Method for extracting motion trail of insulated conductor of overhead line of distribution network
Through the deep learning semantic segmentation algorithm, the problem of monitoring the movement trajectory of conductors in low-voltage distribution network lines was solved, the conductor's own characteristics were quickly and accurately extracted, and the reliability of line operation was improved.
Patent Information
- Application Number
- CN202510653249.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-21
- Publication Date
- 2025-09-19
AI Technical Summary
Existing technologies make it difficult to effectively monitor the movement trajectory of overhead insulated conductors in low-voltage distribution network lines, especially without the installation of feature points, resulting in inaccurate dancing monitoring.
Using a deep learning semantic segmentation algorithm, through image preprocessing, data set annotation and enhancement, SegNet network training and image modification, the semantic segmentation and motion trajectory extraction of three-phase conductors are achieved, and the structural characteristics of the conductors themselves are used for monitoring.
It achieves rapid and accurate extraction of the movement trajectory of insulated conductors in distribution network overhead lines, meets the needs of real-time online monitoring, and improves the reliability of line operation.
Smart Images
Figure CN120673410A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of power distribution analysis, and in particular relates to a method for extracting motion trajectories of overhead insulated conductors based on a deep learning semantic segmentation algorithm. Background Art
[0002] Distribution network overhead lines are installed over a wide area, with complex and diverse wiring methods. They are easily affected by severe weather such as rain, snow, and ice, which can cause overhead line dancing, leading to overhead line operation failures or accidents, affecting people's normal lives and even causing serious economic losses.
[0003] Currently, there are two main methods for online monitoring of overhead line galloping: image processing-based monitoring and sensor-based monitoring. Research focuses on high-voltage transmission lines of 220 kV and above. Image processing-based monitoring methods primarily use prominent structures on the monitoring line as feature points (such as spacers), or manually add feature points. Distribution lines, especially low-voltage distribution lines, are less likely to have spacers installed, and the small diameter of the conductors makes it difficult to add feature points. Summary of the Invention
[0004] The technical problem to be solved by this invention is to provide a method for extracting the motion trajectory of insulated conductors in distribution network overhead lines. This method uses the conductor's inherent structural characteristics, without the need for additional feature points, and processes conductor images using a deep learning semantic segmentation algorithm to extract the motion trajectory of insulated conductors in distribution networks. This method provides technical support for online monitoring of insulated conductor galloping in distribution networks, thereby improving the operational reliability of distribution network lines.
[0005] A method for extracting the motion trajectory of an insulated conductor of a distribution network overhead line comprises the following steps, which are performed in sequence:
[0006] Step 1: Split the video transmitted back by the front monitoring device frame by frame to obtain images, and uniformly adjust the image pixel values;
[0007] Step 2: Perform semantic segmentation on the three-phase conductor images obtained in Step 1 to create three datasets, one for each phase of the conductor. Use Image Labeler to label the images. Perform image enhancement using the Retinex algorithm on the labeled images to avoid trajectory extraction bias caused by low illumination. Combine the original labeled images with the Retinex-enhanced images to form the final dataset for training the network.
[0008] Step 3: Build the SegNet network. The encoder uses convolutional layers and pooling layers to gradually extract the features of the input image and obtain the potential representation.
[0009] Step 4: The three datasets obtained in step 2 are used as inputs to the SegNet network under three conditions, thereby obtaining semantic segmentation images of the left phase, middle phase, and right phase conductors respectively;
[0010] Step 5: Binarize the semantic segmentation image obtained in step 4 and retain the connected area with the largest area in the binary image as the final extracted wire image;
[0011] Step 6: Select the centroid of the wire in the final extracted wire image obtained in step 5 as the monitoring point, and obtain the displacement of the wire through its position change.
[0012] The pooling layer of the SegNet network described in step 3 records the pooling position and performs transposed convolution during depooling.
[0013] The specific numerical value of the displacement of the wire described in step 4 is obtained by estimating the relationship between the actual distance and the pixel value through the relative position of the wire in the image; the position of the wire centroid in the first frame of the image is used as the reference point to calculate the displacement change of the three-phase wires respectively.
[0014] Through the above-mentioned design scheme, the present invention can bring the following beneficial effects: a method for extracting the motion trajectory of the insulated conductor of the distribution network overhead line can provide technical support for the development of the distribution network overhead line dancing monitoring system, and extract the conductor motion trajectory from the structure of the conductor itself without adding any load on the line; at the same time, the motion trajectory of each phase conductor of the three-phase overhead insulated conductor can be extracted simultaneously and separately, and the extraction speed is fast, which fully meets the requirements of real-time online monitoring of the conductor. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:
[0016] Figure 1 The present invention provides a flow chart of conductor motion trajectory extraction for a method of extracting the motion trajectory of insulated conductors in a distribution network overhead line.
[0017] Figure 2 This is a comparison diagram of the marking results and segmentation results of a specific implementation method for extracting the motion trajectory of insulated conductors in distribution network overhead lines of the present invention.
[0018] Figure 3 This is a comparison diagram of the final results and marking results of a specific implementation method of the invention for extracting the motion trajectory of insulated conductors in distribution network overhead lines.
[0019] Figure 4 The invention discloses a method for extracting the motion trajectory of insulated conductors of overhead distribution network lines and a specific implementation method thereof. The invention discloses the extraction result of the motion trajectory of three-phase conductors. DETAILED DESCRIPTION
[0020] In order to more clearly demonstrate the purpose, technical solutions and innovations of the invention, the present invention will be further explained below with reference to the accompanying drawings and embodiments.
[0021] In a specific embodiment of the present invention, the specific steps are:
[0022] Step 1: Image preprocessing
[0023] (1) Split the video transmitted back by the front monitoring device and decompose the video frame by frame into pictures; the video comes from the front monitoring device. The present invention takes the monitoring device installed in the 10kV overhead insulated line group 21 line in Nangang Industrial Zone, Tianjin as an example.
[0024] (2) In order to effectively improve the training speed of the network and ensure the final segmentation accuracy, the pixel value size of the disassembled image is adjusted from 3840×2160 to 768×432.
[0025] Step 2: Create a dataset
[0026] To perform image semantic segmentation, we must first create a dataset. In order to segment the three-phase conductors separately, we need to create three datasets, one dataset corresponding to the semantic segmentation of one phase conductor.
[0027] First, 100 images were labeled using Image Labeler. Then, to increase the amount of training data, the labeled images were enhanced using the Retinex algorithm to avoid potential deviations in trajectory extraction caused by low-light conditions. Finally, the original labeled images and the Retinex-enhanced images were combined to form the final dataset for training the network. The Retinex algorithm works as follows:
[0028] Assume that the original image is represented as P(x, y):
[0029] P(x,y)=I(x,y)×R(x,y)
[0030] Where x and y represent the horizontal and vertical pixel coordinates respectively, I(x, y) is the incident component, and R(x, y) is the reflected component.
[0031] Gaussian filtering of the image can obtain the incident component I(x, y):
[0032] I(x,y)=P(x,y)*G(x,y)
[0033] Where G(x, y) is the Gaussian convolution kernel, and the symbol “*” indicates convolution calculation.
[0034] Based on the above two equations, we can perform logarithmic transformation, convert the multiplication operation into addition operation, and simplify the subsequent calculations to obtain the following expression:
[0035] Lg(R(x,y))=Lg(P(x,y))-Lg(I(x,y))
[0036] =Lg(P(x,y))-Lg(P(x,y)*G(x,y))
[0037] Finally, improve the image contrast by linear stretching:
[0038] P(x,y)=(PV-PV min ) / (PV max -PV min )
[0039] Among them, PV is the initial pixel value, PV max is the maximum iterative pixel result, PV min is the minimum iterative pixel result. The linear stretching is the image enhancement result P(x, y), that is, the reflection component R(x, y).
[0040] Step 3: Train the SegNet network
[0041] The SegNet encoder utilizes convolutional and pooling layers to gradually extract features from the input image and obtain a latent representation. This involves a unique de-pooling process. SegNet records the pooling positions and directly performs transposed convolution during de-pooling, rather than concatenating the layers at the resolution level corresponding to the encoder. This design makes SegNet more efficient at decoding and recovering information. Each SegNet training cycle requires 240 iterations. After a certain number of iterations (less than 50), the SegNet network achieves an accuracy exceeding 99% in all three scenarios.
[0042] Step 4: Split the wires
[0043] The resized images are used as inputs to the SegNet network under three conditions, and the semantic segmentation results of the left phase, middle phase and right phase conductors are obtained respectively. The segmentation results are compared with the labeling results. Figure 2 shown.
[0044] Step 5: Image Retouching
[0045] First, the image obtained by direct segmentation is binarized; then, only the connected area with the largest area in the binary image is retained, which is the final extracted wire. Figure 3 shown.
[0046] Step 6: Motion trajectory extraction
[0047] The centroid of the conductor in the final segmentation image is selected as the monitoring point, and the displacement of the conductor is determined by its position change. To determine the specific value of the conductor displacement, it is necessary to estimate the relationship between the actual distance and pixel value based on the relative position of the conductors in the image. Due to the installation location of the front-end monitoring device, the proportion of the three-phase conductor in the image varies. Therefore, the relationship between the actual distance and pixel value is established for each of the three phase conductors. First, the correspondence between image size and pixel value is determined. The image size is 197.30×110.98 and the pixel size is 768×432. Therefore, the correspondence between image size and pixel value is 1:3.9. Then, the correspondence between the actual size of the conductor and the pixel value is calculated. Under actual conditions, the diameter of the conductor is 25 mm. The calculated correspondence between the pixel value and the actual size of the left phase conductor is 1:0.0067 m; the correspondence between the pixel value and the actual size of the middle phase conductor is 1:0.0078 m; and the correspondence between the pixel value and the actual size of the right phase conductor is 1:0.0063 m.
[0048] Taking the position of the conductor centroid in the first frame image as the reference point, the displacement changes of the three-phase conductors are calculated respectively. The results of the three-phase conductor motion trajectory extraction for the first 100 frames of images are as follows: Figure 4 shown.
Claims
1. A method for extracting the motion trajectory of insulated conductors in distribution network overhead lines, characterized by: The process includes the following steps, which are performed in sequence: Step 1: Split the video transmitted back by the front monitoring device frame by frame to obtain images, and uniformly adjust the image pixel values; Step 2: Perform semantic segmentation on the three-phase conductor images obtained in Step 1 to create three datasets, one for each phase of the conductor. Use Image Labeler to label the images. Perform image enhancement using the Retinex algorithm on the labeled images to avoid trajectory extraction bias caused by low illumination. Combine the original labeled images with the Retinex-enhanced images to form the final dataset for training the network. Step 3: Build the SegNet network. The encoder uses convolutional layers and pooling layers to gradually extract the features of the input image and obtain the potential representation. Step 4: The three datasets obtained in step 2 are used as inputs to the SegNet network under three conditions, thereby obtaining semantic segmentation images of the left phase, middle phase, and right phase conductors respectively; Step 5: Binarize the semantic segmentation image obtained in step 4 and retain the connected area with the largest area in the binary image as the final extracted wire image; Step 6: Select the centroid of the wire in the final extracted wire image obtained in step 5 as the monitoring point, and obtain the displacement of the wire through its position change.
2. The method for extracting the motion trajectory of insulated conductors of overhead power distribution lines according to claim 1, wherein: The pooling layer of the SegNet network described in step 3 records the pooling position and performs transposed convolution during depooling.
3. The method for extracting the motion trajectory of insulated conductors of overhead power distribution lines according to claim 1, wherein: The specific numerical value of the displacement of the wire described in step 4 is obtained by estimating the relationship between the actual distance and the pixel value through the relative position of the wire in the image; the position of the wire centroid in the first frame of the image is used as the reference point to calculate the displacement change of the three-phase wires respectively.