Method and system for detecting icing state of power transmission line

Abstract

The invention discloses a power transmission line icing state detection method and system, and the method comprises the steps: obtaining an original icing image of a power transmission line, and calculating the equivalent icing thickness based on the original icing image; converting the meteorological data into a parameterized expression, and performing iterative optimization on parameters in the expression to obtain a composite correction factor; performing correction optimization on the equivalent icing thickness based on a composite correction factor to obtain a corrected icing thickness; according to the method, the preliminary equivalent icing thickness is corrected and optimized by constructing the environmental meteorological element composite correction factor, so that the icing state detection precision can be ensured in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of line icing detection technology, and in particular to a method and system for detecting icing status of a power transmission line. Background Art

[0002] Icing disasters are a common natural disaster that jeopardizes the safe operation of power systems. In severe cases, they can cause power outages and tower collapses, directly threatening the stable operation of critical power grid infrastructure. Icing disasters are particularly frequent in mountainous and alpine regions with complex terrain and unpredictable climates, posing a severe challenge to power grid operations.

[0003] At present, conventional ice detection mainly relies on two methods: field testing and image detection. Field testing methods such as direct measurement and weighing methods require line patrol personnel to go deep into mountainous areas to manually measure the thickness of ice, or install tension sensors to monitor the force changes on the transmission lines. However, these methods are time-consuming and labor-intensive in complex mountainous terrain and are difficult to adapt to the needs of large-scale line monitoring. In contrast, image-based detection methods are more flexible. For example, by installing image acquisition equipment to remotely monitor transmission lines, the ice condition can be quickly assessed. However, traditional edge detection algorithms are susceptible to interference in the complex natural background of areas such as Anhui, such as the influence of mountain vegetation and weather conditions, making it difficult to obtain high-precision detection results.

[0004] Deep learning-based ice detection technology offers a new solution for ice monitoring in complex terrain. Semantic segmentation models can effectively identify ice-covered areas, significantly outperforming traditional methods in complex backgrounds. Icing in complex terrain is characterized by diverse types and large variations in density. Convolutional neural network-based models not only accurately classify ice types but also leverage density information to improve thickness calculations, providing more scientific data support for line safety assessments. Furthermore, combined with multi-view drone imagery, mountainous lines can be more comprehensively covered, allowing the uniform thickness of ice surrounding the conductors to be estimated.

[0005] Based on this, multimodal data analysis has important practical significance in ice detection. Ice formation is closely related to meteorological factors such as temperature, humidity, and wind speed. Therefore, by fusing multimodal information such as images and meteorological data, the dynamic process of ice formation can be more accurately reflected. Specifically, image data can provide visual information on ice thickness and type, while meteorological data can reveal the external environmental conditions of ice accumulation. This multimodal data analysis can effectively improve the accuracy and real-time performance of ice monitoring, providing strong technical support for the safe operation of power grids in complex areas and their ability to resist ice disasters.

[0006] In related technologies, patent application publication number CN119169535A uses deep learning-based image processing and analysis techniques to detect line icing. However, this approach relies solely on image data. In the preprocessing stage, features are extracted through brightness and roughness features such as histogram equalization and H-LBP texture analysis, emphasizing the extraction of texture features. IceNet-T is then used to extract global, brightness, and roughness features from the icing image. Semantic segmentation utilizes a complete existing segmentation model combined with a SegBoost module to generate multi-scale conditional random field-optimized segmentation results. Patent application publication number CN115994604A uses an RF algorithm to calculate the features of meteorological factors that influence ice growth on overhead transmission lines and constructs a dataset. However, this approach focuses on optimizing the GRNN neural network algorithm and does not involve multimodal data fusion or dynamic correction. Summary of the Invention

[0007] The technical problem to be solved by the present invention is how to improve the accuracy of ice cover state detection in complex environments.

[0008] The present invention solves the above technical problems through the following technical means:

[0009] A method for detecting ice coverage on transmission lines is proposed, including:

[0010] Obtaining the original ice cover image of the transmission line, and calculating the equivalent ice cover thickness based on the original ice cover image;

[0011] The meteorological data is converted into a parameterized expression, and the parameters in the expression are iteratively optimized to obtain a composite correction factor;

[0012] The equivalent ice thickness is corrected and optimized based on the composite correction factor to obtain the corrected ice thickness.

[0013] Furthermore, the meteorological data is converted into a parameterized expression, and the parameters in the expression are iteratively optimized to obtain a composite correction factor, including:

[0014] Convert meteorological data into parameterized expressions:

[0015] f(T,H,V,P)=exp(-αT)·(1+βH)·(1-γV)·(1+δP)

[0016] Where α, β, γ and δ are correction parameters, T, H, V and P are temperature, relative humidity, wind speed and precipitation respectively;

[0017] Iteratively optimizing the correction parameters using the GBDT algorithm to determine the values of the correction parameters;

[0018] The parameterized expression formed by the determined correction parameters is used as a composite correction factor.

[0019] Furthermore, the iterative optimization of the correction parameter using the GBDT algorithm to determine the value of the correction parameter includes:

[0020] The GBDT algorithm was used to construct a regression model, with meteorological factors as input variables and target ice thickness as the target variable for fitting.

[0021] Initialize the regression model prediction value F0(w) to the mean of the target variable, perform gradient boosting iterations, and calculate the pseudo residual value of each iteration process;

[0022] Fit a regression tree using pseudo residuals and update the target variable of the current iteration based on the regression tree and the target variable of the previous iteration.

[0023] The correction parameter is determined by the target variable prediction value obtained when the regression model iterative training is completed.

[0024] Furthermore, the obtaining of the original ice-covered image of the transmission line and calculating the equivalent ice-covered thickness based on the original ice-covered image includes:

[0025] Divide the original ice-covered image into a background sub-image and an ice-covered sub-image;

[0026] Performing feature extraction on the original ice-covered image, the background sub-image, and the ice-covered sub-image respectively to obtain backbone features, background features, and ice-covered features;

[0027] The backbone features, background features and ice cover features are integrated to obtain an ice cover type recognition result;

[0028] Performing semantic segmentation on the original ice-covered image to obtain an area where the transmission line is covered with ice;

[0029] Based on the ice type identification results and the area where the transmission line is iced, the equivalent ice thickness is calculated.

[0030] Furthermore, the extracting features of the original ice-covered image, the background sub-image, and the ice-covered sub-image to obtain backbone features, background features, and ice-covered features respectively includes:

[0031] Inputting the original ice-covered image, the background sub-image and the ice-covered sub-image into the backbone network, the background branch network and the ice-covered branch network respectively to obtain the backbone features, the background features and the ice-covered features;

[0032] Among them, the backbone network adopts the EfficientNet-V2 network, the background branch network and the ice-covered branch network both include an initialization layer, an RDS convolution block group and a linear output layer connected in sequence, and the RDS convolution block group includes several RDS convolution blocks connected in sequence.

[0033] Furthermore, the RDS convolution block includes an inner convolution layer, a first activation function, an outer convolution layer, and a second activation function connected in sequence, and the input features of the RDS convolution block are respectively input to the inner convolution layer, the first activation function, and the second activation function;

[0034] The inner convolution layer and the outer convolution layer each include a depthwise convolution layer, a first batch normalization layer, a pointwise convolution layer, and a second batch normalization layer, which are connected in sequence.

[0035] Furthermore, performing semantic segmentation on the original ice-covered image to obtain the area where the ice is covered on the transmission line includes:

[0036] Inputting the original ice-covered image into a semantic segmentation model for segmentation and recognition to obtain the area where the transmission line is covered with ice;

[0037] The semantic segmentation model includes an encoder, a multi-head attention block and a decoder. The multi-head attention block extracts multi-level feature maps from different depths of the encoder and outputs them to the decoder.

[0038] The decoder includes a first upsampling layer, a second upsampling layer and a third upsampling layer, two convolution blocks are connected between the first upsampling layer and the second upsampling layer, two convolution blocks are connected between the second upsampling layer and the third upsampling layer, and a convolution block and a softmax activation function are sequentially connected after the third upsampling layer;

[0039] The outputs of the first upsampling layer, the second upsampling layer, and the third upsampling layer are channel-wise concatenated with the outputs of the multi-head attention block.

[0040] Furthermore, the feature extraction process formula of the multi-head attention block is expressed as:

[0041]

[0042] In the formula, output is the output feature, X is the input feature, and W O represents linear transformation, Q i , K i 、V i is the output of h attention heads, softmax represents the activation function, is the scaling factor, Concat is the concatenation function, Conv represents the convolution layer, and T represents the transposed symbol.

[0043] Furthermore, the calculation of equivalent ice thickness based on the ice type identification result and the area where the transmission line is iced includes:

[0044] Calculating the major diameter a and minor diameter b of the ice-covered cross section of the transmission line according to the area where the ice-covered transmission line is located;

[0045] determining an ice density according to the ice type identification result;

[0046] The equivalent ice thickness is calculated using the major diameter a, minor diameter b and ice density ρ.

[0047] In addition, the present invention also proposes a transmission line icing status detection system, comprising:

[0048] An ice thickness calculation module is used to calculate equivalent ice thickness based on the original ice image of the transmission line obtained;

[0049] The correction factor calculation module is used to convert meteorological data into parameterized expressions and iteratively optimize the parameters in the expressions to obtain composite correction factors;

[0050] The ice thickness correction module is used to correct and optimize the equivalent ice thickness based on a composite correction factor to obtain a corrected ice thickness.

[0051] The advantages of the present invention are:

[0052] (1) The present invention preliminarily calculates the equivalent ice thickness based on ice image data, comprehensively considers the impact of different meteorological conditions on ice formation, dynamically adapts to environmental changes in different regions and time periods, and uses environmental meteorological elements to construct a composite correction factor to correct the preliminarily calculated ice thickness. This method can compensate for errors in image processing, enhance the credibility and theoretical interpretability of the results, improve the accuracy of thickness calculation and the generalization ability of the model, and provide more accurate decision support for ice trend prediction, extreme weather warning and operation and maintenance of transmission lines, ensuring the accuracy of ice status detection even in complex environments.

[0053] (2) The present invention uses a multi-branch network including a background branch network, an ice cover branch network and a trunk branch network to extract environmental background information, ice cover information and overall feature information. The background branch can extract the characteristics of the background area and reduce the interference of background noise on the model recognition performance. The ice cover branch focuses on extracting detailed features that are strongly related to the ice cover characteristics. The trunk branch extracts global features and high-level feature information, enhancing the model's overall understanding of ice cover types in complex long paths. The multi-branch structure can integrate information from multiple feature dimensions, thereby improving the model's recognition accuracy and robustness for different ice cover types.

[0054] (3) The present invention segments the ice-covered area of the transmission line by constructing a semantic segmentation model, constructs an encoder by migrating the model and a decoder by a custom model, and combines a multi-head attention block to improve the segmentation effect. It captures features of different depths in the encoder and calculates multiple attention heads of different scales in parallel. Each head focuses on different feature representations and local contextual relationships. This method allows the model to focus on input features from different perspectives, which can more fully express the boundary feature information of the ice cover, thereby providing accurate data for the calculation of equivalent thickness.

[0055] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A schematic flow chart of a method for detecting ice coverage on a power transmission line according to an embodiment of the present invention;

[0057] Figure 2 A schematic diagram of the overall model structure for detecting ice coating status of a transmission line in one embodiment of the present invention;

[0058] Figure 3 Schematic diagram of the model structure of ResSepNet in one embodiment of the present invention;

[0059] Figure 4 Schematic diagram of the structure of the RDS module in the ResSepNet model in one embodiment of the present invention;

[0060] Figure 5 Schematic diagram of the MOS-Net model structure in one embodiment of the present invention;

[0061] Figure 6 Schematic diagram of the structure of the multi-head attention block in the MOS-Net model in one embodiment of the present invention;

[0062] Figure 7 The comparison results of the ResSepNet model and other models in the ice type recognition task in one embodiment of the present invention are shown in Figure 1, where (a) is a comparison chart of the training accuracy results, and (b) is a comparison chart of the training loss results;

[0063] Figure 8 This is a diagram showing the segmentation effect of an actual power transmission line according to an embodiment of the present invention;

[0064] Figure 9 The present invention provides a schematic structural diagram of a system for detecting ice coating status of a power transmission line according to an embodiment of the present invention. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0066] like Figure 1 As shown, the first embodiment of the present invention provides a method for detecting the icing state of a transmission line, the method comprising the following steps:

[0067] S10, obtaining an original ice-covered image of the transmission line, and calculating an equivalent ice-covered thickness based on the original ice-covered image;

[0068] It should be noted that, in this embodiment, an online monitoring device and a drone may be used to capture images of ice covering the transmission line for calculating the ice thickness.

[0069] S20, converting the meteorological data into a parameterized expression, and iteratively optimizing the parameters in the expression to obtain a composite correction factor;

[0070] It should be noted that this embodiment can map the longitude and latitude provided by the online monitoring device and the time information of the image shooting to the corresponding meteorological elements in ERA5, and obtain the environmental meteorological element information such as temperature, humidity, wind speed, precipitation, etc. when each original ice cover image is taken. It also comprehensively considers the impact of different meteorological conditions on ice formation, and calculates the composite correction factor of meteorological elements based on the correlation between environmental meteorological elements and their impact on ice cover thickness, which can dynamically adapt to environmental changes in different regions and time periods.

[0071] S30: Correct and optimize the equivalent ice thickness based on the composite correction factor to obtain a corrected ice thickness.

[0072] By using a composite correction factor to correct the initially calculated ice thickness, this method can compensate for errors in image processing, enhance the credibility and theoretical interpretability of the results, improve the accuracy of thickness calculation and the generalization ability of the model, and provide more accurate decision-making support for icing trend prediction, extreme weather warning, and operation and maintenance of transmission lines, ensuring the accuracy of icing status detection even in complex environments.

[0073] In addition, unlike the prior art that uses meteorological characteristics as training data to train the neural network model, the present invention directly uses meteorological factors to construct a composite correction factor to correct the equivalent ice thickness. On the one hand, it can accurately compensate for errors caused by factors such as light changes, image noise, and shooting angle deviation in the image processing process, effectively avoiding the adverse effects of error accumulation on the final result, making the ice thickness data closer to the actual situation, and effectively improving the accuracy of thickness calculation. On the other hand, from a theoretical perspective, this method gives the ice thickness calculation stronger logic and explainability, and no longer relies solely on complex black box neural network training, so that the results of each step have a clear theoretical basis to follow. Not only that, by introducing meteorological factors for direct correction, the model is no longer limited to the characteristic patterns of the training data in a specific scenario, but can adapt to a wider range of environmental conditions, greatly enhancing the generalization ability of the model.

[0074] As a further preferred technical solution, in step S20: converting the meteorological data into a parameterized expression and iteratively optimizing the parameters in the expression to obtain a composite correction factor, specifically includes the following steps:

[0075] S21. Convert meteorological data into parameterized expressions:

[0076] f(T,H,V,P)=exp(-αT)·(1+βH)·(1-γV)·(1+δP)

[0077] Where α, β, γ and δ are correction parameters, T, H, V, P are temperature, relative humidity, wind speed and precipitation respectively, and exp() represents an exponential function with a natural constant as the base;

[0078] Specifically, this embodiment maps the longitude and latitude and time information when the online monitoring device captures the monitoring image to the meteorological data of the corresponding location and time in ERA5, including temperature T (°C), relative humidity H (%), wind speed V (m / s) and precipitation P (mm / h).

[0079] It should be noted that α, β, γ, and δ in the parameterized expression of this embodiment are correction parameters, which are determined by the GBDT algorithm. exp(-αT) indicates that the increase in temperature leads to a decrease in icing, (1+βH) indicates that the increase in humidity increases the possibility of icing, (1-γV) indicates that high wind speed may lead to a decrease in icing, and (1+δP) indicates that the higher the precipitation, the greater the ice thickness.

[0080] S22, iteratively optimizing the correction parameter using the GBDT algorithm to determine the value of the correction parameter;

[0081] S23. Using the parameterized expression formed by the determined correction parameters as a composite correction factor.

[0082] It should be understood that the meteorological data in this embodiment takes into account temperature, relative humidity, wind speed and precipitation, and those skilled in the art may also expand other meteorological data according to actual applications.

[0083] As a further preferred technical solution, the step S22: using the GBDT algorithm to iteratively optimize the correction parameter to determine the value of the correction parameter, specifically includes the following steps:

[0084] S221. Use the GBDT algorithm to build a regression model, using meteorological factor characteristics as input variables and target ice thickness as the target variable for fitting;

[0085] Specifically, the environmental meteorological factors and their correlation with ice thickness are used as input features, and the target ice thickness is used as the output variable. The feature set and target variable set are constructed. The input variable feature set can be expressed as W = [T, H, V, P], and the output target variable can be expressed as Y true , that is, the actual measured ice thickness;

[0086] S222, initialize the regression model prediction value F0(w) to the mean of the target variable, perform gradient boosting iterations, and calculate the pseudo residual value of each iteration process;

[0087] Specifically, this embodiment uses the GBDT algorithm to build a regression model, and the target variable Y true To fit, first set the initial value and initialize the regression model prediction value F0(w) to the mean of the target variable, and then perform gradient boosting iteration. In the mth iteration process, the calculated pseudo residual value can be expressed by the following formula

[0088]

[0089] in, is the true target variable of the i-th sample, F m-1 (w) is the model prediction value of the m-1th iteration, N is the total number of samples, L is the MSE loss function, and the mean square error is obtained by averaging the sum of the squared errors. The formula is:

[0090]

[0091] Where F(w) is the predicted value of the final regression model, and w is the single meteorological element feature of the input variable feature set W.

[0092] S223, fitting a regression tree using pseudo residuals, and updating the target variable of the current iteration process based on the regression tree and the target variable of the previous iteration process;

[0093] Specifically, this embodiment uses the pseudo residual to fit the decision tree. As the target variable, fit a regression tree h(w), and then update the model F m (w):

[0094] F m (w)=F m-1 (w)+ξ·h m (w)

[0095] Among them, ξ is the learning rate, h m (w) is the output of the mth regression tree

[0096] S224. Determine the correction parameter based on the target variable prediction value obtained when the iterative training of the regression model is completed.

[0097] Specifically, this embodiment obtains the final prediction regression model F through model training. M (w), and the correction parameters α, β, γ, and δ are determined using the following formula:

[0098]

[0099] in, Indicates the sensitivity of the model prediction value to the corresponding meteorological factors.

[0100] This embodiment utilizes the GBDT algorithm to iteratively optimize the correction parameters. This method accurately captures the complex nonlinear relationship between meteorological factors and target ice thickness, effectively addressing noise and outliers in the data. Using gradient boosting iterations, the loss function is continuously reduced, allowing the model's predicted values to continuously approach the true values. Furthermore, the algorithm's anti-overfitting properties ensure the generalization of the correction parameters, ensuring stability and reliability across diverse datasets.

[0101] Furthermore, this embodiment can also adjust the model hyperparameters according to the model accuracy to further optimize the parameter correction and obtain the final meteorological factor correction parameters.

[0102] As a further preferred technical solution, this embodiment may also pre-process the input feature set W and perform standard normalization processing, the formula is:

[0103]

[0104] Among them, μ(w) is the mean, σ(w) is the standard deviation, and the mean filling method is used to fill in the missing values. The data set is divided into training set and test set in a ratio of 4:1.

[0105] Accordingly, this embodiment uses the meteorological factor characteristics after standard normalization processing to train the regression model.

[0106] As a further preferred technical solution, in step S10, calculating the equivalent ice thickness based on the original ice image specifically includes the following steps:

[0107] S11, dividing the original ice-covered image into a background sub-image and an ice-covered sub-image;

[0108] Specifically, this embodiment divides the original ice-covered image at the central horizontal line to obtain upper and lower sub-images, wherein the upper sub-image is the background sub-image and the lower sub-image is the ice-covered sub-image.

[0109] S12, performing feature extraction on the original ice-covered image, the background sub-image, and the ice-covered sub-image respectively to obtain backbone features, background features, and ice-covered features;

[0110] S13, fusing the main features, background features and ice cover features to obtain an ice cover type recognition result;

[0111] S14, performing semantic segmentation on the original ice-covered image to obtain an area where the transmission line is covered with ice;

[0112] S15. Calculate equivalent ice thickness based on the ice type identification result and the area where the transmission line is iced.

[0113] It should be noted that this embodiment extracts background, ice cover, and trunk features at different spatial locations, integrates information from multiple feature dimensions, and improves the recognition accuracy and robustness of the model for different ice cover types.

[0114] Specifically, if Figure 2 As shown, this embodiment can construct an ice type recognition model and an ice area segmentation model to process the original ice image respectively to obtain the ice type recognition result and the area where the transmission line is iced.

[0115] The icing type recognition model includes an icing segmentation preprocessing module, a background branch network, a trunk branch network, and an icing branch network, wherein:

[0116] Ice segmentation preprocessing module, used to divide the original ice image into background sub-image and ice sub-image;

[0117] The trunk branch network, the background branch network and the ice-covered branch network are used to process the original ice-covered image, the background sub-image and the ice-covered sub-image respectively to obtain trunk features, background features and ice-covered features.

[0118] like Figure 3As shown, the backbone network adopts the EfficientNet-V2 network, and the background branch network and the ice-covered branch network both include an initialization layer, an RDS convolution block group and a linear output layer connected in sequence, and the RDS convolution block group includes several RDS convolution blocks connected in sequence.

[0119] It should be noted that the RDS convolution block group provided in this embodiment includes 6 RDS convolution blocks connected in sequence. Those skilled in the art may also select other different numbers of RDS convolution blocks to constitute the RDS convolution block group according to actual applications.

[0120] As a further preferred technical solution, Figure 4 As shown, the RDS convolution block includes an inner convolution layer, a first activation function, an outer convolution layer, and a second activation function connected in sequence, and the input features of the RDS convolution block are respectively input to the inner convolution layer, the first activation function, and the second activation function;

[0121] The inner convolution layer and the outer convolution layer each include a depthwise convolution layer, a first batch normalization layer, a pointwise convolution layer, and a second batch normalization layer, which are connected in sequence.

[0122] It should be noted that this embodiment uses a nested residual structure and a depth-separable convolution layer to construct an RDS (Residual Depthwise Separable) convolution block. The specific process is as follows:

[0123] (1) A smaller residual network is embedded inside the conventional residual network to form a nested residual network. The feature extraction process can be expressed as follows:

[0124] Y res =F2(F1(X res )+X res )+X res

[0125] Among them, F1 and F2 represent two consecutive convolution blocks, the network parameters are independent, and X res ∈R H×W×C is the input feature map of the nested residual network, Y res ∈R H×W×D Represents the output feature map of the nested residual network, H and W represent the height and width of the feature map respectively, C represents the number of input channels, and D represents the number of output channels;

[0126] (2) Using the depthwise separable convolutional layer to replace the ordinary convolutional layer inside the nested residual network, the feature extraction process of the depthwise separable layer can be expressed as follows:

[0127] Y dp =PD(X dp)=PW_Conv(DW_Conv(X dp ))

[0128] Among them, X dp ∈R H×W×C represents the input feature map of the depthwise separable convolutional layer, Y dp ∈R H×W×D Represents the output feature map, DW_Conv represents depthwise convolution, PW_Conv represents pointwise convolution, and PD represents full depthwise separable convolution;

[0129] (3) Combining the nested residual structure and depth-wise separable convolution, an RDS convolution block is constructed. The feature processing process of each individual RDS convolution block can be expressed as follows:

[0130] Y rds =φ(F_PD2(φ(F_PD1(X rds )+X rds ))+X rds )

[0131] Among them, φ represents the LeakyReLU activation function, F_PD1 represents the inner RDS convolution layer, F_PD2 represents the outer RDS convolution layer, X rds ∈R H×W×C Represents the input feature map of the RDS convolution block, Y rds ∈R H×W×D Represents the output feature map of the RDS convolution block.

[0132] As a further preferred technical solution, this embodiment uses RDS convolution blocks to construct a background branch network and an ice branch network. The background branch can extract the characteristics of the background area and reduce the interference of background noise on the model recognition performance. The ice branch focuses on extracting detailed features that are strongly related to the ice characteristics. The trunk branch extracts global features and high-level feature information, enhancing the model's overall understanding of ice types in complex long paths. A network with a multi-branch structure is used to extract environmental background information, ice information, and overall feature information. The multi-branch structure can integrate information from multiple feature dimensions to improve the model's recognition accuracy and robustness for different ice types. The formulas for the trunk features, background features, and ice features extracted by the trunk branch network, background branch network, and ice branch network are expressed as follows:

[0133] O main =f Adapter (EfficientNet-V2(X main ))

[0134]

[0135] In the formula, Init represents the initialization layer, which consists of a 1×1 convolution layer and a ReLU activation function. The background branch and the ice branch use the same initialization layer. RDS n Represents n RDS convolution blocks of the background branch and the ice-covered branch, where n is 6, FC represents the linear output layer of the background branch and the ice-covered branch, the subscripts of each module represent the different branches to which they belong, X represents the subgraph of the background branch and the ice-covered branch input, and Y represents the output features of the background branch and the ice-covered branch; X main represents the original ice-covered image input by the trunk branch, EfficientNet-V2 represents the feature extraction process of the migration model EfficientNet-V2, and f Adapter represents an adaptive output layer to adapt to the ice type recognition task, O main Represents the output features of the trunk branch.

[0136] As a further preferred technical solution, the step S13: fusing the main features, background features and ice features to obtain an ice type recognition result, which is expressed as:

[0137] O final =FC final (Concat(Y bg ,Y ice ,O main ))

[0138] Among them, Concat represents the concatenation function, which fuses the output features of the three branches in the channel dimension. final represents the fused linear output layer, O final It represents the fusion features finally output by ResSepNet, that is, the ice type recognition result obtained by the model.

[0139] As a further preferred technical solution, step S14: performing semantic segmentation on the original ice-covered image to obtain the area where the ice is located on the transmission line, specifically includes:

[0140] Inputting the original ice-covered image into a semantic segmentation model for segmentation and recognition to obtain the area where the transmission line is covered with ice; wherein the semantic segmentation model includes an encoder, a multi-head attention block and a decoder, and the multi-head attention block extracts multi-level feature maps from different depths of the encoder and outputs them to the decoder;

[0141] The decoder includes a first upsampling layer, a second upsampling layer and a third upsampling layer, two convolution blocks are connected between the first upsampling layer and the second upsampling layer, two convolution blocks are connected between the second upsampling layer and the third upsampling layer, and a convolution block and a softmax activation function are sequentially connected after the third upsampling layer;

[0142] The outputs of the first upsampling layer, the second upsampling layer, and the third upsampling layer are channel-wise concatenated with the outputs of the multi-head attention block.

[0143] Specifically, if Figure 2 and Figure 5 As shown, this embodiment constructs a semantic segmentation model MOS-Net with a large encoder-small decoder structure, wherein the pre-trained deep network MobileOne is migrated as the encoder of MOS-Net, its fully connected layer is removed, only the convolutional layer is retained as the feature extractor, the weights of the low-level feature extraction part of the model are frozen, and the high-level feature extraction part is fine-tuned to adapt to the ice-covered area recognition and segmentation task.

[0144] A lightweight decoder is constructed, using transposed convolution and bilinear interpolation to gradually restore the encoder output feature map to the input image resolution. Simultaneously, a designed multi-head attention block is used to fuse feature information from different depths in the encoder to preserve more detailed information. During the decoding process, all convolutional layers use the ReLU nonlinear activation function. Finally, a 1×1 convolutional layer compresses the feature map to the number of target categories, namely background and ice-covered areas. A sigmoid activation function is used to determine the segmentation probability of each pixel, which is the final segmentation result output by MOS-Net.

[0145] As a further preferred technical solution, Figure 6 As shown in , the feature extraction process formula of the multi-head attention block is expressed as:

[0146]

[0147] Where output is the output feature, X is the input feature, and X is linearly transformed to obtain the query Q, key K and value V matrices. Q, K, V are divided into h heads, and each head uses its own Q i , K i 、V i , use the softmax function to calculate the Q of each head i and K i The attention weight matrix, where is the scaling factor to prevent the gradient from disappearing, and the attention weight matrix is combined with V i Multiply them together to get the output of each head, that is Next, the Concat function is used to concatenate the attention outputs of each head, and the concatenated outputs are then subjected to a linear transformation W. O, the spliced features are remapped to the original feature space. In order to enhance the local feature extraction capability, they are processed through a convolutional layer Conv after linear transformation to better combine the advantages of convolutional networks in local feature extraction; finally, the output of the multi-head attention block after convolution is added to the original input X to form a residual structure, which retains the original input information, increases the model's ability to extract contextual information, and improves model stability.

[0148] As a further preferred technical solution, step S15: calculating the equivalent ice thickness based on the ice type identification result and the area where the transmission line is iced, specifically includes the following steps:

[0149] S151. Calculating the major diameter a and minor diameter b of the ice-covered cross section of the transmission line according to the ice-covered area of the transmission line;

[0150] Specifically, this embodiment performs a preprocessing operation on the MOS-Net segmentation results by removing the central rectangle to reduce the impact of visibility on the calculation results. The preprocessed segmentation results are used to perform the equal area method to calculate the major diameter a and minor diameter b of the ice-covered conductor cross section. The calculation process of the major diameter a can be expressed as follows:

[0151]

[0152] Where d represents the bare wire diameter, S(x,y) represents the value at position (x,y) in the MOS-Net output segmentation result, and the numerator Indicates the total number of pixels in the ice-covered area in the segmentation result, the denominator It represents the total number of pixels in the bare wire area without ice cover. Similarly, the value of the short diameter b can be calculated.

[0153] S152: determining an ice density according to the ice type identification result;

[0154] S153. Calculate the equivalent ice thickness using the major diameter a, the minor diameter b, and the ice density ρ.

[0155] Specifically, this embodiment uses the long diameter a, the short diameter b, and the ice density ρ to calculate the equivalent ice thickness. The ice density ρ can be inferred from the ice type identified by ResSepNet. The equivalent ice thickness calculation formula is as follows:

[0156]

[0157] Where T' is the required preliminary equivalent ice thickness, a and b are the major and minor diameters, and d is the bare wire diameter.

[0158] As a further preferred technical solution, the step S30: performing correction optimization on the equivalent ice thickness based on the composite correction factor to obtain the corrected ice thickness, which is expressed as:

[0159] T=T ice (a,b,ρ,d)·f(T,H,V,P)

[0160] Where T is the final corrected equivalent ice thickness.

[0161] Specifically, if Figure 2 As shown, the overall process of calculating ice thickness of transmission lines in the present invention includes:

[0162] (1) The original ice-covered image of the transmission line is obtained by using the online monitoring device on the tower. At the same time, the latitude, longitude and time information when the online monitoring device takes the image are matched and mapped to the environmental meteorological data of the corresponding time and place in ERA5;

[0163] (2) The original ice-covered image is input into the ice-covered type recognition model, and the background sub-image and ice-covered sub-image are obtained through the ice-covered segmentation preprocessing module. The background sub-image, ice-covered sub-image and the original ice-covered image are input into the background branch, ice-covered branch and trunk branch of ResSepNet respectively, and the features extracted from the three branches are fused to obtain the corresponding ice-covered type recognition result;

[0164] The original ice-covered image is input into the ice-covered area segmentation model. The segmentation model encoder is constructed using the backbone network of the migrated MobileOne model. The multi-head attention block is combined to extract the spatial feature information of the feature maps of different depths. The custom decoder is then used to upsample and restore the resolution to obtain the final ice-covered area segmentation result.

[0165] (3) In the equivalent thickness calculation module, the corresponding ice density is calculated based on the ice type identification results. Under normal circumstances, the density of rain rime ice is 0.7-0.9 g / cm3, the density of fog rime ice is 0.1-0.4 g / cm3, the density of low-density mixed rime ice is 0.2-0.4 g / cm3, and the density of high-density mixed rime ice is 0.4-0.6 g / cm3. At the same time, the major and minor diameters are calculated based on the ice area segmentation results, and then the equivalent ice thickness is preliminarily calculated according to the equivalent thickness calculation formula;

[0166] (4) The meteorological data obtained by mapping the original ice cover image is used to construct the ice cover thickness correction factor, and the equivalent ice cover thickness obtained by the preliminary calculation is optimized and corrected to obtain the final equivalent ice cover thickness calculation result.

[0167] In this embodiment, the constructed Figure 3The comparative test results of the ResSepNet model shown in the figure and the current advanced classification models ResNeXt, MobileOne, EfficientNet-V2 and MobileNet-V3 on the ice type recognition dataset are as follows: Figure 7 As shown, Figure 7 (a) in the figure is the training accuracy of each model. Figure 7 (b) in the figure is the training loss of each model. It can be seen that the recognition accuracy of ResSepNet can be stabilized at around 96%, which is higher than that of other models. In addition, ResSepNet is also better than other models in terms of model stability and robustness.

[0168] In this embodiment, the detailed model structure of the ice-covered area segmentation model MOS-Net is as follows: Figure 5 As shown in the figure, the multi-head attention block structure for extracting long-distance spatial feature information of feature maps of different scales is as follows: Figure 6 As shown, Figure 5 The detailed process of feature extraction for single-scale feature maps is shown, and the ice-covered area segmentation effect of MOS-Net is shown in Figure 2. Figure 8 At the same time, this embodiment Figure 8 The ice detection results of all original images are shown in Table 1 below, which include the detection results of major diameter a, minor diameter b, equivalent ice thickness and ice type.

[0169] Table 1 Ice detection results

[0170]

[0171]

[0172] As can be seen from the results in Table 1, the transmission line icing status detection method according to the embodiment of the present invention can well detect important information such as the equivalent ice thickness and ice type of the transmission line, and can provide accurate data support for early warning and scientific prevention and control of the power grid, thereby ensuring the stable operation of the power grid.

[0173] In addition, if Figure 9 As shown, the second embodiment of the present invention further provides a system for detecting ice coating status of a power transmission line, the device comprising:

[0174] An ice thickness calculation module 10 is configured to calculate an equivalent ice thickness based on the original ice image of the transmission line obtained;

[0175] The correction factor calculation module 20 is used to convert the meteorological data into a parameterized expression and iteratively optimize the parameters in the expression to obtain a composite correction factor;

[0176] The ice thickness correction module 30 is configured to perform correction and optimization on the equivalent ice thickness based on a composite correction factor to obtain a corrected ice thickness.

[0177] As a further preferred technical solution, the correction factor calculation module 20 specifically includes:

[0178] The parameterization unit is used to convert meteorological data into parameterized expressions:

[0179] f(T,H,V,P)=exp(-αT)·(1+βH)·(1-γV)·(1+δP)

[0180] Where α, β, γ and δ are correction parameters, T, H, V and P are temperature, relative humidity, wind speed and precipitation respectively;

[0181] An iterative unit, configured to iteratively optimize the correction parameter using a GBDT algorithm to determine a value of the correction parameter;

[0182] The parameter determination unit is used to use the parameterized expression composed of the determined correction parameters as a composite correction factor.

[0183] As a further preferred technical solution, the iteration unit is specifically configured to perform the following steps:

[0184] The GBDT algorithm was used to construct a regression model, with meteorological factors as input variables and target ice thickness as the target variable for fitting.

[0185] Initialize the regression model prediction value F0(w) to the mean of the target variable, perform gradient boosting iterations, and calculate the pseudo residual value of each iteration process;

[0186] Fit a regression tree using pseudo residuals and update the target variable of the current iteration based on the regression tree and the target variable of the previous iteration.

[0187] The correction parameter is determined by the target variable prediction value obtained when the regression model iterative training is completed.

[0188] As a further preferred technical solution, the ice thickness calculation module 10 is specifically deployed as follows Figure 2 The model shown includes an ice type recognition model and an ice area segmentation model. The original ice image is processed by the ice type recognition model and the ice area segmentation model respectively to obtain the ice type recognition result and the area where the transmission line is iced; then, based on the ice type recognition result and the area where the transmission line is iced, the equivalent ice thickness is calculated.

[0189] It should be noted that other embodiments or specific implementation methods of the transmission line icing status detection system of the present invention can refer to the above-mentioned method embodiments and will not be repeated here.

[0190] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0191] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0192] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for detecting ice coating status of a transmission line, characterized in that: include: Obtaining the original ice cover image of the transmission line, and calculating the equivalent ice cover thickness based on the original ice cover image; The meteorological data is converted into a parameterized expression, and the parameters in the expression are iteratively optimized to obtain a composite correction factor; The equivalent ice thickness is corrected and optimized based on the composite correction factor to obtain the corrected ice thickness.

2. The method for detecting ice coating status of a power transmission line according to claim 1, wherein: The conversion of meteorological data into parameterized expressions and iterative optimization of the parameters in the expressions to obtain composite correction factors include: Convert meteorological data into parameterized expressions: f(T,H,V,P)=exp(-αT)·(1+βH)·(1-γV)·(1+δP) Where α, β, γ and δ are correction parameters, T, H, V and P are temperature, relative humidity, wind speed and precipitation respectively; Iteratively optimizing the correction parameters using the GBDT algorithm to determine the values of the correction parameters; The parameterized expression formed by the determined correction parameters is used as a composite correction factor.

3. The method for detecting ice coating status of a power transmission line according to claim 2, wherein: The iterative optimization of the correction parameter using the GBDT algorithm to determine the value of the correction parameter includes: The GBDT algorithm was used to construct a regression model, with meteorological factors as input variables and target ice thickness as the target variable for fitting. Initialize the regression model prediction value F0(w) to the mean of the target variable, perform gradient boosting iterations, and calculate the pseudo residual value of each iteration process; Fit a regression tree using pseudo residuals and update the target variable of the current iteration based on the regression tree and the target variable of the previous iteration. The correction parameter is determined by the target variable prediction value obtained when the regression model iterative training is completed.

4. The method for detecting ice coating status of a power transmission line according to claim 1, wherein: The obtaining of the original ice-covered image of the transmission line and calculating the equivalent ice-covered thickness based on the original ice-covered image includes: Divide the original ice-covered image into a background sub-image and an ice-covered sub-image; Performing feature extraction on the original ice-covered image, the background sub-image, and the ice-covered sub-image respectively to obtain backbone features, background features, and ice-covered features; The backbone features, background features and ice cover features are integrated to obtain an ice cover type recognition result; Performing semantic segmentation on the original ice-covered image to obtain an area where the transmission line is covered with ice; Based on the ice type identification results and the area where the transmission line is iced, the equivalent ice thickness is calculated.

5. The method for detecting ice coating status of a power transmission line according to claim 4, wherein: The extracting features of the original ice-covered image, the background sub-image, and the ice-covered sub-image to obtain backbone features, background features, and ice-covered features respectively includes: Inputting the original ice-covered image, the background sub-image and the ice-covered sub-image into the backbone network, the background branch network and the ice-covered branch network respectively to obtain the backbone features, the background features and the ice-covered features; Among them, the backbone network adopts the EfficientNet-V2 network, the background branch network and the ice-covered branch network both include an initialization layer, an RDS convolution block group and a linear output layer connected in sequence, and the RDS convolution block group includes several RDS convolution blocks connected in sequence.

6. The method for detecting ice coating status of a power transmission line according to claim 4, wherein: The RDS convolution block includes an inner convolution layer, a first activation function, an outer convolution layer and a second activation function connected in sequence, and the input features of the RDS convolution block are respectively input to the inner convolution layer, the first activation function and the second activation function; The inner convolution layer and the outer convolution layer each include a depthwise convolution layer, a first batch normalization layer, a pointwise convolution layer, and a second batch normalization layer, which are connected in sequence.

7. The method for detecting ice coating status of a power transmission line according to claim 4, wherein: The performing semantic segmentation on the original ice-covered image to obtain an area where the transmission line is covered with ice includes: Inputting the original ice-covered image into a semantic segmentation model for segmentation and recognition to obtain the area where the transmission line is covered with ice; The semantic segmentation model includes an encoder, a multi-head attention block and a decoder. The multi-head attention block extracts multi-level feature maps from different depths of the encoder and outputs them to the decoder. The decoder includes a first upsampling layer, a second upsampling layer and a third upsampling layer, two convolution blocks are connected between the first upsampling layer and the second upsampling layer, two convolution blocks are connected between the second upsampling layer and the third upsampling layer, and a convolution block and a softmax activation function are sequentially connected after the third upsampling layer; The outputs of the first upsampling layer, the second upsampling layer, and the third upsampling layer are channel-wise concatenated with the outputs of the multi-head attention block.

8. The method for detecting ice coating status of a power transmission line according to claim 7, wherein: The feature extraction process formula of the multi-head attention block is expressed as: In the formula, output is the output feature, X is the input feature, and W O represents linear transformation, Q i , K i 、V i is the output of h attention heads, softmax represents the activation function, is the scaling factor, Concat is the concatenation function, Conv represents the convolution layer, and T represents the transposed symbol.

9. The method for detecting ice coating status of a power transmission line according to claim 4, wherein: The calculating of equivalent ice thickness based on the ice type identification result and the iced area of the transmission line includes: Calculating the major diameter a and minor diameter b of the ice-covered cross section of the transmission line according to the area where the ice-covered transmission line is located; determining an ice density according to the ice type identification result; The equivalent ice thickness is calculated using the major diameter a, minor diameter b and ice density ρ.

10. A transmission line icing status detection system, characterized in that: include: An ice thickness calculation module is used to calculate equivalent ice thickness based on the original ice image of the transmission line obtained; The correction factor calculation module is used to convert meteorological data into parameterized expressions and iteratively optimize the parameters in the expressions to obtain composite correction factors; The ice thickness correction module is used to correct and optimize the equivalent ice thickness based on a composite correction factor to obtain a corrected ice thickness.

Images