An intelligent sensing method for identifying violation of regulations by coal mine workers and analyzing cracks in equipment

Abstract

The present application relates to the field of mine safety monitoring, in particular to a kind of coal mine operating personnel violation identification and equipment crack analysis intelligent sensing method, first through self-cleaning dual-spectrum sensing terminal acquisition coal mine underground multi-modal original data;Visible light video stream is input into lightweight depth curve estimation network and multi-scale feature fusion network, and the restored image sequence is obtained;Violation behavior identification is carried out based on the restored image sequence extraction human key point coordinates;Equipment key stress area is locked to target area, and the authenticity of the crack of the intercepted interest region image is verified to realize equipment crack analysis;At the same time, rely on edge computing gateway to carry out local real-time inference and alarm.The present application can adapt to the complex environment of low illumination, high dust and strong interference in coal mine, improve the accuracy of operating personnel violation identification and the detection ability of equipment micro-crack, reduce false alarm and false negative, and has good real-time, stability and engineering application value.

Description

Technical Field

[0001] This invention relates to the field of smart mine safety monitoring technology, specifically to an intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks. Background Technology

[0002] Ensuring the safety of the underground coal mine production environment has always been a core aspect of the survival and development of mining enterprises. This encompasses real-time monitoring of worker behavior and continuous monitoring of the structural health of large fixed equipment. With the evolution of industrial internet and artificial intelligence technologies, utilizing visual perception technology to replace traditional manual inspections has become an important means of achieving inherent safety in mines.

[0003] Existing coal mine monitoring technologies largely rely on high-definition explosion-proof cameras deployed in roadways and working faces. These cameras transmit video signals to a surface control center, where general target detection algorithms identify simple behaviors such as wearing safety helmets or trespassing into restricted areas. For equipment inspection, traditional methods typically employ periodic manual flaw detection or infrared thermography, attempting to identify safety hazards before obvious equipment malfunctions occur.

[0004] However, the underground environment of coal mines suffers from severe low light levels, high dust levels, and non-uniform illumination, resulting in extremely low contrast and blurred edges in the acquired images. Traditional visual algorithms experience a significant drop in detection rate when processing such degraded images. The uniformity of personnel attire and the confined working space lead to frequent target adhesion, making single-frame recognition patterns highly susceptible to false alarms. For minute, nascent cracks in equipment, algorithms often struggle to capture pixel-level feature changes due to limitations in transmission bandwidth and image resolution. Furthermore, they are highly susceptible to interference from oil, water stains, or weld seams on equipment surfaces, causing frequent false alarms and severely impacting the practicality and reliability of automated monitoring systems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks. This method solves technical problems such as severe image degradation in coal mines, large errors in identifying work behavior, difficulty in detecting minute cracks in equipment, difficulty in coordinating multimodal information, and limitations in edge deployment. As a result, it improves the accuracy, real-time performance, and stability of underground personnel behavior supervision and equipment health monitoring.

[0006] To achieve the above objectives, the present invention provides a method for intelligent sensing of coal mine worker violations and equipment crack analysis, comprising the following steps: S1. Collect raw multimodal data through a self-cleaning dual-spectrum sensing terminal installed in the coal mine. The raw multimodal data includes at least visible light video stream, infrared thermal imaging data, equipment acoustic emission signals, and personnel location information. S2. Input the acquired visible light video stream into a preset lightweight depth curve estimation network and a multi-scale feature fusion network to perform low-light enhancement, defogging, dust removal and edge detail restoration processing to obtain the restored image sequence. S3. Based on the restored image sequence, extract the coordinates of key human body points, construct the human skeleton topology, input the temporal features of the skeleton into the spatiotemporal graph convolutional network and the action recognition network for joint analysis, and combine the personnel location information to perform spatial domain logical verification on the recognition results to obtain the personnel violation recognition results. S4. Target region locking is performed on the key stress area of ​​the equipment in the restored image sequence, and the region of interest image is extracted. The region of interest image is input into a super-resolution generative adversarial network to perform texture reconstruction. Then, the crack length, width, area, direction and edge morphology features are extracted through a pixel-level segmentation model. The authenticity of the crack is verified by combining the acoustic emission signal of the equipment to obtain the crack analysis results of the equipment. S5. Input the personnel violation identification results and equipment crack analysis results into the edge computing gateway to perform local inference and alarm control. At the same time, send back fuzzy samples with confidence levels falling within a preset range to the cloud server, where the cloud completes sample labeling management, incremental model training and remote model updates.

[0007] Preferably, the self-cleaning dual-spectrum sensing terminal includes an explosion-proof housing, a visible light camera component, a long-wave infrared imaging component, a lens contamination detection module, and a physical cleaning module; the physical cleaning module is used to activate an ultrasonic vibration dust removal mechanism or a mechanical wiping mechanism to remove dust, coal sludge, and water mist adhering to the lens surface when a decrease in lens transmittance or an image clarity lower than a set threshold is detected.

[0008] Preferably, the lightweight depth curve estimation network is used to perform pixel-level brightness mapping adjustment on video frames in low-light environments; the multi-scale feature fusion network is used to extract multi-level texture information and restore low-contrast areas caused by dust, water mist and scattered reflection.

[0009] Preferably, the personnel violation identification includes: detecting and tracking human targets in continuous video frames; extracting key points such as the head, shoulders, elbows, wrists, hips, knees, and ankles of the human body; generating a skeleton diagram based on the connection relationship between the key points; assembling temporal skeleton features from the skeleton diagram in chronological order; extracting joint motion patterns using a spatiotemporal graph convolutional network; and then classifying and judging behaviors such as crossing belts, entering restricted areas, staying on duty while sleeping, leaving the post, and abnormal handling by an action recognition network.

[0010] Preferably, the spatial domain logic verification includes: mapping the human body position in the video to the underground area coordinate system, and synchronizing and aligning it with the personnel position information obtained by UWB or an equivalent high-precision positioning system in time and space; when the visual recognition result shows that a person has entered a restricted area, and the positioning result shows that the corresponding person is within the coordinate range of the restricted area, a violation signal is output.

[0011] Preferably, the key stress areas of the equipment include the hydraulic support hinge, pin connection, coal mining machine rocker arm, belt conveyor roller bearing seat, transfer conveyor connection structure, the area adjacent to the weld of the support beam, and other high-load and vulnerable parts.

[0012] Preferably, the super-resolution generative adversarial network is used to restore the region of interest image to a high-resolution image to enhance the details of the microcrack edges; the pixel-level segmentation model is used to segment the crack target in the enhanced region of interest image to obtain a crack contour mask, and further calculate the crack geometric parameters.

[0013] Preferably, the crack authenticity verification includes: calculating the aspect ratio, fractal dimension, edge roughness, continuity, and orientation consistency of the segmented crack candidate regions, and eliminating pseudo-cracks formed by oil stains, watermarks, weld textures, or surface scratches; simultaneously analyzing the amplitude, frequency band energy, event count, and pulse density of the device's acoustic emission signal, and determining it as a real crack when the visual crack candidate target and the acoustic emission abnormal event occur simultaneously within a preset time window.

[0014] Preferably, the edge computing gateway deploys a quantized and compressed perception model and performs asynchronous computing power scheduling for personnel violation identification tasks and equipment crack analysis tasks; wherein, the personnel violation identification task adopts real-time processing at a high frame rate, and the equipment crack analysis task adopts periodic detection or event-triggered detection.

[0015] Preferably, the fuzzy sample is a sample whose identification confidence is within a preset threshold range, or a sample whose multimodal judgment results conflict. After receiving the fuzzy sample, the cloud server performs incremental training in combination with the manual annotation results to generate updated model parameters, and then synchronizes them to the edge computing gateway via remote distribution.

[0016] This invention provides an intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks. It has the following beneficial effects: 1. This invention utilizes a self-cleaning dual-spectrum sensing terminal and introduces a lightweight depth curve estimation network and a multi-scale feature fusion network before the image enters the recognition model to pre-repair images degraded by low illumination, dust, water mist, lens contamination, and localized strong reflections in coal mines. Because the input image is improved in terms of brightness levels, edge sharpness, and texture integrity, subsequent human skeleton extraction and crack candidate region localization have a more stable data foundation. Therefore, it can significantly improve the perception effectiveness in complex underground environments and reduce recognition failures caused by insufficient source image quality.

[0017] 2. This invention employs a combination of skeleton temporal modeling and spatial domain verification in the personnel violation identification process. It utilizes a spatiotemporal graph convolutional network to extract the continuous motion relationships of multiple joints in the human body, and then uses positioning information to cross-validate the visual recognition results. This transforms single-frame static judgment into a temporally continuous and spatially verifiable joint identification process, effectively distinguishing between normal maintenance, temporary avoidance, material handling, and actual violations. This significantly reduces false alarms and false negatives common in traditional single-vision recognition, improving the accuracy of identifying high-risk behaviors such as crossing conveyor belts, entering restricted areas, and abnormal lingering.

[0018] 3. This invention addresses the problem of early-stage microcracks being too small and having weak texture in the original monitoring images by implementing local reinforcement in key stress areas. Furthermore, by combining crack geometry features with acoustic emission events for heterogeneous identification, it effectively eliminates interference from spurious features such as oil stains, watermarks, and weld seams. The system can output reliable early warnings when cracks are still in their initial propagation stage, thereby improving the structural health monitoring capabilities and preventative maintenance levels of key coal mine equipment.

[0019] 4. This invention completes high-real-time task inference through an edge computing gateway and sends fuzzy samples with low confidence levels back to the cloud for active learning and training, forming a continuous optimization closed loop of edge-cloud collaboration. The edge-side model, after quantization and compression, can meet the deployment requirements of the constrained hardware environment underground, ensuring the immediacy of on-site alarms. The cloud is responsible for absorbing new samples, updating model parameters, and sending back optimization results, enabling the system to gradually develop targeted adaptation capabilities in different mines, roadways, lighting conditions, and equipment states. Through this online evolution mechanism, this invention can continuously improve the model's generalization performance and reduce the pressure of manual maintenance during long-term operation.

[0020] 5. This invention integrates personnel violation identification and equipment crack analysis into the same intelligent sensing process, achieving integrated supervision through shared video acquisition, image enhancement, edge inference, cloud training, and alarm linkage infrastructure. This facilitates unified management of safety incidents, maintenance work orders, and model iteration tasks by the mine dispatch center, improving the overall level of integrated safety management. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the method of the present invention. 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. Example 1

[0023] like Figure 1 As shown, this embodiment of the invention provides an intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks, applicable to fully mechanized mining faces, transport roadways, transfer points, electromechanical chambers, pump rooms, and other key underground areas.

[0024] Several self-cleaning dual-spectrum sensing terminals are deployed within the underground monitoring area. Each terminal is electrically or network-connected to an edge computing gateway. The sensing terminals acquire visible light video streams and infrared thermal video streams. The edge computing gateway simultaneously connects to the acoustic emission sensor, UWB positioning base station, and intrinsically safe switching network corresponding to the monitoring area. All data sources achieve time synchronization through a unified timestamp mechanism, with the synchronization accuracy preferably controlled at the millisecond level, at least meeting the correlation requirements between personnel trajectories and video frames, and between acoustic emission events and crack candidate frames.

[0025] After receiving the raw video stream, the edge computing gateway performs an image quality assessment. This assessment can use a combination of indicators such as sharpness, contrast, brightness uniformity, and haze. If the sharpness score is below a preset threshold, it indicates the presence of dust, water vapor, or lens contamination, and the system controls the sensing terminal to initiate a self-cleaning process. This self-cleaning process can be an ultrasonic vibration dust removal procedure or a mechanical brush reciprocating wiping procedure. After physical cleaning, image acquisition continues, and the images are then sent to the low-light enhancement and defogging / dust removal processing link.

[0026] The low-light enhancement link employs a lightweight depth curve estimation network. This network outputs a set of pixel-level or region-level brightness adjustment curves for each frame of image, performing non-linear enhancement on dark areas and suppressing bright areas to avoid overexposure and noise amplification caused by overall brightening. To meet the real-time processing requirements of downhole edge equipment, this network preferably uses a shallow convolutional structure and a lightweight activation function. The defogging and dust removal link employs a multi-scale feature fusion network, extracting spatial features at different scales in parallel to reconstruct low-contrast areas caused by suspended particles, restoring human contours and metal edges. After this processing, a restored image sequence is obtained, serving as the common input for subsequent violation recognition and crack analysis.

[0027] Example 2: This example provides a detailed explanation of the process for identifying personnel violations. In the restored image sequence, the edge computing gateway first uses a human detection model to obtain the human target bounding box in each frame, and then uses a multi-target tracking algorithm to associate the target identity of the same person in consecutive frames to form a continuous trajectory. Considering the problems of narrow underground space, similar workers' clothing, and overlap of personnel and equipment, the multi-target tracking algorithm preferably adopts a combination of motion prediction and appearance feature matching to improve the trajectory continuity under occlusion conditions.

[0028] Within the target human body area, the system invokes a human pose estimation model to extract key point coordinates. These key points include at least the head, neck, left and right shoulders, left and right elbows, left and right wrists, trunk center, left and right hips, left and right knees, and left and right ankles. The key points are arranged according to the natural topological relationships of the human body to form a skeleton diagram. Several consecutive frames of skeleton diagrams are then stitched together in chronological order to form the temporal features of the skeleton.

[0029] Next, the temporal features of the skeleton are input into a spatiotemporal graph convolutional network. This network analyzes the relative positional relationships between joints in the spatial dimension and the joint motion trajectories and velocity changes in the temporal dimension, extracting deep features related to the action category. To further enhance the ability to simultaneously model fast and slow movements, an action recognition network can be superimposed after the spatiotemporal graph convolutional network. This network can employ a two-branch structure, where the fast branch focuses on capturing short, intense movements such as climbing, jumping, and falling, while the slow branch focuses on recognizing continuous movements such as prolonged standing still, sleeping on duty, and prolonged deviation from one's workstation.

[0030] To avoid misjudgments caused by relying solely on visual judgment, the system simultaneously reads location information uploaded by UWB positioning tags. The edge computing gateway pre-stores spatial maps and electronic fence boundaries of the monitoring area, maps the video image coordinate system to the underground spatial coordinate system through calibration parameters, and then aligns the visually tracked target with the UWB personnel tag in time and associates them with their identities. If the action recognition result shows that the target has the action characteristics of crossing the conveyor belt, and its spatial trajectory crosses the preset danger zone boundary, it is judged as a violation of regulations. If the action recognition result shows that the target is in a non-working static posture and the stay time exceeds the preset threshold, and the positioning information shows that it has been staying in an area where it should not stay for a long time, it is judged as sleeping on duty or absenteeism. If the failure to wear the prescribed equipment is detected, the compliance of the protective equipment can be further judged by combining the head area and the wearing recognition model.

[0031] After outputting the violation conclusion, the edge computing gateway executes local linkage control. Linkage control includes, but is not limited to: sending warning messages to on-site voice broadcasting equipment; pushing captured images and short video clips to the dispatch platform; and sending stop or deceleration commands to the PLC or other control systems at a preset severity level. This linkage strategy can be configured according to the mine's safety management regulations.

[0032] Example 3: This example provides a detailed explanation of the equipment crack analysis process. To improve the efficiency of equipment crack detection, the system does not perform full-pixel high-precision analysis of the entire image. Instead, it first identifies key stress areas using a target area locking method. Specifically, the edge computing gateway uses a lightweight detection model to identify the main body of the equipment and its key components. Based on a pre-set structural template or equipment digital model, it locates high-risk areas, such as the connection points of hydraulic support columns, the hinged ends of balancing jacks, the welding transition zone of the coal mining machine rocker arm, the periphery of the belt roller bearing seat, and the transition connection plate of the transfer machine. For these areas, the system extracts regions of interest images to reduce irrelevant background interference and concentrate computing resources.

[0033] A super-resolution generative adversarial network (GAN) is used as input to the region of interest (ROI) image. This network utilizes a generator to reconstruct high-frequency details of the image, while a discriminator constrains the realism of the reconstructed result, thus highlighting the edges of micro-cracks without altering the original structural relationships. Through super-resolution reconstruction, tiny cracks that were originally only a few pixels in the surveillance image and difficult to distinguish can be given a clearer texture representation.

[0034] After super-resolution reconstruction, the high-resolution region of interest image is fed into a pixel-level segmentation model. This model employs a network structure with multi-scale feature extraction and global context modeling capabilities to classify crack targets pixel-by-pixel, outputting a binary crack mask. Based on the mask results, the system further calculates the crack length, average width, maximum width, area, orientation angle, curvature, and edge irregularity. For suspected crack targets, the system performs morphological filtering. For example, if a candidate target is large, blocky, with diffused edges and a low aspect ratio, it can be identified as oil or water stains; if the candidate target has a stable orientation and completely coincides with the orientation of existing weld seams on the equipment, weld seam interference is further excluded based on prior knowledge of the equipment structure; if the candidate target is slender, continuous, and its edge roughness matches crack growth characteristics, it proceeds to the next step of physical verification.

[0035] Physical verification employs acoustic emission signal analysis. Acoustic emission sensors are installed near critical areas of the equipment to collect high-frequency elastic waves generated by the release of internal stress in the metal in real time. An edge computing gateway preprocesses the acoustic emission signals, extracting event amplitude, number of impacts, ring count, energy characteristics, and frequency distribution. If a visual crack candidate area and an abnormal acoustic emission event occur simultaneously within a preset time window, and the acoustic emission characteristics meet preset crack propagation conditions, it is determined to be a real crack. If a candidate area is detected visually but the acoustic emission signal remains stable, the candidate target's risk level is reduced, or it is temporarily recorded as a target awaiting review. This image-physical signal coupling method significantly reduces false alarms caused by pseudo-cracks.

[0036] Once a crack is confirmed, the system generates a structured record containing the crack location, geometric parameters, risk level, equipment number, and timestamp, and pushes it to the cloud platform. The cloud platform can then create a crack evolution archive based on the historical records of the same equipment, used to analyze whether the crack is showing a continuous propagation trend.

[0037] Example 4: This example provides a detailed explanation of model updates and edge-cloud collaborative processing. Edge computing gateways handle real-time inference tasks. To adapt to the intrinsically safe, explosion-proof hardware environment underground, the recognition models deployed at the edge undergo model compression before going online. Compression methods can include channel pruning, parameter sparsity, quantization, and knowledge distillation. A teacher model is trained with high precision in the cloud, and then the knowledge from the teacher model is transferred to the student model. The student model is quantized into INT8 format and deployed on the edge gateway to balance computational speed and recognition accuracy.

[0038] Asynchronous scheduling is implemented at the edge based on task type. Personnel violation identification is a high-time-sensitivity task, typically running continuously at a high frame rate, ideally achieving near real-time processing. Equipment crack analysis is a state evolution task, which can run according to a set cycle, or trigger a high-frequency analysis mode when abnormal vibration, acoustic emission, or temperature is detected. Through task classification and resource allocation, the edge gateway is prevented from performing too many heavy-load calculations at the same time, thus avoiding impacting the on-site alarm response.

[0039] The system calculates a confidence score for each recognition result. If the confidence score of a recognition result falls within a preset fuzzy range, or if there is a conflict between visual judgment and positioning judgment, or between visual judgment and acoustic emission judgment, a fuzzy sample package is automatically generated. The fuzzy sample package includes at least the original image fragment, the enhanced image fragment, the recognition label, the confidence score, relevant positioning data, relevant acoustic emission data, and the corresponding equipment or personnel identification. This sample package is transmitted back to the cloud server via the mine's dedicated network.

[0040] The cloud server categorizes, filters, and labels ambiguous samples. Manual labelers can view suspected violations or crack candidate areas on the cloud platform and provide corrective labels. Subsequently, the cloud training module adds newly labeled samples to the training set, incrementally training both the violation identification model and the crack analysis model. To avoid catastrophic forgetting, historical typical samples are preferentially retained during training, and a hybrid training strategy is employed. After training, a new version of the model parameters is generated and pushed to each edge computing gateway via OTA during off-peak hours in the mine. After completing integrity verification, the edge gateway switches to the new model, thus forming a closed loop of continuous evolution. Example 5

[0041] This embodiment uses the tail area of ​​the conveyor belt in a fully mechanized mining face as an example to illustrate the practical application of the present invention.

[0042] A self-cleaning dual-spectrum sensing terminal was installed above the tail section of the conveyor belt, and a set of acoustic emission sensors was installed near the tail roller bearing housing. UWB positioning base stations were deployed in this area, and workers wore positioning tags. During daily operation, the sensing terminal continuously collected images from the site and uploaded them to the edge computing gateway. At a certain moment, the dust concentration suddenly increased, causing a decrease in image clarity. Based on the image quality evaluation results, the edge computing gateway controlled the sensing terminal to initiate an ultrasonic dust removal program and simultaneously performed low-light enhancement and defogging processing on the image.

[0043] Subsequently, the system detected a person approaching the danger zone at the end of the conveyor belt in the restored footage and extracted key points from multiple consecutive frames of the person's body. Through spatiotemporal graph convolutional network analysis, it was found that the person made a clear leg-lifting movement to cross over; simultaneously, their UWB trajectory crossed the electronic fence of the conveyor belt operating area. Based on this, the system confirmed the illegal crossing, immediately issued a warning message via on-site broadcast, and uploaded the corresponding short video clip to the dispatch center.

[0044] Within the same time period, the system performs crack analysis on the bearing housing area of ​​the belt conveyor's tail roller according to a periodic strategy. The edge computing gateway extracts the region of interest at the bearing housing connection from the video footage, performs super-resolution reconstruction, and the segmentation model identifies a thin, linear crack candidate. This candidate still meets crack characteristics after morphological filtering; the system further compares the data with acoustic emission sensor data and discovers an abnormal high-frequency pulse event within the same time window, thus confirming the presence of an early-stage crack at that location. The platform marks this crack as an early warning and automatically generates maintenance suggestions, notifying maintenance personnel for review and handling.

[0045] As can be seen from the above implementation methods, this invention forms a complete and practical intelligent sensing method for coal mines, encompassing raw data acquisition from underground, degradation image restoration, temporal recognition of personnel actions, detailed analysis of equipment cracks, multimodal cross-validation, and continuous optimization through edge-cloud collaboration. This method provides stable solutions to typical application challenges in the complex environment of coal mines, demonstrating clear engineering feasibility and significant industrial application value.

[0046] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent sensing of coal mine worker violations and equipment crack analysis, characterized in that, Includes the following steps: By deploying self-cleaning dual-spectrum sensing terminals at coal mine sites, raw multimodal datasets containing visible light video streams, infrared thermal imaging data, equipment acoustic emission signals, and personnel location information can be acquired in real time. The preset lightweight depth curve estimation network and multi-scale feature fusion network are invoked to perform synchronous brightness compensation and noise reduction on the visible light video stream, generating a contrast-enhanced restored image sequence. The coordinates of key human body points are extracted from the restored image sequence, a skeleton topology is constructed, the evolution trend of the action is analyzed using a spatiotemporal graph convolutional network, and spatial domain logic verification is performed in conjunction with the personnel location information to determine the personnel's violation of regulations. The target region is locked on the restored image sequence. The captured local stress area image is input into a super-resolution generative adversarial network for texture reconstruction. Then, the geometric features of the crack are extracted through a pixel-level segmentation model, and the fault verification is performed in combination with the energy level features of the acoustic emission signal of the device. A feedback loop for edge-cloud collaboration is constructed, which utilizes the edge computing gateway to perform real-time perception inference and transmits fuzzy samples with recognition confidence within a preset threshold range back to the cloud server for active learning and training, thereby realizing the dynamic updating of the perception model.

2. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, The self-cleaning dual-spectrum sensing terminal includes an explosion-proof housing, a visible light sensor, a long-wave infrared sensor, and a physical cleaning component. The physical cleaning component is configured to automatically perform ultrasonic dust removal or mechanical wiping operations based on the image clarity assessment results.

3. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, The specific process for generating the restored image sequence is as follows: using a zero-reference depth curve estimation network to perform pixel-level brightness adjustment on the dark frames in the visible light video stream, and at the same time, using the multi-scale feature fusion network to filter out scattering noise caused by dust and water mist, thereby restoring the edge details of the target.

4. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, The specific process for determining personnel violations includes: The skeleton topology of multiple consecutive frames is used as human skeleton features and input into a dual-branch action recognition network, where the fast branch is used to capture high-frequency limb movements and the slow branch is used to parse overall posture features. Calculate the relative spatial relationship between the trajectory of key human body points and the preset electronic fence boundary; When the action characteristics are determined to match the violation template and the person's real-time coordinates are within the restricted area, an early warning command is triggered.

5. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, The process of performing target region locking on the restored image sequence specifically includes: identifying key stress points of the device through a lightweight detection network and defining them as regions of interest, and cropping the regions of interest as the local stress area image.

6. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 5, characterized in that, When the captured image of the local stressed area is input into a super-resolution generative adversarial network for texture reconstruction, the resolution of the image of the local stressed area is increased to a preset multiple by the super-resolution generative adversarial network to enhance the visual features of the closed crack.

7. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, After extracting the geometric features of the crack, a morphological filtering step is also included: calculating the length-to-width ratio and fractal dimension of the segmented target, and filtering out non-metallic damage patterns whose geometric features do not conform to the linear evolution law.

8. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, The specific logic for performing fault verification based on the energy level characteristics of the acoustic emission signal of the device is as follows: when the visual algorithm detects a suspected crack based on the crack geometry, and a high-frequency elastic wave pulse exceeding a preset amplitude threshold appears in the synchronously acquired acoustic emission signal of the device, it is confirmed that the device has substantial crack damage.

9. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, The edge computing gateway uses the INT8 quantization algorithm to compress the perception model and asynchronously allocates computing resources to the personnel violation identification task and the equipment crack analysis task according to the real-time requirements of the tasks to be processed.

10. The intelligent sensing method for identifying violations by coal mine workers and analyzing equipment cracks according to claim 1, characterized in that, The specific process of active learning training includes: the cloud server obtains the returned fuzzy samples and retrieves the manual annotation results, stores the annotated new samples into the incremental dataset, generates optimized model weights by fine-tuning the fully connected layer parameters of the perception model, and remotely sends the optimized model weights to the edge computing gateway.

Images