Transmission tower defect identification method based on improved YOLOv11 algorithm

Abstract

A transmission tower defect identification method based on an improved YOLOv11 algorithm belongs to the technical field of defect identification, and designs an improved YOLOv11 network model, firstly designs and adds a HorNet high-order interactive recursive gating network module in a Backbone network, and through gating convolution and recursive structures, fuses information of different widths, and improves the capability of accurately identifying target defects; secondly, a Conv standard convolution module in the Backbone network is replaced by a GhostConv ghost convolution module, so that the calculation amount and the parameter amount are greatly reduced; and finally, a RepViTblock re-parameterization visual transformation module is added in the Neck neck network design, each pixel position in the input image is mapped to a vector space, and the precision and efficiency of target detection are improved. The method is high in identification accuracy and high in identification efficiency, and fully guarantees the safe and stable operation of the power transmission line.

Description

Technical Field

[0001] This invention belongs to the field of defect identification technology, and in particular relates to a method for identifying defects in transmission towers based on an improved YOLOv11 algorithm. Background Technology

[0002] As the scale of power transmission networks continues to expand, transmission towers, as crucial supporting facilities for power transmission, directly impact the safety and reliability of the power system. In some areas, transmission lines traverse complex terrain, are widely distributed, and present significant inspection challenges, placing higher demands on the health monitoring of transmission towers. Regular inspections of transmission towers and transmission lines are essential for timely identification and resolution of problems, ensuring their safe and stable operation.

[0003] Currently, power line inspections are mainly conducted through a combination of manual inspections and drone inspections. Manual inspections are inefficient, unable to cover the vast areas of power transmission lines, and struggle to reach the site promptly in complex terrain or during disasters. They are also difficult to implement in complex terrain and high-altitude areas, posing significant safety hazards. Drone technology is playing an increasingly important role in power line inspections due to its unique advantages. However, drone inspections suffer from several drawbacks: loss of feature information in identifying defects on transmission towers, low accuracy in detecting small-target defects, and frequent missed and false detections. Therefore, deploying a suitable transmission line defect detection model and defect identification method for drone-based power line inspections is of significant practical importance for maintenance and repair work. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for identifying defects in power transmission towers based on an improved YOLOv11 algorithm, which addresses the technical problems in current UAV power inspection processes, such as the easy loss of feature information, low accuracy in detecting small target defects in power transmission towers, and frequent missed and false detections.

[0005] The method for identifying defects in transmission towers based on the improved YOLOv11 algorithm includes the following steps, which are performed sequentially:

[0006] Step 1: Install a power transmission tower defect identification system on a drone equipped with a high-definition visible light and infrared thermal imager camera. The power transmission tower defect identification system includes a dataset storage and preprocessing module, a defect identification module, and an algorithm evaluation module.

[0007] Step 2: The images of tower defects captured by the drone and the corresponding defect categories are combined into a raw dataset. The raw dataset is augmented through the dataset storage and preprocessing module. The defective parts of the transmission towers in the images are labeled using an image data annotation tool. Each labeled image will generate a corresponding text file containing the defect category and the position coordinates of the four corners of the label box. The augmented dataset is divided into a training set and a validation set according to a set ratio.

[0008] Step 3: Construct an improved YOLOv11 network model in the defect identification module.

[0009] The existing YOLOv11 model includes an input layer, a backbone network, a neck network, and a head detection head. The improved YOLOv11 network model is based on the existing YOLOv11 model. First, a HorNet high-order interactive recursive gating network module is added to the backbone network of the existing YOLOv11 model, and then recursively gated convolutions are performed. n Conv enables extended and translationally equivalent high-order spatial interactions, improving the ability to accurately identify target defects. Secondly, the GhostConv convolutional module replaces the Conv convolutional module in the backbone network of the existing YOLOv11 model, reducing computation and parameter count, and reducing model bloat while maintaining recognition accuracy. Finally, the RepViTblock reparameterized visual transformation module is added to the Neck network of the existing YOLOv11 model, separating the token mixer and channel mixer, and transforming the complex multi-branch structure during training into a single-path convolutional layer, eliminating all branch computations and memory overhead during inference.

[0010] Step 4: Train the improved YOLOv11 network model using the training set, and then validate the trained improved YOLOv11 network model using the validation set to obtain the trained improved YOLOv11 network model. Use the trained improved YOLOv11 network model to identify defects in transmission towers.

[0011] The HorNet high-order interactive recursive gating network module includes a recursive gated convolutional module g. n Conv and an FFN feedforward network layer, the recursive gated convolution module g n Conv is a convolutional operation based on recursive and gated convolutions for feature fusion, enabling convolutional neural networks to perform high-order spatial interactions, thereby improving the spatial interaction capabilities of downstream tasks. Recursive gated convolutions g n Conv is built using standard convolution, linear projection, and element-wise multiplication, which can continuously improve dense prediction performance while reducing computational cost.

[0012] Among them, the recursive gated convolution module g n The core method of Conv is gated convolution gConv, where the input feature of gated convolution gConv is x, x∈R. HW×C R HW×C Let gConv represent a real space of dimension HW×C. The output y = gConv(x) of the gated convolution gConv is specifically represented as follows:

[0013]

[0014] Where φ in , φ out It is a linear projection layer that performs channel blending, f is a depthwise convolution, H and W represent the width and height of the input image, respectively, and C represents the number of channels; This represents the initial principal feature in the real space of dimension HW×C; Represents the initial adjacency guiding feature in the real space of HW×C dimensions; R HW×2C Let p0 represent a real number space of dimension HW×2C; p0 represents the initial principal feature; q0 represents the initial adjacent guiding feature.

[0015] The above formula explicitly introduces the random initial principal characteristics through element-wise multiplication. and random initial adjacent guiding features The interactions between them are considered as first-order interactions, because each random initial principal feature... Guided features only with its random initial neighbors The interaction occurs once;

[0016] Recursive gated convolution module g n Conv further enhances model capacity by introducing higher-order interactions;

[0017] Formally, it is first achieved through a linear projection layer φ in A set of projection features and adjacent guided features is obtained, and is specifically represented as follows:

[0018]

[0019] In the formula, C0 represents the set of adjacent guiding features from the initial adjacent guiding features to the (n-1)th recursion, C0 represents the initial number of feature channels, and C... k The number of feature channels after the k-th recursion is represented by n, and the total number of recursions is represented by n. Represents the initial principal features in the real space of dimension HW×C0. Represents the initial adjacent guiding features in the real number space of dimension HW×C0. This indicates that the (n-1)th recursion occurs in HW×C. n-1 Adjacency-guided features in a real space of dimension Represent a real space of dimension HW×C, where C k For the k-th channel dimension of higher-order interactions;

[0020] Then perform gated convolution recursively:

[0021] p k+1 =f k (q k )⊙g k (p k ) / α,k=0,1,…,n-1 (7);

[0022] In the formula, dividing by α in each recursive process is for stable training, k represents the number of times the gated convolution is executed recursively, n represents the total number of recursions, and f k G represents the k-th depth-wise convolution operation. k Let q represent the principal feature transformation function. k p represents the guiding feature of the k-th recursive execution of the gated convolution. k+1 This represents the new principal feature obtained after the k-th recursive execution of the gated convolution;

[0023]

[0024] Identity represents the identity transformation, and Linear represents the transformation of dimension C. k-1 The input is transformed by a linear transformation to a dimension of C. k The output space;

[0025] Finally, output q from the last recursion. n Input to projection layer φ out , thus obtaining the recursive gated convolution g n The result of Conv;

[0026] Each channel dimension C of higher-order interactions k Set to an exponentially decreasing form:

[0027]

[0028] The GhostConv convolutional module includes the following operation steps:

[0029] The first step involves using a 1x1 convolutional layer to generate m intrinsic feature maps through the regular convolutional part, where m = t / s, and t is the final number of feature maps required, and s is the number of transformations.

[0030] The second step involves applying linear transformation operations to each generated intrinsic feature map in parallel s-1 linear operations to generate m*(s-1) ghost feature maps.

[0031] Then, the m intrinsic feature maps generated in the first step are preserved by identity mapping and the m*(s-1) ghost feature maps generated by the second step through linear transformation are concatenated in the channel dimension to finally obtain m*s feature maps as output.

[0032] The RepViTblock module is an improvement on the existing lightweight Vision Transformer module, referred to as the lightweight ViT module. The existing lightweight ViT module simulates the token mixer and channel mixer by splitting them in the high-efficiency mobile module MobileNetV3, and enhances model learning during training through structural reparameterization techniques.

[0033] The high-efficiency mobile module MobileNetV3 uses 1×1 extended convolution and 1×1 projection layer to achieve interaction between channels. After the 1×1 extended convolution, a 3×3 depthwise convolution DWConv is performed to achieve spatial information fusion, which couples the token mixer and the channel mixer together.

[0034] The following improvements were made to the efficient motion module in the RepViTblock reparameterized visual transformation module:

[0035] To separate the token mixer and the channel mixer, a 3×3 deep convolutional (DW) module is first used to extract spatial features, while parallel 1×1 ordinary convolutions are used to process the relationships between channels to form a deep convolutional (DW) layer. In the token mixer, a structure reparameterization technique is used for the DW layer, so that the multi-branch topology is merged into a single 3×3 deep convolutional (DW). The squeeze-excitement network layer (SE) is set after the merged 3×3 deep convolutional (DW) to calculate the importance weight of each channel. Then, the weights of each channel are recalibrated to realize the information exchange between channels. Finally, 1x1 convolutions are used to complete the final channel mixing and dimensionality transformation.

[0036] The categories of defects in transmission towers include infrastructure failures, conductor and ground wire failures, insulator failures, auxiliary facility failures, vibration damper failures, small-sized hardware defects, broken or loose cable strands, foreign objects on the tower, or the tower base being obstructed by debris.

[0037] The defects of the transmission towers include missing pins, pins not being opened, equipotential rings falling off, loose bolts, damaged insulators, or insufficient number of pre-twisted wires in the vibration damper.

[0038] The data augmentation includes the following six methods:

[0039] Flip horizontally;

[0040] Flip vertically;

[0041] Enhance colors by removing a random proportion of colors from the input image;

[0042] To enhance brightness, each image is converted to a color space with a brightness-related channel, that channel is extracted, multiplied by a factor between 0.5 and 1.5, added with a value between -30 and 30, and then converted back to the original color space.

[0043] Pixel-by-pixel random value addition: Random values ​​are added to each pixel of the image. The added pixel value is a random value between -40 and 40, and each value is sampled for each pixel.

[0044] Original color channel histogram equalization.

[0045] The evaluation metrics used in step four include recognition accuracy, recall, and average recognition accuracy.

[0046] Through the above design scheme, the present invention can bring the following beneficial effects:

[0047] This invention combines an improved YOLOv11 algorithm to design an efficient, reliable, and accurate method for identifying defects in power transmission towers, specifically addressing the unique scenarios and requirements of power transmission towers. The main approach involves designing an improved YOLOv11 network model.

[0048] In improving the YOLOv11 network model, firstly, a HorNet (High-Order Interactive Recurrent Gated Network) module was added to the backbone network design. Through gated convolutions and recursive structures, efficient, scalable, and translationally equivalent high-order spatial interactions were achieved. This combined the wide feature maps of the previous layer with the narrow feature maps of the current layer, increasing the model's expressive power by fusing information of different widths, thereby improving the effectiveness and flexibility of feature extraction and enhancing the ability to accurately identify target defects. Secondly, considering the needs for processing efficiency and model lightweighting, a high-efficiency feature extraction module, the GhostConv convolution module, was designed. The Conv convolutional module in the backbone network is replaced by an integral module. Through a strategy of "intrinsic features + inexpensive linear transformation," core features are generated with a small number of standard convolutions, and then similar "ghost features" are expanded through simple operations. This significantly reduces computational cost and parameter count, optimizes the feature extraction process, and reduces computational resource consumption. Finally, the RepViTblock reparameterized visual transformation module is added to the Neck network design. Employing a structural reparameterization strategy, the token mixer and channel mixer are decoupled during training and their parameters are recombined during inference, effectively enhancing the model's feature learning ability during training. SE layers are deployed in an interleaved manner to maximize model accuracy gains at the cost of negligible latency. Attached Figure Description

[0049] The present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0050] Figure 1 This is a schematic diagram of the existing YOLOv11 network structure;

[0051] Figure 2 This is a schematic diagram of the improved YOLOv11 network model in the transmission tower defect identification method based on the improved YOLOv11 algorithm of the present invention;

[0052] Figure 3 This is a schematic diagram of the HorNet high-order interactive recursive gating network module structure in the transmission tower defect identification method based on the improved YOLOv11 algorithm of this invention;

[0053] Figure 4 The recursive gated convolution g in the transmission tower defect identification method based on the improved YOLOv11 algorithm of this invention is... n A structural diagram of Conv;

[0054] Figure 5 The recursive gated convolution g in the transmission tower defect identification method based on the improved YOLOv11 algorithm of this invention is... n A comparison diagram of the principles of Conv, ordinary convolution operations, and dynamic convolution operations;

[0055] Figure 6 This is a flowchart illustrating the operation of the GhostConv convolution module in the transmission tower defect identification method based on the improved YOLOv11 algorithm of this invention.

[0056] Figure 7 This is a schematic diagram of a conventional convolutional module structure;

[0057] Figure 8 This is a schematic diagram of the GhostConv convolutional module in the transmission tower defect identification method based on the improved YOLOv11 algorithm of this invention;

[0058] Figure 9 This is a schematic diagram of the structure of an existing high-efficiency mobile module;

[0059] Figure 10 This is a schematic diagram of the principle structure of the RepViTblock reparameterized visual transformation module for improving the efficient moving module in the transmission tower defect identification method based on the improved YOLOv11 algorithm of this invention.

[0060] Figure 11 This is a comparison diagram of bolt loosening between the improved YOLOv11 network model and the existing YOLOv11 network structure used in this embodiment of the invention;

[0061] Figure 12 This is a comparison chart of the detection results of the pin not being opened using the improved YOLOv11 network model and the existing YOLOv11 network structure in this embodiment of the invention;

[0062] Figure 13 This is a comparison chart of the pin loss detection results using the improved YOLOv11 network model and the existing YOLOv11 network structure in this embodiment of the invention;

[0063] Figure 14 This is a comparison chart of insulator damage detection results using the improved YOLOv11 network model and the existing YOLOv11 network structure in this embodiment of the invention.

[0064] Figure 15 This is a comparison chart of the detection results of insufficient pre-twisted wires in the vibration damper using the improved YOLOv11 network model and the existing YOLOv11 network structure in this embodiment of the invention.

[0065] Figure 16 This is a comparison chart of the detection results of the equalizing ring falling off using the improved YOLOv11 network model and the existing YOLOv11 network structure in this embodiment of the invention. Detailed Implementation

[0066] I. Analysis of Defect Categories and Mechanisms of Transmission Towers

[0067] 1) Basic structure of transmission towers

[0068] In recent years, with the continuous growth of electricity demand and the accelerated construction of power grids, the number of transmission towers installed has shown a significant increasing trend. As key supporting equipment, the number of towers installed is closely related to the scale of power grid construction. Transmission towers, as the main supporting structure in transmission lines, are used for long-distance crossing connections of transmission lines, while maintaining safe distances between conductors and ground wires and other structures, as well as between conductors and ground wires themselves. The main basic structure of a transmission tower includes the foundation, tower body, crossarms, insulator strings, lightning protection wires, and hardware. The foundation is the lowest part of the tower, usually made of poured concrete. Its function is to firmly fix the tower to the ground, bear all the loads transmitted from the tower, and distribute them into the foundation. Crossarms are installed on the tower body to extend and support the insulator strings and conductors, ensuring the necessary safe distances between conductors and between conductors and the tower body. Insulator strings connect the crossarms and conductors; their core function is to maintain electrical insulation between the conductors and the grounded tower. Lightning protection wires are installed at the top of poles and towers, and their main function is to protect the lines from direct lightning strikes and improve the lightning resistance level of the lines.

[0069] Fittings are indispensable and crucial components in power transmission towers and overhead transmission lines. There are many types of fittings, mainly used to connect, secure, and protect power equipment such as conductors, insulators, and towers. To maintain the stability of the line structure, these critical components must withstand significant tensile forces. Furthermore, due to their constant exposure to the complex outdoor environment, they are prone to displacement, tilting, and damage, affecting the stability of the structure between conductors and towers. Failure to detect these issues promptly can lead to line damage accidents.

[0070] 2) Types and hazards of defects in transmission towers

[0071] During long-term operation, transmission towers may develop various defects due to weathering, corrosion, geological changes, floods, lightning strikes, icing, external impacts, and construction quality issues. In power line inspection, common faults and defects in transmission towers can be categorized into nine main types, including: infrastructure faults, conductor / ground wire faults, insulator faults, auxiliary facility faults, vibration damper faults, small-size hardware defects, broken or scattered cable strands, foreign objects on the tower (such as bird nests or beehives), and obstruction of the tower base by debris. The most common and damaging faults include missing pins, pins not being opened, detached equipotential rings, loose bolts, damaged insulators, and insufficient pre-twisted wires in the vibration damper.

[0072] Missing or unopened cotter pins: In transmission towers, cotter pins are anti-loosening devices used in mechanical connections. They primarily prevent relative movement between bolts and nuts, and prevent the nut from loosening due to vibration, wind, or other external forces. Cotter pins form a locking structure by inserting into the slots of the nut and bolt, preventing the nut from loosening under vibration or other external forces. If cotter pins are missing at the connection between the transmission tower and the transmission line, the transmission tower is highly susceptible to collapse, leading to local power grid failures and power outages.

[0073] Equalizing ring detachment: The equalizing ring is located at the connection between the insulator and the conductor, used to control corona discharge on the fittings and protect both the fittings and the insulator. Traditionally, equalizing rings are fixed to the fittings at both ends of the insulator using locking screws and pressure plates. The connection between the connecting rod and the ring body uses only one set of locking screws. This single-point fixing method is prone to loosening under strong winds or intense vibrations. Installing an equalizing ring can make the voltage distribution of the insulator string more uniform. When the equalizing ring detaches, it causes uneven voltage distribution in the insulator string, leading to excessively high local electric field strength, which can easily trigger corona discharge and affect the power supply reliability of the transmission line.

[0074] Loose bolts: For existing transmission towers, angle steel connections are often used at the joints to facilitate efficient assembly by construction workers in complex and changing on-site environments. Bolts, as key components of transmission towers, are mainly used to secure various connecting parts of the tower. In outdoor environments, prolonged exposure to strong winds, rain, and snow can cause them to loosen. While slight loosening of a few bolts may not have a significant impact on the entire tower, if loose bolts remain unkempt for an extended period, they can eventually fall off under the influence of strong winds, ground vibrations, or other external forces. This significantly reduces the tower's load-bearing capacity. Consequently, in the event of strong winds or earthquakes, there is a high probability of severe tower collapse, affecting the stable operation of the transmission system and causing economic losses.

[0075] Insulator damage: Insulators are an indispensable component of overhead transmission lines. Their main function is to support and fix current-carrying conductors, capable of withstanding both mechanical loads and high voltage. Through fixed suspension, they not only extend transmission distances but also reduce capacitive reactance between conductors, thereby lowering current losses. Insulators are exposed to the outdoor environment for extended periods, making them susceptible to damage from icing, lightning strikes, pollution, and other factors. These problems can affect the normal operation of transmission lines and even threaten the safety of the power system. Therefore, to ensure the safe and reliable operation of transmission lines, regular inspections of insulators are essential, and any faults must be repaired promptly. Rapid identification and detection of insulator defects, and timely fault handling, are crucial for ensuring the stable operation of the power system.

[0076] Insufficient prestressed wires in vibration dampers: Vibration dampers are specialized devices used to suppress vibrations in transmission lines caused by wind or other factors, reducing impact on the transmission lines and improving their stability. Vibration is particularly critical in long-distance, high-voltage transmission lines. Factors such as dryness and ice cover can cause line vibrations, potentially damaging supports and towers. Prestressed wires in vibration dampers act as flexible connections, effectively absorbing and reducing conductor vibration energy. Insufficient prestressed wires reduce the damping effect of the damper, leading to increased conductor vibration and further accelerating conductor fatigue and damage. Insufficient prestressed wires can also cause clamps to slip on the conductor, affecting their fixation. Slipping clamps can create additional stress concentration on the conductor, increasing the risk of conductor damage.

[0077] II. Experimental Preparation

[0078] The construction of a defect sample set for power transmission towers is crucial to the quality of power transmission tower defect detection, as it directly affects the training effect of the network model. To ensure the effectiveness of the training process, a large number of high-quality samples are needed to support model learning and extract feature information of the detected targets. During filming, a drone with stable flight capabilities and equipped with high-definition visible light and infrared thermal imagers should be selected.

[0079] III. Examples

[0080] This embodiment uses six types of defects as examples: missing pins, pins not being opened, equipotential rings falling off, loose bolts, damaged insulators, and insufficient pre-twisted wires in vibration dampers. The identification of other types of defects only differs in the training and validation sets used. The transmission tower defect identification method based on the improved YOLOv11 algorithm includes the following steps:

[0081] 1) Dataset preprocessing

[0082] The dataset, composed of 1171 original images of tower defects captured by drones, includes various defects against different backgrounds, such as loose bolts, missing pins, pins not being opened, detached equalizing rings, damaged insulators, and insufficient pre-twisted wires in vibration dampers. However, due to the limited number and scale of defects in the original dataset, overfitting of the model may occur. Therefore, this paper applies six data augmentation techniques to the original dataset: horizontal flipping; vertical flipping; color enhancement (removing a random proportion of color from the input images); brightness enhancement (converting each image to a color space with a brightness-related channel, extracting this channel, multiplying it by a factor between 0.5 and 1.5, adding a value between -30 and 30, and then converting it back to the original color space); pixel-by-pixel random value addition (adding a random value between -40 and 40 to each pixel of the image, sampling each value for each pixel); and histogram equalization of the original color channels. These data augmentation techniques expand the dataset to 3994 images.

[0083] The X-AnyLabeling (2.5.3) image data annotation tool was used to annotate the defective parts of the transmission towers in the images. Each annotated image generated a corresponding text file containing the defect category and the coordinates of the four corners of the label box. The annotation process produced 6314 labels. The image data was divided into training and validation sets in a 7:3 ratio and supplemented by an independent test set including 200 unprocessed images for comprehensive model evaluation.

[0084] 2) Hardware configuration and hyperparameter configuration

[0085] The hardware configuration is shown in Table 1, and the experimental hyperparameter settings are shown in Table 2. All experiments were conducted under the same configuration.

[0086] Table 1 Experimental Hardware Environment

[0087]

[0088] Table 2 Experimental hyperparameters

[0089]

[0090] 3) Algorithm evaluation metrics

[0091] To more objectively and accurately verify the model's effectiveness and stability, we use Precision (P), Recall (R), and Mean Average Precision (mAP) to evaluate the model's performance.

[0092] (1) Precision is a metric that measures how many of the targets predicted as positive (i.e., detected targets) are actually true targets. It primarily focuses on the accuracy of the prediction results. The calculation formula is as follows:

[0093]

[0094] In the formula, TP represents a sample that is predicted to be a defect but is actually a defect; FP represents a sample that is predicted to be a defect but is not actually a defect.

[0095] (2) Recall is an important metric for measuring the coverage of all real-world targets by a model. It assesses how many targets the model successfully detects from all real-world targets, i.e., how many targets are accurately identified. Recall focuses on the model's completeness, reflecting the extent to which the model can avoid missed detections. The calculation formula is as follows:

[0096]

[0097] In the formula, TP represents a sample that is predicted to be a defect but is actually a defect; FN represents a sample that is predicted to be a non-defect but is actually a defect.

[0098] (3) Mean average precision (mAP) reflects the average accuracy of recognition across all categories. The specific calculation formula is as follows.

[0099]

[0100] In the formula, p(r) represents the recognition accuracy under different recall rates; AP is the average recognition accuracy of a certain type of defect; N represents the number of defect categories.

[0101] The intersection-union ratio (IU) measures the degree of overlap between predicted and ground truth bounding boxes. It is calculated by dividing the area of ​​the intersection of the predicted and ground truth bounding boxes by the area of ​​their union. mAP@50 represents the mAP value when the IU threshold is set to 50%.

[0102] Step 3: YOLOv11 is an excellent algorithm in the field of object detection. Compared to the YOLOv8 model, it replaces the CF2 module with C3K2 and adds a feature enhancement C2PSA module after the SPPF module. It also incorporates the detection head concept from YOLOv10 into the YOLOv11 detection head, using a depth-separable method to reduce redundant computation and improve efficiency. YOLOv11 achieves significant improvements in accuracy and speed, providing more accurate bounding box rendering and bringing a powerful and efficient solution to the field of object detection.

[0103] 1) Existing YOLOv11 algorithm model

[0104] The existing YOLOv11 includes an input terminal, a backbone network, a neck network, and a head detection head. The existing YOLOv11 network structure is as follows: Figure 1 As shown.

[0105] The backbone network includes standard convolutional Conv modules, the C3k2 module (a more efficient cross-stage partial bottleneck module), fast spatial pyramid pooling (SPPF), and cross-channel partial spatial attention (C2PSA). The C3k2 module introduces two configurations: C3K (a cross-stage partial bottleneck module with variable kernels) and BottleNeck (a bottleneck module). When the parameter is False, the traditional convolutional structure BottleNeck is maintained; when the parameter is True, the C3K module is used. The C2PSA cross-stage partial pyramid squeeze attention module is a new attention module added in the YOLOv11 network, located at the end of the backbone network. By performing pyramid-shaped splitting and attention weighting of the feature maps, it makes the model focus more on key regions of the image, suppressing interference from irrelevant information, thereby improving the model's ability to capture important features.

[0106] The neck network also uses the C3k2 module to replace the C2F cross-stage dual convolution bottleneck module. By using a more efficient feature fusion strategy, it fully combines feature information from different scales, improving the model's ability to detect targets of different sizes.

[0107] The Head detection head uses a decoupled classification head and regression head to decouple category features and location features, thereby improving the model's detection efficiency. The regression head uses normal convolution operations to predict the location information of the bounding box, while the classification head uses depthwise separable convolution (DWConv) to reduce the number of model parameters and computational cost, thus ensuring that the model's inference speed is accelerated while improving detection accuracy.

[0108] 2) Improve the YOLOv11 network model

[0109] To address the issues of easy loss of feature information, low accuracy in detecting small target defects on transmission towers, and frequent missed and false detections during UAV power line inspections, this paper proposes three improvement measures based on YOLOv11n to enhance the defect detection performance of transmission towers. First, a HorNet high-order recurrent network module was added to the Backbone network design. Through gated convolutions and recurrent structures, efficient, scalable, and translationally equivalent high-order spatial interactions were achieved, enhancing the effectiveness and flexibility of feature extraction and improving the ability to accurately identify target defects. Second, while maintaining recognition accuracy, the model's bloat was reduced by designing a GhostConv convolution module to replace the Conv module in the Backbone network. This module uses an "intrinsic features + inexpensive linear transformation" strategy to generate core features with a small number of standard convolutions, and then expands these into similar "ghost features" through simple operations, significantly reducing computation and parameter count. Finally, a RepViTblock reparameterized visual transformation module was added to the Neck network design. This module separates the token mixer and channel mixer, and transforms the complex multi-branch structure during training into a single-path convolutional layer, eliminating all branch computations and memory overhead during inference. The improved YOLOv11 model is shown below. Figure 2 As shown.

[0110] (1) HorNet high-order interactive recursive gating network module

[0111] The HorNet module primarily addresses the limitations of YOLO in processing horizontal information. In the traditional YOLO network architecture, information is typically passed vertically through stacked convolutional layers, which can limit the model's ability to capture horizontal features. The HorNet module introduces lateral connections to enhance the network's efficiency and accuracy in processing horizontal feature flows. The HorNet module may combine wide feature maps from the previous layer with narrow feature maps from the current layer at some stage, increasing the model's expressive power by fusing information of different widths.

[0112] like Figure 3 As shown, the HorNet module contains a g n Conv recursively gated convolutional modules and an FFN feedforward network layer, where g n The Conv recursive gated convolution module is used for convolutional operations in feature fusion to improve the spatial interaction capabilities of downstream tasks. It proposes a Recursive Gated Convolution (g) based on recursion and gated convolution. n Conv enables convolutional neural networks to perform high-order spatial interactions. Recursive gated convolution g n The structure of Conv is as follows Figure 4As shown, it is constructed using standard convolution, linear projection, and element-wise multiplication, which can continuously improve dense prediction performance while reducing computational cost. The pyramid-like design, which progressively increases channel width during spatial interactions, enables high-order spatial interactions within finite complexity constraints.

[0113] like Figure 5 The diagram illustrates the core idea, analyzing the interaction between features (red blocks) and their surrounding areas (gray blocks) in different operations. Figure 5 In the diagram, (a) represents a standard convolution operation that does not consider spatial information interaction. (b) represents a dynamic convolution operation that utilizes dynamic weights to consider information interaction within the surrounding region, resulting in stronger model performance. (c) represents a convolution operation using recursive gated convolution g. n Conv operations efficiently enable information exchange of arbitrary orders. The fundamental operational trend in visual modeling suggests that the expressive power of a model can be improved by increasing the order of spatial interactions.

[0114] The core method is gated convolution (gConv), let x∈R HW×C Given the input features, the output y = gConv(x) of the gated convolution (gConv) can be written as:

[0115]

[0116] Where φ in , φ out This is a linear projection layer that performs channel blending, where f is a depthwise convolution, H and W represent the width and height of the input image, respectively, and C represents the number of channels. Let represent the initial principal features in the real space of dimension HW×C. R represents the initial adjacent guiding feature in the real space of HW×C dimensions. HW×2C Let p0 represent a real number space of HW×2C dimension, p0 represent the initial main feature, q0 represent the initial adjacent guiding feature, and p1 represent the main feature after the first gated convolution.

[0117] The above formula explicitly introduces the random initial principal characteristics through element-wise multiplication. and random initial adjacent guiding features The interactions between them are considered as first-order interactions, because each random initial principal feature... Guided features only with its random initial neighbors The interaction occurs once;

[0118] Recursive gated convolution g n Conv further enhances model capacity by introducing higher-order interactions. Formally, it first uses a linear projection layer φ inA set of projection features and adjacent guiding features is obtained:

[0119]

[0120] In the formula, C0 represents the set of adjacent guiding features from the initial adjacent guiding features to the (n-1)th recursion, C0 represents the initial number of feature channels, and C... k The number of feature channels after the k-th recursion is represented by n, and the total number of recursions is represented by n. Represents the initial principal features in the real space of dimension HW×C0. Represents the initial adjacent guiding features in the real number space of dimension HW×C0. This indicates that the (n-1)th recursion occurs in HW×C. n-1 Adjacency-guided features in a real space of dimension Represent a real space of dimension HW×C, where C k For the k-th channel dimension of higher-order interactions;

[0121] Then perform gated convolution recursively:

[0122] p k+1 =f k (q k )⊙g k (p k ) / α,k=0,1,…,n-1 (7);

[0123] In the formula, dividing by α in each recursive process is for stable training, k represents the number of times the gated convolution is executed recursively, and p k+1 f represents the new principal feature obtained after the k-th recursive execution of the gated convolution. k G represents the k-th depth-wise convolution operation. k Let q represent the principal feature transformation function. k This represents the guiding feature for the k-th recursive execution of the gated convolution;

[0124]

[0125] Identity represents the identity transformation, and Linear represents the transformation of dimension C. k-1 The input is transformed by a linear transformation to a dimension of C. k The output space.

[0126] Finally, the output of the last recursion is input into the projection layer φ. out , thus obtaining the recursive gated convolution g n The result of Conv. Recursive gated convolution g n Conv can realize n-order interactions between a feature and its surrounding features.

[0127] To ensure that higher-order interactions do not introduce excessive computational overhead, the channel dimension for each order is set to decrease exponentially:

[0128]

[0129] In the formula, C represents the number of channels.

[0130] (2) GhostConv convolution module

[0131] GhostConv is a design for building lightweight convolutional neural networks. Its core idea is to generate more feature maps through cheap operations, thereby significantly reducing the number of parameters and computational cost while maintaining model performance. Figure 6 The diagram shows the module's operation flow.

[0132] The structure of a conventional convolutional module is as follows: Figure 7 As shown, the GhostConv convolutional module structure is as follows: Figure 8 As shown, the difference between the two is obvious. The GhostConv ghost convolution module solves the problem in two steps. The regular convolution part uses a 1x1 convolutional layer to generate m intrinsic feature maps, m = t / s, where t is the final number of feature maps required and s is the number of transformations, usually including one identity transformation. The linear transformation part applies s-1 linear operations in parallel to each of the generated intrinsic feature maps to generate m*(s-1) ghost feature maps. The computational cost of this operation is much smaller than that of regular convolution. The m intrinsic feature maps generated in the first step are preserved by identity mapping and the m*(s-1) ghost feature maps generated by the linear transformation operation in the second step are concatenated in the channel dimension to finally obtain m*s feature maps as output.

[0133] (3) RepViTblock reparameterized visual transformation module

[0134] The core idea behind the RepViTblock reparameterized vision transformation module is derived from the existing lightweight ViT (Vision Transformer) module, which includes an important design feature: independent token mixers and channel mixers. The effectiveness of ViT mainly stems from its general token mixer and channel mixer architecture. The RepViTblock module is designed to simulate the existing lightweight ViT by splitting the token mixer and channel mixer in the efficient mobile module (MobileNetV3) and enhancing model learning during training through structural reparameterization techniques.

[0135] Figure 9It is the existing high-efficiency mobile module (MobileNetV3) with an optional squeeze-and-excitation network (SE) layer. Figure 10 This is a schematic diagram of the progressive improvement process of the RepViTblock module design, which separates the token mixer and the channel mixer through structural reparameterization technology.

[0136] like Figure 9 As shown, the existing high-efficiency mobile module (MobileNetV3) uses 1×1 extended convolution and 1×1 projection layers to achieve inter-channel interaction. After the 1×1 extended convolution, a 3×3 depth-wise separable convolution (DWConv) is performed to achieve spatial information fusion. This design couples the token mixer and the channel mixer together.

[0137] The improvement process of the high-efficiency mobile module in the RepViTblock module is as follows: Figure 10 As shown: To separate the token mixer and the channel mixer, the depthwise convolution (DW) module is first moved up to extract spatial features, and parallel 1×1 ordinary convolutions are used to handle inter-channel relationships. In the token mixer, a structure reparameterization technique is used in the DW layer to merge multi-branch topologies into a single depthwise convolution to enhance model learning during training and eliminate the computational and memory costs associated with skipped connections during inference. The squeeze-and-excitation network (SE) layer is also moved up to be placed after the DW module, calculates the importance weights of each channel, and then recalibrates and weights each channel to achieve strong inter-channel information interaction. Finally, 1×1 convolutions are used to complete the final channel mixing and dimensionality transformation.

[0138] Among them, optimizing the squeeze-and-excitation networks (SE) can compensate for the limitations of convolution in terms of lack of data-driven attributes, resulting in better performance. Compared with the higher resolution stages, the accuracy improvement obtained from the global average pooling operation provided by the SE layer is smaller in the lower resolution stages. Since the SE layer introduces a non-negligible computational cost, in order to achieve a better trade-off between model accuracy and inference latency, a novel deployment strategy for the SE layer is proposed. This strategy deploys the SE layer in an interleaved manner rather than in a block-by-block manner in all network stages, so as to maximize the model accuracy gain at the cost of negligible latency overhead.

[0139] Step 4: Experiments and Results Analysis. Ablation experiments are a common model analysis method in deep learning, primarily used to evaluate the independent contribution of specific modules, components, or techniques to the overall system. The basic idea is to selectively "remove" or "replace" a sub-module in the model while keeping other conditions constant. Compared to simple model comparisons, ablation experiments provide a more interpretive way for researchers to understand the interrelationships between parts of a complex system. Comparative experiments are a commonly used validation method in deep learning research, mainly evaluating the practical significance and effectiveness of the proposed improvements by comparing the performance of the improved model with the original model or other mainstream models.

[0140] 1) Comparison of accuracy across categories

[0141] This paper compares the detection accuracy of the benchmark algorithm and HGR-YOLO on the dataset presented in this paper for various categories. The results are shown in Table 3. Table 3 displays the AP data values ​​for six categories: loose bolts, pins not detached, missing pins, detached equalizing rings, damaged insulators, and insufficient pre-twisted wires in the vibration damper, as well as the overall mAP@50 data for all defect categories. The improved algorithm outperforms the original YOLOv11n algorithm in all six defect categories, demonstrating the adaptability of the HGR-YOLO algorithm designed in this paper to target detection at different scales. The improved accuracy for small target defects such as loose bolts, missing pins, and pins not detached shows that the algorithm can better focus on small targets in the image.

[0142] Table 3 Comparison of Detection Accuracy for Each Category

[0143]

[0144] 2) Ablation experiment

[0145] To demonstrate the effectiveness and rigor of the improved method presented in this paper, ablation experiments were designed and conducted on the dataset presented in this paper. Each group of experiments was conducted on the dataset presented in this paper under the same parameters. The experimental results are shown in Table 4. √ indicates the inclusion of the module. The baseline model used is the YOLOv11n model. Group A experiments added the HorNet module to the baseline model; Group B experiments improved the Conv module in the baseline model to the GhostConv module; Group C experiments added the RepViTblock module to the baseline model; and Group D experiments added all the improvements presented in this paper to the YOLOv11n model.

[0146] Table 4 Ablation Experiment

[0147]

[0148] Based on the ablation results, the following conclusions can be drawn. In experiments A, B, and C, the accuracy, recall, and mAP@50 were all improved. In experiment A, the introduction of the HorNet module improved recall by 6.1% and mAP@50 by 3.3%. In experiment B, the use of the GhostConv module improved the p-value by 1.1%, recall by 5.1%, and mAP@50 by 3.7%. In experiment C, the addition of the RepViTblock module improved mAP@50 by 0.5%. According to experiment D, the improved model proposed in this paper achieves a detection mAP@50 of 92.7%, representing a 5.0% improvement. Each improvement point contributes to the detection performance, meeting the accuracy requirements for defect detection in transmission towers under actual conditions.

[0149] 3) Comparative Experiment

[0150] To further analyze the performance of the improved model, this paper selects YOLOv10n, YOLOv9t, and YOLOv6 for comparative experiments. These models are all excellent algorithms in the field of target detection, possessing high accuracy, high speed, and good adaptability, and are widely used in various detection tasks. Among them, YOLOv10n and YOLOv9t represent the latest advancements in the YOLO series, achieving significant improvements in overall performance; YOLOv6 is a widely used benchmark model in the field of power grid defect detection. By comparing with these excellent models, the superiority of the improved algorithm in terms of detection accuracy when detecting defects in transmission towers can be more comprehensively evaluated. To ensure experimental fairness, all models were tested on the same dataset and under the same experimental environment.

[0151] Table 5 Comparative Experiments

[0152]

[0153] The experimental results are shown in Table 5. The comparative results show that the improved algorithm performs best in terms of mAP@50 compared to other algorithms. Furthermore, the improved model also performs well in terms of recall and detection speed.

[0154] 4) Visualization and analysis of test results

[0155] To more directly verify the detection effect of the improved model, comparison charts of the detection effects for loose bolts, pins not being opened, missing pins, damaged insulators, insufficient pre-twisted wires in the vibration damper, and detached equalizing rings were presented respectively. Figures 11 to 16The results are shown below. Left (a) displays the detection results of the YOLOv11 model, while right (b) displays the detection results of the improved YOLOv11 network model. As can be seen from the comparison, the improved model proposed in this paper detects all defect targets, significantly reducing the probability of false positives and false negatives. It also outperforms the original YOLOv11 model in terms of detection accuracy, demonstrating excellent detection performance. These results show that the improved algorithm presented in this paper has significant value in the field of transmission tower defects, achieving the expected results and can be initially applied to the inspection of transmission towers.

Claims

1. A method for identifying defects in transmission towers based on an improved YOLOv11 algorithm, characterized in that: The steps are as follows, and they are performed sequentially: Step 1: Install a power transmission tower defect identification system on a drone equipped with a high-definition visible light and infrared thermal imager camera. The power transmission tower defect identification system includes a dataset storage and preprocessing module, a defect identification module, and an algorithm evaluation module. Step 2: The images of tower defects captured by the drone and the corresponding defect categories are combined into a raw dataset. The raw dataset is augmented through the dataset storage and preprocessing module. The defective parts of the transmission towers in the images are labeled using an image data annotation tool. Each labeled image will generate a corresponding text file containing the defect category and the position coordinates of the four corners of the label box. The augmented dataset is divided into a training set and a validation set according to a set ratio. Step 3: Construct an improved YOLOv11 network model in the defect identification module. The existing YOLOv11 model includes an input layer, a backbone network, a neck network, and a head detection head. The improved YOLOv11 network model is based on the existing YOLOv11 model. First, a HorNet high-order interactive recursive gating network module is added to the backbone network of the existing YOLOv11 model, and then recursively gated convolutions are performed. n Conv enables extended and translationally equivalent high-order spatial interactions, improving the ability to accurately identify target defects. Secondly, the GhostConv convolutional module replaces the Conv convolutional module in the backbone network of the existing YOLOv11 model, reducing computation and parameter count, and reducing model bloat while maintaining recognition accuracy. Finally, the RepViTblock reparameterized visual transformation module is added to the Neck network of the existing YOLOv11 model, separating the token mixer and channel mixer, and transforming the complex multi-branch structure during training into a single-path convolutional layer, eliminating all branch computations and memory overhead during inference. Step 4: Train the improved YOLOv11 network model using the training set, and then validate the trained improved YOLOv11 network model using the validation set to obtain the trained improved YOLOv11 network model. Use the trained improved YOLOv11 network model to identify defects in transmission towers.

2. The method for identifying transmission tower defects based on the improved YOLOv11 algorithm according to claim 1, characterized in that: The HorNet high-order interactive recursive gating network module includes a recursive gated convolutional module g. n Conv and an FFN feedforward network layer, the recursive gated convolution module g n Conv is a convolutional operation based on recursive and gated convolutions for feature fusion, enabling convolutional neural networks to perform high-order spatial interactions, thereby improving the spatial interaction capabilities of downstream tasks. Recursive gated convolutions g n Conv is built using standard convolution, linear projection, and element-wise multiplication, which can continuously improve dense prediction performance while reducing computational cost. Among them, the recursive gated convolution module g n The core method of Conv is gated convolution gConv, where the input feature of gated convolution gConv is x, x∈R. HW×C R HW×C Let gConv represent a real space of dimension HW×C. The output y = gConv(x) of the gated convolution gConv is specifically represented as follows: p1=f(q0)⊙p0∈R HW×C ,y=φ out (p1)∈R HW×C (5); Where φ in , φ out It is a linear projection layer that performs channel blending, f is a depthwise convolution, H and W represent the width and height of the input image, respectively, and C represents the number of channels; This represents the initial principal feature in the real space of dimension HW×C; Represents the initial adjacency guiding feature in the real space of HW×C dimensions; R HW×2C Let p0 represent a real number space of dimension HW×2C; p0 represents the initial principal feature; q0 represents the initial adjacent guiding feature. The above formula explicitly introduces the random initial principal characteristics through element-wise multiplication. and random initial adjacent guiding features The interactions between them are considered as first-order interactions, because each random initial principal feature... Guided features only with its random initial neighbors The interaction occurs once; Recursive gated convolution module g n Conv further enhances model capacity by introducing higher-order interactions; Formally, it is first achieved through a linear projection layer φ in A set of projection features and adjacent guided features is obtained, and is specifically represented as follows: In the formula, C0 represents the set of adjacent guiding features from the initial adjacent guiding features to the (n-1)th recursion, C0 represents the initial number of feature channels, and C... k The number of feature channels after the k-th recursion is represented by n, and the total number of recursions is represented by n. Represents the initial principal features in the real space of dimension HW×C0. Represents the initial adjacent guiding features in the real number space of dimension HW×C0. This indicates that the (n-1)th recursion occurs in HW×C. n-1 Adjacency-guided features in a real space of dimension Represent a real space of dimension HW×C, where C k For the k-th channel dimension of higher-order interactions; Then perform gated convolution recursively: p k+1 =f k (q k )⊙g k (p k ) / α,k=0,1,...,n-1 (7); In the formula, dividing by α in each recursive process is for stable training, k represents the number of times the gated convolution is executed recursively, n represents the total number of recursions, and f k G represents the k-th depth-wise convolution operation. k Let q represent the principal feature transformation function. k p represents the guiding feature of the k-th recursive execution of the gated convolution. k+1 This represents the new principal feature obtained after the k-th recursive execution of the gated convolution; Identity represents the identity transformation, and Linear represents the transformation of dimension C. k-1 The input is transformed by a linear transformation to a dimension of C. k The output space; Finally, output q from the last recursion. n Input to projection layer φ out , thus obtaining the recursive gated convolution g n The result of Conv; Each channel dimension C of higher-order interactions k Set to an exponentially decreasing form:

3. The method for identifying transmission tower defects based on the improved YOLOv11 algorithm according to claim 1, characterized in that: The GhostConv convolutional module includes the following operation steps: The first step involves using a 1x1 convolutional layer to generate m intrinsic feature maps through the regular convolutional part, where m = t / s, and t is the final number of feature maps required, and s is the number of transformations. The second step involves applying linear transformation operations to each generated intrinsic feature map in parallel s-1 linear operations to generate m*(s-1) ghost feature maps. Then, the m intrinsic feature maps generated in the first step are preserved by identity mapping and the m*(s-1) ghost feature maps generated by the second step through linear transformation are concatenated in the channel dimension to finally obtain m*s feature maps as output.

4. The method for identifying transmission tower defects based on the improved YOLOv11 algorithm according to claim 1, characterized in that: The RepViTblock module is an improvement on the existing lightweight Vision Transformer module, referred to as the lightweight ViT module. The existing lightweight ViT module simulates the token mixer and channel mixer by splitting them in the efficient mobile module MobileNetV3, and enhances model learning during training through structural reparameterization techniques. The high-efficiency mobile module MobileNetV3 uses 1×1 extended convolution and 1×1 projection layer to achieve interaction between channels. After the 1×1 extended convolution, a 3×3 depthwise convolution DWConv is performed to achieve spatial information fusion, so that the token mixer and the channel mixer are coupled together. The following improvements were made to the efficient motion module in the RepViTblock reparameterized visual transformation module: To separate the token mixer and the channel mixer, a 3×3 depthwise convolutional (DW) module is first used to extract spatial features, while parallel 1×1 ordinary convolutions are used to process the inter-channel relationships to form a depthwise convolutional (DW) layer. In the token mixer, a structure reparameterization technique is used for the DW layer, so that the multi-branch topology is merged into a single 3×3 depthwise convolutional (DW). The squeeze-excite network layer SE is set after the merged 3×3 depthwise convolution DW to calculate the importance weight of each channel. Then, the weights of each channel are recalibrated to achieve information exchange between channels. Finally, 1x1 convolution is used to complete the final channel blending and dimensionality transformation.

5. The method for identifying transmission tower defects based on the improved YOLOv11 algorithm according to claim 1, characterized in that: The categories of defects in transmission towers include infrastructure failures, conductor and ground wire failures, insulator failures, auxiliary facility failures, vibration damper failures, small-sized hardware defects, broken or loose cable strands, foreign objects on the tower, or the tower base being obstructed by debris.

6. The method for identifying transmission tower defects based on the improved YOLOv11 algorithm according to claim 1, characterized in that: The defects of the transmission towers include missing pins, pins not being opened, equipotential rings falling off, loose bolts, damaged insulators, or insufficient number of pre-twisted wires in the vibration damper.

7. The method for identifying transmission tower defects based on the improved YOLOv11 algorithm according to claim 1, characterized in that: The data augmentation includes the following six methods: Flip horizontally; Flip vertically; Enhance colors by removing a random proportion of colors from the input image; To enhance brightness, each image is converted to a color space with a brightness-related channel, that channel is extracted, multiplied by a factor between 0.5 and 1.5, added with a value between -30 and 30, and then converted back to the original color space. Pixel-by-pixel random value addition: Random values ​​are added to each pixel of the image. The added pixel value is a random value between -40 and 40, and each value is sampled for each pixel. Original color channel histogram equalization.

8. The method for identifying transmission tower defects based on the improved YOLOv11 algorithm according to claim 1, characterized in that: The evaluation metrics used in step four include recognition accuracy, recall, and average recognition accuracy.

Images