A method for predicting ice thickness of transmission lines based on improved Unet network

Abstract

This invention discloses a method for predicting ice thickness on power transmission lines based on an improved Unet network. The method constructs a dataset of transmission line ice images, performs preprocessing, and labels them. A modified Unet network is constructed and trained using the preprocessed images. The trained Unet network performs semantic segmentation on the images before and after ice coverage, calculates the pixel area of ​​the conductors, and uses an empirical formula to obtain a preliminary prediction of ice thickness. The prediction errors of each group are averaged to obtain an adjustment coefficient, and the final ice thickness is calculated. This method can accurately predict the ice thickness of power transmission lines.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transmission lines, and in particular relates to a method for predicting ice thickness of power transmission lines based on an improved Unet network. Background Art

[0002] Transmission lines are widely distributed, and most are installed in harsh environments such as outdoors and on plateaus. Every spring and winter, overhead transmission lines are prone to ice accumulation, which increases the weight of the lines and can cause conductors to fall off, resulting in significant economic losses. Because different ice thicknesses pose varying risks to conductors, an effective method for predicting conductor ice thickness is urgently needed. Currently, the main method uses empirical formulas to predict ice thickness based on the difference in pixels before and after ice coating. However, images captured by video equipment often have complex backgrounds, and the proportion of conductors in the image is too low, resulting in large errors in the calculated pixel values.

[0003] Thanks to the rapid development of convolutional neural networks, the semantic segmentation network within them can perfectly segment objects in images and completely avoid the influence of the image background. The calculated pixel values ​​are highly reliable, making them suitable for use as the detection method of the present invention. Furthermore, to address the large prediction errors associated with empirical formulas, the present invention uses the average of the prediction errors to obtain the adjustment coefficient α, significantly reducing the prediction error and providing a highly efficient method for predicting ice thickness on transmission lines. Summary of the Invention

[0004] In view of the shortcomings of existing transmission line ice thickness detection methods, the present invention aims to provide a transmission line ice thickness prediction method based on an improved Unet network, which is used to accurately predict the ice thickness of transmission lines and provide a technical reference for inspection personnel to prevent line disconnection accidents caused by icing.

[0005] To achieve this object, the present invention is implemented through the following scheme: a method for predicting ice thickness of transmission lines based on an improved Unet network, comprising the following steps:

[0006] S1. Construct a data set of ice-covered transmission line images and perform preprocessing and labeling.

[0007] S2. Build an improved Unet network and use the preprocessed images for training. The improved Unet network includes a backbone feature extraction network VGG-S, a feature fusion network, and a semantic segmentation module. The backbone feature extraction network VGG-S is constructed using a Focus structure, three Conv2D+Maxpool structures, a Conv2D+SPP structure, and three Conv2D structures. The feature fusion network is constructed using four upsampling operations, four Conv2D operations, and four Concatenate operations. The semantic segmentation module uses convolution to adjust the number of channels and outputs the last feature layer for semantic segmentation.

[0008] S3. Use the trained Unet network to perform semantic segmentation on the images before and after icing, calculate the pixel area of ​​the conductor, and then use the empirical formula to obtain a preliminary prediction of the ice thickness. For the same shooting point, the images before and after icing are grouped together, perform semantic segmentation on each group of images, obtain the segmented images, and calculate the area of ​​the conductor pixels before and after icing. The pixel area of ​​the conductor before icing is S q , the pixel area of ​​the wire after ice covering is S h , the empirical formula is used to obtain the preliminary predicted ice thickness of each group of images. The empirical formula is shown as follows:

[0009]

[0010] λ i is the preliminary predicted ice thickness of the i-th group of images, and d is the diameter of the wire;

[0011] S4. Obtain preliminary predicted ice thickness for multiple sets of images of the same conductor, average the prediction errors of each set to obtain an adjustment coefficient, and calculate the final ice thickness.

[0012] Furthermore, the preprocessing in step S1 is to take a part of the acquired transmission line ice coverage image data set and perform blurring, noise addition, direction adjustment, sharpness change, color decay processing, brightness dimming, brightness enhancement and contrast change on a part of the image, leaving a part of the image without any processing.

[0013] Furthermore, in step S1, labeling is performed using labelme software.

[0014] Furthermore, the feature extraction process of the backbone feature extraction network VGG-S is as follows: the input image passes through a Focus structure to adjust the size of the image feature layer; then passes through three Conv2D+Maxpool structures in sequence to obtain the first effective feature layer F1, the second effective feature layer F2 and the third effective feature layer F3 respectively; the third effective feature layer F3 is further downsampled through a Conv2D+SPP structure to obtain the fourth effective feature layer F4; finally, the fourth effective feature layer F4 undergoes three Conv2D structures, and convolution is used to ensure that the number of channels does not change, to obtain the fifth effective feature layer F5.

[0015] Furthermore, the feature fusion network performs feature fusion in the following process: upsampling the fifth effective feature layer F5 and concatenating it with the fourth effective feature layer F4 to obtain the first enhanced feature layer P1; passing the first enhanced feature layer P1 through a Conv2D structure and then upsampling it and concatenating it with the third effective feature layer F3 to obtain the second enhanced feature layer P2; similarly, passing the second enhanced feature layer P2 through a Conv2D structure and then upsampling it and concatenating it with the second effective feature layer F2 to obtain the third enhanced feature layer P3; in addition, passing the third enhanced feature layer P3 through a Conv2D structure and then upsampling it and concatenating it with the first effective feature layer F1 to obtain the fourth enhanced feature layer P4; passing the fourth enhanced feature layer P4 through a Conv2D structure to obtain the fifth enhanced feature layer P5.

[0016] Furthermore, the semantic segmentation module adjusts the number of channels of the fifth enhanced feature layer P5 to 2 using a 1×1 convolution, and outputs a feature layer of size 256×256×2 for semantic segmentation.

[0017] Furthermore, in step S3, for the same shooting point, the images before and after ice covering are grouped together, and semantic segmentation is performed on each group of images to obtain the segmented images and calculate the area of ​​the wire pixels before and after ice covering; the pixel area of ​​the wire before ice covering is S q , the pixel area of ​​the wire after ice covering is S h , the empirical formula is used to obtain the preliminary predicted ice thickness of 20 sets of images. The empirical formula is shown as follows:

[0018]

[0019] λ i is the preliminary predicted ice thickness of the i-th group of images, where i = 1 to 20 and d is the diameter of the wire.

[0020] Furthermore, before semantic segmentation, in order to calculate the pixel area of ​​the wire more accurately, the color of the target wire is set to green so that the pixel area of ​​the wire is obvious, and the color of the image background is black so that the pixel area of ​​the background is not calculated.

[0021] Furthermore, the adjustment coefficient is calculated in the historical data modeling stage. The calculation of the adjustment coefficient α is shown in the following formula:

[0022]

[0023] λ i is the initial predicted ice thickness of the i-th group of images, R i is the actual ice thickness of the i-th group of images, and N is the number of groups of images before and after ice coverage.

[0024] Furthermore, in the actual prediction stage, the final predicted ice thickness is λ i +α.

[0025] The present invention also provides a transmission line ice thickness prediction device based on an improved Unet network, comprising a data acquisition device and a data processing device; the data acquisition device comprises a camera for capturing images before and after ice coverage and a device equipped with a camera, the device equipped with a camera including but not limited to a drone; the data processing device comprises a preprocessing module, labelme software, an image processing module based on the Unet network, a preliminary calculation module and a correction calculation module.

[0026] The present invention also provides a non-volatile computer storage medium storing computer executable instructions, which can execute the method for predicting ice thickness of transmission lines based on the improved Unet network in any of the above embodiments.

[0027] The present invention also provides a computer program product, which includes a computer program stored on a non-volatile computer storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the method for predicting ice thickness of transmission lines based on the improved Unet network of the above embodiment.

[0028] Compared with the existing technology, the beneficial effects of the present invention include: the present invention simulates the shooting conditions of ice-covered conductors from all angles, and the trained model is more generalized and can accurately segment ice-covered conductors; at the same time, while ensuring that the segmentation accuracy is not reduced, it also provides a method for lightweighting the model; in addition, in response to the problem that traditional ice thickness prediction may have large errors, it proposes the use of adjustment coefficients to reduce the prediction error; the present invention can predict the ice thickness of the target area more quickly and accurately, and can provide a technical reference for power inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Flowchart of the method of the present invention.

[0030] Figure 2 This is the structural diagram of the improved algorithm.

[0031] Figure 3 This is the effect diagram of semantic segmentation. DETAILED DESCRIPTION

[0032] The present invention implements the flow chart as follows Figure 1 The embodiment will be described in detail below. The method for predicting ice thickness of transmission lines based on the improved Unet network includes the following steps:

[0033] S1. Construct a data set of ice-covered transmission line images and perform preprocessing and labeling.

[0034] The transmission line ice coverage image dataset includes images before and after ice coverage at each shooting point (shooting points were taken at multiple locations to ensure the comprehensiveness of the data). The images before and after ice coverage are in one-to-one correspondence. 10% of the images are taken to simulate the image blur caused by image compression during data transmission, 5% of the images are added with salt and pepper noise, 30% of the images are offset left and right, and the direction of the flip is adjusted, and 5% of the images are changed in sharpness with the sharpness value set to 1.2, 5% are subjected to color decay processing with the color value set to 0.8, 5% of the images are darkened to 0.87 of the original image, 5% of the images are enhanced to 1.07 of the original image, and 5% of the images have their contrast set between 0.8 and 1.2 of the original image. The remaining 30% of the images are not processed in any way. The purpose of classifying and preprocessing images is to enable the power transmission line ice image dataset to simulate various weather conditions and situations that may result in poor image quality during camera shooting and image transmission, so that the trained model can be more generalized. The proportion of image classification in the power transmission line ice image dataset is divided according to the proportion of blurred images, low light conditions, noise, etc. in the actual shooting conditions.

[0035] The preprocessed images were labeled using LabelMe software. Ice-covered and un-ice-covered conductors were labeled as the same class, and the rest of the background was treated as the image background. Labeled files were in .json format and converted to .png format before training the UNet network. LabelMe software treated ice-covered and un-ice-covered conductors as the same class of segmentation targets. When calculating ice thickness, ice-covered conductors were given a predicted ice thickness, while un-ice-covered conductors were given an approximate ice thickness of zero.

[0036] S2. Build an improved Unet network and use the preprocessed images for training;

[0037] The structure of the improved Unet network is as follows Figure 2 As shown in the figure, first, a Focus structure, three Conv2D+Maxpool structures, a Conv2D+SPP structure, and three Conv2D structures are used to construct the backbone feature extraction network VGG-S. Secondly, four upsampling, four Conv2D operations, and four Concatenate operations are used to construct a feature fusion network. Finally, a 1×1 convolution (Conv2D) is used to adjust the number of channels to 2, and the last feature layer is output for semantic segmentation. The specific operations are as follows:

[0038] S2.1. Construct the basic unit of the improved Unet network: use two Conv2D operations and one maximum pooling Maxpool to construct the Conv2D+Maxpool structure; use three Conv2D operations and one spatial pyramid pooling SPP structure to construct the Conv2D+SPP structure; in addition, use Conv2D, normalization, activation function ReLU and 4 pooling kernels of sizes 1×1, 5×5, 9×9 and 13×13 for maximum pooling to form the SPP structure.

[0039] S2.2. Construct the backbone feature extraction network of the improved Unet network and perform feature extraction: set the input image (Input) size to 512×512×3, and after a Focus structure, the image feature layer size becomes 256×256×12; then pass through three Conv2D+Maxpool structures in sequence to obtain three first effective feature layers F1, second effective feature layers F2 and third effective feature layers F3 with sizes of 128×128×64, 64×64×128 and 32×32×256 respectively; the third effective feature layer F3 is further downsampled through a Conv2D+SPP structure to obtain a fourth effective feature layer F4 with a size of 16×16×512; finally, the fourth effective feature layer F4 undergoes three Conv2D structures, and uses 1×1 convolution to ensure that the number of channels does not change, and a fifth effective feature layer F5 with a size of 16×16×512 is obtained;

[0040] S2.3. Construct a feature fusion network of the improved Unet network: upsample the fifth effective feature layer F5 and concatenate it with the fourth effective feature layer F4 to obtain a first enhanced feature layer P1 of size 32×32×1024; pass the first enhanced feature layer P1 through a Conv2D structure and then upsample it and concatenate it with the third effective feature layer F3 to obtain a second enhanced feature layer P2 of size 64×64×768; similarly, pass the second enhanced feature layer P2 through a Conv2D structure and then Continue upsampling and concatenate the second effective feature layer F2 to obtain a third enhanced feature layer P3 of size 128×128×384; in addition, the third enhanced feature layer P3 is passed through a Conv2D structure and then further upsampled and concatenated with the first effective feature layer F1 to obtain a fourth enhanced feature layer P4 of size 256×256×192; the fourth enhanced feature layer P4 is resized to 256×256×64 through a Conv2D structure to obtain a fifth enhanced feature layer P5;

[0041] S2.4. Build the semantic segmentation module of the improved Unet network: Finally, the fifth enhanced feature layer P5 uses a 1×1 convolution to adjust the number of channels to 2, and outputs a feature layer of size 256×256×2 for semantic segmentation;

[0042] S2.5. Set the parameters for training the improved Unet network: Before training, the improved Unet network uses a random parameter initialization strategy. Since the backbone of the improved Unet network has changed, the pre-trained weights are unavailable, and the features of the wires in the ice-covered image are very obvious, the improved Unet network can automatically adjust the parameters during training. During training, the batch size Bacthsize = 8, and a total of 500 rounds are trained. The Unet network uses the early stopping method to terminate training after a sufficient number of training rounds. In addition, the frozen training and multi-threading modes are turned off, the initial learning rate is set to 0.0001, and the learning rate of the improved Unet network is then discretely adjusted during training. The Adam optimizer is also enabled to assist in training.

[0043] S3. Use the trained improved Unet network to perform semantic segmentation on the images before and after ice coverage, calculate the pixel area of ​​the wire, and use the empirical formula to obtain a preliminary prediction of ice coverage thickness;

[0044] For the same shooting point, the images before and after ice covering are grouped together, and a total of 20 groups of images before and after ice covering are obtained. After the training, multiple groups of training weights can be obtained, from which the weight with the lowest loss value is taken and input into the improved Unet network to perform semantic segmentation on the 20 groups of images. The effect is as follows: Figure 3 As shown, 20 sets of segmented images can be obtained and the areas of the wire pixels before and after ice coating can be calculated. In addition, before semantic segmentation, in order to calculate the wire pixel area more accurately, the color of the target wire is set to green so that the pixel area of ​​the wire is obvious, and the color of the image background is black so that the pixel area of ​​the background is not calculated.

[0045] The pixel area of ​​the ice-covered front line is S q , the pixel area of ​​the wire after ice covering is S h , the empirical formula is used to obtain the preliminary predicted ice thickness of 20 sets of images. The empirical formula is shown as follows:

[0046]

[0047] λ i is the preliminary predicted ice thickness of the i-th group of images, where i = 1 to 20 and d is the diameter of the wire.

[0048] S4. Obtain preliminary predicted ice thickness for multiple sets of images of the same conductor, average the prediction errors of each set to obtain an adjustment coefficient, and calculate the final ice thickness;

[0049] In the historical data modeling stage, knowing the preliminary predicted ice thickness of each group and the actual ice thickness of the conductor obtained by actual measurement, the specific calculation of the adjustment coefficient α of the ice thickness in this section is as follows:

[0050]

[0051] λ i is the initial predicted ice thickness of the i-th group of images, R i is the actual ice thickness of the i-th group of images.

[0052] During the forecasting phase, the final predicted ice thickness is λ i +α; In addition, the images before and after icing must be of the same area, and the adjustment coefficient α calculated for different areas should be different. Experimental results show that the method of the present invention can accurately predict the thickness of ice covering the transmission line, and the prediction error is smaller than that of traditional methods.

[0053] In this embodiment, based on the possible impacts that may occur during actual operation: when pre-processing the target image, the image cannot be scaled, blocked, or other operations, otherwise it will affect the subsequent calculation of the pixel area, thereby affecting the calculation of the ice thickness; when calculating the adjustment coefficient α, the ice-covered image should be an image of the same ice-covered area, and the number of ice groups i should be increased as much as possible, and the calculated adjustment coefficient α will be more reliable; generally speaking, the ice characteristics of different regions are also different, and the adjustment coefficient α here can only be used to adjust the predicted ice thickness in that area.

[0054] The Unet network is selected in this embodiment because under the same parameter settings, the original Unet network and the DeeplabV3+ network are used to train the models respectively. The experimental results show that the Unet network is more effective in segmenting ice-covered images.

[0055] This embodiment improves the Unet network to make it more lightweight without sacrificing accuracy, and the number of Unet network parameters is reduced by two-thirds. Since the characteristics of the wire are very obvious, it is possible to accurately segment the wire and the image background while reducing the number of convolutions. The Focus and SPP structures are added to improve the scale of the feature layer while reducing the number of convolutions. The number of channels is adjusted to 2 using a 1×1 convolution because the image segmentation categories in the method of the present invention are wires and background.

[0056] This embodiment uses a strategy of random parameter initialization during training. Since the network backbone changes, the pre-trained weights are unavailable, and the features of the wires in the ice-covered image are very obvious, the model can automatically adjust the model parameters during training. The total number of training rounds is 500 so that the model can terminate training using the early stopping method after a sufficient number of training rounds.

[0057] This embodiment provides a transmission line ice thickness prediction device based on an improved Unet network, including a data acquisition device and a data processing device; the data acquisition device includes a camera for capturing images before and after ice coating and a device equipped with a camera, and the device equipped with a camera includes but is not limited to a drone; the data processing device includes a preprocessing module, labelme software, an image processing module based on the Unet network, a preliminary calculation module, and a correction calculation module.

[0058] In this embodiment, the image processing module based on the Unet network is equipped with an improved Unet network. The improved Unet network includes a backbone feature extraction network VGG-S, a feature fusion network and a semantic segmentation module; the backbone feature extraction network VGG-S is constructed using a Focus structure, three Conv2D+Maxpool structures, a Conv2D+SPP structure and three Conv2D structures in sequence; the feature fusion network is constructed using four upsampling, four Conv2D operations and four Concatenate operations; the semantic segmentation module uses convolution to adjust the number of channels and outputs the last feature layer for semantic segmentation.

[0059] In this embodiment, the preprocessing module simulates image blurring caused by image compression during data transmission on 10% of the images, adds salt-and-pepper noise to 5% of the images, shifts and flips 30% of the images, alters sharpness to 1.2 on 5% of the images, applies color decay to 0.8 on 5% of the images, dims the brightness to 0.87 of the original image on 5% of the images, increases the brightness to 1.07 of the original image on 5% of the images, and sets the contrast of 0.8 to 1.2 on 5% of the images. The remaining 30% of the images undergo no processing. Image classification and preprocessing are intended to ensure that the transmission line ice image dataset can simulate various weather conditions, as well as conditions that may result in poor image quality during camera capture and image transmission, thereby enhancing the generalizability of the trained model. The image classification ratios in the transmission line ice image dataset are based on the proportion of images in the dataset that are blurred, dark, or noisy, as encountered in actual capture.

[0060] The preliminary calculation module uses an empirical formula to obtain the preliminary predicted ice thickness of 20 sets of images. The empirical formula is shown as follows:

[0061]

[0062] λ i is the preliminary predicted ice thickness of the i-th group of images, where i = 1 to 20 and d is the diameter of the wire.

[0063] The correction calculation module calculates the adjustment coefficient α and adjusts it according to λ i +α to calculate the final predicted ice thickness.

[0064] In other embodiments, a non-volatile computer storage medium is provided, which stores computer-executable instructions. The computer-executable instructions can execute the method for predicting ice thickness of transmission lines based on the improved Unet network in any of the above embodiments.

[0065] This embodiment also provides a computer program product, which includes a computer program stored on a non-volatile computer storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the transmission line ice thickness prediction method based on the improved Unet network of the above embodiment.

[0066] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0067] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0068] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0069] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0070] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0071] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for predicting ice thickness of transmission lines based on an improved Unet network, characterized by: The steps include: S1. Construct a data set of ice-covered transmission line images and perform preprocessing and labeling. S2. Build an improved Unet network and use the preprocessed images for training. The improved Unet network includes a backbone feature extraction network VGG-S, a feature fusion network, and a semantic segmentation module. The backbone feature extraction network VGG-S is constructed using a Focus structure, three Conv2D+Maxpool structures, a Conv2D+SPP structure, and three Conv2D structures. The feature fusion network is constructed using four upsampling operations, four Conv2D operations, and four Concatenate operations. The semantic segmentation module uses convolution to adjust the number of channels and outputs the last feature layer for semantic segmentation. The feature extraction process of the backbone feature extraction network VGG-S is as follows: the input image is passed through a Focus structure to adjust the size of the image feature layer; Then, three Conv2D+Maxpool structures are sequentially used to obtain the first effective feature layer F1, the second effective feature layer F2, and the third effective feature layer F3; The third effective feature layer F3 is further downsampled through a Conv2D+SPP structure to obtain the fourth effective feature layer F4; Finally, the fourth effective feature layer F4 is subjected to the Conv2D structure three times, and convolution is used to ensure that the number of channels does not change, to obtain the fifth effective feature layer F5; The feature fusion network performs feature fusion as follows: upsampling the fifth effective feature layer F5 and concatenating it with the fourth effective feature layer F4 to obtain the first enhanced feature layer P1; passing the first enhanced feature layer P1 through a Conv2D structure and then upsampling it and concatenating it with the third effective feature layer F3 to obtain the second enhanced feature layer P2; similarly, passing the second enhanced feature layer P2 through a Conv2D structure and then upsampling it and concatenating it with the second effective feature layer F2 to obtain the third enhanced feature layer P3; In addition, the third enhanced feature layer P3 is further upsampled after passing through a Conv2D structure and concatenated with the first effective feature layer F1 to obtain a fourth enhanced feature layer P4; the fourth The enhanced feature layer P4 passes through a Conv2D structure to obtain the fifth enhanced feature layer P5; S3. Use the trained improved Unet network to perform semantic segmentation on the images before and after icing, calculate the pixel area of ​​the conductor, and then use the empirical formula to obtain a preliminary prediction of the ice thickness. For the same shooting point, take the images before and after icing as a group, perform semantic segmentation on each group of images, obtain the segmented images of each group, and calculate the area of ​​the conductor pixels before and after icing. The pixel area of ​​the conductor before icing is S q , the pixel area of ​​the wire after ice covering is S h , the empirical formula is used to obtain the preliminary predicted ice thickness of each group of images. The empirical formula is shown as follows: ； λ i is the preliminary predicted ice thickness of the i-th group of images, d is the diameter of the wire; S4. Obtain preliminary predicted ice thickness for multiple sets of images of the same conductor, average the prediction errors of each set to obtain an adjustment coefficient, and calculate the final ice thickness.

2. The method for predicting ice thickness of transmission lines based on the improved Unet network according to claim 1 is characterized in that: The preprocessing in step S1 is to take a part of the acquired transmission line ice coverage image data set and perform blurring, noise addition, direction adjustment, sharpness change, color decay processing, brightness dimming, brightness enhancement and contrast change on a part of the image, leaving a part of the image without any processing.

3. The method for predicting ice thickness of transmission lines based on the improved Unet network according to claim 1 is characterized in that: The semantic segmentation module adjusts the number of channels of the fifth enhanced feature layer P5 to 2 using a 1×1 convolution, and outputs a feature layer of size 256×256×2 for semantic segmentation.

4. The method for predicting ice thickness of transmission lines based on the improved Unet network according to claim 2 is characterized in that: In step S1, 10% of the images are taken to simulate the image blur caused by image compression during data transmission, 5% of the images are added with salt and pepper noise, 30% of the images are shifted left and right, and the direction of the flip is adjusted, 5% of the images are changed in sharpness, and the sharpness value is set to 1.2, 5% are subjected to color decay processing, and the color value of the color decay processing is set to 0.8, 5% of the images are dimmed to 0.87 of the original image, 5% of the images are enhanced to 1.07 of the original image, and the contrast of 5% of the images is set between 0.8 and 1.2 of the original image. The remaining 30% of the images are not processed in any way.

5. The method for predicting ice thickness of transmission lines based on the improved Unet network according to claim 2 is characterized in that Before semantic segmentation, the color of the target wire is set to green so that the pixel area of ​​the wire is obvious and the color of the image background is black.

6. The method for predicting ice thickness of transmission lines based on the improved Unet network according to claim 1 is characterized in that: The adjustment coefficient is calculated during the historical data modeling phase. The calculation of the adjustment coefficient α is shown in the following formula: ； λ i is the initial predicted ice thickness of the i-th group of images, R i is the actual ice thickness of the i-th group of images, and N is the number of groups of images before and after ice covering; In the actual prediction stage, the final predicted ice thickness is λ i +α.

7. A device for predicting ice thickness on power transmission lines based on an improved Unet network, comprising a data acquisition device and a data processing device; the data acquisition device comprises a camera for capturing images before and after ice coating and a device equipped with the camera, the device equipped with the camera including but not limited to a drone; the device is characterized in that: The data processing device includes a preprocessing module, labelme software, an image processing module based on the Unet network, a preliminary calculation module and a correction calculation module; the data processing device is used to implement the steps of the transmission line ice thickness prediction method based on the improved Unet network as described in any one of claims 1-6.

8. A non-volatile computer storage medium storing computer-executable instructions, characterized in that: The computer-executable instructions can execute the method for predicting ice thickness of transmission lines based on the improved Unet network as described in any one of claims 1-6.

Images