Method and system for judging insulation of electrified exposed point of pole-mounted equipment

Abstract

The invention belongs to the technical field of acceptance inspection after insulation transformation of power equipment, and provides a pole-mounted equipment electrified bare point insulation judgment method and system, and a focal length adjustment method for determining illumination and capture modes based on illumination intensity judgment and image quality judgment based on a multi-feature linear weighted fusion strategy. The shooting requirements under different illumination conditions and complex equipment layout can be met, the problem that the rod shakes to affect the image quality in the manual operation insulation rod shooting process is solved, and the image quality is improved; on the basis, on the basis of an abnormal defect automatic detection model, by considering classification loss, bounding box regression loss and confidence coefficient loss, identification and classification of different defect features can be better realized, the purpose of improving the detection precision can be achieved, and the problem of inaccurate defect identification is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of acceptance after insulation retrofitting of power equipment, and particularly relates to a method and system for determining the insulation of exposed live points of pole-mounted equipment. Background Technology

[0002] 10kV pole-mounted equipment is widely distributed in power systems, and the insulation retrofitting of its exposed live points is crucial for ensuring the safe and stable operation of the power system. Exposed point insulation retrofitting mainly refers to the spraying of insulating coatings onto the pole-mounted equipment. Currently, the acceptance (or judgment) of insulation retrofitting largely relies on manual visual inspection. Given the complex structure of pole-mounted equipment and the live-line working environment, it is difficult to comprehensively and meticulously inspect the insulation retrofitted areas, easily overlooking potential defects and safety hazards.

[0003] Some existing technologies attempt to use insulating rods and camera structures for acceptance testing. While the use of insulating rods can achieve the purpose of insulation protection in live environments, cameras struggle to meet the shooting needs under different lighting conditions and complex equipment layouts, failing to guarantee image quality and affecting defect identification results. Furthermore, manual operation of the insulating rods can lead to rod wobbling during shooting, affecting the camera's image quality and resulting in poor image acquisition. Additionally, defects in the sprayed insulating coating, such as damage, cracks, bubbles, drips, and unevenness, may have indistinct characteristics, and there may be errors in determining the boundaries between different defect types. This leads to inaccurate defect identification when conducting acceptance testing under conditions of poor image quality. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method and system for determining the insulation of exposed live points on pole-mounted equipment. This invention determines the lighting and image capture mode based on light intensity determination, and uses a focus adjustment method based on a multi-feature linear weighted fusion strategy for image quality assessment. This method can meet the shooting needs under different lighting conditions and complex equipment layouts, solving the problem of image quality being affected by pole swaying during manual operation of the insulating pole, thus improving image quality. Furthermore, based on an automatic defect detection model, by considering classification loss, bounding box regression loss, and confidence loss, it can better identify and classify different defect features, thereby improving detection accuracy and solving the problem of inaccurate defect identification.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a method for determining the insulation of live exposed points of pole-mounted equipment, comprising: The light intensity of the equipment on the column is obtained, the light intensity is judged, and the lighting and image capture modes are determined based on the light intensity judgment results; Using a defined lighting and image capture mode, image information of the pole-mounted device is acquired; and a multi-feature linear weighted fusion strategy is used to judge the image quality, and the focal length is adjusted according to the judgment result. The image information of the column-mounted equipment is judged to be qualified, and qualified images are obtained. For qualified images, abnormal defects are detected and classified using a preset automatic abnormal defect detection model. Acceptance is carried out based on the abnormal defect detection results and classification results. The loss in the automatic abnormal defect detection model is equal to the weighted sum of classification loss, bounding box regression loss and confidence loss.

[0006] Furthermore, a preset light intensity determination model is used for judgment, wherein the light intensity determination model is a YOLOv5 deep learning model; in the YOLOv5 deep learning model, EfficientNetV2 is used to replace the backbone network, and the last two layers of the convolutional layer are replaced with depthwise separable convolutional layers to reduce the number of parameters.

[0007] Furthermore, when the lighting conditions are low, the camera used to acquire image information of the pole-mounted equipment selects the supplementary lighting mode; when the lighting conditions are high, the camera selects the strong light suppression mode; when the lighting conditions are normal, the camera selects the default parameter mode; and when the lighting conditions are non-uniform, the camera selects the wide dynamic range mode.

[0008] Furthermore, adjusting the focal length based on the image quality assessment results includes: ; in, This is the final quality score; , and These are the weighting coefficients; , and These are the image quality scores; when the final quality score... When the value exceeds the threshold, it indicates that the image is not sharp and requires automatic fine-tuning of the focus until the final quality score is achieved. The value should not exceed the threshold, and a threshold for the number of focusing attempts should also be set.

[0009] Furthermore, when making a qualification judgment, blurry images, ghost images, and images without target parts are filtered out as unqualified based on a preset judgment model; the judgment model is a deep learning model based on the YOLOv5 network architecture.

[0010] Furthermore, in the judgment model, a Swing Transformer module is introduced into the backbone network, and a convolutional block attention module is inserted to enhance the feature extraction capability and key feature response capability; at the same time, the last layer of the convolutional layer is replaced with a depthwise separable convolutional layer to reduce the number of parameters.

[0011] Furthermore, the automatic detection model for abnormal defects is a deep learning model based on the YOLOv8 network architecture; the YOLOv8 network architecture introduces a weighted bidirectional feature pyramid and utilizes a dynamic label allocation strategy to perform classification and regression tasks.

[0012] Furthermore, the classification loss is: ; in, These are real labels; It is a predicted probability; ,in, The weights of positive and negative samples are balanced; γ is a hyperparameter for adjusting the weights of difficult samples.

[0013] Furthermore, the bounding box regression loss is: ; in, , This is the intersection-union ratio (IU) between the predicted bounding box and the ground truth bounding box. The squared Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; The diagonal length of the smallest bounding rectangle that covers both frames; These are the weighting coefficients; This is the aspect ratio consistency adjustment factor; An adjustable power parameter α is introduced to dynamically adjust the loss weights corresponding to different IoU values, thereby improving the performance of the object detection model; a shape similarity factor is introduced to enhance the detection capability for objects with complex shapes. ,in, Used for balance And the weight of shape similarity, Shape similarity is calculated based on aspect ratio differences; ,in, The width and height of the prediction box. The actual width and height of the bounding box. Hyperparameters for controlling shape similarity sensitivity.

[0014] Secondly, the present invention also provides a system for determining the insulation of exposed live points of pole-mounted equipment, comprising: The lighting and capture mode determination module is configured to: acquire the light intensity of the equipment on the column, determine the light intensity, and determine the lighting and capture mode based on the light intensity determination result; The image acquisition module is configured to: acquire image information of the pole-mounted device using a defined lighting and image capture mode; and to judge the image quality using a multi-feature linear weighted fusion strategy and adjust the focal length based on the judgment result. The acceptance module is configured to: perform a qualification judgment on the image information of the column-mounted equipment to obtain a qualified image; for the qualified image, perform abnormal defect detection and classification using a preset abnormal defect automatic detection model; and perform acceptance based on the abnormal defect detection results and classification results; wherein, the loss in the abnormal defect automatic detection model is equal to the weighted sum of classification loss, bounding box regression loss and confidence loss.

[0015] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for determining the insulation of live exposed points of pole-mounted equipment as described in the first aspect.

[0016] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the pole-mounted equipment live exposed point insulation determination method described in the first aspect.

[0017] Fifthly, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the method for determining the insulation of live exposed points of pole-mounted equipment as described in the first aspect.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention innovatively proposes a method for determining the insulation of exposed live points on pole-mounted equipment. It determines the lighting and image capture mode based on light intensity and uses a focus adjustment method based on a multi-feature linear weighted fusion strategy for image quality assessment. This achieves the shooting requirements under different lighting conditions and complex equipment layouts, solving the problem of image quality being affected by pole swaying during manual operation of the insulating pole, thus improving image quality. Furthermore, based on an automatic defect detection model, by considering classification loss, bounding box regression loss, and confidence loss, it can better identify and classify different defect features, improving detection accuracy and solving problems such as unclear features of defects like breakage, cracks, bubbles, drips, and unevenness, as well as errors in determining the boundaries between different defect types. Attached Figure Description

[0019] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0020] Figure 1 This is a schematic diagram of the acceptance tool structure according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the display according to Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the acceptance process in Embodiment 1 of the present invention; Figure 4 The image acquisition process is described in Embodiment 1 of the present invention; Figure 5 This is the data processing flow of Embodiment 1 of the present invention; The components include: 1. Camera; 2. First antenna; 3. Universal joint; 4. Insulating rod; 5. Second antenna; 6. Display. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0023] Example 1: During the acceptance (judgment) process after insulation modification, some existing technologies attempt to use simple tools to assist in the acceptance, such as ordinary telescopes. However, telescopes can only provide a limited field of view and cannot clearly image the details of the equipment, let alone detect internal defects in the insulation layer. Other methods use ordinary cameras for acceptance, but in a live environment, ordinary cameras lack the necessary insulation protection, posing a significant safety risk.

[0024] As described in the background section, although the purpose of insulation protection in a live environment can be achieved, the camera is difficult to meet the shooting needs under different lighting conditions and complex equipment layouts. Furthermore, manual operation of the insulating rod can cause it to shake during shooting, which in turn affects the camera's image quality and results in poor image quality. In addition, the sprayed insulating coating may have defects such as damage, cracks, bubbles, drips, and unevenness, which are not obvious and there are errors in determining the boundaries between different defect types. This leads to inaccurate defect identification when accepting the product under poor image quality.

[0025] This embodiment provides a method for determining the insulation of live exposed points of pole-mounted equipment, including: S1. Preparation Phase: S1.1 Preparation of Acceptance Tools: The acceptance tools used in this embodiment are as follows: Figure 1 Specifically, a tool for determining the insulation of exposed live points of pole-mounted equipment is provided. The tool includes an insulating rod 4, a universal joint 3 at one end of the insulating rod 4, a camera 1 mounted on the universal joint 3, and a first antenna 2 mounted on the camera 1. The tool also includes a display 6, a second antenna 5 mounted on the display 6, and a processor within the display 6.

[0026] Optionally, the camera 1 can be implemented using a visualization camera. The universal joint is a 360° universal joint, meeting the requirements for rotation in all directions; the insulating rod 4 can be configured as multiple interconnected sub-insulating rods, each sub-insulating rod being 1.5 meters long, and can be extended according to the actual site environment; the first antenna 2 and the second antenna 5 transmit information to each other wirelessly, for example, transmitting the images captured by the camera 1 to the display 6 and its internal processor, the processor containing image processing algorithms, etc.

[0027] S1.2, Operator preparation: The operator dons protective insulated gear, checks the battery level of camera 1 on insulating pole 4, and confirms that all connections between components are normal, ensuring the acceptance tools are in working order. Simultaneously, they gather basic information about the 10kV pole-mounted equipment to be accepted, including equipment type and insulation modification details.

[0028] S2, On-site Operation and Image Processing Stage: The operator holds the insulating rod 4 and slowly approaches the camera 1 towards the insulation modification site of the live exposed point of the 10kV pole-mounted equipment. Based on the equipment structure and location, the length and angle of the insulating rod 4 are adjusted to allow the camera 1 to obtain the optimal shooting angle and achieve the goal of acquiring image information of the pole-mounted equipment.

[0029] To overcome the influence of lighting conditions and capture clear images, this embodiment employs a lighting intensity determination model based on the YOLOv5 deep learning model. This model automatically determines the lighting intensity and adaptively selects different lighting and capture modes for the camera based on the determination results. Then, an image quality assessment algorithm is used to determine the clarity of the captured image. If the image is deemed unclear, the camera's automatic focus adjustment mode is selected for focusing and capturing. Based on the varying lighting conditions in the actual application scenario, four scenarios are categorized: low light, high light, normal light, and non-uniform light.

[0030] The specific process of step S2 is as follows: S2.1 Model Training Phase: First, the collected sample images under four different lighting conditions are classified and labeled. Algorithms such as linear transformation, gamma correction, and histogram equalization are used to modify image brightness to expand the samples. The image samples are then divided into training and test sets according to the proportion. Next, a deep learning model with a YOLOv5 network architecture is used for fine-tuning and training. Finally, the trained model is used to test and evaluate the test set to obtain the optimal or suboptimal lighting intensity judgment model (Model-0). Among these, lighting conditions can be determined by a light sensor, which will not be detailed here. Low lighting, high lighting, and normal lighting can be determined by comparing with lighting intensity thresholds. For example, lighting conditions below the lower threshold are low lighting, lighting conditions above the upper threshold are high lighting, and lighting conditions between the lower and upper thresholds are normal lighting. Lighting conditions where the frequency of light intensity fluctuation is greater than a preset frequency within a preset time are non-uniform lighting conditions, or where the difference in lighting intensity between different areas within the same image is greater than a preset difference are non-uniform lighting conditions.

[0031] Since the deep learning-based illumination intensity determination model needs to be deployed on the mobile camera 1, the YOLOv5 deep learning model is lightweighted by replacing the backbone network with EfficientNetV2 and replacing the last two convolutional layers with depthwise separable convolutions to reduce the number of parameters. This approach can accurately extract features from various samples while reducing the computational cost of the model, thereby improving inference speed while maintaining inference accuracy under limited hardware. The improved YOLOv5 deep learning architecture illumination intensity determination model can improve inference speed while maintaining accuracy, and can be deployed on the mobile camera 1.

[0032] S2.2 Application stage of the light intensity determination algorithm: After the insulation layer of the column-mounted equipment is sprayed, various defects such as damage, cracks, bubbles, drips, and unevenness may exist. However, the characteristics are not obvious. Especially under different lighting conditions, it is easy to confuse the boundary cracks and fissures in the damage, the bubbles and drips, the fissures and unevenness boundaries, and other features. If the image information quality is poor, problems such as incorrect defect identification, missed detection, or false detection may occur.

[0033] Based on this, a light intensity determination model is deployed on the camera mobile terminal. When a capture signal is received, the optimal or suboptimal light intensity determination model is invoked to automatically determine the light intensity. Then, based on the light intensity determination result, different lighting and application modes of the camera are adaptively selected. When the light intensity determination algorithm determines that the lighting condition is low, the camera automatically selects the supplementary lighting mode; when the lighting condition is high, the camera automatically selects the strong light suppression mode; when the lighting condition is normal, the camera selects the default parameter mode; when the lighting condition is non-uniform, the camera automatically selects the wide dynamic range mode. By selecting lighting and capture modes under different lighting conditions, the shooting needs under different lighting conditions and complex equipment layouts are met, ensuring image quality. It provides a guarantee for feature extraction and boundary determination of defects such as damage, cracks, bubbles, drips, and unevenness, and can improve the accuracy of defect identification results.

[0034] S2.3, Application stage of image quality assessment algorithm: After the application mode and parameter settings of camera 1 are completed, an image quality judgment algorithm that fuses multiple features, including BRISQUE (Blind / Referenceless Image Spatial Quality Evaluator), NIQE (Natural Image Quality Evaluator), and / or PIQE (Perception-based Image Quality Evaluator), is used to determine whether the acquired image is clear. If the image is determined to be unclear, the focus is automatically fine-tuned until the camera emits a click sound, indicating that a clear image has been acquired. The number of times the focus is automatically fine-tuned is a configurable parameter, with a default of 3 times. Once a clear image of this area is acquired, the image acquisition of the next target area can begin. This enables adaptive shooting of the insulation modification area from multiple angles, such as the front, side, and top, to ensure image quality and comprehensive coverage of the target area.

[0035] Because the insulating rod 4 requires manual operation, there will be slight shaking of the insulating rod 4 and the camera 1 during the shooting process, which will affect the shooting picture and the image information quality. Therefore, in order to overcome the problems of image blurring and image ghosting caused by slight shaking, a strategy of BRISQUE, NIQE and / or PIQE multi-feature linear weighted fusion is adopted to automatically judge the image quality, and then adjust the focus autonomously according to the image quality judgment result.

[0036] pass The final quality score is obtained by weighted summation of the scores from the three factors. ;in, , and These are the weighting coefficients. , and These are the image quality scores corresponding to BRISQUE, NIQE, and PIQE, respectively.

[0037] when Greater than threshold A ( If the image is not clear enough, it indicates that automatic fine-tuning of the focus is needed to capture a clear image, which in turn helps to accurately identify whether the sprayed insulation layer is qualified, until the final quality score is determined. If the value is not greater than the threshold, it indicates that the image is clear. At the same time, by setting a threshold for the number of focus adjustments, shooting efficiency can be effectively controlled, preventing camera 1 from automatically focusing indefinitely at a certain position.

[0038] Weighted fusion can cover a wider range of distortion types, such as structural distortion, reducing the blind spots of single indicators. The three different indicators have varying sensitivities to noise, lighting changes, and compression; the fusion strategy improves the algorithm's stability in complex and variable outdoor scenarios, ensuring clear target images of the pole-mounted equipment. Fine-tuning the focus is automatic by camera 1. After the initial image capture, the image quality is judged by an image quality assessment algorithm. If the image is deemed clear, the capture of that area ends; if it is deemed unclear, camera 1's autofocus function is activated to refocus and capture the image. To prevent this process from looping indefinitely, a threshold is used to set the number of refocusing attempts.

[0039] By using real-time image quality assessment and automatic focus adjustment, the problem of image quality being affected by slight shaking of camera 1 during manual operation of insulating rod 4 is avoided.

[0040] S3, Data Processing Stage: After shooting, the image information from camera 1 is transmitted to the processor. The processor has built-in image analysis software that automatically analyzes and processes the captured image information. It can efficiently filter out invalid and unqualified images caused by uncontrollable factors during the image acquisition stage, such as blurred images, image ghosting, and images without target areas; as well as unqualified images with actual defects. First, the image qualification judgment model Model-1 is used for preliminary judgment to filter out invalid and unqualified images with blurred images, image ghosting, and images without target areas, so as to facilitate secondary acquisition of images of these areas later. Then, the abnormal defect automatic detection model Model-2 is used to automatically detect defects such as insulation coating damage, cracks, bubbles, drips, and unevenness on all images except invalid and unqualified images, so as to facilitate effective location and rectification of these areas later.

[0041] After the insulation modification of exposed live points, various objective factors make it difficult to avoid defects such as damage, cracks, bubbles, drips, and unevenness of the insulation coating. The main causes are twofold: the performance of the insulating coating and the spraying process. Insufficient content of binders such as epoxy resin and silicone rubber in the coating formulation leads to poor adhesion between the coating and the metal equipment surface. Alternatively, the presence of oil, rust, dust, or moisture on the equipment surface affects coating adhesion, resulting in easy peeling and damage. Insufficient flexibility / extensibility of the insulating coating or excessively short intervals between multiple spray layers result in a hard, brittle coating prone to cracking. Poor defoaming properties of the insulating coating, insufficient stirring after storage, or air incorporation during dilution lead to bubble formation. Spraying too close a distance or too slow a speed, or excessively thick single sprays, causes the coating to drip due to gravity. Poor pigment dispersion or unstable equipment parameters can also lead to uneven spraying.

[0042] In practical applications, the generation of insulation coating defects is random and uncertain, leading to an imbalance in the number of samples of various defects in the collected data. Given the diverse, random, and uncertain causes of these defects, their characteristics are also varied, resulting in problems such as indistinct defect features and variable defect morphologies. The specific workflow for the data processing stage is as follows: S3.1 Model Training Phase: First, the collected sample images after insulation modification are classified and labeled into qualified and unqualified images. Then, algorithms such as linear transformation, Gamma correction, and histogram equalization are used to expand the samples by changing the image brightness. At the same time, the defect types and defect locations in the unqualified images are labeled, and image samples are expanded by methods such as scaling, rotation, flipping, and copy-paste augmentation.

[0043] Then, based on the classified and labeled qualified and unqualified image samples, a deep learning model with a YOLOv5 network architecture was used for fine-tuning and training of the qualification judgment model, resulting in the image qualification judgment model Model-1. This model can effectively filter out blurry images, ghost images, and unqualified images without target parts. Based on the labeled abnormal defect samples, a deep learning model with a YOLOv8 network architecture was used for fine-tuning and training of the abnormal defect automatic detection model, resulting in the abnormal defect automatic detection model Model-2. This model can effectively detect and locate bubbles, uneven defects, and small dripping targets with indistinct features.

[0044] In the aforementioned judgment model, the deep learning model based on the YOLOv5 network architecture introduces a Swing Transformer module and inserts a CBAM convolutional block attention module into the backbone network to enhance feature extraction and key feature response capabilities. At the same time, the last layer of the convolutional layer is replaced with a depthwise separable convolutional layer to reduce the number of parameters, improve inference speed, and ensure model accuracy. The improved YOLOv5 deep learning architecture network model can improve inference speed while maintaining accuracy.

[0045] This deep learning model, based on the YOLOv8 network architecture, introduces a weighted bidirectional feature pyramid (BiFPN) to replace the PAN-FPN to improve the detection capability of small objects. A dynamic label assignment strategy is used with TOOD (Task-aligned One-stage Object Detection) to align classification and regression tasks, improving the matching quality of positive samples. Simultaneously, Alpha-IoU is employed to power-law transform the IoU loss, enhancing the gradient for difficult samples. During model training, a fusion of Cosine LR and Warmup is used to smooth learning rate changes and avoid early oscillations. The improved YOLOv8 deep learning architecture network model can improve inference speed while maintaining accuracy, and can detect small objects relatively well.

[0046] To address the imbalance between positive and negative samples, reduce the weight of easily classified samples, and improve the detection and localization accuracy of targets with varied shapes and small targets, the improved YOLOv8 deep learning model employs an improved loss function during training, namely: ; in, For classification loss; For bounding box regression loss; For confidence loss; , and The weights are dynamic, and The weights can be adjusted based on the training stage and sample difficulty; for example, increasing the weights of the classification loss in the early stages of model training. Increase the weight of the bounding box regression loss in the later stage. Confidence loss weights during training Remain unchanged. (Through) , and The combined effect of these three parts can improve the model's localization accuracy and category recognition ability.

[0047] Due to the performance limitations of insulating coatings, defects such as breakage, cracks, bubbles, drips, and unevenness may occur during actual implementation. Different defects have different characteristics. In this embodiment, classification and labeling based on defect characteristics can better achieve the identification and classification of different defect features, which is beneficial for ultimately determining the defect type. Furthermore, it allows for increased weighting of the bounding box regression loss in later stages. This ensures testing accuracy. Specifically: Classification loss ,in, It's a real label. It is a prediction probability. ,in, The weights of positive and negative samples are balanced, with a default value of 0.25; γ is a hyperparameter that adjusts the weights of hard samples, with a default value of 2. This represents the model's predicted probability for positive samples. Focal Loss can reduce the loss contribution of easily classified samples, allowing the model to focus more on difficult-to-distinguish samples during training, thus addressing the imbalance between positive and negative samples and increasing the weight of difficult samples.

[0048] Bounding box regression loss ,in, , This is the intersection-union ratio (IU) between the predicted bounding box and the ground truth bounding box. The squared Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; The diagonal length of the smallest bounding rectangle that covers both frames; This is a weighting coefficient used to control the impact of aspect ratio loss; The aspect ratio consistency adjustment factor measures the difference between the width and height ratios. CIoU uses complex logarithmic operations when calculating the aspect ratio penalty, which can lead to instability in the training process. Therefore, by introducing center point distance and aspect ratio consistency penalty terms on top of IoU loss, the difference between the predicted box and the ground truth box can be more comprehensively measured, thereby improving the accuracy of model detection.

[0049] In this embodiment, by introducing a penalty term for center point distance and aspect ratio consistency, the difference between the predicted box and the true box can be measured more comprehensively. By reducing the determination error of the boundaries of defects such as damage, cracks, bubbles, drips and unevenness, the accuracy of predicting each defect is improved, as well as the classification accuracy of defects such as damage, cracks, bubbles, drips and unevenness is improved.

[0050] An adjustable power parameter α is introduced to improve the performance of the object detection model by dynamically adjusting the loss weights corresponding to different IoU values. , where parameters The model can be flexibly controlled for different Sensitivity to value, when When the model is more sensitive to the loss of high IoU targets, it encourages fine-tuning of the predicted boxes to align with the true boxes; when At high IoU, the model is more sensitive to loss for low IoU targets, which can balance positive and negative samples. Regional settings This helps improve positioning accuracy; low Regional settings This is used to alleviate the problem of imbalanced samples.

[0051] Due to traditional The loss is insensitive to shape differences; by introducing a shape similarity factor, the detection capability for complex-shaped targets can be improved. ,in Used for balance And the weight of shape similarity, Shape similarity is calculated by the difference in aspect ratio. ,in The width and height of the prediction box. The actual width and height of the bounding box. Hyperparameters for controlling shape similarity sensitivity.

[0052] During the model training phase, a combination of Cosine LR and Warmup is used to smooth the learning rate changes. In the initial training phase (the first N steps), the learning rate is linearly / exponentially increased from 0 to the initial value. To avoid gradient explosion caused by random parameter initialization, Cosine Logistic Regression (CLR) is used to decay the learning rate. Assume the total number of training steps is... Warmup steps are Then the learning rate By constructing a full-cycle learning rate scheduling strategy from "smooth start" to "smooth decay," Warmup can effectively avoid initial parameter oscillations and alleviate instability in the early stages of training. Simultaneously, utilizing a cosine curve, which is smoother than step decay, can reduce convergence fluctuations caused by sudden changes in the learning rate later on. Controlling the minimum learning rate can prevent premature entrapment in local optima.

[0053] S3.2, Model Application Stage: First, a pass / fail judgment model is used to determine the pass / fail status of the images collected during acceptance. Then, an automatic defect detection model is used to automatically detect and identify abnormal defects such as damage, cracks, bubbles, drips, and unevenness. Finally, the judgment and detection results of the two models are fused. When both models determine that the image is passable and there are no abnormal defects, the algorithm determines that it is passable; otherwise, it is considered unqualified. This process ultimately determines whether the appearance meets the requirements. By fusing the pass / fail judgment model with the automatic defect detection model, the false negative rate can be greatly reduced. Combined with manual review, the false positive rate can be significantly reduced while improving work efficiency.

[0054] S4. Acceptance and Judgment Stage: Based on the results of automatic image analysis and processing, and in conjunction with relevant power equipment insulation acceptance standards, images deemed unqualified during the data processing stage undergo manual review. The final determination of whether the insulation modification is qualified is made after review. If the image shows a good insulation layer appearance and all key indicators meet the standard requirements, the acceptance is deemed qualified; otherwise, the acceptance is deemed unqualified, and the location and reason for the unqualified are recorded in detail. Optional acceptance methods include: S4.1 Appearance Inspection: Based on a method that integrates multiple deep learning detection models, the system automatically detects and identifies defects in the images collected for acceptance testing. It automatically marks the insulation layer surface for smoothness, flatness, and the presence of obvious damage, cracks, bubbles, drips, and unevenness. Then, a manual review process is used to verify the accuracy of the automatic detection results, the tightness of the connection between the insulation layer and the equipment, and the absence of gaps or detachment. This allows for a rapid and accurate judgment of whether the insulation layer's appearance is acceptable. This combination of automatic detection and manual review reduces the false negative and positive rates, ensuring accuracy. Furthermore, the accumulation of review image data from the automatic detection results can improve the model's accuracy, ultimately achieving fully automated detection.

[0055] 4.2 Acceptance by voltage detector: 4.2.1 Acceptance Preparation: Personnel requirements: Acceptance personnel should possess a 10kV live-line working certificate for distribution networks, be familiar with the use of 10kV detectors and related electrical safety knowledge, and be in good health without any diseases that would hinder high-voltage work.

[0056] Tools Preparation: Prepare a qualified 10kV voltage detector, insulating gloves, insulating boots, a safety helmet, a multimeter, a thermometer and hygrometer, and other protective equipment. A qualified 10kV voltage detector refers to one that has been periodically calibrated and is within its validity period.

[0057] Documentation preparation: Collect relevant documents such as quality certificates, construction records, and self-inspection reports for the insulating spray coating materials to ensure that the documents are complete and accurate.

[0058] Site conditions: Before acceptance, ensure that the equipment has completed insulation spraying, the spraying area is clean and dry, there are no obstacles that may affect acceptance, the ambient temperature is between 5℃ and 40℃, and the relative humidity is not greater than 80%.

[0059] 4.2.2 Voltage Detector Inspection: Before using a 10kV voltage detector for acceptance testing, perform a self-test on a live 10kV device to confirm that the detector's audible and visual alarms are functioning correctly. Wear insulated gloves and boots, hold the insulated handle of the voltage detector, and gradually bring the detector's probe closer to the insulating sprayed area, keeping the probe perpendicular to the sprayed surface. Start with a distance of 10cm and slowly decrease the distance. When the distance reaches 2-3cm, observe whether the voltage detector emits an audible and visual alarm signal. At least 10 different locations should be selected for testing in each sprayed area. Result Judgment: If the voltage detector does not emit an audible and visual alarm signal during the test, the insulation performance at that test point is qualified; if the voltage detector emits an audible and visual alarm signal at any test point, the insulation performance of that area is deemed unqualified.

[0060] 4.3 Acceptance Criteria: If the appearance inspection, insulation performance test and other inspection items are all qualified, the insulation spraying of the 10kV power distribution exposed point equipment is deemed to be qualified, and the acceptance personnel should sign the acceptance report to confirm.

[0061] If any inspection item fails to meet the requirements, a rectification notice should be issued immediately, specifying the rectification content and deadline. After the construction unit completes the rectification, it should reapply for acceptance until it passes the acceptance.

[0062] Example 2: This embodiment provides a system for determining the insulation of exposed live points of pole-mounted equipment, including: The lighting and capture mode determination module is configured to: acquire the light intensity of the equipment on the column, determine the light intensity, and determine the lighting and capture mode based on the light intensity determination result; The image acquisition module is configured to: acquire image information of the pole-mounted device using a defined lighting and image capture mode; and to judge the image quality using a multi-feature linear weighted fusion strategy and adjust the focal length based on the judgment result. The acceptance module is configured to: perform a qualification judgment on the image information of the column-mounted equipment to obtain a qualified image; for the qualified image, perform abnormal defect detection and classification using a preset abnormal defect automatic detection model; and perform acceptance based on the abnormal defect detection results and classification results; wherein, the loss in the abnormal defect automatic detection model is equal to the weighted sum of classification loss, bounding box regression loss and confidence loss.

[0063] The working method of the system is the same as the insulation determination method for live exposed points of pole-mounted equipment in Embodiment 1, and will not be repeated here.

[0064] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the pole-mounted equipment live exposed point insulation determination method described in Embodiment 1.

[0065] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the pole-mounted equipment live exposed point insulation determination method described in Embodiment 1.

[0066] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the pole-mounted equipment live exposed point insulation determination method described in Embodiment 1.

[0067] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for determining whether to insulate a live exposed point of a pole-mounted device, characterized by, The method comprises the following steps: acquiring the illumination intensity of the column device, determining the illumination intensity, and determining the illumination and image capturing mode according to the illumination intensity determination result; acquiring the image information of the column device by using the determined illumination and image capturing mode; judging the image quality by using a multi-feature linear weighted fusion strategy, and adjusting the focal length according to the judgment result; judging the eligibility of the image information of the column device to obtain qualified images; detecting and classifying abnormal defects of the qualified images by using a preset abnormal defect automatic detection model; and performing acceptance according to the abnormal defect detection result and the classification result; wherein the loss in the abnormal defect automatic detection model is equal to the weighted sum of the classification loss, the bounding box regression loss and the confidence loss.

2. The method of claim 1, wherein the method is characterized by: The illumination intensity determination model is a yolov5 deep learning model; in the yolov5 deep learning model, EfficientNetV2 is used to replace the backbone network, and the last two layers of the convolution layer are replaced by depth separable convolution layers to reduce the parameter amount.

3. The method of claim 2, wherein the method is characterized by: When the illumination condition is low light, the camera used to acquire the image information of the column device selects a light compensation mode; when the illumination condition is high light, the camera selects a strong light suppression mode; and when the illumination condition is normal light, the camera selects a default parameter mode. When the illumination condition is non-uniform light, the camera selects a wide dynamic mode.

4. The method of claim 1, wherein the method is characterized by: Adjusting the focal length according to the judgment result of the image quality comprises: ； wherein, is the final quality score; , and are weight coefficients; , and are image quality scores, respectively; when the final quality score is greater than a threshold value, it indicates that the image is not clear, and automatic fine-tuning of the focal length is required until the final quality score is not greater than the threshold value, while setting a threshold value for the number of times of focusing.

5. The method of claim 1, wherein the method is characterized by: When performing the eligibility judgment, the unqualified images such as blurred images, ghost images and images without target parts are screened out based on a preset judgment model; wherein the judgment model is a deep learning model based on the yolov5 network architecture.

6. The method of claim 5, wherein the method is characterized by: In the judgment model, a Swin Transformer module is introduced into the backbone network, and a convolution block attention module is inserted to enhance the feature extraction capability and key feature response capability; and the last layer of the convolution layer is replaced by a depth separable convolution layer to reduce the parameter amount.

7. The method of claim 5, wherein the method is characterized by: The abnormal defect automatic detection model is a deep learning model based on the yolov8 network architecture; a weighted bidirectional feature pyramid is introduced into the yolov8 network architecture, and a dynamic label allocation strategy is used to classify and regress tasks.

8. The method of claim 7, wherein the method is characterized by: The classification loss is: ； wherein, is the true label, is the predicted probability; wherein, balances the weights of positive and negative samples, and γ is a hyperparameter to adjust the weight of difficult samples, is the predicted probability.

9. The point-on-the-ground live exposure determination method of claim 7, wherein, The bounding box regression loss is: ； wherein, , is the intersection over union of the predicted box and the ground truth box; is the squared Euclidean distance between the centers of the predicted box and the ground truth box; is the length of the diagonal of the minimum enclosing rectangle covering both boxes; is a weight coefficient; is a length-width ratio consistency adjustment factor; An adjustable power parameter a is introduced to improve the performance of the target detection model by dynamically adjusting the loss weight corresponding to different IoU values; a shape similarity factor is introduced to improve the detection ability of complex shape targets, wherein, to balance the weights of and shape similarity, shape similarity is calculated by the height-width ratio difference; wherein, is the height-width ratio of the predicted box, is the height-width ratio of the ground truth box, is a hyperparameter for controlling the sensitivity of shape similarity.

10. A live bare point insulation determination system for a pole-mounted device, characterized by, The method comprises the following steps: The illumination and image capturing mode determination module is configured to acquire the illumination intensity of the column device, determine the illumination intensity, and determine the illumination and image capturing mode according to the illumination intensity determination result; The image acquisition module is configured to acquire the image information of the column device by using the determined illumination and image capturing mode; and judge the image quality by using a multi-feature linear weighted fusion strategy, and adjust the focal length according to the judgment result; The acceptance module is configured to judge the eligibility of the image information of the column device to obtain qualified images; For qualified images, abnormal defect detection and classification are performed by using a preset abnormal defect automatic detection model; acceptance is performed according to the abnormal defect detection result and the classification result; wherein the loss in the abnormal defect automatic detection model is equal to the weighted sum result of the classification loss, the boundary box regression loss and the confidence loss.

Images