Transmission conductor icing monitoring method based on AI image segmentation
By using an AI-based image segmentation method, the YOLOv8-SEG icing model, and deep learning technology, the problem of rapid and accurate identification and measurement of conductor icing was solved, enabling efficient monitoring of conductor icing conditions and ensuring the safe and stable operation of the power system.
Patent Information
- Application Number
- CN202411058923.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-15
AI Technical Summary
Existing technologies make it difficult to quickly and accurately identify and measure icing conditions on power transmission lines, which affects the safety and stability of power transmission lines.
An AI-based image segmentation method is adopted, which uses the yolov8-seg icing model to detect and segment ice on the conductor. The model is optimized by combining the cross-entropy loss function and the Dice loss function. The ice thickness is calculated through high-resolution image capture and deep learning, and an alarm is issued or the de-icing operation is initiated when a set threshold is set.
It enables rapid and accurate identification and measurement of conductor icing conditions, improving monitoring efficiency and accuracy, and ensuring the safe operation and stability of the power grid.
Smart Images

Figure BDA0004977210480000021 
Figure BDA0004977210480000022 
Figure BDA0004977210480000041
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of icing monitoring of power transmission lines, and in particular to a method for monitoring icing of power transmission lines based on AI image segmentation. Background Art
[0002] Wire icing is a phenomenon in which water droplets freeze when they come into contact with the surface of an object under cooling conditions. It is common in winter or under certain climatic conditions. Wire icing is mainly caused by the following reasons; 1. Temperature conditions: When the temperature is below freezing, especially around 0°C, water droplets in the air will condense into ice when they come into contact with the surface of a cold object. 2. Humidity conditions: A high humidity environment promotes the condensation of water vapor, which forms an ice layer when it comes into contact with the cooled surface of the wire. 3. Fog and clouds: In foggy or cloudy conditions, tiny water droplets come into contact with the surface of the wire and freeze quickly, forming ice. 4. Rainfall and snowfall: At temperatures at or below freezing, rain or snow will also quickly freeze after coming into contact with the surface of the wire.
[0003] The negative impacts of icing on power transmission lines are as follows: First, it increases the weight of the line. Icing will significantly increase the weight of the transmission line, which may cause overload of the conductors, poles and suspension devices, resulting in conductor breakage or tower collapse. In extreme cases, severe icing may cause transmission line breakage or power tower collapse. Second, it affects the transmission efficiency: icing may also increase the resistance of the transmission line, thereby affecting the transmission efficiency and quality. Third, the increase in icing may lead to flashover accidents: icing may also cause flashover accidents between conductors or between conductors and the ground, increasing the risk of operation. A good icing monitoring system can prevent and reduce accidents. By accurately monitoring the icing of the transmission line, preventive measures can be taken in time, such as heating, deicing or adjusting the operation mode, thereby reducing accidents caused by icing. Secondly, it can ensure the stable operation of the power grid: accurate icing monitoring can help operation and maintenance personnel to understand the real-time status of the transmission line in a timely manner and take corresponding measures to ensure the stable operation of the power grid. Therefore, the development and implementation of effective icing monitoring technology is of vital significance to ensure the safe, reliable and efficient operation of power transmission lines. Summary of the invention
[0004] The technical problem to be solved by the present invention is to overcome the deficiencies of the prior art and provide a transmission line icing monitoring method based on AI image segmentation for rapid and accurate identification and measurement of icing conditions of the conductors.
[0005] The technical solution adopted by the present invention is that the method comprises the following steps:
[0006] (1) On the transmission line to be monitored, a conductor tower is selected to install a camera and a simulated conductor. The camera is aimed at the simulated conductor and captures clear images of the conductor at a set frequency and transmits them back to the backend server;
[0007] (2) The backend server uses the yolov8-seg icing model to perform AI analysis on the received images to obtain the wire detection border map, wire segmentation mask map, and the value of the detection result;
[0008] (3) Perform pixel calculation on the returned detection frame value and mask value to obtain the pixel height of the wire after ice coating, and calculate the ice coating thickness of the wire according to the pixel height after ice coating and the pixel height before ice coating;
[0009] (4) The conductor ice thickness and the conductor ice image after AI analysis are transmitted back to the backend server, and the backend server displays the real conductor ice image after AI analysis;
[0010] (5) Repeat steps (2) to (4), process the wire image received by the backend server in step (1), and continuously monitor the wire icing condition. When the icing value is greater than or equal to the set deicing threshold, an alarm is issued to de-ice the wire.
[0011] Furthermore, before performing step (2), the yolov8-seg icing model is trained, specifically including the following steps:
[0012] s1. Obtain a large number of on-site pictures of ice-free and ice-covered conductors on the background server, and then perform data preprocessing on these pictures to remove image noise and enhance the generalization ability of the yolov8-seg ice model;
[0013] s2. Combine the cross entropy loss function and the Dice loss function to optimize the model's recognition accuracy of the ice-covered area boundary;
[0014] s3. Use multi-scale training methods to enable the model to learn the characteristics of wire ice coverage at different scales, and improve the model's recognition ability in ice-covered areas of various sizes;
[0015] s4. Use model distillation technology to guide the training of the YoloV8-Seg icing model through a pre-trained larger model to reduce the model's inference time and improve the model's accuracy.
[0016] In step (3), a plurality of pictures of the conductor are taken in advance and AI analysis is performed to obtain the pixel height l of the conductor when it is not covered with ice, and the conductor ice thickness is calculated by the following formula:
[0017]
[0018] Where L is the pixel height of the wire after ice covering, and d is the wire diameter.
[0019] In the step (5), if the ice coverage value is greater than or equal to the deicing threshold set for the conductor, a pop-up window alarm is displayed on the web page on the background server, and the background inspection personnel activate the deicing mechanism to start deicing the conductor; if the ice coverage value is less than the deicing threshold set for the conductor, no pop-up window alarm is displayed.
[0020] More specifically,
[0021] In step s1, data preprocessing includes performing noise reduction processing on the image and performing contrast enhancement processing on the image:
[0022] When denoising an image, smoothing, blurring and filtering techniques are used to remove unwanted noise from the image, and histogram equalization is used to adjust the image contrast;
[0023] When contrast enhancement is performed on images, the images are randomly rotated, flipped, scaled, and cropped to increase the diversity of training data and improve the model's generalization ability to different ice cover conditions.
[0024] Finally, Labelme software is used to label these images and train the model.
[0025] The specific steps of step s2 are as follows:
[0026] The cross entropy loss function is defined as:
[0027]
[0028] The Dice loss can be defined as:
[0029] Where N is the batch size, y i is the true label, y p is the predicted probability;
[0030] Combining the cross entropy loss function and the Dice loss function, the combined loss function can be expressed as:
[0031] L=αL ce +βL dice
[0032] Among them, α and β are balance weight coefficients, which are obtained by continuous debugging during model optimization;
[0033] The combined loss function is used to optimize the model's recognition accuracy of the ice-covered area boundary.
[0034] Furthermore, the pseudo code for calculating the height of the ice-covered pixel point on the wire is as follows:
[0035] Initialize totalHeight to 0
[0036] Initialize count to 0
[0037] For each column x, from 0 to the width of segMaskResize:
[0038] Initialize columnHeight to 0
[0039] For each row of y, from 0 to the height of segMaskResize:
[0040] If segMaskResize's color at position (y,x) is equal to white (255,255,255):
[0041] ColumnHeight increases by 1 (the height count of this column increases by one)
[0042] If columnHeight is greater than 0:
[0043] totalHeight increases by the value of columnHeight
[0044] Count increases by 1 (the number of columns containing wires increases by one)
[0045] If count is greater than 0:
[0046] averageHeight is set to totalHeight divided by count
[0047] otherwise:
[0048] averageHeight is set to 0
[0049] Output averageHeight
[0050] Among them: segMaskResize is the mask value information after the wire is segmented after ice covering, and averageHeight is the calculated height of the ice-covered pixels.
[0051] In addition, the camera captures pictures once per hour.
[0052] The beneficial effects of the present invention are as follows: the present invention combines high-resolution image capture with a deep learning model, and realizes pixel-level segmentation of simulated wires through AI analysis, thereby calculating the ice thickness of real wires; the method of the present invention realizes rapid and accurate identification and measurement of the ice coverage of real transmission wires; experimental results show that compared with traditional monitoring methods, the method of the present invention can effectively improve monitoring efficiency and accuracy, and provide a new technical means for ice damage prevention and control in power systems; in addition, the present invention combines advanced deep learning algorithms and image processing technology, which greatly improves the accuracy and real-time performance of ice detection. Through continuous technology iteration and data accumulation, the AI model will become more and more intelligent, providing strong support for the safe operation and stability of the power grid. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 This is the original image of the conductor ice cover;
[0054] Figure 2 It is the mask image of the wire ice after using the segmentation algorithm. DETAILED DESCRIPTION
[0055] like Figure 1 and Figure 2 As shown, in the present invention, the method comprises the following steps:
[0056] (1) On the transmission line to be monitored, select a conductor tower to install a camera and a simulated conductor. Here, the test results of the simulated conductor are used to reflect the icing condition of the actual transmission line. There is no need to operate the transmission line on site, which ensures safety and reliability. At the same time, the simulated conductor in the same environment is used to reflect the actual transmission line, which can truly reflect the icing condition of the actual transmission line. The camera is aimed at the simulated conductor, and at a set frequency, the camera captures clear pictures of the conductor and transmits them back to the background server. In this embodiment, three cameras are set on the conductor tower at the same time. In addition to one camera aimed at the simulated conductor, there is another camera aimed at the insulator and one camera aimed at the tower to comprehensively monitor the icing condition of the transmission line. In addition, the frequency is set to once an hour by default, and the user can also customize the upload time.
[0057] (2) The backend server uses the yolov8-seg icing model to perform AI analysis on the received images to obtain the wire detection border map, wire segmentation mask map, and the value of the detection result;
[0058] (3) Perform pixel calculation on the returned detection frame value and mask value to obtain the pixel height of the wire after ice coating, and calculate the ice coating thickness of the wire according to the pixel height after ice coating and the pixel height before ice coating;
[0059] (4) The conductor ice thickness and the conductor ice image after AI analysis are transmitted back to the backend server, and the backend server displays the real conductor ice image after AI analysis;
[0060] (5) Repeat steps (2) to (4), process the wire image received by the backend server in step (1), and continuously monitor the wire icing condition. When the icing value is greater than or equal to the set deicing threshold, an alarm is issued to de-ice the wire.
[0061] Specifically, in step (3), a plurality of pictures of the conductor are taken in advance and AI analysis is performed to obtain the pixel height l of the conductor when it is not covered with ice, and the conductor ice thickness is calculated by the following formula:
[0062]
[0063] Where L is the pixel height of the wire after ice covering, and d is the wire diameter.
[0064] In the step (5), if the ice coverage value is greater than or equal to the deicing threshold set for the conductor, a pop-up window alarm is displayed on the web page on the background server, and the background inspection personnel activate the deicing mechanism to start deicing the conductor; if the ice coverage value is less than the deicing threshold set for the conductor, no pop-up window alarm is displayed.
[0065] In step s1, data preprocessing includes performing noise reduction processing on the image and performing contrast enhancement processing on the image:
[0066] When denoising an image, smoothing, blurring and filtering techniques are used to remove unwanted noise from the image, and histogram equalization is used to adjust the image contrast;
[0067] When contrast enhancement is performed on images, the images are randomly rotated, flipped, scaled, and cropped to increase the diversity of training data and improve the model's generalization ability to different ice cover conditions.
[0068] Finally, Labelme software is used to label these images and train the model.
[0069] Through actual model verification, the model using data augmentation technology has better generalization ability than the model without data augmentation technology, and can better segment the areas where the conductor is ice-free and ice-covered.
[0070] Before performing step (2), the yolov8-seg icing model is trained, which specifically includes the following steps:
[0071] s1. Obtain a large number of on-site pictures of ice-free and ice-covered conductors on the background server, and then perform data preprocessing on these pictures to remove image noise and enhance the generalization ability of the yolov8-seg ice model;
[0072] s2. Combine the cross entropy loss function and the Dice loss function to optimize the model's recognition accuracy of the ice-covered area boundary;
[0073] s3. A multi-scale training method is used to enable the model to learn the characteristics of wire ice coverage at different scales, and improve the model's recognition ability in ice-covered areas of various sizes; the core idea of multi-scale training is to randomly change the size of the training image so that the model can see objects of different sizes during the training process. The purpose of doing this is two-fold: First, to enhance generalization ability: by training inputs of different scales, the model not only learns the detailed features of the object, but also learns to recognize the structural features of the object. This helps the model to accurately recognize new and unseen size changes. Second, improve robustness: in practical applications, the size of the object may change due to factors such as distance and viewing angle. Multi-scale training makes the model more robust to size changes and can maintain high recognition accuracy under different conditions.
[0074] s4. Use model distillation technology to guide the training of the YoloV8-Seg icing model through a pre-trained larger model to reduce the model's inference time and improve the model's accuracy. Model distillation is a model compression technology that aims to transfer the knowledge of a large, complex model (usually called a "teacher model") to a smaller, simpler model (called a "student model"). The key advantage of this technology is that it enables the student model to approach or achieve the performance of the teacher model while maintaining a small size and low computational cost. Model distillation is particularly valuable in mobile devices and edge computing applications. Our edge devices use model distillation technology to significantly reduce model inference time while achieving relatively good accuracy.
[0075] The specific steps of step s2 are as follows:
[0076] The cross entropy loss function is defined as:
[0077]
[0078] The Dice loss can be defined as:
[0079] Where N is the batch size, y i is the true label, y p is the predicted probability;
[0080] Combining the cross entropy loss function and the Dice loss function, the combined loss function can be expressed as:
[0081] L=αL ce +βL dice
[0082] Among them, α and β are balance weight coefficients, which are continuously debugged during model optimization. Here, the debugging process of α and β is as follows: the initial values of α and β are preset to 0.5. Through the continuous learning process of the deep learning algorithm, the gradient descent algorithm will automatically adjust to the most appropriate value, which can make the recognition effect of the model achieve the best.
[0083] The combined loss function is used to optimize the model's recognition accuracy of the ice-covered area boundary.
[0084] Here, Cross-Entropy Loss is a loss function used to measure the difference between the probability distribution of the model output and the true label distribution. It is very effective for classification problems, especially when there are two or more categories.
[0085] Dice loss is based on the Dice coefficient, a statistical tool for measuring the similarity of two samples, which is particularly suitable for imbalanced datasets. Dice loss encourages the model to increase the accuracy of positive class predictions, which is especially useful for rare events such as ice cover.
[0086] Combining cross entropy loss and Dice loss can optimize the performance of the model in terms of precision and recall at the same time, which is an advantage for imbalanced datasets.
[0087] The pseudo code for calculating the height of ice-covered pixel points on the wire is as follows:
[0088] Initialize totalHeight to 0
[0089] Initialize count to 0
[0090] For each column x, from 0 to the width of segMaskResize:
[0091] Initialize columnHeight to 0
[0092] For each row of y, from 0 to the height of segMaskResize:
[0093] If segMaskResize's color at position (y,x) is equal to white (255,255,255):
[0094] ColumnHeight increases by 1 (the height count of this column increases by one)
[0095] If columnHeight is greater than 0:
[0096] totalHeight increases by the value of columnHeight
[0097] Count increases by 1 (the number of columns containing wires increases by one)
[0098] If count is greater than 0:
[0099] averageHeight is set to totalHeight divided by count
[0100] otherwise:
[0101] averageHeight is set to 0
[0102] Output averageHeight
[0103] Among them: segMaskResize is the mask value information after the wire is segmented after ice covering, and averageHeight is the calculated height of the ice-covered pixels.
[0104] The method of the present invention combines high-resolution image capture with a deep learning model, and uses AI analysis to segment simulated wires at the pixel level, thereby calculating the ice thickness. The method of the present invention achieves rapid and accurate identification and measurement of the ice coverage of real transmission lines. Experimental results show that compared with traditional monitoring methods, the method of the present invention can effectively improve monitoring efficiency and accuracy, and provides a new technical means for ice damage prevention and control in power systems.
[0105] Finally, it should be emphasized that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and modifications. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for monitoring ice coating on power transmission lines based on AI image segmentation, characterized in that: The method comprises the following steps: (1) On the transmission line to be monitored, a conductor tower is selected to install a camera and a simulated conductor. The camera is aimed at the simulated conductor and captures clear images of the conductor at a set frequency and transmits them back to the backend server; (2) The backend server uses the yolov8-seg icing model to perform AI analysis on the received images to obtain the wire detection border map, wire segmentation mask map, and the value of the detection result; (3) Calculate the pixel point of the returned detection frame value and mask value to obtain the pixel point height of the wire after ice coating, and calculate the ice coating thickness of the wire according to the pixel point height after ice coating and the pixel point height before ice coating; (4) The conductor ice thickness and the conductor ice image after AI analysis are transmitted back to the backend server, and the backend server displays the real conductor ice image after AI analysis; (5) Repeat steps (2) to (4), process the wire image received by the backend server in step (1), and continuously monitor the wire icing condition. When the icing value is greater than or equal to the set deicing threshold, an alarm is issued to de-ice the wire.
2. According to the method for monitoring ice coating on power transmission lines based on AI image segmentation in claim 1, it is characterized in that: Before performing step (2), the yolov8-seg icing model is trained, which specifically includes the following steps: s1. Obtain a large number of on-site pictures of ice-free and ice-covered conductors on the background server, and then perform data preprocessing on these pictures to remove image noise and enhance the generalization ability of the yolov8-seg ice model; s2. Combine the cross entropy loss function and the Dice loss function to optimize the model's recognition accuracy of the ice-covered area boundary; s3. Use multi-scale training methods to enable the model to learn the characteristics of wire ice coverage at different scales, and improve the model's recognition ability in ice-covered areas of various sizes; s4. Use model distillation technology to guide the training of the YoloV8-Seg icing model through a pre-trained larger model to reduce the model's inference time and improve the model's accuracy.
3. According to the method of claim 1, the method is characterized in that: In step (3), a plurality of pictures of the conductor are taken in advance and AI analysis is performed to obtain the pixel height l of the conductor when it is not covered with ice, and the conductor ice thickness is calculated by the following formula: Where L is the pixel height of the wire after ice covering, and d is the wire diameter.
4. According to the method for monitoring ice coating on power transmission lines based on AI image segmentation in claim 1, it is characterized in that: In the step (5), if the ice coverage value is greater than or equal to the deicing threshold set for the conductor, a pop-up window alarm is displayed on the web page on the background server, and the background inspection personnel activate the deicing mechanism to start deicing the conductor; if the ice coverage value is less than the deicing threshold set for the conductor, no pop-up window alarm is displayed.
5. The method for monitoring ice coating on power transmission lines based on AI image segmentation according to claim 2 is characterized in that: In step s1, data preprocessing includes performing noise reduction processing on the image and performing contrast enhancement processing on the image: When denoising an image, smoothing, blurring and filtering techniques are used to remove unwanted noise from the image, and histogram equalization is used to adjust the image contrast; When contrast enhancement is performed on images, the images are randomly rotated, flipped, scaled, and cropped to increase the diversity of training data and improve the model's generalization ability to different ice cover conditions. Finally, Labelme software is used to label these images and train the model.
6. The method for monitoring ice coating on power transmission lines based on AI image segmentation according to claim 2 is characterized in that: The specific steps of step s2 are as follows: The cross entropy loss function is defined as: The Dice loss can be defined as: Where N is the batch size, y i is the true label, y p is the predicted probability; Combining the cross entropy loss function and the Dice loss function, the combined loss function can be expressed as: L=αL ce +βL dice Among them, α and β are balance weight coefficients, which are obtained by continuous debugging during model optimization; The combined loss function is used to optimize the model's recognition accuracy of the ice-covered area boundary.
7. The method for monitoring ice coating on power transmission lines based on AI image segmentation according to claim 3 is characterized in that: The pseudo code for calculating the height of ice-covered pixel points on the wire is as follows: Initialize totalHeight to 0 Initialize count to 0 For each column x, from 0 to the width of segMaskResize: Initialize columnHeight to 0 For each row of y, from 0 to the height of segMaskResize: If segMaskResize's color at position (y,x) is equal to white (255,255,255): ColumnHeight increases by 1 (the height count of this column increases by one) If columnHeight is greater than 0: totalHeight increases by the value of columnHeight Count increases by 1 (the number of columns containing wires increases by one) If count is greater than 0: averageHeight is set to totalHeight divided by count otherwise: averageHeight is set to 0 Output averageHeight Among them: segMaskResize is the mask value information after the wire is segmented after ice covering, and averageHeight is the calculated height of the ice-covered pixels.
8. The method for monitoring ice coating on power transmission lines based on AI image segmentation according to claim 1, characterized in that: The camera captures pictures once per hour.
Citation Information
Patent Citations
Medical image segmentation method based on double segmentation models
CN114998362A
Power transmission line icing monitoring method based on infrared image processing
CN115861221A
3D medical image segmentation system and method
CN117522891A
Method and system for reducing misrecognition rate of power transmission channel forest fire smoke detection model
CN117953372A
Colposcope image segmentation model construction method, image classification method and device
CN118334336A
Cited By
Icing thickness monitoring method based on AI image
CN119915230A
An AI image-based ice thickness monitoring method
CN119915230B
Power grid line ice melting effect detection method based on deep learning
CN120747732A
Calculation method, system and device for actively sensing ice layer thickness in OPGW icing monitoring and medium
CN120760615A
An end-to-end edge side power line icing closed-loop monitoring method and system
CN122887773A
