A belt multi-type tear recognition method and system based on improved YOLOv11

Abstract

The present invention discloses a method and system for identifying multiple types of belt tears based on an improved YOLOv11, and relates to the technical field of belt tear identification. The specific steps of the method are as follows: obtaining belt tear images and performing type annotation to obtain an initial data set; using data enhancement technology to add interference factors to the initial data set to obtain a training data set, training the improved YOLOv11 model with the training data set to obtain a detection model for belt tear detection, collecting actual belt running images in real time and inputting them into the detection model to identify the belt tear type, and also discloses a system based on the above method. The data set is expanded using a data enhancement strategy, and interference factors are added to simulate the actual working environment of the belt running in the coal transportation system of a power plant, thereby enhancing the diversity and robustness of the image data set. The YOLOv11 model is improved, and the generalization ability of the original model is enhanced, and it is adapted to different tear types and different image sizes, with high working efficiency and enhanced accuracy of model training.

Description

Technical Field

[0001] The present invention relates to the technical field of belt tear identification, and in particular to a method and system for identifying multiple types of belt tears based on an improved YOLOv11. Background Art

[0002] Coal conveying systems play a vital role in power generation. As the core equipment of coal conveying systems, belt conveyors, with their simple structure, reliable operation, and convenient maintenance, play a key role in the efficient and continuous transportation of coal and the stability of fuel supply. However, during use, belts may tear or break due to large debris in the raw coal, sharp objects, belt quality issues, uneven material distribution, excessive impact forces, and clogged guide chutes. This not only threatens the stable operation of the coal conveying system, but can also lead to the paralysis of the entire system, resulting in economic losses to the enterprise and posing a threat to the safety of operators. Therefore, scientific management and monitoring of coal conveying systems are necessary.

[0003] There are two primary methods for detecting belt tears: contact and non-contact. Contact methods, including roller, laser, and metal coil detection, are popular due to their low cost and ease of installation, but they are limited in detection range and recognition capabilities. Non-contact detection methods include ultrasonic, X-ray, and machine vision. With technological advancements, machine vision has become the mainstream choice due to its superior detection capabilities. Patent CN119006434A proposes a belt tear detection method and system based on a dual visual state space model. This method inputs the surface image of the belt to be detected into a trained belt tear image defect detection backbone network, which then outputs the target detection result. Patent CN118529439B proposes a belt conveyor belt damage detection method and system based on an improved CenterNet. This invention, based on a deep learning model and utilizing the KDFA-CenterNet target detection loss algorithm, can effectively detect belt tear damage, wear damage, and rupture damage. Patent CN118608767A proposes a belt tear detection method, system, device, and medium based on YOLOv5s. This method uses the YOLOv5s network structure to train preprocessed data to form a belt detection model. The system, device, and medium are used to implement a belt tear detection method based on YOLOv5s. Although the above methods can identify longitudinal belt tears and achieve good detection results in the face of minimal environmental interference, in actual power plant coal transportation systems, belt tears are not only diverse in type, but also face a very complex working environment. Rain, snow, strong light, and smoke can interfere with the identification of belt tears. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for identifying multi-type belt tears based on improved YOLOv11 to solve the above technical problems.

[0005] To achieve the above objectives, the present invention provides a method for identifying multi-type belt tears based on an improved YOLOv11. The specific steps are as follows:

[0006] Step S1: Obtain historical belt image data and perform preprocessing to obtain a belt tear image;

[0007] Step S2: labeling the belt tear images by type to obtain an initial dataset;

[0008] Step S3: Using data enhancement technology to add interference factors related to the belt operating environment to the initial data set to obtain a training data set;

[0009] Step S4: Training an improved YOLOv11 model using the training dataset to obtain a detection model for belt tear detection. The improved YOLOv11 model includes a backbone network embedded with a lightweight multi-scale convolutional module, a neck network using an advanced feature pyramid structure, and a detection head provided with a shared convolutional layer and a detail enhancement layer.

[0010] Step S5: Collect the actual belt running image in real time and input it into the detection model to identify the belt tear type.

[0011] Preferably, in step S1, a video of the belt working is obtained by a camera device, and frame extraction processing is performed to obtain an image of the torn portion of the belt; the obtained belt working video includes an image of the belt torn during operation of a belt conveyor test platform built in a laboratory and an image of the belt torn collected at the coal transportation site of a power plant.

[0012] Preferably, in step S2, Labelimg is used for feature annotation, the tear types include tear and scratch, and the annotated image is saved as a txt text file.

[0013] Preferably, in step S3, interference factors include rain, snow, fog, dust and noise; data enhancement techniques including affine transformation, random cropping and perspective transformation are used, and then over-enhanced images are eliminated through manual screening.

[0014] Preferably, the C3k2 module in the backbone network is embedded with a lightweight multi-scale convolution module, which includes an input module, the input module is connected to a convolution branch for extracting spatial features and an edge branch for capturing edge features of the image, the convolution branch and the edge branch are connected to a feature fusion module, the feature fusion module includes a feature fusion layer and a convolution layer alternately connected in series, and the feature fusion module is connected to an output module.

[0015] Preferably, the input end of the neck network is connected to the output module connected to the feature fusion module, and the neck network includes an advanced feature pyramid structure that uses a coordinate attention module instead of a channel attention module, and the improved advanced feature pyramid structure is used to automatically adapt to different perspectives and changes in the surrounding environment;

[0016] The coordinate attention module includes an input layer, which is connected to the first horizontal average pooling layer and the first vertical average pooling layer. The first horizontal average pooling layer and the second vertical average pooling layer are connected to the feature fusion layer. The feature fusion layer is connected to the first convolutional layer. The first convolutional layer is connected to the normalization layer. The normalization layer is connected to the second horizontal average pooling layer and the second vertical average pooling layer. The second horizontal average pooling layer and the second vertical average pooling layer are respectively connected to the second convolutional layer. The two second convolutional layers are both connected to the Sigmoid activation function layer. The two Sigmoid activation function layers are reweighted with the input layer and output through the output layer after the reweighted operation.

[0017] Preferably, the detection head comprises a shared convolution layer and a detail enhancement layer connected in sequence, the shared convolution layer is connected to an output layer connected to the reweighted operation layer, and the detail enhancement layer is connected to the detection output layer.

[0018] A system based on the above-mentioned improved YOLOv11 belt multi-type tear recognition method includes:

[0019] Image acquisition module, used to obtain belt running images;

[0020] A data processing module, used for processing and annotating the acquired image data;

[0021] Annotation adding module, used to add interference factors according to the actual belt operation conditions;

[0022] The recognition module is embedded with a detection model and is used to identify whether there is belt tearing and the type of belt tearing in the actual belt running image.

[0023] Therefore, the present invention adopts the above-mentioned method and system for identifying multi-type belt tears based on the improved YOLOv11, which has the following beneficial effects:

[0024] (1) The data set is expanded using a data augmentation strategy. By adding rain, snow, and noise to simulate the actual working environment of the belt operation in the coal transportation system of a power plant, the diversity and robustness of the image data set are enhanced.

[0025] (2) The YOLOv11 model is improved. The backbone network is embedded with a lightweight multi-scale convolution module. The multi-scale convolution technology is used to enhance the ability to capture contextual information, thereby improving the accuracy of detection. The neck network of the advanced feature pyramid structure is adopted, and the adaptive pooling technology is combined to reduce the dimension of the feature map, significantly reducing the number of model parameters. The detection head is equipped with a shared convolution layer and detail enhancement layer. The shared convolution layer and detail enhancement layer are used to reduce the computational complexity and can dynamically generate anchor point sizes to adapt to different image sizes and tearing types. The improved YOLOv11 model improves the generalization ability of the original model, adapts to different tearing types and different image sizes, has strong universality and high work efficiency, and significantly enhances the accuracy of model training while maintaining a small memory footprint.

[0026] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a flow chart of a method for identifying multiple types of belt tears based on improved YOLOv11.

[0028] Figure 2 Improved YOLOv11 model diagram for this invention;

[0029] Figure 3 This is a structural diagram of the lightweight multi-scale convolution module of the present invention;

[0030] Figure 4 This is a diagram of the edge branch structure of the present invention;

[0031] Figure 5 This is the structural diagram of the coordinate attention module of the present invention;

[0032] Figure 6 This is the structural diagram of the detection head of the present invention

[0033] Figure 7 This is a comparison graph before and after the accuracy improvement of the present invention;

[0034] Figure 8 This is a comparison curve diagram before and after the average accuracy improvement of the present invention. DETAILED DESCRIPTION

[0035] In the description of the present invention, it should be noted that the terms "upper", "lower", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or the orientations or positional relationships in which the inventive product is usually placed when in use. These are only for the convenience of describing the present invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, they cannot be understood as limitations on the present invention. In the description of the present invention, it should also be noted that, unless otherwise expressly specified and limited, the terms "setting", "installation" and "connection" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be a communication between the internal parts of two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0036] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0037] like Figure 1 As shown in the figure, a belt multi-type tear recognition method based on improved YOLOv11 is shown in the figure. The specific steps are as follows:

[0038] Step S1: Obtain historical belt image data and perform preprocessing to obtain a belt tear image.

[0039] The belt operation video is captured using a camera, frame-by-frame processing is performed, and images of the belt tearing portion are manually screened. The belt operation video captured includes images of belt tearing during operation on a belt conveyor test platform built in the laboratory and images of belt tearing collected at the coal handling site of a power plant.

[0040] Step S2: Label the belt torn images by type to obtain the initial data set. When labeling the belt torn images, the labelimg tool is used to identify and mark the features in the image. Since the belt may produce many small scratches or scratches during operation, these small damages usually do not affect the normal operation of the belt, but their existence may interfere with the model's accurate identification of real tears. Tear types include tears and scratches. All these scratches and scratches that do not affect operation are uniformly classified as "scratch" category, while the tear damage that actually affects operation is marked as "tear". Classification helps the model to distinguish and identify the tearing of the belt more accurately. The divided txt file saves the coordinates of the upper left vertex and the lower right vertex of the target box. The resolution of the labeled image is 640x640x3, and 3 represents the three RGB colors.

[0041] Step S3: Data augmentation techniques are used to add interference factors related to the belt's operating environment to the initial dataset to generate a training dataset. Interference factors include rain, snow, fog, dust, and noise. Data augmentation techniques include affine transformation, random cropping, and perspective transformation. Over-enhanced images are then manually screened to simulate the effects of belts operating in harsh environments. By simulating various real-world distortions and changes, the diversity and robustness of the image dataset are enhanced.

[0042] Step S4: The improved YOLOv11 model is trained using the training dataset to obtain a detection model for belt tear detection. In this embodiment, the computer operating system is Windows 10. The hardware configuration used in the experiment is an AMD Ryzen 9 7950X 16-Core processor and an NVIDIA GeForce RTX 4080 graphics card with 16GB of video memory. Pytorch 2.0.0, CUDA 11.8, and Python 3.9 are used for programming.

[0043] like Figure 2 As shown in the figure, the improved YOLOv11 model includes a backbone network embedded with a lightweight multi-scale convolution module, a neck network with a high-level feature pyramid structure, and a detection head with a shared convolution layer and a detail enhancement layer.

[0044] like Figure 3-Figure 4As shown, the C3k2 module in the backbone network is embedded with a lightweight multi-scale convolutional module. This module replaces the residual block in the original C3k2 module, improving the network's ability to capture various tearing features. The lightweight multi-scale convolutional module includes an input module connected to a convolution branch for extracting spatial features and an edge branch for capturing image edge features. The convolution branch preserves more spatial details. Convolutional neural networks (CNNs) have significant advantages in spatial information processing, but may be inefficient in extracting image edge features. Therefore, an edge branch is introduced to capture image edge features. The edge branch uses Sobel edge detection and standard convolution techniques to effectively identify areas of intensity variation in the image, thereby extracting critical edge information. It can simultaneously capture image structural information and texture details, achieving superior performance in various vision tasks. The convolution branch and the edge branch are connected by a feature fusion module, so that the final features contain both edge and spatial information, thereby more comprehensively describing the image content and improving the efficiency and accuracy of feature extraction. The feature fusion module consists of alternating feature fusion layers and convolutional layers. The fused feature map first undergoes a 1×1 convolution operation to integrate and compress the features and reduce its dimensionality. Another 1×1 convolution operation then maps the compressed feature map back to the output channel number and forms a residual connection with the original input. This not only helps alleviate the vanishing gradient problem in deep network training but also further enhances the model's feature learning capabilities. The feature fusion module is connected to the output module.

[0045] The input of the neck network is connected to the output module of the feature fusion module. The neck network includes an advanced feature pyramid structure that replaces the channel attention module with a coordinate attention module. This enhances the recognition of subtle features in belt tear images, improving the model's performance in tasks such as object detection and semantic segmentation, enabling more accurate target location and identification. To address the interference of belt tear target identification caused by dust, rain, and snow in harsh operating environments, an improved advanced feature pyramid structure automatically adapts to changes in viewpoint and surrounding environment, effectively integrating information about the tear and its surroundings. This enhances the advanced feature fusion pyramid's capabilities in feature selection and fusion, further improving detection accuracy.

[0046] like Figure 5As shown, the coordinate attention module includes an input layer, which is connected to the first horizontal average pooling layer and the first vertical average pooling layer. The first horizontal average pooling layer and the second vertical average pooling layer are connected to the feature fusion layer. The feature fusion layer is connected to the first convolutional layer. The first convolutional layer is connected to the normalization layer. The normalization layer is connected to the second horizontal average pooling layer and the second vertical average pooling layer. The second horizontal average pooling layer and the second vertical average pooling layer are respectively connected to the second convolutional layer. The two second convolutional layers are both connected to the Sigmoid activation function layer. The two Sigmoid activation function layers are reweighted with the input layer and output through the output layer after the reweighted operation.

[0047] There are two main steps to capture the dependencies and location information between channels.

[0048] First, features are collected along one spatial dimension (width or height), and then these features are aggregated along the other dimension. Specifically, the coordinate attention module uses pooling kernels of size (H, 1) and (1, W) to encode each channel along the horizontal and vertical directions of the input respectively. The c-th channel output Z when the height is h c h (h), can be expressed as:

[0049]

[0050] Similarly, the output Z of the cth channel when the width is w c h (h) can be expressed as

[0051]

[0052] Where H represents the height of the input feature map, w represents the width of the input feature map, and X c (h,i) represents the input feature map along the horizontal direction, X c (j,w) represents the input feature map along the vertical direction.

[0053] The feature map with a specific direction is encoded into two feature maps that store position information. Position information refers to the information extracted from the feature map along the horizontal and vertical directions. It is then transformed using a 1×1 convolution as follows:

[0054] f=σ(F1[z h ,z w ])

[0055] g h =σ(F h (f h ))

[0056] g w =σ(Fw (f w ))

[0057] Where [...] represents the concatenation operation connecting the spatial dimensions, F1 represents the convolution, σ represents the Sigmoid activation function, and the spatial information in the horizontal and vertical directions is included in the intermediate feature map f, f h ∈R c / r×h and f w ∈R c / r×h It is two feature maps obtained by splitting f along the horizontal and vertical directions, and the attention weight g h and g w Is through f h and f w Convolution will f h and f w The number of channels of the two feature maps is converted, g h and g w The number of channels is the same as the input feature map.

[0058] Finally, the weight g is multiplied h and g w Applied to the input feature map, the output of the attention module is represented as

[0059]

[0060] where y c (i, j) is the output of the cth channel, x c (i,j) is the input of the cth channel, is the weight of the c-th channel along the horizontal direction, is the weight of the c-th channel along the vertical direction.

[0061] like Figure 6As shown, the detection head consists of a sequentially connected shared convolutional layer and detail enhancement layer. These layers are used to address the interference of scuffs and scratches on the belt caused by coal on torn objects by predicting the object's bounding box and class probability. A decoding step converts the predictions into actual bounding box coordinates, enhancing feature fusion and improving detection accuracy. The use of lightweight convolutional layers and batch normalization layers effectively reduces the model's parameter count and computational complexity, lowering resource usage and making it suitable for real-time detection scenarios. The shared convolutional layer is connected to the output layer connected to the reweighting operation layer, while the detail enhancement layer is connected to the detection output layer. After convolution and deconvolution, the input features are divided into two parts: object bounding box prediction and label class prediction. These two predictions are then fused through feature enhancement. During the training phase, the module directly outputs the processed features. During the recognition phase, a decoding function converts the predicted bounding box into actual coordinates, which are then combined with class probabilities to generate the final detection result. Furthermore, the module can dynamically adjust the anchor point and stride according to different requirements to adapt to diverse detection tasks.

[0062] Step S5: Collect the actual belt running image in real time and input it into the detection model to identify the belt tear type.

[0063] The above-mentioned system based on the improved YOLOv11 belt multi-type tear recognition method includes:

[0064] The image acquisition module is used to obtain belt running images.

[0065] The data processing module is used to process and annotate the acquired image data.

[0066] The annotation adding module is used to add interference factors according to the actual belt running conditions.

[0067] The recognition module is embedded with a detection model and is used to identify whether there is belt tearing and the type of belt tearing in the actual belt running image.

[0068] In order to verify the superiority of the improved YOLOv11 model, simulation tests were performed on the existing YOLOv11 model and the improved YOLOv11 model. The test results are as follows: Figure 7-Figure 8 As shown in the figure, compared with the existing YOLOv11 model, the precision and mAP50 of the improved model in this embodiment are improved by 3.4% and 1.8% respectively, and the weight file size and parameter amount are reduced by 1.6MB and 0.995×10 6 .

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A belt multi-type tear recognition method based on improved YOLOv11, characterized in that: The specific steps are as follows: Step S1: Obtain historical belt image data and perform preprocessing to obtain a belt tear image; Step S2: labeling the belt tear images by type to obtain an initial dataset; Step S3: Using data enhancement technology to add interference factors related to the belt operating environment to the initial data set to obtain a training data set; Step S4: Training an improved YOLOv11 model using the training dataset to obtain a detection model for belt tear detection. The improved YOLOv11 model includes a backbone network embedded with a lightweight multi-scale convolutional module, a neck network using an advanced feature pyramid structure, and a detection head provided with a shared convolutional layer and a detail enhancement layer. The input of the neck network is connected to the output module of the feature fusion module. The neck network includes a high-level feature pyramid structure that uses a coordinate attention module instead of a channel attention module. The improved high-level feature pyramid structure is used to automatically adapt to different viewpoints and changes in the surrounding environment. The coordinate attention module includes an input layer, which is connected to a first horizontal average pooling layer and a first vertical average pooling layer. The first horizontal average pooling layer and the second vertical average pooling layer are connected to a feature fusion layer. The feature fusion layer is connected to a first convolutional layer. The first convolutional layer is connected to a normalization layer. The normalization layer is connected to a second horizontal average pooling layer and a second vertical average pooling layer. The second horizontal average pooling layer and the second vertical average pooling layer are respectively connected to a second convolutional layer. Both second convolutional layers are connected to a Sigmoid activation function layer. Both Sigmoid activation function layers are reweighted with the input layer and output through the output layer after the reweighting operation. Step S5: Collect the actual belt running image in real time and input it into the detection model to identify the belt tear type.

2. The method for identifying multi-type belt tears based on improved YOLOv11 according to claim 1 is characterized in that: In step S1, a video of the belt working is obtained by a camera device, and frame extraction is performed to obtain an image of the torn portion of the belt; the obtained belt working video includes an image of the belt torn during operation of a belt conveyor test platform built in a laboratory and an image of the belt torn collected at the coal transportation site of a power plant.

3. The method for identifying multi-type belt tears based on improved YOLOv11 according to claim 2 is characterized in that: In step S2, Labelimg is used for feature annotation. The tear types include tear and scratch. The annotated image is saved as a txt text file.

4. The method for identifying multi-type belt tears based on improved YOLOv11 according to claim 3 is characterized in that: In step S3, interference factors include rain, snow, fog, dust and noise; data enhancement techniques including affine transformation, random cropping and perspective transformation are used, and then over-enhanced images are eliminated through manual screening.

5. The method for identifying multi-type belt tears based on improved YOLOv11 according to claim 4 is characterized in that: The C3k2 module in the backbone network is embedded with a lightweight multi-scale convolution module, which includes an input module connected to a convolution branch for extracting spatial features and an edge branch for capturing edge features of the image. The convolution branch and the edge branch are connected to a feature fusion module, which includes feature fusion layers and convolution layers alternately connected in series, and the feature fusion module is connected to an output module.

6. The method for identifying multiple types of belt tears based on improved YOLOv11 according to claim 5 is characterized in that: The detection head includes a shared convolutional layer and a detail enhancement layer connected in sequence. The shared convolutional layer is connected to the output layer connected to the reweighted operation layer, and the detail enhancement layer is connected to the detection output layer.

7. A system based on the improved YOLOv11 belt multi-type tear recognition method according to claim 6, characterized in that: include: Image acquisition module, used to obtain belt running images; A data processing module, used for processing and annotating the acquired image data; Annotation adding module, used to add interference factors according to the actual belt operation conditions; The recognition module is embedded with a detection model and is used to identify whether there is belt tearing and the type of belt tearing in the actual belt running image.

Images