A method for detecting construction quality of a power transmission line with adaptive dynamic grouping

Abstract

The application discloses a power transmission line construction quality detection method based on adaptive dynamic grouping, and belongs to the technical field of power engineering construction quality monitoring. The method comprises the following steps: obtaining a power transmission line construction site picture; constructing an improved network model based on YOLOv13; inputting the picture into the model for detection to complete identification of construction components and defects. The improved network model is composed of a backbone network, a neck network and a detection head. The improvement lies in that an adaptive dynamic grouping convolution module with a dynamically adjustable grouping number is introduced into the backbone network and the neck network; dynamic pruning based on gradient feedback and shallow feature enhancement are performed on the neck network; and the detection head is directly connected to the output of a feature splicing module to retain high-frequency detail information. The application effectively solves the problem of high missing detection rate of subtle construction defects such as missing bolts and unopened pins in the traditional method, and realizes high-precision and automatic construction quality detection.

Description

Technical Field

[0001] This invention relates to the field of power engineering construction quality monitoring technology, specifically to an adaptive dynamic grouping method for detecting the construction quality of transmission lines. Background Technology

[0002] Transmission lines are critical infrastructure of the power system, and their construction quality directly affects the safety and reliability of the power grid operation. Traditional transmission line construction quality inspection mainly relies on visual inspection, manual measurement, and post-construction testing by supervisors and construction personnel. This approach has several drawbacks: manual inspection of each tower and component is time-consuming and labor-intensive, especially in challenging construction sections such as mountainous areas and river crossings, resulting in long inspection cycles and making it difficult to achieve real-time inspection throughout the entire process and at all points; inspection results heavily depend on individual experience and responsibility, easily leading to missed inspections and misjudgments, and making it difficult to ensure the uniformity and objectivity of inspection standards; manual close-range inspections during high-altitude operations, live or near-live conductors pose safety risks such as falls from heights and electric shock; some connection points and fasteners are located inside complex structures or in blind spots, making effective manual observation difficult. With the development of computer vision technology, research has attempted to apply target detection algorithms to power equipment defect identification. In recent years, computer vision-based target detection technology has been introduced into this field. However, defect detection in power transmission line construction scenarios presents unique challenges: the scale of the targets to be inspected varies greatly, from tens of meters of conductors and insulator strings to centimeter-level bolts and pins. In particular, defects in small-scale fasteners (such as pins in cotter pin condition) account for a very small proportion in the image, and general detection models are prone to missing them; the background of the construction site is complex and the lighting conditions are varied, which places extremely high demands on the robustness of the model.

[0003] YOLOv13, as an advanced one-stage target detection model, has achieved a good balance between speed and accuracy. However, its original model still falls short in its ability to detect defects in small targets and highly similar backgrounds in the aforementioned construction scenarios. Existing improvements mostly focus on static module replacement or increasing network depth, lacking adaptive feature extraction mechanisms for different defect visual characteristics, such as the "texture interruption" of missing bolts and the "geometric contour breakage" of insulator self-explosion, as well as the collaborative optimization capabilities between components within the model. This makes it difficult to achieve the ultimate detection of diverse construction defects, especially minute defects. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an adaptive dynamic grouping method for inspecting the construction quality of transmission lines. This method significantly improves the detection accuracy of small-scale construction components and defects in transmission lines by constructing an improved YOLOv13 model with an internal feedback adjustment mechanism.

[0005] To achieve the above objectives, this invention provides an adaptive dynamic grouping method for inspecting the construction quality of transmission lines, comprising the following steps: S1. Obtain photos of the power transmission line construction site; S2. Construct an improved network model for YOLOv13; S3. Input the images of the power transmission line construction site into the improved YOLOv13 network model for detection, and output the detection results of construction components and defect categories; The improved YOLOv13 network model consists of a backbone network, a neck network, and a detection head. Images of the power transmission line construction site are input into the backbone network for processing. The output of the backbone network is input into the neck network for processing, and the output of the neck network is input into the detection head for processing, outputting the detection results of the construction components and defect categories. Improvements include: introducing an adaptive dynamic grouping convolution module in the backbone and neck network, which replaces pointwise convolution in the depthwise separable convolution module with grouped convolution; the number of groups in the grouped convolution in the adaptive dynamic grouping convolution module is a dynamically adjustable parameter; performing gradient-feedback-based dynamic pruning on the neck network; optimizing shallow feature enhancement of the neck network; and directly connecting the detection head to the output of the feature splicing module of the neck network.

[0006] Furthermore, the backbone network includes a first ordinary convolutional layer, a second ordinary convolutional layer, a first improved DS-C3K2 module, a third ordinary convolutional layer, a second improved DS-C3K2 module, a first adaptive dynamic group convolutional module, a first improved region attention-enhanced cross-feature module, a second adaptive dynamic group convolutional module, and a second improved region attention-enhanced cross-feature module connected in sequence. The convolutions in the first and second improved DS-C3K2 modules are replaced with group convolutions in the adaptive dynamic group convolutional module. The first and second improved region attention-enhanced cross-feature modules replace the linear position encoding in the original YOLOv13 region attention-enhanced cross-feature module with position encoding based on periodic functions. The position encoding based on periodic functions uses sine waves of different frequencies to encode positions.

[0007] Further, the neck network includes a first upsampling module. The output of the second improved region attention-enhanced cross-feature module and the output of the adaptive association enhancement module, after being concatenated, is input into the first upsampling module for processing. The output of the first upsampling module, the output of the first improved region attention-enhanced cross-feature module, and the output of the adaptive association enhancement module, after being concatenated, is input into a first feature concatenation module for processing. The output of the first feature concatenation module is input into a third improved DS-C3K2 module for processing. The output of the third improved DS-C3K2 module is input into the second upsampling module for processing. The output of the first improved DS-C3K2 module is input into a downsampling module for processing. The output of the downsampling module, the output of the second upsampling module, the output of the adaptive association enhancement module, and the output of the second improved DS-C3K2 module, after being concatenated, is input into the second feature concatenation module for processing. The output of the module is processed by the fourth improved DS-C3K2 module. The output of the fourth improved DS-C3K2 module is concatenated with the output of the adaptive association enhancement module and then processed by the fourth ordinary convolutional layer. The output of the fourth ordinary convolutional layer, the output of the third improved DS-C3K2 module, and the output of the adaptive association enhancement module are concatenated and then processed by the third feature concatenation module. The output of the third feature concatenation module is processed by the fifth improved DS-C3K2 module. The output of the fifth improved DS-C3K2 module is processed by the fifth ordinary convolutional layer. The output of the fifth ordinary convolutional layer, the output of the second improved region attention enhancement cross-feature module, and the output of the adaptive association enhancement module are concatenated and then processed by the fourth feature concatenation module. The outputs of the second, third, and fourth feature concatenation modules serve as the output of the neck network. The first, second, third, fourth, and fifth ordinary convolutional layers have the same structure; the first, second, third, fourth, and fifth improved DS-C3K2 modules have the same structure; the first and second improved DS-C3K2 modules have the same structure; the first and second improved region attention-enhanced cross-feature modules have the same structure; the neck network also includes a full-process aggregation and distribution tunnel.

[0008] Furthermore, the detection head includes a first detection head, a second detection head, and a third detection head. The output of the second feature stitching module is input to the first detection head for processing. The first detection head outputs the detection results of small-scale construction components and defect categories, completing the detection of small targets. The output of the third feature stitching module is input to the second detection head for processing. The second detection head outputs the detection results of medium-scale construction components and defect categories, completing the detection of medium-scale targets. The output of the fourth feature stitching module is input to the third detection head for processing. The third detection head outputs the detection results of large-scale construction components and defect categories, completing the detection of large targets.

[0009] Furthermore, the construction components include: insulators, vibration dampers, equalizing rings, wire clamps, bolts, and pins; the defect categories include: missing bolts, pins with cotter pins not opened, vibration damper slippage, insulator self-explosion, hardware installed backwards, and conductor damage.

[0010] Furthermore, the adaptive dynamic grouped convolution module replaces the pointwise convolution in the depthwise separable convolution module with grouped convolution. The number of parameters and computational cost within the grouped convolution are as follows: ; ; in, For the number of parameters, Input the number of channels. Number of output channels The kernel size is the convolution kernel size. and These represent the height and width of the feature map, respectively. Number of groups For computational quantity; In the adaptive dynamic grouped convolution module, the number of groups G in the grouped convolution is a dynamically adjustable parameter.

[0011] Furthermore, the number of groups G in the adaptive dynamic grouped convolution module is also adjusted according to the feedback signal of the detection head: based on the average accuracy of the defect category, the number of groups G in the adaptive dynamic grouped convolution module on the corresponding target detection path is adjusted.

[0012] Furthermore, the specific process of performing gradient-feedback-based dynamic pruning on the neck network is as follows: During the training of the improved YOLOv13 network model, the importance metric of the full-process aggregation and distribution tunnels in the neck network is calculated periodically. For full-process aggregation and distribution tunnels with an importance metric lower than the dynamic threshold τ, they are masked in subsequent training processes; the formula is as follows: ; in, For importance measurement, N represents the total aggregation and distribution tunnel. The number of associated feature maps For the first Each feature map For loss function, For the first The gradient of each feature map with respect to the loss function. It is the Euclidean norm.

[0013] Furthermore, the enhancement and optimization of shallow features in the neck network specifically involves: during the fusion process of the neck network, the output of the downsampling module, i.e., the shallow features... The output of the second upsampling module is the deep feature. The fusion process is performed using the following formula: ; in, As a feature of fusion, for convolution, For splicing operations, For upsampling operation, For deep features, This is a shallow feature.

[0014] Furthermore, the improved YOLOv13 network model uses a composite loss function during training, as shown in the following formula: ; in, For composite loss function, for loss, For target confidence loss, For classifying losses, , , They are respectively The weighting coefficients for loss, target confidence loss, and classification loss.

[0015] The present invention has the following beneficial effects: (1) The adaptive dynamic grouping convolution module proposed in this invention enables the improved YOLOv13 network model to adaptively optimize feature learning strategies for construction parts and defects of different scales and texture characteristics by dynamically adjusting the number of groups. Without excessively increasing the number of parameters, it significantly improves the feature representation ability of small target detection and subtle defects.

[0016] (2) By using gradient feedback-based dynamic pruning of the neck network and shallow feature enhancement, the feature flow efficiency is optimized and redundant computation is removed. On the other hand, the utilization of high-resolution detail features is strengthened, thus improving the detection accuracy of small targets.

[0017] (3) The detection head direct connection mechanism and closed-loop feedback design organically couple the three core points of YOLOv13 model improvement dynamic grouping, neck optimization and detail preservation, forming an internal collaborative optimization system that can self-adjust according to the detection effect, which goes beyond simple module stacking and enhances the overall intelligence and adaptability of the model. Attached Figure Description

[0018] Figure 1 This is an overall framework diagram of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] This embodiment provides an adaptive dynamic grouping method for inspecting the construction quality of transmission lines, including the following steps: S1. Obtain photos of the power transmission line construction site. S1.1 Data Acquisition A drone equipped with a high-definition camera was used to conduct aerial inspections of the power transmission line construction area. The flight altitude was set at 30-50 meters, the camera resolution at 3840×2160 pixels, and the shooting angle was either vertically downwards or tilted at 45°. Data collection was conducted during the daytime with good lighting conditions, avoiding strong backlighting or shadows. Several images of the power transmission line construction site were collected, covering different construction stages and different component types.

[0021] S1.2 Defect Type Definition and Labeling A team of experienced experts labeled the collected images of transmission line construction sites, identifying the construction components and defect categories. The construction components include, but are not limited to: insulators, vibration dampers, equipotential rings, clamps, bolts, and pins. The defect categories include, but are not limited to: missing bolts, pins with unopened cotter pins, vibration damper slippage, insulator spontaneous explosion, reversed hardware installation, and conductor damage. For example, missing bolts: description: missing connecting bolts; visual characteristics: empty bolt holes, no bolts, the core feature being "a break in the texture at the regular installation hole, contrasting sharply with the surrounding continuous hardware texture." Insulator spontaneous explosion: description: damaged insulator discs; visual characteristics: broken disc edges, exposed internal material, the core feature being "a break or discontinuity in the regular geometric contour." During labeling, appropriate tools were used to draw a rectangular bounding box for each defect instance and label it with the corresponding category label. Multiple defect instances in the same image were labeled separately. After labeling, cross-validation was performed to ensure labeling quality.

[0022] S1.3 Dataset Partitioning and Preprocessing The labeled images were divided into training, validation, and test sets in a 7:2:1 ratio. Preprocessing operations such as size normalization, color space normalization, and data augmentation were then performed on the labeled images.

[0023] Size normalization adjusts all images to a uniform 640×640 pixels. After normalizing the input image, the edges of the normalized image are filled with gray pixels (RGB values ​​128, 128, 128) to achieve a resolution of 640×640.

[0024] Color space normalization normalizes the pixel values ​​of an image from [0, 255] to the range [0, 1]. ;in, The original image matrix, This is the normalized image matrix.

[0025] Furthermore, the following data augmentation operations are applied in real time during training (applied only on the training set):

[0026] Random horizontal flip: probability 0.5; Random rotation: angle range [-10°, 10°]; Random scaling: scaling ratio range [0.8, 1.2]; Random brightness adjustment: brightness variation range [-0.2, 0.2]; Random contrast adjustment: contrast variation range [0.8, 1.2]; Add Gaussian noise: noise standard deviation range [0, 0.05].

[0027] S2. Construct an improved YOLOv13 network model.

[0028] like Figure 1As shown, the improved network model of YOLOv13 consists of a backbone network, a neck network, and a head, but the internal components and connections have been reconstructed for construction defect detection tasks.

[0029] The backbone network comprises, in sequence, a first ordinary convolutional layer, a second ordinary convolutional layer, a first improved DS-C3K2 module, a third ordinary convolutional layer, a second improved DS-C3K2 module, a first adaptive dynamic group convolutional module, a first improved region attention-enhanced cross-feature module, a second adaptive dynamic group convolutional module, and a second improved region attention-enhanced cross-feature module; the convolutions in the first and second improved DS-C3K2 modules are replaced with group convolutions in the adaptive dynamic group convolutional module; the first and second improved region attention-enhanced cross-feature modules replace the linear positional encoding in the original YOLOv13 region attention-enhanced cross-feature module with positional encoding based on periodic functions.

[0030] The neck network includes a first upsampling module. The output of a second improved region attention-enhanced cross-feature module and an adaptive association enhancement module (Hyper ACE module) is concatenated and then input into the first upsampling module for processing. The output of the first upsampling module, the first improved region attention-enhanced cross-feature module, and the adaptive association enhancement module (Hyper ACE module) is concatenated and then input into a first feature concatenation module for processing. The output of the first feature concatenation module is input into a third improved DS-C3K2 module for processing. The output of the third improved DS-C3K2 module is input into the second upsampling module for processing. The output of the first improved DS-C3K2 module is input into a downsampling module for processing. The output of the downsampling module, the second upsampling module, the adaptive association enhancement module (Hyper ACE module), and the second improved DS-C3K2 module is concatenated and then input into the second feature concatenation module for processing. The output of the second feature concatenation module is input into a fourth improved DS-C3K2 module for processing. The output of the fourth improved DS-C3K2 module is then concatenated with the output of the adaptive association enhancement module (Hyper ACE module). The output of the ACE module is concatenated and then fed into the fourth ordinary convolutional layer for processing. The output of the fourth ordinary convolutional layer, the output of the third improved DS-C3K2 module, and the output of the Hyper ACE module are concatenated and then fed into the third feature concatenation module for processing. The output of the third feature concatenation module is fed into the fifth improved DS-C3K2 module for processing. The output of the fifth improved DS-C3K2 module is fed into the fifth ordinary convolutional layer for processing. The output of the fifth ordinary convolutional layer, the output of the second improved region attention-enhanced cross-feature module, and the output of the Hyper ACE module are concatenated and then fed into the fourth feature concatenation module for processing. The outputs of the second, third, and fourth feature concatenation modules serve as the output of the neck network. The first, second, third, fourth, and fifth ordinary convolutional layers have the same structure; the first, second, third, fourth, and fifth improved DS-C3K2 modules have the same structure; the first and second improved DS-C3K2 modules have the same structure; the first and second improved region attention-enhanced cross-feature modules have the same structure; the neck network also includes a full-process aggregation and distribution tunnel.

[0031] The detection head includes a first detection head, a second detection head, and a third detection head. The output of the second feature splicing module is input into the first detection head for processing. The first detection head outputs the detection results of small-scale construction components and defect categories, completing the detection of small targets. The output of the third feature splicing module is input into the second detection head for processing. The second detection head outputs the detection results of medium-scale construction components and defect categories, completing the detection of medium-scale targets. The output of the fourth feature splicing module is input into the third detection head for processing. The third detection head outputs the detection results of large-scale construction components and defect categories, completing the detection of large targets.

[0032] S2.1 Backbone Network Improvement An adaptive dynamic grouped convolution module is introduced into the backbone network. This module replaces the pointwise convolution in the depthwise separable convolution module with grouped convolution. The implementation of the adaptive dynamic grouped convolution module is as follows: First, the input feature map is divided into G groups along the channel dimension. For each group, depthwise convolution and grouped pointwise convolution are performed independently. The grouped pointwise convolution is a 1x1 grouped convolution with G groups. Finally, the outputs of each grouped convolution are concatenated along the channel dimension. In this embodiment, based on the shallow layers of the backbone network (the first and second improved DS-C3K2 modules) and the path for small object detection in the neck network (the third improved DS-C3K2 module), the first group number G1=8 is set; based on the deep layers of the backbone network (the second improved region attention-enhanced cross-feature module), the second group number G2=4 is set, where G1>G2.

[0033] S2.2 Neck Network Improvement (1) Dynamic pruning: During training initialization, all full-process aggregation and distribution tunnels (FullPAD) are retained. After training begins, the importance metric of each full-process aggregation and distribution tunnel is calculated every 50 training epochs. In this embodiment, a dynamic threshold τ = 0.1 is set. Importance is measured. The weights and gradients corresponding to the aggregation and distribution tunnels throughout the entire process are set to zero and masked in subsequent forward and backward propagation. The dynamic threshold τ decays linearly to 0.05 with increasing training epochs, gradually tightening the pruning criteria.

[0034] (2) Shallow feature injection: During the fusion process of the neck network, the output of the downsampling module, i.e., the shallow features, is injected. The output of the second upsampling module is the deep feature. The data is then fused. The output of the downsampling module is the shallow feature. After a 1x1 convolution, the number of channels is adjusted to match the output of the second upsampling module, i.e., the deep features. The upsampling process results in the same number of channels, which is then concatenated with the upsampling deep features. Finally, a 1x1 convolution is performed to fuse the resulting fused features. .

[0035] S2.3 Detector head connection optimization The three detector heads that were originally connected to the output of the DS-C3k2 module in the neck network of the YOLOv13 model are now directly connected to the output of the feature concatenation module. This modification is achieved at the code level by rewriting the model's forward propagation function, bypassing further processing by the DS-C3k2 module.

[0036] S2.4 Feedback Mechanism Establishment During the training cycle, the average precision (AP) of each defect category on the validation set is monitored. If the average precision (AP) of the "pin cotter pin not open" category does not improve significantly for 10 consecutive training epochs and is lower than the set benchmark, a feedback mechanism is triggered. The number of groups in the group convolution in the first adaptive dynamic group convolution module is temporarily increased by 2 (e.g., from 8 to 10), and then restored after 20 training epochs to attempt to stimulate more fine-grained feature learning.

[0037] S3, Model Training and Evaluation The improved YOLOv13 network model constructed in step S2 was trained end-to-end using the training and validation sets prepared in step S1. The training hyperparameters were set as follows: input size 640x640, batch size 16, and total training epochs 300. The AdamW optimizer was used with an initial learning rate of 0.001 and weight decay of 0.0005. Cosine annealing was employed for learning rate scheduling.

[0038] During training, a composite loss function is used, as shown in the following formula: ; in, For composite loss function, for loss, For target confidence loss, For classification loss.

[0039] The formula for calculating CIoU loss is: ; in, For intersection, union, and comparison, For Euclidean distance, and These are the center points of the predicted bounding box and the ground truth bounding box, respectively. The length of the diagonal of the smallest bounding rectangle. As a measure of aspect ratio consistency, These are the weighting coefficients.

[0040] Training was conducted on a workstation equipped with an NVIDIA GeForce RTX 4090 graphics card, using the PyTorch 2.0 framework.

[0041] After training, the results were evaluated on the test set. The evaluation metric was the mean accuracy (mAP@0.5). To demonstrate the effectiveness of this invention, a comparative experiment was conducted, and the results are shown in Table 1. Table 1. Performance comparison of different models on the transmission line construction defect test set.

[0042] This invention proposes an adaptive dynamic grouping method for transmission line construction quality inspection. By constructing an improved YOLOv13 model, it solves the problems of low efficiency, high false negative rate, and insufficient identification of small target defects in traditional manual inspection and general models. Experiments show that, with similar parameter count (2.6M) and FLOPs (5.9G) to the benchmark model, this invention achieves an mAP@0.5 of 92.1%, a 3.6% improvement over the benchmark model. This demonstrates that the performance improvement mainly stems from the qualitative leap in feature extraction quality brought about by the synergy of dynamic grouping, neck optimization, and direct connection of the inspection head, rather than simply from weight reduction.

[0043] Further comparisons were made using the defect category identification method for transmission line construction equipment, and the average accuracy (AP value) comparisons are shown in Table 2: Table 2 Comparison of Average Precision (AP) for Some Defect Categories (%)

[0044] As can be seen from Tables 1 and 2, the method proposed in this invention improves the overall average accuracy (mAP@0.5) by 3.6% compared to the baseline model while keeping the computational cost and number of parameters basically unchanged. In the most challenging defect category of "cotter pin not opened", the average accuracy (AP) of this invention is improved by 13.3 percentage points, which is significantly better than the model variant with a single improvement, proving the effectiveness of the synergistic work of various improvements.

[0045] S4, Model Deployment and Application The trained optimal model weights are converted to ONNX format and deployed on an UAV-borne computing platform. In actual inspection missions, the UAV flies along a preset route and captures images in real time. The onboard computer runs the improved YOLOv13 network model of this invention to analyze the video stream in real time. Detection results (bounding boxes, categories, confidence scores) are transmitted back to the ground station in real time via wireless link, and an inspection report containing defect location, defect type, and image snapshots is automatically generated for inspection personnel to review and process. Field trials have shown that the system can effectively identify hidden dangers such as missing high-altitude bolts and unopened pins that are difficult to detect manually, greatly improving the efficiency and quality of construction equipment defect detection.

[0046] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for inspecting the construction quality of transmission lines using adaptive dynamic grouping, characterized in that the method... include: S1. Obtain photos of the power transmission line construction site; S2. Construct an improved network model for YOLOv13; S3. Input the images of the power transmission line construction site into the improved YOLOv13 network model for detection, and output the detection results of construction components and defect categories; The improved YOLOv13 network model consists of a backbone network, a neck network, and a detection head. Images of the power transmission line construction site are input into the backbone network for processing. The output of the backbone network is input into the neck network for processing, and the output of the neck network is input into the detection head for processing. The detection results of the construction components and defect categories are then output. Improvements include: introducing an adaptive dynamic grouping convolution module in the backbone and neck network, which replaces pointwise convolution in the depthwise separable convolution module with grouped convolution; the number of groups in the grouped convolution in the adaptive dynamic grouping convolution module is a dynamically adjustable parameter; performing gradient-feedback-based dynamic pruning on the neck network; optimizing shallow feature enhancement of the neck network; and directly connecting the detection head to the output of the feature splicing module of the neck network.

2. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 1, characterized in that, The backbone network includes a first ordinary convolutional layer, a second ordinary convolutional layer, a first improved DS-C3K2 module, a third ordinary convolutional layer, a second improved DS-C3K2 module, a first adaptive dynamic grouping convolutional module, a first improved region attention-enhanced cross-feature module, a second adaptive dynamic grouping convolutional module, and a second improved region attention-enhanced cross-feature module, all connected in sequence. The convolutions in the first and second improved DS-C3K2 modules are replaced with grouped convolutions in the adaptive dynamic grouped convolution module; the first and second improved region attention-enhanced cross-feature modules replace the linear positional encoding in the original YOLOv13 region attention-enhanced cross-feature module with positional encoding based on periodic functions.

3. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 2, characterized in that, The neck network includes a first upsampling module. The output of a second improved region attention-enhanced cross-feature module and an adaptive association enhancement module, concatenated, is input to the first upsampling module for processing. The output of the first upsampling module, the first improved region attention-enhanced cross-feature module, and the adaptive association enhancement module, concatenated, is input to a first feature concatenation module for processing. The output of the first feature concatenation module is input to a third improved DS-C3K2 module for processing. The output of the third improved DS-C3K2 module is input to the second upsampling module for processing. The output of the first improved DS-C3K2 module is input to a downsampling module for processing. The output of the downsampling module, the second upsampling module, the adaptive association enhancement module, and the second improved DS-C3K2 module, concatenated, is input to the second feature concatenation module for processing. The output of the module is processed by the fourth improved DS-C3K2 module. The output of the fourth improved DS-C3K2 module is concatenated with the output of the adaptive association enhancement module and then processed by the fourth ordinary convolutional layer. The output of the fourth ordinary convolutional layer, the output of the third improved DS-C3K2 module, and the output of the adaptive association enhancement module are concatenated and then processed by the third feature concatenation module. The output of the third feature concatenation module is processed by the fifth improved DS-C3K2 module. The output of the fifth improved DS-C3K2 module is processed by the fifth ordinary convolutional layer. The output of the fifth ordinary convolutional layer, the output of the second improved region attention-enhanced cross-feature module, and the output of the adaptive association enhancement module are concatenated and then processed by the fourth feature concatenation module. The outputs of the second, third, and fourth feature concatenation modules serve as the output of the neck network. The first, second, third, fourth, and fifth ordinary convolutional layers have the same structure; the first, second, third, fourth, and fifth improved DS-C3K2 modules have the same structure; the first and second improved DS-C3K2 modules have the same structure; the first and second improved region attention-enhanced cross-feature modules have the same structure; the neck network also includes a full-process aggregation and distribution tunnel.

4. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 3, characterized in that, The detection head includes a first detection head, a second detection head, and a third detection head. The output of the second feature splicing module is input into the first detection head for processing. The first detection head outputs the detection results of small-scale construction components and defect categories, completing the detection of small targets. The output of the third feature splicing module is input into the second detection head for processing. The second detection head outputs the detection results of medium-scale construction components and defect categories, completing the detection of medium-scale targets. The output of the fourth feature splicing module is input into the third detection head for processing. The third detection head outputs the detection results of large-scale construction components and defect categories, completing the detection of large targets.

5. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 1, characterized in that, The construction components include: insulators, vibration dampers, equalizing rings, wire clamps, bolts, and pins; the defect categories include: missing bolts, pins with cotter pins not opened, vibration damper slippage, insulator spontaneous explosion, hardware installed backwards, and conductor damage.

6. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 1, characterized in that, The adaptive dynamic grouped convolution module replaces the pointwise convolution in the depthwise separable convolution module with grouped convolution. The number of parameters and computational cost within the grouped convolution are as follows: ； ； in, For the number of parameters, Input the number of channels. Number of output channels The kernel size is the convolution kernel size. and These represent the height and width of the feature map, respectively. Number of groups For computational quantity; In the adaptive dynamic grouped convolution module, the number of groups G in the grouped convolution is a dynamically adjustable parameter.

7. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 1, characterized in that, The number of groups G in the adaptive dynamic grouped convolution module is also adjusted according to the feedback signal of the detection head: based on the average accuracy of the defect category, the number of groups G in the adaptive dynamic grouped convolution module on the corresponding target detection path is adjusted.

8. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 1, characterized in that, The specific process of performing gradient-feedback-based dynamic pruning on the neck network is as follows: During the training of the improved YOLOv13 network model, the importance metric of the full-process aggregation and distribution tunnels in the neck network is calculated periodically. For full-process aggregation and distribution tunnels with an importance metric lower than the dynamic threshold τ, they are masked in subsequent training processes; the formula is as follows: ； in, For importance measurement, N represents the total aggregation and distribution tunnel. The number of associated feature maps For the first Each feature map For loss function, For the first The gradient of each feature map with respect to the loss function. It is the Euclidean norm.

9. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 1, characterized in that, The aforementioned shallow feature enhancement and optimization for the neck network specifically involves: during the fusion process of the neck network, the output of the downsampling module, i.e., the shallow features... The output of the second upsampling module is the deep feature. The fusion process is performed using the following formula: ； in, As a feature of fusion, for convolution, For splicing operations, For upsampling operation, For deep features, This is a shallow feature.

10. The method for inspecting the construction quality of transmission lines using adaptive dynamic grouping according to claim 1, characterized in that, The improved YOLOv13 network model uses a composite loss function during training, as shown in the following formula: ； in, For composite loss function, for loss, For target confidence loss, For classifying losses, , , They are respectively The weighting coefficients for loss, target confidence loss, and classification loss.

Images