Artificial intelligence-based power transmission channel safety online monitoring method

Abstract

The application provides a power transmission channel safety online monitoring method based on artificial intelligence. The method is characterized in that: through acquiring power transmission channel monitoring image and laser point cloud data, the image is preprocessed; the coordinate conversion relationship is acquired by using camera calibration and pose estimation algorithm; a single-stage detector is constructed, and the training and parameter adjustment of the model are completed; the minimum distance from the target object to the power line is calculated by using the target detection result and the coordinate conversion relationship; finally, according to the regulation of the safety distance of overhead line in the enterprise standard of the power grid company, the safety level is judged, and the target object is timely warned. The method helps the inspection department to judge the risk level of hidden dangers, realizes the automatic monitoring and early warning of the power line, and has important research significance for effectively guaranteeing the safety of personnel life, the safety production of power enterprises and the stability of resident power consumption.

Description

Technical Field

[0001] This invention belongs to the field of power equipment safety monitoring technology, and relates to an online monitoring method for power transmission channel safety based on artificial intelligence. Background Technology

[0002] Traditional power transmission channel inspections mainly rely on periodic manual inspections, which have many drawbacks such as a lack of responsibility among inspectors, limited equipment levels, and reliance on experience for misjudgment. With the development of science and technology and the advancement of intelligent inspections, the monitoring mode based on visual terminals can better solve the above problems. However, relying solely on remote monitoring to obtain the status of power transmission channels has a lag and cannot automatically warn of accidents.

[0003] Existing technologies for monitoring the safety of overhead power lines primarily rely on deep learning-based methods for detecting foreign objects in transmission channels. However, in overhead line scenarios, while many potential hazards are captured and identified, the number of truly meaningful alarms is limited. Furthermore, relevant safety regulations stipulate safe distances for overhead lines at different voltage levels, and relying solely on target detection algorithms cannot determine whether a foreign object will pose a danger to the transmission line. Additionally, some existing research uses binocular vision-based 3D reconstruction methods, which monitor the spatial position of foreign objects to determine whether they have entered a defined alarm area, thus achieving real-time monitoring and early warning. However, binocular cameras require high precision in stereo matching during calibration and matching, and relative positional shifts during installation can reduce recognition accuracy. Therefore, an effective method for monitoring and issuing early warnings regarding safe distances to power lines is lacking. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an online monitoring method for power transmission channel safety based on artificial intelligence. This method combines deep learning and multi-sensor fusion technology to solve the problem of measuring and monitoring the safe distance of power lines. It extracts and matches features from three-dimensional point cloud data and two-dimensional images, and combines them with a single-stage detector target recognition algorithm to obtain the three-dimensional distance from foreign objects to power lines. This helps inspection departments to judge the risk level of hidden dangers and realize automatic monitoring and early warning of power transmission channels.

[0005] To achieve the above-mentioned technical effects, the technical solution adopted by this invention is: an online monitoring method for the safety of power transmission channels based on artificial intelligence, which includes the following steps:

[0006] S1. Acquire surveillance images and laser point cloud data of the power transmission channel, filter and preprocess the images, and establish a dedicated database of foreign object images of the power transmission channel:

[0007] 2D images were acquired using surveillance cameras mounted on power poles; laser point cloud data were collected using drones equipped with LiDAR, with the data following the State Grid Corporation's "Regulations on the Application and Management of Laser Scanning Technology for Overhead Transmission Lines (Trial)"; the acquired 2D images were filtered, labeled, and expanded, and image data and labels containing foreign objects in the transmission channel were randomly extracted from the benchmark dataset COCO-datasets. Combined with a self-built dataset for power scenarios, a dedicated database of foreign object images in the transmission channel was established.

[0008] S2. Construct an optimized YOLOX single-stage detector, complete the model training and parameter tuning, and detect vehicles in the power transmission channel. The specific method is as follows: improve the backbone network of YOLOX and optimize the positive sample selection strategy.

[0009] Furthermore, the original backbone network CSPDarkNet was replaced with the higher-performing ConvNeXt. The ConvNeXt module primarily employs an inverse bottleneck structure, performing convolution operations in an order of dimensionality increase followed by dimensionality decrease, thus avoiding information loss. The depthwise separable convolutions within the module consist of both depthwise and pointwise convolutions, ensuring that each convolutional kernel is responsible for the feature map of one channel, significantly reducing the network's computational cost and parameter count. Additionally, ConvNeXt greatly reduces the number of downsampling operations, setting a downsampling layer only before each ConvNeXt block. This is because excessive downsampling operations can lead to excessively low spatial resolution of the feature map, making recovery difficult, while completely removing downsampling would increase network computation. These structural designs allow the ConvNeXt network to achieve a good balance between detection accuracy and computational complexity.

[0010] Furthermore, the positive sample selection strategy of the SimOTA algorithm is optimized to ensure that the network selects higher-quality positive samples for loss function calculation. The implementation steps of the SimOTA algorithm are as follows:

[0011] (1) Pre-filter the predicted boxes near the label boxes based on the center point position;

[0012] (2) Calculate the IoU between each label box and each prediction box;

[0013] (3) Calculate the sum k of the top 10 IoUs. i , which is the number of prediction boxes required for each label box i;

[0014] (4) Calculate the cost value between each prediction box and the label box, and select the box with the smallest cost value. iOne predicted bounding box is used as a positive sample for label box i, and the rest are negative samples; the final selected positive samples are used to calculate the classification loss and regression loss in the loss function, and all samples are used to calculate the confidence loss; the cost matrix is ​​calculated using the following formula:

[0015] cost = c × loss cls +r×loss reg =BCE_loss + 3 × IoU_loss

[0016] The purpose of the cost matrix is ​​to adaptively fit the true bounding boxes to the predicted bounding box samples. The principle is that samples with smaller IoU, more accurate classification, and closer center points have smaller cost values. In 2D-YOLO, the cost matrix is ​​optimized. Its classification loss is replaced with varifocal loss, and the regression loss is replaced with the more advanced giouloss, as shown in the following equation.

[0017] cost = vfl_loss + 3 × GIoU_loss

[0018] in,

[0019]

[0020]

[0021] vfl_loss and GIoU_loss are the varifocal loss and GIoU loss, respectively; q is the IoU between the predicted bounding box and the label bounding box, q>0 indicates a positive sample, q=0 indicates a negative sample, and the larger the sample IoU, the greater its contribution to the loss; p is the prediction score; α is the balancing coefficient; γ is the loss scaling factor to reduce the loss contribution of negative samples and retain the learning information of positive samples; A m Let represent the area of ​​the smallest bounding rectangle of the two boxes, and U represent the area of ​​the union of the two boxes. The goal of the SimOTA algorithm is to select positive sample prediction boxes from a large number of pre-selected boxes. This requires a greater focus on positive samples. On the one hand, varifocal loss reduces the contribution of negative samples to the loss by adding a scaling factor to the loss of negative samples. On the other hand, it assigns greater weight to positive samples with larger IoU, which can encourage the model to focus its training on higher-quality positive samples.

[0022] Furthermore, the improved YOLOX model outputs category information, target location information, and target confidence information. Therefore, for the model's training task, its total loss L YOLOX It consists of three parts: classification loss L cls Regression loss L reg Target confidence loss Lobj As shown in the following formula:

[0023] L YOLOX =a1L cls +a2L reg +a3L obj =BCE_loss+5×IoU_loss+BCE_loss

[0024] in,

[0025] BCE_loss=-(tlog(p)+(1-t)log(1-p))

[0026]

[0027] BCE_loss and IoU_loss represent the binary cross-entropy loss and IoU loss, respectively; p is the predicted probability of each class; t indicates whether it is the correct class or target, with a value of 0 or 1; IoU is the ratio of the area of ​​the intersection region to the area of ​​the union region of two boxes; I(gt,pred) represents the area of ​​the intersection of the label box and the predicted box, and U(gt,pred) represents the area of ​​the union of the label box and the predicted box; only the positive sample predicted boxes selected by SimOTA participate in the calculation of classification loss and regression loss, while confidence loss involves all samples; in addition, the training process replaces the original Cosine learning mechanism with a piecewise decay strategy to optimize the learning rate decay method; the optimization scheme is verified using ablation experiments; the trained model is tested and verified on the test set and compared with advanced detectors.

[0028] S3. Obtain the coordinate transformation relationship using camera calibration and attitude estimation algorithms. The specific method is as follows:

[0029] S301 uses camera calibration technology to obtain the intrinsic parameters and distortion coefficients of a monocular camera. To ensure higher calibration accuracy, calibration plates of different sizes and materials were tested. During the calibration process, the calibration plate was fixed and the camera was moved to calibrate multiple images from different positions and angles, so as to ensure that the set of positions of the calibration plate covered the entire field of view of the camera in all calibration images.

[0030] S302 utilizes a pose estimation algorithm to jointly match feature points from a 2D image and a 3D point cloud, calculating their coordinate transformation relationship. The coordinate transformation relationship for the pinhole camera model is as follows:

[0031]

[0032] Among them, (X) c ,Y c Z c (Z) represents the camera coordinate system coordinates.c Also represents the depth value; (uv) is the pixel coordinate system coordinate; (X) w ,Y w Z w ) represents the world coordinate system coordinates; dx and dy represent the physical dimensions of each pixel in the x and y directions of the image coordinate system, respectively; f is the camera focal length in mm; f x f y These are the focal lengths of the camera on the x and y axes, respectively; (u0, v0) are the pixel coordinates corresponding to the camera's optical center; f x ,f y u0 and v0 form the camera's intrinsic parameter matrix M; R is a 3×3 rotation matrix, and T is a 3×1 translation vector. R and T together form the extrinsic parameter E for the transformation from world coordinates to image polar coordinates. In addition, the radial and tangential distortions of the camera need to be considered in the coordinate transformation, namely the distortion coefficients D, which include five key parameters (k1, k2, k3, p1, p2).

[0033] S303, based on the above mapping relationship from world coordinates to pixel coordinates, select more than 4 matching points in the 2D image and 3D point cloud for parameter solving of the back projection transformation.

[0034] S4. Using the target detection results and coordinate transformation relationships, calculate the minimum distance from the top of the target object to the electric field line:

[0035] Using the target detection results and coordinate transformation relationships, the minimum three-dimensional distance from the top of the target detection box to the electric field line is calculated. The formula for the three-dimensional Euclidean distance is:

[0036]

[0037] S5. Based on the regulations on safe distances for overhead lines in the power grid company's enterprise standards, determine the safety level and provide timely warnings to target objects, as detailed below:

[0038] The algorithm compares the predicted three-dimensional distance with the power grid company's enterprise standards, and intelligently determines the safety level based on the safety distance of overhead lines at different voltage levels. For targets that are less than or close to the safety distance, it issues timely warnings.

[0039] The beneficial effects of this invention are as follows:

[0040] This invention provides an artificial intelligence-based online monitoring method for power transmission channel safety. It establishes a model for measuring and monitoring the safe distance of power transmission channels. By analyzing deep learning-based multi-target image detection algorithms, it focuses on three key indicators: speed, accuracy, and computational complexity of foreign object image recognition algorithms for power transmission channels. It studies a regression-based detection algorithm network structure to achieve rapid and accurate localization and identification of foreign objects. The method explores the mapping relationship between camera 2D coordinates and LiDAR 3D coordinates, using camera calibration and pose estimation algorithms to obtain the 2D-3D coordinate transformation relationship. Combining the target detection box of the image with the 2D-3D mapping relationship, it further realizes the task of measuring the shortest 3D distance between foreign objects and power lines in power transmission channels under a monocular camera. Based on relevant enterprise standards, it ultimately achieves safe distance measurement and monitoring early warning for power transmission channels in on-site environments. The proposed optimized single-stage detector can accurately identify and locate foreign object targets. Using LiDAR data as prior information, it utilizes monocular vision to achieve effective distance measurement and monitoring, significantly reducing manpower and material consumption. Attached Figure Description

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0043] Figure 2 This is a flowchart illustrating the technical concept of an embodiment of the present invention;

[0044] Figure 3 This is a diagram of the backbone network structure according to an embodiment of the present invention;

[0045] Figure 4 This is a training loss diagram of the improved YOLOX according to an embodiment of the present invention;

[0046] Figure 5 and Figure 6 This is a diagram showing the YOLOX detection results of an embodiment of the present invention;

[0047] Figure 7 and Figure 8 This is a diagram showing the detection results of the method proposed in this embodiment of the invention;

[0048] Figure 9 The test diagram is selected for the calibration plate in this embodiment of the invention;

[0049] Figure 10 This is a diagram showing the shooting position and angle of the calibration camera in an embodiment of the present invention.

[0050] Figure 11 This is a corner detection diagram according to an embodiment of the present invention;

[0051] Figure 12This refers to the calibration reprojection error in this embodiment of the invention.

[0052] Figure 13 This is a 2D-3D joint matching graph according to an embodiment of the present invention;

[0053] Figure 14 This is a coordinate transformation diagram of the camera imaging model in an embodiment of the present invention;

[0054] Figure 15 and Figure 16 This is a diagram showing the minimum distance from the van to power lines ② and ③ in the video frames of an embodiment of the present invention. Detailed Implementation

[0055] like Figure 1 The present invention discloses an artificial intelligence-based online monitoring method for power transmission channel safety, which includes the following steps:

[0056] S1. Acquire surveillance images and laser point cloud data of the power transmission channel, filter and preprocess the images, and establish a dedicated database of foreign object images of the power transmission channel;

[0057] S2. Construct an optimized YOLOX single-stage detector, complete the model training and parameter tuning, and detect vehicles in the power transmission channel;

[0058] S3. Obtain coordinate transformation relationships using camera calibration and attitude estimation algorithms;

[0059] S4. Using the target detection results and coordinate transformation relationship, calculate the minimum distance from the top of the target object to the electric field line;

[0060] Preferably, such as Figure 2 As shown, it mainly consists of two parts: a 2D single-stage detector and a 3D ranging system.

[0061] S5. Determine the safety level and provide timely warnings to target objects based on the safety distance regulations for overhead lines in the power grid company's enterprise standards.

[0062] In the preferred scheme, in S1, 2D images are acquired by monitoring cameras mounted on the towers; laser point cloud data is collected using a drone equipped with LiDAR, and the data follows the "Regulations on the Application and Management of Laser Scanning Technology for Overhead Transmission Lines (Trial)" issued by the State Grid Equipment Department; the collected 2D images are filtered, labeled and expanded, and image data and labels containing foreign objects in the transmission channel are randomly extracted from the benchmark dataset COCO-datasets. Combined with a self-built dataset in the power scenario, a dedicated database of foreign object images in the transmission channel is established.

[0063] In the preferred embodiment, in S2, the backbone network of YOLOX is improved, and the positive sample selection strategy is optimized:

[0064] S201 replaces the original backbone network CSPDarkNet with the better-performing ConvNeXt;

[0065] Preferably, ConvNeXt combines the structural characteristics of ResNet and Vision Transformer, achieving state-of-the-art performance on the ImageNet classification dataset. Based on model performance or complexity, ConvNeXt is categorized into four forms. Considering both detection performance, speed, and computational cost, ConvNeXt-S is selected as the backbone network, with the following structure: Figure 3 As shown.

[0066] Furthermore, the ConvNeXt network mainly consists of a basic input module C1 and four parts C2-C5 with ConvNeXt stacked blocks, with stacking times of (3, 3, 27, 3). The ConvNeXt modules primarily employ an inverse bottleneck structure, performing convolution operations in an order of dimensionality increase followed by dimensionality decrease, thus avoiding information loss. The depthwise separable convolutions within the modules consist of depthwise and pointwise convolutions. This structure ensures that one convolutional kernel is responsible for the feature map of one channel, significantly reducing the network's computational cost and parameter count. Additionally, ConvNeXt greatly reduces the number of downsampling operations, placing a downsampling layer only before each ConvNeXt block. This is because excessive downsampling operations can lead to excessively low spatial resolution of the feature map, making recovery difficult. Completely removing downsampling would also increase the network's computational cost. These structural designs allow the ConvNeXt network to achieve a good balance between detection accuracy and computational complexity.

[0067] S202 optimizes the positive sample selection strategy of the SimOTA algorithm for the original YOLOX model to ensure that the network selects higher quality positive samples for loss function calculation.

[0068] Furthermore, the implementation steps of the SimOTA algorithm are as follows:

[0069] (1) Pre-filter the predicted boxes near the label boxes based on the center point position;

[0070] (2) Calculate the IoU between each label box and each prediction box;

[0071] (3) Calculate the sum k of the top 10 IoUs. i , which is the number of prediction boxes required for each label box i;

[0072] (4) Calculate the cost value between each prediction box and the label box, and select the box with the smallest cost value. iOne predicted bounding box is used as a positive sample for label box i, and the rest are negative samples. The final selected positive samples are used to calculate the classification loss and regression loss in the loss function, and all samples are used to calculate the confidence loss; the cost matrix is ​​calculated using the following formula:

[0073] cost = c × loss cls +r×loss reg =BCE_loss + 3 × IoU_loss

[0074] Preferably, the purpose of the Cost matrix is ​​to adaptively fit the true bounding box to the predicted bounding box sample. The principle followed is that the smaller the IoU, the more accurate the classification, and the closer the center point, the smaller the cost value. In 2D-YOLO, the cost matrix is ​​optimized. Its classification loss is replaced by varifocal loss, and the regression loss is replaced by the more advanced giouloss, as shown in the following formula.

[0075] cost = vfl_loss + 3 × GIoU_loss

[0076] in,

[0077]

[0078]

[0079] vfl_loss and GIoU_loss are the varifocal loss and GIoU loss, respectively; q is the IoU between the predicted bounding box and the label bounding box, q>0 indicates a positive sample, q=0 indicates a negative sample, and the larger the sample IoU, the greater its contribution to the loss; p is the prediction score; α is the balancing coefficient; γ is the loss scaling factor to reduce the loss contribution of negative samples and retain the learning information of positive samples; A m Let U represent the area of ​​the smallest bounding rectangle of the two boxes, and let U represent the area of ​​the union of the two boxes.

[0080] Furthermore, the goal of the SimOTA algorithm is to select positive sample prediction boxes from a large number of pre-screened boxes, which requires a greater focus on positive samples. On the one hand, varifocal loss reduces the contribution of negative samples to the loss by adding a scaling factor to the loss of negative samples; on the other hand, it assigns greater weight to positive samples with larger IoU, which can encourage the model to focus its training on higher quality positive samples.

[0081] In the preferred embodiment, the improved YOLOX model outputs category information, target location information, and target confidence information. Therefore, for the model's training task, its total loss L YOLOX It consists of three parts: classification loss L cls Regression loss Lreg Target confidence loss L obj As shown in the following formula:

[0082] L YOLOX =a1L cls +a2L reg +a3L obj =BCE_loss+5×IoU_loss+BCE_loss

[0083] in,

[0084] BCE_loss=-(tlog(p)+(1-t)log(1-p))

[0085]

[0086] BCE_loss and IoU_loss represent the binary cross-entropy loss and IoU loss, respectively; p is the predicted probability of each class; t indicates whether it is the correct class or target, with a value of 0 or 1; IoU is the ratio of the area of ​​the intersection region to the area of ​​the union region of two boxes; I(gt,pred) represents the area of ​​the intersection of the label box and the predicted box, and U(gt,pred) represents the area of ​​the union of the label box and the predicted box; only the positive sample predicted boxes selected by SimOTA participate in the calculation of classification loss and regression loss, while confidence loss involves all samples.

[0087] Furthermore, the training process replaces the original Cosine learning mechanism with a Piecewise Decay strategy to optimize the learning rate decay method; the optimization scheme is validated using ablation experiments; the trained model is tested and validated on the test set and compared with advanced detectors.

[0088] Preferably, the single-stage detector is initialized for training starting with ConvNeXt weights; stochastic gradient descent with Nesterov momentum and L2 regularization terms are used to optimize the objective function, with momentum and weight decay factors of 0.9 and 0.0005, respectively; the batch size of each training input is set to 2, and training is performed for a total of 200 epochs, followed by training with Mosaic & Mixup image augmentation up to 185 epochs. Furthermore, a series of data augmentation strategies are incorporated into the model training, such as color gamut enhancement, scale transformation, and Mosaic & Mixup image augmentation.

[0089] Preferably, the optimized scheme of YOLOX was subjected to ablation experiments. Different schemes were trained separately and tested on the test set. When the IoU threshold was 0.5, the proposed scheme not only achieved the highest detection accuracy of 0.855, but also the fastest detection speed of 23 FPS.

[0090] Preferably, the training loss curve of the proposed method and cosine learning mechanism training model is as follows: Figure 4 As shown, the piecewise learning mechanism allows the model to converge faster and to a lower loss value.

[0091] Preferably, such as Figures 5-8 As shown, through comparison Figures 5-8 For complex environments and small targets, the proposed method can still accurately identify and locate the target. However, due to background occlusion and target overlap, YOLOX may produce false detections; for example... Figure 5 and Figure 6 The presence of extra detection boxes demonstrates that the optimized model is more robust.

[0092] In the preferred embodiment, in S3, the intrinsic parameters and distortion coefficients of the monocular camera are obtained using camera calibration technology; and the coordinate transformation relationship is calculated by jointly matching the feature points of the 2D image and the 3D point cloud using a pose estimation algorithm.

[0093] Preferably, to ensure higher calibration accuracy, calibration plates of different sizes and materials were tested, such as... Figure 9 As shown; during the calibration process, the calibration plate is fixed and the camera is moved to calibrate multiple images from different positions and angles, so as to ensure that the combined position of the calibration plate under all calibration images covers the entire field of view of the camera. Figure 10 The location and angle from which the experiment was filmed were shown; Figure 11 and Figure 12 The results of corner detection and reprojection error after calibration are shown respectively.

[0094] Preferably, pose estimation of the camera image and laser point cloud is achieved using 2D-3D joint matching, selecting more than 4 pairs of matching points based on the principle of edge point priority, such as... Figure 13 As shown.

[0095] Preferably, Figure 14 This paper demonstrates the scene of video surveillance of a power transmission channel and the four coordinate systems of the camera imaging model. 2D images of the transmission channel are acquired using cameras mounted on power towers, while 3D point clouds are acquired by a LiDAR-equipped UAV. The LiDAR point cloud serves as prior information, used to jointly acquire extrinsic parameters from the surveillance images. Joint matching adjusts the 3D point cloud perspective of the LiDAR based on the 2D image viewpoint. Through rotation, translation, and projection, the mapping of 3D points coincides with the 2D points to obtain transformation parameters. The coordinate transformation relationship of the pinhole camera model is as follows:

[0096]

[0097] Among them, (X) c ,Yc Z c (Z) represents the camera coordinate system coordinates. c Also represents the depth value; (uv) is the pixel coordinate system coordinate; (X) w ,Y w Z w ) represents the world coordinate system coordinates; dx and dy represent the physical dimensions of each pixel in the x and y directions of the image coordinate system, respectively; f is the camera focal length in mm; f x f y These are the focal lengths of the camera on the x and y axes, respectively; (u0, v0) are the pixel coordinates corresponding to the camera's optical center; f x ,f y u0 and v0 form the camera's intrinsic parameter matrix M; R is a 3×3 rotation matrix, and T is a 3×1 translation vector. R and T together form the extrinsic parameter E for the transformation from world coordinates to image polar coordinates. In addition, the radial and tangential distortions of the camera need to be considered in the coordinate transformation, namely the distortion coefficients D, which include five key parameters (k1, k2, k3, p1, p2).

[0098] Preferably, based on the aforementioned mapping relationship from world coordinates to pixel coordinates, more than four matching points are selected in the 2D image and 3D point cloud for parameter solving of the back projection transformation. Using the target detection results and coordinate transformation relationships, the minimum three-dimensional distance from the top of the target detection box to the electric field line is calculated. The formula for the three-dimensional Euclidean distance is:

[0099]

[0100] Preferably, the distance from the electric field lines in the laser point cloud to the ground is measured and used as the distance label for the experiment. The experiment uses monitored video data as input, and the model outputs the detection results frame by frame. Figures 15-16 The minimum distance from vehicles to power lines in consecutive frames was shown; the predicted distance from power lines to the ground was calculated based on the standard vehicle height, and the measurement error was further calculated to verify the ranging accuracy.

[0101] In the preferred scheme, in S5, the algorithm compares the predicted three-dimensional distance with the relevant regulations of the power grid company's enterprise standards, and intelligently judges the safety level based on the safety distance of overhead lines at different voltage levels. For targets that are less than or close to the safety distance, a timely warning is issued.

Claims

1. A method for online safety monitoring of power transmission channels based on artificial intelligence, characterized in that: Includes the following steps: S1. Acquire surveillance images and laser point cloud data of the power transmission channel, filter and preprocess the images, and establish a dedicated database of foreign object images of the power transmission channel; S2. Construct an optimized YOLOX single-stage detector, complete model training and parameter tuning, to detect vehicles in power transmission channels; the optimized YOLOX single-stage detector improves the YOLOX backbone network and optimizes the positive sample selection strategy, outputting category information, target location information, and target confidence information; the specific method is as follows: S201, the original backbone network CSPDarkNet is replaced with ConvNeXt; the depthwise separable convolutions in the ConvNeXt module consist of depthwise convolutions and pointwise convolutions; S202, optimize the positive sample selection strategy of the SimOTA algorithm to ensure that the network selects higher quality positive samples for loss function calculation. The implementation steps of the SimOTA algorithm are as follows: (1) Pre-select the predicted boxes near the label boxes based on the center point position; (2) Calculate the IoU between each label box and each prediction box; (3) Calculate the sum of the top 10 IoUs. k i That is, each label box i The required number of prediction boxes; (4) Calculate the cost value between each prediction box and the label box, and select the box with the lowest cost value. k i Each prediction box serves as a label box. i The positive samples are selected first, and the rest are negative samples. The final selected positive samples are used to calculate the classification loss and regression loss in the loss function, and all samples are used to calculate the confidence loss. The cost matrix is ​​calculated using the following formula: In the cost matrix, samples with smaller IoU, more accurate classification, and closer centroids have smaller cost values. In 2D-YOLO, the cost matrix is ​​optimized, with the classification loss replaced by varifocal loss and the regression loss replaced by giouloss, as shown in the following equation: in, In the formula, vfl_loss and GIoU_loss are the varifocal loss and GIoU loss, respectively; q It is the IoU between the prediction bounding box and the label bounding box. q >0 indicates a positive sample. q =0 indicates a negative sample; the larger the sample IoU, the greater its contribution to the loss. p It is the predicted score; α It is the balance coefficient; γ It is a loss scaling factor to reduce the loss contribution of negative samples and preserve the learning information of positive samples; Am This represents the area of ​​the smallest bounding rectangle of the two boxes. U This represents the area of ​​the union of the two boxes; S3. Use camera calibration and pose estimation algorithms to obtain the coordinate transformation relationship between the 2D image pixel coordinate system and the 3D world coordinate system; S4. Using the target detection results and coordinate transformation relationship, calculate the minimum distance from the top of the target to the electric field line. The target detection result is the position information output by the optimized YOLOX single-stage detector, i.e. the detection box. S5. Determine the safety level and provide timely warnings for target objects.

2. The method for online safety monitoring of power transmission channels based on artificial intelligence according to claim 1, characterized in that: In step S1, 2D images are acquired by monitoring cameras mounted on the towers; laser point cloud data are collected using a drone equipped with LiDAR; the collected 2D images are filtered, labeled and expanded, and image data and labels containing foreign objects in the power transmission channel are randomly extracted from the benchmark dataset COCO-datasets. Combined with the self-built dataset in the power scenario, a dedicated database of foreign object images in the power transmission channel is established.

3. The method for online safety monitoring of power transmission channels based on artificial intelligence according to claim 1, characterized in that: The improved YOLOX model outputs category information, target location information, and target confidence information; therefore, for the model's training task, its total loss is... L YOLOX It consists of three parts: classification loss L cls Regression loss L reg and target confidence loss L obj The calculation formula is as follows: in: In the formula, BCE_loss and IoU_loss Let represent the binary cross-entropy loss and IoU loss, respectively; p' is the predicted probability for each class. t Indicates whether it is a correct category or target; its value is 0 or 1. IoU This is the ratio of the area of ​​the intersection region to the area of ​​the union region of the two boxes; I ( gt, pred U( represents the area of ​​intersection between the label box and the prediction box.) gt, pred ) represents the area of ​​the union of the label box and the prediction box.

4. The method for online safety monitoring of power transmission channels based on artificial intelligence according to claim 1, characterized in that: Step S3 includes the following steps: S301 uses camera calibration technology to obtain the intrinsic parameters and distortion coefficients of a monocular camera. To ensure higher calibration accuracy, experiments are conducted using calibration plates of different sizes and materials. During the calibration process, the calibration plate is fixed while the camera is moved to calibrate multiple images from different positions and angles, so that the combined position of the calibration plate under all calibration images covers the entire camera field of view. S302, using a pose estimation algorithm, jointly matches feature points from a 2D image and a 3D point cloud to calculate their coordinate transformation relationship; the coordinate transformation relationship of the pinhole camera model is as follows: ； in,( X c , Y c , Z c () represents the camera coordinate system coordinates. Z c It is also a depth value; u , v () are pixel coordinates; X w , Y w , Z w () are world coordinates; dx, dy Each pixel in the image coordinate system x, y Physical dimensions of the direction; f The focal length is the camera's focal length, in mm. f x， f y The camera is at x Focal length on the y-axis; u 0 , v 0) represents the pixel coordinates corresponding to the camera's optical center; f x , f y , u 0 , v 0s constitute the camera's intrinsic parameter matrix. M ; R It is a 3×3 rotation matrix. T It is a 3×1 translation vector. R and T The extrinsic parameters that form the transition from world coordinates to polar coordinates E Coordinate transformation requires consideration of the camera's radial and tangential distortions, i.e., the distortion coefficients. D Including five key parameters ( k1, k2, k3, p1, p2 ); S303, based on the above mapping relationship from world coordinates to pixel coordinates, select more than 4 matching points in the 2D image and 3D point cloud for parameter solving of the back projection transformation.

5. The method for online safety monitoring of power transmission channels based on artificial intelligence according to claim 1, characterized in that: In step S4, using the target detection results and coordinate transformation relationships, the minimum three-dimensional distance from the target detection box to the electric field line is calculated. The formula for calculating the three-dimensional Euclidean distance is: 。 6. The method for online safety monitoring of power transmission channels based on artificial intelligence according to claim 1, characterized in that: In step S5, the algorithm compares the predicted three-dimensional distance with the relevant regulations of the power grid company's enterprise standards. Based on the safety distance of overhead lines at different voltage levels, it intelligently determines the safety level and issues timely warnings for targets that are less than or close to the safety distance.

Images