Coal mine intrinsic safety variable structure open space robot catastrophe detection system and method

Abstract

The invention discloses a coal mine intrinsic safety variable structure open space robot catastrophe detection system, which belongs to the technical field of catastrophe robots and comprises a coal mine intrinsic safety variable structure open space robot consisting of a rack unit, a rotor wing mechanism, a battery unit, a control unit, an airborne sensor unit and the like. The invention further discloses a coal mine intrinsic safety variable structure open space robot catastrophe detection method. The method comprises the steps of collecting underground coal mine disaster photos, making a data set, building a YOLO10m network model and building an underground coal mine catastrophe detection system. And finally, a YOLO10m algorithm is improved, a TSample module is constructed, an up-sampling module in the neck network is replaced, the number of model parameters is reduced, the operation load is reduced, and a low-illumination loss function is provided to optimize the detection precision of the network on the underground coal mine catastrophe, so that the underground coal mine fire, flood and roof catastrophe are detected in the underground coal mine low-illumination environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of disaster robots, in particular to a variable-structure air-ground robot. BACKGROUND

[0002] The safety problem of coal mining has a great impact on the development of the coal industry. The early intelligent detection system of underground coal mine disasters mainly relies on physical and chemical sensors such as methane, carbon monoxide, temperature and wind speed for fixed-point monitoring, and uses a threshold-based alarm mechanism. Although this traditional detection mechanism has laid the foundation framework for safety monitoring, it has significant limitations such as blind area in monitoring coverage, high false alarm rate, and insufficient forward-looking in early warning.

[0003] In recent years, with the breakthrough of multi-source sensor fusion technology and artificial intelligence algorithm, underground disaster detection is rapidly developing towards intelligence, networking and precision. The current system widely introduces heterogeneous perception units such as visible light cameras, thermal imagers, laser radars and acoustic sensors, combined with deep learning models, to realize real-time identification and risk analysis of multiple disaster signs such as gas accumulation, conveyor belt tearing, early-stage open fire, smoke diffusion, equipment abnormal sound and personnel violation operation, greatly improving the early warning and active prevention and control capabilities of mine disasters.

[0004] However, in actual production, it is found that the extremely special environment in the underground mine still poses a serious challenge to intelligent detection technology: low light and high dust working conditions lead to serious degradation of the imaging quality of visual sensors, restricting the robustness of the algorithm; although mobile monitoring platforms such as mine unmanned aerial vehicles and robots have been introduced to expand the coverage range, their actual application is still limited by the single through function and insufficient perception accuracy. SUMMARY

[0005] In view of at least one of the above technical problems, the present application provides a coal mine intrinsic safety variable-structure air-ground robot disaster detection system and method, which can detect in a low-light environment in the underground mine, and the intrinsic safety variable-structure air-ground robot can increase the obstacle crossing ability and perform accurate disaster detection. The specific technical solutions are as follows: A coal mine intrinsically safe variable-configuration air-to-ground robot disaster detection system includes a frame with four arms capable of rotating 90°. The four arms are arranged in pairs on either side of the frame. A control compartment is located in the central area of ​​the frame, housing a control module and onboard sensors. A battery unit is located below the control compartment. Each arm is equipped with a rotor mechanism, which is controlled by the control module. The rotor mechanism includes a motor and a propeller. A wheel cover is fitted around the propeller, and a toothed ring is coaxially mounted on the inner surface of the wheel cover. The machine's output shaft is linked to the blade shaft, the upper end of which is connected to the propeller. The lower end of the blade shaft is connected to the transmission gear set via a clutch. The transmission gear set is connected to the drive shaft, which is equipped with a transmission gear that drives the gear ring. The clutch includes a release fork, which is equipped with a return spring and a traction rope. The outer end of the traction rope is controlled by a traction servo motor. When the boom is not rotating, the traction rope is in a tightened state, and the clutch is in a disengaged state. When the boom is in an outward-tilted state, the traction rope is in a relaxed state, and the release fork is engaged under the action of the return spring.

[0006] In some embodiments of this disclosure, the control module includes an intrinsically safe flight control board, an intrinsically safe electronic speed controller, an intrinsically safe airborne computer, and a control unit management module.

[0007] In some embodiments of this disclosure, the battery cell includes a battery pack, intrinsically safe protection circuitry, and a distributed power management module.

[0008] In some embodiments of this disclosure, the airborne sensors include a binocular structured light 3D camera, a millimeter-wave radar, an illumination unit, and a multi-component gas sensor.

[0009] In some embodiments of this disclosure, the frame further includes two side plates mirror-displayed on both sides, each side plate having a traction servo. The traction servo is equipped with a rotating disk, which is connected to a variable-structure rope and the traction rope. The rotating arm is rotated by pulling the variable-structure rope.

[0010] In some embodiments of this disclosure, the wheel cover further includes a cover frame, which includes an annular protective frame and several support strips.

[0011] Intrinsically safe variable-structure air-ground robot disaster detection method for coal mines.

[0012] In some embodiments of this disclosure, the following steps are included: S1 collects disaster information from underground coal mines using airborne sensors and creates a dataset. S2, Construct a YOLO10m network model; the YOLO10m network model includes a backbone network, a neck network, and a head network; the backbone network is used to extract features from photos of coal mine disasters and output multiple disaster feature maps in groups; the neck network is used to continuously downsample the multiple disaster feature maps to obtain various fused feature maps; the head network is used to detect the various fused feature maps and output the result map; the dataset is fed into the constructed YOLO10m network model for training; S3 improves the YOLO10m algorithm by constructing the TSample module to replace the upsampling module in the neck network, reducing the number of model parameters and computational load. It also proposes a low-illumination loss function to optimize the network's accuracy in detecting coal mine disasters.

[0013] In some embodiments of this disclosure, in step S3, the TSample module is constructed to replace the upsampling module in the neck network, reducing the number of model parameters and lowering the computational load. Specifically, the TSample module performs multi-scale linear projection processing on the input H×W×C features, mapping the number of feature channels to 2gs through three linear layers: Linear1, Linear2, and Linear3. 2 Multiple initial features are obtained, and the expression is parsed using the following equation 1: <Formula 1> in, The initial features are intermediate features that have been adjusted after the input features have undergone the i-th linear layer transformation, with dimensions such as the number of channels adjusted. As input features, ; This is the weight matrix. ; For bias, This initializes the channel dimensions of multi-scale features. The intermediate features are offset by the offsetO module, then multiplied by 0.25σ, 0.50σ, and 0.75σ for scale modulation to generate multi-scale features. The expression is analyzed using Equation 2 below: <Formula 2> Where Fi represents the multi-scale feature; σ is the learnable scale parameter; offsetO is the offset operation; k iThe scaling factor is used. The module utilizes the pixel shuffle operation to process multi-scale features. This operation rearranges the channel dimension information to the spatial dimension, achieving upsampling of the feature map and expanding the spatial size of the feature by a factor of s. At the same time, the number of channels is compressed to 2g. The features of the upper path after offset by offsetO are dynamically fused with the features obtained from the initial sampling through element-wise multiplication to obtain the final features that can be used for subsequent processing, thus enhancing the expressive power of the features. The expression is analyzed by the following equation 3: <Formula 3> Where M represents intermediate features; G represents the initial sampling operation; and PS represents the pixel rearrangement operation. These are three groups of multi-scale feature sets obtained after pixel rearrangement. .

[0014] In some embodiments of this disclosure, the low illumination loss function is defined as L dark It consists of a one-to-many branch loss L o2m One-to-one branch loss L o2o and the two-domain consistency loss L d The weighted composition, with weights controlled by a balancing hyperparameter λ, adapts to the complex requirements of low-light scenarios, and employs a dual-domain consistency loss L. d It is further subdivided into feature contrast loss L contrast and frequency reconstruction loss L frequency The feature contrast loss constrains the foreground average feature vector F through logarithmic and exponential operations. fg and background average eigenvector F bg The difference between the foreground and background features facilitates effective separation; the frequency reconstruction loss utilizes a high-pass filter operation ψ to calculate the feature map F and the reference feature map F. ref The norm of the result after ψ processing ensures the reconstruction accuracy of the frequency domain features. The combination of the two can simultaneously constrain the similarity of the feature domain and the reconstruction accuracy of the frequency domain. , , , , Among them, L dark For low illumination loss function, L o2m For a one-to-many branch loss function, L o2o The loss function is a one-to-one branch; L d Let L be the two-domain consistency loss function; λ is the balancing hyperparameter that controls the weight of the auxiliary loss; contrast F is the feature contrast loss function; fg F is the average eigenvector of the foreground; bgL is the average feature vector of the background. frequency ψ is the frequency reconstruction loss function; ψ is the high-pass filter operation; F is the feature map; F ref For reference feature map.

[0015] Compared with existing technologies, the above-mentioned intrinsically safe coal mine variable structure air-ground robot disaster detection system and method have the following advantages: 1. The intrinsically safe flight control board connects to the motor via an intrinsically safe ESC to control the motor speed, thereby enabling the intrinsically safe variable-structure air-to-ground robot in the coal mine to change its flight attitude. The customized intrinsically safe onboard computer connects to the onboard sensors, frame, or intrinsically safe flight control board via the control unit management module to collect environmental and body attitude information, thereby processing and analyzing the underground coal mine environmental information and driving the flight control board to achieve obstacle avoidance, navigation, and underground coal mine disaster detection for the intrinsically safe variable-structure air-to-ground robot in the coal mine. 2. The TSample module uses the Pixel shuffle operation to process multi-scale features. This operation rearranges the information in the channel dimension to the spatial dimension, realizes the upsampling of the feature map, expands the spatial size of the feature by a factor of s, and compresses the number of channels to 2g. The features of the upper path after offset by offsetO are dynamically fused with the features obtained by the initial sampling through element-wise multiplication to obtain the final features that can be used for subsequent processing, thus enhancing the expressive power of the features. 3. The TSample module adopts an efficient upsampling method, which can dynamically learn the coordinates of sampling points to upsample the feature map. While preserving the basic features of the input data, it reduces the computational load and the size of the YOLO10m algorithm model. 4. The low-light loss function can comprehensively address the complex needs of low-light scenes by combining and weighting multiple loss terms. The dual-domain consistency loss combines feature contrast loss and frequency reconstruction loss, which enhances target recognition by constraining the separation of foreground and background features, and retains key details by accurately reconstructing frequency domain features, thereby improving the robustness of feature extraction in low-light environments. The integration of multi-branch losses can take into account constraints of different dimensions, allowing the model to learn more stably under complex lighting conditions. Attached Figure Description

[0016] Figure 1 This is a top view schematic diagram of the intrinsically safe variable-structure air-ground robot for coal mines according to the present invention; Figure 2 This is a schematic diagram of the rotor mechanism of the present invention; Figure 3 This is a schematic diagram of the structure of the traction servo motor of the present invention; Figure 4 A flowchart illustrating the implementation of the intrinsically safe variable-structure air-ground robot disaster detection method for coal mines according to the present invention; Figure 5This is a network model diagram of the improved YOLO10m algorithm implemented in the inherently safe variable-structure air-ground robot disaster detection method for coal mines of the present invention; Figure 6 This is a diagram of the Tsample module implemented in the intrinsically safe variable-structure air-ground robot disaster detection method for coal mines of the present invention; The following are the labels in the diagram: 1. Frame; 11. Arm; 12. Side plate; 2. Control compartment; 3. Rotor mechanism; 31. Motor; 32. Propeller; 33. Wheel cover; 331. Gear ring; 34. Blade shaft; 35. Gearbox; 36. Drive shaft; 361. Drive gear; 37. Separation fork; 38. Traction rope; 41. Traction rope servo; 42. Variable structure rope; 43. Rope post. Detailed Implementation

[0017] To better understand the purpose, structure, and function of this invention, the technical solutions in the embodiments of this invention will be clearly and completely described below. It should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing specific embodiments only and is not intended to limit this application. The terms "comprising" and "having," and any variations thereof, are open-ended and intended to cover non-exclusive inclusion. In the description of this application, it should be understood that the directional terms "up," "down," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate directional or positional relationships based on the appendix. Figure 2 The orientations or positional relationships shown are for ease of description only and are not intended to indicate or imply that the device or unit referred to must have a specific orientation, be constructed or operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0018] As shown in the attached diagram. Figures 1 to 3 As shown, this embodiment discloses an intrinsically safe variable-structure air-ground robot for coal mines, including a frame 1. The frame 1 includes four arms 11 that can rotate 90°. The four arms 11 are arranged in pairs on both sides of the frame 1. A control compartment 2 is installed in the central area of ​​the frame 1. Further optimization is that a center plate is installed in the central area of ​​the frame 1, and the control compartment 2 is connected to the frame 1 through the center plate. The control compartment 2 is an explosion-proof shell. The control compartment 2 is equipped with a control module and onboard sensors. A battery unit is installed below the control compartment 2.

[0019] The control module includes an intrinsically safe flight control board, an intrinsically safe electronic speed controller (ESC), an intrinsically safe onboard computer, and a control unit management module. The intrinsically safe flight control board connects to motor 31 via the intrinsically safe ESC to control the motor's speed, enabling the intrinsically safe variable-structure air-to-ground robot to change its flight attitude. The intrinsically safe onboard computer connects to onboard sensors, the frame, and the intrinsically safe flight control board via the control unit management module to collect environmental and body attitude information. This allows for the processing and analysis of underground environmental information and the driving of the intrinsically safe flight control board to achieve obstacle avoidance, navigation, and underground disaster monitoring for the intrinsically safe variable-structure air-to-ground robot.

[0020] The battery unit includes a battery pack, an intrinsically safe protection circuit, and a dynamically balanced distributed power intelligent management module. The battery pack supplies power to the rotor mechanism 3, frame 1, control unit, and airborne sensors through the intrinsically safe protection circuit. At the same time, the dynamically balanced distributed power intelligent management module monitors parameters such as voltage, current, and temperature in the battery pack and individual cells in real time, realizing first-level protection for the battery pack. It also aggregates and centrally controls the battery pack data.

[0021] The airborne sensors include a binocular structured light 3D camera, a millimeter-wave radar, an illumination unit, and a multi-component gas sensor. The illumination unit uses low-power LED lights, and the multi-component gas sensor is a miniature multi-component gas sensor. When the intrinsically safe, variable-structure air-to-ground robot collects environmental information in this coal mine, the low-power LED lights provide efficient illumination, the miniature binocular structured light 3D camera acquires depth and RGB maps of the coal mine, the miniature millimeter-wave radar acquires point cloud information from the coal mine, and the miniature multi-component gas sensor can collect data on methane concentration, dust concentration, ambient humidity, and temperature in the coal mine.

[0022] The arm 11 is equipped with a rotor mechanism 3, and the arm 11 and the rotor mechanism 3 are controlled by a control module. The rotor mechanism 3 includes a motor 31 and a propeller 32. The propeller 32 is equipped with a wheel cover 33 on its outside. The inner surface of the wheel cover 33 is equipped with a toothed ring 331 coaxially. The wheel cover 33 also includes a cover frame, which includes a circular protective frame and several support strips.

[0023] The output shaft of the motor 31 is linked to the blade shaft 34. The upper end of the blade shaft 34 is connected to the propeller 32. The lower end of the blade shaft 34 is linked to the speed change gear set 35 through the clutch. The speed change gear set 35 is linked to the transmission shaft 36. The transmission shaft 36 is equipped with the transmission gear 361 that is linked to the gear ring 331.

[0024] The clutch includes a release fork 37, which is equipped with a return spring and a traction rope 38. The outer end of the traction rope 38 is controlled by a traction servo motor 41, and the operation of the traction servo motor 41 is controlled by a control module.

[0025] In this embodiment, it can be further disclosed that, as Figure 3As shown, the frame 1 also includes two side plates 12 mirror-imagely disposed on both sides. A traction servo 41 is provided in the central area of ​​the side plate 12, positioning the traction servo 41 between the two rotor mechanisms on the same side. The traction servo 41 is equipped with a rotating disk, which is connected to a variable-structure rope 42 and the traction rope 38. The variable-structure rope 42 is as follows... Figure 3 As shown by the red line, the traction rope is 38mm long. Figure 3 As shown by the blue line, when the traction servo motor 41 is running, the rotating disk rotates, thereby pulling the variable-structure rope 42 and the traction rope 38. The variable-structure rope 42 is wound around the outer surface of the arm 11. When the variable-structure rope 42 is pulled, the friction between the variable-structure rope 42 and the outer surface of the arm 11 causes the arm 11 to rotate. Ideally, the variable-structure rope 42 should be a loop-shaped rope, with one side of its middle section coiled around the rotating disk, and each end wrapped around and contacting the outer surface of one arm 11. This allows both arms 11 on the same side of the frame 1 to rotate simultaneously when the variable-structure rope 42 is driven by the rotating disk. A further optimization in this embodiment is that the side plate 12 has a rope post 43 between the rotating disk and the arm 11, making the middle of the variable-structure rope 42 concave, thereby increasing the contact angle between the variable-structure rope 42 and the outer surface of the arm 11, thus increasing the friction. The middle of the traction rope 38 also contacts the rope post. In this design, the rotating disk can be equipped with two grooves to avoid interference between the variable rope 42 and the traction rope 38.

[0026] It should be noted that in this embodiment, the two rope-pulling servos 41 are controlled by the control module to operate synchronously.

[0027] When the arm 11 is not rotating, the propeller 32 is in a horizontal position, the traction rope 38 is in a tightened state, the clutch is in a disengaged state, and the inherently safe variable structure air-ground robot of this coal mine is in flight mode, relying on the propeller 32 to achieve overall displacement.

[0028] When the boom 11 is in the outward tilted position, the propeller 32 is in a vertical position, the traction rope 38 is in a relaxed state, and the release fork is in the engaged state under the action of the return spring. At this time, the motor 31 drives the gear ring 331 to rotate through the clutch, the gear set 35, the drive shaft 36, and the drive gear 361 in sequence, thereby driving the wheel cover 33 to rotate. The outer surface of the wheel cover 33 contacts the ground to function as a wheel, realizing the ground mode.

[0029] The intrinsically safe flight control board connects to the motor via an intrinsically safe ESC to control the motor speed, enabling the intrinsically safe variable-structure air-to-ground robot in the coal mine to change its flight attitude. The customized intrinsically safe onboard computer connects to the onboard sensors, frame 1, and intrinsically safe flight control board through the control unit management module to collect environmental and body attitude information, realize the processing and analysis of underground coal mine environmental information, and drive the flight control board to realize obstacle avoidance, navigation, and underground coal mine disaster monitoring of the intrinsically safe variable-structure air-to-ground robot in the coal mine.

[0030] like Figures 4 to 6As shown, a detection method based on the aforementioned intrinsically safe variable-structure air-ground robot for coal mines is also disclosed.

[0031] The main steps include: S1 collects disaster information from underground coal mines using airborne sensors and creates a dataset. Multiple photos can be acquired using a miniature binocular structured light 3D camera. These photos can be annotated using the LabelImg annotation tool, and corresponding text annotation files can be generated. The files are then divided into training and validation sets in an 8:2 ratio, and a data.yaml configuration file is created to specify the image paths, number of categories, and name list. The intrinsically safe variable-structure air-to-ground robot uses a miniature multi-component gas sensor to collect dust and gas concentration information and determine the dust and gas concentration thresholds when actual disasters occur. S2. Construct a YOLO10m network model. The YOLO10m network model includes a backbone network, a neck network, and a head network. The backbone network is used to extract features from photos of coal mine disasters, outputting multiple disaster feature maps in groups. The neck network is used to continuously downsample the multiple disaster feature maps to obtain various fused feature maps. The head network is used to detect the various fused feature maps and output the result map. The dataset is then fed into the constructed YOLO10m network model for training. The ROS operating system was installed in the intrinsically safe airborne computer of the control module, and the trained YOLO10m network model was deployed in the intrinsically safe airborne computer. The API interfaces of the customized miniature binocular structured light 3D camera, miniature millimeter-wave radar, and miniature multi-component gas sensor in the airborne sensors were connected, and their connection with the data buffer was configured to ensure that the algorithm can correctly receive input data and start the processing process.

[0032] After the initial connection of all modules of the disaster detection system, interface debugging is carried out to verify whether the data format, communication protocol, and call frequency are compatible, ensuring that each module can work together as expected and realize the core detection function. Dust and gas concentration thresholds are set, and miniature multi-component gas sensors collect underground environmental information in coal mines and transmit it to the onboard computer to determine whether dust or gas disasters have occurred. A customized miniature binocular structured light 3D camera collects underground environmental information in coal mines and transmits the collected photos to the YOLO10m network model on the onboard computer to determine whether floods, fires, and roof collapses have occurred.

[0033] S3 improves the YOLO10m algorithm by constructing the TSample module to replace the upsampling module in the neck network, reducing the number of model parameters and computational load. It also proposes a low-light loss function to optimize the network's accuracy in detecting coal mine disasters, thereby enabling the detection of coal mine disasters in low-light environments.

[0034] In step S3, the TSample module is constructed to replace the upsampling module in the neck network, reducing the number of model parameters and lowering the computational load. Specifically, the TSample module performs multi-scale linear projection processing on the input H×W×C features, mapping the number of feature channels to 2gs through three linear layers: Linear1, Linear2, and Linear3. 2 Multiple initial features are obtained, and the expression is parsed using the following equation 1: <Formula 1> in, The initial features are intermediate features that have been adjusted after the input features have undergone the i-th linear layer transformation, with dimensions such as the number of channels adjusted. As input features, ; This is the weight matrix. ; For bias, This initializes the channel dimensions of multi-scale features. The intermediate features are offset by the offsetO module, then multiplied by 0.25σ, 0.50σ, and 0.75σ for scale modulation to generate multi-scale features. The expression is analyzed using Equation 2 below: <Formula 2> Where Fi represents the multi-scale feature; σ is the learnable scale parameter; offsetO is the offset operation; k i The scaling factor is used. The module utilizes the pixel shuffle operation to process multi-scale features. This operation rearranges the channel dimension information to the spatial dimension, achieving upsampling of the feature map and expanding the spatial size of the feature by a factor of s. At the same time, the number of channels is compressed to 2g. The features of the upper path after offset by offsetO are dynamically fused with the features obtained from the initial sampling through element-wise multiplication to obtain the final features that can be used for subsequent processing, thus enhancing the expressive power of the features. The expression is analyzed by the following equation 3: <Formula 3> Where M represents intermediate features; G represents the initial sampling operation; and PS represents the pixel rearrangement operation. These are three groups of multi-scale feature sets obtained after pixel rearrangement. The TSample module employs an efficient upsampling method that dynamically learns the coordinates of sampling points to upsample the feature map. This reduces computational load and decreases the size of the YOLO10m algorithm model while preserving the basic features of the input data.

[0035] YOLOv10's original loss function is significantly inadequate in low-light environments. Its bounding box regression loss, confidence loss, and classification loss are mainly designed for normal lighting, making it difficult to adapt to the characteristics of low image contrast and blurred features under low light conditions. This can easily lead to difficulties in distinguishing between positive and negative samples and a decrease in localization accuracy. Furthermore, it lacks specific constraints on frequency domain features and foreground-background separation, making it unable to effectively capture the key features of targets under low light conditions. Its detection performance will significantly degrade in complex dark light scenes.

[0036] To address this, a low-illumination loss function is proposed to optimize the network's accuracy in detecting underground coal mine disasters. In step S3, the low-illumination loss function is defined as L... dark It consists of a one-to-many branch loss L o2m One-to-one branch loss L o2o and the two-domain consistency loss L d The weighted composition, with weights controlled by a balancing hyperparameter λ, adapts to the complex requirements of low-light scenarios, and employs a dual-domain consistency loss L. d It is further subdivided into feature contrast loss L contrast and frequency reconstruction loss L frequency The feature contrast loss constrains the foreground average feature vector F through logarithmic and exponential operations. fg and background average eigenvector F bg The difference between the foreground and background features facilitates effective separation; the frequency reconstruction loss utilizes a high-pass filter operation ψ to calculate the feature map F and the reference feature map F. ref The norm of the result after ψ processing ensures the reconstruction accuracy of the frequency domain features. The combination of the two can simultaneously constrain the similarity of the feature domain and the reconstruction accuracy of the frequency domain. , , , , Among them, L dark For low illumination loss function, L o2m For a one-to-many branch loss function, L o2o The loss function is a one-to-one branch; L d Let L be the two-domain consistency loss function; λ is the balancing hyperparameter that controls the weight of the auxiliary loss; contrast F is the feature contrast loss function; fg F is the average eigenvector of the foreground;bg L is the average feature vector of the background. frequency ψ is the frequency reconstruction loss function; ψ is the high-pass filter operation; F is the feature map; F ref For reference feature map.

[0037] It is understood that the above description is only for illustrating specific embodiments of this application, but the protection scope of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of disclosure of this application.

Claims

1. A coal mine intrinsically safe variable-structure air-to-ground robot disaster detection system, comprising a frame (1), characterized in that: The frame (1) includes four arms (11) that can rotate 90°. The four arms (11) are arranged in pairs on both sides of the frame (1). The central area of ​​the frame (1) is equipped with a control compartment (2). The control compartment (2) is equipped with a control module and airborne sensors. A battery unit is located below the control compartment (2). The arms (11) are equipped with rotor mechanisms (3). The arms (11) and rotor mechanisms (3) are controlled by the control module. The rotor mechanism (3) includes a motor (31) and a propeller (32). The propeller (32) is equipped with a wheel cover (33). The inner surface of the wheel cover (33) is coaxially equipped with a toothed ring (331). The output shaft of the motor (31) is linked to the blade shaft (34). The upper end of the blade shaft (34) is connected to the propeller (32), and the lower end of the blade shaft (34) is connected to the speed change gear set (35) through the clutch. The speed change gear set (35) is connected to the transmission shaft (36), and the transmission shaft (36) is equipped with the transmission gear (361) of the connecting gear ring (331). The clutch includes a release fork (37), which is equipped with a return spring and a traction rope (38). The outer end of the traction rope (38) is controlled by the traction servo motor (41). When the arm (11) is not rotating, the traction rope (38) is in a tightened state, and the clutch is in a disengaged state. When the arm (11) is in an outward-folded state, the traction rope (38) is in a relaxed state, and the release fork is in an engaged state under the action of the return spring.

2. The intrinsically safe variable-structure air-to-ground robot disaster detection system for coal mines according to claim 1, characterized in that, The control module includes an intrinsically safe flight control board, an intrinsically safe electronic speed controller, an intrinsically safe airborne computer, and a control unit management module.

3. The intrinsically safe variable-structure air-to-ground robot disaster detection system for coal mines according to claim 1, characterized in that, The battery unit includes a battery pack, an intrinsically safe protection circuit, and a distributed power management module.

4. The intrinsically safe variable-structure air-to-ground robot disaster detection system for coal mines according to claim 1, characterized in that, The airborne sensors include a binocular structured light 3D camera, a millimeter-wave radar, an illumination unit, and a multi-component gas sensor.

5. The intrinsically safe variable-structure air-to-ground robot disaster detection system for coal mines according to claim 1, characterized in that, The frame (1) also includes two side plates (12) mirrored on both sides. The side plates (12) are equipped with a traction servo (41). The traction servo (41) is equipped with a rotating disk. The rotating disk is connected to the variable structure rope (42) and the traction rope (38). The arm (11) is rotated by pulling through the variable structure rope (42).

6. The intrinsically safe variable-structure air-to-ground robot disaster detection system for coal mines according to claim 1, characterized in that, The wheel cover (33) also includes a cover frame, which includes a circular protective frame and several support strips.

7. A disaster detection method for an intrinsically safe variable-structure air-to-ground robot in coal mines, characterized in that, The method of using the intrinsically safe variable structure air-ground robot disaster detection system for coal mines as described in any one of claims 1 to 6.

8. The intrinsically safe variable-structure air-to-ground robot disaster detection system for coal mines according to claim 7, characterized in that, Includes the following steps: S1 collects disaster information from underground coal mines using airborne sensors and creates a dataset. S2, Construct a YOLO10m network model; the YOLO10m network model includes a backbone network, a neck network, and a head network; the backbone network is used to extract features from photos of coal mine disasters and output multiple disaster feature maps in groups; the neck network is used to continuously downsample the multiple disaster feature maps to obtain various fused feature maps; the head network is used to detect the various fused feature maps and output the result map; the dataset is fed into the constructed YOLO10m network model for training; S3 improves the YOLO10m algorithm by constructing the TSample module to replace the upsampling module in the neck network, reducing the number of model parameters and computational load. It also proposes a low-illumination loss function to optimize the network's accuracy in detecting coal mine disasters.

9. The intrinsically safe variable-structure air-to-ground robot disaster detection method for coal mines according to claim 8, characterized in that, In step S3, the TSample module is constructed to replace the upsampling module in the neck network, reducing the number of model parameters and lowering the computational load. Specifically, the TSample module performs multi-scale linear projection processing on the input H×W×C features, mapping the number of feature channels to 2gs through three linear layers: Linear1, Linear2, and Linear3. 2 Multiple initial features are obtained, and the expression is parsed using the following equation 1: <Formula 1> in, The initial features are intermediate features that have been adjusted after the input features have undergone the i-th linear layer transformation, with dimensions such as the number of channels adjusted. As input features, ; This is the weight matrix. ; For bias, This initializes the channel dimensions of multi-scale features. The intermediate features are offset by the offsetO module, then multiplied by 0.25σ, 0.50σ, and 0.75σ for scale modulation to generate multi-scale features. The expression is analyzed using Equation 2 below: <Formula 2> Where Fi represents the multi-scale feature; σ is the learnable scale parameter; offsetO is the offset operation; k i The scaling factor is used. The module utilizes the pixel shuffle operation to process multi-scale features. This operation rearranges the channel dimension information to the spatial dimension, achieving upsampling of the feature map and expanding the spatial size of the feature by a factor of s. At the same time, the number of channels is compressed to 2g. The features of the upper path after offset by offsetO are dynamically fused with the features obtained from the initial sampling through element-wise multiplication to obtain the final features that can be used for subsequent processing, thus enhancing the expressive power of the features. The expression is analyzed by the following equation 3: <Formula 3> Where M represents intermediate features; G represents the initial sampling operation; and PS represents the pixel rearrangement operation. These are three groups of multi-scale feature sets obtained after pixel rearrangement. .

10. The intrinsically safe variable-structure air-to-ground robot disaster detection system for coal mines according to claim 9, characterized in that, In step S3, the low illumination loss function is defined as L dark It consists of a one-to-many branch loss L o2m One-to-one branch loss L o2o and the two-domain consistency loss L d The weighted composition, with weights controlled by a balancing hyperparameter λ, adapts to the complex requirements of low-light scenarios, and employs a dual-domain consistency loss L. d It is further subdivided into feature contrast loss L contrast and frequency reconstruction loss L frequency The feature contrast loss constrains the foreground average feature vector F through logarithmic and exponential operations. fg and background average eigenvector F bg The differences enable effective separation of foreground and background features; The frequency reconstruction loss is calculated by operating ψ with a high-pass filter to compare the feature map F with the reference feature map F. ref The norm of the result after ψ processing ensures the reconstruction accuracy of the frequency domain features. The combination of the two can simultaneously constrain the similarity of the feature domain and the reconstruction accuracy of the frequency domain. ， ， ， ， Among them, L dark For low illumination loss function, L o2m For a one-to-many branch loss function, L o2o The loss function is a one-to-one branch; L d Let L be the two-domain consistency loss function; λ is the balancing hyperparameter that controls the weight of the auxiliary loss; contrast F is the feature contrast loss function; fg F is the average eigenvector of the foreground; bg L is the average feature vector of the background. frequency ψ is the frequency reconstruction loss function; ψ is the high-pass filter operation; F is the feature map; F ref For reference feature map.

Images