A real-time health detection method and system for transmission line insulators

Abstract

The present invention discloses a real-time health detection method and system for transmission line insulators, which relates to the technical field of insulator detection. The image is preprocessed and then the preprocessed image is input into a pre-trained YOLO-F model. The YOLO-F model extracts features through a backbone network, and uses a feature fusion network to perform multi-scale feature fusion. The backbone network extracted features and fused features are transmitted to a free detection head for analysis. The free detection head analyzes the backbone network extracted features and fused features. The analysis results are further processed by a non-maximum suppression algorithm, and finally the insulator health detection results are obtained. According to the insulator health detection results, corresponding management suggestions are generated. The detection system accurately captures the true shape and position of the object through the YOLO-F model. When processing the detection of insulators at different angles, the actual contour of the object is better fitted by directly regressing the bounding box and the rotation amount, thereby improving the detection accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of insulator detection, and in particular to a real-time health detection method and system for insulators of a transmission line. Background Art

[0002] As an indispensable component of the power system, insulators are mainly used to support and isolate high-voltage wires, prevent current leakage through air or other media, and ensure the stability and safety of power transmission. With the continuous growth of power demand and the increasing complexity of power systems, higher requirements are placed on the maintenance and detection of insulators.

[0003] Traditional insulator inspection methods mainly include manual inspection and regular maintenance. Manual inspection relies on experienced technicians to inspect through naked eye observation and manual tools. This method is not only inefficient, but also easily affected by human factors, making it difficult to detect potential faults in time. Although regular maintenance can prevent faults to a certain extent, it cannot monitor the status of insulators in real time and has a large lag.

[0004] With the rapid development of computer vision and deep learning technologies, image-based target detection technology has gradually become a new trend in insulator detection. Target detection technology can automatically identify and locate insulators in images and monitor their status in real time, greatly improving the accuracy and efficiency of detection. Especially in the daily inspection of transmission lines, target detection technology can quickly detect defects in insulators, such as cracks, breakage, pollution, etc., and take timely measures to avoid power outages and safety accidents caused by insulator failures.

[0005] The prior art has the following deficiencies:

[0006] Although target detection technology has shown great potential in insulator detection, it still faces some challenges. For example, the transmission line environment is complex, there are many background interferences, and the shapes and colors of insulators vary. These factors increase the difficulty of detection. The scale of the target in the image varies greatly. The existing detection algorithms are unstable when processing multi-scale targets. High-performance target detection models are usually complex in structure, consume a lot of computing resources, and are difficult to deploy on resource-constrained devices. Although some algorithms perform well in terms of accuracy, they are insufficient in real-time and cannot meet the needs of real-time applications. When the rotation and posture of the target in the image change greatly, the detection performance decreases. Summary of the invention

[0007] The purpose of the present invention is to provide a real-time health detection method and system for transmission line insulators. Through the YOLO-F model, the real shape and position of the object can be captured more accurately. When processing the detection of insulators at different angles, the actual contour of the object can be better fitted by directly regressing the bounding box and the rotation amount, thereby improving the detection accuracy.

[0008] In order to achieve the above object, the present invention provides the following technical solution: a real-time health detection method for transmission line insulators, the detection method comprising the following steps:

[0009] The detection system controls the UAV to fly along a predetermined path and uses the visible light imager on the UAV to collect images along the transmission line. The collected images are transmitted to the detection system in real time.

[0010] The detection system first preprocesses the image, and then inputs the preprocessed image into the pre-trained YOLO-F model. The YOLO-F model includes a backbone network, a feature fusion network, and a free detection head. The YOLO-F model extracts features through the backbone network, and uses the feature fusion network to perform multi-scale feature fusion to obtain fused features. The backbone network extracted features and fused features are then passed to the free detection head for analysis.

[0011] The free detection head analyzes the backbone network to extract features and fuse features. The analysis results are further processed by the non-maximum suppression algorithm to finally obtain the insulator health detection results. After generating corresponding management suggestions based on the insulator health detection results, the detection results and management suggestions are mapped to images, and the images are saved to the local storage device and uploaded to the cloud at the same time.

[0012] In a preferred embodiment, the detection system preprocesses the image, comprising the following steps:

[0013] The LabelImg software is used to label the insulators in the image. The insulator labeling information includes the coordinates of the insulator center, the width and height of the insulator bounding box, and the rotation angle of the insulator bounding box. The eccentricity of all pixels in the insulator bounding box is also calculated: , where is the eccentricity of pixel (i, j), is the horizontal coordinate position of the pixel point, is the central horizontal coordinate of the real frame where the pixel (i, j) is located, is the ordinate of the pixel point, y is the ordinate of the real frame of the (i, j) pixel point, is the width of the real frame where the pixel (i, j) is located, is the height of the real box where the pixel (i, j) is located, is a disturbance greater than zero, ensuring that the entire formula is not zero, Is the natural base.

[0014] In a preferred embodiment, the preprocessed image is input into a pre-trained YOLO-F model, and the YOLO-F model extracts features through a backbone network, including the following steps:

[0015] Pass the preprocessed image into the input layer of the YOLO-F model;

[0016] In the convolution layer, the image first undergoes multiple layers of convolution operations. Each convolution layer extracts local features on the image by sliding the convolution kernel. Each convolution kernel is used to detect features in the image. An activation function is applied after each convolution layer, and negative values ​​are set to zero.

[0017] Between the convolutional layers, the pooling layer is used to reduce the size of the feature map. The maximum pooling selects the maximum value in the feature map. The pooling operation uses a preset step size to reduce the spatial dimension of the feature map layer by layer.

[0018] After several layers of convolution and pooling, the backbone network outputs a high-dimensional feature map, which contains multi-level feature information of insulators in the image.

[0019] In a preferred embodiment, the YOLO-F model uses a feature fusion network to perform multi-scale feature fusion to obtain fusion features, including the following steps:

[0020] The multiple high-dimensional feature maps output by the backbone network are used as the input of the feature fusion network, and the multiple high-dimensional feature maps are concatenated in the channel dimension;

[0021] The concatenated high-dimensional feature map is further processed through the convolution layer to extract the comprehensive feature map, and after the convolution layer, the activation function is used for nonlinear transformation;

[0022] After fusing the feature maps, a pooling layer is applied for downsampling to reduce the size of the comprehensive feature map;

[0023] After feature fusion processing, multiple feature maps of different scales are output, and feature maps of different layers are selected for subsequent target detection;

[0024] The fused feature map is passed to the free detection head of the YOLO-F model for subsequent target detection and classification.

[0025] In a preferred embodiment, the free detection head analyzes the backbone network to extract features and fuse features, and the analysis results are further processed by a non-maximum suppression algorithm to finally obtain the insulator health detection results, including the following steps:

[0026] The feature maps extracted from the backbone network and the feature maps processed by the feature fusion network are passed into the free detection head of the YOLO-F model;

[0027] The free detection head generates predicted bounding boxes based on the input feature map. Each bounding box consists of the center coordinates, width, height and rotation angle of the bounding box.

[0028] Output the category probability distribution for each predicted bounding box, which indicates the possibility of containing different categories of objects in the bounding box;

[0029] The free detection head generates a set of candidate boxes based on the predicted bounding boxes and category probabilities, and calculates the confidence for each candidate box;

[0030] Sort all candidate boxes according to their confidence from large to small, start with the candidate box with the highest confidence, keep the candidate box and calculate the overlapping area with other candidate boxes, and determine the degree of overlap by the intersection-over-union ratio threshold:

[0031] If the overlap of a candidate box exceeds the intersection-over-union ratio threshold, the candidate box is deleted and the process is repeated until all candidate boxes are processed;

[0032] After processing, the remaining candidate boxes are the final detection results, and the final retained bounding boxes are used as the output results of the YOLO-F model.

[0033] In a preferred embodiment, the free detection head generates a set of candidate boxes according to the predicted bounding box and category probability, and calculates the confidence for each candidate box, including the following steps:

[0034] For each candidate box, the confidence calculation expression is:

[0035] , where is the confidence level, represents the probability of the existence of insulators in the candidate box, is the overlap between the predicted bounding box and the true bounding box, is the focal length offset of the visible light imager, , , are adjustment coefficients, and , , Both are greater than 0.

[0036] In a preferred embodiment, the candidate box content includes the bounding box position and size, category label, category probability, overlap degree, and insulator health detection result;

[0037] The steps to obtain the insulator health test results are as follows:

[0038] The damage ratio value and the total deviation of the center coordinate are normalized to map the value range of the damage ratio value and the total deviation of the center coordinate to the range of [0,1], and the normalized value of the damage ratio value and the normalized value of the total deviation of the center coordinate are obtained. The normalized value of the damage ratio value and the normalized value of the total deviation of the center coordinate are summed to obtain the health impact factor of the insulator, and the obtained health impact factor is compared with the preset impact threshold. If the health impact factor is less than or equal to the impact threshold, the insulator is evaluated to be in a normal health state. If the health impact factor is greater than the impact threshold, the insulator is evaluated to be in an abnormal health state.

[0039] In a preferred embodiment, the bounding box with the insulator is divided into a number of 1×1 grids, all cracks and holes on the surface of the insulator are identified by an image recognition algorithm, the number of grids occupied by the cracks and holes is recorded, the number of grids occupied by the cracks and holes is divided by the total number of grids in the bounding box to obtain a damage ratio value, multiple frames of images with the insulator are obtained within a period of time, the center coordinates of the insulator in the multiple frames of images are obtained, and the total deviation of the center coordinates of the insulator in the multiple frames of images is calculated based on the Euclidean distance calculation formula, and the expression is: , where is the total deviation of the center coordinate, is the number of image frames, For the The center coordinates of the frame image, For the The center coordinates of the frame image.

[0040] A real-time health detection system for transmission line insulators, comprising an image acquisition module, a feature analysis module, and a health detection module;

[0041] Image acquisition module: controls the drone to fly along a predetermined path and uses the visible light imager on the drone to collect images along the transmission line. The collected images are transmitted to the feature analysis module in real time.

[0042] Feature analysis module: First, the image is preprocessed, and then the preprocessed image is input into the pre-trained YOLO-F model. The YOLO-F model includes a backbone network, a feature fusion network, and a free detection head. The YOLO-F model extracts features through the backbone network, and uses the feature fusion network to perform multi-scale feature fusion to obtain fused features. The backbone network extracted features and fused features are transmitted to the free detection head for analysis, and the free detection head sends the analysis results to the health detection module.

[0043] Health detection module: The analysis results are further processed through the non-maximum suppression algorithm to finally obtain the insulator health detection results. After generating corresponding management suggestions based on the insulator health detection results, the detection results and management suggestions are mapped to images, and the images are saved to the local storage device and uploaded to the cloud at the same time.

[0044] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0045] The present invention preprocesses the image and then inputs the preprocessed image into the pre-trained YOLO-F model. The YOLO-F model extracts features through the backbone network and uses the feature fusion network to perform multi-scale feature fusion. The backbone network extracted features and fused features are transmitted to the free detection head for analysis. The free detection head analyzes the backbone network extracted features and fused features. The analysis results are further processed by the non-maximum suppression algorithm to finally obtain the insulator health detection results, and corresponding management suggestions are generated based on the insulator health detection results. The detection system accurately captures the true shape and position of the object through the YOLO-F model. When processing the detection of insulators at different angles, the actual contour of the object is better fitted by directly regressing the bounding box and the rotation amount, thereby improving the detection accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0047] Figure 1 is a flow chart of the method of the present invention;

[0048] Figure 2 This is the YOLO-F model structure diagram;

[0049] Figure 3 It is the schematic diagram of GBRS module;

[0050] Figure 4 It is the schematic diagram of MCBRN module;

[0051] Figure 5 This is a schematic diagram of the NS module;

[0052] Figure 6 Schematic diagram of SA module. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] Example 1: Please refer to Figure 1 As shown, this embodiment provides a real-time health detection method for transmission line insulators, and the detection method includes the following steps:

[0055] The detection system controls the UAV to fly along the predetermined path, and collects images along the transmission line through the visible light imager on the UAV. The collected images are transmitted to the detection system in real time. The detection system first preprocesses the images, and then inputs the preprocessed images into the pre-trained YOLO-F model. The YOLO-F model includes a backbone network, a feature fusion network and a free detection head. The YOLO-F model extracts features through the backbone network, and uses the feature fusion network to perform multi-scale feature fusion to obtain fused features. The backbone network extracted features and fused features are transmitted to the free detection head for analysis. The free detection head analyzes the backbone network extracted features and fused features. The analysis results are further processed by the non-maximum suppression algorithm to finally obtain the insulator health detection results. After generating corresponding management suggestions based on the insulator health detection results, the detection results and management suggestions are mapped to images, and the images are simultaneously saved to the local storage device and uploaded to the cloud.

[0056] This application preprocesses the image and then inputs the preprocessed image into the pre-trained YOLO-F model. The YOLO-F model extracts features through the backbone network and uses the feature fusion network to perform multi-scale feature fusion. The features extracted from the backbone network and the fused features are passed to the free detection head for analysis. The free detection head analyzes the features extracted from the backbone network and the fused features. The analysis results are further processed by the non-maximum suppression algorithm to finally obtain the insulator health detection results, and corresponding management recommendations are generated based on the insulator health detection results. The detection system accurately captures the true shape and position of the object through the YOLO-F model. When processing the detection of insulators at different angles, the actual contour of the object is better fitted by directly regressing the bounding box and rotation amount, thereby improving the detection accuracy.

[0057] By constructing a low-complexity model YOLO-F model to significantly reduce the size of the model and the consumption of computing resources, the low-complexity YOLO-F model has a faster forward propagation speed. When drones inspect transmission lines, it is necessary to quickly detect the status of insulators. The YOLO-F model can complete the detection task in a short time and meet the real-time requirements. The free detection frame can better match all objects. At the same time, the free detection frame with rotation can directly predict the bounding box and rotation angle of the object without matching and adjusting through the anchor frame. This enables the model to capture the true shape and position of the object more accurately, especially when dealing with rotating objects, such as the detection of insulators at different angles. By directly regressing the bounding box and rotation amount, the model can better fit the actual contour of the object and improve the detection accuracy. Since there is no need to preset the anchor frame, the free detection frame reduces a lot of computational overhead during the training process. The traditional anchor frame-based method needs to calculate the matching degree between each anchor frame and the real frame, which increases the complexity and time of training. The free detection frame method directly predicts the position and attributes of the object from the feature map, simplifies the training process, and improves the training efficiency.

[0058] Embodiment 2: The detection system controls the drone to fly along a predetermined path, and collects images along the transmission line through the visible light imager on the drone, and the collected images are transmitted to the detection system in real time, including the following steps:

[0059] The detection system controls the drone to fly along the power transmission line according to the pre-set flight path to ensure that the shooting range covers all target areas. Path planning optimizes the route design by analyzing the geographical distribution of the power transmission line, environmental factors (such as terrain and obstacles), and combining factors such as the drone's endurance. The preset path can ensure that the drone can achieve accurate image collection along the line during uninterrupted flight, improve inspection efficiency and avoid missed inspection areas.

[0060] Suppose the detection system plans a transmission line flight path from station A to station B, passing through hills, rivers, and a few residential areas along the way. To ensure that all detection points are covered and obstacles are avoided, the system first imports geographic information along the transmission line, including terrain undulations, obstacle locations, and area restrictions.

[0061] Path optimization: In hilly areas, the drone may be affected by wind speed due to uneven terrain. The path planning algorithm will increase the flight altitude to ensure a complete field of view while maintaining flight stability to avoid visual blind spots due to insufficient altitude.

[0062] Endurance considerations: Considering the battery life of the drone, the system will set a midway recharging point, Station C, on the route, evenly distributed from Station A and Station B. The drone will make a short landing at Station C and replace the battery to ensure sufficient energy and avoid power shortage during image acquisition.

[0063] Obstacle avoidance in residential areas: When passing through residential areas, in order to avoid flight interference, the system plans a detour path so that the drone can fly from the outside of the residential area, while adjusting the lens angle to ensure that image acquisition of the transmission lines is not affected.

[0064] Through such optimized planning, the UAV can maintain a reasonable flight altitude within the path and accurately complete image collection along the route without the need to frequently adjust the path in the middle, thereby improving inspection efficiency and ensuring the safety of the UAV in complex environments.

[0065] The detection system preprocesses the image, including the following steps:

[0066] The LabelImg software is used to carefully annotate the insulators in the image, paying special attention to the insulator position, outline, and size. The specific annotation information of a single insulator includes the coordinates of the insulator center, the width and height of the insulator bounding box (rectangle), and the rotation angle of the insulator bounding box. In addition, the eccentricity of all pixels in the bounding box of a single insulator is calculated using the following formula: , where is the eccentricity of pixel (i, j), is the horizontal coordinate position of the pixel point, is the central horizontal coordinate of the real frame where the pixel (i, j) is located, is the ordinate of the pixel point, y is the ordinate of the real frame of the (i, j) pixel point, is the width of the real frame where the pixel (i, j) is located, is the height of the real box where the pixel (i, j) is located, is a small perturbation greater than zero, ensuring that the entire formula is not zero, Is the natural base.

[0067] Description of shape characteristics: Eccentricity provides a quantitative description of the insulator shape, which is particularly effective in identifying different types of insulators or distinguishing between normal and abnormal insulators. The smaller the eccentricity value, the closer the insulator is to a symmetrical shape; while a larger eccentricity indicates a stretched or deformed shape. In this way, eccentricity can help the inspection system identify standard insulators and abnormal insulators that may have problems such as damage, cracks, etc.

[0068] Assisted posture and angle calibration: The calculation of eccentricity can further support the calibration of rotation angle. When the insulator is rotated, the change in eccentricity value can help the detection system optimize the rotation angle of the bounding box so that the bounding box fits the actual shape of the insulator more accurately, thereby improving the accuracy of annotation and detection.

[0069] Deformation or damage detection: Dynamic monitoring of eccentricity can reflect the deformation of insulators. When the insulator changes its physical shape due to aging, external forces or environmental factors, the eccentricity will deviate significantly from the original value, which can serve as a basis for the detection system to identify potential insulator failures.

[0070] On the basis of marking the position, outline and size of the insulator, the addition of eccentricity calculation further enhances the detailed characterization of the insulator shape characteristics and provides a reliable shape reference for subsequent detection, monitoring and fault warning.

[0071] The preprocessed image is input into the pre-trained YOLO-F model. The YOLO-F model extracts features through the backbone network, including the following steps:

[0072] Image input: The preprocessed image (resized to the input size required by the YOLO-F model, such as 416x416 or 608x608) is passed to the input layer of the YOLO-F model. The input image needs to be normalized, usually by scaling the pixel values ​​to between 0 and 1 to meet the requirements of model training.

[0073] Convolutional layer processing:

[0074] Convolution operation: The image first goes through multiple layers of convolution operations. Each convolution layer extracts local features on the image by sliding the convolution kernel (filter). Each convolution kernel is responsible for detecting specific features in the image, such as edges, corners, etc.

[0075] Application of activation function: After each convolutional layer, an activation function (such as ReLU) is usually applied to increase the nonlinear expression ability of the model by setting negative values ​​to zero. This step helps the model better capture complex features.

[0076] Pooling layer:

[0077] Dimensionality reduction: Between convolutional layers, pooling layers (such as max pooling or average pooling) are used to reduce the size of feature maps, reduce computational complexity, and make the model more robust to position changes. Max pooling selects the maximum value in the feature map to retain important features while reducing the amount of information.

[0078] Step size setting: The pooling operation usually uses a small step size (such as 2x2) to ensure that the size of the feature map after pooling is reduced by half, reducing the spatial dimension of the feature map layer by layer.

[0079] Feature map output: After several layers of convolution and pooling, the backbone network will output a high-dimensional feature map. This feature map contains multi-level feature information of the insulators in the image, which serves as the output of subsequent processing by the YOLO-F model.

[0080] The YOLO-F model significantly reduces the amount of computation by reducing the number of convolutional layers and introducing a free detection box mechanism. In addition, the YOLO-F model uses the MCBRA module and the GBRS module for efficient feature extraction and decoding, which effectively improves the detection accuracy. The YOLO-F model structure is as follows: Figure 2 As shown, including:

[0081] Input layer: The YOLO-F model accepts 416×416×3 images as model input.

[0082] The structure of the backbone network is as follows: the GBRS module is connected to the four MCBRN modules in series, the four MCBRN modules in series are connected to the two MCBRN modules in series, the two MCBRN modules in series are connected to the GBRS module, and the GBRS modules are connected to the GBRS modules.

[0083] The feature fusion network includes an Up module and a Concat module. The Up module is connected to the last GBRS module in the backbone network, the Concat module is connected to the Up module, and the Concat module is connected to the four MCBRN modules connected in series in the backbone network.

[0084] The free detection head includes two paths: the two paths have the same structure, both of which are composed of GBRS modules and convolution modules in sequence. The output of GBRS of the backbone network is used as one path input of the free detection head, and the output of the convolution module in this path is used as the first output (Output1) of the free detection head; the output of the Concat module of the feature fusion network is used as another path input of the free detection head, and the output of the convolution module in this path is used as the second output (Output2) of the free detection head.

[0085] The output dimension of the free detection head is (S, S, B × (6 + C)), where S of Output1 is 26 and S of Output2 is 13. B is the number of detection boxes for a pixel, which defaults to 3. C is the type of object that the target detection algorithm can identify. 6 represents the six elements of the channel dimension, which represent the offset g, q of the free detection box, the width and height w, h of the free detection box, the confidence and rotation p of the free detection box. .

[0086] Module description in the YOLO-F model structure:

[0087] like Figure 3 As shown, the GBRS module includes grouped convolution, BN module, ReLU activation function and SA module, where the grouped convolution is connected to the BN module (batch normalization), the BN module is connected to the ReLU activation function, and the ReLU activation function is connected to the SA module.

[0088] like Figure 4 As shown, the MCBRN module includes a maximum pooling module, a convolution module, a BN module, a ReLU activation function and a NS module, wherein the maximum pooling module is connected to the convolution module, the convolution module is connected to the BN module, the BN module is connected to the ReLU activation function, and the ReLU activation function is connected to the NS module.

[0089] like Figure 5 As shown in the figure, the NS module includes a BN module, a Weights module, a multiplication module, a maximum pooling module and an average pooling module. The BN module is connected to the Weights module, and both the Weights module and the BN module are connected to the multiplication module. The multiplication module is connected to the maximum pooling module, and the maximum pooling module is connected to the average pooling module. Among them, the Weights module is a channel weight extraction module, which uses the scaling factor in BN. The scaling factor reflects the size of the change in each channel and also indicates the importance of the channel.

[0090] like Figure 6 As shown, the SA module includes a BN module, a Weights module, a maximum pooling module, an average pooling module and a multiplication module. The BN module is connected to the Weights module, the maximum pooling module is connected to the average pooling module, and the BN module, the Weights module and the average pooling module are all connected to the multiplication module.

[0091] The loss functions of the YOLO-F model include: classification loss function, detection box regression loss function, detection box rotation loss function, and confidence loss function:

[0092]

[0093] Classification loss function:

[0094]

[0095] Detection box regression loss function:

[0096]

[0097] Detection box rotation loss function:

[0098]

[0099] Confidence loss function:

[0100]

[0101] The symbols of the loss function are as follows:

[0102] is the loss function of the YOLO-F model;

[0103] is the classification loss function of the YOLO-F model;

[0104] is the confidence loss function of the YOLO-F model;

[0105] is the rotation loss function of the YOLO-F model;

[0106] is the bounding box regression loss function of the YOLO-F model;

[0107] It is the weight value of the classification loss function, the default value is 1;

[0108] The weight value of the center point loss function, the default value is 1;

[0109] The weight value of the rotation loss function, the default value is 1;

[0110] It is the weight value of the bounding box regression loss function, the default value is 1;

[0111] is the width of the feature map;

[0112] is the height of the feature map;

[0113] The number of detection boxes, the default is 3;

[0114] Indicates The first grid cell Whether the bounding box contains the object. If so, is 1, otherwise it is 0;

[0115] Indicates The eccentricity of a grid cell in the bounding box of an object, given by the dataset label information;

[0116] Indicates The grid cell The predicted probability that a bounding box belongs to category c;

[0117] Indicates grid cells belong to category c;

[0118] Represents the set of all categories that the YOLO-F model can recognize;

[0119] Indicates The grid cell The width of the bounding box;

[0120] Indicates The width of the ground-truth bounding box of grid cells;

[0121] Indicates The grid cell The height of the bounding box;

[0122] Indicates The height of the ground-truth bounding box of each grid cell, given by the dataset label information;

[0123] Indicates The grid cell The horizontal coordinate of the center of the bounding box;

[0124] Indicates The horizontal coordinate of the center of the ground-truth bounding box of the grid cell;

[0125] Indicates The grid cell The ordinate of the center of the bounding box;

[0126] Indicates The ordinate of the center of the ground-truth bounding box of the grid cell;

[0127] is the cosine function;

[0128] Indicates The grid cell The rotation of the bounding box;

[0129] Indicates The rotation of the ground-truth bounding box of each grid cell;

[0130] It is the rotation weight ratio coefficient, the default value is 1;

[0131] Indicates The grid cell The confidence of the bounding box;

[0132] Indicates The confidence of the ground-truth bounding box for each grid cell.

[0133] The preprocessed images are used as training data for the YOLO-F model. The hyperparameters use the default settings and the YOLO-F model is iterated for 300 times. After the training, the optimal weight of the YOLO-F model is taken as the weight of the YOLO-F model. The insulator image to be tested is input into the YOLO-F model to obtain the detection result.

[0134] The YOLO-F model is designed to be a low-complexity model, which means that its model size is small and has low requirements for computing resources. This feature enables it to run on devices with limited resources, such as drones, which is very important for power line inspections, as drones usually have limited computing power and battery life. Due to its low-complexity design, the YOLO-F model is able to achieve faster forward propagation speeds, which is critical for application scenarios that require real-time detection, such as when drones inspect power lines and need to quickly and accurately identify the status of insulators. The YOLO-F model uses a mechanism called "free detection box", which allows the model to directly predict the bounding box of an object and its rotation angle instead of relying on pre-defined anchor boxes. This approach can not only more accurately match the actual shape and position of the object, especially for those objects with large rotation or posture changes in the image, but also reduce the computational overhead during training and improve training efficiency. By directly regressing the bounding box and rotation, the YOLO-F model is able to better fit the actual contours of the object, thereby improving the accuracy of detection. This is especially important for detecting targets of different shapes and colors, such as insulators, because they may appear at different angles on power lines. Due to the use of the free detection box method, the YOLO-F model avoids the large amount of computational overhead required by the traditional anchor box-based method during training, that is, it does not need to calculate the matching degree between each anchor box and the real box. This not only simplifies the training process, but also speeds up the training speed and makes the model easier to optimize and adjust.

[0135] The YOLO-F model uses a feature fusion network to fuse multi-scale features to obtain fused features, including the following steps:

[0136] The multiple high-dimensional feature maps (usually from different levels) output by the backbone network are used as inputs to the feature fusion network. These feature maps represent features at different scales and have different spatial resolutions, which can capture the performance of insulators of different sizes and positions in the image.

[0137] Feature map connection: Multiple feature maps from the backbone network are concatenated in the channel dimension. This usually includes high-resolution shallow feature maps and low-resolution deep feature maps. The concatenation operation retains information from different levels by increasing the dimension of the feature.

[0138] Specific operation: For example, three feature maps are concatenated to form a new feature map, the number of channels of which is the sum of the number of channels of the original feature maps.

[0139] Feature processing: The concatenated feature maps are further processed by convolutional layers to extract comprehensive features. These convolutional layers can learn the spatial relationship of the fused features and enhance the expressiveness of the features.

[0140] Application of activation function: After the convolution layer, an activation function (such as ReLU) is usually used for nonlinear transformation to help the model learn more complex feature representations.

[0141] Pooling layer: After fusing the feature maps, you can selectively apply the pooling layer for downsampling to further reduce the size of the feature maps. Downsampling can help the model focus on higher-level feature representations and reduce computational complexity.

[0142] Preserve information: When downsampling, make sure to preserve key information as much as possible and avoid excessive loss of details.

[0143] Generate multi-scale features: After feature fusion processing, multiple feature maps of different scales are output. These feature maps can cover insulators of different sizes, ensuring good response to objects of different sizes during the detection process.

[0144] Feature map selection: Feature maps of different layers are usually selected for subsequent target detection to form the final detection output.

[0145] The fused feature maps are passed to the detection head of the YOLO-F model for subsequent object detection and classification. These fused feature maps will provide the model with more comprehensive context information and help improve the detection accuracy of insulators.

[0146] The backbone network extracted features and fused features are passed to the free detection head for analysis. The free detection head analyzes the backbone network extracted features and fused features. The analysis results are further processed by the non-maximum suppression algorithm to finally obtain the insulator health detection results, including the following steps:

[0147] The feature maps extracted from the backbone network and processed by the feature fusion network are passed into the free detection head of the YOLO-F model. These feature maps contain rich contextual information and multi-level features, which can effectively support subsequent target detection.

[0148] Bounding box regression: The free detection head generates predicted bounding boxes based on the input feature map through a series of convolutional layers. Each bounding box consists of its center coordinates (x, y), width (w), height (h), and rotation angle (if applicable).

[0149] Category probability prediction: At the same time, the model also outputs a category probability distribution for each predicted bounding box, indicating the possibility of containing objects of different categories in the bounding box.

[0150] In the free detection head, a set of candidate boxes are generated based on the predicted bounding boxes and class probabilities. The candidate box content includes the class probability and its corresponding insulator health status. These candidate boxes are potential insulator locations identified by the YOLO-F model based on the feature map, and there may be overlap.

[0151] Confidence is calculated for each candidate box, which is usually a comprehensive evaluation of the category probability and the regression error of the bounding box. Confidence is used to measure the credibility of the model's prediction for each bounding box.

[0152] Candidate box sorting: Sort all candidate boxes according to their confidence, from high to low.

[0153] Box screening: Start with the box with the highest confidence, keep it and calculate the overlapping area with other boxes. The degree of overlap is determined by the Intersection over Union (IoU) threshold:

[0154] If the overlap of a candidate box exceeds the intersection-over-union ratio threshold (usually 0.5), it is considered to overlap too much with the current box and is deleted.

[0155] Repeat this process until all candidate boxes have been processed.

[0156] Retained boxes: After NMS processing, the remaining candidate boxes are the final detection results. These boxes have high confidence and less overlap with each other, indicating that they effectively indicate the location of the insulator.

[0157] Output detection results: The final retained bounding box is used as the output result of the YOLO-F model.

[0158] Suppose there are multiple insulators in an image of a transmission line, some of which may be damaged. We want to detect these insulators and assess their health status through the YOLO-F model;

[0159] Assume that the backbone network extracts two feature maps:

[0160] Feature A (high resolution, contains details): provides detailed information on the profile and edges of the insulator.

[0161] Feature map B (low resolution, contains contextual information): helps the model understand the positional relationship of insulators in the transmission line.

[0162] The free detection head analyzes feature map A and feature map B and generates the following predictions:

[0163] Candidate box 1: center coordinates (100, 150), width 30, height 60, confidence 0.9, category probability is "normal".

[0164] Candidate box 2: center coordinates (102, 155), width 32, height 62, confidence 0.85, category probability is "normal".

[0165] Candidate box 3: center coordinates (300, 400), width 40, height 80, confidence 0.75, category probability is "damaged".

[0166] Candidate box 1 and candidate box 2 are very close in position and may refer to the same insulator, while candidate box 3 is located at another insulator.

[0167] The confidence scores of the candidate boxes have been obtained in the prediction, which are 0.9, 0.85 and 0.75 respectively.

[0168] Sort by confidence:

[0169] Candidate box 1 (0.9);

[0170] Candidate box 2 (0.85);

[0171] Candidate box 3 (0.75);

[0172] Start processing from the candidate box with the highest confidence:

[0173] Process candidate box 1: keep.

[0174] Processing candidate box 2: Calculate the overlap (IoU) with candidate box 1. Assume IoU = 0.8, which exceeds the set threshold (0.5), so candidate box 2 is deleted.

[0175] Processing candidate box 3: The overlapping IoU with candidate box 1 may be small, so it is retained.

[0176] The remaining boxes are:

[0177] Candidate box 1: center coordinates (100, 150), category is "normal".

[0178] Candidate box 3: center coordinates (300, 400), category is "damaged".

[0179] The final output detection results are:

[0180] Insulator 1 is at position (100, 150) and its status is Normal.

[0181] Insulator 2 is at position (300, 400) and has a status of "broken".

[0182] Draw the detection box on the original image:

[0183] Draw a green box at (100, 150) and label it "Normal".

[0184] Draw a red box at (300, 400) and label it "Damage".

[0185] The images that end users see include clear annotations, showing the status and health of each insulator, helping them make timely maintenance decisions.

[0186] This method not only improves the accuracy of detection, but also reduces the probability of false detection and missed detection, providing strong support for power maintenance work.

[0187] Output the category probability distribution for each predicted bounding box. The category probability distribution indicates the possibility of containing different categories of objects in the bounding box, including the following steps:

[0188] For each predicted bounding box, the probability distribution of the soft-max output represents the likelihood of different categories of objects in the box. For example, if the model has three categories (such as "insulator", "wire", and "background"), then the output of each bounding box is the probability distribution of these three categories, such as [0.7, 0.2, 0.1]. The larger the probability, the more "confident" the model is that the bounding box belongs to the corresponding category.

[0189] Use the category probability distribution and confidence threshold for filtering. If the probability of a category is higher than the threshold (such as 0.5), the bounding box is predicted to belong to this category. If the category probabilities are all lower than the threshold, it can be determined that the bounding box does not detect the object.

[0190] The free detection head generates a set of candidate boxes based on the predicted bounding boxes and category probabilities, and calculates the confidence for each candidate box, including the following steps:

[0191] For each candidate box, the confidence calculation expression is: , where is the confidence level, represents the probability of the existence of insulators in the candidate box, is the overlap between the predicted bounding box and the true bounding box, is the focal length offset of the visible light imager, , , are adjustment coefficients, and , , Both are greater than 0.

[0192] The calculation logic of the focal length offset of the visible light imager is as follows: obtain the focal length measured during the actual shooting process and the ideal focal length of the imager, subtract the focal length measured during the actual shooting process from the ideal focal length of the imager to obtain the focal length difference, and take the absolute value of the focal length difference as the focal length offset of the visible light imager.

[0193] The probability of an insulator existing in a candidate box directly affects the confidence value of the candidate box containing the insulator. This probability usually represents the model's classification confidence in the object category in the candidate box. If the probability is high (close to 1), the model has strong confidence that the candidate box does contain an insulator; if the probability is low (close to 0), the model's confidence is weak.

[0194] IoU measures the ratio of the overlap area between the candidate box and the real box, and is a key indicator of the accuracy of bounding box positioning. The larger the IoU, the more accurate the bounding box prediction of the model, thereby improving the credibility of the detection results.

[0195] The focal offset affects the image clarity, which in turn affects the positioning accuracy and classification confidence of the bounding box. A large focal offset will cause the image to be blurred, making the model inaccurate in judging the insulator boundary, resulting in reduced confidence.

[0196] The candidate box content includes the bounding box position and size, category label, category probability, overlap degree, and insulator health detection results;

[0197] The steps to obtain the insulator health test results are as follows:

[0198] The bounding box with the insulator is divided into several 1×1 grids. All cracks and holes on the surface of the insulator are identified by the existing image recognition algorithm. The number of grids occupied by cracks and holes is recorded. The number of grids occupied by cracks and holes is divided by the total number of grids in the bounding box to obtain the damage ratio value. Multiple frames of images with insulators are obtained within a period of time. The center coordinates of the insulators in the multiple frames are obtained. The total deviation of the center coordinates of the insulators in the multiple frames is calculated based on the Euclidean distance calculation formula. The expression is: , where is the total deviation of the center coordinate, is the number of image frames, For the The center coordinates of the frame image, For the The center coordinates of the frame image;

[0199] The damage ratio value and the total deviation of the central coordinate are normalized to map the value range of the damage ratio value and the total deviation of the central coordinate to the range of [0,1], and the normalized value of the damage ratio value and the normalized value of the total deviation of the central coordinate are obtained. The normalized value of the damage ratio value and the normalized value of the total deviation of the central coordinate are summed to obtain the health impact factor of the insulator. The larger the health impact factor, the worse the overall health status of the insulator. The obtained health impact factor is compared with the preset impact threshold. If the health impact factor is less than or equal to the impact threshold, the insulator is evaluated to be in a normal health state. If the health impact factor is greater than the impact threshold, the insulator is evaluated to be in an abnormal health state.

[0200] Sort all candidate boxes according to their confidence from large to small, starting from the candidate box with the highest confidence, retain the candidate box and calculate the overlapping area with other candidate boxes, and determine the degree of overlap by the intersection-over-union ratio threshold. If the overlap degree of a candidate box exceeds the intersection-over-union ratio threshold, delete the candidate box and repeat this process until all candidate boxes are processed, including the following steps:

[0201] Collect a list of all candidate boxes, each containing bounding box coordinates, class label, and confidence.

[0202] Sort the candidate box list in descending order according to confidence, ensuring that the candidate box with the highest confidence is at the front of the list.

[0203] Determine an intersection-over-union (IoU) threshold, which is used to judge the degree of overlap between candidate boxes and is usually set between 0.3 and 0.7.

[0204] Create a new list to store the final retained candidate boxes.

[0205] Select the candidate box with the highest confidence from the sorted candidate box list, add it to the final retained list, and delete the selected candidate box from the sorted candidate box list.

[0206] Loop through the remaining candidate boxes (from the sorted list) and calculate the candidate boxes and candidate boxes The overlapping area between the candidate boxes and candidate boxes The overlapping area between them is greater than the intersection-over-union ratio threshold, then the candidate box delete.

[0207] After generating corresponding management suggestions based on the insulator health test results, the test results and management suggestions are mapped to images, which are saved to a local storage device and uploaded to the cloud at the same time, including the following steps:

[0208] According to the insulator health test results, the test data is analyzed and corresponding management recommendations are generated. These recommendations may include:

[0209] Recommendations for periodic inspection or maintenance.

[0210] List of insulators that need to be replaced.

[0211] Recommendations for cleaning or treatment of existing insulators.

[0212] Prepare information to be displayed in the image, including:

[0213] Insulator health status (normal / abnormal).

[0214] Specific management recommendations.

[0215] The confidence level or related metrics of the detection.

[0216] Select Image: Select the original image associated with the detection result.

[0217] Mark the insulator position: Mark the detected insulator position on the image, usually using a bounding box.

[0218] Add text information: Add text information, including health status and management suggestions, to appropriate locations in the image. Make sure the text is clearly visible and has appropriate color contrast.

[0219] For example, display "Health Status: Abnormal" next to the bounding box along with specific recommendations.

[0220] Save the image with the inspection results and management suggestions added to your local storage device. Make sure the file name clearly indicates the content of the image (for example, include the inspection date and insulator ID).

[0221] Connect to cloud storage services and upload processed images to the cloud for remote access and backup.

[0222] Make sure the uploaded file format and naming convention complies with the cloud storage requirements.

[0223] Monitor the upload status during the upload process to ensure that the upload is successful.

[0224] Record log information during the upload process, including:

[0225] Upload time.

[0226] The local file path.

[0227] The access link or storage path of the cloud storage.

[0228] After the upload is complete, check whether the files in the cloud storage are consistent with the files in the local storage to ensure data integrity.

[0229] Embodiment 3: A real-time health detection system for transmission line insulators, comprising an image acquisition module, a feature analysis module, and a health detection module;

[0230] Image acquisition module: controls the drone to fly along a predetermined path and uses the visible light imager on the drone to collect images along the transmission line. The collected images are transmitted to the feature analysis module in real time.

[0231] Feature analysis module: First, the image is preprocessed, and then the preprocessed image is input into the pre-trained YOLO-F model. The YOLO-F model includes a backbone network, a feature fusion network, and a free detection head. The YOLO-F model extracts features through the backbone network, and uses the feature fusion network to perform multi-scale feature fusion to obtain fused features. The backbone network extracted features and fused features are transmitted to the free detection head for analysis, and the free detection head sends the analysis results to the health detection module.

[0232] Health detection module: The analysis results are further processed through the non-maximum suppression algorithm to finally obtain the insulator health detection results. After generating corresponding management suggestions based on the insulator health detection results, the detection results and management suggestions are mapped to images, and the images are saved to the local storage device and uploaded to the cloud at the same time.

[0233] It should be understood that the term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. A and B can be singular or plural. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship, but it may also indicate an "and / or" relationship. Please refer to the context for specific understanding.

[0234] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0235] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application. Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0236] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A real-time health detection method for transmission line insulators, characterized in that: The detection method comprises the following steps: The detection system controls the UAV to fly along a predetermined path and uses the visible light imager on the UAV to collect images along the transmission line. The collected images are transmitted to the detection system in real time. The detection system first preprocesses the image, and then inputs the preprocessed image into the pre-trained YOLO-F model. The YOLO-F model includes a backbone network, a feature fusion network, and a free detection head. The YOLO-F model extracts features through the backbone network, and uses the feature fusion network to perform multi-scale feature fusion to obtain fused features. The backbone network extracted features and fused features are then passed to the free detection head for analysis. The free detection head analyzes the backbone network to extract features and fuse features. The analysis results are further processed by the non-maximum suppression algorithm to finally obtain the insulator health detection results. After generating corresponding management suggestions based on the insulator health detection results, the detection results and management suggestions are mapped to images, which are saved to the local storage device and uploaded to the cloud at the same time. The free detection head analyzes the backbone network to extract features and fuse features. The analysis results are further processed by the non-maximum suppression algorithm to finally obtain the insulator health detection results, including the following steps: The feature maps extracted from the backbone network and the feature maps processed by the feature fusion network are passed into the free detection head of the YOLO-F model; The free detection head generates predicted bounding boxes based on the input feature map. Each bounding box consists of the center coordinates, width, height and rotation angle of the bounding box. Output the category probability distribution for each predicted bounding box, which indicates the possibility of containing different categories of objects in the bounding box; The free detection head generates a set of candidate boxes based on the predicted bounding boxes and category probabilities, and calculates the confidence for each candidate box; Sort all candidate boxes according to their confidence from large to small, start with the candidate box with the highest confidence, keep the candidate box and calculate the overlapping area with other candidate boxes, and determine the degree of overlap by the intersection-over-union ratio threshold: If the overlap of a candidate box exceeds the intersection-over-union ratio threshold, the candidate box is deleted and the process is repeated until all candidate boxes are processed; After processing, the remaining candidate boxes are the final detection results, and the final retained bounding boxes are used as the output results of the YOLO-F model; The free detection head generates a set of candidate boxes based on the predicted bounding boxes and category probabilities, and calculates the confidence for each candidate box, including the following steps: For each candidate box, the confidence calculation expression is: , where is the confidence level, represents the probability of the existence of insulators in the candidate box, is the overlap between the predicted bounding box and the true bounding box, is the focal length offset of the visible light imager, , , are adjustment coefficients, and , , All are greater than 0; The candidate box content includes the position and size of the bounding box, the category label, the category probability, the degree of overlap, and the insulator health detection result; The steps to obtain the insulator health test results are as follows: Normalize the damage ratio value and the total deviation of the center coordinates, map the value range of the damage ratio value and the total deviation of the center coordinates to the range of [0,1], obtain the normalized value of the damage ratio value and the normalized value of the total deviation of the center coordinates, sum the normalized value of the damage ratio value and the normalized value of the total deviation of the center coordinates to obtain the health impact factor of the insulator, compare the obtained health impact factor with the preset impact threshold, if the health impact factor is less than or equal to the impact threshold, evaluate the insulator to be in a normal health state, if the health impact factor is greater than the impact threshold, evaluate the insulator to be in an abnormal health state; The bounding box with insulators is divided into several 1×1 grids. All cracks and holes on the surface of the insulator are identified by the image recognition algorithm. The number of grids occupied by cracks and holes is recorded. The number of grids occupied by cracks and holes is divided by the total number of grids in the bounding box to obtain the damage ratio value. Multiple frames of images with insulators are obtained during the monitoring period. The center coordinates of the insulators in the multiple frames are obtained. The total deviation of the center coordinates of the insulators in the multiple frames is calculated based on the Euclidean distance calculation formula. The expression is: , where is the total deviation of the center coordinates, is the number of image frames, For the The center coordinates of the frame image, For the The center coordinates of the frame image.

2. A real-time health detection method for transmission line insulators according to claim 1, characterized in that: The detection system preprocesses the image, including the following steps: The LabelImg software is used to label the insulators in the image. The insulator labeling information includes the coordinates of the insulator center, the width and height of the insulator bounding box, and the rotation angle of the insulator bounding box. The eccentricity of all pixels in the insulator bounding box is also calculated: , where is a pixel ( i, j ) of the eccentricity, is the horizontal coordinate position of the pixel point, for( i , j ) is the central horizontal coordinate of the real frame where the pixel point is located, is the vertical coordinate of the pixel, y for( i, j ) The vertical coordinate of the real frame of the pixel point, for( i, j ) The width of the real frame where the pixel point is located, for( i， j ) The height of the real frame where the pixel is located, is a disturbance greater than zero, ensuring that the entire formula is not zero, Is the natural base.

3. A real-time health detection method for transmission line insulators according to claim 2, characterized in that: The preprocessed image is input into the pre-trained YOLO-F model. The YOLO-F model extracts features through the backbone network, including the following steps: Pass the preprocessed image into the input layer of the YOLO-F model; In the convolution layer, the image first undergoes multiple layers of convolution operations. Each convolution layer extracts local features on the image by sliding the convolution kernel. Each convolution kernel is used to detect features in the image. An activation function is applied after each convolution layer, and negative values ​​are set to zero. Between the convolutional layers, the pooling layer is used to reduce the size of the feature map. The maximum pooling selects the maximum value in the feature map. The pooling operation uses a preset step size to reduce the spatial dimension of the feature map layer by layer. After several layers of convolution and pooling, the backbone network outputs a high-dimensional feature map, which contains multi-level feature information of insulators in the image.

4. A real-time health detection method for transmission line insulators according to claim 3, characterized in that: The YOLO-F model uses a feature fusion network to fuse multi-scale features to obtain fused features, including the following steps: The multiple high-dimensional feature maps output by the backbone network are used as the input of the feature fusion network, and the multiple high-dimensional feature maps are concatenated in the channel dimension; The concatenated high-dimensional feature map is further processed through the convolution layer to extract the comprehensive feature map, and after the convolution layer, the activation function is used for nonlinear transformation; After fusing the feature maps, a pooling layer is applied for downsampling to reduce the size of the comprehensive feature map; After feature fusion processing, multiple feature maps of different scales are output, and feature maps of different layers are selected for subsequent target detection; The fused feature map is passed to the free detection head of the YOLO-F model for subsequent target detection and classification.

5. A real-time health detection system for transmission line insulators, used to implement the detection method according to any one of claims 1 to 4, characterized in that: Including image acquisition module, feature analysis module and health detection module; Image acquisition module: controls the drone to fly along a predetermined path and uses the visible light imager on the drone to collect images along the transmission line. The collected images are transmitted to the feature analysis module in real time. Feature analysis module: First, the image is preprocessed, and then the preprocessed image is input into the pre-trained YOLO-F model. The YOLO-F model includes a backbone network, a feature fusion network, and a free detection head. The YOLO-F model extracts features through the backbone network, and uses the feature fusion network to perform multi-scale feature fusion to obtain fused features. The backbone network extracted features and fused features are transmitted to the free detection head for analysis, and the free detection head sends the analysis results to the health detection module. Health detection module: The analysis results are further processed through the non-maximum suppression algorithm to finally obtain the insulator health detection results. After generating corresponding management suggestions based on the insulator health detection results, the detection results and management suggestions are mapped to images, and the images are saved to the local storage device and uploaded to the cloud at the same time.

Images