Method for detecting icing type of power transmission line, model training method and device

Abstract

This application provides a method, model training method, and apparatus for detecting icing types on power transmission lines. The method includes: acquiring image data of the power transmission line; performing label smoothing on the image labels of the image data to obtain training data; the image labels indicating the icing type of the image data; inputting the training data into an initial neural network for processing, using a feature extraction module in the initial neural network to extract features from the training data, obtaining a feature set; processing the feature set based on the output module in the initial neural network to obtain a power transmission line icing detection model; the feature set represents a set of feature data of different dimensions of the image data of the power transmission line; the power transmission line icing detection model is used to process the image data of the power transmission line to obtain the detection result of the icing type of the power transmission line. This method achieves accurate and efficient detection of the icing type of power transmission lines.

Description

Technical Field

[0001] This application relates to the field of power transmission line technology, and in particular to a method, model training method and apparatus for detecting icing types on power transmission lines. Background Technology

[0002] Icing disasters on transmission lines are a major safety hazard in the operation of power systems. Especially in high-altitude and cold regions, under extreme weather conditions, different types of icing such as rime, hoarfrost, and mixed ice are easily formed on the surface of conductors, which leads to increased line load, decreased insulation performance, and even accidents such as line breakage and tripping.

[0003] Traditional manual inspections rely on visual judgment of icing conditions, which suffers from low efficiency, high subjectivity, and delayed response, making it difficult to meet the real-time monitoring needs of large-scale power transmission networks.

[0004] Therefore, there is an urgent need for a solution that can accurately and efficiently detect the type of icing on transmission lines. Summary of the Invention

[0005] This application provides a method, model training method, and apparatus for detecting icing types on power transmission lines, so as to achieve accurate and efficient detection of icing types on power transmission lines.

[0006] In a first aspect, embodiments of this application provide a training method for a transmission line icing detection model, comprising:

[0007] Image data of transmission lines is acquired; and image labels of the image data are smoothed to obtain training data; wherein, the image data includes images of transmission lines and image labels; the image labels indicate the icing type of the image data;

[0008] The training data is input into an initial neural network for processing. The feature extraction module in the initial neural network performs feature extraction on the training data to obtain a feature set. The output module in the initial neural network processes the feature set to obtain a transmission line icing detection model.

[0009] The feature set represents the feature data set of different dimensions of the image data of the transmission line; the icing detection model of the transmission line is used to process the image data of the transmission line to obtain the detection result of the icing type of the transmission line.

[0010] In one possible implementation, the training data is input into an initial neural network for processing, to perform feature extraction processing on the training data based on the feature extraction module in the initial neural network, to obtain a feature set, including:

[0011] The training data is input into an initial neural network for processing. Based on the first module in the feature extraction module, the training data is processed to obtain a first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line.

[0012] Based on the second module in the feature extraction module, the first feature is processed to obtain the second feature; wherein, the second feature represents the mid-layer shape features of the image data of the transmission line;

[0013] Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line;

[0014] The feature set is determined based on the first feature, the second feature, and the third feature.

[0015] In one possible implementation, the training data is processed based on the first module in the feature extraction module to obtain a first feature, including:

[0016] Based on the first module in the feature extraction module, the training data is sequentially processed by 3×3 convolution, the first fusion inverse residual module, the second fusion inverse residual module, and the third fusion inverse residual module to obtain a fourth feature map; wherein, the fourth feature represents the primary shallow texture features of the image data of the transmission line.

[0017] The fourth feature map is processed based on the Laplacian operator to obtain the first feature.

[0018] In one possible implementation, the first feature is processed based on the second module in the feature extraction module to obtain the second feature, including:

[0019] Based on the second module in the feature extraction module, the first feature is sequentially processed by 1×1 convolution, 3×3 depthwise separable convolution, coordinate attention mechanism processing, and 1×1 convolution to obtain the fifth feature; wherein, the fifth feature represents the primary mid-level shape features of the image data of the transmission line.

[0020] The first feature and the fifth feature are concatenated to obtain the second feature.

[0021] In one possible implementation, the second feature is processed based on the third module in the feature extraction module to obtain the third feature, including:

[0022] Based on the third module in the feature extraction module, the second feature is processed sequentially by the first inverse residual module and the second inverse residual module to obtain the third feature.

[0023] In one possible implementation, the feature set includes a first feature, a second feature, and a third feature; wherein the first feature represents the shallow texture features of the image data of the transmission line; the second feature represents the mid-level shape features of the image data of the transmission line; and the third feature represents the deep semantic features of the image data of the transmission line.

[0024] The feature set is processed based on the output module of the initial neural network to obtain a transmission line icing detection model, including:

[0025] Based on the output module in the initial neural network, the first feature, the second feature, and the third feature are spliced ​​together to obtain spliced ​​features; wherein, the spliced ​​features characterize the multi-dimensional features of the image data of the transmission line.

[0026] Based on the output module in the initial neural network, the spliced ​​features are sequentially processed by 1×1 convolution, pooling, and fully connected layers to obtain the transmission line icing detection model.

[0027] In one possible implementation, before inputting the training data into an initial neural network for processing, and performing feature extraction processing on the training data based on the feature extraction module in the initial neural network to obtain a feature set, the method further includes:

[0028] The training data is subjected to image enhancement processing to obtain a processed image set.

[0029] Secondly, embodiments of this application provide a method for detecting the type of icing on transmission lines, including:

[0030] Acquire image data of power transmission lines;

[0031] The image data is input into the transmission line icing detection model, and the image data is processed based on the feature extraction module in the transmission line icing detection model to obtain a feature set;

[0032] Based on the output module of the transmission line icing detection model, the feature set is processed to obtain the detection result of the icing type of the transmission line. The transmission line icing detection model is trained by the first aspect and / or various possible training methods of the first aspect.

[0033] In one possible implementation, the image data is input into a power transmission line icing detection model, and the image data is processed based on the feature extraction module in the power transmission line icing detection model to obtain a feature set, including:

[0034] The image data is input into the transmission line icing detection model for processing. Based on the first module in the feature extraction module, the image data is processed to obtain a first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line.

[0035] Based on the second module in the feature extraction module, the first feature is processed to obtain the second feature; wherein, the second feature represents the mid-layer shape features of the image data of the transmission line;

[0036] Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line;

[0037] The feature set is determined based on the first feature, the second feature, and the third feature.

[0038] Thirdly, embodiments of this application provide a training device for a transmission line icing detection model, comprising:

[0039] An acquisition module is used to acquire image data of transmission lines; and to perform label smoothing on the image labels of the image data to obtain training data; wherein, the image data includes images of transmission lines and image labels; the image labels indicate the icing type of the image data;

[0040] The training module is used to input the training data into an initial neural network for processing, and to perform feature extraction processing on the training data based on the feature extraction module in the initial neural network to obtain a feature set; and to process the feature set based on the output module in the initial neural network to obtain a transmission line icing detection model.

[0041] The feature set represents the feature data set of different dimensions of the image data of the transmission line; the icing detection model of the transmission line is used to process the image data of the transmission line to obtain the detection result of the icing type of the transmission line.

[0042] In one possible implementation, the training module includes:

[0043] The first training module is used to input the training data into the initial neural network for processing, and to process the training data based on the first module in the feature extraction module to obtain a first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line;

[0044] The second training module is used to process the first feature based on the second module in the feature extraction module to obtain the second feature; wherein the second feature represents the mid-level shape features of the image data of the transmission line;

[0045] The third training module is used to process the second feature based on the third module in the feature extraction module to obtain the third feature; wherein the third feature represents the deep semantic features of the image data of the transmission line.

[0046] The feature set is determined based on the first feature, the second feature, and the third feature.

[0047] In one possible implementation, the first training module includes:

[0048] Based on the first module in the feature extraction module, the training data is sequentially processed by 3×3 convolution, the first fusion inverse residual module, the second fusion inverse residual module, and the third fusion inverse residual module to obtain a fourth feature map; wherein, the fourth feature represents the primary shallow texture features of the image data of the transmission line.

[0049] The fourth feature map is processed based on the Laplacian operator to obtain the first feature.

[0050] In one possible implementation, the second training module includes:

[0051] Based on the second module in the feature extraction module, the first feature is sequentially processed by 1×1 convolution, 3×3 depthwise separable convolution, coordinate attention mechanism processing, and 1×1 convolution to obtain the fifth feature; wherein, the fifth feature represents the primary mid-level shape features of the image data of the transmission line.

[0052] The first feature and the fifth feature are concatenated to obtain the second feature.

[0053] In one possible implementation, the third training module includes:

[0054] Based on the third module in the feature extraction module, the second feature is processed sequentially by the first inverse residual module and the second inverse residual module to obtain the third feature.

[0055] In one possible implementation, the feature set includes a first feature, a second feature, and a third feature; wherein the first feature represents the shallow texture features of the image data of the transmission line; the second feature represents the mid-level shape features of the image data of the transmission line; and the third feature represents the deep semantic features of the image data of the transmission line.

[0056] The training module includes:

[0057] Based on the output module in the initial neural network, the first feature, the second feature, and the third feature are spliced ​​together to obtain spliced ​​features; wherein, the spliced ​​features characterize the multi-dimensional features of the image data of the transmission line.

[0058] Based on the output module in the initial neural network, the spliced ​​features are sequentially processed by 1×1 convolution, pooling, and fully connected layers to obtain the transmission line icing detection model.

[0059] In one possible implementation, prior to the training module, the following is also included:

[0060] The image enhancement module is used to perform image enhancement processing on the training data to obtain a processed image set.

[0061] Fourthly, embodiments of this application provide a detection device for the type of icing on transmission lines, comprising:

[0062] The acquisition module is used to acquire image data of power transmission lines;

[0063] The processing module is used to input the image data into the transmission line icing detection model, and process the image data based on the feature extraction module in the transmission line icing detection model to obtain a feature set;

[0064] The detection module is used to process the feature set based on the output module of the transmission line icing detection model to obtain the detection result of the icing type of the transmission line, wherein the transmission line icing detection model is trained by the first aspect and / or various possible training methods of the first aspect as described above.

[0065] In one possible implementation, the processing module includes:

[0066] The image data is input into the transmission line icing detection model for processing. Based on the first module in the feature extraction module, the image data is processed to obtain a first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line.

[0067] Based on the second module in the feature extraction module, the first feature is processed to obtain the second feature; wherein, the second feature represents the mid-layer shape features of the image data of the transmission line;

[0068] Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line;

[0069] The feature set is determined based on the first feature, the second feature, and the third feature.

[0070] Fifthly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0071] The memory stores computer-executed instructions;

[0072] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0073] In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0074] In a seventh aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0075] This application provides a method, model training method, and apparatus for detecting icing types on power transmission lines. It acquires image data of the power transmission line and corresponding image labels, where the image labels clearly indicate the icing type. The image labels undergo label smoothing processing; by adjusting the label confidence distribution, the model's sensitivity to label noise is reduced, overfitting risk is decreased, and generalization ability is improved. After the processed training data is input into the initial neural network, the feature extraction module uses a convolutional neural network structure to extract low-dimensional features such as texture, edges, and shapes from the image through multiple convolutional kernels, and fuses them into a high-dimensional feature set, effectively capturing subtle differences in the icing area. The output module performs classification decisions based on the feature set, mapping to the icing type probability distribution through a fully connected layer and a softmax activation function, ultimately outputting an accurate icing detection result. This solves the problem of misjudgment caused by differences in image quality and complex icing morphology in traditional detection methods, achieving high-precision icing type identification in complex environments, improving detection accuracy, and significantly enhancing the differentiation of different icing types such as rime, hoarfrost, and mixed icing, effectively ensuring the safe operation of power transmission lines. Attached Figure Description

[0076] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0077] Figure 1 A flowchart illustrating a training method for a transmission line icing detection model provided in this application embodiment. Figure 1 ;

[0078] Figure 2 A flowchart illustrating a training method for a transmission line icing detection model provided in this application embodiment. Figure 2 ;

[0079] Figure 3 A flowchart illustrating a method for detecting icing types on power transmission lines, provided in an embodiment of this application;

[0080] Figure 4 A schematic diagram of the structure of a training device for a transmission line icing detection model provided in an embodiment of this application;

[0081] Figure 5 This is a schematic diagram of the structure of a detection device for icing type of transmission lines provided in an embodiment of this application;

[0082] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0083] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0084] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0085] Icing disasters on transmission lines are a major safety hazard in the operation of power systems. Especially in high-altitude and cold regions, under extreme weather conditions, different types of icing such as rime, hoarfrost, and mixed ice are easily formed on the surface of conductors, which leads to increased line load, decreased insulation performance, and even accidents such as line breakage and tripping.

[0086] Traditional manual inspections rely on visual assessment of icing conditions, resulting in low efficiency, high subjectivity, and delayed response, making them unsuitable for the real-time monitoring needs of large-scale power transmission networks. In recent years, image recognition-based intelligent inspection technology has gradually become mainstream. By deploying cameras or drones near transmission lines to collect images and combining them with edge computing devices for real-time analysis, icing types can be quickly identified and warnings issued. However, real-world scenarios present challenges such as uneven lighting, complex backgrounds, blurred boundaries between icing and conductors, and weak target textures, leading to insufficient accuracy of existing algorithms in complex environments. Furthermore, the limited computing and storage resources of edge terminals necessitate models that are both lightweight and highly accurate.

[0087] Therefore, the method for detecting icing type of transmission lines provided in this application embodiment can solve the above-mentioned problems.

[0088] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0089] Figure 1 A flowchart illustrating a training method for a transmission line icing detection model provided in this application embodiment. Figure 1 ,like Figure 1 As shown, the method includes:

[0090] S101. Acquire image data of the transmission line; and perform label smoothing on the image labels of the image data to obtain training data; wherein, the image data includes images of the transmission line and image labels; the image labels indicate the icing type of the image data.

[0091] For example, image data of power transmission lines refers to a collection of images containing power transmission lines acquired through various image acquisition devices (such as cameras, cameras mounted on drones, etc.). These images record the appearance characteristics of power transmission lines under different environments and different icing conditions, and are the basic material for models to learn icing type recognition.

[0092] Image labels are markers used to identify the type of icing on power transmission lines in image data. They provide the model with the correct classification result for each image, enabling the model to accurately determine the type of icing by learning the association between these labels and image features.

[0093] Label smoothing is a data preprocessing technique that transforms raw hard labels (i.e., explicit category markers, such as icing type A, B, C, etc.) into soft labels (i.e., labels with a certain probability distribution). This process prevents the model from being overconfident in the labels in the image data, thereby improving the model's generalization ability and reducing the risk of overfitting.

[0094] Choose appropriate image acquisition equipment based on actual needs and scenario characteristics. For example, when large-scale, high-efficiency image acquisition is required, a drone equipped with a high-definition camera can be used for shooting; at fixed monitoring points, cameras can be installed for long-term, continuous image acquisition.

[0095] Based on the distribution of transmission lines, the data acquisition area should be rationally planned to ensure coverage of transmission lines in different geographical locations and environmental conditions. Simultaneously, considering the differences in icing conditions on transmission lines under different seasons and weather conditions, appropriate times should be selected for image acquisition to obtain images of diverse icing types.

[0096] According to the planned acquisition area and time, use the selected image acquisition equipment to carry out the actual image acquisition work. During the acquisition process, pay attention to adjusting the equipment parameters, such as focal length and exposure time, to ensure that the acquired images are clear and have distinct features.

[0097] The acquired image data is processed to remove blurry, duplicate, or invalid images. Then, the image data is stored in a database or file system according to certain naming rules and storage formats for subsequent processing and use.

[0098] Common label smoothing methods include uniform label smoothing and prior knowledge-based label smoothing. Uniform label smoothing adjusts the original hard labels evenly according to a certain ratio; prior knowledge-based label smoothing adjusts the labels of different categories at different ratios based on prior knowledge of the icing type.

[0099] Taking uniform label smoothing as an example, assuming the original hard label is in one-hot encoding form (for example, for icing type A, the label is [1,0,0]), and the smoothing parameter is α (0<α<1), then the formula for calculating the smoothed label is: New label = (1-α)×original label + α / n (where n is the total number of icing types). For example, when α=0.1 and n=3, for the original label [1,0,0] of icing type A, the smoothed label is (1-0.1)×[1,0,0]+0.1 / 3×[1,1,1]=[0.9+0.033,0+0.033,0+0.033]=[0.933,0.033,0.033].

[0100] The calculated smoothed labels are associated with the corresponding image data and stored to form a training dataset containing images and corresponding smoothed labels.

[0101] S102. Input the training data into the initial neural network for processing, and perform feature extraction processing on the training data based on the feature extraction module in the initial neural network to obtain a feature set; process the feature set based on the output module in the initial neural network to obtain the transmission line icing detection model.

[0102] Among them, the feature set represents the feature data set of different dimensions of the image data of the transmission line; the transmission line icing detection model is used to process the image data of the transmission line to obtain the detection result of the icing type of the transmission line.

[0103] For example, an initial neural network is a neural network model with a basic architecture and parameters that has not yet been fully trained. It typically contains basic structures such as a feature extraction module and an output module, providing a basic framework for subsequent training and optimization to learn patterns and features from the input data.

[0104] Preprocessed training data (such as label smoothing) is input into the initial neural network. The training data includes image data of the transmission line and corresponding smoothed labels. The image data serves as the network input, and the smoothed labels are used for subsequent model training and evaluation.

[0105] The initial feature extraction module in the neural network begins its work. This module typically consists of multiple convolutional layers, pooling layers, etc. Convolutional layers slide their kernels across the image to extract local features such as edges and textures; pooling layers downsample the feature maps output by the convolutional layers, reducing the amount of data while retaining important features. Through alternating processing by multiple convolutional and pooling layers, high-level features of the image are gradually extracted, ultimately resulting in a feature set.

[0106] The feature set output by the feature extraction module is passed to the output module. The output module typically contains a fully connected layer and a classifier (such as a Softmax classifier). The fully connected layer integrates and transforms the feature set to make it more suitable for the classifier's processing; the classifier, based on the output of the fully connected layer and combined with the smoothed labels in the training data, calculates the probability of each icing type, and finally outputs the icing type with the highest probability as the detection result.

[0107] During the initial processing of training data by the neural network, backpropagation and optimization algorithms (such as stochastic gradient descent) are used to adjust the parameters of the neural network based on the error between the output and the true label (smooth label). Through continuous iterative training, the error is gradually reduced, improving the model's accuracy in judging the type of icing on transmission lines, ultimately resulting in a well-trained transmission line icing detection model.

[0108] This application provides a training method for a transmission line icing detection model. It acquires image data of the transmission line and corresponding image labels, where the image labels clearly indicate the icing type. Label smoothing is performed on the image labels; by adjusting the label confidence distribution, the model's sensitivity to label noise is reduced, overfitting risk is decreased, and generalization ability is improved. After the processed training data is input into the initial neural network, the feature extraction module uses a convolutional neural network structure to extract low-dimensional features such as texture, edges, and shape from the image through multiple convolutional kernels, and fuses them into a high-dimensional feature set, effectively capturing subtle differences in the icing area. The output module performs classification decisions based on the feature set, mapping to the icing type probability distribution through a fully connected layer and a softmax activation function, ultimately outputting an accurate icing detection result. This method solves the misjudgment problem caused by image quality differences and complex icing morphologies in traditional detection methods, achieving high-precision icing type identification in complex environments, improving detection accuracy, and significantly enhancing the differentiation of different icing types such as rime, hoarfrost, and mixed icing, effectively ensuring the safe operation of transmission lines.

[0109] Figure 2 A flowchart illustrating a training method for a transmission line icing detection model provided in this application embodiment. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1Based on the examples, a training method for a transmission line icing detection model is described in detail. The method includes:

[0110] S201. Acquire image data of the transmission line; and perform label smoothing on the image labels of the image data to obtain training data.

[0111] For example, this step can be referred to step S101, and will not be described again.

[0112] In one example, image augmentation is performed on the training data to obtain a processed set of images.

[0113] For example, raw image data of transmission lines are collected, and these images undergo preliminary preprocessing, such as resizing them to a uniform size, to facilitate subsequent image enhancement and model training. Simultaneously, it is ensured that each image has a corresponding accurate icing type label.

[0114] Based on the characteristics of icing images of transmission lines and the training requirements, appropriate image enhancement methods should be selected. Multiple methods, such as geometric transformation, color transformation, and noise addition, can be combined to fully expand the training dataset.

[0115] Rotation processing uses rotation functions in image processing libraries (such as OpenCV) to rotate each original image by different angles, generating a rotated image.

[0116] The flipping process involves flipping the image horizontally and vertically to obtain a mirrored image.

[0117] Scaling is the process of scaling an image by a certain ratio (such as 0.8, 1.2, etc.) to generate images of different scales.

[0118] Brightness adjustment is achieved by changing the brightness values ​​of image pixels, such as increasing or decreasing the brightness values ​​of all pixels by a fixed value.

[0119] Contrast adjustment processing uses contrast adjustment algorithms, such as histogram equalization, to enhance the contrast of an image.

[0120] Saturation adjustment involves manipulating the color channels of an image to change the saturation of colors, making the image colors more vibrant or duller.

[0121] Adding noise involves selecting an appropriate noise type (such as Gaussian noise) and adding noise to the image based on set noise parameters (such as the mean and variance of the noise), thus generating a noisy image.

[0122] The images obtained after the above image enhancement processes are integrated to form a new image set, which serves as the expanded training dataset.

[0123] S202. Input the training data into the initial neural network for processing. Based on the first module in the feature extraction module, process the training data to obtain a first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line. Based on the second module in the feature extraction module, process the first feature to obtain a second feature; wherein, the second feature represents the mid-level shape features of the image data of the transmission line. Based on the third module in the feature extraction module, process the second feature to obtain a third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line. Determine the feature set based on the first feature, the second feature, and the third feature.

[0124] For example, prepared training data is fed into the initial neural network. This training data is typically pre-processed images of transmission lines, ensuring that the image size, format, etc., meet the input requirements of the neural network.

[0125] The first module in the feature extraction module performs convolution operations on the input training data. By selecting appropriate parameters such as kernel size, number, and stride, the shallow texture features of the image are extracted by sliding the convolution kernel across the image to obtain the first feature. Optionally, a 3x3 convolution kernel with a stride of 1 is used to perform convolution operations on the image to capture local texture information.

[0126] The first feature is input into the second module. The second module first performs pooling operations, such as max pooling or average pooling, selecting an appropriate pooling window size and stride to reduce the dimensionality of the first feature, preserving important shape features. Then, some convolutional operations may be performed to further extract features, ultimately obtaining the second feature. Optionally, a 2x2 max pooling window with a stride of 2 can be used to pool the first feature, reducing the amount of data while highlighting shape features.

[0127] The second feature is then fed into the third module. This third module typically contains multiple fully connected layers. The second feature is flattened and then input into these fully connected layers. The fully connected layers integrate and perform in-depth analysis of the preceding features, extracting deep semantic features through connections between neurons and weight adjustments, thus obtaining the third feature. Optionally, two fully connected layers can be used: the first layer has 256 neurons, and the second layer has 128 neurons, to perform in-depth analysis of the second feature.

[0128] The obtained first, second, and third features are combined in a certain way to form a complete feature set. This feature set will serve as input data for subsequent model training, used to adjust the model's parameters and improve its performance.

[0129] In one example, based on the first module in the feature extraction module, the training data is sequentially processed by 3×3 convolution, the first fusion inverse residual module, the second fusion inverse residual module, and the third fusion inverse residual module to obtain the fourth feature map; wherein, the fourth feature represents the primary shallow texture features of the image data of the transmission line; based on the Laplacian operator, the fourth feature map is processed to obtain the first feature.

[0130] For example, the training data is first input into the first module of the feature extraction module. A 3×3 convolution process is then performed, which involves moving a 3×3 convolution kernel across the image step by step to illuminate each local region of the image and extract basic information such as local texture and edges.

[0131] Next, the data is processed sequentially through the first, second, and third fusion inverse residual modules. In the fusion inverse residual module, the inverse residual structure first expands the number of channels, essentially opening more channels for feature information, allowing more feature information to flow. Then, depthwise separable convolution and other operations are performed. This convolution method significantly reduces computational cost compared to traditional convolution while effectively extracting features. Finally, the number of channels is compressed to remove redundant information, making the features more refined. The fusion operation integrates feature information from different levels, much like combining information observed from different angles. After processing by these three modules, a fourth feature map is obtained, representing the primary shallow texture features of the transmission line image data.

[0132] Finally, the fourth feature map is processed based on the Laplacian operator. The Laplacian operator acts as a sharpening tool, highlighting areas of abrupt grayscale changes, i.e., edges and texture details, by calculating the second derivative of the image.

[0133] In one possible implementation, the first fusion inverse residual module can be Fused-MBConv; the second fusion inverse residual module can also be Fused-MBConv.

[0134] In one example, based on the second module in the feature extraction module, the first feature is sequentially processed by 1×1 convolution, 3×3 depthwise separable convolution, coordinate attention mechanism processing, and 1×1 convolution to obtain the fifth feature; wherein, the fifth feature represents the primary mid-level shape features of the image data of the transmission line; the first feature and the fifth feature are concatenated to obtain the second feature.

[0135] For example, the first feature is input into the second module of the feature extraction module. The first feature typically exists in the form of a feature map with a specific number of channels, height, and width.

[0136] 1×1 convolution processing refers to selecting an appropriate number of 1×1 convolution kernels, the number of which determines the number of channels in the output feature map. For example, if the input feature map has 64 channels and you want to adjust the number of channels to 128 through 1×1 convolution, then you choose 128 1×1 convolution kernels. The weight parameters of the convolution kernels are initialized using random initialization methods, such as generating random weight values ​​using a Gaussian distribution. The 1×1 convolution kernels are slid across the input feature map. At each position, the convolution kernel is multiplied by the pixel value of the corresponding local region in the feature map, and then all the product results are summed to obtain the output value at that position. The convolution kernels are slid along a set stride (usually 1) until the entire feature map is covered, resulting in the feature map after 1×1 convolution processing.

[0137] 3×3 depthwise separable convolution processing refers to preparing an independent 3×3 convolution kernel for each channel of the input feature map. For example, if the input feature map has 128 channels, then 128 3×3 convolution kernels are prepared. These kernels are used to perform convolution operations on the corresponding channels, extracting local features for each channel, resulting in 128 intermediate feature maps. An appropriate number of 1×1 convolution kernels are selected, the number of which determines the number of channels in the output feature map. For example, if 256 channels are desired, then 256 1×1 convolution kernels are selected. These 1×1 kernels are slid across the 128 intermediate feature maps obtained from depthwise convolution. At each position, the kernel is multiplied by the corresponding 128 pixel values, and all products are summed to obtain the output value at that position. The kernels are slid with a stride of 1 to cover the entire feature map, resulting in the feature map after 3×3 depthwise separable convolution processing.

[0138] The coordinate attention mechanism includes horizontal global average pooling, vertical global average pooling, and feature fusion.

[0139] Horizontal global average pooling is a process that performs global average pooling on the input feature map in the horizontal direction to obtain a feature vector. Each element of this vector represents the average feature value of the corresponding row of the feature map in the horizontal direction.

[0140] Vertical global average pooling is a process that performs global average pooling on the input feature map in the vertical direction to obtain another feature vector. Each element of this vector represents the average feature value of the corresponding column of the feature map in the vertical direction.

[0141] Feature fusion involves concatenating two feature vectors obtained from the horizontal and vertical directions, then fusing them through a fully connected layer to obtain a composite feature vector. The fused feature vector is then passed through an activation function (such as the sigmoid function) to generate attention weights, which incorporate spatial location information in both the horizontal and vertical directions. These attention weights are then multiplied by the original input feature map along the channel dimension, causing the model to focus more on important spatial location information, resulting in a feature map processed by the coordinate attention mechanism.

[0142] The feature map, after being processed by the coordinate attention mechanism, is then subjected to another 1×1 convolution to further adjust the feature dimensions, resulting in the fifth feature.

[0143] Optionally, check if the height and width of the first and fifth features are the same. If they are different, make appropriate adjustments (such as cropping or padding) to make their dimensions consistent.

[0144] The first feature and the fifth feature are concatenated along the channel dimension to obtain the second feature.

[0145] In one example, based on the third module in the feature extraction module, the second feature is processed sequentially by the first inverse residual module and the second inverse residual module to obtain the third feature.

[0146] For example, the second feature is input into the third module of the feature extraction module. The second feature typically exists in the form of a feature map with a specific number of channels, height, and width.

[0147] Based on the third module in the feature extraction module, the second feature is processed sequentially by the first inverse residual module and the second inverse residual module to obtain the third feature.

[0148] In one possible implementation, the first inverse residual module processing includes 1×1 convolution processing, 3×3 depthwise separable convolution processing, coordinate attention mechanism processing, and 1×1 convolution processing; the second inverse residual module processing includes 1×1 convolution processing, 3×3 depthwise separable convolution processing, coordinate attention mechanism processing, and 1×1 convolution processing.

[0149] S203. Based on the output module of the initial neural network, the first feature, the second feature, and the third feature are spliced ​​together to obtain the spliced ​​feature; wherein, the spliced ​​feature represents the multi-dimensional features of the image data of the transmission line; based on the output module of the initial neural network, the spliced ​​feature is sequentially processed by 1×1 convolution, pooling, and fully connected layer processing to obtain the transmission line icing detection model.

[0150] For example, the dimensional information of the first feature, the second feature, and the third feature is determined, including the number of channels, height, and width. Assume the first feature has dimensions H1×W1×C1, the second feature has dimensions H2×W2×C2, and the third feature has dimensions H3×W3×C3.

[0151] To enable stitching, the feature maps must be identical in both height and width. If they are not identical, the feature maps can be resized using interpolation or other methods to ensure that H1=H2=H3=H and W1=W2=W3=W.

[0152] The three resized feature maps are spliced ​​together along the channel dimension to obtain the spliced ​​feature, which has a size of H×W×(C1+C2+C3).

[0153] Based on the output module in the initial neural network, the spliced ​​features are sequentially processed by 1×1 convolution, pooling, and fully connected layers to obtain the transmission line icing detection model.

[0154] The transmission line icing detection model training method provided in this application acquires transmission line image data and performs label smoothing processing to adjust the label confidence distribution to reduce noise sensitivity, reduce overfitting, and improve generalization ability. The feature extraction module adopts a layered design: the first module extracts shallow texture features through edge detection convolutional kernels to capture subtle texture differences on the icing surface; the second module extracts mid-level shape features through a shape analysis layer to identify the geometric shape of the icing contour; and the third module extracts deep semantic features through a semantic understanding network to analyze the overall structure and type association information of the icing. The three-stage features are concatenated to form a multi-dimensional feature set. Feature fusion is achieved by adjusting the channel dimension through 1×1 convolution, pooling layers compress spatial dimensions to reduce computational load, and fully connected layers combined with a softmax classifier complete type decision-making. This design solves the problem of insufficient accuracy in identifying icing types in complex environments by using hierarchical feature extraction and multi-dimensional fusion. It achieves accurate differentiation of different forms of icing, such as rime, hoarfrost, and mixed icing, improving detection accuracy, reducing false positive rate, and effectively ensuring the real-time performance and reliability of icing monitoring of transmission lines, thus providing key technical support for the safe operation of the power grid.

[0155] Figure 3 A flowchart illustrating a method for detecting icing types on transmission lines provided in this application embodiment is shown below. Figure 3 As shown, the method includes:

[0156] S301. Acquire image data of the power transmission line.

[0157] For example, image data of power transmission lines can be acquired by high-definition cameras deployed on inspection drones or fixed monitoring points, such as using visible light or infrared cameras to capture real-time images of the lines. Power transmission lines refer to facilities such as conductors and towers used to transmit electrical energy in a power system, while image data is visual information stored in digital format, such as an RGB pixel matrix.

[0158] S302. Input the image data into the transmission line icing detection model, and process the image data based on the feature extraction module in the transmission line icing detection model to obtain a feature set.

[0159] For example, image data is input into the transmission line icing detection model, and the image data is processed based on the feature extraction module in the transmission line icing detection model to obtain a feature set.

[0160] In one example, image data is input into a power transmission line icing detection model for processing. Based on the first module in the feature extraction module, the image data is processed to obtain a first feature, which represents the shallow texture features of the power transmission line image data. Based on the second module in the feature extraction module, the first feature is processed to obtain a second feature, which represents the mid-level shape features of the power transmission line image data. Based on the third module in the feature extraction module, the second feature is processed to obtain a third feature, which represents the deep semantic features of the power transmission line image data. Based on the first feature, the second feature, and the third feature, a feature set is determined.

[0161] For example, the input image data is processed based on a first module in the feature extraction module. This first module includes 3×3 convolution processing, a first fusion inverse residual module, a second fusion inverse residual module, and a third fusion inverse residual module.

[0162] Input a 256x256 pixel RGB image of a power transmission line. After processing by the first module, the first feature is obtained. The first feature is a set of feature maps, which mainly characterize shallow texture features in the image, such as the edges of conductors, the outline of insulator strings, and the rough texture of background trees. These features are the most basic structural units of visual elements.

[0163] The first feature is processed based on the second module. The second module may be an inverse residual module; alternatively, the second module includes 1×1 convolution, 3×3 depthwise separable convolution, coordinate attention mechanism processing, and 1×1 convolution.

[0164] The second module performs multiple convolutions and non-linear mappings on the first feature, but directly adds the input to the output through residual connections.

[0165] After processing by the second module, the size of the feature map is further reduced, but the original image region (i.e., the receptive field) sensed by the feature vector at each location is significantly increased. Therefore, the resulting second feature begins to integrate broader contextual information, representing mid-level shape features composed of shallow textures. For example, it may be able to represent the arcuate outline formed by icing on a conductor, the complete geometry of an insulator string, or the initial shape of an icicle.

[0166] Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature. The third feature includes the first inverse residual module and the second inverse residual module.

[0167] Through the deep abstraction of the third module, the resulting third feature is highly semantic. It strips away specific spatial details and focuses on high-level conceptual representations, namely deep semantic features. For example, this feature can powerfully represent high-level category information such as severely iced conductor segments, insulators with ice ridges, or normally ice-free lines.

[0168] The final feature set is determined based on the first feature, the second feature, and the third feature.

[0169] In one possible implementation, the third feature (deep semantic vector) is used as the core, while the first and second features are adjusted to appropriate sizes and channel compressed using techniques such as upsampling or spatial pyramid pooling. These features are then concatenated or weighted and summed across different dimensions. This resulting feature set integrates shallow detail textures, mid-level geometry, and high-level semantic information, greatly enriching the information sources the model uses to determine icing conditions. This allows the model to grasp both global semantics and local details, resulting in stronger robustness and higher detection accuracy in scenarios with complex backgrounds, varying lighting conditions, or uneven icing levels.

[0170] S303. The output module based on the transmission line icing detection model processes the feature set to obtain the detection results of the icing type of the transmission line.

[0171] For example, the output module of the transmission line icing detection model processes the feature set that integrates multi-level information to finally obtain the detection result of the icing type of the transmission line, such as no ice, rime, hoarfrost, or mixed rime.

[0172] This application provides a method for detecting the icing type of transmission lines. By acquiring image data of the transmission line, inputting the image data into the transmission line icing detection model, processing the image data based on the feature extraction module to obtain a feature set, and then processing the feature set through the output module to obtain the icing type detection result, the method improves the detection accuracy and efficiency and reduces the cost of manual intervention.

[0173] First, by acquiring real-time image data of the transmission lines, the timeliness and intuitiveness of the detection data are ensured. Second, the image data is processed in multiple dimensions using a feature extraction module to extract key features such as ice thickness, morphology, and distribution, forming a feature set. This process uses deep learning algorithms to automatically extract and optimize features, avoiding the limitations of manual feature design. Finally, the output module makes classification decisions based on the feature set and outputs the detection results of the icing type (such as rime, hoarfrost, mixed rime, etc.).

[0174] This technical solution, through a model-based and automated processing flow, solves the problems of low efficiency, strong subjectivity, and poor environmental adaptability of traditional manual inspection. It enables rapid and accurate identification of icing types on transmission lines, provides reliable data support for power grid operation and maintenance, effectively improves the power grid's ability to resist icing disasters, and ensures the safe and stable operation of transmission lines.

[0175] Figure 4 This is a schematic diagram of the structure of a training device for a transmission line icing detection model provided in an embodiment of this application, as shown below. Figure 4 As shown, the training device 40 for a transmission line icing detection model provided in this embodiment includes:

[0176] The acquisition module 401 is used to acquire image data of the transmission line; and to perform label smoothing on the image labels of the image data to obtain training data; wherein, the image data includes images of the transmission line and image labels; the image labels indicate the icing type of the image data;

[0177] The training module 402 is used to input training data into the initial neural network for processing, and to perform feature extraction processing on the training data based on the feature extraction module in the initial neural network to obtain a feature set; and to process the feature set based on the output module in the initial neural network to obtain the transmission line icing detection model.

[0178] Among them, the feature set represents the feature data set of different dimensions of the image data of the transmission line; the transmission line icing detection model is used to process the image data of the transmission line to obtain the detection result of the icing type of the transmission line.

[0179] In one possible implementation, the training module 402 includes:

[0180] The first training module 4021 is used to input training data into the initial neural network for processing. Based on the first module in the feature extraction module, the training data is processed to obtain the first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line.

[0181] The second training module 4022 is used to process the first feature based on the second module in the feature extraction module to obtain the second feature; wherein, the second feature represents the mid-level shape feature of the image data of the transmission line;

[0182] The third training module 4023 is used to process the second feature based on the third module in the feature extraction module to obtain the third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line;

[0183] The feature set is determined based on the first feature, the second feature, and the third feature.

[0184] In one possible implementation, the first training module 4021 includes:

[0185] Based on the first module in the feature extraction module, the training data is processed sequentially by 3×3 convolution, the first fusion inverse residual module, the second fusion inverse residual module, and the third fusion inverse residual module to obtain the fourth feature map; wherein, the fourth feature represents the primary shallow texture features of the image data of the transmission line.

[0186] The fourth feature map is processed using the Laplacian operator to obtain the first feature.

[0187] In one possible implementation, the second training module 4022 includes:

[0188] Based on the second module in the feature extraction module, the first feature is sequentially processed by 1×1 convolution, 3×3 depthwise separable convolution, coordinate attention mechanism processing, and 1×1 convolution to obtain the fifth feature; wherein, the fifth feature represents the primary and intermediate shape features of the image data of the transmission line.

[0189] The first feature and the fifth feature are concatenated to obtain the second feature.

[0190] In one possible implementation, the third training module 4023 includes:

[0191] Based on the third module in the feature extraction module, the second feature is processed sequentially by the first inverse residual module and the second inverse residual module to obtain the third feature.

[0192] In one possible implementation, the feature set includes a first feature, a second feature, and a third feature; wherein the first feature represents the shallow texture features of the image data of the transmission line; the second feature represents the mid-level shape features of the image data of the transmission line; and the third feature represents the deep semantic features of the image data of the transmission line.

[0193] Training module 402 includes:

[0194] Based on the output module of the initial neural network, the first feature, the second feature, and the third feature are concatenated to obtain the concatenated feature; wherein, the concatenated feature represents the multi-dimensional features of the image data of the transmission line.

[0195] Based on the output module in the initial neural network, the spliced ​​features are sequentially processed by 1×1 convolution, pooling, and fully connected layers to obtain the transmission line icing detection model.

[0196] In one possible implementation, prior to training module 402, the following is also included:

[0197] The image enhancement module 403 is used to perform image enhancement processing on the training data to obtain a set of processed images.

[0198] This embodiment provides a training device for a transmission line icing detection model, which can execute the method provided in the above-described method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0199] Figure 5 This is a schematic diagram of the structure of a detection device for icing type of transmission lines provided in an embodiment of this application, as shown below. Figure 5 As shown, the transmission line icing detection device 50 provided in this embodiment includes:

[0200] The acquisition module 501 is used to acquire image data of the transmission line;

[0201] Processing module 502 is used to input image data into the transmission line icing detection model, and process the image data based on the feature extraction module in the transmission line icing detection model to obtain a feature set;

[0202] The detection module 503 is used as the output module based on the transmission line icing detection model to process the feature set and obtain the detection result of the icing type of the transmission line.

[0203] The transmission line icing detection model is trained using the training methods described in the various possible embodiments above.

[0204] In one possible implementation, the processing module 502 includes:

[0205] Image data is input into the transmission line icing detection model for processing. Based on the first module in the feature extraction module, the image data is processed to obtain the first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line.

[0206] Based on the second module in the feature extraction module, the first feature is processed to obtain the second feature; wherein, the second feature represents the mid-level shape features of the image data of the transmission line;

[0207] Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line;

[0208] The feature set is determined based on the first feature, the second feature, and the third feature.

[0209] This embodiment provides a detection device for icing type of transmission lines, which can perform the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0210] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus 604.

[0211] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0212] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0213] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0214] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0215] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0216] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0217] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0218] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0219] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0220] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0221] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0222] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0223] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0224] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0225] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A training method for a transmission line icing detection model, characterized in that, include: Image data of transmission lines is acquired; and image labels of the image data are smoothed to obtain training data; wherein, the image data includes images of transmission lines and image labels; the image labels indicate the icing type of the image data; The training data is input into an initial neural network for processing. The feature extraction module in the initial neural network performs feature extraction on the training data to obtain a feature set. The output module in the initial neural network processes the feature set to obtain a transmission line icing detection model. The feature set represents the feature data set of different dimensions of the image data of the transmission line; the icing detection model of the transmission line is used to process the image data of the transmission line to obtain the detection result of the icing type of the transmission line.

2. The method according to claim 1, characterized in that, The training data is input into an initial neural network for processing, and feature extraction is performed on the training data based on the feature extraction module in the initial neural network to obtain a feature set, including: The training data is input into an initial neural network for processing. Based on the first module in the feature extraction module, the training data is processed to obtain a first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line. Based on the second module in the feature extraction module, the first feature is processed to obtain the second feature; wherein, the second feature represents the mid-layer shape features of the image data of the transmission line; Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line; The feature set is determined based on the first feature, the second feature, and the third feature.

3. The method according to claim 2, characterized in that, Based on the first module in the feature extraction module, the training data is processed to obtain the first feature, including: Based on the first module in the feature extraction module, the training data is sequentially processed by 3×3 convolution, the first fusion inverse residual module, the second fusion inverse residual module, and the third fusion inverse residual module to obtain a fourth feature map; wherein, the fourth feature represents the primary shallow texture features of the image data of the transmission line. The fourth feature map is processed based on the Laplacian operator to obtain the first feature.

4. The method according to claim 2, characterized in that, Based on the second module in the feature extraction module, the first feature is processed to obtain the second feature, including: Based on the second module in the feature extraction module, the first feature is sequentially processed by 1×1 convolution, 3×3 depthwise separable convolution, coordinate attention mechanism processing, and 1×1 convolution to obtain the fifth feature; wherein, the fifth feature represents the primary mid-level shape features of the image data of the transmission line. The first feature and the fifth feature are concatenated to obtain the second feature.

5. The method according to claim 2, characterized in that, Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature, including: Based on the third module in the feature extraction module, the second feature is processed sequentially by the first inverse residual module and the second inverse residual module to obtain the third feature.

6. The method according to any one of claims 1-5, characterized in that, The feature set includes a first feature, a second feature, and a third feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line; the second feature represents the mid-level shape features of the image data of the transmission line; and the third feature represents the deep semantic features of the image data of the transmission line. The feature set is processed based on the output module of the initial neural network to obtain a transmission line icing detection model, including: Based on the output module in the initial neural network, the first feature, the second feature, and the third feature are spliced ​​together to obtain spliced ​​features; wherein, the spliced ​​features characterize the multi-dimensional features of the image data of the transmission line. Based on the output module in the initial neural network, the spliced ​​features are sequentially processed by 1×1 convolution, pooling, and fully connected layers to obtain the transmission line icing detection model.

7. The method according to any one of claims 1-5, characterized in that, Before inputting the training data into an initial neural network for processing, and performing feature extraction processing on the training data based on the feature extraction module in the initial neural network to obtain a feature set, the process further includes: The training data is subjected to image enhancement processing to obtain a processed image set.

8. A method for detecting icing types on transmission lines, characterized in that, include: Acquire image data of power transmission lines; The image data is input into the transmission line icing detection model, and the image data is processed based on the feature extraction module in the transmission line icing detection model to obtain a feature set; Based on the output module of the transmission line icing detection model, the feature set is processed to obtain the detection result of the icing type of the transmission line, wherein the transmission line icing detection model is trained by the training method of any one of claims 1-7.

9. The method according to claim 8, characterized in that, The image data is input into the transmission line icing detection model. Based on the feature extraction module in the transmission line icing detection model, the image data is processed to obtain a feature set, including: The image data is input into the transmission line icing detection model for processing. Based on the first module in the feature extraction module, the image data is processed to obtain a first feature; wherein, the first feature represents the shallow texture features of the image data of the transmission line. Based on the second module in the feature extraction module, the first feature is processed to obtain the second feature; wherein, the second feature represents the mid-layer shape features of the image data of the transmission line; Based on the third module in the feature extraction module, the second feature is processed to obtain the third feature; wherein, the third feature represents the deep semantic features of the image data of the transmission line; The feature set is determined based on the first feature, the second feature, and the third feature.

10. A training device for a transmission line icing detection model, characterized in that, include: An acquisition module is used to acquire image data of transmission lines; and to perform label smoothing on the image labels of the image data to obtain training data; wherein, the image data includes images of transmission lines and image labels; the image labels indicate the icing type of the image data; The training module is used to input the training data into an initial neural network for processing, and to perform feature extraction processing on the training data based on the feature extraction module in the initial neural network to obtain a feature set; and to process the feature set based on the output module in the initial neural network to obtain a transmission line icing detection model. The feature set represents the feature data set of different dimensions of the image data of the transmission line; the icing detection model of the transmission line is used to process the image data of the transmission line to obtain the detection result of the icing type of the transmission line.

Images