function, The Kalman gain at time t is The prior estimation error covariance is The posterior estimation error covariance is ,function exist The Jacobian matrix is , It is the adjustment coefficient. It is a standard matrix. It is the average density of the point cloud, and the observation noise covariance matrix is... .

[0012] Preferably, step S5 specifically includes: S5a, Spatiotemporal tag node creation: the robot in time Arrival Location At that time, the sensor is triggered; S5b, construct the feature vector: create a four-dimensional vector containing location information, gas concentration information, environmental information, and the body state. .

[0013] Preferably, step S6 specifically includes: S6a, defining the spatiotemporal neighborhood of the node: considering the size of the underground tunnel and the typical gas diffusion rate, according to the formula and formula Only calculate spatial Euclidean distance less than and the time difference of adoption is less than Sampling point data; S6b, multilayer perceptron optimization: using formula and formula , Scalar output of the network Normalization is performed; S6c and gated loop units are dynamically adjusted using formulas. Define the average environmental characteristics, and then use the formula and Environmental factors are corrected; in the formula Indicates the points being scanned. Indicates and Adjacent points, nodes and its adjacent points The correlation strength between them is set to Time set to and , These are learnable parameters. It is a network output scalar. Average environmental characteristics of all active nodes at any given time , It represents the average hidden state of all nodes at present. This represents element-wise multiplication, and the gate vector is... .

[0014] Preferably, step S7 specifically includes: S7a, calculating node attention weights: , , Each node receives an attention weight; S7b, the probability distribution of the outflow source space is calculated using the formula... , , Calculate probability density; S7c, Risk assessment: using the formula Calculate the risk score coefficient; where the query vector is in the formula. Key vector Value vector , It is the input feature matrix of all nodes. For attention weights, It is a learnable weight matrix. It is the dimension of the key vector. Represents the Gaussian function. The smoothness is represented by the overall attention weight. , The reconstructed concentration field uses pre-set weighting coefficients. The probability map of the outflow source is as follows The overall risk score is , It is the highest methane concentration in the region; The concentration exceeds the safety threshold. volume, It is an indicator function. It is the magnitude of the concentration gradient. It is the total intensity of the probability of the gushing source. Indicates the first Measured values ​​of gas concentration at each sampling point.

[0015] Preferably, step S9 specifically involves the following steps: S9a, extracting the target point cloud cluster: extracting the target point cloud cluster based on the semantic labels defined in S8; S9b, calculating the unit normal vector: using the formula... Perform eigenvalue decomposition; calculate S9c and key parameters using formulas. Calculate the slope of the target point cloud cluster using the formula Calculate the height of the target point cloud cluster using the formula Calculate the surface roughness of the target point cloud cluster; where It is the center of the point cloud. Represents the covariance matrix of the point cloud. Represents the slope, and the horizontal plane normal vector is... The height is Surface roughness is , and This represents the maximum and minimum Z-coordinate values ​​of all points in the cloud cluster.

[0016] Preferably, the specific method of step S10 is as follows: S10a, comparing geometric parameters with robot physical capability thresholds: comparing the calculated slope... ,high Roughness Compare with the robot's physical capability threshold; S10b, Passage Decision: Calculated using a formula.

[0017] Based on the threshold comparison results, the robot is guided on whether to proceed; where Indicates the slope, with a height of Roughness is , max This indicates the maximum slope the robot can climb. max Indicates the maximum height the robot can climb. This represents the maximum permissible roughness for the robot.

[0018] Preferably, an inspection device employing the aforementioned collaborative decision-making method for underground gas inspection in coal mines includes a drone base, a gas sensor, a walking assembly, a drone casing, a control module, a visual sensing module, a data sensing module, and a communication module.

[0019] The advantages of this invention are: (1) At the methodological level, the algorithm proposes a gas emission source localization method that integrates dynamic spatiotemporal graph modeling and physical information fusion. By constructing a node network with spatiotemporal constraints and using a learnable neural network to calculate node association weights, combined with a gating modulation mechanism for environmental factors such as wind speed, efficient fusion and inference of sparse and dynamic gas data are achieved. Furthermore, by introducing an attention mechanism and kernel density estimation with optimizable bandwidth, a continuous and interpretable gas emission probability field is generated, breaking through the limitations of traditional point-based monitoring.

[0020] (2) In terms of engineering implementation, the algorithm has significant adaptive and collaborative capabilities. The SLAM module improves robustness in harsh environments through online zero-bias calibration and dynamic noise adjustment; the traversability analysis module directly couples terrain semantic parameters with robot body physical parameters (such as track length) to achieve embodied intelligent decision-making. The entire system seamlessly integrates localization, gas perception, semantic segmentation and risk assessment, forming a complete closed loop from perception and mapping to risk warning and path decision-making, providing a highly reliable solution for unmanned underground inspection. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the structure of the amphibious inspection robot according to the first embodiment of the present invention; Figure 2 This is a schematic diagram of the amphibious inspection robot sensor according to the first embodiment of the present invention; Figure 3 This is a schematic diagram of the walking component of the amphibious inspection robot according to the first embodiment of the present invention; Figure 4 This is a schematic diagram of the tension component of the amphibious inspection robot according to the first embodiment of the present invention; Figure 5This is a flowchart of the gas detection method of the amphibious inspection robot according to the first embodiment of the present invention; Numbering on the map: 1. Drone base; 2. Gas sensor; 3. Walking assembly; 31. Vertical plate; 32. Support wheel; 33. Tension component; 331. Limiting plate; 332. Miniature cylinder; 333. Sliding plate; 334. Length plate; 335. Length shell; 336. Tension wheel; 337. Rotating shaft; 338. Support spring; 34. Track body; 35. Drive wheel; 36. Drive seat; 37. Drive shaft; 4. Servo motor; 5. Drone shell; 6. Angle seat; 7. Flight arm; 8. Motor seat; 9. Brushless motor; 10. Control module; 11. Propeller; 12. Sensor module; 13. Power module; 14. Data sensor module; 15. Communication module; 16. Alarm; 17. Charging port. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: See Figures 1-4 This invention provides a multi-functional amphibious inspection robot for underground coal mines, including a drone base 1, a gas sensor 2, a walking component 3, a servo motor 4, a drone shell 5, an angle mount 6, a flight arm 7, a motor mount 8, a brushless motor 9, a control module 10, a propeller 11, a sensing module 12, a power module 13, a data sensing module 14, a communication module 15, an alarm 16, and a charging port 17.

[0024] See Figure 1 The drone base 1 is a rectangular plate-shaped component. A servo motor 4 is fixedly mounted on its upper surface, and a gas sensor 2 is fixedly mounted on its lower surface to detect smoke conditions in the mine. Walking components 3 are also fixedly mounted on the left and right ends of the drone base 1. (See reference...) Figure 3 The two walking components 3 include a vertical plate 31, support wheels 32, a tensioning element 33, a track body 34, a drive wheel 35, a drive seat 36, and a drive shaft 37. The vertical plate 31 is a rectangular plate-shaped component, and multiple support wheels 32 are rotatably connected to the upper and lower sides of the vertical plate 31, with more support wheels 32 on the lower side of the vertical plate 31 than on the upper side. The tensioning element 33 is located on the left side of the vertical plate 31 and is fixedly connected to the vertical plate 31.

[0025] See Figure 4The tensioning component 33 includes a limiting plate 331, a micro cylinder 332, a sliding plate 333, a length plate 334, a length shell 335, tension wheels 336, a rotating shaft 337, and a support spring 338. The limiting plate 331 is a single-piece component fixed to the upright plate 31. The fixed end of the micro cylinder 332 is fixedly connected to the limiting plate 331. The movable end of the micro cylinder 332 is fixedly mounted with a sliding plate 333 that is slidably connected to the upright plate 31. A length plate 334 is fixedly mounted on one side of the sliding plate 333. A length shell 335 is slidably connected to one end of the length plate 334. A rotating shaft 337 is rotatably connected inside the length shell 335. Two symmetrically arranged tension wheels 336 are fixedly mounted on the surface of the rotating shaft 337. The outer sides of the two tension wheels 336 are in contact with the side of the track body 34 facing each other. A support spring 338 is fixedly mounted on the surface of the length shell 335. One end of the support spring 338 is fixedly connected to the side of the sliding plate 333 facing each other.

[0026] The track body 34 surrounds and contacts the support wheel 32, tension wheel 336, and drive wheel 35, forming a ring. The drive wheel 35 is located on the right side of the upright plate 31 and contacts the track body 34. The drive seat 36 is fixed to the right side of the upright plate 31. The drive shaft 37 is rotatably connected inside the drive seat 36. Two symmetrically arranged drive wheels 35 are fixedly mounted on the surface of the drive shaft 37. The outer sides of the two drive wheels 35 are in contact with the side of the track body 34 facing each other. The top of one side of the upright plate 31 is fixedly connected to the UAV base 1.

[0027] See Figure 1 The output end of the servo motor 4 is fixedly mounted on the drone housing 5, which is a hollow cylindrical component. Four rectangular angle seats 6 are fixedly mounted on the top side of the drone housing 5. A flight arm 7 is rotatably connected to the center of each of the four angle seats 6. A motor mount 8 is fixedly mounted at one end of each of the four flight arms 7. Brushless motors 9 are fixedly mounted on both sides of each of the four motor mounts 8. Propellers 11 are fixedly mounted on the output ends of each of the eight brushless motors 9. The control module 10 is located at the top of the inner wall of the drone housing 5 and is used to process signals collected by the amphibious inspection drone and issue action commands to the robot. The sensing module 12 is fixed to the bottom of one side of the drone housing 5 and is used to collect environmental information from the mine. Specifically, it includes a six-axis inertial measurement unit (IMU), a gas concentration sensor, a temperature and humidity sensor, and an ultrasonic wind speed sensor. The power module 13 is fixedly installed inside the drone base 1, and a charging port 17 communicating with the power module 13 is provided on the surface of the drone housing 5. A communication module 15 is fixedly installed on one side of the top of the drone shell 5, a data sensing module 14 is fixedly installed in the middle of the top of the drone shell 5, and an alarm 16 is fixedly installed on the other side of the top of the drone shell 5. The communication module 15 adopts wireless communication based on ultra-wideband technology to achieve high-speed and stable data transmission.

[0028] The amphibious inspection robot of this application ensures the real-time transmission of remote control commands by establishing a stable and reliable communication network. In the complex electromagnetic environment underground, anti-interference technology is particularly critical. It effectively filters various interference signals, ensuring the stability and accuracy of communication, so that the control commands issued by the operator can be quickly and accurately transmitted to the robot. At the same time, this device can be equipped with a variety of sensors to collect various environmental variables in the mine in real time, realizing the automation of robot actions.

[0029] The intelligent functions of the amphibious inspection robot in this application rely not only on high-performance hardware, but more importantly, on a series of algorithms specifically optimized for the underground environment deployed in the control module. The input data for these algorithms comes directly from various sensors mounted on the robot, and through specific preprocessing and fusion processes, ultimately outputs decision commands or early warning information. Based on the coal mine underground gas inspection device of this application, a coal mine underground gas detection method is also provided, as detailed below: (1) Raw data collection and state prediction The purpose of this step is to acquire basic motion data from the robot's built-in six-axis inertial measurement unit (IMU) and predict the robot's next position and orientation. In an environment like a coal mine, where there is no GPS, high dust levels, and strong electromagnetic interference, directly using raw data is prone to errors, potentially causing the robot to deviate from its course or make inaccurate judgments. Therefore, by collecting raw data and predicting the robot's next position and orientation, the robot can roughly "perceive" its own position and motion state, providing a foundation for subsequent mapping and gas detection. The specific steps are as follows: Raw data acquisition: Relying on the six-axis inertial measurement unit (IMU) integrated inside the inspection robot, the IMU unit collects data hundreds of times per second while working synchronously with the robot, and outputs the angular velocities of the X, Y, and Z axes. and linear acceleration linear acceleration Used to determine the orientation and speed of movement of equipment; angular velocity Used to accurately capture the rotational movements of equipment; through the IMU unit, the inspection robot can accurately calculate its own attitude, angular velocity, and acceleration. (This will be...) A set of angular velocity and acceleration data collected at any given time is recorded as the attitude, and is used as... Indicates, that is .

[0030] Because IMU sensors are affected by temperature and have an inherent bias (initial deviation), directly using raw data will cause positioning drift. Therefore, two compensation calculations must be performed before inputting data into the algorithm: ① Temperature compensation: Based on the real-time temperature read from the internal temperature sensor of the IMU. Query the pre-stored temperature-error comparison table to obtain the compensation coefficient. and ,right angular velocity in and acceleration Make corrections:

[0031]

[0032] ② Zero-bias correction: After each power-on and before the robot begins to move, it is kept stationary for 30 seconds. The algorithm calculates the posture during these 30 seconds. average This average value is the "zero bias" under the current environment. Subsequently, all real-time attitude data collected will be... Subtracting this "zero bias" value yields the corrected attitude data. :

[0033] State prediction: The previous moment As a reference point, calculate The robot's state at any given moment, i.e. The robot pose prediction value at time t, let's assume They are respectively Position, velocity, and attitude angle at any given moment Indicates the sampling time interval. express Given the position, velocity, and attitude angle at any given moment:

[0034] This section will include Yes, robot The state at time is denoted as .

[0035] (2) Point cloud acquisition and update Coal mine tunnels are complex, dusty, and poorly lit, making it easy to accumulate errors when relying directly on IMU (Insulated Measurement Unit) for location prediction. This step uses lidar to scan the surrounding environment and generate 3D point cloud data, like drawing a map. Figure 1 The system records data from walls, the ground, and other structures to help the robot correct its position. This is crucial for preventing data drift, ensuring the robot clearly understands its location and avoids collisions or getting lost, while also providing a spatial framework for gas detection. In practice, it allows the robot to clearly see its path, much like a miner's lamp, preventing obstacles and safety hazards. The specific steps are as follows: Point cloud acquisition: The inspection robot system controls its onboard LiDAR to rotate and scan, thereby obtaining a set of "point cloud" data composed of tens of thousands of three-dimensional points, denoted as... Due to the harsh environment and high dust content in coal mines, dust will also be included in the scanning data. Therefore, obtain Then, noise reduction processing is required. Let the laser points in each point cloud data be... , for Let each point be an adjacent point. Its nearest neighbor The average distance between the points is Standard deviation is If the distance of a point satisfy If a point is found to be an outlier caused by dust or noise, it is considered an "outlier" and is removed from the point cloud, resulting in... The average distance is Standard deviation is The calculation formula is as follows:

[0036] Motion distortion correction: In addition to noise reduction, point cloud data also needs to be addressed with motion distortion correction. This is because a single scan by a LiDAR scanner takes time. During this period, the robot itself is also moving, so the data generated by the lidar scan is subject to motion distortion. Let the laser point... The relative acquisition time within the scanning cycle is .

[0037] To address the point cloud distortion problem caused by carrier motion during lidar data acquisition, this embodiment employs a motion compensation algorithm based on an inertial measurement unit (IMU) for correction. The specific principle is as follows: Because lidar uses a rotating scanning mechanism, the resulting frame... Point cloud data is not generated instantaneously, but requires a scanning cycle (e.g., ...). , recorded as At this time, the inspection robot is in a continuous moving state, which causes the laser points in the same frame of the point cloud to be in a state of continuous motion. It is obviously at different times. The points are collected sequentially. If the robot's motion is ignored and these points are directly applied to the same coordinate system, the surrounding environment (such as vertical walls or alleyways) will appear tilted, distorted, or ghosted in the point cloud, severely affecting the accuracy of subsequent mapping. To solve the above problems, this step performs the following operations: First, the system accurately records each laser point. Relative acquisition time when triggered (This time is in) to (between); secondly, using high-frequency sampled IMU data (angular velocity and acceleration), the robot's acquisition time from the laser point is calculated through numerical and integral interpolation. Until the end of the current frame scan The relative motion trajectory during this brief period. Based on the relative motion trajectory, construct the pose transformation matrix. This matrix describes the translational changes of the robot relative to the end of the scan when acquiring the point; finally, this pose transformation matrix is ​​used to transform the original laser point. The coordinates are subjected to a rigid body transformation and mapped to a unified time reference (i.e., the end time of the scan). In the coordinate system of ), the corrected laser point is obtained.

[0038] Through processing, this method effectively eliminates the Doppler effect and geometric deformation caused by the robot's own motion, transforming dynamically acquired non-rigid point clouds into spatially consistent static point clouds, thereby significantly improving the geometric fidelity of subsequent point cloud registration and environmental map construction. This applies to all initial point cloud data. After transformation, the corrected point cloud data is obtained, denoted as... .

[0039] Map Update: In the control module of the inspection robot, a map is generated based on the initial point cloud. Accumulated local environment map , will correct With map Perform matching using a rotation matrix. Representing rotation operations in space (where the rotation matrix is ​​a 3x3 orthogonal matrix), using translation vectors. It represents a translation operation in space (the translation vector is a 3-dimensional vector that represents the displacement in the X, Y, and Z directions). Indicates the first The original coordinates of the laser points Indicates the first The target coordinates of each laser point are calculated using the following formula:

[0040] This formula represents all possible rotation matrices. Translation vector In the given information, find the set of parameters that minimizes the objective function value, and denote this set of parameters as . ,but This refers to the real-time "observation" after updating the original map; the state at this moment is recorded as... .

[0041] The inspection robot's control module includes a pre-set "observation model." This model is designed to address the issue of time-accumulated gaps arising from simply relying on IMUs for state extrapolation. By installing this model, the system can map the predicted values ​​of previous state moments to the installation space, allowing for comparison with the current actual sensor installation data to obtain more accurate values.

[0042] The theoretical observation value at this moment is predicted using the "observation model," which assumes... For the first The predicted state value at time step 1 is substituted into the "prediction model" function. , Indicates the first The observation residual at time (i.e., the difference between the actual and predicted observations) is then:

[0043] The residual determines the deviation between the robot's "self-calculated pose" and "the pose measured by the sensor," which is a key factor for subsequent algorithms to determine whether the robot has deviated from its course and how to correct it.

[0044] Based on this, set The Kalman gain at time t is In order to obtain Let the prior estimation error covariance be... "Observation model" function exist The Jacobian matrix is Let the observation noise covariance matrix be... (representing the uncertainty of the observed values), then Calculation formula

[0045] get Then, using the initial residual, the first... Observation residuals at time A weighted correction is performed to eliminate accumulated bias, yielding the optimal pose estimate for the current moment. This is then combined with the updated state based on the observed values. The calculation formula is:

[0046] Posterior estimation error covariance The calculation formula is: ( (For standard matrices).

[0047] Then Through transformation Integrate In the middle, update the map. This includes the observation noise covariance. Dynamically adjust based on the current point cloud matching fitting error and point cloud density:

[0048] in It is the root mean square error of the match. It is the average density of point clouds (points per cubic meter). It is the adjustment coefficient. It is a standard matrix.

[0049] (3) Gas sensing and inference of gas source Gas sensing: Based on multi-source data collected during the movement of the inspection robot, which is sparsely distributed in both time and space, a continuous three-dimensional gas concentration field underground is reconstructed in real time, and the location of the most likely gas outburst (probability distribution map) is calculated. The following steps are required for calculation: Data collection: The robot in time Arrival Location (This location is provided in real time by the above calculations) Simultaneously, the multi-source sensing sensors carried by the inspection robot are triggered to collect data, creating a data node with a spatiotemporal label. This node contains data from its sensors, including IMU, gas concentration sensor, temperature and humidity sensor, ultrasonic anemometer, etc.

[0050] Based on the created data nodes, data is collected from these nodes to construct node feature vectors. . It is a four-dimensional vector that includes location information, gas concentration information, environmental information, and the robot's own state. The location information is... Gas concentration information is set to Environmental information is set to The robot's body state is set to .

[0051] To avoid interference from momentary fluctuations in the sensor, the algorithm continuously collects readings. Taking gas concentration information as an example, once the reading stabilizes within a threshold, the arithmetic mean of all readings within the following 3 seconds is taken as the final value.

[0052] in It is the sensor in Readings at time, This is the moment when the readings begin to stabilize. It is the number of sampling points within 3 seconds.

[0053] Simultaneously read the temperature value from the digital temperature and humidity sensor. (°C), humidity value (%RH), and the wind speed value from the ultrasonic anemometer. (m / s), forming a vector Record the robot pitch angle provided by the IMU at this time. Roll angle And the drive motor current value G, forming a vector .

[0054] For efficient computation, not all historical nodes are related to the current node. To optimize the algorithm and reduce computational load, the method in this application only considers nodes related to the current node. Historical nodes that meet the following two conditions As its adjacent points: ① Spatial distance m; ② Time interval The specific spatial distance and time interval depend on the size of the underground tunnel. In this embodiment, the values ​​used are based on engineering experience with typical underground tunnel sizes and typical gas diffusion rates.

[0055] Gas concentration diffuses spatially and varies over time. Node and its adjacent points The correlation strength between them is set to Using a learnable multilayer perceptron To calculate: Calculate the difference between the feature vectors of two nodes. Calculate spatial distance Calculate the time difference ;Will As input, it is fed into the multilayer perceptron. The network has Layers, each with an activation function of: :

[0056] in These are learnable parameters. It is the network output scalar. For nodes All adjacent points Perform the above calculations to obtain a series of outputs. Then use Softmax The function normalizes these values:

[0057] Environmental factors, especially wind speed, can significantly affect gas distribution. The algorithm introduces a gated cyclic unit to dynamically adjust the model.

[0058] First, set Average environmental characteristics of all active nodes at any given time .calculatet Average environmental characteristics of all active nodes at any given time Especially the average wind speed Context modulation is achieved through the following formula:

[0059] in It is the sigmoid function. It represents the average hidden state of all nodes at present. This represents element-wise multiplication; an MLP is a multilayer perceptron. This bias term is used in the gating vector calculation process to correct the initial response level of the gating function. When the changes in downhole environmental parameters are small or the sensor sampling data is within a stable range, introducing this bias term can prevent the gating vector output from being overly biased towards a single feature source, thereby ensuring the continuity and stability of historical hidden states and current environmental features in the information fusion process. As an adjustable parameter, this bias term is updated collaboratively with the weight parameters during system operation and model training, enabling the environmental context modulation process to adapt to different roadway structures and wind speed conditions, improving the reliability of gas concentration modeling and outburst source inference. When the sensor detects a sudden increase in environmental wind speed, the gating vector... It will automatically reduce the weight of historical concentration characteristics while increasing the dependence on current environmental characteristics, thereby adjusting the impact of environmental factors on gas concentration.

[0060] Efferent source inference: Through the "multi-head attention" module in the graph neural network, the algorithm automatically analyzes which sampling points are more likely to be located near the efferent source. For the th Attention head, query vector Key vector Sum value vector Through linear transformation, we obtain:

[0061] in It is the input feature matrix of all nodes, a two-dimensional real number with dimensions N×Q, where N represents the total number of nodes in the current graph neural network, and Q represents the dimension of the features managed by each node. It is a learnable weight matrix. The attention weights are calculated as follows:

[0062] in This is the dimension of the key vector. The output of multi-head attention is:

[0063] Through the attention mechanism, each node receives a comprehensive attention weight. .

[0064] To generate a continuous probability distribution map from discrete, weighted sampling points, a kernel density estimation method is employed.

[0065] Select kernel function: Choose the standard Gaussian kernel function. .

[0066] Determine bandwidth: bandwidth This determines the smoothness. Initial bandwidth. Hydraulic diameter of the tunnel Proportional:

[0067] in It is an empirical coefficient (usually taken as 0.1~0.3), hydraulic diameter Calculations based on the map constructed earlier show that for rectangular alleyways, Where a and b are the width and height of the tunnel cross-section; for a circular tunnel, It is equal to the diameter. During the model training phase, It is also an optimizable parameter that can be adjusted.

[0068] Calculate the gas probability density: for any point in space It is the probability density of the outflow source. The result is obtained by summing the weighted contributions of all sampling points:

[0069] Calculate the reconstructed concentration field : Represents an arbitrary location in space reconstructed based on discrete sampling points. The continuous gas concentration field function at the location, where Indicates the first Measured values ​​of gas concentration at each sampling point (or data node).

[0070]

[0071] Combined with the reconstructed concentration field and the probability map of the outflow source Calculate a certain region Comprehensive risk score :

[0072] in: These are pre-set weighting coefficients (usually) ); It is the highest methane concentration in the region; The concentration exceeds the safety threshold. volume, It is an indicator function. It is the magnitude of the concentration gradient, calculated using the finite difference method.

[0073] It is the total intensity of the outflow source probability, calculated through numerical integration.

[0074] (4) Three-dimensional semantic understanding and accessibility analysis To facilitate the rapid understanding of the geometric and semantic information of a scene in post-disaster or complex environments, and to perform quantitative analysis of terrain accessibility, it is necessary to compare the information identified by the inspection robot with the robot's own performance and issue corresponding instructions to the robot.

[0075] Semantic understanding: The acquired real-time point cloud data Input a pre-trained deep learning network, RandLA-Net. This network consists of an encoder and a decoder, including a local feature aggregation module and an attention-based upsampling module. The network outputs a class probability distribution for each point in the point cloud. Where B is the number of categories (e.g., ground, wall, machinery, rubble pile, large rock mass, etc.), the point cloud data is segmented to distinguish different categories. After segmentation, the category with the highest probability is taken as the semantic label for that point:

[0076] From the segmentation results, extract all points marked as "rubble piles (climbable)". These points form one or more point cloud clusters. For one of the point cloud clusters Contains multiple points Principal component analysis is used to fit an optimal plane. First, the covariance matrix of the point cloud is calculated:

[0077] in It is the center of the point cloud. (Yes / No) Eigenvalue decomposition reveals that the eigenvector corresponding to the smallest eigenvalue is the unit normal vector of the plane. .slope It is the angle between the plane and the horizontal plane. The normal vector of the horizontal plane is... The formula for calculating slope is:

[0078] The closer the absolute value of the dot product is to 1, the more level the plane is and the smaller the slope.

[0079] high It is the difference between the maximum and minimum Z-coordinates (elevation coordinates) of all points in the point cloud cluster:

[0080] Surface roughness This measures the unevenness of the point cloud cluster surface. It calculates the perpendicular distance from each point within the cluster to the fitted plane. Then calculate the root mean square of this distance:

[0081] Accessibility analysis: The calculated slope ,high Roughness Compare with the robot's physical capability threshold: Maximum passable slope: (The actual climbing slope of the amphibious inspection robot in this application); Maximum climbing height: (in (This refers to the track ground contact length). Maximum permissible roughness: (This value was determined by the amphibious inspection robot of this application in conjunction with laboratory tests to ensure that the treads can effectively grip the ground.)

[0082] Based on the comparison results, a binary decision is given:

[0083] In summary, this application proposes a method for locating gas emission sources by fusing dynamic spatiotemporal graph modeling with physical information. By constructing a node network with spatiotemporal constraints and utilizing a learnable neural network to calculate node association weights, combined with a gating modulation mechanism for environmental factors such as wind speed, efficient fusion and inference of sparse, dynamic gas data are achieved. Furthermore, by introducing an attention mechanism and kernel density estimation with optimizable bandwidth, a continuous and interpretable gas emission probability field is generated, overcoming the limitations of traditional point-based monitoring.

[0084] In terms of engineering implementation, the algorithm exhibits significant adaptive and collaborative capabilities. The SLAM module enhances robustness in harsh environments through online zero-bias calibration and dynamic noise adjustment; the traversability analysis module directly couples terrain semantic parameters with robot body physical parameters (such as track length) to achieve intelligent decision-making. The entire system seamlessly integrates localization, gas sensing, semantic segmentation, and risk assessment, forming a complete closed loop from perception and mapping to risk warning and path decision-making, providing a highly reliable solution for unmanned underground inspection.

[0085] Example 2: Figure 5This is a flowchart illustrating a collaborative decision-making method for underground gas inspection in coal mines, according to the second embodiment of the present invention.

[0086] This invention provides a collaborative decision-making method for underground gas inspection in coal mines, comprising: S1. Amphibious Inspection Robot System Initialization: Complete LiDAR startup, gas sensor preheating, and algorithm loading. The specific process of step S1 is as follows: The amphibious inspection robot is started, and the LiDAR, gas sensor, wind speed sensor, IMU sensor, and temperature and humidity sensor carried by the robot are simultaneously activated. Algorithm loading and mechanical preheating are also completed.

[0087] S2. Inertial Data Preprocessing and State Prediction: S2a. Raw Data Acquisition; S2b. Raw Data Correction and Compensation; S2c. State Prediction. The specific process of step S2 is as follows: The system controls the robot's integrated six-axis inertial measurement unit (IMU sensor) to collect the angular velocity and acceleration of three axes (X, Y, and Z directions) in real time; using the formula... and formula Temperature compensation corrections are applied to angular velocity and acceleration; using the formula Perform initial state bias correction on the robot; then use the formula Calculate separately The robot's position, speed, and attitude angle at all times.

[0088] S3. Point Cloud Acquisition and State Update: S3a. Point Cloud Data Acquisition; S3b. Noise Reduction Filtering and Motion Distortion Correction; S3c. State Update. The specific process of step S3 is as follows: The robot system controls the carried LiDAR to rotate and scan, obtaining a point cloud composed of three-dimensional points; using the formula... and formula Remove outliers caused by dust; use the formula Motion distortion correction is performed on all points, and the corrected point cloud data is obtained after the transformation. The current frame point cloud is compared with the map formed by the initial point cloud inside the system, and then the formula is used. Calculate the rotation matrix R and translation vector describing the pose of the point cloud relative to the map in the current frame. .

[0089] S4. Multi-source data fusion and map update: S4a. Kalman gain calculation; S4b. Map update. The specific process of step S4 is as follows: using the formula... Calculate the observation residuals, and then use the formula Calculate the Kalman gain; use the formula Update the robot's position status using the formula Update the covariance, Through transformation Integrate into the system's initial map to update the map; utilize formulas Adjust the reliability of observations.

[0090] S5. Spatiotemporal Feature Node Construction: S5a. Spatiotemporal Label Node Creation; S5b. Feature Vector Construction. The specific process of step S5 is as follows: the robot in time... Arrival Location At that time, the sensor is triggered to collect location information. ( Gas concentration information, including environmental vectors containing temperature, humidity, and wind speed values. A vector containing the robot's pitch angle, roll angle, and motor current values. Together they form a four-dimensional vector .

[0091] S6. Dynamic Spatiotemporal Graph Modeling and Information Propagation: S6a. Defining the Spatiotemporal Neighborhood of Nodes; S6b. Multilayer Perceptron Optimization; S6c. Dynamic Adjustment of Gated Cyclic Units. The specific process of step S6 is as follows: Considering the dimensions of the underground roadway and the typical gas diffusion rate, according to the formula... and formula Only calculate spatial Euclidean distance less than and the time difference of adoption is less than The sampling point data; using the formula and formula , Scalar output of the network Perform normalization; use the formula To represent the average environmental characteristics, and then use the formula and Corrective measures are taken to address environmental factors.

[0092] S7. Generation of the Emission Source Probability Field: S7a. Calculation of Node Attention Weights; S7b. Calculation of Spatial Probability Distribution of Emission Sources; S7c. Risk Assessment. The specific process of step S7 is as follows: , , Each node receives an attention weight; using the formula , , Calculate the probability density; use the formula Calculate risk score coefficient S8. 3D Semantic Understanding: S8a. Point Cloud Recognition and Segmentation; S8b. Point Cloud Semantic Label Assignment. The specific process of step S8 is as follows: The real-time point cloud obtained in step S4... Input a pre-trained deep learning network, RandLA-Net; this network outputs a class probability distribution for each point in the point cloud. According to the formula The category with the highest probability is taken as the semantic label for that point.

[0093] S9. Terrain Geometric Parameter Extraction: S9a. Extract target point cloud clusters; S9b. Calculate unit normal vectors; S9c. Calculate key parameters. The specific process of step S9 is as follows: Extract the target point cloud clusters based on the semantic labels defined in S8; use the formula... Perform eigenvalue decomposition; use the formula Calculate the slope of the target point cloud cluster using the formula Calculate the height of the target point cloud cluster using the formula Calculate the surface roughness of the target point cloud cluster.

[0094] S10, Accessibility Quantification Decision: S10a, Comparison of geometric parameters with robot physical capability thresholds; S10b, Accessibility Decision Judgment. The specific process of step S10 is as follows: Comparison of geometric parameters with robot physical capability thresholds: The calculated slope... ,high Roughness Compare with the robot's physical capability threshold; through algorithms

[0095] The threshold comparison results guide the robot on whether to proceed.

[0096] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Terms such as "upper," "lower," "left," "right," "front," and "rear" used in the invention are merely for clarity of description and are not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.

Claims

1. A collaborative decision-making method for underground gas inspection in coal mines, characterized in that, Includes the following steps: S1. Amphibious Inspection Robot System Initialization: Complete lidar startup, gas sensor preheating, and algorithm loading; S2. Inertial Data Preprocessing and State Prediction: Raw data acquisition; raw data correction and compensation; state prediction; S3. LiDAR Data Preprocessing and Observation Generation: Point Cloud Data Acquisition; Noise Reduction Filtering and Motion Distortion Correction; State Update; S4. Multi-source data fusion and state update: Calculate Kalman gain; map update; S5. Spatiotemporal Feature Node Construction: Spatiotemporal label node creation; feature vector construction; S6. Dynamic Spatiotemporal Graph Modeling and Information Propagation: Defining the spatiotemporal neighborhood of nodes; Multilayer Perceptron Optimization; Dynamic Adjustment of Gated Cyclic Units; S7. Generation of Emission Source Probability Field: Calculate node attention weights; Calculation of spatial probability distribution of outflow sources; risk assessment; S8. 3D Semantic Understanding: Point Cloud Recognition and Segmentation; Point Cloud Semantic Label Assignment; S9. Terrain geometric parameter extraction: Extract target point cloud clusters; calculate unit normal vectors; Key parameter calculation; S10. Accessibility Quantification Decision: Comparison of geometric parameters with robot physical capability thresholds; accessibility decision discrimination.

2. The collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, The specific method for step S2 is as follows: S2a, Raw data acquisition: The six-axis inertial measurement unit integrated into the system-controlled robot collects the angular velocity and acceleration in the three-axis directions in real time; S2b, Original Data Correction and Compensation: ① Temperature Compensation, using the formula and formula Corrections are made for angular velocity and acceleration; ② Zero bias correction is performed using the formula Calculate the corrected real-time state; S2c, State Prediction: Using the formula Calculate separately k Position, velocity, and attitude angle at any given moment; In the formula Indicates the corrected angular velocity, express k angular velocity at time, and Indicates the compensation coefficient, Indicates the corrected acceleration, express k acceleration at any moment They are respectively Position, velocity, and attitude angle at any given moment Indicates the sampling time interval. express Position, velocity, and attitude angle at any given moment This indicates the corrected acceleration. This indicates the corrected angular velocity.

3. The collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, The specific methods for step S3 include: S3a, Point Cloud Data Acquisition: The system controls the LiDAR to rotate and scan to obtain a point cloud composed of three-dimensional points; S3b, Noise Reduction Filtering and Motion Distortion Correction: Noise reduction filtering, using the formula... and formula Remove outliers caused by dust; correct motion distortion using formulas. After transforming all points, the corrected point cloud is obtained; S3c, State Update: Compare the current frame point cloud with the initial map inside the system, and then use the formula Calculate the rotation matrix R and translation vector describing the pose of the point cloud relative to the map in the current frame. ; In the formula Indicates the points being scanned. Indicates and Adjacent points, Represents each point and its nearest neighbors. The average distance between points The standard deviation is , and the corrected laser point is . The collection time is The end time of this frame scan is ; Represents the rotation matrix. Indicates the first The original coordinates of the laser points Indicates the first The target coordinates of the laser points are given by the translation vector. .

4. The collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, The specific methods for step S4 include: S4a, Kalman gain calculation: using the formula Calculate the observation residuals, and then use the formula Calculate the Kalman gain; S4b, Map Update: Using Formulas Update the height status using the formula Update the covariance, Through transformation Integrate into the system's initial map to update the map; utilize formulas Adjust the reliability of observations; In the formula For the first The predicted state value at time 10:

00. Indicates the first The observation residual at time, express k The state at any given moment, h For the observation function, The Kalman gain at time t is The prior estimation error covariance is The posterior estimation error covariance is ,function exist The Jacobian matrix is , It is the adjustment coefficient. It is a standard matrix. It is the average density of the point cloud, and the observation noise covariance matrix is... .

5. The collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, Step S5 includes the following specific methods: S5a, Spatiotemporal Tag Node Creation: The robot in time... Arrival Location When the sensor is triggered; S5b, Constructing Feature Vectors: Create a four-dimensional vector containing location information, gas concentration information, environmental information, and the entity's state. .

6. The collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, Step S6 includes the following specific methods: S6a. Define the spatiotemporal neighborhood of the node: Given the size of the underground tunnel and the typical gas diffusion rate, according to the formula... and formula Only spatial Euclidean distances less than 1 are calculated. and the time difference of adoption is less than Sampling point data; S6b, Multilayer Perceptron Optimization: Using Formula and formula , Scalar output of the network Perform normalization processing; S6c, Dynamic Adjustment of Gated Loop Unit: Using Formula To represent the average environmental characteristics, and then use the formula and Corrective measures for environmental factors; In the formula Indicates the points being scanned. Indicates and Adjacent points, nodes and its adjacent points The correlation strength between them is set to Time set to and , These are learnable parameters. It is a network output scalar. Average environmental characteristics of all active nodes at any given time , It represents the average hidden state of all nodes at present. This represents element-wise multiplication, with the gate vector being... .

7. The collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, The specific method for step S7 is as follows: S7a, Calculate node attention weights: , , Each node receives an attention weight. S7b, Calculation of Spatial Probability Distribution of Emission Sources: Using the formula , , Calculate the probability density; use the formula Calculate the reconstructed concentration field ; S7c, Risk Assessment: Using Formulas Calculate the risk score coefficient; In the formula, the query vector Key vector Value vector , It is the input feature matrix of all nodes. For attention weights, It is a learnable weight matrix. It is the dimension of the key vector. Represents the Gaussian function. The smoothness is represented by the overall attention weight. , The reconstructed concentration field uses pre-set weighting coefficients. The probability map of the outflow source is as follows The overall risk score is , It is the highest methane concentration in the region; The concentration exceeds the safety threshold. volume, It is an indicator function. It is the magnitude of the concentration gradient. It is the total intensity of the probability of the gushing source. Indicates the first Measured values ​​of gas concentration at each sampling point.

8. The collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, The specific method for step S9 is as follows: S9a. Extract target point cloud clusters: Extract target point cloud clusters based on the semantic labels defined in S8; S9b, Calculation of Unit Normal Vector: Using the formula Perform eigenvalue decomposition; S9c, Key Parameter Calculation: Using Formulas Calculate the slope of the target point cloud cluster using the formula Calculate the height of the target point cloud cluster using the formula Calculate the surface roughness of the target point cloud cluster; In the formula It is the center of the point cloud. Represents the covariance matrix of the point cloud. Represents the slope, and the horizontal plane normal vector is... The height is Surface roughness is , and This represents the maximum and minimum Z-coordinate values ​​of all points in the cloud cluster.

9. A collaborative decision-making method for underground gas inspection in coal mines according to claim 1, characterized in that, The specific method for step S10 is as follows: S10a, Comparison of geometric parameters with robot physical capability thresholds: Calculate the slope... ,high Roughness Compare with the robot's physical capability threshold; S10b, Passage Decision: Decision is made using the following formula... Based on the threshold comparison results, the robot is guided on whether to proceed; In the formula Indicates the slope, with a height of Roughness is , max This indicates the maximum slope the robot can climb. max Indicates the maximum height the robot can climb. This represents the maximum permissible roughness for the robot.

10. An inspection device employing the collaborative decision-making method for underground gas inspection in coal mines as described in any one of claims 1 to 9, characterized in that, It includes a drone base, gas sensor, walking assembly, drone shell, control module, sensing module, data sensing module, and communication module.

Images