Power transmission line icing thickness identification method based on global micro-sensing network

Abstract

The invention discloses a power transmission line icing thickness identification method based on a global micro-sensing network, and the method specifically comprises the steps: carrying out the image data collection of a power transmission line, and forming a data set; preprocessing the icing image, and dividing the icing image into a training set, a verification set and a test set; labeling the icing images in the training set and the verification set by using a Label labeling tool, and generating a json configuration file and an 8-bit grey-scale map; constructing an icing thickness identification model; training, optimizing and evaluating the model by using a training set, a verification set and a test set respectively until the accuracy error between an evaluation result and a training result is less than 10%; and deploying the trained icing thickness identification model to an edge server, transmitting the to-be-detected image to the edge server for identification through a wireless or wired network, and outputting the icing thickness. According to the power transmission line icing thickness identification method based on the global micro-sensing network, the problem that the accuracy of the calculated icing thickness is not high in the prior art is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transmission line monitoring, and in particular relates to a method for identifying ice thickness of a power transmission line based on a global micro-sensing network. Background Art

[0002] Transmission lines are an important part of the power system, responsible for power transmission between power plants and power-consuming equipment, so their transmission stability is crucial. Ice coating on conductors is a common dangerous phenomenon. When the lines are covered with ice, their weight will increase significantly. Once the carrying capacity is exceeded, the conductors will be deformed and may even cause the towers at both ends of the transmission lines to collapse. In severe cases, it may even affect the normal operation of the power system and cause great inconvenience to residents' daily lives. Therefore, it is very important to monitor the icing status of transmission lines in real time. If the thickness of the ice coating can be accurately identified, the icing problem of the transmission lines can be effectively warned. Therefore, how to accurately identify the thickness of the ice coating is an urgent problem that needs to be solved.

[0003] Traditional image recognition methods use artificially designed feature extraction methods to calculate the geometric, edge and texture information of images. This method performs well in simple scenarios, but cannot effectively identify the thickness of ice in complex real-world environments. In recent years, the continuous development of deep learning technology has gradually replaced traditional image processing methods in the field of computer vision, and neural networks have demonstrated their powerful feature extraction capabilities. At present, the technology of using deep learning to detect the thickness of ice on transmission lines is mainly divided into three categories: image segmentation, target detection and edge detection. Among them, the image segmentation algorithm can determine the ice and snow covered area and calculate the ice thickness by segmenting the ice and snow covered area and the background area of ​​the conductor. However, due to the uneven upper and lower edges of the ice and snow covered area, the existing image segmentation method can only obtain the fuzzy range of ice thickness and cannot calculate the accurate ice thickness, and the calculated ice thickness is not accurate enough. Summary of the invention

[0004] The purpose of the present invention is to provide a method for identifying ice thickness of a power transmission line based on a global micro-sensing network, which solves the problem of low accuracy of ice thickness calculated in the prior art.

[0005] The technical solution adopted by the present invention is a method for identifying ice thickness of a transmission line based on a global micro-sensing network, which specifically includes the following steps: S1: collect image data of the transmission line to form a data set; S2: Preprocess the ice-covered images in the dataset and divide them into training set, validation set and test set according to the ratio; S3: Label the ice-covered images in the training set and validation set using the Labelme annotation tool to generate a json configuration file and an 8-bit grayscale image; S4: Construct ice thickness identification model; S5: Use the training set, validation set and test set to train, optimize and evaluate the ice thickness recognition model respectively until the accuracy error between the evaluation result and the training result is less than 10%, and then stop training and evaluation; S6: Deploy the trained ice thickness recognition model to the edge server, transmit the image to be detected to the edge server through a wireless or wired network for recognition, and output the ice thickness.

[0006] The present invention is also characterized in that: S1 specifically includes: installing cameras on the transmission lines to collect ice-covered images of the transmission lines to form a data set.

[0007] In S2, the ice-covered images in the data set are preprocessed, including adjusting the images to a uniform size and performing histogram equalization, and then performing data cleaning, denoising, and normalization on the images; The preprocessed dataset is divided into training set, validation set and test set in a ratio of 3:2:2.

[0008] S3 is as follows: for the ice-covered images in the training set and the validation set, the Labelme annotation tool is used to annotate the ice-covered areas. After the annotation is completed, a json configuration file is automatically generated. A 24-bit grayscale image is generated based on the json configuration file, and then the 24-bit grayscale image is batch converted into an 8-bit grayscale image through the image pixel conversion method.

[0009] In S4, an ice thickness recognition model is constructed, which includes two mixed strip convolution modules, two convolution modules, a global micro-sensing module, four deconvolution modules, an image segmentation module and a thickness calculation module connected in sequence; Each mixed strip convolution module specifically includes 1 5×5 convolution, and 1 5×5 convolution is connected to 1 1×3 convolution with an interval of 1, 1 1×3 convolution with an interval of 2, and 1 1×3 convolution with an interval of 5; and 1×3 convolution with an interval of 1, 1×3 convolution with an interval of 2, and 1×3 convolution with an interval of 5 are connected to 1 3×1 convolution with an interval of 1, 1 3×1 convolution with an interval of 2, and 1 3×1 convolution with an interval of 5 respectively; Each convolution module consists of two 3×3 convolutions connected in sequence; The global micro-perception module includes a convolutional layer, a micro-perception module and a global perception module. The micro-perception module includes a maximum pooling layer of sizes k5, k9 and k13, connected to a cascade fusion layer; the global perception module includes a linear projection layer, a normalization layer, a multi-head self-attention module, a normalization layer, and a feature fusion module connected in sequence; Each deconvolution module consists of an input layer, a 3×3 transposed convolution, an activation function, and an output layer; The thickness calculation module includes a data transmission part, a data analysis and calculation part, and a result output part.

[0010] The ice-covered image is input into the constructed ice-covered thickness recognition model, and the processing process is as follows: The ice-covered image is subjected to a mixed strip convolution module composed of two convolutions with different convolution kernels to extract the external structural features of the ice-covered image, and feature maps of different sizes corresponding to different convolution kernel sizes are obtained; the obtained feature maps are then subjected to feature extraction and abstraction through two ordinary convolution modules; the processed operational feature maps are then subjected to global and microscopic interaction and transformation of semantic information through the global micro-perception module, and features related to the ice-covered properties are gradually extracted, including the texture and structural information of the ice-covered properties; four deconvolution operations are performed again to restore the size and details of the image, and the extracted features are adjusted and optimized to obtain an image with the same size as the original input and integrating and refining the features extracted by different modules; the processed image is then segmented through the image segmentation module to generate an ice-covered segmentation image with foreground pixel values ​​of 255 and background pixel values ​​of 0; finally, the pixel values ​​in the image are calculated through the thickness calculation module, and the ice thickness in the ice-covered image is identified and obtained in combination with the proportional relationship between the actual thickness and the pixel value.

[0011] The specific process of mixed strip convolution of ice-covered images is as follows: The ice-covered image is first subjected to a 5×5 ordinary convolution to enhance the convolution channel, and then 1×3 and 3×1 convolutions are used respectively. The external structural features of the ice-covered image are extracted by using 1×3 and 3×1 convolutions with an interval of 2 and 1×3 and 3×1 convolutions with an interval of 5. Finally, the image features of different receptive fields are fused to obtain the feature map.

[0012] The global micro-perception module performs global and microscopic interaction and conversion of semantic information on the processed operational feature map, and gradually extracts features related to ice cover properties. The specific process is as follows: First, the feature dimension is transformed through the convolution layer to reduce the channel dimension; then the feature perception is performed at the global and micro levels respectively. Among them, the micro perception module perceives the semantic information locally through pooling layers of different sizes, and then performs cascade fusion, and finally restores it to the original dimension through the convolution layer to extract local information; the global perception module uses a multi-head self-attention module to divide the features into image blocks, add a learnable position vector, and add the image vector to the position vector to obtain the position information of the divided image blocks. The specific calculation formula is as follows:

[0013] In the formula, is the location information of the image block, N represents the number of image blocks, is the image block, E pos is the position vector, Then, the three learning variables Q, K, and V are used to calculate the attention representation between variables and the global self-attention. The specific calculation formula is as follows:

[0014] In the formula, Represents the first j Size; represents the attention mechanism function; denote the query vector, key vector and value vector respectively, represents the transpose of the key vector; represents the normalization function; d represents the scaling factor; represents the concatenation function; Z represents the total number of heads in the multi-head attention mechanism; is a weight matrix, which is used to linearly transform, adjust and integrate the concatenated multi-head attention output; After micro-perception and global perception, the respective channel dimensions are first enhanced to restore them to the original feature dimensions; then, the original input is added to the micro-perception features and global perception features, and information aggregation is achieved through the convolutional layer to obtain the output of the global micro-perception module.

[0015] The ice segmentation image output by the image segmentation module is input into the thickness calculation module. Combined with the measured diameter parameter of the transmission line, the proportional relationship between the actual thickness and the pixel value is obtained. The ice thickness is obtained by calculating the pixel area difference before and after ice coating and combining the proportional relationship. The calculation formula is as follows:

[0016] In the formula, D 1 is the actual diameter of the transmission line in the non-icing condition; d 1 is the number of pixels of the transmission line image in the non-icing condition; d 2 is the number of pixels of the transmission line image under ice coverage; h is the actual ice thickness.

[0017] In S5, during the training of the ice thickness recognition model, the number of iterations was set to 50, the learning rate was 0.0001, the non-maximum suppression threshold was 0.6, and the model label confidence threshold was 0.5.

[0018] The beneficial effects of the present invention are: The present invention is based on a method for identifying ice thickness of transmission lines based on a global micro-perception network. A hybrid strip convolution module is proposed to replace the ordinary convolution module in the original neural network. The external structural features of ice on transmission lines are extracted by using strip convolutions of different receptive fields. This not only improves the ability of convolutional neural networks to extract structural features of ice on transmission lines, but also effectively reduces the amount of calculation and improves the detection speed to a certain extent. At the same time, a global micro-perception module for interaction of global and local features is proposed. This module performs feature perception on both global and micro aspects, fully explores the connection between semantic information of objects in the image, and improves the accuracy of subsequent image segmentation, thereby improving the accuracy of ice thickness calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of a method for identifying ice thickness of a power transmission line based on a global micro-sensing network of the present invention; Figure 2 It is a structural schematic diagram of an ice thickness identification model in a method for identifying ice thickness of a power transmission line based on a global micro-sensing network of the present invention; Figure 3 It is a structural schematic diagram of a hybrid strip convolution module in an ice thickness recognition model in a method for recognizing ice thickness of a power transmission line based on a global micro-sensing network of the present invention; Figure 4 It is a structural schematic diagram of a hybrid strip convolution module in an ice thickness recognition model in a method for recognizing ice thickness of a power transmission line based on a global micro-sensing network of the present invention; Figure 5 This is a diagram showing the recognition result of an ice-covered image under a complex background by the method for recognizing ice-covered thickness of a transmission line based on a global micro-sensing network of the present invention. DETAILED DESCRIPTION

[0020] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] The present invention provides a method for identifying ice thickness of transmission lines based on a global micro-sensing network. Figure 1 As shown, the specific steps include: S1: Collect image data of the transmission line to form a data set.

[0022] Considering the complexity and harmfulness of the transmission line icing scene, a high-precision camera is installed on the transmission line to collect the icing image of the transmission line. This method has the characteristics of high sampling rate, high reliability, low power consumption, real-time transmission, and good anti-interference ability and stability, which can meet the requirements of icing scene data collection. The collected icing images constitute the data set.

[0023] S2: Preprocess the ice-covered images in the dataset and divide them into training set, validation set and test set according to the ratio.

[0024] Since the directly collected ice-covered images of transmission lines have large pixels, they cannot be directly used for training or testing models. Therefore, it is necessary to preprocess and resize the collected ice-covered images. After obtaining the ice-covered images of the ice-covered scene, they are uniformly resized to 512×512 and preprocessed with histogram equalization to make them suitable for training deep learning models.

[0025] Then perform data cleaning, denoising and normalization on the image; Data cleaning is used to delete some useless pictures, such as pictures of power lines not covered with ice, pictures with low clarity, pictures with too cluttered backgrounds, and pictures of lines not covered with ice, in order to prevent adverse effects on model training results.

[0026] Denoising includes the use of mean filtering, median filtering, and Gaussian filtering methods to remove noise interference in the image and enhance image details.

[0027] Normalization includes linear normalization and nonlinear normalization to reduce the numerical complexity of the calculation process and also help reduce the registration error caused by the differences between different images.

[0028] The preprocessed dataset is divided into training set, validation set and test set in a ratio of 3:2:2.

[0029] S3: Label the ice-covered images in the training set and validation set using the Labelme annotation tool to generate a json configuration file and an 8-bit grayscale image.

[0030] Considering that the constructed model requires a large number of 8-bit grayscale images during the training process, the ice-covered images of the training set and the validation set are first annotated as a preliminary process. The Labelme annotation tool is used to annotate the ice-covered area. After the annotation is completed, a json configuration file can be automatically generated. Based on the json configuration file, a 24-bit grayscale image is generated, and then the 24-bit grayscale image is converted into an 8-bit grayscale image in batches through the image pixel conversion method.

[0031] S4: Construct an ice thickness identification model.

[0032] The division of the ice-covered area of ​​the transmission line is a prerequisite for calculating the ice thickness. Only by ensuring the accuracy of the ice-covered area segmentation can the accurate ice thickness be obtained. Therefore, we proposed an ice thickness recognition model based on the global micro-perception network, and adopted an encoder-decoder architecture to segment the ice and snow covered area, and used the global micro-perception module (GMAM) to enhance the semantic information conversion capability at the bottleneck.

[0033] The ice thickness recognition model constructed is as follows: Figure 2 As shown, it includes two mixed strip convolution modules, two convolution modules, a global micro-perception module, four deconvolution modules, an image segmentation module and a thickness calculation module connected in sequence.

[0034] The ice-covered image is input into the constructed ice-covered thickness recognition model, and the processing process is as follows: The ice-covered image is subjected to a mixed strip convolution module composed of two convolutions with different convolution kernels to extract the external structural features of the ice-covered image, and feature maps of different sizes corresponding to different convolution kernel sizes are obtained; the obtained feature maps are then subjected to feature extraction and abstraction through two ordinary convolution modules; the processed operational feature maps are then subjected to global and microscopic interaction and transformation of semantic information through the global micro-perception module, and features related to the ice-covered properties are gradually extracted, including the texture and structural information of the ice-covered properties; four deconvolution operations are performed again to restore the size and details of the image, and the extracted features are adjusted and optimized to obtain an image with the same size as the original input and integrating and refining the features extracted by different modules; the processed image is then segmented through the image segmentation module to generate an ice-covered segmentation image with foreground pixel values ​​of 255 and background pixel values ​​of 0; finally, the pixel values ​​in the image are calculated through the thickness calculation module, and the ice thickness in the ice-covered image is identified and obtained in combination with the proportional relationship between the actual thickness and the pixel value.

[0035] The ordinary convolution module is a convolution with the same length and width. When ordinary convolution is used to detect the image of the power transmission line with ice, it is easy to cause the detection area to contain a large number of background-irrelevant areas, thereby extracting redundant image edge features. To address this problem, the present invention proposes a mixed strip convolution module (MSCM) to replace the ordinary convolution module in the original neural network. The mixed strip convolution module extracts the external structural features of the ice-covered power transmission line by using strip convolutions with different receptive fields. It is more suitable for detecting strip-type objects such as transmission lines than ordinary convolution modules.

[0036] Each mixed strip convolution module, such as Figure 3 As shown in the figure, it specifically includes 1 5×5 convolution, and 1 5×5 convolution is respectively connected to 1 1×3 convolution with an interval number of 1, 1×3 convolution with an interval number of 2, and 1×3 convolution with an interval number of 5; the 1×3 convolution with an interval number of 1, the 1×3 convolution with an interval number of 2, and the 1×3 convolution with an interval number of 5 are respectively connected to 1 3×1 convolution with an interval number of 1, 1 3×1 convolution with an interval number of 2, and 1 3×1 convolution with an interval number of 5.

[0037] The ice-covered image first undergoes a 5×5 ordinary convolution to enhance the convolution channel, and then uses 1×3 and 3×1 convolutions respectively, and uses 1×3 and 3×1 convolutions with an interval of 2 and 1×3 and 3×1 convolutions with an interval of 5 to extract the external structural features of the ice-covered image. Finally, the image features of different receptive fields are fused to obtain the feature map. This operation not only improves the ability of the convolutional neural network to extract the structural features of ice-covered transmission lines, but also effectively reduces the amount of calculation and improves the detection speed to a certain extent.

[0038] The mixed strip convolution module and the ordinary convolution module constitute the encoder of the ice thickness recognition model, and each convolution module includes two 3×3 convolutions connected in sequence.

[0039] The global micro-perception module is a bottleneck area located between the encoder and the decoder. It is responsible for converting the semantic information generated by the encoder and removing irrelevant information such as background, and retaining useful semantic features for the decoder. This area is related to the accuracy of subsequent image segmentation and plays a vital role. Figure 4 As shown, it specifically includes convolutional layers, micro-perception modules and global perception modules. The micro-perception module includes maximum pooling layers with sizes of k5, k9 and k13, respectively, connected to cascade fusion layers; the global perception module includes a linear projection layer, a normalization layer, a multi-head self-attention module, a normalization layer, and a feature fusion module connected in sequence.

[0040] First, the feature dimension is transformed through the convolution layer to reduce the channel dimension to obtain more compact feature information and accelerate model training; then the feature perception is performed at the global and micro levels respectively. Among them, the micro perception module uses the feature pyramid pooling module to perceive the semantic information locally through pooling layers of different sizes, and then performs cascade fusion, and finally restores it to the original dimension through the convolution layer to extract local information; the global perception part uses the self-attention module for perception, learns the global relationship between different features, and uses the multi-head self-attention module to divide the features into image blocks, add a learnable position vector, and add the image vector to the position vector to obtain the position information of the divided image blocks. The specific calculation formula is as follows:

[0041] In the formula, is the location information of the image block, N represents the number of image blocks, is the image block, E pos is the position vector, Then, the three learning variables Q, K, and V are used to calculate the attention representation between variables and the global self-attention. The specific calculation formula is as follows:

[0042] In the formula, Represents the first j Size; represents the attention mechanism function; denote the query vector, key vector and value vector respectively, represents the transpose of the key vector; represents the normalization function; d represents the scaling factor; represents the concatenation function; Z represents the total number of heads in the multi-head attention mechanism; is a weight matrix, which is used to linearly transform, adjust and integrate the concatenated multi-head attention output; After micro-perception and global perception, the respective channel dimensions are first enhanced to restore them to the original feature dimensions; then, the original input is added to the micro-perception features and global perception features, and information aggregation is achieved through the convolutional layer to obtain the output of the global micro-perception module.

[0043] Each deconvolution module consists of an input layer, a 3×3 transposed convolution, an activation function, and an output layer; The image segmentation module includes an image segmentation algorithm module, which uses a segmentation algorithm based on edge detection. Specifically, it uses a Canny operator for edge detection to achieve image segmentation. This method can detect as many real edges in the image as possible, while misjudging non-edge points as edge points as possible, and the detected edge position is as close to the actual edge position as possible, and each edge point is only responded to once to avoid repeated detection. This image segmentation method has good edge detection performance, can detect weak edges, and has good noise suppression ability.

[0044] The thickness calculation module includes a data transmission part, a data analysis and calculation part, and a result output part. The ice segmentation image output by the image segmentation module is input into the thickness calculation module, and combined with the measured diameter parameter of the transmission line, the proportional relationship between the actual thickness and the pixel value is obtained. The ice thickness is obtained by calculating the pixel area difference before and after ice coating and combining the proportional relationship. The calculation formula is as follows:

[0045] In the formula, D 1 is the actual diameter of the transmission line in the non-icing condition; d 1 is the number of pixels of the transmission line image in the non-icing condition; d 2 is the number of pixels of the transmission line image under ice coverage; h is the actual ice thickness.

[0046] S5: Use the training set, validation set and test set to train, optimize and evaluate the ice thickness recognition model respectively until the accuracy error between the evaluation result and the training result is less than 10%, then stop training and evaluation.

[0047] The constructed ice thickness recognition model is trained using the preprocessed data set. The parameters of the model are adjusted according to the performance indicators in the training process, such as accuracy, recall rate, loss rate, etc., so that it can accurately identify the ice thickness of the transmission line and achieve the best recognition effect. The specific operation is: in the training process, a large number of ice-covered images of transmission lines in different scenes in the training set that have been preprocessed and divided in advance are used to train the model. At the same time, some images in the validation set are used to optimize the training process of the model, so that it can finally output the optimal weight with the highest accuracy, the highest recall rate, and the lowest loss rate. During the training of the ice thickness recognition model, the number of iterations is set to 50, the learning rate is 0.0001, the non-maximum suppression threshold is 0.6, and the model label confidence threshold is 0.5. Finally, the trained model is tested and evaluated using ice images in the test set that are completely different from the training set and validation set. When the accuracy error between the test result and the training result is less than 10%, it means that the model training and evaluation are completed, and the training is stopped.

[0048] S6: Deploy the trained ice thickness recognition model to the edge server, transmit the image to be detected to the edge server through a wireless or wired network for recognition, and output the ice thickness.

[0049] The trained ice thickness recognition model is deployed on the edge server. With its powerful computing power and storage resources, it can process a large amount of image data and transmit the detection results to the cloud through a high-speed and stable wireless network, so as to achieve real-time monitoring of the ice coverage of the transmission line, reduce the delay of data transmission, and improve the real-time performance of detection. It is especially suitable for some scenarios with high requirements for response speed, and provide timely and accurate information for the safe operation of the power system. Figure 5 As shown, this is a test result diagram of the method of the present invention under a complex background, which can clearly present the ice coverage state.

[0050] Example 1 This embodiment provides a method for identifying ice thickness of a transmission line based on a global micro-sensing network. Figure 1 As shown, the specific steps include: S1: collect image data of the transmission line to form a data set; S2: Preprocess the ice-covered images in the dataset and divide them into training set, validation set and test set according to the ratio; S3: Label the ice-covered images in the training set and validation set using the Labelme annotation tool to generate a json configuration file and an 8-bit grayscale image; S4: Construct ice thickness identification model; S5: Use the training set, validation set and test set to train, optimize and evaluate the ice thickness recognition model respectively until the accuracy error between the evaluation result and the training result is less than 10%, and then stop training and evaluation; S6: Deploy the trained ice thickness recognition model to the edge server, transmit the image to be detected to the edge server through a wireless or wired network for recognition, and output the ice thickness.

[0051] Example 2 On the basis of Example 1, S1 specifically includes: collecting ice-covered images of the transmission line by installing a camera on the transmission line to form a data set.

[0052] In S2, the ice-covered images in the data set are preprocessed, including adjusting the images to a uniform size and performing histogram equalization, and then performing data cleaning, denoising, and normalization on the images; The preprocessed dataset is divided into training set, validation set and test set in a ratio of 3:2:2.

[0053] S3 is as follows: for the ice-covered images in the training set and the validation set, the Labelme annotation tool is used to annotate the ice-covered areas. After the annotation is completed, a json configuration file is automatically generated. A 24-bit grayscale image is generated based on the json configuration file, and then the 24-bit grayscale image is batch converted into an 8-bit grayscale image through the image pixel conversion method.

[0054] In S5, during the training of the ice thickness recognition model, the number of iterations was set to 50, the learning rate was 0.0001, the non-maximum suppression threshold was 0.6, and the model label confidence threshold was 0.5.

[0055] Example 3 On the basis of Example 2, in S4, an ice thickness recognition model is constructed. The ice thickness recognition model is as follows: Figure 2 As shown, it includes two mixed strip convolution modules, two convolution modules, a global micro-perception module, four deconvolution modules, an image segmentation module and a thickness calculation module connected in sequence; Each mixed strip convolution module is as follows Figure 3 As shown in the figure, it specifically includes 1 5×5 convolution, and 1 5×5 convolution is connected to 1 1×3 convolution with an interval number of 1, 1 1×3 convolution with an interval number of 2, and 1×3 convolution with an interval number of 5; and the 1×3 convolution with an interval number of 1, the 1×3 convolution with an interval number of 2, and the 1×3 convolution with an interval number of 5 are respectively connected with 1 3×1 convolution with an interval number of 1, 1 3×1 convolution with an interval number of 2, and 1 3×1 convolution with an interval number of 5; Each convolution module consists of two 3×3 convolutions connected in sequence; Global micro-sensing module such as Figure 4 As shown, it includes a convolutional layer, a micro-perception module and a global perception module. The micro-perception module includes a maximum pooling layer with sizes of k5, k9 and k13, connected to a cascade fusion layer; the global perception module includes a linear projection layer, a normalization layer, a multi-head self-attention module, a normalization layer, and a feature fusion module connected in sequence; Each deconvolution module consists of an input layer, a 3×3 transposed convolution, an activation function, and an output layer; The thickness calculation module includes a data transmission part, a data analysis and calculation part, and a result output part.

[0056] The ice-covered image is input into the constructed ice-covered thickness recognition model, and the processing process is as follows: The ice-covered image is passed through a mixed strip convolution module composed of two convolutions with different convolution kernels to extract the external structural features of the ice-covered image and obtain feature maps of different sizes corresponding to different convolution kernel sizes; The ice-covered image is first subjected to a 5×5 ordinary convolution to enhance the convolution channel, and then 1×3 and 3×1 convolutions are used respectively. The external structural features of the ice-covered image are extracted by using 1×3 and 3×1 convolutions with an interval of 2 and 1×3 and 3×1 convolutions with an interval of 5. Finally, the image features of different receptive fields are fused to obtain the feature map.

[0057] Then, two common convolution modules are used to extract and abstract the obtained feature map; Then, the global micro-perception module performs global and microscopic interaction and conversion of semantic information on the processed operational feature map, and gradually extracts features related to ice properties, including ice texture and structural information. First, the feature dimension is transformed through the convolution layer to reduce the channel dimension; then the feature perception is performed at the global and micro levels respectively. Among them, the micro perception module perceives the semantic information locally through pooling layers of different sizes, and then performs cascade fusion, and finally restores it to the original dimension through the convolution layer to extract local information; the global perception module uses a multi-head self-attention module to divide the features into image blocks, add a learnable position vector, and add the image vector to the position vector to obtain the position information of the divided image blocks. The specific calculation formula is as follows:

[0058] In the formula, is the location information of the image block, N represents the number of image blocks, is the image block, E pos is the position vector, Then, the three learning variables Q, K, and V are used to calculate the attention representation between variables and the global self-attention. The specific calculation formula is as follows:

[0059] In the formula, Represents the first j Size; represents the attention mechanism function; denote the query vector, key vector and value vector respectively, represents the transpose of the key vector; represents the normalization function; d represents the scaling factor; represents the concatenation function; Z represents the total number of heads in the multi-head attention mechanism; is a weight matrix, which is used to linearly transform, adjust and integrate the concatenated multi-head attention output; After micro-perception and global perception, the respective channel dimensions are first enhanced to restore them to the original feature dimensions; then, the original input is added to the micro-perception features and global perception features, and information aggregation is achieved through the convolutional layer to obtain the output of the global micro-perception module.

[0060] After four deconvolution operations, the size and details of the image are restored, and the extracted features are adjusted and optimized to obtain an image with the same size as the original input and integrating and refining the features extracted by different modules; Then, the processed image is segmented by the image segmentation module to generate an ice segmentation image with a foreground pixel value of 255 and a background pixel value of 0. Finally, the pixel value in the image is calculated by the thickness calculation module, and the proportional relationship between the actual thickness and the pixel value is combined to identify and obtain the ice thickness in the ice image.

[0061] The ice segmentation image output by the image segmentation module is input into the thickness calculation module. Combined with the measured diameter parameter of the transmission line, the proportional relationship between the actual thickness and the pixel value is obtained. The ice thickness is obtained by calculating the pixel area difference before and after ice coating and combining the proportional relationship. The calculation formula is as follows:

[0062] In the formula, D 1 is the actual diameter of the transmission line in the non-icing condition; d 1 is the number of pixels of the transmission line image in the non-icing condition; d 2 is the number of pixels of the transmission line image under ice coverage; h is the actual ice thickness.

Claims

1. A method for identifying ice thickness of transmission lines based on a global micro-sensing network, characterized in that: The specific steps include: S1: collect image data of the transmission line to form a data set; S2: Preprocess the ice-covered images in the dataset and divide them into training set, validation set and test set according to the ratio; S3: Label the ice-covered images in the training set and validation set using the Labelme annotation tool to generate a json configuration file and an 8-bit grayscale image; S4: Construct ice thickness identification model; S5: using the training set, validation set and test set to train, optimize and evaluate the ice thickness recognition model respectively, until the accuracy error between the evaluation result and the training result is less than 10%, and then stopping the training and evaluation; S6: Deploy the trained ice thickness recognition model to the edge server, transmit the image to be detected to the edge server through a wireless or wired network for recognition, and output the ice thickness.

2. The method for identifying ice thickness of power transmission lines based on global micro-sensing network according to claim 1 is characterized in that: S1 specifically includes: installing cameras on the transmission lines to collect ice-covered images of the transmission lines to form a data set.

3. The method for identifying ice thickness of power transmission lines based on a global micro-sensing network according to claim 1 is characterized in that: In S2, the ice-covered images in the data set are preprocessed, including adjusting the images to a uniform size and performing histogram equalization, and then performing data cleaning, denoising, and normalization on the images; The preprocessed dataset is divided into training set, validation set and test set in a ratio of 3:2:

2.

4. The method for identifying ice thickness of power transmission lines based on a global micro-sensing network according to claim 1 is characterized in that: S3 is as follows: for the ice-covered images in the training set and the validation set, the Labelme annotation tool is used to annotate the ice-covered areas. After the annotation is completed, a json configuration file is automatically generated. A 24-bit grayscale image is generated based on the json configuration file, and then the 24-bit grayscale image is batch converted into an 8-bit grayscale image through the image pixel conversion method.

5. The method for identifying ice thickness of power transmission lines based on global micro-sensing network according to claim 1 is characterized in that: In S4, an ice thickness recognition model is constructed, wherein the ice thickness recognition model includes two mixed strip convolution modules, two convolution modules, a global micro-perception module, four deconvolution modules, an image segmentation module, and a thickness calculation module connected in sequence; Each of the mixed strip convolution modules specifically includes one 5×5 convolution, and one of the 5×5 convolutions is respectively connected to one 1×3 convolution with an interval number of 1, one 1×3 convolution with an interval number of 2, and one 1×3 convolution with an interval number of 5; and the 1×3 convolution with an interval number of 1, the 1×3 convolution with an interval number of 2, and the 1×3 convolution with an interval number of 5 are respectively connected to one 3×1 convolution with an interval number of 1, one 3×1 convolution with an interval number of 2, and one 3×1 convolution with an interval number of 5; Each of the convolution modules includes two 3×3 convolutions connected in sequence; The global micro-perception module includes a convolutional layer, a micro-perception module and a global perception module. The micro-perception module includes a maximum pooling layer with sizes of k5, k9 and k13, connected to a cascade fusion layer; the global perception module includes a linear projection layer, a normalization layer, a multi-head self-attention module, a normalization layer, and a feature fusion module connected in sequence; Each of the deconvolution modules includes an input layer, a 3×3 transposed convolution, an activation function, and an output layer; The image segmentation module adopts a segmentation algorithm based on edge detection; The thickness calculation module includes a data transmission part, a data analysis and calculation part, and a result output part.

6. The method for identifying ice thickness of power transmission lines based on global micro-sensing network according to claim 5 is characterized in that: The ice-covered image is input into the constructed ice-covered thickness recognition model, and the processing process is as follows: The ice-covered image is subjected to a mixed strip convolution module composed of two convolutions with different convolution kernels to extract the external structural features of the ice-covered image, and feature maps of different sizes corresponding to different convolution kernel sizes are obtained; the obtained feature maps are then subjected to feature extraction and abstraction through two ordinary convolution modules; the processed operational feature maps are then subjected to global and microscopic interaction and conversion of semantic information through the global micro-perception module, and features related to the ice-covered properties are gradually extracted, and the features related to the ice-covered properties include the texture and structural information of the ice-covered property; the image size and details are restored through four deconvolution operations again, and the extracted features are adjusted and optimized to obtain an image with the same size as the original input and integrating and refining the features extracted by different modules; the processed image is then segmented through the image segmentation module to generate an ice-covered segmentation image with a foreground pixel value of 255 and a background pixel value of 0; finally, the pixel values ​​in the image are calculated through the thickness calculation module, and the ice thickness in the ice-covered image is identified and obtained in combination with the proportional relationship between the actual thickness and the pixel value.

7. The method for identifying ice thickness of power transmission lines based on global micro-sensing network according to claim 6 is characterized in that: The specific process of mixed strip convolution of ice-covered images is as follows: The ice-covered image is first subjected to a 5×5 ordinary convolution to enhance the convolution channel, and then 1×3 and 3×1 convolutions are used respectively. The external structural features of the ice-covered image are extracted by using 1×3 and 3×1 convolutions with an interval of 2 and 1×3 and 3×1 convolutions with an interval of 5. Finally, the image features of different receptive fields are fused to obtain the feature map.

8. The method for identifying ice thickness of power transmission lines based on global micro-sensing network according to claim 6 is characterized in that: The global micro-perception module performs global and microscopic interactions and conversions on the processed operational feature maps, and gradually extracts features related to ice cover properties. The specific process is as follows: First, the feature dimension is transformed through the convolution layer to reduce the channel dimension; then the feature perception is performed at the global and micro levels respectively. Among them, the micro perception module perceives the semantic information locally through pooling layers of different sizes, and then performs cascade fusion, and finally restores it to the original dimension through the convolution layer to extract local information; the global perception module uses a multi-head self-attention module to divide the features into image blocks, add a learnable position vector, and add the image vector to the position vector to obtain the position information of the divided image blocks. The specific calculation formula is as follows: In the formula, is the location information of the image block, N represents the number of image blocks, is the image block, E pos is the position vector, Then, the three learning variables Q, K, and V are used to calculate the attention representation between variables and the global self-attention. The specific calculation formula is as follows: In the formula, Represents the first j Size; represents the attention mechanism function; denote the query vector, key vector and value vector respectively, represents the transpose of the key vector; represents the normalization function; d represents the scaling factor; represents the concatenation function; Z represents the total number of heads in the multi-head attention mechanism; is a weight matrix, which is used to linearly transform, adjust and integrate the concatenated multi-head attention output; After micro-perception and global perception, the respective channel dimensions are first enhanced to restore them to the original feature dimensions; then, the original input is added to the micro-perception features and global perception features, and information aggregation is achieved through the convolutional layer to obtain the output of the global micro-perception module.

9. The method for identifying ice thickness of power transmission lines based on global micro-sensing network according to claim 6 is characterized in that: The ice segmentation image output by the image segmentation module is input into the thickness calculation module. Combined with the measured diameter parameter of the transmission line, the proportional relationship between the actual thickness and the pixel value is obtained. The ice thickness is obtained by calculating the pixel area difference before and after ice coating and combining the proportional relationship. The calculation formula is as follows: In the formula, D 1 is the actual diameter of the transmission line in the non-icing condition; d 1 is the number of pixels of the transmission line image in the non-icing condition; d 2 is the number of pixels of the transmission line image under ice coverage; h is the actual ice thickness.

10. The method for identifying ice thickness of power transmission lines based on global micro-sensing network according to claim 1 is characterized in that: In S5, during the training of the ice thickness recognition model, the number of iterations was set to 50, the learning rate was 0.0001, the non-maximum suppression threshold was 0.6, and the model label confidence threshold was 0.5.

Images