Coal gangue recognition system and method based on improved YOLOv11 for coal mine conveying belt

Abstract

The application relates to the technical field of image processing, and particularly discloses a coal mine conveying belt coal gangue recognition system and method based on an improved YOLOv11, which comprises an image acquisition unit for acquiring real-time images of a coal mine conveying belt; an edge computing processing unit loads and runs an improved YOLOv11 algorithm model, and performs real-time coal gangue detection on the conveying belt images acquired by the image acquisition unit; the improved YOLOv11 algorithm model comprises a backbone network, a neck network and a detection head connected in sequence; the backbone network comprises at least one adaptive down-sampling module for down-sampling input feature maps; the neck network is used for fusing multi-scale features output by the backbone network to generate a fused feature map; the detection head is a P2-P5 lightweight detection head, which is used for receiving four layers of feature maps output by the neck network respectively, and after the feature maps are fused through an adaptive fusion layer, the detection head is input into a bounding box regression branch, a class prediction branch and a confidence prediction branch respectively for decoupled output.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, specifically relating to a coal mine conveyor belt coal gangue identification system and method based on an improved YOLOv11. Background Technology

[0002] Coal gangue mixed in during coal conveyor transport can seriously affect coal quality, damage crushing equipment, and cause safety accidents. Traditional manual visual inspection is inefficient and highly subjective; image processing methods based on edge detection and color segmentation have poor generalization ability under complex working conditions such as changes in lighting and coal dust coverage. In recent years, deep learning has become the mainstream technology path, among which the YOLO series algorithms have received widespread attention due to their balance between real-time performance and accuracy. For example, the "Foreign Object Detection Method for Conveyor Belts Based on IML-YOLOv11" published in 2025 uses the standard YOLOv11n as the basic network, including a backbone composed of C3k2 modules, a neck with an FPN / PAN structure, and a multi-scale detection head; it is trained on a public dataset, inputting a 640×640 image, and outputting coal gangue bounding boxes and categories; it is trained on an NVIDIA RTX 3090 with an SGD optimizer, an initial learning rate of 0.001, and 50 training epochs.

[0003] However, this technology has the following problems: (1) Feature information decay: Downsampling uses stride convolution or pooling, which leads to irreversible loss of shallow texture details, and key discrimination features such as small cracks in coal gangue cannot be transmitted to deep networks; (2) Lack of adaptive feature focusing: The standard Neck uses fixed weight fusion, which cannot dynamically adjust the importance of features according to the similarity between coal gangue and coal blocks, resulting in a high false detection rate; (3) Weak small target detection mechanism: The P3-P5 detection layer has a weak response to targets with pixels smaller than 8×8, resulting in serious missed detections; (4) Computational redundancy: There is a lot of parameter redundancy in the C3k2 module with a large number of 3×3 convolutions, resulting in slow inference speed and a frame rate of less than 15 FPS on edge devices such as Jetson Nano.

[0004] Other studies have introduced attention modules in YOLOv7 to enhance the representation of small targets, or adopted the Ghost module for lightweighting, but these improvements have not systematically solved the problem of multi-scale feature fusion and information decay in YOLOv11 in coal mine scenarios. Summary of the Invention

[0005] To address the shortcomings and problems of existing technologies for coal gangue identification, which fail to optimize the structure for the small size, diverse shapes, and high speed of coal gangue, resulting in low accuracy and a large number of model parameters that are difficult to deploy at the edge, this invention provides a coal gangue identification system and method for coal mine conveyor belts.

[0006] The coal gangue identification system for coal conveyor belts based on the improved YOLOv11 provided by this invention includes an image acquisition unit and an edge computing processing unit; the image acquisition unit is used to acquire real-time images of the coal conveyor belt. The edge computing processing unit loads and runs an improved YOLOv11 algorithm model to perform real-time coal gangue detection on the conveyor belt image acquired by the image acquisition unit; the improved YOLOv11 algorithm model includes a backbone network, a neck network, and a detection head connected in sequence. The backbone network includes at least one adaptive downsampling module for downsampling the input feature map; The neck network is used to fuse multi-scale features output by the backbone network to generate a fused feature map. The detection head is a lightweight detection head, P2-P5, which is used to receive four layers of feature maps output from the neck network. After the feature maps of each layer are fused by an adaptive fusion layer, they are input into the bounding box regression branch, the category prediction branch and the confidence prediction branch respectively for decoupled output.

[0007] The aforementioned coal gangue identification system for coal conveyor belts based on improved YOLOv11 includes a backbone network comprising an initial convolutional unit, three sets of feature enhancement and downsampling units, a fourth feature enhancement unit, and a multi-scale feature fusion unit. The initial convolutional unit consists of two cascaded convolutional layers, performing initial convolution operations on the input image to complete preliminary feature extraction and channel dimension adjustment. The feature enhancement and downsampling unit sets employ an alternating cascaded structure of feature enhancement and downsampling modules. The feature enhancement module is a C3K2 module, which uses residual connections and grouped convolutional structures. The downsampling module is an ADown module, which uses stride convolution to downsample the feature map size.

[0008] The aforementioned coal gangue identification system for coal conveyor belts based on improved YOLOv11 includes an adaptive downsampling module comprising a left branch and a right branch. The left branch includes an average pooling layer and a 1×1 convolutional layer, while the right branch includes a 3×3 convolutional layer and a max pooling layer. The outputs of the left and right branches are fused through channel splicing and then passed through a 3×3 convolutional layer to output multi-scale features.

[0009] The aforementioned coal gangue identification system for coal conveyor belts based on improved YOLOv11 includes a stacked attention-enhanced residual module in the FPN / PAN structure of the neck network. The stacked attention-enhanced residual module includes a parallel attention branch and an enhanced feedforward network. The parallel attention branch is used to perform spatial attention and channel attention calculations on the input feature map and fuse them. Its output is added to the input feature map through a residual connection. The enhanced feedforward network includes a first 1×1 convolutional layer, a 3×3 depth-separable convolutional layer, and a second 1×1 convolutional layer.

[0010] This invention also provides a method for identifying coal gangue in coal mine conveyor belts based on an improved YOLOv11, which is based on the aforementioned coal gangue identification system for coal mine conveyor belts and includes the following steps: Step 1: Collect and acquire real-time images of coal mine conveyor belts containing coal gangue targets; Step 2: Input the real-time image into the improved YOLOv11 algorithm model for processing. The processing steps of the improved YOLOv11 algorithm model include: (1) Adaptive downsampling: The input feature map is processed by at least one adaptive downsampling module. The adaptive downsampling module performs the left branch processing of average pooling and 1×1 convolution and the right branch processing of 3×3 convolution and max pooling on the input feature map in parallel, and then concatenates the results of the two branches in the channel dimension and outputs them by 3×3 convolution. (2) Feature enhancement: In the feature fusion process of the Neck network, the feature map is enhanced by the C2PSA-EDFFN feature enhancement module. The C2PSA-EDFFN feature enhancement module first calculates spatial attention and channel attention through parallel attention branches and fuses them, then adds them to the input feature map through residual connections, and then performs nonlinear transformation through an enhanced feedforward network containing depthwise separable convolutions. (3) Multi-scale detection: The P2-P5 lightweight detection head is used to adaptively fuse the four feature maps of P2, P3, P4 and P5 output by the Neck network, and the bounding box position, target category and confidence are decoupled and predicted for the fused features respectively. Step 3: Post-process the detection results output by the improved YOLOv11 algorithm model and output them.

[0011] The above-mentioned coal gangue identification method for coal mine conveyor belts based on improved YOLOv11 includes the following steps in step one: Acquiring and obtaining real-time images of coal mine conveyor belts containing coal gangue targets. (1) Real-time acquisition of video stream of coal flow from conveyor belt, uniform extraction of frame images from the acquired video stream at time intervals, and retention of positive sample images containing coal gangue targets after manual screening; (2) LabelImg professional annotation tool was used to accurately annotate all positive sample images. The annotation process strictly followed the rule of completely surrounding the target and reserving a 5-pixel edge, and the target was clearly distinguished into two categories: coal and gangue. (3) Random brightness adjustment simulates the low illumination conditions in the well, adds Gaussian noise with a mean square error σ=0.01, uses a random scaling factor of 0.8 to 1.2 to simulate the change in detection distance, and implements a Mosaic enhancement strategy of random stitching of four images to enhance the data of the image; (4) The enhanced image is uniformly scaled to a resolution of 640×640 pixels, normalized, and then converted to RGB tensor format.

[0012] In the above-mentioned coal gangue identification method based on improved YOLOv11, in step three, the decoupled prediction information of the output bounding box position, target category and confidence level is filtered by the NMS non-maximum suppression algorithm to filter overlapping detection boxes, and only the output results with a confidence level higher than 0.5 are retained.

[0013] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention improves upon existing technologies through modular hardware-level algorithm structure and customized training processes. By replacing standard stride convolution with a dual-branch parallel pooling-convolution structure in the YOLOv11 backbone network, it preserves global context and local detail features, enabling the effective transmission of minute features of coal gangue. Parallel spatial-channel attention and enhanced feedforward networks are embedded in the neck network FPN / PAN structure to adaptively focus on the coal gangue region and suppress coal background interference. A shallow detection layer P2 is added, and shared-weight depthwise separable convolution is used to achieve multi-scale feature adaptive fusion, specifically optimizing the detection of small targets at the 8×8 pixel level. It can improve mAP50 to 78.03% with a 14.3% reduction in parameters, achieving a real-time processing capability of 25 FPS on NVIDIA Jetson Nano. It can identify coal gangue with a minimum size of 0.1mm (corresponding to an 8×8 pixel image) and can adapt to harsh working conditions such as coal mine dust, low illumination (minimum 10 lux), and vibration. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the image-based intelligent coal gangue detection system of the present invention; Figure 2 This is a structural diagram of the detection system model of the present invention; Figure 3 This is a structural diagram of the adaptive downsampling module (ADown downsampling module) of the present invention; Figure 4 This is a structural diagram of the attention-enhanced residual module (C2PSA_EDFFN module) of the present invention; Figure 5This is a structural diagram of the detection head module of the present invention; Figure 6 This is a curve comparing the accuracy of the model of this invention with that of the traditional YOLOv11 model. Detailed Implementation

[0015] To address the problems of feature information attenuation, lack of adaptive feature focusing, weak small target detection mechanism, and computational redundancy in existing methods for detecting foreign objects on conveyor belts, this invention provides a coal gangue identification method and system based on improved YOLOv11. The invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0016] Example 1: This example provides a coal gangue identification system for coal mine conveyor belts based on an improved YOLOv11, such as... Figure 1 As shown, the overall system architecture includes an image acquisition unit, an edge computing processing unit, a storage unit, and a human-computer interaction unit. The image acquisition unit (such as an industrial-grade high-definition camera device (recommended: Basler acA2500-20gm 5-megapixel global shutter camera equipped with an 8mm focal length industrial lens)) is installed 1.5 meters above the conveyor belt in the coal mine, and the coal flow video stream is acquired in real time through a gigabit network port.

[0017] The edge computing processing unit loads and runs an improved YOLOv11 algorithm model to perform real-time coal gangue detection on the conveyor belt image acquired by the image acquisition unit. The improved YOLOv11 algorithm model includes a backbone network, a neck network, and a head connected in sequence. The backbone network includes at least one adaptive downsampling module for downsampling the input feature map. The adaptive downsampling module includes a left branch and a right branch. The left branch includes an average pooling layer and a 1×1 convolutional layer in sequence. The right branch includes a 3×3 convolutional layer and a max pooling layer in sequence. The outputs of the left and right branches are concatenated and fused through channels, and then passed through a 3×3 convolutional layer to output multi-scale features.

[0018] As a specific example, such as Figure 2 As shown, the backbone network includes an initial convolutional unit, three sets of feature enhancement and downsampling units, a fourth feature enhancement unit, and a multi-scale feature fusion unit. The initial convolutional unit consists of two concatenated convolutional layers, which perform initial convolution operations on the input image to complete the preliminary feature extraction and channel dimension adjustment of the image, providing a basic feature map for subsequent feature enhancement.

[0019] The feature enhancement and downsampling unit group consists of three groups, employing an alternating cascaded structure of feature enhancement and downsampling modules. The feature enhancement module is a C3K2 module, which uses residual connections and grouped convolutional structures. The downsampling module is an ADown module, which uses strided convolution to downsample the feature map size. The first group of C3K2+ADown is connected to the output of the initial convolutional unit; the second group of C3K2+ADown is connected to the output of the first group; and the third group of C3K2+ADown is connected to the output of the second group.

[0020] As an example, such as Figure 3 As shown, the adaptive downsampling module is used to downsample the input feature map (640×640×3 RGB image tensor). The adaptive downsampling module includes a left branch and a right branch. The left branch includes an average pooling layer (window 2×2, stride 2) and a 1×1 convolutional layer (output channel C / 2) in sequence. The right branch includes a 3×3 convolutional layer (output channel C / 2, stride 1) and a max pooling layer (2×2, stride 2) in sequence. The outputs of the left and right branches are concatenated and fused through channels, and then passed through a 3×3 convolutional layer to output multi-scale features (output channel C).

[0021] Technical parameters: Input channel C=64 (first layer) or 128 / 256 / 512 (deep layer); All convolutional layers are followed by BatchNorm and SiLU activation functions; This module replaces the standard stride convolutions in layers 2, 4, and 6 of YOLOv11 Backbone. By using a two-branch parallel processing method, the left branch retains global illumination background information (suppressing dust interference), while the right branch enhances local details at the edge of coal gangue. The two features are complementary and fused in the channel domain, resulting in a 42% reduction in information entropy loss.

[0022] The multi-scale fusion unit uses one SPPF (Fast Spatial Pyramid Pooling) module to fuse local features of different scales through multi-scale pooling operations.

[0023] The Backbone network contains four feature output nodes, each corresponding to a feature map at a different scale, providing multi-scale basic features to the subsequent Neck module and supporting multi-scale feature fusion and object detection tasks.

[0024] The neck network's FPN / PAN structure includes a stacked attention-enhanced residual module (C2PSA-EDFFN) to replace the standard C3k2 module, such as... Figure 4As shown, the C2PSA-EDFFN module consists of two parts: a parallel attention branch and an enhanced feedforward network. The parallel attention branch is used to perform spatial attention and channel attention calculations on the input feature map and then fuse them. Its output is added to the input feature map through a residual connection.

[0025] The module input is a feature map with a size of 80×80 and 256 channels. The overall structure adopts a residual connection design, and the specific implementation method is as follows: (1) Attention branching The input feature map is first processed by 1×1 standard convolution to reduce the number of channels from 256 to 128, thereby reducing the amount of subsequent computation.

[0026] The core parallel PSA attention mechanism then proceeds, simultaneously performing feature recalibration in both spatial and channel dimensions. In the spatial attention path, a 7×7 large-window average pooling is applied to the 128-channel feature map to aggregate local contextual information. This is followed by 1×1 convolution to reduce the dimensionality to a single channel. Finally, a spatial weight map in the 0-1 range is generated using the Sigmoid function to emphasize the coal gangue target region and suppress background noise from the conveyor belt. In the channel attention path, global average pooling is performed on the feature map to obtain a 128-dimensional vector. This vector is then mapped sequentially through fully connected layers (128→16→128) to achieve information exchange between channels. A Sigmoid activation is then used to generate channel weight vectors, automatically identifying feature channels that significantly contribute to coal gangue classification. The two attention outputs are fused through element-wise multiplication, enabling adaptive weighting of the feature map in both spatial location and channel dimension.

[0027] Finally, a 1×1 convolution is used to restore the number of channels to 256, matching the size of the input feature map.

[0028] (2) Residual connection: To alleviate gradient vanishing and preserve the original feature information, a direct skip branch is set up to perform element-wise addition of the unprocessed input feature map with the output of the attention branch, thereby realizing residual mapping. This design ensures that the network can still maintain its basic feature transfer capability even if the attention mechanism fails.

[0029] (3) EDFFN feedforward network processing: The fused feature map enters the enhanced feedforward network and passes through three convolutional layers in sequence to achieve nonlinear transformation.

[0030] First, a 1×1 convolution expands the channels to 512, enhancing feature representation. Then, a 3×3 depthwise separable convolution is used for efficient feature extraction while maintaining spatial resolution. This convolution decomposes the standard convolution into depthwise and pointwise convolutions, reducing the number of parameters to 1 / 9 of the traditional convolution. Finally, a 1×1 convolution compresses the channels back to 256, restoring the original dimensions. The output feature map maintains a size of 80×80×256 and serves as the input to the next network layer.

[0031] Technical parameters: PSA module compression ratio r=16; Depth-separable convolution kernel size 3×3, stride 1; It is applied to the FPN / PAN structure of the Neck network, replacing the original C3k2 module.

[0032] PSA dynamically assigns weights to coal gangue regions (the peak of the weight map appears at the center of the target), and EDFFN enhances the nonlinear expression capability through channel expansion-compression, amplifying the difference features between coal gangue and coal by 3.2 times.

[0033] The detection head is a P2-P5 lightweight detection head, with the structure as follows: Figure 5 As shown, the four feature maps (P2, P3, P4, and P5) from the neck network are received respectively. After the feature maps from each layer are fused through an adaptive fusion layer, they are input into the bounding box regression branch, the class prediction branch, and the confidence prediction branch for decoupled output. Specifically: Input: 4-level feature maps P2 (160×160), P3 (80×80), P4 (40×40), P5 (20×20); Unified channel: Each layer is reduced to 128 channels through 1×1 convolution; Adaptive fusion: P2 layer: Upsample by 2 times → Add to P3; P3 layer: After merging with P2, downsample by 2 times → add with P4; P4 layer: After merging with P3, downsample by 2 times → add with P5; After each layer is fused, a 3×3 depth-separable convolution (with shared weights) is applied. Decoupling output: Branch 1: 3×3 convolution → 1×1 convolution → bounding box regression (4 channels); Branch 2: 3×3 convolution → 1×1 convolution → category prediction (2 channels: coal / gangue); Branch 3: 3×3 convolution → 1×1 convolution → confidence prediction (1 channel).

[0034] Technical parameters: Number of depthwise separable convolutional groups = 128 All use BN+SiLU Output anchor frame dimensions: P2 layer corresponds to 4×8 pixels, P3 layer to 8×16 pixels, P4 layer to 16×32 pixels, and P5 layer to 32×64 pixels. The shallow P2 layer (160×160) directly detects extremely small targets smaller than 8×8, avoiding semantic dilution in deep layers; the shared convolutional kernel reduces the number of parameters by 0.15M and improves inference speed by 18%.

[0035] The improved system of this invention was compared with the standard YOLOv11 model based on the same dataset, and the results are as follows: Figure 6 As shown, the improved model of this invention rises extremely steeply in the early stage of training, indicating that the model can quickly learn effective features from the data and has a high convergence efficiency. As the number of training rounds increases, it gradually enters a stable stage with minimal curve fluctuations and no obvious performance oscillations, indicating that the model training process is stable and more robust to changes in data distribution. Finally, with a smaller number of parameters, the mAP50 is close to 0.8, and the detection accuracy is significantly better than the standard YOLOv11 model.

[0036] Example 2: Based on the conveyor belt coal gangue identification system of Example 1, this example provides a coal mine conveyor belt coal gangue identification method based on the improved YOLOv11, including: Step 1: Construction and preprocessing of a coal mine scenario-specific dataset, specifically as follows: a. First, a high-quality dataset is constructed specifically for the unique working conditions of coal mine conveyor belts to ensure the relevance and generalization ability of the model training. In practical implementation, such as... Figure 1 As shown, an industrial-grade high-definition camera (recommended: Basler acA2500-20gm 5-megapixel global shutter camera with an 8mm industrial lens) is installed 1.5 meters directly above the conveyor belt in the coal mine. The camera then collects real-time video streams of the coal flow via a gigabit network port. The video capture should cover different production shifts, different time periods (morning, noon, and evening shifts), and different coal quality conditions to fully encompass factors such as changes in lighting, fluctuations in dust concentration, and equipment vibration under actual working conditions.

[0037] b. Frame images were uniformly extracted from the acquired video stream at time intervals. After manual screening, 7342 positive sample images containing coal gangue targets were retained, with the original resolution maintained at 2592×2048 pixels to ensure detail integrity. Subsequently, the LabelImg professional annotation tool was used to accurately annotate all images. The annotation process strictly followed the rule of completely surrounding the target and reserving a 5-pixel edge, clearly distinguishing the targets into two categories: "coal" and "gangue," with label codes of 0 and 1, respectively.

[0038] c. To improve the model's adaptability to harsh environments, a dedicated data augmentation process was designed: random brightness adjustment was used to simulate low-light conditions in the well (brightness coefficient range was set from 0.6 to 1.4), Gaussian noise with a root mean square error σ=0.01 was added to simulate dust particle interference, random scaling factors from 0.8 to 1.2 were used to simulate changes in detection distance, and a four-image random stitching Mosaic enhancement strategy was implemented to enrich background diversity.

[0039] d. All enhanced images are uniformly scaled to a resolution of 640×640 pixels, normalized, and then converted into the RGB tensor format required for model training. Finally, they are divided into training, validation, and test sets in a 7:2:1 ratio to ensure the rationality of data distribution and the objectivity of evaluation.

[0040] Step Two: Improved YOLOv11 Model Training and Optimization: The real-time image is input into the improved YOLOv11 algorithm model for processing. The processing procedure of the improved YOLOv11 algorithm model includes: (1) Adaptive downsampling: The input feature map is processed by at least one adaptive downsampling module. The adaptive downsampling module performs the left branch processing of average pooling and 1×1 convolution and the right branch processing of 3×3 convolution and max pooling on the input feature map in parallel, and then concatenates the results of the two branches in the channel dimension and outputs them by 3×3 convolution. (2) Feature enhancement: In the feature fusion process of the Neck network, the feature map is enhanced by the C2PSA-EDFFN feature enhancement module. The C2PSA-EDFFN feature enhancement module first calculates spatial attention and channel attention through parallel attention branches and fuses them, then adds them to the input feature map through residual connections, and then performs nonlinear transformation through an enhanced feedforward network containing depthwise separable convolutions. (3) Multi-scale detection: The P2-P5 lightweight detection head is used to adaptively fuse the four feature maps of P2, P3, P4 and P5 output by the Neck network, and the bounding box position, target category and confidence are decoupled and predicted for the fused features respectively. Step 3: Post-process the detection results output by the improved YOLOv11 algorithm model and output them.

[0041] After completing the data preparation, the core model training phase begins. The training environment is based on the Ubuntu 20.04 operating system, using the Python 3.8 programming language and the PyTorch 2.0.0 deep learning framework, along with the CUDA 11.8 toolkit for GPU-accelerated computing.

[0042] During model initialization, the pre-trained weights provided by the official YOLO11n.pt were loaded as the starting point, and the parameters of the first three convolutional network layers were strategically frozen to retain general feature extraction capabilities and avoid overfitting on specific datasets. For training parameters, the SGD stochastic gradient descent optimizer was selected, with a momentum factor of 0.937 and a weight decay coefficient of 0.0005 to control model complexity. The initial learning rate was set to 0.001 and dynamically adjusted using a cosine annealing strategy, decreasing to a minimum of 0.0001 for fine-grained convergence. The training batch size was set to 32 based on hardware conditions (on an NVIDIA RTX 3090 with 24GB of VRAM), and the entire training process lasted 50 rounds, with 460 iterations per round. The loss function was designed as a linear combination of CIoU Loss and Focal Loss, where the Focal Loss parameters α=0.25 and γ=2, effectively alleviating the coal gangue class imbalance problem. During training, the loss is monitored in real time. The training loss is recorded and a PR curve is plotted every 10 batches. If the mAP50 metric on the validation set shows no improvement for three consecutive rounds, an early termination mechanism is triggered to prevent ineffective training. After the model training converges, it is first exported as an intermediate model in the ONNX open format. Then, NVIDIA TensorRT 8.5 inference optimization tool is used for FP16 precision quantization and layer fusion optimization. Finally, the model is compressed to 14.7MB, and the inference speed is improved by more than 3 times, laying the foundation for edge deployment.

[0043] Step 3: Edge Deployment and Real-time Detection Applications After model optimization, it is deployed to edge computing devices for industrial applications. The TensorRT engine file is burned onto an NVIDIA Jetson Nano 4GB development board, which is connected to a front-end industrial camera via a USB 3.0 high-speed interface to build a complete embedded inspection system. After system startup, the camera continuously acquires images at a rate of 30 frames per second, using DMA (Direct Memory Access) technology to transfer image data to GPU memory with zero copies, avoiding CPU bottlenecks. Image preprocessing, including scaling and normalization, is all executed in parallel on the Jetson Nano's CUDA cores, with a single frame processing time of only 8 milliseconds.

[0044] The preprocessed image tensors are input into the TensorRT engine, where end-to-end inference is performed on the GPU using FP16 computational precision, with a single inference time controlled within 32 milliseconds. The inference output includes bounding box coordinates, class probabilities, and confidence information. In the post-processing stage, the Non-Maximum Suppression (NMS) algorithm (IoU threshold set to 0.45) filters overlapping detection boxes, retaining only high-quality results with a confidence level higher than 0.5. The final detection results are displayed in real-time on a 5-inch LCD screen, with coal targets marked with green rectangles and gangue targets highlighted with red rectangles. The system also displays the class label, confidence value, and estimated gangue size. All raw images and detection logs are automatically saved to a 128GB TF card for easy post-processing and quality analysis. This deployment achieves a stable processing frame rate of 25 FPS on a Jetson Nano, with a total power consumption of less than 10 watts, fully meeting the engineering requirements for long-term continuous operation in underground coal mines.

[0045] The above description is only a preferred embodiment of the present invention and does not limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A coal mine conveyor belt coal gangue identification system based on an improved YOLOv11, characterized in that: Includes an image acquisition unit and an edge computing processing unit; The image acquisition unit is used to acquire real-time images of the coal mine conveyor belt; The edge computing processing unit loads and runs an improved YOLOv11 algorithm model to perform real-time coal gangue detection on the conveyor belt image acquired by the image acquisition unit; the improved YOLOv11 algorithm model includes a backbone network, a neck network, and a detection head connected in sequence. The backbone network includes at least one adaptive downsampling module for downsampling the input feature map; The neck network is used to fuse multi-scale features output by the backbone network to generate a fused feature map. The detection head is a lightweight detection head, P2-P5, which is used to receive four layers of feature maps output from the neck network. After the feature maps of each layer are fused by an adaptive fusion layer, they are input into the bounding box regression branch, the category prediction branch and the confidence prediction branch respectively for decoupled output.

2. The coal gangue identification system for coal conveyor belts based on the improved YOLOv11 as described in claim 1, characterized in that: The backbone network includes an initial convolutional unit, three sets of feature enhancement and downsampling unit groups, a fourth feature enhancement unit, and a multi-scale feature fusion unit. The initial convolutional unit consists of two cascaded convolutional layers, which perform initial convolution operations on the input image to complete the preliminary feature extraction and channel dimension adjustment of the image. The feature enhancement and downsampling unit groups adopt an alternating cascaded structure of feature enhancement modules and downsampling modules. The feature enhancement module is a C3K2 module, which uses residual connections and grouped convolutional structures. The downsampling module is an ADown module, which uses stride convolution to downsample the feature map size.

3. The coal gangue identification system for coal conveyor belts based on the improved YOLOv11 as described in claim 1, characterized in that: The adaptive downsampling module includes a left branch and a right branch. The left branch includes an average pooling layer and a 1×1 convolutional layer in sequence, and the right branch includes a 3×3 convolutional layer and a max pooling layer in sequence. The outputs of the left and right branches are concatenated and fused through channels, and then passed through a 3×3 convolutional layer to output multi-scale features.

4. The coal mine conveyor belt gangue identification system based on improved YOLOv11 according to claim 1, characterized in that: The FPN / PAN structure of the neck network includes a stacked attention-enhanced residual module, which includes a parallel attention branch and an enhanced feedforward network. The parallel attention branch is used to perform spatial attention and channel attention calculations on the input feature map and fuse them. Its output is added to the input feature map through a residual connection. The enhanced feedforward network comprises, in sequence, a first 1×1 convolutional layer, a 3×3 depth-separable convolutional layer, and a second 1×1 convolutional layer.

5. A method for identifying coal gangue in coal mine conveyor belts based on an improved YOLOv11, characterized in that, Based on the coal mine conveyor belt coal gangue identification system based on the improved YOLOv11 as described in any one of claims 1-4, the process includes the following steps: Step 1: Collect and acquire real-time images of coal mine conveyor belts containing coal gangue targets; Step 2: Input the real-time image into the improved YOLOv11 algorithm model for processing. The processing steps of the improved YOLOv11 algorithm model include: (1) Adaptive downsampling: The input feature map is processed by at least one adaptive downsampling module. The adaptive downsampling module performs the left branch processing of average pooling and 1×1 convolution and the right branch processing of 3×3 convolution and max pooling on the input feature map in parallel, and then concatenates the results of the two branches in the channel dimension and outputs them by 3×3 convolution. (2) Feature enhancement: In the feature fusion process of the Neck network, the feature map is enhanced by the C2PSA-EDFFN feature enhancement module. The C2PSA-EDFFN feature enhancement module first calculates spatial attention and channel attention through parallel attention branches and fuses them, then adds them to the input feature map through residual connections, and then performs nonlinear transformation through an enhanced feedforward network containing depthwise separable convolutions. (3) Multi-scale detection: The P2-P5 lightweight detection head is used to adaptively fuse the four feature maps of P2, P3, P4 and P5 output by the Neck network, and the bounding box position, target category and confidence are decoupled and predicted for the fused features respectively. Step 3: Post-process the detection results output by the improved YOLOv11 algorithm model and output them.

6. The method for identifying coal gangue in coal mine conveyor belts based on improved YOLOv11 according to claim 1, characterized in that: Step one involves acquiring real-time images of coal mine conveyor belts containing coal gangue targets, which includes the following steps: (1) Real-time acquisition of video stream of coal flow from conveyor belt, uniform extraction of frame images from the acquired video stream at time intervals, and retention of positive sample images containing coal gangue targets after manual screening; (2) LabelImg professional annotation tool was used to accurately annotate all positive sample images. The annotation process strictly followed the rule of completely surrounding the target and reserving a 5-pixel edge, and the target was clearly distinguished into two categories: coal and gangue. (3) Random brightness adjustment simulates the low illumination conditions in the well, adds Gaussian noise with a mean square error σ=0.01, uses a random scaling factor of 0.8 to 1.2 to simulate the change in detection distance, and implements a Mosaic enhancement strategy of random stitching of four images to enhance the data of the image; (4) The enhanced image is uniformly scaled to a resolution of 640×640 pixels, normalized, and then converted to RGB tensor format.

7. The coal gangue identification method for coal mine conveyor belts based on improved YOLOv11 according to claim 5, characterized in that: In step three, the decoupled prediction information of the output bounding box position, target category and confidence level is filtered by the NMS non-maximum suppression algorithm to filter overlapping detection boxes, and only the output results with a confidence level higher than 0.5 are retained.

Images