Method for detecting defects of stockbridge damper of power transmission line through unmanned aerial vehicle

Abstract

The invention provides a method for detecting defects of a vibration damper of a power transmission line through an unmanned aerial vehicle. According to the detection method, a data acquisition module, a deep learning detection module and a subsequent logic module are arranged; the data acquisition module obtains related data of a working site through an unmanned aerial vehicle and inputs the related data into the deep learning detection module; the deep learning detection module is improved on the basis of a Yolov12 model, performs deep learning and training through multi-scale fusion enhancement, an adaptive data enhancement strategy, Backbone network improvement, an IoU-based adaptive positioning loss function and category and confidence joint loss, and is used for analysis processing; the deep learning detection module identifies the damper and related objects thereof from the input image, and transmits a detection result to the subsequent logic module; and the subsequent logic module analyzes the detection result and outputs the state information of the damper in the current image so as to judge whether the damper has the defects such as corrosion, damage, falling, loosening or slippage or not.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video recognition analysis, in particular to a method for detecting defects of a power transmission line shock absorber by using a UAV. BACKGROUND

[0002] The damper is an important anti-vibration device of the transmission line, mainly used to suppress the fatigue damage of the conductor caused by wind vibration; its core function is to consume vibration energy, change vibration frequency and protect key parts; when the damper has defects, it will cause the loosening of the line clamp, damage to the conductor, increase the risk of discharge and other adverse factors to the safe operation of the transmission line; the traditional damper detection mainly relies on manual tower climbing or ground telescope observation, which has low efficiency, high safety risk, limited detection accuracy and other problems, and it is difficult to find early and subtle defects; at present, some technical solutions have been applied, such as patent CN110458798A "Damper defect visual detection method and system based on key point detection and medium", the steps of this scheme are: 1) input the transmission line image containing the damper; 2) based on the transmission line image, detect the damper target to get the width and height of the damper boundary box, and detect the left top point, right top point and line clamp body midpoint of the damper; 3) according to the existence state value and coordinate value of the left top point, right top point and line clamp body midpoint of the damper, judge whether the damper has the defects of missing hammer head, loose hammer head and deformation, and complete the detection of the damper defects; Patent CN110458798A uses key point detection technology to identify the defects of the damper; however, in actual application, key point detection can only show good effect in close-up pictures of the damper, and for non-close-up pictures, this method often causes a large number of false detections. Patent CN112950620A, "Transmission line damper deformation defect detection method based on cascade R-CNN algorithm"; the steps of this scheme are: step 1: pre-process the damper image obtained from the transmission line inspection image set to obtain the preliminary pre-processed damper image data set; step 2: based on the pre-processed image data, construct a damper target detection training set, a verification set and a test set; step 3: use the data in the training set to iteratively train the Cascade R-CNN model; step 4: input the test set into the trained Cascade R-CNN model to detect the damper image and test the model precision; step 5: input the damper image to be detected into the trained model to judge whether the damper has deformed, and label the final obtained damper deformation defect area to be detected.The patent CN112950620A adopts the method of R-CNN cascade, which has the phenomenon of low real-time performance and insufficient detection accuracy. With the development of unmanned aerial vehicle technology, non-contact detection with high efficiency and high precision can be realized through unmanned aerial vehicle detection, which avoids the risk of high-altitude operation and supports the establishment of digital archives for intelligent analysis. Unmanned aerial vehicle technology can also detect small displacements and initial corrosion that are difficult to detect by traditional methods, greatly improving the intelligent level of line operation. It has become the mainstream way of detecting the damper on the current transmission line. However, such patents have certain limitations. For example, patent CN112991252A discloses a method for detecting the absence of a damper on a transmission line. The algorithm first uses a UAV to collect aerial images containing a damper, then uses a Haar feature-based AdaBoost classifier to identify and segment the damper region in the UAV aerial image, then performs straight line detection on the damper region, corrects the damper image horizontally by the average angle of the detected straight line, then binarizes and vertically projects the corrected image, and determines the absence of the damper based on the distance relationship between the two end weight positions and the line clamp positions in the vertical projection image. Patent CN213535129U proposes a high-voltage transmission line damper homing robot, which includes a UAV, an insulating partition is arranged on the upper part of the UAV body, an operating manipulator and a walking mechanism are arranged on one side of the insulating partition, a damper homing controller is arranged on the lower part of the UAV body, and the damper homing controller is connected with the operating manipulator and the walking mechanism respectively. The damper homing controller is connected with the UAV, and the UAV is connected with a ground remote controller. Patent CN116805398A discloses a method for detecting defects of a damper on a transmission line, which includes the following steps: collecting damper images of a transmission line by a UAV device and preprocessing, establishing a sample data set; using a K-means clustering algorithm to perform clustering analysis on the sample data set, predicting anchor box size; constructing a defect detection network; the defect detection network includes an input preprocessing network, a feature extraction network, a feature fusion network, and a detector; using the sample data set and the predicted anchor box size, training the defect detection network to obtain a defect detection model; using the defect detection model to detect defects of the damper on the transmission line to obtain detection results. The above patents have certain defects or unsatisfactory precision or are difficult to implement. Therefore, the applicant proposes a method for detecting defects of a damper on a transmission line by a UAV, which uses a deep learning algorithm, collects video data on the transmission line by a UAV, and analyzes the state of the damper in real time to determine whether it has defects such as rust, damage, falling off, loosening, or sliding. SUMMARY

[0003] To solve the above technical problems, the application provides a method for detecting defects of a damper for a power transmission line by using a UAV.

[0004] To achieve the above object, the application adopts the technical scheme of:

[0005] The method for detecting defects of a damper for a power transmission line by using a UAV comprises a data acquisition module, a deep learning detection module and a subsequent logic module.

[0006] Further, the deep learning detection module of the method for detecting defects of a damper for a power transmission line by using a UAV is improved on the basis of a Yolov12 model and comprises multi-scale fusion enhancement, an adaptive data enhancement strategy, Backbone network improvement, an adaptive positioning loss function based on IoU and a combined loss of class and confidence.

[0007] The multi-scale fusion enhancement comprises the following steps:

[0008] Adaptive data augmentation strategy; YoLov12 based on the training of adaptive data augmentation mechanism, the strategy according to the training phase characteristics and model convergence state intelligent adjustment of enhancement intensity; In the initial training stage, the model parameters have not yet stabilized, the conservative enhancement strategy is adopted to ensure the stability of training; With the gradual convergence of the model, the validation set loss is used as the basis for adjustment, and high-order enhancement methods including random affine transformation and MixUp mixed enhancement are dynamically introduced, and the enhancement strength is gradually increased to the preset upper limit;

[0009] Backbone network improvement; The original Backbone of YoLov12 is composed of standard convolutional layers, C3k2 auxiliary feature extraction modules and A2C2f feature refinement modules; Now introduce Triplet Attention module, the module uses lightweight structure to realize cross-channel-space dimension interaction modeling, without dimension reduction, through the triple attention mechanism "channel-height-width" to capture long-range dependencies, significantly enhance the feature representation ability; Specifically, in the Triplet Attention module, Permute is used to convert the dimension order, and Z-Pool pooling is the maximum pooling and average pooling in series, and its formula is:

[0010]

[0011] Where: CxHxW represents the feature map with channel number C, height H and width W;

[0012] ⊗ represents matrix multiplication;

[0013] ⊙ represents broadcast element multiplication;

[0014] ⊕ represents broadcast element addition;

[0015] The formula of Triplet Attention module is:

[0016]

[0017] Where: represents the sigmoid activation function;

[0018] , , represents the standard two-dimensional convolution layer defined by the kernel size k in the three branches of triple attention;

[0019] Its y can be changed to:

[0020]

[0021] Where: , and is the three cross-dimensional attention weight of triple attention calculation, and the equation is and In the above equation, represents a 90-degree clockwise rotation to maintain the original input shape of (C H W);

[0022] The adaptive positioning loss function based on IoU; in view of the scale sensitivity defect of the traditional mean square error loss function in the boundary box regression, that is, the imbalance problem of insufficient punishment for small target positioning error and excessive punishment for large target, YoL0v12 adopts an adaptive positioning loss mechanism weighted by IoU; the core of the mechanism is to establish a dynamic mapping relationship between the IoU value and the loss weight; when the IoU of the predicted box and the real box is low, it indicates that the positioning deviation is significant, and the loss weight is automatically increased to strengthen the model correction; on the contrary, for the prediction box with high IoU and good positioning, the weight is reduced to avoid overfitting; through the scale-aware weighting strategy, the positioning optimization process of targets of different sizes is effectively balanced, and the detection accuracy of small targets is significantly improved;

[0023] Category and confidence combined loss: YoL0v12 designs a combined loss function to weight and combine the category cross-entropy loss and the binary classification loss of confidence, and replaces the cross-entropy loss with an adaptive threshold focal loss, which is expressed as:

[0024] During the training process, the weights of the two are dynamically adjusted according to different training stages and task requirements, so that the model can more accurately estimate the confidence of the predicted box while learning the category features, and improve the accuracy and reliability of the overall detection.

[0025] Further, the method for detecting the defect of the damper of the power transmission line by the unmanned aerial vehicle is provided with a post-logic module, and the logic for judging whether the damper slips is:

[0026] 1) When there is only one damper, it can only be judged whether the damper is too close to other fittings on the line, and the specific method is to detect the rectangular frame of the damper, and whether there is an intersection between the left and right horizontal direction of 1 times extension and the rectangular frame of other fittings, if there is, it is judged that the damper slips, otherwise it is normal;

[0027] 2) When there are multiple damping hammers, first determine whether the damping hammer has horizontal slip. The specific method is to determine whether the "left and right horizontal extension damping hammer rectangle frame" of each damping hammer intersects with the rectangle frame of other fittings on the line and the rectangle frame of other damping hammers without extension. If there is intersection, the damping hammer has a slip phenomenon; second, determine whether the damping hammer on the side-by-side line has a slip. The specific method is to extend the rectangle frame of each damping hammer by 1 times in the left and right horizontal directions, and expand to the Y axis of the picture in the up and down vertical directions. If the damping hammer intersects with other damping hammers in the vertical direction, it is considered that the damping hammers are side by side, and the damping hammers in the picture that do not intersect with other damping hammers indicate that the damping hammers have a slip phenomenon.

[0028] Further, the training step of the deep learning detection module of the method for detecting defects of damping hammers of a power transmission line by using a drone is specifically:

[0029] Step one, collect the video of the power transmission line by using a drone, and convert the video stream into an image;

[0030] Step two, label each object in the image, and divide the labeled image into a training set, a validation set, and a test set according to a ratio of 8:1:1;

[0031] Step three, train the improved yolov12 model using the training set and the validation set, and verify the model performance using the test set after the training is completed.

[0032] Further, the specific steps of the method for detecting defects of damping hammers of a power transmission line by using a drone are:

[0033] Step one, the data acquisition module transmits the real-time image of the power transmission line to the deep learning detection module by using a drone;

[0034] Step two, the deep learning detection module analyzes the image and transmits the result to the subsequent processing module;

[0035] Step three, the subsequent processing module determines whether there is damping hammer slip, and outputs the final result.

[0036] The application brings the following benefits:

[0037] 1. The method for detecting defects of damping hammers of a power transmission line by using a drone is based on video data analysis of a drone, which realizes real-time detection of damping hammers on a power transmission line and is efficient and fast;

[0038] 2. The method for detecting defects of damping hammers of a power transmission line by using a drone uses an improved deep learning model, which effectively improves the detection accuracy of damping hammers, and the subsequent processing module more accurately determines damping hammer slip;

[0039] 3. A method for detecting vibration damper defects in transmission lines using drones: By adding a TripletAttention module to YOLOv12 and changing the loss to adaptive threshold focus loss, defects in vibration dampers of transmission lines can be detected quickly. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0041] Figure 2 This is a schematic diagram of the Triplet Attention module used in this invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0043] like Figure 1 As shown, this is a method for detecting defects in vibration dampers of power transmission lines using a drone. The method comprises a data acquisition module, a deep learning detection module, and a subsequent logic module. The data acquisition module uses the drone to acquire relevant data from the work site, converts the video stream into image frames for annotation, and generates an annotated dataset for analysis by the deep learning detection module. The deep learning detection module improves upon the Yolov12 model by employing multi-scale fusion enhancement, adaptive data augmentation strategies, improved Backbone networks, an IoU-based adaptive localization loss function, and a joint loss of category and confidence for deep learning and training, which is then used for subsequent analysis. The deep learning detection module identifies the vibration dampers and related objects from the input images and passes the detection results to the subsequent logic module. The subsequent logic module analyzes the detection results and ultimately outputs the state information of the vibration dampers in the current image.

[0044] The deep learning detection module shown in the method for detecting vibration damper defects in power transmission lines using drones is an improvement on the Yolov12 model. It supports single-GPU training, with an input image size of 1280*1280. The module utilizes multi-scale fusion enhancement, adaptive data augmentation strategies, improved backbone networks, an IoU-based adaptive localization loss function, and a joint loss function combining category and confidence.

[0045] Multi-scale fusion enhancement; YoLov12 uses a cross-scale feature fusion enhancement mechanism in the data preprocessing stage, which is different from traditional single-scale data enhancement methods. The algorithm first constructs a multi-scale image pyramid, and then independently implements standardization enhancement operations, including random flipping and brightness adjustment, on each scale. Then, the features of each scale level are extracted through the backbone network, and finally the cross-scale feature aggregation is realized through the feature fusion module, which significantly improves the multi-scale feature representation ability of the model and makes it more effective in detecting target instances of different sizes.

[0046] Adaptive data augmentation strategy; YoLov12 is based on a dynamic adaptive data augmentation mechanism. The strategy intelligently adjusts the enhancement strength according to the characteristics of the training phase and the convergence state of the model. In the initial training phase, when the model parameters are not yet stable, a conservative enhancement strategy is used, such as ±15° random rotation, 10% scale translation, and other low disturbance operations to ensure training stability. As the model gradually converges, the validation set loss is used as the basis for adjustment, and high-order enhancement methods including random affine transformation and MixUp mixed enhancement are dynamically introduced to gradually increase the enhancement strength to the preset upper limit. This learning-based enhancement strategy achieves an optimal balance between model robustness and data utilization efficiency, significantly improving the completeness of feature extraction and the generalization performance of the model.

[0047] Backbone network improvement; the original Backbone of YoLov12 is composed of standard convolution layers, C3k2 auxiliary feature extraction modules, and A2C2f feature refinement modules. The Triplet Attention module is introduced as shown in Figure 2 The module uses a lightweight structure to realize cross-channel-spatial dimension interaction modeling, and captures long-range dependencies through the triple attention mechanism "channel-height-width" without dimension reduction, significantly enhancing the feature representation ability. In the Triplet Attention module, Permute is used to convert the dimension order, and Z-Pool pooling is a combination of maximum and average pooling, whose formula is:

[0048]

[0049] where CxHxW represents a feature map with channel number C, height H, and width W.

[0050] ⊗ represents matrix multiplication;

[0051] ⊙ represents broadcast element multiplication;

[0052] ⊕ represents broadcast element addition;

[0053] The formula of the Triplet Attention module is:

[0054]

[0055] wherein: represents a sigmoid activation function;

[0056] , , represents a standard two-dimensional convolutional layer defined by the kernel size k in the three branches of triple attention;

[0057] Its y can become:

[0058]

[0059] wherein: , and are three cross-dimensional attention weights in triple attention calculation, and and In the above equation, represents a 90-degree clockwise rotation to maintain the original input shape of (C H W);

[0060] IoU-based adaptive positioning loss function; In view of the scale sensitivity defect of the traditional mean square error loss function in the boundary box regression, that is, the imbalance problem of insufficient punishment for small target positioning error and excessive punishment for large target, YoL0v12 adopts an adaptive positioning loss mechanism weighted by IoU; The core of the mechanism is to establish a dynamic mapping relationship between the IoU value and the loss weight; When the IoU of the predicted box and the real box is low, it indicates that the positioning deviation is significant, and the loss weight is automatically increased to strengthen the model correction; On the contrary, for the prediction box with high IoU and good positioning, the weight is reduced to avoid overfitting; Through the scale-aware weighting strategy, the positioning optimization process of targets of different sizes is effectively balanced, and the detection accuracy of small targets is significantly improved;

[0061] Category and confidence combined loss: YoL0v12 designs a combined loss function to combine the category cross-entropy loss and the binary classification loss of confidence by weighting, replaces the cross-entropy loss with an adaptive threshold focal loss, and its formula is expressed as:

[0062] In the training process, according to different training stages and task requirements, the weights of the two are dynamically adjusted, so that the model can more accurately estimate the confidence of the prediction box while learning the category features, and improve the accuracy and reliability of the overall detection.

[0063] The detection method of the anti-vibration hammer defect of the power transmission line by the unmanned aerial vehicle is provided with a rear logic module, and the logic of the rear logic module for judging whether the anti-vibration hammer slips is:

[0064] 1) When there is only one damper, it can only be determined whether the damper is too close to other fittings on the line, and the specific method is to detect the rectangular frame of the damper, and whether the left and right horizontal directions are each extended by 1 times whether there is an intersection with the rectangular frame of other fittings, if there is, it is judged that the damper slips, otherwise it is normal;

[0065] 2) When there are multiple dampers, first determine whether the damper is horizontally slipped, the specific method is to determine whether the rectangular frame of each damper "left and right horizontal extension damper rectangular frame" is intersected with the rectangular frame of other fittings and other dampers without extension, if there is, then the damper slips; Secondly, determine whether the damper on the parallel line slips, the specific method is to extend the rectangular frame of each damper by 1 times in the left and right horizontal directions, and expand to the Y axis of the picture in the up and down vertical directions, if the damper in the vertical direction intersects with the rectangular frame of other dampers, it is considered that the dampers are parallel, and the damper in the picture has no intersection with other dampers, which means that the damper slips.

[0066] The training steps of the deep learning detection module of the damper defect detection method of the power transmission line shown by the unmanned aerial vehicle are as follows:

[0067] Step one, collect the power transmission line video through the unmanned aerial vehicle, and convert the video stream into an image;

[0068] Step two, label each object in the image, and divide the labeled image into a training set, a validation set and a test set according to the ratio of 8:1:1;

[0069] Step three, train the improved yolov12 model using the training set and the validation set, and verify the model performance using the test set after training.

[0070] The specific steps of the damper defect detection method of the power transmission line shown by the unmanned aerial vehicle are as follows:

[0071] Step one, the data acquisition module transmits the real-time image of the power transmission line to the deep learning detection module through the unmanned aerial vehicle;

[0072] Step two, the deep learning detection module analyzes the image and transmits the result to the subsequent processing module;

[0073] Step three, the subsequent processing module determines whether there is damper slip, and outputs the final result.

[0074] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in any other form, and any modification or equivalent change made according to the technical essence of the present application still belongs to the scope claimed by the present application.

Claims

1. A method for detecting defects in vibration dampers of power transmission lines using unmanned aerial vehicles (UAVs), characterized in that: The method for detecting defects in vibration dampers of power transmission lines using drones is configured with a data acquisition module, a deep learning detection module, and a subsequent logic module. The data acquisition module acquires relevant data from the work site using drones, converts video streams into image frames for annotation processing, and finally generates an annotated dataset for analysis and processing by the deep learning detection module, which then inputs the dataset into the deep learning detection module. The deep learning detection module is an improvement on the Yolov12 model. It uses multi-scale fusion enhancement, adaptive data augmentation strategy, Backbone network improvement, IoU-based adaptive localization loss function and joint loss of category and confidence for deep learning and training, and is used for subsequent analysis and processing. The deep learning detection module identifies the vibration damper and its associated objects from the input image and passes the detection results to the subsequent logic module. The subsequent logic module analyzes the detection results and finally outputs the state information of the vibration damper in the current image.

2. The method for detecting defects in vibration dampers of power transmission lines using a drone according to claim 1, characterized in that: The deep learning detection module of the method for detecting defects in transmission line vibration dampers using drones is an improvement on the Yolov12 model. Specifically, it employs multi-scale fusion enhancement, adaptive data augmentation strategies, Backbone network improvements, an IoU-based adaptive localization loss function, and a joint loss function combining category and confidence. Multi-scale fusion enhancement; YoLov12 uses a cross-scale feature fusion enhancement mechanism in the data preprocessing stage, which is different from the traditional single-scale data enhancement method. The algorithm first constructs a multi-scale image pyramid and performs standardized enhancement operations independently on each scale, including random flipping and brightness adjustment; then, the backbone network extracts the feature representations of each scale level respectively, and finally uses a feature fusion module to achieve cross-scale feature aggregation. Adaptive data augmentation strategy; YoLov12 is based on a training-dynamic adaptive data augmentation mechanism. The strategy intelligently adjusts the augmentation intensity according to the characteristics of the training stage and the model convergence state. When the model parameters are not yet stable in the initial training stage, a conservative augmentation strategy is adopted to ensure training stability. As the model gradually converges, higher-order enhancement methods, including random affine transformation and MixUp hybrid enhancement, are dynamically introduced based on the validation set loss, gradually increasing the enhancement intensity to a preset upper limit. Backbone Network Improvement: The original YoLov12 Backbone consists of standard convolutional layers, a C3k2 auxiliary feature extraction module, and an A2C2f feature refinement module. A Triplet Attention module is introduced, employing a lightweight structure to achieve cross-channel-spatial dimension interactive modeling. Without dimensionality reduction, it captures long-range dependencies through a triple attention mechanism ("channel-height-width"), significantly enhancing feature representation capabilities. Specifically, in the Triplet Attention module, Permute is used to transform the dimensionality order, and Z-Pool pooling is a concatenation of max pooling and average pooling, expressed by the following formula: ； Where: C×H×W represents the feature mapping with channel number C, height H, and width W; ⊗ represents matrix multiplication; ⊙ indicates broadcast element multiplication; ⊕ indicates addition of broadcast elements; The Triplet Attention module can be expressed by the following formula: ； in: Represents the sigmoid activation function; , , The standard two-dimensional convolutional layer, defined by kernel size k, represents the three branches of triple attention; Its y can be transformed into: ； in: , and The attention weights in the three cross-dimensional attention calculations are in the equation. and In the above equation, represents a 90-degree clockwise rotation to maintain (C). H The original input shape of W); An IoU-based adaptive localization loss function addresses the scale-sensitive limitation of traditional mean squared error loss functions in bounding box regression—the imbalance of insufficient penalty for localization errors of small targets and excessive penalty for large targets. YoL0v12 employs an IoU-weighted adaptive localization loss mechanism. The core of this mechanism lies in establishing a dynamic mapping relationship between IoU values ​​and loss weights. When the IoU between the predicted box and the ground truth box is low, indicating a significant localization deviation, the loss weight is automatically increased to strengthen model correction. Conversely, for predicted boxes with high IoU—those with good localization—the weight is reduced to avoid overfitting. This scale-aware weighting strategy effectively balances the localization optimization process for targets of different sizes, significantly improving the detection accuracy of small targets. Joint Loss Based on Category and Confidence: YoL0v12 designs a joint loss function that weights and combines the category cross-entropy loss and the binary classification loss based on confidence, replacing the cross-entropy loss with an adaptive threshold focus loss. The formula is as follows: ; During training, the weights of both are dynamically adjusted according to different training stages and task requirements, so that while the model learns category features, it can more accurately estimate the confidence of the predicted box, thereby improving the overall accuracy and reliability of detection.

3. The method for detecting defects in vibration dampers of power transmission lines using a drone according to claim 1, characterized in that: The method for detecting defects in vibration dampers of power transmission lines using drones includes a post-logic module. The logic of the post-logic module for determining whether the vibration damper has slipped is as follows: 1) When there is only one vibration damper, it can only be determined whether the vibration damper is too close to other hardware on the line. The specific method is to detect the rectangular frame of the vibration damper and extend it horizontally by 1 times in both directions to see if it intersects with the rectangular frames of other hardware. If it does, it is determined that the vibration damper is slipping; otherwise, it is normal. 2) When there are multiple vibration dampers, first, determine whether the vibration damper has horizontal slippage. The specific method is to check whether the "horizontal extension of the vibration damper rectangle" of each vibration damper intersects with the rectangles of other hardware on the line and the rectangles of other vibration dampers without extension. If so, then the vibration damper has slipped. Second, determine whether the vibration dampers on parallel lines have slippage. The specific method is to extend the rectangle of each vibration damper by 1 times in the horizontal direction and extend it vertically to the Y-axis of the image. If the vibration damper intersects with the rectangles of other vibration dampers in the vertical direction, then parallel vibration dampers are considered to exist. If there is a vibration damper in the image that does not intersect with other vibration dampers, then the vibration damper has slipped.

4. The method for detecting defects in vibration dampers of power transmission lines using a drone according to claim 1, characterized in that: The specific training steps for the deep learning detection module of the method for detecting defects in transmission line vibration dampers using drones are as follows: Step 1: Collect video of the power transmission line using a drone and convert the video stream into images; Step 2: Label each object in the image accordingly, and divide the labeled images into training set, validation set and test set in a ratio of 8:1:1; Step 3: Train the improved YOLOv12 model using the training and validation sets. After training, use the test set to verify the model's performance.

5. A method for detecting defects in vibration dampers of power transmission lines using a drone, as described in claim 1, characterized in that: The specific steps of the method for detecting defects in vibration dampers of power transmission lines using drones are as follows: Step 1: The data acquisition module transmits real-time images of the power transmission line to the deep learning detection module via a drone; Step 2: The deep learning detection module analyzes the image and inputs the results into the subsequent processing module; Step 3: The subsequent processing module determines whether there is slippage of the vibration damper and outputs the final result.

Images