Method and device for ice risk assessment of power transmission line based on ice type recognition

Abstract

The present application relates to the field of power transmission line icing disaster prevention, and specifically provides a power transmission line icing risk assessment method and device based on icing type identification. The method comprises: obtaining a target image representing the icing state of the power transmission line, and meteorological data corresponding to the target image; inputting the meteorological data into a first feature extraction sub-model to obtain a first target feature vector; inputting the target image into a second feature extraction sub-model to obtain a second target feature vector; the second feature extraction sub-model comprises a plurality of mixed attention enhancement modules and a plurality of efficient channel attention modules; inputting the first target feature vector and the second target feature vector into a fusion feature network to obtain an icing type corresponding to the target image and a confidence probability corresponding to the icing type; and determining the icing risk level of the power transmission line according to the icing type. The technical solution provided by the present application can improve the accuracy of the risk assessment of the power transmission line caused by icing.

Description

Technical Field

[0001] This invention relates to the field of transmission line icing disaster prevention technology, and in particular to a method and apparatus for assessing transmission line icing risk based on icing type identification. Background Technology

[0002] Ice accumulation not only significantly increases the weight of power transmission lines but can also alter their shape and mechanical properties, posing a serious threat to the safe and stable operation of transmission lines. In related technologies, ice accumulation detection methods primarily employ deep learning to identify the visual features of ice accumulation. However, this method, which focuses solely on the visual characteristics of ice accumulation, is highly susceptible to misjudging complex backgrounds and noise interference as ice accumulation, or failing to distinguish between different types of ice accumulation with similar shapes, thus leading to errors in the risk assessment of transmission lines caused by ice accumulation.

[0003] Therefore, existing technologies suffer from the technical problem of low accuracy in assessing the risks to transmission lines caused by icing. Summary of the Invention

[0004] The present invention provides a method, apparatus, electronic device, storage medium, and computer program product for assessing the risk of icing on transmission lines based on icing type identification, which can improve the accuracy of risk assessment of transmission lines caused by icing to a certain extent.

[0005] In a first aspect, the present invention provides a method for assessing the icing risk of transmission lines based on icing type identification, the method comprising:

[0006] Acquire a target image characterizing the icing state of a transmission line, and meteorological data corresponding to the target image; wherein the meteorological data includes ambient temperature data and ambient humidity data;

[0007] The meteorological data is input into the first feature extraction sub-model to obtain the first target feature vector;

[0008] The target image is input into the second feature extraction sub-model to obtain the second target feature vector; wherein, the second feature extraction sub-model includes multiple hybrid attention enhancement modules and multiple efficient channel attention modules; each of the hybrid attention enhancement modules includes at least one global context space attention module;

[0009] The first target feature vector and the second target feature vector are input into a fusion feature network to obtain the icing type corresponding to the target image and the confidence probability corresponding to the icing type.

[0010] The icing risk level of the transmission line is determined based on the icing type.

[0011] Acquire a target image characterizing the icing state of a transmission line; wherein, the target image is printed with environmental parameters of the environment in which the transmission line is located;

[0012] Environmental parameters on the target image are identified to obtain meteorological data corresponding to the target image.

[0013] In some embodiments of the present invention, the meteorological data is input into a first feature extraction sub-model to obtain a first target feature vector, including:

[0014] The meteorological data is mapped to a first original feature vector through the first fully connected layer;

[0015] The first original feature vector is mapped to a first high-dimensional feature vector through a second fully connected layer; wherein the dimension of the first high-dimensional feature vector is greater than that of the first original feature vector.

[0016] The first high-dimensional feature vector is reduced to a first transformed feature vector by a third fully connected layer; wherein the first transformed feature vector has the same dimension as the first original feature vector.

[0017] The first original feature vector and the first transformed feature vector are residually concatenated to obtain the first target feature vector.

[0018] In some embodiments of the present invention, the target image is input into a second feature extraction sub-model to obtain a second target feature vector, including:

[0019] Convolutional feature extraction, normalization, and activation function operations are performed on the target image to obtain an initial visual feature matrix;

[0020] The spatial features of the initial visual feature matrix are extracted by multiple hybrid attention enhancement modules to obtain the target spatial enhancement matrix; wherein the multiple hybrid attention enhancement modules use different convolution kernels; each hybrid attention enhancement module is reused at least once;

[0021] The semantic features of the target space enhancement matrix are extracted by multiple efficient channel attention modules to obtain a high-level semantic feature matrix; wherein the multiple efficient channel attention modules use different convolution kernels; each efficient channel attention module is used at least once;

[0022] The high-level semantic feature matrix is ​​subjected to convolutional compression to obtain the second target feature vector.

[0023] In some embodiments of the present invention, spatial features of the initial visual feature matrix are extracted by multiple hybrid attention enhancement modules to obtain a target spatial enhancement matrix, including:

[0024] Based on the initial visual feature matrix, channel attention weights are extracted to obtain the first channel attention vector, and the initial visual feature matrix is ​​multiplied by the first channel attention vector to obtain the current channel enhancement feature matrix;

[0025] Feature extraction is performed based on the current channel enhancement feature matrix to obtain the spatial attention weight matrix, and the current channel enhancement feature matrix is ​​multiplied by the spatial attention weight matrix to obtain the current spatial enhancement feature matrix;

[0026] Based on the convolution kernels corresponding to each hybrid attention enhancement module, repeat the steps of iterating the current channel enhancement feature matrix and the current spatial enhancement feature matrix until the preset number of iterations is reached to obtain the target spatial enhancement matrix.

[0027] In some embodiments of the present invention, channel attention weights are extracted based on the initial visual feature matrix to obtain a first channel attention vector, and the initial visual feature matrix is ​​multiplied by the first channel attention vector to obtain a current channel enhancement feature matrix, including:

[0028] Extract the global features from the initial visual feature matrix, normalize the global features using the L2 norm, and multiply the normalized result by the embedding weights to obtain intermediate features;

[0029] Calculate the L2 norm of the channel dimension for the intermediate features, and divide the L2 norm by... The target norm is obtained, and the intermediate feature is divided by the target norm to obtain the target intermediate feature;

[0030] The result of multiplying the target intermediate features with the gate weights, adding the gate bias, and then passing the tanh activation function, is used to obtain the gate weight vector representing the gate weights of each channel.

[0031] Multiply the initial visual feature matrix by the gating weight vector to obtain a weighted feature, and add the weighted feature to the initial visual feature to obtain the first channel attention vector;

[0032] The initial visual feature matrix is ​​multiplied by the first channel attention vector to obtain the current channel enhancement feature matrix.

[0033] In some embodiments of the present invention, feature extraction is performed based on the current channel enhancement feature matrix to obtain a spatial attention weight matrix, and the current channel enhancement feature matrix is ​​multiplied by the spatial attention weight matrix to obtain a current spatial enhancement feature matrix, including:

[0034] The current channel enhancement feature matrix is ​​subjected to average pooling and max pooling in the channel dimension to obtain the average pooling feature matrix and the max pooling feature matrix.

[0035] The average pooling feature matrix and the max pooling feature matrix are concatenated to obtain the target pooling feature matrix.

[0036] The target pooling features are processed through two layers of two-dimensional convolution operations to obtain a spatial attention weight matrix; wherein, the first convolution operation is followed by a ReLU activation function, and the second convolution operation is followed by a Sigmoid activation function.

[0037] Multiply the current channel enhancement feature matrix by the spatial attention weight matrix to obtain the current spatial enhancement feature matrix.

[0038] In some embodiments of the present invention, semantic features of the target space enhancement matrix are extracted through multiple efficient channel attention modules to obtain a high-level semantic feature matrix, including:

[0039] The target space enhancement matrix is ​​subjected to global average pooling in each channel dimension to obtain the second channel feature vector;

[0040] Perform a one-dimensional convolution operation on the second channel feature vector to obtain the target second channel feature vector;

[0041] The target second channel feature vector is transformed into a target second channel weight vector by using the Sigmoid activation function;

[0042] Multiply the target space enhancement matrix by the target second channel weight vector to obtain the high-level semantic feature matrix.

[0043] In some embodiments of the present invention, the first target feature vector and the second target feature vector are input into a fusion feature network to obtain the icing type corresponding to the target image and the probability corresponding to the icing type, including:

[0044] The first target feature vector and the second target feature vector are concatenated to obtain the target feature vector;

[0045] The target feature vector is processed by two fully connected layers to obtain the ice type corresponding to the target image and the probability corresponding to the ice type.

[0046] In some embodiments of the present invention, determining the icing risk level of the transmission line based on the icing type includes:

[0047] When the icing type is frost, the icing risk level of the transmission line is determined to be high risk.

[0048] In the case where the icing type is mixed frost, the icing risk level of the transmission line is determined to be medium risk.

[0049] When the icing type is rime, the icing risk level of the transmission line is determined to be low risk.

[0050] When the icing type is no icing, the icing risk level of the transmission line is determined to be a safe level.

[0051] Secondly, the present invention provides a transmission line icing risk assessment device based on icing type identification, the transmission line icing risk assessment device based on icing type identification comprising:

[0052] The data acquisition module is used to acquire a target image representing the icing state of the transmission line, and meteorological data corresponding to the target image; wherein, the meteorological data includes ambient temperature data and ambient humidity data;

[0053] The first feature extraction module is used to input the meteorological data into the first feature extraction sub-model to obtain the first target feature vector;

[0054] The second feature extraction module is used to input the target image into the second feature extraction sub-model to obtain the second target feature vector; wherein, the second feature extraction sub-model includes multiple hybrid attention enhancement modules and multiple efficient channel attention modules; each of the hybrid attention enhancement modules includes at least one global context space attention module;

[0055] The feature fusion module is used to input the first target feature vector and the second target feature vector into the fusion feature network to obtain the icing type corresponding to the target image and the confidence probability corresponding to the icing type;

[0056] The risk level determination module is used to determine the icing risk level of the transmission line based on the icing type.

[0057] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the transmission line icing risk assessment method based on icing type identification as described above.

[0058] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the transmission line icing risk assessment method based on icing type identification as described in any of the preceding claims.

[0059] Fifthly, the present invention provides a computer program product, including a computer program, which, when executed by a processor, causes the computer to perform the transmission line icing risk assessment method based on icing type identification as described in any of the preceding claims.

[0060] This invention provides a method for assessing the icing risk of transmission lines based on icing type identification. First, it simultaneously acquires target images of the icing state of the transmission line and corresponding temperature and humidity meteorological data. Meteorological feature vectors and image feature vectors are extracted using a first feature extraction sub-model and a second feature extraction sub-model, respectively. Then, the dual feature vectors are input into a fusion feature network to complete feature splicing and classification, outputting the icing type and its corresponding confidence probability. Finally, the icing risk level of the transmission line is determined by combining the icing type with preset grading rules. By fusing visual images of icing transmission lines with environmental meteorological data and optimizing the attention mechanism structure of the deep learning feature extraction module, accurate identification of icing types is achieved. Simultaneously, based on the actual hazardous characteristics of different icing types, differentiated risk levels are classified, which improves the accuracy of risk assessment for transmission lines caused by icing to a certain extent. Attached Figure Description

[0061] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart illustrating a method for assessing the icing risk of transmission lines based on icing type identification, provided in an embodiment of the present invention.

[0063] Figure 2 These are partial images from an ice classification dataset provided in an embodiment of the present invention.

[0064] Figure 3 This is a network architecture diagram for icing type identification provided in an embodiment of the present invention.

[0065] Figure 4 This is a schematic diagram of the GCSA module architecture, which is a hybrid attention mechanism provided in an embodiment of the present invention.

[0066] Figure 5 This is a schematic diagram of the ECA attention mechanism module architecture provided in an embodiment of the present invention.

[0067] Figure 6 This is a Grad-CAM visualization analysis diagram provided in an embodiment of the present invention.

[0068] Figure 7 This is a confusion matrix diagram of different models provided in an embodiment of the present invention.

[0069] Figure 8 This is a comparison chart of the prediction effects of different models provided in an embodiment of the present invention.

[0070] Figure 9 This is a schematic diagram of the structure of a transmission line icing risk assessment device based on icing type identification provided in an embodiment of the present invention.

[0071] Figure 10 This is a schematic diagram of an electronic device provided according to an embodiment of the present invention. Detailed Implementation

[0072] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0073] Ice accumulation not only significantly increases the weight of power transmission lines but can also alter their shape and mechanical properties, posing a serious threat to the safe and stable operation of transmission lines. This is especially true during extreme weather events such as freezing rain, strong winds, dense fog, and blizzards, as well as under the influence of micro-topography and micro-meteorological factors, leading to a series of serious accidents including conductor icing, galloping, line breakage, and tower collapse.

[0074] When transmission lines become icy, a risk assessment and early warning system are needed to guide subsequent handling procedures. The icing density varies greatly among different types of transmission lines. Higher icing density results in a greater additional load on the line for the same volume, posing a greater threat to the safe operation of the transmission line. Based on the appearance and shape of the icing, it is mainly classified into rime, mixed frost, and hoarfrost. Rime density ranges from 0.7 to 0.9 g / cm³. 3 It appears transparent and has extremely strong adhesion, making it difficult to peel off. Frost is the most threatening type of icing. The density of mixed frost is approximately 0.2 g / cm³. 3 Up to 0.6 g / cm 3 Rime ice, typically composed of alternating layers of transparent and opaque ice, is a composite ice-like structure formed by the alternating freezing of rime and hoarfrost. Due to its strong adhesion and solidification properties, it easily forms large-sized ice deposits, making it second only to rime ice in terms of its destructive power. The density of hoarfrost is approximately 0.1 g / cm³. 3 up to 0.3 g / cm 3Between different types of icing (such as rime, hoarfrost, and mixed frost), the density of rime is lower than that of ice glaze, resulting in weaker adhesion and condensation, and thus posing less of a threat to the normal operation of the power system. Therefore, accurate identification and detection of different icing types on transmission lines (such as ice glaze, hoarfrost, and mixed frost) and the assessment and tiered early warning of icing risks are crucial for ensuring the safe operation of transmission lines.

[0075] In related technologies, icing detection mainly employs two types of methods. One is a physical model based on parameter discrimination criteria, such as establishing type classification curves through meteorological parameters like wind speed and temperature, or deriving critical conditions for dry and wet growth based on thermodynamic principles. However, this type of method has limitations such as vague judgment criteria and failure to consider all influencing factors.

[0076] Secondly, there are sensor-based physical models that input meteorological data and mechanical parameters into the physical model to deduce the degree of icing. However, these models cannot distinguish between different types of icing and are significantly affected by instrument aging and strong winds. In contrast, image-based methods can directly identify icing types through visual analysis and have the advantage of non-contact measurement. The development of image-based icing classification technology can be divided into two stages: traditional machine learning and deep learning methods.

[0077] With the development of machine learning, image-based icing classification technology has emerged. In the traditional machine learning stage, icing classification mainly relies on manually designed image features to complete type identification. Researchers extract texture, geometric, and grayscale features from icing images and combine them with classifiers to classify icing types and assess their severity. Such methods rely on pre-designed rules; however, when faced with complex outdoor backgrounds, camera distortion, and mixed growth of multiple icing types, the representational power of pre-designed rules drops sharply, resulting in low classification accuracy, extremely poor robustness, and difficulty in adapting to complex working conditions in actual engineering.

[0078] With the popularization of deep learning technology, intelligent icing classification methods based on convolutional neural networks are gradually replacing traditional machine learning, relying on neural networks to autonomously extract deep features from images. However, existing deep learning methods mostly focus on feature mining in a single visual dimension, only paying attention to the visual features of icing without combining meteorological environment or other multi-dimensional information. Under complex background interference, they are prone to misjudging complex backgrounds and noise interference as icing, or failing to distinguish between different types of icing with similar morphologies. At the same time, for scenarios such as blurred images under extreme weather conditions and mixed icing types, the model's feature capture ability and generalization ability are insufficient, resulting in problems such as low recognition accuracy, high misjudgment rate, and difficulty in balancing real-time performance and stability. These methods cannot meet the long-term, stable, and accurate icing type monitoring needs of power systems.

[0079] Therefore, the current field of transmission line icing detection still faces two major technical bottlenecks: First, the single visual image data dimension is too thin, making it difficult to take into account the correlation between icing morphology characteristics and external environment and physical properties, resulting in one-sided feature representation; Second, there are many interference factors in complex field backgrounds, making it difficult for existing models to accurately lock the key features of icing and to achieve fine-grained classification of different types of icing such as rime, hoarfrost, and mixed icing, which directly restricts the accuracy of subsequent icing risk assessment.

[0080] Based on this, the present invention provides a method for assessing the icing risk of transmission lines based on icing type identification. First, it simultaneously acquires target images of the icing state of the transmission line and corresponding temperature and humidity meteorological data. Meteorological feature vectors and image feature vectors are extracted using a first feature extraction sub-model and a second feature extraction sub-model, respectively. Then, the dual feature vectors are input into a fusion feature network to complete feature splicing and classification, outputting the icing type and corresponding confidence probability. Finally, the icing risk level of the transmission line is determined by combining the icing type with preset grading rules. By fusing visual images of icing on transmission lines with environmental meteorological data and optimizing the attention mechanism structure of the deep learning feature extraction module, accurate identification of icing types is achieved. Simultaneously, based on the actual hazardous characteristics of different icing types, differentiated risk levels are classified, which improves the accuracy of risk assessment for transmission lines caused by icing to a certain extent.

[0081] Please see Figure 1 The present invention provides a method for assessing the icing risk of transmission lines based on icing type identification. This method may include the following steps.

[0082] Step S110: Obtain a target image representing the icing state of the transmission line, and meteorological data corresponding to the target image; wherein the meteorological data includes ambient temperature data and ambient humidity data.

[0083] In this embodiment, the target image is a high-definition visual image of the icing morphology of conductors and insulators, collected by transmission line monitoring equipment, serving as a carrier of the icing appearance characteristics. Meteorological data, including ambient temperature and humidity data, are environmental parameters that determine the formation and type of icing. Specifically, the target image can be automatically captured periodically by a high-precision imaging device, the "Ice Observation Elf," installed on the transmission line conductors. The high-resolution images can keenly capture minute changes during the icing process, making them highly suitable for detailed analysis. Meteorological data can also be derived from real-time monitoring data from the "Ice Observation Elf."

[0084] In this embodiment, please refer to Figure 2Each target image has the following information printed in the top left corner: time, BV device voltage, PC device current, LC conductor current, T temperature, H humidity, N longitude, E latitude, CSQ signal strength, and ON heating switch. The environmental parameters in the target images can be identified using Tesseract-OCR, an open-source optical character recognition engine that extracts text from images into editable text. This engine is managed and maintained by Google. Any missing temperature and humidity information was manually checked and supplemented to create the meteorological data (Metedata.json) corresponding to the power transmission line icing classification image dataset.

[0085] The temperature and humidity data is stored in JSON (JavaScript Object Notation) format. JSON is a lightweight data interchange format that represents structured data in text form, making it easy to read and parse. It consists of objects (collections of key-value pairs) and arrays (ordered collections of values), supporting various data types such as strings, numbers, booleans, and null values. Metedata.json stores meteorological data in key-value pairs, where each key represents an image name, and the value is an array storing the temperature and humidity information associated with that image. For example, the array [-2.7, 92.4] corresponding to "fogIce_1.jpg" indicates that the image has a temperature of -2.7°C and a humidity of 92.4%. This structure is concise and clear, suitable for storing and quickly querying structured data related to temperature and humidity of images. This data was used to create a multimodal transmission line icing classification dataset.

[0086] Step S120: Input the meteorological data into the first feature extraction sub-model to obtain the first target feature vector.

[0087] In this embodiment, the first feature extraction sub-model is a deep learning model used to process one-dimensional meteorological data, responsible for converting discrete temperature and humidity data into standardized feature vectors. Specifically, the first feature extraction sub-model can be a lightweight residual fully connected network (Lite-ResFC) model, a batch normalized optimized residual MLP (BN-ResMLP) model, a micro-attention residual MLP (Att-ResMLP) model, etc.

[0088] Step S130: Input the target image into the second feature extraction sub-model to obtain the second target feature vector; wherein, the second feature extraction sub-model includes multiple hybrid attention enhancement modules and multiple efficient channel attention modules; each of the hybrid attention enhancement modules includes at least one global context space attention module.

[0089] In this embodiment, the second feature extraction model is a deep learning model for processing two-dimensional image data, incorporating a hybrid attention enhancement module and a high-efficiency channel attention module to strengthen feature extraction capabilities. The hybrid attention enhancement module is a composite module that integrates global contextual space attention, taking into account both local icing details and global line morphology features. The high-efficiency channel attention module can be a lightweight channel attention module, used to quickly extract core semantic features across image channels, reducing computational load.

[0090] For details, please refer to Figure 3 In this embodiment, the second feature extraction sub-model is based on the EfficientNetV2-S model as the backbone network. The EfficientNetV2-S model includes seven different feature extraction modules (stages 0-7), three hybrid attention enhancement modules (stages 1-3), and three efficient channel attention (ECA) modules (stages 4-6). Each hybrid attention module includes a Global Context Space Attention (GCSA Fused-MBConv) module, which is an improvement on the Fused-MBConv module in the EfficientNetV2-S model. This helps the network better understand long-range dependencies in the image, enabling the model to not only focus on local details but also capture important patterns in the image at the global level, effectively reducing the interference of complex backgrounds on the classification results.

[0091] The ECA module is an improvement on the MBConv module in the EfficientNetV2-S model, featuring an enhanced attention mechanism. The ECA attention mechanism module is integrated into the MBConv module, replacing the original SE module, allowing the network to more accurately focus on feature channels that significantly contribute to the classification task.

[0092] Step S140: Input the first target feature vector and the second target feature vector into the fusion feature network to obtain the icing type corresponding to the target image and the confidence probability corresponding to the icing type.

[0093] In this embodiment, the fusion feature network is a network module that performs splicing and classification of meteorological features and image features, and is used to output the icing type and the confidence probability corresponding to the icing type.

[0094] Step S150: Determine the icing risk level of the transmission line based on the icing type.

[0095] In this embodiment, the icing risk level is a risk level classified based on the type of icing, which intuitively reflects the degree of damage to the transmission line caused by icing.

[0096] In one specific embodiment, at a monitoring point of a 220kV transmission line in a mountainous area, a target image of the conductor surface covered with a transparent ice layer was collected. The synchronous meteorological data were an ambient temperature of -2℃ and a relative humidity of 96%. The meteorological data was input into the first feature extraction sub-model to obtain a one-dimensional meteorological feature vector, and the target image was input into the second feature extraction sub-model to obtain a two-dimensional image feature vector. After the two features were fused, the icing type was identified as rime ice with a recognition probability of 98.2%. Finally, the icing risk of the line was determined to be high risk.

[0097] The above embodiments, by fusing meteorological data and visual images from two sources, fully combine environmental factors and appearance features, significantly improving the basic accuracy of icing type identification. Then, different feature extraction models are used for meteorological data and image data respectively, avoiding feature distortion caused by mixing different types of data. Furthermore, the attention module combination is optimized in the image processing module, taking into account both spatial and channel feature extraction, thereby improving the accuracy of icing category identification. Further, based on the icing type, the fault risk level of transmission lines is determined, which also improves the accuracy of risk assessment of transmission lines caused by icing to a certain extent, providing support for real-time early warning.

[0098] In some embodiments, step S120, inputting the meteorological data into the first feature extraction sub-model to obtain the first target feature vector, may include the following steps.

[0099] Step S121: Map the meteorological data into a first original feature vector through the first fully connected layer.

[0100] Step S122: Map the first original feature vector to a first high-dimensional feature vector through the second fully connected layer; wherein the dimension of the first high-dimensional feature vector is greater than that of the first original feature vector;

[0101] Step S123: The first high-dimensional feature vector is reduced to a first transformed feature vector by passing through the third fully connected layer; wherein the first transformed feature vector has the same dimension as the first original feature vector.

[0102] Step S124: Perform a residual concatenation between the first original feature vector and the first transformed feature vector to obtain the first target feature vector.

[0103] In this embodiment, the fully connected layer is a network layer in deep learning that implements feature mapping, responsible for mapping low-dimensional data to high-dimensional features, or reducing the dimensionality of high-dimensional features.

[0104] For details, please refer to Figure 3In this embodiment, the meteorological data feature branch network mainly processes meteorological data (temperature and humidity) step by step through three fully connected layers, and optimizes the feature extraction process through residual connections. The meteorological data is first mapped to a 64-dimensional space through the first fully connected layer, and then undergoes nonlinear transformation and regularization through the ReLU activation function and Dropout layer, resulting in the first original feature vector. Then, the features are further mapped to a 128-dimensional space through a second fully connected layer, and processed using ReLU activation and Dropout, resulting in the first high-dimensional feature vector. Finally, the third fully connected layer reduces the feature dimension back to 64 dimensions, and adds the output of the first layer to the output of the last layer through a residual connection, thus forming the first target feature vector. .

[0105] The above embodiments achieve higher accuracy, training speed, and parameter efficiency by replacing traditional MLP with residual-based MLP. Simultaneously, residual connections help the model better utilize original input features during training, improving its generalization ability and avoiding overfitting, which is of great significance for meteorological data feature extraction tasks.

[0106] In some embodiments, step S130, inputting the target image into the second feature extraction sub-model to obtain the second target feature vector, may include the following steps.

[0107] Step S131: Perform convolution feature extraction, normalization, and activation function operations on the target image to obtain an initial visual feature matrix.

[0108] Step S132: Extract spatial features from the initial visual feature matrix through multiple hybrid attention enhancement modules to obtain the target spatial enhancement matrix; wherein, the multiple hybrid attention enhancement modules use different convolution kernels; each hybrid attention enhancement module is used at least once.

[0109] Step S133: Extract the semantic features of the target space enhancement matrix through multiple efficient channel attention modules to obtain a high-level semantic feature matrix; wherein, the multiple efficient channel attention modules use different convolution kernels; each efficient channel attention module is used at least once.

[0110] Step S134: Perform convolutional compression on the high-level semantic feature matrix to obtain the second target feature vector.

[0111] In this embodiment, convolutional feature extraction, normalization, and activation function operations constitute the three steps of image preprocessing. Convolution extracts basic texture features, normalization unifies the data scale, and the activation function introduces nonlinear features to obtain an initial visual feature matrix. The initial visual feature matrix is ​​the basic feature matrix after image preprocessing, containing the basic appearance and texture features of icing.

[0112] In this embodiment, the target spatial enhancement matrix is ​​a feature matrix with enhanced spatial features after processing by the hybrid attention enhancement module, highlighting the spatial morphology of the icy region. Different hybrid attention modules employ convolutional kernels of different sizes (e.g., 3×3, 5×5) to accommodate both large- and small-scale icing feature extraction.

[0113] In this embodiment, the high-level semantic feature matrix is ​​a high-order feature containing the core semantics of icing type after being processed by the efficient channel attention module.

[0114] Specifically, in this embodiment, please refer to Figure 3 The target image is subjected to 2D convolution, batch normalization, and ReLU activation to obtain an initial visual feature matrix. A hybrid attention enhancement module with three different convolutional kernels is then used, with each kernel iterating 2, 4, and 4 times respectively to extract spatial features, resulting in a target spatial enhancement matrix. Next, a high-efficiency channel attention module is used with convolutional kernels of different sizes, iterating 6, 9, and 15 times respectively to extract semantic features, resulting in a high-level semantic feature matrix. Finally, the matrix is ​​convolved and compressed using 2D convolution, batch normalization, and ReLU activation to obtain the second target feature vector.

[0115] The above embodiments, by combining channel attention and spatial attention mechanisms, enable sufficient optimization of features in both the channel and spatial dimensions, thereby enhancing the overall expressive power of the model. Then, a hybrid attention mechanism, the GCSA module, is proposed to improve the Fused-MBConv module. This attention-enhanced GCSA Fused-MBConv module helps the network better understand long-range dependencies in images, allowing the model to not only focus on local details but also capture important patterns in the image at the global level, effectively reducing the interference of complex backgrounds on classification results. Finally, the MBConv module is improved by designing an attention-enhanced ECA MBConv module. Integrating the ECA attention mechanism module into the MBConv module replaces the original SE module, enabling the network to more accurately focus on feature channels that significantly contribute to the classification task.

[0116] In some embodiments, step S132, extracting spatial features of the initial visual feature matrix through multiple hybrid attention enhancement modules to obtain the target spatial enhancement matrix may include the following steps.

[0117] Step S1321: Extract channel attention weights based on the initial visual feature matrix to obtain the first channel attention vector, and multiply the initial visual feature matrix with the first channel attention vector to obtain the current channel enhancement feature matrix.

[0118] Step S1322: Perform feature extraction based on the current channel enhancement feature matrix to obtain the spatial attention weight matrix, and multiply the current channel enhancement feature matrix with the spatial attention weight matrix to obtain the current spatial enhancement feature matrix.

[0119] Step S1323: Repeat the steps of iterating the current channel enhancement feature matrix and the current spatial enhancement feature matrix according to the convolution kernels corresponding to each hybrid attention enhancement module until the preset number of iterations is reached to obtain the target spatial enhancement matrix.

[0120] In this embodiment, channel attention weight extraction involves extracting weight coefficients along the image channel dimension to distinguish between effective and redundant feature channels. The current channel enhancement feature matrix is ​​the matrix resulting from multiplying the original features by the channel weights, thus enhancing the channel features. The spatial attention weight matrix is ​​a weight matrix along the spatial dimension, highlighting the core icing region and minimizing background interference. The current spatial enhancement feature matrix is ​​the matrix resulting from multiplying the channel enhancement features by the spatial weights, thus enhancing the spatial features.

[0121] In this embodiment, the number of iterations refers to the number of times the channel-spatial attention weighting is repeated, gradually strengthening the core features. Specifically, for example, for the initial visual feature matrix, channel attention weights are first extracted to obtain channel attention vectors, which are then multiplied with the original features to obtain a channel enhancement matrix; then, spatial weight matrices are extracted from the channel enhancement matrix and multiplied to obtain a spatial enhancement matrix; the above iterations are repeated twice using a 3×3 convolution kernel and twice using a 5×5 convolution kernel, finally obtaining a target spatial enhancement matrix that takes into account multiple scales, significantly enhancing the features of the icing area and weakening the interference of the line background.

[0122] The above embodiments design an attention-enhanced fusion moving convolutional block GCSA Fused-MBConv module. GCSA is integrated into the Fused-MBConv module to address the limitations of using the Fused-MBConv module alone when processing global information with long-range dependencies. GCSA, by integrating local details and global contextual information, helps the network better understand long-range dependencies in images. Since icing on transmission lines often exhibits a continuous distribution along the power line, GCSA aggregates global contextual information in each channel by embedding global context, helping to avoid local ambiguity and capture the transmission line's orientation. Combined with the local icing features extracted by the convolutional layers, a long-range feature association network is formed. For transmission line images with complex backgrounds, GCSA, through a dynamic weighting strategy, can automatically enhance the sharp local features of icing edges (such as icing texture and transparency) and the global geometric features of the overall conductor orientation. To address image blurring and contrast issues caused by varying weather conditions and day-night cycles, GCSA enhances the sensitivity to local features (such as abrupt changes in brightness at the boundary between iced areas and the background of transmission lines) and the adaptability to global illumination distribution features (such as day-night illumination variations). The Fused-MBConv module originally focused on feature extraction from local receptive fields, extracting local image information through efficient convolution, making it particularly suitable for tasks requiring detailed processing of local details. However, using the Fused-MBConv module alone can be limiting when processing global information with long-range dependencies. Therefore, GCSA is introduced to address this deficiency. Especially when facing complex backgrounds, combining global and local information improves the network's classification performance for iced transmission line images. GCSA combines global information with channel attention to adjust channel relationships and enhances local feature learning capabilities through spatial attention. Integrating GCSA into the Fused-MBConv module allows the model to not only focus on local image details but also capture important global features. This approach effectively reduces noise from complex environments unrelated to icing and enhances the network's stability and accuracy when handling complex backgrounds. The introduction of this module enables it to stably and effectively perform feature extraction and classification tasks under various complex backgrounds.

[0123] In some embodiments, step S1321 involves extracting channel attention weights based on the initial visual feature matrix to obtain a first channel attention vector, and multiplying the initial visual feature matrix with the first channel attention vector to obtain a current channel enhancement feature matrix. This may include the following steps.

[0124] Step S13211: Extract the global features of the initial visual feature matrix, normalize the global features using the L2 norm, and multiply the normalized result with the embedding weights to obtain intermediate features.

[0125] Step S13212: Calculate the L2 norm of the channel dimension for the intermediate feature, and divide the L2 norm by... The target norm is obtained, and the intermediate feature is divided by the target norm to obtain the target intermediate feature.

[0126] Step S13213: Multiply the target intermediate features by the gate weights, add the gate bias, and pass the tanh activation function to obtain the gate weight vector representing the gate weights of each channel.

[0127] Step S13214: Multiply the initial visual feature matrix with the gating weight vector to obtain a weighted feature, and add the weighted feature to the initial visual feature to obtain the first channel attention vector.

[0128] Step S13215: Multiply the initial visual feature matrix with the first channel attention vector to obtain the current channel enhancement feature matrix.

[0129] In this embodiment, L2 norm normalization is used to standardize the global features, unifying the feature numerical scale and avoiding weight imbalance caused by numerical differences. For details, please refer to [link to relevant documentation]. Figure 4 Let the initial visual features be... , , , L2 norm normalization of global features can be expressed as: H and W are the height and width of the feature map, respectively, and C is the number of channels. For each channel c, the intermediate feature S of that channel is calculated. c :

[0130]

[0131] in, For embedding weights, it is responsible for controlling the weight of each channel; For each channel of x, ε is a minimal constant to avoid the problem of finding the derivative at the zero point.

[0132] In this embodiment, the target intermediate features are normalized using L2 standardization for cross-channel operation. , represents an input feature vector with C channels, and its normalized output is calculated using the following formula:

[0133]

[0134] Where ϵ is a local constant. Introducing... Used for standardization This avoids the problem of the value being too small when the number of channels C is large. Compared with the fully connected layer used in the SE module, this channel normalization method has a lower computational cost.

[0135] In this embodiment, the gating adaptation mechanism aims to dynamically adjust the weights of each channel, controlling the relative importance of feature maps across different channels. Through the gating adaptation mechanism, the responses of each channel can be adaptively adjusted according to the input features, enhancing the representational power of important features and suppressing unimportant features. When the gating weights of a channel are positively activated, the attention mechanism promotes competition between that channel and other channels; conversely, the attention mechanism encourages cooperation between that channel and other channels. The gating adaptation mechanism introduces a 1+tanh(x) gate to avoid problems such as gradient vanishing caused by the Sigmoid function, making the training process more stable. The tanh activation function maps the gating weights to the [-1,1] interval, achieving flexible control of channel weights. Gating bias Responsible for adjusting the scale of the input feature channels, the channel attention output is , Z1 is obtained by element-wise multiplication with the input x, and then fed into the spatial attention mechanism:

[0136]

[0137]

[0138] In the formula, This indicates that the two are multiplied together.

[0139] It should be noted that embedding weights, gated weights, and gated biases are trainable parameters of the model, used to optimize the accuracy of channel weight extraction.

[0140] In some embodiments, in step S1322, feature extraction is performed based on the current channel enhancement feature matrix to obtain a spatial attention weight matrix, and the current channel enhancement feature matrix is ​​multiplied by the spatial attention weight matrix to obtain a current spatial enhancement feature matrix, which may include the following steps.

[0141] Step S13221: Perform average pooling and max pooling on the current channel enhancement feature matrix in the channel dimension to obtain the average pooling feature matrix and the max pooling feature matrix.

[0142] Step S13222: Concatenate the average pooling feature matrix and the max pooling feature matrix to obtain the target pooling feature matrix.

[0143] Step S13223: The target pooling features are processed through two layers of two-dimensional convolution operations to obtain a spatial attention weight matrix; wherein, the first convolution operation is followed by a ReLU activation function, and the second convolution operation is followed by a Sigmoid activation function.

[0144] Step S13224: Multiply the current channel enhancement feature matrix with the spatial attention weight matrix to obtain the current spatial enhancement feature matrix.

[0145] In this embodiment, average pooling and max pooling are two pooling operations. Average pooling extracts overall spatial features, while max pooling extracts prominent local features. Then, the average pooling feature matrices and max pooling feature matrices obtained from average pooling and max pooling are concatenated to combine the advantages of both pooling methods, resulting in the target pooling feature matrix.

[0146] In this embodiment, the ReLU activation function is activated after the first convolution layer to remove negative features and enhance positive features. The Sigmoid activation function is activated after the second convolution layer to map the spatial weights to the [0,1] interval, thereby achieving spatial weight standardization.

[0147] For details, please refer to Figure 4 Spatial attention employs a two-branch spatial attention mechanism. First, average pooling and max pooling layers are used to extract spatial information, resulting in feature maps. , Then splice them along the channel dimension. and Obtain feature map :

[0148] ,

[0149] ,

[0150] .

[0151] In the formula: , These represent the height and width of the feature map, respectively. , ; This is the i-th channel of the input feature map. It then passes through a 3×3 convolutional layer, a ReLU activation layer, another 3×3 convolution, and finally a Sigmoid activation layer. .

[0152] In some embodiments, step S133, extracting semantic features of the target space enhancement matrix through multiple efficient channel attention modules to obtain a high-level semantic feature matrix, may include the following steps.

[0153] Step S1331: Perform global average pooling on each channel dimension of the target space enhancement matrix to obtain the second channel feature vector.

[0154] Step S1332: Perform a one-dimensional convolution operation on the second channel feature vector to obtain the target second channel feature vector.

[0155] Step S1333: The target second channel feature vector is transformed into the target second channel weight vector by using the Sigmoid activation function.

[0156] Step S1334: Multiply the target space enhancement matrix with the target second channel weight vector to obtain the high-level semantic feature matrix.

[0157] In this embodiment, global average pooling is performed on each channel, compressing the channel features into a one-dimensional vector. The one-dimensional convolution operation is a lightweight convolution, extracting local correlation features from the channel vectors to obtain the target second channel feature vector. Then, this target second channel feature vector is passed through a sigmoid activation function to obtain the target second channel weight vector, which is used to weight the target space augmentation matrix. Finally, the target space augmentation matrix is ​​multiplied by the target second channel weight vector to obtain the high-level semantic feature matrix.

[0158] For details, please refer to Figure 5 This paper proposes an attention-enhanced moving convolutional module, ECA MBConv, to replace the original Squeeze and Excitation (SE) module. Unlike the traditional SE module, the ECA module removes fully connected layers and uses 1D convolution operations to directly capture channel dependencies on the globally averaged feature map. Through a more flexible channel attention mechanism, the ECA module enables the model to better identify the contribution of different channels to the classification results. This reduces noise interference from irrelevant information in complex backgrounds when processing iced images, highlighting important features. This improvement not only enhances classification accuracy but also reduces the model's computational resource requirements. The ECA MBConv module enhances the model's ability to extract features from iced areas and optimizes computational efficiency, providing a more accurate and efficient solution for transmission line icing classification tasks, achieving superior performance.

[0159] In some embodiments, step S140, inputting the first target feature vector and the second target feature vector into a fusion feature network to obtain the icing type corresponding to the target image and the probability corresponding to the icing type, may include the following steps.

[0160] Step S141: Concatenate the first target feature vector and the second target feature vector to obtain the target feature vector.

[0161] Step S142: Process the target feature vector through two fully connected layers to obtain the ice type corresponding to the target image and the probability corresponding to the ice type.

[0162] In this embodiment, the first target feature vector (meteorological data) and the second target feature vector (image data) are merged along the dimension to form a fused feature. The target feature vector is the complete feature vector after the fusion of the two source data, containing all the core information of the meteorological and image data. The two fully connected layers are used for optimization processing before classification of the fused feature. The first layer extracts the association of the fused features, and the second layer outputs the icing type and probability.

[0163] Specifically, in this embodiment, please refer to Figure 3 The model combines image and meteorological features and processes them through two fully connected layers. First, image and geographic features are concatenated and then input into a fully connected layer, passing through a ReLU activation function and a Dropout regularization layer to improve non-linear fitting and prevent overfitting. The fused features are then passed through another fully connected layer, and finally through a Softmax layer to output the final classification result and obtain a score for each category. This structure effectively integrates multimodal information, improving the model's performance in complex tasks.

[0164] In some embodiments, determining the icing risk level of the transmission line in step S150 based on the icing type may include the following steps.

[0165] Step S151: If the icing type is frost, the icing risk level of the transmission line is determined to be high risk level.

[0166] Step S152: In the case that the icing type is mixed frost, the icing risk level of the transmission line is determined to be medium risk level.

[0167] Step S153: If the icing type is rime, the icing risk level of the transmission line is determined to be low risk level.

[0168] Step S154: If the icing type is no icing, the icing risk level of the transmission line is determined to be a safe level.

[0169] In this embodiment, the icing risk level of transmission lines is divided into four categories, corresponding to three icing types: rime, mixed rime, and hoarfrost, as well as a non-icing state. The risk level assessment and graded early warning for each type are as follows:

[0170] 1. Ice frost: High risk level

[0171] Considering the physical characteristics of frost (density 0.7 g / cm³~0.9 g / cm³, transparent, extremely adhesive, and difficult to detach), its hazards to transmission lines are mainly manifested in three aspects: First, its high density causes the weight of the icing to increase rapidly, easily exceeding the bearing capacity limits of the transmission line conductors, insulators, and towers, leading to serious accidents such as conductor breakage and tower collapse; second, its extremely strong adhesion makes it difficult for the icing to detach naturally, and prolonged adhesion accelerates conductor aging, reduces conductor mechanical strength, and may also cause insulator flashover, affecting the insulation performance of the line; third, frost often forms in extreme environments of low temperature and high humidity, and the icing growth rate is rapid, forming a thick icing layer in a short time, posing a great challenge to emergency response. Based on the above hazardous characteristics, this invention classifies frost as a high-risk level. When the system detects frost accretion on a transmission line, it immediately triggers a high-risk warning and simultaneously pushes the information to relevant system platforms and personnel. The handling requirements are as follows: maintenance personnel must arrive at the scene within 1 hour, initiate emergency de-icing operations, suspend power supply to some non-critical loads on the line, monitor the ice thickness and line operation status in real time throughout the process, and restore normal line operation only after the ice is completely removed and the risk level is reduced to a safe level.

[0172] 2. Mixed frost: Medium risk level

[0173] Mixed frost has a density between 0.2 g / cm³ and 0.6 g / cm³, and is composed of alternating layers of transparent ice (rime component) and non-transparent ice (hoarse component). It is a composite ice-like structure formed by the alternating freezing of rime and hoarse frost, exhibiting strong adhesion and solidification properties, and growing faster than both rime and hoarse frost. It easily forms large-sized ice layers, posing a hazard to transmission lines second only to rime. Its main hazards include: uneven ice layer structure leading to conductor stress imbalance, causing conductor galloping, and consequently loosening of line joints and damage to hardware; prolonged adhesion causing conductor thermal expansion and contraction imbalance, accelerating line fatigue aging, and potentially affecting the insulation performance of insulators, increasing the risk of flashover. This invention classifies mixed frost as a medium-risk level. When mixed frost icing is detected on a transmission line, a medium-risk warning is triggered, and relevant system platforms and personnel are simultaneously notified. The handling requirements are as follows: maintenance personnel must complete on-site inspections within 2 hours, monitor the ice thickness, growth rate and line operating parameters in real time, and prepare for de-icing. If the ice thickness continues to increase and approaches the high-risk threshold, emergency de-icing operations should be initiated immediately. At the same time, monitoring of the surrounding environment of the line should be strengthened to prevent secondary risks caused by ice falling off.

[0174] 3. Rime ice: Low risk level

[0175] The density of rime ice is 0.1 g / cm³ to 0.3 g / cm³, much lower than that of frost and mixed frost. It often appears radially, has weak adhesion and condensation properties, and is not easily deposited on the surface of transmission lines. Furthermore, it is easily detached naturally due to changes in ambient temperature, posing a relatively small threat to the normal operation of the power system. Its main potential hazards are: when the rime ice thickness is too large (exceeding 10 mm), it may slightly increase the weight of the conductor, affecting line stability. Also, when the temperature rises sharply, the rapid melting of rime ice may cause ice to form on the conductor surface, indirectly increasing the risk of insulator flashover. However, the probability of these hazards occurring is low, and the impact is limited. This invention classifies rime ice as a low-risk level. When rime ice is detected on a transmission line and the ice thickness reaches the warning threshold, a low-risk warning is triggered, and the warning is simultaneously pushed to the relevant system platform. The handling requirements are as follows: maintenance personnel should monitor the line operation status at the regular monitoring frequency (once every 4 hours). There is no need to start special de-icing operations. They only need to pay attention to changes in ambient temperature and prevent the risk of secondary icing after the rime melts. When the ice falls off naturally or the thickness drops below the warning threshold, the low-risk warning will be automatically lifted.

[0176] 4. No icing: Safety level

[0177] When there is no ice adhering to the surface of the transmission line, or the ice thickness is ≤1mm (negligible), it is considered to be in a non-icing state, corresponding to the safety level. At this time, the transmission line conductors, insulators, and towers are all in normal working condition, with no safety risks related to ice accumulation, and the line can operate stably. Maintenance personnel conduct routine inspections according to daily maintenance procedures, monitoring line operating parameters and the surrounding environment, promptly identifying and addressing non-icing-related safety hazards to ensure the normal and stable operation of the transmission line.

[0178] The above embodiments, with the icing type identification results as the core, combined with the physical characteristics, growth patterns and environmental factors of icing, construct a multi-indicator, differentiated risk assessment system. This system achieves accurate classification and scientific early warning of icing risks on transmission lines, providing clear handling guidelines for operation and maintenance personnel. It enables differentiated de-icing operations and safety control work, effectively avoiding resource waste, and minimizing the threat of icing disasters to the safe operation of transmission lines. This overcomes the shortcomings of existing technologies, such as inaccurate icing risk assessment, weak early warning targeting, and low handling efficiency.

[0179] In one embodiment of the present invention, the dataset is first cleaned, removing abnormal datasets such as blurry images, overexposed images, and images with obstructed cameras, and the images are renamed. Then, OCR technology is used to extract the temperature and humidity meteorological information from the upper left corner of the transmission line icing image, and manual error checking is performed. The temperature and humidity information is stored in a JSON file in the form of key-value pairs, where the key is the image name and the value is an array storing the temperature and humidity information. The dataset is then classified into four categories: bare lines, rime ice, hoarfrost, and mixed rime ice. Thus, a multimodal transmission line icing classification dataset is created. Figure 2 This is a partially classified dataset. Bare Wire represents bare wire, Rime represents rime, Mixed Rime represents mixed rime, and Glaze represents rime.

[0180] A meteorological data feature-based MFEN network is constructed. The meteorological data feature branch processes temperature and humidity data through three fully connected layers and utilizes residual connections to optimize feature extraction. The network maps meteorological data features sequentially to 64-dimensional, 128-dimensional, and 64-dimensional channels, and improves training efficiency and generalization ability through ReLU activation, Dropout, and residual connections. The residual-based MLP design significantly improves the model's accuracy and parameter efficiency, facilitating more effective extraction of meteorological data features.

[0181] To address the complex background issues often encountered in images of icing on power transmission lines, a hybrid attention mechanism, GCSA, is proposed. This module consists of two parts: channel attention and spatial attention. It aims to improve the network's robustness in complex scenes by fusing local and global information. The channel attention mechanism, combined with a global information embedding module, adaptively embeds global information into each channel and adjusts the weights of each channel before normalization. By calculating the global information embedding for each channel and normalizing it using L2 normalization, the utilization of information from a large receptive field is enhanced, thus avoiding local blurring caused by a small receptive field. Specifically, for each channel, the global information embedding is first calculated, followed by normalization and parameter control of the channel weights, thereby optimizing the model's ability to handle different channel features. Furthermore, a gating mechanism further enhances the model's adaptive modeling of inter-channel relationships. By adjusting the weights of each channel using 1 + tanh(x), the vanishing gradient problem caused by using the sigmoid function is avoided, making the training process more stable. The spatial attention mechanism extracts spatial information through a two-branch pooling operation. First, average pooling and max pooling are used to obtain two feature maps, F_max and F_avg, respectively. These two feature maps are then concatenated along the channel dimension to obtain the feature map F_cat. Next, the feature map is processed through two layers of 3×3 convolutions and ReLU activation functions, and finally, a sigmoid activation function is applied to obtain the spatial attention map F_GCSA. This spatial attention mechanism adaptively focuses on important regions in the image, adjusting channel weights by combining a gating mechanism, thereby enhancing the model's spatial perception capability. The GCSA module, through joint optimization of channels and space, effectively reduces local noise and irrelevant feature interference when processing images with complex backgrounds, improving the ability to capture important patterns. This not only enhances the model's understanding of long-range dependencies in images but also effectively improves classification accuracy and robustness in complex environments.

[0182] An improvement to the Fused-MBConv module is proposed: the GCSAFused-MBConv module, based on the GCSA hybrid attention mechanism. This invention integrates the GCSA attention module into the Fused-MBConv module. After integrating GCSA, the Fused-MBConv module not only extracts local features better but also enhances its understanding of long-range dependencies through weighted processing of the global context, thereby improving the model's robustness and accuracy in complex backgrounds. Through this improvement, the GCSAFused-MBConv module enables EfficientNetV2 to exhibit stronger stability and efficiency when processing images with complex backgrounds, reducing the interference of background noise on classification results and improving accuracy in practical applications.

[0183] An improved version of the MBConv module, the ECA MBConv module, was designed: This invention integrates the ECA module into the MBConv module, replacing the original SE attention mechanism module. The ECA module captures the dependencies between channels through 1D convolution operations, optimizing the channel attention mechanism, reducing computational complexity, and improving the model's ability to focus on important features. This improvement significantly enhances the model's learning ability in the transmission line icing classification task, strengthens the extraction of icing features, optimizes computational efficiency, and improves classification accuracy and overall performance.

[0184] Based on the GCSA Fused-MBConv and ECA MBConv modules, a feature-enhanced attention mechanism network, FAENet, is constructed. FAENet uses EfficientNetV2-S as its backbone and, through enhanced attention mechanisms, fully utilizes local and global features to improve the network's robustness and accuracy in complex environments. The GCSA Fused-MBConv module combines channel attention and spatial attention mechanisms, dynamically adjusting the weights of each channel and embedding them using global contextual information. The ECA MBConv module, by optimizing the channel attention mechanism, replaces the traditional SE module, further improving the model's ability to focus on key image features. This allows the network to more accurately focus on feature channels that contribute to the classification task, thus better identifying important features in images against complex backgrounds. By improving the Fused-MBConv and MBConv modules and proposing attention-enhanced GCSA Fused-MBConv and ECA MBConv modules, FAENet can improve classification accuracy and efficiency when handling complex tasks such as power transmission line icing. The network can not only effectively extract local features from images, but also capture key patterns in images globally, making the model significantly robust and stable when facing challenging images with complex backgrounds.

[0185] A feature fusion module, Fusion_Fc, is implemented to fuse image and geographic features. First, the image and geographic features are concatenated together, then input into a fully connected layer, followed by a ReLU activation function and a Dropout regularization layer. The fused features pass through another fully connected layer, and finally a Softmax layer to output the final classification result and obtain a score for each class. This structure effectively fuses multimodal information, improving the model's performance in complex tasks.

[0186] Finally, an MFAENet for transmission line icing is constructed by using the attention mechanism enhanced by image features, the network FAENet based on image features, the network MFEN based on meteorological data features, and the feature fusion module Fusion_Fc.

[0187] Then, the dataset is divided into training, validation, and test sets, and the training and validation sets are input into the constructed MFAENet model for training and validation. This invention uses accuracy, precision, recall, and F1 score as evaluation metrics to comprehensively assess the model's performance on classification tasks. The formulas for these metrics are shown below:

[0188] Accuracy = (TP + TN) / (TP + TN + FP + FN)

[0189] Precision = TP / (TP + FP)

[0190] Recall = TP / (TP + FN)

[0191] F1-score = 2 × (Precision × Recall) / (Precision + Recall)

[0192] Wherein, TP represents the number of positive samples that are correctly classified as positive (True Positive), that is, the number of samples correctly predicted as positive by the model; TN represents the number of negative samples that are correctly classified as negative (True Negative), that is, the number of samples correctly predicted as negative by the model; FP represents the number of negative samples that are incorrectly classified as positive (Flase Positive), that is, the number of samples incorrectly classified as positive by the model; and FN represents the number of positive samples that are incorrectly classified as negative (Flase Negative), that is, the number of samples incorrectly classified as negative by the model.

[0193] Grad-CAM is an effective method for interpreting the decision-making process of convolutional neural networks, particularly suitable for image classification tasks. This method determines which parts of the image are most important for the classification result by calculating the gradient of the score for a specific class relative to the feature map of the last convolutional layer of the network. By employing Grad-CAM to overlay heatmaps onto the original image, it's possible to observe whether the model captures features related to icing on power transmission lines during the classification process. Figure 6 As shown, compared to the baseline model EfficientNetV2, the attention-enhanced model is able to capture the features of the icing area of ​​the transmission line better.

[0194] from Figure 7 The confusion matrix shown demonstrates that the proposed model significantly improves classification performance compared to the baseline model (EfficientNetV2-S). The proposed model achieves a higher number of correct classifications across all categories compared to the baseline model, with a significantly reduced number of misclassifications, particularly in the easily confused mixed frost and hoarfrost categories. This indicates that by enhancing the attention mechanism and using multimodal meteorological data branches, the proposed model can more effectively extract and utilize features, significantly improving the classification accuracy of icing on transmission lines.

[0195] from Figure 8 The comparison chart of prediction results of different models shows that the proposed model (MFAENET) has significantly improved classification accuracy compared with the EfficientNet, MobileNet V3 and ShuffleNet models, and has higher confidence in its classification results compared with the EfficientNetV2-S model.

[0196] Please see Figure 9 One embodiment of the present invention provides a transmission line icing risk assessment device based on icing type identification. The transmission line icing risk assessment device based on icing type identification may include: a data acquisition module, a first feature extraction module, a second feature extraction module, a feature fusion module, and a risk level determination module.

[0197] The data acquisition module is used to acquire a target image representing the icing state of the transmission line, and meteorological data corresponding to the target image; wherein the meteorological data includes ambient temperature data and ambient humidity data.

[0198] The first feature extraction module is used to input the meteorological data into the first feature extraction sub-model to obtain the first target feature vector.

[0199] The second feature extraction module is used to input the target image into the second feature extraction sub-model to obtain the second target feature vector; wherein, the second feature extraction sub-model includes multiple hybrid attention enhancement modules and multiple efficient channel attention modules; each of the hybrid attention enhancement modules includes at least one global context space attention module.

[0200] The feature fusion module is used to input the first target feature vector and the second target feature vector into the fusion feature network to obtain the icing type corresponding to the target image and the confidence probability corresponding to the icing type.

[0201] The risk level determination module is used to determine the icing risk level of the transmission line based on the icing type.

[0202] The specific functions and effects of the transmission line icing risk assessment device based on icing type identification can be explained by referring to other embodiments in this specification, and will not be repeated here. Each module in the transmission line icing risk assessment device based on icing type identification can be implemented entirely or partially through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0203] Please see Figure 10 One embodiment of the present invention can provide an electronic device, the electronic device comprising:

[0204] A memory, and one or more processors communicatively connected to the memory;

[0205] The memory stores instructions that can be executed by the one or more processors, which, when executed by the one or more processors, enable the one or more processors to implement the transmission line icing risk assessment method based on icing type identification as described in any of the above embodiments.

[0206] One embodiment of the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the transmission line icing risk assessment method based on icing type identification as described in any of the above embodiments.

[0207] The embodiments of this specification also provide a computer program product containing instructions that, when executed by a computer, cause the computer to perform the transmission line icing risk assessment method based on icing type identification as described in any of the above embodiments.

[0208] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments described herein, and are not intended to limit the scope of the invention.

[0209] It is understood that in the various embodiments described in this specification, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments described in this specification.

[0210] It is understood that the various implementation methods described in this specification can be implemented individually or in combination, and the implementation methods in this specification are not limited in this respect.

[0211] Unless otherwise stated, all technical and scientific terms used in the embodiments of this specification have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this specification. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0212] It is understood that the processor in this invention can be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method implementation can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this specification. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this specification can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.

[0213] It is understood that the memory in this invention can be volatile memory or non-volatile memory, or may include both. Specifically, the non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM). It should be noted that the memory in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0214] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in the embodiments of this invention are subject to strict compliance with relevant laws, regulations, and regulatory requirements in their collection, storage, use, processing, transmission, provision, and disclosure, and adhere to the principles of legality, legitimacy, necessity, and good faith. The acquisition of relevant information and data is premised on the user's explicit consent or other legitimate reasons, and a clear and convenient authorization management approach is provided to the user, allowing the user to independently choose to consent, withdraw consent, or refuse to provide relevant information. For functions that rely on user information, if the user does not authorize or withdraws authorization, the corresponding technical function cannot be implemented, and the technical solution of this invention is not applicable in this scenario.

[0215] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this specification.

[0216] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the aforementioned method implementations, and will not be repeated here.

[0217] In the several embodiments provided in this specification, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and 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 through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0218] 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, depending on actual needs.

[0219] 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.

[0220] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of this specification, in essence, or the parts that contribute to the prior art, or parts of the technical solutions, can be embodied in the form of software products. These computer software products are stored in a storage medium and include 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 described in the various embodiments of this specification. 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.

[0221] The above description is merely a specific embodiment of this specification, but the scope of protection of this invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this specification should be included within the scope of protection of this specification. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A power transmission line icing risk assessment method based on icing type identification, characterized in that, The method includes: Acquire a target image characterizing the icing state of a transmission line, and meteorological data corresponding to the target image; wherein the meteorological data includes ambient temperature data and ambient humidity data; The meteorological data is input into the first feature extraction sub-model to obtain the first target feature vector; The target image is input into the second feature extraction sub-model to obtain the second target feature vector; wherein, the second feature extraction sub-model includes multiple hybrid attention enhancement modules and multiple efficient channel attention modules; each of the hybrid attention enhancement modules includes at least one global context space attention module; The first target feature vector and the second target feature vector are input into a fusion feature network to obtain the icing type corresponding to the target image and the confidence probability corresponding to the icing type. The icing risk level of the transmission line is determined based on the icing type.

2. The method of claim 1, wherein, The meteorological data is input into the first feature extraction sub-model to obtain the first target feature vector, including: The meteorological data is mapped to a first original feature vector through the first fully connected layer; The first original feature vector is mapped to a first high-dimensional feature vector through a second fully connected layer; wherein the dimension of the first high-dimensional feature vector is greater than that of the first original feature vector. The first high-dimensional feature vector is reduced to a first transformed feature vector by a third fully connected layer; wherein the first transformed feature vector has the same dimension as the first original feature vector. The first original feature vector and the first transformed feature vector are residually concatenated to obtain the first target feature vector.

3. The method according to claim 1, characterized in that, The target image is input into the second feature extraction sub-model to obtain the second target feature vector, including: Convolutional feature extraction, normalization, and activation function operations are performed on the target image to obtain an initial visual feature matrix; The spatial features of the initial visual feature matrix are extracted by multiple hybrid attention enhancement modules to obtain the target spatial enhancement matrix; wherein the multiple hybrid attention enhancement modules use different convolution kernels; each hybrid attention enhancement module is reused at least once; The semantic features of the target space enhancement matrix are extracted by multiple efficient channel attention modules to obtain a high-level semantic feature matrix; wherein the multiple efficient channel attention modules use different convolution kernels; each efficient channel attention module is used at least once; The high-level semantic feature matrix is ​​subjected to convolutional compression to obtain the second target feature vector.

4. The method according to claim 3, characterized in that, Spatial features of the initial visual feature matrix are extracted using multiple hybrid attention enhancement modules to obtain the target spatial enhancement matrix, including: Based on the initial visual feature matrix, channel attention weights are extracted to obtain the first channel attention vector, and the initial visual feature matrix is ​​multiplied by the first channel attention vector to obtain the current channel enhancement feature matrix; Feature extraction is performed based on the current channel enhancement feature matrix to obtain the spatial attention weight matrix, and the current channel enhancement feature matrix is ​​multiplied by the spatial attention weight matrix to obtain the current spatial enhancement feature matrix; Based on the convolution kernels corresponding to each hybrid attention enhancement module, repeat the steps of iterating the current channel enhancement feature matrix and the current spatial enhancement feature matrix until the preset number of iterations is reached to obtain the target spatial enhancement matrix.

5. The method according to claim 4, characterized in that, Channel attention weights are extracted based on the initial visual feature matrix to obtain a first channel attention vector. The initial visual feature matrix is ​​then multiplied by the first channel attention vector to obtain the current channel enhancement feature matrix, including: Extract the global features from the initial visual feature matrix, normalize the global features using the L2 norm, and multiply the normalized result by the embedding weights to obtain intermediate features; Calculate the L2 norm of the channel dimension for the intermediate features, and divide the L2 norm by... The target norm is obtained, and the intermediate feature is divided by the target norm to obtain the target intermediate feature; The result of multiplying the target intermediate features with the gate weights, adding the gate bias, and then passing the tanh activation function, is used to obtain the gate weight vector representing the gate weights of each channel. Multiply the initial visual feature matrix by the gating weight vector to obtain a weighted feature, and add the weighted feature to the initial visual feature to obtain the first channel attention vector; The initial visual feature matrix is ​​multiplied by the first channel attention vector to obtain the current channel enhancement feature matrix.

6. The method according to claim 4, characterized in that, Feature extraction is performed based on the current channel enhancement feature matrix to obtain a spatial attention weight matrix. The current channel enhancement feature matrix is ​​then multiplied by the spatial attention weight matrix to obtain the current spatial enhancement feature matrix, which includes: The current channel enhancement feature matrix is ​​subjected to average pooling and max pooling in the channel dimension to obtain the average pooling feature matrix and the max pooling feature matrix. The average pooling feature matrix and the max pooling feature matrix are concatenated to obtain the target pooling feature matrix. The target pooling features are processed through two layers of two-dimensional convolution operations to obtain a spatial attention weight matrix; wherein, the first convolution operation is followed by a ReLU activation function, and the second convolution operation is followed by a Sigmoid activation function. Multiply the current channel enhancement feature matrix by the spatial attention weight matrix to obtain the current spatial enhancement feature matrix.

7. The method according to claim 3, characterized in that, The semantic features of the target space enhancement matrix are extracted using multiple efficient channel attention modules to obtain a high-level semantic feature matrix, including: The target space enhancement matrix is ​​subjected to global average pooling in each channel dimension to obtain the second channel feature vector; Perform a one-dimensional convolution operation on the second channel feature vector to obtain the target second channel feature vector; The target second channel feature vector is transformed into a target second channel weight vector by using the Sigmoid activation function; Multiply the target space enhancement matrix by the target second channel weight vector to obtain the high-level semantic feature matrix.

8. The method according to claim 1, characterized in that, The first target feature vector and the second target feature vector are input into a fusion feature network to obtain the icing type corresponding to the target image and the probability corresponding to the icing type, including: The first target feature vector and the second target feature vector are concatenated to obtain the target feature vector; The target feature vector is processed by two fully connected layers to obtain the ice type corresponding to the target image and the probability corresponding to the ice type.

9. The method according to claim 1, characterized in that, Based on the type of icing, the icing risk level of the transmission line is determined, including: When the icing type is frost, the icing risk level of the transmission line is determined to be high risk. In the case where the icing type is mixed frost, the icing risk level of the transmission line is determined to be medium risk. When the icing type is rime, the icing risk level of the transmission line is determined to be low risk. When the icing type is no icing, the icing risk level of the transmission line is determined to be a safe level.

10. A transmission line icing risk assessment device based on icing type identification, characterized in that, The transmission line icing risk assessment device based on icing type identification includes: The data acquisition module is used to acquire a target image representing the icing state of the transmission line, and meteorological data corresponding to the target image; wherein, the meteorological data includes ambient temperature data and ambient humidity data; The first feature extraction module is used to input the meteorological data into the first feature extraction sub-model to obtain the first target feature vector; The second feature extraction module is used to input the target image into the second feature extraction sub-model to obtain the second target feature vector; wherein, the second feature extraction sub-model includes multiple hybrid attention enhancement modules and multiple efficient channel attention modules; each of the hybrid attention enhancement modules includes at least one global context space attention module; The feature fusion module is used to input the first target feature vector and the second target feature vector into the fusion feature network to obtain the icing type corresponding to the target image and the confidence probability corresponding to the icing type; The risk level determination module is used to determine the icing risk level of the transmission line based on the icing type.

11. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the transmission line icing risk assessment method based on icing type identification as described in any one of claims 1 to 9.

12. A computer storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the transmission line icing risk assessment method based on icing type identification as described in any one of claims 1 to 9.

Images