Power transmission line icing detection method based on multi-feature collaborative enhanced self-adjustment

Abstract

The invention discloses a power transmission line icing detection method based on multi-feature collaborative enhancement self-adjustment, relates to the technical field of image processing, and aims to solve the problem that in the prior art, when an unmanned aerial vehicle is used for image shooting, a target image is possibly shielded due to the view angle problem. In order to solve the problem that due to the fact that the power transmission line icing detection accuracy is low, the application of the method in a YOLOv8 model integrates SPD-Conv convolution, FocalNets and SEAM technologies, self-adaptive selective kernel size and receptive field adjustment of different targets is achieved, and the difference of the different targets in required background information is effectively processed. The method can effectively deal with the condition of target shielding, and obviously reduces the conditions of missing detection and false detection in the detection process. Therefore, the accuracy of power transmission line icing detection is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method for detecting icing on power transmission lines based on multi-feature collaborative enhanced self-adjustment. Background Art

[0002] With the continuous growth of electricity demand and the expansion of power grids, overhead transmission lines are playing an increasingly important role in power systems. They shoulder the important task of efficiently transmitting electricity from power plants to consumers, making them an indispensable component of the power system. However, in severe weather conditions, especially in cold regions, icing is a frequent occurrence, posing a serious challenge to the safe operation of transmission lines.

[0003] However, traditional inspection methods rely heavily on manual labor, which not only poses safety risks but also makes it difficult to guarantee accuracy. While recent technological advances have gradually replaced manual inspections with intelligent methods like drones and inspection robots, the perspective of drones can obstruct the target image, leading to low accuracy in ice detection on transmission lines. Summary of the Invention

[0004] The purpose of this invention is to address the problem in the prior art of low accuracy in power transmission line icing detection due to the obstruction of target images caused by viewing angle issues when using drones for image capture. This invention provides a transmission line icing detection method based on multi-feature collaborative enhancement and self-adjustment.

[0005] The technical solution adopted by the present invention to solve the above technical problems is:

[0006] A method for detecting ice coating on a transmission line based on multi-feature collaborative enhanced self-regulation includes the following steps:

[0007] Obtaining a transmission line image to be identified, and inputting the transmission line image to be identified into a trained neural network to obtain an output detection result;

[0008] The neural network is a YOLOv8-SFS model, which is improved based on YOLOv8. The improvement steps are as follows:

[0009] Step 1: Replace the traditional convolution in YOLOv8 with SPD-Conv convolution;

[0010] Step 2: Replace the SPPF module in the Backbone network with the FocalNets module;

[0011] Step 3: Introduce the SEAM attention mechanism in the head end and combine the SEAM attention mechanism with the detection head to build the YOLOv8-SFS model.

[0012] Furthermore, the SPD-Conv convolution performs the following steps:

[0013] First, the input feature map is transformed from space to depth, and then the transformed feature map is convolved to obtain the feature map.

[0014] Furthermore, the FocalNets module performs the following steps:

[0015] Step 1: The input feature map undergoes linear transformation to generate the initial feature representation;

[0016] Step 2: The initial feature representation is divided into two paths after feature aggregation, one path undergoes hierarchical context processing, and the other path undergoes gated aggregation processing;

[0017] Step 3: The features after hierarchical context processing and the features after gated aggregation processing are aggregated into the feature interaction module to generate the final output features.

[0018] Furthermore, the SEAM attention mechanism performs the following steps:

[0019] The input features are extracted through multiple CSMM modules, and then the output features of each CSMM module are additively aggregated and average pooled. The pooled features are then channel-expanded through a fully connected layer, and finally the expanded features are fused with the input features.

[0020] Furthermore, the space-to-depth conversion is expressed as:

[0021] SPD(X)=concat(X 0,0 ,X 0,1 ,X 1,0 ,X 1,1 )

[0022] Among them, X i,j Represents the sub-block of feature map X at position (i, j).

[0023] Furthermore, the converted feature map is further extracted using non-step convolution to represent the features:

[0024] Conv(X)=Conv2d(X,kernel,stride=1,padding=autopad(kernel)),

[0025] Among them, autopad(kernel) represents the automatically calculated padding value.

[0026] Furthermore, the output of the SEAM attention mechanism is expressed as:

[0027]

[0028] Among them, x i and x j Indicates the xth i and xth j input variables, C(x i ) represents the normalization function, f(x i ,x j ) represents the weight between the two, g(x j ) represents the value of x j Perform feature extraction.

[0029] Furthermore, the trained neural network is obtained by the following steps:

[0030] Step 1: Acquire the transmission line image;

[0031] Step 2: Mark the snow granular area on the transmission line image;

[0032] Step 3: Use the transmission line image and the labeled image to train the neural network to obtain a trained neural network.

[0033] Furthermore, the labeling is performed through labelimg.

[0034] Furthermore, the specific steps of marking the snow granular area on the transmission line image are as follows:

[0035] For insulators that are completely covered by snow, the entire insulator is marked and its label is named "insulator-icing ring". For transmission lines where granular snow occurs, the specific granular snow area on the transmission line is marked and its label is named "snowline".

[0036] The beneficial effects of the present invention are:

[0037] This application, in the YOLOv8 model, integrates SPD-Conv convolution, FocalNets, and SEAM technologies to achieve adaptive selective kernel size and receptive field adjustment for different targets, effectively addressing differences in required background information for different targets. This approach effectively handles target occlusion and significantly reduces missed and false detections during the detection process, thereby improving the accuracy of transmission line ice detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a schematic diagram of the application process;

[0039] Figure 2 This is a schematic diagram of the network structure of this application;

[0040] Figure 3 This is a schematic diagram of the SPD-Conv convolution structure of this application;

[0041] Figure 4 This is a schematic diagram of the FocalNets module structure for this application;

[0042] Figure 5 This is a schematic diagram of the SEAM structure for this application;

[0043] Figure 6 This is a schematic diagram of ice detection for this application;

[0044] Figure 7 This is a comparison chart of the detection results of the YOLOv8 model and the YOLOv8-SFS model. DETAILED DESCRIPTION

[0045] It should be noted that, unless there is any conflict, the various embodiments disclosed in this application can be combined with each other.

[0046] Specific embodiment 1: This embodiment describes a method for detecting ice coating on a transmission line based on multi-feature collaborative enhanced self-regulation, including:

[0047] Step 1: Obtain an ice cover image taken by a drone;

[0048] Step 2: Use the labelimg tool to preprocess the ice-covered image and output a dataset in YOLO format;

[0049] Step 3: Combine the SPD-Conv convolution, FocalNets module, and SEAM attention mechanism with the original YOLOv8 model to build the YOLOv8-SFS model;

[0050] Step 4: Input the dataset into the YOLOv8-SFS neural network model for training;

[0051] Step 5: Use the trained YOLOv8-SFS model to detect the ice-covered image to be identified;

[0052] The specific steps in step 1 are: Use a drone to capture aerial images of the transmission lines outdoors to capture images of ice covering the lines. Leveraging the drone's high flexibility, capture images of ice covering the transmission lines from multiple angles and distances, ensuring high resolution and diversity. These images are then screened and used to create training and test sets for the improved YOLOv8 model.

[0053] The specific steps in step 2 are as follows: Use the Labelimg annotation tool in the YOLO format to annotate the manually screened images. For insulators completely encased in snow, we annotated the entire insulator and named it "insulator-icing ring." For transmission lines experiencing severe snow granulation, we annotated specific areas of snow granulation on the transmission lines and labeled them "snowline." The label names and location information are saved in a .txt file. After the annotation process is complete, the dataset is randomly divided into a training set and a validation set in an 8:2 ratio.

[0054] The specific operation steps in step 3 are: replacing the traditional convolution in YOLOv8 with SPD-Conv convolution, replacing the SPPF module in the Backbone network with the FocalNets module, and combining the SEAM attention mechanism with the detection head in the Head end to build the YOLOv8-SFS model.

[0055] Replace traditional convolution in the YOLOv8 model with SPD-Conv convolution: In the YOLOv8 model, SPD-Conv (spatial depth conversion convolution) is used to solve the problem of fine-grained information loss caused by traditional strided convolution and pooling layers, especially when dealing with low-resolution images and small objects. SPD-Conv enhances the model's detection ability for low-resolution images and small objects through a three-step process. First, the input feature map is transformed from spatial dimensions (width and height) to depth dimensions (number of channels) through spatial-to-depth (SPD) conversion. The process is: SPD(X) = concat(X 0,0 ,X 0,1 ,X 1,0 ,X 1,1 ) Among them, X i,j Represents the sub-block of feature map X at position (i, j). Next, a non-strided convolution (i.e., a convolution with a kernel size of 1 or 3) is applied to the converted feature map to further extract features. The process is: Conv(X) = Conv2d(X, kernel, stride = 1, padding = autopad(kernel)), where autopad(kernel) is an automatically calculated padding value that ensures that the output feature map has the same size as the input feature map. Finally, feature fusion technology is used to combine the feature map processed by SPD-Conv with other feature maps to enhance the model's feature expression capabilities. This entire process makes YOLOv8 more effective at detecting small objects in complex environments.

[0056] Replacing the SPPF module with FocalNets on the Backbon side: In the YOLOv8 model, FocalNets are used to further improve the model's ability to identify key features in images, particularly when dealing with images of ice-covered power lines captured in natural environments, which are often affected by lighting, occlusion, and complex backgrounds, making image recognition difficult. FocalNets introduces a novel attention mechanism through Focal Modulation Networks. This mechanism automatically selects and integrates useful information from different levels of granularity based on input label interactions, potentially representing features at different levels. FocalNets' Focal Modulation uses an early aggregation procedure to generate refined feature representations. The core idea is to extract a modulator at each query position and implement the interaction between the query and the modulator in a much simpler manner. This approach allows the model to focus more attention on specific regions, thereby improving the ability to identify important features. First, the contextual features at each position i are aggregated using M2. This step involves aggregating information from different regions in the image, allowing the model to extract useful contextual information from labels at different granularity levels. The query then interacts with the aggregated features based on T2 to form a feature representation y. This step allows the model to adjust its behavior based on the aggregated feature information when executing the query, thereby better understanding and identifying key areas in the image.

[0057] Introducing the SEAM attention mechanism on the Head side and combining it with the detection head: The application of SEAM (Occlusion-Aware Attention Network) is intended to improve the model's detection performance for small and occluded targets, especially in ice detection in complex environments. This module combines spatial and edge attention mechanisms, which can effectively enhance the feature response of unoccluded areas while suppressing the features of occluded areas, thereby improving the model's ability to detect ice on transmission lines. Specifically, first, the input feature map is processed by spatial and edge attention to extract more detailed feature information. Then, after feature fusion, the model can better focus on key areas in the image. The process is as follows: in, g(x j ) gives the input signal x at each position j , and aggregated to position i using a given similarity weight. This is especially true when the target is small or occluded. Ultimately, the output feature map of the Detect_SEAM module will be passed to the detection head for target detection, thereby improving the overall detection accuracy and robustness.

[0058] The specific operation steps of step 4 are as follows: input the labeled ice-covered image dataset into the YOLOv8-CBS model for training to obtain the YOLOv8-CBS.pt weight file.

[0059] The specific operation steps of step 5 are as follows: Use the trained YOLOv8-SFS.pt weight file to detect and locate the ice-covered image to be identified.

[0060] This application is implemented through the following technical solutions: Figure 1 As shown, a method for detecting ice coating on a transmission line based on an improved YOLOv8 neural network includes the following steps:

[0061] Step 1: Obtain an aerial image of ice cover from a drone.

[0062] The process involves capturing images of ice-covered transmission lines outdoors using drones. The drones are deployed above the transmission lines, using their high-resolution cameras to capture a comprehensive view of the lines. The drones' flexibility allows them to capture images of ice-covered transmission lines from various angles and distances. Because these images may be affected by lighting, occlusion, and background complexity in natural environments, they require preliminary manual screening to prepare the dataset.

[0063] Step 2: Use the labelimg tool to preprocess the ice-covered image and output a dataset in YOLO format.

[0064] The data labeling method of the labelimg tool in this embodiment is as follows Figure 3 Specifically, in this experiment, we used the Labelimg annotation tool to annotate the ice-covered images obtained in step 1: insulators covered in granular snow and ice-covered transmission lines were labeled with class labels 0 and 1, respectively, and named "insulator-icing ring" and "snowline." We converted these annotated defect images into a YOLO dataset and randomly partitioned the dataset into an 80% training set and a 20% validation set for subsequent model training.

[0065] Step 3: Combine the SPD-Conv convolution, FocalNets module, and SEAM attention mechanism with the original YOLOv8 model to build the YOLOv8-SFS model.

[0066] The process is to replace the traditional convolutional layer in the model with the SPD-Conv convolutional layer. Secondly, in the backbone network, we replace the original SPPF module with the FocalNets module. Finally, in the detection head, we integrate the SEAM attention mechanism. The improved YOLOv8 network structure is as follows: Figure 2 As shown in Figure 1, it includes the backbone network, neck network and head network.

[0067] 1) Replace the traditional convolution in the model with SPD-Conv convolution. For the SPD-Conv convolution structure, see Figure 3 First, the input feature map X is of size S×S×C1, where S is the spatial dimension and C1 is the number of channels. The input feature map is converted to "space-to-depth". This process decomposes the information in the spatial dimension and rearranges it into the channel dimension. The result is that the original S×S×C1 feature map is converted into four smaller S / 2×S / 2×C1 feature blocks. These feature blocks are merged together to form a new feature map whose number of channels increases to 4C1 but the spatial dimension is reduced to S / 2×S / 2. This operation retains more local information by increasing the number of channels. The merged feature map is then convolved. A standard convolution with a stride of 1 is used to generate the output feature maps, which are restored to the size of S / 2×S / 2×C1. The entire process effectively increases the receptive field while maintaining the spatial resolution, thereby improving the richness and effectiveness of feature extraction.

[0068] 2) The SPPF module in Backbone is replaced by the FocalNets module, which participates in Figure 4 First, the input data undergoes a lightweight linear transformation to generate an initial feature representation. Part of the features enter the feature aggregation module for processing, which is further processed separately. Part of the features undergo hierarchical contextualization processing, which processes the features layer by layer through multiple layers (l=1,2,3 in the figure). Each layer extracts features by focusing on different contextual information, gradually enhancing the richness of the features. The other part undergoes gated aggregation processing. In this process, the features on each level are processed by gated units. These units control the flow of information. The gated units perform selective aggregation of features through operations such as weighted summation to generate more representative feature representations. The output of each layer is accumulated with the output of the upper layer to form a gradually enriched feature representation Z. outFinally, it converges to the feature interaction module, where the initial feature q interacts with the modulator feature after linear transformation, and the features adjusted by the interaction module are passed to the output layer to generate the final output.

[0069] 3) Introduce the SEAM attention mechanism at the head end and combine it with the detection head. See the SEAM structure for details. Figure 5 First, the SEAM input data is processed by multiple CSMM (Convolutional Spatial Multi-head Module) modules. Each module uses a different patch size (such as 6, 7, and 8) for feature extraction to capture spatial information at different scales. In the CSMM module, the input is first divided into small blocks (patches) and embedded. The embedded features are activated with GELU and batch normalized to stabilize the feature distribution. Deep convolution operations are used to extract inter-channel features. At the same time, skip connections are added after the deep convolution to retain low-level feature information. Finally, point-by-point convolution is performed to integrate features. The output features of each CSMM module are additively aggregated to merge information at different scales. The aggregated features are average pooled to reduce the feature dimension and extract important information. The pooled features are then channel-expanded through a fully connected layer to increase the feature expression capability. Finally, the expanded features are operated on with the input features to achieve interaction and update of self-enhanced features.

[0070] Step 4: Input the dataset into the YOLOv8-SFS neural network model for training.

[0071] The labeled ice-covered image dataset in YOLO format, divided into an 8:2 ratio, is input into the YOLOv8-SFS model for training to obtain the YOLOv8-SFS.pt weight file.

[0072] Step 5: Use the trained YOLOv8-SFS model to detect the ice-covered image to be identified.

[0073] Use the trained YOLOv8-SFS.pt weight file to detect and locate the new ice-covered image. The detection results are shown in Figure 6 .

[0074] This example uses the deep learning framework PyTorch 2.1 for experiments. The hardware configuration includes an RTX 4060 GPU and a 12th-generation Intel(R) Core(TM) i7-12700KF 3.60GHz. In this experiment, the learning rate is 0.001, the final learning rate is 0.1, the batch size is 16, the number of iterations is 200, the mosaic data augmentation factor is 1.0, and the image size is 640.

[0075] The comparison of the model detection results of this example with other models is shown in the following table:

[0076]

[0077] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solutions of the present invention and cannot be used to limit the scope of protection. Any minor changes made based on the claims and description of the present invention shall still fall within the scope of protection of the present invention.

Claims

1. A transmission line icing detection method based on multi-feature collaborative enhanced self-adjustment, characterized by The following steps are involved: Obtaining a transmission line image to be identified, and inputting the transmission line image to be identified into a trained neural network to obtain an output detection result; The neural network is a YOLOv8-SFS model, which is improved based on YOLOv8. The improvement steps are as follows: Step 1: Replace the traditional convolution in YOLOv8 with SPD-Conv convolution; Step 2: Replace the SPPF module in the Backbone network with the FocalNets module; Step 3: Introduce the SEAM attention mechanism in the head end and combine the SEAM attention mechanism with the detection head to build the YOLOv8-SFS model.

2. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 1 is characterized in that The SPD-Conv convolution performs the following steps: First, the input feature map is transformed from space to depth, and then the transformed feature map is convolved to obtain the feature map.

3. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 2 is characterized in that The FocalNets module performs the following steps: Step 1: The input feature map undergoes linear transformation to generate the initial feature representation; Step 2: The initial feature representation is divided into two paths after feature aggregation, one path undergoes hierarchical context processing, and the other path undergoes gated aggregation processing; Step 3: The features after hierarchical context processing and the features after gated aggregation processing are aggregated into the feature interaction module to generate the final output features.

4. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 3 is characterized in that The SEAM attention mechanism performs the following steps: The input features are extracted through multiple CSMM modules, and then the output features of each CSMM module are additively aggregated and average pooled. The pooled features are then channel-expanded through a fully connected layer, and finally the expanded features are fused with the input features.

5. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 4 is characterized in that The space-to-depth conversion is expressed as: SPD(X)=concat(X 0,0 ,X 0,1 ,X 1,0 ,X 1,1 ) Among them, X i,j Represents the sub-block of feature map X at position (i, j).

6. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 5 is characterized in that The converted feature map is further extracted using non-step convolution to represent the features: Conv(X)=Conv2d(X,kernel,stride=1,padding=autopad(kernel)), Among them, autopad(kernel) represents the automatically calculated padding value.

7. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 6 is characterized in that The output of the SEAM attention mechanism is expressed as: Among them, x i and x j Indicates the xth i and xth j input variables, C(x i ) represents the normalization function, f(x i ,x j ) represents the weight between the two, g(x j ) represents the value of x j Perform feature extraction.

8. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 1 is characterized in that The trained neural network is obtained by the following steps: Step 1: Acquire the transmission line image; Step 2: Mark the snow granular area on the transmission line image; Step 3: Use the transmission line image and the labeled image to train the neural network to obtain a trained neural network.

9. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 8 is characterized in that The labeling is performed by labelimg.

10. The method for detecting ice coating on power transmission lines based on multi-feature collaborative enhanced self-regulation according to claim 9 is characterized in that The specific steps of marking the granular snow area on the transmission line image are as follows: For insulators that are completely covered by snow, mark the entire insulator and name it "insulator-icingring". For transmission lines where granular snow occurs, mark the specific granular snow area on the transmission line and name it "snowline".

Images