YOLOv10-based 10kV pole-mounted equipment defect detection method

Abstract

The invention provides a 10kV pole-mounted equipment defect detection method based on YOLOv10, and the method comprises the steps: constructing two sample data sets of 10kV pole-mounted equipment and comparison before and after equipment defect maintenance, and constructing a YOLOv10 and improved YOLOv10 cascaded neural network structure; compared with the YOLOv10 baseline model, the improved YOLOv10 baseline model is additionally provided with an AKV composite module and a dynamic feature sampling module; using YOLOv10 to train a data set of the equipment on the 10kV column; training an equipment insulation damage defect comparison data set by using the improved YOLOv10; the detected defects are dynamically tracked by integrating DeepSort, and spraying maintenance is performed by using an unmanned aerial vehicle; and the improved YOLOv10 is used for detecting the maintenance effectiveness for the maintained defects. And the power grid is helped to detect defects more accurately and quickly and maintain the defects in time, so that the workload and risk of manual inspection are reduced.

Description

Background Art

[0002] In recent years, with the accelerated construction of the new power system, the distribution network is gradually transforming from a traditional power distribution network to a distributed intelligent grid with interactive power sources, grids, loads, and energy storage. The distribution network under the new power system is not only a channel for power distribution but also an important platform for realizing energy transformation, improving power supply reliability, promoting user interaction, and intelligent management. As the core component of the medium-voltage distribution network, 10kV pole-mounted equipment undertakes the important functions of power transmission and distribution. However, it is long-term exposed to the outdoor environment and is vulnerable to factors such as bad weather and biological gnawing, resulting in defects such as damaged insulation skins and exposed metals. This seriously affects the quality of power supply and increases the risks of electrical faults and power outages. Such defects not only significantly reduce the power supply quality but may also trigger chain electrical faults, threatening the safe and stable operation of the power grid. Therefore, how to achieve efficient and accurate detection of pole-mounted equipment defects has become a technical problem urgently to be solved in the field of intelligent grid operation and maintenance.

[0003] The traditional manual inspection method is limited by subjective judgment and has limitations such as a high false detection rate and a high risk of missed detection. In recent years, the integration of deep learning technology and drone collaborative detection has provided new ideas for power grid defect identification, and detection methods based on neural networks have become a research hotspot in power grid maintenance. For example, the patent with the publication number CN118351420A detects transmission line insulators by lightweight improving the YOLO algorithm; for example, the patent with the publication number CN118505672A adds an SA attention module to the neck network structure of YOLOv8, uses the attention mechanism to efficiently capture global information, and strengthens the network's attention to small insulation defect targets; for example, the patent with the publication number CN116128820A establishes an anchor-free end-to-end object detection architecture by introducing a decoupled detection head (DecoupledHead) and various data augmentations (DataAug), improving the detection accuracy. However, all of the above methods improve the performance of defect detection by continuously enhancing the model's information extraction ability. Although they have achieved great advantages compared with manual inspection, they lack pertinence for the special working conditions of the 10kV pole-mounted equipment to be detected in this invention. Summary of the Invention

[0004] The main purpose of this invention is to provide a 10kV pole-mounted equipment defect detection method based on YOLOv10 to effectively solve the above problems mentioned in the background art.

[0005] The technical solution of this invention is as follows:

[0006] A 10kV pole-mounted equipment defect detection method based on YOLOv10 is proposed, and this method includes the following steps:

[0007] S1. Take pictures of the 10kV pole-mounted equipment through drone inspections, and respectively construct a pole-mounted equipment dataset and a comparison dataset before and after defect repair;

[0008] S2. Build a YOLOv10 neural network model and an improved YOLOv10 neural network model;

[0009] S3. Use the pole-mounted equipment dataset to train the YOLOv10 neural network model, and use the comparison dataset before and after defect repair to train the improved YOLOv10 neural network model;

[0010] S4. Use the YOLOv10 neural network model to identify the pole-mounted equipment in the pictures of the pole-mounted equipment transmitted back by the drone, and use the improved YOLOv10 neural network model to perform dynamic tracking detection of insulation breakage defects on the identified pole-mounted equipment;

[0011] S5. Use the drone to spray and repair the detected defects, and use the improved YOLOv10 neural network model to detect whether the spraying is effective for the repaired defects.

[0012] A further improvement of the present invention is that the improved YOLOv10 neural network model in S2 includes an AKV composite module added to the original YOLOv10 baseline model.

[0013] A further improvement of the present invention is that the improved YOLOv10 neural network model in S2 further includes a DySample dynamic upsampling module.

[0014] A further improvement of the present invention is that the AKV composite module combines AKConv with Vanilla Net, and Vanilla Net is only composed of convolutional layers and pooling layers, and replaces the second convolutional layer in AKConv and Vanilla Net.

[0015] A further improvement of the present invention is that the dynamic tracking detection of insulation breakage defects in S4 uses the DeepSort object tracking algorithm.

[0016] A further improvement of the present invention is that the improved YOLOv10 neural network model in S2 performs point sampling calculation using the DySample dynamic upsampling module; the specific steps of the point sampling calculation are as follows:

[0017] S201. Input a feature tensor X of size C×H×W, and given an upsampling scale factor s, calculate an offset O of size 2gs2×H×W through a linear layer with input and output channels of C and 2gs2 respectively, and the expression is:

[0018] O = linear(X);

[0019] S202. Pixel recombination is performed to obtain O with a high resolution of 2g×sH×sW, and the elements with an offset of G are added to obtain δ. The expression is:

[0020] δ = O + G;

[0021] S203. The Grid_sample function is used to map the elements at each position in the point sampling set δ with a size of 2gs×H1×W1 to the input feature tensor X with a size of 2g×H1×W1, and resample it to the output tensor X' with a size of C×H2×W2. The expression is:

[0022] X' = grid_sample(X, δ).

[0023] A further improvement of the present invention is that the DeepSort target tracking algorithm includes the following specific steps:

[0024] Target detection initialization: In the first frame of the video sequence, the cascaded YOLOv10 algorithm is used to detect defective targets, and initial target trajectories are generated according to the flight state of the drone. The Kalman filter is used to predict the target trajectory and predict the possible positions of the defective targets in the next frame;

[0025] Association matching: The predicted trajectories are divided into two categories: confirmed status and unconfirmed status. The detection results of the second frame are associated and matched with the trajectories in the confirmed status. By calculating the appearance similarity and motion consistency of the targets, the successfully matched trajectories, unmatched detection boxes, and unmatched trajectory boxes are obtained; the successfully matched trajectories are updated through the Kalman filter and integrated into the defective trajectory set;

[0026] DIoU association degree matching: For the unmatched detection boxes and trajectory boxes, the DIoU method is used for secondary association. The calculation formula of DIoU is:

[0027]

[0028] DIoU loss = 1 - DIoU;

[0029] Among them, b represents the predicted box, b gt represents the ground truth box, c represents the length of the diagonal of the minimum circumscribed rectangle of the two boxes; ρ 2 (b, b gt ) represents the Euclidean distance between the center points of the two boxes;

[0030] Processing of unmatched detection boxes: New trajectories are generated for the unmatched detection boxes and confirmed. If it is confirmed as a defective target, it is updated to the existing trajectories;

[0031] Unmatched trajectory box processing: Further confirm the unmatched trajectory boxes. If it is confirmed that the box does not belong to the defective target, delete it; if it is a defective target, retain the trajectory and match it with the detection results of subsequent frames. If the match is successful within 30 frames, update the trajectory; if the match fails, delete the saved trajectory.

[0032] The technical effects of the present invention are as follows:

[0033] A 10kV pole-mounted equipment defect detection method based on YOLOv10 is constructed. The present invention proposes an improved 10kV pole-mounted equipment defect cascade detection method for YOLOv10 to solve the problems of difficult positioning of pole-mounted equipment in complex dynamic backgrounds and difficult capture of insulation breakage defect information. This method is divided into two stages: In the first stage, the YOLOv10 detection algorithm is used to quickly identify and extract various pole-mounted equipment, and the extracted pole-mounted equipment pictures are used as the input of the second stage; in the second stage, the improved YOLOv10 is used to accurately detect the insulation breakage defects of the equipment output in the first stage. To address the problems of diverse defect morphologies and difficult extraction of defect features, an AKV module and a dynamic feature sampling (DySample) module are designed to enhance the model's perception ability of multi-scale defects and the focusing accuracy of small target areas respectively; to address the possible defect loss problems during the UAV flight detection process, an improved DeepSort algorithm is fused to achieve robust tracking of dynamic defects. Finally, to verify the effectiveness of the spraying repair of the experimental equipment, a comparison dataset is constructed based on the color feature differences before and after spraying. The method in this paper can accurately detect and repair the 10kV pole-mounted equipment defect problems, providing double support for the theoretical innovation and technical practice of intelligent inspection of the distribution network. Description of the Drawings

[0034] By reading the following detailed description of the non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0035] Figure 1 It is a flowchart of the present invention;

[0036] Figure 2 It is a structural diagram of the improved YOLOv10 neural network model in Embodiment 1 of the present invention;

[0037] Figure 3 It is a structural diagram of the Vanilla Net in Embodiment 1 of the present invention;

[0038] Figure 4 It is a structural diagram of the AKConv fused Vanilla Net in Embodiment 1 of the present invention;

[0039] Figure 5 It is a schematic diagram of the initial sampling coordinates of AKConv in Embodiment 1 of the present invention;

[0040] Figure 6 Schematic diagram of DySample point sampling calculation for Embodiment 1 of the present invention;

[0041] Figure 7 Schematic diagram of the upsampling module of DySample for Embodiment 1 of the present invention;

[0042] Figure 8 Flow chart of dynamic tracking of insulation breakage defects for Embodiment 1 of the present invention;

[0043] Figure 9 Schematic diagram of the DIoU principle for Embodiment 1 of the present invention;

[0044] Figure 10 Picture of the pole-mounted equipment after maintenance spraying for Embodiment 1 of the present invention;

[0045] Figure 11 Schematic diagram of inspection of maintenance effect for Embodiment 1 of the present invention. Detailed implementation manners

[0046] The present invention aims to propose a 10kV pole-mounted equipment defect detection method based on YOLOv10. In view of the problems that it is difficult to locate pole-mounted equipment in a complex dynamic background and it is difficult to capture insulation breakage defect information, the present invention proposes an improved cascaded detection method for 10kV pole-mounted equipment defects based on YOLOv10. This method is divided into two stages: in the first stage, the YOLOv10 detection algorithm is used to quickly identify and extract various pole-mounted equipment, and the extracted pictures of pole-mounted equipment are used as the input of the second stage; in the second stage, the improved YOLOv10 is used to accurately detect the insulation breakage defects of the equipment in the pictures output by the first stage. In view of the problems of diverse defect morphologies and difficult extraction of defect features, the AKV module and the dynamic feature sampling (DySample) module are designed to strengthen the model's perception ability of multi-scale defects and the focusing accuracy of small target areas respectively; in view of the possible defect loss problems in the process of drone flight detection, the improved DeepSort algorithm is fused to achieve robust tracking of dynamic defects. Finally, to verify the effectiveness of the spraying maintenance of the experimental equipment, a comparison data set is constructed based on the color feature differences before and after spraying. The method in this paper can accurately detect and repair the 10kV pole-mounted equipment defect problems, providing double support for the theoretical innovation and technical practice of intelligent inspection of the distribution network.

[0047] Embodiment 1:

[0048] This embodiment proposes a 10kV pole-mounted equipment defect detection method based on YOLOv10, as Figures 1-11 shown, including the following specific steps:

[0049] S1. Take pictures of 10kV pole-mounted equipment through drone inspections, and respectively construct a pole-mounted equipment dataset and a comparison dataset before and after defect repair;

[0050] S2. Build a YOLOv10 neural network model and an improved YOLOv10 neural network model;

[0051] S3. Use the pole-mounted equipment dataset to train the YOLOv10 neural network model, and use the comparison dataset before and after defect repair to train the improved YOLOv10 neural network model;

[0052] S4. Use the YOLOv10 neural network model to identify the pole-mounted equipment in the pictures of the pole-mounted equipment transmitted back by the drone, and use the improved YOLOv10 neural network model to dynamically track and detect the insulation breakage defects of the identified pole-mounted equipment;

[0053] S5. Use the drone to spray and repair the detected defects, and use the improved YOLOv10 neural network model to detect whether the spraying is effective for the repaired defects.

[0054] In this embodiment, 4234 images of 10kV pole-mounted equipment are obtained through drone inspections. The image resolution is adjusted to 1152*864, and data augmentation is performed through color adjustment, adding complex weather, and geometric transformation, increasing the number of images to 5081. The so-called geometric transformation includes, but is not limited to, translation, rotation, and cropping. Then, 3381 images before and after the insulation breakage defect repair are selected from them, and all the images are marked using the Labelimg tool to obtain the corresponding XML files, establishing a pole-mounted equipment dataset and a comparison dataset including before and after the insulation breakage defect repair.

[0055] In this embodiment, a neural network model cascading YOLOv10 and the improved YOLOv10 is constructed; the neural network model of the improved YOLOv10 adds a composite module of AKV (AKConv and Vanilla Net) and a dynamic feature sampling (DySample) module to the original YOLOv10 baseline model, as Figure 2 .

[0056] It should be noted that the present invention proposes an AKV module based on the combination of Adaptive Kernel Convolution (AKConv) and Vanilla Net. The Vanilla Net network structure is as Figure 3As shown, its core feature is that it consists only of convolutional layers and pooling layers. This simple design significantly reduces the number of model parameters and computational complexity. In the initial stage of training, Vanilla Net replaces a single convolutional layer with two convolutional layers containing activation functions. As the number of training cycles increases, these activation functions are gradually simplified into identity mappings. At the end of training, these two convolutional layers can be quickly merged to shorten the inference time. However, the 1*1 square convolutional kernel used in Vanilla Net can only extract features from local regions, making it difficult to capture global context information, and its sampling pattern is fixed and lacks flexibility. To better adapt to defect detection, in this paper, AKConv is used to replace the second convolutional layer in Vanilla Net, as Figure 4 shown. AKConv introduces a flexible convolution method that allows the convolutional kernel to have a variable number of parameters. This design enables the convolutional kernel to adaptively adjust its shape according to the characteristics of the image task and target features, thus breaking through the operation limit of traditional convolutional kernels that are only limited to fixed local regions. When dealing with defect detection tasks, the convolutional kernel of AKConv can use a new coordinate generation algorithm to automatically adjust its sampling shape, as Figure 5 shown.

[0057] This algorithm can generate initial sampling coordinates for convolutional kernels of different sizes and shapes, providing the initial sampling shape of an irregular convolutional kernel to more accurately match and cover the irregular regions of the target defect. Through this design, AKConv can more effectively adapt to complex and variable image features. In addition, AKConv also allows the number of convolutional kernel parameters to be adjusted according to requirements, which enables AKConv to improve performance by reducing redundant parameters according to hardware conditions and task requirements. It not only reduces the pressure on computing and storage resources but also speeds up the inference speed of the model.

[0058] In addition, in this embodiment, a dynamic feature sampling (DySample) module is also designed to guide the sampling points to focus on the defect area to achieve pixel-level feature optimization;

[0059] In the task of detecting defects in pole-mounted equipment, the insulated damaged area at the connection usually appears as a target with a relatively small pixel proportion in the image, making it difficult to extract features with sufficient discriminability. In addition, due to the dynamic change in the distance between the drone during inspection and the pole-mounted equipment, the scale of the defect points in the image will also change significantly, which poses higher requirements for the multi-scale small target adaptation ability of the detection model. To address the above problems, the present invention introduces the DySample dynamic upsampling module, which can adaptively generate dynamic interpolation weights based on the local information of the input feature map to achieve more accurate feature reconstruction. Compared with the previously used nearest neighbor interpolation method, DySample effectively avoids the boundary imbalance problem caused by the single interpolation method, shows higher smoothness when dealing with feature boundaries, and significantly reduces the boundary blurring phenomenon. At the same time, the dynamic adjustment characteristic of DySample further enhances the ability to capture detailed features, especially showing excellent performance in the small target segmentation task. The DySample sampling process is as Figure 6 、 7 shown. As Figure 6 , first, calculate the point sampling set δ according to the input feature tensor X; then, input δ and the input feature tensor X into the upsampling module together, as Figure 7 and obtain the entire output feature tensor X' through upsampling calculation.

[0060] Among them, the calculation process of δ is as Figure 6 shown: First, input the feature tensor X with a size of C×H×W, and given the upsampling ratio factor s, calculate the offset O with a size of 2gs2×H×W through a linear layer with input and output channels of C and 2gs2 respectively, and the expression is as formula (1):

[0061] O = linear(X) (1)

[0062] Then, perform pixel recombination to add the high-resolution O with a size of 2g×sH×sW and the element with an offset of G to obtain δ, as formula (2):

[0063] δ = O + G (2)

[0064] Figure 7 The calculation process for DySample is as follows: The Grid_sample function maps the elements at each position in the point sampling set δ with a size of 2gs×H1×W1 to the input feature tensor X with a size of 2g×H1×W1 and resamples them to the output tensor X' with a size of C×H2×W2, as formula (3):

[0065] X' = grid_sample(X, δ) (3)

[0066] The DySample dynamic sampling module is used to replace the original upsampling module, which can effectively guide the sampling points to focus on the defective target area and significantly improve the accuracy of defect detection.

[0067] To verify the effectiveness of the improved modules in this paper, ablation experiments are conducted on the YOLOv10 baseline model, and the experimental results are shown in Table 1:

[0068] Table 1 Ablation Experiment Results

[0069]

[0070]

[0071] The experimental results show that the improved modules in this paper significantly improve the performance of the model. After adding the AKV module, the mean average precision mAP@0.5 of the model is improved by 0.7% compared with the baseline model. In addition, the AKV module effectively reduces redundant parameters through an adaptive mechanism, increasing the FPS from 86 frames per second to 93 frames per second. After replacing UpSample with DySample, the mAP@0.5 of the model is improved by 4.5%, enhancing the model's ability to extract features from small target defect areas. When adding and using both the AKV module and DySample, the mAP@0.5 of the model is improved by 8.7% and the FPS is also relatively fast. This experimental result fully demonstrates the effectiveness of the module improvement.

[0072] In this implementation, a detection and repair integrated UAV platform is adopted. To effectively solve the problem of target loss during the UAV flight detection of pole-mounted equipment defects, the DeepSort (Deep SimpleOnline and Realtime Tracking) object tracking algorithm based on deep learning is introduced. This algorithm extracts the appearance features of the target through a deep convolutional neural network and combines the motion features of the target for data association, thus significantly improving the accuracy and robustness of object tracking. The specific implementation process is as Figure 8 :

[0073] (1) Object detection initialization: In the first frame of the video sequence, the cascaded YOLOv10 algorithm is used to detect defective targets, and the initial target trajectory is generated according to the flight state of the UAV. Subsequently, the Kalman filter is used to predict the target trajectory to estimate the possible position of the defective target in the next frame.

[0074] (2) Association matching: Divide the predicted trajectories into two categories: confirmed status and unconfirmed status. Associate and match the detection results of the second frame with the trajectories in the confirmed status. By calculating the appearance similarity and motion consistency of the targets, obtain the successfully matched trajectories, unmatched detection boxes, and unmatched trajectory boxes. The successfully matched trajectories are updated through the Kalman filter and integrated into the defective trajectory set.

[0075] (3) IoU (Intersection over Union) matching: For the unmatched detection boxes and trajectory boxes, use the IoU method for secondary association. The IoU value reflects the overlapping degree of the two bounding boxes, and the larger the value, the higher the matching probability. Screen out the new matching results through a preset IoU threshold, and update them to the defective trajectories using the Kalman filter; those that are not successfully matched enter the subsequent processing flow.

[0076] (4) Processing of unmatched detection boxes: When new targets appear in the detection boxes or targets reappear after being occluded for a long time, it may cause continuous matching failures of the detection boxes. In response to this situation, the system will generate new trajectories for the unmatched detection boxes and confirm them. If it is confirmed as a defective target, update it to the existing trajectories.

[0077] (5) Processing of unmatched trajectory boxes: Considering that there may be a small number of missed detections in the cascade detection algorithm, it is necessary to further confirm the unmatched trajectory boxes. If it is confirmed that the box does not belong to a defective target, delete it directly; if it is a defective target, retain the trajectory and match it with the detection results of subsequent frames. If the match is successful within 30 frames, update the trajectory; if the match fails, delete the saved trajectory.

[0078] To further improve the pertinence and tracking accuracy of the algorithm, this paper uses DIoU (Distance-IoU) to replace the traditional IoU (Intersection over Union) measurement method. Specifically, in the third step of the target tracking process, for the association problem of unmatched detection boxes and trajectory boxes, the traditional DeepSort algorithm only calculates based on the overlapping area of the bounding boxes, ignoring the distance information between the center points of the bounding boxes. Therefore, this paper introduces DIoU as a new association metric, and its calculation principle is as Figure 9 shown.

[0079] Among them, the calculation formula of DIoU is as follows:

[0080]

[0081] DIoU loss = 1 - DIoU (5)

[0082] In the formula: b represents the predicted box ( Figure 9Black rectangular part); b gt Represents the real box ( Figure 9 Gray rectangle part); c represents the diagonal length of the minimum circumscribed rectangle of the two frames; ρ 2 (b,b gt ) represents the Euclidean distance between the center points of the two boxes. DIoU can more accurately reflect the intersection of the two boxes by comprehensively considering the overlapping area and center point distance of the bounding boxes. Applying it to the loss function calculation of the network can directly minimize the center point distance between the predicted box and the true box, thereby accelerating the convergence of the model and improving the robustness of target tracking.

[0083] The present invention conducts comparative experiments after integrating the DeepSort target tracking algorithm, as shown in Table 2:

[0084] Table 2 Target tracking comparison experimental results

[0085]

[0086] The experimental results show that the target tracking accuracy (MOTA) and tracking precision (MOTP) increased by 2% and 3.2% respectively after the improved YOLOv10 and DeepSort method was adopted. After DIoU optimization, MOTA and MOTP increased by 1.3% and 0.9% respectively. Therefore, the improved method improves the accuracy of defect tracking of pole-mounted equipment and better meets the task requirements.

[0087] In this embodiment, the drone used is equipped with a special insulating coating, which will be sprayed to repair the defect after it is detected. Figure 10 As shown in the figure, after the insulating coating covers the defect area and solidifies, its surface appears white, which is significantly different from the color characteristics before spraying. To verify the effectiveness of spraying, this paper conducts comparative learning training on the cascade network model, and the visualization results are shown in the figure. Figure 11 As shown in the figure. In the figure, RDQ represents fuse, ZK represents pole-mounted equipment, ZB represents pole-mounted transformer, PTQ represents defect before spraying, and PTH represents defect after spraying. Practice has proved that the cascade algorithm first detects the pole-mounted equipment and cuts off most of the background interference, and then performs defect repair detection on the pole-mounted equipment. This method has high detection accuracy and fast detection rate.

[0088] Embodiment 2:

[0089] This embodiment provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor; the processor executes the above-mentioned 10kV pole-mounted equipment defect detection method based on YOLOv10 by calling the computer program stored in the memory.

[0090] The electronic device can vary significantly due to different configurations or performances, and can include one or more processors (Central Processing Units, CPUs) and one or more memories. Among them, at least one computer program is stored in the memory, and this computer program is loaded and executed by the processor to implement the 10kV pole-mounted equipment defect detection method based on YOLOv10 provided by the above method embodiment. The electronic device can also include other components for implementing the device functions. For example, the electronic device can also have components such as wired or wireless network interfaces and input / output interfaces for data input and output. This embodiment will not be elaborated here.

[0091] Those skilled in the art know that the present invention can be implemented as a system, a method, or a computer program product. Therefore, the present disclosure can be specifically implemented in the following forms: it can be completely hardware, can be completely software (including firmware, resident software, microcode, etc.), or can also be a combination of hardware and software, which is generally referred to as "circuit", "module", or "system" in this article. In addition, in some embodiments, the present invention can also be implemented in the form of a computer program product in one or more computer-readable media, and the computer-readable media contains computer-readable program code.

[0092] Any combination of one or more computer-readable media can be adopted. The computer-readable media can be a computer-readable signal medium or a computer-readable storage medium. More specific examples (non-exhaustive list) of the computer-readable storage medium include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, the computer-readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or combined with an instruction execution system, device, or device.

[0093] The present invention is described with reference to the flowcharts and block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process or block in the flowchart and block diagram can be implemented by computer program instructions, and the combination of the processes and blocks in the flowchart or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate for implementing in the process Figure 1 one process or multiple processes and blocks Figure 1means for the functions specified in one or more boxes.

[0094] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide for implementing in the process Figure 1 one process or more processes and boxes Figure 1 steps for the functions specified in one box or more boxes.

[0095] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative rather than restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit and scope protected by the claims of the present invention. All of these fall within the protection scope of the present invention.

Claims

1. A 10kV pole-mounted equipment defect detection method based on YOLOv10, characterized by: The specific steps include: S1. Take pictures of 10kV pole-mounted equipment through drone inspections, and build a pole-mounted equipment dataset and a comparison dataset before and after defect repair. S2. Build and improve the YOLOv10 neural network model; S3. Use the pole-mounted equipment dataset to train the YOLOv10 neural network model, and use the defect repair before and after comparison dataset to train and improve the YOLOv10 neural network model; S4. Use the YOLOv10 neural network model to identify the pole-mounted equipment in the pole-mounted equipment images sent back by the drone, and use the improved YOLOv10 neural network model to dynamically track and detect insulation damage defects of the identified pole-mounted equipment; S5. Use drone spraying to repair the detected defects and use the improved YOLOv10 neural network model to detect whether the spraying is effective.

2. The 10kV pole equipment defect detection method based on YOLOv10 according to claim 1 is characterized in that: The improved YOLOv10 neural network model in S2 includes an AKV composite module added to the original YOLOv10 baseline model.

3. The 10kV pole equipment defect detection method based on YOLOv10 according to claim 2 is characterized in that: The improved YOLOv10 neural network model in S2 further includes a DySample dynamic upsampling module.

4. The 10kV pole equipment defect detection method based on YOLOv10 according to claim 2 is characterized in that: The AKV composite module combines AKConv with Vanilla Net, wherein the Vanilla Net consists of only convolutional layers and pooling layers, and replaces the second convolution layer in AKConv and Vanilla Net.

5. The 10kV pole equipment defect detection method based on YOLOv10 according to claim 4 is characterized in that: The dynamic tracking detection of insulation damage defects in S4 adopts the DeepSort target tracking algorithm.

6. The 10kV pole equipment defect detection method based on YOLOv10 according to claim 3 is characterized in that: The improved YOLOv10 neural network model in S2 uses the DySample dynamic upsampling module to perform point sampling calculation; the specific steps of the point sampling calculation are: S201, input feature tensor X of size C×H×W, and given upsampling factor s, calculate offset O through linear layer with input and output channels C and 2gs2 respectively, with size 2gs2×H×W, expressed as: O = linear(X); S202, the pixel reorganization obtains the high-resolution size of 2g×sH×sW, and the element O with the offset G is added to obtain δ, which is expressed as: δ = O + G; S203, using the Grid_sample function, maps the elements at each position in the point sampling set δ of size 2gs×H1×W1 to the input feature tensor X of size 2g×H1×W1, and resamples it to the output tensor X' of size C×H2×W2, expressed as: X'=grid_sample(X,δ).

7. The 10kV pole equipment defect detection method based on YOLOv10 according to claim 5 is characterized in that: The DeepSort target tracking algorithm includes the following specific steps: Target detection initialization: In the first frame of the video sequence, the cascaded YOLOv10 algorithm is used to detect defective targets and generate the initial target trajectory according to the flight status of the drone. The Kalman filter is used to predict the trajectory of the target and predict the possible position of the defective target in the next frame; Association matching: The predicted trajectory is divided into two categories: confirmed state and unconfirmed state. The detection result of the second frame is associated with the trajectory in the confirmed state. By calculating the appearance similarity and motion consistency of the target, the matched trajectory, unmatched detection frame and unmatched trajectory frame are obtained. The successfully matched trajectory is updated through the Kalman filter and integrated into the defect trajectory set; DIoU correlation matching: For the detection boxes and trajectory boxes that have not been successfully matched, the DIoU method is used for secondary correlation. The calculation formula of DIoU is: God loss =1-DIoU; Among them, b represents the prediction box, b gt represents the real box, c represents the diagonal length of the minimum circumscribed rectangle of the two boxes; ρ 2 (b,b gt ) represents the Euclidean distance between the center points of the two frames; Unmatched detection frame processing: Generate a new trajectory for the unmatched detection frame and confirm it. If it is confirmed to be a defect target, update it to the existing trajectory; Processing of unmatched trajectory frames: Further confirm the unmatched trajectory frames. If it is confirmed that the frame does not belong to a defective target, delete it; if it is a defective target, retain the trajectory and match it with the detection results of subsequent frames. If the match is successful within 30 frames, update the trajectory; if the match fails, delete the saved trajectory.

Images