Ice detection method for power transmission line based on pwl-det detection algorithm model

Abstract

The application discloses a power transmission line icing detection method based on a PWL-Det detection algorithm model, first, a practical application scene reference example of the PWL-Det is given to illustrate the characteristics of the algorithm; then, a MobileNet v3 which is more balanced in speed and accuracy is selected through experiments as a backbone layer network for feature extraction, a lightweight coordinated efficient layer aggregation network is proposed to improve the detection ability of small targets, and a CAFPN is designed to increase the receptive field; finally, the performance of the PWL-Det is verified through comparative experiments and ablation experiments. Experiments show that the PWL-Det has high recognition accuracy and lower requirements for the performance of a computing platform, and is more suitable for edge computing devices.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a method for detecting icing on power transmission lines using the PWL-Det lightweight algorithm. Background Technology

[0002] Transmission lines have certain unique characteristics. They are often distributed across different regions, exhibiting significant dispersion, and are affected by various factors such as terrain and weather conditions, making maintenance and management challenging. Furthermore, high-voltage transmission systems, due to their large capacity, long-distance transmission, and wide coverage area, have a significant impact on politics and the economy; therefore, their reliability must meet extremely high standards. However, these lines often traverse harsh environments such as high altitudes, icing, and acid rain. Icing is an unavoidable problem and can lead to performance degradation and icing accidents.

[0003] An insulator is a device installed between conductors at different potentials or between a conductor and a grounding component. It is capable of withstanding voltage and mechanical stress and is a special type of insulating control widely used in various power systems, including overhead transmission lines, power plants, and substations. Its main function is to support and fix conductors while providing electrical insulation, ensuring the safe and stable operation of the power system. Prolonged exposure to harsh natural environments can lead to insulator contamination, flashover, damage, spontaneous explosion, and missing caps, posing significant hidden dangers to the safe and stable operation of the power system. Therefore, identifying problems in insulators in advance to ensure the reliable and safe operation of electrical energy has become a primary task.

[0004] Ice accumulation on transmission lines and insulators poses a significant risk to the power system, as the weight of the ice can cause line breakage, tower collapse, and insulator rupture.

[0005] Edge computing in the power Internet of Things (IoT) embeds artificial intelligence algorithms into edge devices, preprocessing, computing locally, and performing inference and judgment on data close to the data source. This reduces the communication bandwidth requirements for uploading to the server, lowering transmission latency and power consumption. Edge computing offers high real-time performance and low communication requirements; however, the performance of edge computing devices is often insufficient, necessitating lightweight algorithms. Lightweight algorithms offer advantages such as low computational resource consumption, low power consumption, and fast inference speed. Therefore, this invention proposes the PWL-Det lightweight algorithm for detecting icing on transmission lines. The final experimental results demonstrate that this high-performance monitoring network lays a solid foundation for detecting icing targets on transmission lines. Summary of the Invention

[0006] The purpose of this invention is to provide a transmission line icing detection method based on the PWL-Det algorithm. This method utilizes a lightweight version of the PWL-Det algorithm for transmission line icing detection, significantly reducing the number of parameters and computational load while maintaining accuracy, thus lowering the performance requirements of operating equipment. It overcomes the shortcomings of existing technologies that do not consider the detection performance of targets in complex environments. Experiments show that PWL-Det has high recognition accuracy and lower requirements for computing platform performance, making it more suitable for edge computing devices.

[0007] First, images of icing on transmission lines were acquired, preprocessed, labeled, and divided into training, validation, and test sets. A transmission line icing detection network model was built based on the PWL-Det detection algorithm. Next, the PWL-Det network model was designed, selecting MobileNet v3, which offers a better balance between speed and accuracy, as the backbone network for feature extraction. A lightweight, coordinated, and efficient layer aggregation network was proposed to improve the detection capability for small targets, and CAFPN was designed to increase the receptive field. Then, partial convolution and coordinated attention were combined with the efficient layer aggregation network to make the network focus more on small target features, thereby improving overall performance. Images from the test set were input into the trained network model for detection, obtaining the transmission line icing detection result images. Finally, comparative and ablation experiments were conducted to verify the performance of PWL-Det. The experiments show that PWL-Det has high recognition accuracy and lower requirements for computing platform performance, making it more suitable for edge computing devices.

[0008] The method for detecting icing on transmission lines based on the PWL-Det detection algorithm model is implemented according to the following steps:

[0009] Step 1: Collect images of icing on transmission lines, preprocess the collected images, label them, and divide them into training, validation, and test sets.

[0010] Step 2: Design a PWL-Det network model for transmission line icing detection based on the PWL-Det detection algorithm, including the following steps:

[0011] Step 2.1: Use MobileNet v3 instead of CSPDarknet as the backbone network for feature extraction;

[0012] Step 2.2: Introduce new optimization strategies to optimize the network;

[0013] Step 2.3: Use the lightweight content-aware feature reconstruction module CARAFE to replace the upsampling operation in the original feature pyramid FPN of YOLOv7-tiny to obtain a high-quality icing feature map of transmission lines.

[0014] Step 2.4: In the Coordinated Efficient Layer Aggregation Network (CELAN), Coordinated Attention (CA) is used to make the model pay more attention to the target region, avoid the loss of small target features, and enhance the model's ability to capture cross-channel information in safety wear images. At the same time, partial convolution (PConv) is used as convolution to extract target features, reducing the number of parameters without reducing performance. Furthermore, focal loss and SIoU are used to improve the loss function to solve the problem of imbalanced samples.

[0015] Step 3: Use partial convolution and coordinated attention combined with a coordinated efficient layer aggregation network to train the PWL-Det network model, so that the PWL-Det network model can focus more on small target features, thereby improving the overall performance.

[0016] Step 4: Input the images in the test set into the trained PWL-Det network model for detection to obtain the icing detection result image of the transmission line.

[0017] The innovation of the PWL-Det detection algorithm model in this invention is that it replaces the upsampling operation in the original FPN of YOLOv7-tiny with Content-aware Reassembly of Features (CARAFE) to obtain high-quality icing feature maps of transmission lines. The algorithm proposes a Content-aware Feature Pyramid Network (CAFPN). CAFPN uses Content-aware Reassembly of Features (CARAFE) for upsampling within the FPN. This module adjusts the reassembly process based on the semantic information of the input features, thereby more accurately capturing key information in the image. Furthermore, CARAFE achieves a larger receptive field, and the entire operator introduces fewer parameters and computational overhead, making it more suitable for lightweight object detection models.

[0018] The beneficial effects of this invention are that it proposes a Coordinate Efficient Layer Aggregation Network (CELAN). CELAN uses Coordinate Attention (CA) to make the model focus more on the target region. Simultaneously, it uses Partial Convolution (PConv) as the convolution to extract target features, reducing the number of parameters without sacrificing performance. While the MobileNet v3 used in this invention reduces the mAP value by 1.73%, it reduces FLOPs by nearly half and the number of parameters by 30.73%. Due to the reduced memory access efficiency caused by grouping operations, the FPS of this invention's algorithm is slightly lower than other comparative algorithms, but it still meets the requirements. The changes in the total loss function value during the training process of each algorithm are shown below. Figure 10 As shown, lightweight models are crucial for resource-constrained edge devices. CSPDarkNet has a large computational load, but high memory access efficiency and good support from platforms supporting parallel processing, such as GPUs. Therefore, models using CSPDarkNet as the backbone layer achieve the highest FPS on GPU platforms. MobileNet v3 has the smallest computational load; however, due to its use of a grouped convolutional structure, its MAC access efficiency is lower, and GPU support for it is generally limited, resulting in slightly lower FPS on GPU platforms. However, on CPU platforms, MobileNet v3 achieves the highest FPS. For resource-constrained edge computing devices, lower computational load and parameter count place lower demands on platform computing performance, allowing for the deployment of more devices, while FPS is sufficient for practical use. Therefore, using MobileNetv3 as the backbone layer in YOLOv7-tiny achieves a better balance between speed, accuracy, and lightweight design. Attached Figure Description

[0019] Figure 1 It is designed for the PWL-Det platform;

[0020] Figure 2 This is a diagram of the PWL-Det network structure;

[0021] Figure 3 It is the MobileNetv3 network architecture;

[0022] Figure 4 It is a content-aware feature reorganization structure diagram;

[0023] Figure 5 It is a diagram of a coordinated and efficient layered aggregation network structure;

[0024] Figure 6 It is a partial convolutional structure diagram;

[0025] Figure 7 It is a comparison of the backbone network;

[0026] Figure 8 It is a comparison of ablation test results;

[0027] Figure 9 It is a comparison of different detection algorithms;

[0028] Figure 10 This refers to the change in the total training loss function value of each algorithm in this invention; Detailed Implementation

[0029] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1 , Figure 2 As shown, the platform design diagram and network structure diagram of the PWL-Det algorithm are presented. This embodiment provides a transmission line icing detection method based on the PWL-Det detection algorithm model, which is implemented according to the following steps:

[0031] Step 1: Collect images of icing on transmission lines, preprocess the collected images, label them, and divide them into training, validation, and test sets.

[0032] Step 2: Design a PWL-Det network model for transmission line icing detection based on the PWL-Det detection algorithm, including the following steps:

[0033] Step 2.1: Use MobileNet v3 instead of CSPDarknet as the backbone network for feature extraction;

[0034] Step 2.2: Introduce new optimization strategies to optimize the network;

[0035] Step 2.3: Use the lightweight content-aware feature reconstruction module CARAFE to replace the upsampling operation in the original feature pyramid FPN of YOLOv7-tiny to obtain a high-quality icing feature map of transmission lines.

[0036] Step 2.4: In the Coordinated Efficient Layer Aggregation Network (CELAN), Coordinated Attention (CA) is used to make the model pay more attention to the target region, avoid the loss of small target features, and enhance the model's ability to capture cross-channel information in safety wear images. At the same time, partial convolution (PConv) is used as convolution to extract target features, reducing the number of parameters without reducing performance. Furthermore, focal loss and SIoU are used to improve the loss function to solve the problem of imbalanced samples.

[0037] Step 3: Use partial convolution and coordinated attention combined with a coordinated efficient layer aggregation network to train the PWL-Det network model, so that the PWL-Det network model can focus more on small target features, thereby improving the overall performance.

[0038] Step 4: Input the images in the test set into the trained PWL-Det network model for detection to obtain the icing detection result image of the transmission line.

[0039] Step 1 is implemented in the following steps:

[0040] Step 1.1: Collect a large number of images through tower cameras and inspection drones, including images of various transmission lines covered with ice and snow. By using close-up and extreme close-up images, keep the number of images of each type of transmission line covered with ice and snow consistent to obtain images of the transmission lines covered with ice.

[0041] Step 1.2: Perform data enhancement and augmentation on the icing image of the transmission line obtained in Step 1.1, including mirroring, reversal, contrast and brightness processing, to obtain an augmented icing image of the transmission line.

[0042] Step 1.3: Combine the icing images of the transmission lines obtained in Step 1.1 and the expanded icing images of the transmission lines obtained in Step 1.2 into a transmission line icing image set, and divide it into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0043] Step 2.1 is implemented in the following steps:

[0044] like Figure 3As shown, Input represents the size variation of each feature layer in MobileNetV3; Operator represents the block structure that each feature layer will go through. It can be seen that feature extraction in MobileNetV3 involves many bottleneck (bneck) structures. A bneck represents a feature extraction block composed of inverse residual structures with linear bottlenecks; expsize and #out represent the number of channels after the inverse residual structure rises within the bneck and the number of channels in the feature layer when input to the bneck, respectively; SE represents whether an attention mechanism is introduced at this layer; NL represents the type of activation function, HS represents h-swish, and RE represents ReLU; s represents the stride used for each block structure. The bneck mainly consists of channel-separable convolutions, the SE channel attention mechanism, and inverse residual connection structures, and is the basic and core module of this network. PWL-Det uses MobileNetv3 instead of CSPDarknet as its backbone layer. MobileNet v3, as a lightweight network, has the advantages of fewer parameters, lower computational cost, and faster detection speed, making it more suitable for edge computing and other scenarios with limited processor performance, storage space, and power consumption. The algorithm has two versions, small and large, and the algorithm of this invention adopts the more lightweight small version. MobileNet v3 employs a variety of techniques to improve network performance, including using the Network Architecture Search (NAS) method to optimize each module to obtain the general network structure, and then using the Network Adaptation (NetAdapt) algorithm to obtain the optimal number of convolutional kernels and channels, so that the model achieves a balance between lightweight and performance.

[0045] Step 2.2 introduces a new optimization strategy to optimize the network, which is implemented according to the following steps.

[0046] (1) Use a lightweight attention module SE (Squeeze-and-Excitation) to improve the feature extraction capability of the model, and use the h-swish activation function to speed up the calculation and improve the accuracy of the network;

[0047] (2) A squeezing operation is used to compress each two-dimensional feature map into corresponding real numbers using global average pooling. Each real number is the importance coefficient of each channel, reflecting the response intensity of that channel in the global range. The real number vector composed of all real numbers can reflect the global distribution of the feature map. The squeezing operation is formalized as shown in Equation 1.

[0048]

[0049] In the formula: z represents the global feature, fsq This represents the squeeze operation, where x represents the input feature map, H is the height of the feature map, and W is the width of the feature map. i,j Let be the feature vector of the pixel in the i-th row and j-th column;

[0050] (3) An activation operation is used to learn the importance coefficients through a fully connected layer, thereby obtaining a weight vector. The weights represent the influence of each channel on feature extraction. The activation operation is formalized as shown in Equation 2.

[0051] s = f ex (z,W)=σ(W2δ(W1z)) Equation 2

[0052] In the formula: s represents the motivation score vector, f ex This is represented as an incentive operation. This represents the weight matrix of row C. σ represents the weight matrix in column C, where σ is the hardsigmoid function and δ is the ReLU activation function. The hardsigmoid function is a function proposed by MobileNetv3 to solve the problems of computational complexity and quantization difficulty of the sigmoid function. Its formula is shown in Equations 3 and 4.

[0053] ReLU6(x) = min(max(x,0),6) Equation 3

[0054]

[0055] (4) Finally, the weights are applied to the two-dimensional feature map to obtain the final features. The feature map of each channel is weighted, and the channel with the most feature information is retained while the channel with less information is suppressed.

[0056] Step 2.3 is implemented in the following steps.

[0057] like Figure 4 As shown, input feature channel compression is performed in the upsampling kernel prediction part, compressing the channels from C to Cupsampling kernel. m This reduces the number of input feature mapping channels while decreasing the parameters and computational cost of subsequent steps, thus improving the efficiency of CARAFE. Content encoding uses an upsampling factor of σ and an upsampling size of K. up ×K up The convolution operation will input channel C m This is transformed into σ×k implementation, and then the channel is expanded in the spatial dimension as follows: The upsampling kernel is then normalized using Softmax so that the sum of the kernel weights is 1. During feature reconstruction, each position in the output feature map is mapped back to the input feature map, and K values ​​centered at that position are extracted. up ×K upThe original feature map region. Subsequently, this original feature map region N(x) l ,k up ) and the corresponding predicted upsampling kernel w′ l Performing a dot product operation, we finally obtain a new feature map σH×σW×C.

[0058] The YOLOv7-tiny original Feature Pyramid Network (FPN) uses a lightweight Content-aware Reassembly of Features (CARAFE) module to replace the upsampling operation, resulting in high-quality icing feature maps of transmission lines. CARAFE adjusts the reconstruction process based on the semantic information of the input features, thereby more accurately capturing key information in the image. Furthermore, CARAFE achieves a larger receptive field and introduces fewer parameters and computational overhead, making it more suitable for lightweight object detection models.

[0059] Step 2.4 is implemented in the following manner.

[0060] To avoid losing features of small targets, a CA attention mechanism is used before the last convolutional module in CELAN. The CA mechanism combines channel attention with orientation-related positional information, enhancing the model's ability to capture cross-channel information in safety wearable images and incorporating orientation- and position-sensitive information. This allows the model to more accurately locate and identify target regions. The CELAN network structure diagram is shown below. Figure 5 .

[0061] Its network structure is as follows Figure 6 As shown, C is the number of image channels, H is the image height, W is the image width, r is the downsampling ratio, XAvgPool is the horizontal average pooling, YAvgPool is the vertical average pooling, Nonlinear is the non-linear activation function, Sigmoid is the activation function, and Re-weight is the weight acquisition.

[0062] The CA attention mechanism first uses pooling kernels of size (H,1) and (1,W) to encode each channel along the horizontal and vertical directions, respectively, as shown in Equations 5 and 6.

[0063]

[0064]

[0065] Let x represent the height feature map and width feature map of the c-th channel output, respectively. c (h,i), xc (j, w) represent the input of the c-th channel of the input feature map along the H and W directions, respectively; then, the feature maps in the two directions are concatenated, dimensionality reduced, normalized, and subjected to a nonlinear activation function to obtain the intermediate feature map, calculated as shown in Equation 7.

[0066] f=δ(F1([z h ,z w Formula 7

[0067] In the formula: δ is the nonlinear activation function, F1 is the convolution transformation function, and represents the concatenation operation along the spatial dimension. Then, the feature map f is split into two separate tensors f along the spatial dimension. h and f w After passing through two 1x1 convolutional kernels and a sigmoid activation function, the attention weights g" in the height direction and g" in the width direction of the feature map are obtained, as shown in Equations 8 and 9.

[0068] g h =σ(F h (f h Formula 8

[0069] g w =σ(F w (f w Formula 9

[0070] In the formula: σ is the Sigmoid activation function, F h and F w This is a convolution operation. Finally, attention weights g are used. h and g w The image input to the model is multiplied and weighted to obtain a feature map with attention weights in the width and height directions, as shown in Equation 10.

[0071]

[0072] In the formula: x c and y c These are the input and output of the c-th channel, respectively. and These represent the attention weights along the H and W directions in the c-th channel, respectively.

[0073] Step 3 is implemented in the following steps:

[0074] Step 3.1: Input the training set images divided in Step 2 into the PWL-Det algorithm model for training, set the batch_size to 32, train with stochastic gradient descent (SGD) for 300 rounds, and obtain the training results;

[0075] Step 3.2: Test the PWL-Det algorithm model using the validation set, obtain the validation results, and analyze the differences between the training results and the validation results;

[0076] Step 3.3: By fine-tuning the hyperparameters of the PWL-Det algorithm on the validation set, the learning rate is fine-tuned to 0.001 and the batch_size is set to 32 to obtain the optimized PWL-Det algorithm model.

[0077] Step 4 is implemented in the following steps:

[0078] The test set is input into the optimized PWL-Det algorithm model obtained in step 2.3 for optimization. The specific process is as follows: PWL-Det algorithm model hyperparameter settings: input image pixel size 640*640, training 300 rounds, batch size set to 32, momentum size 0.9, weight decay set to 0.0005, and learning rate set to 0.001.

[0079] To effectively compare the advantages of the PWL-Det algorithm with various other algorithms, we used consistent training and testing methods. The PWL-Det algorithm significantly reduces FLOPs and parameter count by improving the backbone layer and using PConv, making it far more lightweight than the four algorithms. For example, the computational cost and parameter count of the model in this invention are only 36.71% and 53.23% of YOLOv5-s, respectively, demonstrating a significant improvement in lightweight design. Furthermore, the addition of the CA attention mechanism and the use of the CARAFE upsampling method slightly improves the mAP value, resulting in the highest mAP value among the four algorithms. Due to the decreased memory access efficiency caused by grouping operations, the FPS of the algorithm in this invention is slightly lower than the other compared algorithms, but it still meets the requirements. The changes in the total loss function value during the training process of each algorithm are shown below. Figure 10 As shown.

Claims

1. A method for detecting icing on a power transmission line based on a PWL-Det detection algorithm model, characterized in that, The method is implemented according to the following steps: Step 1: Collecting the transmission line icing image, preprocessing the collected transmission line icing image, marking and dividing into a training set, a validation set and a test set; Step 2: Designing a transmission line icing detection PWL-Det network model based on the PWL-Det detection algorithm, including the following steps: Step 2.1: Using MobileNet v3 instead of CSPDarknet as the backbone layer network for feature extraction; Step 2.2: Introducing a new optimization strategy to optimize the network; Step 2.3: Using a lightweight content-aware feature reorganization module CARAFE instead of the up-sampling operation in the original feature pyramid FPN of YOLOv7-tiny to obtain high-quality transmission line icing feature maps; Step 2.4: Using coordinated attention CA in the coordinated efficient layer aggregation network CELAN to make the model pay more attention to the target area, avoid the loss of small target features, enhance the model's ability to capture cross-channel information in the safety wearing image, use partial convolution PConv as the convolution to extract target features, reduce the parameter amount without reducing the performance, and use focal loss and SIoU to improve the loss function, solve the problem of sample imbalance; Step 3: Training the PWL-Det network model using partial convolution and coordinated attention combined with the coordinated efficient layer aggregation network, making the PWL-Det network model more focused on small target features, thereby improving the overall performance; Step 4: Inputting the images in the test set into the trained PWL-Det network model for detection to obtain the transmission line icing detection result image.

2. The method for transmission line icing detection based on the PWL-Det detection algorithm model according to claim 1, characterized in that, The step 1 is implemented according to the following steps: Step 1.1: Collecting a large number of images through tower cameras and inspection unmanned aerial vehicles, including various transmission line icing and snow images, through close-up and close-up images, keeping the number of images of each type of transmission line icing and snow consistent, and obtaining transmission line icing images; Step 1.2: Data augmentation and expansion of the transmission line icing images obtained in step 1.1, including mirror, inversion, contrast, brightness processing, thereby obtaining expanded transmission line icing images; Step 1.3: Collecting the transmission line icing images obtained in step 1.1 and the expanded transmission line icing images obtained in step 1.2 as a transmission line icing image set, and dividing them into a training set, a validation set and a test set according to the ratio of 8:1:

1.

3. The method for transmission line icing detection based on the PWL-Det detection algorithm model according to claim 1, characterized in that, The step 2.1 is implemented according to the following steps: PWL-Det uses MobileNetv3 instead of CSPDarknet as the backbone layer, adopts a lighter small version, uses a network architecture search NAS method to optimize each module to obtain the general network structure, and then uses a network adaptation NetAdapt algorithm to obtain the best number of convolution kernels and channels, so that the model balances lightness and performance.

4. The transmission line icing detection method based on the PWL-Det detection algorithm model as described in claim 1, characterized in that, Step 2.2 introduces a new optimization strategy to optimize the network, which is implemented according to the following steps (1) using a lightweight attention module SE to improve the feature extraction capability of the model, and using a h-swish activation function to speed up the calculation and improve the network accuracy; (2) using a squeeze operation to compress each two-dimensional feature map into a corresponding real number in a global average pooling manner, each real number is an importance coefficient of each channel, reflecting the response intensity of the channel in the global range, and the real number vector composed of all real numbers can reflect the global distribution of the feature map; the squeeze operation is formalized as formula 1, wherein: z represents a global feature, f sq represents a squeezing operation, x represents an input feature map, H is the height of the feature map, W is the width of the feature map, x i,j is a feature vector of the i-th row and j-th column pixel; (3) using an excitation operation to learn the importance coefficients through a fully connected layer, so as to obtain a weight vector, the weight representing the influence of each channel on feature extraction; the excitation operation is formalized as formula 2, s = f ex (z,W) = σ(W2δ(W1z)) Equation 2 In the formula: s represents the excitation score vector, f ex is expressed as an excitation operation, represents a weight matrix of the C row, represents a weight matrix of the C column, σ is a hardsigmoid function, and δ is a ReLU activation function, and the formula of the hardsigmoid function is shown in formula 3 and formula 4. ReLU6(x)=min(max(x,0),6) Formula 3 (4) finally, the weight is applied to the two-dimensional feature map to obtain the final feature, each channel of the feature map is weighted, the channel with the largest feature information is retained, and the channel with small information quantity is suppressed.

5. The method for transmission line icing detection based on the PWL-Det detection algorithm model according to claim 1, characterized in that, Step 2.4 is specifically implemented in the following manner The CA attention mechanism first uses a pooling kernel with a size of (H, 1) and (1, W) to encode each channel in the horizontal and vertical directions respectively, as shown in formula 5 and formula 6, respectively represent the height feature map and the width feature map output by the cth channel, x c (h, i), x c (j, w) respectively represent the input of the cth channel of the input feature map along the H direction and the W direction; then, after splicing, dimension reduction, normalization and non-linear activation function of the feature maps in the two directions, the intermediate feature map is obtained, and the calculation formula is as formula 7, f = δ(F1([z h ,z w ])) Equation 7 In the formula: δ is the nonlinear activation function, F1 is the convolution transformation function, representing the concatenation operation along the spatial dimension. Then, the feature map f is split into two separate tensors f along the spatial dimension. h and f w The attention weights g in the height direction of the feature map are obtained after passing through two 1x1 convolutional kernels and a sigmoid activation function. h Attention weights g in the width direction w The formulas are shown in Equations 8 and 9. g h = σ (F h (f h )) Equation 8 g w = σ (F w (f w )) Equation 9 where σ is a sigmoid activation function, F h and F w are convolution operations; finally, using attention weights g h and g w and the image input to the model, a feature map with attention weights in the width direction and the height direction is obtained, as shown in equation 10, where x c and y c are the input and output of the cth channel, respectively, and denote the attention weights along the H and W directions in the cth channel, respectively.

6. The method for transmission line icing detection based on the PWL-Det detection algorithm model according to claim 1, characterized in that, The step 3 is specifically implemented in the following steps: Step 3.1, input the training set image divided in step 2 into the PWL-Det algorithm model for training, set batch_size to 32, use stochastic gradient descent (SGD) to train for 300 rounds, and obtain the training result; Step 3.2, test the PWL-Det algorithm model using the validation set to obtain the validation result, and analyze the difference between the training result and the validation result; Step 3.3, by fine-tuning the hyperparameters of the PWL-Det algorithm on the validation set, the learning rate is fine-tuned to 0.001, and the batch_size is set to 32, to obtain the optimized PWL-Det algorithm model.

7. The method for transmission line icing detection based on the PWL-Det detection algorithm model according to claim 1, characterized in that, Step 4 is specifically implemented in the following steps: Input the test set into the optimized PWL-Det algorithm model obtained in step 2.3 for optimization, the specific process is: the PWL-Det algorithm model hyperparameter setting: input image pixel size 640*640, training 300 rounds, batch size set to 32, momentum size 0.9, weight decay set to 0.0005, and learning rate set to 0.001.

Images