Non-spark forcible entry processing system for coal mine based on image processing
By using the YOLOv4 focused detection network and multi-module collaborative design in underground coal mines, the risk of spark explosions during coal gangue identification and demolition was solved, achieving robust identification and accurate positioning in complex environments, and ensuring the safety and efficiency of the demolition process.
Patent Information
- Application Number
- CN202511304145.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-12
- Publication Date
- 2025-11-28
AI Technical Summary
Existing technologies are insufficient to effectively identify coal gangue and prevent sparks from triggering gas/coal dust explosions in the complex environment of underground coal mines, especially under conditions of low light, dust, and water mist.
A YOLOv4-based focusing detection network is adopted, combined with the MobileNet-v3 backbone network, pixel space focusing module and SE channel attention mechanism to suppress interference in non-significant areas and enhance the features of key areas; combined with explosion-proof imaging and gimbal components, robotic arm control module and hazardous gas monitoring module, spark-free demolition is achieved.
Robust coal and gangue identification and precise alignment were achieved in complex underground environments, reducing the rate of false detections and missed detections, ensuring the safety and efficiency of the demolition process, and preventing sparks through high-pressure spraying and intrinsically safe interlocking.
Smart Images
Figure CN121024688A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of image processing, and more particularly, to a coal mine spark-free breaking treatment system based on image processing. BACKGROUND
[0002] The removal of underground coal mine sealing walls, large coal gangue and roof fall area obstacles must not only be efficient, but also must avoid sparks causing gas / coal dust explosions. Traditional manual or general visual solutions are difficult to stably identify breaking objects in low-illumination, dust, water mist and other underground working conditions.
[0003] Existing public document 1 (Research on coal gangue image recognition based on image processing technology and InceptionV3 neural network, 2023) proposes a support vector machine coal gangue image recognition system, as shown in Figure 2 The support vector machine coal gangue image recognition system uses image preprocessing and feature extraction methods to perform grayscale and mean filter noise reduction processing on the collected coal and gangue images, thereby extracting the gray variance reflecting the image difference as a classification feature, and inputting it into the support vector machine (SVM) classifier. The SVM realizes the binary classification of coal and gangue images by constructing an optimal hyperplane, and divides the coal gangue images into different category spaces, thereby realizing recognition. However, the support vector machine has insufficient generalization ability in complex underground environments, and when there is coal dust and water interference, the recognition effect will be significantly reduced.
[0004] Existing public document 2 (Design and experimental study of coal gangue recognition system based on machine vision, 2021) proposes a coal gangue recognition system based on machine vision, as shown in Figure 3 The coal gangue recognition system based on machine vision uses an industrial camera to collect coal and gangue images in real time on a conveyor belt, removes noise and background after image preprocessing (filtering, sharpening, threshold segmentation and morphological processing), extracts gray mean value, peak gray value, energy, entropy, contrast and inverse difference matrix, and then uses a particle swarm optimization algorithm (PSO) to optimize and train the penalty parameter C and kernel function width parameter g of the support vector machine (SVM), to realize the optimal performance of the classifier. However, the gray features of the coal gangue surface will change due to factors such as coal dust and water, and the robustness of fixed threshold or single feature in complex working conditions is insufficient.
[0005] Therefore, there is an urgent need for a coal mine spark-free breaking treatment system that can suppress the interference of non-significant areas of coal dust and water mist when detecting coal gangue. SUMMARY
[0006] To overcome the aforementioned deficiencies of the prior art, this invention provides a sparkless demolition system for coal mines based on image processing. This system proposes a focus detection network using YOLOv4 as the overall framework and MobileNet-v3 as the backbone network. A collaborative attention mechanism, formed by a pixel spatial focusing module and SE channel attention, is introduced into the backbone network to suppress spatially irrelevant regions and enhance key regions, thereby solving the problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: A sparkless demolition system for coal mines based on image processing includes a sparkless demolition actuator and a hazardous gas monitoring and safety interlock module; it also includes an explosion-proof imaging and pan-tilt unit for real-time acquisition of roadway target images and line-of-sight orientation. The target identification and tracking module identifies and tracks coal and gangue. The coordinate deviation-servo angle conversion module converts the pixel deviation between the target center and the image center into gimbal horizontal / tilt angle commands in real time. The robotic arm control module drives the multi-degree-of-freedom hydraulic arm to adjust its posture. The target recognition and tracking module is based on a focused detection network. It identifies coal gangue by suppressing interference from insignificant areas of coal dust and water mist attached to the surface of the coal gangue. A collaborative attention mechanism, formed by a pixel spatial focusing module and SE channel attention, is introduced into the second, fourth, and fifth bneck modules of the focused detection network backbone. The input features of the pixel spatial focusing module first undergo pixel regularization. The scaling factor of the first batch normalization layer in the bneck module is used as a measure of pixel importance and normalization is completed to obtain a steady-state normalized output. A pixel weight map is constructed based on the scaling factor to represent the importance of each spatial location. The pixel weight map is multiplied pixel-by-pixel by the first batch normalization layer, thereby selectively amplifying and suppressing spatial features: the weights of locations relevant to the target and with clearer information are enhanced, while the weights of locations unrelated to the target are suppressed.
[0008] As a further aspect of the present invention, the pixel spatial focusing module is used to suppress spatially irrelevant regions and enhance key regions. It is based on the premise that the parameters of the first batch normalization layer can characterize spatial importance. The first batch normalization layer is located after a 3x3 depthwise convolution. The scaling factor of the first batch normalization layer is used to normalize the features pixel-by-pixel, and then these are aggregated into a pixel weight matrix on the entire feature map. The weight at each position characterizes the importance of that pixel in the overall image context. The pixel spatial focusing module performs a one-to-one spatial dot product between the normalized feature output and the pixel weight matrix. For positions with weights greater than the average level, their corresponding features are explicitly suppressed; for positions with weights less than the average level, their corresponding features are explicitly suppressed.
[0009] As a further embodiment of the present invention, an explosion-proof imaging and gimbal assembly is used to acquire target images in the tunnel in real time and achieve line-of-sight orientation, including the following specific contents: The explosion-proof imaging and gimbal assembly consists of a mining explosion-proof camera, a two-degree-of-freedom servo gimbal, and a gimbal controller, and is installed at the front end of the demolition robot as a vision base; the camera continuously acquires video frames in the tunnel and sends them to the upper processing unit. After initializing and selecting the demolition target, based on the target center pixel coordinates detected in each frame and the offset of the image center, the horizontal angle and pitch angle are calculated by the coordinate deviation-servo angle conversion module and sent to the gimbal for execution in real time, thereby achieving line-of-sight alignment and stable tracking.
[0010] As a further aspect of the present invention, the coordinate deviation-servo angle conversion module converts the pixel deviation between the target center and the image center into gimbal horizontal / tilt angle commands in real time. This includes the following specific details: The coordinate deviation-servo angle conversion module takes the target center pixel coordinates and timestamp output by the focus detection network as input. Combined with the image principal point obtained from power-on calibration, camera intrinsic parameters, lens field of view, and gimbal zero position, it first sets pixel-level overlap criteria and dead zone thresholds: when the target center and image center fall into the dead zone, they are considered aligned; if not aligned, the pixel deviations in the horizontal and vertical directions are calculated, and the pixel deviations are converted into line-of-sight deviation angles based on imaging geometry to obtain the target angle increments for gimbal horizontal and pitch.
[0011] As a further aspect of the present invention, the robotic arm control module drives a multi-degree-of-freedom hydraulic arm to perform pose adjustment, including the following specific content: The robotic arm control module takes the target pose requirement output from the "detection / tracking—coordinate deviation—gimbal orientation" link as input, and drives the multi-degree-of-freedom hydraulic arm to complete the continuous action of "coarse alignment—fine alignment—attitude fine adjustment": In the coarse alignment stage, the robotic arm control module, based on hand-eye calibration and the roadway safety boundary, calls the inverse solver to plan the end effector TCP from the standby position to the working transition position, allocates the displacement, velocity, and acceleration of each joint, and performs obstacle avoidance; Entering the fine alignment stage, combined with the target surface normal constraint and spray envelope limitation, through small step trajectory and velocity feedforward, the end effector is gradually centered and the included angle between the end effector tool and the target surface is controlled to be lower than a set threshold; During the attitude fine-tuning phase, the proportional / servo valve is controlled in conjunction with the position-velocity dual-loop and hydraulic pressure limiting to implement valve dead zone compensation and backlash compensation, suppressing low-speed crawling and vibration. If necessary, a "micro-motion" command is superimposed to ensure stable contact of the end effector within the confined space. Joint limit, speed / acceleration limiting, hydraulic pressure and temperature rise monitoring are performed throughout the process, and abnormalities such as overload, jamming, and excessive position deviation are triggered to quickly retract to the safe position. At the same time, it is tightly coupled with the safety interlock: if methane or carbon monoxide exceeds the limit, spray pressure / flow is abnormal, or the end effector leaves the spray envelope, the valve group is immediately frozen, the proximity command is canceled, and an alarm is triggered. The module continuously corrects zero-point drift and thermal drift, and adaptively corrects the trajectory by combining the online estimated target depth and body attitude error to ensure stable end effector centering and attitude control even under low light, dust and water mist interference.
[0012] As a further aspect of the present invention, a sparkless demolition actuator is used for dust suppression, cooling, and fire source isolation during sparkless demolition of large coal and gangue blocks underground. The specific details are as follows: The sparkless demolition actuator consists of a composite-coated hydraulic breaker, hydraulic shears, and a high-pressure spray system. The high-pressure spray system comprises a water tank, a high-pressure pump, a spray device, and a control unit, with a designed flow rate of no less than 20 liters per minute. High-pressure nozzles are arranged around the demolition end, atomizing water into fine droplets and forming a dense water mist envelope in the impact zone. This mist collides, adsorbs, and condenses with coal dust and rock powder particles in the air, causing them to increase in weight and settle, thereby reducing dust concentration, inertizing flammability, and cooling the end and near-field environment, suppressing heat accumulation and secondary hazards caused by surface incandescence.
[0013] As a further embodiment of the present invention, the hazardous gas monitoring and safety interlocking module performs real-time monitoring of methane and carbon monoxide and triggers intrinsically safe interlocking according to graded thresholds, including the following specific contents: The hazardous gas monitoring and safety interlocking module consists of two types of intrinsically safe sensing units, namely methane and carbon monoxide, and is equipped with temperature, humidity and air pressure compensation channels, sampling and communication interfaces, an integrated sound, light and voice alarm and an interlocking output relay. The hazardous gas monitoring and safety interlock module includes sensor self-testing and fault-safe strategies (disconnection, drift exceeding limits, overdue calibration, and abnormal noise are all handled according to the worst-case scenario and trigger shutdown). It supports zero-point / span calibration and periodic verification, and automatically shields interlock outputs during verification to prevent malfunctions. All key quantities (multi-point concentration, rate of change, threshold comparison results, interlock trigger / release timestamps, and manual confirmation records) and operating environment quantities (temperature, humidity, wind speed / direction if connected) are archived for a long period to facilitate post-event traceability and parameter tuning. The hazardous gas monitoring and safety interlock module is also linked with "coordinate deviation - servo angle conversion" and "robotic arm control": once an early warning is triggered, the front end is only allowed to make minor attitude corrections and is prohibited from approaching further; if the limit is exceeded, it is forcibly withdrawn to a safe position, and the demolition circuit and quick oil change circuit are shut down. Through the closed-loop design of "multi-point sensing - threshold grading - intrinsically safe interlock - traceable recording", the intrinsically safe control of the demolition process is achieved under the condition of rapid changes in the risk of underground gas and carbon monoxide.
[0014] The technical effects and advantages of this invention's image processing-based sparkless demolition system for coal mines are as follows: This invention significantly increases the weight of salient areas (coal and gangue outlines and main texture) and decreases the weight of insignificant areas to address interference caused by coal dust and water mist, thereby achieving more robust detection and fewer false positives / false negatives. Coordinate deviation-angle conversion combined with amplitude limiting, filtering, and closed-loop control enables the gimbal and multi-degree-of-freedom hydraulic arm to complete continuous movements of "coarse approach—fine approach—attitude fine adjustment," ensuring precise centering of the end effector within the safety spray envelope. The use of a composite-coated breaker / hydraulic shear reduces metal... To mitigate the risk of direct contact with ignition sources, a high-pressure spray with a design flow rate of no less than 20 L / min is used to create a dense water mist envelope in the impact zone, achieving dust suppression, cooling, and an inert atmosphere, thus achieving essentially "spark-free" demolition. Multi-point monitoring of methane and carbon monoxide triggers intrinsically safe interlocks based on graded thresholds; exceeding these limits immediately cuts off the demolition output, freezes the operation, and maintains or increases the spray. Combined with a recovery logic based on hysteresis and pressure holding time, and a fail-safe strategy, a complete "early warning—alarm—shutdown—recovery" protection chain is formed. Simultaneously, real-time recording of key quantities and interlock events enhances traceability and parameter tuning efficiency. In summary, this invention achieves robust identification, precise alignment, and safe and efficient spark-free demolition in complex downhole environments. Attached Figure Description
[0015] Figure 1This is a schematic diagram of the structure of a sparkless demolition system for coal mines based on image processing, according to the present invention.
[0016] Figure 2 This is a diagram of a coal gangue image recognition system based on existing technology, which uses support vector machines.
[0017] Figure 3 This refers to an existing machine vision-based coal and gangue identification system.
[0018] Figure 4 This is a schematic diagram of the backbone of the focused detection network of the present invention.
[0019] Figure 5 This is a schematic diagram of the bneck structure after the introduction of the collaborative attention mechanism in this invention.
[0020] Figure 6 This is the pixel space focusing module of the present invention.
[0021] Figure 7 This is a partial dataset. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] Example 1 See Figure 1 The schematic diagram shown illustrates a sparkless demolition system for coal mines based on image processing, comprising an explosion-proof imaging and pan-tilt unit for real-time acquisition of target images in the roadway and achieving line-of-sight orientation. The target identification and tracking module identifies and tracks coal and gangue. The coordinate deviation-servo angle conversion module converts the pixel deviation between the target center and the image center into gimbal horizontal / tilt angle commands in real time. The robotic arm control module drives the multi-degree-of-freedom hydraulic arm to adjust its posture. Sparkless demolition actuators are used for dust suppression, cooling, and fire source isolation when performing sparkless demolition of large coal and gangue blocks underground. The hazardous gas monitoring and safety interlock module monitors methane and carbon monoxide in real time and triggers intrinsically safe interlocks according to graded thresholds.
[0024] Furthermore, the explosion-proof imaging and gimbal assembly, used for real-time acquisition of target images in the tunnel and achieving line-of-sight orientation, includes: the explosion-proof imaging and gimbal assembly consists of a mining explosion-proof camera, a two-degree-of-freedom servo gimbal, and a gimbal controller, installed at the front end of the demolition robot as a vision base; the camera continuously acquires video frames in the tunnel and sends them to the upper-level processing unit; after initializing and selecting the demolition target, based on the target center pixel coordinates detected in each frame and the offset of the image center, the horizontal angle and pitch angle are calculated through the coordinate deviation-servo angle conversion module and sent to the gimbal for execution in real time, thereby achieving line-of-sight alignment and stable tracking. The gimbal control applies amplitude limiting and hysteresis filtering to the angle commands to suppress jitter, and maintains the previous steady-state pointing and quickly resamples until the target is relocked when the target is briefly obscured or dust suddenly increases.
[0025] Furthermore, the target recognition and tracking module identifies and tracks coal and gangue, including: the target recognition and tracking module uses a proposed focused detection network for coal and gangue identification. The focused detection network uses YOLOv4 as the overall framework and MobileNet-v3 as the backbone network, such as... Figure 4 As shown, a collaborative attention mechanism formed by the pixel space focusing module and the SE channel attention is introduced into the second, fourth, and fifth bneck modules of the focusing detection network backbone. The first and third bneck modules are composed of lightweight 1x1 and 3x3 convolutions. Figure 5 As shown, the input features of the bneck module after introducing the collaborative attention mechanism first enter a 3x3 depthwise convolution and a first batch normalization layer. Two branches are output from the first batch normalization layer: one branch enters the SE channel attention module to recalibrate the channel weights, and the other branch proceeds along the main branch and is multiplied pixel-by-pixel with the spatial weights generated by the pixel spatial focusing module. The result of the multiplication is then passed through a sigmoid function and followed by a 1x1 convolution and a second batch normalization layer to obtain the output features.
[0026] The pixel space focusing module, as shown Figure 6 As shown, the input features first enter pixel regularization, which takes the scaling factor of the first batch normalization layer, regards the scaling factor as the source of pixel importance measurement and completes normalization to obtain a steady-state normalized output; then, a pixel weight map is constructed based on these scaling factors to represent the importance of each spatial location; the pixel weight map is multiplied pixel by pixel by the first batch normalization layer, thereby selectively amplifying and suppressing spatial features: the weights of locations that are relevant to the target and have clearer information are enhanced, while the weights of locations that are irrelevant to the target are suppressed.
[0027] The pixel spatial focusing module is used to suppress spatially irrelevant regions and enhance key regions. It is based on the premise that the parameters of the first batch normalization layer can characterize spatial importance. The first batch normalization layer is located after the "channel-wise 3x3 depthwise convolution". The scaling factor of the first batch normalization layer is used to normalize the features pixel by pixel, and then the features are summarized into a pixel weight matrix on the entire feature map. The weight of each position describes the importance of the pixel in the overall image context. Subsequently, the pixel spatial focusing module performs a one-to-one spatial dot product between the normalized feature output and the pixel weight matrix. For positions with weights greater than the average level, the corresponding features are explicitly suppressed; for positions with weights less than the average level, the corresponding features are explicitly suppressed. The contour edges and main texture of coal gangue carry distinguishing clues, while coal dust and water mist are mostly weak texture, low contrast or randomly distributed. In the pixel weight matrix, the coal gangue area is given a higher weight, while the coal dust and water mist areas are given a lower weight. This systematically increases the signal proportion of the "salient area - coal gangue" and attenuates the interference proportion of the "non-salient area - coal dust and water mist" at the output feature level.
[0028] In this embodiment, the dataset used by the focused detection network consists of 18 segments of 1080p high-definition video (total duration approximately 510 minutes) collected by monitoring equipment. Keyframes containing the target are extracted from these videos and uniformly set to a resolution of 1920×1080, resulting in 981 real-world images of the mine. These images are then labeled using LabelImg (using rectangular bounding boxes) according to the Pascal VOC2012 standard. Figure 7 The image shows a portion of the dataset.
[0029] In this embodiment, the focus detection network divides the dataset into training, validation, and test sets in a ratio of 7:1:2. Evaluation is based on precision, recall, and mAP, while location accuracy is determined by an intersection-union threshold of 0.5 (IoU ≥ 0.5 is considered correct). The focus detection network uses the CIoU loss function for location regression. The focus detection network retains Mosaic as the primary enhancement method, supplemented by random scaling, horizontal flipping, random brightness / contrast, HSV jitter, and mild motion blur. Its optimizer is SGD with a momentum of 0.9 and a weight decay of 5 × 10⁻⁻⁻⁶. 4 The initial learning rate was 0.01, the batch size was 16, and the total number of epochs was 300. Table 1 below compares the performance of the focused detection network and the original YOLOv4 in detecting coal gangue: Table 1 Comparison of Coal and Gangue Detection Performance in Mining Scenarios
[0030] As shown in Table 1, for the target coal gangue, the focusing detection network outperforms YOLOv4 (also a lightweight backbone of MobileNet-v3) in all three metrics: mAP@0.5 increases from approximately 57.3% to approximately 57.5%, while the simultaneous slight improvement in precision and recall indicates that the pixel space focusing module enhances the salient regions (coal gangue outline and main texture) and suppresses the non-salient regions (coal dust and water mist) in the spatial dimension, resulting in a more stable detection curve and fewer false positives / false negatives. After the focused detection network detects coal and gangue, the target recognition and tracking module starts online tracking (low-latency tracking based on correlation filtering) and continuously outputs the target center pixel coordinates.
[0031] Furthermore, the coordinate deviation-servo angle conversion module converts the pixel deviation between the target center and the image center into gimbal horizontal / tilt angle commands in real time. This includes: the coordinate deviation-servo angle conversion module takes the target center pixel coordinates and timestamp output by the focus detection network as input, and combines the image principal point obtained from power-on calibration with camera intrinsic parameters, lens field of view, and gimbal zero position. First, it sets pixel-level overlap criteria and dead zone thresholds: when the target center and image center fall into the dead zone, it is considered aligned; if not aligned, it calculates the pixel deviation in the horizontal and vertical directions, and converts the pixel deviation into a line-of-sight deviation angle based on imaging geometry, obtaining the target angle increments for gimbal horizontal and pitch. To avoid high-frequency jitter caused by low light and dust, the coordinate deviation-servo angle conversion module... The servo angle conversion module imposes amplitude, speed, and acceleration constraints on the angle command, and superimposes a first-order low-pass filter and short-time prediction. The angle increment is then sent to the servo driver, employing closed-loop control (the angle error fed back from the encoder is adjusted via proportional-derivative or proportional-integral-derivative with anti-saturation), and hysteresis and static friction compensation to ensure rapid gimbal convergence without significant overshoot. Once the line of sight is stably aligned, the module issues a "coarse approach" command to the travel control, moving the tracked vehicle closer at a constant or segmented speed based on the target frame's scale change in the frame and the camera-gimbal attitude steady-state error. This is achieved by integrating ultrasonic / laser ranging or obstacle avoidance sensing to maintain a safe distance. In the "fine approach" phase, the module decomposes the residual line-of-sight deviation with the end effector's pose. By combining the calculation matrix (hand-eye calibration results) with the micro-increments of the robotic arm's end effector in the workpiece's normal and tangential directions, the module drives the hydraulic arm to perform fine-tuning of its posture. This ensures the end effector is centered within a predetermined spray envelope and that its angle with the target surface's normal meets a set threshold, thereby reducing the risk of secondary sparks during impact or shearing. Throughout the process, coordinate deviation and angle control logic are tightly coupled with safety interlocks: if methane or carbon monoxide concentrations exceed limits, spray pressure or flow rates fall below thresholds, or the end effector is outside the spray envelope, the module immediately freezes the servo motor and hydraulic valve group, cancels the proximity command, and triggers an alarm. If a target is lost due to short-term obstruction or strong light spots, the module activates a hold-search strategy (sweeping within a small angle range and re-identifying within a limited time); if the time limit is exceeded, it reverts to a safe approach. The module is always in standby position and records events. To improve dynamic stability and alignment accuracy, it continuously corrects minor eccentricities and zero-point drift of the camera-gimbal axis. The power-on self-test covers gimbal limits, encoder consistency, camera exposure / gain, and time synchronization. All key quantities (pixel deviation, angle commands and feedback, speed / acceleration limiting triggers, interlock status changes, and target confidence) are recorded in real time for post-event traceability and parameter tuning. Through the above integrated process of "judgment-conversion-filtering-closed-loop-linkage", the module ensures that the camera is always facing the target in complex underground environments and forms a continuous operation chain with the "coarse engagement" of the track and the "fine engagement / attitude fine adjustment" of the robotic arm. The sparkless demolition command is only released when interlock conditions such as spray and gas monitoring are met.
[0032] Furthermore, the robotic arm control module drives the multi-degree-of-freedom hydraulic arm to perform pose adjustment, including: the robotic arm control module takes the target pose requirement output from the "detection / tracking—coordinate deviation—gimbal orientation" link as input, and drives the multi-degree-of-freedom hydraulic arm to complete the continuous action of "coarse alignment—fine alignment—attitude fine adjustment": In the coarse alignment stage, the robotic arm control module, based on hand-eye calibration and the roadway safety boundary, calls the inverse kinematics solver to plan the end effector TCP from the standby position to the working transition position, allocates the displacement, velocity, and acceleration of each joint, and performs obstacle avoidance; entering the fine alignment stage, combined with the target surface normal constraint and spray envelope limitation, through small step trajectory and velocity feedforward, the end effector is gradually centered and the included angle between the end effector tool and the target surface is controlled to be lower than the set threshold; in the attitude fine adjustment stage... The system utilizes a position-velocity dual-loop control system and hydraulic pressure limiting to coordinate the control of proportional / servo valves, implementing valve dead zone compensation and backlash compensation to suppress low-speed crawling and vibration. When necessary, it superimposes "micro-motion" commands to ensure stable contact of the end effector within confined spaces. Throughout the process, it performs joint limit, speed / acceleration limiting, hydraulic pressure and temperature rise monitoring, and quickly retracts to a safe position in response to abnormalities such as overload, jamming, and excessive positional deviation. It is also tightly coupled with a safety interlock: if methane or carbon monoxide exceeds the limit, spray pressure / flow is abnormal, or the end effector leaves the spray envelope, the valve group is immediately frozen, the proximity command is canceled, and an alarm is triggered. The module continuously corrects zero-point drift and thermal drift, and adaptively corrects the trajectory by combining online estimated target depth and body attitude error, ensuring stable end effector centering and attitude control even under low light, dust, and water mist interference.
[0033] Furthermore, the sparkless demolition actuator, used for dust suppression, cooling, and ignition source isolation during sparkless demolition of large coal gangue underground, comprises: a composite-coated hydraulic breaker, a hydraulic shear, and a high-pressure spray system. The hammerhead matrix of the composite-coated hydraulic breaker adopts a lath-like martensitic structure, providing support for the tungsten carbide particles in the composite layer. Due to its high hardness, wear resistance, and impact resistance, it significantly reduces the ignition source generated by direct metal-to-metal contact during gangue crushing and sealed wall demolition, thereby achieving smooth operation. The process essentially achieves "spark-free" control; the high-pressure spray system consists of a "water tank - high-pressure pump - spray device - control unit", with a designed flow rate of not less than 20 liters per minute; high-pressure nozzles are arranged around the crushing end, atomizing water into fine droplets and forming a dense water mist envelope in the impact zone, which collides, adsorbs and condenses with coal dust and rock powder particles in the air, causing them to increase in weight and settle, thereby reducing dust concentration, inertizing combustibility and cooling the end and near-field environment, suppressing the accumulation of heat sources and secondary hazards caused by surface incandescence.
[0034] Furthermore, the hazardous gas monitoring and safety interlock module performs real-time monitoring of methane and carbon monoxide and triggers intrinsically safe interlocks according to graded thresholds. This module comprises intrinsically safe sensing units for methane and carbon monoxide, and is equipped with temperature, humidity, and pressure compensation channels, sampling and communication interfaces, an integrated audible, visual, and voice alarm, and an interlock output relay. The sensors are arranged in zones from "operating end—middle of vehicle body—exhaust side," and multi-point data is sent to the electronic control unit for noise reduction and consistency verification (moving average + median removal of anomalies) in conjunction with time synchronization. Concentration curves and rate-of-change criteria are generated in real time. When any channel reaches the warning threshold, the hazardous gas monitoring and safety interlock module first limits the speed, increases the spray, and issues voice and indicator light prompts; when the alarm / over-limit threshold is reached, the electronic control unit immediately executes the intrinsically safe interlock: cuts off the hydraulic output of the demolition execution end, freezes the movement of the robotic arm and gimbal, maintains or increases the spray flow, and the walking mechanism enters braking or slow-moving mode. At the same time, an alarm window pops up on the driving terminal and the remote host computer and records the event; the recovery logic adopts a dual condition of "hysteresis + pressure holding time": the interlock can only be released after the concentration continuously drops below the recovery threshold and meets the minimum stability time, and after manual confirmation that there is no hidden danger. The hazardous gas monitoring and safety interlock module includes sensor self-testing and fault-safe strategies (disconnection, drift exceeding limits, overdue calibration, and abnormal noise are all handled according to the worst-case scenario and trigger shutdown). It supports zero-point / span calibration and periodic verification, and automatically shields interlock outputs during verification to prevent malfunctions. All key quantities (multi-point concentration, rate of change, threshold comparison results, interlock trigger / release timestamps, and manual confirmation records) and operating environment quantities (temperature, humidity, wind speed / direction if connected) are archived for a long period to facilitate post-event traceability and parameter tuning. The hazardous gas monitoring and safety interlock module is also linked with "coordinate deviation - servo angle conversion" and "robotic arm control": once an early warning is triggered, the front end is only allowed to make minor attitude corrections and is prohibited from approaching further; if the limit is exceeded, it is forcibly withdrawn to a safe position, and the demolition circuit and quick oil change circuit are shut down. Through the closed-loop design of "multi-point sensing - threshold grading - intrinsically safe interlock - traceable recording", the intrinsically safe control of the demolition process is achieved under the condition of rapid changes in the risk of underground gas and carbon monoxide.
[0035] This invention significantly increases the weight of significant areas (coal gangue outline and main texture) and decreases the weight of insignificant areas to address interference caused by coal dust and water mist, thereby achieving more robust detection and fewer false positives / false negatives. Coordinate deviation-angle conversion combined with amplitude limiting, filtering, and closed-loop control enables the gimbal and multi-degree-of-freedom hydraulic arm to complete continuous movements of "coarse alignment—fine alignment—attitude fine adjustment," ensuring precise centering of the end effector within the safety spray envelope. The use of a composite-coated hydraulic breaker / shear reduces the risk of direct metal-to-metal ignition, and ensures a design flow rate of not less than [a certain value]. A high-pressure spray of 20 L / min forms a dense water mist envelope in the impact zone, achieving dust suppression, cooling, and an inert atmosphere, thus achieving essentially "spark-free" demolition. Multi-point monitoring of methane and carbon monoxide triggers intrinsically safe interlocks based on graded thresholds; exceeding these limits immediately cuts off the demolition output, freezes the operation, and maintains or increases the spray. Combined with a recovery logic based on hysteresis and pressure holding time, and a fail-safe strategy, this forms a complete "early warning—alarm—shutdown—recovery" protection chain. Simultaneously, real-time recording of key quantities and interlock events enhances traceability and parameter tuning efficiency. In summary, this invention achieves robust identification, precise alignment, and safe and efficient spark-free demolition in complex downhole environments.
[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
[0037] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A sparkless demolition system for coal mines based on image processing, comprising a sparkless demolition actuator and a hazardous gas monitoring and safety interlock module, characterized in that, It also includes explosion-proof imaging and pan-tilt-zoom (PTZ) components for real-time acquisition of roadway target images and line-of-sight orientation; The target identification and tracking module identifies and tracks coal and gangue. The coordinate deviation-servo angle conversion module converts the pixel deviation between the target center and the image center into gimbal horizontal / tilt angle commands in real time. The robotic arm control module drives the multi-degree-of-freedom hydraulic arm to adjust its posture. The target recognition and tracking module is based on a focus detection network. It identifies coal gangue by suppressing interference from insignificant areas of coal dust and water mist attached to the surface of the coal gangue. A collaborative attention mechanism, formed by a pixel spatial focus module and SE channel attention, is introduced into the second, fourth, and fifth bneck modules of the focus detection network backbone. The input features of the pixel spatial focus module first undergo pixel regularization, taking the scaling factor from the first batch normalization layer in the bneck module. This scaling factor is used as a measure of pixel importance and normalization is completed to obtain a steady-state normalized output. A pixel weight map is constructed based on the scaling factor to represent the importance of each spatial location. The pixel weight map is multiplied pixel by pixel by the first batch normalization layer, thereby selectively amplifying and suppressing spatial features: the weights of positions that are relevant to the target and have clearer information are enhanced, while the weights of positions that are irrelevant to the target are suppressed.
2. The image processing-based sparkless demolition system for coal mines according to claim 1, characterized in that, The pixel spatial focusing module is used to suppress spatially irrelevant regions and enhance key regions. It is based on the premise that the parameters of the first batch normalization layer can characterize spatial importance. The first batch normalization layer is located after the 3x3 depth convolution. The scaling factor of the first batch normalization layer is used to normalize the features pixel by pixel, and then the features are summarized into a pixel weight matrix on the entire feature map. The weight of each position describes the importance of the pixel in the overall image context. The pixel spatial focusing module performs a dot product between the normalized feature output and the pixel weight matrix, with each position corresponding to a specific spatial location. For positions with weights greater than the average level, the corresponding features are explicitly defined. For positions with weights below the average level, the corresponding features are explicitly suppressed.
3. The image processing-based sparkless demolition system for coal mines according to claim 1, characterized in that, The explosion-proof imaging and gimbal assembly consists of a mining explosion-proof camera, a two-degree-of-freedom servo gimbal, and a gimbal controller, and is installed at the front end of the demolition robot. The camera captures video frames and sends them to the upper processing unit. The gimbal calculates the horizontal and pitch angles based on the offset between the target center and the image center through the coordinate deviation-servo angle conversion module. It performs amplitude limiting and hysteresis filtering on the angle commands. When the target is temporarily obscured or dust suddenly increases, it maintains the previous steady-state pointing until the target is relocked.
4. The image processing-based sparkless demolition system for coal mines according to claim 1, characterized in that, The target recognition and tracking module adopts a focused detection network, which uses YOLOv4 as the framework and MobileNet-v3 as the backbone network. The second, fourth and fifth bneck modules of the backbone network introduce a collaborative attention mechanism consisting of a pixel space focusing module and SE channel attention. The first and third bneck modules are composed of 1×1 convolution and 3×3 convolution.
5. The sparkless demolition system for coal mines based on image processing according to claim 1, wherein the workflow of the coordinate deviation-servo angle conversion module is as follows: input the target center pixel coordinates and timestamp, combine the camera intrinsic parameters, lens field of view, and gimbal zero position, and set the pixel dead zone threshold; when misaligned, calculate the pixel deviation and convert it into the line-of-sight deviation angle, implement amplitude limiting, speed limiting, acceleration limiting and first-order low-pass filtering on the angle command, drive the gimbal through closed-loop control, and compensate for hysteresis and static friction.
6. The sparkless demolition system for coal mines based on image processing according to claim 1, wherein the working process of the robotic arm control module is as follows: according to the target pose requirements, the multi-degree-of-freedom hydraulic arm is driven to complete coarse alignment, fine alignment, and attitude fine adjustment, and the joint limit, hydraulic pressure and temperature rise are monitored throughout the process, and the arm is retracted to a safe position when abnormal.
7. A sparkless demolition system for coal mines based on image processing according to claim 1, wherein the sparkless demolition actuator consists of a composite-coated breaker, a hydraulic shear, and a high-pressure spray system; the hammerhead matrix of the composite-coated breaker is lath-shaped martensitic, and the composite layer contains tungsten carbide particles; the high-pressure spray system has a flow rate greater than or equal to 20 L / min, and the nozzles are arranged around the crushing end to form a dense water mist envelope in the impact zone, thereby achieving dust suppression, cooling, and fire source isolation.
8. The image processing-based sparkless demolition system for coal mines according to claim 1, wherein the hazardous gas monitoring and safety interlocking module consists of intrinsically safe methane and carbon monoxide sensing units, a temperature / humidity / pressure compensation channel, an audible and visual alarm, and an interlocking relay. The sensors are arranged in zones according to "working end - vehicle body middle - exhaust side". The data is processed by moving average + median removal to generate a concentration curve.
9. The image processing-based sparkless demolition system for coal mines according to claim 1, comprising a hazardous gas monitoring and safety interlock module, which performs real-time monitoring of methane and carbon monoxide and triggers intrinsically safe interlocks according to graded thresholds.