Method for detecting damage to a ground wire at a concealed location of a power transmission line
By employing a Transformer model with multi-scale edge-texture separation, dual-domain dynamic selection, and intra-scale interaction module for decoupling high and low frequency features in the detection of conductor and ground wire damage in concealed parts of transmission lines, the problems of blurred edges, unclear textures, and inconsistent scales in the detection model were solved, achieving higher detection accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGSHA UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-04-03
- Publication Date
- 2026-06-19
AI Technical Summary
Existing deep learning detection models suffer from problems such as blurred edges and unclear textures in the detection of conductor and ground wire damage in concealed parts of transmission lines, difficulty in fusing global and local contextual information, and missed or false detections due to inconsistent defect scales.
A multi-scale edge-texture separation and dual-domain dynamic selection module is designed using the Transformer model. Combined with an intra-scale interaction module that decouples high and low frequency features and a multi-scale bidirectional information flow pyramid network, the feature interaction and fusion strategies are optimized to improve detection accuracy.
It effectively improves the accuracy and robustness of conductor and ground wire damage detection in concealed parts of transmission lines, and can quickly and accurately identify the location of damage in complex backgrounds.
Smart Images
Figure CN122243992A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power line inspection, and in particular to a method for detecting damage to conductors and ground wires in concealed parts of transmission lines. Background Technology
[0002] With the rapid development of my country's power industry, the voltage levels of transmission lines are becoming increasingly higher, and the amount of electricity transmitted is increasing. Defects in transmission equipment can lead to major power safety accidents, directly threatening the stable operation of the power grid. As key metal accessories for transmission lines, fittings bear the functions of conductor support, fixation, and splicing. Due to long-term tensile stress and exposure to complex outdoor environments, they often experience faults such as positional displacement and structural damage. Internal conductor and ground wire damage in components such as splicing pipes, suspension clamps, and tension clamps is particularly hidden and difficult to detect using conventional methods. In recent years, drones equipped with X-ray imaging equipment, combined with deep learning detection technology, have significantly improved the efficiency of intelligent inspection of transmission lines.
[0003] However, the physical structure of transmission lines determines the implicit positional relationships between various types of fittings, resulting in widespread occlusion between fittings in inspection images. Coupled with environmental interference and other issues leading to blurred edges, unclear textures, and inconsistent scales of defects in X-ray images, existing deep learning detection models have the following limitations in this specific application scenario: High-frequency detail feature representation collapse; the low-pass filtering effect of conventional convolution causes the loss of minute damage features with blurred edges and extremely low contrast during transmission, making robust extraction difficult; Scale adaptation imbalance; a single-scale extraction mechanism struggles to balance the detection accuracy of targets with different scales under a lightweight architecture, limiting its effective application in on-site inspections.
[0004] Therefore, those skilled in the art are dedicated to developing a method for detecting damage to conductors and ground wires in concealed parts of transmission lines. Summary of the Invention
[0005] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is that the intelligent detection of conductor and ground wire damage in the concealed parts of transmission lines has problems such as blurred edge of defect target, unclear texture, difficulty in fusion of global and local context information, and missed detection and false detection caused by different defect target scales.
[0006] To achieve the above objectives, this invention provides a method for detecting conductor and ground wire damage in concealed parts of transmission lines. The method comprises the following steps: Step 1: Acquiring an image of the target object and constructing a dataset of conductor and ground wire damage in concealed parts of the transmission line. Step 2: Based on a Transformer model, designing a multi-scale edge-texture separation and dual-domain dynamic selection module to reconstruct its backbone network. Step 3: Proposing an intra-scale interaction module based on high- and low-frequency feature decoupling to optimize the feature interaction process. Step 4: Proposing a multi-scale bidirectional information flow pyramid network to optimize the fusion strategy of multi-scale features in the model. Step 5: Outputting the damage location, category, and confidence level based on the detection results. This invention effectively improves the accuracy and robustness of conductor and ground wire damage detection in concealed parts of transmission lines under complex backgrounds.
[0007] Technical effect
[0008] The present invention provides a method for detecting conductor and ground wire damage in concealed parts of transmission lines. The method is reasonably designed and highly practical, and can effectively and quickly and accurately detect and identify conductor and ground wire damage in concealed parts of existing transmission lines in complex environments.
[0009] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description
[0010] Figure 1 This is a flowchart of the Transformer-based method for detecting conductor and ground wire damage in concealed parts of power transmission lines according to the present invention.
[0011] Figure 2 This is a network structure diagram of the Transformer-based transmission line conductor and ground wire damage detection model.
[0012] Figure 3 This is a schematic diagram of the multi-scale edge-texture separation and dual-domain dynamic selection module structure of the present invention;
[0013] Figure 4 This is a schematic diagram of the dual-domain selection mechanism structure of the present invention;
[0014] Figure 5 This is a structural diagram of the high- and low-frequency separation feature attention mechanism of the present invention;
[0015] Figure 6 This is a diagram of the multi-scale bidirectional information flow pyramid network structure of the present invention. Detailed Implementation
[0016] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.
[0017] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.
[0018] Specific implementation method one: Combining Figures 1 to 6 This embodiment describes the following steps:
[0019] Step 1: Obtain the image of the target to be tested and construct a dataset of conductor and ground wire damage in concealed parts of the transmission line;
[0020] Step 2: Based on the RT_DETR model, design a multi-scale edge-texture separation and dual-domain dynamic selection module to reconstruct the backbone network, such as... Figure 2 , Figure 3 and Figure 4 As shown;
[0021] Step 3: Propose an intra-scale interaction module based on high- and low-frequency feature decoupling to optimize the feature interaction process, such as... Figure 5 As shown;
[0022] Step 4: Propose a multi-scale bidirectional information flow pyramid network and optimize the fusion strategy of multi-scale features of the model;
[0023] Step 5: Output the damage location, category, and confidence level based on the detection results.
[0024] Specific Implementation Method Two: Combining Figure 3 and Figure 4 This implementation method is described below. Step 2 of this implementation method involves designing a multi-scale edge-texture separation and dual-domain dynamic selection module (MEDS) based on the Transformer model, and reconstructing the backbone network (MEDS_Net) to extract features. The Transformer target detection model is the RT-DETR model. Due to the complex structure of the concealed parts of transmission lines, X-ray images often have metal artifacts and low contrast due to minor damage. Traditional backbone networks tend to smooth out high-frequency details during multi-scale fusion. Therefore, this step introduces a multi-scale edge enhancer (MSEE) and a dual-domain selection mechanism (DSM) into the bottleneck sub-network of the RT-DETR-Resnet18 backbone network. The reconstruction of the transmission path from shallow features to the deep network includes the following steps:
[0025] Step 2.1: Extract and fuse multi-scale edge features using a multi-scale edge enhancer;
[0026] Step 2.2: Adaptively filter features in the spatial and frequency domains using a dual-domain selection mechanism;
[0027] Step 2.3: Perform global feature fusion to complete the backbone network reshaping. The reshaped backbone network is as follows: Figure 1 As shown in MEDS_Net.
[0028] Specific implementation method three, combined with Figure 3 This embodiment describes the following steps in step 2.1 of this embodiment, which involves extracting and fusing multi-scale edge features using a multi-scale edge enhancer:
[0029] Step 2.1.1: Perform adaptive average pooling on the input original feature map to map it into feature maps of various sizes (such as 3×3, 6×6, 9×9, 12×12).
[0030] Step 2.1.2: The pooled feature map is passed through 1×1 and 3×3 convolutional layers for non-linear refinement, and then upsampling is used to restore the feature map to the original input resolution;
[0031] Step 2.1.3: For each scale feature, perform average pooling on the original feature map to obtain a smooth feature map, and subtract the smooth feature map from the original feature map to accurately extract the high-frequency edge feature map;
[0032] Step 2.1.4: After performing 1×1 convolution to enhance details on the high-frequency edge feature map, perform residual fusion with the original feature map to obtain edge enhancement features at each scale;
[0033] Step 2.1.5: After restoring the edge enhancement features at each scale to the input resolution through bilinear interpolation, the features are stitched together along the channel dimension to generate multi-scale representation features.
[0034] Specific implementation method four, combined with Figure 4 This embodiment describes the following steps in step 2.2 of this embodiment, which utilizes a dual-domain selection mechanism to adaptively filter features in the spatial and frequency domains:
[0035] Step 2.2.1: Input the multi-scale representation features into the spatial selection module, and generate channel-level feature representations by sequentially passing them through pooling layers, convolutional layers, and cascaded deep convolutions. Then, modulate the representations with the input features to generate spatial selection features.
[0036] Step 2.2.2: Input the spatial selection feature into the frequency selection module, generate low-frequency features through channel global average pooling, and subtract the low-frequency features from the input to obtain high-frequency features;
[0037] Step 2.2.3: Multiply the high-frequency features element-wise with the input features and combine them with residual connections to generate a local attention map.
[0038] Specific implementation method five, combined with Figure 1 and Figure 5 This embodiment describes a scale-based interaction module based on high- and low-frequency feature decoupling, which optimizes the feature interaction process. Figure 2 As shown in FDIIM, the step involves reconstructing the intra-scale feature interaction module using a high-low frequency separation feature attention mechanism to enhance the perception of the global structure and local features of transmission line damage images.
[0039] Specific implementation method six, combined with Figure 5 This embodiment describes a high- and low-frequency separation feature attention mechanism, specifically step 3, which includes the following steps:
[0040] Step 3.1: Perform local window self-attention calculation through high-frequency paths to extract high-frequency local fine-grained features;
[0041] Step 3.2: Perform pooling dimensionality reduction and global attention calculation through low-frequency paths to extract low-frequency global context features;
[0042] Step 3.3: Perform semantic aggregation and module integration of high and low frequency features.
[0043] Specific implementation method seven, combined with Figure 5 This embodiment describes the following steps in step 3.1: Performing local window self-attention calculation via high-frequency paths to extract high-frequency local fine-grained features.
[0044] Step 3.1.1: Use a non-overlapping window segmentation strategy to divide the input feature map into multiple local windows;
[0045] Step 3.1.2: Independently apply a self-attention mechanism within each local window to capture fine-grained details and limit computational complexity to the local window. The self-attention calculation process within the local window is represented as follows:
[0046]
[0047] In the formula, Q, K, and V are the query, key, and value matrices within a local window under a high-frequency path, respectively, and d is the feature dimension.
[0048] Specific implementation method eight, combined with Figure 5This embodiment describes the following steps in step 3.2: performing pooling dimensionality reduction and global attention calculation via low-frequency paths to extract low-frequency global context features.
[0049] Step 3.2.1: Perform average pooling operation on the features within each local window to filter out high-frequency detail noise and condense the low-frequency signal representing the global topology and contextual semantics of the image;
[0050] Step 3.2.2: Project the pooled and compressed low-frequency signals onto low-frequency keys K and low-frequency values V, respectively, keeping the query Q in the low-frequency path aligned with the original input feature map X space, and apply standard attention calculation to the low-frequency key-value pairs to efficiently model long-range dependencies with reduced sequence length.
[0051]
[0052] In the formula, Q, , and represent the query, key, and value matrices within the local window under the low-frequency path, respectively, and d is the feature dimension.
[0053] Specific implementation method nine, combined with Figure 2 and Figure 6 This embodiment describes the multi-scale bidirectional information flow pyramid network proposed in step 3 of this embodiment. The optimization strategy for the fusion of multi-scale features of the model includes the following steps:
[0054] Step 4.1: Obtain the multi-resolution feature map output by the backbone network, use downsampling blocks containing pooling and convolution to reduce the dimension of the high-resolution feature map and align it, and then concatenate the channels with the low-resolution deep feature map enhanced by convolution.
[0055] Step 4.2: Process the concatenated feature map using a set of parallel multi-scale depthwise separable convolutions, and perform residual connection between the output and the concatenated feature map to obtain a focused feature map rich in contextual information.
[0056] Step 4.3: Input the focused feature map into the bidirectional fusion pyramid network. Perform bidirectional cross-fusion of cross-scale features through top-down upsampling and bottom-up downsampling operations, and output the final feature pyramid set, such as... Figure 2 As shown in MSBFPN.
[0057] Specific implementation method ten, combined with Figure 6 This embodiment describes the following steps in step 4.1: obtaining the multi-resolution feature map output by the backbone network, performing dimensionality reduction and alignment on the high-resolution feature map using downsampling blocks containing pooling and convolution, and channel concatenating it with the convolution-enhanced low-resolution deep feature map.
[0058] Step 4.1.1: First, extract multi-resolution feature maps from the backbone network (including shallow high-resolution features S3 and deep low-resolution features S4 and S5). The input features are processed using convolutional functions in the feature focusing module to enhance key features and suppress background noise interference. Specifically, for the high-resolution shallow features S3, a downsampling block (Adown) containing multiple pooling and convolutional layers is used for dimensionality reduction and alignment. This Adown module effectively reduces the spatial resolution of the feature maps to obtain a wider receptive field, while simultaneously enhancing the network's feature extraction capability by expanding feature channels and further reducing the number of model parameters. The specific calculation formula is as follows: , , ;
[0059] Step 4.1.2: Subsequently, channel concatenation technology is used to concatenate the three dimensionally aligned feature maps to construct an initial feature representation with multi-dimensional information, as shown in the following formula: .
[0060] Specific implementation method eleven, combined with Figure 6 This embodiment describes a method where step 4.2 uses a set of parallel multi-scale depthwise separable convolutions to process the concatenated feature maps, and then performs a residual connection between the output and the concatenated feature maps to obtain a focused feature map rich in contextual information. This includes the following steps:
[0061] Step 4.2.1: Drawing on the design concept of multi-core initial networks, a set of parallel multi-scale depthwise separable convolutional operations are used to process the cascaded initial feature maps to comprehensively capture contextual information at multiple scales. The calculation formula is as follows:
[0062]
[0063] Step 4.2.2: The S3, S4, and S5 feature maps, which have undergone independent convolutional enhancement processing, are fused with the output of the parallel multi-scale deep convolution path described above. This dual-path fusion mechanism fully integrates the complementary information between different processing paths, ensuring the preservation and enhancement of multi-scale features, thereby obtaining a focused feature map rich in contextual information and more discriminative. The formula is as follows: .
[0064] Specific implementation method twelve, combined with Figure 2 This embodiment describes step 4.3, where the focused feature map is input into a bidirectional fusion pyramid network. Cross-scale feature fusion is performed through top-down upsampling and bottom-up downsampling operations to output the final feature pyramid set. The steps include:
[0065] Step 4.3.1: Top-down upsampling fusion, upsampling the low-resolution, high-semantic deep features (S4, S5) to the same spatial resolution as S3, and integrating them with the high-resolution S3 features through splicing or weighted fusion.
[0066] Step 4.3.2: Bottom-up downsampling fusion. The high-resolution S3 features are downsampled through stride convolution or pooling operations and then deeply fused with the S4 and S5 features.
[0067] Specific implementation method thirteen, combined with Figure 1 and Figure 2 This embodiment describes step 5, which generates a report containing the category, location, and confidence level of the defective target based on the detection results to support power inspection decisions. The report includes the location, classification, and confidence level value of each defective target, and is ultimately output visually for operators to make judgments.
[0068] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A method for detecting damage to conductors and ground wires in concealed parts of transmission lines, characterized in that, The method includes the following steps: Step 1: Obtain the image of the target to be tested and construct a dataset of conductor and ground wire damage in concealed parts of the transmission line; Step 2: Based on a Transformer model, design a multi-scale edge-texture separation and dual-domain dynamic selection module, and reconstruct its backbone network; Step 3: Propose an intra-scale interaction module based on high- and low-frequency feature decoupling to optimize the feature interaction process; Step 4: Propose a multi-scale bidirectional information flow pyramid network and optimize the fusion strategy of multi-scale features of the model; Step 5: Output the damage location, category, and confidence level based on the detection results.
2. The method for detecting damage to conductors and ground wires in concealed parts of transmission lines according to claim 1, characterized in that, Step 2, based on the Transformer model, designs a multi-scale edge-texture separation and dual-domain dynamic selection module, and reconstructs the backbone network. The Transformer object detection model is an RT-DETR model, which includes the following steps: Step 2.1: Extract and fuse multi-scale edge features using a multi-scale edge enhancer; Step 2.2: Adaptively filter features in the spatial and frequency domains using a dual-domain selection mechanism; Step 2.3: Perform global feature fusion to complete the backbone network reshaping.
3. The method for detecting damage to conductors and ground wires in concealed parts of transmission lines according to claim 2, characterized in that, Step 2.1, which involves extracting and fusing multi-scale edge features using a multi-scale edge enhancer, includes the following steps: Step 2.1.1: Perform adaptive average pooling on the input original feature map to map it into feature maps of various sizes (3×3, 6×6, 9×9, 12×12). Step 2.1.2: The pooled feature map is passed through 1×1 and 3×3 convolutional layers for non-linear refinement, and then upsampling is used to restore the feature map to the original input resolution; Step 2.1.3: For each scale feature, perform average pooling on the original feature map to obtain a smooth feature map, and subtract the smooth feature map from the original feature map to accurately extract the high-frequency edge feature map; Step 2.1.4: After performing 1×1 convolution to enhance details on the high-frequency edge feature map, perform residual fusion with the original feature map to obtain edge enhancement features at each scale; Step 2.1.5: After restoring the edge enhancement features at each scale to the input resolution through bilinear interpolation, the features are stitched together along the channel dimension to generate multi-scale representation features.
4. The method for detecting damage to conductors and ground wires in concealed parts of transmission lines according to claim 2, characterized in that, Step 2.2, which uses a dual-domain selection mechanism to adaptively filter features in the spatial and frequency domains, includes the following steps: Step 2.2.1: Input the multi-scale representation features into the spatial selection module, and generate channel-level feature representations by sequentially passing them through pooling layers, convolutional layers, and cascaded deep convolutions. Then, modulate the representations with the input features to generate spatial selection features. Step 2.2.2: Input the spatial selection feature into the frequency selection module, generate low-frequency features through channel global average pooling, and subtract the low-frequency features from the input to obtain high-frequency features; Step 2.2.3: Multiply the high-frequency features element-wise with the input features and combine them with residual connections to generate a local attention map.
5. The method for detecting damage to conductors and ground wires in concealed parts of transmission lines according to claim 1, characterized in that, Step 3 proposes an intra-scale interaction module based on high- and low-frequency feature decoupling to optimize the feature interaction process. The intra-scale feature interaction module is reconstructed through a high- and low-frequency separation feature attention mechanism to enhance the model's ability to perceive the global structure and local features of transmission line damage images.
6. The method for detecting damage to conductors and ground wires in concealed parts of transmission lines according to claim 5, characterized in that, The steps described above for reconstructing the intra-scale feature interaction module through a high- and low-frequency separation feature attention mechanism include the following steps: Step 3.1: Perform local window self-attention calculation through high-frequency paths to extract high-frequency local fine-grained features; Step 3.2: Perform pooling dimensionality reduction and global attention calculation through low-frequency paths to extract low-frequency global context features; Step 3.3: Perform semantic aggregation and module integration of high and low frequency features.
7. The method for detecting damage to conductors and ground wires in concealed parts of transmission lines according to claim 1, characterized in that, The multi-scale bidirectional information flow pyramid network proposed in step 4, and the optimization strategy for the fusion of multi-scale features of the model, includes the following steps: Step 4.1: Obtain the multi-resolution feature map output by the backbone network, use downsampling blocks containing pooling and convolution to reduce the dimension of the high-resolution feature map and align it, and then concatenate the channels with the low-resolution deep feature map enhanced by convolution. Step 4.2: Process the concatenated feature map using a set of parallel multi-scale depthwise separable convolutions, and perform residual connection between the output and the concatenated feature map to obtain a focused feature map rich in contextual information; Step 4.3: Input the focused feature map into the bidirectional fusion pyramid network, and perform bidirectional cross-fusion of cross-scale features through top-down upsampling and bottom-up downsampling operations to output the final feature pyramid set.
8. The method for detecting damage to conductors and ground wires in concealed parts of transmission lines according to claim 1, characterized in that, The detection results of the method can output comprehensive information including the location, type, and confidence level of the target defect, which can be used for subsequent power inspection decision support.