Underground coal mine belt conveyor belt damage detection method based on deep learning

Abstract

The invention provides an underground coal mine belt conveyor belt damage detection method based on deep learning. The method comprises the following steps: constructing an image data set close to a real underground scene through brightness transformation, Gamma correction, dust noise simulation and motion blur processing; secondly, replacing a C3k2 / C2f structure of an original YOLOv11 by adopting a MobileNetv4 lightweight backbone network, introducing a GAM global attention mechanism to enhance long-distance feature association, and improving the detection robustness in a multi-scale and complex scene by adopting a Dynamic Head detection head; and finally, designing dual-stage bounding box loss (CIoU-SIOU) and class balance Focal Loss, and optimizing damage positioning and rare damage classification performance. According to the method, on the premise that the real-time performance is guaranteed, the detection precision and robustness of damage such as belt cracks, scratches and delamination are remarkably improved, and the method is suitable for continuous, automatic and high-reliability belt health state monitoring under a coal mine.

Description

Technical Field

[0001] This invention relates to the field of image recognition and industrial intelligent detection technology, specifically a method for detecting belt damage in underground coal mine conveyors based on deep learning. Background Technology

[0002] Belt conveyors are core equipment for material transportation in underground coal mines, undertaking the continuous conveying tasks throughout the entire process of coal mining, transportation, and storage. Their operational stability directly affects underground production efficiency and operational safety. As a key load-bearing component of the conveyor, the belt is constantly exposed to complex environments such as high dust, high humidity, low light, and heavy impact, making it highly susceptible to various structural damages such as cracks, scratches, delamination, edge breakage, and pitting wear. If belt damage is not detected and addressed in a timely manner, it may lead to tearing accidents, transmission interruptions, or even serious safety hazards. Therefore, conducting high-precision, real-time belt damage monitoring is of great significance.

[0003] Traditional methods for detecting belt damage mainly include manual inspection and various non-destructive testing (NDT) techniques. Manual inspection relies on experience, is highly subjective, and is significantly affected by factors such as insufficient underground lighting and dust, resulting in a high rate of missed detections. NDT methods such as ultrasonic, magnetic particle, and X-ray testing are heavily limited by equipment and environmental conditions, have high testing costs, and are difficult to implement for online monitoring under continuous operation. Furthermore, destructive testing requires cutting or locally damaging the belt, making it unsuitable for actual underground working environments. Against this backdrop, image analysis-based detection methods have gradually been introduced into the field of belt monitoring. However, due to the common problems of insufficient lighting, strong noise, and multi-grayscale interference in underground images, traditional image processing techniques struggle to extract effective features when dealing with complex belt surface textures, leading to insufficient detection accuracy and robustness, and failing to meet the needs of continuous monitoring in coal mines.

[0004] With the rapid development of artificial intelligence technology, deep learning has provided a new path for belt damage detection. Deep learning models possess powerful feature representation capabilities, enabling them to autonomously learn rich texture and structural information from a large number of belt images, avoiding the limitations of manually designed features in traditional methods. By constructing an end-to-end deep learning detection model, high-precision identification and location regression of belt damage can be achieved, significantly improving the efficiency and reliability of automated monitoring.

[0005] Currently, deep learning object detection algorithms are mainly divided into two categories. One category consists of two-stage methods, such as R-CNN, Faster-RCNN, and Mask-RCNN. These methods complete the detection task through two stages: candidate box generation and object recognition. While they perform well in terms of accuracy, their complex network structure and high computational cost make them unsuitable for real-time monitoring scenarios in underground coal mines. The other category consists of one-stage methods, such as SSD, YOLOv8, and YOLOv11. These methods integrate the detection task into a single stage, offering advantages such as high speed and convenient deployment, and are widely used in industrial visual inspection. However, when dealing with targets commonly found in underground coal mine environments, such as small-sized damage and weak-contrast cracks, these methods still tend to suffer from missing edge features and insufficient detail recognition capabilities, limiting their practical application in underground conveyor belt damage detection.

[0006] Therefore, there is an urgent need to develop a deep learning-based belt damage detection method that can operate effectively under complex underground coal mine conditions, possesses both high precision and robustness, and has stronger detection capabilities, in order to improve the intelligent monitoring level and safety assurance capabilities of underground conveying systems. Summary of the Invention

[0007] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a deep learning-based method for detecting belt damage in underground coal mine conveyors, improving the detection accuracy and robustness of belt cracks, scratches, delamination, and other damages. This method is suitable for continuous, automated, and highly reliable monitoring of belt health status in underground coal mines.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A deep learning-based method for detecting belt damage in underground coal mine conveyors includes the following steps:

[0010] S1: Acquire images of belt damage;

[0011] S2: Use a multi-strategy data augmentation method based on downhole scenarios to preprocess the acquired images;

[0012] S3: Classify and label the belt damage in the images to form a dataset for model training; divide the dataset according to the ratio of training set: validation set: test set = 7:2:1;

[0013] S4: Construct a coal mine underground conveyor belt damage detection model based on the improved YOLOv11;

[0014] S5: Use the above training set to train the improved YOLOv11 coal mine underground conveyor belt damage detection model; use the validation set to validate the model, and use the training and validation results to calculate the loss function of the model for each training session.

[0015] S6: Determine whether the model performance meets expectations or whether the maximum number of training iterations has been reached. If it does, proceed to step S7. If it does not, optimize the model's training parameters based on the loss function and sample allocation mechanism, and then repeat step S5 to train the model.

[0016] S7: Output the trained model and evaluate the performance of the trained belt damage detection model using the test set. The evaluation metrics are mAP@0.5, mAP@0.5:0.95, Precision, Recall, and FPS.

[0017] Preferably, in step S2, the specific steps of image preprocessing include:

[0018] S21: Simulation of low illumination and uneven local lighting. The brightness of the entire image is linearly transformed to simulate insufficient light and changes in lighting conditions in the well, generating belt images under various low illumination and polarized light conditions.

[0019] S22: Nonlinear Gamma transformation is used to further simulate different light source intensities and illumination attenuation in the well. The transformed image is then subjected to Gamma correction to simulate insufficient illumination and strong light irradiation.

[0020] S23: Add random noise to the image to enhance the model's robustness to belt damage texture under coal dust interference;

[0021] S24: Perform motion blur convolution on the image along the belt running direction to simulate the blurring of the belt image at different speeds, thereby improving the model's ability to recognize motion blur damage images.

[0022] Preferably, in step S4, the specific steps for constructing the model include:

[0023] S41: Use MobileNetv4 network to replace the C3k2 / C2f stacked structure in the backbone layer of the original YOLOv11 object detection model for feature extraction;

[0024] S42: Replace the C2PSA attention module in the backbone layer of the original network with the GAM module;

[0025] S43: Replace the detection head of the Head layer in the YOLOv11 model with the Dynamic Head detection head.

[0026] Preferably, in step S21, the formula for performing a linear brightness transformation on the entire image is:

[0027] ;

[0028] Where 𝐼(𝑥,𝑦) and 𝐼1(𝑥,𝑦) represent the pixel values ​​at pixel position (𝑥,𝑦) in the original image and the transformed image, respectively, with a value range of [0,255]. a∈[0.5,0.9] is the brightness scaling factor, used to simulate weak or distant lighting conditions of mine lamps; b∈[−20,10] controls the local illumination shift, used to simulate reflective positions and shadow areas; clip(⋅) is the pixel truncation function; in addition, to more realistically simulate local overexposure phenomena underground, regional brightness perturbations can also be added:

[0029] ;

[0030] M(x,y) is a local bright spot generated using a random shape mask, with a value range of [0, 𝛿], where 𝛿 is the maximum bright spot intensity, used to simulate the high-brightness area caused by direct sunlight from a miner's lamp.

[0031] Preferably, in step S22, the formula for simulating Gamma correction on the transformed image is:

[0032] ;

[0033] ;

[0034] Where γ∈[0.6,1.4] is randomly sampled according to the actual underground lighting intensity; when γ>1: enhances the details of the dark area, indicating insufficient lighting, aging of the miner's lamp, or a long distance; when γ<1: enhances the contrast of the bright area, simulating the direct illumination of the miner's lamp.

[0035] Preferably, in step S23, random noise is added to the image:

[0036] ;

[0037] ;

[0038] Where, σ n ∈[5,15] is used to simulate different dust concentrations. This indicates that the expression follows a zero mean and a variance of . Independent and identically distributed Gaussian noise. Additionally, salt-and-pepper noise can be added to simulate particulate coal dust adhesion.

[0039] ;

[0040] Where, p s The value ∈ [0.001, 0.01] depends on the dust content at the scene; in addition, to simulate camera lens contamination, random Gaussian blur or small-area occlusion texture overlay can be used to improve realism.

[0041] Preferably, in step S24, the formula for performing motion blur convolution on the image along the belt running direction is:

[0042] ;

[0043] Among them, K motion The kernel is a linear convolution kernel of length L∈{5,7,9,11} and direction φ, where "*" represents a two-dimensional discrete convolution operation. Furthermore, variable-speed blur can be added to randomly change the continuity of the blur length or direction.

[0044] Preferably, in step S42, the GAM module adopts a serial architecture of "channel attention + spatial attention", and the core calculation process can be fully characterized by the formula:

[0045] ;

[0046] ;

[0047] Where, 𝐹1 represents the input feature map with dimensions 𝐶×𝐻×𝑊, 𝐹2 represents the output feature map after channel attention weighting, and 𝐹3 represents the output feature map after spatial attention weighting of 𝐹2. That is, 𝐹1 represents the output feature map after weighting by the GAM module, and M... c and M s These represent channel attention and spatial attention, respectively. The input feature map, representing element-wise multiplication, has dimensions C×H×W. It first enters the channel attention branch, and after 3D arrangement to adjust the dimension order to W×H×C, it is input into two layers of MLP for non-linear transformation. Then, it is restored to its original dimensions through reverse arrangement and activated by Sigmoid to generate a channel attention map. After element-wise multiplication with the original feature map, it enters the spatial attention branch, where global spatial features are fused through two 7×7 convolutions. Then, it is activated by Sigmoid to generate a spatial attention map, and finally, the enhanced feature map is output.

[0048] Preferably, in step S43, the DynamicHead detection head adopts a three-level unified attention modeling structure of "scale awareness - spatial awareness - task awareness", the core of which is based on the input feature 𝐹 and through scale attention 𝜋 𝐿 (I), spatial attention 𝜋 𝑆 (I) and task attention 𝐶 (𝐹) gains enhanced features through progressive action. The formula is as follows:

[0049] ;

[0050] Among them, 𝜋 𝐿 (·), 𝜋 𝑆(·), 𝜋 𝐶 (·) represent the scale, spatial, and task attention mapping functions, respectively, and "·" indicates element-wise multiplication. The resulting enhanced features... The three branches of classification, center point regression, and bounding box regression enable a unified and efficient detection head design; the scale-aware part uses the following formula:

[0051] ;

[0052] The weights of features at different scales are dynamically adjusted so that the network can automatically select the optimal level based on the actual size of belt damage.

[0053] The spatial perception part utilizes deformable convolution:

[0054] ;

[0055] Among them, 𝑝 𝑘 Δi is the nth sampling offset of the standard convolution kernel. 𝑘 This represents the offset for network self-learning. Corresponding to the convolution weights. Adaptively sample the damage region, "avoiding interference and aligning with damage" from dust, stains, shadows, or motion blur, enhancing the robustness of downhole scenes;

[0056] The task awareness part passed:

[0057] ;

[0058] To establish differentiated channel preferences for classification and regression, the classification-localization conflict caused by the complex background of underground coal mines is alleviated, enabling the model to have a clearer representation ability for different types of damage.

[0059] Preferably, in step S5, during the initial training phase, Complete-IoU loss is used to simultaneously optimize the predicted bounding boxes in terms of overlapping regions, center point distance, and aspect ratio consistency. Its mathematical form is:

[0060] ;

[0061] Where, 𝐼𝑜𝑈 represents the intersection-union ratio of the predicted box and the ground truth box; 𝜌 represents the Euclidean distance between their center points; 𝑐 is the diagonal length of the minimum bounding box of the two boxes; 𝑣 is the aspect ratio consistency index; and 𝛼 is the weight coefficient.

[0062] Once the model parameters gradually stabilize, the training phase switches to Scylla-IoU loss for further parameter updates. SIoU introduces an angle factor into traditional IoU, ensuring that the predicted bounding boxes not only approximate the ground truth boxes but also align directionally. Its core idea is to achieve more refined bounding box regression through joint constraints of distance, angle, and shape. Its overall representation can be written as:

[0063] ;

[0064] in , , These correspond to penalties for center distance, frame shape differences, and angle differences, respectively.

[0065] The model parameters are optimized using class balance (FocalLoss), and its mathematical expression is as follows:

[0066] ;

[0067] Among them, 𝑝 𝑡 is the model's predicted probability for the true class; is the hard example adjustment factor; is the class balance weight automatically calculated based on the number of samples, adaptively calculated according to the inverse relationship between the number of samples in each class. This allows rare damages to receive higher importance during training.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] By integrating multi-strategy data augmentation with low-light transformation, dust noise, nonlinear illumination, and motion blur, the model adapts to complex working conditions such as uneven lighting, low visibility, and dynamic blur during the training phase, improving its generalization ability and robustness in harsh real-world environments. Secondly, targeted optimizations were made to the network structure: a lightweight and efficient MobileNetv4 backbone network was adopted, balancing feature extraction capabilities and computational efficiency for easy deployment on downhole edge devices; a GAM global attention mechanism was introduced to enhance the capture and association of long-distance, fragmented damage features; and a DynamicHead detection head was employed to dynamically adapt to damage targets of different sizes and shapes, improving detection accuracy in complex backgrounds. Finally, a two-stage bounding box loss optimization strategy from CIoU to SIoU was designed during the training phase, combined with FocalLoss for class balancing to address sample imbalance, strengthening the learning of rare damage categories and achieving a balance between high-precision localization and high-recall classification. Attached Figure Description

[0070] Figure 1 This is a flowchart of the present invention.

[0071] Figure 2This is a diagram of the improved YOLOv11 network structure in this invention.

[0072] Figure 3 This is a diagram of the MobileNetv4 network structure in this invention.

[0073] Figure 4 This is a structural diagram of the GAM module in this invention.

[0074] Figure 5 This is a structural diagram of the Dynamic Head module in this invention. Detailed Implementation

[0075] The invention will now be further described with reference to the accompanying drawings.

[0076] like Figures 1 to 5 As shown, a deep learning-based method for detecting belt damage in underground coal mine conveyors includes the following steps:

[0077] S1: Acquire images of belt damage.

[0078] S2: Use a multi-strategy data augmentation method based on the downhole scenario to preprocess the acquired images.

[0079] S21: Simulation of Low Illumination and Uneven Local Lighting (Linear Brightness Transformation). A linear brightness transformation is performed on the entire image. Underground coal mines commonly use point or line-based local lighting, resulting in overall low image brightness and the presence of excessively bright or dark areas. Therefore, this invention performs a linear brightness transformation on the original image:

[0080] ;

[0081] Where 𝐼(𝑥,𝑦) and 𝐼1(𝑥,𝑦) represent the pixel values ​​at pixel position (𝑥,𝑦) in the original image and the transformed image, respectively, with a value range of [0,255]; a∈[0.5,0.9] is the brightness scaling factor, used to simulate weak or distant lighting conditions of mine lamps; b∈[−20,10] controls the local illumination shift, used to simulate reflective positions and shadow areas; clip(⋅) is the pixel truncation function. Furthermore, to more realistically simulate local overexposure phenomena underground, regional brightness perturbations can also be added:

[0082] ;

[0083] M(x,y) represents a local bright spot generated using a random shape mask, with values ​​ranging from [0, 𝛿]. 𝛿 represents the maximum intensity of the bright spot, used to simulate the reflective high-brightness area caused by direct sunlight from a miner's lamp. This step can simulate complex underground lighting conditions of "overall darkness + local bright spots + local shadows," enabling the network to learn the ability to identify damage even under drastic changes in lighting.

[0084] S22: Downhole illumination exhibits significant nonlinear attenuation, meaning there is a substantial difference in contrast between bright and dark areas. Therefore, Gamma correction is used to simulate this nonlinear phenomenon.

[0085] ;

[0086] ;

[0087] Where γ∈[0.6,1.4] is randomly sampled based on the actual underground lighting intensity. When γ>1: details in dark areas are enhanced, indicating insufficient lighting, aging of the miner's lamp, or a long distance. When γ<1: contrast in bright areas is enhanced, simulating direct illumination from the miner's lamp. This step can significantly improve the model's ability to extract damage features under different lighting distributions, especially for low-contrast cracks and wear areas.

[0088] S23: A large amount of coal dust particles are often present in the air underground in coal mines. To simulate the random grayscale perturbation introduced by dust, random noise is superimposed on the image to enhance the model's robustness to conveyor belt damage textures caused by coal dust interference. The high concentration of coal dust in the air underground often results in random noise or smudges obscuring images in camera imaging. To simulate this characteristic, random noise is added to the image in this embodiment:

[0089] ;

[0090] ;

[0091] Where, σ n ∈[5,15] is used to simulate different dust concentrations. This indicates that the expression follows a zero mean and a variance of . Independent and identically distributed Gaussian noise; in addition, salt-and-pepper noise can be added to simulate particulate coal dust adhesion:

[0092] ;

[0093] Where, p s The range [0.001, 0.01] depends on the dust content at the site. Furthermore, to simulate camera lens contamination, random Gaussian blur or small-area occlusion texture overlay can be used to improve realism. This step enables the model to adapt to conditions such as coal dust obstruction, random particle interference, and camera contamination, helping to reduce false positives and false negatives.

[0094] S24: To simulate motion blur caused by the high-speed operation of the belt and the camera exposure time, the image is subjected to motion blur convolution along the belt's direction of movement.

[0095] ;

[0096] Among them, K motion The model uses a linear convolution kernel of length L∈{5,7,9,11} and direction φ (±10° of the belt running direction). "∗" represents a two-dimensional discrete convolution operation. The convolution operation can simulate rapid belt movement or camera equipment jitter. Furthermore, variable-speed blur (blurring caused by uneven speed) can be added, randomly changing the continuity of the blur length or direction. This step enhances the model's ability to learn weak-texture damage (fine cracks, shallow scratches) under motion blur, maintaining high detection accuracy even in high-speed belt environments.

[0097] S3: Classify and label the belt damage in the images to form a dataset for model training. Divide the dataset into training set: validation set: test set in a ratio of 7:2:1.

[0098] S4: Construct a coal mine underground conveyor belt damage detection model based on the improved YOLOv11.

[0099] S41: The C3k2 / C2f stacked structure in the backbone layer of the original YOLOv11 object detection model is replaced with MobileNetv4 for feature extraction, reducing environmental interference in industrial scenarios while achieving model lightweighting. Its structure diagram is shown below. Figure 3As shown, the Convbn module is the basic convolutional unit of MobileNet, which is responsible for extracting fine-grained features such as low-level texture and edges of the belt image layer by layer; the Uib module stacked subsequently is the lightweight bottleneck unit of MobileNet (usually adopting a structure of depthwise separable convolution combined with residual connections), which further abstracts high-level semantic features while retaining defect details through residual mechanism, and at the same time significantly reduces the number of computational parameters and inference overhead. Belt inspection often requires deployment on edge devices such as embedded cameras and industrial control computers. The lightweight design of this module allows YOLOv11 to achieve real-time detection with limited computing power, while avoiding overloading device resources. Belt damage (such as small cracks and localized wear) is often characterized by weak texture and small size. The layer-by-layer convolution of the Convbn module can effectively extract such details, while the residual connections of Uib avoid feature loss in deep networks, significantly improving the recall rate of small defects. Finally, this network can also enhance the ability to resist background interference. In industrial scenarios, belts often contain stains, complex textures, and other interfering information. The "bottom-to-high-level" feature abstraction process of this module can help YOLOv11 distinguish between "defect features" and "background noise," reducing false positive detections caused by background misidentification. Ultimately, it achieves a better balance between detection accuracy and deployment efficiency, meeting the actual industrial needs of belt damage detection.

[0100] S42: The C2PSA attention module in the backbone layer of the original network is replaced with the GAM (Global Attention Mechanism) module to strengthen the global correlation of features, solve the problem of feature dispersion of long-distance targets, and improve the detection accuracy and integrity recognition ability of large-sized defects. The GAM module adopts a serial architecture of "channel attention + spatial attention", and the core calculation process can be fully represented by the formula:

[0101] ;

[0102] ;

[0103] Where, 𝐹1 represents the input feature map with dimensions 𝐶×𝐻×𝑊, 𝐹2 represents the output feature map after channel attention weighting, and 𝐹3 represents the output feature map after spatial attention weighting of 𝐹2. In other words, 𝐹1 represents the output feature map after weighting by the GAM module; M c and M s These represent channel attention and spatial attention, respectively. This represents element-wise multiplication. The input feature map has dimensions C×H×W. It first enters the channel attention branch, where the dimensionality is adjusted to W×H×C via 3D arrangement. It is then input into two layers of MLP for non-linear transformation, followed by a reverse arrangement to restore the original dimensions. After Sigmoid activation, a channel attention map is generated, which is then element-wise multiplied with the original features before entering the spatial attention branch. This branch fuses global spatial features through two 7×7 convolutions, and then generates a spatial attention map via Sigmoid activation, ultimately outputting enhanced features. In the task of detecting belt conveyor damage in underground coal mines, the 3D arrangement and reverse arrangement of the channel branch avoid the spatial information loss caused by traditional global pooling, fully preserving the spatial details of minor damage such as belt cracks and adhesive peeling. The MLP strengthens the cross-dimensional correlation between channels and space, enabling the model to accurately distinguish the channel response differences between the damaged area and dust / background. The 7×7 large convolution kernel of the spatial branch uses a pooling-free design, through Conv... 7×7 It can cover a larger receptive field, adapt to long-distance feature associations of dispersed damage, and solve the problem of blurred damage edge features caused by pooling operations under low illumination. Overall, GAM, through its formula design that preserves all dimensions of features, not only solves the information loss problem of traditional attention modules, but also improves upon the M... c and M s The dual weighting enhances the perception of damage features in complex environments. Using this module, the detection recall rate of fine belt damage can be effectively improved by outputting more accurate damage feature representations, and the false negative rate under low light and dust interference can be reduced, making it more suitable for the actual needs of downhole belt damage detection.

[0104] S43: The detection head of the YOLOv11 model's Head layer is replaced with the DynamicHead detection head to enhance the model's detection generalization in complex scenarios, enabling the model to dynamically adapt to belt damage detection requirements with large scale differences, complex spatial deformations, and high noise interference. The DynamicHead detection head adopts a three-level unified attention modeling structure of "scale awareness - spatial awareness - task awareness," with its core based on input features 𝐹 and scale attention 𝜋. 𝐿 (I), spatial attention 𝜋 𝑆 (I) and task attention 𝐶 (𝐹) gains enhanced features through progressive action. The formula is as follows:

[0105] ;

[0106] Among them, 𝜋 𝐿 (·), 𝜋 𝑆 (·), 𝜋 𝐶(·) represents the scale, spatial, and task attention mapping functions, respectively; "·" indicates element-wise multiplication, resulting in enhanced features. The three branches of classification, center point regression, and bounding box regression enable a unified and efficient detection head design. The scale-aware part uses the following formula:

[0107] ;

[0108] The weights of features at different scales are dynamically adjusted so that the network can automatically select the optimal level based on the actual size of belt damage (such as millimeter-level cracks to large-area tears).

[0109] The spatial perception part utilizes deformable convolution:

[0110] ;

[0111] Among them, 𝑝 𝑘 Δi is the nth sampling offset of the standard convolution kernel. 𝑘 This represents the offset for network self-learning. To correspond to the convolution weights, the system adaptively samples the damage region, "avoiding interference and aligning with damage" from dust, stains, shadows, or motion blur, thus enhancing the robustness of the downhole scene.

[0112] The task awareness part passed:

[0113] ;

[0114] To establish differentiated channel preferences for classification and regression, the classification-localization conflict caused by the complex background in coal mines (coal blocks, rollers, supports, etc.) is alleviated, and the model has a clearer ability to represent different types of damage (tears, delamination, burns, etc.).

[0115] Overall, DynamicHead's triple attention structure introduces stronger multi-scale perception, robustness to complex scenes, and task decoupling capabilities to YOLOv11, significantly improving the detection accuracy and stability of belt damage without significantly increasing computational load. When replacing the original YOLOv11 detection head with DynamicHead, the model exhibits higher recall for small cracks, more accurate localization of large-area damage, and significantly enhanced adaptability to coal dust and fuzzy conditions, making it highly suitable as the core detection structure for underground belt conveyor belt damage detection.

[0116] S5: Train the improved YOLOv11 model using the training set described above. Validate the model using the validation set, and calculate the loss function for each training iteration based on the training and validation results. Specifically, this step uses a training loss function specifically designed for the characteristics of belt conveyor damage in underground coal mines, to update model parameters during training, thereby optimizing damage localization accuracy and damage classification performance. In the initial training phase, Complete-IoU (CIoU) loss is used to simultaneously optimize the predicted bounding boxes in terms of overlapping areas, center point distance, and aspect ratio consistency. Its mathematical form is:

[0117] ;

[0118] Where CIoU represents the intersection-union ratio (IU) between the predicted and ground truth bounding boxes; φ represents the Euclidean distance between their center points; φ is the diagonal length of the minimum bounding box; φ is the aspect ratio consistency index; and φ is the weighting coefficient. Since damage boundaries in downhole images are often blurred due to insufficient lighting and dirt occlusion, CIoU can provide a more stable gradient signal, avoiding large oscillations in the predicted bounding boxes in the early stages of training, allowing the model to converge to a reasonable position faster and more stably.

[0119] Once the model parameters stabilize, the training process switches to Scylla-IoU (SIoU) loss for further parameter updates. SIoU introduces an angle factor into traditional IoU, ensuring that the predicted bounding boxes not only approximate the ground truth bounding boxes but also align directionally. This is particularly suitable for fitting damage commonly found in underground coal mines, such as thin, elongated cracks and narrow scratches. Its core idea is to achieve more refined bounding box regression through joint constraints of distance, angle, and shape. Its overall representation can be written as:

[0120] ;

[0121] Each term corresponds to a penalty for center distance, frame shape difference, and angle difference. Through further constraints of SIoU, the model's bounding box can more accurately fit the actual shape of the belt surface damage, significantly reducing the positioning error of weak edge damage in downhole drilling.

[0122] In damage classification, due to the highly uneven frequency of belt conveyor damage in underground coal mines—for example, cracks have a large number of samples, while delamination, bulging, and edge breakage have very few—if ordinary cross-entropy loss is used, the network will tend to predict high-frequency categories while ignoring low-frequency damage, resulting in serious missed detections. Therefore, this invention employs class balancing FocalLoss to optimize the model parameters, and its mathematical expression is:

[0123] ;

[0124] Among them, 𝑝 𝑡 is the model's predicted probability for the true class; is the hard-case adjustment factor; is the class balance weight automatically calculated based on the number of samples, adaptively calculated according to the inverse relationship between the number of samples in each class, so that rare damages receive higher importance during training. Through this loss function, the model prioritizes learning damage patterns that are difficult to identify, have few samples, or have extremely low contrast when updating the classification layer parameters, thus significantly improving the ability to identify rare damage categories. These two types of loss functions continuously affect the model parameters during training, enabling the network to simultaneously optimize in both localization and classification: the localization part gradually transitions from stable bounding box selection to high-precision fitting, while the classification part shifts from high-frequency damage dominance to multi-class balanced learning. Ultimately, the model can maintain accurate and robust damage detection capabilities in the actual environment of low illumination, high noise, and high interference in coal mines, significantly reducing missed and false detections, and ensuring the reliability and safety of the conveyor belt damage monitoring system.

[0125] S6: Determine whether the model performance meets expectations or whether the maximum number of training iterations has been reached. If it does, proceed to step S7. If it does not, optimize the model's training parameters based on the loss function and sample allocation mechanism, and then repeat step S5 to train the model.

[0126] S7: Output the trained model and evaluate the performance of the trained belt damage detection model using the test set. Evaluation metrics include mAP@0.5, mAP@0.5:0.95, Precision, Recall, and FPS.

[0127] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for detecting belt damage in underground coal mine conveyors based on deep learning, characterized in that, Includes the following steps: S1: Acquire images of belt damage; S2: Use a multi-strategy data augmentation method based on downhole scenarios to preprocess the acquired images; S3: Classify and label the belt damage in the images to form a dataset for model training; divide the dataset according to the ratio of training set: validation set: test set = 7:2:1; S4: Construct a coal mine underground conveyor belt damage detection model based on the improved YOLOv11; S5: Use the above training set to train the improved YOLOv11 coal mine underground conveyor belt damage detection model; use the validation set to validate the model, and use the training and validation results to calculate the loss function of the model for each training session. S6: Determine whether the model performance meets expectations or whether the maximum number of training iterations has been reached. If it does, proceed to step S7. If it does not, optimize the model's training parameters based on the loss function and sample allocation mechanism, and then repeat step S5 to train the model. S7: Output the trained model and evaluate the performance of the trained belt damage detection model using the test set. The evaluation metrics are mAP@0.5, mAP@0.5:0.95, Precision, Recall, and FPS.

2. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 1, characterized in that, In step S2, the specific steps of image preprocessing include: S21: Simulation of low illumination and uneven local lighting. The brightness of the entire image is linearly transformed to simulate insufficient light and changes in lighting conditions in the well, generating belt images under various low illumination and polarized light conditions. S22: Nonlinear Gamma transformation is used to further simulate different light source intensities and illumination attenuation in the well. The transformed image is then subjected to Gamma correction to simulate insufficient illumination and strong light irradiation. S23: Add random noise to the image to enhance the model's robustness to belt damage texture under coal dust interference; S24: Perform motion blur convolution on the image along the belt running direction to simulate the blurring of the belt image at different speeds, thereby improving the model's ability to recognize motion blur damage images.

3. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 1, characterized in that, In step S4, the specific steps for constructing the model include: S41: Use MobileNetv4 network to replace the C3k2 / C2f stacked structure in the backbone layer of the original YOLOv11 object detection model for feature extraction; S42: Replace the C2PSA attention module in the backbone layer of the original network with the GAM module; S43: Replace the detection head of the Head layer in the YOLOv11 model with the Dynamic Head detection head.

4. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 2, characterized in that, In step S21, the formula for performing a linear brightness transformation on the entire image is: ； Where, 𝐼(𝑥,𝑦) and 𝐼1(𝑥,𝑦) represent the pixel values ​​at pixel position (𝑥,𝑦) in the original image and the transformed image, respectively, with a value range of [0,255]; a∈[0.5,0.9] is the brightness scaling factor, used to simulate weak or distant lighting conditions of mine lamps; b∈[−20,10] controls the local illumination shift, used to simulate reflective positions and shadow areas; clip(⋅) is the pixel truncation function; in addition, to more realistically simulate local overexposure phenomena underground, regional brightness perturbations can also be added: ； M(x,y) is a local bright spot generated using a random shape mask, with a value range of [0, 𝛿], where 𝛿 is the maximum bright spot intensity, used to simulate the high-brightness area caused by direct sunlight from a miner's lamp.

5. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 4, characterized in that, In step S22, the formula for simulating Gamma correction on the transformed image is: ； ； Where γ∈[0.6,1.4] is randomly sampled according to the actual underground lighting intensity; when γ>1: enhances the details of the dark area, indicating insufficient lighting, aging of the miner's lamp, or a long distance; when γ<1: enhances the contrast of the bright area, simulating the direct illumination of the miner's lamp.

6. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 5, characterized in that, In step S23, random noise is added to the image: ； ； Where, σ n ∈[5,15] is used to simulate different dust concentrations. This indicates that the expression follows a zero mean and a variance of . Independent and identically distributed Gaussian noise; in addition, salt-and-pepper noise can be added to simulate particulate coal dust adhesion: ； Where, p s The value ∈ [0.001, 0.01] depends on the dust content at the scene; in addition, to simulate camera lens contamination, random Gaussian blur or small-area occlusion texture overlay can be used to improve realism.

7. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 6, characterized in that, In step S24, the formula for performing motion blur convolution on the image along the belt running direction is: ； Among them, K motion A linear convolution kernel with length L∈{5,7,9,11} and direction φ, where "∗" represents a two-dimensional discrete convolution operation; in addition, variable speed blur can be added to randomly change the continuity of the blur length or direction.

8. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 3, characterized in that, In step S42, the GAM module adopts a serial architecture of "channel attention + spatial attention", and the core calculation process can be fully represented by the formula: ； ； Where, 𝐹1 represents the input feature map with dimensions 𝐶×𝐻×𝑊, 𝐹2 represents the output feature map after channel attention weighting, and 𝐹3 represents the output feature map after spatial attention weighting of 𝐹2. In other words, 𝐹1 represents the output feature map after weighting by the GAM module; M c and M s These represent channel attention and spatial attention, respectively. This represents element-wise multiplication. The input feature map has dimensions C×H×W. It first enters the channel attention branch, and after 3D arrangement to adjust the dimension order to W×H×C, it is input into two layers of MLP for non-linear transformation. Then, it is restored to the original dimension by reverse arrangement, activated by Sigmoid to generate the channel attention map, and multiplied element-wise with the original feature map before entering the spatial attention branch. After two 7×7 convolutions to fuse global spatial features, it is activated by Sigmoid to generate the spatial attention map, and finally outputs the enhanced feature.

9. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 8, characterized in that, In step S43, the DynamicHead detection head adopts a three-level unified attention modeling structure of "scale awareness - spatial awareness - task awareness". Its core is based on the input feature 𝐹, and is implemented through scale attention 𝜋 𝐿 (I), spatial attention 𝜋 𝑆 (I) and task attention 𝐶 (𝐹) gains enhanced features through progressive action. The formula is as follows: ； Among them, 𝜋 𝐿 (·), 𝜋 𝑆 (·), 𝜋 𝐶 (·) represent the scale, spatial, and task attention mapping functions, respectively, and "·" indicates element-wise multiplication; the resulting enhanced features are then... The three branches of classification, center point regression, and bounding box regression enable a unified and efficient detection head design; the scale-aware part uses the following formula: ； The weights of features at different scales are dynamically adjusted so that the network can automatically select the optimal level based on the actual size of belt damage. The spatial perception part utilizes deformable convolution: ； Among them, 𝑝 𝑘 Δi is the nth sampling offset of the standard convolution kernel. 𝑘 This represents the offset for network self-learning. To correspond to the convolution weights; adaptively sample the damage region to "avoid interference and align with damage" from dust, stains, shadows or motion blur, thereby enhancing the robustness of the downhole scene; The task awareness part passed: ； To establish differentiated channel preferences for classification and regression, the classification-localization conflict caused by the complex background of underground coal mines is alleviated, enabling the model to have a clearer representation ability for different types of damage.

10. The method for detecting belt damage in underground coal mine conveyors based on deep learning as described in claim 1, characterized in that, In step S5, during the initial training phase, Complete-IoU loss is used to simultaneously optimize the predicted bounding boxes in terms of overlapping regions, center point distance, and aspect ratio consistency. Its mathematical form is: ； Where, 𝐼𝑜𝑈 represents the intersection-union ratio of the predicted box and the ground truth box; 𝜌 represents the Euclidean distance between their center points; 𝑐 is the diagonal length of the minimum bounding box of the two boxes; 𝑣 is the aspect ratio consistency index; and 𝛼 is the weight coefficient. Once the model parameters gradually stabilize, the training phase switches to Scylla-IoU loss for further parameter updates. SIoU introduces an angle factor into traditional IoU, ensuring that the predicted bounding boxes not only approximate the ground truth boxes but also align directionally. Its core idea is to achieve more refined bounding box regression through joint constraints of distance, angle, and shape. Its overall representation can be written as: ； in , , These correspond to penalties for center distance, frame shape differences, and angle differences, respectively. The model parameters are optimized using class balance FocalLoss, and its mathematical expression is as follows: ； Among them, 𝑝 𝑡 is the model's predicted probability of the true class; is the hard example adjustment factor; is the class balance weight automatically calculated based on the number of samples, which is adaptively calculated according to the inverse relationship between the number of samples in each class, so that rare damages get higher importance in training.

Images