An icing detection method, device and storage medium for a transmission line

Abstract

An ice coating detection method, device and storage medium for a transmission line according to an embodiment of the present invention include: obtaining a first downsampling result corresponding to a current ice coating image, determining a transformation matrix according to the number of channels of the first downsampling result, and performing a set number of matrix transformations on the first downsampling result using the transformation matrix to obtain matrix transformation results respectively corresponding to each matrix transformation; performing matrix transposition on a target matrix transformation result among the matrix transformation results, multiplying the transposed matrix by the remaining matrix transformation results to obtain a current feature extraction result; obtaining a first upsampling result corresponding to the current feature extraction result, and performing downsampling a plurality of preset times on the first upsampling result according to a second convolutional kernel to obtain spare feature maps respectively corresponding to the preset times; splicing the spare feature maps, and determining an ice coating area of the transmission line according to the splicing result, thereby improving the accuracy and efficiency of detecting the ice coating area.

Description

Technical Field

[0001] The present invention relates to the technical field of power transmission engineering, and particularly to a method, device, and storage medium for detecting icing on transmission lines. Background Art

[0002] Icing on transmission lines is likely to occur in mid-high altitude areas and mountainous regions with high humidity and low temperatures in winter. Severe icing can lead to a sharp decline in the mechanical and electrical performance of transmission lines, causing icing disasters such as insulator flashover, line tripping, wire breakage, tower collapse, conductor galloping, and communication interruption, resulting in large-scale power outages. Therefore, in order to ensure the stable operation of the power grid, it is necessary to monitor the icing condition of transmission lines.

[0003] In the prior art, an edge detection algorithm is usually used to detect the edges of the icing image of the transmission line, and then the icing area in the image is determined based on the edge detection result.

[0004] However, since the prior art relies more on local features, when the resolution of the icing image of the transmission line is low and the shape of the icing area is irregular, problems such as poor handling of icing edge details and failure to detect small icing areas are likely to occur, thus affecting the accuracy of the detection result. Summary of the Invention

[0005] The present invention provides a method, device, equipment, and storage medium for detecting icing on transmission lines, which can improve the accuracy and efficiency of detecting the icing area.

[0006] In a first aspect, an embodiment of the present invention provides a method for detecting icing on a transmission line, including:

[0007] Obtain the current icing image of the transmission line, and perform downsampling on the current icing image according to a predefined first convolution kernel and a second convolution kernel to obtain a first downsampling result;

[0008] Obtain the number of channels corresponding to the first downsampling result, determine the number of rows and columns of the transformation matrix according to the number of channels, and use the transformation matrix to perform a set number of matrix transformations on the first downsampling result to obtain matrix transformation results corresponding to each matrix transformation respectively;

[0009] Determine a target matrix transformation result among the matrix transformation results, perform matrix transposition on the target matrix transformation result to obtain a transposed matrix, multiply the transposed matrix by the remaining matrix transformation results in the matrix transformation results to obtain a current feature extraction result;

[0010] Perform upsampling on the current feature extraction result according to a set number of first convolution kernels and second convolution kernels to obtain a first upsampling result;

[0011] According to the second convolutional kernel, perform downsampling on the first upsampling result for the first preset number of times, the second preset number of times, and the third preset number of times respectively to obtain a first spare feature map, a second spare feature map, and a third spare feature map;

[0012] Stitch the first spare feature map, the second spare feature map, and the third spare feature map, and determine the icing area of the transmission line according to the stitching result.

[0013] Optionally, obtaining the current icing image of the transmission line includes: obtaining a multi-channel original icing image, and dividing the original icing image in each channel into blocks according to a set size; flattening the block result of the original icing image along the channel dimension direction to obtain the current icing image; wherein, the number of channels corresponding to the current icing image is greater than or equal to the number of channels corresponding to the original icing image.

[0014] Optionally, perform downsampling on the current icing image according to a predefined first convolutional kernel and a second convolutional kernel to obtain a first downsampling result, including: obtaining the current number of channels corresponding to the current icing image, and performing convolution on the current icing image using the first convolutional kernel with the current number of channels to obtain a first feature map; determining a target number of channels according to the current number of channels and a set value, and performing downsampling on the current icing image using the second convolutional kernel with the target number of channels to obtain a second feature map; stitching the first feature map and the second feature map to obtain the first downsampling result.

[0015] Optionally, after multiplying the transposed matrix by the remaining matrix transformation result in the matrix transformation result to obtain the current feature extraction result, it further includes: performing pooling on the current feature extraction result to obtain a target pooling result, and performing dimensionality reduction processing on the target pooling result according to a predefined third convolutional kernel to obtain a target processing result; using a predefined activation function to process the target processing result to obtain an attention weight corresponding to the current feature extraction result; determining a feature enhancement result according to the attention weight and the current feature extraction result; using the first convolutional kernel and the second convolutional kernel to perform downsampling on the current feature extraction result to obtain a second downsampling result; using a transformation matrix to perform feature extraction on the second downsampling result to obtain a target feature extraction result; stitching the feature enhancement result and the target feature extraction result to obtain an updated current feature extraction result.

[0016] Optionally, performing pooling on the current feature extraction result to obtain a target pooling result includes: performing maximum pooling and hybrid pooling on the current feature extraction result respectively to obtain a maximum pooling result and a hybrid pooling result; stitching the maximum pooling result and the hybrid pooling result to obtain the target pooling result.

[0017] Optionally, splice the first standby feature map, the second standby feature map, and the third standby feature map, and determine the icing area of the transmission line according to the splicing result, including: performing convolution on the first standby feature map using a second convolution kernel to obtain a first feature map to be spliced; performing upsampling on the second standby feature map and the third standby feature map respectively using the second convolution kernel to obtain a second feature map to be spliced corresponding to the second standby feature map and a third feature map to be spliced corresponding to the third standby feature map; splicing the first feature map to be spliced, the second feature map to be spliced, and the third feature map to be spliced, and performing convolution on the splicing result using the second convolution kernel to obtain a convolution feature map; performing upsampling on the convolution feature map using a bilinear interpolation algorithm to obtain an icing detection feature map; determining the icing area of the transmission line according to the pixel point values in the icing detection feature map; wherein, the pixel point values in the icing area are different from the pixel point values in the non-icing area.

[0018] Optionally, perform upsampling on the current feature extraction result according to a set number of first convolution kernels and second convolution kernels to obtain a first upsampling result, including: inputting the current feature extraction result into a set number of convolution modules for convolution, and performing dimensionality reduction processing on the convolution result using a first convolution kernel to obtain a dimensionality reduction feature map; wherein, the convolution module includes a second convolution kernel, a preset normalization layer, and a preset activation function; performing upsampling on the dimensionality reduction feature map using a predefined upsampling module to obtain a size increase result, and performing dimensionality increase processing on the size increase result using a first convolution kernel to obtain a first upsampling result; wherein, the upsampling module is used to increase the size of the dimensionality reduction feature map on the basis of not reducing the number of channels of the dimensionality reduction feature map.

[0019] Optionally, after performing upsampling on the current feature extraction result to obtain a first upsampling result, it further includes: performing upsampling on the first upsampling result according to a set number of first convolution kernels and second convolution kernels to obtain a second upsampling result; inputting the second upsampling result into a channel expansion module for upsampling to obtain a channel expansion result, and performing convolution on the channel expansion result using a second convolution kernel to obtain a feature map to be detected, and using the feature map to be detected as the updated first upsampling result; wherein, the channel expansion module is used to increase the number of channels corresponding to the second upsampling result and reduce the size corresponding to the second upsampling result.

[0020] In a second aspect, an embodiment of the present invention further provides an electronic device, which includes:

[0021] At least one processor; and

[0022] A memory communicatively connected to at least one processor; wherein,

[0023] The memory stores a computer program that can be executed by at least one processor. When the computer program is executed by the at least one processor, the at least one processor is enabled to execute the icing detection method for a transmission line provided in any embodiment of the present invention.

[0024] In a third aspect, an embodiment of the present invention further provides a computer-readable storage medium. The computer-readable storage medium stores computer instructions for causing a processor to implement the icing detection method for a transmission line in any embodiment of the present invention when executed.

[0025] The technical solution provided by the embodiment of the present invention avoids the problem of inaccurate icing area detection caused by detecting the icing area based on the edge detection result of the transmission line icing image by downsampling and feature extraction of the current icing image of the transmission line and upsampling and icing area detection of the feature extraction result, and improves the accuracy of the icing area detection result.

[0026] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. Description of the Drawings

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0028] Figure 1 is a flowchart of an icing detection method for a transmission line according to Embodiment 1 of the present invention;

[0029] Figure 2 is a flowchart of a matrix transformation method according to an embodiment of the present invention;

[0030] Figure 3 is a schematic structural diagram of a model for detecting the icing area of a transmission line according to Embodiment 1 of the present invention;

[0031] Figure 4 is a flowchart of another icing detection method for a transmission line according to Embodiment 2 of the present invention;

[0032] Figure 5 is a flowchart of a method for determining a current icing image according to an embodiment of the present invention;

[0033] Figure 6It is a flowchart of a method for determining a downsampling result provided by an embodiment of the present invention;

[0034] Figure 7 It is a flowchart of a method for determining a feature enhancement result provided by an embodiment of the present invention;

[0035] Figure 8 It is a flowchart of a method for determining an upsampling result provided by an embodiment of the present invention;

[0036] Figure 9 It is a flowchart of a method for detecting an icing area provided by an embodiment of the present invention;

[0037] Figure 10 It is a schematic structural diagram of another model for detecting an icing area of a transmission line provided by Embodiment II of the present invention;

[0038] Figure 11 It is a schematic structural diagram of a downsampling and feature extraction network provided by an embodiment of the present invention;

[0039] Figure 12 It is a schematic structural diagram of an icing detection device for a transmission line provided by Embodiment III of the present invention;

[0040] Figure 13 It is a schematic structural diagram of an electronic device provided by Embodiment IV of the present invention. Detailed implementation manners

[0041] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0042] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily need to be limited to those clearly listed steps or units, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0043] Embodiment 1

[0044] Figure 1 It is a flowchart of an icing detection method for a transmission line provided according to Embodiment 1 of the present invention. This embodiment is applicable to the situation of detecting the icing condition of a transmission line. This method can be executed by an icing detection device of the transmission line, and the icing detection device of the transmission line can be implemented in the form of hardware and / or software, and the icing detection device of the transmission line can be configured in an electronic device such as a computer.

[0045] As Figure 1 shown, an icing detection method for a transmission line disclosed in this embodiment includes:

[0046] S110. Obtain the current icing image of the transmission line, and perform downsampling on the current icing image according to a predefined first convolution kernel and a second convolution kernel to obtain a first downsampling result.

[0047] In this embodiment, the transmission line can be a line for transmitting electric energy, such as an overhead transmission line of a power grid. The current icing image can be an image containing the transmission line. The first convolution kernel and the second convolution kernel can be convolution kernels with different sizes. For example, the first convolution kernel can be a 1*1 convolution kernel, and the second convolution kernel can be a 3*3 convolution kernel.

[0048] In this step, specifically, the current icing image can be convolved using the first convolution kernel, and the convolution result can be downsampled using the second convolution kernel to obtain a first downsampling result. Or, the current icing image can be convolved using the first convolution kernel, the current icing image can be downsampled using the second convolution kernel, and the convolution result and the downsampling result can be concatenated to obtain a first downsampling result. Among them, the number of channels corresponding to the current icing image can be multiple.

[0049] S120. Obtain the number of channels corresponding to the first downsampling result, determine the number of rows and columns of the transformation matrix according to the number of channels, and use the transformation matrix to perform a set number of matrix transformations on the first downsampling result to obtain matrix transformation results corresponding to each matrix transformation respectively.

[0050] In this embodiment, the number of rows and columns of the transformation matrix is the same as the number of channels. The matrix transformation result can be the multiplication result of the first downsampling result and the transformation matrix.

[0051] In this step, specifically, the number of transformation matrices can be determined according to the set number, and each transformation matrix is multiplied by the first downsampling result respectively to obtain a set number of matrix transformation results.

[0052] Figure 2It is a flowchart of a matrix transformation method provided according to an embodiment of the present invention.

[0053] Exemplarily, as Figure 2 shown, a first downsampling result with dimensions of H×W×C can be obtained and the first downsampling result is multiplied by three C×C transformation matrices respectively to obtain three matrix transformation results with dimensions of (H×W)×C (i.e., matrices D, E, and F).

[0054] S130. Determine a target matrix transformation result among the matrix transformation results, perform matrix transposition on the target matrix transformation result to obtain a transposed matrix, and multiply the transposed matrix by the remaining matrix transformation results in the matrix transformation results to obtain a current feature extraction result.

[0055] In this embodiment, the target matrix transformation result can be any one of the matrix transformation results.

[0056] In this step, specifically, the transposed matrix can be successively multiplied by each of the remaining matrix transformation results in the matrix transformation results, the multiplication result is projected onto a low-dimensional vector space, and the mapping result is used as the current feature extraction result.

[0057] Continuing with the above example, as Figure 2 shown, matrix E can be transposed to obtain E-1, and matrix E-1 is multiplied by matrix D to obtain a matrix G with dimensions of (H×W)×(H×W). Then, matrix G can be multiplied by matrix F to obtain a matrix H with dimensions of (H×W)×C. Finally, matrix H can be projected (i.e., proj mapping) to obtain the current feature extraction result with dimensions still being H×W×C.

[0058] Specifically, first, the matrix transformation result can be determined by the following specific calculation formula:

[0059]

[0060] where F4 is the first downsampling result, P, Q, and R are all transformation matrices, and D, E, and F are matrix transformation results.

[0061] Then, the current feature extraction result can be determined by the following specific calculation formula:

[0062]

[0063] where E -1 is the transposed matrix, G is the result of multiplying E -1 by D, H is the result of multiplying G by F, F 5 is the current feature extraction result, and PROJ() is a projection mapping function.

[0064] The advantage of such a setting is that, compared with the prior art that extracts features from the first downsampling result through multiple convolutional layers, the technical solution of this embodiment performs a matrix transformation on the first downsampling result through a transformation matrix, and performs a transpose and multiplication operation on the matrix transformation result to obtain the current feature extraction result, reducing the complexity of feature extraction, reducing the amount of calculation, and improving the efficiency of feature extraction.

[0065] S140. Upsample the current feature extraction result according to a set number of first convolutional kernels and second convolutional kernels to obtain a first upsampling result.

[0066] In this step, specifically, after performing convolutional processing on the current feature extraction result using a set number of first convolutional kernels, perform upsampling on the current feature extraction result using a set number of second convolutional kernels to obtain a first upsampling result.

[0067] S150. Downsample the first upsampling result by a first preset number of times, a second preset number of times, and a third preset number of times respectively according to the second convolutional kernel to obtain a first backup feature map, a second backup feature map, and a third backup feature map.

[0068] In this embodiment, the first backup feature map may be a feature map obtained by downsampling the first upsampling result by a first preset number of times. The second backup feature map may be a feature map obtained by downsampling the first upsampling result by a second preset number of times. The third backup feature map may be a feature map obtained by downsampling the first upsampling result by a third preset number of times.

[0069] In this step, specifically, the second convolutional kernel may be used to downsample the first upsampling result to obtain an updated first upsampling result. Then, the second convolutional kernel may be used to downsample the updated first upsampling result by a first preset number of times, a second preset number of times, and a third preset number of times respectively.

[0070] S160. Concatenate the first backup feature map, the second backup feature map, and the third backup feature map, and determine the icing area of the transmission line according to the concatenation result.

[0071] In this embodiment, the number of channels corresponding to the concatenation result is equal to the sum of the number of channels corresponding to the first backup feature map, the second backup feature map, and the third backup feature map.

[0072] In this step, specifically, the second convolutional kernel may be used to perform an upsampling operation on the concatenation result, and perform a pooling operation or a bilinear interpolation operation on the upsampling result of the concatenation result, and determine the icing area according to the operation result.

[0073] Figure 3It is a schematic structural diagram of a model for detecting icing areas of a transmission line according to Embodiment 1 of the present invention. As a specific implementation, as Figure 3 shown, first, the current icing image of the transmission line is input into the downsampling and feature extraction network. The downsampling and feature extraction network downsamples the current icing image according to a predefined first convolution kernel and a second convolution kernel to obtain a first downsampling result, and extracts features from the first downsampling result to obtain the current feature extraction result. Secondly, the current feature extraction result can be input into the upsampling network. The upsampling network upsamples the current feature extraction result according to a set number of first convolution kernels and second convolution kernels to obtain a first upsampling result. Finally, the first upsampling result can be input into the icing area detection network. The icing area detection network downsamples the first upsampling result a different number of preset times, and splices the first backup feature map, the second backup feature map, and the third backup feature map obtained by the downsampling, and determines the icing area of the transmission line according to the splicing result.

[0074] The technical solution of the embodiment of the present invention downsamples the current icing image according to a predefined first convolution kernel and a second convolution kernel to obtain a first downsampling result; determines the number of rows and columns of the transformation matrix according to the number of channels of the first downsampling result, and performs a set number of matrix transformations on the first downsampling result using the transformation matrix to obtain matrix transformation results corresponding to each matrix transformation respectively; transposes the target matrix transformation result among the matrix transformation results to obtain a transposed matrix, multiplies the transposed matrix by the remaining matrix transformation results in the matrix transformation results to obtain the current feature extraction result; upsamples the current feature extraction result according to a set number of first convolution kernels and second convolution kernels to obtain a first upsampling result; respectively downsamples the first upsampling result a first preset number of times, a second preset number of times, and a third preset number of times according to the second convolution kernel to obtain a first backup feature map, a second backup feature map, and a third backup feature map; splices the first backup feature map, the second backup feature map, and the third backup feature map, and determines the icing area of the transmission line according to the splicing result. This technical means solves the problem that in the prior art, when determining the icing area based on the edge detection result, it is easy to have poor handling of the icing edge details, resulting in difficulty in detecting small icing areas, and improves the accuracy of detecting the icing area.

[0075] Embodiment 2

[0076] Figure 4 It is a flowchart of another icing detection method for a transmission line according to Embodiment 2 of the present invention. This embodiment is a further optimization and expansion based on the above embodiments and can be combined with each optional technical solution in the above implementation manners.

[0077] As Figure 4 shown, a method for detecting ice coating on a transmission line disclosed in this embodiment includes:

[0078] S210. Obtain the current ice-coated image of the transmission line and the corresponding current number of channels, and perform convolution on the current ice-coated image using a first convolutional kernel with the current number of channels to obtain a first feature map.

[0079] In this embodiment, the number of channels corresponding to the first feature map is the same as the current number of channels.

[0080] In this step, specifically, the multi-channel original ice-coated image can be used as the current ice-coated image. Alternatively, the original ice-coated image can be divided into blocks according to a set size for each channel, and the divided results of the original ice-coated image are flattened along the channel dimension direction to obtain the current ice-coated image. Among them, the number of channels corresponding to the current ice-coated image is greater than or equal to the number of channels corresponding to the original ice-coated image, and the source of the original ice-coated image can be a network, a handheld camera, a drone, etc.

[0081] Figure 5 is a flowchart of a method for determining a current ice-coated image provided according to an embodiment of the present invention, Figure 6 is a flowchart of a method for determining a downsampling result provided according to an embodiment of the present invention.

[0082] Exemplarily, as Figure 5 shown, an original ice-coated image with a dimension of 4×4×3 can be obtained, and the original ice-coated image for each channel is divided into blocks according to a size of 2×2 to obtain a current ice-coated image with a dimension of 2×2×12 (that is, the size of the original ice-coated image is reduced to 1 / 2 of the original, and the number of channels is expanded to four times the original). By dividing the original ice-coated image into blocks and flattening the divided results along the channel direction, the number of channels can be increased, and thus the number of extracted features can be increased, thereby improving the accuracy of detecting the ice-coated area.

[0083] As Figure 6 shown, a current ice-coated image with a dimension of H×W×C can be obtained, and the current number of channels is determined to be C. Then, C first convolutional kernels can be used to perform convolution on the current ice-coated image to obtain a first feature map with a dimension of H×W×C.

[0084] S220. Determine the target number of channels according to the current number of channels and a set value, and perform downsampling on the current ice-coated image using a second convolutional kernel with the target number of channels to obtain a second feature map, and splice the first feature map and the second feature map to obtain a first downsampling result.

[0085] Among them, the number of target channels can be obtained by multiplying the current number of channels by a set value.

[0086] Continuing with the above example, as Figure 6 shown, assuming the set value is 2, then the number of target channels can be determined to be 2C. Then, 2C second convolutional kernels can be used to downsample the current icing image to obtain a second feature map with dimensions H×W×2C. Finally, the first feature map and the second feature map can be concatenated to obtain a first downsampling result with a channel dimension of H×W×3C.

[0087] Specifically, the first downsampling result can be determined through the following specific calculation formula:

[0088]

[0089] Among them, F2 is the current icing image, is the first feature map, is the second feature map, and F3 is the first downsampling result.

[0090] S230. Obtain the number of channels corresponding to the first downsampling result, determine the size of the transformation matrix according to the number of channels, and use the transformation matrix to perform a set number of matrix transformations on the first downsampling result to obtain matrix transformation results corresponding to each matrix transformation respectively.

[0091] S240. Determine the target matrix transformation result among the matrix transformation results, and perform matrix transposition on the target matrix transformation result to obtain a transposed matrix. Multiply the transposed matrix by the remaining matrix transformation results in the matrix transformation results to obtain the current feature extraction result.

[0092] S250. Perform pooling on the current feature extraction result to obtain a target pooling result, and perform dimensionality reduction processing on the target pooling result according to a predefined third convolutional kernel to obtain a target processing result.

[0093] In this embodiment, the target processing result can be the result obtained after performing dimensionality reduction processing on the target pooling result. The size of the third convolutional kernel can be various, such as 7*7.

[0094] In this step, specifically, maximum pooling can be performed on the current feature extraction result, and the maximum pooling result can be used as the target pooling result. Or, mixed pooling can be performed on the current feature extraction result, and the mixed pooling result can be used as the target pooling result. Or, maximum pooling and mixed pooling can be respectively performed on the current feature extraction result to obtain a maximum pooling result and a mixed pooling result, and the maximum pooling result and the mixed pooling result can be concatenated to obtain the target pooling result.

[0095] Figure 7It is a flowchart of a method for determining a feature enhancement result provided according to an embodiment of the present invention.

[0096] Exemplarily, as Figure 7 shown, the current feature extraction result with dimensions of H×W×C can be subjected to max pooling and hybrid pooling respectively to obtain a max pooling result and a hybrid pooling result both with dimensions of H×W×1. Then, the max pooling result and the hybrid pooling result can be concatenated to obtain a target pooling result with dimensions of H×W×2. After that, a 7*7 convolutional kernel can be used to perform dimensionality reduction processing on the target pooling result to obtain a target processing result with dimensions of H×W×1.

[0097] S260. Process the target processing result using a predefined activation function to obtain an attention weight corresponding to the current feature extraction result, and determine the feature enhancement result according to the attention weight and the current feature extraction result.

[0098] In this embodiment, there can be various types of activation functions, such as the Sigmoid function, the Relu function, and the Tanh function, etc. The attention weight can be any value between 0 and 1.

[0099] Continuing with the above example, the Sigmoid function can be used to process the target processing result to obtain the attention weight. Among them, the size of the attention weight is H×W×1. Then, the attention weight and the current feature extraction result can be multiplied to obtain the feature enhancement result.

[0100] Specifically, the attention weight can be determined through the following specific calculation formula:

[0101] Q M =σ(Cov 7×7 (Conc(MixPool(F 7 ),MaxPool(F 7 )))

[0102] where Q M is the attention weight, σ is the Sigmoid function, Cov 7×7 is a 7*7 convolutional operation, MixPool is hybrid pooling, MaxPool is max pooling, Conc is a feature map concatenation operation, and F 7 is the current feature extraction result.

[0103] Then, the feature enhancement result can be determined through the following specific calculation formula:

[0104] F 8 =Q M ×F 7

[0105] where F8 is the result of feature enhancement.

[0106] The advantage of this setting is that by determining the attention weights corresponding to the current feature extraction result and determining the feature enhancement result based on the attention weights and the current feature extraction result, it is possible to focus on the feature regions that contribute greatly to the icing detection task and suppress the feature regions with small contributions, thereby improving the efficiency and accuracy of detecting the icing region.

[0107] S270. Use the first convolutional kernel and the second convolutional kernel to downsample the current feature extraction result to obtain a second downsampled result, and use a transformation matrix to perform feature extraction on the second downsampled result to obtain a target feature extraction result.

[0108] In this step, specifically, the number of channels corresponding to the first downsampled result can be obtained, and the first convolutional kernel with the above number of channels is used to perform convolution on the first downsampled result to obtain a third feature map. Then, the above number of channels can be updated according to a set value, and the second convolutional kernel with the updated number of channels is used to downsample the first downsampled result to obtain a fourth feature map. After that, the third feature map and the fourth feature map can be concatenated to obtain a second downsampled result. Finally, the second downsampled result can be subjected to a set number of matrix transformations using the transformation matrix to obtain matrix transformation results corresponding to each matrix transformation respectively, and the matrix transformation results are transposed and multiplied to obtain a target feature extraction result.

[0109] S280. Concatenate the feature enhancement result and the target feature extraction result to obtain an updated current feature extraction result, and perform upsampling on the current feature extraction result according to a set number of first convolutional kernels and second convolutional kernels to obtain a first upsampled result.

[0110] In this step, specifically, the current feature extraction result can be input into a set number of convolutional modules for convolution, and the first convolutional kernel is used to perform dimensionality reduction processing on the convolution result to obtain a dimensionality reduction feature map. Among them, the convolutional module includes a second convolutional kernel, a preset normalization layer, and a preset activation function. Preferably, the preset activation function can be a Relu function. Then, a predefined upsampling module can be used to perform upsampling on the dimensionality reduction feature map to obtain a size increase result, and the first convolutional kernel is used to perform dimensionality increase processing on the size increase result to obtain a first upsampled result. Among them, the upsampling module is used to increase the size of the dimensionality reduction feature map on the basis of keeping the number of channels of the dimensionality reduction feature map unchanged.

[0111] Figure 8 is a flowchart of a method for determining an upsampled result according to an embodiment of the present invention.

[0112] Exemplarily, such asFigure 8 As shown, a second convolutional kernel can be used to perform convolution on the current feature extraction result with dimensions of H×W×C. The convolution result is input into a preset normalization layer for normalization, and the ReLU function is used to activate the normalization result to obtain a first convolutional feature map with dimensions of H×W×C. Then, the first convolutional feature map can be input into the next convolutional module to repeat the above operations to obtain a second convolutional feature map with dimensions of H×W×C. After that, a first convolutional kernel can be used to perform dimensionality reduction on the second convolutional feature map to obtain a dimensionality-reduced feature map with dimensions of H×W×C / 4, and the dimensionality-reduced feature map is input into a predefined upsampling module for upsampling to obtain a size-increased result with dimensions of 2H×2W×C / 4. Finally, a first convolutional kernel can be used to perform dimensionality increase on the size-increased result to obtain a first upsampling result with dimensions of 2H×2W×C / 2.

[0113] The advantage of such a setting is that by concatenating the feature enhancement result and the target feature extraction result, an updated current feature extraction result is obtained. While focusing on the feature regions that contribute significantly to the icing detection task, the original feature regions are retained, thereby improving the accuracy and reliability of detecting the icing region.

[0114] In an optional implementation manner of the embodiment of the present invention, after upsampling the current feature extraction result to obtain a first upsampling result, it further includes: performing upsampling on the first upsampling result according to a set number of first convolutional kernels and second convolutional kernels to obtain a second upsampling result; inputting the second upsampling result into a channel expansion module for upsampling to obtain a channel expansion result, and using a second convolutional kernel to perform convolution on the channel expansion result to obtain a feature map to be detected, and using the feature map to be detected as the updated first upsampling result; wherein, the channel expansion module is used to increase the number of channels corresponding to the second upsampling result and reduce the size corresponding to the second upsampling result.

[0115] In a specific implementation manner, the first upsampling result can be input into a set number of convolutional modules for convolution, and a first convolutional kernel is used to perform dimensionality reduction on the convolution result to obtain a new dimensionality-reduced feature map. Then, a predefined upsampling module can be used to perform upsampling on the new dimensionality-reduced feature map to obtain a new size-increased result, and a first convolutional kernel is used to perform dimensionality increase on the new size-increased result to obtain a second upsampling result.

[0116] Exemplarily, the second upsampling result with dimensions of 4H×4W×C / 4 can be input into a channel expansion module for upsampling to obtain a channel expansion result with dimensions of 2H×2W×C. Then, a second convolutional kernel can be used to perform convolution on the channel expansion result to obtain a feature map to be detected with dimensions of 4H×4W×C.

[0117] The advantage of such a setting is that by performing an upsampling operation on the current feature extraction result the same number of times as the downsampling times, the resolution of the feature map for icing area detection can be made consistent with the current icing image.

[0118] S290. According to the second convolution kernel, perform downsampling for the first preset number of times, downsampling for the second preset number of times, and downsampling for the third preset number of times on the first upsampling result respectively to obtain a first standby feature map, a second standby feature map, and a third standby feature map; splice the first standby feature map, the second standby feature map, and the third standby feature map, and determine the icing area of the transmission line according to the splicing result.

[0119] In this step, specifically, the second convolution kernel can be used to perform convolution on the first standby feature map to obtain a first feature map to be spliced, and the second convolution kernel is used to perform upsampling on the second standby feature map and the third standby feature map respectively to obtain a second feature map to be spliced corresponding to the second standby feature map and a third feature map to be spliced corresponding to the third standby feature map. Then, the first feature map to be spliced, the second feature map to be spliced, and the third feature map to be spliced can be spliced, and the second convolution kernel is used to perform convolution on the splicing result to obtain a convolution feature map. After that, the bilinear interpolation algorithm can be used to perform upsampling on the convolution feature map to obtain an icing detection feature map. Finally, the icing area of the transmission line can be determined according to the pixel point values in the icing detection feature map. Among them, the pixel point values in the icing area are different from those in the non-icing area.

[0120] Figure 9 It is a flowchart of an icing area detection method provided by an embodiment of the present invention.

[0121] Exemplarily, such as Figure 9As shown, a 3×3 convolution operation can be performed on the first upsampling result, average pooling is performed on the convolution result through a preset pooling layer, and ReLU function is used to activate the average pooling to obtain the updated first upsampling result (that is, the first downsampling result is updated through the convolution pooling activation module). Then, the updated first upsampling result can be successively fed into three convolution pooling activation modules for downsampling operations to obtain a first spare feature map with dimensions of H / 4×W / 4×4C, a second spare feature map with dimensions of H / 8×W / 8×8C, and a third spare feature map with dimensions of H / 16×W / 16×16C. After that, a convolution operation can be performed on the first spare feature map to obtain a first feature map to be concatenated with unchanged dimensions, an upsampling operation is performed on the second spare feature map to obtain a second feature map to be concatenated with dimensions changed to H / 4×W / 4×2C, and an upsampling operation is performed on the third spare feature map to obtain a second feature map to be concatenated with dimensions changed to H / 4×W / 4×C. The first feature map to be concatenated, the second feature map to be concatenated, and the third feature map to be concatenated can be concatenated to obtain a feature map with dimensions of H / 4×W / 4×7C. Finally, a 3×3 convolution operation and a bilinear interpolation operation can be performed on the feature map with dimensions of H / 4×W / 4×7C above to upsample the above feature map to an ice detection feature map with dimensions of H / 2×W / 2×1. Among them, each pixel value in the ice detection feature map is 0 or 1, where 0 indicates that the pixel at this position is a non-icing area, and 1 indicates that the pixel at this position is an icing area.

[0122] Specifically, the ice detection feature map can be determined by the following specific calculation formula:

[0123] F 17 =Conv 3×3 AvgPoolingReLuF 16 ;

[0124]

[0125]

[0126] F22=BIConv3×3F21.

[0127] Among them, F16 is the first upsampling result, Conv3×3AvgPoolingReLu is the convolution pooling activation module, F17, F18, F19, and F20 are feature maps obtained through the convolution pooling activation module, Conv3×3 is a 3×3 convolution operation, is the convolution upsampling operation, Concat is the concatenation operation, BI is the bilinear interpolation operation, and F22 is the ice detection feature map.

[0128] Figure 10It is a schematic structural diagram of another model for detecting icing areas of transmission lines according to Embodiment 2 of the present invention. Figure 11 It is a schematic structural diagram of a downsampling and feature extraction network according to an embodiment of the present invention.

[0129] As a specific implementation manner, as Figure 10 shown, first, the current icing image of the transmission line is input into the feature map block network. The feature map block network divides the original icing image under each channel into blocks according to the set size, and flattens the block result of the original icing image along the channel dimension direction to obtain the current icing image. Secondly, the current icing image can be input into the first downsampling and feature extraction network to obtain the current feature extraction result, and the current feature extraction result is input into the second downsampling and feature extraction network to obtain the target feature extraction result.

[0130] Taking the downsampling and feature extraction network directly connected to the feature map block network as an example, the processing flow of the downsampling and feature extraction network is described, as Figure 11 shown:

[0131] The current icing image can be input into the downsampling layer. The downsampling layer uses the first convolutional kernel with the current number of channels to perform convolution on the current icing image to obtain the first feature map; according to the current number of channels and the set value, the target number of channels is determined, and the second convolutional kernel with the target number of channels is used to perform downsampling on the current icing image to obtain the second feature map; the first feature map and the second feature map are spliced to obtain the first downsampling result. Then, the first downsampling result can be sequentially input into the normalization layer and the feature extraction layer. The feature extraction layer performs matrix transformation on the first downsampling result for a set number of times to obtain matrix transformation results corresponding to each matrix transformation respectively; the target matrix transformation result is determined among the matrix transformation results, and the target matrix transformation result is transposed to obtain a transposed matrix, and the transposed matrix is multiplied by the remaining matrix transformation results in the matrix transformation results to obtain the current feature extraction result. After that, the first downsampling result and the current feature extraction result are added, and the added result is sequentially input into the normalization layer and the multi-layer perceptron to obtain the updated current feature extraction result.

[0132] Again, the current feature extraction result can be input into the feature enhancement network. The feature enhancement network performs pooling on the current feature extraction result to obtain the target pooling result, and performs dimensionality reduction processing on the target pooling result according to the predefined third convolutional kernel to obtain the target processing result. The predefined activation function is used to process the target processing result to obtain the attention weight corresponding to the current feature extraction result. According to the attention weight and the current feature extraction result, the feature enhancement result is determined.

[0133] The concatenation result of the feature enhancement result and the target feature extraction result can be input into two upsampling networks in sequence. The concatenation result is input into a set number of convolutional modules through the upsampling network for convolution, and the first convolutional kernel is used to perform dimensionality reduction processing on the convolution result to obtain a dimensionality-reduced feature map. The predefined upsampling module is used to perform upsampling on the dimensionality-reduced feature map to obtain a result with increased size, and the first convolutional kernel is used to perform dimensionality increase processing on the result with increased size to obtain an upsampling result. The upsampling result can be input into a channel expansion network. The channel expansion network performs channel expansion on the upsampling result to obtain a channel expansion result, and the second convolutional kernel is used to perform convolution on the channel expansion result to obtain a feature map to be detected.

[0134] Finally, the feature map to be detected can be input into an icing area detection network. The icing area detection network performs downsampling for the first preset number of times, downsampling for the second preset number of times, and downsampling for the third preset number of times on the feature map to be detected respectively to obtain a first standby feature map, a second standby feature map, and a third standby feature map. The first standby feature map, the second standby feature map, and the third standby feature map are concatenated, and the icing area of the transmission line is determined according to the concatenation result.

[0135] Optionally, before detecting the icing area through the model for detecting the icing area of the transmission line, it further includes: using a handheld camera or a drone to take the original icing images of the transmission line at the substation, or obtaining the original icing images under the actual scenario on the network, and forming a dataset with the original icing images obtained from various sources. Then, the above dataset can be subjected to data cleaning. The content of data cleaning includes: modifying the names of the original icing images to a unified format, such as source_category_sequence number; deleting the unclear images in the dataset; using a sampling tool to label the icing areas in the original icing images. Finally, the dataset after data cleaning can be divided into a training set, a validation set, and a test set according to the ratio of 7:2:1, and a model for detecting the icing area of the transmission line can be trained using the training set, the validation set, and the test set. Specifically, the training set can be input into a pre-constructed neural network model for training. During the training process, the number of samples used in each iteration can be set to 2, the number of iterations can be set to 1000 rounds, and the random gradient descent algorithm can be used to complete the backpropagation during the training process, and the Adam optimizer can be used to adaptively adjust the learning rate to enable the pre-constructed neural network model to converge quickly. Then, the validation set can be input into the trained neural network model to obtain the accuracy of the network model for detecting the icing of the transmission line, and the accuracy can be used to evaluate the performance of the network model on the validation set. Among them, the binary cross-entropy loss function can be used to measure the difference between the model prediction result and the true label. After each validation is completed, the hyperparameters (such as the learning rate and regularization parameters, etc.) and the structure of the network model can be adjusted according to the validation result to further optimize the performance of the network model.

[0136] The technical solution of this embodiment can focus on the feature regions that contribute greatly to the icing detection task and suppress the feature regions with small contributions by determining the attention weights corresponding to the current feature extraction results and determining the feature enhancement results according to the attention weights and the current feature extraction results, thereby improving the efficiency and accuracy of detecting the icing area. Secondly, by splicing the feature enhancement result and the target feature extraction result to obtain the updated current feature extraction result, while focusing on the feature regions that contribute greatly to the icing detection task, the original feature regions are retained, further improving the accuracy and reliability of detecting the icing area.

[0137] On the basis of the above embodiments, to illustrate the icing area detection method and effect in this solution in detail, the following is an illustration with a most detailed embodiment:

[0138] Assume that the dimension of the current icing image is 1024×1024×3 (width×height×channels). First, take the current icing image as the feature map A and input it into the feature map partitioning network for partitioning. Divide the feature map A into 2×2 block images in each channel, and then splice them in the channel dimension direction to reduce the feature map size to 1 / 2 of the original and expand the number of channels to 4 times the original, obtaining a feature map B with a dimension of 512×512×12.

[0139] Input the feature map B into the downsampling and feature extraction network for downsampling and feature extraction. First, input the feature map B into the downsampling layer. Perform a 1×1 convolution on the feature map B through the downsampling layer to obtain the feature map B1 (with a dimension of 512×512×12), then perform a 3×3 convolution on the feature map B to obtain the feature map B2 (with a dimension of 512×512×24), and finally splice the feature map B1 and the feature map B2 to obtain the feature map C (with a dimension of 512×512×36). Then input the feature map C into the normalization layer for normalization processing to obtain the feature map D (with a dimension of 512×512×36). Then input the feature map D into the feature extraction layer for self-attention learning to extract features.

[0140] In the feature extraction layer, first multiply the feature map D by three 36×36 transformation matrices respectively to obtain three matrices denoted as D1, D2, and D3 (with a dimension of (512×512)×36). Then transpose the matrix D2 to obtain D21, multiply the matrix D21 and the matrix D1 to obtain a matrix D4 with a dimension of (512×512)×(512×512), then multiply the matrix D4 and the matrix D3 to obtain a matrix D5 with a dimension of (512×512)×36. Finally, perform a proj mapping on the matrix D5 to obtain a feature map E with a dimension still of 512×512×36. Then add the feature map E and the feature map C to obtain a feature map F with a dimension of 512×512×36. Then input the feature map F into the normalization layer and the multi-layer perceptron layer to obtain a feature map G with a dimension still of 512×512×12. Input the feature map G into the downsampling and feature extraction network again for one-time downsampling and feature extraction to obtain a feature map G1 with a dimension of 512×512×12.

[0141] Input the feature map G into the feature enhancement network for feature enhancement. In the feature enhancement network, first perform max pooling and mixed pooling on the feature map G to obtain two feature maps with dimensions of 512×512×1. Then, concatenate the two feature maps to obtain a feature map with dimensions of 512×512×2, and transform it into a feature map with dimensions of 512×512×1 through 7×7 Conv (convolution). Then, pass it through a sigmoid function for activation to obtain the attention weight Qm of the feature map G, with a size of 512×512×1. Finally, multiply the attention weight Qm by the feature map G to obtain a feature map G2 with dimensions of 512×512×12. Finally, add the feature map G2 after feature enhancement to the feature map G1 to obtain a feature map H with dimensions of 512×512×12.

[0142] Input the feature map H after feature enhancement into the upsampling network for upsampling operations to double the input feature size and reduce the number of feature channels to 1 / 2 of the original. First, perform a 3×3 convolution on the feature map H, input the result into the preset pooling layer for normalization and activate it using the ReLu function to obtain the feature map I (with dimensions of 512×512×12). Repeat the above operations, perform another 3×3 convolution, pass through the preset pooling layer, and activate it through the ReLu function to obtain the feature map O (with dimensions of 512×512×12). Then, use a 1×1 convolution operation to reduce the dimension of the feature map I to obtain a feature map with dimensions of 512×512×3, and then use the predefined upsampling module for upsampling to double the feature map size to obtain a feature map with dimensions of 1024×1024×3. Finally, use a 1×1 convolution for dimension increase operation to obtain the output feature map J (with dimensions of 1024×1024×6). Input the feature map J into the upsampling network again to obtain the feature map J_1 (with dimensions of 2048×2048×3).

[0143] Input the feature map J1 (with dimensions of 2048×2048×3) into the channel expansion module for channel expansion to obtain the feature map K (with dimensions of 1024×1024×12), and then perform a 1×1 convolution to obtain the feature map L (with dimensions of 2048×2048×12).

[0144] Input the feature map L (with a dimension of 2048×2048×12) into the icing area detection network. First, input it into the convolutional pooling activation module for downsampling. In the convolutional pooling activation module, first perform a 3×3 convolutional operation and an average pooling operation, and use the ReLU function to activate to obtain the feature map M (with a dimension of 1024×1024×24). Then, sequentially input the feature map M into three convolutional pooling activation modules for downsampling operations to obtain three feature maps M1 (with a dimension of 512×512×48), M2 (with a dimension of 256×256×96), and M3 (with a dimension of 128×128×192). Then, perform a 3×3 convolutional operation with an unchanged dimension on the feature map M1, and perform a 3×3 convolutional operation and an upsampling operation on the feature maps M2 and M3. The dimension of the feature map M2 becomes (512×512×24), and the dimension of the feature map M3 becomes (512×512×121). Then, perform a concatenation operation on the three obtained feature maps to obtain the feature map N (with a dimension of 512×512×84). Finally, perform a 3×3 convolutional operation and bilinear interpolation on the feature map N to upsample the feature map to a feature map P with a dimension of 1024×1024×1. The feature map P is the output feature map. Each pixel value in the feature map P is 0 or 1. 0 indicates that the pixel at that position is a non-icing area, and 1 indicates that the pixel at that position is an icing area.

[0145] Embodiment III

[0146] Figure 12 It is a schematic structural diagram of an icing detection device for a transmission line provided according to Embodiment III of the present invention. This embodiment is applicable to the situation of detecting the icing condition of a transmission line. The icing detection device for the transmission line can be implemented in the form of hardware and / or software and can be configured in an electronic device such as a computer.

[0147] As Figure 12 shown, the icing detection device for the transmission line disclosed in this embodiment includes:

[0148] The downsampling module 121 is used to obtain the current icing image of the transmission line and perform downsampling on the current icing image according to a predefined first convolution kernel and a second convolution kernel to obtain a first downsampling result;

[0149] The matrix transformation module 122 is used to obtain the number of channels corresponding to the first downsampling result, determine the number of rows and columns of the transformation matrix according to the number of channels, and use the transformation matrix to perform a set number of matrix transformations on the first downsampling result to obtain matrix transformation results corresponding to each matrix transformation respectively;

[0150] The feature extraction module 123 is configured to determine a target matrix transformation result among the matrix transformation results, transpose the target matrix transformation result to obtain a transposed matrix, multiply the transposed matrix by the remaining matrix transformation results in the matrix transformation results, and obtain a current feature extraction result;

[0151] The upsampling module 124 is configured to perform upsampling on the current feature extraction result according to a set number of first convolution kernels and second convolution kernels to obtain a first upsampling result;

[0152] The spare feature map determination module 125 is configured to perform downsampling on the first upsampling result for a first preset number of times, a second preset number of times, and a third preset number of times respectively according to the second convolution kernel to obtain a first spare feature map, a second spare feature map, and a third spare feature map;

[0153] The icing area determination module 126 is configured to splice the first spare feature map, the second spare feature map, and the third spare feature map, and determine the icing area of the transmission line according to the splicing result.

[0154] In the technical solution of this embodiment, through the mutual cooperation of the downsampling module, the matrix transformation module, the feature extraction module, the upsampling module, the spare feature map determination module, and the icing area determination module, the problem that in the prior art, when determining the icing area based on the edge detection result, it is easy to have poor processing of the icing edge details, resulting in difficulty in detecting small icing areas, is solved, and the accuracy of detecting the icing area is improved.

[0155] Optionally, the downsampling module 121 includes:

[0156] The image block unit is configured to obtain a multi-channel original icing image and block the original icing image in each channel according to a set size;

[0157] The block result flattening unit is configured to flatten the block result of the original icing image along the channel dimension direction to obtain a current icing image;

[0158] Wherein, the number of channels corresponding to the current icing image is greater than or equal to the number of channels corresponding to the original icing image;

[0159] The first feature map determination unit is configured to obtain the current number of channels corresponding to the current icing image, and perform convolution on the current icing image using the first convolution kernel with the current number of channels to obtain a first feature map;

[0160] The second feature map determination unit is configured to determine a target number of channels according to the current number of channels and a set value, and perform downsampling on the current icing image using the second convolution kernel with the target number of channels to obtain a second feature map;

[0161] A feature map splicing unit, configured to splice a first feature map and a second feature map to obtain a first downsampling result.

[0162] Optionally, the apparatus further includes a feature enhancement module, which includes:

[0163] A target processing result determination unit, configured to perform pooling on the current feature extraction result to obtain a target pooling result, and perform dimensionality reduction processing on the target pooling result according to a predefined third convolution kernel to obtain a target processing result;

[0164] An attention weight determination unit, configured to process the target processing result using a predefined activation function to obtain an attention weight corresponding to the current feature extraction result;

[0165] A feature enhancement result determination unit, configured to determine a feature enhancement result according to the attention weight and the current feature extraction result;

[0166] A second downsampling result determination unit, configured to perform downsampling on the current feature extraction result using a first convolution kernel and a second convolution kernel to obtain a second downsampling result;

[0167] A target extraction result determination unit, configured to perform feature extraction on the second downsampling result using a transformation matrix to obtain a target feature extraction result;

[0168] A feature extraction result update unit, configured to splice the feature enhancement result and the target feature extraction result to obtain an updated current feature extraction result;

[0169] A pooling unit, configured to perform maximum pooling and hybrid pooling on the current feature extraction result respectively to obtain a maximum pooling result and a hybrid pooling result;

[0170] A pooling result splicing unit, configured to splice the maximum pooling result and the hybrid pooling result to obtain a target pooling result.

[0171] Optionally, the icing area determination module 126 includes:

[0172] A convolution processing unit, configured to perform convolution on a first standby feature map using a second convolution kernel to obtain a first feature map to be spliced;

[0173] An upsampling unit, configured to perform upsampling on a second standby feature map and a third standby feature map respectively using a second convolution kernel to obtain a second feature map to be spliced corresponding to the second standby feature map and a third feature map to be spliced corresponding to the third standby feature map;

[0174] A convolution feature map determination unit, configured to splice a first feature map to be spliced, a second feature map to be spliced, and a third feature map to be spliced, and perform convolution on the splicing result using a second convolution kernel to obtain a convolution feature map;

[0175] An icing feature map determination unit, configured to perform upsampling on the convolution feature map using a bilinear interpolation algorithm to obtain an icing detection feature map;

[0176] An icing area determination unit, configured to determine an icing area of a transmission line according to pixel point values in the icing detection feature map;

[0177] Wherein, pixel point values in the icing area are different from pixel point values in the non-icing area.

[0178] Optionally, the upsampling module 124 includes:

[0179] A dimensionality reduction processing unit, configured to input a current feature extraction result into a set number of convolution modules for convolution, and perform dimensionality reduction processing on the convolution result using a first convolution kernel to obtain a dimensionality reduction feature map;

[0180] Wherein, the convolution module includes a second convolution kernel, a preset normalization layer, and a preset activation function;

[0181] An upsampling processing unit, configured to perform upsampling on the dimensionality reduction feature map using a predefined upsampling module to obtain a size increase result, and perform upsampling processing on the size increase result using a first convolution kernel to obtain a first upsampling result;

[0182] Wherein, the upsampling module is configured to increase the size of the dimensionality reduction feature map on the basis of not reducing the number of channels of the dimensionality reduction feature map.

[0183] Optionally, the apparatus further includes an upsampling result update module, and the module includes:

[0184] An upsampling result determination unit, configured to perform upsampling on the first upsampling result according to a set number of first convolution kernels and second convolution kernels to obtain a second upsampling result;

[0185] An upsampling result update unit, configured to input the second upsampling result into a channel expansion module for upsampling to obtain a channel expansion result, and perform convolution on the channel expansion result using a second convolution kernel to obtain a feature map to be detected, and use the feature map to be detected as the updated first upsampling result;

[0186] Wherein, the channel expansion module is configured to increase the number of channels corresponding to the second upsampling result and reduce the size corresponding to the second upsampling result.

[0187] The icing detection device for a transmission line provided by an embodiment of the present invention can execute the icing detection method for a transmission line provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution of the method. The content not described in detail in this embodiment can be referred to the description in any method embodiment of this application.

[0188] Embodiment 4

[0189] Figure 13 FIG. shows a schematic structural diagram of an electronic device 20 that can be used to implement an embodiment of the present invention. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0190] As Figure 13 shown, the electronic device 20 includes at least one processor 21, and a memory communicatively connected to at least one processor 21, such as a read-only memory (ROM) 22, a random access memory (RAM) 23, etc. Among them, the memory stores a computer program executable by at least one processor. The processor 21 can execute various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 22 or the computer program loaded from the storage unit 28 into the random access memory (RAM) 23. In the RAM 23, various programs and data required for the operation of the electronic device 20 can also be stored. The processor 21, the ROM 22, and the RAM 23 are connected to each other through a bus 24. The input / output (I / O) interface 25 is also connected to the bus 24.

[0191] A plurality of components in the electronic device 20 are connected to the I / O interface 15, including: an input unit 26, such as a keyboard, a mouse, etc.; an output unit 27, such as various types of displays, speakers, etc.; a storage unit 28, such as a magnetic disk, an optical disc, etc.; and a communication unit 29, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 29 allows the electronic device 20 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0192] The processor 21 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the processor 21 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 21 executes the various methods and processes described above, such as the icing detection method for a transmission line.

[0193] In some embodiments, the method for detecting icing on a transmission line can be implemented as a computer program tangibly embodied in a computer-readable storage medium, such as storage unit 28. In some embodiments, part or all of the computer program can be loaded and / or installed onto the electronic device 20 via the ROM 22 and / or the communication unit 29. When the computer program is loaded into the RAM 23 and executed by the processor 21, one or more steps of the method for detecting icing on a transmission line described above can be performed. Alternatively, in other embodiments, the processor 21 can be configured to perform the method for detecting icing on a transmission line by any other suitable means (e.g., by means of firmware).

[0194] The various embodiments of the systems and techniques described above in this document can be implemented in digital electronic circuitry, integrated circuit systems, field programmable gate arrays (FPGA), application specific integrated circuits (ASIC), application specific standard products (ASSP), systems on a chip (SOC), complex programmable logic devices (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special-purpose or general-purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit the data and instructions to the storage system, the at least one input device, and the at least one output device.

[0195] The computer programs for implementing the methods of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when the computer programs are executed by the processor, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The computer programs can be executed entirely on the machine, partially on the machine, as a stand-alone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0196] In the context of the present invention, a computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. The computer-readable storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any suitable combination of the foregoing. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of the machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0197] To provide for interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the electronic device. Other kinds of devices can also be used to provide for interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0198] The systems and techniques described herein can be implemented in a computing system that includes backend components (such as, for example, a data server), or a computing system that includes middleware components (such as, for example, an application server), or a computing system that includes frontend components (such as, for example, a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected by any form or medium of digital data communication (such as, for example, a communication network). Examples of communication networks include: local area network (LAN), wide area network (WAN), blockchain network, and the Internet.

[0199] A computing system may include a client and a server. The client and the server are generally far from each other and usually interact via a communication network. The client-server relationship is created by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system, and solves the defects of difficult management and weak business scalability existing in traditional physical hosts and VPS services.

[0200] It should be understood that various forms of processes shown above can be used, steps can be reordered, added or deleted. For example, the steps recited in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and no limitation is made herein.

[0201] The above specific embodiments do not constitute a limitation on the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting ice coating on a power transmission line, characterized in that: The method comprises: Acquire a current ice-covered image of the transmission line, and downsample the current ice-covered image according to a predefined first convolution kernel and a second convolution kernel to obtain a first downsampling result; Obtaining the number of channels corresponding to the first downsampling result, determining the number of rows and columns of a transformation matrix according to the number of channels, and using the transformation matrix to perform a set number of matrix transformations on the first downsampling result to obtain matrix transformation results corresponding to each matrix transformation; Determine a target matrix transformation result from among the matrix transformation results, perform matrix transposition on the target matrix transformation result to obtain a transposed matrix, and multiply the transposed matrix by the remaining matrix transformation results from the matrix transformation result to obtain a current feature extraction result; Inputting the current feature extraction result into a set number of convolution modules for convolution, and performing dimensionality reduction processing on the convolution result using a first convolution kernel to obtain a dimensionality reduction feature map; Wherein, the convolution module includes a second convolution kernel, a preset normalization layer and a preset activation function; Upsampling the reduced-dimensional feature map using a predefined upsampling module to obtain a size-increased result, and performing a dimension-increased processing on the size-increased result using a first convolution kernel to obtain a first upsampling result; The upsampling module is used to increase the size of the dimensionality reduction feature map while maintaining the number of channels of the dimensionality reduction feature map. According to the second convolution kernel, downsampling the first upsampling result by a first preset number of times, downsampling by a second preset number of times, and downsampling by a third preset number of times respectively, to obtain a first spare feature map, a second spare feature map, and a third spare feature map; Convolving the first spare feature map to obtain a first feature map to be spliced, upsampling the second spare feature map to obtain a second feature map to be spliced, and upsampling the third spare feature map to obtain a third feature map to be spliced; Splicing the first feature map to be spliced, the second feature map to be spliced, and the third feature map to be spliced, and performing a convolution operation and a bilinear interpolation operation on the splicing result to obtain an ice detection feature map; Determine the ice-covered area of ​​the transmission line according to the pixel point values ​​in the ice-covered detection feature map; The pixel values ​​in the ice-covered area are different from those in the non-ice-covered area.

2. The method according to claim 1, characterized in that Get current ice coverage images of transmission lines, including: Acquire multi-channel original ice-covered images, and divide the original ice-covered images under each channel into blocks according to a set size; Flatten the block results of the original ice-covered image along the channel dimension direction to obtain the current ice-covered image; The number of channels corresponding to the current ice-covered image is greater than or equal to the number of channels corresponding to the original ice-covered image.

3. The method according to claim 1, characterized in that Downsampling the current ice-covered image according to a predefined first convolution kernel and a second convolution kernel to obtain a first downsampling result includes: Acquire a current number of channels corresponding to the current ice-covered image, and convolve the current ice-covered image using a first convolution kernel of the current number of channels to obtain a first feature map; Determine a target number of channels according to the current number of channels and a set value, and downsample the current ice-covered image using a second convolution kernel of the target number of channels to obtain a second feature map; The first feature map and the second feature map are concatenated to obtain a first downsampling result.

4. The method according to claim 1, characterized in that After multiplying the transposed matrix with the remaining matrix transformation results in the matrix transformation results to obtain the current feature extraction result, the method further includes: Pooling the current feature extraction result to obtain a target pooling result, and performing dimensionality reduction processing on the target pooling result according to a predefined third convolution kernel to obtain a target processing result; Processing the target processing result using a predefined activation function to obtain an attention weight corresponding to the current feature extraction result; Determining a feature enhancement result according to the attention weight and the current feature extraction result; Downsampling the current feature extraction result using the first convolution kernel and the second convolution kernel to obtain a second downsampling result; Perform feature extraction on the second downsampling result using the transformation matrix to obtain a target feature extraction result; The feature enhancement result and the target feature extraction result are spliced ​​to obtain an updated current feature extraction result.

5. The method according to claim 4, characterized in that Pooling the current feature extraction result to obtain a target pooling result includes: Performing maximum pooling and mixed pooling on the current feature extraction result respectively to obtain a maximum pooling result and a mixed pooling result; The maximum pooling result and the mixed pooling result are concatenated to obtain a target pooling result.

6. The method according to claim 1, characterized in that Splicing the first standby characteristic map, the second standby characteristic map, and the third standby characteristic map, and determining the ice-covered area of ​​the transmission line according to the splicing result, including: Using a second convolution kernel to convolve the first spare feature map to obtain a first feature map to be spliced; Using a second convolution kernel to upsample the second spare feature map and the third spare feature map respectively, to obtain a second feature map to be spliced ​​corresponding to the second spare feature map, and a third feature map to be spliced ​​corresponding to the third spare feature map; Splicing the first feature map to be spliced, the second feature map to be spliced, and the third feature map to be spliced, and convolving the splicing result using a second convolution kernel to obtain a convolution feature map; Upsampling the convolution feature map by using a bilinear interpolation algorithm to obtain an ice detection feature map; Determine the ice-covered area of ​​the transmission line according to the pixel point values ​​in the ice-covered detection feature map; The pixel values ​​in the ice-covered area are different from those in the non-ice-covered area.

7. The method according to claim 1, characterized in that After upsampling the current feature extraction result to obtain a first upsampling result, the method further includes: Upsampling the first upsampling result according to a set number of first convolution kernels and second convolution kernels to obtain a second upsampling result; Inputting the second upsampling result into a channel expansion module for upsampling to obtain a channel expansion result, and convolving the channel expansion result with a second convolution kernel to obtain a feature map to be detected, and using the feature map to be detected as the updated first upsampling result; The channel expansion module is used to increase the number of channels corresponding to the second up-sampling result and reduce the size corresponding to the second up-sampling result.

8. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the method for detecting ice coating on a power transmission line according to any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the method for detecting ice coating on a power transmission line according to any one of claims 1 to 7 when executed.

Images