Power transmission line insulator defect detection method and system

Abstract

The invention provides a power transmission line insulator defect detection method and system, and belongs to the technical field of power transmission line detection, and the method comprises the steps: obtaining a to-be-recognized power transmission line image; and inputting a to-be-identified power transmission line image into the pre-trained FDR-YOLO network, and processing to obtain the position and category of an insulator defect target. Wherein the FDR-YOLO network takes a YOLO11s model as a basis, a backbone network of an original YOLO11s model is replaced by a Faster Net structure, a DWD (dynamic weight downsampling) module is utilized to replace a convolution module in the neck network, and a RepFusion module is utilized to replace a C3k2 module in the neck network. The algorithm real-time performance and the detection precision are balanced, and a lightweight solution with balanced precision and real-time performance is provided for an electric power inspection scene.

Description

Technical Field

[0001] This invention belongs to the field of power transmission line testing technology, specifically relating to a method and system for detecting defects in power transmission line insulators. Background Technology

[0002] A safe and stable power system is crucial for national economic development and people's social lives. Insulation is a key factor in the operation of the power system, and insulators, as the weakest link in transmission lines, directly affect the quality of power grid operation. Due to the complex and ever-changing natural environment, insulators are highly susceptible to defects such as cracks, breakage, and contamination, which can even lead to insulator breakdown or string breakage, threatening power grid safety. Therefore, timely and accurate detection of insulator defects is of great significance to ensure the safe and stable operation of the power system.

[0003] With the rapid development of deep learning technology, computer vision-based insulator defect detection methods have gradually become a research hotspot. Among them, the YOLO series algorithms have been widely used in the field of target detection due to their advantage of balancing detection speed and accuracy. Although the YOLO series models are known for their efficiency and real-time performance in target detection, and can quickly and accurately complete target detection and localization, they face two major challenges in the task of detecting insulator defects in transmission lines. On the one hand, insulator defect detection usually needs to run on edge devices such as power line inspection drones, but the original YOLO11 network structure has a large number of parameters and high computational complexity, making it difficult to meet the deployment requirements of edge devices. On the other hand, power line inspection tasks have high real-time requirements, and the original network is difficult to balance accuracy and real-time performance on edge devices. Summary of the Invention

[0004] To address the shortcomings of existing YOLO models in balancing real-time performance and detection accuracy, this invention provides a method and system for detecting defects in transmission line insulators.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting defects in transmission line insulators includes the following steps: Acquire images of the transmission line to be identified; The image of the transmission line to be identified is input into a pre-trained FDR-YOLO network to obtain the location and category of the insulator defect target. The FDR-YOLO network is based on the YOLO11s model. The backbone network of the original YOLO11s model is replaced with a FasterNet structure. The dynamic weight downsampling (DWD) module is used to replace the convolutional module in the neck network, and the reparameterized feature fusion module RepFusion is used to replace the C3k2 module in the neck network. The FasterNet structure is used to downsample and extract features from the input image of the transmission line to be identified, resulting in a multi-scale feature map. The multi-scale feature map is then processed by the DWD module to generate dynamic weights. These dynamic weights are used to perform weighted adaptive fusion of the multi-scale features to obtain a defect-enhanced feature map. The defect-enhanced feature map is then input into the RepFusion module for multi-scale feature fusion. The fused features are then processed by the detection head to obtain the location and category of the insulator defect target.

[0006] Preferably, the step of using the FasterNet structure to downsample and extract features from the input transmission line image to be identified to obtain a multi-scale feature map is as follows: First, the input transmission line image to be identified is downsampled by 4 times through the Embedding layer, and features are extracted using the basic FasterNet module; then, the Merging layer composed of 2×2 standard convolutions is used to perform 2 times downsampling and feature extraction using the basic FasterNet module again, and the feature extraction results are normalized and processed by conv1×1 to obtain a multi-scale feature map.

[0007] Preferably, in the RepFusion module, the input features are first expanded through a 1×1 convolution, then divided into two parts. The first part retains the original information, and the second part undergoes a non-linear transformation through RepConv. The two parts of features are then spliced ​​and fused using a dynamic channel allocation strategy, and finally compressed to the target dimension through a 1×1 convolution to obtain the fused features.

[0008] Preferably, the DWD module includes three stages: dynamic weight generation, feature rearrangement, and adaptive fusion; in the dynamic weight generation, spatial attention is used to dynamically allocate feature fusion weights, and in the feature rearrangement, grouped convolution is introduced.

[0009] Preferably, in the DWD module, dynamic weights are generated for the input feature map. Global context information is extracted through an average pooling layer, and then a preliminary weight matrix is ​​formed by compressing the channels through a 1x1 convolution. Next, the feature map is divided into 2x2 local windows through a rearrange operation and expanded into four groups of sub-features along the channel dimension. Then, the weights are normalized along the window dimension using softmax, generating a dynamic weight matrix for each input feature. In the feature rearrangement stage, the input feature map is downsampled through grouped convolution, and its output channels are expanded by 4 times before being decomposed into four groups of sub-features through channel rearrangement. The four groups of sub-features are multiplied element-wise with the corresponding dynamic weight matrix and summed along the window dimension to achieve adaptive feature fusion based on local context, resulting in a feature map for defect enhancement.

[0010] Preferably, images of transmission line operation and maintenance scenarios including various environmental conditions are selected, and the images are labeled with insulator defects using the Labelimg tool to obtain a training dataset. The FDR-YOLO network is then trained using the training dataset to obtain a pre-trained FDR-YOLO network.

[0011] Preferably, the detection accuracy of the pre-trained FDR-YOLO network is evaluated using the mean average accuracy (mAP).

[0012] This invention also provides a defect detection system for transmission line insulators, specifically comprising: The data acquisition module is used to acquire images of the transmission lines to be identified.

[0013] The defect detection module is used to input the image of the transmission line to be identified into a pre-trained FDR-YOLO network to obtain the location and category of the insulator defect target. The FDR-YOLO network is based on the YOLO11s model, replacing the backbone network of the original YOLO11s model with a FasterNet structure, using the Dynamic Weight Downsampling (DWD) module to replace the convolutional module in the neck network, and using the RepFusion reparameterized feature fusion module to replace the C3k2 module in the neck network. The FasterNet structure is used to downsample and extract features from the input image of the transmission line to be identified, resulting in a multi-scale feature map. The multi-scale feature map is then processed by the DWD module to generate dynamic weights. These dynamic weights are used to perform weighted adaptive fusion of the multi-scale features to obtain a defect-enhanced feature map. The defect-enhanced feature map is then input into the RepFusion module for multi-scale feature fusion. The fused features are then processed by the detection head to obtain the location and category of the insulator defect target.

[0014] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps described in the method for detecting defects in transmission line insulators.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute the steps described in the method for detecting defects in transmission line insulators.

[0016] The method for detecting defects in transmission line insulators provided by this invention has the following beneficial effects: This invention is based on the original YOLO11s model, employing a lightweight FasterNet backbone network and utilizing partial convolutions to improve feature extraction efficiency. Secondly, a Dynamic Weight Downsampling (DWD) module is designed, combining grouped convolutions and spatial attention mechanisms to achieve adaptive fusion of multi-scale features. Finally, a reparameterized feature fusion module is proposed, which improves inference efficiency while maintaining accuracy through multi-branch enhancement during training and single-path transformation techniques during inference. Attached Figure Description

[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a structural diagram of the improved YOLO network model in an embodiment of the present invention.

[0019] Figure 2 This is a diagram of the FasterNet network structure in an embodiment of the present invention.

[0020] Figure 3 This is a structural diagram of the DWD module in an embodiment of the present invention.

[0021] Figure 4 This is a structural diagram of the RepFusion module in an embodiment of the present invention.

[0022] Figure 5 This is a partial image of the insulator dataset in an embodiment of the present invention.

[0023] Figure 6 This is a performance comparison chart before and after the improvement in an embodiment of the present invention.

[0024] Figure 7 These are comparison images of insulator defect recognition in some embodiments of the present invention. Detailed Implementation

[0025] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0026] Example This invention provides a method for detecting defects in insulators of power transmission lines, specifically including the following steps: Step 1: Improve the YOLO11s model by replacing the backbone network with a FasterNet structure. A Dynamic Weight Downsampling (DWD) module replaces the convolutional modules in the neck network, and a Reparameterized Feature Fusion (RepFusion) module replaces the C3k2 module in the neck network, resulting in an improved YOLO network. The network structure is as follows: Figure 1 As shown.

[0027] (1) FasterNet backbone network, which performs hierarchical downsampling and feature extraction on the input image as the backbone network, and outputs multi-scale feature maps for subsequent use.

[0028] FasterNet is a high-efficiency, lightweight convolutional neural network architecture. Its core innovation lies in its novel partial convolution (PConv), which extracts spatial features more efficiently by simultaneously reducing redundant computations and memory accesses. This network not only reduces floating-point operations (FLOPs) but also improves floating-point operations per second (FLOPS) by optimizing memory access patterns, thus achieving high inference performance on various hardware platforms while maintaining high accuracy.

[0029] The structure of partial convolution is as follows Figure 2 As shown in the upper right section, this convolution improves computational efficiency by performing convolution operations only on a subset of the input channels. For consecutive or regular memory accesses, it selects the channel of the first or last consecutive cp as a representative of the entire feature map for computation. It assumes the number of input and output channels are the same. The computational cost (FLOPs) of PConv can be expressed as:

[0030] (1); Where h and w are the height and width of the feature map, k is the size of the convolution kernel, and c p This represents the number of channels involved in the partial convolution. In practice, generally... Where c is the number of channels used in a regular convolution. Therefore, PConv's FLOPs are only 1 / 16 of those of a regular convolution. Furthermore, PConv has lower memory access requirements, i.e.:

[0031] (2); For r=1 / 4, its memory access volume is only 1 / 4 of that of a regular Conv.

[0032] like Figure 3As shown, FasterNet has four stages. Before the initial stage, there is an Embedding layer (composed of 4×4 standard convolutions with a stride of 4) to downsample the input image by a factor of 4. Each subsequent stage is preceded by a Merging layer (composed of 2×2 standard convolutions with a stride of 2) to downsample by a factor of 2. Each stage consists of multiple FasterNet basic modules stacked together. Each FasterNet basic module uses partial convolutions (PConv) for efficient feature extraction, followed by two 1×1 pointwise convolutions to form a feature transformation layer. Besides the aforementioned operators, normalization layers and activation functions are also essential for high-performance neural networks. This network adopts a simplified strategy in the use of normalization layers and activation functions, placing them only after the pointwise convolutions in the intermediate layers. This design maintains feature diversity while avoiding the impact of excessive normalization operations on computational efficiency, thus achieving low inference latency while ensuring the model's expressive power.

[0033] (2) Dynamic weight downsampling module DWD is used to generate dynamic weights, and then perform downsampling through group convolution and weighted adaptive fusion of multi-path sub-features to obtain a feature map of "downsampling + emphasis on key regions".

[0034] Traditional downsampling methods, employing fixed-weight fusion strategies, struggle to adapt to the multi-scale characteristics of insulator defect regions. To address this issue, a Dynamic Weighted Downsampling (DWD) module is proposed. This module utilizes spatial attention to dynamically allocate feature fusion weights, enhancing the network's ability to perceive defect regions. Furthermore, group convolution is introduced within the module to reduce computational overhead. Together, these two approaches achieve both a lightweight model and effective fusion of multi-scale features.

[0035] Grouped convolution, such as Figure 3 The lower half shows a convolution operation that reduces the number of parameters and improves computational efficiency by grouping channels. Its core idea is to divide the channels of the input feature map into multiple subgroups, perform convolution operations on each group independently, and then concatenate the results.

[0036] like Figure 4As shown, the workflow of the DWD module can be divided into three stages: dynamic weight generation, feature rearrangement, and adaptive fusion. In the dynamic weight generation stage, the input feature map first extracts global context information through an average pooling layer, then compresses the channels using a 1x1 convolution to form a preliminary weight matrix. Next, a rearrange operation divides the feature map into 2x2 local windows, expands it into four groups of sub-features along the channel dimension, and then normalizes the weights along the window dimension using softmax. The goal of this stage is to generate a dynamic weight matrix for each input feature. Using this weight matrix, the model can automatically adjust the weights of each region based on the spatial information of the input features to adapt to different input features. The dimensionality transformation at this point is: → Where C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map. In the feature rearrangement stage, the input feature map undergoes basic downsampling through grouped convolution, and its output channels are expanded by a factor of 4 before being decomposed into four groups of sub-features through channel rearrangement. The purpose of this stage is to achieve efficient downsampling of the input feature map through grouped convolution and to effectively combine different channels. Grouped convolution effectively reduces computational complexity while preserving the validity of spatial information. The dimensionality transformation at this point is: → Finally, in the adaptive fusion stage, the four sets of sub-features are multiplied element-wise with their corresponding dynamic weights and summed along the window dimension to achieve adaptive feature fusion based on local context. In this stage, the generated dynamic weights are used to weight and fuse the sub-features, effectively strengthening the features of important regions and suppressing unimportant regions. This design uses grouped convolutions to reduce computation and leverages dynamic weights to enhance the contribution of key regions, effectively mitigating the feature degradation problem in traditional downsampling.

[0037] (3) Reparameterized feature fusion module (RepFusion) is used to further integrate the concatenated features. During the training phase, multi-branch enhancement is used, and during the inference phase, the multi-branch is equivalently folded into a single-path convolution to improve speed, outputting more efficient fused features and sending them into the detection head to complete classification and regression.

[0038] To address the issue of high redundancy in intermediate feature maps in traditional CNNs, this module employs inexpensive computation to generate partially redundant feature maps, thereby reducing computational complexity. Furthermore, by abandoning the BottleNeck structure commonly used in the YOLO series, it introduces reparameterized convolutions (RepConv) to construct gradient flow branches, enhancing feature representation capabilities while reducing the number of parameters.

[0039] RepConv as Figure 4As shown in the right half, this convolution, as the core component of the module, effectively balances model performance and computational efficiency through a reparameterized structure design. During training, RepConv enhances feature representation through 3×3 convolutions, 1×1 convolutions, and identity branches. During inference, RepConv converts multiple branches into a single 3×3 convolution through weight fusion, significantly reducing computational latency. This design gives RepConv the advantage of multi-scale feature fusion during training, while maintaining the same computational efficiency as standard convolutions during deployment, greatly reducing the complexity of hardware adaptation.

[0040] The RepFusion module, for example Figure 5 As shown, the input features are first expanded through a 1×1 convolution, then divided into two parts. The first part retains the original information, while the second part undergoes a non-linear transformation using RepConv. Then, a dynamic channel allocation strategy is used to concatenate and fuse the two parts of the features, finally compressing them to the target dimension using a 1×1 convolution. This design maintains high feature expressiveness while optimizing inference efficiency, making it suitable for lightweight model deployment scenarios.

[0041] Step Two: Integrate the insulator dataset using the open-source Chinese transmission line insulator dataset and other public datasets through a screening and manual annotation process. The dataset contains 848 insulator fault samples, covering various environmental conditions in typical transmission line operation and maintenance scenarios. These include different lighting conditions such as sunny and cloudy days; and different spatial backgrounds such as plains, rivers, mountains, and forests, effectively verifying the robustness and generalization ability of the algorithm under real-world conditions. Image annotation uses the Labelimg tool, uniformly labeling insulator defects as the "insulator_defect" category. To achieve objective evaluation of model performance, a random sampling method is used to divide the dataset into training, validation, and test sets in an 8:1:1 ratio. Some images from the insulator dataset are shown below. Figure 5 As shown.

[0042] Step 3: Train the improved YOLO11s network using the training set data. The specific hardware configuration and hyperparameter settings for model training are shown in Table 1.

[0043] Table 1 conditions To evaluate the detection performance of the algorithm, commonly used evaluation metrics in the field of object detection are adopted: mean average precision (mAP) is used to measure the overall detection accuracy; at the same time, the complexity and real-time performance of the model are comprehensively analyzed by considering the number of parameters (M), computational cost (GFLOPs, G), and inference frame rate (FPS, f / s). The specific calculation formula for mAP is shown in Equation (5).

[0044] (3); (4); (5); Where P represents precision, R represents recall, TP represents the number of positive samples correctly detected by the model, FP represents the number of negative samples incorrectly detected by the model, mAP represents average detection precision, and N represents the number of classes.

[0045] The accuracy of the improved YOLO11s network after training was tested using test set data. The accuracy requirements were met, and the FDR-YOLO network was obtained.

[0046] Step 3: Input the image of the transmission line to be identified into the pre-trained FDR-YOLO network to obtain the location and category of the insulator defect target.

[0047] To verify the superiority of FasterNet, the backbone network selected for the insulator defect detection model, different backbone networks such as ConvNeXtV2, LSKNet, MobileNetV4, and ResNet were selected for comparison. The results are shown in Table 2.

[0048] Table 2 Backbone Network Selection The data in Table 2 shows that, compared with other backbone networks, FasterNet maintains high detection accuracy while having fewer parameters and lower computational cost, and also boasts the highest inference speed. These data demonstrate FasterNet's superiority in balancing accuracy and efficiency, and also indicate that FasterNet is suitable as the backbone network architecture for insulator defect detection algorithms.

[0049] To verify the effectiveness of the FasterNet backbone network, DWD module, and RepFusion module introduced into the insulator defect detection model in improving the YOLO11s algorithm, eight ablation trials were performed, and the results are shown in Table 3. "√" indicates that the module was added to the network.

[0050] Table 3. Comparison of ablation methods for different structures introduced into the YOLO11s algorithm As shown in Table 3, when using only FasterNet, the model achieved a 38.3% frame rate improvement while maintaining high detection accuracy, while reducing the number of parameters and computational cost by 20.2% and 26.8%, respectively, validating its dual advantages in model lightweighting and accelerated inference. When using only DWD, although the detection frame rate decreased, the reduction in model parameters and computational cost indicates that this module is suitable for lightweight scenarios. When using only RepFusion, the model reduced both the number of parameters and computational cost while improving the detection frame rate, confirming its unique advantages in reducing computational complexity and optimizing inference efficiency through its reparameterized structure. Using both FasterNet and DWD simultaneously resulted in a 25.5% reduction in parameters and a 27.7% reduction in computational cost, while incurring only a 0.1% loss in accuracy and achieving a 27.6% frame rate improvement, demonstrating excellent overall performance. While using both RepFusion and DWD maintained the detection accuracy of the original model, the lightweighting effect was relatively limited. Using both FasterNet and RepFusion simultaneously achieved a 28.7% reduction in parameters and a 33.3% reduction in computation, but with a 1.1% decrease in accuracy. Finally, employing all three methods simultaneously yielded the best performance, achieving a 33.0% reduction in parameters and a 34.3% reduction in computation, while incurring only a 0.5% loss in accuracy and a 40.7% increase in frame rate, achieving the optimal balance between model lightweighting and performance maintenance. Performance comparison data between the improved model using the three methods and the baseline model are shown below. Figure 6 As shown.

[0051] To further verify the detection performance of FDR-YOLO, it was compared with the current mainstream object detection algorithms, and the results are shown in Table 4.

[0052] Table 4 Performance comparison with different mainstream object detection algorithms As shown in Table 4, FDR-YOLO achieves higher detection accuracy than most models while having lower parameter count and computational cost compared to mainstream YOLO models, and its inference speed is only lower than YOLOv5. While YOLOv5s boasts the highest FPS, its parameter count and computational cost are higher than FDR-YOLO. Although YOLOv7-tiny is close to FDR-YOLO in terms of parameters and computational cost, its detection frame rate is significantly lower. While YOLOv8s and YOLOv9s have slightly higher accuracy than FDR-YOLO, their parameter count and computational cost are far greater. Furthermore, FDR-YOLO outperforms the latest YOLOv12 and classic object detection algorithms such as Tood, Rtmdet, ATSS, and DINO. This indicates that FDR-YOLO, through structural optimization, effectively balances model complexity and inference efficiency while maintaining high detection accuracy, making it particularly suitable for detection scenarios sensitive to real-time performance and computational resources.

[0053] To more intuitively compare the performance improvement of FDR-YOLO over the original model, the network models before and after the improvement are visualized and compared. Some comparison results are shown below. Figure 7 As shown in the figure, the first row is the original image, the second row is the detection result image of the YOLO11s model, and the third row is the detection result image of the FDR-YOLO model. From the figure, it can be seen that YOLO11s has high detection performance for insulator defects. After the lightweight improvement of the original model, the detection performance of the model remains almost unchanged. This indicates that the proposed FDR-YOLO model maintains the high accuracy of the original model while significantly reducing the computational burden, making it suitable for lightweight application scenarios.

[0054] The improved model provided by this invention has the following advantages in handling transmission line insulator testing: To address the challenges of high model complexity and difficult deployment of edge devices in transmission line insulator defect detection tasks, a lightweight improved model, FDR-YOLO, is proposed. This model achieves a balance between accuracy and efficiency through three improvements: (1) FasterNet is used to replace the original backbone network. The feature extraction process is optimized by partial convolution (PConv). This reduces redundant computation while enhancing the ability to express local features, significantly reducing model complexity and computational resource consumption, and improving inference efficiency.

[0055] (2) A Dynamic Weighted Downsampling (DWD) module is proposed, which combines grouped convolution and spatial attention mechanisms to adaptively allocate fusion weights for multi-scale features. This module reduces the computational burden through a lightweight strategy, enhances the perception of key regions of insulator defects, and effectively alleviates the feature degradation problem of traditional downsampling methods in complex backgrounds.

[0056] (3) The RepFusion module is designed to replace the traditional feature fusion structure. It uses reparameterization technology to achieve multi-branch feature enhancement in the training stage and efficient single-path computation in the inference stage. This module improves hardware adaptability and significantly optimizes model deployment efficiency by reducing the number of parameters through gradient flow optimization and structural simplification.

[0057] The results show that FDR-YOLO maintains a high detection accuracy of 95.9% while reducing the number of parameters by 33.0% and the computational cost by 34.3%, and increasing the inference speed to 169.5f / s. This achieves a synergistic optimization of reducing model complexity and improving inference speed, providing a lightweight solution for real-time detection of defects in transmission line insulators.

[0058] The present invention also provides a defect detection system for transmission line insulators, comprising: The data acquisition module is used to acquire images of the transmission lines to be identified.

[0059] The defect detection module is used to input the image of the transmission line to be identified into a pre-trained FDR-YOLO network to obtain the location and category of the insulator defect target. The FDR-YOLO network is based on the YOLO11s model, replacing the backbone network of the original YOLO11s model with a FasterNet structure, using the Dynamic Weight Downsampling (DWD) module to replace the convolutional module in the neck network, and using the RepFusion reparameterized feature fusion module to replace the C3k2 module in the neck network. The FasterNet structure is used to downsample and extract features from the input image of the transmission line to be identified, resulting in a multi-scale feature map. The multi-scale feature map is then processed by the DWD module to generate dynamic weights. These dynamic weights are used to perform weighted adaptive fusion of the multi-scale features to obtain a defect-enhanced feature map. The defect-enhanced feature map is then input into the RepFusion module for multi-scale feature fusion. The fused features are then processed by the detection head to obtain the location and category of the insulator defect target.

[0060] The modules in the aforementioned power transmission line insulator defect detection system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0061] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of a method for detecting defects in transmission line insulators. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0062] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of a method for detecting defects in transmission line insulators. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0063] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0064] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for detecting defects in transmission line insulators, characterized in that, Includes the following steps: Acquire images of the power transmission line to be identified; The image of the transmission line to be identified is input into a pre-trained FDR-YOLO network to obtain the location and category of the insulator defect target. The FDR-YOLO network is based on the YOLO11s model. The backbone network of the original YOLO11s model is replaced with a FasterNet structure. The dynamic weight downsampling (DWD) module is used to replace the convolutional module in the neck network, and the reparameterized feature fusion module RepFusion is used to replace the C3k2 module in the neck network. The FasterNet structure is used to downsample and extract features from the input image of the transmission line to be identified, resulting in a multi-scale feature map. The multi-scale feature map is then processed by the DWD module to generate dynamic weights. These dynamic weights are used to perform weighted adaptive fusion of the multi-scale features to obtain a defect-enhanced feature map. The defect-enhanced feature map is then input into the RepFusion module for multi-scale feature fusion. The fused features are then processed by the detection head to obtain the location and category of the insulator defect target.

2. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The method of using the FasterNet structure to downsample and extract features from the input transmission line image to be identified, and obtain a multi-scale feature map, specifically involves: first, downsampling the input transmission line image to be identified by 4 times through the Embedding layer, and then extracting features using the basic FasterNet module; then, repeating the downsampling by 2 times and feature extraction using the Merging layer composed of 2×2 standard convolutions, and then normalizing and processing the feature extraction results with conv1×1 to obtain a multi-scale feature map.

3. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, In the RepFusion module, the input features are first expanded through a 1×1 convolution, and then divided into two parts. The first part retains the original information, and the second part undergoes a non-linear transformation through RepConv. The two parts of features are then concatenated and fused using a dynamic channel allocation strategy, and finally compressed to the target dimension through a 1×1 convolution to obtain the fused features.

4. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The DWD module comprises three stages: dynamic weight generation, feature rearrangement, and adaptive fusion. In dynamic weight generation, spatial attention is used to dynamically allocate feature fusion weights, and in feature rearrangement, grouped convolution is introduced.

5. The method for detecting defects in transmission line insulators according to claim 4, characterized in that, In the DWD module, dynamic weights are generated for the input feature map. Global context information is extracted through an average pooling layer, and then a preliminary weight matrix is ​​formed by compressing the channels through a 1x1 convolution. Next, the feature map is divided into 2x2 local windows through a rearrange operation and expanded into four groups of sub-features along the channel dimension. Then, the weights are normalized along the window dimension using softmax, generating a dynamic weight matrix for each input feature. In the feature rearrangement stage, the input feature map is downsampled through grouped convolution, and its output channels are expanded by 4 times before being decomposed into four groups of sub-features through channel rearrangement. The four groups of sub-features are multiplied element-wise with the corresponding dynamic weight matrix and summed along the window dimension to achieve adaptive feature fusion based on local context, resulting in a feature map for defect enhancement.

6. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, Images of power transmission line operation and maintenance scenarios under various environmental conditions were selected, and insulator defects were labeled in the images using the Labelimg tool to obtain a training dataset. The FDR-YOLO network was then trained using the training dataset to obtain a pre-trained FDR-YOLO network.

7. The method for detecting defects in transmission line insulators according to claim 1, characterized in that, The detection accuracy of the pre-trained FDR-YOLO network is evaluated using the mean average precision (mAP).

8. A defect detection system for transmission line insulators, characterized in that, include: The data acquisition module is used to acquire images of the transmission lines to be identified. The defect detection module is used to input the image of the transmission line to be identified into a pre-trained FDR-YOLO network to obtain the location and category of the insulator defect target. The FDR-YOLO network is based on the YOLO11s model, replacing the backbone network of the original YOLO11s model with a FasterNet structure, using the Dynamic Weight Downsampling (DWD) module to replace the convolutional module in the neck network, and using the RepFusion reparameterized feature fusion module to replace the C3k2 module in the neck network. The FasterNet structure is used to downsample and extract features from the input image of the transmission line to be identified, resulting in a multi-scale feature map. The multi-scale feature map is then processed by the DWD module to generate dynamic weights. These dynamic weights are used to perform weighted adaptive fusion of the multi-scale features to obtain a defect-enhanced feature map. The defect-enhanced feature map is then input into the RepFusion module for multi-scale feature fusion. The fused features are then processed by the detection head to obtain the location and category of the insulator defect target.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.

Images