Power transmission line icing detection method based on wide-area perception dynamic convolution model

Abstract

The application discloses a kind of transmission line icing detection methods based on wide-area perception dynamic convolution model, first, multi-dimensional attention module (Multi-Dimensional Attention Module, MDAM) is used to optimize the feature sharing between convolution kernel, more accurately obtain the detailed information of small target;Second, a kind of Thin-head feature extraction network fusing lightweight model (Lightweight model, LM) is proposed, which retains the hidden connection of spatial feature information gradually transmitted to the channel;Using a loss function (Shape-NWD) that combines attention shape cross ratio with normalization, reduces the underfitting caused by positive and negative sample imbalance.Finally, the experiment fully proves the efficiency of the method of the application.

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 based on a wide-area perception dynamic convolution model. 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. Summary of the Invention

[0005] The purpose of this invention is to provide a transmission line icing detection method based on a wide-area sensing dynamic convolution model, which overcomes the shortcomings of existing technologies that do not consider the detection performance of targets in complex environments.

[0006] The technical solution adopted in this invention is a transmission line icing detection method based on a wide-area sensing dynamic convolution model, which is implemented according to the following steps:

[0007] Step 1: Obtain images of icing on transmission lines in different environments through drone inspections, expand the obtained images of icing on transmission lines, and use the expanded images of icing on transmission lines as a real insulator dataset.

[0008] Step 2: Divide the transmission line icing dataset obtained in Step 1 into a training set and a validation set, with a ratio of 9:1 between the training set and the validation set.

[0009] Step 3: Use the Labellmg software to label the icing categories of the transmission lines in the training set in Step 2. Use VOC format for image labeling and save the icing image information of the transmission lines after labeling to obtain the transmission line icing database.

[0010] Step 4: Replace the ordinary convolution in the backbone network with the MDAM multi-dimensional attention module. MDAM extracts feature information from three different scales to obtain feature maps, and then uses convolution and upsampling operations to fuse the feature maps.

[0011] The Thin-head feature extraction network, which integrates a lightweight model, introduces hybrid convolutional RS Conv, including shuffling and DW Conv. RS Conv uses a learnable weight matrix to compute a weighted sum of the features of a node's neighboring nodes, concatenates the weighted sum with the node's own features, and then processes the concatenated features through a non-linear activation function to obtain the updated node representation. In the encoder head, normalized convolutions are replaced with RS Conv. Thin-head also introduces RS-bottleneck and VoV-RSCSPC modules, replacing the RepC3 module in the baseline model with the VoV-RSCSPC module.

[0012] By incorporating the idea of ​​Shape-IoU into normalized NWD, we obtain Shape-NWD;

[0013] The MDAL-DETR object detection network is composed of MDAM, a Thin-head feature extraction network that integrates a lightweight model, and the Shape-NWD loss function.

[0014] Step 5: Use the optimal weight network obtained after training in Step 4 to predict the validation set in Step 2, and obtain the corresponding transmission line icing detection results.

[0015] Step 6: The detection results obtained in Step 5 are evaluated using precision, recall, and mAP values ​​to efficiently determine the efficiency of each model in detecting icing on transmission lines and the practicality of the improved algorithm.

[0016] Furthermore, step 1 is specifically implemented according to the following steps:

[0017] Step 1.1: Use drones to inspect and capture images of ice accretion on power transmission lines. Perform horizontal flipping, random extraction, rotation transformation, brightness enhancement, and contrast enhancement operations on the obtained images of ice accretion on power transmission lines to expand the dataset.

[0018] Step 1.2: Reintegrate the icing images of the transmission lines captured in Step 1.1 with the expanded icing images of the transmission lines into a new dataset as a sample library.

[0019] Furthermore, step 3 is specifically implemented according to the following steps:

[0020] Step 3.1: Select the icing image of the transmission line to be marked, accurately identify and mark the icing area of ​​the transmission line, and obtain the marking box of the icing area of ​​the transmission line. According to the icing image of the transmission line in the marking box, it is divided into two types of labels: transmission line icing (ice transmission) and insulator icing (ice insulator). Finally, the marking boxes of transmission line icing and insulator icing area with category labels are obtained.

[0021] Step 3.2: Use the Labellmg annotation software to generate XML tag files corresponding to the icing images of transmission lines, and store the annotated sample data in Pascal VOC format. The sample data folder is named VOC devkit, which includes three folders: Annotations, JPEG Images, and Image Sets. The Annotations folder contains all the annotated XML tag files for icing of transmission lines and insulators. The XML tag file includes the image ID, image path, image name, and the image's pixel height and width. The image's pixel height and width are represented by the four coordinates of a rectangle, including x, y, and g. min ,y min ,x max ,y max , where (x min ,y min (x) is the coordinate of the top-left vertex of the rectangle. max ,y max The coordinates of the bottom right vertex of the rectangle are given, and each labeled file corresponds to an original image in JPEG Images. The JPEG Images folder contains all the labeled original images. The Image Sets folder has a Main subfolder, where the train document contains the names of all training set images, and the val document contains the names of all validation set images.

[0022] Furthermore, step 4 is specifically implemented according to the following steps:

[0023] Step 4.1: Replace the ordinary convolutions in the backbone network with the MDAM multidimensional attention module, as follows:

[0024] The multidimensional attention mechanism comprises three branches, let F inTo obtain the initial feature map F for the input dimension attention mechanism, features are first extracted using a convolutional layer with ReLU activation. r Then feature map F r After pooling, the data is fed into an existing multilayer perceptron (MLP) for further feature extraction. The extracted features are then multiplied by the channel statistical coefficients using a sigmoid activation function to obtain the feature map. Finally, the feature map Input feature map F with multi-dimensional attention accumulation mechanism in The output of each branch is obtained by adding elements one by one.

[0025] The feature map is calculated as shown in equation (1):

[0026] F r =f r (F in Equation (1)

[0027] In the formula: f r (·) indicates the first step of feature extraction, which includes two convolutional operations and one ReLU activation operation; F in The input feature map representing the multi-dimensional attention mechanism; F r This represents the feature map obtained after the initial feature extraction of the r-th branch, where r = 1, 2, 3;

[0028] For feature map F r Further feature extraction is performed using the steps shown in equations (2), (3), and (4):

[0029] f ca (F r )=σ(MLP(Aνgpool(F r ))+MLP(Maxpool(F r Formula (2)

[0030] coef = f ca (F r Equation (3)

[0031]

[0032] In the formula: f ca (·) represents the execution operation of the multi-dimensional attention mechanism; coef represents the multi-dimensional statistical coefficient; The r-th branch of the multi-dimensional attention mechanism initially extracts the features and then multiplies them by the activation function and coef, where r takes the values ​​1, 2, or 3; Avgpool(·) represents the average pooling operation; Maxpool(·) represents the max pooling operation.

[0033] Ultimately, each output of the multi-dimensional attention module As shown in equation (5):

[0034]

[0035] In the formula: The final output feature map of the r-th branch representing the multi-dimensional attention mechanism;

[0036] The multidimensional attention mechanism has three branches with the same structure but different kernel sizes: Conv1, Conv2, and Conv3, with kernel sizes of 3*3, 5*5, and 7*7, respectively. Therefore, the multidimensional attention mechanism extracts feature information from three different scales to obtain feature maps. Use F out Convolution and upsampling operations fuse feature maps As shown in equation (6):

[0037]

[0038] In the formula: f fuse (·) represents a fusion operation, [·] represents a join operation, F out Output feature map representing a multi-dimensional attention mechanism;

[0039] Step 4.2: The Thin-head feature extraction network integrating the lightweight model introduces hybrid convolutional RS Conv, including shuffle and DW Conv; the Thin-head feature extraction network integrating the lightweight model is as follows:

[0040] RS Conv uses a learnable weight matrix to compute a weighted sum of the features of a node's neighboring nodes, concatenates the weighted sum with the node's own features, and then processes the concatenated features through a non-linear activation function to obtain the updated node representation. Specifically, in the encoder head, the normalized convolution is replaced with RS Conv. Based on RS Conv, the Thin-head introduces the RS-bottleneck and VoV-RSCSPC modules; the RepC3 module in the baseline model is replaced with the VoV-RSCSPC module.

[0041] Step 4.3: The Shape-IoU loss function is used to address the issue of slow convergence during sample training. Simultaneously, the Normalized Wasserstein Distance (NWD) strategy is combined to address the baseline model's GIoU neglecting pixel-level similarity and the global structure and distribution of images. The Shape-NWD loss function calculation formula is as follows:

[0042] The Shape-IoU formula is shown below:

[0043]

[0044]

[0045] In the formula: w and h are the width and height of the bounding box, respectively. gt h gt This refers to the actual width and height of the frame; w c and h c represents the width and height of the minimum detection box covering the ground truth box and the anchor box, respectively. scale is a scaling factor that is related to the size of the target in the dataset. ww and hh represent the weight coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the ground truth box.

[0046] The corresponding bounding box shape loss is as follows:

[0047] L shape-IoU =1-IoU+distance shape +0.5×Ω shape Equation (13)

[0048] By incorporating the concept of Shape-IoU into normalized NWD, we obtain Shape-NWD, whose formula is as follows:

[0049]

[0050] In the formula: D is the Euclidean distance between the center point of the GT box and the center point of the anchor box, weight = 2, and C is a constant related to the dataset;

[0051] Step 4.4: Combine the MDAM from Step 4.1, the Thin-head feature extraction network that integrates the lightweight model from Step 4.2, and the Shape-NWD loss function from Step 4.3 to form the MDAL-DETR object detection network with better detection performance.

[0052] Furthermore, step 5 is specifically implemented according to the following steps:

[0053] Step 5.1: Input the training set images divided in Step 2 into the MDAL-DETR object detection network model obtained in Step 4.4 for training; finally, the optimized wide-area perception dynamic convolution model is obtained, and the best weight data is obtained.

[0054] Step 5.2: Predicting the training results requires two files: mdal-detr.py and predict.py. First, you need to modify model_path and classes_path in mdal-detr.py. model_path points to the trained optimal weight file, which is located in the logs folder, and classes_path points to the txt file corresponding to the detected classes. After making the modifications, start the prediction.

[0055] Furthermore, the specific method for step 5.1, which involves inputting the training set images divided in step 2 into the MDAL-DETR object detection network model obtained in step 4.4 for training, is as follows:

[0056] In the `train.py` file, stochastic gradient descent with a momentum of 0.9 is used for 250 training epochs. The input image pixel size is set to 640*640. The batch size is frozen for 50 epochs to 32, and then unfrozen for 200 epochs to 4. `num_workers2` is used, with the Adam optimizer, a decaying weight coefficient of 5*10⁻⁴, and an initial learning rate of 1*10⁻⁵. An IoU threshold of 0.5 is set for testing on the training set. During training, the learning rate is fine-tuned to 0.003 for better robustness. The first 50 epochs of frozen training result in rapid loss reduction, while the last 200 epochs of unfrozen training allow for continuous fine-tuning of the network. After 200 epochs, the loss on the training set gradually decreases, ultimately resulting in an optimized wide-area-aware dynamic convolutional model with the best weight data.

[0057] The beneficial effects of this invention are:

[0058] This paper presents a method for detecting icing on power transmission lines based on a wide-area perception dynamic convolutional model. A detection network capable of effectively identifying icing on power transmission lines under complex conditions is constructed. First, a multi-dimensional attention module (MDAM) is employed to optimize feature sharing among convolutional kernels, more accurately acquiring detailed information about small targets. Second, a thin-head feature extraction network integrating a lightweight model (LM) is proposed, preserving hidden connections that progressively transmit spatial feature information to the channel. Then, a loss function combining shape crossover ratio and normalization (Shape-NWD) is used to reduce underfitting caused by imbalanced positive and negative samples, improving the completeness of target detection. Finally, experimental data demonstrate that the detection network designed in this invention can accurately locate icing areas on power transmission lines in harsh environments, effectively reducing false alarms and omissions in icing problems on power transmission lines and insulators, achieving an mAP of 97.85%. This lays a solid foundation for the detection of icing targets on power transmission lines. Attached Figure Description

[0059] Figure 1 This invention is based on the wide-area perceptual dynamic convolution algorithm;

[0060] Figure 2 This is a network structure diagram of the multi-dimensional attention module;

[0061] Figure 3 It is a Thin-head feature extraction network that integrates a lightweight model;

[0062] Figure 4 This is an example diagram of a loss function that combines morphological crossover ratio and normalization;

[0063] Figure 5 This is an example dataset diagram of the transmission line icing detection method based on the wide-area perception dynamic convolution model of this invention;

[0064] Figure 6 This is a three-dimensional comparison of the mAP values ​​of various models in the transmission line icing detection method based on the wide-area perception dynamic convolution model of this invention;

[0065] Figure 7 is a comparison of the detection performance of the network of the present invention with that of other representative algorithms; Figure 7(a) is the original image, Figure 7(b) is SwinTransformer, Figure 7(c) is YOLOv7-detr; Figure 7(d) is YOLOv8-detr; Figure 7(e) is the model used in the present invention.

[0066] Figure 8 shows the evolution of the training and validation loss values ​​of the model proposed in this invention and various SOTA algorithms during 250 iterations; Figure 8(a) is a comparison of the changes in the loss values ​​of the training set of each algorithm, and Figure 8(b) is a comparison of the changes in the loss values ​​of the validation set of each algorithm. Detailed Implementation

[0067] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0068] like Figure 1 As shown, the transmission line icing detection method based on a wide-area sensing dynamic convolution model is implemented according to the following steps:

[0069] Step 1: Obtain images of icing on transmission lines in different environments through drone inspections, expand the obtained images of icing on transmission lines, and use the expanded images of icing on transmission lines as a real insulator dataset.

[0070] Step 2: Divide the transmission line icing dataset obtained in Step 1 into a training set and a validation set, with a ratio of 9:1 between the training set and the validation set.

[0071] Step 3: Use the Labellmg software to label the icing categories of the transmission lines in the training set in Step 2. Use VOC format for image labeling and save the icing image information of the transmission lines after labeling to obtain the transmission line icing database.

[0072] Step 4: Replace the ordinary convolution in the backbone network with the MDAM multi-dimensional attention module. MDAM extracts feature information from three different scales to obtain feature maps, and then uses convolution and upsampling operations to fuse the feature maps.

[0073] The Thin-head feature extraction network, which integrates a lightweight model, introduces hybrid convolutional RS Conv, including shuffling and DW Conv. RS Conv uses a learnable weight matrix to compute a weighted sum of the features of a node's neighboring nodes, concatenates the weighted sum with the node's own features, and then processes the concatenated features through a non-linear activation function to obtain the updated node representation. In the encoder head, normalized convolutions are replaced with RS Conv. Thin-head also introduces RS-bottleneck and VoV-RSCSPC modules, replacing the RepC3 module in the baseline model with the VoV-RSCSPC module.

[0074] By incorporating the idea of ​​Shape-IoU into normalized NWD, we obtain Shape-NWD;

[0075] The MDAL-DETR object detection network is composed of MDAM, a Thin-head feature extraction network that integrates a lightweight model, and the Shape-NWD loss function.

[0076] Step 5: Use the optimal weight network obtained after training in Step 4 to predict the validation set in Step 2, and obtain the corresponding transmission line icing detection results.

[0077] Step 6: The detection results obtained in Step 5 are evaluated using precision, recall, and mAP values ​​to efficiently determine the efficiency of each model in detecting icing on transmission lines and the practicality of the improved algorithm.

[0078] Step 1 is implemented in the following steps:

[0079] Step 1.1: Use drones to inspect and capture images of ice accretion on power transmission lines. Perform horizontal flipping, random extraction, rotation transformation, brightness enhancement, and contrast enhancement operations on the obtained images of ice accretion on power transmission lines to expand the dataset.

[0080] Step 1.2: Reintegrate the icing images of the transmission lines captured in Step 1.1 with the expanded icing images of the transmission lines into a new dataset as a sample library, such as... Figure 5 The image shows a partial dataset example of the present invention.

[0081] Step 3 is implemented in the following steps:

[0082] Step 3.1: Select the icing image of the transmission line to be marked, accurately identify and mark the icing area of ​​the transmission line, and obtain the marking box of the icing area of ​​the transmission line. According to the icing image of the transmission line in the marking box, it is divided into two types of labels: transmission line icing (ice transmission) and insulator icing (ice insulator). Finally, the marking boxes of transmission line icing and insulator icing area with category labels are obtained.

[0083] Step 3.2: Use the Labellmg annotation software to generate XML tag files corresponding to the icing images of transmission lines, and store the annotated sample data in Pascal VOC format. The sample data folder is named VOC devkit, which includes three folders: Annotations, JPEG Images, and Image Sets. The Annotations folder contains all the annotated XML tag files for icing of transmission lines and insulators. The XML tag file includes the image ID, image path, image name, and the image's pixel height and width. The image's pixel height and width are represented by the four coordinates of a rectangle, including x, y, and g. min ,ymin ,x max ,y max , where (x min ,y min (x) is the coordinate of the top-left vertex of the rectangle. max ,y max The coordinates of the bottom right vertex of the rectangle are given, and each labeled file corresponds to an original image in JPEG Images. The JPEG Images folder contains all the labeled original images. The Image Sets folder has a Main subfolder, where the train document contains the names of all training set images, and the val document contains the names of all validation set images.

[0084] like Figure 2 As shown, step 4 is implemented in the following steps:

[0085] Step 4.1: The MDAM (Multi-Dimensional Attention Module) aims to improve the network's image representation and feature extraction capabilities. Replacing ordinary convolutions in the backbone network with the MDAM module allows for better capture of complex structural information in images. By focusing on key information and ignoring irrelevant information, computational resources are saved. The MDAM module is detailed below:

[0086] The multidimensional attention mechanism comprises three branches, let F in To obtain the initial feature map F for the input dimension attention mechanism, features are first extracted using a convolutional layer with ReLU activation. r Then feature map F r After pooling, the data is fed into an existing multilayer perceptron (MLP) for further feature extraction. The extracted features are then multiplied by the channel statistical coefficients using a sigmoid activation function to obtain the feature map. Finally, the feature map Input feature map F with multi-dimensional attention accumulation mechanism in The output of each branch is obtained by adding elements one by one.

[0087] The feature map is calculated as shown in equation (1).

[0088] F r =f r (F in Equation (1)

[0089] In the formula: f r (·) indicates the first step of feature extraction, which includes two convolutional operations and one ReLU activation operation; F inThe input feature map representing the multi-dimensional attention mechanism; F r Let r represent the feature map obtained after the initial feature extraction of the r-th branch, where r = 1, 2, 3.

[0090] For feature map F r Further feature extraction is performed as shown in equations (2), (3), and (4).

[0091] f ca (F r )=σ(MLP(Aνgpool(F r ))+MLP(Maxpool(F r Formula (2)

[0092] coef = f ca (F r Equation (3)

[0093]

[0094] In the formula: f ca (·) represents the execution operation of the multi-dimensional attention mechanism; coef represents the multi-dimensional statistical coefficient; The r-th branch of the multi-dimensional attention mechanism initially extracts the features and then multiplies them by the activation function and coef, where r takes the values ​​1, 2, or 3; Avgpool(·) represents the average pooling operation; Maxpool(·) represents the max pooling operation.

[0095] Ultimately, each output of the multi-dimensional attention module It is represented as shown in equation (5).

[0096]

[0097] In the formula: The r-th branch represents the final output feature map of the multi-dimensional attention mechanism.

[0098] The multidimensional attention mechanism has three branches with the same structure but different kernel sizes: Conv1, Conv2, and Conv3, with kernel sizes of 3*3, 5*5, and 7*7, respectively. Therefore, the multidimensional attention mechanism extracts feature information from three different scales to obtain feature maps. Use F out Convolution and upsampling operations fuse feature maps As shown in equation (6).

[0099]

[0100] In the formula: f fuse(·) represents a fusion operation, [·] represents a join operation, F out The output feature map represents the multi-dimensional attention mechanism.

[0101] like Figure 3 As shown in step 4.2, the Thin-head feature extraction network that integrates the lightweight model introduces hybrid convolution RSConv, including shuffle and DW Conv operations. The Thin-head feature extraction network that integrates the lightweight model is as follows:

[0102] RS Conv uses a learnable weight matrix to compute a weighted sum of the features of a node's neighboring nodes, concatenates this weighted sum with the node's own features, and then processes the concatenated features through a non-linear activation function to obtain an updated node representation. In this paper, normalized convolutions are replaced with RS Conv in the encoder head. This multi-layer stacking enhances the node feature representation capability, solving the problem of excessive stacking of ordinary convolutions required for multi-branch network structures, filtering out redundant information, and thus improving the detection and recognition capabilities of multi-scale objects or features.

[0103] To reduce network layer depth and inference time, Thin-head introduces RS-bottleneck and VoV-RSCSPC modules on top of RS Conv. The RepC3 module introduces multiple repetitive convolutional layers and cross-layer connections, increasing the number of parameters and training complexity. Replacing the RepC3 module in the baseline model with the VoV-RSCSPC module effectively integrates contextual information at different scales and levels, enabling the model to better understand the target's surrounding environment and semantic features. The VoV-RSCSPC module reduces computational and network structure complexity while maintaining sufficient accuracy.

[0104] like Figure 4 As shown in step 4.3, this invention utilizes the Shape-IoU loss function to address the issue of slow convergence during sample training; simultaneously, it combines the Normalized Wasserstein Distance (NWD) strategy to address the baseline model's GIoU neglecting pixel-level similarity and the global structure and distribution of images; the Shape-NWD loss function calculation formula is as follows:

[0105] The Shape-IoU formula is shown below:

[0106]

[0107]

[0108] In the formula: w and h are the width and height of the bounding box, respectively. gt h gt This refers to the actual width and height of the frame; w c and h c represents the width and height of the minimum detection box covering the ground truth box and the anchor box, respectively. scale is a scaling factor that is related to the size of the target in the dataset. ww and hh represent the weight coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the ground truth box.

[0109] The corresponding bounding box shape loss is as follows:

[0110] L shape-IoU =1-IoU+distance shape +0.5×Ω shape Equation (13)

[0111] By incorporating the concept of Shape-IoU into normalized NWD, we obtain Shape-NWD, whose formula is as follows:

[0112]

[0113] In the formula: D is the Euclidean distance between the center point of the GT box and the center point of the anchor box, weight = 2, and C is a constant related to the dataset.

[0114] Step 4.4: Combine the MDAM from Step 4.1, the Thin-head feature extraction network that integrates the lightweight model from Step 4.2, and the Shape-NWD loss function from Step 4.3 to form the MDAL-DETR object detection network with better detection performance.

[0115] Table 1 shows the relevant hardware and software parameter configurations for the transmission line icing detection method based on the wide-area perception dynamic convolution model of this invention on the PyCharm platform.

[0116]

[0117]

[0118] Step 5 is implemented in the following steps:

[0119] Step 5.1: Using the hardware configuration shown in Table 1, input the training set images divided in Step 2 into the MDAL-DETR object detection network model obtained in Step 4.4 for training. In the file train.py, set stochastic gradient descent with momentum of 0.9, train for 250 epochs, set the input image pixel size to 640*640, freeze training for 50 epochs with batch_size 32, unfreeze training for 200 epochs with batch_size 4, num_workers 2, Adam optimizer, decay weight coefficient 5*10-4, initial learning rate 1*10-5. When testing the training set, set the IoU threshold to 0.5 for experimentation. During training, fine-tune the learning rate to 0.003 to obtain better robustness. Freeze training for the first 50 epochs for rapid loss reduction, and unfreeze training for the last 200 epochs for continuous fine-tuning of the network. In the total of 250 epochs, the loss of the training set gradually decreases after 200 epochs, and finally obtains the optimized wide-area perception dynamic convolutional model with the best weight data.

[0120] Step 5.2: Predicting the training results requires two files: mdal-detr.py and predict.py. First, you need to modify model_path and classes_path in mdal-detr.py. model_path points to the trained optimal weight file, which is located in the logs folder, and classes_path points to the txt file corresponding to the detected classes. After making the modifications, start the prediction.

[0121] Step 6 is implemented in the following steps:

[0122] The evaluation criteria for this algorithm mainly use precision, recall, mean average precision (mAP), frames per second (FPS), number of parameters, and number of floating point operations (FLOPs) for evaluation.

[0123] Precision reflects the proportion of cases that the model identifies as positive and that are actually positive, as shown in formula (17):

[0124]

[0125] In the formula, p represents precision; TP represents the detected positive examples; and FP represents the detected negative examples.

[0126] Recall reflects the proportion of positive examples judged by the model out of all positive test images, as shown in formula (18):

[0127]

[0128] In the formula, R represents recall; TP represents detected positive examples; and FN represents undetected positive examples.

[0129] The mean average precision (mAP) represents the average of the average precision of each category, as shown in formula (19):

[0130]

[0131] Frames Per Second (FPS) represents the number of frames of an image or video that a model can process per second, and is a standard for measuring the processing power and efficiency of a model. Generally speaking, the higher the FPS, the faster the model's inference speed and the faster it processes input data, as shown in formula (20):

[0132]

[0133] In the formula: T represents the detection time of a single image; FPS is the number of images detected per second.

[0134] The number of parameters refers to the total number of parameters that a network model needs to train, affecting the model's complexity and capability. Generally, a larger number of parameters indicates a stronger fit, but it also increases the computational burden and training difficulty.

[0135] Floating point operations (FLOPs) are a metric that measures the amount of computation required for a neural network to perform one round of forward propagation. They are used to evaluate the computational complexity and performance of a model. The higher the FLOPs, the more time the model consumes.

[0136] The conclusion is drawn based on the detection effect diagram in step 5.2 and the various evaluation indicators in step 6.1.

[0137] To effectively compare the advantages of our algorithm with other state-of-the-art (SOTA) algorithms, we used consistent training and testing methods to conduct comparative experiments on other SOTA algorithms and the model proposed in this invention. The specific results are shown in Table 2.

[0138] Table 2 shows the comparative experimental results of different algorithms for the transmission line icing detection method based on the wide-area sensing dynamic convolution model of this invention.

[0139]

[0140] like Figure 6 As shown, the target detection network based on the wide-area perception dynamic convolutional model achieves an mAP value of 97.85% for detecting both transmission line icing and insulator icing. This metric significantly surpasses various state-of-the-art algorithms, highlighting the superior accuracy of our method. Compared to other algorithms, the model in this invention exhibits a more stable growth curve in mAP without significant fluctuations, indicating higher detection stability and anti-interference capabilities. Furthermore, its precision and recall far exceed other attention mechanisms. Despite having more parameters and a frame rate of only 20.33%, the algorithm in this invention demonstrates superior accuracy and recognition performance, indicating its immense potential and broad application prospects in detecting transmission line icing.

[0141] Figure 7 shows the detection results of the target detection network based on the wide-area perception dynamic convolution model for icing of transmission lines and insulators in different environments. As can be seen from the figure, the MDAL-DETR target detection network can accurately detect icing of transmission lines and insulators in complex environments. Moreover, its detection accuracy and recall are higher and its robustness is stronger than other popular algorithms, making it more suitable for detecting icing of transmission lines in complex environments.

[0142] Figure 8 shows the evolution of training and validation losses for a target detection network based on a wide-area perception dynamic convolutional model and various state-of-the-art (SOTA) algorithms over 250 iterations. The SOTA algorithms exhibit wider convergence threshold intervals, reflecting poorer model performance and indicating instability during backpropagation in training and validation. However, the model presented in this study shows a reduction in training set loss from 0.55 to approximately 0.042, and a convergence set loss to approximately 0.07, demonstrating the smallest and most stable convergence value. This strongly suggests that the improved loss function effectively mitigates the adverse effects of positive-negative sample imbalance, enabling our model to more accurately backpropagate errors along the direction of minimum gradient, thereby correcting and updating weight values ​​and improving the algorithm's performance in detecting ice accretion.

Claims

1. A method for detecting icing on transmission lines based on a wide-area sensing dynamic convolution model, characterized in that, The specific steps are as follows: Step 1: Obtain images of icing on transmission lines in different environments through drone inspections, expand the obtained images of icing on transmission lines, and use the expanded images of icing on transmission lines as a real insulator dataset. Step 2: Divide the transmission line icing dataset obtained in Step 1 into a training set and a validation set, with a ratio of 9:1 between the training set and the validation set. Step 3: Use the Labellmg software to label the icing categories of the transmission lines in the training set in Step 2. Use VOC format for image labeling and save the icing image information of the transmission lines after labeling to obtain the transmission line icing database. Step 4: Replace the ordinary convolution in the backbone network with the MDAM multi-dimensional attention module. MDAM extracts feature information from three different scales to obtain feature maps, and uses convolution and upsampling operations to fuse the feature maps. The Thin-head feature extraction network, which integrates a lightweight model, introduces hybrid convolution RS Conv, including shuffling and DWConv. RS Conv uses a learnable weight matrix to compute a weighted sum of the features of a node's neighboring nodes, concatenates the weighted sum with the node's own features, and then processes the concatenated features through a non-linear activation function to obtain the updated node representation. In the encoder head, normalized convolutions are replaced with RS Conv. Thin-head also introduces RS-bottleneck and VoV-RSCSPC modules, replacing the RepC3 module in the baseline model with the VoV-RSCSPC module. By incorporating the idea of ​​Shape-IoU into normalized NWD, we obtain Shape-NWD; The MDAL-DETR object detection network is composed of MDAM, a Thin-head feature extraction network that integrates a lightweight model, and the Shape-NWD loss function. Step 5: Use the optimal weight network obtained after training in Step 4 to predict the validation set in Step 2, and obtain the corresponding transmission line icing detection results. Step 6: Evaluate the detection results obtained in Step 5 using precision, recall, and mAP value to determine the efficiency of each model in detecting icing on transmission lines and the practicality of the improved algorithm.

2. The method for detecting icing on transmission lines based on a wide-area sensing dynamic convolution model according to claim 1, characterized in that, Step 1 is implemented in the following steps: Step 1.1: Use drones to inspect and capture images of ice accretion on power transmission lines. Perform horizontal flipping, random cropping, rotation transformation, brightness enhancement, and contrast enhancement operations on the obtained images of ice accretion on power transmission lines to expand the dataset. Step 1.2: Reintegrate the icing images of the transmission lines captured in Step 1.1 with the expanded icing images of the transmission lines into a new dataset as a sample library.

3. The method for detecting icing on transmission lines based on a wide-area sensing dynamic convolution model according to claim 1, characterized in that, Step 3 is implemented in the following steps: Step 3.1: Select the icing image of the transmission line to be marked, accurately identify and mark the icing area of ​​the transmission line, and obtain the marking box of the icing area of ​​the transmission line. According to the icing image of the transmission line in the marking box, it is divided into two types of labels: transmission line icing and insulator icing. Finally, the marking boxes of transmission line icing and insulator icing area with category labels are obtained. Step 3.2: Use the Labellmg annotation software to generate XML tag files corresponding to the icing images of transmission lines, and store the annotated sample data in Pascal VOC format. The sample data folder is named VOC devkit, which includes three folders: Annotations, JPEG Images, and Image Sets. The Annotations folder contains all the annotated XML tag files for icing of transmission lines and insulators. The XML tag file includes the image ID, image path, image name, and the image's pixel height and width. The image's pixel height and width are represented by the four coordinates of a rectangle. , , , ,in( , ) is the coordinate of the top-left vertex of the rectangle. , The coordinates of the bottom right vertex of the rectangle are given, and each labeled file corresponds to an original image in JPEG Images. The JPEG Images folder contains all the labeled original images. The Image Sets folder has a Main subfolder, where the train document contains the names of all training set images, and the val document contains the names of all validation set images.

4. The method for detecting icing on transmission lines based on a wide-area sensing dynamic convolution model according to claim 1, characterized in that, Step 4 is implemented in the following steps: Step 4.1: Replace the ordinary convolutions in the backbone network with the MDAM multidimensional attention module, as follows: The multidimensional attention mechanism comprises three branches, let's assume... To obtain the feature map from the input feature map of the multi-dimensional attention mechanism, features are first extracted using a convolutional layer with ReLU activation function. Then the feature map After pooling, the data is fed into an existing multilayer perceptron (MLP) for further feature extraction. The extracted features are then multiplied by the channel statistical coefficients using a sigmoid activation function to obtain the feature map. Finally, the feature map Input feature maps with multi-dimensional attention mechanisms The output of each branch is obtained by adding elements one by one. ; The feature map is calculated as shown in Equation 1: Formula 1 In the formula: This indicates that the first step of feature extraction is being performed, which includes two convolutional operations and one ReLU activation operation. Input feature map representing a multi-dimensional attention mechanism; Indicates the first The feature map obtained after the initial feature extraction of the branches. ; For feature maps Further feature extraction is performed as shown in Equations 2, 3, and 4: Formula 2 Formula 3 Formula 4 In the formula: It is the execution operation of a multi-dimensional attention mechanism; Represents multidimensional statistical coefficients; Representing the The features initially extracted by the branch-based multi-dimensional attention mechanism are then processed by the activation function and The result after multiplication Choose 1, 2, 3; This represents the average pooling operation; This represents the max pooling operation; Ultimately, each output of the multi-dimensional attention module As shown in Equation 5: Formula 5 In the formula: The first representing a multi-dimensional attention mechanism The final output feature map of the branch; The multidimensional attention mechanism has three branches with the same structure but different kernel sizes, where Conv1, Conv2, and Conv3 have kernel sizes of [missing information]. , , Therefore, the multi-dimensional attention mechanism extracts feature information from three different scales to obtain feature maps. , , and use Convolution and upsampling operations fuse feature maps , , As shown in Equation 6: Formula 6 In the formula: Represents fusion operation, Represents a connection operation. Output feature map representing a multi-dimensional attention mechanism; Step 4.2: The Thin-head feature extraction network integrating the lightweight model introduces hybrid convolutional RS Conv, including shuffling operations and DW Conv; the Thin-head feature extraction network integrating the lightweight model is as follows: RS Conv uses a learnable weight matrix to compute a weighted sum of the features of a node's neighboring nodes, concatenates the weighted sum with the node's own features, and then processes the concatenated features through a non-linear activation function to obtain the updated node representation. Specifically, in the encoder head, the normalized convolution is replaced with RS Conv. Based on RS Conv, Thin-head also introduces RS-bottleneck and VoV-RSCSPC modules; and the RepC3 module in the baseline model is replaced with the VoV-RSCSPC module. Step 4.3: The Shape-IoU loss function is used to address the issue of slow convergence during sample training. Simultaneously, a normalization strategy is employed to address the baseline model's GIoU neglecting pixel-level similarity and the global structure and distribution of images. The Shape-NWD loss function calculation formula is as follows: The Shape-IoU formula is shown below: Formula 7 Formula 8 Formula 9 Formula 10 Formula 11 Formula 12 In the formula: , This refers to the width and height of the bounding box. , This refers to the actual width and height of the frame; and These represent the width and height of the ground truth (GT) box and the anchor box covered by the minimum detection bounding box, respectively. scale It is a scaling factor, related to the size of the target in the dataset. ww and hh These represent the weighting coefficients in the horizontal and vertical directions, respectively, and their values ​​are related to the shape of the ground truth box. The corresponding bounding box shape loss is as follows: Formula 13 By incorporating the concept of Shape-IoU into normalized NWD, we obtain Shape-NWD, whose formula is as follows: Formula 14 Formula 15 Formula 16 In the formula: D is the Euclidean distance between the center point of the GT box and the center point of the anchor frame. =2, C is a constant related to the dataset; Step 4.4: Combine the MDAM from Step 4.1, the Thin-head feature extraction network that integrates the lightweight model from Step 4.2, and the Shape-NWD loss function from Step 4.3 to form the MDAL-DETR object detection network.

5. The method for detecting icing on transmission lines based on a wide-area sensing dynamic convolution model according to claim 4, characterized in that, Step 5 is implemented in the following steps: Step 5.1: Input the training set images divided in Step 2 into the MDAL-DETR object detection network obtained in Step 4.4 for training, and finally obtain the optimized wide-area perception dynamic convolution model to obtain the best weight data. Step 5.2: Predicting the training results requires two files: mdal-detr.py and predict.py. First, you need to modify model_path and classes_path in mdal-detr.py. model_path points to the trained optimal weight file, which is located in the logs folder, and classes_path points to the txt file corresponding to the detected classes. After making the modifications, start the prediction.

6. The method for detecting icing on transmission lines based on a wide-area sensing dynamic convolution model according to claim 5, characterized in that, The specific method for step 5.1, which involves inputting the training set images divided in step 2 into the MDAL-DETR object detection network obtained in step 4.4 for training, is as follows: In the `train.py` file, we set up stochastic gradient descent with a momentum of 0.9, trained for 250 epochs, with an input image pixel size of 640*640. We froze the training for 50 epochs with a batch size of 32, then unfroze it for 200 epochs with a batch size of 4. We used `num_workers2`, the Adam optimizer, a decaying weight coefficient of 5*10⁻⁴, and an initial learning rate of 1*10⁻⁵. During training set testing, we set the IoU threshold to 0.

5. During training, we fine-tuned the learning rate to 0.003 for better robustness. The first 50 epochs of frozen training resulted in rapid loss reduction, while the last 200 epochs of unfrozen training allowed for continuous fine-tuning of the network. After 200 epochs, the loss on the training set gradually decreased, ultimately yielding an optimized wide-area-aware dynamic convolutional model with the best weight data.

Images