Bird excrement flashover insulation sleeve damage early warning method for power transmission conductor
By constructing the ISDI-Net network and combining multi-level feature fusion and data augmentation techniques, the feature fusion and multi-scale problems in the detection of bird droppings flashover insulation sleeve damage in power transmission lines were solved, achieving higher accuracy and robust damage detection.
Patent Information
- Application Number
- CN202510466674.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-15
- Publication Date
- 2025-12-05
- Estimated Expiration
- 2045-04-15
AI Technical Summary
Existing methods for detecting damage to the insulation sleeves of transmission lines that prevent bird droppings from flashing out are difficult to accurately identify under long-distance monitoring and complex backgrounds. They suffer from insufficient feature fusion, inadequate multi-scale feature capture capabilities, and poor adaptability to complex backgrounds, resulting in insufficient detection accuracy and robustness.
The ISDI-Net network is constructed, employing a multi-level adaptive feature fusion module, a Ghost pointwise convolutional multi-scale feature fusion module, and a grouped convolutional and coordinated attention residual expansion module. Combined with data augmentation techniques, it enhances feature extraction and fusion capabilities, thereby improving the detection of features in complex environments and small targets.
It significantly improves the accuracy and robustness of insulation sleeve damage detection, reduces the interference of complex environments on detection, enhances the ability to detect fine-grained damage such as micro-cracks and wear, and improves detection accuracy and model adaptability.
Smart Images
Figure CN120298803B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power transmission conductor technology, specifically to a method for early warning of damage to the insulation sleeve of power transmission conductors caused by bird droppings flashover. Background Technology
[0002] Bird droppings-proof flashover insulation sleeves on transmission lines are crucial protective devices for power lines, widely used in power transmission and distribution systems to reduce the impact of bird droppings flashover on power grid safety. However, during long-term operation, the insulation sleeves may break down due to environmental corrosion, mechanical damage, aging, and cracking, leading to a decline in insulation performance and increasing the risk of power equipment failure. Therefore, in the fields of power grid operation and maintenance and fault early warning, how to efficiently and accurately detect the damage to bird droppings-proof flashover insulation sleeves on transmission lines has become key to ensuring the safe and stable operation of the power system. However, due to the subtlety of insulation sleeve damage and its complex operating environment, traditional detection methods often struggle to achieve accurate identification under long-distance monitoring or complex background interference, necessitating new technologies to improve the accuracy and robustness of damage detection.
[0003] In practical applications, insulation sleeve damage detection faces multiple challenges: Firstly, the damaged area is usually small with blurred edges and may blend closely with background structures such as wires and supports, making it difficult for traditional image processing methods (such as edge detection and color segmentation) to effectively separate the damage features. Secondly, the damage features may vary significantly under different environments due to changes in natural lighting, occlusion, and weather factors (such as rain and fog), further increasing the detection difficulty. Furthermore, under long-distance monitoring conditions, the image resolution is low, and the damaged area occupies a small proportion, making it easy for traditional detection methods to overlook these subtle structures, leading to a decrease in detection accuracy.
[0004] Traditional methods for detecting insulation sleeve damage, such as manual inspection or image-processing-based automated detection techniques, often suffer from low detection efficiency, poor adaptability, and high false detection rates, making them unsuitable for large-scale power grid inspections. Furthermore, these methods typically cannot accommodate multiple damage modes and lack robustness to complex backgrounds. In recent years, with the rapid development of deep learning technology, especially the successful application of models such as convolutional neural networks (CNNs) in object detection, deep learning-based damage detection methods have demonstrated significant advantages. Deep learning can automatically extract multi-level features through end-to-end learning, effectively improving the ability to identify damaged areas and providing a novel solution for early warning of bird droppings flashover protection on power transmission conductors.
[0005] Despite significant advancements in target detection using deep learning technology, detecting damage to insulation sheaths in power transmission lines caused by bird droppings flashover still faces numerous technical bottlenecks. For example, the limited image quality from long-distance photography can make it difficult for deep learning models to accurately distinguish damaged areas from the background during feature extraction, leading to false positives and false negatives. Furthermore, damaged areas are often small and varied in shape, and existing detection methods still have limitations in multi-scale feature learning, particularly in low-resolution images where crucial information is easily lost. Therefore, designing deep learning models that are more adaptable to complex environments and small target features for power grid inspection scenarios, and improving their accuracy in detecting insulation sheath damage, is a crucial research direction.
[0006] Based on the above, existing deep learning-based methods for detecting bird droppings flashover damage to transmission line insulation sleeves have the following problems:
[0007] 1. Insufficient Feature Fusion Leads to Inadequate Identification of Damaged Insulation Sheath Areas: Current detection methods often suffer from insufficient feature fusion during the feature extraction process for damaged insulation sheath areas. Due to the complex morphology of damaged areas, traditional methods struggle to fully fuse shallow and deep features at different scales, resulting in poor model adaptability to different types of damage. Particularly in the extraction of detailed features such as contamination, aging, or micro-cracks, existing methods have limited feature representation capabilities, easily leading to unclear identification of damaged areas, false detections, or missed detections.
[0008] 2. Insufficient multi-scale feature capture capability affects the detection of damaged area boundaries: The size of the damaged area of the insulating sleeve is usually small and intertwined with the surrounding background, increasing the difficulty of detection. Traditional models have shortcomings in multi-scale feature extraction, especially for micro-cracks or irregular damaged areas, which are easily overlooked due to feature blurring. Existing methods are prone to losing key information due to downsampling operations when processing subtle damaged features, affecting the detection accuracy of the model.
[0009] 3. Insufficient integration of feature extraction and attention mechanisms affects the model's adaptability to complex backgrounds: In complex power grid environments, the background information of insulation sleeves is complex, with interference factors such as intersecting wires, changing lighting, and dirt obstruction, which can easily affect the accuracy of damage detection. Existing methods lack specificity in feature extraction for damaged areas, making the model susceptible to background noise interference, resulting in blurred boundaries of damaged areas. While grouped convolution can improve computational efficiency, if it is not effectively combined with an attention mechanism, it may reduce the model's responsiveness to target areas. Furthermore, an improperly designed residual expansion mechanism may lead to information redundancy, affecting the model's generalization ability and causing unstable detection performance in different environments.
[0010] 4. Existing methods have limited ability to detect multi-scale damage, making it difficult to balance detection accuracy and computational efficiency: Since damage to insulation sleeves can manifest in different forms (such as cracks, peeling, wear, etc.), existing methods often rely on fixed convolutional structures or simple feature fusion strategies when processing multi-scale targets, making it difficult to accurately adapt to damage areas of different scales. Furthermore, in large-scale inspection tasks, models need to ensure both detection accuracy and computational efficiency, but traditional detection methods often struggle to balance accuracy and real-time performance, affecting their application value in actual power grid inspections.
[0011] Therefore, a method for early warning of damage to the insulation sleeve of power transmission lines due to bird droppings flashover was invented. Summary of the Invention
[0012] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:
[0013] A method for early warning of bird droppings flashover and insulation damage on power transmission lines includes the following specific steps:
[0014] S1: Construct the original dataset for detecting damage to the insulation sleeve of transmission lines to prevent bird droppings flashover, or use the existing dataset for detecting damage to the insulation sleeve. Collect image data of insulation sleeve damage under different environments, and manually annotate the collected image data. Use the annotation tool to select the location of the damaged area and label it.
[0015] S2: Preprocess the original dataset for detecting bird droppings flashover damage to transmission line insulation sleeves to obtain a preprocessed dataset of insulation sleeve damage; divide the dataset of insulation sleeve damage into training set, validation set and test set.
[0016] S3: Constructing the ISDI-Net network for detecting damage to insulation sleeves of power transmission lines to prevent bird droppings flashover;
[0017] S4: Based on the constructed ISDI-Net model of the transmission line anti-bird droppings flashover insulation sleeve damage early warning network, train and update the parameters of each layer, initialize all neural network parameters, and set the model-related hyperparameters, including training rounds, batch size, weight decay coefficient, learning rate and total number of iterations.
[0018] S5: After the early warning model for bird droppings flashover and insulation sleeve damage of transmission lines is trained, the trained ISDI-Net model is used to detect insulation sleeve damage in transmission line images. The final output of the detection results includes the bounding box coordinates of the damaged area, the damage type label, and the confidence score, providing accurate early warning information for power equipment operation and maintenance.
[0019] As a preferred embodiment of the method for early warning of bird droppings flashover and insulation sleeve damage in power transmission lines according to the present invention, the specific steps of S2 are as follows:
[0020] S21: The insulation sleeve images in the original dataset for insulation sleeve damage detection are cropped to a uniform size, and then data augmentation processing is performed on all insulation sleeve images to obtain the data-augmented insulation sleeve damage dataset. The purpose of data augmentation processing on the insulation sleeve images is to increase the size of the dataset, enhance the robustness of the damage detection model, and reduce the model's sensitivity to environmental changes.
[0021] S22: Divide the data augmented dataset of insulation sleeve damage into training, validation and test sets for subsequent training of the deep learning-based insulation sleeve damage detection model.
[0022] As a preferred embodiment of the early warning method for preventing bird droppings flashover and insulation sleeve damage in power transmission lines according to the present invention, the specific steps of S3 are as follows:
[0023] S31: Input the images in the data-enhanced insulation sleeve damage dataset into the multi-level adaptive feature fusion module to obtain the damage enhancement feature map of the insulation sleeve damage area;
[0024] S32: Input F17 into the Ghost pointwise convolutional multi-scale feature fusion module;
[0025] S33: Input T9 into the grouped convolution and coordinated attention residual extension module.
[0026] As a preferred embodiment of the early warning method for preventing bird droppings flashover and insulation sleeve damage in transmission lines according to the present invention, the construction and operation process of the multi-level adaptive feature fusion module is as follows:
[0027] S311: The image of the bird droppings flashover protection insulation sleeve of the transmission line is processed by CBS operation to obtain the damage enhancement feature map F1. Then, F1 is input into the first adaptive feature fusion module for feature extraction to generate the damage enhancement feature map F2. Further local features are extracted through Conv3×3 convolution operation to obtain the damage enhancement feature map F3. At the same time, Conv3×3 convolution operation is performed on F1 to obtain the damage enhancement feature map F5.
[0028] S312: Using the residual connection method, F3 and F5 are added element by element to form the damaged enhancement feature map F6. Then, F6 is input again into the first adaptive feature fusion module for feature extraction to obtain the damaged enhancement feature map F7. F7 is subjected to Conv3×3 convolution operation to obtain the damaged enhancement feature map F8. Finally, F8 is subjected to Conv1×1 convolution operation to obtain the damaged enhancement feature map F9.
[0029] S313: Normalize F5 using the Sigmoid activation function to obtain attention weights. Weight F5 with these attention weights to obtain the damage enhancement feature map F10. Further extract global information from F10 using Conv5×5 convolution to obtain the damage enhancement feature map F11.
[0030] S314: Using the residual connection method, F11 and F9 are added element by element to form the damaged enhancement feature map F12. F12 is input into the second adaptive feature fusion module for deep adaptive feature extraction to generate the damaged enhancement feature map F13. F13 is then convolved with 5×5 to enhance the feature expression ability, resulting in the damaged enhancement feature map F14.
[0031] S315: Input F11 into the first adaptive feature fusion module for feature extraction to obtain the damaged enhanced feature map F15. Perform Conv1×1 convolution operation on F15 to generate the damaged enhanced feature map F16. Finally, fuse F14 and F16 by element-wise addition to obtain the final output damaged enhanced feature map F17.
[0032] In a preferred embodiment of the method for early warning of bird droppings flashover and insulation sleeve damage in transmission lines according to the present invention, the construction and operation process of the first adaptive feature fusion module is as follows:
[0033] Perform a Conv1×1 convolution operation on F1 to extract preliminary features and generate an adaptive fusion feature map X1. Then, X1 undergoes feature processing through two different paths:
[0034] In path one, X1 extracts local features through deformable convolution to obtain an adaptive fusion feature map X2. Then, X2 is processed by the Sigmoid activation function to generate an adaptive weight map. X2 is multiplied element-wise with the adaptive weight map to obtain an adaptive fusion feature map X3.
[0035] In path two, X1 first extracts features through deformable convolution to obtain an adaptive fusion feature map X4. Then, X4 is subjected to max pooling to enhance the robustness of the features, resulting in an adaptive fusion feature map X5. Next, X5 is upsampled to obtain an adaptive fusion feature map X6. X3 and X6 are concatenated along the channel dimension to obtain an adaptive fusion feature map X7. Subsequently, X7 is subjected to a Conv1×1 convolution operation to obtain the final output adaptive fusion feature map F2.
[0036] In a preferred embodiment of the method for early warning of bird droppings flashover and insulation sleeve damage in transmission lines according to the present invention, the construction and operation process of the second adaptive feature fusion module is as follows:
[0037] Perform a Conv3×3 convolution operation on F12 to extract preliminary features and generate an adaptive fusion feature map Z1. Then, Z1 undergoes feature processing through two different paths:
[0038] In path one, Z1 extracts local features through deformable convolution to obtain an adaptive fusion feature map Z2. Then, Z2 is processed by the Sigmoid activation function to generate an adaptive weight map. Z2 is multiplied element-wise with the adaptive weight map to obtain an adaptive fusion feature map Z3.
[0039] In path two, Z1 first extracts features through deformable convolution to obtain an adaptive fusion feature map Z4. Then, Z4 is enhanced with average pooling to obtain an adaptive fusion feature map Z5. Next, Z5 is upsampled to obtain an adaptive fusion feature map Z6. Z3 and Z6 are concatenated along the channel dimension to obtain an adaptive fusion feature map Z7. Finally, Z7 is subjected to a Conv1×1 convolution operation to obtain the final output adaptive fusion feature map F13.
[0040] As a preferred embodiment of the early warning method for preventing bird droppings flashover and insulation sleeve damage in power transmission lines according to the present invention, the construction and execution process of the Ghost point-by-point convolutional multi-scale feature fusion module is as follows:
[0041] S321: Perform Conv 1×1 convolution operation, batch normalization, and LeakyReLU activation function on the input feature map F17 to generate a preliminary insulation sleeve damage feature map T1;
[0042] S322: Split T1 into three feature processing paths: In path one, after a Conv 1×1 convolution operation, an insulating sleeve damage feature map T2 is generated; a DWConv 5×5 convolution operation is performed on T2 to obtain an insulating sleeve damage feature map T5; In path two, a DWConv 3×3 convolution operation is performed on T1 to generate an insulating sleeve damage feature map T3; In path three, a GhostConv convolution operation is performed on T1 to obtain an insulating sleeve damage feature map T4; a Conv 3×3 convolution operation is performed on T4 to obtain an insulating sleeve damage feature map T6; simultaneously, a Conv 1×1 convolution operation is performed on T4 to obtain an insulating sleeve damage feature map T7;
[0043] S323: T2, T3, T5, T6 and T7 are spliced along the channel dimension by the Concat operation to obtain the fused insulation sleeve damage feature map T8;
[0044] S324: The convolution operation GhostConv is used to further process T8 to obtain a more compact and efficient insulating sleeve damage feature map T9, thereby completing the fusion and optimization of multi-scale features.
[0045] As a preferred embodiment of the early warning method for preventing bird droppings flashover and insulation sleeve damage in power transmission lines according to the present invention, the construction and execution process of the grouped convolution and coordinated attention residual expansion module is as follows:
[0046] S331: Perform CBS processing on the input feature map T9 to generate the insulation sleeve defect perception feature map A1. Then, perform feature splitting on A1, dividing it into three feature processing paths: A2, A3, and A4. In path one, A2 is processed by a Conv 1×1 convolution operation to generate the insulation sleeve defect perception feature map A5. In path two, A5 and A3 are added element-wise using a residual connection method to form the insulation sleeve defect perception feature map A6. A6 is then processed by a Conv 1×1 convolution operation to generate feature map A7. In path three, A4 and A7 are added element-wise using a residual connection method to form the insulation sleeve defect perception feature map A8. A8 is then input into the first group convolution and coordinated attention layer for further processing to generate the insulation sleeve defect perception feature map A9. A9 is then input into the second GCA layer for further processing to generate the insulation sleeve defect perception feature map A10. A10 is then input into the third GCA layer for further processing to generate the insulation sleeve defect perception feature map A11.
[0047] S332: Perform a concatenation operation on A5, A7, and A11 to obtain the insulation sleeve defect perception feature map A12. Perform a concat operation on feature maps A5, A7, and A11 along the channel dimension to obtain the final fused insulation sleeve defect perception feature map A12. Perform a Conv 1×1 convolution operation on A12 to generate the insulation sleeve defect perception feature map A13. Input A13 into the target detection head Head. Use the Head to detect A13 and output a tensor containing detection information. Each row corresponds to a detection result. The detection result is the detection result of the bird droppings flashover prevention insulation sleeve damage of the transmission line, including the bounding box coordinates of the damaged area, the damage type label, and the confidence score.
[0048] As a preferred embodiment of the early warning method for preventing bird droppings flashover and insulation sleeve damage in power transmission lines according to the present invention, the construction and operation process of the grouped convolution and coordinated attention layer is as follows:
[0049] S3311: Perform a Conv 1×1 convolution operation on the input feature map A8, adjust the number of channels, extract preliminary features, and generate the attention-optimized feature map S1;
[0050] S3312: Perform grouped convolution operation on feature map S1, further extract features from different channels through grouped convolution, generate attention-optimized feature map S2, perform Conv 1×1 convolution operation on S2, and further optimize the representation of feature map to generate attention-optimized feature map S3;
[0051] S3313: S3 is processed by a coordinated attention mechanism to enhance the features of key regions by coordinating the spatial and channel information of the features, generating an attention-optimized feature map S4. S4 is then averaged to generate the final attention-optimized feature map A9, thereby completing the integration and optimization of multi-scale features.
[0052] Compared with existing technologies:
[0053] This invention, when handling the task of detecting damage to insulation sleeves for bird droppings flashover prevention in power transmission lines, can more accurately capture and identify the damage characteristics of the insulation sleeve, effectively reducing the interference of complex environments on damage characteristics, and solving the problem of unclear damage feature representation caused by complex backgrounds, changes in lighting, or dirt obscuring the surface in existing technologies. Through multi-scale feature extraction and multi-level feature fusion, this invention can more accurately distinguish damaged areas from normal insulation sleeves, solving the problem of damaged areas being easily confused with background information in complex environments. Furthermore, this invention enhances the detection capability for fine-grained damage such as micro-cracks and surface wear, significantly reducing the problem of damage feature loss caused by excessive pooling and convolution operations in existing technologies, thus significantly improving the accuracy and robustness of insulation sleeve damage detection. Attached Figure Description
[0054] Figure 1 This is a schematic diagram of the process of the present invention;
[0055] Figure 2 This is a schematic diagram of the main structure of the ISDI-Net of the present invention;
[0056] Figure 3 This is a schematic diagram of the multi-level adaptive feature fusion module of the present invention;
[0057] Figure 4 This is a schematic diagram of the AFM-1 submodule structure of the present invention;
[0058] Figure 5 This is a schematic diagram of the AFM-2 submodule structure of the present invention;
[0059] Figure 6 This is a schematic diagram of the Ghost pointwise convolutional multi-scale feature fusion module of the present invention;
[0060] Figure 7 This is a schematic diagram of the grouped convolution and coordinated attention residual expansion module of the present invention;
[0061] Figure 8 This is a schematic diagram of the grouped convolution and coordinated attention layer of the present invention. Detailed Implementation
[0062] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0063] This invention provides a method for early warning of bird droppings flashover damage to insulation sleeves in power transmission lines, used to detect and identify damage to insulation sleeves in power lines. This invention designs an insulation sleeve damage detection network called ISDI-Net (Insulation Sleeve Damage Identification Network). This network can effectively suppress complex background interference, deeply analyze insulation sleeve images, accurately capture and extract features of damaged areas, and finally generate a damage detection result image. In the damage detection result image, the detected damaged areas are marked with bounding boxes, while intact areas remain unchanged. Please refer to [link / reference]. Figures 1-8 The overall technical solution flowchart of this invention is as follows: Figure 1 As shown; the main structure of ISDI-Net is as follows Figure 2 As shown;
[0064] The specific steps are as follows:
[0065] S1: Construct the original dataset for detecting damage to the insulation sleeve of transmission lines to prevent bird droppings flashover, or use the existing dataset for detecting damage to the insulation sleeve. Collect image data of insulation sleeve damage under different environments, and manually annotate the collected image data. Use the annotation tool to select the location of the damaged area and label it.
[0066] S1 includes, but is not limited to, the following embodiments:
[0067] The more diverse the types of insulation sleeve damage in the dataset, the more types of damage the model can identify. Based on our self-collected original dataset for insulation sleeve damage detection, there are a total of 6000 images, including 2400 images with minor damage, 2000 images with moderate damage, 1000 images with severe damage, and 600 images with other abnormal conditions (such as contamination, aging, etc.). The labels here can be divided into four categories: minor damage, moderate damage, severe damage, and other abnormalities. The LabelImg tool is used to label the dataset and export an XML label file, which contains the label name and annotation box information of the damaged area.
[0068] S2: Preprocess the original dataset for detecting bird droppings flashover damage to transmission line insulation sleeves to obtain a preprocessed dataset of insulation sleeve damage; divide the dataset of insulation sleeve damage into training set, validation set and test set.
[0069] The specific steps of S2 are as follows:
[0070] S21: The insulation sleeve images in the original dataset for insulation sleeve damage detection are cropped to a uniform size, and then data augmentation processing is performed on all insulation sleeve images to obtain the data-augmented insulation sleeve damage dataset. The purpose of data augmentation processing on the insulation sleeve images is to increase the size of the dataset, enhance the robustness of the damage detection model, and reduce the model's sensitivity to environmental changes.
[0071] S22: Divide the data augmented insulation sleeve damage dataset into training set, validation set and test set so that the deep learning-based insulation sleeve damage detection model can be trained in the future.
[0072] S2 includes, but is not limited to, the following embodiments:
[0073] The original dataset for detecting bird droppings flashover damage to insulation sleeves on power transmission lines contained 6000 images. The images were first cropped to a uniform size of 512×512 pixels. The original dataset was then augmented using data augmentation methods, including adding random noise, Gaussian blur, simulating illumination changes, enhancing contrast, random rotation, and image flipping. These methods can be implemented using Python. Random noise can be added using the `cv2.randu()` function; Gaussian blur can be applied using the `PIL.ImageFilter.GaussianBlur()` method; illumination change simulation can be performed using `cv2.convertScaleAbs()` for brightness adjustment; contrast enhancement can be achieved using the `ImageEnhance.Contrast()` method; rotation transformations can be performed using `cv2.getRotationMatrix2D()` and `cv2.warpAffine()`; and image flipping can be performed using the `cv2.flip()` function. Finally, the dataset was augmented to 18000 images and divided in an 8:1:1 ratio, resulting in 14400 images for training, 1800 images for validation, and 1800 images for testing.
[0074] S3: Constructing the ISDI-Net network for detecting damage to insulation sleeves of power transmission lines to prevent bird droppings flashover; The process of using ISDI-Net to detect damage to conductor insulation sleeve images consists of three steps: global and local feature extraction; multi-scale feature enhancement; and deep feature fusion and damage prediction. The above three main steps correspond to the three modules of the ISDI-Net network model: multi-level adaptive feature fusion module, Ghost pointwise convolution multi-scale feature fusion module, and grouped convolution and coordinated attention residual expansion module.
[0075] The specific steps of S3 are as follows:
[0076] S31: The images in the augmented insulation sleeve damage dataset are input into the multi-level adaptive feature fusion module to obtain the damage enhancement feature map of the damaged area of the insulation sleeve. This feature map can highlight the salient features of the damaged area while suppressing interference from complex backgrounds. The main function of the multi-level adaptive feature fusion module is to improve the perception capability of damaged areas at different scales through multi-level feature extraction and fusion. The adaptive feature fusion module achieves adaptive feature enhancement through a multi-path feature extraction and fusion mechanism, which helps to improve the feature expression capability of small targets or key areas, making the details such as tiny cracks and slight wear on the insulation sleeve clearer, thereby improving the accuracy of damage warning. Its module structure is as follows: Figure 3 As shown;
[0077] The construction and operation process of the multi-level adaptive feature fusion module is as follows:
[0078] S311: Perform CBS operation on the image of the bird droppings flashover protection insulation sleeve of the transmission line to obtain the damage enhancement feature map F1. Then, input F1 into the first adaptive feature fusion module (i.e. Figure 3 Feature extraction is performed on AFM-1 to generate a damage enhancement feature map F2. Then, a deeper local feature is extracted through Conv3×3 convolution operation to obtain a damage enhancement feature map F3. At the same time, Conv3×3 convolution operation is performed on F1 to obtain a damage enhancement feature map F5.
[0079] S312: Using a residual connection method, F3 and F5 are added element-wise to form a damaged enhanced feature map F6. Then, F6 is input again into the first adaptive feature fusion module (i.e., Figure 3 Feature extraction is performed using AFM-1 to obtain the damage enhancement feature map F7. Conv3×3 convolution operation is performed on F7 to obtain the damage enhancement feature map F8. Finally, Conv1×1 convolution operation is performed on F8 to obtain the damage enhancement feature map F9.
[0080] S313: Normalize F5 using the Sigmoid activation function to obtain attention weights. Weight F5 with these attention weights to obtain the damage enhancement feature map F10. Further extract global information from F10 using Conv5×5 convolution to obtain the damage enhancement feature map F11.
[0081] S314: Using a residual connection method, F11 and F9 are added element-wise to form a damaged enhancement feature map F12. F12 is then input into the second adaptive feature fusion module (i.e., Figure 3 Deep adaptive feature extraction is performed on AFM-2 in the model to generate a damaged enhanced feature map F13. Then, 5×5 convolution is performed on F13 to enhance the feature expression ability, resulting in a damaged enhanced feature map F14.
[0082] S315: Input F11 into the first adaptive feature fusion module (i.e. Figure 3 Feature extraction is performed using AFM-1 to obtain the damage enhancement feature map F15. Conv1×1 convolution operation is performed on F15 to generate the damage enhancement feature map F16. Finally, F14 and F16 are fused by element-wise addition to obtain the final output damage enhancement feature map F17.
[0083] CBS operation is a basic convolutional module that extracts local features of an image through a 3×3 convolutional kernel, then performs batch normalization (BatchNorm) to stabilize the feature distribution, and introduces nonlinearity through an activation function to enhance the model's feature representation ability.
[0084] The construction and operation process of the first adaptive feature fusion module is as follows: Figure 4 As shown:
[0085] Perform a Conv1×1 convolution operation on F1 to extract preliminary features and generate an adaptive fusion feature map X1. Then, X1 undergoes feature processing through two different paths:
[0086] In path one, X1 is processed by deformable convolution (i.e., Figure 4 Local features are extracted using DConv to obtain an adaptive fusion feature map X2. X2 is then processed through a Sigmoid activation function to generate an adaptive weight map. X2 is then multiplied element-wise with the adaptive weight map (i.e.,...). Figure 4 The * operation in the middle) is used to obtain the adaptive fusion feature map X3;
[0087] In path two, X1 first extracts features through deformable convolution to obtain an adaptive fused feature map X4, and then X4 is processed by max pooling (i.e., ... Figure 4 The MaxPool operation in the model enhances the robustness of the features, resulting in an adaptive fused feature map X5; then, X5 is upsampled (i.e., ...). Figure 4 Upsample (in the process of sampling), to obtain the adaptive fusion feature map X6, and then concatenate X3 and X6 along the channel dimension (i.e., Figure 4 The Concat operation is performed to obtain the adaptive fusion feature map X7. Then, Conv1×1 convolution is performed on X7 to obtain the final output adaptive fusion feature map F2.
[0088] The construction and operation process of the second adaptive feature fusion module is as follows: Figure 5 As shown:
[0089] Perform a Conv3×3 convolution operation on F12 to extract preliminary features and generate an adaptive fusion feature map Z1. Then, Z1 undergoes feature processing through two different paths:
[0090] In path one, Z1 is achieved through deformable convolution (i.e., Figure 5 In path two, Z1 first extracts features through deformable convolution to obtain adaptive fusion feature map Z4. Then, Z4 is processed through the sigmoid activation function to generate an adaptive weight map. Z2 is then multiplied element-wise with the adaptive weight map to obtain adaptive fusion feature map Z3. In path two, Z1 first extracts features through deformable convolution to obtain adaptive fusion feature map Z4. Then, Z4 is processed through average pooling to enhance the robustness of the features to obtain adaptive fusion feature map Z5. Finally, Z5 is upsampled (i.e., ...). Figure 4 Upsample (in the process of sampling), to obtain the adaptive fusion feature map Z6, and then concatenate Z3 and Z6 along the channel dimension (i.e., Figure 5 The Concat operation is used to obtain the adaptive fusion feature map Z7. Then, Conv1×1 convolution is performed on Z7 to obtain the final output adaptive fusion feature map F13.
[0091] S32: Input F17 into the Ghost pointwise convolutional multi-scale feature fusion module, such as... Figure 6 As shown;
[0092] The construction and execution process of the Ghost pointwise convolutional multi-scale feature fusion module is as follows:
[0093] S321: Perform Conv 1×1 convolution operation, batch normalization (BN), and LeakyReLU activation function on the input feature map F17 to generate a preliminary insulation sleeve damage feature map T1;
[0094] S322: Split T1 into three feature processing paths: In path one, after a Conv 1×1 convolution operation, an insulating sleeve damage feature map T2 is generated; a DWConv 5×5 convolution operation is performed on T2 to obtain an insulating sleeve damage feature map T5; In path two, a DWConv 3×3 convolution operation is performed on T1 to generate an insulating sleeve damage feature map T3; In path three, a GhostConv convolution operation is performed on T1 to obtain an insulating sleeve damage feature map T4; a Conv 3×3 convolution operation is performed on T4 to obtain an insulating sleeve damage feature map T6; simultaneously, a Conv 1×1 convolution operation is performed on T4 to obtain an insulating sleeve damage feature map T7;
[0095] S323: T2, T3, T5, T6 and T7 are spliced along the channel dimension by the Concat operation to obtain the fused insulation sleeve damage feature map T8;
[0096] S324: The convolution operation GhostConv is used to further process T8 to obtain a more compact and efficient insulating sleeve damage feature map T9, thereby completing the fusion and optimization of multi-scale features;
[0097] GhostConv convolution is an efficient convolution operation designed to reduce computational cost and parameter count while maintaining feature extraction capabilities. It generates a subset of features through a small number of standard convolutions and then uses a series of inexpensive linear transformations to generate "ghost features," thereby reducing computational overhead.
[0098] S33: Input T9 into the grouped convolution and coordinated attention residual expansion module, such as Figure 7 As shown;
[0099] The construction and execution process of the grouped convolution and coordinated attention residual expansion module is as follows:
[0100] S331: Perform CBS processing on the input feature map T9 to generate an insulation sleeve defect sensing feature map A1. Then, perform feature splitting on A1 (i.e., ... Figure 7 The Split operation in the model is divided into three feature processing paths: A2, A3, and A4. In path one, A2 is processed by a Conv 1×1 convolution to generate the insulation sleeve defect perception feature map A5. In path two, A5 and A3 are added element-wise using a residual connection method to form the insulation sleeve defect perception feature map A6. A6 is then processed by a Conv 1×1 convolution to generate feature map A7. In path three, A4 and A7 are added element-wise using a residual connection method to form the insulation sleeve defect perception feature map A8. A8 is then input into the first group convolution and the coordinating attention layer (i.e., ... Figure 7 The GCA Layer in the middle is further processed to generate the insulation sleeve defect perception feature map A9. A9 is then input into the second GCA Layer for further processing to generate the insulation sleeve defect perception feature map A10. A10 is then input into the third GCA Layer for further processing to generate the insulation sleeve defect perception feature map A11.
[0101] S332: Perform a splicing operation on A5, A7, and A11 (i.e., ... Figure 7The concat operation is used to obtain the insulation sleeve defect perception feature map A12. The feature maps A5, A7, and A11 are then concatenated along the channel dimension using the concat operation to obtain the final fused insulation sleeve defect perception feature map A12. A12 is then convolved using a Conv 1×1 operation to generate the insulation sleeve defect perception feature map A13. A13 is then input into the target detection head Head, and the Head is used to detect A13. The output is a tensor containing detection information, with each row corresponding to a detection result. The detection result is the detection result of the bird droppings flashover prevention insulation sleeve damage detection result for the transmission line, including the bounding box coordinates of the damaged area, the damage type label, and the confidence score.
[0102] The construction and operation process of the grouped convolution and coordinated attention layer is as follows: Figure 8 As shown:
[0103] S3311: Perform a Conv 1×1 convolution operation on the input feature map A8, adjust the number of channels, extract preliminary features, and generate the attention-optimized feature map S1;
[0104] S3312: Perform grouped convolution operation on feature map S1 (i.e. Figure 8 The Group Conv operation in the model further extracts features from different channels through group convolution, generating attention-optimized feature map S2. S2 is then subjected to Conv 1×1 convolution operation, and the representation of the feature map is further optimized to generate attention-optimized feature map S3.
[0105] S3313: By coordinating attention mechanisms (i.e. Figure 8 The Coordinate Attention layer processes S3 by coordinating the spatial and channel information of the features to enhance the features of key regions, generating an attention-optimized feature map S4. S4 is then subjected to average pooling to generate the final attention-optimized feature map A9, thus completing the integration and optimization of multi-scale features. The construction and operation process of the grouped convolution and coordinated attention layer is as follows:
[0106] S3311: Perform a Conv 1×1 convolution operation on the input feature map A8, adjust the number of channels, extract preliminary features, and generate the attention-optimized feature map S1;
[0107] S3312: Perform grouped convolution operation on feature map S1, further extract features from different channels through grouped convolution, generate attention-optimized feature map S2, perform Conv 1×1 convolution operation on S2, and further optimize the representation of feature map to generate attention-optimized feature map S3;
[0108] S3313: S3 is processed through a coordinated attention mechanism to enhance the features of key regions by coordinating the spatial and channel information of the features, generating an attention-optimized feature map S4, and then performing average pooling on S4 (i.e., ...). Figure 8 The AveragePooling operation in the process generates the final attention-optimized feature map A9, thereby completing the integration and optimization of multi-scale features;
[0109] Among them, coordinated attention is an improved attention mechanism that aims to capture spatial and channel information simultaneously to enhance feature representation capabilities. This method performs well in tasks such as object detection, image classification, and semantic segmentation, and is particularly suitable for scenes with complex backgrounds or fine-grained features.
[0110] The head is responsible for directly detecting the object's class, location (including bounding box coordinates), and presence confidence from the feature map; the output of the detection head is a three-dimensional tensor, where each row contains the class probability, bounding box coordinates, and object presence confidence for each detection box at each spatial location (relative to the feature map);
[0111] (1) Bounding box prediction: Each cell is responsible for predicting the coordinates and confidence scores of several bounding boxes; the coordinates of the bounding boxes include: center point (Position relative to the feature map cell), width and height (Scaled relative to the width and height of the entire image);
[0112] (2) Category and confidence prediction: In addition to coordinate prediction, each bounding box also has a confidence prediction, which is used to indicate the probability that there is an object of a specific category within the box;
[0113] S3 includes, but is not limited to, the following embodiments:
[0114] Images from the wire insulation sleeve damage detection dataset are input into a multi-level adaptive feature fusion module. The images are 512 x 512 pixels in size and have 3 channels. First, a 3×3 convolution kernel is applied to the input image to perform a convolution operation on the haze image. Then, batch normalization and activation function processing (CBS) are applied to obtain a damage enhancement feature map F1, which is 512 x 512 pixels in size and has 64 channels. Next, F1 is input into the first adaptive feature fusion module (AFM-1) for feature extraction, generating a damage enhancement feature map F2, which is 512 x 512 pixels in size and has 64 channels. A 3x3 convolution operation is then used to further extract local features, resulting in a damage enhancement feature map F3, which is 512 x 512 pixels in size and has 64 channels. Simultaneously, a 3x3 convolution operation is performed on F1 to obtain a damage enhancement feature map F5, which is 512 x 512 pixels in size and has 64 channels. Finally, residual connections are used to connect F3 and F5. Element-wise addition is performed to form the damaged enhancement feature map F6, which has a size of 512×512 pixels and 64 channels. Then, F6 is re-input into AFM-1 for feature extraction to obtain the damaged enhancement feature map F7, which also has a size of 512×512 pixels and 64 channels. A 3x3 convolution operation is performed on F7 to obtain the damaged enhancement feature map F8, which also has a size of 512×512 pixels and 64 channels. A 1x1 convolution operation is then performed on F8 to generate the damaged enhancement feature map F9, which also has a size of 512×512 pixels and 64 channels. Simultaneously, F5 is normalized using the Sigmoid activation function to obtain attention weights. These weights are then used to weight F5 to obtain the damaged enhancement feature map F10, which has a size of 512×512 pixels and 64 channels. Finally, F10 is subjected to a 5x5 convolution operation. Further global feature extraction is performed to form a damaged enhancement feature map F11, with a size of 512×512 pixels and 64 channels. Using a residual connection method, F11 and F9 are added element-wise to obtain a damaged enhancement feature map F12, with a size of 512×512 pixels and 64 channels. Then, F12 is input into the second adaptive feature fusion module (AFM-2) for deep adaptive feature extraction, generating a damaged enhancement feature map F13, with a size of 512×512 pixels and 64 channels. The feature representation capability is further enhanced through a 5x5 convolution to obtain a damaged enhancement feature map F14, with a size of 512×512 pixels and 64 channels. Simultaneously, F11 is input into the first adaptive feature fusion module (AFM-1) for feature extraction, obtaining a damaged enhancement feature map F15, with a size of 512×512 pixels. The F15 image is then convolved with 1x1 to obtain the damaged enhancement feature map F16, which is 512×512 pixels and has 64 channels.Finally, F14 and F16 are fused element-wise to obtain the final output damaged enhancement feature map F17, which is 512×512 pixels in size and has 64 channels.
[0115] In the AFM-1 submodule, a 1x1 convolution operation is performed on the input feature map F1 to extract preliminary features, generating an adaptive fusion feature map X1 with a size of 512×512 pixels and 64 channels. Then, X1 undergoes feature processing via two different paths. In path one, X1 extracts local features through deformable convolution (DConv) to obtain an adaptive fusion feature map X2 with a size of 512×512 pixels and 64 channels. X2 is then processed using a Sigmoid activation function to generate an adaptive weight map. Finally, X2 is multiplied element-wise with the adaptive weight map to obtain an adaptive fusion feature map X3 with a size of 512×512 pixels and 64 channels. In path two, X1 first extracts features through deformable convolution to obtain an adaptive fusion feature map X4 with a size of 512×512 pixels. The first feature map, X4, is first pixel-256 with 64 channels. Then, X4 is enhanced with max pooling to obtain an adaptive fusion feature map X5, which has a size of 256×256 pixels and 64 channels. Next, X5 is upsampled to its original size of 512 x 512 pixels to restore it to X6, which has a size of 512×512 pixels and 64 channels. Then, X6 and X3 are concatenated along the channel dimension to obtain an adaptive fusion feature map X7, which has a size of 512×512 pixels and 128 channels. Finally, X7 is dimensionality reduced back to 64 channels by 1x1 convolution to obtain the final output adaptive fusion feature map F2, which has a size of 512×512 pixels and 64 channels.
[0116] In the AFM-2 submodule, a 3x3 convolution operation is performed on the input feature map F12 to extract preliminary features, generating an adaptive fusion feature map Z1 with a size of 512×512 pixels and 64 channels. Then, Z1 undergoes feature processing via two different paths. In path one, Z1 extracts local features through deformable convolution, resulting in an adaptive fusion feature map Z2 with a size of 512×512 pixels and 64 channels. Z2 is then processed using a Sigmoid activation function to generate an adaptive weight map. Next, Z2 is element-wise multiplied with the adaptive weight map to obtain an adaptive fusion feature map Z3 with a size of 512×512 pixels and 64 channels. In path two, Z1 first extracts features through deformable convolution, resulting in an adaptive fusion feature map Z4 with a size of 512×512 pixels and 64 channels. Z4 is then enhanced with average pooling (AvgPool) to improve feature robustness, resulting in an adaptive fusion feature map Z5 with a size of 256×256 pixels. The Z5 map is first 512x512 pixels with 64 channels. Then, the Z5 map is upsampled back to its original size of 512x512 pixels to restore the feature map Z6, which is 512×512 pixels with 64 channels. Subsequently, Z6 and Z3 are concatenated along the channel dimension to obtain the adaptive fusion feature map Z7, which is 512×512 pixels with 128 channels. Finally, Z7 is 1x1 convolutionally reduced to 64 channels to obtain the final output adaptive fusion feature map F13, which is 512×512 pixels with 64 channels.
[0117] In the Ghost pointwise convolutional multi-scale feature fusion module, the input feature map F17 undergoes a 1x1 convolution operation and is processed by batch normalization (BN) and LeakyReLU activation functions to generate a preliminary insulation sleeve damage feature map T1, which is 512×512 pixels and has 64 channels. Then, T1 is split into three feature processing paths: In path one, a 1x1 convolution operation generates an insulation sleeve damage feature map T2, which is 512×512 pixels and has 64 channels; a depthwise separable 5x5 convolution (DWConv5x5) operation is performed on T2 to obtain the insulation sleeve damage feature map T5, which is 512×512 pixels and has 64 channels; In path two, a depthwise separable 3x3 convolution (DWConv3x3) operation is performed on T1 to generate the insulation sleeve damage feature map T3, which is 512×512 pixels and has 64 channels; In the third path, T1 is processed by Ghost... A convolution (GhostConv) operation yields the insulation sleeve damage feature map T4, with a size of 512×512 pixels and 64 channels. Subsequently, a 3x3 convolution operation is performed on T4 to obtain the insulation sleeve damage feature map T6, also with a size of 512×512 pixels and 64 channels. Simultaneously, a 1x1 convolution operation is performed on T4 to obtain the insulation sleeve damage feature map T7, with a size of 512×512 pixels and 64 channels. Then, feature maps T2, T3, T5, T6, and T7 are concatenated along the channel dimension to obtain the fused insulation sleeve damage feature map T8, with a size of 512×512 pixels and 320 channels. Finally, GhostConv is used to further process T8, generating a more compact and highly expressive insulation sleeve damage feature map T9, with a size of 512×512 pixels and 64 channels. This completes the fusion and optimization of multi-scale features.
[0118] In the grouped convolution and coordinated attention residual extension module, the input feature map T9 is processed by CBS to generate an insulation sleeve defect-aware feature map A1, which is 512×512 pixels in size and has 64 channels. Next, A1 is split into three feature processing paths, generating A2, A3, and A4, which are also 512×512 pixels in size and have 64 channels. In path one, A2 is processed by a 1×1 convolution to generate an insulation sleeve defect-aware feature map A5, which is also 512×512 pixels in size and has 64 channels. In path two, A5 is added element-wise to A3 using a residual connection method to obtain feature map A6, which is also 512×512 pixels in size and has 64 channels. Finally, A6 is processed by a 1×1 convolution to generate feature map A7, which is also 512×512 pixels in size. In path three, a residual connection is used to add A4 and A7 element-wise to obtain feature map A8, which is 512×512 pixels and has 64 channels. Then, A8 is processed through the first grouped convolution and Coordinating Attention (GCA) layer to generate feature map A9, which is also 512×512 pixels and has 64 channels. A9 is further processed through the second GCA layer to obtain feature map A10, which is also 512×512 pixels and has 64 channels. A10 is then processed again through the third GCA layer to obtain feature map A11, which is also 512×512 pixels and has 64 channels. The first three paths output feature maps A5, A7, and A11 are then added together to obtain feature map A12, which is 512×512 pixels and has 64 channels. A11, A5, and A7 are then concatenated along the channel dimension using a Concat operation to obtain the final fused feature map A12, which is 512×512 pixels and has 192 channels. A12 is then convolved with a 1×1 convolution to generate feature map A13, which is 512×512 pixels and has 64 channels. A13 is then input into the target detection head (Head), which performs detection on A13 and outputs a tensor containing detection information. Each row of the tensor corresponds to a detection result, including the bounding box coordinates of the damaged area, the damage type label, and the confidence score.
[0119] In the grouped convolution and coordinated attention layer, a 1×1 convolution operation is performed on the input feature map A8 to adjust the number of channels and extract preliminary features, generating an attention-optimized feature map S1 with a size of 512×512 pixels and 64 channels. Next, a group convolution operation is performed on S1 to further extract features from different channels, generating an attention-optimized feature map S2 with a size of 512×512 pixels and 64 channels. Subsequently, a 1×1 convolution operation is performed on S2 to further optimize the representation of the feature map, generating an attention-optimized feature map S3 with a size of 512×512 pixels and 64 channels.
[0120] Finally, S3 is processed through a Coordinate Attention mechanism to coordinate the spatial and channel information of the features to enhance the features of key regions, generating an attention-optimized feature map S4 with a size of 512×512 pixels and 64 channels. Then, S4 is subjected to Average Pooling to reduce the spatial dimension, generating the final attention-optimized feature map A9 with a size of 512×512 pixels and 64 channels.
[0121] S4: Based on the constructed ISDI-Net model of the transmission line anti-bird droppings flashover insulation sleeve damage early warning network, train and update the parameters of each layer, initialize all neural network parameters, and set the model-related hyperparameters, including training rounds, batch size, weight decay coefficient, learning rate and total number of iterations.
[0122] S4 includes, but is not limited to, the following embodiments:
[0123] The ISDI-Net network for early warning of bird droppings flashover and insulation sheath damage on power transmission lines was trained. The entire training process was set to 120 epochs, with a weight decay coefficient of 0.005, an initial learning rate of 0.0001, and a batch size of 16. The first 60 epochs had 400 iterations, and the last 60 epochs had 500 iterations, for a total of 54,000 iterations. In each training round, data augmentation strategies were employed, including random cropping, horizontal flipping, rotation, contrast adjustment, and noise interference, to improve the robustness and generalization ability of the model. The resolution of all input images was uniformly adjusted to 640×640 to meet the network input requirements. After parameter initialization, the training and validation sets of insulation sheath damage data, which had been preprocessed in Step 2, were divided into multiple batches. Each batch of data was input into ISDI-Net for training, and the training loss value (loss) for that batch was calculated. The model parameters were then updated based on the loss value. After all batches of data in the entire training set had been trained, the validation phase began.
[0124] During the validation phase, validation set data is input into ISDI-Net in batches, and the validation loss (batch_loss) for each batch is calculated. ISDI-Net dynamically adjusts the learning rate based on the training loss and validation batch_loss values, and adopts an adaptive optimization strategy to improve training efficiency and model performance. The entire training process continues through multiple iterations until the batch_loss value tends to converge, indicating that the model has good insulation sheath damage detection capabilities and has completed training. Finally, the trained ISDI-Net can achieve high-precision insulation sheath damage early warning in complex environments, providing technical support for the safe operation of transmission lines.
[0125] S5: After the early warning model for bird droppings flashover and insulation sleeve damage of transmission lines is trained, the trained ISDI-Net model is used to detect insulation sleeve damage in transmission line images. The final output of the detection results includes the bounding box coordinates of the damaged area, the damage type label, and the confidence score, providing accurate early warning information for power equipment operation and maintenance.
[0126] Based on the above, this invention designs an insulating sleeve damage detection network ISDI-Net, whose structure includes three modules: a multi-level adaptive feature fusion module, a Ghost pointwise convolution multi-scale feature fusion module, and a grouped convolution and coordinated attention residual expansion module.
[0127] Multi-level adaptive feature fusion module: This module addresses the issues of blurred features and indistinct edges in the damaged area of the insulation sleeve. By introducing an adaptive feature fusion mechanism, this module can dynamically adjust the weights of features at different scales, improving the ability to perceive features in the damaged area. Simultaneously, by combining different features, the model can extract local detail information from the damaged area, enhancing its ability to identify different types of damage.
[0128] Ghost Pointwise Convolutional Multi-Scale Feature Fusion Module: This module employs Ghost pointwise convolution technology to reduce computational complexity while enhancing the model's ability to detect damaged areas of small targets. Through multi-scale feature fusion, it strengthens the edge contour information of damaged areas, improves the distinction between damaged areas and the background, and effectively reduces false positives and false negatives. This module is particularly suitable for damage detection tasks in complex power grid backgrounds, improving the precision of detection.
[0129] Grouped Convolution and Coordinated Attention Residual Extension Module: This module combines grouped convolution and coordinated attention mechanisms to achieve efficient feature extraction and key region focus. Grouped convolution reduces the number of parameters and computational overhead, improving the model's lightweight characteristics; the coordinated attention mechanism enhances the model's responsiveness to damaged areas and improves its adaptability to damaged areas of different scales. The residual extension structure further optimizes feature learning, ensuring that the model maintains high-precision damage detection capabilities in long-distance inspection scenarios, thereby improving the robustness and stability of detection.
[0130] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A method for early warning of bird droppings flashover and insulation sleeve damage on power transmission lines, characterized in that, The specific steps are as follows: S1: Construct the original dataset for detecting damage to the insulation sleeve of transmission lines to prevent bird droppings flashover, or use the existing dataset for detecting damage to the insulation sleeve. Collect image data of insulation sleeve damage under different environments, and manually annotate the collected image data. Use the annotation tool to select the location of the damaged area and label it. S2: Preprocess the original dataset for detecting bird droppings flashover damage to transmission line insulation sleeves to obtain a preprocessed dataset of insulation sleeve damage; divide the dataset of insulation sleeve damage into training set, validation set and test set. S3: Constructing the ISDI-Net network for detecting damage to insulation sleeves of power transmission lines to prevent bird droppings flashover; S4: Based on the constructed ISDI-Net model of the transmission line anti-bird droppings flashover insulation sleeve damage early warning network, train and update the parameters of each layer, initialize all neural network parameters, and set the model-related hyperparameters, including training rounds, batch size, weight decay coefficient, learning rate and total number of iterations. S5: After the early warning model for bird droppings flashover insulation sleeve damage of transmission line is trained, the trained ISDI-Net model is used to detect insulation sleeve damage in transmission line images. The final output of the detection results includes the bounding box coordinates of the damaged area, the damage type label and the confidence score, providing accurate early warning information for power equipment operation and maintenance. The specific steps for constructing the ISDI-Net network for detecting bird droppings flashover and insulation damage to power transmission lines using S3 are as follows: S31: Input the images in the data-enhanced insulation sleeve damage dataset into the multi-level adaptive feature fusion module to obtain the damage enhancement feature map F17 of the insulation sleeve damage area; S32: Input F17 into the Ghost pointwise convolutional multi-scale feature fusion module to obtain the insulation sleeve damage feature map T9; S33: Input T9 into the grouped convolution and coordinated attention residual expansion module; The construction and execution process of the Ghost pointwise convolutional multi-scale feature fusion module is as follows: S321: Perform Conv 1×1 convolution operation, batch normalization, and LeakyReLU activation function on the input feature map F17 to generate a preliminary insulation sleeve damage feature map T1; S322: Split T1 into three feature processing paths: In path one, after a Conv 1×1 convolution operation, an insulating sleeve damage feature map T2 is generated; a DWConv 5×5 convolution operation is performed on T2 to obtain an insulating sleeve damage feature map T5; In path two, a DWConv 3×3 convolution operation is performed on T1 to generate an insulating sleeve damage feature map T3; In path three, a GhostConv convolution operation is performed on T1 to obtain an insulating sleeve damage feature map T4; a Conv 3×3 convolution operation is performed on T4 to obtain an insulating sleeve damage feature map T6; simultaneously, a Conv 1×1 convolution operation is performed on T4 to obtain an insulating sleeve damage feature map T7; S323: T2, T3, T5, T6 and T7 are spliced along the channel dimension by the Concat operation to obtain the fused insulation sleeve damage feature map T8; S324: The convolution operation GhostConv is used to further process T8 to obtain a more compact and efficient insulating sleeve damage feature map T9, thereby completing the fusion and optimization of multi-scale features.
2. The method for early warning of bird droppings flashover and insulation sleeve damage in transmission lines according to claim 1, characterized in that, The specific steps of S2 are as follows: S21: Crops the insulation sleeve images in the original dataset of insulation sleeve damage detection to a uniform size, and then performs data augmentation processing on all insulation sleeve images to obtain the data-augmented insulation sleeve damage dataset; The purpose of data augmentation processing on insulating sleeve images is to increase the size of the dataset, enhance the robustness of the damage detection model, and reduce the model's sensitivity to environmental changes. S22: Divide the data augmented dataset of insulation sleeve damage into training, validation and test sets for subsequent training of the deep learning-based insulation sleeve damage detection model.
3. The method for early warning of bird droppings flashover and insulation sleeve damage on transmission lines according to claim 1, characterized in that, The construction and operation process of the multi-level adaptive feature fusion module is as follows: S311: The image of the bird droppings flashover protection insulation sleeve of the transmission line is processed by CBS operation to obtain the damage enhancement feature map F1. Then, F1 is input into the first adaptive feature fusion module for feature extraction to generate the damage enhancement feature map F2. Further local features are extracted through Conv3×3 convolution operation to obtain the damage enhancement feature map F3. At the same time, Conv3×3 convolution operation is performed on F1 to obtain the damage enhancement feature map F5. S312: Using the residual connection method, F3 and F5 are added element by element to form the damaged enhancement feature map F6. Then, F6 is input again into the first adaptive feature fusion module for feature extraction to obtain the damaged enhancement feature map F7. F7 is subjected to Conv3×3 convolution operation to obtain the damaged enhancement feature map F8. Finally, F8 is subjected to Conv1×1 convolution operation to obtain the damaged enhancement feature map F9. S313: Normalize F5 using the Sigmoid activation function to obtain attention weights. Weight F5 with these attention weights to obtain the damage enhancement feature map F10. Further extract global information from F10 using Conv5×5 convolution to obtain the damage enhancement feature map F11. S314: Using the residual connection method, F11 and F9 are added element by element to form the damaged enhancement feature map F12. F12 is input into the second adaptive feature fusion module for deep adaptive feature extraction to generate the damaged enhancement feature map F13. F13 is then convolved with 5×5 to enhance the feature expression ability, resulting in the damaged enhancement feature map F14. S315: Input F11 into the first adaptive feature fusion module for feature extraction to obtain the damaged enhanced feature map F15. Perform Conv1×1 convolution operation on F15 to generate the damaged enhanced feature map F16. Finally, fuse F14 and F16 by element-wise addition to obtain the final output damaged enhanced feature map F17.
4. The method for early warning of bird droppings flashover and insulation sleeve damage in transmission lines according to claim 3, characterized in that, The construction and operation process of the first adaptive feature fusion module is as follows: Perform a Conv1×1 convolution operation on F1 to extract preliminary features and generate an adaptive fusion feature map X1. Then, X1 undergoes feature processing through two different paths: In path one, X1 extracts local features through deformable convolution to obtain an adaptive fusion feature map X2. Then, X2 is processed by the Sigmoid activation function to generate an adaptive weight map. X2 is multiplied element-wise with the adaptive weight map to obtain an adaptive fusion feature map X3. In path two, X1 first extracts features through deformable convolution to obtain an adaptive fusion feature map X4. Then, X4 is subjected to max pooling to enhance the robustness of the features, resulting in an adaptive fusion feature map X5. Next, X5 is upsampled to obtain an adaptive fusion feature map X6. X3 and X6 are concatenated along the channel dimension to obtain an adaptive fusion feature map X7. Subsequently, X7 is subjected to a Conv1×1 convolution operation to obtain the final output adaptive fusion feature map F2.
5. The method for early warning of bird droppings flashover and insulation sleeve damage on transmission lines according to claim 3, characterized in that, The construction and operation process of the second adaptive feature fusion module is as follows: Perform a Conv3×3 convolution operation on F12 to extract preliminary features and generate an adaptive fusion feature map Z1. Then, Z1 undergoes feature processing through two different paths: In path one, Z1 extracts local features through deformable convolution to obtain an adaptive fusion feature map Z2. Then, Z2 is processed by the Sigmoid activation function to generate an adaptive weight map. Z2 is multiplied element-wise with the adaptive weight map to obtain an adaptive fusion feature map Z3. In path two, Z1 first extracts features through deformable convolution to obtain an adaptive fusion feature map Z4. Then, Z4 is enhanced with average pooling to obtain an adaptive fusion feature map Z5. Next, Z5 is upsampled to obtain an adaptive fusion feature map Z6. Z3 and Z6 are concatenated along the channel dimension to obtain an adaptive fusion feature map Z7. Finally, Z7 is subjected to a Conv1×1 convolution operation to obtain the final output adaptive fusion feature map F13.
6. The method for early warning of bird droppings flashover and insulation sleeve damage in transmission lines according to claim 1, characterized in that, The construction and execution process of the grouped convolution and coordinated attention residual expansion module is as follows: S331: Perform CBS processing on the input feature map T9 to generate the insulation sleeve defect perception feature map A1. Then, perform feature splitting on A1, dividing it into three feature processing paths: A2, A3, and A4. In path one, A2 is processed by a Conv 1×1 convolution operation to generate the insulation sleeve defect perception feature map A5. In path two, A5 and A3 are added element-wise using a residual connection method to form the insulation sleeve defect perception feature map A6. A6 is then processed by a Conv 1×1 convolution operation to generate feature map A7. In path three, A4 and A7 are added element-wise using a residual connection method to form the insulation sleeve defect perception feature map A8. A8 is then input into the first group convolution and coordinating attention layer for further processing to generate the insulation sleeve defect perception feature map A9. A9 is then input into the second group convolution and coordinating attention layer for further processing to generate the insulation sleeve defect perception feature map A10. A10 is then input into the third group convolution and coordinating attention layer for further processing to generate the insulation sleeve defect perception feature map A11. S332: Perform a concatenation operation on A5, A7, and A11 to obtain the insulation sleeve defect perception feature map A12. Perform a concat operation on feature maps A5, A7, and A11 along the channel dimension to obtain the final fused insulation sleeve defect perception feature map A12. Perform a Conv 1×1 convolution operation on A12 to generate the insulation sleeve defect perception feature map A13. Input A13 into the target detection head Head. Use the Head to detect A13 and output a tensor containing detection information. Each row corresponds to a detection result. The detection result is the detection result of the bird droppings flashover prevention insulation sleeve damage detection result of the transmission line, including the bounding box coordinates of the damaged area, the damage type label, and the confidence score.
7. The method for early warning of bird droppings flashover and insulation sleeve damage on transmission lines according to claim 6, characterized in that, The construction and operation process of the grouped convolution and coordinated attention layers is as follows: S3311: Perform a Conv 1×1 convolution operation on the input feature map A8, adjust the number of channels, extract preliminary features, and generate the attention-optimized feature map S1; S3312: Perform grouped convolution operation on feature map S1, further extract features from different channels through grouped convolution, generate attention-optimized feature map S2, perform Conv 1×1 convolution operation on S2, and further optimize the representation of feature map to generate attention-optimized feature map S3; S3313: S3 is processed by a coordinated attention mechanism to enhance the features of key regions by coordinating the spatial and channel information of the features, generating an attention-optimized feature map S4. S4 is then averaged to generate the final attention-optimized feature map A9, thereby completing the integration and optimization of multi-scale features.
Citation Information
Patent Citations
Fault insulator detection method based on candidate box generation and attention coordination mechanism
CN114359541A
Underwater small target detection method based on multilevel feature interactive attention
CN117392525A