Power transmission line insulator defect detection method
By improving the YOLOv1 framework and combining the LEGM, MLCA modules and the Focaler-EIoU loss function, the problem of efficient identification of insulator defects in transmission lines under complex environments is solved, achieving high-precision detection with a low false negative rate.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-22
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies struggle to efficiently identify defects in transmission line insulators in complex environments, especially small-sized, blurry-outline, and low-contrast defects, and suffer from high rates of missed detection and false detection.
An improved YOLOv1 framework is adopted, introducing a global feature extraction module with local feature embedding (LEGM) and a multi-scale lightweight cross-attention module (MLCA), combined with the Focaler-EIoU comprehensive loss function, to enhance the model's ability to identify insulator defects and improve its localization stability.
It significantly improves the accuracy and robustness of defect detection in transmission line insulators, reduces the rate of missed and false detections, and improves detection efficiency.
Smart Images
Figure CN121746348A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of power transmission line intelligent inspection, in particular to a power transmission line insulator defect detection method. BACKGROUND
[0002] As a key component in power systems, power transmission line insulators play an important role in supporting conductors and preventing arc occurrence. However, their performance and reliability are easily affected by factors such as manufacturing defects, environmental erosion, and operational aging, leading to problems such as insulator damage, shedding, and flashover. Therefore, timely and accurate detection and repair of insulator defects are crucial for ensuring the safe operation of power grids.
[0003] Insulator images captured by drones often face multiple challenges such as complex environments, varying perspectives, and natural background interference, resulting in small-sized, fuzzy-contoured, and low-contrast defect targets. Additionally, multiple scale targets often coexist in images, further increasing the difficulty of detection. Deep learning has shown significant advantages in such tasks, as its multi-layer neural network can automatically extract features without relying on human design, efficiently recognize complex defect patterns, and have good scale and class adaptability.
[0004] Therefore, based on the deep learning framework, the present application selects the classic target detection algorithm YOLO and improves the algorithm for actual difficulties in insulator defect detection, aiming to improve the detection performance in complex scenarios and achieve higher recognition accuracy, lower miss rate, and lower false detection rate. SUMMARY
[0005] In view of the defects in the prior art, the present application aims to construct a power transmission line insulator defect detection method.
[0006] To achieve the above technical purposes, the present application adopts the following technical solutions:
[0007] A power transmission line insulator defect detection method, comprising the following steps:
[0008] Step S01: Data acquisition and construction
[0009] Use the imaging device mounted on the inspection robot or unmanned aerial vehicle device to obtain insulator raw images in different poses, scales, and environmental conditions; use image annotation tools to accurately annotate insulator targets and their damage, shedding, and other defect types, and divide the annotated data set into a training set and a validation set according to a predetermined ratio to ensure the generalization and complex environment adaptability of the model;
[0010] Step S02: Improved model construction
[0011] An insulator defect detection model is built based on the YOLOv11 architecture, with the following improvements:
[0012] (1) Introduce a global feature extraction module (LEGM module) with local feature embedding in the backbone network, and fuse the fine-grained texture information of the insulator through local gradient enhanced convolution and hybrid domain feature enhancement mechanism to improve the gradient response capability of the model to micro-crack type defects;
[0013] (2) Introduce a multi-scale lightweight cross-attention module (MLCA module) into the feature fusion network, and use cross-scale feature interaction and lightweight attention weight adaptive allocation strategy to enhance the model’s ability to identify insulator detachment areas with drastic scale changes.
[0014] (3) Replace the bounding box regression loss with the Focaler-EIoU comprehensive loss function. By introducing a hard sample focusing factor and enhanced cross-union ratio geometric constraints, the stability and convergence performance of the model in defect localization under complex background interference are improved.
[0015] Step S03: Model Training
[0016] The improved YOLOv11 model is trained based on the training set and validation set to obtain the trained transmission line insulator defect detection model.
[0017] Step S04: Defect Detection
[0018] The acquired image of the insulator of the transmission line to be inspected is input into the defect detection model trained in step S03, and the detection results of the target box position and defect type are output to realize the automatic identification and location of defects such as insulator breakage and detachment.
[0019] Further, in step S02, the LEGM module is integrated into the C3k2 module in YOLOv11 to construct the C3k2-LEGM module, and all C3k2 modules in the YOLOv11 network are replaced with C3k2-LEGM modules.
[0020] Furthermore, the specific content of the LEGM module is as follows:
[0021] (1) The aim is to integrate the local feature extraction capability of convolutional networks with the global dependency modeling capability of self-attention mechanisms to provide a more comprehensive feature representation for object detection;
[0022] (2) The theoretical basis of the LEGM module is that the features extracted by the convolutional network contain a lot of local information. By combining the convolutional layer with the self-attention mechanism, local and global features can be obtained at the same time, and the effective fusion of features can be achieved.
[0023] (3) The LEGM module integrates features from different sources, explores the potential relationships between them, and highlights image-related features, thereby helping to protect the detailed information of the image and improve the detection effect.
[0024] Furthermore, in step S02, the MLCA module is integrated into the C2PSA-MLCA module in YOLOv11.
[0025] Furthermore, the MLCA module undergoes two-step pooling, including: (1) converting the input to local pooling. (2) The input is converted into a one-dimensional vector using two branches. The first branch contains global information and the second branch contains local spatial information. After one-dimensional convolution, the original resolution of the two vectors is restored by unpooling. Then the information is fused to achieve a hybrid attention mechanism.
[0026] Furthermore, in step S02, the Focaler-EIoU comprehensive loss function is composed of the Focaler loss function and the EIoU loss function, and its definition is as follows:
[0027]
[0028] In the formula: L EIoU This is the expression for the EIoU loss function, where IoU is the object detection evaluation metric. Focaler This is the expression for the Focaler loss function.
[0029] Furthermore, the proposed EIoU loss function directly penalizes the width and height of the predicted bounding box. It uses width-height loss to directly calculate the difference between the width and height of the predicted bounding box and the ground truth bounding box, thus accelerating convergence. The function is defined as follows:
[0030]
[0031] In the formula: w c and h c ... This represents the Euclidean distance between the center points of the actual border and the predicted border. The squared width difference directly measures the difference between the predicted bounding box and the ground truth bounding box in the width dimension. The squared difference in height measures the regression error in the height dimension.
[0032] Furthermore, the Focaler loss function ensures that high-quality and low-quality anchor boxes receive appropriate attention and adjustment during training, improving the model's performance and robustness in detection tasks. Its definition is as follows:
[0033]
[0034] Where [d, u] ∈ [0, 1], by adjusting the values of d and u, we can make IoU more focused on different regression samples.
[0035] Further, in step S03, the improved YOLOv11 model is trained based on the training set and validation set to obtain a trained transmission line insulator defect detection model. The specific parameters are: 200 training iterations, 16 batches, and 640×640 pixel image size.
[0036] Compared with the prior art, the present invention has the following technical advantages:
[0037] (1) This invention utilizes inspection robots or drones to inspect and photograph insulators, quickly and accurately identifying defects in insulators in transmission lines, which can greatly reduce the workload of staff and improve work efficiency.
[0038] (2) The present invention uses the LEGM module to solve the problem of insufficient feature expression caused by the separation of local features and global features. It is superior to convolutional neural networks, which are only good at extracting local features of images and have limited ability to capture global context information.
[0039] (3) The present invention uses the MLCA module to achieve a balance between performance and complexity, which is better than most channel attention mechanisms that only contain channel feature information and ignore spatial feature information, resulting in poor model representation or target detection performance, and the spatial attention module is often complex and expensive.
[0040] (4) The Focaler-EIoU loss function of this invention can give reasonable attention and adjustment to high-quality and low-quality anchor boxes during training, thereby improving the model’s performance and robustness in detection tasks. This is better than the traditional loss function, which only considers the overlapping area and center point distance of the bounding box regression. As a result, the penalty term for measuring aspect ratio consistency is not the real difference between the width and height of the predicted box and the real box, which will hinder model training.
[0041] (5) The present invention can effectively address the problem of detecting small target defects in complex backgrounds, significantly improve the accuracy and robustness of the model in target localization, effectively improve accuracy, and reduce the rate of missed detection and false detection. Attached Figure Description
[0042] Figure 1 This is a flowchart of the method for detecting defects in transmission line insulators according to the present invention.
[0043] Figure 2 This is a schematic diagram of the overall structure of the model of the present invention.
[0044] Figure 3 This is a schematic diagram of the C3k2-LEGM module of the present invention.
[0045] Figure 4 This is a schematic diagram of the C2PSA_MLCA module of the present invention. Detailed Implementation
[0046] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0047] like Figure 1 As shown, a method for detecting defects in transmission line insulators includes the following steps:
[0048] Step S01: Data Acquisition and Construction
[0049] The original images of insulators in different postures, scales and environmental conditions are obtained by using imaging devices mounted on inspection robots or drones; the insulator targets and their defect types such as damage and detachment are accurately labeled by image annotation tools, and the labeled dataset is divided into training set and validation set according to a preset ratio to ensure the generalization and adaptability of the model to complex environments.
[0050] Step S02: Improve model construction
[0051] An insulator defect detection model based on the YOLOv11 architecture is constructed as follows: Figure 2 As shown, the following improvements are adopted:
[0052] (1) Introduce a global feature extraction module (LEGM module) with local feature embedding in the backbone network, and fuse the fine-grained texture information of the insulator through local gradient enhanced convolution and hybrid domain feature enhancement mechanism to improve the gradient response capability of the model to micro-crack type defects;
[0053] (2) Introduce a multi-scale lightweight cross-attention module (MLCA module) into the feature fusion network, and use cross-scale feature interaction and lightweight attention weight adaptive allocation strategy to enhance the model’s ability to identify insulator detachment areas with drastic scale changes.
[0054] (3) Replace the bounding box regression loss with the Focaler-EIoU comprehensive loss function. By introducing a hard sample focusing factor and enhanced cross-union ratio geometric constraints, the stability and convergence performance of the model in defect localization under complex background interference are improved.
[0055] Step S03: Model Training
[0056] The improved YOLOv11 model was trained based on the training set and validation set to obtain a trained transmission line insulator defect detection model. The specific parameters were: 200 training iterations, 16 batch size, and 640×640 pixel image size.
[0057] Step S04: Defect Detection
[0058] The acquired image of the insulator of the transmission line to be inspected is input into the defect detection model trained in step S03, and the detection results of the target box position and defect type are output to realize the automatic identification and location of defects such as insulator breakage and detachment.
[0059] The LEGM module is integrated into the C3k2 module in YOLOv11 to build the C3k2-LEGM module. The newly built module is as follows: Figure 3 As shown, in C3k2-LEGM, if C3k=True, then Bottleneck-LEGM feature fusion is used; if it is False, then ordinary Bottleneck feature extraction is used.
[0060] The LEGM module is designed to integrate the local feature extraction capabilities of convolutional networks with the global dependency modeling capabilities of self-attention mechanisms, providing a more comprehensive feature representation for object detection.
[0061] The theoretical basis of LEGM lies in the fact that the features extracted by convolutional networks contain a large amount of local information. By combining convolutional layers with self-attention mechanisms, local and global features can be obtained simultaneously, achieving effective feature fusion.
[0062] LEGM helps preserve image details and improve detection performance by fusing features from different sources and uncovering their potential relationships. For example, it takes as input features from the U-Net output after 1×1 convolution, features from the 3×3 convolution, and features from the dilated residual dense block (DRDB) after the depth estimation network (DE), and combines these local and global features for subsequent processing.
[0063] The MLCA module is integrated into the C2PSA_MLCA module in YOLOv11, and the newly constructed module is as follows: Figure 4 As shown, the MLCA module is specifically combined with the PSA module in C2PSA to construct a new PSA-MLCA, which realizes a hybrid attention mechanism and improves the performance of the object detection network.
[0064] The MLCA module specifically improves the performance of the target detection network by integrating channel information and spatial information.
[0065] The MLCA module requires two-step pooling. First, the input is transformed through local pooling. The input is first transformed into a one-dimensional vector to extract local spatial information. Based on this, two branches are used to convert the input into a one-dimensional vector: the first branch contains global information, and the second branch contains local spatial information. After one-dimensional convolution, unpooling is used to restore the original resolution of the two vectors, and then the information is fused to achieve a hybrid attention mechanism.
[0066] The Focaler-EIoU loss function is a combination of the Focaler loss function and the EIOU loss function. Its definition is as follows:
[0067]
[0068] In the formula: L EIoU This is the expression for the EIoU loss function, where IoU is the object detection evaluation metric. Focaler This is the expression for the Focaler loss function.
[0069] The proposed EIOU loss function directly penalizes the width and height of the predicted bounding box. It uses width-height loss to directly calculate the difference between the width and height of the predicted bounding box and the ground truth bounding box, thus accelerating convergence. It is defined as follows:
[0070]
[0071] In the formula: w c and h c ... This represents the Euclidean distance between the center points of the actual border and the predicted border. The squared width difference directly measures the difference between the predicted bounding box and the ground truth bounding box in the width dimension. The squared difference in height measures the regression error in the height dimension.
[0072] The Focaler loss function allows for reasonable attention and adjustment of high-quality and low-quality anchor boxes during training, thereby improving the model's performance and robustness in detection tasks. The definition is as follows:
[0073]
[0074] Where [d, u] ∈ [0, 1], by adjusting the values of d and u, we can make IoU more focused on different regression samples.
[0075] Integrating the MLCA module into YOLOv11 improved the MAP@50 by 2.5% compared to the original model. Further integrating the LEGM module into YOLOv11 improved the MAP@50 by 3.1% compared to the original model. Finally, replacing the CIOU in YOLOv11 with the Focaler-EIoU loss function improved the MAP@50 by 4.7% compared to the original model. As can be seen from the above, the model proposed in this invention reduces the false positive and false negative rates of the original model and improves the detection accuracy of defects in transmission line insulators.
Claims
1. A method for detecting defects in transmission line insulators, characterized in that, Includes the following steps: Step S01: Data Acquisition and Construction The original images of insulators in different postures, scales and environmental conditions are obtained by using imaging devices mounted on inspection robots or drones; the insulator targets and their defect types such as damage and detachment are accurately labeled by image annotation tools, and the labeled dataset is divided into training set and validation set according to a preset ratio to ensure the generalization and adaptability of the model to complex environments. Step S02: Improve model construction An insulator defect detection model is built based on the YOLOv11 architecture, with the following improvements: (1) Introduce the LEGM module into the backbone network, and fuse the fine-grained texture information of the insulator through local gradient enhancement convolution and hybrid domain feature enhancement mechanism to improve the gradient response capability of the model to micro-crack type defects; (2) Introduce the MLCA module into the feature fusion network and use cross-scale feature interaction and lightweight attention weight adaptive allocation strategy to enhance the model’s ability to identify insulator shedding areas with drastic scale changes. (3) Replace the bounding box regression loss with the Focaler-EIoU comprehensive loss function. By introducing a hard sample focusing factor and enhanced cross-union ratio geometric constraints, the stability and convergence performance of the model in defect localization under complex background interference are improved. Step S03: Model Training The improved YOLOv11 model is trained based on the training set and validation set to obtain the trained transmission line insulator defect detection model. Step S04: Defect Detection The acquired image of the insulator of the transmission line to be inspected is input into the defect detection model trained in step S03, and the detection results of the target box position and defect type are output to realize the automatic identification and location of defects such as insulator breakage and detachment.
2. The method for detecting defects in transmission line insulators according to claim 1, characterized in that: In step S02, the LEGM module is integrated into the C3k2 module in YOLOv11 to construct the C3k2-LEGM module, and all C3k2 modules in the YOLOv11 network are replaced with the C3k2-LEGM module.
3. The method for detecting defects in transmission line insulators according to claim 2, characterized in that, The specific content of the LEGM module is as follows: (1) The aim is to integrate the local feature extraction capability of convolutional networks with the global dependency modeling capability of self-attention mechanisms to provide a more comprehensive feature representation for object detection; (2) The theoretical basis of the LEGM module is that the features extracted by the convolutional network contain a lot of local information. By combining the convolutional layer with the self-attention mechanism, local and global features can be obtained at the same time, and the effective fusion of features can be achieved. (3) The LEGM module integrates features from different sources, explores the potential relationships between them, and highlights image-related features, thereby helping to protect the detailed information of the image and improve the detection effect.
4. The method for detecting defects in transmission line insulators according to claim 1, characterized in that: In step S02, the MLCA module is integrated into the C2PSA-MLCA module in YOLOv11.
5. The method for detecting defects in transmission line insulators based on the improved YOLOv11 according to claim 4, characterized in that: The MLCA module performs two-step pooling, including: (1) converting the input to local pooling. (2) The input is converted into a one-dimensional vector using two branches. The first branch contains global information and the second branch contains local spatial information. After one-dimensional convolution, the original resolution of the two vectors is restored by unpooling. Then the information is fused to achieve a hybrid attention mechanism.
6. The method for detecting defects in transmission line insulators according to claim 1, characterized in that: In step S02, the Focaler-EIoU comprehensive loss function is composed of the Focaler loss function and the EIoU loss function, and its definition is as follows: In the formula: L EIoU This is the expression for the EIoU loss function, where IoU is the object detection evaluation metric. Focaler This is the expression for the Focaler loss function.
7. The method for detecting defects in transmission line insulators according to claim 6, characterized in that: The proposed EIoU loss function directly penalizes the width and height of the predicted bounding box. It uses width-height loss to directly calculate the difference between the width and height of the predicted bounding box and the ground truth bounding box, thus accelerating convergence. It is defined as follows: In the formula: w c and h c ... This represents the Euclidean distance between the center points of the actual border and the predicted border. The squared width difference directly measures the difference between the predicted bounding box and the ground truth bounding box in the width dimension. The squared difference in height measures the regression error in the height dimension.
8. The method for detecting defects in transmission line insulators according to claim 6, characterized in that: The Focaler loss function ensures that high-quality and low-quality anchor boxes receive appropriate attention and adjustment during training, improving the model's performance and robustness in detection tasks. Its definition is as follows: Where [d, u] ∈ [0, 1], by adjusting the values of d and u, IoU can be made to focus more on different regression samples.
9. The method for detecting defects in transmission line insulators according to claim 1, characterized in that: In step S03, the improved YOLOv11 model is trained based on the training set and validation set to obtain the trained transmission line insulator defect detection model. The specific parameters are: 200 training iterations, 16 batches, and 640×640 image size.