A differential prediction method for icing thickness of transmission lines

Abstract

The present invention discloses a method for differentially predicting the icing thickness of a transmission line, which includes the following steps: constructing an icing data set of the transmission line and performing preprocessing; constructing an improved PSPnet semantic segmentation network, inputting the made icing data set into the semantic segmentation network for model training, and using the trained model to perform semantic segmentation on the iced image; when performing semantic segmentation, setting the color of the wire to red, the icing part of the wire to green, the iced insulator to blue, and the rest is regarded as the picture background and set to black; performing a three-classification on the segmented image according to the color of the segmented picture; differentially predicting the icing thickness for different icing states, and setting an error weight to adjust the calculation formula in real time; this method can quickly and accurately predict the icing thickness of the transmission line, and can provide a technical reference for predicting the icing thickness.

Description

Technical Field

[0001] The present invention belongs to the technical field of transmission lines, and particularly relates to a method for differentially predicting the icing thickness of transmission lines. Background Art

[0002] Transmission lines are widely distributed, and most of them are distributed in areas prone to icing such as plateaus and lakes. If icing occurs on the transmission lines, it will cause an increase in the weight of the lines, which is extremely likely to cause the collapse of the lines and result in huge economic losses. Spring and winter are the peak periods for conductor icing every year. Facing a large number of icing conductors, it is impossible to detect them one by one manually. Therefore, there is an urgent need for a method to predict the icing thickness of conductors. Currently, when calculating the icing thickness of conductors, all icing conductors are usually grouped into one category and the average icing thickness is calculated, and it is impossible to classify the icing types and grades, resulting in additional prediction errors. Summary of the Invention

[0003] Aiming at the deficiencies in the existing detection of the icing thickness of transmission lines, the present invention aims to provide a method for differentially predicting the icing thickness of transmission lines, which is used to accurately predict the icing thickness of conductors and insulators of transmission lines. First, a set of image preprocessing schemes are provided to remove the noise in the icing images and the unevenness of the image edges, laying a foundation for calculating the diameter of the conductors and the pixel area of the insulators in the images in the later stage. Secondly, an improved scheme for the image segmentation algorithm is provided, which not only improves the accuracy of image segmentation but also can be integrated into a system with poor hardware performance. Then, the use of error weights is proposed to balance the possible prediction errors. Finally, a differential prediction scheme for the icing thickness is provided, which can not only accurately classify the icing state but also predict the icing thickness of the conductors and insulators. This method can provide a technical reference for inspection personnel to monitor the icing state of transmission lines.

[0004] To achieve the purpose of this invention, the present invention is realized through the following solutions: A method for differentially predicting the icing thickness of transmission lines, comprising the following steps:

[0005] S1. Construct a dataset of icing images of transmission lines and perform preprocessing to remove the noise in the icing images and smooth the image edges;

[0006] S2. Use the improved PSPNet semantic segmentation network to extract the icing conductors and insulators in the images;

[0007] S3. Classify the icing categories according to the three primary color values existing in the images;

[0008] S4. Differentially predict the icing thickness of the insulators and conductors according to the icing categories, and use the ratio of the true value to the predicted value of the conductors as the error weight to adjust the calculation formula of the icing thickness in real time.

[0009] Further, in step S1, the data set is preprocessed to remove the noise in the icing image and the roughness of the image edge, laying a foundation for calculating the diameter of the conductor and the pixel area of the insulator in the later stage.

[0010] Further, the preprocessing method adopted in step S1 includes: applying first-order filtering to the icing image, and the filtering coefficient of the first-order filtering = 0.35; sharpening the filtered image using the Roberts gradient algorithm, and the threshold T of the gradient is 7; performing adaptive smoothing on the sharpened image, where the adaptive smoothing is composed of linear smoothing and non-linear smoothing. Linear smoothing includes neighborhood averaging method, selective averaging method, and Wiener filtering. Non-linear smoothing includes median filtering, Prewitt operator, and Sobel operator. One of the modules is automatically selected for image smoothing processing by calculating the difference in pixel values at the image edge, and the module with smaller pixel values at the image edge is preferentially selected.

[0011] Further, the reason for choosing the PSPnet network for semantic segmentation in step S2 is that, first, under the same environment configuration, the image segmentation effect of using the PSPnet network is better than that of the Unet, DeeplabV3+, and Mask RCNN networks; second, the PSPnet network has fewer parameters and uses a large number of poolings, which can increase the receptive field of the image; therefore, the PSPnet network is optimized and improved on this basis.

[0012] Further, in step S2, an improved PSPNet semantic segmentation network is constructed and trained using the data set; the VGG network is used as the backbone network, and the dilated convolution Conv_rate is used to replace the max pooling in the backbone network; the PSP module is improved, and average pooling is performed using a convolution kernel matching the size of the feature layer. Each feature layer is divided into three regions according to the size of the pooled feature layer; the divided regions are integrated, and the number of channels is adjusted to 4; the operations include:

[0013] S2.1. Set the size of the input image to 512×512×3. After two ordinary convolutions, the size of the feature layer becomes 512×512×64. At this time, through a dilated convolution Conv_rate with a dilation rate of 1, the size of the feature layer F1 becomes 256×256×64; after downsampling F1 and passing through two ordinary convolutions, the size of the feature layer becomes 256×256×128. At this time, through a dilated convolution Conv_rate with a dilation rate of 3, the size of the feature layer F2 becomes 128×128×128; the feature layer F2 continues to be downsampled. After three ordinary convolutions, a dilated convolution Conv_rate with a dilation rate of 6 is used to change the size of the feature layer F3 to 64×64×256; similarly, the feature layer F3 continues to be downsampled. After three ordinary convolutions, a dilated convolution Conv_rate with a dilation rate of 12 is used to change the size of the feature layer F4 to 32×32×512; the feature layer F4 continues to be downsampled and passes through three ordinary convolutions to obtain a feature layer F5 with a size of 32×32×512.

[0014] S2.2. Extract the feature layer F5 from the improved backbone network VGG. Use a 1×1 convolution to adjust the size of F5 to 32×32×1024. At this time, through average pooling with convolution kernel sizes of 8, 16, and 32 respectively, F5 is divided into three regions with sizes of 4×4, 2×2, and 1×1; stack the feature layers after average pooling to obtain the enhanced feature layer P5, and use the resize operation to change the size of the feature layer P5 to 512×512. Use a 1×1 convolution to change the number of channels to 4. At this time, the final feature layer P6 is obtained, and P6 is output for semantic segmentation.

[0015] Furthermore, the improvement of the PSPnet semantic segmentation network in step S2 is because the proportion of the wire in the image is small, but the span is often large. Therefore, in order to increase the receptive field of the feature layer and enable it to more effectively segment the icing image from the entire image, dilated convolution is introduced for improvement; the improvement of the PSP module is considered that the features of the wire and insulator are very obvious, and average pooling in three regions can effectively complete feature enhancement.

[0016] Further, in step S3 for classifying the icing categories, the operations include: when performing semantic segmentation on the icing image, setting the color of the non-iced conductor to red, with the values of the three primary colors being R = 255, G = 0, B = 0; setting the icing conductor part to green, with the values of the three primary colors being R = 0, G = 255, B = 0; setting the icing insulator to blue, with the values of the three primary colors being R = 0, G = 0, B = 255; regarding the rest as the picture background and setting it to black; determining whether the image belongs to conductor icing, insulator icing, or both conductor and insulator icing based on the values of the three primary colors in the image; and further determining whether the conductor is fully iced, semi-iced, or non-iced based on the values of the three primary colors in the image on the basis of determining the existence of conductor icing.

[0017] If only blue exists in the image or both blue and red exist simultaneously, it is classified into the insulator icing category.

[0018] If blue does not exist in the image, it is classified into the conductor icing category, and a three-classification is performed again based on the target color in the picture: if only red exists in the segmented image, it is classified into the non-iced conductor category, and the icing thickness is not calculated; if only green exists in the segmented image, it is classified into the fully iced conductor category, the conductor diameter is set, and its icing thickness is calculated; if both red and green exist in the segmented image, it is classified into the semi-iced conductor category, and the icing thickness is calculated using the diameter difference between different color regions.

[0019] If both blue and green exist in the image, it is classified into the category of both insulator and conductor icing.

[0020] Further, in step S3, setting the conductor segmentation color to red, the icing conductor segmentation color to green, the insulator segmentation color to blue, and the picture background color to black can effectively reflect the icing condition of the transmission line, and calculate the icing thickness for different icing conditions.

[0021] Further, in step S4, using the ratio of the true value to the predicted value of the conductor as the error weight to adjust the icing thickness calculation formula in real time, the purpose is to verify the rationality of the icing thickness calculation formula using the actual conductor diameter, so as to adjust the calculation formula in real time; the calculation formula is as follows

[0022]

[0023] First, use the actual conductor diameter d 1 and the average diameter of the red region in the figure D 1 as the error weight σ , if σ is less than 1, it means that the predicted value of the formula is too large, so multiplying by σ can linearly regress and reduce the predicted value; if σIf it is greater than 1, it means that the predicted value of the formula is too small, so multiply by σ to linearly regress and increase the predicted value.

[0024] Furthermore, in step S4, the ice-covered thickness of the insulator and the wire is predicted differentially according to the ice-covered category, and the operations include:

[0025] If the image is of the insulator ice-covered category, calculate the pixel area of the blue region in the segmented image, and use the empirical formula to predict the ice-covered thickness of the insulator H J , and the expression of its formula is:

[0026]

[0027] where J1 is the preset average thickness of the insulator, S J is the pixel area of the blue region in the image, Q J is the pixel area before the insulator is ice-covered; the pixel area before the insulator is ice-covered is preset, the insulator is photographed from multiple angles before being ice-covered, and multiple groups of pixel areas are calculated respectively, and the average of them is the pixel area before the insulator is ice-covered Q J ;

[0028] If the image is of the wire ice-covered category, for the fully ice-covered wire category, calculate the average diameter of the green region in the figure, and the calculation formula is as follows:

[0029]

[0030] where, d 1 is the actual diameter of the wire, d 2 is the average diameter of the green region in the figure, σ is the error weight;

[0031] If the image is of the wire ice-covered category, for the semi-ice-covered wire category, calculate the average diameter of the red part and the green part in the figure, and the calculation formula is as follows:

[0032]

[0033] where, d 2 is the average diameter of the green region in the figure, D 1 is the average diameter of the red region in the figure, σ is the error weight;

[0034] If the image is of the category where both the insulator and the wire are ice-covered, calculate the ice-covered thickness of the wire and the insulator respectively.

[0035] Further, partition calculations are performed on the image using different values of the three primary colors; for the fully-iced conductor image, the actual diameter of the conductor is set d 1. Use the Laplacian operator to extract the edges of the green region, and take the edges as the image coordinate axes to obtain multiple sets of edge coordinates , for the edge points with the same coordinates, the difference in their coordinates β is the calculated diameter after icing. Use multiple sets of β the average value of the differences to find the average diameter of the green region in the figure d 2 is the diameter of the conductor after icing;

[0036] For the semi-iced conductor image, use the Canny operator to extract the edges of the red region in the image, and take the edges as the image coordinate axes to obtain multiple sets of edge coordinates ( x i , y i ). For multiple sets of x the edge points with the same coordinates, take the average of the differences in their respective vertical coordinates y to calculate the average diameter of the red region in the figure D 1 is the diameter of the conductor before icing; use the Laplacian operator to extract the edges of the green region in the image to obtain multiple sets of edge coordinates , for the edge points with the same coordinates, take the average of the differences in their respective vertical coordinates β to find the average diameter of the green region in the figure d 2 is the diameter of the conductor after icing.

[0037] Further, in step S4, different icing thickness calculation methods are adopted for the fully-iced and semi-iced conductor images, which can effectively reduce the errors generated when classified into one category for calculation.

[0038] Compared with the prior art, the beneficial effects of the present invention include:

[0039] The present invention provides an idea for differentially calculating the ice coating thickness and makes a detailed classification of the ice coating conditions; proposes a set of image preprocessing schemes to remove the noise in the ice coating images and the unevenness of the image edges, laying a foundation for calculating the diameter of the conductor and the pixel area of the insulator in the later stage; provides an improved scheme for the image segmentation algorithm, which not only improves the accuracy of image segmentation but also can be incorporated into a system with poor hardware performance; aiming at the deficiency of calculating the ice coating thickness by using the diameter difference before and after icing, proposes to obtain the coordinate axes of the target edge by using the edge algorithm; proposes an idea for differentially predicting the ice coating thickness of the transmission line by using the color difference in the image; proposes to use the error weight to balance the possible prediction errors; the present invention can classify three states of conductor icing, insulator icing, and both conductor and insulator icing, and when there is conductor icing, it classifies the conductor icing conditions into three levels: no icing, half icing, and full icing, and provides three calculation methods for the ice coating thickness for different ice coating conditions, which can quickly and accurately predict the ice coating thickness of the transmission line and provide a technical reference for power inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flowchart of the method of the present invention;

[0041] Figure 2 is a structural diagram of the improved PSPnet network. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] The implementation flowchart of the present invention is as Figure 1 shown, and the embodiments will be described in detail below. The prediction of the ice coating thickness of the transmission line includes the following steps:

[0043] S1. Construct a dataset of transmission line ice coating images and perform preprocessing and label making: Construct a dataset of transmission line ice coating images, first perform first-order filtering on the ice coating images, where the filtering coefficient = 0.35; secondly, sharpen the filtered image, and the sharpening gradient threshold T = 7; finally, perform image adaptive smoothing on the sharpened image, where the adaptive smoothing is composed of linear smoothing and non-linear smoothing. Linear smoothing includes neighborhood averaging method, selective averaging method, and Wiener filtering, and non-linear smoothing includes median filtering, Prewitt operator, and Sobel operator. One of the modules is automatically selected for image smoothing by calculating the difference in pixel values of the image edges. The smaller the pixel value difference, the better the image smoothing effect;

[0044] The labels of the dataset are made into four categories: conductors, ice-covered conductors, insulators, and the background. Use the Labelme software to make labels for the ice-covered images. When the ice-covered shape is extremely irregular, the ice-covered shape should be carefully marked. Since the characteristics of the insulator before and after icing are not very different, the ice-covered insulators and non-iced insulators are labeled as one category. When calculating the ice thickness of the insulator, the ice thickness of the non-iced insulator will be approximately equal to 0. The format of the labeled label file is.json format, and it should be uniformly converted to.png format before model training. In addition, the ice-covered dataset is required to be no less than 200 images, and the non-iced dataset is no less than 100 images. In this embodiment, 1840 images are used as training samples.

[0045] S2. As Figure 2 shown, construct an improved PSPNet semantic segmentation network and train it using the dataset. Replace the VGG network as the backbone network and use dilated convolutions Conv_rate with dilation rates of 1, 3, 6, and 12 to replace the max pooling in the backbone network to increase the receptive field of the network. After feature extraction by the backbone network, an effective feature layer F5 with a size of 32×32×512 can be obtained. At this time, improve the PSP module. After F5 passes through average pooling with convolution kernel sizes of 8, 16, and 32 respectively, it is divided into regions of 4×4, 2×2, and 1×1 respectively, and these three regions are stacked to obtain an enhanced feature layer P5, so as to aggregate the feature information of different regions, thereby further improving the global receptive field. Use a 1×1 convolution to adjust the number of channels of the enhanced feature layer P5 to 4. At this time, input the preprocessed picture into the improved network for training.

[0046] S2.1. Set the size of the input picture to 512×512×3. After two ordinary convolutions, the size of the feature layer becomes 512×512×64. At this time, through a dilated convolution Conv_rate with a dilation rate of 1, the size of the feature layer F1 becomes 256×256×64. After F1 is downsampled and passes through two ordinary convolutions, the size of the feature layer becomes 256×256×128. At this time, through a dilated convolution Conv_rate with a dilation rate of 3, the size of the feature layer F2 becomes 128×128×128. The feature layer F2 continues to be downsampled. After three ordinary convolutions, use a dilated convolution Conv_rate with a dilation rate of 6 to change the size of the feature layer F3 to 64×64×256. Similarly, the feature layer F3 continues to be downsampled. After three ordinary convolutions, use a dilated convolution Conv_rate with a dilation rate of 12 to change the size of the feature layer F4 to 32×32×512. The feature layer F4 continues to be downsampled and passes through three ordinary convolutions to obtain a feature layer F5 with a size of 32×32×512.

[0047] S2.2. Extract the feature layer F5 from the improved backbone network VGG, and use a 1×1 convolution to adjust the size of F5 to 32×32×1024. At this time, average pooling with convolutional kernel sizes of 8, 16, and 32 is performed respectively to divide F5 into three regions with sizes of 4×4, 2×2, and 1×1; Stack the feature layers after average pooling to obtain the enhanced feature layer P5, and use the resize operation to change the size of the feature layer P5 to 512×512, and use a 1×1 convolution to change the number of channels to 4. At this time, the final feature layer P6 is obtained, and P6 is output for semantic segmentation; At this time, the improved PSPNet semantic segmentation network is constructed;

[0048] S2.3. During model training, the entire process uses random parameter initialization and does not require the use of other tools for assistance; In this embodiment, the model learning rate is 0.001, the batch size Batchsize is 16, and the number of multi-threads num_workers = 2.

[0049] The construction of the dilated convolution Conv_rate in steps S2.1 and S2.2 specifically includes: introducing a hyperparameter dilation rate on the basis of ordinary convolution to increase the receptive field. The dilation rates of the dilated convolution in S2.1 are 1, 3, 6, and 12 respectively.

[0050] S3. Use the trained model to perform semantic segmentation on the input image and classify the icing categories: In this embodiment, 250 model weights can be obtained after training, and the model weight with the smallest loss value is taken to perform semantic segmentation on the icing image; When performing semantic segmentation, set the color of the wire to red, and the values of the three primary colors of red are R = 255, G = 0, B = 0; Set the icing part of the wire to green, and the values of the three primary colors of green are R = 0, G = 255, B = 0; Set the insulator part to blue, and the values of the three primary colors of blue are R = 0, G = 0, B = 255; The rest is regarded as the picture background and set to black; Determine wire icing, insulator icing, and icing of both wire and insulator based on the values of the three primary colors in the image; On the basis of determining the existence of wire icing, further use the values of the three primary colors in the image to determine whether the wire is fully iced, semi-iced, or non-iced;

[0051] If only blue or both blue and red exist in the image, it is classified as the insulator icing category;

[0052] If there is no blue in the image, it is classified as icing on the conductor. Then, a three-classification is performed again based on the target color in the picture: if the segmented image only has the base color R = 255, it is classified as no icing on the conductor, and the icing thickness is not calculated; if the segmented image only has the base color G = 255, it is classified as full icing on the conductor, the conductor diameter is set, and its icing thickness is calculated; if the segmented image has both red and green, it is classified as semi-icing, and the icing thickness is calculated using the diameter difference between different color regions;

[0053] If both blue and green exist in the image, it is classified as icing on both the insulator and the conductor.

[0054] S4. Differentially predict the icing thickness of the insulator and the conductor according to the icing category. The operations include:

[0055] S4.1. If the image is of the icing-on-insulator category, calculate the pixel area of the blue region in the segmented image, and use the empirical formula to predict the icing thickness of the insulator H J , and the expression of the formula is:

[0056]

[0057] where J1 is the preset average thickness of the insulator, S J is the pixel area of the blue region in the image, Q J is the pixel area of the insulator before icing; the pixel area of the insulator before icing is preset. The insulator is photographed from multiple angles before icing, and multiple groups of pixel areas are calculated respectively. The average of them is the pixel area of the insulator before icing Q J ;

[0058] S4.2. Use the different values of the three primary colors to perform zonal calculations on the image; for the full-icing image of the conductor, set the actual diameter of the conductor d 1. Use the Laplacian operator to extract the edges of the green region, and use the edges as the image coordinate axes to obtain multiple groups of edge coordinates , for the edge points with the same coordinates, the difference of their coordinates β is the diameter after calculating the icing. Use the average of multiple groups of β differences to find the average diameter of the green region in the figure d 2, which is the diameter of the conductor after icing;

[0059] For the semi-icing image of the conductor, use the Canny operator to extract the edges of the red region in the image, and use the edges as the image coordinate axes to obtain multiple groups of edge coordinates( x i ,y i ), for multiple groups x of edge points with the same coordinates, the average of the differences in their respective vertical coordinates y is taken to calculate the average diameter of the red area in the figure D 1 is the diameter of the wire before icing; the Laplacian operator is used to extract the edges of the green area in the image to obtain multiple groups of edge coordinates , for edge points with the same coordinates, the average of the differences in their respective vertical coordinates β is taken to find the average diameter of the green area in the figure d 2 is the diameter of the wire after icing.

[0060] S4.2. Use the ratio of the true value to the predicted value of the wire as the error weight to adjust the calculation formula of the ice thickness in real time. The operations include: before predicting the ice thickness of the wire, first calculate the actual diameter of the wire d 1 and the average diameter of the red area in the figure D 1, and use it as the error weight. The calculation formula is as follows:

[0061]

[0062] This error weight is not a fixed value, and its size is adjusted in real time according to different predicted wires.

[0063] S4.3. If the image is of the wire icing type, for the fully-iced wire type, calculate the average diameter of the green area in the figure. The calculation formula is as follows:

[0064]

[0065] Among them, d 1 is the actual diameter of the wire, d 2 is the average diameter of the green area in the figure, σ is the error weight;

[0066] For the semi-iced wire type, calculate the average diameter of the red part and the green part in the figure. The calculation formula is as follows:

[0067]

[0068] Among them, d 2 is the average diameter of the green area in the figure, D 1 is the average diameter of the red area in the figure, σ is the error weight;

[0069] S4.4. If the image is of the type where both the insulator and the wire are iced, calculate the ice thicknesses of the wire and the insulator respectively.

[0070] In this embodiment, according to the possible influencing situations in actual operation: when preprocessing the target image, regardless of the image quality, the above-mentioned preprocessing steps need to be performed on it; when improving the PSPNet semantic segmentation network, using pooling and dilated convolution should make the change in the size of the feature layer conform to the setting of the entire network; when setting the colors of the wire, insulator, and icing area in semantic segmentation, the color difference should be maximized to make the partition more obvious; the setting of the error weight must rely on the actual diameter of the wire as the known diameter.

[0071] The above only expresses the preferred embodiments of the present invention, and its description is relatively specific and detailed, but it cannot be understood as a limitation to the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations, improvements, and substitutions can be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. A method for predicting the differential icing thickness of a transmission line, characterized in that: S1. Construct a dataset of icing images of the transmission line and perform preprocessing to remove icing image noise and smooth the image edges; S2. Use an improved PSPNet semantic segmentation network to extract the icing conductors and insulators in the image; S3. Classify the icing categories according to the three primary color values existing in the image; S4. Differentially predict the icing thickness of the insulators and conductors according to the icing categories, and use the ratio of the true value to the predicted value of the conductor as the error weight to adjust the calculation formula of the icing thickness in real time; In S2, the operation of using the improved PSPNet semantic segmentation network to extract the icing conductors and insulators in the image includes: changing to the VGG network as the backbone network, and using the dilated convolution Conv_rate to replace the max pooling in the backbone network; improving the PSP module, using a convolution kernel matching the size of the feature layer for average pooling, and dividing each feature layer into three regions according to the size of the feature layer after pooling; integrating the divided regions and adjusting the number of channels to 4; In S3, the operation of classifying the icing categories according to the three primary color values existing in the image includes: when performing semantic segmentation on the icing image, setting the color of the non-iced conductor to red, the part of the iced conductor to green, the iced insulator to blue, and the rest as the image background and setting it to black; judging whether the image belongs to conductor icing, insulator icing, and both conductor and insulator icing according to the values of the three primary colors in the image; further judging whether the conductor is fully iced, semi-iced, and non-iced based on the values of the three primary colors in the image on the basis of judging the existence of conductor icing; The operation of differentially predicting the icing thickness of the insulators and conductors according to the icing categories includes: If the image is of the insulator icing type, calculate the pixel area of the blue region in the segmented image and use the empirical formula to predict the insulator icing thickness H J , and the expression of the formula is as follows: Among them, J1 is the preset average thickness of the insulator, and S J is the pixel area of the blue region in the image, and Q J is the pixel area before the insulator is iced; the pixel area before the insulator is iced is preset. The insulator is photographed from multiple angles before icing, and multiple groups of pixel areas are calculated respectively. After taking the average, it is the pixel area Q J ; If the image is of the conductor icing type, for the fully iced conductor type, calculate the average diameter of the green area in the figure, and the calculation formula is as follows: where d1 is the actual diameter of the conductor, d2 is the average diameter of the green area in the figure, and σ is the error weight; If the image is of the conductor icing type, for the semi-iced conductor type, calculate the average diameter of the red part and the green part in the figure, and the calculation formula is as follows: where d2 is the average diameter of the green area in the figure, D1 is the average diameter of the red area in the figure, and σ is the error weight; If the image is of the type where both the insulator and the conductor are iced, calculate the icing thickness of the conductor and the insulator respectively.

2. The differential prediction method for ice coating thickness of a transmission line according to claim 1, characterized in that In S1, the preprocessing method includes: applying first-order filtering to the icing image; sharpening the filtered image using the Roberts gradient algorithm; performing adaptive smoothing on the sharpened image, where the adaptive smoothing is composed of linear smoothing and non-linear smoothing. Linear smoothing includes neighborhood averaging method, selective averaging method, and Wiener filtering. Non-linear smoothing includes median filtering, Prewitt operator, and Sobel operator. Automatically select one of the modules for image smoothing processing by calculating the difference in the pixel values of the image edges.

3. A method for differentially predicting the icing thickness of a transmission line according to claim 1, characterized in that On the basis of determining the existence of conductor icing, further use the numerical values of the three primary colors in the image to judge whether the conductor is fully iced, semi-iced, or non-iced. If only blue exists in the image or both blue and red exist simultaneously, it is classified as insulator icing; If blue does not exist in the image, it is classified as conductor icing, and a three-classification is performed again based on the target color in the picture: if only red exists in the segmented image, it is classified as non-iced conductor, and the icing thickness is not calculated; if only green exists in the segmented image, it is classified as fully iced conductor, the conductor diameter is set, and its icing thickness is calculated; If both red and green exist in the segmented image, it is classified as semi-iced conductor, and the icing thickness is calculated using the diameter difference between different color regions; If both blue and green exist in the image, it is classified as both insulator and conductor iced.

4. A method for predicting the differential ice thickness of a transmission line according to claim 1, characterized in that, In S4, the ratio of the true value to the predicted value of the conductor is used as the error weight to adjust the calculation formula of the icing thickness in real time. The operations include: before predicting the conductor icing thickness, first calculate the ratio of the actual diameter d1 of the conductor to the average diameter D1 of the red region in the picture, and use it as the error weight. The calculation formula is as follows: This error weight is not a fixed value, and its size is adjusted in real time according to different predicted conductors.

5. A method for predicting the differential icing thickness of a transmission line according to claim 1, characterized in that Regarding the calculation of the diameters of different color regions, partitioning calculations are performed on the image using the different values of the three primary colors; for the fully-iced conductor image, the actual diameter d1 of the conductor is set, and the Laplacian operator is used to extract the edges of the green region. Using the edges as the image coordinate axes, multiple sets of edge coordinates (α i , β i ) are obtained. For the edge points with the same α coordinate, the difference in their β coordinates is the calculated diameter after icing. The average diameter d2 of the green region in the figure is obtained using the average value of multiple sets of β differences, which is the diameter of the conductor after icing; For the semi-iced conductor image, use the Canny operator to extract the edges of the red region in the image, and use the edges as the image coordinate axes to obtain the coordinates (x i , y i ) of multiple groups of edges. For the edge points with the same x coordinate in multiple groups, take the average of the differences in their respective y coordinates to calculate the average diameter D1 of the red region in the image, which is the diameter of the conductor before icing; use the Laplacian operator to extract the edges of the green region in the image to obtain multiple groups of edge coordinates (α i , β i ). For the edge points with the same α coordinate, take the average of the differences in their respective β coordinates to calculate the average diameter d2 of the green region in the image, which is the diameter of the conductor after icing.

Images