A deep learning-based power transmission line anti-hammer patrol inspection method

Abstract

The application discloses to the technical field of power equipment intelligent operation and maintenance, in particular to a power transmission line anti-vibration hammer inspection method based on deep learning, first, build a power transmission line anti-vibration hammer inspection image dataset Dataset; Then build a hidden danger feature deep extraction unit, take any anti-vibration hammer inspection image F1 in Dataset as input, after multi-level, deep-level feature extraction, get anti-vibration hammer inspection feature map F4, anti-vibration hammer inspection feature map F7, anti-vibration hammer inspection feature map F11; After that, build a multi-level hidden danger feature fusion unit, input F4, F7 and F11 obtained from the hidden danger feature deep extraction unit into the multi-level hidden danger feature fusion unit, and carry out multi-level path feature fusion and refining operation; Based on this, the application uses hidden danger feature cross fusion module and three-branch attention enhancer module to actively amplify weak hidden danger features submerged by complex background from the hidden danger source.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power equipment, specifically a method for inspecting vibration dampers for transmission lines based on deep learning. Background Technology

[0002] Vibration dampers, as crucial power fittings in transmission lines that absorb vibration energy from transmission conductors and prevent conductor breakage, are core components ensuring the safe and stable operation of transmission lines. However, during long-term operation, vibration dampers are susceptible to various hazards due to continuous mechanical vibration, environmental corrosion, and micro-wind stress, such as displacement, hammer head detachment, strand unwinding, and surface corrosion. These hazards can lead to power grid accidents such as conductor wear, line breaks, and power outages, posing a direct threat to power grid safety. Currently, vibration damper hazard detection mainly relies on manual interpretation of images collected by drones or inspection robots. However, vibration dampers typically occupy a small area in images and are easily obstructed by vegetation, intersecting transmission conductors, complex backgrounds, and metallic reflections, making it difficult to accurately identify early hazard characteristics.

[0003] Therefore, developing an automated and intelligent method for inspecting vibration dampers, and constructing a model for perceiving potential hazards of vibration dampers in transmission lines under power grid operating scenarios, is of vital practical significance and engineering application value for ensuring the safe operation of key infrastructure such as transmission lines and improving the level of intelligent power operation and maintenance.

[0004] Currently, the detection methods used for potential hazards in vibration dampers for power transmission lines can be mainly classified into the following categories: (1) Detection methods based on traditional image processing. These methods typically employ edge detection, threshold segmentation, contour extraction, and other techniques to determine whether there are any abnormalities such as detachment or displacement based on the shape contour or pixel gradient changes of the vibration damper. This method has a relatively low computational cost, but it is highly dependent on image quality. When there are changes in lighting, background interference, or power transmission line obstructions in the inspection scene, its detection effect is easily affected.

[0005] (2) A method based on general deep learning for object detection: Faster R-CNN, YOLO series and other object detection models are used to extract features and regress bounding boxes in the inspection images of vibration dampers through convolutional neural networks. This method achieves automated detection to a certain extent, but due to the relatively fixed feature extraction and fusion mechanism and receptive field size of general models, it is difficult to deal with the problem of extremely large scale span of the features of vibration damper hazards; at the same time, when facing environmental pseudo features such as strong reflection and dirt on metal surfaces, the single decision dimension easily leads to a high rate of missed detection and false judgment.

[0006] Based on the above, the current method has the following drawbacks: 1. Existing detection models are insufficient in extracting subtle hazard features from vibration dampers, potentially leading to a high rate of missed detections for early-stage hazards. In actual power grid inspection scenarios, early-stage hazard features such as minor strand loosening of the steel strands, small cracks in the clamps, and initial surface rust spots on vibration dampers are weak and have low contrast with the surrounding complex background, making them easily confused with background noise and hindering the model from accurately extracting effective features.

[0007] 2. Existing detection models lack the ability to identify and suppress false features in complex environments, leading to frequent misjudgments. Inspection images contain interfering information such as metallic reflections, dirt adhesion, vegetation obstruction, and intersecting wires. Existing models struggle to adequately distinguish between genuine defect features and environmental false features, affecting the reliability of detection results.

[0008] 3. Existing models have poor adaptive perception of the morphology and scale of vibration damper hazards, affecting their ability to generalize detection of irregular and multi-sized vibration damper hazards. Vibration damper hazards vary greatly in form, such as the entire hammerhead falling off, significant displacement, or a single broken strand of steel wire, with extremely wide ranges in physical size and aspect ratio. Existing models typically use fixed convolution kernels or fixed feature fusion methods, making it difficult to dynamically adjust the feature extraction range according to the actual shape of the hazard target, thus limiting the model's generalization ability in complex scenarios. Summary of the Invention

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A deep learning-based method for inspecting vibration dampers on power transmission lines includes the following steps: S1, Construct a Dataset of images for the inspection of vibration dampers on power transmission lines; S2, Construct a hidden danger feature deep extraction unit, take any vibration damper inspection image F1 in the Dataset as input, and after multi-level and deep feature extraction, obtain vibration damper inspection feature map F4, vibration damper inspection feature map F7 and vibration damper inspection feature map F11; S3, construct a multi-level hidden danger feature fusion unit, input F4, F7 and F11 obtained from the hidden danger feature depth extraction unit into the multi-level hidden danger feature fusion unit, and obtain the vibration hammer inspection feature map F27, vibration hammer inspection feature map F29 and vibration hammer inspection feature map F32 through multi-level path feature fusion and refinement operation; S4. Construct a hazard feature adaptive output unit, inputting F27, F29 and F32 obtained from the multi-level hazard feature fusion unit into the hazard feature adaptive output unit to obtain a hazard detection result image containing the hazard target bounding box, category and confidence level; S5. Based on the hidden danger feature deep extraction unit, multi-level hidden danger feature fusion unit, and hidden danger feature adaptive output unit, a vibration damper hidden danger perception model is constructed. The training set is input into the vibration damper hidden danger perception model for training. The model is optimized according to the training results. Then, the validation set is used to validate the trained vibration damper hidden danger perception model. Finally, the test set is used to test the validated vibration damper hidden danger perception model. S6 applies the tested anti-vibration hammer hazard perception model to the actual power grid inspection scenario for intelligent identification and detection, and obtains hazard detection result images containing the hazard target bounding box, category, and confidence level.

[0010] As a preferred embodiment of the deep learning-based transmission line vibration damper inspection method described in this invention, the specific steps of step S2 are as follows: S21, Perform a DSConv3×3 depth-separable convolution operation on the vibration damper inspection image F1 to obtain the vibration damper inspection feature map F2; At the same time, perform a Conv3×3 convolution operation on F1 to obtain the vibration damper inspection feature map F3. Input F2 and F3 into the hidden danger feature cross-fusion module to perform a weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F4. S22, perform DSConv5×5 depth separable convolution operation on F4 to obtain the vibration damper inspection feature map F5; at the same time, perform Conv5×5 convolution operation on F4 to obtain the vibration damper inspection feature map F6. Input F5 and F6 into the hidden danger feature cross-fusion module to perform weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F7. S23, perform a DSConv7×7 depthwise separable convolution operation on F7 to obtain the vibration damper inspection feature map F8; at the same time, perform a Conv7×7 convolution operation on F7 to obtain the vibration damper inspection feature map F9. Input F8 and F9 into the hidden danger feature cross-fusion module to perform a weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F10. Input F10 into the SPPF spatial pyramid pooling module to perform a multi-scale spatial context information aggregation operation to obtain the vibration damper inspection feature map F11.

[0011] As a preferred embodiment of the deep learning-based transmission line vibration damper inspection method described in this invention, the execution process of the hidden danger feature cross-fusion module is as follows: First, the input vibration damper inspection feature maps F2 and F3 are concatenated along the channel dimension to obtain a cross-fused feature map A1. Then, A1 is convolved using a Conv3×3 method to obtain a cross-fused feature map A2. Finally, A2 is processed using parallel path processing. In path one, the cross-fusion feature map A2 is input to the first TBAE submodule to perform multi-path, deep-level hidden danger feature extraction operation to obtain the cross-fusion feature map A3. Then, A3 is subjected to average pooling operation to obtain the cross-fusion feature map A4. At the same time, A3 is subjected to max pooling operation to obtain the cross-fusion feature map A5. Then, A4 and A5 are added element by element to obtain the cross-fusion feature map A6. Finally, A6 is processed by the Sigmoid activation function to obtain the cross-fusion feature map A7. In path two, the cross-fusion feature map A2 is input to the second TBAE submodule to perform multi-path, deep-level hidden danger feature extraction operation to obtain the cross-fusion feature map A8. Then, A8 is subjected to minimum pooling operation to obtain the cross-fusion feature map A9. Subsequently, A9 is subjected to Conv1×1 convolution operation to obtain the cross-fusion feature map A10. Then, A10 is processed by the Sigmoid activation function to obtain the cross-fusion feature map A11. After parallel path processing is completed, element-wise multiplication is performed on A7 and A10 to obtain cross-fusion feature map A12; element-wise multiplication is performed on A11 and A6 to obtain cross-fusion feature map A13; finally, feature concatenation is performed on A12 and A13, that is, concatenation is performed along the feature map channel dimension to obtain vibration damper inspection feature map F4.

[0012] As a preferred embodiment of the deep learning-based transmission line vibration damper inspection method described in this invention, the specific steps of step S3 are as follows: S31, perform a Conv1×1 convolution operation on the vibration damper inspection feature map F11 to obtain the vibration damper inspection feature map F12; perform parallel processing operations on F12: First, upsampling is performed on F12 to obtain the vibration damper inspection feature map F13. Then, a Conv3×3 convolution operation is performed on F13 to obtain the vibration damper inspection feature map F14. F7 and F14 are input into the multi-dimensional hidden danger collaborative refinement module to perform feature refinement and fusion operations on the feature maps, resulting in the vibration damper inspection feature map F15. Upsampling is performed on F15 to obtain the vibration damper inspection feature map F16. F4 and F16 are input into the multi-dimensional hidden danger collaborative refinement module to perform feature refinement and fusion operations on the feature maps, resulting in the vibration damper inspection feature map F17. A Conv3×3 convolution operation is then performed on F17. The following operations are performed to obtain the vibration damper inspection feature map F18: F17 and F18 are then concatenated along the feature map channel dimension to obtain the vibration damper inspection feature map F19. A Conv3×3 convolution operation is then performed on F19 to obtain the vibration damper inspection feature map F20. Similarly, F15 and F20 are concatenated along the feature map channel dimension to obtain the vibration damper inspection feature map F21. A Conv3×3 convolution operation is then performed on F21 to obtain the vibration damper inspection feature map F22. Finally, a Conv3×3 convolution operation is performed on F22 to obtain the vibration damper inspection feature map F23. S32, perform a Conv5×5 convolution operation on F12 to obtain the vibration damper inspection feature map F24. Upsample F24 to obtain the vibration damper inspection feature map F25. Simultaneously, perform a Conv3×3 convolution operation on F12 to obtain the vibration damper inspection feature map F26. Input F25 and F26 into the multi-dimensional hidden danger collaborative refinement module to perform a fusion and refinement operation on the feature maps, obtaining the vibration damper inspection feature map F27. Upsample F27 to obtain the vibration damper inspection feature map F27. Feature map F28; input F23 and F28 into the multi-dimensional hidden danger collaborative refinement module, perform fusion refinement operation on the feature map to obtain vibration damper inspection feature map F29, perform Conv3×3 convolution operation on F29 to obtain vibration damper inspection feature map F30, perform upsampling operation on F30 to obtain vibration damper inspection feature map F31, input F19 and F31 into the multi-dimensional hidden danger collaborative refinement module, perform fusion refinement operation on the feature map to obtain vibration damper inspection feature map F32.

[0013] As a preferred embodiment of the deep learning-based transmission line vibration damper inspection method described in this invention, the execution process of the multi-dimensional hidden danger collaborative refinement module is as follows: First, a Conv3×3 convolution operation is performed on the input vibration damper inspection feature map F7 to obtain the refined hazard feature map X1. Simultaneously, a Conv3×3 convolution operation is performed on the input vibration damper inspection feature map F14 to obtain the refined hazard feature map X2. Then, X1 and X2 are added element-wise to obtain the refined hazard feature map X3. Next, a Conv3×3 convolution operation is performed on X3 to obtain the refined hazard feature map X4. Finally, parallel branch processing operations are performed on X4 in two separate branches: In branch one, X4 is input into the first MRFI submodule to perform multi-scale receptive field feature extraction to obtain the hazard refinement feature map X5; then, a channel attention mechanism is applied to X5 to perform channel-by-channel weighted adjustment to obtain the hazard refinement feature map X6; then, X1 and X6 are multiplied element-by-element to obtain the hazard refinement feature map X7. In branch two, X4 is input into the second MRFI submodule to perform multi-scale receptive field feature extraction to obtain the hazard refinement feature map X8; then, a spatial attention mechanism is applied to X8 to perform a weighted adjustment operation on each spatial position of X8 to obtain the hazard refinement feature map X9; then, X2 and X9 are multiplied element-wise to obtain the hazard refinement feature map X10. After the two branches are processed, feature concatenation is performed on X7 and X10, that is, concatenation is performed along the feature map channel dimension to obtain the hidden danger refined feature map X11. Then, X11 is processed by the Softmax activation function to obtain the hidden danger refined feature map X12. Then, parallel branching operations are performed on X12: In branch three, X12 is subjected to a DSConv3×3 depthwise separable convolution operation to obtain the hazard refinement feature map X13; then X13 is subjected to a high-average pooling operation to obtain the hazard refinement feature map X14. In branch four, a DSConv3×3 depthwise separable convolution operation is performed on X12 to obtain the hazard refinement feature map X15; then, a width average pooling operation is performed on X15 to obtain the hazard refinement feature map X16. After the branching process is completed, feature concatenation is performed on X14 and X16, that is, concatenation is performed along the feature map channel dimension to obtain the hazard refinement feature map X17; then, reweighting is performed on X17 and X11 to obtain the hazard refinement feature map X18; finally, Conv3×3 convolution is performed on X18 to obtain F15.

[0014] As a preferred embodiment of the deep learning-based transmission line vibration damper inspection method described in this invention, the specific steps of step S4 are as follows: S41, perform a Conv1×1 convolution operation on the vibration damper inspection feature map F32 output by the multi-level hidden danger feature fusion unit to obtain the vibration damper inspection feature map F33. Then, input F33 into the multi-scale hidden danger morphology adaptive module to perform a hidden danger feature adaptive focusing operation to obtain the vibration damper inspection feature map F34. Perform a feature concatenation operation between the vibration damper inspection feature map F29 output by the multi-level hidden danger feature fusion unit and F34, that is, concatenate along the feature map channel dimension to obtain the vibration damper inspection feature map F35. Perform a Conv3×3 convolution operation on F35 to obtain the vibration damper inspection feature map. Figure F36 is then input into the multi-scale hazard morphology adaptive module for adaptive focusing of hazard features, resulting in vibration damper inspection feature map F37. The vibration damper inspection feature map F27 output by the multi-level hazard feature fusion unit is then combined with F37 by a feature stitching operation, i.e., stitching along the feature map channel dimension, to obtain vibration damper inspection feature map F38. Conv5×5 convolution operation is then performed on F38 to obtain vibration damper inspection feature map F39. Subsequently, F39 is input into the multi-scale hazard morphology adaptive module for adaptive focusing of hazard features, resulting in vibration damper inspection feature map F40. S42, input F34 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R1 containing information on the detection of vibration damper hazards; input F37 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R2 containing information on the detection of vibration damper hazards; input F40 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R3 containing information on the detection of vibration damper hazards; then, use the non-maximum suppression algorithm (NMS) to filter out the predicted bounding boxes of the hazard targets from R1, R2, and R3, determine the location of the hazard through the predicted bounding boxes, and obtain a hazard detection result image containing the hazard target bounding boxes, category, and confidence level based on the hazard location.

[0015] As a preferred embodiment of the deep learning-based transmission line vibration damper inspection method described in this invention, the execution process of the multi-scale hazard morphology adaptive module is as follows: Parallel branching processing is performed on the input vibration damper inspection feature map F33: In branch one, a Conv3×3 convolution operation is performed on F33 to obtain the hazard morphology feature map N1; then, a max pooling operation is performed on N1 to obtain the hazard morphology feature map N2; then, N1 and N2 are added element by element to obtain the hazard morphology feature map N3; then, N3 is processed by the Sigmoid activation function to obtain the hazard morphology feature map N4. In branch two, a Conv7×7 convolution operation is performed on F33 to obtain the hazard morphology feature map N5; then, a minimum pooling operation is performed on N5 to obtain the hazard morphology feature map N6; then, N5 and N6 are subtracted element by element to obtain the hazard morphology feature map N7; then, N7 is processed by the Sigmoid activation function to obtain the hazard morphology feature map N8. In branch three, a Conv5×5 convolution operation is performed on F33 to obtain the hazard morphology feature map N9; then, an efficient channel attention mechanism is applied to N9 to perform channel feature adaptive weighting operation on the feature map to obtain the hazard morphology feature map N10; next, N3, N7 and N10 are concatenated, that is, concatenated along the channel dimension of the feature map to obtain the hazard morphology feature map N11; then, a Conv1×1 convolution operation is performed on N11 to obtain the hazard morphology feature map N12; After the parallel branch processing is completed, N4, N8 and N12 are multiplied element by element to obtain the vibration damper inspection feature map F34.

[0016] Compared with existing technologies: 1. To address the problem of difficulty in identifying subtle hazard features due to long-distance shooting or complex background interference in vibration damper inspection scenarios, this method uses a hazard feature deep extraction unit, a hazard feature cross-fusion (HCFF) module and a three-branch attention enhancement (TBAE) sub-module to actively amplify subtle hazard features submerged by complex backgrounds from the source of the hazard. At the same time, this unit combines with the SPPF module to aggregate global multi-scale contextual information, solving the problem of missed detection caused by low contrast and complex background in traditional detection models, and improving the detection accuracy of the vibration damper hazard perception model. 2. To address the interference from false features such as strong reflections from metal surfaces, complex intersecting power transmission lines, and vegetation shadows during vibration damper hazard detection, this method utilizes a multi-level hazard feature fusion unit. Within a multi-level path interaction structure, it employs a multi-dimensional hazard collaborative refinement (MDHCR) module and a multi-receptive field interaction (MRFI) sub-module. Through a dual attention mechanism of channels and space, and a coordinate perception reweighting mechanism of bidirectional average pooling of height and width, it actively suppresses false hazard features, thereby improving the accuracy and reliability of the vibration damper hazard perception model. 3. To address the problem of hazard detection caused by the large differences in the shape and size of hazard in vibration dampers, this method designs an adaptive output unit for hazard features. Through layer-by-layer feature stitching and multi-scale convolution operations, multi-scale feature extraction and channel integration are performed on hazard feature maps of different sizes. Furthermore, the multi-scale hazard morphology adaptive (MSHA) module is used to perform adaptive focusing operations on hazard features, which solves the problem that traditional detection models are difficult to adapt to different hazard shapes under a fixed single receptive field. This improves the generalization recognition capability of the vibration damper hazard perception model for real hazards of multiple scales and shapes. Attached Figure Description

[0017] Figure 1 This is a flowchart of the technical solution of the present invention; Figure 2 This is a structural diagram of the HCFF module of the present invention; Figure 3 This is a structural diagram of the TBAE submodule of the present invention; Figure 4 This is a structural diagram of the MDHCR module of the present invention; Figure 5 This is a structural diagram of the MRFI submodule of the present invention; Figure 6 This is a structural diagram of the MSHA module of the present invention; Figure 7 This is a structural diagram of the vibration damper hazard perception model of the present invention; Figure 8 This is an image depicting a power grid inspection scene according to the present invention. Figure 9 This is an image showing the hazard detection results of the present invention. Detailed Implementation

[0018] 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.

[0019] This invention provides a deep learning-based method for inspecting vibration dampers on power transmission lines. The technical solution process is as follows: Figure 1 As shown.

[0020] S1, Construct a dataset of images for the inspection of vibration dampers on power transmission lines.

[0021] S11 utilizes a drone equipped with a high-definition visible light camera to conduct multi-angle and multi-distance inspections and photographs of the vibration dampers along the power transmission corridor in complex field environments, acquiring inspection images of the vibration dampers.

[0022] S12. Using an image annotation tool, the potential hazard areas in each vibration damper inspection image are annotated, and then all images are used to construct an original vibration damper inspection image dataset.

[0023] S13 involves preprocessing all original vibration damper inspection images in the original vibration damper inspection image dataset: First, image standardization is performed on the original vibration damper inspection images to eliminate scale and viewpoint inconsistencies caused by differences in drone shooting height and angle. Then, various data augmentation strategies such as image cropping, rotation, flipping, and contrast transformation are used to enhance the image sample diversity and robustness under different lighting and background conditions, resulting in enhanced vibration damper inspection images. All enhanced vibration damper inspection images are then constructed into a transmission line vibration damper inspection image dataset.

[0024] S14. Divide the Dataset into training set, validation set and test set according to a certain ratio.

[0025] S2, Construct a hidden danger feature deep extraction unit, take any vibration damper inspection image F1 in the Dataset as input, and after multi-level and deep feature extraction, obtain vibration damper inspection feature map F4, vibration damper inspection feature map F7 and vibration damper inspection feature map F11; The hazard feature deep extraction unit employs a design combining parallel extraction paths and a hazard feature cross-fusion (HCFF) module to progressively extract hazard features. This allows the vibration damper hazard perception model to extract rich hazard feature information. Furthermore, during the step-by-step downsampling operation, the HCFF module amplifies weak hazard signals, actively enhancing low-contrast hazard features such as slight strand breakage and surface rust in the vibration damper's steel strands. This improves the sensitivity of the vibration damper hazard perception model for detecting early-stage, weak hazards. Its structure is as follows: Figure 7 Deep extraction unit for hidden danger characteristics.

[0026] The specific process of obtaining the vibration damper inspection feature maps F4, F7, and F11 through the hidden danger feature depth extraction unit is as follows: S21, perform a DSConv3×3 depth-separable convolution operation on the vibration damper inspection image F1 to obtain the vibration damper inspection feature map F2; at the same time, perform a Conv3×3 convolution operation on F1 to obtain the vibration damper inspection feature map F3. Input F2 and F3 into the hidden danger feature cross-fusion (HCFF) module to perform a weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F4.

[0027] S22, perform a DSConv5×5 depth-separable convolution operation on F4 to obtain the vibration damper inspection feature map F5; at the same time, perform a Conv5×5 convolution operation on F4 to obtain the vibration damper inspection feature map F6. Input F5 and F6 into the hidden danger feature cross-fusion module to perform a weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F7.

[0028] S23, perform a DSConv7×7 depthwise separable convolution operation on F7 to obtain the vibration damper inspection feature map F8; at the same time, perform a Conv7×7 convolution operation on F7 to obtain the vibration damper inspection feature map F9. Input F8 and F9 into the hidden danger feature cross-fusion module to perform a weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F10. Input F10 into the SPPF spatial pyramid pooling module to perform a multi-scale spatial context information aggregation operation to obtain the vibration damper inspection feature map F11.

[0029] This invention designs a Hazard Cross-Feature Fusion (HCFF) module. This module employs a dual-path cross-interaction structure, extracting hazard features through a combination of max pooling and average pooling, and capturing weak hazard features using min pooling. This enables the vibration damper hazard perception model to amplify and enhance weak hazard features and actively suppress environmental noise, thereby improving the detection accuracy of the vibration damper hazard perception model for weak hazard features. The HCFF module structure diagram is shown below. Figure 2 As shown.

[0030] The execution process of the HCFF module is described below, with inputs F2 and F3 and output F4 used for the flow description: First, the input vibration damper inspection feature maps F2 and F3 are concatenated (concat) along the channel dimension to obtain a cross-fused feature map A1. Then, A1 is convolved (Conv3×3) to obtain a cross-fused feature map A2. Finally, A2 is processed by parallel path processing. In path one, the cross-fusion feature map A2 is input to the first TBAE submodule to perform multi-path, deep-level hidden danger feature extraction operation to obtain the cross-fusion feature map A3. Then, A3 is subjected to average pooling (AvgPool) operation to obtain the cross-fusion feature map A4. At the same time, A3 is subjected to max pooling (MaxPool) operation to obtain the cross-fusion feature map A5. Then, A4 and A5 are subjected to element-wise addition (Add) operation to obtain the cross-fusion feature map A6. Finally, A6 is processed by the Sigmoid activation function to obtain the cross-fusion feature map A7. In path two, the cross-fusion feature map A2 is input into the second TBAE submodule to perform multi-path, deep-level hidden danger feature extraction operation to obtain the cross-fusion feature map A8. Then, A8 is subjected to a minimum pooling operation to obtain the cross-fusion feature map A9. Subsequently, A9 is subjected to a Conv1×1 convolution operation to obtain the cross-fusion feature map A10. Then, A10 is processed by the Sigmoid activation function to obtain the cross-fusion feature map A11. After parallel path processing is completed, element-wise multiplication is performed on A7 and A10 to obtain cross-fusion feature map A12; element-wise multiplication is performed on A11 and A6 to obtain cross-fusion feature map A13; finally, feature concat operation is performed on A12 and A13, that is, concatenation is performed along the feature map channel dimension to obtain vibration damper inspection feature map F4.

[0031] The first TBAE submodule and the second TBAE submodule have the same structure.

[0032] This invention designs a Triple-Branch Attention Enhancement (TBAE) submodule. This submodule employs a three-way parallel collaboration of attention feature enhancement, local smoothing feature extraction, and multi-scale deep convolutional feature interaction. This enables the vibration damper hazard perception model to perform cross-dimensional feature amplification and enhancement on subtle hazard features present in complex field inspection backgrounds, thereby improving the detection accuracy of the vibration damper hazard perception model for multi-scale subtle hazard features. The TBAE submodule structure diagram is shown in Figure 3.

[0033] The execution process of the TBAE submodule is described below, with input A2 and output A3 used for the process description: First, perform parallel branching processing on the input cross-fusion feature map A2: In branch one, a Conv1×1 convolution operation is performed on the cross-fusion feature map A2 to obtain the hazard enhancement feature map B1. Then, a channel attention mechanism is applied to B1, and a channel-by-channel weighted adjustment operation is performed on each channel of B1 to obtain the hazard enhancement feature map B2. B1 and B2 are multiplied element-by-element to obtain the hazard enhancement feature map B3. Next, a spatial attention mechanism is applied to B3, and a weighted adjustment operation is performed on each spatial position of B3 to obtain the hazard enhancement feature map B4. Then, B3 and B4 are multiplied element-by-element to obtain the hazard enhancement feature map B5. In branch two, a Conv3×3 convolution operation is performed on the cross-fusion feature map A2 to obtain the hazard enhancement feature map B6. Then, batch normalization (BN) is performed on B6 to obtain the hazard enhancement feature map B7. Subsequently, ReLU activation function is applied to B7 to obtain the hazard enhancement feature map B8. Feature concatenation (Concat) is performed on B5 and B8, that is, concatenating along the channel dimension of the feature maps to obtain the hazard enhancement feature map B9. Then, a Conv1×1 convolution operation is performed on B9 to obtain the hazard enhancement feature map B10. In branch three, a Conv5×5 convolution operation is performed on the cross-fusion feature map A2 to obtain the hazard enhancement feature map B11. Then, B11 is processed in parallel: a DWConv3×3 depthwise convolution operation is performed on B11 to obtain the hazard enhancement feature map B12, followed by a DWConv1×1 depthwise convolution operation on B12 to obtain the hazard enhancement feature map B13; simultaneously, a DWConv5×5 depthwise convolution operation is performed on B11 to obtain the hazard enhancement feature map B14, followed by Softmax activation function processing on B14 to obtain the hazard enhancement feature map B15; after the parallel processing of B11 is completed, element-wise multiplication is performed on B13 and B15 to obtain the hazard enhancement feature map B16. After the parallel branching process is completed, B10 and B16 are added element by element to obtain the cross-fusion feature map A3.

[0034] S3. Construct a multi-level hidden danger feature fusion unit. Input F4, F7 and F11 obtained from the hidden danger feature depth extraction unit into the multi-level hidden danger feature fusion unit. Through feature fusion and refinement operations of multi-level paths, obtain vibration hammer inspection feature map F27, vibration hammer inspection feature map F29 and vibration hammer inspection feature map F32.

[0035] The multi-level hazard feature fusion unit uses the Multi-Dimensional Hazard Collaborative Refinement (MDHCR) module to perform feature fusion and refinement on multi-level hazard features in multi-level paths. This enables the vibration damper hazard perception model to use the MDHCR module to filter environmental pseudo-features during the detection process, actively suppressing them and thus improving the anti-interference capability and stability of the vibration damper hazard perception model in field inspection environments. Its structure is as follows: Figure 7 Multi-level hidden danger characteristics integration unit.

[0036] The specific process of obtaining the vibration damper inspection feature maps F27, F29, and F32 through the multi-level hidden danger feature fusion unit is as follows: S31, perform a Conv1×1 convolution operation on the vibration damper inspection feature map F11 to obtain the vibration damper inspection feature map F12; perform parallel processing operations on F12: First, upsampling is performed on F12 to obtain the vibration damper inspection feature map F13. Then, a Conv3×3 convolution operation is performed on F13 to obtain the vibration damper inspection feature map F14. F7 and F14 are input into the multi-dimensional hidden danger collaborative refinement module to perform feature refinement and fusion operations on the feature maps, resulting in the vibration damper inspection feature map F15. Upsampling is performed on F15 to obtain the vibration damper inspection feature map F16. F4 and F16 are input into the multi-dimensional hidden danger collaborative refinement module to perform feature refinement and fusion operations on the feature maps, resulting in the vibration damper inspection feature map F17. A Conv3×3 convolution operation is then performed on F17. The following operations are performed to obtain the vibration damper inspection feature map F18: F17 and F18 are then concatenated along the feature map channel dimension to obtain the vibration damper inspection feature map F19. A Conv3×3 convolution operation is then performed on F19 to obtain the vibration damper inspection feature map F20. Similarly, F15 and F20 are concatenated along the feature map channel dimension to obtain the vibration damper inspection feature map F21. A Conv3×3 convolution operation is then performed on F21 to obtain the vibration damper inspection feature map F22. Finally, a Conv3×3 convolution operation is performed on F22 to obtain the vibration damper inspection feature map F23.

[0037] S32, perform a Conv5×5 convolution operation on F12 to obtain the vibration damper inspection feature map F24. Then, perform an upsampling operation on F24 to obtain the vibration damper inspection feature map F25. Simultaneously, perform a Conv3×3 convolution operation on F12 to obtain the vibration damper inspection feature map F26. Input F25 and F26 into the MDHCR module, perform a fusion and refinement operation on the feature maps to obtain the vibration damper inspection feature map F27, and then perform an upsampling operation on F27. Obtain the vibration damper inspection feature map F28; input F23 and F28 into the MDHCR module, perform a fusion and refinement operation on the feature maps to obtain the vibration damper inspection feature map F29; perform a Conv3×3 convolution operation on F29 to obtain the vibration damper inspection feature map F30; perform an upsampling operation on F30 to obtain the vibration damper inspection feature map F31; input F19 and F31 into the MDHCR module, perform a fusion and refinement operation on the feature maps to obtain the vibration damper inspection feature map F32.

[0038] This invention designs a Multi-Dimensional Hazard Collaborative Refinement (MDHCR) module. This module employs a dual attention mechanism of channel and space, as well as a coordinate perception reweighting mechanism of bidirectional average pooling of height and width. This enables the vibration damper hazard perception model to actively suppress environmental pseudo-features such as strong reflections on the metal surface of the vibration damper and complex crossing transmission lines in the field inspection background, thereby improving the detection accuracy of the vibration damper hazard perception model in complex field scenarios. The MDHCR module structure diagram is shown below. Figure 4 As shown: The execution process of the MDHCR module is described below, with inputs F7 and F14 and output F15 used for process description: First, a Conv3×3 convolution operation is performed on the input vibration damper inspection feature map F7 to obtain the refined hazard feature map X1. Simultaneously, a Conv3×3 convolution operation is performed on the input vibration damper inspection feature map F14 to obtain the refined hazard feature map X2. Next, X1 and X2 are added element-wise to obtain the refined hazard feature map X3. Then, a Conv3×3 convolution operation is performed on X3 to obtain the refined hazard feature map X4. Finally, parallel branch processing operations are performed on X4 in two separate branches: In branch one, X4 is input into the first multi-receptive field interaction (MRFI) submodule to perform multi-scale receptive field feature extraction, resulting in a refined feature map X5 for potential hazards. Then, channel attention is applied to X5 to perform a channel-by-channel weighted adjustment operation, resulting in a refined feature map X6 for potential hazards. Finally, element-by-element multiplication is performed on X1 and X6 to obtain a refined feature map X7 for potential hazards. In branch two, X4 is input into the second multi-receptive field interaction (MRFI) submodule to perform multi-scale receptive field feature extraction, resulting in a refined feature map X8 for potential hazards. Then, spatial attention is applied to X8 to perform weighted adjustment operations on each spatial location of X8, resulting in a refined feature map X9 for potential hazards. Finally, element-wise multiplication is performed on X2 and X9 to obtain a refined feature map X10 for potential hazards. After the two branches are processed, feature concatenation (Concat) is performed on X7 and X10, that is, concatenation is performed along the channel dimension of the feature map to obtain the hazard refinement feature map X11. Then, the Softmax activation function is applied to X11 to obtain the hazard refinement feature map X12. Then, parallel branching operations are performed on X12: In branch three, X12 is subjected to a DSConv3×3 depthwise separable convolution operation to obtain the hazard refinement feature map X13; then X13 is subjected to a high-average pooling (H-AvgPool) operation to obtain the hazard refinement feature map X14.

[0039] In branch four, a DSConv3×3 depthwise separable convolution operation is performed on X12 to obtain the hazard refinement feature map X15; then, a width average pooling (W-AvgPool) operation is performed on X15 to obtain the hazard refinement feature map X16.

[0040] After the branch processing is completed, a feature concatenation operation is performed on X14 and X16, that is, concatenation is performed along the channel dimension of the feature map to obtain the hazard refinement feature map X17; then, a re-weighting operation is performed on X17 and X11 to obtain the hazard refinement feature map X18; finally, a Conv3×3 convolution operation is performed on X18 to obtain F15.

[0041] The first MRFI submodule and the second MRFI submodule have the same structure.

[0042] This invention designs a Multi-Receptive Field Interaction (MRFI) submodule. This submodule employs multi-branch parallel convolutional layers to extract hazard features from different receptive fields. Furthermore, it combines an adaptive fusion mechanism to fuse the hazard features extracted from different branches. This enables the vibration damper hazard perception model to accurately capture multi-scale hazards with large scale spans and diverse shapes in complex field inspection scenarios. It achieves adaptive focusing on hazard features at different scales, thereby improving the vibration damper hazard perception model's ability to extract and detect multi-scale real-world hazard features in complex field scenes. The MRFI submodule structure diagram is shown below. Figure 5 As shown.

[0043] The execution process of the Multi-Receptive Field Interaction (MRFI) submodule is described below, with input X4 and output X5 used for flow description: First, the input hazard refinement feature map X4 is processed in parallel branching mode: In branch one, a Conv1×1 convolution operation is performed on the hazard refinement feature map X4 to obtain the multi-receptive field hazard feature map H1; then, batch normalization (BN) is performed on H1 to obtain the multi-receptive field hazard feature map H2; then, the Sigmoid activation function is applied to H2 to obtain the multi-receptive field hazard feature map H3; finally, H1 and H3 are multiplied element-wise to obtain the multi-receptive field hazard feature map H4. In branch two, a Conv3×3 convolution operation is performed on the refined feature map X4 to obtain the multi-receptive-field feature map H5; then, a feature concatenation operation is performed on H3 and H5, that is, concatenating along the channel dimension of the feature maps to obtain the multi-receptive-field feature map H6; then, a Conv3×3 convolution operation is performed on H6 to obtain the multi-receptive-field feature map H7; finally, an adaptive fusion operation is performed on H4 and H7 to obtain the multi-receptive-field feature map H8. In branch three, a Conv5×5 convolution operation is performed on the hazard refinement feature map X4 to obtain the multi-receptive field hazard feature map H9; then, a Conv5×5 convolution operation is performed on H9 again to obtain the multi-receptive field hazard feature map H10; next, an upsampling operation is performed on H10 to obtain the multi-receptive field hazard feature map H11; and an element-wise addition operation is performed on H11 and H8 to obtain the multi-receptive field hazard feature map H12. In branch four, a Conv7×7 convolution operation is performed on the hazard refinement feature map X4 to obtain the multi-receptive field hazard feature map H13; then, H11 and H13 are concatenated (concat) along the feature map channel dimension to obtain the multi-receptive field hazard feature map H14; then, a Conv1×1 convolution operation is performed on H14 to obtain the multi-receptive field hazard feature map H15; then, H12 and H15 are multiplied element-wise to obtain the multi-receptive field hazard feature map H16. After the parallel branch processing is completed, H8 and H16 are concatenated, that is, concatenated along the channel dimension of the feature map to obtain the multi-receptive field hazard feature map H17; finally, H17 is convolved with Conv1×1 to obtain the hazard refined feature map X5.

[0044] S4. Construct a hazard feature adaptive output unit. Input F27, F29 and F32 obtained from the multi-level hazard feature fusion unit into the hazard feature adaptive output unit to obtain a hazard detection result image containing the hazard target bounding box, category and confidence level.

[0045] The adaptive output unit for hazard features, through cross-scale feature concatenation, enables the vibration damper hazard perception model to obtain feature maps adapted to hazard targets of different sizes through layer-by-layer feature stitching and multi-scale convolution operations during the hazard detection output stage. Furthermore, before the feature maps are fed into the detection head, the multi-scale hazard morphology adaptation (MSHA) module performs adaptive operations on the hazard features, ensuring accurate capture of real hazard features with large scale spans and varied shapes in vibration damper inspection images. Combined with non-maximum suppression (NMS) algorithm for deduplication, this enhances the generalization ability of the vibration damper hazard perception model for identifying vibration damper hazards in complex field scenarios. Its structure is as follows: Figure 7 Adaptive output unit for hidden danger characteristics.

[0046] The specific process of obtaining the hazard detection result image through the hazard feature adaptive output unit is as follows: S41, perform a Conv1×1 convolution operation on the vibration damper inspection feature map F32 output by the multi-level hidden danger feature fusion unit to obtain the vibration damper inspection feature map F33. Then, input F33 into the multi-scale hidden danger morphology adaptive (MSHA) module to perform a hidden danger feature adaptive focusing operation to obtain the vibration damper inspection feature map F34. Perform a feature concatenation operation (Concat) on the vibration damper inspection feature map F29 output by the multi-level hidden danger feature fusion unit and F34, that is, concatenate along the feature map channel dimension to obtain the vibration damper inspection feature map F35. Perform a Conv3×3 convolution operation on F35 to obtain the vibration damper inspection feature map F35. Feature map F36 is inspected and then input into the multi-scale hazard morphology adaptive module for adaptive focusing of hazard features to obtain vibration damper inspection feature map F37. Vibration damper inspection feature map F27 output by the multi-level hazard feature fusion unit is concatted with F37, i.e., concatenated along the feature map channel dimension to obtain vibration damper inspection feature map F38. Conv5×5 convolution operation is performed on F38 to obtain vibration damper inspection feature map F39. Then, F39 is input into the multi-scale hazard morphology adaptive module for adaptive focusing of hazard features to obtain vibration damper inspection feature map F40. S42, input F34 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R1 containing information on the detection of vibration damper hazards; input F37 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R2 containing information on the detection of vibration damper hazards; input F40 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R3 containing information on the detection of vibration damper hazards; in each prediction tensor, each row corresponds to a prediction result, which includes the predicted bounding box of the hazard target, the category, and the confidence level; then, the non-maximum suppression algorithm (NMS) is used to select the predicted bounding boxes of hazard targets with high confidence from R1, R2, and R3, and the location of the hazard is determined by the predicted bounding boxes of the hazard targets. Based on the location of the hazard, a hazard detection result image containing the hazard target bounding box, the category, and the confidence level is obtained.

[0047] This invention designs a Multi-Scale Hazard Adaptive (MSHA) module. This module employs multi-scale convolutional parallel feature extraction and a dynamic collaborative weighted design using efficient channel attention (ECA). This enables the vibration damper hazard perception model to accurately capture spatial local details and deeply align cross-channel global context information of vibration damper hazards with large scale spans and diverse morphological features in complex field inspection backgrounds before hazard detection output. This achieves adaptive focusing and amplification of key hazard features of different morphologies and active suppression of environmental background noise, thereby improving the localization accuracy of the vibration damper hazard perception model for real hazard features. The MSHA module structure diagram is shown below. Figure 6 As shown.

[0048] The execution process of the MSHA module is described below, with input F33 and output F34 used for the process description: Parallel branching processing is performed on the input vibration damper inspection feature map F33: In branch one, a Conv3×3 convolution operation is performed on F33 to obtain the hazard morphology feature map N1; then, a max pooling operation is performed on N1 to obtain the hazard morphology feature map N2; next, N1 and N2 are added element-wise to obtain the hazard morphology feature map N3; then, N3 is processed by the Sigmoid activation function to obtain the hazard morphology feature map N4. In branch two, a Conv7×7 convolution operation is performed on F33 to obtain the hazard morphology feature map N5; then, a min-pooling operation is performed on N5 to obtain the hazard morphology feature map N6; then, N5 and N6 are subtracted element-wise to obtain the hazard morphology feature map N7; then, N7 is processed by the Sigmoid activation function to obtain the hazard morphology feature map N8. In branch three, a Conv5×5 convolution operation is performed on F33 to obtain the hazard morphology feature map N9; then, an efficient channel attention (ECA) mechanism is applied to N9 to perform channel feature adaptive weighting operation on the feature map to obtain the hazard morphology feature map N10; next, N3, N7 and N10 are concatenated along the channel dimension of the feature map to obtain the hazard morphology feature map N11; then, a Conv1×1 convolution operation is performed on N11 to obtain the hazard morphology feature map N12. After the parallel branch processing is completed, N4, N8 and N12 are multiplied element by element to obtain the vibration damper inspection feature map F34.

[0049] S5, based on a hazard feature deep extraction unit, a multi-level hazard feature fusion unit, and a hazard feature adaptive output unit, a vibration damper hazard perception model is constructed. The training set is input into the vibration damper hazard perception model for training, and the model is optimized based on the training results. Then, the trained vibration damper hazard perception model is validated using a validation set. Finally, the validated vibration damper hazard perception model is tested using a test set. The structure diagram of the vibration damper hazard perception model is shown below. Figure 7 As shown.

[0050] S6 applies the tested anti-vibration hammer hazard perception model to the actual power grid inspection scenario for intelligent identification and detection, and obtains hazard detection result images containing the hazard target bounding box, category, and confidence level.

[0051] S6 Example: Inputting the power grid inspection scene image of the anti-vibration hammer hazard perception model, as shown in the example... Figure 8 As shown.

[0052] Will Figure 8 As input to the vibration damper hazard perception model, the output is an image of the hazard detection result, including the target bounding box, category, and confidence level. Figure 9 As shown.

[0053] in: Hazard type: Hammer head detachment, confidence level: 0.93 Hazard type: Location displacement; Confidence level: 0.87 Hazard type: Hammer head detached, confidence level: 0.95 Hazard type: normal, confidence level: 0.95.

[0054] 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 inspecting vibration dampers on transmission lines based on deep learning, characterized in that, Includes the following steps: S1, Construct a Dataset of images for the inspection of vibration dampers on power transmission lines; S2, Construct a hidden danger feature deep extraction unit, take any vibration damper inspection image F1 in the Dataset as input, and after multi-level and deep feature extraction, obtain vibration damper inspection feature map F4, vibration damper inspection feature map F7 and vibration damper inspection feature map F11; S3, construct a multi-level hidden danger feature fusion unit, input F4, F7 and F11 obtained from the hidden danger feature depth extraction unit into the multi-level hidden danger feature fusion unit, and obtain the vibration hammer inspection feature map F27, vibration hammer inspection feature map F29 and vibration hammer inspection feature map F32 through multi-level path feature fusion and refinement operation; S4. Construct a hazard feature adaptive output unit, inputting F27, F29 and F32 obtained from the multi-level hazard feature fusion unit into the hazard feature adaptive output unit to obtain a hazard detection result image containing the hazard target bounding box, category and confidence level; S5. Based on the hidden danger feature deep extraction unit, multi-level hidden danger feature fusion unit, and hidden danger feature adaptive output unit, a vibration damper hidden danger perception model is constructed. The training set is input into the vibration damper hidden danger perception model for training. The model is optimized according to the training results. Then, the validation set is used to validate the trained vibration damper hidden danger perception model. Finally, the test set is used to test the validated vibration damper hidden danger perception model. S6 applies the tested anti-vibration hammer hazard perception model to the actual power grid inspection scenario for intelligent identification and detection, and obtains hazard detection result images containing the hazard target bounding box, category, and confidence level.

2. The method for inspecting vibration dampers for transmission lines based on deep learning according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, Perform a DSConv3×3 depth-separable convolution operation on the vibration damper inspection image F1 to obtain the vibration damper inspection feature map F2; At the same time, perform a Conv3×3 convolution operation on F1 to obtain the vibration damper inspection feature map F3. Input F2 and F3 into the hidden danger feature cross-fusion module to perform a weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F4. S22, perform DSConv5×5 depth separable convolution operation on F4 to obtain the vibration damper inspection feature map F5; at the same time, perform Conv5×5 convolution operation on F4 to obtain the vibration damper inspection feature map F6. Input F5 and F6 into the hidden danger feature cross-fusion module to perform weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F7. S23, perform a DSConv7×7 depthwise separable convolution operation on F7 to obtain the vibration damper inspection feature map F8; at the same time, perform a Conv7×7 convolution operation on F7 to obtain the vibration damper inspection feature map F9. Input F8 and F9 into the hidden danger feature cross-fusion module to perform a weak hidden danger feature enhancement operation to obtain the vibration damper inspection feature map F10. Input F10 into the SPPF spatial pyramid pooling module to perform a multi-scale spatial context information aggregation operation to obtain the vibration damper inspection feature map F11.

3. The method for inspecting vibration dampers for transmission lines based on deep learning according to claim 2, characterized in that, The execution process of the hidden danger feature cross-fusion module is as follows: First, the input vibration damper inspection feature maps F2 and F3 are concatenated along the channel dimension to obtain a cross-fused feature map A1. Then, A1 is convolved using a Conv3×3 method to obtain a cross-fused feature map A2. Finally, A2 is processed using parallel path processing. In path one, the cross-fusion feature map A2 is input to the first TBAE submodule to perform multi-path, deep-level hidden danger feature extraction operation to obtain the cross-fusion feature map A3. Then, A3 is subjected to average pooling operation to obtain the cross-fusion feature map A4. At the same time, A3 is subjected to max pooling operation to obtain the cross-fusion feature map A5. Then, A4 and A5 are added element by element to obtain the cross-fusion feature map A6. Finally, A6 is processed by the Sigmoid activation function to obtain the cross-fusion feature map A7. In path two, the cross-fusion feature map A2 is input to the second TBAE submodule to perform multi-path, deep-level hidden danger feature extraction operation to obtain the cross-fusion feature map A8. Then, A8 is subjected to minimum pooling operation to obtain the cross-fusion feature map A9. Subsequently, A9 is subjected to Conv1×1 convolution operation to obtain the cross-fusion feature map A10. Then, A10 is processed by the Sigmoid activation function to obtain the cross-fusion feature map A11. After parallel path processing is completed, element-wise multiplication is performed on A7 and A10 to obtain cross-fusion feature map A12; element-wise multiplication is performed on A11 and A6 to obtain cross-fusion feature map A13; finally, feature concatenation is performed on A12 and A13, that is, concatenation is performed along the feature map channel dimension to obtain vibration damper inspection feature map F4.

4. The method for inspecting vibration dampers for transmission lines based on deep learning according to claim 1, characterized in that, The specific steps of S3 are as follows: S31, perform a Conv1×1 convolution operation on the vibration damper inspection feature map F11 to obtain the vibration damper inspection feature map F12; perform parallel processing operations on F12: First, upsampling is performed on F12 to obtain the vibration damper inspection feature map F13. Then, a Conv3×3 convolution operation is performed on F13 to obtain the vibration damper inspection feature map F14. F7 and F14 are input into the multi-dimensional hidden danger collaborative refinement module to perform feature refinement and fusion operations on the feature maps, resulting in the vibration damper inspection feature map F15. Upsampling is performed on F15 to obtain the vibration damper inspection feature map F16. F4 and F16 are input into the multi-dimensional hidden danger collaborative refinement module to perform feature refinement and fusion operations on the feature maps, resulting in the vibration damper inspection feature map F17. A Conv3×3 convolution operation is then performed on F17. The following operations are performed to obtain the vibration damper inspection feature map F18: F17 and F18 are then concatenated along the feature map channel dimension to obtain the vibration damper inspection feature map F19. A Conv3×3 convolution operation is then performed on F19 to obtain the vibration damper inspection feature map F20. Similarly, F15 and F20 are concatenated along the feature map channel dimension to obtain the vibration damper inspection feature map F21. A Conv3×3 convolution operation is then performed on F21 to obtain the vibration damper inspection feature map F22. Finally, a Conv3×3 convolution operation is performed on F22 to obtain the vibration damper inspection feature map F23. S32, perform a Conv5×5 convolution operation on F12 to obtain the vibration damper inspection feature map F24. Upsample F24 to obtain the vibration damper inspection feature map F25. Simultaneously, perform a Conv3×3 convolution operation on F12 to obtain the vibration damper inspection feature map F26. Input F25 and F26 into the multi-dimensional hidden danger collaborative refinement module to perform a fusion and refinement operation on the feature maps, obtaining the vibration damper inspection feature map F27. Upsample F27 to obtain the vibration damper inspection feature map F27. Feature map F28; input F23 and F28 into the multi-dimensional hidden danger collaborative refinement module, perform fusion refinement operation on the feature map to obtain vibration damper inspection feature map F29, perform Conv3×3 convolution operation on F29 to obtain vibration damper inspection feature map F30, perform upsampling operation on F30 to obtain vibration damper inspection feature map F31, input F19 and F31 into the multi-dimensional hidden danger collaborative refinement module, perform fusion refinement operation on the feature map to obtain vibration damper inspection feature map F32.

5. The method for inspecting vibration dampers for transmission lines based on deep learning according to claim 4, characterized in that, The execution process of the multi-dimensional hidden danger collaborative refinement module is as follows: First, a Conv3×3 convolution operation is performed on the input vibration damper inspection feature map F7 to obtain the refined hazard feature map X1. Simultaneously, a Conv3×3 convolution operation is performed on the input vibration damper inspection feature map F14 to obtain the refined hazard feature map X2. Then, X1 and X2 are added element-wise to obtain the refined hazard feature map X3. Next, a Conv3×3 convolution operation is performed on X3 to obtain the refined hazard feature map X4. Finally, parallel branch processing operations are performed on X4 in two separate branches: In branch one, X4 is input into the first MRFI submodule to perform multi-scale receptive field feature extraction to obtain the hazard refinement feature map X5; then, a channel attention mechanism is applied to X5 to perform channel-by-channel weighted adjustment to obtain the hazard refinement feature map X6; then, X1 and X6 are multiplied element-by-element to obtain the hazard refinement feature map X7. In branch two, X4 is input into the second MRFI submodule to perform multi-scale receptive field feature extraction to obtain the hazard refinement feature map X8; then, a spatial attention mechanism is applied to X8 to perform a weighted adjustment operation on each spatial position of X8 to obtain the hazard refinement feature map X9; then, X2 and X9 are multiplied element-wise to obtain the hazard refinement feature map X10. After the two branches are processed, feature concatenation is performed on X7 and X10, that is, concatenation is performed along the feature map channel dimension to obtain the hidden danger refined feature map X11. Then, X11 is processed by the Softmax activation function to obtain the hidden danger refined feature map X12. Then, parallel branching operations are performed on X12: In branch three, X12 is subjected to a DSConv3×3 depthwise separable convolution operation to obtain the hazard refinement feature map X13; then X13 is subjected to a high-average pooling operation to obtain the hazard refinement feature map X14. In branch four, a DSConv3×3 depthwise separable convolution operation is performed on X12 to obtain the hazard refinement feature map X15; then, a width average pooling operation is performed on X15 to obtain the hazard refinement feature map X16. After the branching process is completed, feature concatenation is performed on X14 and X16, that is, concatenation is performed along the feature map channel dimension to obtain the hazard refinement feature map X17; then, reweighting is performed on X17 and X11 to obtain the hazard refinement feature map X18; finally, Conv3×3 convolution is performed on X18 to obtain F15.

6. The method for inspecting vibration dampers of transmission lines based on deep learning according to claim 1, characterized in that, The specific steps of S4 are as follows: S41, perform a Conv1×1 convolution operation on the vibration damper inspection feature map F32 output by the multi-level hidden danger feature fusion unit to obtain the vibration damper inspection feature map F33. Then, input F33 into the multi-scale hidden danger morphology adaptive module to perform a hidden danger feature adaptive focusing operation to obtain the vibration damper inspection feature map F34. Perform a feature concatenation operation between the vibration damper inspection feature map F29 output by the multi-level hidden danger feature fusion unit and F34, that is, concatenate along the feature map channel dimension to obtain the vibration damper inspection feature map F35. Perform a Conv3×3 convolution operation on F35 to obtain the vibration damper inspection feature map. Figure F36 is then input into the multi-scale hazard morphology adaptive module for adaptive focusing of hazard features, resulting in vibration damper inspection feature map F37. The vibration damper inspection feature map F27 output by the multi-level hazard feature fusion unit is then combined with F37 by a feature stitching operation, i.e., stitching along the feature map channel dimension, to obtain vibration damper inspection feature map F38. Conv5×5 convolution operation is then performed on F38 to obtain vibration damper inspection feature map F39. Subsequently, F39 is input into the multi-scale hazard morphology adaptive module for adaptive focusing of hazard features, resulting in vibration damper inspection feature map F40. S42, input F34 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R1 containing information on the detection of vibration damper hazards; input F37 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R2 containing information on the detection of vibration damper hazards; input F40 into the target detection head to perform a prediction operation on the feature map, obtaining a prediction tensor R3 containing information on the detection of vibration damper hazards; then, use the non-maximum suppression algorithm (NMS) to filter out the predicted bounding boxes of the hazard targets from R1, R2, and R3, determine the location of the hazard through the predicted bounding boxes, and obtain a hazard detection result image containing the hazard target bounding boxes, category, and confidence level based on the hazard location.

7. The method for inspecting vibration dampers for transmission lines based on deep learning according to claim 6, characterized in that, The execution process of the multi-scale hidden danger morphology adaptive module is as follows: Parallel branching processing is performed on the input vibration damper inspection feature map F33: In branch one, a Conv3×3 convolution operation is performed on F33 to obtain the hazard morphology feature map N1; then, a max pooling operation is performed on N1 to obtain the hazard morphology feature map N2; then, N1 and N2 are added element by element to obtain the hazard morphology feature map N3; then, N3 is processed by the Sigmoid activation function to obtain the hazard morphology feature map N4. In branch two, a Conv7×7 convolution operation is performed on F33 to obtain the hazard morphology feature map N5; then, a minimum pooling operation is performed on N5 to obtain the hazard morphology feature map N6; then, N5 and N6 are subtracted element by element to obtain the hazard morphology feature map N7; then, N7 is processed by the Sigmoid activation function to obtain the hazard morphology feature map N8. In branch three, a Conv5×5 convolution operation is performed on F33 to obtain the hazard morphology feature map N9; Then, an efficient channel attention mechanism is applied to N9 to perform adaptive weighting of channel features on the feature map, resulting in the hazard morphology feature map N10. Next, N3, N7 and N10 are concatenated, that is, concatenated along the channel dimension of the feature map, to obtain the hazard morphology feature map N11. Then, N11 is convolved with Conv1×1 to obtain the hazard morphology feature map N12. After the parallel branch processing is completed, N4, N8 and N12 are multiplied element by element to obtain the vibration damper inspection feature map F34.

Images