Power transmission line galloping detection method based on improved YOLO11-OBB

Abstract

The power transmission line galloping detection method based on the improved YOLO11-OBB comprises the following steps: processing a power transmission line galloping monitoring video to obtain a power transmission line galloping image; preprocessing the power transmission line galloping image; the improved YOLO11-OBB deep learning network model is trained, and a weight is generated; detecting the power transmission line galloping monitoring video; and summarizing and sorting detection results, and drawing a power transmission line galloping track. According to the method, the power transmission line in the monitoring video is detected by using the rotating detection frame, so that the detection frame is closer to the outline of the power transmission line, and meanwhile, the YOLO11 network model is improved, so that the detection accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of transmission line galloping detection, and in particular to a transmission line galloping detection method based on improved YOLO11-OBB. Background Art

[0002] Transmission lines exposed to the elements for extended periods of time, often exposed to ice and strong winds, can experience line galloping, characterized by low-frequency (approximately 0.1-3 Hz) and large-amplitude (approximately 5-300 times the conductor diameter) self-oscillations. This can lead to conductor strand breakage, wire disconnection, accessory damage, phase short circuits, and tower collapses, resulting in line paralysis and widespread power outages, causing significant economic losses.

[0003] At present, the commonly used methods for detecting the dancing of transmission lines are mainly the acceleration sensor method and the video monitoring method. The acceleration sensor method measures the acceleration or displacement of multiple positions on the line by arranging multiple sensors to achieve dancing monitoring. For example, the Chinese patent "CN112179476B" discloses a sensor for measuring the dancing of transmission lines. A sensor for measuring the dancing of transmission lines is designed. The fiber-optic dancing sensor can measure the dancing of overhead transmission lines. However, due to the characteristics of the sensor itself, there are disadvantages in the detection process: the number of sensors required for monitoring is large and the cost is high, the sensor's own survival ability cannot guarantee its long-term operation, and the data collected by the sensor is easily disturbed by the environment, affecting the detection results. Therefore, it is difficult to be widely used.

[0004] The video monitoring method uses a camera to monitor the dancing of the transmission line. It is widely used in various occasions due to its non-contact measurement characteristics. It can greatly save the time and economic cost of installing and fixing equipment, and reduce the workload of researchers on regular maintenance of equipment. For example, the literature [1]: Yin Hui, Zhang Xiaoming, Li Xiaoxiang, et al. Transmission line dancing information extraction and spectrum analysis based on video monitoring [J]. High Voltage Technology, 2017, 43(09): 2889-2895. A transmission line dancing information extraction and spectrum analysis method based on video monitoring is proposed to effectively extract the transmission line dancing information. However, factors such as the installation position of the camera, the selection of dancing feature points, and the angle between the camera and the conductor have a great influence on the accuracy of the monitoring system. At the same time, the detection method has low accuracy and the detection frame cannot tightly surround the transmission line area, resulting in a lot of irrelevant information in the detection frame.

[0005] In summary, the current technology has the following shortcomings: the current transmission line galloping detection method has problems such as high cost, great environmental impact and low detection accuracy. Summary of the Invention

[0006] To solve the above technical problems, the present invention proposes a transmission line galloping detection method based on improved YOLO11-OBB. This method detects the transmission lines in the surveillance video by using a rotating detection frame, so that the detection frame is closer to the transmission line contour. At the same time, the YOLO11 network model is improved to improve the detection accuracy.

[0007] The technical solution adopted by the present invention is:

[0008] The transmission line galloping detection method based on the improved YOLO11-OBB includes the following steps:

[0009] Step 1: Process the transmission line dancing monitoring video to obtain a transmission line dancing image;

[0010] Step 2: Preprocess the transmission line dancing image;

[0011] Step 3: Train the improved YOLO11-OBB deep learning network model and generate weights;

[0012] Step 4: Detect the transmission line dancing monitoring video;

[0013] Step 5: Summarize and organize the detection results of step 4 and draw the dancing trajectory of the transmission line.

[0014] The transmission line dancing monitoring video is processed, and the transmission line dancing image is extracted at intervals of 1 frame to obtain the transmission line dancing image.

[0015] The extracted power line dancing images were preprocessed to generate a power line dancing dataset. This process involved using the contrast-limited adaptive histogram equalization method, part of an image enhancement algorithm, to preprocess the power line dancing images. The histogram h(k) is defined as the number of pixels in the image with a grayscale value of k. The adaptive histogram equalization method divides the image into blocks and performs independent equalization on each local region, enhancing local contrast while limiting the contrast enhancement to avoid excessive noise amplification.

[0016] The algorithm steps of the adaptive histogram equalization method are as follows:

[0017] First, the image is divided into multiple small blocks, and the histogram is calculated for each small block;

[0018] Then, the histogram is clipped to limit the number of pixels of a single gray level to not exceed the preset contrast limit threshold, and the clipped pixels are evenly distributed to other gray levels;

[0019] Next, histogram equalization is performed on each small block to generate a grayscale mapping table;

[0020] Finally, bilinear interpolation is used to remove discontinuities at the boundaries between blocks and generate the final enhanced image.

[0021] CLAHE effectively suppresses noise while enhancing local details through local processing and contrast limitation.

[0022] The transmission line galloping dataset was divided into training and test sets according to the appropriate ratio. The sets were then fed into the improved YOLO11-OBB deep learning network model for training. The weights with the highest accuracy were finally selected and saved for subsequent transmission line galloping detection tasks.

[0023] The improved YOLO11-OBB deep learning network model replaces the first two C3k2 modules in the YOLO11-OBB model backbone network with the C3_BiFormer module, thereby enhancing the feature extraction effect of the improved YOLO11-OBB network model on small targets and improving the detection accuracy of the improved YOLO11-OBB network model for small targets. The original fast spatial pyramid pooling structure is replaced with the SimSPPF module, so that the improved YOLO11-OBB network model can achieve better detection speed and improve performance. The original loss function CIoU is improved to PIoU loss function, which strengthens the effect of detecting small targets and improves the accuracy and stability of network regression.

[0024] The C3 module is integrated with the BiFormer attention mechanism to replace the two C3k2 modules with the C3k parameter set to False in the shallow layer of the original network to enhance the network's attention to small target features.

[0025] The SimSPPF module replaces the SPPF module in the original network. The SimSPPF module uses the ReLU activation function. The SimSPPF module first compresses the input feature map through a convolution module. It then performs pooling through three identical MaxPool2d modules. These pooling modules are then stacked in series to obtain feature maps with different receptive fields. The resulting four-dimensional feature maps are then fused. Finally, the fused feature map is convolved again to increase its dimensionality, resulting in the final feature extraction result.

[0026] During network training, the improved YOLO11-OBB deep learning network model uses the PIoU loss function. The PIoU loss function guides the target box to regress along a valid path, thereby converging faster than other loss functions. The PIoU loss function proposes an adaptive target size penalty factor P, which is only associated with the size of the target box and is defined as follows:

[0027]

[0028] Where P is the penalty factor for adaptive target size, dw1, dw2, dh1 and dh2 are the absolute values of the distances between the corresponding edges of the prediction box and the target box, and w gt and h gt Indicates the width and height of the target box.

[0029] In order to further improve the loss function, the PIoU loss function proposes a penalty function f(P) that is adaptive to the quality of the predicted box. Its formula is as follows:

[0030]

[0031] Where f(P) is the penalty function and P is the penalty factor for adaptive target size.

[0032] A larger P value indicates a more significant difference between the predicted box and the target box, and the smaller the f(P) function value, suppressing harmful gradients from low-quality predicted boxes. When the P value is near 1, the predicted box is close to the target box, and a larger f(P) value accelerates the regression of the loss function. When the P value approaches 0, the f(P) function value decreases as the quality of the predicted box increases, continuously optimizing the predicted box so that it can overlap with the target box. The penalty function f(P) can maximize the gradient of medium-quality predicted boxes, allowing these predicted boxes to quickly regress to the vicinity of the target box.

[0033] The PIoU loss function is calculated as follows:

[0034] PIoU=IoU-f(P),-1≤PIoU≤1

[0035] L PIoU =1-PIoU=L IoU +f(P),0≤L PIoU ≤2

[0036] Where PIoU is the intersection over union ratio in polar coordinates, IoU is the intersection over union ratio, f(P) is the penalty function, and L PIoU is the PIoU loss function, L IoU As basic loss.

[0037] The improved YOLO11-OBB network model is repeatedly trained for multiple rounds, and the weight with the highest accuracy on the test set is finally selected and saved. During the training process, the training parameters are set to generate a weight file that can be used for subsequent detection tasks after every ten rounds of training. In order to achieve higher subsequent detection accuracy, each generated weight is tested on the test set to verify the correctness of the detection, and then the weight accuracy p is obtained:

[0038]

[0039] TP indicates that the improved YOLO11-OBB network model predicts positive examples as positive, and FP indicates that the improved YOLO11-OBB network model predicts negative examples as positive. The weight with the highest accuracy is finally selected for subsequent dance detection.

[0040] In step 4, the weight file generated in step 3 is used to detect the transmission line dancing monitoring video, and the detection results are as follows: Figure 9 As shown in the figure, the transmission line area can be effectively detected and the coordinates of the detection frame can be extracted. The coordinate extraction results are shown in the figure. Figure 10 As shown, the coordinates of the center point of the detection frame are accurately extracted;

[0041] In step 5, the detection results and coordinate extraction results obtained in step 4 are summarized and integrated to draw the dancing trajectory of the transmission line. The drawn dancing trajectory is as follows: Figure 11 As shown, the galloping of the transmission line can be accurately described.

[0042] The present invention provides a method for detecting power line galloping based on improved YOLO11-OBB, and the technical effects are as follows:

[0043] 1) The present invention adopts rotating target detection and uses a rotating detection frame to detect the transmission line area, so that the detection frame can be closer to the transmission line contour, resulting in better effect.

[0044] 2) The present invention uses an improved YOLO11-OBB deep learning network model to train a power line dancing dataset to obtain detection weights. The improved YOLO11-OBB deep learning network model replaces the first two C3k2 modules in the YOLO11-OBB network model backbone network with C3_BiFormer modules, enhancing the model's small target feature extraction effect and improving the model's small target detection accuracy; replaces the original fast spatial pyramid pooling (SPPF) structure with the SimSPPF structure, enabling the model to achieve better detection speed and improve performance; improves the original loss function CIoU to PIoU, strengthening the effect of small target detection, improving the accuracy and stability of network regression, and having good practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention will be further described below with reference to the accompanying drawings and examples:

[0046] Figure 1 Schematic diagram of the process of the present invention.

[0047] Figure 2 This is the dancing image of the power transmission line described in the present invention.

[0048] Figure 3This is a structural diagram of the improved YOLO11-OBB deep learning network model described in the present invention.

[0049] Figure 4 This is a schematic diagram of the C3k2 module described in the present invention.

[0050] Figure 5 Schematic diagram of the C3_BiFormer module described in the present invention.

[0051] Figure 6 This is a schematic diagram of the SimSPPF module described in the present invention.

[0052] Figure 7 Schematic diagram of the PIoU loss function described in the present invention.

[0053] Figure 8 This is a screenshot of a transmission line dancing monitoring video according to an embodiment of the present invention.

[0054] Figure 9 This is a transmission line area detection result according to an embodiment of the present invention.

[0055] Figure 10 This is the detection box coordinate extraction result of an embodiment of the present invention.

[0056] Figure 11 This is an image of the transmission line's dancing trajectory according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] A power line galloping detection method based on an improved YOLO11-OBB algorithm detects power line galloping and plots its trajectory. This method, based on an improved YOLO11-OBB deep learning network model, performs preprocessing on the power line galloping image, including dehazing. It then uses a deep learning object detection network to detect the power line area within the galloping image. The detected coordinates are then extracted, and finally, the power line galloping trajectory is plotted, visually demonstrating the power line galloping situation.

[0058] A transmission line galloping detection method based on improved YOLO11-OBB is characterized by comprising the following steps:

[0059] Step 1: Process the transmission line dancing monitoring video to obtain a transmission line dancing image;

[0060] Step 2: Preprocess the transmission line dancing image;

[0061] Step 3: Train the improved YOLO11-OBB deep learning network model and generate weights;

[0062] Step 4: Detect the transmission line dancing monitoring video;

[0063] Step 5: Summarize and organize the detection results of step 4 and draw the dancing trajectory of the transmission line.

[0064] In step 1, the transmission line dancing monitoring video is processed, and the transmission line dancing images are extracted at intervals of 1 frame, and a total of 1800 transmission line dancing images are obtained. Figure 2 shown.

[0065] In step 2, the extracted transmission line dancing images are preprocessed. The purpose of preprocessing is to effectively improve the model performance, accelerate the model training speed, and enhance the model generalization ability of the images after preprocessing. The main image preprocessing methods include normalization, standardization, data enhancement, denoising, and color correction. By preprocessing the transmission line dancing images, a transmission line dancing dataset is obtained;

[0066] The preprocessing of the present invention mainly uses the contrast-limited adaptive histogram equalization method (CLAHE) in the image enhancement algorithm to preprocess the transmission line dancing image. Since transmission line dancing often occurs in some extreme weather conditions, preprocessing the image and removing interference factors such as fog in the monitoring image are important for improving the performance of subsequent models. Histogram equalization is a global contrast enhancement algorithm that redistributes the grayscale values of image pixels to make its histogram as evenly distributed as possible, thereby expanding the dynamic range of the image and improving the visual effect. It is particularly suitable for processing low-contrast images (such as those shot in foggy weather or scenes with insufficient light).

[0067] The histogram of an image represents the frequency of each grayscale level in the image. For a grayscale image with a grayscale range of [0, L-1], such as L=256 in an 8-bit image, the histogram h(k) is defined as the number of pixels in the image with a grayscale value of k. By adjusting the pixel values, the histogram is made as "flat" as possible, that is, the number of pixels at each grayscale level is close to equal, thereby maximizing the information entropy and contrast of the image. However, if the original image is noisy, equalization will enhance the noise, so contrast-limited adaptive histogram equalization is used. The adaptive histogram equalization method (CLAHE) divides the image into blocks and equalizes each local area separately, enhancing the local contrast and limiting the contrast enhancement amplitude of the local area to avoid excessive noise amplification.

[0068] The algorithm steps of the adaptive histogram equalization method (CLAHE) are as follows:

[0069] First, the image is divided into multiple small blocks (such as 8×8 or 16×16 pixels), and the histogram is calculated for each small block. Then, the histogram is cropped to limit the number of pixels in a single gray level to not exceed the preset contrast limit threshold, and the cropped pixels are evenly distributed to other gray levels.

[0070] Next, histogram equalization is performed on each small block to generate a grayscale mapping table;

[0071] Finally, bilinear interpolation is used to remove discontinuities at the boundaries between blocks and generate the final enhanced image.

[0072] CLAHE effectively suppresses noise while enhancing local details through local processing and contrast limitation.

[0073] In step 3, the power line galloping dataset was divided into a training set and a test set in a ratio of 8:2. The dataset was then fed into the improved YOLO11-OBB deep learning network model for training. The weights with the highest accuracy were finally selected and saved for subsequent power line galloping detection tasks.

[0074] The transmission line dancing dataset contains 1440 training images and 360 test images, both with a resolution of 300×300 pixels. The network training parameters are shown in Table 1 below:

[0075] Table 1 Network training parameters

[0076]

[0077] The improved YOLO11-OBB deep learning network model structure is as follows Figure 3 As shown. By replacing the first two C3k2 modules in the YOLO11-OBB model backbone network with the C3_BiFormer module, the feature extraction effect of the improved YOLO11-OBB network model for small targets is enhanced, and the detection accuracy of the improved YOLO11-OBB network model for small targets is improved; the original fast spatial pyramid pooling (SPPF) structure is replaced with the SimSPPF module, so that the improved YOLO11-OBB network model can achieve better detection speed and improve performance; the original loss function CIoU is improved to PIoU loss function, which strengthens the effect of detecting small targets and improves the accuracy and stability of network regression. The details are as follows:

[0078] The C3k2 module is an important feature extraction component in the Yolo11 model. It is an improved design based on the traditional C3 module. It provides more powerful feature extraction capabilities by combining variable convolution kernels and channel separation strategies. It is especially suitable for more complex scenes and deep feature extraction tasks. The two structures of the C3k2 module (C3k = True and C3k = False) are as follows: Figure 4 As shown in Figure 1, compared to the standard C3 module, the C3k2 module introduces a multi-scale convolution kernel C3k, where K is an adjustable convolution kernel size, such as 3x3, 5x5, etc. This design can expand the receptive field, enabling the model to capture a wider range of contextual information, and is particularly suitable for detecting large objects or scenes with complex backgrounds.

[0079] In the YOLO11-OBB deep learning network model, the backbone network uses convolution and C3k2 modules for downsampling, and in the shallow network, the C3k parameter is set to False. At this time, the C3k2 module is the same as the C2f module in the Yolov8 network. This will lead to a certain loss of contextual features, resulting in a reduction in feature map resolution and loss of detection target space information. In order to reduce the loss of small target information during downsampling, the present invention enhances the network's attention to small target features by fusing the C3 module with the BiFormer attention mechanism to replace the two C3k2 modules with the C3k parameter set to False in the shallow layer of the original network. The C3_BiFoemer module structure is as follows: Figure 5 shown.

[0080] The simplified fast spatial pyramid pooling (Simplified SPPF, SimSPPF) module is used to replace the SPPF module in the original network. The difference between SimSPPF and SPPF lies in the activation function used. Among them, SPPF uses SiLU activation function, while SimSPPF uses ReLU activation function. Compared with SPPF, using SimSPPF can achieve better detection speed and improve network performance. The structure of the SimSPPF module is as follows Figure 6 As shown in the figure, the SimSPPF module first performs a convolution operation to compress the input feature map. It then performs pooling through three identical MaxPool2d modules, and obtains feature maps with different receptive field information from these pooling modules in a serial stacking manner. The four-dimensional feature maps obtained by pooling are then fused. Finally, the fused feature map is convolved again to increase its dimension, obtaining the final feature extraction result.

[0081] During network training, the improved YOLO11-OBB deep learning network model uses the PIoU loss function. PIoU addresses the problem that the existing IoU-based loss function is easily affected by unreasonable penalty factors, which causes the target box to expand during the regression process and significantly slows down the convergence speed. It combines a target size adaptive penalty factor and a gradient adjustment function based on the target box quality. The PIoU loss function guides the target box to regress along a valid path, thereby converging faster than other loss functions. In order to solve the problem of target box expansion and the limitations of the IoU-based loss function, the PIoU loss function proposes an adaptive target size penalty factor P, which is only associated with the size of the target box and has nothing to do with the size of the predicted box and the minimum external box of the target box. Even if the predicted box expands, it will not affect the size of P, thereby improving the adaptability to the target size. It is defined as follows:

[0082]

[0083] Where P is the penalty factor for adaptive target size, dw1, dw2, dh1 and dh2 are the absolute values of the distances between the corresponding edges of the prediction box and the target box, and w gt and h gt Indicates the width and height of the target box, such as Figure 7 shown.

[0084] In order to further improve the loss function, the PIoU loss function proposes a penalty function f(P) that is adaptive to the quality of the predicted box. Its formula is as follows:

[0085]

[0086] Where f(P) is the penalty function and P is the penalty factor for adaptive target size.

[0087] A larger P value indicates a more significant difference between the predicted box and the target box, resulting in a smaller f(P) function value, which suppresses harmful gradients from low-quality predictions. When the P value is near 1, the predicted box is close to the target box, and a larger f(P) value accelerates the regression of the loss function. When the P value approaches 0, the f(P) function value decreases as the quality of the predicted box increases, continuously optimizing the predicted box so that it coincides with the target box. The penalty function f(P) maximizes the gradients of medium-quality predictions, allowing them to quickly regress toward the target box, transforming these medium-quality predictions into high-quality ones. This allows the object detection network to focus more on anchor boxes of medium quality.

[0088] The PIoU loss function is calculated as follows:

[0089] PIoU=IoU-f(P),-1≤PIoU≤1

[0090] L PIoU =1-PIoU=L IoU +f(P),0≤L PIoU ≤2

[0091] Where PIoU is the intersection over union ratio in polar coordinates, IoU is the intersection over union ratio, f(P) is the penalty function, and L PIoU is the PIoU loss function, L IoU As basic loss.

[0092] The improved YOLO11-OBB network model is repeatedly trained for multiple rounds, and the weight with the highest accuracy on the test set is finally selected and saved. During the training process, the training parameters are set. Every ten rounds of training will generate a weight file that can be used for subsequent detection tasks. In order to achieve higher subsequent detection accuracy, each generated weight will be tested on the test set to verify the correctness of the detection, and then the weight accuracy p is obtained:

[0093]

[0094] TP indicates that the improved YOLO11-OBB network model predicts positive examples as positive, and FP indicates that the improved YOLO11-OBB network model predicts negative examples as positive. The weight with the highest accuracy is finally selected for subsequent dance detection.

[0095] In step 4, the weight file generated in step 3 is used to detect the transmission line dancing monitoring video, and the detection results are as follows: Figure 9 As shown in the figure, the transmission line area can be effectively detected and the coordinates of the detection frame can be extracted. The coordinate extraction results are shown in the figure. Figure 10 As shown, the coordinates of the center point of the detection frame are accurately extracted;

[0096] In step 5, the detection results and coordinate extraction results obtained in step 4 are summarized and integrated to draw the dancing trajectory of the transmission line. The drawn dancing trajectory is as follows: Figure 11 As shown, the galloping of the transmission line can be accurately described.

[0097] Verification example:

[0098] In order to prove the accuracy of the power line dancing detection method of the present invention in detecting the power line area and the intuitiveness of the dancing trajectory, the power line dancing monitoring video is detected and the dancing trajectory is drawn.

[0099] The screenshot of the monitoring video of the power transmission line dancing detected in this embodiment is as follows: Figure 8 The improved YOLO11-OBB deep learning network model detects the transmission line galloping area. Figure 9 As shown by Figure 9 It can be seen that the improved YOLO11-OBB deep learning network model of the present invention can effectively detect the transmission line area and make the detection frame more closely fit the direction of the transmission line. Figure 10 This is the transmission line area coordinate extraction result of this embodiment. Figure 11 This is the dancing trajectory of the power transmission line drawn in this embodiment. Figure 10 、 Figure 11 It can be seen that the method of the present invention can extract the coordinates of the detection frame and then draw the dancing trajectory of the transmission line, which can intuitively represent the dancing situation of the transmission line.

[0100] The power line gallop detection method disclosed herein can effectively detect the transmission line area and fit the line's direction and width with good accuracy. Furthermore, by extracting detection frame coordinates and plotting galloping trajectories, the galloping condition can be intuitively expressed. This method can effectively and intuitively detect power line galloping, reducing the probability of galloping accidents.

[0101] This method incorporates rotating target detection and an improved YOLO11-OBB deep learning network model into the transmission line gallop detection process, ensuring that the detection frame more closely matches the actual transmission line location compared to other detection algorithms. Furthermore, network improvements enhance the accuracy and reliability of the detection method. Furthermore, by extracting the detection frame coordinates and plotting the transmission line gallop trajectory, the method provides a visual representation of the transmission line gallop.

Claims

1. A transmission line galloping detection method based on improved YOLO11-OBB is characterized by The following steps are involved: Step 1: Process the transmission line dancing monitoring video to obtain a transmission line dancing image; Step 2: Preprocess the transmission line dancing image; Step 3: Train the improved YOLO11-OBB deep learning network model and generate weights; Step 4: Detect the transmission line dancing monitoring video; Step 5: Summarize and organize the detection results of step 4 and draw the dancing trajectory of the transmission line.

2. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 1 is characterized in that: In the step 1, the power line dancing monitoring video is processed, and the power line dancing image is extracted at intervals of one frame to obtain the power line dancing image.

3. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 1 is characterized in that: In the step 2, the extracted transmission line dancing image is preprocessed to obtain a transmission line dancing dataset; Specifically include: The contrast-limited adaptive histogram equalization method in the image enhancement algorithm is used to preprocess the transmission line dancing image. The histogram h(k) is defined as the number of pixels with grayscale value k in the image. The adaptive histogram equalization method divides the image into blocks and equalizes each local area separately, thereby enhancing the local contrast and limiting the contrast enhancement amplitude of the local area to avoid excessive noise amplification.

4. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 3 is characterized by: The algorithm steps of the adaptive histogram equalization method are as follows: First, the image is divided into multiple small blocks, and the histogram is calculated for each small block; Then, the histogram is clipped to limit the number of pixels of a single gray level to not exceed the preset contrast limit threshold, and the clipped pixels are evenly distributed to other gray levels; Next, histogram equalization is performed on each small block to generate a grayscale mapping table; Finally, bilinear interpolation is used to eliminate discontinuities at the boundaries between blocks and generate the final enhanced image; CLAHE effectively suppresses noise while enhancing local details through local processing and contrast limitation.

5. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 1 is characterized in that: In step 3, the power line galloping dataset is divided into a training set and a test set according to a certain ratio, and is input into the improved YOLO11-OBB deep learning network model for training. Finally, the weight with the highest accuracy is selected and saved for performing the subsequent power line galloping detection task.

6. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 5, characterized in that: The improved YOLO11-OBB deep learning network model replaces the first two C3k2 modules in the YOLO11-OBB model backbone network with the C3_BiFormer module, thereby enhancing the feature extraction effect of the improved YOLO11-OBB network model on small targets and improving the detection accuracy of the improved YOLO11-OBB network model for small targets. The original fast spatial pyramid pooling structure is replaced with the SimSPPF module, so that the improved YOLO11-OBB network model can achieve better detection speed and improve performance. The original loss function CIoU is improved to PIoU loss function, which strengthens the effect of detecting small targets and improves the accuracy and stability of network regression.

7. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 6, characterized in that: The C3 module is integrated with the BiFormer attention mechanism to replace the two C3k2 modules with the C3k parameter set to False in the shallow layer of the original network to enhance the network's attention to small target features. The SimSPPF module is used to replace the SPPF module in the original network. The SimSPPF module uses the ReLU activation function. The SimSPPF module first performs a convolution module operation to compress the input feature map. Then it is pooled through three identical MaxPool2d modules, and feature maps of different receptive field information are obtained from these pooling modules in a serial superposition manner. The four-dimensional feature maps obtained by pooling are then fused. Finally, the fused feature map is convolved again to increase its dimension to obtain the final feature extraction result.

8. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 7, characterized in that: In network training, the improved YOLO11-OBB deep learning network model uses the PIoU loss function; the PIoU loss function guides the target box to regress along the effective path, thereby converging faster than other loss functions; the PIoU loss function proposes an adaptive target size penalty factor P, which is only associated with the size of the target box and is defined as follows: Where P is the penalty factor for adaptive target size, dw1, dw2, dh1 and dh2 are the absolute values of the distances between the corresponding edges of the prediction box and the target box, and w gt and h gt Indicates the width and height of the target box.

9. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 8, characterized in that: In order to further improve the loss function, the PIoU loss function proposes a penalty function f(P) that is adaptive to the quality of the predicted box. Its formula is as follows: Where f(P) is the penalty function, and P is the penalty factor of the adaptive target size; When the P value is larger, the difference between the predicted box and the target box is more significant, and the function f(P) value is smaller, which suppresses the harmful gradient from the low-quality predicted box. When the P value is close to 1, it indicates that the predicted box is close to the target box. The larger f(P) value will accelerate the regression of the loss function. When the P value approaches 0, the function f(P) value decreases as the quality of the predicted box increases, and the predicted box is continuously optimized so that it can overlap with the target box. The penalty function f(P) can make the medium-quality predicted box have the largest gradient, allowing these predicted boxes to quickly regress to the vicinity of the target box. The PIoU loss function is calculated as follows: PIoU=IoU-f(P),-1≤PIoU≤1 L PIoU =1-PIoU=L IoU +f(P),0≤L PIoU ≤2 Where PIoU is the intersection over union ratio in polar coordinates, IoU is the intersection over union ratio, f(P) is the penalty function, and L PIoU is the PIoU loss function, L IoU As basic loss.

10. The method for detecting power line galloping based on the improved YOLO11-OBB according to claim 9, characterized in that: Repeat the training of the improved YOLO11-OBB network model for multiple rounds, and finally save the weight with the highest accuracy on the test set; During the training process, the training parameter settings will generate a weight file that can be used for subsequent detection tasks every ten rounds of training. In order to make the subsequent detection accuracy higher, each generated weight will be tested in the test set to verify the correctness of the detection, and then the weight accuracy p will be obtained: Among them, TP means that the improved YOLO11-OBB network model predicts positive examples as positive examples, and FP means that the improved YOLO11-OBB network model predicts negative examples as positive examples. Finally, the weight with the highest accuracy is selected for subsequent dance detection.

Images