Lightweight YOLO-based power transmission line fault inspection method for unmanned aerial vehicle

Abstract

The invention relates to the technical field of target detection, in particular to a light-weight YOLO-based power transmission line fault inspection method for an unmanned aerial vehicle, which introduces RFCACConv convolution, the convolution introduces a receptive field optimization mechanism, provides a K3RFCA structure, effectively fuses shallow feature information and deep feature information, and provides a K3RFCA structure in a feature fusion stage. An improved KBiFPN is adopted to enhance the fusion capacity of distinguishing features, meanwhile, the module is further combined with K3Star, the adaptive capacity of the KBiFPN to objects with different scales is remarkably enhanced, the invention provides a novel multi-scale multi-dimensional cooperative attention module MMCA, the module fuses multi-scale convolution and space-channel multi-dimensional information, and the algorithm has the advantages that the algorithm is simple, the algorithm is easy to implement, and the algorithm is easy to popularize and use. According to the light-weight RKM-YOLO power transmission line fault inspection method, the detection precision reaches 90.7%, compared with an initial YOLOv8n model, the parameter quantity is reduced by 18.27%, meanwhile, the recall rate and mAP50 are both improved by 3.1%, and the detection speed meets the inspection requirement of an unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the field of target detection and testing technology, and in particular to a transmission line fault inspection method using a drone based on lightweight YOLO. Background Art

[0002] Transmission lines, as key components of the power system, are crucial for their proper operation. Therefore, ensuring their reliability is crucial. Insulators and anti-vibration hammers are crucial components of transmission lines. However, due to long-term exposure to harsh environments, these lines are susceptible to adverse factors such as ice damage, wind damage, acid rain, and lightning strikes, leading to failures. Common failures include insulator failure, insulator breakage, insulator contamination flashover, and anti-vibration hammer and anti-vibration hammer rust. Defects in insulators and anti-vibration hammers can lead to unstable power supply and even widespread power outages, resulting in severe economic losses. Therefore, timely inspection of key electrical equipment on transmission lines to identify potential defects is essential.

[0003] As power grids expand in size and complexity, traditional inspection methods are no longer able to meet the demand for efficient and accurate fault detection. Manual inspections are not only time-consuming and labor-intensive, but also prone to missed or false positives. Helicopter inspections, while offering wide coverage, are also costly and significantly affected by environmental factors such as weather. In contrast, drone-based inspection technology, with its advantages of wide coverage, high efficiency, safety, cost-effectiveness, and environmental friendliness, is gradually replacing manual and helicopter inspections and becoming a widely used inspection method. However, drone inspections generate a large number of aerial images, and relying on manual recognition of these images is not only time-consuming and costly, but can also lead to inaccurate inspection results.

[0004] Therefore, it is urgent to invent a UAV-based lightweight YOLO-based transmission line fault inspection method to solve the above problems. Summary of the Invention

[0005] The present invention provides a transmission line fault inspection method using an unmanned aerial vehicle (UAV) based on lightweight YOLO. The present invention introduces RFCAConv convolution, which introduces a receptive field optimization mechanism and adopts an improved KBiFPN module further combined with a K3_Star structure. This solves the problem in the prior art that UAV inspections of transmission lines generate a large number of aerial images, which relies on manual recognition, is time-consuming, labor-intensive, and costly, and may also lead to inaccurate detection results.

[0006] In order to achieve the above objectives, the following technical solutions are provided:

[0007] A transmission line fault inspection method based on lightweight YOLO using a drone includes the following steps:

[0008] S1: Obtain aerial images of transmission lines taken by a UAV;

[0009] S2: performing data enhancement processing on the aerial image, including horizontal flipping, brightness adjustment, and adding Gaussian noise and salt and pepper noise;

[0010] S3: Build an improved lightweight RKM-YOLO model, where the YOLO model is a YOLOv8 model. The lightweight RKM-YOLO model includes:

[0011] (a) The K3_RFCA module is used in the YOLOv8 backbone network to replace the original C2f module in YOLOv8. The K3_RFCA module extracts multi-scale features through parallel branched RFCAConv convolution, fuses shallow and deep features, and optimizes feature weights;

[0012] (b) A KBiFPN structure is used in the YOLOv8 neck network. The KBiFPN structure is integrated into the BiFPN structure using the K3_Star module. The BiFPN structure and the K3_Star module perform cross-scale feature fusion.

[0013] (c) A multi-scale multi-dimensional collaborative attention module (MMCA) is introduced into the YOLOv8 head network to enhance key area features through deep separable convolution and adaptive weighting strategy;

[0014] S4: Train the RKM-YOLO model and deploy it to the drone to perform real-time fault detection on the power transmission line.

[0015] Preferably, the RFCAConv convolution extracts receptive field features, and the specific RFCAConv convolution includes: extracting spatial features from the input feature map through group convolution, and after normalization and ReLU activation, the RFCAConv introduces the CA channel attention mechanism in combination with group convolution, generates a weight matrix through the Softmax function, and finally weights the feature map. The channel attention mechanism generates spatial weights by decomposing the feature maps in the height and width directions.

[0016] Preferably, the K3_RFCA module includes the following structure: the input feature map is processed by two parallel 1×1 convolution branches, one of which extracts deep features through multiple series-parallel mixed RFCAConv layers, and the other retains the original shallow features, and finally the outputs of the two branches are spliced ​​through residual connection.

[0017] Preferably, the KBiFPN structure operation includes: reducing the number of output layers of BiFPN from five layers to three layers, and integrating the K3_Star module to optimize feature correlation by element-by-element multiplication. The KBiFPN structure dynamically adjusts the importance of cross-scale feature fusion by learning weights.

[0018] Preferably, the MMCA module uses depth-wise separable 5×5 convolution to extract local detail features, and the implementation of the MMCA module includes: using 1×7, 7×1, 1×11, 11×1, 1×21, and 21×1 multi-scale convolution kernels to extract local and global information.

[0019]

[0020] Where X represents the input features, Att and F are the attention map and output respectively, represents the element-wise matrix multiplication operation, DWConv represents the depthwise separable convolution, Scalei(i∈{0, 1, 2, 3}), the i-th branch in the MMCA attention mechanism diagram, and Scale0 is the residual connection.

[0021] Preferably, the MMCA module generates spatial-channel collaborative attention weights by spatial rotation operation (C×H×W) and channel compression of the computing unit, and enhances the interactive information between spatial and channel dimensions through adaptive pooling features. The two pooling of the mth channel of the MMCA module are specifically expressed as follows:

[0022]

[0023] represents the global average pooling after the mth channel, Represents the standard pooling after the mth channel, H / W represents the height and width of the feature map, i / j is the spatial index variable used to traverse the pixel position of the feature map, is the eigenvalue of the mth channel at position (i, j),

[0024] After squeezing transformation, feature representation is obtained as follows:

[0025]

[0026] Among them, α and β are two trainable parameters. and These are two output matrices after pooling operations, Tsq represents compression transformation, and Tex represents extrusion transformation.

[0027] After the MMCA module performs spatial dimension interaction on the computing units, the output is a feature map optimized by the triple-dimensional interaction and global attention mechanism, as shown in the following formula:

[0028]

[0029] Among them, FOUT∈RC×H×W represents the feature map output, F″ W Indicates that after F'W The feature map rotated 90 degrees clockwise along the axis H, F″ H The enhanced feature map F′ H Rotate 90° counterclockwise along the width axis W, F″ C The enhanced feature map F′ C No need to rotate to get it.

[0030] Preferably, the spatial-channel collaborative mechanism of the MMCA is specifically as follows: rotating the input feature map along the height axis, generating the width dimension attention weight Aw through channel compression transformation; multiplying Aw by the rotated feature map element by element and then rotating it inversely to restore the spatial dimension; fusing the mean and standard deviation pooling features through parameters α and β to enhance the context expression of the occluded area.

[0031] Preferably, the transmission line defect images collected by the drone include five types of targets: insulators, broken insulators, insulator contamination flashovers, shock absorbers, and rusted shock absorbers.

[0032] Preferably, a power transmission line defect detection system includes: a high-definition camera device carried by a drone, a processor and an output module, and the processor is used to execute the technical method mentioned in the present invention.

[0033] The beneficial effects of the present invention are:

[0034] This paper introduces RFCAConv convolution, which combines the receptive field spatial attention mechanism and the CA channel attention mechanism to solve the problem of insufficient local feature extraction caused by parameter sharing.

[0035] The present invention designs the K3_RFCA module, which effectively integrates multi-layer processed deep features and relatively original shallow features through effective attention weighting, enabling the backbone network to simultaneously utilize local and global multi-scale information, thereby enhancing the feature extraction capability of the backbone network.

[0036] In order to improve the defect detection capability of multi-target objects in transmission lines, this paper proposes a lightweight feature fusion network (KBiFPN) to enhance the ability to extract distinguishing features. This structure can efficiently capture key feature channels and enhance the correlation between features through an element-by-element multiplication mechanism, thereby generating a more accurate feature map. It significantly reduces the number of model parameters while ensuring accuracy.

[0037] This paper proposes a multi-scale multi-dimensional attention module MMCA, which uses depthwise separable convolutions of different scales to capture multi-scale local and global information, enhances the model's perception of details and long-distance dependencies, and uses an adaptive weighting strategy to dynamically adjust the weights of channel features, so that the model can adaptively strengthen its attention to key channel features while suppressing irrelevant information.

[0038] The lightweight RKM-YOLO transmission line fault inspection method proposed in this paper has an average detection accuracy of 90.7%. Compared with the initial YOLOv8n model, the number of parameters is reduced by 18.27%. At the same time, the recall rate and mAP50 are both improved by 3.1%. The detection speed meets the requirements of drone inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the RKM-YOLOv8 network structure in the present invention;

[0040] Figure 2 Schematic diagram of the RFCAConv convolution module structure in the present invention;

[0041] Figure 3 Schematic diagram of the three modules of YOLOv8's K3 structure, YOLOv8's C2f structure and YOLOv5's C3;

[0042] Figure 4 Schematic diagram of the K3_RFCA module structure in the present invention;

[0043] Figure 5 Schematic diagram of the neck fusion structure of FPN structure, PAN structure and BiFPN structure;

[0044] Figure 6 Schematic diagram of the application of the three output layers of BiFPN in the present invention on yolov8;

[0045] Figure 7 It is a structural diagram of the K3_Star structure in the present invention;

[0046] Figure 8 Schematic diagram of the MMCA module and its unit structure in the present invention;

[0047] Figure 9 Schematic diagram of the improved mAP algorithm curve in the present invention;

[0048] Figure 10 A comparison chart of the performance impact of different attention mechanisms;

[0049] Figure 11 Schematic diagram for comparing mAP performance results of different target detection algorithms;

[0050] Figure 12 A schematic diagram comparing the training curves of the RKM-YOLO model and the YOLOv8n model;

[0051] Figure 13 A schematic diagram comparing the visualization effects of different algorithms; DETAILED DESCRIPTION

[0052] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions of the structural diagrams in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] The following is based on Figure 1-13 A transmission line fault inspection method based on lightweight YOLO using a drone according to an embodiment of the present invention is described in detail.

[0054] The present invention proposes a transmission line fault inspection method based on lightweight YOLO using a UAV, comprising the following steps:

[0055] S1: Obtain aerial images of transmission lines taken by a UAV;

[0056] S2: performing data enhancement processing on the aerial image, including horizontal flipping, brightness adjustment, and adding Gaussian noise and salt and pepper noise;

[0057] S3: Build an improved lightweight RKM-YOLO model, where the YOLO model is a YOLOv8 model. The lightweight RKM-YOLO model includes:

[0058] (a) The K3_RFCA module is used in the YOLOv8 backbone network to replace the original C2f module in YOLOv8. The K3_RFCA module extracts multi-scale features through parallel branched RFCAConv convolution, fuses shallow and deep features, and optimizes feature weights;

[0059] (b) A KBiFPN structure is used in the YOLOv8 neck network. The KBiFPN structure is integrated into the BiFPN structure using the K3_Star module. The BiFPN structure and the K3_Star module perform cross-scale feature fusion.

[0060] (c) A multi-scale and multi-dimensional collaborative attention module (MMCA) is introduced into the head network to enhance key area features through depth-wise separable convolution and adaptive weighting strategy;

[0061] S4: Train the RKM-YOLO model and deploy it to the drone to perform real-time fault detection on the power transmission line.

[0062] In some embodiments, RKM-YOLO is used to enhance its ability to identify defects in transmission line devices based on the characteristics of small targets, multi-scale targets and complex backgrounds.

[0063] In some embodiments, an improved RFCAConv convolution based on RFAConv is used as a receptive field attention convolution. RFCAConv convolution allows the network to generate specific weights for each receptive field, enabling the network to learn local features in the image more carefully. The input features of this module are first extracted through group convolution to extract rich receptive field features to capture a wider range of spatial information. The features are normalized and activated with ReLU to enhance the nonlinear representation ability of the features.

[0064] The CA channel attention mechanism is introduced in the RFCAConv convolution, which embeds position information into the channel attention, enhances the module's ability to capture spatial dependencies, and further optimizes feature extraction.

[0065] By emphasizing the importance of each receptive field feature through the softmax function, the RFCAConv convolution reweights the input features, captures global features in different dimensions, reduces the damage of transmission line noise to long-range dependencies, and improves the robustness of the model in complex backgrounds, thereby more accurately capturing the subtle differences of abnormal targets when processing each pixel.

[0066] In some embodiments, the K3 structure is adopted in the backbone network of YOLOv8. The K3 structure inherits the gradient flow mechanism of the C2f module and combines it with a stronger local feature capture capability. It can more effectively retain fine-grained spatial information while improving the processing capability of long-distance dependencies and global context. The present invention reduces the number of parameters by introducing RFCAConv convolution to replace the Bottleneck structure in the K3 module, while optimizing the fusion capability of multi-scale features, further improving the overall performance of the network.

[0067] In some embodiments, the K3_RFCA module adopts a two-parallel 1×1 convolution branch design, which can efficiently extract information from input features and maintain a high feature expression capability. Replacing Bottleneck with RFCAConv significantly enhances the network feature extraction capability.

[0068] In some embodiments, the K3_RFCA module employs two parallel 1×1 convolution branches, where the features from one path through the 1×1 convolution are further processed through multiple RFCAConv layers in a series-parallel hybrid. This adaptively assigns weights based on the input features, highlighting the features of key areas and suppressing noisy areas. Defects in power transmission line equipment often have complex and diverse scales. By employing the RFCAConv convolution extraction process with layer-by-layer gradient flow, the model can more deeply extract useful features, ensuring that feature expressions at different levels effectively represent the rich information of the input image. This rich information is then concatenated with another 1×1 convolution, effectively fusing deep features with the original shallow features, ensuring synergy between semantic information and detailed features.

[0069] In some embodiments, the K3_RFCA module introduces a residual connection mechanism, which, on the one hand, alleviates the gradient vanishing problem and enhances the training stability of deep networks; on the other hand, it retains the key information in the original input features, ensuring that multi-layer processing does not lose important details, and reduces the channel information degradation problem caused by noise such as illumination changes. This improvement can significantly improve the robustness and detection accuracy of the model and reduce missed detections and false detections.

[0070] In some embodiments, the neck network of the YOLO algorithm usually adopts PAN (path aggregation network) and FPN (feature pyramid network) for feature fusion, and uses lateral connections and pyramid structures to integrate feature information of different scales. The BiFPN structure has the bidirectional information flow mechanism of FPN, and also introduces a learnable weighted fusion mechanism so that the features of each layer can be weighted and adjusted according to their importance. The BiFPN structure can adaptively adjust the weights of different scales according to the different features of the input image, avoiding the information fusion imbalance problem that may occur in FPN.

[0071] In some embodiments, the number of output layers of the BiFPN structure is reduced from five to three to better be compatible with the backbone network in YOLO. Fewer output layers not only reduce feature redundancy and improve the efficiency of feature fusion, thereby improving the final recognition accuracy, but also significantly reduce the number of model parameters. In BiFPN, each bidirectional path (top-down and bottom-up) is regarded as a separate feature network layer, and then these layers can be repeated multiple times to promote higher-level feature fusion. In addition, BiFPN weights different input features so that the network understands the importance of each feature, thereby better integrating features of different scales. Using a fast normalized fusion method, the expression is:

[0072]

[0073] Among them, w iis a learnable weight, each w i There is a ReLU activation function to ensure that w i ≥0, the minimum value ε=0.0001 is used to avoid numerical instability, each 0≤w i ≤1, vector (per channel) or multidimensional tensor (per pixel), O is the output feature after fusion, which retains the most useful information of the input through weighted summation, I i Represents the i-th feature input.

[0074] In some embodiments, in a single-layer neural network of the StarNet structure, w1,w2,x∈R are defined (d+1)×1 , it can be easily expanded to accommodate multiple output channels and process multiple feature elements. Its star operation can be further written as:

[0075]

[0076]

[0077] Where w1, w2, x represent two independent weights w1 and w2 that linearly transform the input x feature, d is the number of input channels, i and j are channel indices, and α is the coefficient of each sub-item.

[0078] In some embodiments, the Neck part of the YOLOv8 model is optimized using the K3_Star module. The K3_Star module is as follows: Figure 10 As shown in the figure, integrating K3_Star into the BiFPN structure is the KBiFPN structure, which can effectively focus on important feature channels and optimize the relationship between features by element-by-element multiplication to generate a more accurate feature map.

[0079] In some embodiments, as Figure 11 As shown, the MMCA module solid line multi-scale multi-dimensional collaborative attention mechanism is used. The MMCA module improves the MCA module. The MMCA module uses depth-separable 5×5 convolution to extract local features and capture fine-grained information of the target. Then, the present invention uses 1×7, 7×1, 1×11, 11×1, 1×21, and 21×1 convolution to extract multi-scale information. The calculation formula is as follows:

[0080]

[0081] Among them, X represents the input feature, Att and F are attention map and output respectively, represents element-wise matrix multiplication, DWConv represents depthwise separable convolution, and Scalei(i∈{0, 1, 2, 3}) represents Figure 11(a) The i-th branch in the MSCA attention mechanism diagram, where Scale0 is a residual connection.

[0082] Considering the MCA unit as a computational unit, the input feature (C×H×W) is rotated by 90° to obtain the rotated feature map. Then, after channel compression, the aggregated feature weights are generated. and Apply the sigmoid function to convert these weights into attention weights AW∈RW×1×1. The calculation formula is as follows;

[0083]

[0084]

[0085] Among them, H / W represents the height and width of the feature map, Tsq represents compression transformation, Tex represents squeezing transformation, σ represents Sigmoid activation function, PMH represents the counterclockwise rotation of the feature map by 90° along the H axis, and A W represents the attention matrix, F′ W Indicates the element-wise multiplication operation to generate the enhanced feature map, F″ W Indicates that after F' W The feature map is rotated 90 degrees clockwise along the axis H to restore it to its original shape. F represents the input feature, PMH represents the operation of rotating the feature map 90 degrees counterclockwise along the H axis, Tsq represents compression transformation, and Tex represents extrusion transformation.

[0086] The MMCA module then uses an adaptive mechanism to aggregate global average and standard deviation information through squeezing transformation. The credit values ​​of the average pooling and standard deviation pooling features are different, so this mechanism aims to enhance the interactive information between spatial and channel dimensions, overcoming the sub-optimization problem that may be caused by fixed pooling weights in traditional methods. The two pooling methods of the mth channel are specifically expressed as follows:

[0087]

[0088] in, represents the global average pooling after the mth channel, Represents standard pooling after the m-th channel, i / j is the spatial index variable used to traverse the pixel position of the feature map, i / j is the spatial index variable used to traverse the pixel position of the feature map, and is the eigenvalue of the m-th channel at position (i, j).

[0089] After squeezing transformation, feature representation is obtained as follows:

[0090]

[0091] Among them, α and β are two trainable parameters, and are two output matrices after pooling operations, Tsq represents compression transformation, and Tex represents extrusion transformation.

[0092] After spatial dimension interaction in the MCA unit, the output is a feature map optimized by the three-dimensional interaction and global attention mechanism, as shown in the following formula.

[0093]

[0094] Among them, FOUT∈RC×H×W represents the feature map output, F″ W Indicates that after F' W The feature map rotated 90 degrees clockwise along the axis H, F″ H The enhanced feature map F′ H Rotate 90° counterclockwise along the width axis W, F″ C The enhanced feature map F′ C No need to rotate to get it.

[0095] In some embodiments, the dataset used is derived from a dataset captured by a drone and a dataset obtained from the Roboflow website. Due to the lack of a public dataset to solve various types of anomalies on the power sector's transmission lines, the present invention constructs a comprehensive dataset of 1,911 images, each of which contains one or more targets and is labeled with five types of labels: insulator, broken insulator, insulator contamination flashover, shock absorber, and rusted shock absorber.

[0096] The present invention divides the entire dataset into training set, test set and validation set in a ratio of 7:2:1. This division ensures that the model is trained on a sufficiently diverse dataset, while providing separate validation and test sets for evaluating the generalization ability and performance of the model in real-world applications.

[0097] The data augmentation operations used in this paper include horizontal flipping, brightness adjustment, Gaussian noise, and salt-and-pepper noise. The resulting data sets reach 3,618, increasing the diversity of training samples and simulating the complex environmental conditions likely encountered in real-world applications, thereby ensuring the effectiveness and robustness of the model.

[0098] Data verification principle of the present invention:

[0099] To verify the effectiveness of the proposed algorithm, the experimental evaluation used three key metrics: accuracy, recall, and average precision. Accuracy measures the model's reliability in predicting positive samples, while recall represents the ratio of correctly identified positive samples to the total number of actual positive samples, reflecting the model's ability to detect real targets. The mAP metric combines precision and recall to comprehensively evaluate model performance. This metric is widely used in comparative analysis of algorithm models and optimization of hyperparameters. Its calculation formula is as follows:

[0100]

[0101]

[0102] Among them, TP: true positive example FP is false positive example, FN: false negative example, N: category, Precision: precision rate, Tp: number of correctly predicted positive samples, Fp: number of negative samples incorrectly predicted as positive, Recall: recall rate, F1: harmonic mean of precision and recall rate, AP: average value of precision rate at different recall rate thresholds, mAP: average value of AP of all categories.

[0103] In some embodiments, different lightweight backbone networks are compared and analyzed, and Backone, Baseline, GhostNetV2, ShuttleNetV2, and MobilenNetV4 are compared with the lightweight backbone network used in the present invention, and their characteristics are compared and analyzed. The specific analysis data is shown in Table 1:

[0104] Table 1: Comparative analysis of feature extraction of different lightweight backbone networks

[0105]

[0106] The proposed model achieves a mAP50 score of 90.3%, significantly outperforming MobileNetV4's 82.3% and GhostNetV2's 83.7%. The model's parameters are reduced to 2.79M, significantly reducing computational resource requirements. Furthermore, the model achieves a precision of 90.8% and a recall of 82.6%, demonstrating its superior performance in complex backgrounds and multi-target detection scenarios. Experimental results demonstrate that the improved model offers significant advantages in computational efficiency and practical performance, providing reliable technical support for practical engineering applications such as power line inspection.

[0107] In some embodiments, a comparative analysis of different neck network feature fusion structures is performed using Baseline, AFPN, GFPN, ASF-YOLO, and the KBiFPN structure of the present invention to perform a systematic performance evaluation comparison. The specific analysis data is shown in Table 2:

[0108] Table 2: Comparison of experimental results of different feature fusion networks

[0109]

[0110] Comparative analysis shows that the improved KBiFPN structure used in this paper reduces the number of parameters by 14.28% while achieving a 0.6% improvement in the mAP50 indicator. Compared with other feature fusion networks, KBiFPN demonstrates significant advantages in both detection accuracy and mAP50, with a parameter count of only 2.58M. This achieves significant results while using a relatively low number of parameters.

[0111] In some embodiments, different attention mechanisms are compared and analyzed, using YOLOv8, YOLOv8+MSCA, YOLOv8+EMA, YOLOv8+MCA and the YOLOv8+MMCA attention mechanism of the present invention for comparison. The specific analysis data is shown in Table 3:

[0112] Table 3: Comparison results of attention mechanisms

[0113]

[0114]

[0115] Comparative analysis shows that the proposed MMCA module demonstrates significant advantages in object detection. Compared with the YOLOv8 baseline model and attention mechanisms such as MSCA, EMA, and MCA, the MMCA module achieves significant improvements in multiple key metrics. Specifically, the MMCA module achieves 88.7% on mAP50, a significant improvement over MCA's 87.8% and EMA's 87.9%, demonstrating its ability to extract multi-scale feature information. Although MMCA's accuracy slightly decreases to 97.6% in the Insulator category, it significantly improves performance in the Broken, Flashover, Harmmer, and Harmmer-defect categories. Compared with the MSCA module, the MMCA module in this paper not only significantly improves accuracy, but also has only 3.11M parameters, which is almost the same as the MSCA module, ensuring low computational overhead. This verifies that the lightweight MMCA module has strong multi-scale feature extraction capabilities and can accurately locate faults.

[0116] In some embodiments, comparative ablation experiments were conducted on various properties of the present invention, and the specific analysis data are shown in Table 4:

[0117] Table 4: Ablation comparison test results

[0118]

[0119]

[0120] A comparative analysis showed that before adding the module, the initial model achieved precision of 89.9%, recall of 80.7%, mAP0.5 of 87.6%, and a parameter count of 3.01M. When only the RFCA backbone module was replaced, precision increased to 90.8%, recall reached 82.6%, mAP50 reached 90.3%, and the parameter count decreased by 7.31%. This improvement was significant across the board, with precision increasing by 0.9%, recall by 1.9%, mAP by 2.7%, and mAP50-90 by 2.3%, demonstrating that the backbone RFCA significantly enhances the model's feature extraction capabilities. When the KBiFPN module was introduced alone, mAP50 increased by 0.6%, and the parameter count decreased by 14.28%. Furthermore, the addition of a lightweight MMCA attention mechanism module improves precision by 1.9%, recall by 0.7%, and mAP50 by 1.1%. This module effectively captures long-term spatial and channel dependencies in multi-scale defect features, thereby enhancing the model's attention to defect information at different scales, significantly improving detection accuracy with only a slight increase in parameter count. When the backbone RFCA and KBiFPN are combined, precision reaches 92.2%, an improvement of 2.3%. mAP reaches 90.2, an improvement of 2.6%. The number of parameters is reduced by 21.59%. Finally, integrating these three improvements into YOLOv8 increases Precision, Recall, mAP50, and mAP50-90 by 2.1%, 3.1%, 3.1%, and 0.9%, respectively. Furthermore, the number of parameters decreases by 18.27%, validating that the proposed improvements can effectively improve defect detection performance.

[0121] In some embodiments, in order to verify the comprehensive performance of the RKM-YOLO algorithm, the target detection network of the present invention is compared with other target detection networks. The specific analysis data is shown in Table 5:

[0122] Table 5: Performance comparison between RKM-YOLO target detection network and classic model

[0123]

[0124]

[0125] Comparative analysis shows that the second-stage FasterR-CNN algorithm lags significantly behind the first-stage YOLO series algorithms, particularly for small targets like anti-vibration hammers, anti-vibration hammer corrosion, and insulator flashover, with mAP50 scores of 41.8%, 60.6%, and 45.5%, respectively. This demonstrates that Faster-RCNN faces significant challenges in small object detection, particularly in complex environments. In contrast, the YOLO series demonstrates superior small object detection performance. From YOLOv3-Tiny to YOLOv11n, the model family has progressively reduced parameter count and computational complexity while simultaneously improving mAP50. As shown in the table, YOLOv5n, with 2.51M parameters and 7.2G GFLOPs, achieves an mAP50 of 85.1%. YOLOv8n further optimizes performance, increasing its mAP to 87.6%, demonstrating particularly strong performance in small object detection. While YOLOv10n and YOLOv11n have reduced parameter counts, mAP50 improvements are limited. The RKM-YOLO model proposed in this paper has achieved further breakthroughs in accuracy and lightweighting. It demonstrates excellent detection results for small targets such as insulator damage, flashover, anti-vibration hammer, and anti-vibration hammer corrosion, achieving accuracies of 89.7%, 87.5%, 88.9%, and 89.5%, respectively. This significantly outperforms other YOLO models.

[0126] In some embodiments, the training metrics of the RKM-YOLO and YOLOv8n proposed in the present invention were compared. During the training process, it can be seen that the convergence speed of RKM-YOLO is significantly faster than that of YOLOv8n, especially in the early stages of training. This shows that the improved model optimization efficiency is higher and can find better parameters more quickly. As can be seen from the Precision and Recall curves, the performance of RKM-YOLO during training is better than that of YOLOv8n, indicating that the improved model performs better in reducing false detections and missed detections. The improved model has a significant improvement in mAP, verifying the feasibility of the improved algorithm in this paper.

[0127] In some embodiments, comparative tests based on drone aerial photography test sets verified the detection performance of the improved model. For the detection difficulties of small targets such as insulator faults and shock hammer defects, the improvement of the model successfully captured more detailed features. Even when the defect area is small in size and similar in color to the background, the model can still accurately locate and identify it.

[0128] exist Figure 13 The models in the first and third columns can accurately identify the damage and flashover defects of insulators, solving the missed detection problem of the above YOLO algorithm and proving its effectiveness. Figure 13In the second column, faced with complex backgrounds, such as scenes with multiple interference factors such as high-voltage lines, metal structures, and vegetation, the improved model still maintains a high detection accuracy and can accurately distinguish between targets and backgrounds.

[0129] In the case of increased noise, such as Figure 13 In the fourth column, noise interference further exacerbates the challenge to model performance. YOLOv5 and YOLOv11 both experience missed detections, and YOLOv8 also suffers from false detections. The improved model of the present invention successfully overcomes these interferences, can accurately identify targets and output reliable detection results.

[0130] The improved model of the present invention solves the difficulty of small target detection in UAV aerial photography scenes, outperforms other algorithms in accuracy and robustness, and has good engineering application prospects.

Claims

1. A UAV-based transmission line fault inspection method based on lightweight YOLO, characterized in that: The following steps are involved: S1: Obtain aerial images of transmission lines taken by a UAV; S2: performing data enhancement processing on the aerial image, including horizontal flipping, brightness adjustment, and adding Gaussian noise and salt and pepper noise; S3: Build an improved lightweight RKM-YOLO model, where the YOLO model is a YOLOv8 model. The lightweight RKM-YOLO model includes: (a) The K3_RFCA module is used in the YOLOv8 backbone network to replace the original C2f module in YOLOv8. The K3_RFCA module extracts multi-scale features through parallel branched RFCAConv convolution, fuses shallow and deep features, and optimizes feature weights; (b) A KBiFPN structure is used in the YOLOv8 neck network. The KBiFPN structure is integrated into the BiFPN structure using the K3_Star module. The BiFPN structure and the K3_Star module perform cross-scale feature fusion. (c) A multi-scale multi-dimensional collaborative attention module (MMCA) is introduced into the YOLOv8 head network to enhance key area features through deep separable convolution and adaptive weighting strategy; S4: Train the RKM-YOLO model and deploy it to the drone to perform real-time fault detection on the power transmission line.

2. The method according to claim 1, characterized in that The RFCAConv convolution extracts receptive field features. The specific RFCAConv convolution includes: extracting spatial features from the input feature map through group convolution. After normalization and ReLU activation, the RFCAConv introduces the CA channel attention mechanism in combination with group convolution, generates a weight matrix through the Softmax function, and finally weights the feature map. The channel attention mechanism generates spatial weights by decomposing the feature maps in the height and width directions.

3. The method according to claim 1, characterized in that The K3_RFCA module has the following structure: the input feature map is processed by two parallel 1×1 convolution branches, one of which extracts deep features through multiple series-parallel mixed RFCAConv layers, and the other retains the original shallow features. Finally, the outputs of the two branches are spliced ​​through a residual connection.

4. The method according to claim 1, wherein The KBiFPN structure operation includes: reducing the number of BiFPN output layers from five to three, integrating the K3_Star module, optimizing feature correlation by element-by-element multiplication, and dynamically adjusting the importance of cross-scale feature fusion by learning weights.

5. The method according to claim 1, wherein The MMCA module uses depth-wise separable 5×5 convolution to extract local detail features. The implementation of the MMCA module includes: using 1×7, 7×1, 1×11, 11×1, 1×21, and 21×1 multi-scale convolution kernels to extract local and global information. Where X represents the input features, Att and F are the attention map and output respectively, represents the element-wise matrix multiplication operation, DWConv represents the depthwise separable convolution, Scalei(i∈{0, 1, 2, 3}), the i-th branch in the MMCA attention mechanism diagram, and Scale0 is the residual connection.

6. The method according to claim 5, characterized in that The MMCA module generates spatial-channel collaborative attention weights by combining the computational unit with spatial rotation operations (C×H×W) and channel compression, and enhances the interactive information between spatial and channel dimensions through adaptive pooling features. The two poolings of the mth channel of the MMCA module are specifically expressed as follows: represents the global average pooling after the mth channel, Represents the standard pooling after the mth channel, H / W represents the height and width of the feature map, i / j is the spatial index variable used to traverse the pixel position of the feature map, is the eigenvalue of the mth channel at position (i, j), After squeezing transformation, feature representation is obtained as follows: Among them, α and β are two trainable parameters. and These are two output matrices after pooling operations, Tsq represents compression transformation, and Tex represents extrusion transformation. After the MMCA module performs spatial dimension interaction on the computing units, the output is a feature map optimized by the triple-dimensional interaction and global attention mechanism, as shown in the following formula: Among them, FOUT∈RC×H×W represents the feature map output, F″ W Indicates that after F' W The feature map rotated 90 degrees clockwise along the axis H, F″ H The enhanced feature map F′ H Rotate 90° counterclockwise along the width axis W, F″ C The enhanced feature map F′ C No need to rotate to get it.

7. The method according to claim 6, characterized in that The spatial-channel collaborative mechanism of MMCA is specifically as follows: the input feature map is rotated along the height axis, and the width dimension attention weight Aw is generated through channel compression transformation; Aw is multiplied element-by-element with the rotated feature map and then reversely rotated to restore the spatial dimension; the mean and standard deviation pooling features are fused through parameters α and β to enhance the contextual expression of the occluded area.

8. The method according to claim 1, characterized in that The transmission line defect images collected by drones include five types of targets: insulators, broken insulators, insulator contamination flashovers, shock absorbers, and rusted shock absorbers.

9. A power transmission line defect detection system, characterized in that: include: A high-definition camera device, a processor, and an output module carried by a drone, wherein the processor is used to execute the method according to any one of claims 1 to 8.

Images