Power transmission line ice melting method and device based on multi-scale icing diagnosis

Abstract

The invention provides a power transmission line ice melting method and device based on multi-scale icing diagnosis, and relates to the field of high-voltage power transmission. The method comprises the following steps: acquiring a power transmission line sampling image, and marking the power transmission line sampling image to obtain a first power transmission line icing data set; performing image transformation data enhancement on the first power transmission line icing data set to obtain a second power transmission line icing data set; performing model training based on the second power transmission line icing data set to obtain a multi-scale power transmission line icing diagnosis network; a power transmission line image of the target power transmission line is acquired, power transmission line icing diagnosis is performed on the power transmission line image by using the multi-scale power transmission line icing diagnosis network, and an icing diagnosis result is output and comprises position coordinates of a target position needing ice melting, an icing type label and a confidence score; and based on an icing diagnosis result, carrying out ice melting operation through the directional induction ice melting device.

Description

Technical Field

[0001] This application relates to the field of high-voltage power transmission, and in particular to a method and device for de-icing transmission lines based on multi-scale icing diagnosis. Background Technology

[0002] In the operation and maintenance of high-voltage transmission systems, accurately identifying the type of icing and implementing targeted de-icing are core aspects of ensuring power grid safety. Current detection methods primarily rely on manual inspections, fixed sensor monitoring, or conventional image algorithms for diagnosing icing on transmission lines. These methods have the following drawbacks:

[0003] Manual inspections are difficult to distinguish between different types of icing, such as rime and glaze, and are inefficient. Tension sensors cannot identify the physical properties of ice. Conventional image algorithms have a high error rate when detecting fine-grained icing data, and their diagnostic efficiency is also low.

[0004] The icing environment of power transmission lines presents challenges such as variable ice patterns, large differences in scale, and complex backgrounds. Existing detection methods cannot achieve satisfactory detection results and efficiency. Summary of the Invention

[0005] In view of this, this application provides a method and equipment for de-icing transmission lines based on multi-scale icing diagnosis, which solves the problem that the detection of icing on transmission lines in related technologies cannot achieve good detection results and efficiency.

[0006] In a first aspect, embodiments of this application provide a method for de-icing transmission lines based on multi-scale icing diagnosis, including:

[0007] Acquire sampled images of the transmission line and annotate them to obtain the first transmission line icing dataset;

[0008] Image transformation and data augmentation were performed on the first transmission line icing dataset to obtain the second transmission line icing dataset.

[0009] A multi-scale transmission line icing diagnosis network was obtained by training the model based on the second transmission line icing dataset.

[0010] The system acquires images of the target transmission line, uses a multi-scale transmission line icing diagnosis network to diagnose icing in the images, and outputs icing diagnosis results, including the location coordinates of the target location that needs to be melted, the icing type label, and the confidence score.

[0011] Based on the icing diagnosis results, icing operations are carried out using a directional induction icing melting device.

[0012] The method described in the embodiments of this application may also have the following additional technical features:

[0013] In the above technical solution, optionally, the sampled images of the transmission line are annotated, including:

[0014] The icing type and location coordinates of the icing targets in the sampled images of the transmission line are labeled. The icing types include: no icing, icing and rime, icing and mixed rime, icing and wet snow, and icing and rime rain.

[0015] Optionally, in any of the above technical solutions, image transformation data enhancement is performed on the first transmission line icing dataset to obtain a second transmission line icing dataset, including:

[0016] For any one of the transmission line sampling images in the first transmission line icing dataset, image transformation data enhancement is performed, and the image after image transformation data enhancement is combined with the first transmission line icing dataset to form an enhanced transmission line icing dataset. The image transformation data enhancement method includes at least one of the following: rotation, flipping, cropping, random occlusion, contrast adjustment, random scaling, and noise addition.

[0017] The enhanced transmission line icing dataset is randomly divided into training data, validation data, and test data according to a preset ratio to obtain a second transmission line icing dataset containing training data, validation data, and test data.

[0018] A multi-scale transmission line icing diagnosis network was obtained by training the model based on the second transmission line icing dataset, including:

[0019] The training and validation data are input into the multi-scale transmission line icing diagnosis network for training until the network loss function converges.

[0020] The trained multi-scale transmission line icing diagnosis network is tested using test data. If the accuracy of the test results is greater than or equal to the preset accuracy threshold, the multi-scale transmission line icing diagnosis network is obtained.

[0021] In any of the above technical solutions, optionally, the multi-scale transmission line icing diagnosis network includes an image patch embedding layer, a model backbone module, a feature focusing diffusion pyramid, an average pooling layer, and a regression classifier. The model backbone module includes a multi-scale scene feature interaction module.

[0022] The initial image is input into the image patch embedding layer to convert the initial image into serialized data, resulting in the first transmission line icing diagnostic feature map Z1. The first transmission line icing diagnostic feature map Z1 is then input into the model backbone module and the feature focusing diffusion pyramid to obtain the second transmission line icing diagnostic feature map Z2. The initial image is either the second transmission line icing dataset or a transmission line image.

[0023] The icing diagnostic feature map Z2 of the second transmission line is downsampled to obtain the icing diagnostic feature map Z3 of the third transmission line;

[0024] The icing diagnostic feature map Z3 of the third transmission line is sequentially input into the main module of the model and the feature focusing diffusion pyramid to obtain the icing diagnostic feature map Z4 of the fourth transmission line.

[0025] The icing diagnostic feature map Z4 of the fourth transmission line is downsampled to obtain the icing diagnostic feature map Z5 of the fifth transmission line. The icing diagnostic feature map Z5 of the fifth transmission line is then input into the main module of the model and the feature focusing diffusion pyramid to obtain the icing diagnostic feature map Z6 of the sixth transmission line.

[0026] The icing diagnosis feature map Z6 of the sixth transmission line is averaged using an average pooling layer and then input into a regression classifier to obtain the icing diagnosis feature map Z7 of the seventh transmission line. The icing diagnosis feature map Z7 of the seventh transmission line is then used as the icing diagnosis result.

[0027] In any of the above technical solutions, optionally, the main module of the model includes a first DW convolutional and batch normalization layer, a first convolutional layer, a first tensor summation unit, a multi-scale scene feature interaction module, a second tensor summation unit, a second DW convolutional and batch normalization layer, a second convolutional layer, and a third tensor summation unit connected in sequence.

[0028] The first transmission line icing feature map Y0 is sequentially input into the first DW convolutional and batch normalization layer and the first convolutional layer to obtain the first scene feature interaction backbone feature map Y1.

[0029] The first tensor summation unit adds the first transmission line icing feature map Y0 and the first scene feature interaction backbone feature map Y1 to obtain the second scene feature interaction backbone feature map Y2.

[0030] The second scene feature interaction backbone feature map Y2 is input into the multi-scale scene feature interaction module to obtain the third scene feature interaction backbone feature map Y3. Then, through the second tensor summation unit, the second scene feature interaction backbone feature map Y2 and the third scene feature interaction backbone feature map Y3 are tensor-added to obtain the fourth scene feature interaction backbone feature map Y4.

[0031] The fourth scene feature interaction backbone feature map Y4 is sequentially input into the second DW convolutional and batch normalization layer and the second convolutional layer to obtain the fifth scene feature interaction backbone feature map Y5.

[0032] The third tensor summation unit adds the fourth scene feature interaction backbone feature map Y4 and the fifth scene feature interaction backbone feature map Y5 into a tensor to obtain the sixth scene feature interaction backbone feature map Y6. The sixth scene feature interaction backbone feature map Y6 is the output of the model backbone module.

[0033] In any of the above technical solutions, optionally, the multi-scale scene feature interaction module includes a channel segmentation unit, a multi-kernel deep convolutional layer, a long-range information enhancement transformation unit, and a channel splicing unit;

[0034] The second transmission line icing feature map X0 is divided into three equal parts along the channel dimension by the channel segmentation unit to obtain the first scene feature interaction feature map X1, the second scene feature interaction feature map X2 and the third scene feature interaction feature map X3 respectively.

[0035] The first scene feature interaction feature map X1 is input into a multi-kernel deep convolutional layer to enhance the perception ability of different receptive fields, resulting in the fourth scene feature interaction feature map X4.

[0036] The second scene feature interaction feature map X2 is transformed by the long-range information enhancement transformation unit to obtain the fifth scene feature interaction feature map X5.

[0037] The third scene feature interaction feature map X3, the fourth scene feature interaction feature map X4, and the fifth scene feature interaction feature map X5 are stitched together along the channel dimension through the channel stitching unit to obtain the sixth scene feature interaction feature map X6. The sixth scene feature interaction feature map X6 is the output of the multi-scale scene feature interaction module.

[0038] In any of the above technical solutions, optionally, the feature focusing diffusion pyramid includes a first branch, a second branch, a third branch, a channel merging unit, a fourth branch, a fifth branch, a sixth branch, a seventh branch, a fourth tensor summing unit, a pointwise convolution unit, and a fifth tensor summing unit;

[0039] The icing feature map F0 of the third transmission line is input into the first branch, the second branch and the third branch respectively. The first branch performs upsampling and convolution operations in sequence to obtain the first focused diffusion feature map F1. The second branch performs convolution operation to obtain the second focused diffusion feature map F2. The third branch performs adaptive downsampling operation to obtain the third focused diffusion feature map F3.

[0040] The first focusing diffusion feature map F1, the second focusing diffusion feature map F2, and the third focusing diffusion feature map F3 are merged along the channel dimension through the channel merging unit to obtain the fourth focusing diffusion feature map F4.

[0041] The fourth focus diffusion feature map F4 is input into the fourth, fifth, sixth, and seventh branches respectively. In the fourth branch, a depthwise convolution with a 5×5 kernel is performed to obtain the fifth focus diffusion feature map F5; in the fifth branch, a depthwise convolution with a 7×7 kernel is performed to obtain the sixth focus diffusion feature map F6; in the sixth branch, a depthwise convolution with a 9×9 kernel is performed to obtain the seventh focus diffusion feature map F7; and in the seventh branch, a depthwise convolution with an 11×11 kernel is performed to obtain the eighth focus diffusion feature map F8.

[0042] The fourth tensor summation unit is used to add the tensors of the fourth focusing diffusion feature map F4, the fifth focusing diffusion feature map F5, the sixth focusing diffusion feature map F6, the seventh focusing diffusion feature map F7, and the eighth focusing diffusion feature map F8 to obtain the ninth focusing diffusion feature map F9.

[0043] The ninth focusing diffusion feature map F9 is input into a pointwise convolutional unit for feature extraction to obtain the tenth focusing diffusion feature map F. 10 ;

[0044] Using the fifth tensor summation unit, the third transmission line icing feature map F0 and the tenth focusing diffusion feature map F are combined. 10 Tensor addition yields the eleventh focusing diffusion feature map F. 11 Eleventh Focused Diffusion Feature Map F 11 This is the output of the feature-focused diffusion pyramid.

[0045] In any of the above technical solutions, optionally, the directional induction de-icing device is used to calculate de-icing parameters based on the icing diagnosis results, and uses the principle of high-frequency electromagnetic induction to induce eddy currents inside the target transmission line to heat up the target transmission line itself, thereby melting the surface icing of the target transmission line.

[0046] The directional induction ice melting device includes a servo mechanism and an induction coil array, with the servo mechanism used to drive the induction coil array to move.

[0047] The induction coil array includes a hinge and a coil. The coil includes a coil housing, a transmission conductor, a copper tube winding unit, a U-shaped magnetic core conductor, a ceramic energy-concentrating groove, and magnetic lines of force. The copper tube winding unit includes a multi-turn copper tube. The distance between the innermost copper tube and the surface of the transmission conductor is fixed. The U-shaped magnetic core conductor covers the transmission conductor. The ceramic energy-concentrating groove is bonded to the copper tube winding unit with epoxy adhesive. The magnetic lines of force are used to guide and promote the generation of induced eddy currents.

[0048] In any of the above technical solutions, optionally, the directional induction de-icing device is specifically used for:

[0049] Calculate the heating power P required for melting the ice based on the location coordinates and ice type label.total And the time for ice melting;

[0050] The induction coil array is moved to align with the target power transmission line via a servo mechanism;

[0051] Apply heating power P to the induction coil array total Based on the location coordinates and the de-icing time, induction heating is applied to the target power transmission line.

[0052] Among them, the optimal heating power P total The calculation formula is:

[0053] p total =p base ×L ice ×K temp

[0054] p base Based on power density, L ice K is the length of the icing layer. temp This is the temperature compensation coefficient;

[0055] Melting time t heat The calculation formula is:

[0056] t heat =max(Th melt ,a×k density )×(D wire / a) ^ b

[0057] Where, k density D is the density coefficient of the icing type. wire Th is the diameter of the transmission line of the induction coil array. melt To ensure the shortest melting time, 'a' is the unit density calibration factor and 'b' is the diameter-heat capacity correlation index.

[0058] Secondly, embodiments of this application provide a transmission line de-icing device based on multi-scale icing diagnosis, characterized in that it includes:

[0059] An image capturing device for acquiring sampled images of transmission lines and images of target transmission lines;

[0060] An icing diagnostic device is used for: annotating sampled images of transmission lines to obtain a first transmission line icing dataset; performing image transformation and data augmentation on the first transmission line icing dataset to obtain a second transmission line icing dataset; training a model based on the second transmission line icing dataset to obtain a multi-scale transmission line icing diagnostic network; and using the multi-scale transmission line icing diagnostic network to perform transmission line icing diagnosis on transmission line images and outputting icing diagnosis results.

[0061] A directional induction de-icing device is used to perform de-icing operations based on icing diagnosis results.

[0062] The transmission line de-icing method and equipment based on multi-scale icing diagnosis in this application establish a multi-scale transmission line icing diagnosis network with strong generalization, high accuracy, and automation. It can efficiently identify whether icing exists on the transmission line, as well as the type and location coordinates of the icing. At the same time, based on the design of a directional induction de-icing device, it performs directional and timed de-icing on the target transmission line where icing occurs. After the network training is completed, the de-icing method designed in this application can be deployed on each transmission line to realize real-time diagnosis and de-icing of transmission line icing.

[0063] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0064] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0065] Figure 1 A flowchart illustrating the transmission line de-icing method based on multi-scale icing diagnosis according to an embodiment of this application is shown.

[0066] Figure 2 This paper illustrates the working logic of a multi-scale transmission line icing diagnostic network according to an embodiment of this application.

[0067] Figure 3 This paper illustrates a schematic diagram of the working logic of the main module of the model in an embodiment of this application.

[0068] Figure 4 This diagram illustrates the working logic of the multi-scale scene feature interaction module according to an embodiment of this application.

[0069] Figure 5 This paper illustrates a schematic diagram of the working logic of the feature focusing diffusion pyramid according to an embodiment of this application.

[0070] Figure 6 A top cross-sectional view of an induction coil array according to an embodiment of this application is shown. Detailed Implementation

[0071] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0072] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0073] The following description, in conjunction with the accompanying drawings, details the transmission line de-icing method and equipment based on multi-scale icing diagnosis provided in this application through specific embodiments and application scenarios. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0074] This application combines deep learning networks, monitoring equipment, and a directional induction de-icing module to achieve real-time diagnosis and de-icing of transmission line icing. An embodiment of this application provides a method for de-icing transmission lines based on multi-scale icing diagnosis, such as... Figure 1 As shown, the method includes:

[0075] Step 101: Obtain sampled images of the transmission line and annotate the sampled images of the transmission line to obtain the first transmission line icing dataset.

[0076] In this step, sampling images of the transmission lines with and without icing are acquired using an image capturing device, and the sampling images are then filtered to identify data showing icing in various types of transmission lines.

[0077] After the transmission line sampling images are filtered, the transmission line sampling images are labeled. In one embodiment of this application, labeling the transmission line sampling images includes: labeling the icing type and location coordinates of the icing targets in the transmission line sampling images, wherein the icing type includes: no icing, icing and rime, icing and mixed rime, icing and wet snow, and icing and rime rain.

[0078] In this embodiment, an image annotation tool is used to annotate the icing type and location coordinates of all selected transmission line sample images. The icing type is divided into no icing, icing with rime, icing with mixed rime, icing with wet snow, and icing with rain rime, etc. Through annotation, the first transmission line icing dataset D1 is obtained.

[0079] In one implementation, the image capturing device may be a power transmission line monitoring camera, a power transmission line inspection drone, etc., and the image annotation tools may include LabelImg, LabelME, etc.

[0080] Step 102: Perform image transformation data enhancement on the first transmission line icing dataset to obtain the second transmission line icing dataset.

[0081] In most cases, due to numerous limitations in collecting images of icing on power transmission lines at the field, the number of collected images is generally less than 1000, which is insufficient to cover the diversity and complexity required for training the detection task, making it difficult to train the model effectively. Therefore, this application performs image transformation data augmentation on the first power transmission line icing dataset to increase the diversity of samples, resulting in a second power transmission line icing dataset.

[0082] In one embodiment of this application, step 102 involves performing image transformation data enhancement on the first transmission line icing dataset to obtain a second transmission line icing dataset, including:

[0083] Step 1.1: Take any one of the transmission line sampling images in the first transmission line icing dataset D1, perform image transformation data enhancement, and combine the image after image transformation data enhancement with the first transmission line icing dataset D1 to form the enhanced transmission line icing dataset D2. The image transformation data enhancement method includes at least one of the following: rotation, flipping, cropping, random occlusion, contrast adjustment, random scaling, and noise addition.

[0084] Step 1.2: Randomly divide the enhanced transmission line icing dataset D2 into training data D4, validation data D5, and test data D6 according to a preset ratio to obtain a second transmission line icing dataset D7 containing training data D4, validation data D5, and test data D6.

[0085] In another embodiment, after obtaining the enhanced transmission line icing dataset D2, scene data augmentation is performed on the enhanced transmission line icing dataset D2 to obtain the enhanced transmission line icing dataset D3. Then, the enhanced transmission line icing dataset D3 is randomly divided into training data D4, validation data D5, and test data D6 according to a preset ratio. Scene data augmentation is a technique that expands the diversity and robustness of the dataset by simulating different scene conditions or generating new scene samples. Its core lies in applying scene-related transformations such as lighting and color adjustments, rotation and translation, scaling and cropping, noise and occlusion to the original data, or using generative models to create synthetic scenes, thereby covering a wider range of scene variations. This application uses scene data augmentation to simulate complex real-world scenes and improve the robustness of the sample training network.

[0086] Step 103: Train the model based on the second transmission line icing dataset to obtain a multi-scale transmission line icing diagnosis network.

[0087] In one embodiment of this application, a multi-scale transmission line icing diagnosis network is obtained by training a model based on a second transmission line icing dataset, including:

[0088] The training and validation data are input into the multi-scale transmission line icing diagnosis network for training until the network loss function converges.

[0089] The trained multi-scale transmission line icing diagnosis network is tested using test data. If the accuracy of the test results is greater than or equal to the preset accuracy threshold, the multi-scale transmission line icing diagnosis network is obtained.

[0090] In this embodiment, the parameters and hyperparameters of the MTICD-Net Multi-scale (Transmission Line Icing Diagnosis Net) are initialized, including the number of training iterations, batch size, optimizer, learning rate type, and initial learning rate.

[0091] After completing the above preparations, the training data D4 and validation data D5 are input into the multi-scale transmission line icing diagnosis network for training until the network loss function converges. In each training round, the network uses transmission line icing data from the training set at preset batch processing values ​​to update the loss value and various parameters in the network. After each training round, the network uses transmission line icing data from the validation set at preset batch processing values ​​to verify the effectiveness of each training round. The training of the multi-scale transmission line icing diagnosis network ends when the loss value tends to converge.

[0092] After the multi-scale transmission line icing diagnosis network is trained, test data D6 is used for testing. Test data D6 is input into the trained multi-scale transmission line icing diagnosis network for performance evaluation. The test result evaluation metric for the multi-scale transmission line icing diagnosis network is accuracy, which is the percentage of transmission line icing types correctly identified from the test data out of the total test data. When the accuracy reaches the user's expectation (i.e., the preset accuracy threshold set by the user), the multi-scale transmission line icing diagnosis network test is complete, and the multi-scale transmission line icing diagnosis network is used to diagnose the types of transmission line icing.

[0093] In one embodiment of this application, the multi-scale transmission line icing diagnosis network includes a feature focusing and diffusion pyramid, a multi-scale scene feature interaction module, and a model backbone module with the multi-scale scene feature interaction module as its core. The feature focusing and diffusion pyramid has a high efficiency in extracting information for fine-grained scenes. The multi-scale scene feature interaction module and the model backbone module with the multi-scale scene feature interaction module as its core can avoid feature duplication and redundancy, enhance the computational efficiency of the network, and strengthen the ability to extract fine-grained information based on global modeling. The multi-scale transmission line icing diagnosis network also includes an image patch embedding layer, an average pooling layer, and a regression classifier. The working logic of the MTICD-Net network is as follows: Figure 2 As shown, it includes:

[0094] Step 2.1: Input the initial image into the image patch embedding layer to convert the initial image into serialized data, obtaining the first transmission line icing diagnostic feature map Z1. Then, input the first transmission line icing diagnostic feature map Z1 into the model backbone module and the feature focusing diffusion pyramid in sequence to obtain the second transmission line icing diagnostic feature map Z2. The initial image is either the second transmission line icing dataset or a transmission line image. When training the model, the initial image is the second transmission line icing dataset. When inference, the initial image is the transmission line image of the target transmission line.

[0095] Step 2.2: Downsample the second transmission line icing diagnostic feature map Z2 to obtain the third transmission line icing diagnostic feature map Z3; input the third transmission line icing diagnostic feature map Z3 into the model backbone module and the feature focusing diffusion pyramid in sequence to obtain the fourth transmission line icing diagnostic feature map Z4.

[0096] Step 2.3: Downsample the fourth transmission line icing diagnostic feature map Z4 to obtain the fifth transmission line icing diagnostic feature map Z5, and input the fifth transmission line icing diagnostic feature map Z5 into the model backbone module and the feature focusing diffusion pyramid in sequence to obtain the sixth transmission line icing diagnostic feature map Z6.

[0097] Step 2.4: The icing diagnosis feature map Z6 of the sixth transmission line is averaged using an average pooling layer and then input into a regression classifier to obtain the icing diagnosis feature map Z7 of the seventh transmission line. The icing diagnosis feature map Z7 of the seventh transmission line is used as the icing diagnosis result. The output icing diagnosis result contains a tensor of prediction information. Each row corresponds to a prediction. The icing diagnosis result includes the location coordinates of the target location that needs to be melted, the icing type label, and the confidence score.

[0098] In one embodiment, this application provides a model backbone module (MSFIB, Multi-Scale SceneFeature Interaction Backbone) for feature extraction in a multi-scale transmission line icing diagnostic network. The model backbone module includes a first DW convolutional and batch normalization layer, a first convolutional layer, a first tensor summation unit, a multi-scale scene feature interaction module, a second tensor summation unit, a second DW convolutional and batch normalization layer, a second convolutional layer, and a third tensor summation unit, connected sequentially. The working logic of the model backbone module is as follows: Figure 3 The following are included:

[0099] Step 3.1: Input the first transmission line icing feature map Y0 into the first DW convolutional and batch normalization layer and the first convolutional layer in sequence to obtain the first scene feature interaction backbone feature map Y1; wherein, the first transmission line icing feature map Y0 is the input of the model backbone module, which can be the first transmission line icing diagnosis feature map Z1, the third transmission line icing diagnosis feature map Z3, and the fifth transmission line icing diagnosis feature map Z5.

[0100] Step 3.2: Using the first tensor summation unit, the first transmission line icing feature map Y0 and the first scene feature interaction backbone feature map Y1 are tensor-summed to obtain the second scene feature interaction backbone feature map Y2.

[0101] Step 3.3: Input the second scene feature interaction backbone feature map Y2 into the multi-scale scene feature interaction module to obtain the third scene feature interaction backbone feature map Y3. Then, through the second tensor summation unit, add the second scene feature interaction backbone feature map Y2 and the third scene feature interaction backbone feature map Y3 by tensor summation to obtain the fourth scene feature interaction backbone feature map Y4.

[0102] Step 3.4: Input the fourth scene feature interaction backbone feature map Y4 into the second DW convolutional layer and batch normalization layer and the second convolutional layer in sequence to obtain the fifth scene feature interaction backbone feature map Y5.

[0103] Step 3.5: Using the third tensor summation unit, the fourth scene feature interaction backbone feature map Y4 and the fifth scene feature interaction backbone feature map Y5 are tensor-summed to obtain the sixth scene feature interaction backbone feature map Y6. The sixth scene feature interaction backbone feature map Y6 is the output of the model backbone module.

[0104] In one embodiment, this application provides a Multi-Scale Scene Feature Interaction (MSFI) module. This module avoids feature redundancy and enhances the ability to extract fine-grained information based on global modeling. The fine-grained information refers to details such as the geometric properties and brightness of the icing edge. The MSFI module includes a channel segmentation unit, a multi-kernel deep convolutional layer, a long-range information enhancement transform unit, and a channel stitching unit. The working logic of the MSFI module is as follows: Figure 4 The following are included:

[0105] Step 4.1: Using the channel segmentation unit, the second transmission line icing feature map X0 is divided into 3 equal parts along the channel dimension to obtain the first scene feature interaction feature map X1, the second scene feature interaction feature map X2, and the third scene feature interaction feature map X3 respectively; wherein, the second transmission line icing feature map X0 is the input of the multi-scale scene feature interaction module and can be the second scene feature interaction backbone feature map Y2.

[0106] Step 4.2: Input the first scene feature interaction feature map X1 into a multi-kernel deep convolutional layer to enhance the perception ability of different receptive fields and obtain the fourth scene feature interaction feature map X4.

[0107] Step 4.3: The second scene feature interaction feature map X2 is transformed by the long-range information enhancement transformation unit to obtain the fifth scene feature interaction feature map X5.

[0108] The formula for long-range information enhancement transformation is as follows:

[0109] X 21 =σ(Conv(Linear(X2)[:0.5c]))

[0110] X 22 =σ(Linear(X2)[0.5c:])

[0111]

[0112] In the formula, σ represents the activation function, Linear represents the linear layer, c represents the number of channels, and Conv represents the convolution transformation.

[0113] Step 4.4: Using the channel splicing unit, the third scene feature interaction feature map X3, the fourth scene feature interaction feature map X4, and the fifth scene feature interaction feature map X5 are spliced ​​along the channel dimension to obtain the sixth scene feature interaction feature map X6. The sixth scene feature interaction feature map X6 is the output of the multi-scale scene feature interaction module.

[0114] It should be noted that no operation is performed on the third scene feature interaction feature map X3 before splicing along the channel dimension, which can reduce feature redundancy in high-dimensional space and reduce computational complexity.

[0115] This application provides a multi-scale scene feature interaction module that avoids feature duplication and redundancy, enhances the computational efficiency of the network, and strengthens the ability to extract fine-grained information based on global modeling, such as details of icing edges.

[0116] In one embodiment, this application provides a FeatureFocused Diffusion Pyramid (FFDP), which focuses the input features and uses a set of parallel deep convolutions to capture rich information across multiple scales. Finally, a diffusion mechanism diffuses the features with rich contextual information to various detection scales. The feature focusing module and feature diffusion mechanism of the FeatureFocused Diffusion Pyramid ensure that features at each scale have detailed contextual information, which is more conducive to subsequent target detection. The FeatureFocused Diffusion Pyramid includes a first branch 11, a second branch 12, a third branch 13, a channel merging unit, a fourth branch 21, a fifth branch 22, a sixth branch 23, a seventh branch 24, a fourth tensor summation unit, a pointwise convolution unit, and a fifth tensor summation unit. The working logic of the FeatureFocused Diffusion Pyramid is as follows: Figure 5 The following are included:

[0117] Step 5.1: Input the icing feature map F0 of the third transmission line into the first branch 11, the second branch 12, and the third branch 13 respectively. The first branch 11 performs upsampling and convolution operations in sequence to obtain the first focused diffusion feature map F1. The second branch 12 performs convolution operations to obtain the second focused diffusion feature map F2. The third branch 13 performs adaptive downsampling operations to obtain the third focused diffusion feature map F3. Among them, the icing feature map F0 of the third transmission line is the input of the feature focused diffusion pyramid and the output of the main module of the model.

[0118] Step 5.2: Through the channel merging unit, the first focusing diffusion feature map F1, the second focusing diffusion feature map F2, and the third focusing diffusion feature map F3 are merged along the channel dimension to obtain the fourth focusing diffusion feature map F4.

[0119] Step 5.3: Input the fourth focus diffusion feature map F4 into a parallel four-branch DW convolution (depth convolution) structure, that is, input it into the fourth branch 21, the fifth branch 22, the sixth branch 23, and the seventh branch 24 respectively. Perform a depth convolution with a kernel of 5×5 in the fourth branch 21 to obtain the fifth focus diffusion feature map F5; perform a depth convolution with a kernel of 7×7 in the fifth branch 22 to obtain the sixth focus diffusion feature map F6; perform a depth convolution with a kernel of 9×9 in the sixth branch 23 to obtain the seventh focus diffusion feature map F7; and perform a depth convolution with a kernel of 11×11 in the seventh branch 24 to obtain the eighth focus diffusion feature map F8.

[0120] Step 5.4: Using the fourth tensor summation unit, the fourth focusing diffusion feature map F4, the fifth focusing diffusion feature map F5, the sixth focusing diffusion feature map F6, the seventh focusing diffusion feature map F7, and the eighth focusing diffusion feature map F8 are tensor-summed to obtain the ninth focusing diffusion feature map F9.

[0121] Step 5.5: Input the ninth focusing diffusion feature map F9 into the pointwise convolutional unit for feature extraction to obtain the tenth focusing diffusion feature map F. 10 .

[0122] Step 5.6: Using the fifth tensor summation unit, combine the third transmission line icing feature map F0 and the tenth focusing diffusion feature map F... 10 Tensor addition yields the eleventh focusing diffusion feature map F. 11 Eleventh Focused Diffusion Feature Map F 11 This is the output of the feature-focused diffusion pyramid.

[0123] It should be noted that Pointwise convolution (PW convolution, also known as pointwise convolution unit) is one of the core components of depthwise separable convolution. It achieves feature transformation in the channel dimension through a 1×1 convolution kernel. The operation is very similar to the regular convolution operation. Its convolution kernel size is 1×1×M, where M is the number of channels in the previous layer.

[0124] Step 104: Obtain the transmission line image of the target transmission line, use a multi-scale transmission line icing diagnosis network to perform icing diagnosis on the transmission line image, and output the icing diagnosis results, which include the location coordinates of the target location that needs to be melted, the icing type label, and the confidence score.

[0125] In this step, for the target transmission line that needs to be iced, the transmission line image is captured by an image capturing device, and then the multi-scale transmission line icing diagnosis network is used to perform icing diagnosis on the transmission line image. The location coordinates of the target location that needs to be melted, the icing type label, and the confidence score are output, thus realizing the icing diagnosis of the transmission line through the multi-scale transmission line icing diagnosis network.

[0126] Step 105: Based on the icing diagnosis results, perform icing operation using a directional induction icing melting device.

[0127] In this step, after obtaining the icing diagnosis result, if the target transmission line is iced, the directional induction de-icing device is controlled to perform de-icing operations on the target transmission line.

[0128] The directional induction de-icing device is installed on transmission line towers. It receives icing diagnosis results from a multi-scale transmission line icing diagnosis network, calculates de-icing parameters, and uses the principle of high-frequency electromagnetic induction to induce eddy currents inside the target transmission line (mainly conductors or ground wires). The resulting Joule effect heats up the target transmission line itself, thus melting the surface ice. The directional induction de-icing device includes a servo mechanism and a high-frequency, high-power induction coil array. The servo mechanism drives the induction coil array, moving it to the icing location for de-icing, avoiding the energy waste and interference with non-iced areas associated with traditional global heating.

[0129] In one embodiment, this application provides a directional induction ice-melting device, including a servo mechanism and a high-frequency, high-power induction coil array. The induction coil array includes a hinge and a coil. The coil includes a coil housing, a power transmission conductor, a copper tube winding unit, a U-shaped magnetic core conductor, a ceramic energy-concentrating groove, and magnetic lines of force. The copper tube winding unit includes a multi-turn copper tube. The innermost copper tube of the multi-turn copper tube is spaced at a fixed distance from the surface of the power transmission conductor. The U-shaped magnetic core conductor covers the power transmission conductor. The ceramic energy-concentrating groove is bonded to the copper tube winding unit with epoxy adhesive. The magnetic lines of force are used to guide and promote the generation of induced eddy currents.

[0130] Specifically, such as Figure 6 As shown, the induction coil array of the directional induction ice-melting device adopts an openable dumbbell-shaped symmetrical structure and includes the following core components:

[0131] (1) Quick-release hinge 201, connected to the coil housing.

[0132] (2) Coil housing, including left coil housing 202 and right coil housing 203.

[0133] Made of high-strength glass fiber reinforced polyamide by injection molding, the shell wall thickness is 0.2 times the diameter of the conductor (typical value 5±0.5mm). The two halves of the shell are connected by quick-release hinges (105c) to achieve ±90° opening and closing movement. The parallelism error between the hinge axis and the center axis of the conductor is ≤0.05mm.

[0134] (3) Transmission conductor 204.

[0135] (4) Copper tube winding unit, each half coil has 4 turns of rectangular copper tube 205 inside, the rectangular copper tube 205 has a cross-sectional size of 5mm×3mm and a wall thickness of 1mm. The copper tubes are arranged in an Archimedean spiral structure, the distance between the innermost turn and the surface of the transmission conductor 204 is fixed and can be 8mm, the insulation gap between adjacent turns is ≥1mm, and DN6 quick-connect water nozzles are welded to both ends of the copper tube.

[0136] (5) The U-shaped magnetic core conductor 206 is made of 0.1mm thick nanocrystalline soft magnetic alloy strips stacked together. The magnetic core is U-shaped and covers the transmission conductor 204. The magnetic pole gap is 1mm. The ratio of the effective cross-sectional area of ​​the magnetic circuit to the cross-sectional area of ​​the transmission conductor 204 is 1:1.25.

[0137] (6) Ceramic energy-concentrating groove 207. The ceramic energy-concentrating groove 207 is precision sintered from 99% alumina ceramic. The groove curvature radius R = 0.6, which is the diameter of the power transmission conductor 204. The surface roughness Ra of the ceramic energy-concentrating groove 207 is ≤ 0.8 μm. The ceramic layer thickness is 2 mm. It is bonded and fixed to the copper tube winding unit by high-temperature epoxy adhesive.

[0138] (7) Magnetic field lines 208 guide and promote the generation of induced eddy currents.

[0139] In one embodiment, the operating logic of the directional sensing de-icing device includes:

[0140] Step 6.1: Based on the location coordinates P of the target location requiring de-icing output by the multi-scale transmission line icing diagnosis network. melt (x,y) and the corresponding icing type P type Calculate the optimal heating power P required for the target location. total And melting time t heat .

[0141] Optimal heating power P total The calculation formula is as follows:

[0142] p total =p base ×L ice ×K temp

[0143] Where, p base Based on power density, L iceThe length of the ice layer is K, which is proportionally converted from the output position coordinates. temp This is the temperature compensation coefficient.

[0144] Base power density p base The calculation formula is as follows:

[0145] p base =R type (d / D wire ) ^ e

[0146] Among them, R type D is the thermal resistance coefficient. wire The parameter d is the diameter of the conductor, which is a fixed reference diameter and can be 20. The parameter e is the default diameter-thermal efficiency attenuation coefficient and can be 1.5.

[0147] Temperature compensation coefficient K temp The calculation formula is as follows:

[0148] K temp =f + g × max(Th) Comp -T env ,0)

[0149] Where, parameter f is the default baseline coefficient, which can be 1; parameter g is the default temperature-power compensation slope, which can be 0.02; and power needs to be increased by 2% for every 1°C decrease in temperature. Comp To compensate for the start-up threshold, T env This is the current temperature.

[0150] Melting time t heat The calculation formula is as follows:

[0151] t heat =max(Th melt ,a×k density )×(D wire / a) ^ b

[0152] Where, k density D is the density coefficient of the icing type. wire Th is the diameter of the transmission line of the induction coil array. melt To ensure the shortest melting time, parameter a is the default unit density calibration factor, which can be 30, and parameter b is the default diameter-heat capacity correlation index, which can be 2.

[0153] Step 6.2: Control the servo mechanism of the induction coil array to precisely align the induction coil array with the target transmission line that has been detected to be covered with ice.

[0154] Step 6.3: Turn on the high-frequency power supply and apply the calculated power p to the induction coil array. total Based on the location coordinates P output by the multi-scale transmission line icing diagnosis network melt (x,y) and the corresponding icing type P type And the melting time t heat It performs induction heating at fixed points and timed intervals.

[0155] The icing diagnosis results output from the multi-scale transmission line icing diagnosis network are input into the directional induction de-icing device. The output position coordinates are converted into actual icing position information proportionally to obtain the icing length. At the same time, the optimal heating power and de-icing time required for the target position are calculated, and the induction coil array is activated to achieve fixed-point and timed induction heating of the target position of the target transmission line icing.

[0156] The technical advantages of directional induction de-icing devices include: The internal quick-release hinge and coil design facilitates flexible opening and closing, making the induction coil array easy to maintain and suitable for transmission lines in icing environments requiring frequent maintenance. Furthermore, the directional induction de-icing device directly induces eddy currents within the target transmission line through high-frequency electromagnetic induction, resulting in a short energy transmission path and low loss. On the other hand, due to the high operating frequency required in icing scenarios, the directional induction de-icing device incorporates a U-shaped magnetic core to ensure high permeability and low loss during high-frequency applications.

[0157] This application establishes a multi-scale transmission line icing diagnosis network with strong generalization, high accuracy, and automation. It can efficiently identify whether icing exists on transmission lines, as well as the type and location coordinates of icing. Simultaneously, based on the design of a directional induction de-icing device, it performs directional and timed de-icing on target transmission lines with icing. After the network training is completed, the de-icing method designed in this application can be deployed on various transmission lines to achieve real-time diagnosis and de-icing of transmission line icing.

[0158] This application provides a transmission line de-icing device based on multi-scale icing diagnosis, characterized in that it includes:

[0159] An image capturing device for acquiring sampled images of transmission lines and images of target transmission lines;

[0160] An icing diagnostic device is used for: annotating sampled images of transmission lines to obtain a first transmission line icing dataset; performing image transformation and data augmentation on the first transmission line icing dataset to obtain a second transmission line icing dataset; training a model based on the second transmission line icing dataset to obtain a multi-scale transmission line icing diagnostic network; and using the multi-scale transmission line icing diagnostic network to perform transmission line icing diagnosis on transmission line images and outputting icing diagnosis results.

[0161] A directional induction de-icing device is used to perform de-icing operations based on icing diagnosis results.

[0162] It should be noted that the icing diagnosis device has the same or similar icing diagnosis process as the transmission line de-icing method based on multi-scale icing diagnosis in the above embodiments, and the directional induction de-icing device has the same or similar structure and principle as the directional induction de-icing device in the transmission line de-icing method based on multi-scale icing diagnosis in the above embodiments, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0163] In one specific embodiment of this application, sampling images of transmission lines containing icing are acquired using an image capturing device, and the data is filtered. To ensure a sufficient amount of data, data containing various types of icing on transmission lines are filtered, and the images are cropped to a size of 640×640 pixels to obtain a first transmission line icing dataset D1{i1,i2,…,i 300 Subsequently, the Label Img image annotation tool was used to annotate the icing targets in all transmission line sample images with icing type and location coordinates: A pre-defined VOC annotation format was used to select and mark the icing areas of the transmission lines, creating bounding boxes and icing type labels. For the icing type labels, this application classifies the transmission line icing types into {no icing, icing and rime, icing and mixed rime, icing and wet snow, and icing and rain rime}, corresponding to type indices {0, 1, 2, 3, 4} respectively. After annotation, a dataset D1 = {i1, i2, ..., i...} is generated, corresponding to the first transmission line icing dataset. 300 The annotation file L1 = {l1, l2, ..., l} corresponds one-to-one with the {l1, l2, ..., l}. 300 The data format is .xml. The labeled dataset D1 = {i1,i2,…,i...} is used to represent the icing dataset of the first transmission line. 300}, L1{l1,l2,…,l 300 Perform data augmentation operations, including data rotation, flipping, cropping, random occlusion, contrast adjustment, random scaling, and noise addition, each affecting D1 = {i1, i2, ..., i...} 300} and L1={l1,l2,…,l 300 The data augmentation operations will generate double the amount of augmented data. Finally, the augmented transmission line icing dataset D2 = {i1, i2, ..., i} is obtained, which includes the original data and all the augmented images. 2400} and L2={l1,l2,…,l 2400}

[0164] After the dataset preprocessing is completed, the enhanced transmission line icing dataset D2 is divided into training data D4, validation data D5, and test data D6 in a 7:2:1 ratio. After the division, a second transmission line icing dataset D7 containing 1680 training data, 480 validation data, and 240 test data can be obtained.

[0165] The dataset after data partitioning is trained according to the model designed in this application: First, the various parameters and hyperparameters of the MTICD-Net model training are initialized. This embodiment explains several important parameters in network training: the corresponding training batch size is initialized according to the hardware environment, the optimizer is initialized to optimizer, the batch size is set to 4, the number of training rounds is initialized to 200, the learning rate type is cosine learning rate, and the initial learning rate is 0.001. These parameters need to be adjusted according to the effect of multiple trainings of the network until the multi-scale transmission line icing diagnosis network converges.

[0166] After initializing the basic parameters for training the multi-scale transmission line icing diagnosis network, model training begins. Within this network, the input second transmission line icing dataset D7 is first converted into sequential data suitable for model processing, yielding the first transmission line icing diagnosis feature map Z1. Z1 is then sequentially input into the model backbone module and the feature focusing and diffusion pyramid to obtain the second transmission line icing diagnosis feature map Z2. Z2 is downsampled to obtain the third transmission line icing diagnosis feature map Z3. Z3 is then sequentially input into the model backbone module and the feature focusing and diffusion pyramid to obtain the fourth transmission line icing diagnosis feature map Z4. Z4 is downsampled to obtain the fifth transmission line icing diagnosis feature map Z5. Z5 is then sequentially input into the model backbone module and the feature focusing and diffusion pyramid to obtain the sixth transmission line icing diagnosis feature map Z6. Z6 is average pooled and then input into the regression and classifier to obtain the seventh transmission line icing diagnosis feature map Z7, which serves as the output of the multi-scale transmission line icing diagnosis network. The output icing diagnosis results contain a tensor of prediction information, with each row corresponding to a prediction, including the location coordinates of the target location that needs to be melted, the icing type label, and the confidence score.

[0167] For training the multi-scale transmission line icing diagnosis network, parameters are updated internally through backpropagation using a loss function. In the loss function described in this application, the location coordinate loss of the icing target on the transmission line uses the DFL loss, the confidence loss of the bounding box uses the WIoU loss, and the classification loss uses the cross-entropy loss. These loss functions are updated epoch-by-epoch during gradient backpropagation during network training, allowing the network to gradually converge and improve detection accuracy. After each training epoch, the network uses the partitioned validation data to validate its accuracy (P), recall (R), and mean precision (mAP), serving as evaluation criteria for the effectiveness of the training. After training the MTICD-Net in this application, 240 partitioned transmission line icing image datasets are used as test data. This test data is input into the optimally trained MTICD-Net to test its performance.

[0168] A multi-scale icing diagnosis-based transmission line de-icing device was deployed. The device includes a multi-scale transmission line icing diagnosis network and a directional induction de-icing device. Images of the target transmission line are periodically captured using an image acquisition device and input into the multi-scale transmission line icing diagnosis network. The network outputs the icing diagnosis result for the target transmission line. This icing diagnosis result contains an information tensor, with each row corresponding to a prediction, including the location coordinates P of the target location requiring de-icing. melt (x, y) (representing the top-left and bottom-right coordinates respectively), icing type label P type And confidence score.

[0169] Subsequently, based on the pre-set confidence score threshold S Comp Filter out those exceeding the confidence score threshold S Comp The icing diagnosis results are input into the directional induction de-icing device. Since the directional induction de-icing device is fixed on the transmission line tower, the output position coordinates can be proportionally converted into actual icing position information P. melt According to P melt Calculate the length of icing L ice In this embodiment, L is set ice =2.4m, assuming environmental parameter T env = -12℃, compensation start-up threshold Th Comp = -15℃, let the current P type The corresponding thermal resistance coefficient R is "rime". type =0.6, wire diameter D wire =36.0mm, from which the basic power density p can be obtained. base =0.6×(20 / 36) ^ 1.5 = 0.284 kW / m, then calculate the temperature compensation coefficient K. temp=1 + 0.02 × max(-15 + 12, 0) = 1, then calculate the heating power p total =0.284×2.4×1=0.682kW, and finally, calculate the ice melting time t. heat Set the shortest melting time to guarantee Th melt The density coefficient k of the rime ice type is 90s. density If t is 4, then t heat =max(90,30×4)×(36 / 30) ^ 2 = 172.8s.

[0170] Finally, an induction coil array is activated to achieve targeted and timed induction heating of the icy target transmission line.

[0171] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0172] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for de-icing a power transmission line based on multi-scale icing diagnosis, characterized in that, The method comprises the following steps: acquiring power line sampling images and labeling the power line sampling images to obtain a first power line icing data set; performing image transformation data enhancement on the first power line icing data set to obtain a second power line icing data set; training a model based on the second power line icing data set to obtain a multi-scale power line icing diagnosis network; acquiring a power line image of a target power line, using the multi-scale power line icing diagnosis network to perform power line icing diagnosis on the power line image, and outputting an icing diagnosis result, wherein the icing diagnosis result comprises the position coordinates of the target position that needs to be deiced, an icing type label and a confidence score; based on the icing diagnosis result, performing deicing operation through a directional induction deicing device.

2. The method of claim 1, wherein the step of applying the electric field to the transmission line is performed by a plurality of electrodes. The labeling of the power line sampling images comprises: labeling the icing type and position coordinates of the icing target of the power line sampling images, wherein the icing type comprises: no icing, icing and rime, icing and mixed rime, icing and wet snow, and icing and glaze.

3. The method of claim 2, wherein, The image transformation data enhancement on the first power line icing data set to obtain a second power line icing data set comprises: performing image transformation data enhancement on any one of the power line sampling images in the first power line icing data set, and using the images after the image transformation data enhancement and the first power line icing data set to constitute an enhanced power line icing data set, wherein the image transformation data enhancement mode comprises at least one of the following: rotation, flipping, cropping, random occlusion, contrast adjustment, random scaling and noise addition; performing random data division of training data, verification data and test data according to a preset proportion on the enhanced power line icing data set to obtain the second power line icing data set containing the training data, the verification data and the test data; training a model based on the second power line icing data set to obtain a multi-scale power line icing diagnosis network comprises: inputting the training data and the verification data into the multi-scale power line icing diagnosis network for training until the network loss function converges; testing the trained multi-scale power line icing diagnosis network using the test data, and if the accuracy of the test result is greater than or equal to a preset accuracy threshold, the multi-scale power line icing diagnosis network is obtained.

4. The method of claim 1, wherein, The multi-scale power line icing diagnosis network comprises an image block embedding layer, a model backbone module, a feature focusing diffusion pyramid and an average pooling layer, and a regression classifier, and the model backbone module comprises a multi-scale scene feature interaction module. The initial image is input into the image block embedding layer to convert the initial image into serialized data, to obtain a first power line icing diagnosis feature map Z1, and the first power line icing diagnosis feature map Z1 is sequentially input into the model backbone module and the feature focusing diffusion pyramid to obtain a second power line icing diagnosis feature map Z2; wherein the initial image is the second power line icing data set or the power line image; The second power line icing diagnosis feature map Z2 is down-sampled to obtain a third power line icing diagnosis feature map Z3; The third power line icing diagnosis feature map Z3 is sequentially input into the model backbone module and the feature focusing diffusion pyramid to obtain a fourth power line icing diagnosis feature map Z4; The fourth power line icing diagnosis feature map Z4 is down-sampled to obtain a fifth power line icing diagnosis feature map Z5, and the fifth power line icing diagnosis feature map Z5 is sequentially input into the model backbone module and the feature focusing diffusion pyramid to obtain a sixth power line icing diagnosis feature map Z6; The sixth power line icing diagnosis feature map Z6 is averaged by the average pooling layer and then input into the regression classifier to obtain a seventh power line icing diagnosis feature map Z7, and the seventh power line icing diagnosis feature map Z7 is taken as an icing diagnosis result.

5. The method of claim 4, wherein, The model backbone module comprises a first DW convolution and batch normalization layer, a first convolution layer, a first tensor sum unit, the multi-scale scene feature interaction module, a second tensor sum unit, a second DW convolution and batch normalization layer, a second convolution layer, and a third tensor sum unit connected in sequence; The first power line icing feature map Y0 is sequentially input into the first DW convolution and batch normalization layer and the first convolution layer to obtain a first scene feature interaction backbone feature map Y1; The first power line icing feature map Y0 and the first scene feature interaction backbone feature map Y1 are added by tensor through the first tensor sum unit to obtain a second scene feature interaction backbone feature map Y2; The second scene feature interaction backbone feature map Y2 is input into the multi-scale scene feature interaction module to obtain a third scene feature interaction backbone feature map Y3, and the second scene feature interaction backbone feature map Y2 and the third scene feature interaction backbone feature map Y3 are added by tensor through the second tensor sum unit to obtain a fourth scene feature interaction backbone feature map Y4; The fourth scene feature interaction backbone feature map Y4 is sequentially input into the second DW convolution and batch normalization layer and the second convolution layer to obtain a fifth scene feature interaction backbone feature map Y5; The fourth scene feature interaction backbone feature map Y4 and the fifth scene feature interaction backbone feature map Y5 are added by tensor through the third tensor sum unit to obtain a sixth scene feature interaction backbone feature map Y6, and the sixth scene feature interaction backbone feature map Y6 is the output of the model backbone module.

6. The method of claim 4, wherein, The multi-scale scene feature interaction module comprises a channel segmentation unit, a multi-core deep convolution layer, a long-range information enhancement variable unit and a channel splicing unit; The second transmission line icing feature map X0 is segmented into three equal parts along the channel dimension by the channel segmentation unit, and first, second and third scene feature interaction feature maps X1, X2 and X3 are obtained respectively; The first scene feature interaction feature map X1 is input into the multi-core deep convolution layer to enhance the perception ability of different receptive fields, and a fourth scene feature interaction feature map X4 is obtained; The second scene feature interaction feature map X2 is subjected to long-range information enhancement transformation by the long-range information enhancement variable unit, and a fifth scene feature interaction feature map X5 is obtained; The third, fourth and fifth scene feature interaction feature maps X3, X4 and X5 are spliced along the channel dimension by the channel splicing unit, and a sixth scene feature interaction feature map X6 is obtained, which is the output of the multi-scale scene feature interaction module.

7. The method of claim 4, wherein, The feature focusing diffusion pyramid comprises a first branch, a second branch, a third branch, a channel merging unit, a fourth branch, a fifth branch, a sixth branch, a seventh branch, a fourth tensor summation unit, a point-by-point convolution unit and a fifth tensor summation unit; The third transmission line icing feature map F0 is input into the first, second and third branches respectively, the first branch sequentially performs upsampling and convolution operation to obtain a first focusing diffusion feature map F1, the second branch performs convolution operation to obtain a second focusing diffusion feature map F2, and the third branch performs adaptive downsampling operation to obtain a third focusing diffusion feature map F3; The first, second and third focusing diffusion feature maps F1, F2 and F3 are merged along the channel dimension by the channel merging unit to obtain a fourth focusing diffusion feature map F4; The fourth focusing diffusion feature map F4 is input into the fourth, fifth, sixth and seventh branches respectively, a fifth focusing diffusion feature map F5 is obtained by performing deep convolution with a convolution kernel of 5x5 in the fourth branch, a sixth focusing diffusion feature map F6 is obtained by performing deep convolution with a convolution kernel of 7x7 in the fifth branch, a seventh focusing diffusion feature map F7 is obtained by performing deep convolution with a convolution kernel of 9x9 in the sixth branch, and an eighth focusing diffusion feature map F8 is obtained by performing deep convolution with a convolution kernel of 11x11 in the seventh branch; The fourth, fifth, sixth, seventh and eighth focusing diffusion feature maps F4, F5, F6, F7 and F8 are added by tensor addition by the fourth tensor summation unit to obtain a ninth focusing diffusion feature map F9. The ninth focus diffusion feature map F9 is input into the point-by-point convolution unit for feature extraction, and a tenth focus diffusion feature map F 10 ; By the fifth tensor summing unit, the third transmission line icing feature map F0 and the tenth focused diffusion feature map F 10 Tensor addition is performed to obtain an eleventh focused diffusion feature map F 11 , the eleventh focused diffusion feature map F 11 is the output of the feature focused diffusion pyramid.

8. The method according to any one of claims 1 to 7, characterized in that, The directional induction ice melting device is used for calculating ice melting parameters according to the ice coating diagnosis result, generating eddy current in the target power transmission line by using the high-frequency electromagnetic induction principle, making the target power transmission line heat itself, and thus melting the surface ice coating of the target power transmission line. The directional induction ice melting device comprises a servo mechanism and an induction coil array, and the servo mechanism is used for driving the induction coil array to move. The induction coil array comprises a hinge and a coil, and the coil comprises a coil shell, a power transmission conductor, a copper pipe winding unit, a U-shaped magnetic core magnetic conductor, a ceramic energy-gathering groove, and a magnetic force line.

9. The method of claim 8, wherein, The copper pipe winding unit comprises a plurality of turns of purple copper pipe. based on the position coordinates and the ice type label, calculate the heating power P required for ice melting total and ice melting time; The innermost turn of the purple copper pipe has a fixed distance from the surface of the power transmission conductor. applying the heating power P to the array of induction coils total induction heating the target power transmission line according to the position coordinates and the ice melting time; Wherein, the optimal heating power P total The calculation formula is: p total = p base x L ice x K temp p base is the base power density, L ice is the ice length, K temp is the temperature compensation coefficient; The ice melting time t heat The calculation formula is: t heat = max(Th melt ,a×k density ) × (D wire / a)^b wherein k density is a density coefficient of ice type, D wire is a wire diameter of the power transmission wire of the induction coil array, Th melt is a shortest ice-melting time guarantee, a is a unit density calibration factor, and b is a diameter-thermal capacity correlation index.

10. A power transmission line ice-melting device based on multi-scale icing diagnosis, characterized in that, The U-shaped magnetic core magnetic conductor covers the power transmission conductor. The ceramic energy-gathering groove is fixedly connected with the copper pipe winding unit by epoxy adhesive. The magnetic force line is used for guiding and promoting the generation of induced eddy current. The directional induction ice melting device is particularly used for: moving the induction coil array to align with the target power transmission line by using the servo mechanism; comprising: an image shooting device used for acquiring a power transmission line sampling image and a power transmission line image of a target power transmission line; an ice coating diagnosis device used for labeling the power transmission line sampling image to obtain a first power transmission line ice coating data set; performing image transformation data enhancement on the first power transmission line ice coating data set to obtain a second power transmission line ice coating data set; training a model based on the second power transmission line ice coating data set to obtain a multi-scale power transmission line ice coating diagnosis network; using the multi-scale power transmission line ice coating diagnosis network to perform power transmission line ice coating diagnosis on the power transmission line image to output an ice coating diagnosis result; and a directional induction ice melting device used for performing ice melting operation based on the ice coating diagnosis result.

Images