A method for quickly identifying appearance defects of an insulating device

Abstract

This invention relates to the field of equipment defect detection, and more particularly to a method for rapid identification of appearance defects in insulating equipment. The method includes the following steps: Using a visible light imaging device mounted on a drone or inspection robot, images of the insulating equipment and associated geographical location information are collected and sent to an edge server; the edge server uses a general defect detection model for insulating equipment migrated from a cloud server, combined with local samples, to perform localized training, generating an adapted localized defect detection model, completing real-time detection, and providing feedback on the results, while also uploading the images and detection data to the cloud server; the cloud side constructs a resource library containing edge data and public datasets, trains and optimizes the general model, and periodically updates the edge model with parameters. This application effectively solves the core problems of poor real-time performance and insufficient adaptation of pure cloud-side detection, as well as the weak generalization ability of edge-side detection alone, achieving real-time detection adapted to local scenarios and continuous model optimization.

Description

Technical Field

[0001] This invention relates to the field of equipment defect detection, and in particular to a method for rapid identification of appearance defects in insulating equipment. Background Technology

[0002] In the safe and stable operation of a power system, the appearance integrity of insulating equipment (including key components such as insulators and insulating terminals) directly determines the insulation performance and power supply reliability of transmission lines. Rapid and accurate identification of defects is one of the core tasks of power line inspection. Early power line inspections relied primarily on manual labor. However, transmission lines often traverse complex geographical environments such as mountains and uninhabited areas. Inspection personnel not only face safety risks such as working at heights and snake / insect bites, but manual inspection is also affected by subjective judgment and physical limitations, resulting in low efficiency and a high rate of missed or false defects. This makes it difficult to meet the requirements of modern power systems for timeliness and accuracy in inspections.

[0003] With the development of intelligent technology, equipment such as drones and inspection robots are gradually being applied to the field of power inspection, and related detection technologies are constantly evolving. For example, the existing technology CN117830216A discloses a method for detecting lightning damage defects in the insulation terminals of transmission lines, which improves the detection efficiency to a certain extent. However, existing technologies still have significant limitations: First, pure cloud-side detection architectures rely on transmitting large amounts of image data to the cloud platform. Even with the help of 5G networks, transmission delays are still likely to occur in remote mountainous areas with weak signals. Furthermore, the concentrated computing power of the cloud platform can easily lead to congestion when multiple edge nodes access the platform simultaneously, making it difficult to meet real-time detection requirements. Second, insulation equipment in different regions may have individual differences due to model specifications, operating environments (such as humidity and types of contaminants). The general detection model deployed on the pure cloud side lacks a localization adaptation process, making it difficult to accurately identify defect features adapted to local scenarios, thus limiting detection accuracy. Third, some edge computing solutions simply migrate the model to the edge without combining it with local samples for targeted training, resulting in weak model generalization ability. Moreover, there is a lack of efficient data interaction and model update mechanisms between the edge and cloud sides. The optimization results of the cloud-side model cannot be synchronized to the edge side in a timely manner, and the personalized data from the edge side cannot feed back into the iteration of the cloud-side model, making it difficult to continuously improve the overall detection performance. Therefore, how to construct a rapid identification scheme for appearance defects of insulation equipment that takes into account real-time performance, accuracy and local adaptability, and solve the core problems of poor real-time performance, insufficient adaptability and weak generalization ability of edge-side detection in existing technologies, has become a technical bottleneck that urgently needs to be overcome in the current power inspection field. Summary of the Invention

[0004] This invention provides a method for rapid identification of appearance defects in insulation equipment, aiming to solve the problems of poor real-time performance, insufficient adaptability, and weak generalization ability of edge-side detection in the prior art.

[0005] To achieve the above objectives, the following technical solution is adopted.

[0006] A method for rapid identification of appearance defects in insulating equipment includes the following steps: Visible light images of insulating equipment are acquired by visible light imaging devices deployed on drones and inspection robots, and the visible light images and their associated geographic location information are sent to an edge server. On the edge server side, the visible light image is received, and a general detection model for insulation equipment defects migrated from the cloud server side is trained locally using locally stored insulation equipment sample data to obtain a localized defect detection model. This localized defect detection model is then used to detect defects in the received visible light image, obtaining real-time detection results, which are then fed back to the client. Simultaneously, the visible light image and the real-time detection results are uploaded to the cloud server. On the cloud server side, a cloud-side resource library containing image data from various edge servers and public datasets is constructed. Based on this cloud-side resource library, the general detection model for insulation equipment defects is trained and optimized, and the optimized model parameters are periodically distributed to the edge server to update the localized defect detection model on the edge server side.

[0007] Optionally, the data interaction between the edge server side and the cloud server side specifically includes: The edge server periodically uploads locally stored personalized insulation device sample images to the cloud server to supplement and update the training set and validation set in the cloud-side resource library; After the cloud server updates the cloud-side resource library, it retrains the general detection model for insulation equipment defects to optimize its parameters, and then migrates the optimized model parameters to the edge server to correct the localized defect detection model on the edge server side.

[0008] Optionally, the localized training on the edge server side includes: Receive initial detection model parameters migrated from the cloud server side; Based on the image features of the visible light image received by the edge server, the image size input to the initial detection model is adjusted; Based on the computing resources of the edge server and the number of locally stored insulating device samples, a matching neural network configuration is selected; Using the adjusted image size and selected neural network configuration, the initial detection model is trained with the local insulation equipment samples to generate a localized defect detection model adapted to the local scene.

[0009] Optionally, both the general detection model for insulation equipment defects and the localized defect detection model are target detection models built on an improved YOLOv3-tiny network. The improvement includes adding a third-scale prediction layer on top of the original two prediction layers of the YOLOv3-tiny network to detect small target defects. The feature map of the third-scale prediction layer is generated in the following way: the low-dimensional feature map from the shallow layer of the YOLOv3-tiny network is concatenated with the feature map from the middle layer network in the same dimension, and the residual operation is performed on the concatenated feature map to obtain the fused feature. The fused feature is then input into the newly added YOLO prediction layer.

[0010] Optionally, the feature map scale corresponding to the third-scale prediction layer is 52×52, and the feature map scales corresponding to the two prediction layers of the original YOLOv3-tiny network are 13×13 and 26×26, respectively.

[0011] Optionally, residual operations can be performed on the concatenated feature maps by introducing a residual module. This residual module includes a constant mapping branch and a residual branch, and their operational relationship is expressed as follows: ; ; Where y represents the output of the residual module, This represents the residual mapping function learned through training, where x is the input. Ws represents the network parameters.

[0012] Optionally, the residual module consists of a 1×1 convolutional kernel and a 3×3 convolutional kernel. The 1×1 convolutional kernel is used to reduce the dimensionality of the input feature map channels and introduce nonlinear transformations.

[0013] Optionally, training and optimizing the general detection model for insulation equipment defects on the cloud server side includes: In the initial stage, the model was pre-trained using a publicly available dataset of insulation equipment defects to generate an initial general detection model; During the operation phase, visible light images and detection results uploaded from the edge server are continuously received and added to the cloud-side resource library; The general detection model is retrained periodically using the updated cloud-side resource library, and the model parameters are updated through the backpropagation algorithm to achieve continuous optimization of the model.

[0014] Optionally, the step of periodically sending the optimized model parameters to the edge server specifically means that after the cloud server completes the optimized training of the general detection model each time, it automatically sends the updated neural network parameters to all the edge servers that cooperate with it. The edge servers receive the parameters and update their local defect detection model parameters accordingly.

[0015] Optionally, after using the localized defect detection model to perform defect detection on the received visible light image and obtaining the real-time detection result, the method further includes the following steps: Based on the type and geographical location information of the defects in the real-time detection results, inspection instructions or alarm information are generated and fed back to the control terminal of the drone and inspection robot for subsequent processing.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This application employs an edge-cloud collaborative detection architecture. It uses drones and inspection robots to collect visible light images of insulation equipment and their associated geographic location information, sending these to an edge server. The edge side utilizes a general defect detection model for insulation equipment migrated from the cloud server, combined with locally stored sample data, for localized training. This results in a detection model adapted to the local scenario, enabling real-time detection. Simultaneously, the cloud side constructs a multi-source data resource library to continuously optimize the general model and periodically distributes updates. This effectively solves the core technical problems of poor real-time performance, insufficient adaptability, and weak generalization ability of edge-side detection in existing technologies. Under this architecture, localized training on the edge server significantly reduces the amount of image data transmitted to the cloud, avoiding the impact of network transmission latency and cloud computing power congestion, thus significantly improving detection real-time performance. Localized training allows the model to fully learn the personalized characteristics of local insulation equipment, while the continuously optimized general model on the cloud side provides high-quality foundational support for the edge side. The two work together to simultaneously improve detection accuracy and localized adaptability. Furthermore, the association of geographic location information provides precise location data for subsequent defect handling, further enhancing the practicality of inspection work. Other technical designs in this solution further enhance the overall effect: the data interaction mechanism between the edge and cloud servers allows for the uploading of personalized samples from the edge to supplement the cloud-side resource library, and the back-migrating of the optimized model from the cloud to the edge, forming a two-way iteration of data and model, continuously improving the model's generalization ability and localization adaptation effect; by adjusting the image size and matching the neural network configuration, localized training can be adapted to the computing resources and local sample size of different edge servers, ensuring both model performance and training efficiency; the improved design of the YOLOv3-tiny network adds a third prediction layer with a scale of 52×52, combining shallow and mid-layer feature stitching By incorporating residual computation, the problem of insufficient accuracy in detecting small target defects is specifically addressed. The application of 1×1 convolution kernels introduces nonlinear transformations while achieving channel dimensionality reduction, balancing detection performance with computational overhead. The process design of cloud-side model pre-training, continuous optimization, and periodic parameter distribution ensures that the general model can continuously iterate based on multi-source data, and the edge-side model can be updated in a timely manner to maintain stable detection performance. By generating inspection instructions or alarm information and feeding them back to the equipment control terminal, seamless linkage between defect detection and subsequent handling is achieved, constructing a closed loop for inspection work, effectively improving defect handling efficiency, and providing more comprehensive protection for the safe and stable operation of the power system. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a cloud-edge collaborative ultra-high voltage external insulation equipment defect detection model, which is an embodiment of a rapid identification method for appearance defects in insulation equipment according to the present invention.

[0018] Figure 2This is a schematic diagram of the edge-end algorithm detection and update process in an embodiment of the rapid identification method for appearance defects in insulation equipment according to the present invention.

[0019] Figure 3 This is a schematic diagram of the residual module of an embodiment of the method for rapid identification of appearance defects in insulation equipment according to the present invention.

[0020] Figure 4 This is a micronetwork structure diagram of an improved YOLO v3 according to an embodiment of a method for rapid identification of appearance defects in insulation equipment according to the present invention.

[0021] Figure 5 This is a schematic diagram illustrating the cloud-based detection model training and updating process in an embodiment of the rapid identification method for appearance defects in insulation equipment according to the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0023] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0024] Example 1 This embodiment discloses a method for rapid identification of appearance defects in insulating equipment, applicable to appearance defect detection scenarios of various insulating equipment such as insulators and insulating terminals in ultra-high voltage transmission lines. Through the deep integration of visible light imaging and cloud-edge collaborative technology, combined with an improved target detection model, rapid and accurate defect identification is achieved. The specific implementation process of this method is described in detail below.

[0025] First, the acquisition and transmission of visible light images of the insulation equipment were carried out. The imaging team consisted of multiple drones and inspection robots, all equipped with high-definition visible light imaging devices. These devices support high-resolution imaging of 2048 pixels × 2448 pixels, clearly capturing minute defect features on the surface of the insulation equipment, meeting the image detail requirements for small target defect detection. During equipment deployment, inspection routes were predefined for the drones and inspection robots based on the geographical information of the area to be inspected, the distribution of power transmission lines, and the installation location of the insulation equipment. The onboard processor and the ground control center collaboratively planned the patrol paths to ensure coverage of all insulation equipment to be inspected, with no blind spots. Simultaneously, each drone and inspection robot was equipped with a high-precision GPS module to acquire its own geographical location information in real time. This geographical location information was linked to the acquired visible light images, accurately marking the actual installation location of the insulation equipment corresponding to each image, providing precise spatial reference for subsequent defect location and handling.

[0026] During the inspection process, drones and inspection robots autonomously fly or move within a designated area according to predefined inspection commands. Visible light imaging equipment automatically acquires images of the insulation equipment based on preset shooting frequencies and angles. The shooting angles cover multiple directions, including the front, side, and top, ensuring comprehensive capture of the appearance of all surfaces of the insulation equipment. After acquisition, the visible light images and their associated GPS geographic location information are transmitted in real time to a ground edge server via a wireless communication network. The wireless communication network can be either a 5G network or a WiFi 6 network. 5G networks are suitable for large-scale inspection scenarios such as remote mountainous areas, relying on their low latency and high bandwidth characteristics to ensure the real-time performance and stability of image transmission. WiFi 6 networks are suitable for short-range, concentrated inspection scenarios such as substations, reducing transmission costs. During transmission, the image data and geographic location information are encapsulated to ensure the integrity of data transmission and avoid data loss or corruption due to network fluctuations.

[0027] Next, the localized training and real-time detection process on the edge server side is executed. The edge server side consists of multiple edge servers, each configured with a hardware environment adapted to local computing resources, including processors, memory, and storage units, capable of meeting the computing power requirements for model training and real-time detection. The edge server first receives visible light images and related information transmitted from the drone and inspection robot, performing preliminary preprocessing on the received data, including data format verification and redundant data removal, to ensure the validity of the input data. Subsequently, the edge server migrates a general detection model for insulation equipment defects from the cloud server side. This migration process is implemented through a secure communication protocol, and the migrated content includes the model's network structure parameters, pre-trained weights, anchor box configuration, and other initialization parameters of the detection model, ensuring that the model acquired by the edge server has basic defect detection capabilities.

[0028] During the localization training phase, the initial detection model is first adapted. The first step is adjusting the input image size. The edge server performs feature analysis on the received visible light images, extracting image features such as resolution, pixel distribution, and the proportion of defect areas. Based on these features, the optimal input image size is determined. For example, when the pixel proportion of insulating equipment in the received image is small, the image size can be adjusted to 640 pixels × 640 pixels to retain more information about small target defects. When the edge server's computing resources are limited, the image size can be adjusted to 416 pixels × 416 pixels to reduce computational overhead while maintaining detection performance. The second step is selecting a matching neural network configuration. The edge server evaluates its own computing resources, including processor speed, memory capacity, and GPU computing power, while also counting the number of locally stored insulating equipment samples. If the edge server has sufficient computing power and a large number of local samples, a neural network configuration with more convolutional layers and feature fusion modules can be selected to improve the model's feature extraction capability. If the edge server has limited computing power or a small number of local samples, a lightweight neural network configuration is selected to reduce the model parameter size and ensure efficient and stable training.

[0029] After completing the above adjustments, the initial detection model was trained locally using locally stored insulation equipment sample data. The local sample data consisted of images collected during past inspections within the coverage area of ​​the edge server. These images included insulation equipment with different defect types, lighting conditions, and operating years, and were all manually annotated with information such as defect location and type to ensure the effectiveness and accuracy of the training data. During training, batch training was employed, with the batch size appropriately set based on the number of local samples. Iterative training optimized the model parameters, allowing the model to fully learn the personalized characteristics of local insulation equipment, such as defect characteristics caused by specific types of contaminants in a particular area and the aging characteristics of insulation equipment under local climatic conditions. Ultimately, this resulted in a localized defect detection model adapted to the local scenario.

[0030] Once the localized defect detection model is trained, it can be deployed for real-time detection applications. The edge server inputs real-time received visible light images into the localized model, which performs defect detection, quickly identifying defects such as cracks, breaks, stains, and lightning strike damage on the surface of insulating equipment. Simultaneously, it outputs the specific location coordinates and defect type of the defect. After detection, the edge server feeds back the real-time detection results to clients via wired or wireless communication. Clients include monitoring terminals at ground control centers and mobile terminals of inspection personnel. The detection results are presented in a combination of image annotations and text descriptions, facilitating quick understanding of the defect situation by relevant personnel. Simultaneously, the edge server uploads the original visible light images and corresponding real-time detection results to the cloud server. The upload frequency can be set according to the inspection intensity to ensure the cloud server can obtain the latest detection data in a timely manner, providing data support for the optimization of the general model.

[0031] The cloud server side undertakes the training, optimization, and parameter distribution tasks of the general detection model. Relying on ample storage and powerful computing resources, it constructs a comprehensive cloud-based resource library. The data sources of this library mainly consist of two parts: one part is visible light images and detection results uploaded by various edge servers. This data covers defect characteristics of different regions and types of insulation equipment, exhibiting strong personalization and scenario-based features; the other part is publicly available insulation equipment defect datasets, including public datasets of transmission line insulation equipment defects and PV module defect datasets. These datasets, accumulated over a long period, are rich in defect types and have a large number of samples, providing a solid foundation for training the general model. The cloud-based resource library adopts a distributed storage architecture, classifying and managing all data, and indexing it according to dimensions such as insulation equipment type, defect type, collection area, and collection time, facilitating rapid retrieval of required data during model training.

[0032] The training and optimization of the general detection model for insulation equipment defects is carried out in two phases. In the initial phase, the model is pre-trained using a public dataset from a cloud-based resource library. During pre-training, the public dataset is divided into a training set and a validation set in an 8:2 ratio. The training set is used to learn the model parameters, and the validation set is used to evaluate the model's detection performance. Stochastic gradient descent is used as the optimizer during training, with appropriate learning rates and iteration counts. Through multiple iterations, the model initially grasps the general characteristics of insulation equipment defects, generating an initial general detection model. In the operational phase, the cloud server continuously receives visible light images and detection results uploaded from various edge servers. This new data is added to the cloud-based resource library, and the data in the resource library is regularly updated and organized, including data cleaning, duplicate data removal, and annotation correction, to ensure the quality of the training data. Subsequently, the general detection model is retrained according to a preset cycle, which can be set based on the data update speed and changes in model performance, for example, once a month. During retraining, the Adam optimizer is used in conjunction with the backpropagation algorithm. The updated cloud-based resource library is used to adjust and optimize the model parameters. By minimizing the loss function, the model's accuracy in identifying various defects and its generalization ability are improved, thus achieving continuous iterative upgrades of the model.

[0033] An efficient two-way data interaction mechanism is established between the cloud server and the edge server to ensure timely updates of model parameters and effective supplementation of sample data. The edge server periodically uploads locally stored personalized insulation device sample images to the cloud server according to a preset cycle. Uploaded sample images are screened, prioritizing those containing locally unique defect types, rare defects, or samples with high detection difficulty. These samples effectively supplement the diversity of the cloud-side resource library and solve the detection bias problem caused by the limited sample pool in the general model. After receiving the sample images, the cloud server performs format conversion, annotation verification, and other processing. Qualified samples are added to the training and validation sets of the cloud-side resource library, completing the resource library update. Once the cloud-side resource library is updated, the cloud server automatically triggers the retraining process of the general detection model. Through the above retraining process, the model parameters are optimized. After training, the optimized neural network parameters are automatically distributed to all collaborating edge servers. The distributed parameters include core parameters such as model weights, biases, and anchor boxes. Encryption protocols are used during transmission to ensure data security. After receiving the parameters, the edge server selects either incremental or full update based on the current state of its own model to correct the parameters of the localized defect detection model, ensuring that the edge model always maintains the same optimization direction as the cloud side and maintains high detection accuracy.

[0034] In this method, both the general detection model and the localized defect detection model for insulation equipment are built based on the improved YOLOv3-tiny network. This network, as a lightweight real-time target detection algorithm, originally consists of 11 CBL modules, 6 max-pooling layers, 2 convolutional layers, 2 routing layers, 1 upsampling layer, and 2 YOLO prediction layers, totaling 24 network layers. It boasts advantages such as low computational cost and fast execution speed, making it highly suitable for deployment on edge servers and other devices with limited computing power. To address the issues of high false negative rates and insufficient feature extraction capabilities in small target defect detection of the original YOLOv3-tiny network, targeted improvements are made. The core improvement is the addition of a third-scale prediction layer specifically for the detection of small target defects.

[0035] The feature map scale corresponding to the third-scale prediction layer is 52×52, while the feature map scales corresponding to the two prediction layers of the original YOLOv3-tiny network are 13×13 and 26×26, respectively. The 13×13 scale prediction layer is suitable for detecting large targets, and the 26×26 scale prediction layer is suitable for detecting medium-sized targets. The newly added 52×52 scale prediction layer captures finer-grained feature information and is specifically adapted to the needs of detecting small target defects with a small pixel ratio in insulation equipment images. The feature map of the third-scale prediction layer is generated by combining cross-layer feature fusion and residual operation. The specific process is as follows: First, the shallow and middle layer feature output nodes of the network are determined. The feature map output of the 5th layer of the shallow network is selected. This layer's feature map contains rich low-dimensional detail features, which is beneficial for small target localization. The feature map output of the 13th layer of the middle network is selected. This layer's feature map has certain semantic information and can help distinguish defect types. Subsequently, channel number matching processing is performed on the shallow low-dimensional feature map and the middle-level feature map to ensure that the two dimensions are consistent. Then, tensor concatenation operation is used to achieve same-dimensional concatenation, resulting in a concatenated feature map that integrates detailed features and semantic features.

[0036] To further enhance feature extraction capabilities and address the gradient vanishing problem in deep networks, a residual module is introduced after concatenating the feature maps to perform residual operations. This residual module consists of a 1×1 convolutional kernel and a 3×3 convolutional kernel. The 1×1 convolutional kernel first performs channel dimensionality reduction on the concatenated feature map, reducing computational complexity while introducing a non-linear transformation. The ReLU activation function is used to enhance the model's ability to express complex features. Subsequently, feature extraction is performed through the 3×3 convolutional kernel, deepening the mining of feature information. The residual module includes a constant mapping branch and a residual branch, and their operational relationship is specifically represented in two forms. The first is... Where x is the input feature map of the residual module, W is the residual mapping function learned through training with 1×1 and 3×3 convolutional kernels. iThe first type is the network parameters of the convolution kernel, and the second type is the output of the residual module. W s The network parameters for the constant mapping branch are used to adapt to the dimensionality differences between the input and output feature maps, ensuring the effectiveness of the computation. Residual operations effectively preserve the feature information of small targets, preventing its loss during deep network propagation, while also improving the model's training efficiency and enabling it to converge quickly to the optimal state. Finally, the fused features obtained through residual operations are input into a newly added YOLO prediction layer. This prediction layer has the same structure as the original network's prediction layer, and through processes such as anchor box matching and confidence calculation, accurate detection of defects in small targets is achieved.

[0037] Once the edge server uses a localized defect detection model to detect defects in visible light images and obtains real-time detection results, subsequent defect handling and coordination work needs to be carried out. First, the real-time detection results are analyzed to clarify the specific type and severity of the defects. Defect types include cracks, damage, stains, lightning strike ablation, etc., and the severity is classified into different levels based on the defect area, location, and impact on insulation performance. By combining the geographic location information associated with the images, corresponding inspection instructions or alarm information are generated: For minor defects, a routine inspection instruction is generated, including the defect location, defect type, and recommended inspection cycle, which is fed back to the control terminal of the drone and inspection robot. The control terminal adjusts the subsequent inspection plan according to the instruction, increasing the inspection frequency of the area; for moderate defects, a key inspection instruction is generated, controlling the drone to adjust its flight path to the defect location for close-up, multi-angle shooting, supplementing the collection of detailed defect images to provide a basis for further evaluation; for severe defects, a high-level alarm information is immediately generated, and the alarm information is simultaneously pushed to the ground control center monitoring terminal and the mobile terminals of relevant inspection personnel. At the same time, an emergency response instruction is generated, specifying the defect location, defect severity, and recommended handling plan, guiding inspection personnel to quickly go to the site for emergency repair or replacement, avoiding power safety accidents caused by the expansion of defects.

[0038] Throughout the implementation of this method, the cloud server and edge server collaborate continuously, forming a closed-loop optimization mechanism. The edge server adapts to local scenarios through localized training, ensuring the real-time performance and accuracy of detection; the cloud server optimizes the general model by integrating multi-source data, providing high-quality model support for the edge server; through bidirectional interaction of sample uploading and parameter distribution, both parties achieve continuous iteration of data and models, continuously improving the performance of the entire detection system. This enables it to adapt to the defect detection needs of different regions, environments, and types of insulation equipment, providing a reliable guarantee for the safe and stable operation of ultra-high voltage power systems.

[0039] Example 2 Step 1: Construct a cloud-edge collaborative model This study constructs a cloud-edge collaborative defect detection model for ultra-high voltage external insulation equipment, such as... Figure 1 As shown, the core lies in achieving efficient data processing through the collaborative operation of edge servers and the cloud. Its network architecture is clearly divided into three parts: the shooting side, the edge side, and the cloud side. Figure 1 The specific relationships and functional positioning of these three ports are clearly presented, providing a framework support for the overall construction of the model.

[0040] The edge side, serving as the core of the model's intermediate processing, consists of M edge servers. Its construction focuses on achieving real-time response and personalized adaptation. Functionally, each edge server undertakes two core tasks: first, the migration and localization training of the general model. During construction, the general defect detection model generated in the cloud is first migrated to the edge side, and then targeted training is conducted using samples of ultra-high voltage and extra-high voltage external insulation equipment stored locally on the edge. By adjusting model parameters to adapt to the edge scenario, a localized defect detection model is ultimately generated, solving the adaptability problem of the general model in edge scenarios. Second, real-time data processing and result interaction. After receiving real-time data uploaded by drones and inspection robots, the edge server completes the detection using the locally trained model and outputs the results. On the one hand, the detection results are fed back to the client, enabling real-time early warning and response to defects; on the other hand, the original real-time data and detection results are synchronously uploaded to cloud storage, providing data support for cloud model optimization and achieving the goal of real-time detection of defects in ultra-high voltage and extra-high voltage external insulation equipment.

[0041] The cloud side serves as the global resource hub and core of model optimization, its construction relying on ample storage resources and powerful computing capabilities. At the data level, the construction of the cloud-side resource repository requires integrating two types of data: first, processed and collected data uploaded from various edge servers; and second, public datasets in the field of defect detection for ultra-high voltage and extra-high voltage external insulation equipment. Through data fusion, a training resource pool with broad coverage and diverse scenarios is formed. Based on this resource pool, the cloud side can initially generate a general defect detection model. Simultaneously, as edge data is continuously uploaded, the cloud side dynamically updates the dataset and retrains the model, continuously optimizing the model's transfer decision-making capabilities. After training, the optimized neural network parameters are fed back to the edge side to update the parameters of the localized defect detection model, improving the processing efficiency of subsequent tasks on the edge servers and forming a closed-loop collaborative mechanism.

[0042] The core of cloud-edge collaboration lies in building an efficient two-way data interaction mechanism, from Figure 1It is evident that this mechanism identifies two key tasks in model construction: first, edge-to-cloud sample supplementation. To improve the generalization ability of the cloud-based general model, the edge server needs to periodically upload local personalized samples to cloud storage during construction to supplement the cloud's training and validation sets, addressing the detection bias issue caused by the limited sample pool in the general model; second, cloud-to-edge parameter optimization. Once the cloud dataset is updated, the parameters of the general detection model will be optimized through retraining. A timed migration mechanism needs to be designed during construction to synchronize the optimized parameters to the edge side, correcting the localized defect detection model and ensuring that the edge model always maintains consistency with the cloud's optimization direction, thus maintaining high detection accuracy.

[0043] In the cloud-edge collaborative technology system, the construction of edge-side algorithms and training processes must adhere to the principle of "unified framework and characteristic adaptation" to balance detection accuracy and edge computing power. Edge-side algorithms are based on cloud-side algorithms, using transfer learning to migrate general model parameters from the cloud to the edge. However, they require targeted optimization based on edge-side characteristics. Specific optimization strategies include: adjusting the input image size according to the resolution and lighting conditions of the images acquired at the edge to avoid redundant data consuming computing power; and selecting a matching neural network configuration based on the edge server's computing power limit, memory size, and the number of local samples to ensure efficient algorithm operation at the edge. The update and application process of edge-side algorithms is as follows: Figure 2 As shown, the initial edge detection model is the initialization detection model; the edge detection model is the localized defect detection model.

[0044] Furthermore, the training process for the localized defect detection model involves three steps: first, obtaining an initial model from the cloud via parameter transfer; then, deploying the model to the edge controller and conducting localized training using local samples to optimize the model's ability to identify defects in local equipment; finally, generating a localized detection model that can be directly applied to the edge. In practical applications, simply inputting real-time images of ultra-high voltage external insulation equipment into the model will generate detection results in real time, and the response can be quickly scheduled based on the results, completing the closed-loop construction at the edge.

[0045] Step 2: Improve the accuracy of small target defect detection In the field of defect detection for ultra-high voltage and extra-high voltage external insulation equipment, the core challenge lies in improving the detection capability for minute targets. To address this issue, this research proposes an improved architecture as follows: by fusing shallow features and introducing a residual module, a third-scale prediction layer is added to the original YOLOv3-tiny network. Furthermore, during feature fusion and scale expansion, the integration of shallow low-dimensional features and deep semantic features is completed in the mid-layer feature processing stage of the network. Notably, the newly added feature fusion path does not alter the backbone structure of the original network, thus achieving compatibility between the improved module and the base network. This design approach, coupled with the method of expanding network functionality only through additional layers, allows the model to perform transfer learning based on the pre-trained weights of YOLOv3-tiny while retaining the core advantages of the original algorithm: lightweight performance and real-time capability.

[0046] Compared to the traditional YOLOv3-tiny mechanism that relies on only two-scale prediction layers, this study proposes a third-scale prediction strategy that draws on the research approach of multi-scale feature detection. This strategy uses a 52×52 scale feature map to specifically capture the feature information of small targets. In the key step of feature processing, the residual enhancement feature fusion module performs same-dimensional concatenation of shallow and mid-level features and residual operations. The feature generation method of the third-scale prediction layer is as follows: it adopts a combination of cross-layer feature fusion and residual connection to focus on enhancing the representation ability of small target features. Finally, the obtained fused features are input into the newly added YOLO prediction layer to complete the target detection task.

[0047] YOLOv3-tiny, as a lightweight real-time object detection algorithm, perfectly aligns with the aforementioned design criteria. Its lightweight network architecture is better suited for edge deployment requirements; simultaneously, YOLOv3-tiny is a simplified version of the YOLOv3 network structure, reducing memory usage during model training, accelerating object detection speed, and thus better meeting the demands for real-time detection in complex environments. The network comprises 24 layers, specifically consisting of 11 CBL modules, 6 maxpooling layers, 2 convolutional layers, 2 routing layers, 1 upsampling layer, and 2 YOLO prediction layers.

[0048] In real-world inspection scenarios, the pixel ratio of ultra-high voltage external insulation equipment captured by drones is extremely low, directly resulting in a persistently high false negative rate for small target detection. To address this issue, this study constructs a third-scale prediction layer by fusing shallow features and introducing a residual module, achieving a targeted improvement to the YOLOv3-tiny algorithm.

[0049] From the perspective of feature extraction characteristics, shallow networks have an advantage in locating small targets, but their semantic expression ability is weak; while deep networks, although containing rich semantic information, are prone to losing the localization details of small targets. When the image size input to the network is 416 pixels × 416 pixels, YOLOv3-tiny only has two target detection scales, namely 13×13 and 26×26. The 13×13 detection layer is suitable for detecting large targets, and the 26×26 detection layer is suitable for detecting medium-sized targets. Since the camera of a drone is installed on the bottom of the fuselage to collect equipment information, the image resolution is usually 2048 pixels × 2448 pixels, resulting in the extremely small size of defects in equipment components in the image. When the YOLOv3-tiny model detects such small targets, it is very easy to miss detections.

[0050] To address the aforementioned issues with the model and overcome the limitation of YOLOv3-tiny in losing shallow low-dimensional features in deeper layers, leading to missed detections of small targets, this study fuses shallow feature information with the concatenation layer of the second detection scale, adding a new 52×52 detection scale layer as the third prediction layer. The feature information of this third prediction layer originates from the feature map resulting from the concatenation of the shallow and second layers at the same dimension, containing richer low-dimensional image features. This helps the model improve its ability to detect small targets, thereby minimizing the probability of missed detections of small targets.

[0051] Step 3: Introduce a deep residual framework Inspired by the design principles of ResNet, this study introduces a deep residual framework to address the vanishing gradient problem. This framework allows convolutional networks to learn residual mappings, rather than requiring each stacked layer in the network to perfectly fit the latent mapping. The core building block of a residual network is the residual module, which contains two branches—a constant mapping branch and a residual branch, as shown below. Figure 3 As shown.

[0052] In the residual learning module, if the input is x and its basis mapping is denoted as H(x), then the residual function that the network is expected to learn is F(x) = H(x) - x, and correspondingly, the learned feature output by the module is F(x) + x. If the residual is 0, this building block only performs constant mapping, and the detection accuracy of the network model will at least not decrease; however, in reality, the residual function cannot be 0, which allows the residual learning unit to further learn new feature information based on the constant mapping, thereby improving the learning efficiency of the network. The mathematical expression of the residual network is as follows: Where y represents the output of the residual module. This represents the residual mapping function obtained after training. and These are the input parameters of the network.

[0053] The introduction of residual modules can prevent the loss of small target features due to gradient vanishing as network depth increases, thus better improving the high false negative rate problem in small target detection. This study constructs residual modules by introducing 1×1 and 3×3 convolutional kernels: the 1×1 kernel not only achieves channel dimensionality reduction—reducing model complexity when computation is high—but also introduces more nonlinear transformations into the neural network. This design significantly increases the depth of the residual blocks, effectively improving the feature representation capability of the residual network. The final improved YOLOv3-tiny network model is shown below. Figure 4 As shown.

[0054] like Figure 5 As shown, in the initial stage of the task, the publicly available PV component dataset is mainly used for general model training. As the task progresses, local samples from real-world application scenarios will be generated at the edge. These samples will be incorporated into the database for continuous model updates and training. During training, all samples will be divided into training and validation sets according to a preset ratio. With the help of ample computing resources in the cloud, continuous and efficient general model parameter update services can be provided for different task scenarios throughout the entire detection process, and the model parameters are updated through the backpropagation algorithm in each training session.

[0055] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative in all respects and are not the only ones. All modifications within the scope of this invention or its equivalents are included in this invention.

Claims

1. A method for rapid identification of appearance defects in insulating equipment, characterized in that, Includes the following steps: Visible light images of insulating equipment are acquired by visible light imaging devices deployed on drones and inspection robots, and the visible light images and their associated geographic location information are sent to an edge server. On the edge server side, the visible light image is received, and a general detection model for insulation equipment defects migrated from the cloud server side is trained locally using locally stored insulation equipment sample data to obtain a localized defect detection model. This localized defect detection model is then used to detect defects in the received visible light image, obtaining real-time detection results, which are then fed back to the client. Simultaneously, the visible light image and the real-time detection results are uploaded to the cloud server. On the cloud server side, a cloud-side resource library containing image data from various edge servers and public datasets is constructed. Based on this cloud-side resource library, the general detection model for insulation equipment defects is trained and optimized, and the optimized model parameters are periodically distributed to the edge server to update the localized defect detection model on the edge server side.

2. The method for rapid identification of appearance defects in insulating equipment according to claim 1, characterized in that, The data interaction between the edge server side and the cloud server side specifically includes: The edge server periodically uploads locally stored personalized insulation device sample images to the cloud server to supplement and update the training set and validation set in the cloud-side resource library; After the cloud server updates the cloud-side resource library, it retrains the general detection model for insulation equipment defects to optimize its parameters, and then migrates the optimized model parameters to the edge server to correct the localized defect detection model on the edge server side.

3. The method for rapid identification of appearance defects in insulating equipment according to claim 1, characterized in that, The localized training on the edge server side includes: Receive initial detection model parameters migrated from the cloud server side; Based on the image features of the visible light image received by the edge server, the image size input to the initial detection model is adjusted; Based on the computing resources of the edge server and the number of locally stored insulating device samples, a matching neural network configuration is selected; Using the adjusted image size and selected neural network configuration, the initial detection model is trained with the local insulation equipment samples to generate a localized defect detection model adapted to the local scene.

4. The method for rapid identification of appearance defects in insulating equipment according to claim 1, characterized in that, Both the general detection model for insulation equipment defects and the localized defect detection model are target detection models built on an improved YOLOv3-tiny network. The improvement includes adding a third-scale prediction layer to the original two prediction layers of the YOLOv3-tiny network to detect small target defects. The feature map of the third-scale prediction layer is generated in the following way: the low-dimensional feature map from the shallow layer of the YOLOv3-tiny network is concatenated with the feature map from the middle layer network in the same dimension, and the residual operation is performed on the concatenated feature map to obtain the fused feature. The fused feature is then input into the newly added YOLO prediction layer.

5. The method for rapid identification of appearance defects in insulating equipment according to claim 4, characterized in that, The feature map scale corresponding to the third-scale prediction layer is 52×52, and the feature map scales corresponding to the two prediction layers of the original YOLOv3-tiny network are 13×13 and 26×26, respectively.

6. The method for rapid identification of appearance defects in insulating equipment according to claim 4, characterized in that, Performing residual operations on the stitched feature maps is achieved by introducing a residual module, which includes a constant mapping branch and a residual branch, and their operational relationship is expressed as follows: ； ； Where y represents the output of the residual module, This represents the residual mapping function learned through training, where x is the input. Ws represents the network parameters.

7. The method for rapid identification of appearance defects in insulating equipment according to claim 6, characterized in that, The residual module consists of a 1×1 convolution kernel and a 3×3 convolution kernel. The 1×1 convolution kernel is used to reduce the dimensionality of the input feature map channels and introduce nonlinear transformations.

8. The method for rapid identification of appearance defects in insulating equipment according to claim 1, characterized in that, The training and optimization of the general detection model for insulation equipment defects on the cloud server side includes: In the initial stage, the model was pre-trained using a publicly available dataset of insulation equipment defects to generate an initial general detection model; During the operation phase, visible light images and detection results uploaded from the edge server are continuously received and added to the cloud-side resource library; The general detection model is retrained periodically using the updated cloud-side resource library, and the model parameters are updated through the backpropagation algorithm to achieve continuous optimization of the model.

9. The method for rapid identification of appearance defects in insulating equipment according to claim 1, characterized in that, The step of periodically sending the optimized model parameters to the edge server specifically involves the cloud server automatically sending the updated neural network parameters to all the edge servers that are working with it after each optimization training of the general detection model. The edge servers receive the parameters and use them to update their local defect detection model parameters.

10. The method for rapid identification of appearance defects in insulating equipment according to claim 1, characterized in that, After using the localized defect detection model to perform defect detection on the received visible light image and obtaining the real-time detection result, the following steps are also included: Based on the type and geographical location information of the defects in the real-time detection results, inspection instructions or alarm information are generated and fed back to the control terminal of the drone and inspection robot for subsequent processing.

Images