Distribution line insulator partial discharge defect inversion identification method

Abstract

The invention relates to a distribution line insulator partial discharge defect inversion identification method. The method comprises the following steps: S1, establishing a small sample data set; s2, adding the same condition model to a generator and a judger of the recursive residual generative adversarial network R3GAN to obtain a CRGAN network; s3, performing iterative training on the CRGAN network, and after training is completed, supplementing the generated samples to the small sample data set to obtain an expanded data set; s4, a SeaFormer module is introduced into the YOLOv8seg network, an SYOLOv8seg network is constructed, and an improved loss function is set; and S5, inputting the expanded data set into the SYOLOv8seg network, and training the SYOLOv8seg network in combination with the improved loss function to obtain a distribution line insulator partial discharge defect inversion identification model. Compared with the prior art, the method has the advantages of realizing high-precision identification and inversion evaluation in a small sample environment and the like.

Description

Technical Field

[0001] This invention relates to the field of insulator inversion technology, and in particular to a method for inverting and identifying partial discharge defects in distribution line insulators. Background Technology

[0002] Insulators in power distribution lines bear the critical responsibility of isolating high-voltage currents and supporting conductors, and their working condition directly affects the safe and stable operation of the power grid. However, because insulators are exposed to the outdoor environment for a long time, they are susceptible to the effects of wind, rain, snow, high temperatures, dust, and other factors, making them prone to defects such as damage, contamination, foreign object adhesion, and partial discharge. In severe cases, this can lead to flashover faults and even cause widespread power outages.

[0003] Current insulator inspection methods mainly rely on manual line patrols or ground-based handheld devices. These methods suffer from drawbacks such as low efficiency, potential for missed detections, high labor intensity, and high safety risks. In recent years, although drone inspections have begun to be applied to power line inspection operations, they are still primarily limited to image acquisition and have not achieved efficient and accurate automated identification and intelligent assessment. Especially in the area of ​​partial discharge detection, traditional methods such as ultraviolet imaging still rely on manual interpretation, which is insufficient to meet the requirements of modern power distribution systems for refined and intelligent management.

[0004] In practical engineering applications, due to the relatively low probability of discharge defect events, coupled with the limitations of weather, flight conditions, time windows, and safety regulations on on-site data acquisition, the number of ultraviolet images with defect characteristics that can be obtained is very limited, often resulting in only small sample datasets. Small sample datasets are common in power line inspection tasks, which not only leads to insufficient model training data but also easily results in imbalanced class distribution, making it difficult to meet the requirements of high-precision identification and inversion. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying and retrieving partial discharge defects in power distribution line insulators to achieve high-precision identification and inversion evaluation under small sample conditions.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for inverting and identifying partial discharge defects in insulators of power distribution lines, the method comprising the following steps: S1. Use a drone equipped with an ultraviolet imager system to collect ultraviolet image data of insulators under different discharge states, perform grayscale conversion and noise reduction processing, and establish a small sample dataset; S2. Add the same conditional model to the generator and judge of the recursive residual generative adversarial network R3GAN to obtain the CRGAN network. S3. Based on the forward modeling law of ultraviolet imaging, the classification of discharge intensity by the light spot area threshold range is determined and used as the input of the classification label. The CRGAN network is iteratively trained. After the training is completed, the generated samples are added to the small sample dataset to obtain the expanded dataset. S4. Introduce the SeaFormer module into the YOLOv8_seg network, construct the S_YOLOv8_seg network, and set the improved loss function; S5. Input the expanded dataset into the S_YOLOv8_seg network, train the S_YOLOv8_seg network with the improved loss function to obtain the inversion identification model of partial discharge defects in distribution line insulators, and perform actual inversion identification of partial discharge defects in distribution line insulators based on the inversion identification model of partial discharge defects in distribution line insulators.

[0007] Furthermore, the ultraviolet imager system carried by the UAV includes a UAV platform, a ground control system, and a data processing system.

[0008] Furthermore, the specific steps for grayscale image conversion and denoising are as follows: The ultraviolet image data of the insulator was converted into a grayscale image, and noise was removed using a 3×3 median filter.

[0009] Furthermore, the SeaFormer module is introduced into the YOLOv8_seg network. The specific steps for constructing the S_YOLOv8_seg network are as follows: The S_YOLOv8_seg network is obtained by replacing the YOLOv8-seg backbone module with the SeaFormer module.

[0010] Furthermore, the improved loss function is as follows: in, To improve the loss function, To compare the expected loss with the expected loss, The minimum point distance intersection-union ratio loss.

[0011] Furthermore, the expected intersection-union loss is: For prediction boxes and truth boxes; For intersection, union, and comparison; Euclidean distance squared; The coordinates of the center points of the prediction box and the truth box; To predict the width and height of the bounding box; Truth box width and height; Let the width and height be the minimum bounding rectangle; The length of the diagonal of the minimum bounding rectangle that covers both frames.

[0012] Furthermore, the minimum point distance intersection-union ratio loss is: .

[0013] Furthermore, the classification labels include discharge intensity level labels.

[0014] Furthermore, a discharge intensity rating label of "Good" indicates a continuous discharge area < 250; a discharge intensity rating label of "Weak Discharge" indicates a continuous discharge area that meets the 250 standard. <Area <30000; Discharge intensity rating label indicates continuous discharge area when flashover> 30000.

[0015] Furthermore, a mini-batch training mode is adopted during the iterative training of the CRGAN network.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The CRGAN model designed in this invention is based on adversarial generative network, combined with regularized relative loss function and conditional model, which can generate high-quality defect samples, further improving the generalization ability and recognition accuracy of the model; the partial discharge defect inversion network adopts S_YOLOv8_seg network as defect recognition and segmentation module, and introduces SeaFormer module to improve feature extraction capability, realizing the best trade-off between segmentation accuracy and inference delay, which is suitable for actual intelligent inspection and online status monitoring of power distribution lines; (2) This invention proposes a combination and loss function This effectively improves the boundary quality of discharge spot segmentation in the S_YOLOv8_seg network, resulting in faster and smoother convergence. (3) The partial discharge defect inversion network designed in this invention achieves high-precision identification and morphological inversion in one step. The model can not only identify but also invert defect features. Attached Figure Description

[0017] Figure 1 : A schematic diagram of the overall process of the method of this invention; Figure 2 : A structural diagram of the image acquisition system of this invention; Figure 3 Flowchart of ultraviolet image preprocessing; Figure 4 : S_YOLOv8_seg network structure diagram; Figure 5 : Structure diagram of the augmentation model for small sample ultraviolet image data; Figure 6 : Structure diagram of partial discharge defect inversion model. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0019] Achieving accurate identification and reliable inversion of partial discharge defects in insulators under small sample conditions has become a pressing technical challenge. This invention addresses this problem by proposing a comprehensive method combining dataset expansion, identification network improvement, and inversion modeling, enabling high-precision identification and inversion evaluation in small sample environments.

[0020] This invention proposes a method for inverting and identifying partial discharge defects in power distribution line insulators based on a small sample dataset, belonging to the field of intelligent detection and condition monitoring technology for power equipment. The methods include: using a drone equipped with an ultraviolet imager to collect ultraviolet image data of insulators under different discharge states to establish a small sample dataset; building a CRGAN network based on a recursive residual generative adversarial network by adding the same conditional model to the generator and judge to improve the quality and stability of the generated images; determining the classification of discharge intensity based on the forward modeling law of ultraviolet imaging by the threshold range of spot area, using it as the input of classification labels, and continuously iterating and training the CRGAN model. When the training is stable, the generated samples are supplemented to the original small sample dataset as needed to establish an expanded dataset; introducing the SeaFormer module into the YOLOv8_seg network to construct the S_YOLOv8_seg network, proposing a loss function suitable for discharge spot segmentation, improving the model recall and Dice coefficient, and realizing the inversion and identification of partial discharge defects; training the model with the expanded dataset, validating it with the original dataset, and outputting the spot area and discharge intensity level to achieve accurate identification and inversion of partial discharge defects in insulators.

[0021] Compared with existing technologies, the method of this invention achieves high-precision identification and inversion evaluation of partial discharge defects in distribution line insulators under small sample conditions, significantly improving identification accuracy and inversion performance, and can be widely applied to intelligent inspection and online condition monitoring of distribution lines.

[0022] A method for inverting and identifying partial discharge defects in distribution line insulators based on a small sample dataset includes the following steps: S1. Using a drone equipped with an ultraviolet imager, collect ultraviolet image data of insulators under different discharge states, perform grayscale conversion and noise reduction processing, and establish a small sample dataset. S2. Based on the R3GAN network, the same conditional model is added to the generator and the judge to build the CRGAN network, which improves the quality and stability of the generated images. S3. Iterate and train the CRGAN model continuously. Once the training is stable, supplement the original small sample dataset with the generated samples as needed to generate an expanded dataset. S4. The SeaFormer module is introduced into the YOLOv8_seg network to construct the S_YOLOv8_seg network, and a combination of... and loss function This improves the model recall and Dice coefficient, enabling the inversion and identification of partial discharge defects. S5. Train the model using the expanded dataset, test the model using the original dataset, and output the spot area and discharge intensity level.

[0023] Compared with the prior art, the present invention has the following advantages: (1) The CRGAN model designed in this invention is based on adversarial generative network, combined with regularized relative loss function and conditional model, which can generate high-quality defect samples, further improving the generalization ability and recognition accuracy of the model. (2) The partial discharge defect inversion network designed in this invention uses the S_YOLOv8_seg network as the defect identification and segmentation module, and introduces the SeaFormer module to improve the feature extraction capability, thus achieving the best trade-off between segmentation accuracy and inference delay. It is suitable for actual intelligent inspection and online status monitoring of power distribution lines. (3) This invention proposes a combination and loss function This effectively improves the boundary quality of discharge spot segmentation in the S_YOLOv8_seg network, resulting in faster and smoother convergence. (4) The partial discharge defect inversion network designed in this invention achieves high-precision identification and morphological inversion in one step. The model can not only identify but also invert defect features.

[0024] like Figure 1 As shown, a method for inverting and identifying partial discharge defects in distribution line insulators based on a small sample dataset is proposed, aiming to solve the problems of data scarcity, insufficient model generalization ability, and poor real-time performance in partial discharge detection of distribution line insulators. This method combines UAV ultraviolet image acquisition, data preprocessing, CRGAN data augmentation, and a lightweight insulator partial discharge defect inversion network model, making it suitable for intelligent inspection and online monitoring of actual distribution lines.

[0025] 1. Establish an ultraviolet image acquisition system, such as... Figure 2As shown, an inspection of a 10 kV distribution line is being conducted. The image acquisition system includes: a ground control system for UAV flight control and inspection route planning; a UAV platform supporting preset path flight and obstacle avoidance functions; an ultraviolet imaging device capable of detecting very weak corona discharge phenomena, with adaptable focusing distance and variable gain; and a data processing system that acquires ultraviolet image data in real time and wirelessly transmits it to the ground workstation for analysis and processing.

[0026] 2. Perform the following preprocessing operations on the acquired raw ultraviolet images to create a small sample dataset, such as... Figure 3 As shown.

[0027] (1) Grayscale image: Converting ultraviolet images into grayscale images (2) Noise removal: Use 3×3 median filtering to remove noise; in, The original image at pixel position The value, The pixel values ​​of the filtered image. For The system uses a 3×3 window centered on the neighborhood, where median is the median grayscale value of the neighboring pixels.

[0028] 3. To address the problem of small sample data, a CRGAN model is designed based on the R3GAN structure, such as... Figure 4 As shown, discharge intensity classification labels were set based on the threshold range of discharge spot area determined by the forward modeling law of ultraviolet imaging. Five types of insulators with different insulation states (normal, good, microporous, deteriorated, and severely aged) were selected in the field, and the spot area was extracted. Through MATLAB digital simulation, the trend of spot area variation with insulation state was obtained: good (<250 pixels); weak discharge (250~30000 pixels); flashover (>30000 pixels). An augmented dataset based on traditional data augmentation and the discharge intensity classification labels were used as input to the CRGAN model for offline learning, enabling the CRGAN generator network to fit the mapping relationship between partial discharge intensity and different spot areas well.

[0029] Based on the R3GAN network, the RpGAN+R1+R2 loss function is introduced. When R1 and R2 are used simultaneously, even in the early stages of training, the gradient norms of the discriminator on real and pseudo data tend to be equal. The combined use of R1 and R2 can effectively suppress overfitting of the discriminator. in, These are the parameters of the discriminator; is the penalty coefficient, used to control the strength of the R1 penalty term; x is the actual sample data; This represents the true data distribution. Let x be the gradient of the discriminator at the real data x. in, These are the parameters of the generator; Pseudo-data generated by the generator; The pseudo-data distribution generated by the generator; For the discriminator in pseudo data The gradient at that point. 4. In the mini-batch (batchsize=4) training mode, the CRGAN model was trained for 200 rounds to achieve a high-quality UV image dataset expansion based on the original data.

[0030] 5. After data augmentation, design a partial discharge defect inversion model, such as... Figure 5 As shown, defect inversion is performed. The S_YOLOv8seg network is used to segment and classify the partial discharge region. It is more computationally efficient and achieves an optimal trade-off between segmentation accuracy and inference latency, balancing speed and accuracy, which is beneficial for model training and performance improvement.

[0031] To solve the loss function To address the issues of incomplete alignment with the segmentation target and gradient saturation in high IoU regions, a combined approach is proposed. (Expected Intersection over Union) and Loss function for (Minimum Point Distance Intersection over Union) . The tighter and more accurate bounding box results in less mask clipping error, effectively improving boundary quality while alleviating gradient saturation and leading to faster and smoother convergence.

[0032] loss function The formula is shown below: in, For prediction boxes and truth boxes; For intersection, union, and comparison; Euclidean distance squared; The coordinates of the center points of the prediction box and the truth box; To predict the width and height of the bounding box; Truth box width and height; Let the width and height be the minimum bounding rectangle; The length of the diagonal of the minimum bounding rectangle that covers both frames.

[0033] 6. Based on the partial discharge defect inversion model, an area decision interval curve is constructed, and the continuous discharge area is divided into three intensity level labels to indicate the severity of discharge in the insulator. The discharge intensity level labels include: I: Good (Area < 250), II: Weak discharge (Area < 250). <Area <30000, ), Ⅲ: Flashover (Area> (30000). Based on the discharge area, the system automatically determines the region to which it belongs and assigns a corresponding discharge intensity level label, outputting the qualitative and quantitative inversion results of the partial discharge defect, thus achieving an accurate assessment that integrates identification and inversion.

[0034] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for inverting and identifying partial discharge defects in insulators of power distribution lines, characterized in that, The method includes the following steps: S1. Use a drone equipped with an ultraviolet imager system to collect ultraviolet image data of insulators under different discharge states, perform grayscale conversion and noise reduction processing, and establish a small sample dataset; S2. Add the same conditional model to the generator and judge of the recursive residual generative adversarial network R3GAN to obtain the CRGAN network. S3. Based on the forward modeling law of ultraviolet imaging, the classification of discharge intensity by the light spot area threshold range is determined and used as the input of the classification label. The CRGAN network is iteratively trained. After the training is completed, the generated samples are added to the small sample dataset to obtain the expanded dataset. S4. Introduce the SeaFormer module into the YOLOv8_seg network, construct the S_YOLOv8_seg network, and set the improved loss function; S5. Input the expanded dataset into the S_YOLOv8_seg network, train the S_YOLOv8_seg network with the improved loss function to obtain the inversion identification model of partial discharge defects in distribution line insulators, and perform actual inversion identification of partial discharge defects in distribution line insulators based on the inversion identification model of partial discharge defects in distribution line insulators.

2. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 1, characterized in that, The UAV-equipped ultraviolet imager system includes a UAV platform, a ground control system, and a data processing system.

3. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 1, characterized in that, The specific steps for grayscale image conversion and noise reduction are as follows: The ultraviolet image data of the insulator was converted into a grayscale image, and noise was removed using a 3×3 median filter.

4. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 1, characterized in that, The specific steps for constructing the S_YOLOv8_seg network by introducing the SeaFormer module are as follows: The S_YOLOv8_seg network is obtained by replacing the YOLOv8-seg backbone module with the SeaFormer module.

5. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 1, characterized in that, The improved loss function is: in, To improve the loss function, To compare the expected loss with the expected loss, The minimum point distance intersection-union ratio loss.

6. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 5, characterized in that, The expected crossover ratio loss is: For prediction boxes and truth boxes; For intersection, union, and comparison; Euclidean distance squared; The coordinates of the center points of the prediction box and the truth box; To predict the width and height of the bounding box; Truth box width and height; Let the width and height be the minimum bounding rectangle; The length of the diagonal of the minimum bounding rectangle that covers both frames.

7. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 6, characterized in that, The minimum point distance intersection-union ratio loss is: 。 8. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 1, characterized in that, The classification labels include discharge intensity level labels.

9. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 8, characterized in that, A discharge intensity rating label of "Good" indicates a continuous discharge area <250; a discharge intensity rating label of "Weak Discharge" indicates a continuous discharge area that meets the 250 standard. <Area <30000; Discharge intensity rating label indicates continuous discharge area when flashover> 30000.

10. The method for inverting and identifying partial discharge defects in distribution line insulators according to claim 1, characterized in that, Mini-batch training mode is used during the iterative training of the CRGAN network.

Images