A transmission line suspension clamp defect detection method based on unmanned aerial vehicle inspection

Abstract

The application discloses the technical field of power transmission line intelligent inspection and computer vision, and particularly relates to a power transmission line suspension clamp defect detection method based on unmanned aerial vehicle inspection, which comprises the following steps: firstly, constructing a strong anti-interference multi-scale deformation perception feature extraction unit; then, constructing a cross-layer re-calibration guided feature fusion unit; subsequently, constructing a dynamic rotation decoupling prediction unit; finally, constructing a strong anti-interference deformation adaptive fine-grained suspension clamp defect detection model; based on the above, after the input features are encoded by the strong anti-interference multi-scale deformation perception feature extraction unit, the parallel heterogeneous receptive field distortion adaptive module (HRDA) dynamically adheres to the component boundary through cross-dimension pooling to adapt to the perspective deformation and retain the tiny abnormal features; then, the multi-scale receptive field cascade background suppression module (CRBS) enlarges the receptive field through a hollow convolution and captures the structural features, so that the defect omission rate is reduced.

Description

Technical Field

[0001] This invention relates to the fields of intelligent inspection of power transmission lines and computer vision technology, specifically a method for detecting defects in suspension clamps of power transmission lines based on UAV inspection. Background Technology

[0002] Suspension clamps, as key connecting hardware in ultra-high voltage and extra-high voltage transmission lines, reliably suspend conductors from the lower end of the insulator string on the tower and bear vertical loads. They are core components ensuring the safe and stable operation of transmission lines. However, during long-term operation, suspension clamps are susceptible to defects such as micro-cracks and missing cotter pins due to continuous mechanical vertical loads, environmental corrosion, and severe wind deflection. These defects can lead to power grid accidents such as conductor detachment and line tripping, posing a direct threat to power grid safety.

[0003] Current methods for detecting defects in suspension clamps mainly rely on manual inspection or traditional visual target detection algorithms to judge inspection images captured by drones and other equipment. However, the drawbacks of this approach are becoming increasingly apparent. For example, suspension clamps themselves have a complex configuration involving multiple components such as U-shaped hanging rings and connecting plates, and are easily obscured by insulator strings, dense hardware, and complex background vegetation. In addition, the severe perspective distortion and highly isomorphic metal texture interference caused by drone flight make minor defects such as cracks and missing pins extremely inconspicuous in inspection images. Furthermore, traditional visual models are limited by rigid convolutional sampling, coarse pyramid fusion that swallows weak details, and the lack of cross-dimensional feature focusing mechanisms, making it difficult to overcome these extreme scale differences and visual interference, resulting in a high risk of missed detections and false alarms.

[0004] Therefore, there is an urgent need for a novel visual perception and defect detection method that integrates an adaptive deformation feature extraction mechanism, a high-resolution multi-scale feature decoupling architecture, and cross-dimensional spatial attention reconstruction. This method can achieve adaptive and accurate positioning of complex and heavily overlapping suspension clamps in complex high-altitude environments, and solve the problem of capturing fine-grained micro-defects in conventional algorithms under conditions of multiple interwoven fittings, severe perspective distortion, and strong textured background interference.

[0005] Currently, the detection methods applied to suspension clamp defects in transmission lines can be mainly divided into the following categories: (1) Detection method based on traditional digital image processing: Classical computer vision algorithms are used, and edge detection operators and threshold segmentation techniques are employed to extract the geometric contours or edge pixel gradient abruptness features of the suspension clamp to determine whether macroscopic displacement or structural damage has occurred. This method has a relatively low computational cost, but its robustness and generalization ability are extremely weak in actual field environments of power grids, such as imaging scenarios with drastic changes in lighting and complex background textures, especially when the background contains intricate insulator skirts or conductor textures.

[0006] (2) A method for object detection based on general deep learning: mainstream object detection network models such as Faster R-CNN and YOLO series are used to extract features and regress bounding boxes in the inspection images of suspended clamps 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, and the use of rigid convolutional sampling, it is difficult to deal with the problem of the large scale span between macroscopic components of suspended clamps and millimeter-level micro-defects; at the same time, when faced with environmental pseudo-features such as strong reflection of metal surfaces and isomorphic textures of hardware, the single decision dimension easily leads to a high rate of missed detection and false judgment.

[0007] (3) Detection method based on multi-sensor and spatial modeling: Combining high-definition visible light images with point cloud data from sensors such as lidar, three-dimensional spatial joint modeling is used to accurately measure the spatial pose and surface deformation of the suspension clamp, thereby comprehensively diagnosing whether it has micro-cracks or missing parts. Although this method has high spatial positioning accuracy, it relies on expensive lidar and heavy sensor equipment, the computational cost of data acquisition and subsequent processing is high, and the deployment is difficult, making it difficult to meet the business needs of modern power grids to conduct large-scale, routine, and high-frequency inspections using lightweight UAVs.

[0008] Based on the above, existing technologies have drawbacks in the following dimensions: 1. Insufficient Capability of Feature Extraction Backbone Networks for Capturing Minor Defect Features and Adapting to Deformation: Existing object detection models generally rely on conventional rigid convolutions with fixed square receptive fields during the feature extraction stage. In UAV high-altitude inspection scenarios, due to real-time changes in the shooting perspective and the influence of wind, suspended clamps exhibit severe geometric perspective distortion in two-dimensional images. Simultaneously, the clamps themselves contain nested structures of multiple components such as U-shaped hanging rings, connecting plates, and the hull, resulting in highly irregular physical boundaries. Traditional rigid convolutions cannot break through grid constraints to dynamically fit these complex non-rigid edges for sampling, leading to a large amount of background clutter being included in the receptive field, severely reducing the signal-to-noise ratio of feature extraction. More importantly, during continuous downsampling, the backbone network is highly susceptible to losing high-frequency, high-resolution sub-pixel-level abnormal features such as minute cracks and millimeter-level missing cotter pins in the deep network, resulting in an extremely high false negative rate for minor defects.

[0009] 2. Insufficient cross-layer feature fusion leads to inadequate decoupling and recombination capabilities for extreme scale differences: There exists an extremely asymmetric polarization scale difference between the overall macroscopic framework of the dangling clamp and its local micro-defects. Existing feature fusion networks, such as FPN or PANet structures, typically employ simple nearest-neighbor interpolation and coarse concat fusion methods when handling multi-scale features across layers. This shallow linear fusion, lacking deep decoupling, not only easily leads to severe jagged distortion of the geometric edges of high-resolution micro-defects during cross-layer transmission, but also causes extremely weak defect anomaly responses at lower layers to be directly diluted by the strong macroscopic semantic features transmitted from deeper networks. Due to the lack of pixel-level recalibration and fine recombination mechanisms for weak features, existing models struggle to simultaneously address macroscopic localization and microscopic flaw detection within a unified feature space.

[0010] 3. The real-world environment of power transmission lines is complex and multifaceted. The metal material of suspension clamps exhibits high visual isomorphism with the closely connected insulator strings, vibration dampers, and conductors in the background. The grayscale gradients created by extremely minor cracks or missing components in the suspension clamp itself can easily be confused with the normal casting textures, rust patches, and pseudo-edges generated by complex light and shadow overlaps in the background. Existing models are typically limited by single feature representations and lack effective guidance from cross-dimensional spatial attention mechanisms, making it impossible to dynamically suppress environmental pseudo-features and cleanly separate genuine intrinsic defects of the clamps from the high-frequency, complex, isomorphic metal background noise. This leads to the model easily misjudging normal textures as cracks under extreme conditions of low contrast or multiple overlapping and obstructing hardware, generating numerous false alarms and failing to meet the high reliability requirements of intelligent closed-loop inspection of the power grid. Summary of the Invention

[0011] This invention addresses the problem of accurate positioning and fine-grained defect detection of suspension clamps in complex high-altitude environments. It proposes a method for detecting defects in transmission line suspension clamps based on UAV inspection, which extracts the complex and variable physical boundaries and defect features of suspension clamps, comprehensively improving the accuracy of target perception and robustness against isomorphic background interference in complex power grid high-altitude inspection environments.

[0012] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for detecting defects in suspension clamps of power transmission lines based on unmanned aerial vehicle (UAV) inspection includes the following steps: S1, Construct a dataset of images of suspension clamps for power transmission lines; S2, construct a strong anti-disturbance multi-scale deformation sensing feature extraction unit, input the image F1 of any transmission line suspension clamp to the strong anti-disturbance multi-scale deformation sensing feature extraction unit, perform feature extraction, and output the fine-grained defect feature map F12, the fine-grained defect feature map F13, and the fine-grained defect feature map F14 of the directional decoupled suspension clamp; S3, construct a cross-level recalibration guided feature fusion unit, input F12, F13, and F14 into the cross-level recalibration guided feature fusion unit, perform feature fusion, and output the fine-grained defect feature map of the directional decoupling suspension clamp F27, the fine-grained defect feature map of the directional decoupling suspension clamp F28, and the fine-grained defect feature map of the directional decoupling suspension clamp F29. S4, construct a dynamic rotational decoupling prediction unit, input F27, F28, and F29 into the dynamic rotational decoupling prediction unit, perform multi-scale target regression, classification prediction, draw directional detection boxes and confidence scores on F1, and output the hanging clamp defect detection map P; S5. Construct a strong disturbance rejection deformation adaptive fine-grained suspension clamp defect detection model. Use the transmission line suspension clamp image dataset to train and update the parameters of the strong disturbance rejection deformation adaptive fine-grained suspension clamp defect detection model to obtain the optimized strong disturbance rejection deformation adaptive fine-grained suspension clamp defect detection model. S6 applies the optimized strong anti-disturbance deformation adaptive fine-grained suspension clamp defect detection model to the actual UAV transmission line inspection operation scenario: taking the transmission line image containing suspension clamps captured by the UAV as input, it realizes the location and detection of suspension clamp defects, and outputs the suspension clamp defect detection results.

[0013] As a preferred embodiment of the method for detecting defects in suspension clamps of transmission lines based on UAV inspection as described in this invention, the specific construction process of the highly disturbance-resistant multi-scale deformation sensing feature extraction unit is as follows: S21, Input the image F1 of any transmission line suspension clamp into a standard convolutional submodule to obtain the fine-grained defect feature map F2 of the directional decoupled suspension clamp; In step S22, input F2 into the first HRDA module. Utilizing its internal three-branch parallel structure, perform multi-level cross-dimensional feature extraction to extract the boundary deformation features of the complex components of the drooping clamp, resulting in the fine-grained defect feature map F3 of the directional decoupled drooping clamp. Input F2 into the first CRBS module. Utilize its internal cascaded dilated convolution to expand the receptive field and combine it with dual-pooling spatial attention to suppress complex backgrounds and capture the macroscopic contextual structural features of the drooping clamp, resulting in the fine-grained defect feature map F4 of the directional decoupled drooping clamp. Perform a Concat operation on F3 and F4, concatenating them along the channel dimension to fuse the multi-scale deformation and macroscopic structural information of the drooping clamp, resulting in the fine-grained defect feature map F5 of the directional decoupled drooping clamp. Perform a Conv1×1 convolution operation on F5 to achieve channel dimensionality reduction and fusion, resulting in the fine-grained defect feature map F12 of the directional decoupled drooping clamp. S23, input F3 into the second HRDA module to further adaptively extract deeper features of the droop clamp and its defects, resulting in the fine-grained defect feature map F6 of the directional decoupled droop clamp; input F4 into the second CRBS module to remove the confusing features between the normal metal texture and the defect fracture surface of the droop clamp, enhancing its anti-interference ability against complex backgrounds, resulting in the fine-grained defect feature map F7 of the directional decoupled droop clamp; perform a Concat operation on F6 and F7, splicing them along the channel dimension to obtain the fine-grained defect feature map F8 of the directional decoupled droop clamp; perform a Conv1×1 convolution operation on F8 to achieve channel dimensionality reduction and fusion, resulting in the fine-grained defect feature map F13 of the directional decoupled droop clamp; In step S24, input F6 into the third HRDA module for final deformation adaptive extraction, extracting weak overhang clamp defect features to obtain the fine-grained defect feature map F9 of the directional decoupled overhang clamp; input F7 into the third CRBS module to complete the extraction of strong semantic features of the overhang clamp and the removal of complex backgrounds, to obtain the fine-grained defect feature map F10 of the directional decoupled overhang clamp; perform a Concat operation on F9 and F10, concatenating them along the channel dimension to achieve efficient aggregation of deep high semantic features, to obtain the fine-grained defect feature map F11 of the directional decoupled overhang clamp; perform a Conv1×1 convolution operation on F11 to achieve cross-channel information integration, and output the fine-grained defect feature map F14 of the directional decoupled overhang clamp.

[0014] As a preferred embodiment of the method for detecting defects in suspension clamps of transmission lines based on UAV inspection as described in this invention, the construction and execution process of the HRDA module is as follows: A1. Perform Conv3×3 convolution operation on the adaptive feature map X1 of the suspended clamp distortion to achieve preliminary feature extraction and downsampling, and output the adaptive feature map X2 of the suspended clamp distortion. A2, inputs X2 in parallel into three branches for feature extraction: In branch one: Perform a Conv3×3 convolution operation on X2 to capture the basic local geometric edges and deformation features of the hanging clamp, and output the hanging clamp distortion adaptive feature map X7. In branch two: First, a Conv3×3 convolution operation is performed on X2 to output the drooping clamp distortion adaptive feature map X3; then, X3 is input into a dilated convolutional layer with a dilation rate of 3 to output the drooping clamp distortion adaptive feature map X4; a Concat operation is performed on X3 and X4, concatenating them along the channel dimension to obtain the drooping clamp distortion adaptive feature map X5; finally, a Conv1×1 convolution operation is performed on X5 to achieve the initial fusion and dimensionality reduction aggregation of multi-scale heterogeneous features under different receptive fields, and the drooping clamp distortion adaptive feature map X6 is output. In branch three: First, X2 is simultaneously subjected to horizontal and vertical global average pooling operations, and one-dimensional feature encoding is performed along different spatial dimensions to extract the structural edge distribution features of the hanging clamp in the horizontal and vertical directions, outputting the hanging clamp distortion adaptive feature map X9 and X10 respectively; then, X9 is subjected to dimensional rearrangement operation to obtain the hanging clamp distortion adaptive feature map X11; X10 and X11 are concatted along the channel dimension to obtain the hanging clamp distortion adaptive feature map X11. The drooping clamp distortion adaptive feature map X12 is generated. A Conv1×1 convolution operation is performed on X12 to achieve feature dimension fusion and dimensionality reduction, resulting in the drooping clamp distortion adaptive feature map X13. Finally, X13 is non-linearly normalized using the Sigmoid activation function to generate a cross-dimensional deformation boundary response weight map H1 that significantly enhances the drooping clamp and its defect features. Simultaneously, X9 is non-linearly normalized using the Sigmoid activation function to generate a cross-dimensional deformation boundary response weight map H2. A3. First, X7 and X6 are added element-wise to achieve deep synergy and fusion of basic deformation features and multi-scale heterogeneous receptive field features, resulting in the suspended clamp distortion adaptive feature map X8. Then, based on tensor broadcasting, X8 and H1 are multiplied element-wise for preliminary spatial positioning guidance, outputting the suspended clamp distortion adaptive feature map X14. Finally, based on tensor broadcasting, X14 and H2 are multiplied element-wise to suppress complex environmental background noise, outputting the suspended clamp distortion adaptive feature map X15.

[0015] As a preferred embodiment of the method for detecting defects in suspension clamps of transmission lines based on UAV inspection as described in this invention, the construction and execution process of the CRBS module is as follows: B1 inputs the multi-receptive-field collaboratively purified feature map Y1 into two branches in parallel, performing multi-scale macroscopic feature extraction and generating weights to suppress spatial background, respectively. The specific construction process of the two branches is as follows: In branch one: Y1 is input into a dilated convolutional layer with a dilation rate of 3 to initially expand the receptive field without sacrificing spatial resolution, outputting a multi-receptive field collaboratively refined feature map Y2; Y2 is input into a dilated convolutional layer with a dilation rate of 5 to output a multi-receptive field collaboratively refined feature map Y3; Y3 is input into a dilated convolutional layer with a dilation rate of 7 to further expand the receptive field, outputting a multi-receptive field collaboratively refined feature map Y4; a Concat operation is performed on Y2, Y3, and Y4, concatenating them along the channel dimension to obtain a multi-receptive field collaboratively refined feature map Y5; a Conv1×1 convolution operation is performed on Y5 to output a multi-receptive field collaboratively refined feature map Y6; In branch two: Channel average pooling and channel max pooling operations are performed on Y1 in parallel, and spatial feature compression is performed along the channel dimension to capture the overall representation characteristics of the global background and local edge details, outputting multi-receptive field co-purified feature maps Y7 and Y8 respectively; Y7 and Y8 are concatted along the channel dimension to obtain multi-receptive field co-purified feature map Y9; Y9 is convolved with Conv7×7 to output multi-receptive field co-purified feature map Y10; Y10 is non-linearly normalized using the Sigmoid activation function to generate a cascaded background suppression weight map W1 that has been stripped of interference from similar metallic materials and textured backgrounds; B2 performs element-wise multiplication of Y6 and W1 to achieve adaptive weighting of multi-scale macro features, suppress interference from similar metal materials and textured backgrounds, and outputs multi-receptive field collaborative purified feature map Y11; Conv3×3 convolution operation is performed on Y11 to achieve local spatial smoothing and nonlinear reconstruction of the weighted purified features, and outputs multi-receptive field collaborative purified feature map Y12.

[0016] As a preferred embodiment of the method for detecting defects in suspension clamps of transmission lines based on UAV inspection according to the present invention, the specific construction process of the cross-level recalibration guided feature fusion unit is as follows: S31, input F12 into the Conv 1×1 convolutional layer for channel feature alignment to obtain the fine-grained defect feature map F16 of the directional decoupling overhang clamp; perform downsampling on F16 to compress and align the spatial feature scale to obtain the fine-grained defect feature map F17 of the directional decoupling overhang clamp; simultaneously, input F12 into Conv A 3×3 convolutional layer is used to perform spatial feature fusion to obtain the fine-grained defect feature map F15 of the directional decoupling overhang clamp. A weighted fusion operation is then performed on F12 and F15 to reconstruct the shallow high-resolution geometric details, resulting in the fine-grained defect feature map F23 of the directional decoupling overhang clamp. F23 is then input into the second MDFG module, where its internal mechanisms dynamically amplify the weights of fine-grained anomalous responses, outputting the fine-grained defect feature map F24 of the directional decoupling overhang clamp. F24 is then input into the cross-stage local bottleneck submodule C2f for information extraction and channel information fusion, outputting the fine-grained defect feature map F27 of the directional decoupling overhang clamp. S32, input F14 into the Conv 3×3 convolutional layer to obtain the fine-grained defect feature map F18 of the directional decoupling overhang clamp; perform upsampling operation on F18 to realize the spatial scale magnification and obtain the fine-grained defect feature map F20 of the directional decoupling overhang clamp. S33, perform a weighted fusion operation on F13, F17, and F20 to avoid the jagged distortion caused by traditional coarse channel splicing, and obtain the fine-grained defect feature map F21 of the directional decoupled suspension clamp; input F21 into the first MDFG module to obtain the fine-grained defect feature map F22 of the directional decoupled suspension clamp. S34, perform a weighted fusion operation on F20, F21, and F22 to complete feature recombination and obtain the fine-grained defect feature map F25 of the directional decoupling overhang clamp; input F25 into the cross-stage local bottleneck submodule C2f, aggregate contextual semantic information and channel information to fuse and output the fine-grained defect feature map F28 of the directional decoupling overhang clamp. In step S35, F18 is input into the Conv 1×1 convolutional layer to restore the channels and obtain the fine-grained defect feature map F19 of the directional decoupled hanging clamp. F14 and F19 are weighted and fused to enhance and reconstruct the deep features of the hanging clamp and obtain the fine-grained defect feature map F26 of the directional decoupled hanging clamp. F26 is input into the cross-stage local bottleneck submodule C2f to refine the deep feature information and fuse the channel information and obtain the fine-grained defect feature map F29 of the directional decoupled hanging clamp.

[0017] As a preferred embodiment of the method for detecting defects in suspension clamps of power transmission lines based on UAV inspection as described in this invention, the construction and execution process of the MDFG module is as follows: C1 inputs the multi-scale recalibrated feature map Z1 into two branches, branch one and branch two, in parallel for feature fusion. The specific construction process of the two branches is as follows: In branch one: Conv3×3 convolution operation is performed on Z1 to capture high-resolution geometric features of small defect edges, and multi-scale recalibrated feature map Z2 is output; Z2 is input into a dilated convolutional layer with an expansion rate of 3 to expand the local receptive field without losing spatial resolution, and multi-scale recalibrated feature map Z3 is output; Z2 and Z3 are concat operation and concatenated along the channel dimension to achieve the fusion of basic local features and expanded receptive field features, and multi-scale recalibrated feature map Z4 is obtained. In branch two: A Conv5×5 convolution operation is performed on Z1 to capture low-resolution geometric features of the defect edges, resulting in a multi-scale recalibrated feature map Z5; Z5 is input into a dilated convolutional layer with an expansion rate of 5 to capture fine-grained contextual structural features around the defect with a larger receptive field, resulting in a multi-scale recalibrated feature map Z6; A Concat operation is performed on Z5 and Z6, concatenating them along the channel dimension to achieve feature fusion at different expansion scales, resulting in a multi-scale recalibrated feature map Z7; C2 performs element-wise addition of Z4 and Z7 to achieve deep fusion of the dual-branch multi-scale expansion features, resulting in a multi-scale recalibration feature map Z8. Max pooling is then applied to Z8 to extract the most significant drooping clamps and their defect features in the spatial dimension, yielding a multi-scale recalibration feature map Z9. A Conv7×7 convolution operation is then performed on Z9 to achieve spatial feature smoothing and depth extraction under a large receptive field, outputting a multi-scale recalibration feature map Z10. Finally, Z10 is non-linearly normalized using the Sigmoid activation function to generate a feature recalibration guiding weight map O1 with dynamically amplified subtle defect features. C3 performs element-wise multiplication of Z1 and O1 to obtain the multi-scale recalibrated feature map Z11.

[0018] As a preferred embodiment of the method for detecting defects in suspension clamps of transmission lines based on UAV inspection according to the present invention, the specific construction process of the dynamic rotation decoupling prediction unit is as follows: S41, input F27, F28, and F29 in parallel into the three directional bounding box detection heads, adaptively fit any tilt angle of the suspended clamp and its local defect area under the UAV inspection view, and output the initial orientation prediction results R1, R2, and R3 of the suspended clamp respectively. S42 performs weighted frame fusion and rotation soft nonmaximum suppression operations on R1, R2, and R3 to suppress redundant false alarms under the conditions of hardware overlap and light and shadow distortion, realizes deep decoupling and collaborative optimization of the prediction frame of macro components of the suspension clamp and millimeter-level micro defects, and outputs the final orientation prediction result R4 of the suspension clamp. S43, based on R4, draw the directional detection box and confidence score on F1 to obtain the directional detection box containing the outline of the suspension clamp and its confidence score, and the directional detection box containing the defect of the suspension clamp and its confidence score, which is the suspension clamp defect detection map P.

[0019] Compared with existing technologies: To address the issue of insufficient ability of feature extraction backbone networks to capture minute defect features and adapt to deformation, this invention designs a robust multi-scale deformation-aware feature extraction unit. After the input features are encoded, the parallel heterogeneous receptive field distortion adaptive module HRDA dynamically fits the component boundaries through cross-dimensional pooling to adapt to perspective deformation and retain minute abnormal features. Then, the multi-scale receptive field cascaded background suppression module CRBS expands the receptive field and captures structural features through dilated convolution, thereby reducing the defect false detection rate. To address the problem of insufficient cross-layer feature fusion leading to inadequate decoupling and recombination capabilities for extreme scale differences, this invention designs a cross-layer recalibration-guided feature fusion unit. This unit constructs a multi-directional interlaced structure to converge and align features from different stages and provide bidirectional feedback, avoiding jagged distortion. Simultaneously, a multi-scale hole feature recalibration guidance module (MDFG) is introduced to dynamically adjust the abnormal response weights for pixel-level recalibration before fusion, thereby achieving decoupling for target size differences. To address the issue of insufficient anti-interference capability of features against complex isomorphic texture backgrounds, this invention combines a multi-scale receptive field cascaded background suppression module (CRBS), a multi-scale hole feature recalibration guidance module (MDFG), and a dynamic rotation decoupling prediction unit to jointly eliminate clutter. During the feature processing stage, similar metallic clutter from the surrounding area is filtered out. In the output stage, the dynamic rotation decoupling prediction unit adaptively wraps the defect region, supplemented by weighted box fusion and rotation soft nonmaximum suppression operations, to solve the false alarm problem under component overlap and lighting conditions, thereby improving detection robustness. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall framework of the present invention; Figure 2 This is a flowchart of the technical solution of the present invention; Figure 3 This is a structural diagram of the strong anti-disturbance multi-scale deformation sensing feature extraction unit of the present invention; Figure 4 This is a structural diagram of the HRDA module of the present invention; Figure 5This is a structural diagram of the CRBS module of the present invention; Figure 6 This is a structural diagram of the cross-level recalibration guidance feature fusion unit of the present invention; Figure 7 This is a structural diagram of the MDFG module of the present invention; Figure 8 This is a structural diagram of the dynamic rotational decoupling prediction unit of the present invention; Figure 9 This is a diagram of the defect detection model for the highly disturbance-resistant deformation adaptive fine-grained suspension clamp of the present invention. Detailed Implementation

[0021] 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.

[0022] This invention provides a method for detecting defects in suspension clamps of power transmission lines based on unmanned aerial vehicle (UAV) inspection. Please refer to [link / reference]. Figures 1-9 This includes the following steps: S1, Construct a dataset of images of suspension clamps for power transmission lines.

[0023] Drones equipped with high-definition imaging equipment were used to conduct inspections along the power transmission corridor, capturing high-definition images of suspension clamps from multiple angles and under various operating conditions. Image annotation tools were then used to locate and annotate the suspension clamps and their defective areas, such as cracks and missing cotter pins, in the images using rotated bounding boxes, thus constructing a dataset of images of suspension clamps for power transmission lines.

[0024] The high-definition images of the suspended clamps are uniformly cropped and scaled to a fixed resolution size. Then, data augmentation operations are performed on the images. Data augmentation strategies include, but are not limited to, random rotation, multi-angle tilting, horizontal and vertical flipping, adaptive adjustment of contrast and brightness, injection of random noise, and simulated stain occlusion, in order to expand the scale of training samples and improve the data diversity and generalization ability of the model under complex perspective distortion.

[0025] The preprocessed and data-enhanced image dataset of transmission line suspension clamps was proportionally divided into training, validation, and test subsets.

[0026] S2, construct a strong anti-disturbance multi-scale deformation sensing feature extraction unit, input the image F1 of any transmission line suspension clamp into the strong anti-disturbance multi-scale deformation sensing feature extraction unit, perform feature extraction, and output the fine-grained defect feature map F12, the fine-grained defect feature map F13, and the fine-grained defect feature map F14 of the directional decoupled suspension clamp.

[0027] The main function of the robust anti-interference multi-scale deformation sensing feature extraction unit is to adaptively capture the macroscopic topology and fine boundaries of nested suspension clamps in complex high-altitude inspection scenarios through the parallel collaboration of the HRDA and CRBS modules, overcoming the problem of missing minor defects caused by UAV perspective distortion. Simultaneously, this unit utilizes cross-dimensional spatial perception and adaptive background suppression to amplify the location of suspension clamps and the features of weak cracks and pin holes, suppressing complex background interference. This addresses the issues of high overlap between the metal texture of the suspension clamps and overlapping hardware, and the easy confusion between minor cracks and missing cotter pins in the clamp body and the normal casting texture and light and shadow edges of the hardware. This facilitates high-precision and robust detection of suspension clamps and their fine-grained defects. The structural diagram of the robust anti-interference multi-scale deformation sensing feature extraction unit is shown below. Figure 3 As shown.

[0028] The specific construction process of the robust multi-scale deformation sensing feature extraction unit is as follows: An image F1 of any transmission line suspension clamp is input into a standard convolutional submodule CBS. The CBS submodule is composed of convolutional layers, batch normalization layers and SiLU activation functions to achieve nonlinear extraction of initial features and high-dimensional channel projection mapping, thereby obtaining a fine-grained defect feature map F2 of the directional decoupled suspension clamp. F2 is input into the first HRDA module, which utilizes its internal three-branch parallel structure to perform multi-level cross-dimensional feature extraction operations, extracting the boundary deformation features of the complex components of the drooping clamp, resulting in the fine-grained defect feature map F3 of the directional decoupled drooping clamp; F2 is input into the first CRBS module, which utilizes its internal cascaded dilated convolution to expand the receptive field and combines it with a dual-pooling spatial attention method to suppress complex backgrounds and capture the macroscopic contextual structural features of the drooping clamp, resulting in the fine-grained defect feature map F4 of the directional decoupled drooping clamp; F3 and F4 are concatted along the channel dimension to fuse the multi-scale deformation and macroscopic structural information of the drooping clamp, resulting in the fine-grained defect feature map F5 of the directional decoupled drooping clamp; F5 is conv1×1 convolution to achieve channel dimensionality reduction and fusion, resulting in the fine-grained defect feature map F12 of the directional decoupled drooping clamp; Inputting F3 into the second HRDA module further adaptively extracts deeper features of the droop clamps and their defects, resulting in the fine-grained defect feature map F6 of the directional decoupled droop clamps. Inputting F4 into the second CRBS module removes the confusing features between the normal metal texture and the defect fracture surface of the droop clamps, enhancing the anti-interference ability against complex backgrounds, resulting in the fine-grained defect feature map F7 of the directional decoupled droop clamps. Performing a Concat operation on F6 and F7, splicing them along the channel dimension, results in the fine-grained defect feature map F8 of the directional decoupled droop clamps. Performing a Conv1×1 convolution operation on F8 achieves channel dimensionality reduction and fusion, resulting in the fine-grained defect feature map F13 of the directional decoupled droop clamps.

[0029] Input F6 into the third HRDA module for final deformation adaptive extraction, extracting weak overhang clamp defect features to obtain the fine-grained defect feature map F9 of the directional decoupled overhang clamp; input F7 into the third CRBS module to complete the extraction of strong semantic features of the overhang clamp and the removal of complex backgrounds to obtain the fine-grained defect feature map F10 of the directional decoupled overhang clamp; perform a Concat operation on F9 and F10, concatenating them along the channel dimension to achieve efficient aggregation of deep high semantic features, to obtain the fine-grained defect feature map F11 of the directional decoupled overhang clamp; perform a Conv1×1 convolution operation on F11 to achieve cross-channel information integration, and output the fine-grained defect feature map F14 of the directional decoupled overhang clamp.

[0030] This invention designs a Parallel Heterogeneous Receptive Field Distortion Adaptive Module (HRDA). The HRDA module aims to address the problem that, under the complex perspective of UAVs, the perspective distortion of hanging clamps caused by the shooting angle limits the feature extraction of conventional defect detection models, leading to the easy omission of small defects.

[0031] Conventional convolutional networks use a fixed grid for feature extraction, which is ill-suited to the severe perspective distortion caused by the shooting angle of suspended clamps during high-altitude inspections. Therefore, the HRDA module does not employ a rigid feature extraction method, but instead uses a three-branch structure and a cross-dimensional scanning approach. This method breaks the limitations of a square receptive field, adaptively and closely conforming to the complex and realistic contours of metal components. Simultaneously, through deep fusion of the extracted features, the module can suppress complex background interference, focusing on features such as suspended clamps and their minute cracks and missing cotter pins, thereby improving the detection rate of fine-grained defects. The structure diagram of the HRDA module is shown below. Figure 4 As shown The construction and execution process of the HRDA module is as follows: Perform a Conv3×3 convolution operation on the suspended clamp distortion adaptive feature map X1 to achieve preliminary feature extraction and downsampling, and output the suspended clamp distortion adaptive feature map X2. X2 is input in parallel to three branches for feature extraction: In branch one: Perform a Conv3×3 convolution operation on X2 to capture the basic local geometric edges and deformation features of the hanging clamp, and output the hanging clamp distortion adaptive feature map X7. In branch two: First, a Conv3×3 convolution operation is performed on X2 to output the adaptive feature map of the hanging clamp distortion X3; then, X3 is input into a dilated convolutional layer with a dilation rate of 3 to expand the receptive field without losing spatial resolution, so as to capture the macroscopic contextual features of multiple nested components such as U-shaped hanging rings and connecting plates, and output the adaptive feature map of the hanging clamp distortion X4; a Concat operation is performed on X3 and X4 to concatenate them along the channel dimension to obtain the adaptive feature map of the hanging clamp distortion X5; finally, a Conv1×1 convolution operation is performed on X5 to achieve the preliminary fusion and dimensionality reduction aggregation of multi-scale heterogeneous features under different receptive fields, and output the adaptive feature map of the hanging clamp distortion X6. In branch three: First, X2 is simultaneously subjected to horizontal global average pooling (X-GAP) and vertical global average pooling (Y-GAP) operations, and one-dimensional feature encoding is performed along different spatial dimensions to extract the structural edge distribution features of the dangling clamp in the horizontal and vertical directions, outputting the dangling clamp distortion adaptive feature map X9 and X10 respectively; then, X9 is subjected to dimensional rearrangement operation to obtain the dangling clamp distortion adaptive feature map X11; and X10 and X11 are subjected to concat operation along the channel dimension. The features of the suspended clamp distortion adaptive feature map X12 are obtained by concatenating the features. Then, a Conv1×1 convolution operation is performed on X12 to achieve feature dimensionality fusion and reduction, resulting in the suspended clamp distortion adaptive feature map X13. Finally, X13 is non-linearly normalized using the Sigmoid activation function to generate a cross-dimensional deformation boundary response weight map H1 that significantly enhances the features of the suspended clamp and its defects. At the same time, X9 is non-linearly normalized using the Sigmoid activation function to generate a cross-dimensional deformation boundary response weight map H2. First, X7 and X6 are added element-wise to achieve deep synergy and fusion of basic deformation features and multi-scale heterogeneous receptive field features, resulting in the suspended clamp distortion adaptive feature map X8. Then, based on tensor broadcasting, X8 and H1 are multiplied element-wise for preliminary spatial positioning guidance, outputting the suspended clamp distortion adaptive feature map X14. Finally, based on tensor broadcasting, X14 and H2 are multiplied element-wise to suppress complex environmental background noise, outputting the suspended clamp distortion adaptive feature map X15.

[0032] This invention designs a Cascaded Multi-Scale Receptive Field Background Suppression Module (CRBS). The CRBS module aims to solve the problem that under the complex inspection perspective of UAVs, visual interference caused by the extremely similar material and appearance of closely connected metal components makes it difficult for conventional networks to accurately separate effective targets, leading to the problem that hanging clamps and their subtle defects are missed due to interference from the surrounding similar background.

[0033] Conventional deep convolutional networks struggle to simultaneously remove complex macroscopic backgrounds and preserve subtle local semantic features during feature extraction and downsampling. To address this, the CRBS module employs a dual-branch structure and a multi-scale dilated convolutional cascade approach. This allows it to expand the receptive field without sacrificing spatial resolution, adaptively and completely encompassing the macroscopic framework of the entire suspension clamp assembly. This module guides the network to remove interference from overlapping and overlapping hardware elements such as closely connected normal casting textures and rust patches, focusing computation on sub-pixel-level anomalies in the suspension clamp and its minute cracks and missing cotter pins. This improves the model's robustness to defect detection in low-contrast and complex background conditions in the field. The CRBS module structure diagram is shown below. Figure 5 As shown The construction and execution process of the CRBS module is as follows: The multi-receptive-field collaboratively refined feature map Y1 is input in parallel into two branches, which are used for multi-scale macroscopic feature extraction and weight generation to suppress spatial background, respectively. The specific construction process of the two branches is as follows: In branch one: Y1 is input to a dilated convolutional layer (DilatedConv) with a dilation rate of 3 to initially expand the receptive field without sacrificing spatial resolution, outputting a multi-receptive field collaboratively refined feature map Y2; Y2 is input to a dilated convolutional layer (DilatedConv) with a dilation rate of 5, outputting a multi-receptive field collaboratively refined feature map Y3; Y3 is input to a dilated convolutional layer (DilatedConv) with a dilation rate of 7 to further expand the receptive field, outputting a multi-receptive field collaboratively refined feature map Y4; Y2, Y3, and Y4 are concatenated along the channel dimension to obtain a multi-receptive field collaboratively refined feature map Y5; Y5 is convolved with a Conv1×1 convolution to achieve macro-contextual structural feature fusion and dimensionality reduction aggregation of the overall component of the dangling clamp, outputting a multi-receptive field collaboratively refined feature map Y6; In branch two: Channel average pooling and channel max pooling operations are performed on Y1 in parallel, and spatial feature compression is performed along the channel dimension to capture the overall representation characteristics of the global background and local edge details, outputting multi-receptive field collaborative purification feature maps Y7 and Y8 respectively; Concat operation is performed on Y7 and Y8, and they are concatenated along the channel dimension to obtain multi-receptive field collaborative purification feature map Y9; Conv7×7 convolution operation is performed on Y9 to achieve spatial feature smoothing and depth fusion under a large receptive field, outputting multi-receptive field collaborative purification feature map Y10; Y10 is non-linearly normalized by the Sigmoid activation function to generate a cascaded background suppression weight map W1 that has been stripped of interference from similar metallic materials and textured backgrounds; Element-wise multiplication of Y6 and W1 is performed to adaptively weight the macroscopic features at multiple scales, suppress interference from similar metal materials and textured backgrounds, and output the multi-receptive-field collaborative purified feature map Y11; Conv3×3 convolution is performed on Y11 to achieve local spatial smoothing and nonlinear reconstruction of the weighted purified features, and output the multi-receptive-field collaborative purified feature map Y12.

[0034] S3. Construct a cross-level recalibration guided feature fusion unit, input F12, F13, and F14 into the cross-level recalibration guided feature fusion unit, perform feature fusion, and output the fine-grained defect feature map F27, the fine-grained defect feature map F28, and the fine-grained defect feature map F29 of the directional decoupling pendant clamp.

[0035] The main function of the cross-level recalibration-guided feature fusion unit is to address the size differences between the macroscopic component structure and minute defects of the suspended clamp in UAV high-altitude inspection scenarios, preventing detailed features from being covered by macroscopic semantics. The cross-level recalibration-guided feature fusion unit converges and aligns features extracted at different stages by the robust multi-scale deformation perception feature extraction unit after downsampling and upsampling, and then bidirectionally feeds the fused features back to each network branch, achieving decoupling of different dimensions of the target features. Simultaneously, this unit introduces a multi-scale hole feature recalibration guidance module (MDFG) into the feature fusion chain. Utilizing multi-branch parallel expansion of feature extraction and max pooling, it dynamically adjusts the weights of fine-grained crack and pin-missing anomaly responses before final fusion for pixel-level recalibration guidance, thereby improving the model's detection performance for suspended clamps and their fine-grained defects. The structure diagram of the cross-level recalibration-guided feature fusion unit is shown below. Figure 6 As shown.

[0036] The specific construction process of the cross-level recalibration guided feature fusion unit is as follows: F12 is input into a Conv 1×1 convolutional layer for channel feature alignment, resulting in the fine-grained defect feature map F16 of the directional decoupling overhang clamp. F16 is then downsampled to compress and align the spatial feature scale, yielding the fine-grained defect feature map F17 of the directional decoupling overhang clamp. Simultaneously, F12 is input into Conv... A 3×3 convolutional layer is used to perform spatial feature fusion to obtain the fine-grained defect feature map F15 of the directional decoupling overhang clamp. A weighted fusion operation is then performed on F12 and F15 to reconstruct the shallow high-resolution geometric details, resulting in the fine-grained defect feature map F23 of the directional decoupling overhang clamp. F23 is then input into the second MDFG module, where its internal mechanisms dynamically amplify the weights of fine-grained anomalous responses, outputting the fine-grained defect feature map F24 of the directional decoupling overhang clamp. F24 is then input into the cross-stage local bottleneck submodule C2f for information extraction and channel information fusion, outputting the fine-grained defect feature map F27 of the directional decoupling overhang clamp. F14 is input into a Conv 3×3 convolutional layer to obtain the fine-grained defect feature map F18 of the directional decoupling overhang clamp; F18 is upsampled to achieve spatial scale magnification and obtain the fine-grained defect feature map F20 of the directional decoupling overhang clamp. F13, F17, and F20 are weighted and fused to avoid the jagged distortion caused by traditional coarse channel splicing, resulting in a fine-grained defect feature map F21 for directional decoupling suspension clamps. F21 is then input into the first MDFG module to obtain a fine-grained defect feature map F22 for directional decoupling suspension clamps. F20, F21, and F22 are weighted and fused to complete feature recombination, resulting in the fine-grained defect feature map F25 of the directional decoupling overhang clamp. F25 is then input into the cross-stage local bottleneck submodule C2f, where contextual semantic information and channel information are aggregated and fused to output the fine-grained defect feature map F28 of the directional decoupling overhang clamp. Inputting F18 into a Conv 1×1 convolutional layer restores the channels, yielding the fine-grained defect feature map F19 for the directional decoupled hanging clamp. Weighted fusion of F14 and F19 enhances and reconstructs the deep features of the hanging clamp, resulting in the fine-grained defect feature map F26 for the directional decoupled hanging clamp. Inputting F26 into the cross-stage local bottleneck submodule C2f refines the deep feature information and fuses the channel information, yielding the fine-grained defect feature map F29 for the directional decoupled hanging clamp.

[0037] This invention designs a Multi Scale Dilated Feature Recalibration Guidance Module (MDFG). The MDFG module aims to address the problem in complex high-altitude inspection scenarios using UAVs, where the scale difference between the macroscopic component structure and minute defects of the suspension clamp leads to the easy obscuring of high-resolution minute cracks and pin hole edge geometric details by deep semantics during cross-level feature fusion, and interference from complex isomorphic texture backgrounds.

[0038] Conventional deep convolutional networks often employ channel concatenation or element-wise addition when performing cross-layer, multi-scale feature fusion, lacking targeted protection for weak anomalous response features and failing to decouple the scale differences of drooping clamps. The MDFG module dynamically amplifies the weights of fine-grained defects belonging to drooping clamps before feature fusion, while simultaneously using attention to suppress isomorphic clutter such as closely connected normal casting textures and rust patches. This prevents the high-frequency geometric features of minute defects from being covered by the semantics of the deep macroscopic framework, thus avoiding jagged distortion caused by fusion and providing features for capturing fine-grained defects. The MDFG module structure diagram is shown below. Figure 7 As shown.

[0039] The construction and execution process of the MDFG module is as follows: The multi-scale recalibrated feature map Z1 is input in parallel into two branches, Branch 1 and Branch 2, for feature fusion. The specific construction process of the two branches is as follows: In branch one: Z1 is subjected to a Conv3×3 convolution operation to capture high-resolution geometric features of small defect edges, and the multi-scale recalibrated feature map Z2 is output; Z2 is input into a dilated convolutional layer (DilatedConv) with a dilation rate of 3 to expand the local receptive field without losing spatial resolution, and the multi-scale recalibrated feature map Z3 is output; Z2 and Z3 are subjected to a Concat operation and concatenated along the channel dimension to achieve the fusion of basic local features and expanded receptive field features, and the multi-scale recalibrated feature map Z4 is obtained. In branch two: A Conv5×5 convolution operation is performed on Z1 to capture low-resolution geometric features of the defect edges, resulting in a multi-scale recalibrated feature map Z5; Z5 is input into a dilated convolutional layer with a dilation rate of 5 to capture fine-grained contextual structural features around the defect with a larger receptive field, resulting in a multi-scale recalibrated feature map Z6; A Concat operation is performed on Z5 and Z6, concatenating them along the channel dimension to achieve feature fusion at different dilation scales, resulting in a multi-scale recalibrated feature map Z7; Element-wise addition of Z4 and Z7 is performed to achieve deep fusion of the dual-branch multi-scale expansion features, resulting in a multi-scale recalibrated feature map Z8. Max pooling is then applied to Z8 to extract the most significant drooping clamps and their defect features in the spatial dimension, resulting in a multi-scale recalibrated feature map Z9. Conv7×7 convolution is then performed on Z9 to achieve spatial feature smoothing and depth extraction under a large receptive field, outputting a multi-scale recalibrated feature map Z10. Z10 is then non-linearly normalized using the Sigmoid activation function to generate a feature recalibration guiding weight map O1 with dynamically amplified subtle defect features. Element-wise multiplication of Z1 and O1 is performed to prevent high-resolution crack and pin hole details from being swallowed by deep strong semantics, while suppressing clutter interference from isomorphic texture background, resulting in a multi-scale recalibrated feature map Z11.

[0040] S4. Construct a dynamic rotational decoupling prediction unit. Input F27, F28, and F29 into the dynamic rotational decoupling prediction unit to perform multi-scale target regression, classification prediction, draw directional detection boxes and confidence scores on F1, and output the hanging clamp defect detection map P.

[0041] The Dynamic Rotation Decoupling Prediction Unit (ROD) enables the localization and defect diagnosis of suspended clamp components in UAV inspection scenarios of power transmission lines by decoupling and predicting multi-level features. This unit utilizes an Oriented Bounding Box Header (OBBHead) to adaptively wrap suspended clamps and their defect regions at arbitrary tilt and rotation angles, reducing the interference from sky, trees, or background conductor pixels within the traditional horizontal detection box and eliminating clutter interference from surrounding similar metal components. When faced with overlapping and occluded targets, the unit can suppress light and shadow edge interference and improve the model's robustness to defect detection in low-contrast and complex background conditions through weighted box fusion and Rotation Soft-NMS operations. The Dynamic Rotation Decoupling Prediction Unit is described below. Figure 8 As shown.

[0042] The specific construction process of the dynamic rotational decoupling prediction unit is as follows: F27, F28, and F29 are input in parallel into three directional bounding box detection heads (OBB Heads) to adaptively fit any tilt angle of the suspended clamp and its local defect area from the UAV inspection view, and output the initial orientation prediction results R1, R2, and R3 of the suspended clamp respectively. The R1, R2, and R3 mentioned therein each include a confidence score, the absolute coordinates of the center points of all predicted orientation detection boxes, the physical representation width of the orientation detection box, and the orientation tilt angle of the length orientation detection box relative to the horizontal coordinate axis of the image. Weighted box fusion and Rotation Soft-NMS operation are performed on R1, R2, and R3 to suppress redundant false alarms under the conditions of hardware overlap and light and shadow distortion. This achieves deep decoupling and collaborative optimization of the prediction box of macro components of the suspension clamp and millimeter-level micro defects, and outputs the final orientation prediction result R4 of the suspension clamp. The R4 includes the confidence scores after filtering R1, R2, and R3, the absolute coordinates of the center points of all predicted orientation detection boxes, the physical representation width of the orientation detection boxes, and the orientation tilt angle of the length orientation detection boxes relative to the horizontal coordinate axis of the image. Based on R4, directional detection boxes and confidence scores are drawn on F1 to obtain a directional detection box containing the outline of the suspension clamp and its confidence score, and a directional detection box containing the defect of the suspension clamp and its confidence score, which is a suspension clamp defect detection map P.

[0043] S5. A robust anti-interference deformation adaptive fine-grained suspension clamp defect detection model (RADFC) is constructed. The model is trained and its parameters are updated using a transmission line suspension clamp image dataset to obtain an optimized RADFC model. The RADFC model structure diagram is shown below. Figure 9 As shown.

[0044] The training process of the RADFC model using the transmission line suspension clamp image dataset is as follows: Before training the RADFC model, all weight parameters are initialized, and corresponding hyperparameters, including learning rate, batch size, optimizer type, and momentum, are configured according to the characteristics of the fine-grained defect detection task for hanging clamps. After setting the hyperparameters, a batch iterative training strategy is adopted. Each training batch extracts a certain number of data-augmented labeled hanging clamp images from the training subset and inputs them into the RADFC model. The images are then forward-propagated sequentially through a robust multi-scale deformation-aware feature extraction unit, a cross-level recalibration-guided feature fusion unit, and a dynamic rotation decoupling prediction unit to obtain the predicted directional detection boxes for the hanging clamp body and minor defects, as well as the target confidence. Subsequently, the error between the predicted geometric position and confidence of the directional detection boxes and the true labels in the training subset is calculated using a loss function, and the network weights are dynamically updated using backpropagation and gradient descent strategies to optimize model performance.

[0045] During training, the model performance is monitored in real time using a validation subset. The model is judged to be overfitting based on the changes in the mean rotation average precision (Rotation mAP) and loss value on the validation subset. If the performance of the validation subset does not improve within a certain period, the learning rate is dynamically adjusted or early stopping is enabled through a learning rate scheduling strategy to prevent network overfitting.

[0046] When the loss of the validation subset no longer decreases significantly over multiple training epochs or the model performance stabilizes, the training process is terminated, and the weights and parameters of the best-performing model on the validation subset are saved. An independent test subset is then input into this optimal model to evaluate and verify the model's generalization ability for high-precision orientation and fine-grained defect detection of suspended clamps in UAV inspection scenarios with complex isomorphic background interference. Finally, a trained RADFC model that can be used for target detection of suspended clamps in complex field conditions is obtained.

[0047] S6 applies the optimized strong anti-disturbance deformation adaptive fine-grained suspension clamp defect detection model to the actual UAV transmission line inspection operation scenario: taking the transmission line image containing suspension clamps captured by the UAV as input, it realizes the location and detection of suspension clamp defects, and outputs the suspension clamp defect detection results.

[0048] The results include oriented detection boxes that fit the outline of the suspension clamp and their confidence scores, and oriented detection boxes that fit the outline of the suspension clamp defects and their confidence scores; the above results are labeled in the suspension clamp defect image.

[0049] 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 detecting defects in suspension clamps of power transmission lines based on unmanned aerial vehicle (UAV) inspection, characterized in that, Includes the following steps: S1, Construct a dataset of images of suspension clamps for power transmission lines; S2, construct a strong anti-disturbance multi-scale deformation sensing feature extraction unit, input the image F1 of any transmission line suspension clamp to the strong anti-disturbance multi-scale deformation sensing feature extraction unit, perform feature extraction, and output the fine-grained defect feature map F12, the fine-grained defect feature map F13, and the fine-grained defect feature map F14 of the directional decoupled suspension clamp; S3, construct a cross-level recalibration guided feature fusion unit, input F12, F13, and F14 into the cross-level recalibration guided feature fusion unit, perform feature fusion, and output the fine-grained defect feature map of the directional decoupling suspension clamp F27, the fine-grained defect feature map of the directional decoupling suspension clamp F28, and the fine-grained defect feature map of the directional decoupling suspension clamp F29. S4, construct a dynamic rotational decoupling prediction unit, input F27, F28, and F29 into the dynamic rotational decoupling prediction unit, perform multi-scale target regression, classification prediction, draw directional detection boxes and confidence scores on F1, and output the hanging clamp defect detection map P; S5. Construct a strong disturbance rejection deformation adaptive fine-grained suspension clamp defect detection model. Use the transmission line suspension clamp image dataset to train and update the parameters of the strong disturbance rejection deformation adaptive fine-grained suspension clamp defect detection model to obtain the optimized strong disturbance rejection deformation adaptive fine-grained suspension clamp defect detection model. S6 applies the optimized strong anti-disturbance deformation adaptive fine-grained suspension clamp defect detection model to the actual UAV transmission line inspection operation scenario: taking the transmission line image containing suspension clamps captured by the UAV as input, it realizes the location and detection of suspension clamp defects, and outputs the suspension clamp defect detection results.

2. The method for detecting defects in suspension clamps of power transmission lines based on UAV inspection according to claim 1, characterized in that, The specific construction process of the highly disturbance-resistant multi-scale deformation sensing feature extraction unit is as follows: S21, Input the image F1 of any transmission line suspension clamp into a standard convolutional submodule to obtain the fine-grained defect feature map F2 of the directional decoupled suspension clamp; In step S22, input F2 into the first HRDA module. Utilizing its internal three-branch parallel structure, perform multi-level cross-dimensional feature extraction to extract the boundary deformation features of the complex components of the drooping clamp, resulting in the fine-grained defect feature map F3 of the directional decoupled drooping clamp. Input F2 into the first CRBS module. Utilize its internal cascaded dilated convolution to expand the receptive field and combine it with dual-pooling spatial attention to suppress complex backgrounds and capture the macroscopic contextual structural features of the drooping clamp, resulting in the fine-grained defect feature map F4 of the directional decoupled drooping clamp. Perform a Concat operation on F3 and F4, concatenating them along the channel dimension to fuse the multi-scale deformation and macroscopic structural information of the drooping clamp, resulting in the fine-grained defect feature map F5 of the directional decoupled drooping clamp. Perform a Conv1×1 convolution operation on F5 to achieve channel dimensionality reduction and fusion, resulting in the fine-grained defect feature map F12 of the directional decoupled drooping clamp. S23, input F3 into the second HRDA module to further adaptively extract deeper features of the droop clamp and its defects, resulting in the fine-grained defect feature map F6 of the directional decoupled droop clamp; input F4 into the second CRBS module to remove the confusing features between the normal metal texture and the defect fracture surface of the droop clamp, enhancing its anti-interference ability against complex backgrounds, resulting in the fine-grained defect feature map F7 of the directional decoupled droop clamp; perform a Concat operation on F6 and F7, splicing them along the channel dimension to obtain the fine-grained defect feature map F8 of the directional decoupled droop clamp; perform a Conv1×1 convolution operation on F8 to achieve channel dimensionality reduction and fusion, resulting in the fine-grained defect feature map F13 of the directional decoupled droop clamp; In step S24, input F6 into the third HRDA module for final deformation adaptive extraction, extracting weak overhang clamp defect features to obtain the fine-grained defect feature map F9 of the directional decoupled overhang clamp; input F7 into the third CRBS module to complete the extraction of strong semantic features of the overhang clamp and the removal of complex backgrounds, to obtain the fine-grained defect feature map F10 of the directional decoupled overhang clamp; perform a Concat operation on F9 and F10, concatenating them along the channel dimension to achieve efficient aggregation of deep high semantic features, to obtain the fine-grained defect feature map F11 of the directional decoupled overhang clamp; perform a Conv1×1 convolution operation on F11 to achieve cross-channel information integration, and output the fine-grained defect feature map F14 of the directional decoupled overhang clamp.

3. The method for detecting defects in suspension clamps of transmission lines based on UAV inspection according to claim 2, characterized in that, The construction and execution process of the HRDA module is as follows: A1. Perform Conv3×3 convolution operation on the adaptive feature map X1 of the suspended clamp distortion to achieve preliminary feature extraction and downsampling, and output the adaptive feature map X2 of the suspended clamp distortion. A2, X2 is input in parallel to the three branches for feature extraction: In branch one: Perform a Conv3×3 convolution operation on X2 to capture the basic local geometric edges and deformation features of the hanging clamp, and output the hanging clamp distortion adaptive feature map X7. In branch two: First, a Conv3×3 convolution operation is performed on X2 to output the drooping clamp distortion adaptive feature map X3; then, X3 is input into a dilated convolutional layer with a dilation rate of 3 to output the drooping clamp distortion adaptive feature map X4; a Concat operation is performed on X3 and X4, concatenating them along the channel dimension to obtain the drooping clamp distortion adaptive feature map X5; finally, a Conv1×1 convolution operation is performed on X5 to achieve the initial fusion and dimensionality reduction aggregation of multi-scale heterogeneous features under different receptive fields, and the drooping clamp distortion adaptive feature map X6 is output. In branch three: First, X2 is simultaneously subjected to horizontal and vertical global average pooling operations, and one-dimensional feature encoding is performed along different spatial dimensions to extract the structural edge distribution features of the hanging clamp in the horizontal and vertical directions, outputting the hanging clamp distortion adaptive feature map X9 and X10 respectively; then, X9 is subjected to dimensional rearrangement operation to obtain the hanging clamp distortion adaptive feature map X11; X10 and X11 are concatted along the channel dimension to obtain the hanging clamp distortion adaptive feature map X11. The drooping clamp distortion adaptive feature map X12 is generated. A Conv1×1 convolution operation is performed on X12 to achieve feature dimension fusion and dimensionality reduction, resulting in the drooping clamp distortion adaptive feature map X13. Finally, X13 is non-linearly normalized using the Sigmoid activation function to generate a cross-dimensional deformation boundary response weight map H1 that significantly enhances the drooping clamp and its defect features. Simultaneously, X9 is non-linearly normalized using the Sigmoid activation function to generate a cross-dimensional deformation boundary response weight map H2. A3. First, X7 and X6 are added element-wise to achieve deep synergy and fusion of basic deformation features and multi-scale heterogeneous receptive field features, resulting in the suspended clamp distortion adaptive feature map X8. Then, based on tensor broadcasting, X8 and H1 are multiplied element-wise for preliminary spatial positioning guidance, outputting the suspended clamp distortion adaptive feature map X14. Finally, based on tensor broadcasting, X14 and H2 are multiplied element-wise to suppress complex environmental background noise, outputting the suspended clamp distortion adaptive feature map X15.

4. The method for detecting defects in suspension clamps of transmission lines based on UAV inspection according to claim 2, characterized in that, The construction and execution process of the CRBS module is as follows: B1 inputs the multi-receptive-field collaboratively purified feature map Y1 into two branches in parallel, performing multi-scale macroscopic feature extraction and generating weights to suppress spatial background, respectively. The specific construction process of the two branches is as follows: In branch one: Y1 is input into a dilated convolutional layer with a dilation rate of 3 to initially expand the receptive field without sacrificing spatial resolution, outputting a multi-receptive field collaboratively refined feature map Y2; Y2 is input into a dilated convolutional layer with a dilation rate of 5 to output a multi-receptive field collaboratively refined feature map Y3; Y3 is input into a dilated convolutional layer with a dilation rate of 7 to further expand the receptive field, outputting a multi-receptive field collaboratively refined feature map Y4; a Concat operation is performed on Y2, Y3, and Y4, concatenating them along the channel dimension to obtain a multi-receptive field collaboratively refined feature map Y5; a Conv1×1 convolution operation is performed on Y5 to output a multi-receptive field collaboratively refined feature map Y6; In branch two: Channel average pooling and channel max pooling operations are performed on Y1 in parallel, and spatial feature compression is performed along the channel dimension to capture the overall representation characteristics of the global background and local edge details, outputting multi-receptive field co-purified feature maps Y7 and Y8 respectively; Y7 and Y8 are concatted along the channel dimension to obtain multi-receptive field co-purified feature map Y9; Y9 is convolved with Conv7×7 to output multi-receptive field co-purified feature map Y10; Y10 is non-linearly normalized using the Sigmoid activation function to generate a cascaded background suppression weight map W1 that has been stripped of interference from similar metallic materials and textured backgrounds; B2 performs element-wise multiplication of Y6 and W1 to achieve adaptive weighting of multi-scale macro features, suppress interference from similar metal materials and textured backgrounds, and outputs multi-receptive field collaborative purified feature map Y11; Conv3×3 convolution operation is performed on Y11 to achieve local spatial smoothing and nonlinear reconstruction of the weighted purified features, and outputs multi-receptive field collaborative purified feature map Y12.

5. The method for detecting defects in suspension clamps of transmission lines based on UAV inspection according to claim 1, characterized in that, The specific construction process of the cross-level recalibration guided feature fusion unit is as follows: S31, input F12 into the Conv 1×1 convolutional layer for channel feature alignment to obtain the fine-grained defect feature map F16 of the directional decoupling overhang clamp; perform downsampling on F16 to compress and align the spatial feature scale to obtain the fine-grained defect feature map F17 of the directional decoupling overhang clamp; simultaneously, input F12 into Conv A 3×3 convolutional layer is used to perform spatial feature fusion to obtain the fine-grained defect feature map F15 of the directional decoupling overhang clamp. A weighted fusion operation is then performed on F12 and F15 to reconstruct the shallow high-resolution geometric details, resulting in the fine-grained defect feature map F23 of the directional decoupling overhang clamp. F23 is then input into the second MDFG module, where its internal mechanisms dynamically amplify the weights of fine-grained anomalous responses, outputting the fine-grained defect feature map F24 of the directional decoupling overhang clamp. F24 is then input into the cross-stage local bottleneck submodule C2f for information extraction and channel information fusion, outputting the fine-grained defect feature map F27 of the directional decoupling overhang clamp. S32, input F14 into the Conv 3×3 convolutional layer to obtain the fine-grained defect feature map F18 of the directional decoupling overhang clamp; perform upsampling operation on F18 to realize the spatial scale magnification and obtain the fine-grained defect feature map F20 of the directional decoupling overhang clamp. S33, perform a weighted fusion operation on F13, F17, and F20 to avoid the jagged distortion caused by traditional coarse channel splicing, and obtain the fine-grained defect feature map F21 of the directional decoupled suspension clamp; input F21 into the first MDFG module to obtain the fine-grained defect feature map F22 of the directional decoupled suspension clamp. S34, perform a weighted fusion operation on F20, F21, and F22 to complete feature recombination and obtain the fine-grained defect feature map F25 of the directional decoupling overhang clamp; input F25 into the cross-stage local bottleneck submodule C2f, aggregate contextual semantic information and channel information to fuse and output the fine-grained defect feature map F28 of the directional decoupling overhang clamp. In step S35, F18 is input into the Conv 1×1 convolutional layer to restore the channels and obtain the fine-grained defect feature map F19 of the directional decoupled hanging clamp. F14 and F19 are weighted and fused to enhance and reconstruct the deep features of the hanging clamp and obtain the fine-grained defect feature map F26 of the directional decoupled hanging clamp. F26 is input into the cross-stage local bottleneck submodule C2f to refine the deep feature information and fuse the channel information and obtain the fine-grained defect feature map F29 of the directional decoupled hanging clamp.

6. A method for detecting defects in suspension clamps of transmission lines based on UAV inspection, as described in claim 5, is characterized in that, The construction and execution process of the MDFG module is as follows: C1 inputs the multi-scale recalibrated feature map Z1 into two branches, branch one and branch two, in parallel for feature fusion. The specific construction process of the two branches is as follows: In branch one: Conv3×3 convolution operation is performed on Z1 to capture high-resolution geometric features of small defect edges, and multi-scale recalibrated feature map Z2 is output; Z2 is input into a dilated convolutional layer with an expansion rate of 3 to expand the local receptive field without losing spatial resolution, and multi-scale recalibrated feature map Z3 is output; Z2 and Z3 are concat operation and concatenated along the channel dimension to achieve the fusion of basic local features and expanded receptive field features, and multi-scale recalibrated feature map Z4 is obtained. In branch two: A Conv5×5 convolution operation is performed on Z1 to capture low-resolution geometric features of the defect edges, resulting in a multi-scale recalibrated feature map Z5; Z5 is input into a dilated convolutional layer with an expansion rate of 5 to capture fine-grained contextual structural features around the defect with a larger receptive field, resulting in a multi-scale recalibrated feature map Z6; A Concat operation is performed on Z5 and Z6, concatenating them along the channel dimension to achieve feature fusion at different expansion scales, resulting in a multi-scale recalibrated feature map Z7; C2 performs element-wise addition of Z4 and Z7 to achieve deep fusion of the dual-branch multi-scale expansion features, resulting in a multi-scale recalibration feature map Z8. Max pooling is then applied to Z8 to extract the most significant drooping clamps and their defect features in the spatial dimension, yielding a multi-scale recalibration feature map Z9. A Conv7×7 convolution operation is then performed on Z9 to achieve spatial feature smoothing and depth extraction under a large receptive field, outputting a multi-scale recalibration feature map Z10. Finally, Z10 is non-linearly normalized using the Sigmoid activation function to generate a feature recalibration guiding weight map O1 with dynamically amplified subtle defect features. C3 performs element-wise multiplication of Z1 and O1 to obtain the multi-scale recalibrated feature map Z11.

7. The method for detecting defects in suspension clamps of transmission lines based on UAV inspection according to claim 1, characterized in that, The specific construction process of the dynamic rotational decoupling prediction unit is as follows: S41, input F27, F28, and F29 in parallel into the three directional bounding box detection heads, adaptively fit any tilt angle of the suspended clamp and its local defect area under the UAV inspection view, and output the initial orientation prediction results R1, R2, and R3 of the suspended clamp respectively. S42 performs weighted frame fusion and rotation soft nonmaximum suppression operations on R1, R2, and R3 to suppress redundant false alarms under the conditions of hardware overlap and light and shadow distortion, realizes deep decoupling and collaborative optimization of the prediction frame of macro components of the suspension clamp and millimeter-level micro defects, and outputs the final orientation prediction result R4 of the suspension clamp. S43, based on R4, draw the directional detection box and confidence score on F1 to obtain the directional detection box containing the outline of the suspension clamp and its confidence score, and the directional detection box containing the defect of the suspension clamp and its confidence score, which is the suspension clamp defect detection map P.

Images