Defect detection method, system, identification terminal and unmanned aerial vehicle for power distribution line

Abstract

The application discloses a power distribution line defect detection method and system, a recognition terminal and a UAV, and belongs to the technical field of power distribution line defect detection. The inspection image of the power distribution line is detected by a defect detection model based on a YOLO network. The YOLO network determines the number of positive samples by using an intersection-over-union ratio, and screens the positive samples and negative samples by using a weighted loss, so that the distribution speed of the positive and negative samples is improved, the training time is shortened, and the parameters that need to be additionally optimized are avoided.

Description

Technical Field

[0001] This invention relates to the field of power distribution line defect detection technology, specifically to power distribution line defect detection methods, systems, identification terminals, and drones. Background Technology

[0002] With social development and technological advancements, the demand for electricity is increasing daily. Power distribution lines, especially overhead lines in the distribution network, are a crucial component of electricity consumption, and their safety and stability directly impact the normal operation of the power system. However, distribution lines are widely distributed, and during long-term continuous operation, faults such as broken strands or corrosion of conductors, hardware, or insulators can lead to large-scale power outages. To ensure the safe and stable operation of overhead distribution lines, regular inspections are necessary. Traditional distribution network inspections primarily rely on on-site surveys by personnel. Due to the long inspection lines, high workload, and limitations imposed by geographical conditions such as mountainous areas or construction sites, the quality and efficiency of inspections are very low.

[0003] With the maturity of drone technology, using drones to replace manual inspections can improve the speed and efficiency of power maintenance and repair, and reduce the labor intensity of workers. In particular, drones have the characteristics of fast line inspection speed and instantaneous emergency response, which can effectively ensure the safe operation of power distribution network lines. At present, drones store the inspection images or videos captured by cameras on SD cards, and perform identification and analysis on the stored images after the inspection is completed. On the one hand, the timeliness of inspection is low. If an emergency such as a line break or fire occurs, the fault can only be identified and located after the drone has finished flying, which is not timely enough to ensure the safety of the line in real time. On the other hand, the analysis models used in the existing inspection image analysis methods require a long training time. Summary of the Invention

[0004] To address the aforementioned technical problems in the existing technology, this invention provides a method, system, identification terminal, and drone for detecting defects in power distribution lines. By using a positive and negative sample allocation method, the computational speed of training is improved and the training time is shortened.

[0005] This invention discloses a defect detection method for power distribution lines. The method includes: acquiring inspection images or videos of power distribution lines; acquiring a defect detection model based on a YOLO network; detecting defects in the inspection images or videos using the defect detection model to obtain defects in the power distribution lines; wherein, the positive and negative sample allocation method of the YOLO network includes: obtaining prior boxes in the training set; matching the prior boxes with ground truth boxes; obtaining the prior boxes of the cell containing the ground truth box and its adjacent cells as candidate boxes; obtaining the intersection-union ratio (IU / I) of the ground truth boxes and the candidate boxes; sorting the candidate boxes in descending order according to the IU / I, and taking the top P IU / I; summing the top P IU / I and rounding down to obtain the number of positive samples M; obtaining the weighted loss of the classification loss and regression loss of the candidate boxes; sorting the weighted loss in ascending order, taking the top M candidate boxes as positive samples, and the remaining candidate boxes as negative samples.

[0006] Preferably, the case where a candidate box is removed as a positive sample of multiple real target boxes; Images or videos of power distribution lines are collected by drones.

[0007] Preferred methods for removing a candidate bounding box from being a positive sample of multiple true target bounding boxes include: Determine whether the candidate box is a positive sample of two or more real target boxes; If so, obtain the weighted loss between the candidate box and the actual target box; Based on the weighted loss, the true target boxes that match the candidate boxes are determined, and the positive sample relationships between the candidate boxes and the remaining true target boxes are removed.

[0008] The preferred crossover-union ratio is expressed as: in, IoU Represented as intersection-union ratio, A This is represented as a candidate box. D Represented as the true target bounding box; The weighted loss is expressed as: in, Cost Indicated as weighted loss, L_coord Represented as regression box loss L_cls Represented as classification loss; Loss function of defect detection model loss Represented as: in, L_conf This is expressed as confidence loss. in,N It is the sample size. It is the size of the feature map. B It is the predicted border number for each cell. It is the first indicator function, used to represent the first... i The first sample j The first grid k The border is the target box. , For the first i The center coordinates of the true bounding box of each sample , They are the first i The width and height of the true border of each sample. , For the first i The first sample j The first grid k The predicted coordinates of each border. , It is the first i The first sample j The first grid k The predicted width and height of each border; in, It is the first i The first sample j The first grid k The confidence prediction value of each bounding box. It is the second indicator function, used to represent the first i The first sample j The first grid k The border is the background border, and λ is the weight; in, c For category number, C It is the number of categories. It is the first i The first sample c Real labels for each category It is the first i The first sample c Predicted labels for each category.

[0009] Preferred methods for obtaining prior bounding boxes in the training set include: The K-means algorithm is used to cluster the bounding boxes in the training set to obtain multiple prior boxes arranged from smallest to largest.

[0010] Preferred methods for training the defect detection model include: Acquire inspection images of power lines and annotate the inspection images to obtain a dataset; The dataset is augmented by random scaling, random cropping, and random arrangement. The dataset is randomly distributed to obtain a training set, a validation set, and a test set; A defect detection model is obtained by training a training set based on a lightweight YOLOv7 network with an attention mechanism. The defect detection model was validated using a validation set. The defect detection model was tested using a test set.

[0011] Preferably, the YOLO network uses a lightweight YOLOv7 network with an attention mechanism. The lightweight YOLOv7 network includes a basic feature extraction module, an ELANs module, an attention module, a CMC module, and an MP module. The basic feature extraction module includes a convolutional layer, a normalization layer, and an activation function; The ELANs module includes a first basic feature extraction module, a second basic feature extraction module, a third basic feature extraction module, a fourth basic feature extraction module, and a fifth basic feature extraction module. The inputs of the ELANs module are connected to the first basic feature extraction module and the fourth basic feature extraction module, respectively. The first basic feature extraction module, the second basic feature extraction module, and the third basic feature extraction module are connected in series. The outputs of the first basic feature extraction module, the second basic feature extraction module, the third basic feature extraction module, and the fourth basic feature extraction module are connected to the fifth basic feature extraction module after feature fusion. The CMC module includes a basic feature extraction module. The output of the basic feature extraction module and the results of its max pooling operations at various sizes are fused together and then connected to another basic feature extraction module. The CBC module includes a basic feature extraction module and a CMC module. The CBC input is processed by the CMC module and the basic feature extraction module, respectively, and then the input is fused before being processed by another basic feature extraction module. The input of the attention module is connected to the upstream ELANs module or the upstream CBC module through the basic feature extraction module. The output of the attention module is fused with the input features of the downstream ELANS module and then sent to the downstream ELANS module. The MP module includes a first branch and a second branch. The first branch includes two basic feature extraction modules connected in sequence, and the second branch includes a max pooling layer and a basic feature extraction module. The outputs of the first branch and the second branch are used for feature fusion. The lightweight YOLOv7 network has three branches downstream, with a dominant header at the end of each branch. The branch has an ELANs module, the output of which has an auxiliary head. The auxiliary head is used to assist training by utilizing the loss calculation in the middle of the network. The auxiliary head has more positive samples than the dominant head, and the auxiliary head pays more attention to recall.

[0012] The present invention also provides a system for implementing the above-described defect detection method, comprising a monitoring module, a defect analysis module, and an alarm module. The monitoring module is used to acquire inspection images or videos of power distribution lines; The defect analysis module is used to obtain a defect detection model based on the YOLO network; the defect detection model is used to detect defects in the inspection images or videos to obtain the defects of the power distribution lines. The alarm module is used to generate reports and alarms based on the defects of the power distribution line. The report includes the line name, tower number, ID information of the inspected equipment, and file information attributes.

[0013] The present invention also provides an identification terminal, wherein the identification terminal stores instructions for implementing the defect detection method.

[0014] The present invention also provides a drone, which is used to acquire inspection images or videos of power distribution lines and transmit the inspection images or videos to the identification terminal, which is used to execute the instructions and identify the position of the target object in the inspection image; After receiving the location of the target object in the inspection image sent by the identification terminal, the drone adjusts its flight direction, distance from the target object, and shooting angle; The drone is also used to mount the identification terminal.

[0015] Compared with the prior art, the beneficial effects of the present invention are: the number of positive samples is determined by the crossover and union ratio, and positive and negative samples are screened by weighted loss, thereby improving the allocation speed of positive and negative samples, shortening the training time, and avoiding additional parameters that need to be optimized. Attached Figure Description

[0016] Figure 1 This is a flowchart of the defect detection method for power distribution lines according to the present invention; Figure 2 This is a flowchart of the positive and negative sample allocation method; Figure 3 This is a schematic diagram of a lightweight YOLOv7 network; Figure 4 This is a schematic diagram of the ELANs module; Figure 5This is a schematic diagram of the CMC module; Figure 6 This is a schematic diagram of the CBC module; Figure 7 This is a schematic diagram of the MP module; Figure 8 This is the system logic block diagram of Embodiment 2; Figure 9 This is a diagram illustrating the allocation of positive and negative samples; Figure 10 This is a schematic diagram of the attention module. Detailed Implementation

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

[0018] The present invention will now be described in further detail with reference to the accompanying drawings: YOLO (You Only Look Once) is a unified, real-time object detection network that treats object detection as a single regression task. This application uses a YOLO network to detect defects in power distribution lines.

[0019] A defect detection method for power distribution lines, such as Figure 1 As shown, the method includes: Step 101: Obtain inspection images or videos of the power distribution lines. Inspection images or videos can be captured using cameras installed on the power distribution lines or by drones.

[0020] Step 102: Obtain the defect detection model based on the YOLO network.

[0021] Step 103: Detect the inspection images or videos using a defect detection model to obtain the defects in the power distribution lines.

[0022] Step 104: Generate a report or alarm based on the detected defects. In one specific embodiment, the report includes the line name, tower number, ID information of the inspected equipment, and file information attributes, and generates an alarm when fault information appears in the report.

[0023] In step 102, such as Figure 2 The methods for assigning positive and negative samples in the YOLO network include: Step 201: Obtain the prior bounding boxes of the training set. Specifically, the labeled boxes in the training set can be clustered using the K-means algorithm to obtain multiple prior bounding boxes arranged in ascending order.

[0024] Step 202: Match the prior bounding box with the real target bounding box, and obtain the prior bounding boxes of the cell where the real target bounding box is located and the adjacent cells, as candidate bounding boxes.

[0025] Step 204: Obtain the intersection-union ratio (IU) of the ground truth bounding box and the candidate bounding box. The IU represents the degree of overlap or similarity between the prior bounding box A and the ground truth bounding box D.

[0026] Specifically, the intersection-union ratio is expressed as: (5) in, IoU Represented as intersection-union ratio, A This is represented as a candidate box. D This is represented as the true target bounding box. In one specific embodiment, if the IOU is greater than 0.7, the candidate bounding box can be regarded as the target; if it is less than 0.3, it can be regarded as the background, but it is not limited thereto.

[0027] Step 205: Sort the candidate boxes in descending order according to the intersection-union ratios (IU), and take the first P IU, for example, the first 10 or 20 IU. Alternatively, sort the candidate boxes in ascending order and take the last P IU.

[0028] Step 206: Sum the first P crossovers and comparisons, then round down to obtain the number of positive samples M.

[0029] Step 207: Obtain the weighted loss of the classification loss and regression loss of the candidate boxes.

[0030] The weighted loss is expressed as: (6) in, Cost Indicated as weighted loss, L_coord Represented as regression box loss L_cls This is represented as classification loss.

[0031] Step 208: Sort the weighted loss in ascending order, take the top M candidate boxes as positive samples, and the remaining candidate boxes as negative samples. Where P and M are positive integers.

[0032] Step 209: Remove the case where a candidate box is used as a positive sample of multiple real target boxes.

[0033] The number of positive samples is determined by crossover ratio (CUP), and positive and negative samples are filtered by weighted loss, which improves the speed of positive and negative sample allocation, shortens training time, and avoids additional parameters that need to be optimized.

[0034] In step 209, cases where a candidate box is used as a positive sample for multiple real target boxes can be manually removed; alternatively, the following methods can be used: Step 211: Determine whether the candidate box is a positive sample of two or more real target boxes.

[0035] If so, proceed to step 212: obtain the weighted loss of the candidate box and the real target box.

[0036] Step 213: Based on the weighted loss, determine the true target boxes that match the candidate boxes. For example, the true target box with the smallest weighted loss best matches the candidate box, and the association between this candidate box and the other true target boxes is deleted / removed.

[0037] If not, no further processing will be performed.

[0038] The loss function of the YOLO network or defect detection model is expressed as: (1) in, L_conf This is expressed as confidence loss. L_coord Represented as regression box loss L_cls This is represented as classification loss.

[0039] (2) in, N It is the sample size. It is the size of the feature map. B It is the predicted border number for each cell. It is the first indicator function, used to represent the first... i The first sample j The first grid k Each border is a target box, for example, a target value of 1 and a background value of 0; , For the first i The center coordinates of the true bounding box of each sample , They are the first i The width and height of the true border of each sample. , For the first i The first sample j The first grid k The predicted coordinates of each border. , It is the first i The first sample j The first grid k The predicted width and height of each border; (3) in, It is the first i The first sample j The first grid k The confidence prediction value of each bounding box. It is the second indicator function, used to represent the first i The first sample j The first grid k The border is the background border, and λ is the weight coefficient, which can be set based on experience. Where c is the category number and C is the number of categories. It is the first i The first sample c Real labels for each category It is the first i The first sample c Predicted labels for each category.

[0040] Training methods for defect models include: Step 301: Obtain inspection images of the power lines and annotate the images to obtain a dataset. Annotation can be performed using LabelMe software.

[0041] Step 302: Perform data augmentation on the dataset, which includes splicing the dataset by random scaling, random cropping, and random arrangement.

[0042] Step 303: Randomly distribute the dataset to obtain the training set, validation set, and test set.

[0043] Step 304: Based on the YOLO network, train the model using the training set to obtain a defect detection model. In one specific embodiment, a lightweight YOLOv7 network with an attention mechanism is used.

[0044] Step 305: Validate the defect detection model using a validation set.

[0045] Step 306: Test the defect detection model using the test set.

[0046] Example 1 Figure 3A lightweight YOLOv7 network is shown, which includes a basic feature extraction module, an ELANs module, an attention module, a CMC module, and an MP module.

[0047] The basic feature extraction (CBL) module includes convolutional layers (conv), normalization layers (BN), and activation functions (leaky_relu). The grayscale CBL module employs different convolutional kernels and strides.

[0048] like Figure 4 The ELANs module includes a first basic feature extraction module CBL1, a second basic feature extraction module CBL2, a third basic feature extraction module CBL3, a fourth basic feature extraction module CBL4, and a fifth basic feature extraction module CBL5. The inputs of the ELANs module are connected to the first basic feature extraction module CBL1 and the fourth basic feature extraction module CBL4, respectively. The first basic feature extraction module CBL1, the second basic feature extraction module CBL2, and the third basic feature extraction module CBL3 are connected in series. The outputs of the first basic feature extraction module CBL1, the second basic feature extraction module CBL2, the third basic feature extraction module CBL3, and the fourth basic feature extraction module CBL4 are connected to the fifth basic feature extraction module CBL5 after feature fusion (Concat). The ELANs module is an improvement and refinement of the ELAN module. It combines the extracted features together through four CBL modules. By controlling the shortest and longest gradient paths, the network can learn more features and has stronger robustness. It also reduces network parameters and improves speed while ensuring effective feature extraction. It uses fewer convolutional blocks to control the shortest and longest gradient paths, allowing deeper networks to learn and converge more efficiently.

[0049] like Figure 5 The CMC module includes a basic feature extraction module (CBL). The output of the CBL and the results of max pooling operations of various sizes are fused and then connected to another basic feature extraction module. Specifically, the sizes of the max pooling operations are 5, 9, and 13.

[0050] like Figure 6 The CBC module includes a basic feature extraction module (CBL) and a CMC module. The CBC input passes through both the CMC module and the basic feature extraction module, undergoes feature fusion, and then passes through another basic feature extraction module. The CBC module reduces computational complexity through a lightweight structural design.

[0051] Figure 3In this process, the input of the attention module CBAM is connected to the upstream ELANs module or the upstream CBC module through the basic feature extraction module CBL. The output of the attention module CBAM is fused with the input features of the downstream ELANS module and then sent to the downstream ELANS module.

[0052] like Figure 10 The CBAM module consists of two modules: a channel attention module and a spatial attention module. The channel attention module allows the network to focus on the foreground of the image, making it more attentive to meaningful real-world target regions. The spatial attention module allows the network to focus on locations within the entire image that are rich in contextual information. The channel attention mechanism compresses the feature map in the spatial dimension, resulting in a one-dimensional vector before processing. Specifically, the input features are processed by global max pooling (based on width) and average pooling (based on height) to aggregate the spatial information of the feature map. Then, the data is fed into a shared network (shared fully connected layers) to compress the spatial dimension of the input feature map. Finally, the output features are subjected to element-wise summation and an activation function to generate the feature map for the channel attention mechanism.

[0053] The spatial attention mechanism compresses channels by taking the output feature map of the channel attention mechanism module as input features and performing average pooling (Maxpool) and maximum pooling (Avgpool) on the channel dimension. The results are then concatenated based on the channels. Finally, convolutional dimensionality reduction and activation functions are applied to obtain the feature map of the spatial attention mechanism.

[0054] like Figure 7 The MP module includes a first branch and a second branch. The first branch includes two sequentially connected basic feature extraction modules (CBLs), and the second branch includes a max pooling layer and a basic feature extraction module. The features from the first and second branches are then fused using Concat. The gray CBL module has a stride of 2.

[0055] The YOLO network also includes an upsampling module, Upsample, and a REP module. The Upsample module uses nearest neighbor interpolation for upsampling. Figure 3 The lightweight YOLOv7 network is also shown to have three branches downstream, each branch ending with a dominant head. Each branch contains ELANs modules, and the outputs of the ELANs modules have auxiliary heads. These auxiliary heads are used to assist training by utilizing loss calculations in the middle of the network. During positive sample allocation, the auxiliary head receives more positive samples than the dominant head (e.g., 10 positive samples for the dominant head and 20 positive samples for the auxiliary head), and the auxiliary head focuses more on recall. Figure 9The positive sample allocation for the auxiliary head and the dominant head is shown: the positive samples for the dominant head are the prior box containing the center point of the real target box and the two prior boxes adjacent to the center point; the positive samples for the auxiliary head are increased by four prior boxes, which is twice that of the main and auxiliary heads.

[0056] There is a certain difference between training and inference. During training, the output of the three branches is summed. During inference / inspection image detection, the parameters of the branches are reparameterized onto the main branch.

[0057] The training parameters were set as follows: each input image was 640×640 pixels, the initial learning rate was 0.001, the cosine annealing decay strategy was used to update the learning rate of gradient descent, and 50 images were trained per batch.

[0058] Based on the lightweight YOLOv7 network of this embodiment, a defect detection model is trained on the training set to obtain a defect detection model. Testing shows that the defect detection model can accurately and in real-time identify power distribution line defects in images containing various complex backgrounds, such as precisely identifying insulator damage and flashover defects.

[0059] Example 2 This embodiment provides a system for implementing the above-described defect detection method, such as... Figure 8 It includes monitoring module 1, defect analysis module 2, and alarm module 3. Monitoring module 1 is used to acquire inspection images or videos of power distribution lines; The defect analysis module 2 is used to acquire a defect detection model based on the YOLO network; and to detect defects in the inspection images or videos by using the defect detection model. The alarm module 3 is used to generate reports and alarms based on the defects of the power distribution line. The reports include the line name, tower number, ID information of the inspected equipment, and file information attributes.

[0060] Example 3 This embodiment provides an identification terminal and a drone. The identification terminal stores instructions or programs for implementing the above-described defect detection method and is capable of executing the instructions or programs.

[0061] The drone is used to acquire inspection images or videos of power distribution lines and transmit the inspection images or videos to the identification terminal. The identification terminal is used to execute the instructions and identify the location of the target object in the inspection image. After receiving the location of the target object in the inspection image sent by the identification terminal, the drone adjusts its flight direction, distance from the target object, and shooting angle.

[0062] The identification terminal can be mounted on the UAV, facilitating real-time fault analysis and alarm generation for power distribution lines. The UAV / identification terminal uploads the analysis results to the cloud / server via a wireless network, where the results are visualized. The UAV should be capable of automatic flight control, such as speed control, direction change control, and distance control, and should be equipped with a high-precision RTK or BeiDou positioning device to accurately locate and plan the inspection flight route, collecting inspection images and video information from different perspectives and along different routes.

[0063] In one specific embodiment, aerial data from the drone is transmitted to the identification terminal via the Onboard SDK. The identification terminal then transmits the detected defect and fault image and video stream data to the cloud platform via 4G / 5G communication. This effectively ensures reliable communication for defect detection during drone inspections.

[0064] The lightweight YOLOv7 network of this invention can extract rich features, significantly improving the efficiency and accuracy of power distribution line image detection and recognition. At the same time, the lightweight structure reduces the power consumption of the front-end recognition terminal and extends the inspection time of the UAV.

[0065] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting defects in power distribution lines, characterized in that, The method includes: Obtain inspection images or videos of power distribution lines; Obtain a defect detection model based on the YOLO network; The defects in the power distribution lines are obtained by detecting the inspection images or videos using a defect detection model. The positive and negative sample assignment methods of the YOLO network include: Obtain the prior bounding boxes of the training set; Match the prior bounding box with the actual target bounding box; Obtain the prior bounding boxes of the cell containing the actual target bounding box and its adjacent cells, and use them as candidate bounding boxes; Obtain the intersection-union ratio (IUU) of the ground truth bounding boxes and candidate bounding boxes; Based on the intersection-union ratio, the candidate boxes are sorted in descending order, and the top P intersection-union ratios are taken. Sum the first P intersections and comparisons and round down to obtain the number of positive samples M. Obtain the weighted loss of the classification loss and the regression loss of the candidate bounding box; The weighted loss is sorted in ascending order, and the top M candidate boxes are taken as positive samples, while the remaining candidate boxes are taken as negative samples, where P and M are positive integers; Removing a candidate box from being a positive sample of multiple ground truth boxes includes: determining whether the candidate box is a positive sample of two or more ground truth boxes; if so, obtaining the weighted loss between the candidate box and the ground truth boxes; determining the ground truth boxes that match the candidate box based on the weighted loss, and removing the positive sample relationship between the candidate box and the remaining ground truth boxes. The YOLO network described uses a lightweight YOLOv7 network with an attention mechanism. The lightweight YOLOv7 network includes a basic feature extraction module, an ELANs module, an attention module, a CMC module, and an MP module. The basic feature extraction module includes a convolutional layer, a normalization layer, and an activation function; The ELANs module includes a first basic feature extraction module, a second basic feature extraction module, a third basic feature extraction module, a fourth basic feature extraction module, and a fifth basic feature extraction module. The inputs of the ELANs module are connected to the first basic feature extraction module and the fourth basic feature extraction module, respectively. The first basic feature extraction module, the second basic feature extraction module, and the third basic feature extraction module are connected in series. The outputs of the first basic feature extraction module, the second basic feature extraction module, the third basic feature extraction module, and the fourth basic feature extraction module are connected to the fifth basic feature extraction module after feature fusion. The CMC module includes a basic feature extraction module. The output of the basic feature extraction module and the results of its max pooling operations at various sizes are fused together and then connected to another basic feature extraction module. The CBC module includes a basic feature extraction module and a CMC module. The CBC input is processed by the CMC module and the basic feature extraction module, respectively, and then the feature is fused before entering another basic feature extraction module. The attention module input is connected to the upstream ELANs module or the upstream CBC module through the basic feature extraction module. The attention module output is fused with the input features of the downstream ELANs module and then sent to the downstream ELANs module. The MP module includes a first branch and a second branch. The first branch includes two basic feature extraction modules connected in sequence, and the second branch includes a max pooling layer and a basic feature extraction module. The outputs of the first branch and the second branch are used for feature fusion. The lightweight YOLOv7 network has three branches downstream, with a dominant header at the end of each branch. The branch has an ELANs module, the output of which has an auxiliary head. The auxiliary head is used to assist training by utilizing the loss calculation in the middle of the network. The auxiliary head has more positive samples than the dominant head, and the auxiliary head pays more attention to recall.

2. The defect detection method according to claim 1, characterized in that, Images or videos of power distribution lines are collected by drones.

3. The defect detection method according to claim 1, characterized in that, The intersection-union ratio is expressed as: in, IoU Represented as intersection-union ratio, A This is represented as a candidate box. D Represented as the true target bounding box; The weighted loss is expressed as: in, Cost Indicated as weighted loss, L_coord Represented as regression box loss L_cls Represented as classification loss; Loss function of defect detection model loss Represented as: in, L_conf This is expressed as confidence loss. in, N It is the sample size. It is the size of the feature map. B It is the predicted border number for each cell. It is the first indicator function, used to represent the first... i The first sample j The first grid k The border is the target box. , For the first i The center coordinates of the true bounding box of each sample , They are the first i The width and height of the true border of each sample. , For the first i The first sample j The first grid k The predicted coordinates of each border. , It is the first i The first sample j The first grid k The predicted width and height of each border; in, It is the first i The first sample j The first grid k The confidence prediction value of each bounding box. It is the second indicator function, used to represent the first i The first sample j The first grid k The border is the background border, and λ is the weight; in, c For category number, C It is the number of categories. It is the first i The first sample c Real labels for each category It is the first i The first sample c Predicted labels for each category.

4. The defect detection method according to claim 1, characterized in that, Methods for obtaining prior bounding boxes in the training set include: The K-means algorithm is used to cluster the bounding boxes in the training set to obtain multiple prior boxes arranged from smallest to largest.

5. The defect detection method according to claim 1, characterized in that, Training methods for defect detection models include: Acquire inspection images of power distribution lines and annotate the inspection images to obtain a dataset; The dataset is augmented by random scaling, random cropping, and random arrangement. The dataset is randomly distributed to obtain a training set, a validation set, and a test set; A defect detection model is obtained by training a training set based on a lightweight YOLOv7 network with an attention mechanism. The defect detection model was validated using a validation set. The defect detection model was tested using a test set.

6. A defect detection system for power distribution lines, characterized in that, For implementing the defect detection method as described in any one of claims 1-5, the system includes a monitoring module, a defect analysis module, and an alarm module. The monitoring module is used to acquire inspection images or videos of power distribution lines; The defect analysis module is used to obtain a defect detection model based on the YOLO network; The defects in the power distribution lines are obtained by detecting the inspection images or videos using a defect detection model. The alarm module is used to generate reports and alarms based on the defects of the power distribution line. The report includes the line name, tower number, ID information of the inspected equipment, and file information attributes.

7. An identification terminal, characterized in that, The identification terminal stores instructions for implementing the defect detection method as described in any one of claims 1-5.

8. A drone, characterized in that, The drone is used to acquire inspection images or videos of power distribution lines and transmit the inspection images or videos to the identification terminal as described in claim 7. The identification terminal is used to execute the instructions and identify the position of the target object in the inspection image. After receiving the location of the target object in the inspection image sent by the identification terminal, the drone adjusts its flight direction, distance from the target object, and shooting angle; The drone is also used to mount the identification terminal.

Images