A lightweight yolo26 insulator defect monitoring method based on ghostconv

Abstract

The present application relates to the technical field of power system transmission line inspection, in particular to a lightweight YOLO26 insulator defect monitoring method based on ghostconv, comprising the following steps: S1: standardizing and pretreating the 4K high-definition insulator images collected by the unmanned aerial vehicle; through systematic lightweight modification of the YOLO26 model based on GhostConv, replacing the ordinary convolution of the backbone and neck with the GhostConv module, while retaining the key feature fusion and detection structure, the model parameter quantity and the calculation complexity are greatly reduced without reducing the insulator defect detection accuracy, and through the transfer learning and optimization training strategy, fast convergence is realized, and combined with the low-power edge device deployment and the standardized system architecture, a set of lightweight, high-precision, high-real-time and easy-to-deploy insulator defect detection scheme for the transmission line aerial scene is constructed, which effectively solves the problems that the traditional model is difficult to run stably on the edge, the real-time performance is insufficient and the adaptability to complex scenes is poor.

Description

Technical Field

[0001] This invention relates to the field of power system transmission line inspection technology, specifically to a lightweight YOLO26 insulator defect monitoring method based on Ghostconv. Background Technology

[0002] In recent years, with the widespread application of technologies such as autonomous drone inspection and intelligent monitoring in power systems, automated inspection of aerial images of transmission lines has become an important means to ensure the safe operation of the power grid. As a key piece of equipment in transmission lines, insulators are directly affected by defects such as damage, spontaneous explosion, and contamination. Deep learning-based visual inspection methods have become the mainstream technology for intelligent identification of insulator defects due to their advantages such as fast detection speed and strong adaptability. YOLO26, as a representative lightweight, end-to-end target detection algorithm, has the characteristics of no NMS and high inference efficiency, and has good application potential in the field of transmission line defect detection. It is the preferred model for realizing real-time detection at airborne and edge terminals. However, the original YOLO26 model still has the problems of large number of parameters and high computational complexity, making it difficult to deploy directly on low-computing-power platforms such as drones and embedded devices. At the same time, when facing aerial photography scenarios of transmission lines, the model fails to fully adapt to the characteristics of small targets, dense arrangement, and complex backgrounds of insulators, making it difficult to simultaneously achieve detection accuracy and real-time performance in practical engineering applications.

[0003] To address the aforementioned issues, academia and industry have proposed various lightweight improvement methods, such as using depthwise separable convolution, channel pruning, and model quantization to reduce network size. However, these methods generally suffer from shortcomings such as insufficient specificity, reduced feature representation capabilities, and weak engineering adaptability. For example, general lightweight algorithms do not take into account the morphology and distribution characteristics of insulator defects, which can easily lead to missed detection of small targets and a significant decrease in accuracy. Existing improvements often focus on model structure compression, neglecting the balance between computational load, parameter quantity, and real-time performance required for edge deployment, resulting in limited improvement in the efficiency of the model on actual hardware. Secondly, most improvement schemes use a general convolutional structure to replace the existing model, failing to form a dedicated lightweight architecture for power line inspection scenarios. The model redundancy remains high. Furthermore, while ensuring lightweight design, existing methods struggle to maintain the feature extraction capabilities required for insulator defect detection, especially exhibiting insufficient robustness under complex lighting and background conditions. Traditional lightweight strategies often sacrifice detection accuracy, failing to meet the high-precision and high-real-time detection requirements of transmission line insulators. Against this backdrop, there is an urgent need for an improved YOLO26 detection method that is tailored to insulator defect scenarios and balances lightweight design with detection accuracy, thereby enhancing the model's deployment capabilities at edge computing and its engineering practicality. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a lightweight YOLO26 insulator defect monitoring method based on Ghostconv.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A lightweight YOLO26 insulator defect monitoring method based on Ghostconv, comprising the following steps: S1: Standardize and preprocess the 4K high-definition insulator images of power transmission lines captured by drones; S2: Construct a lightweight YOLO26 insulator defect detection model based on GhostConv. The computational cost is reduced by completely replacing the original YOLO26 backbone network and the ordinary convolution of the downsampling convolution of the neck PAN structure with GhostConv modules, while retaining the C2f module, multi-scale feature fusion and end-to-end NMS-Free detection head. S3: Employ a transfer learning strategy, load YOLO26 official pre-trained weights for fine-tuning, set training parameters, and use label smoothing and early stopping strategies. Simultaneously, conduct comparative and ablation experiments using mAP, number of parameters, GFLOPs, and FPS as indicators. S4: Export the model to a common deployment format for ONNX and TensorRT, deploy it to low-computing hardware such as Jetson Nano, Raspberry Pi, and airborne computing units, retain the end-to-end NMS-Free inference process, and output the insulator location, defect category, and confidence level. S5: Connect the detection system to the power grid inspection system, and achieve visualization and automatic alarm by uploading defect data to the power grid intelligent inspection platform in real time. Generate reports based on defect location and category, and extend to the detection of power components such as conductors, fittings and vibration dampers. Form a reproducible, easy-to-train and easy-to-deploy lightweight power vision detection system, and build a full-process intelligent solution from data collection to defect elimination.

[0006] In a preferred embodiment, in step S1, the 4K high-definition insulator images of the transmission line collected by the UAV are standardized and preprocessed. The collected images are initially screened to remove invalid samples that are blurry, severely overexposed, have large areas of obstruction, or have no effective insulator targets. Valid images containing normal insulators and various defects such as spontaneous explosion, breakage, and dirt are selected. The images are standardized by interpolation algorithms, and edge filling is used to fill any insufficient areas. All images are made to meet the same specifications. Adaptive brightness correction, contrast enhancement, defogging and noise reduction are applied to images of different scenes to restore the details of the insulator body and defects. Pixel value normalization is performed to map image pixel values ​​to a fixed range, eliminating pixel distribution differences caused by different acquisition devices and shooting parameters. Data augmentation is performed through geometric transformation and pixel perturbation to expand sample diversity.

[0007] In a preferred embodiment, in step S2, a systematic replacement operation is performed on all ordinary convolutional modules in the model backbone network. The GhostConv module replaces the original standard convolutional units. The GhostConv module generates core features through a small number of basic convolutions. Basic convolution operations are performed on the input data to generate partial core feature maps to capture key basic features in the data. Low-cost linear transformation operations are performed on the generated partial core feature maps to generate redundant feature maps associated with the core features. The core feature maps and redundant feature maps are concatenated along the channel dimension to form a complete feature output. The mechanism of generating redundant features through low-cost linear transformation is then implemented.

[0008] In a preferred embodiment, the downsampling convolutional modules in the neck network PAN structure are lightweighted and adapted by replacing the original ordinary downsampling convolutional units with GhostConv modules. In the neck network PAN structure, all ordinary convolutional units responsible for downsampling operations are identified. Based on the number of input and output channels and the downsampling stride of the replaced convolutional units, the corresponding GhostConv module parameters are configured. The original ordinary downsampling convolutional units are removed from the network structure, and the configured GhostConv modules are inserted. The downsampling stride of the GhostConv modules is set, while keeping the input and output interfaces of the modules completely consistent with the original units. After completing the lightweight transformation of the backbone and neck, the C2f modules in the original network are retained to maintain the efficiency and stability of feature fusion. The multi-scale feature fusion mechanism is retained, enabling the model to simultaneously capture insulator body and defect features of different sizes, while retaining the end-to-end NMS-Free detection head.

[0009] In a preferred embodiment, in step S3, model weights that have been pre-trained on a large-scale general image dataset are loaded as initial training parameters. A transfer learning strategy is adopted to fine-tune the model for the specific data distribution of the insulator defect detection task. The pre-trained weight parameters are loaded as the initial state of the model, preserving its general visual feature extraction capability. The dedicated dataset for insulator defect detection is input into the model. Guided by task-related loss metrics, the model parameters are iteratively updated during training. During the update, some parameters in the model related to the insulator target and defect features are adjusted, and the distribution characteristics of the insulator data are gradually adapted. The model learns the exclusive feature patterns of the insulator body and various defects, so that the model can learn the exclusive feature patterns of the insulator and various defects while retaining its general visual feature extraction capability. Configure standardized training parameters, including key hyperparameters such as training epochs, input image size, batch size, optimizer type, and learning rate, to build a stable and controllable training process. During training, introduce a label smoothing strategy to perform soft label processing on sample labels, and adopt an early stopping strategy to monitor validation set performance indicators. When the validation performance does not improve within a preset number of epochs, terminate training early.

[0010] In a preferred embodiment, after training is completed, multi-dimensional key indicators are selected to evaluate the model performance, including the average precision reflecting detection accuracy, the number of parameters reflecting model size, the number of floating-point operations reflecting computational complexity, and the frame rate reflecting inference speed. Through comparative experiments, the performance of the lightweight model is compared with that of the benchmark model to verify the effect of lightweight modification. Through ablation experiments, the impact of each lightweight module on model performance is verified. Combining the results of comparative and ablation experiments, the balance between accuracy and efficiency of the model after lightweight modification is verified. The specific process for evaluating model performance using multi-dimensional key indicators is as follows: Key evaluation indicators are selected from four dimensions: detection accuracy, model size, computational complexity, and inference speed. A dedicated test dataset for insulator defect detection is used to run the lightweight model and the benchmark model. Measured data for each indicator is collected. The indicators of the lightweight model and the benchmark model are compared horizontally. The impact of lightweight modification on accuracy, size, computational load, and speed is quantitatively analyzed to verify the overall effect of lightweight modification. The impact of each module on model performance is tested individually by replacing lightweight modules. Combined with comparative and ablation experiments, the balance between accuracy and efficiency of the model is analyzed.

[0011] In a preferred embodiment, in step S4, the trained model is converted to an intermediate representation file in ONNX format. Based on the ONNX format, it is converted to the TensorRT optimization format. The model is then optimized for inference through operator fusion and accuracy calibration. The optimized model file is deployed to low-computing-power hardware platforms, including Jetson Nano, Raspberry Pi, and UAV onboard computing units. Adaptive operating environments and inference frameworks are configured for the computing power characteristics of different hardware platforms.

[0012] In a preferred embodiment, during deployment, the original end-to-end NMS-Free inference process of the model is retained, and the inference process is the model outputting detection results. After deployment, inference operations are performed on the input image data. The model performs lightweight feature extraction and detection operations on the input image through a lightweight feature extraction and detection process. Using the replaced lightweight convolution module, the insulator body and defect features are extracted and represented with low computational overhead. Based on the extracted lightweight features, the detection process is executed. Through multi-scale feature interaction and end-to-end detection logic, the insulator target is located and the defect category is determined. The output is a structured detection result, including the location coordinates of the insulator target, the defect category and the corresponding confidence information.

[0013] In a preferred embodiment, in step S5, the deployed insulator defect detection system is connected to the existing power grid inspection system through a standardized data interface. During the operation of the detection system, the output insulator defect data is uploaded to the power grid intelligent inspection platform in real time. After receiving the data, the platform visualizes the defect location, category and confidence information. At the same time, according to the preset defect level and alarm rules, the automatic alarm mechanism is triggered to push defect information to the operation and maintenance personnel. Based on the defect location and category information uploaded to the platform, standardized defect reports are automatically generated in accordance with power grid operation and maintenance specifications. The reports contain key information such as the geographical location of the defect, equipment number, defect type, and severity level. Based on existing insulator defect detection capabilities, the detection scope is expanded to other power transmission line components such as conductors, fittings, and vibration dampers through data expansion and process reuse. Corresponding detection logic is adapted for the characteristics and defect types of different components.

[0014] In a preferred embodiment, a standardized lightweight power vision inspection system is constructed. The existing insulator defect detection process is deconstructed and abstracted, and reusable core modules are extracted. These modules include standardized data processing procedures, lightweight model training methods, low-computing-power hardware deployment specifications, and business interfaces with the power grid inspection system. The modules are standardized and encapsulated to form unified process specifications and interface standards. Through data expansion and process reuse, the insulator detection capabilities are transferred to other power components such as conductors, fittings, and vibration dampers. The universality and reproducibility of the system in different scenarios are verified. The standardized modules are integrated to form a complete lightweight power vision inspection system. Based on the deep integration of the detection system and the power grid inspection system, a full-process intelligent closed-loop solution has been constructed, which includes drone image acquisition, automatic defect detection, real-time data uploading, visual alarms, report generation and maintenance and defect elimination.

[0015] The beneficial effects of this invention are as follows: By systematically lightweighting the YOLO26 model based on GhostConv, replacing the ordinary convolutions in the backbone and neck layers with GhostConv modules while retaining key feature fusion and detection structures, the number of model parameters and computational complexity are significantly reduced without compromising the accuracy of insulator defect detection. Furthermore, rapid convergence is achieved through transfer learning and optimized training strategies. Combined with low-computing-power edge device deployment and a standardized system architecture, a lightweight, high-precision, high-real-time, and easily deployable insulator defect detection solution for aerial photography scenarios of transmission lines is constructed. This solution effectively solves the problems of traditional models being difficult to run stably at the edge, lacking real-time performance, and having poor adaptability to complex scenarios. It improves the model's lightweight effect and inference speed while ensuring detection accuracy, enhancing edge deployment capabilities and engineering practicality, and forming a fully intelligent closed loop from image acquisition to maintenance and defect elimination. It possesses broad applicability to power scenarios and significant practical value. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the insulator defect detection method based on the lightweight improvement YOLO26 using GhostConv, as described in this invention. Figure 2 This is a flowchart of the preprocessing and standardization steps for aerial images of insulators of transmission lines according to the present invention; Figure 3 This is a flowchart illustrating the deployment of the GhostConv lightweight improved YOLO26 insulator defect detection model of the present invention. Figure 4 This is a flowchart of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0020] like Figures 1-4 This embodiment provides a lightweight YOLO26 insulator defect monitoring method based on Ghostconv, comprising the following steps: S1: Standardize and preprocess the 4K high-definition insulator images of power transmission lines captured by drones; S2: Construct a lightweight YOLO26 insulator defect detection model based on GhostConv. The computational cost is reduced by completely replacing the original YOLO26 backbone network and the ordinary convolution of the downsampling convolution of the neck PAN structure with GhostConv modules, while retaining the C2f module, multi-scale feature fusion and end-to-end NMS-Free detection head. S3: Employ a transfer learning strategy, load YOLO26 official pre-trained weights for fine-tuning, set training parameters, and use label smoothing and early stopping strategies. Simultaneously, conduct comparative and ablation experiments using mAP, number of parameters, GFLOPs, and FPS as indicators. S4: Export the model to a common deployment format for ONNX and TensorRT, deploy it to low-computing hardware such as Jetson Nano, Raspberry Pi, and airborne computing units, retain the end-to-end NMS-Free inference process, and output the insulator location, defect category, and confidence level. S5: Connect the detection system to the power grid inspection system, and achieve visualization and automatic alarm by uploading defect data to the power grid intelligent inspection platform in real time. Generate reports based on defect location and category, and extend to the detection of power components such as conductors, fittings and vibration dampers. Form a reproducible, easy-to-train and easy-to-deploy lightweight power vision detection system, and build a full-process intelligent solution from data collection to defect elimination.

[0021] In S1, the 4K high-definition insulator images of transmission lines collected by the UAV are standardized and preprocessed. The collected images are initially screened to remove invalid samples that are blurry, severely overexposed, have large areas of obstruction, or have no effective insulator targets. Valid images containing normal insulators and various defects such as spontaneous explosion, breakage, and dirt are selected. The images are standardized by interpolation algorithms, and edge filling is used to fill any insufficient areas. All images are made to meet the same specifications. Adaptive brightness correction, contrast enhancement, defogging and noise reduction are applied to images of different scenes to restore the details of the insulator body and defects. Pixel value normalization is performed to map image pixel values ​​to a fixed range, eliminating pixel distribution differences caused by different acquisition devices and shooting parameters. Data enhancement is performed through geometric transformation and pixel perturbation to expand sample diversity. It should be noted that the aerial images of insulators captured by drones have high resolution, varying scales, and are subject to lighting and noise interference, requiring standardized preprocessing. The original 4K aerial images were uniformly scaled down to a standard size: ; Pixel value normalization is used to map the image to the [0, 1] interval: ; Where: x: original pixel value; xnorm: normalized value; xmin: minimum pixel value; xmax: maximum pixel value; Random flipping, random cropping, and brightness / contrast perturbation enhance the model's generalization ability.

[0022] In S2, a systematic replacement operation is performed on all ordinary convolutional modules in the backbone network of the model. The GhostConv module is used to replace the original standard convolutional units. The GhostConv module generates core features through a small number of basic convolutions. Basic convolution operations are performed on the input data to generate some core feature maps to capture the key basic features in the data. Low-cost linear transformation operations are performed on the generated core feature maps to generate redundant feature maps associated with the core features. The core feature maps and redundant feature maps are concatenated along the channel dimension to form a complete feature output. The mechanism of generating redundant features through low-cost linear transformation is then used. It should be noted that this invention replaces ordinary convolution with GhostConv to achieve lightweighting while maintaining the stability of the overall network structure. GhostConv module definition: ; ; ; Double features are generated by one main convolution and one depthwise convolution, reducing the number of parameters while maintaining expressive power; Lightweight YOLO26 network architecture: Backbone network: GhostConv is used for downsampling and feature extraction throughout; Neck network: Retain the PAN structure, and replace the downsampling layer with GhostConv; Detection head: The original YOLO26 end-to-end NMS-Free detection head is retained to ensure detection stability; The overall structure can be restructured in a lightweight manner without changing the output logic.

[0023] The downsampling convolutional modules in the neck network PAN structure are lightweighted and adapted by replacing the original ordinary downsampling convolutional units with GhostConv modules. In the neck network PAN structure, all ordinary convolutional units responsible for downsampling operations are identified. Based on the number of input and output channels and downsampling stride of the replaced convolutional units, the corresponding GhostConv module parameters are configured. The original ordinary downsampling convolutional units are removed from the network structure and the configured GhostConv modules are inserted. The downsampling stride of the GhostConv modules is set, while keeping the input and output interfaces of the modules completely consistent with the original units. After completing the lightweight transformation of the backbone and neck, the C2f modules in the original network are retained to maintain the efficiency and stability of feature fusion. The multi-scale feature fusion mechanism is retained so that the model can simultaneously capture insulator body and defect features of different sizes. The end-to-end NMS-Free detection head is also retained. It should be noted that the specific process for identifying all ordinary convolutional units responsible for downsampling operations is as follows: traverse all convolutional units in the neck network PAN structure, read the configuration information of each convolutional unit, and judge based on the stride parameter of the convolutional unit. When the stride of a unit is greater than 1, it is determined to be an ordinary convolutional unit responsible for downsampling operations. Combined with the position and functional description of the unit in the network structure, it is confirmed that it belongs to part of the downsampling path. Non-downsampling convolutional units used only for feature transformation are excluded. Finally, all target units are screened and marked.

[0024] In step S3, the model weights that have been pre-trained on a large-scale general image dataset are loaded as initial training parameters. A transfer learning strategy is adopted to fine-tune the model for the specific data distribution of the insulator defect detection task. The pre-trained general weight parameters are loaded as the initial state of the model, preserving its general visual feature extraction capability. The dedicated dataset for insulator defect detection is input into the model. Guided by the task-related loss index, the model parameters are iteratively updated during the training process. During the update, some parameters in the model related to the insulator target and defect features are adjusted, and the distribution characteristics of the insulator data are gradually adapted. The model learns the exclusive feature patterns of the insulator body and various defects, so that the model can learn the exclusive feature patterns of the insulator and various defects while retaining the general visual feature extraction capability. It should be noted that: the lightweight YOLO26 was trained using transfer learning, and the effect was verified through multiple metrics; Loss function (using the native YOLO26 loss): ; Where: Losscls: classification loss; Lossreg: regression loss; Lossobj: target confidence loss; Training parameter settings: epochs=100, imgsz=640, batch=8, lr0=0.001; Evaluation indicators: Accuracy: mAP0.5, mAP0.5:0.95; Lightweight design: fewer parameters (Params) and fewer computational costs (GFLOPs); Real-time performance: FPS, single-frame inference time; In the model performance verification and effect comparison stage, this invention designed and carried out a complete ablation control experiment. The original YOLO26 target detection model was used as the benchmark control scheme, and the YOLO26 model improved by introducing the GhostConv lightweight structure was used as the technical solution proposed in this invention. Multiple sets of repeated experiments and results statistics were carried out under the same insulator defect detection dataset, the same training parameter configuration, and the same running hardware environment. The differences between the models in key indicators such as parameter quantity, computational complexity, detection accuracy, and inference speed were compared and analyzed. By strictly controlling variables and taking the average value of multiple experiments, the effectiveness and stability of the improvement strategy proposed in this invention were fully verified. It was ensured that after the model was reconstructed with the GhostConv module for lightweighting, the detection accuracy of the model for normal insulator targets and defect targets such as spontaneous explosion, breakage, and contamination would not decrease significantly. At the same time, the model size and computational overhead were significantly reduced. Finally, it was verified that the solution of this invention can achieve a good balance between lightweighting and detection performance, and has outstanding practicality and technical advantages. Configure standardized training parameters, including key hyperparameters such as training epochs, input image size, batch size, optimizer type, and learning rate, to build a stable and controllable training process. During training, introduce a label smoothing strategy to perform soft label processing on sample labels, and adopt an early stopping strategy to monitor validation set performance indicators. When the validation performance does not improve within a preset number of epochs, terminate training early.

[0025] After training, multi-dimensional key indicators were selected to evaluate the model performance, including the mean precision reflecting detection accuracy, the number of parameters reflecting model size, the number of floating-point operations reflecting computational complexity, and the frame rate reflecting inference speed. Through comparative experiments, the performance of the lightweight model was compared with that of the benchmark model to verify the effect of the lightweight transformation. Through ablation experiments, the impact of each lightweight module on the model performance was verified. Combining the results of comparative and ablation experiments, the balance between accuracy and efficiency of the model after lightweight transformation was verified. The specific process for evaluating model performance using multi-dimensional key indicators is as follows: Key evaluation indicators are selected from four dimensions: detection accuracy, model size, computational complexity, and inference speed. A dedicated test dataset for insulator defect detection is used to run the lightweight model and the benchmark model. Measured data for each indicator is collected. The indicators of the lightweight model and the benchmark model are compared horizontally. The impact of lightweight modification on accuracy, size, computational load, and speed is quantitatively analyzed to verify the overall effect of lightweight modification. The impact of each module on model performance is tested individually by replacing lightweight modules. Combined with comparative and ablation experiments, the balance between accuracy and efficiency of the model is analyzed.

[0026] In S4, the trained model is converted to an intermediate representation file in ONNX format. Based on the ONNX format, it is converted to the TensorRT optimization format. The model is optimized for inference through operator fusion and accuracy calibration. The optimized model file is deployed to low-computing-power hardware platforms, including Jetson Nano, Raspberry Pi and UAV onboard computing units. Adaptive operating environments and inference frameworks are configured for the computing power characteristics of different hardware platforms. It should be noted that the lightweight model can be directly implemented for real-time detection on embedded platforms. Reasoning process: ; Output: bounding box, category, confidence score; Model export: Supports ONNX / TensorRT formats: ; Deployment requirements: It can run stably on low-computing-power devices such as Jetson Nano, Raspberry Pi, and airborne chips, enabling real-time airborne detection.

[0027] During deployment, the original end-to-end NMS-Free inference process of the model is retained, and the inference process is based on the model outputting detection results; After deployment, inference operations are performed on the input image data. The model performs lightweight feature extraction and detection operations on the input image through a lightweight feature extraction and detection process. Using the replaced lightweight convolution module, the insulator body and defect features are extracted and represented with low computational overhead. Based on the extracted lightweight features, the detection process is executed. Through multi-scale feature interaction and end-to-end detection logic, the insulator target is located and the defect category is determined. The output is a structured detection result, including the location coordinates of the insulator target, the defect category and the corresponding confidence information.

[0028] In S5, the deployed insulator defect detection system will be connected to the existing power grid inspection system through a standardized data interface. During the operation of the detection system, the output insulator defect data will be uploaded to the power grid intelligent inspection platform in real time. After receiving the data, the platform will visualize the defect location, category and confidence information. At the same time, according to the preset defect level and alarm rules, the automatic alarm mechanism will be triggered to push defect information to the operation and maintenance personnel. Based on the defect location and category information uploaded to the platform, standardized defect reports are automatically generated in accordance with power grid operation and maintenance specifications. The reports contain key information such as the geographical location of the defect, equipment number, defect type, and severity level. Based on existing insulator defect detection capabilities, the detection scope is expanded to other power transmission line components such as conductors, fittings, and vibration dampers through data expansion and process reuse. Corresponding detection logic is adapted for the characteristics and defect types of different components. It should be noted that the system outputs standardized test results and connects to the inspection platform. Defect Judgment Rules: ; when A value greater than 0.5 is considered a defective target. It can support the detection of various defects in insulators, such as spontaneous explosion, breakage, and contamination; In practical engineering applications, the method proposed in this invention can form a complete, efficient, and practical intelligent closed-loop process for detecting insulator defects in transmission lines: First, an inspection drone collects aerial images of the transmission line in real time from high altitude. Then, the images are directly input into an onboard edge computing device, which uses the lightweight improved YOLO26 model proposed in this invention to perform end-to-end real-time defect detection, quickly identifying and locating abnormal targets such as insulator self-explosion, damage, and contamination. The key data such as the location, type, and confidence level of the detected defects are uploaded to the power grid intelligent operation and maintenance backend platform in real time through a wireless communication module. The backend system centrally stores, visualizes, and automatically alarms the defect information, presenting the detection results in the form of intuitive charts, spatial positioning, and defect lists. Finally, maintenance personnel make precise defect elimination, hidden danger investigation, and maintenance decisions based on the defect information displayed on the visualized interface. This achieves intelligent operation throughout the entire process, from drone on-site collection, edge-end intelligent identification, data cloud uploading, backend visualized management, to front-line operation and maintenance defect elimination, significantly improving the efficiency of transmission line inspection, reducing labor costs, and improving the safety and reliability of power grid operation, forming an end-to-end intelligent inspection closed loop.

[0029] To construct a standardized lightweight power vision inspection system, the existing insulator defect detection process is deconstructed and abstracted, and reusable core modules are extracted. These modules include standardized data processing procedures, lightweight model training methods, low-computing-power hardware deployment specifications, and business interfaces with the power grid inspection system. The modules are then standardized and encapsulated to form unified process specifications and interface standards. Through data expansion and process reuse, the insulator detection capabilities are transferred to other power components such as conductors, fittings, and vibration dampers. The system's universality and reproducibility in different scenarios are verified. Finally, the standardized modules are integrated to form a complete lightweight power vision inspection system. Based on the deep integration of the detection system and the power grid inspection system, a full-process intelligent closed-loop solution has been constructed, which includes drone image acquisition, automatic defect detection, real-time data uploading, visual alarms, report generation and maintenance and defect elimination.

[0030] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0031] Those skilled in the art will understand that embodiments of the present invention can be provided as 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.

[0032] 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, and 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 computer, 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 illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0033] 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.

[0034] 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.

[0035] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0036] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A lightweight YOLO26 insulator defect monitoring method based on Ghostconv, characterized in that, Includes the following steps: S1: Standardize and preprocess the 4K high-definition insulator images of power transmission lines captured by drones; S2: Construct a lightweight YOLO26 insulator defect detection model based on GhostConv. The computational cost is reduced by completely replacing the original YOLO26 backbone network and the ordinary convolution of the downsampling convolution of the neck PAN structure with GhostConv modules, while retaining the C2f module, multi-scale feature fusion and end-to-end NMS-Free detection head. S3: Employ a transfer learning strategy, load YOLO26 official pre-trained weights for fine-tuning, set training parameters, and use label smoothing and early stopping strategies. Simultaneously, conduct comparative and ablation experiments using mAP, number of parameters, GFLOPs, and FPS as indicators. S4: Export the model to a common deployment format for ONNX and TensorRT, deploy it to low-computing hardware such as Jetson Nano, Raspberry Pi, and airborne computing units, retain the end-to-end NMS-Free inference process, and output the insulator location, defect category, and confidence level. S5: Connect the detection system to the power grid inspection system, and achieve visualization and automatic alarm by uploading defect data to the power grid intelligent inspection platform in real time. Generate reports based on defect location and category, and extend to the detection of power components such as conductors, fittings and vibration dampers. Form a reproducible, easy-to-train and easy-to-deploy lightweight power vision detection system, and build a full-process intelligent solution from data collection to defect elimination.

2. The lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 1, characterized in that, In S1, the 4K high-definition insulator images of transmission lines collected by the UAV are standardized and preprocessed. The collected images are initially screened to remove invalid samples that are blurry, severely overexposed, have large areas of obstruction, or have no effective insulator targets. Valid images containing normal insulators and various defects such as spontaneous explosion, breakage, and dirt are selected. The images are standardized by interpolation algorithms, and edge filling is used to fill any insufficient areas. All images are made to meet the same specifications. Adaptive brightness correction, contrast enhancement, defogging and noise reduction are applied to images of different scenes to restore the details of the insulator body and defects. Pixel value normalization is performed to map image pixel values ​​to a fixed range, eliminating pixel distribution differences caused by different acquisition devices and shooting parameters. Data augmentation is performed through geometric transformation and pixel perturbation to expand sample diversity.

3. The lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 1, characterized in that, In step S2, a systematic replacement operation is performed on all ordinary convolutional modules in the backbone network of the model. The GhostConv module is used to replace the original standard convolutional units. The GhostConv module generates core features through a small number of basic convolutions. Basic convolution operations are performed on the input data to generate partial core feature maps to capture key basic features in the data. Low-cost linear transformation operations are performed on the generated partial core feature maps to generate redundant feature maps associated with the core features. The core feature maps and redundant feature maps are concatenated along the channel dimension to form a complete feature output. The mechanism of generating redundant features through low-cost linear transformation is then implemented.

4. The lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 3, characterized in that, The downsampling convolutional modules in the neck network PAN structure are lightweighted and adapted by replacing the original ordinary downsampling convolutional units with GhostConv modules. In the neck network PAN structure, all ordinary convolutional units responsible for downsampling operations are identified. Based on the number of input and output channels and downsampling stride of the replaced convolutional units, the corresponding GhostConv module parameters are configured. The original ordinary downsampling convolutional units are removed from the network structure and the configured GhostConv modules are inserted. The downsampling stride of the GhostConv modules is set, while keeping the input and output interfaces of the modules completely consistent with the original units. After completing the lightweight transformation of the backbone and neck, the C2f modules in the original network are retained to maintain the efficiency and stability of feature fusion. The multi-scale feature fusion mechanism is retained, enabling the model to simultaneously capture insulator body and defect features of different sizes. The end-to-end NMS-Free detection head is also retained.

5. A lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 1, characterized in that, In step S3, the model weights that have been pre-trained on a large-scale general image dataset are loaded as initial training parameters. A transfer learning strategy is adopted to fine-tune the model for the specific data distribution of the insulator defect detection task. The pre-trained general weight parameters are loaded as the initial state of the model, preserving its general visual feature extraction capability. The dedicated dataset for insulator defect detection is input into the model. Guided by the task-related loss index, the model parameters are iteratively updated during the training process. During the update, some parameters in the model related to the insulator target and defect features are adjusted, and the distribution characteristics of the insulator data are gradually adapted. The model learns the exclusive feature patterns of the insulator body and various defects, so that the model can learn the exclusive feature patterns of the insulator and various defects while retaining the general visual feature extraction capability. Configure standardized training parameters, including key hyperparameters such as training epochs, input image size, batch size, optimizer type, and learning rate, to build a stable and controllable training process. During training, introduce a label smoothing strategy to perform soft label processing on sample labels, and adopt an early stopping strategy to monitor validation set performance indicators. When the validation performance does not improve within a preset number of epochs, terminate training early.

6. A lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 5, characterized in that, After training, multi-dimensional key indicators were selected to evaluate the model performance, including the mean precision reflecting detection accuracy, the number of parameters reflecting model size, the number of floating-point operations reflecting computational complexity, and the frame rate reflecting inference speed. Through comparative experiments, the performance of the lightweight model was compared with that of the benchmark model to verify the effect of the lightweight transformation. Through ablation experiments, the impact of each lightweight module on the model performance was verified. Combining the results of comparative and ablation experiments, the balance between accuracy and efficiency of the model after lightweight transformation was verified. The specific process for evaluating model performance using multi-dimensional key indicators is as follows: Key evaluation indicators are selected from four dimensions: detection accuracy, model size, computational complexity, and inference speed. A dedicated test dataset for insulator defect detection is used to run the lightweight model and the benchmark model. Measured data for each indicator is collected. The indicators of the lightweight model and the benchmark model are compared horizontally. The impact of lightweight modification on accuracy, size, computational load, and speed is quantitatively analyzed to verify the overall effect of lightweight modification. The impact of each module on model performance is tested individually by replacing lightweight modules. Combined with comparative and ablation experiments, the balance between accuracy and efficiency of the model is analyzed.

7. A lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 1, characterized in that, In step S4, the trained model is converted to an intermediate representation file in ONNX format. Based on the ONNX format, it is converted to the TensorRT optimization format. The model is then optimized for inference through operator fusion and accuracy calibration. The optimized model file is deployed to low-computing-power hardware platforms, including Jetson Nano, Raspberry Pi, and UAV onboard computing units. Adaptive operating environments and inference frameworks are configured for the computing power characteristics of different hardware platforms.

8. A lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 7, characterized in that, During deployment, the original end-to-end NMS-Free inference process of the model is retained, and the inference process is based on the model outputting detection results; After deployment, inference operations are performed on the input image data. The model performs lightweight feature extraction and detection operations on the input image through a lightweight feature extraction and detection process. Using the replaced lightweight convolution module, the insulator body and defect features are extracted and represented with low computational overhead. Based on the extracted lightweight features, the detection process is executed. Through multi-scale feature interaction and end-to-end detection logic, the insulator target is located and the defect category is determined. The output is a structured detection result, including the location coordinates of the insulator target, the defect category and the corresponding confidence information.

9. A lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 1, characterized in that, In S5, the deployed insulator defect detection system will be connected to the existing power grid inspection system through a standardized data interface. During the operation of the detection system, the output insulator defect data will be uploaded to the power grid intelligent inspection platform in real time. After receiving the data, the platform will visualize the defect location, category and confidence information. At the same time, according to the preset defect level and alarm rules, the automatic alarm mechanism will be triggered to push defect information to the operation and maintenance personnel. Based on the defect location and category information uploaded to the platform, standardized defect reports are automatically generated in accordance with power grid operation and maintenance specifications. The reports contain key information such as the geographical location of the defect, equipment number, defect type, and severity level. Based on existing insulator defect detection capabilities, the detection scope is expanded to other power transmission line components such as conductors, fittings, and vibration dampers through data expansion and process reuse. Corresponding detection logic is adapted for the characteristics and defect types of different components.

10. A lightweight YOLO26 insulator defect monitoring method based on Ghostconv according to claim 9, characterized in that, To construct a standardized lightweight power vision inspection system, the existing insulator defect detection process is deconstructed and abstracted, and reusable core modules are extracted. These modules include standardized data processing procedures, lightweight model training methods, low-computing-power hardware deployment specifications, and business interfaces with the power grid inspection system. The modules are then standardized and encapsulated to form unified process specifications and interface standards. Through data expansion and process reuse, the insulator detection capabilities are transferred to other power components such as conductors, fittings, and vibration dampers. The system's universality and reproducibility in different scenarios are verified. Finally, the standardized modules are integrated to form a complete lightweight power vision inspection system. Based on the deep integration of the detection system and the power grid inspection system, a full-process intelligent closed-loop solution has been constructed, which includes drone image acquisition, automatic defect detection, real-time data uploading, visual alarms, report generation and maintenance and defect elimination.

Images