Convolutional neural network-based normative monitoring algorithm, equipment and system before tunneling operation

Abstract

The invention discloses a normalization monitoring algorithm, equipment and system before tunneling operation based on a convolutional neural network. According to the invention, the video stream is continuously monitored through the intrinsic safety type camera, and the monitoring process is automatically started when the camera is recovered from the offline state to the online state. An improved YOLOv8 detection network is adopted, a PP-HGNetV2 structure is introduced into a backbone network to reduce the calculation amount, and a CBAM attention module is added to a feature fusion layer to improve the recognition precision. The system performs preprocessing such as illumination compensation, histogram equalization and denoising on the acquired image, judges the relevance between the gas detector and the personnel through a target matching algorithm, identifies the state of the gas detector by using an RGB value range, confirms the same personnel through intersection-to-union ratio calculation and detects the wearing normalization of the protective articles. The system realizes automatic intelligent monitoring, the detection accuracy exceeds 90%, the response time is controlled within 3 seconds, and the system has excellent environmental adaptability and provides reliable technical support for safe production of mines.

Description

Technical Field

[0001] This invention relates to the field of mine safety monitoring technology, and in particular to a standardized monitoring algorithm, equipment and system based on convolutional neural networks before tunneling operations. Background Technology

[0002] Most mines in my country are characterized by complex geological conditions, high mining difficulty, diverse types of disasters, and wide distribution. In recent years, under the guidance of national governments at all levels and relevant policies, the total number of mine safety accidents has shown a downward trend, and the overall safety situation is generally stable and improving. However, during tunnel excavation, especially after each blasting operation, the high concentration of dust, residual methane and other toxic gases, and the impact of vibration on the support structure create a compounding risk. If not properly managed, accidents can easily occur once personnel enter the work area. Therefore, standardized monitoring before tunneling operations has become a crucial link in ensuring safe mine production.

[0003] Currently, standardized monitoring before tunneling operations mainly relies on on-site personnel manually inspecting according to procedures, filling out checklists, and manually scheduling video monitoring. These traditional monitoring methods require workers to undergo gas testing before entering the tunneling area to confirm that the gas concentration meets safety standards. They also require workers to correctly wear safety helmets, safety harnesses, standard protective gear, and other protective equipment. During monitoring, the on-site safety officer needs to check the implementation of each safety measure one by one and record the inspection results by filling out a safety checklist. Meanwhile, personnel in the dispatch room need to monitor the situation on-site in real time through a video monitoring system.

[0004] Traditional pre-tunneling monitoring methods have significant shortcomings. Manual inspections are labor-intensive, with limited monitoring frequency and coverage, making it difficult to promptly identify problems at a particular working face. Manual recording is prone to omissions, delays, or inaccuracies, and lacks objective, traceable video data. Video surveillance requires real-time monitoring by operators, whose attention is easily distracted, and it cannot automatically identify whether personnel have undergone gas detection or are wearing protective equipment correctly. These problems lead to inadequate pre-operation checks and difficulty in timely detection and handling of safety risks. There is an urgent need for an automated and intelligent pre-tunneling monitoring method that can automatically initiate before tunneling operations, remotely collect real-time video information from each working face, and automatically determine the gas detection process and whether personnel are wearing protective equipment correctly using convolutional neural networks. This would provide timely warnings, reduce manual monitoring costs, and improve safety management efficiency. Summary of the Invention

[0005] The object of the present invention is to solve the technical problems that the existing standardized monitoring method before tunneling operations relies on manual inspections, the monitoring process is cumbersome and vulnerable to subjective factors, and the traditional monitoring method cannot achieve automatic and accurate determination of the usage of on-site gas detectors and the wearing of personal protective equipment.

[0006] To achieve the above object, the present invention provides a standardized monitoring algorithm for tunneling operations based on a convolutional neural network. The algorithm includes the following steps: Step 1: Continuously monitor the video stream status through an intrinsically safe camera. When no valid video frames can be obtained for 30 consecutive seconds, it is determined that the camera is in an offline state. When the subsequent video stream resumes and clear video frames are obtained for 10 consecutive seconds, it is determined that the blasting operation is completed and the monitoring process is started; Step 2: Preprocess the collected images, including performing illumination compensation using the Retinex algorithm, histogram equalization using the contrast-limited adaptive histogram equalization algorithm, denoising using the bilateral filter, and geometric correction using perspective transformation; Step 3: Use a convolutional neural network model based on the YOLOv8 detection network for object detection. The backbone network of the model adopts the PP-HGNetV2 structure, and a CBAM attention module is added to the feature fusion layer; Step 4: Start timing for 3 minutes after a person is detected entering the monitoring area. Detect the status of the gas detector, and judge whether the gas detector is within the operation range of the person by calculating the positional relationship between the center point coordinates of the gas detector detection frame and the person detection frame; Step 5: When the gas detector is successfully matched, judge the status of the gas detector through the RGB value range. The red light status corresponds to R>200 and G<100 and B<100, and the green light status corresponds to R<100 and G>200 and B<100; Step 6: After the gas detection is qualified, confirm the same person by calculating the intersection over union of the person targets in adjacent frames. When the IoU is greater than 0.5, it is considered the same person, and detect the wearing conditions of three types of protective equipment for the person, namely safety helmets, safety back clips, and work clothes.

[0007] Further, the matching judgment method between the gas detector and the person in Step 4 is as follows: The coordinates of the gas detector detection frame are (x1, y1, x2, y2), and the coordinates of the person detection frame are (x3, y3, x4, y4). Calculate the center point of the gas detector target Xcenter=(x1 + x2) / 2, Ycenter=(y1 + y2) / 2. When x3<Xcenter<x4 and y3<Ycenter<y4 are satisfied, it is determined that the match is successful.

[0008] Furthermore, in step 6, the method for calculating the intersection-union ratio (IoU) of personnel targets in adjacent frames is as follows: calculate the intersection area Areaint of the overlapping regions and the union area Areaunion of their respective detection boxes, IoU = Areaint / Areaunion.

[0009] Furthermore, the CBAM attention module includes a channel attention submodule and a spatial attention submodule. Channel attention obtains channel weights through global average pooling and global max pooling, while spatial attention calculates spatial weights through a 7×7 convolution kernel.

[0010] The present invention also provides a pre-tunneling operation standardization monitoring system for implementing the above algorithm, the system comprising: Intrinsically safe cameras are installed at a height of 2.5-3 meters above the ground at the top of the tunneling face, supporting a resolution of 1920×1080, and equipped with a dustproof transparent cover and a three-axis anti-vibration bracket; An edge inference device, an industrial host equipped with a GPU computing unit, configured with no less than 8GB of memory and 256GB of solid-state storage, is used to run the convolutional neural network model of claim 1; The voice broadcast unit uses an explosion-proof speaker with a power of no less than 15W and is connected to the edge inference device via an audio signal cable; The visualization terminal uses a 15-inch industrial display screen with a resolution of 1920×1080 to display monitoring images in real time. The network communication module includes an underground industrial Ethernet switch that connects cameras and edge inference devices.

[0011] Furthermore, the installation location of intrinsically safe cameras should ensure that the monitoring coverage area is not less than 5×8 meters, and that the display size of the operator in the image is not less than 50×50 pixels, and the display size of the gas detector is not less than 30×30 pixels.

[0012] Furthermore, the system also includes an alarm module, which adopts a multi-level alarm mechanism. The first-level alarm triggers a flashing yellow warning light when gas detection is not performed, the second-level alarm triggers a solid red warning light when gas detection fails, and the third-level alarm triggers a flashing yellow warning light when protective equipment is not worn correctly.

[0013] The present invention further provides a pre-tunneling standardization monitoring device, the device comprising: The image acquisition unit includes an intrinsically safe camera and an image preprocessing module; The data processing unit includes an edge inference device equipped with a convolutional neural network model. The model adopts the YOLOv8 detection network architecture, the backbone network is the PP-HGNetV2 structure, and the feature fusion layer adds the CBAM attention module. The result output unit includes a detection result plotting module and an alarm output module; The equipment automatically monitors gas detection and protective equipment wearing before tunneling operations by executing the algorithm of claim 1.

[0014] Furthermore, the training of the convolutional neural network model uses a stochastic gradient descent optimizer with an initial learning rate of 0.01, a momentum parameter of 0.937, a weight decay coefficient of 0.0005, a training batch size of 16, and a cosine annealing learning rate scheduling strategy.

[0015] The beneficial effects of this invention are: 1. This invention employs intelligent recognition technology based on convolutional neural networks. It continuously acquires video streams using an intrinsically safe underground camera, combined with real-time AI inference, completely replacing traditional manual inspection and monitoring methods. The system can accurately classify and identify the status of gas detectors, distinguishing between red, green, and off-light states. Simultaneously, it performs high-precision detection and positioning of personnel's protective equipment, such as safety helmets, safety harnesses, and work clothes. Through training and optimization of a lightweight YOLOv8 detection model, the accuracy rate for detecting various targets exceeds 90%, effectively avoiding subjective misjudgments and omissions in manual inspections. Compared to traditional monitoring methods, the system achieves 24 / 7 uninterrupted monitoring, increasing monitoring frequency to second-level response, and achieving 100% spatial coverage, significantly reducing labor costs and greatly improving the comprehensiveness and timeliness of safety monitoring.

[0016] 2. This invention constructs a multi-layered intelligent early warning system. When the system detects critical safety hazards such as "gas detection not performed," "gas detection failure," or "improper wearing of protective equipment," it immediately triggers corresponding levels of visual and voice alarms. Explosion-proof speakers provide on-site voice prompts, helping workers understand the safety situation immediately and correct any improper operations. The system employs a tiered alarm strategy, automatically adjusting alarm intensity and notification scope based on risk level to ensure that important safety events are quickly communicated to relevant management personnel. The entire detection and alarm process, from target identification to early warning output, has a response time controlled within 3 seconds. Compared to traditional manual discovery and reporting processes, this represents a several-fold increase in early warning response speed, greatly enhancing the effectiveness and relevance of on-site safety management.

[0017] 3. This invention features specialized technical optimizations tailored to the unique characteristics of underground working environments. Advanced image preprocessing techniques, such as Retinex multi-scale illumination compensation, contrast-limited adaptive histogram equalization, bilateral filtering denoising, and perspective transformation geometric correction, significantly enhance the system's detection stability under harsh conditions including high dust levels, low light, and lens occlusion. Data augmentation techniques are combined to expand sample diversity, and a CBAM attention mechanism is introduced into the convolutional neural network to improve focus on key targets, enabling the model to maintain high-precision recognition capabilities even under complex background interference. The system employs intrinsically safe explosion-proof equipment and industrial-grade hardware configuration, possessing excellent vibration resistance, dust resistance, and high-temperature resistance, ensuring long-term stable operation in the harsh working environment of mines and providing reliable technical support for safe mine production. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall system workflow according to an embodiment of the present invention; Figure 2 This is an example of camera installation according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the original image and the model detection results (gas detection failure) of an embodiment of the present invention; Figure 4 This is a schematic diagram of the original image and the model detection results (gas qualified, wearing standard) of an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] It should be noted that the use of terms such as "an embodiment," "an embodiment," "an exemplary embodiment," and "some embodiments" in the specification indicates that the described embodiment may include a specific feature, structure, or characteristic, but not every embodiment necessarily includes that specific feature, structure, or characteristic. Furthermore, when a specific feature, structure, or characteristic is described in connection with an embodiment, implementing such a feature, structure, or characteristic in conjunction with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the art.

[0022] Generally, terms can be understood at least partly from their use in context. For example, depending at least partly on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in a singular sense, or a combination of features, structures, or characteristics in a plural sense. Additionally, the term "based on" can be understood not necessarily to convey an exclusive set of factors, but rather, alternatively, depending at least partly on the context, to allow for the presence of other factors that are not necessarily explicitly described.

[0023] like Figure 1 As shown, the pre-tunneling standardization monitoring system based on convolutional neural networks provided by this invention mainly comprises three core components: a hardware deployment module, a data acquisition and model training module, and an online model detection and alarm output module. The entire system's workflow begins with offline camera detection, determines the completion of blasting through video stream recovery, automatically initiates the standardization monitoring process, sequentially completes gas detection and judgment, protective equipment wearing detection, and finally outputs the detection results and alarm information.

[0024] Example 1: System Hardware Deployment and Startup Signal Acquisition like Figure 2 As shown, intrinsically safe cameras are installed at a height of 2.5-3 meters above the ground at the top of the tunneling face. The cameras are high-definition network cameras supporting 1920×1080 resolution, equipped with explosion-proof certification, and suitable for underground working environments. The camera installation position ensures that the monitored area covers a critical working area of ​​no less than 5×8 meters, guaranteeing that the display size of personnel in the image is no less than 50×50 pixels, and the display size of the gas detector is no less than 30×30 pixels. The camera lens is equipped with a dustproof transparent cover and a three-axis vibration-damping bracket to ensure image clarity and stability in high-dust environments.

[0025] The camera is connected to the edge inference device via an underground industrial Ethernet switch. The edge inference device uses an industrial host equipped with a GPU computing unit, and is configured with no less than 8GB of RAM and 256GB of solid-state storage. The system continuously monitors the video stream status. If no valid video frames are acquired for 30 consecutive seconds, the system automatically determines that the camera is offline, indicating that blasting operations are in progress. When the video stream resumes and clear video frames are acquired for 10 consecutive seconds, the system determines that the blasting operation is complete and automatically starts the standard monitoring procedures before tunneling operations.

[0026] The voice broadcast unit uses an explosion-proof speaker with a power of no less than 15W, supporting clear voice broadcasting. It is connected to the edge inference device via an audio signal cable. The visualization terminal uses a 15-inch industrial display screen with a resolution of 1920×1080, supporting touch operation for real-time display of monitoring images and historical record queries.

[0027] Example 2: Data Acquisition, Preprocessing, and Labeling The data acquisition phase focuses on two aspects: normal operation data and abnormal detection data. Normal data collection includes images of gas detectors in normal working condition and personnel wearing standard protective equipment, with no fewer than 5000 images in each category. Abnormal data collection includes images of gas detectors in malfunction condition and personnel wearing protective equipment improperly, ensuring that the ratio of abnormal samples to normal samples is between 1:2 and 2:1. All images are uniformly adjusted to 1280×720 pixels to ensure that personnel and gas detectors are clearly visible in the images, without motion blur or severe obstruction.

[0028] Image preprocessing employs the following specific steps: Illumination compensation utilizes the Retinex algorithm for multi-scale enhancement, estimating and compensating illumination using three scales with Gaussian kernel radii of 15, 80, and 200; Histogram equalization employs a contrast-limited adaptive histogram equalization algorithm, with a cropping limit parameter set to 2.0 and a grid size of 8×8; Denoising utilizes a bilateral filter, with a spatial kernel size of 9 and a color space kernel size of 75; Geometric correction corrects lens distortion through perspective transformation, and the camera intrinsic parameter matrix is ​​obtained using a checkerboard calibration method.

[0029] Data augmentation operations include random rotation angles ranging from -15° to 15°, a horizontal flip probability of 0.5, color jittering including brightness adjustment range of 0.8-1.2, contrast adjustment range of 0.8-1.2, saturation adjustment range of 0.8-1.2, and random cropping size range of 0.8-1.0 times the original image.

[0030] The labeling was performed using the LabelImg tool. The labeling categories included: gas detector status was categorized into three types: "detector_off" (no light), "detector_red" (red light), and "detector_green" (green light); personnel targets were categorized as "person"; and personal protective equipment (PPE) was categorized into three types: "helmet," "backclip," and "workwear." All labeled data was randomly divided into training, validation, and test sets at a ratio of 80%, 10%, and 10%, and then converted to both COCO and VOC formats for simultaneous storage.

[0031] Example 3: Training and Inference Verification of Convolutional Neural Network Model The model architecture is an improvement on the YOLOv8n detection network. The backbone network adopts the PP-HGNetV2 structure instead of the original CSPDarknet53, reducing the number of model parameters to 60% of the original. A CBAM attention module is added after the C2f module of the feature fusion layer. The CBAM module includes two sub-modules: channel attention and spatial attention. Channel attention obtains channel weights through global average pooling and global max pooling, while spatial attention calculates spatial weights through 7×7 convolutional kernels. Finally, the two attention weights are multiplied to obtain the enhanced features.

[0032] The model training employs a unified optimization strategy, using stochastic gradient descent as the optimizer. The initial learning rate is set to 0.01, the momentum parameter to 0.937, and the weight decay coefficient to 0.0005. The training batch size is set to 16, and the total number of training epochs is 300. A cosine annealing strategy is used for learning rate scheduling, with linear warm-up in the first three epochs, gradually increasing the warm-up learning rate from 0.0001 to 0.01. Data loading is processed in parallel using four threads, and mixed-precision training is employed to reduce memory usage and accelerate the training process.

[0033] The inference validation phase evaluates model performance on the validation set, with key metrics including mean precision (mAP50), accuracy, recall, and inference speed. After validation, the model achieved an mAP50 of 92.3% across all categories, with an accuracy of 94.6% for gas detector status recognition, 95.2% for personnel detection, and 91.8% for protective equipment detection. The single-frame inference time is 23 milliseconds, meeting real-time detection requirements. The model weight file is exported in ONNX format and deployed to edge inference devices for online inference.

[0034] Example 4: Determination of Personnel Entry and Gas Detection Process The system starts a 20 - minute countdown monitoring after the camera status changes from offline to online, continuously detecting whether there are people entering the monitored area. For people detection, an improved YOLOv8 model is used for real - time inference. When the detected target category is "person" and the confidence level is greater than 0.5, it is determined that a person has entered. The system records the timestamp of the first detection of a person, which is used as the start time of the gas detection process.

[0035] The gas detection process is immediately started after a person enters, and a 3 - minute detection window period is set. The system simultaneously detects the person target and the gas detector target, and judges whether the gas detector is within the operation range of the person through the target box center point position matching algorithm. The specific calculation method is as follows: The coordinates of the gas detector detection box are recorded as (x1, y1, x2, y2), and the coordinates of the person detection box are recorded as (x3, y3, x4, y4). Calculate the center point of the gas detector target Xcenter=(x1 + x2) / 2, Ycenter=(y1 + y2) / 2. When the conditions x3 < Xcenter < x4 and y3 < Ycenter < y4 are met, it is determined that the gas detector and the person are successfully matched.

[0036] After successful matching, the status of the gas detector is judged. The RGB values of the display area of the detector are analyzed through a color recognition algorithm. The RGB value range corresponding to the red - light state is R > 200 and G < 100 and B < 100. The RGB value range corresponding to the green - light state is R < 100 and G > 200 and B < 100. The state of no light is that the values of all three RGB channels are less than 150. When it is recognized as the red - light state, the system immediately triggers the "gas detection unqualified" alarm and issues a voice prompt of "Harmful gas detected, entry to the working surface is prohibited" through the voice broadcast unit. When it is recognized as the green - light state, the system broadcasts "Gas detection qualified, please check the wearing of protective equipment" and enters the next detection link.

[0037] Example 5: Detection of the specification of wearing protective equipment After the gas detection is qualified, the system automatically enters the stage of detecting the wearing of protective equipment. To ensure a comprehensive detection of the wearing situation of personnel, the system requires the person to stay facing the camera for 3 seconds first, and then turn around and stay with the back facing the camera for 3 seconds. The system confirms that it is the same person being detected by calculating the intersection over union (IOU) of the person target in adjacent frames.

[0038] The specific implementation of IOU calculation is as follows: Assume the coordinates of the person target detection box in the previous frame are (xA1, yA1, xB1, yB1), and the coordinates of the person target detection box in the current frame are (xA2, yA2, xB2, yB2). First, calculate the intersection area. The coordinates of the top left corner of the overlapping area are xint1 = max(xA1, xB1) and yint1 = max(yA1, yB1). The coordinates of the bottom right corner of the overlapping area are xint2 = min(xA2, xB2) and yint2 = min(yA2, yB2). The width of the overlapping area is wint = max(0, xint2 - xint1), the height is hint = max(0, yint2 - yint1), and the intersection area is Areaint = wint × hint. Then calculate the union area: AreaA = (xA2 - xA1) × (yA2 - yA1), AreaB = (xB2 - xB1) × (yB2 - yB1), and the union area Areaunion = AreaA + AreaB - Areaint. Finally, IoU = Areaint / Areaunion. When IoU is greater than 0.5, the individuals are considered to be the same person.

[0039] The protective equipment (PE) testing employs a multi-category target detection algorithm, performing separate tests on three categories of PE: safety helmets, safety back clips, and workwear. Safety helmet testing focuses on the head area; the detection frame should completely encompass the helmet with a confidence level greater than 0.6. Safety back clip testing focuses on the central area of ​​the person's back, requiring the detection of clearly visible reflective stripes. Workwear testing analyzes the color and style of the person's entire clothing, requiring the detection of standard workwear color characteristics (typically highly visible blue or orange clothing).

[0040] The system records the wearing status of three types of protective equipment (PE) in the detection result list, represented by Boolean values: helmet_detected, backclip_detected, and workwear_detected. When all three status values ​​are True, it is determined that "PE is worn correctly," and the system announces "PE inspection passed, you can enter the work area." When any status value is False, the system records the abnormal personnel information and triggers a "PE not worn correctly" alarm, announcing "PE not worn correctly, please re-inspect and test again."

[0041] Example 6: Detection Result Plotting and Alarm Output The detection results were plotted using the OpenCV image processing library. Different colored rectangles were used to label various detection targets in each frame. Gas detectors were represented by yellow rectangles (RGB values ​​255, 255, 0), safety helmets by blue rectangles (RGB values ​​0, 0, 255), safety harnesses by green rectangles (RGB values ​​0, 255, 0), workwear by red rectangles (RGB values ​​255, 0, 0), and personnel by purple rectangles (RGB values ​​255, 0, 255). The category name and confidence score were displayed in the upper left corner of each detection box, using the Arial font, 16 pixels in size, and white in color.

[0042] Set a status display area in the upper left corner of the screen, with a size of 300×200 pixels and a semi-transparent black background (RGB value 0,0,0, transparency 0.7). The status display includes information such as the progress of the inspection process, completed inspection items, and the current inspection status. The text color is differentiated according to the status: green for normal status, yellow for warning status, and red for abnormal status.

[0043] like Figure 3 As shown, when the gas detection result is unqualified, the system highlights the abnormal information in the detection result image. A flashing red border appears around the gas detector, the status area in the upper left corner of the screen displays the red text "Gas Detection: Unqualified," and a warning banner "Do Not Enter the Work Area" appears at the bottom of the screen. The system automatically saves the abnormal image, with the file name formatted as "Abnormal_Gas Unqualified_YYYYMMDD_HHMMSS.jpg."

[0044] like Figure 4 As shown, when all test items are normal, the system displays the complete test results. The gas detector displays a green border indicating a qualified status, and the personnel target displays the complete test results for protective equipment. Safety helmets, safety back clips, and work clothes are all marked with corresponding colored borders. The status area in the upper left corner of the screen displays the green text "Gas Detection: Qualified" and "Protective Equipment: Compliant," and the bottom of the screen displays the confirmation message "Inspection Completed, Entry to Work Area Permitted."

[0045] The alarm output module employs a multi-level alarm mechanism, categorized into three levels based on the severity of the anomaly. Level 1 alarms correspond to "gas detection not performed," triggering a flashing yellow warning light and a voice prompt. Level 2 alarms correspond to "gas detection failure," triggering a continuous red warning light, a voice alarm, and an SMS notification to relevant management personnel. Level 3 alarms correspond to "improper wearing of protective equipment," triggering a flashing yellow warning light and a voice prompt. All alarm information is simultaneously pushed to the visualization system and the dispatch center to ensure timely detection and handling of anomalies.

[0046] Through the above specific implementation methods, the present invention realizes fully automated intelligent monitoring of standardized monitoring before tunneling operations, effectively improving the level of mine safety management and operational efficiency.

[0047] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0048] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A pre-tunneling operation standardization monitoring algorithm based on convolutional neural networks, characterized in that... It includes the following steps: Step 1: Continuously monitor the video stream status through an intrinsically safe camera. When no valid video frames can be obtained for 30 consecutive seconds, it is determined that the camera is offline. When the subsequent video stream resumes and clear video frames are obtained for 10 consecutive seconds, it is determined that the blasting operation is completed and the monitoring process is started; Step 2: Preprocess the collected images, including performing illumination compensation using the Retinex algorithm, histogram equalization using the contrast-limited adaptive histogram equalization algorithm, denoising using the bilateral filter, and geometric correction using perspective transformation; Step 3: Use a convolutional neural network model based on the YOLOv8 detection network. The backbone network of the model uses the PP-HGNetV2 structure, and a CBAM attention module is added to the feature fusion layer; Step 4: Start timing for 3 minutes after detecting that a person enters the monitoring area. Detect the status of the gas detector, and determine whether the gas detector is within the operation range of the person by calculating the positional relationship between the center point coordinates of the gas detector detection box and the person detection box coordinates; Step 5: When the gas detector is successfully matched, determine the status of the gas detector through the RGB value range. The red light status corresponds to R>200 and G<100 and B<100, and the green light status corresponds to R<100 and G>200 and B<100; Step 6: After the gas detection is qualified, confirm the same person by calculating the intersection over union (IoU) of the person targets in adjacent frames. When the IoU is greater than 0.5, it is considered the same person, and detect the wearing conditions of three types of protective equipment for the person, namely safety helmets, safety back clips, and work clothes.

2. The algorithm according to claim 1, characterized in that: The method for matching and judging the gas detector and the person in Step 4 is as follows: The coordinates of the gas detector detection box are (x1, y1, x2, y2), and the coordinates of the person detection box are (x3, y3, x4, y4). Calculate the center point of the gas detector target Xcenter=(x1 + x2) / 2, Ycenter=(y1 + y2) / 2. When x3<Xcenter<x4 and y3<Ycenter<y4 are satisfied, it is determined that the match is successful.

3. The algorithm according to claim 1, characterized in that: The method for calculating the intersection over union of adjacent frame person targets in Step 6 is as follows: Calculate the intersection area Areaint of the overlapping area and the union area Areaunion of their respective detection boxes. IoU = Areaint / Areaunion.

4. The algorithm according to claim 1, characterized in that: The CBAM attention module includes a channel attention sub-module and a spatial attention sub-module. The channel attention obtains channel weights through global average pooling and global maximum pooling, and the spatial attention calculates spatial weights through a 7×7 convolutional kernel.

5. A pre-tunneling operation standardization monitoring system based on convolutional neural networks, characterized in that... It includes: An intrinsically safe camera, installed at a height of 2.5 - 3 meters above the ground at the top of the tunneling face, supporting a resolution of 1920×1080, equipped with a dust-proof transparent cover and a three-axis anti-vibration bracket; An edge inference device, an industrial host equipped with a GPU computing unit, with a memory configuration of not less than 8GB and a solid-state storage of 256GB, used to run the convolutional neural network model described in claim 1; A voice broadcast unit, using an explosion-proof speaker with a power of not less than 15W, connected to the edge inference device through an audio signal line; The visualization terminal uses a 15-inch industrial display screen with a resolution of 1920×1080 to display monitoring images in real time. The network communication module includes an underground industrial Ethernet switch that connects cameras and edge inference devices.

6. The system according to claim 5, characterized in that: The installation location of the intrinsically safe camera ensures that the monitoring coverage area is not less than 5×8 meters, and that the display size of the operator in the screen is not less than 50×50 pixels, and the display size of the gas detector is not less than 30×30 pixels.

7. The system according to claim 5, characterized in that... It also includes an alarm module, which adopts a multi-level alarm mechanism. The first-level alarm triggers a flashing yellow warning light when gas detection is not performed, the second-level alarm triggers a solid red warning light when gas detection fails, and the third-level alarm triggers a flashing yellow warning light when protective equipment is not worn correctly.

8. A pre-tunneling standardization monitoring device based on convolutional neural networks, characterized in that... include: The image acquisition unit includes an intrinsically safe camera and an image preprocessing module; The data processing unit includes an edge inference device equipped with a convolutional neural network model. The model adopts the YOLOv8 detection network architecture, the backbone network is the PP-HGNetV2 structure, and the feature fusion layer adds a CBAM attention module. The result output unit includes a detection result plotting module and an alarm output module; The device automatically monitors gas detection and the wearing of protective equipment before tunneling operations by executing the algorithm described in claim 1.

9. The device according to claim 8, characterized in that: The convolutional neural network model was trained using a stochastic gradient descent optimizer with an initial learning rate of 0.01, a momentum parameter of 0.937, a weight decay coefficient of 0.0005, a training batch size of 16, and a cosine annealing learning rate scheduling strategy.

Images