Power transmission line defect detection method based on improved Yolov11n model

Abstract

The invention discloses a power transmission line defect detection method based on an improved Yolovlln model. The method comprises the following steps: firstly, inputting a defect image training data set into an improved Yolovlln network model for training to obtain a target detection model; the camera collects real-time video image information and sends the real-time video image information to the defect detection unit; and finally, obtaining a defect category and position identification result by a target detection module in the defect detection unit. The improvement method of the Yolovlln is as follows: a loss function CIoU is replaced by ShapeIoU; an optimizer SGD is replaced by AdamW; the method comprises the following steps: replacing a C3k2 module with a C3k2-H module for a structure in a backbone network; a PSA attention mechanism in the C2PSA is replaced with a BRA attention mechanism; an original Yolo detection head is replaced by a task dynamic alignment detection head, so that the recognition capability of the model on a multi-scale target is improved. The defect detection method adopted by the invention can accurately and efficiently detect various abnormal defects in the power transmission line, and has relatively high practicability and application prospect.

Description

Technical Field

[0001] The present invention relates to the field of target detection, and in particular to a method for detecting transmission line defects based on an improved Yolov11n model. Background Art

[0002] With the rapid development of artificial intelligence (AI), people are hoping for more engaging interactions between AI and users, providing a better user experience. Compared to traditional manual inspection methods, which are time-consuming, labor-intensive, and inaccurate, deep learning has achieved significant breakthroughs in computer vision in recent years and is widely used in defect detection tasks. Deep learning models can be trained with large amounts of labeled data, learning rich feature representations and accurately detecting and classifying defects.

[0003] Defects in transmission lines can easily disrupt the normal operation of the power system, potentially causing safety incidents such as power outages and equipment damage. They also increase the difficulty and cost of emergency response, resulting in significant economic losses. To ensure stable grid operation, regular inspection and maintenance of transmission lines must be strengthened to ensure early detection and resolution. Traditional manual inspections alone are inefficient and unsafe, necessitating the introduction of more intelligent and automated detection technologies to improve inspection quality and safety. Summary of the Invention

[0004] The object of the present invention is to provide a method for detecting transmission line defects based on an improved Yolov11n model to solve the problems raised in the above background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: First, a defect image training dataset is input into an improved Yolov11 network model for training to obtain an object detection model; then, a camera captures real-time video image information and transmits it to a defect detection unit; finally, the object detection module in the defect detection unit obtains defect category and location recognition results. The recognition process is performed on the trained improved Yolov11n network model.

[0006] The improved Yolo11n network model training method specifically includes: inputting the training set images into the improved Yolo11n network after data enhancement operations such as flipping, changing brightness, cropping, translating, and adding noise, changing the loss function to Shape-IoU, and changing the optimizer to AdamW for training, so as to obtain a target detection model.

[0007] The improved YoloV11 network implementation method specifically includes replacing the C3k2 module with the C3k2-HFERB module in the Backbone network structure and replacing the PSA attention mechanism in C2PSA with the BRA attention mechanism. In the Head section, the original Yolo detection head is replaced with a task-dynamic alignment detection head to improve the model's ability to recognize multi-scale objects.

[0008] The transmission line abnormal defect detection scheme specifically includes: taking the collected transmission line defect image as input to the defect detection unit, and processing it with the improved Yolov11n model in the target detection module of the defect detection unit and outputting the defect detection result;

[0009] Compared with the prior art, the present invention has the following advantages: an improved Yolov11n model is used to detect transmission line defects. The improved model uses the C3k2-HFERB module. Compared with the original C3k2 module, it can more accurately restore the texture and detail features of the image by explicitly extracting and enhancing high-frequency information (such as edges and details). At the same time, it combines residual learning to alleviate the gradient vanishing problem, thereby improving the reconstruction quality and efficiency of tasks such as super-resolution and denoising. The BRA attention mechanism combines local attention and global routing mechanisms to achieve accurate modeling of detail information and effective capture of long-distance dependencies, while significantly reducing computational complexity and having higher efficiency and scalability. The task dynamic alignment detection head not only achieves a lightweight design, but also significantly improves multi-scale perception capabilities and task interaction capabilities, thereby achieving higher detection accuracy and better overall performance in classification and positioning tasks. Using Shape-IoU as the loss function improves model accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a schematic diagram of the functional modules of the present invention.

[0011] Figure 2 This is the improved C3k2-HFERB module network flow chart of the present invention.

[0012] Figure 3 This is the improved detection head network flow chart of the present invention.

[0013] Figure 4 This is the improved Yolov11n network structure diagram of the present invention. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0015] The present invention provides a power transmission line defect detection method based on an improved Yolov11n model. The method comprises the following steps: inputting a transmission line defect image training data set into the improved Yolov11n network model for training to obtain a defect detection model; sending the collected transmission line defect image information to a defect detection unit; and performing defect detection on the input image by a target detection module in the defect detection unit, thereby obtaining defect classification and detection results. The recognition process is performed on the trained improved Yolov11n network model.

[0016] The dataset used for training in this study is a collection of images captured by drone inspections. It is a dataset used to benchmark abnormal defect detection methods, with a focus on transmission line defect detection. The dataset contains 6,714 images, including abnormal defect images from seven different categories.

[0017] The improved Yolov11n network model is implemented as follows: an image of 640×640×3 is input. After feature extraction by the improved backbone network, the output feature map is fed into the Neck feature fusion network. The Neck network uses a bottom-up unidirectional feature pyramid network (FPN) to connect shallow features in the backbone network with deep semantic features, enhancing the semantic information of low-level features while maintaining the integrity of spatial information. The image is then input into the improved Yolov11n network for classification, obtaining the classification results of the object detection module.

[0018] The Yolov11 network is improved as follows: for the structure in the Backbone network, the C3k2 module is replaced by the C3k2-HFERB module. HFERB is designed to enhance high-frequency information. It includes a local feature extraction (LFE) branch and a high-frequency enhancement (HFE) branch. For the LFE branch, we use a 3×3 convolution layer and then a GELU activation function to extract local high-frequency features. For the HFE branch, we use a maximum pooling layer to extract high-frequency information from the input feature FHFE. Then, we use a 1×1 convolution layer and then a GELU activation function to enhance the high-frequency features. The outputs of the two branches are then connected and fed into a 1×1 convolution layer to thoroughly fuse the information. The whole process can be expressed as:

[0019] The Yolov11n network is improved by replacing the PSA attention mechanism in C2PSA with the BRA attention mechanism. BRA is a highly efficient attention mechanism that divides input features into multiple regions. It first performs local attention independently within each region to capture detailed information. A dynamic routing mechanism then selectively establishes long-range dependencies between regions to achieve global feature interaction. This balances the model's representational power and computational efficiency, making it particularly suitable for processing high-resolution images and lightweight vision tasks.

[0020] The improvement method of the Yolov11n network is as follows: for the Head detection network, the traditional Yolo detection head is replaced with a TDADH module. Its traditional single-scale prediction structure only predicts from one scale of the feature map, ignoring the contribution of features at other scales, resulting in a decrease in the accuracy of target detection. To address the above problems, a TDADH module is proposed. The shared parameter detection head structure is improved, and the original dual-branch design is replaced with a shared parameterized design to construct a lighter detection head, significantly reducing the number of parameters and reducing computational redundancy. A task alignment structure is customized on the detection head, so that the detection head can learn task interaction features from multiple convolutional layers through a feature extractor, thereby enhancing the interaction ability between tasks. This is conducive to improving the detection effect of transmission line defects.

[0021] The improved Yolov11n network training is implemented as follows: loss function ShapeIoU + optimizer AdamW. The training set is input into the model, the loss update gradient is calculated using the loss function Shape-IoU, and the parameters are updated using the AdamW optimizer.

[0022] The improved Yolov11n network training method is as follows: the loss function is replaced by CIoU with Shape-IoU. The main advantage of Shape-IOU over CIoU is that it more comprehensively measures the shape similarity between the predicted box and the real box in target detection. CIoU focuses on the center point distance, overlapping area and aspect ratio, while Shape-IoU introduces the concept of shape similarity, which can capture the matching degree of irregular targets more carefully, especially in the detection of complex targets (such as non-rectangular or curved shapes). It performs better and avoids the limitations of relying solely on rectangular areas, thereby improving positioning accuracy and robustness. Its specific implementation formula is as follows: L Shhapc-IoU =1-IoU+distance shape +0.5×Ω shape

[0023] The improved Yolov11n network training method uses the AdamW optimizer to update parameters. The AdamW optimizer is an improvement on the Adam optimizer that adds L2 regularization to constrain parameter values. Its implementation is simple, computationally efficient, and memory-efficient. Parameter updates are unaffected by gradient scaling, and hyperparameters typically require minimal or no adjustment. The update step size can be constrained to a rough range, and the learning rate can be automatically adjusted. Hyperparameters are easily interpretable, requiring little to no tuning.

[0024] Table 1 below shows a comparative experiment using Yolov10n, Yolov11n, and the improved Yolov11n provided in this embodiment on a dataset, where mAP@0.5, detection speed, and model size are selected as parameter performance evaluation indicators of the model.

[0025] Table 1: Neural network model mAP@0.5.0.95 Detection speed (FPS) Model size Yolov10n 80.4％ 57 35.36M Yolov11n 82.1％ 62 42.41M Improve Yolov11n 85.2％ 83 37.6M Table 1 shows that the improved Yolov11n has a faster detection speed than the original model due to the addition of the attention mechanism and dynamic task alignment detection head. The model size is reduced by 4.81M compared to Yolov11n, and the average accuracy of mAP@0.5 and 0.95 are both higher than those of Yolov10n and Yolov11n, and the accuracy is improved by 3.1% compared to the Yolov11n model.

[0026] Figure 1 This is a schematic diagram of the functional modules of the present invention, which specifically includes: collecting transmission line defect image data, and inputting the collected data into the defect detection unit; the target detection module in the defect detection unit obtains the defect detection result after processing by the improved Yolov11n model.

[0027] Figure 2 The improved C3k2-H module network flowchart is as follows: For the LFE branch, we use a 3×3 convolutional layer followed by a GELU activation function to extract local high-frequency features. For the HFE branch, we use a max pooling layer to extract high-frequency information from the input features FHFE. We then use a 1×1 convolutional layer followed by a GELU activation function to enhance the high-frequency features. The outputs of the two branches are then concatenated and fed into a 1×1 convolutional layer to thoroughly fuse the information.

[0028] Figure 3The improved detection head network structure diagram of the present invention is as follows: feature extraction is performed on feature maps of three different scales through two shared 3×3 convolutions, and the extracted features are spliced ​​in the channel dimension to generate task interaction features. Then, the interaction features are decomposed into classification features and positioning features through the TAP module. The positioning branch uses deformable convolution (DCNv2) to process the positioning features, and generates the offset and mask required by DCNv2 through the interaction features to enhance the adaptability to target deformation and positioning accuracy. The classification branch generates a classification mask through the interaction features, and multiplies it with the classification features to achieve classification alignment, thereby realizing dynamic feature selection and strengthening the interaction between classification and positioning tasks.

[0029] Figure 4 The invention improves the Yolov11n network structure diagram, specifically: Backbone network, Neck network, and Head output terminal. The Backbone network introduces a high-frequency information extraction mechanism through the C3k2-H operation, effectively enhancing the model's perception of image details (such as edges and textures). Subsequently, C2BRA achieves accurate modeling of detail information and effective capture of long-range dependencies. The Neck network uses a bottom-up unidirectional feature pyramid network (FPN) for connection, fusing shallow features in the backbone network with deep semantic features, enhancing the semantic information of low-level features while maintaining the integrity of spatial information. The Detect output terminal serves as the final detection component, improving the model's ability to recognize multi-scale objects.

[0030] The solution proposed in the present invention can be loaded onto a computer or other programmable data processing device, so that after executing a series of operation steps, part or all of the functions of the solution proposed in this article can be implemented on the computer or other programmable data processing device.

[0031] The above description of the present invention is illustrative only. The specific implementation of the present invention is not limited to the aforementioned methods. Any person skilled in the art can readily conceive of variations and substitutions within the scope of the calculations disclosed herein. Modifications to the dataset, number of commodity categories, weight parameters, and other methods are all within the scope of protection of this application. Therefore, the scope of protection of this application shall be subject to the scope of protection of the claims.

Claims

1. A photovoltaic cell defect detection method based on an improved Yolov11n model, characterized by: The transmission line defect image training dataset is input into the improved Yolov11n network model for training to obtain a defect detection model; the collected transmission line defect image information is sent to the defect detection unit, and the target detection module in the defect detection unit performs defect detection on the input image to obtain defect classification and detection results. The recognition process is performed on the trained improved Yolov11n network model.

2. The method for detecting power line defects based on the improved Yolov11n model according to claim 1, wherein: The improved Yolov11n model training method is to input the training set images into the improved Yolov11n network after data enhancement operations such as flipping, changing brightness, cropping, translating, and adding noise, and then change the loss function to Shape-IoU and the optimizer to AdamW for training to obtain the target detection model.

3. The method for detecting power line defects based on the improved Yolov11n model according to claim 1, characterized in that: For the structure in the Backbone network, the C3k2 module is replaced by the C3k2-H module, which explicitly extracts and enhances high-frequency information (such as edges and details) to more accurately restore the texture and detail features of the image.

4. The method for detecting power line defects based on the improved Yolov11n model according to claim 1, characterized in that: The PSA attention mechanism in C2PSA is replaced by the BRA attention mechanism. The BRA attention mechanism combines local attention and global routing mechanisms to achieve accurate modeling of detailed information and effective capture of long-distance dependencies, while significantly reducing computational complexity and achieving higher efficiency and scalability.

5. The method for detecting power line defects based on the improved Yolov11n model according to claim 1, characterized in that: The original Yolo detection head is replaced with a task dynamic alignment detection head to improve the model's recognition ability for multi-scale targets.

Images