Distribution network power transmission line insulator damage detection method based on YOLOv8 improvement

Abstract

The invention provides a distribution network power transmission line insulator damage detection method based on a YOLO algorithm, and the method comprises the steps: collecting a plurality of insulator defect images, and carrying out the data enhancement of the images; performing image detail enhancement on an insulator defect area in the image by using a super-resolution reconstruction algorithm; constructing a composite loss function to guide an image detail enhancement process; constructing an insulator defect detection model based on a YOLOv8 framework; constructing a multi-task joint loss function as an insulator defect detection model training target; and a real-time feedback mechanism is introduced, a weighted combination loss function is constructed, gradient updating optimization is executed, and the detection capability of the insulator defect detection model is improved. The method has high detection precision and robustness, can effectively improve the efficiency and accuracy of fault detection, significantly reduces the burden and cost of manual inspection, improves the safety and reliability of a power system, and provides powerful support for the operation and maintenance of power equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of image detection and processing, and relates to a method for detecting insulator damage in a distribution network transmission line based on a YOLO algorithm. Background Art

[0002] Insulators are key components in power transmission lines. Their condition directly affects the stability and safety of the power system. Regular inspection and maintenance are required for insulators. Defects must be promptly replaced to avoid impacting power transmission efficiency. However, traditional insulator defect detection methods are inefficient and pose safety risks. Specifically:

[0003] 1. Manual visual inspection: As the most traditional inspection method, insulators are inspected by climbing manually or using tools such as telescopes. This method is inefficient and easily affected by human factors such as vision and experience. It is subjective and prone to missed inspections or misjudgments. Visual inspection requires careful observation of each insulator, which is labor-intensive and poses safety risks during high-altitude operations.

[0004] 2. Infrared thermal imager detection: Infrared thermal imagers identify damaged insulators by detecting the temperature distribution on the insulator surface. This method is sensitive to changes in ambient temperature, target surface material, target size, and distance. In some cases, precise calibration is required, otherwise false positives or missed positives are likely to occur. Complex environments such as strong light interference, water vapor, and smoke can also easily affect the accuracy of detection results.

[0005] 3. Ultrasonic testing: Ultrasonic testing uses the propagation characteristics of ultrasonic waves in materials to detect internal damage of insulators. This method requires contact testing and has high requirements for operating skills. It is also costly and difficult to implement.

[0006] 4. Drone-based inspection: Drones are used to carry cameras or sensors to inspect insulators. This method improves inspection efficiency and safety, but places higher demands on image processing and analysis technologies.

[0007] The above methods are all capable of detecting the damage of insulators on distribution network transmission lines, but there are still some technical defects, and the detection accuracy needs to be improved.

[0008] Currently, deep learning technology, especially target detection algorithms, provides new solutions for insulator damage detection. Among them, YOLOv8, as the latest target detection model, has been applied in many fields due to its high speed and accuracy. However, it still has problems such as insufficient detection accuracy and poor robustness when dealing with insulator damage detection in distribution network transmission lines.

[0009] Based on the above defects, the present invention provides an improved distribution network transmission line insulator damage detection method based on YOLOv8 to improve the accuracy and robustness of detection. Its technical core is the field of image processing technology, and an insulator detection method based on target detection algorithm and attention mechanism is proposed.

[0010] After searching, three existing Chinese invention patent application documents are relatively similar to the technical means of the present invention. The following briefly describes the technical features that distinguish the present invention from Comparisons 1 to 3.

[0011] Comparative Document 1: Small Target Detection Method Based on Super-Resolution-YOLO Network (Application No.: 202310255946.1), which differs from the present invention in the following ways: ① Different image enhancement methods: The present invention adopts ESRT for super-resolution reconstruction, which can better restore the details of the image, especially for small defects (such as cracks, discharge marks, etc.). It has stronger recovery ability, which is different from the super-resolution technology in Comparative Document 1. The latter mainly enhances image quality by combining YOLO with super-resolution, but does not emphasize detail recovery and enhancement of small targets; ② Different super-resolution reconstruction methods: The present invention uses deep learning super-resolution methods (such as E SRT), captures global contextual information through the Transformer architecture, enhances the details and contextual dependencies of the image, and has higher performance than the traditional super-resolution method of Comparative Document 1; ③ Whether to optimize small defects: The present invention further improves the detection capability of the YOLO model by decoupling the detection head and fusing multi-scale features, especially in small target detection and complex backgrounds, and can more accurately identify small defects. In contrast, Comparative Document 1 focuses on the application of the attention mechanism, but does not mention task decoupling and multi-scale fusion, and lacks the strategy for optimizing small targets and small defects in the present invention.

[0012] Comparative Document 2: Remote Sensing Ship Target Detection Method (Application No.: 202411105065.2), which differs from the present invention in that: ① The present invention adopts MAP-CA (Maximum and Average Pooling Co ② The focus of Comparative Document 2 is on remote sensing ship target detection, while the application scenario of the present invention is transmission line insulator defect detection. The scale and complexity of the processed targets are significantly different. Specifically, the present invention places particular emphasis on small defect detection, while Comparative Document 2 emphasizes more on how to process ship targets of different scales. ③ The present invention further improves detection accuracy through decoupling the detection head and multi-scale fusion, especially for small targets. It processes small targets more meticulously and can accurately identify small defects. It emphasizes the combination of task decoupling and multi-scale fusion, improving the performance of the model in complex environments. The optimization of Comparative Document 2 mainly focuses on lightweighting and reducing model parameters, which helps the model perform efficiently in terms of computational efficiency, but is not as good as the present invention in detail recovery and small target detection.

[0013] Comparative Document 3: Retail Alcohol Recognition Method for Complex Environments (Application No.: 202310418884.1), which differs from the present invention in the following ways: ① Comparative Document 3 mainly processes images through low-light enhancement and de-reflection, which is only effective for improving image quality under poor lighting conditions, but may not be as effective as the present invention for recovering small targets. Specifically, the present invention can not only improve image resolution but also recover details, especially for details such as small cracks on insulators. ② Comparative Document 3 is mainly used for commodity classification, while the present invention is mainly used for defect detection of transmission line insulators. It needs to deal with small physical defects such as cracks and discharges. The two have different requirements for image features and detail processing. Therefore, the present invention is more creative in detail recovery and small target detection. ③ The present invention more effectively identifies insulator surface defects by decoupling the detection head and fusing multi-scale features, while the creativity of Comparative Document 3 focuses more on enhancing the loss function and image enhancement. Although it is helpful for improving image quality, it lacks a specific optimization strategy for defect detection.

[0014] In summary, compared with comparative documents 1 to 3, the present invention has obvious innovations in processing tiny defects, detail recovery and robustness under complex backgrounds. It is especially suitable for application scenarios with high precision requirements such as transmission line insulator defects, and has more outstanding technical advantages in tiny defect detection, multi-scale fusion and real-time feedback mechanism. Summary of the Invention

[0015] The purpose of the present invention is to address the problems existing in the prior art and provide a distribution network transmission line insulator damage detection method based on the YOLO algorithm, which solves the problems of existing insulator defect methods such as low detection accuracy, poor ability to identify minor defects, inability to stably detect under complex backgrounds, and lack of real-time feedback and dynamic adjustment detection mechanism.

[0016] To this end, the present invention adopts the following technical solutions:

[0017] A method for detecting insulator damage in a distribution network transmission line based on a YOLO algorithm comprises the following steps:

[0018] Step 1: Collecting a number of insulator defect images, wherein the insulator defect includes at least one of insulator damage, discharge marks on the insulator surface, or damage to the shock-absorbing hammer structure;

[0019] The insulator defect images are resized, denoised, and normalized to obtain standardized pre-processed images to ensure that the images meet the model input requirements and have consistent format and quality. Specifically:

[0020] The normalized expression is:

[0021]

[0022] Where, I norm (x, y) is the pixel value of the normalized image; I orig (x,y) is the pixel value of the original image; x,y is the spatial coordinate of the image pixel.

[0023] Step 2: To enhance the model's adaptability to defects in different environments and postures, perform data augmentation on the standardized preprocessed images. Data augmentation includes at least one of rotation, flipping, or cropping, which is used to simulate diverse scenarios and expand the training sample set to improve the model's generalization ability.

[0024] Step 3: Based on data enhancement, a super-resolution reconstruction algorithm is used to enhance the image details of the insulator defect area in the image to improve the recognizability of small targets. A composite loss function is constructed to guide the image restoration process, ensuring that the restored image has high quality in terms of pixel accuracy and semantic consistency. The composite loss function includes a reconstruction loss function and a perceptual loss function. Specifically:

[0025] Based on standardized image preprocessing, this paper introduces a super-resolution reconstruction mechanism to improve feature expression capabilities through high-quality image reconstruction. The expression of the super-resolution reconstruction algorithm is as follows:

[0026] Irec =ESRT(I lowres ;θ)

[0027] Where, I rec is the restored image; I lowres is the standardized preprocessed image; θ is a parameter; ESRT (Efficient Super-Resolution Transformer) is an efficient super-resolution reconstruction algorithm based on deep learning, designed for single-image super-resolution tasks. By combining a lightweight convolutional neural network (CNN) with an improved Transformer architecture, it significantly reduces computational cost and video memory usage while improving image resolution. It is a type of super-resolution reconstruction algorithm and is a technology known to those skilled in the art.

[0028] The image detail enhancement of the insulator defect area in the image using the super-resolution reconstruction algorithm is as follows: the standardized pre-processed image I is enhanced by using the parameter θ lowres The deep features of the image are fused and reconstructed by deconvolution to output a high-resolution image I rec , to restore the edge details and texture information lost due to low-resolution acquisition; among them, the CNN part is used to capture local texture information, and the Transformer module is used to model the global dependencies in the image.

[0029] To ensure that the super-resolution reconstructed image has high fidelity at both the pixel level and the semantic level, a composite loss function is constructed to guide model training. The composite loss function can take into account the pixel-level reconstruction error and the semantic difference at the perceptual level of the image, thereby improving the overall image quality while ensuring the clarity of details. The expression of the composite loss function is as follows:

[0030] L SR =λ1L reconstruction +λ2L perception

[0031] Where, L SR L is the composite loss function used to train the model, guiding the model to generate higher quality images; reconstruction is the reconstruction loss function, which is used to calculate the mean square error between image pixels to measure the difference between the super-resolution restored image and the original high-resolution image at the pixel level; L perception is a perceptual loss function used to measure the difference between the reconstructed image and the high-resolution image in the high-dimensional feature space. It is constructed based on the Euclidean distance (MSE) between the feature maps of the intermediate layers of the VGG network. It can capture the perceptual differences of images at the semantic and texture levels and improve the image generation quality. λ1 and λ2 are both weight coefficients, and λ1+λ2=1.

[0032] in:

[0033]

[0034] Where n is the total number of pixels in the image; is the pixel value of the i-th pixel in the high-resolution real image; is the pixel value of the i-th pixel in the reconstructed image;

[0035]

[0036] Where l is the layer number used to extract feature maps in the VGG network; φ l (*) is the output feature extraction function of the lth layer of the VGG network; I HR is a high-resolution real image; I SR Generate images for super-resolution models; L SR To measure I SR with I HR The pixel-level and perceptual-level differences between is the Euclidean distance, which is used to measure the difference between two feature maps.

[0037] Step 4: Build an insulator defect detection model based on the YOLOv8 architecture and introduce the MAP-CA attention mechanism module to strengthen the focus on the image detail enhancement area. At the same time, embed the GSConv module in the Neck structure in the YOLOv8 architecture to improve feature extraction efficiency. To improve detection accuracy and robustness, continue to build a multi-task joint loss function including the category loss function, bounding box regression loss function, IOU loss function and attention mechanism loss function as the training target of the insulator defect detection model.

[0038] Specifically, the insulator defect detection model is expressed as follows:

[0039] (x,y,w,h,p)=YOLOv8(F gsconv (F att (I aug )))

[0040] Where: x, y are the position coordinates of the target frame; w, h are the width and height of the target frame; p is the predicted probability that the target frame belongs to the defect category; YOLOv8 is the target detection model; I aug is the enhanced input image; F att (I aug ) is image I aug Image feature map after processing by the coordinate attention module MAP-CA.

[0041] Among them, YOLOv8 is an existing target detection model that uses a single neural network to simultaneously complete target position prediction and category recognition. The feature extraction network and feature fusion network parts in the YOLOv8 algorithm are used to extract feature information from the input image and fuse it into the final output.

[0042] After image enhancement, the image is first processed by the coordinate attention module MAP-CA to extract the semantic expression of key areas, which serves as the core link of image feature modeling. Specifically, the coordinate attention module MAP-CA is based on the existing CoordinateAttention (CA) structure, combined with maximum pooling and average pooling operations, and embeds richer spatial information in the channel attention to improve the model's attention to tiny defect areas in the image. It extracts local salient features and global background information of the image respectively, thereby improving the model's attention to key areas.

[0043] The working mechanism of the coordinate attention module MAP-CA is as follows:

[0044] First, the input feature map A is subjected to maximum pooling and average pooling operations along the height dimension and width dimension respectively to obtain feature representations in two one-dimensional directions respectively; the output feature map after maximum pooling along the height dimension is recorded as Figure B (with a size of height × 1 × number of channels), which is used to highlight the significant response area in the image; the output feature map after average pooling along the width dimension is recorded as Figure C (with a size of 1 × width × number of channels), which is used to extract the global context information of the image.

[0045] Secondly, one-dimensional convolution operations are performed on Figures B and C respectively to extract their directional features. The convolution kernel size is 1, the step size is 1, the number of output channels is set to the channel compression ratio, and the convolved feature maps are spliced ​​in the channel dimension to obtain the fused feature tensor.

[0046] Next, another set of one-dimensional convolution operations is performed on the fused feature tensor to restore the number of channels to the original input channel number, and split operations are performed along the height dimension and width dimension respectively to obtain weight feature maps in two directions.

[0047] Finally, the input feature map A is subjected to channel-by-channel weighted multiplication operations according to the weight feature maps in the two directions mentioned above, thereby generating an enhanced feature map. The enhanced feature map can highlight the key areas in the input feature map A while retaining the global structural information.

[0048] The coordinate attention module MAP-CA implements the coordinate attention mechanism through the combined use of maximum pooling and average pooling in different directions, convolution to extract directional features, channel splicing and weighted enhancement operations, effectively improving the feature expression ability of the model in small target recognition and complex background conditions.

[0049] The expression of the coordinate attention module MAP-CA is as follows:

[0050] F att (I aug )=MAP-CA(I aug )

[0051] Where, F att (I aug ) is the input image I aug The enhanced feature map obtained after processing by the coordinate attention module MAP-CA; MAP is an attention mapping mechanism combining maximum pooling and average pooling, which is used to extract the response information of the salient area from the original image; CA (I aug ) is the weight map generated by the coordinate attention module MAP-CA, which strengthens the model's attention to key areas by modeling the key spatial position information and inter-channel correlation in the image; specifically, the CA module belongs to a composite attention mechanism, which includes channel attention and spatial attention. Channel attention expresses the importance of features by applying different weights to the feature maps of each channel, while spatial attention strengthens the perception of the target area by allocating weights in the spatial dimension.

[0052] The MAP-CA module builds on the traditional coordinate attention mechanism by introducing a parallel structure of maximum pooling and average pooling to capture global saliency and background information, respectively. It then generates one-dimensional attention weight vectors in both directions. By fusing these weights in the height and width dimensions, the resulting attention weight map is multiplied element-wise with the original input feature map as a weighting factor, enhancing the input features and significantly improving the model's ability to represent key regions within the image structure. With its lightweight structure and strong expressiveness, the MAP-CA module has been widely used in lightweight object detection and image recognition scenarios.

[0053] Subsequently, in order to improve the efficiency of feature extraction and reduce computational overhead, the above feature map is input into the fused convolution module GSConv. The fused convolution module GSConv is a lightweight convolution fusion structure proposed in the prior art and has been applied in models such as YOLOv6 and YOLOv7. It combines standard convolution with depthwise separable convolution to achieve efficient extraction of spatial information and low-overhead fusion of channel information. Its expression is as follows:

[0054]

[0055] Where: F input is the input feature map; F depthwise It is the feature map after depth convolution; F gsconvThe output feature map after the image feature map is input into the GSConv module; DepthwiseConv(·) is a depth-wise separable convolution operation, which is used to independently extract spatial features for each channel of the image feature map. It belongs to the existing technology and is a lightweight convolution operation proposed by Google in MobileNet in 2017; PointwiseConv(·) is a point-by-point convolution operation, which is used to achieve cross-channel fusion. It belongs to the existing technology. Pointwise Convolution (1×1 convolution) is the second step in the depth-wise separable convolution structure and is used for cross-channel feature fusion; the symbol ⊕ represents the feature fusion operation, which specifically adopts channel splicing. Specifically, channel splicing refers to: splicing feature maps from different branches or processing paths on the channel dimension such as C dimension to form a new feature map with richer feature representation. The splicing operation is essentially to stack each input tensor in the NCHW format direction. The number of channels of the resulting feature map is the sum of the number of channels of each feature map before splicing; ⊕ represents the feature fusion operation, which adopts any calculation method of channel splicing or weighted summation.

[0056] The introduction of the MAP-CA attention mechanism can improve the robustness of YOLOv8 in complex backgrounds. Specifically, the MAP-CA attention mechanism effectively improves the focus on key information areas by combining the advantages of maximum pooling and average pooling. It can focus on significant areas of the image at different scales, especially areas with small defects, and enhance the model's sensitivity to defects. Among them, maximum pooling helps extract the most significant features in the image, and average pooling helps the model obtain global context information. MAP-CA improves attention to the target area by coordinating the fusion of the two.

[0057] The GSConv module, namely the depthwise separable convolution module, is introduced in the Neck layer to reduce the amount of model computation and improve feature extraction efficiency. The GSConv module combines ordinary convolution and depthwise separable convolution, which can effectively reduce the amount of computation while maintaining efficient feature extraction capabilities. In addition, the GSConv module first performs separate convolution processing on each input channel, and then performs convolution fusion on the output, thereby greatly reducing the amount of computation.

[0058] Specifically, the Neck layer in the present invention refers to the Neck layer structure in the improved insulator defect detection model based on YOLOv8; the target detection framework model YOLOv8 structure includes: Backbone layer, Neck layer and Head layer, among which the Backbone layer serves as the backbone network for extracting basic image features; the Neck layer serves as the intermediate structure connecting the backbone and the detection head for feature fusion and enhancement; the Head layer is used to output the final detection results, including the output target box, classification probability, etc.

[0059] In order to take into account the classification accuracy, target positioning accuracy and effectiveness of the attention mechanism, a multi-task joint loss function is constructed as the training target of the insulator defect detection model; the multi-task joint loss function includes the category loss function, the bounding box regression loss function, the IOU loss function and the attention mechanism loss function.

[0060] Among them: the category loss function adopts cross entropy or binary cross entropy; the bounding box regression loss function adopts CIoU loss function or DIoU loss function, and the Smooth L1 loss function can be additionally selected to enhance the robustness of the predicted box regression; the IoU loss function adopts GIoU loss function or EIoU loss function, both of which introduce geometric compensation terms on the basis of traditional IoU calculation, and improve the regression stability and convergence speed through the minimum enclosing rectangle and width and height difference terms respectively; the attention mechanism loss function adopts one of the perceptual loss or attention heatmap matching loss; it should be noted that the category loss function, bounding box regression loss function, IOU loss function and attention mechanism loss function are all existing technologies.

[0061] Specifically, the expression of the multi-task joint loss function is as follows:

[0062] L total =α det L cls +β det L box +γL iou +δL att

[0063] in:

[0064]

[0065] L box =1-CIoU(B pred ,B true )

[0066] L iou =1-GIoU(B pred ,B true )

[0067]

[0068] Where, L total is the multi-task joint loss function; L cls is the category loss function, and the cross entropy loss function is used to measure the deviation between the model's predicted category probability and the true category label; L box is the bounding box regression loss function, which can comprehensively consider the factors of bounding box position overlap, center point distance and aspect ratio through CIoU loss measurement; L iou is the intersection-over-union loss function; Latt is the attention mechanism loss function; α det , β det , γ, and δ are weight coefficients used to balance the contribution of different loss items; C is the number of categories; y c is the true label; p c is the predicted category probability; CIoU is the loss function for bounding box regression in target detection tasks; B pred is the predicted bounding box; B true is the real border; L iou is the intersection-over-union loss function, used to measure the prediction box B pred With the real box B true The degree of matching in geometric position uses the generalized intersection-over-union function. The generalized intersection-over-union function introduces the area difference compensation term of the minimum enclosing rectangle based on the traditional IoU. It can provide effective gradient information even when the predicted box and the target box have no intersection, thereby improving the regression stability and convergence speed. att is the attention mechanism loss function, which is used to measure the image generated by the super-resolution model I SR With high-resolution real image I HR The difference in semantic feature space is computed by the super-resolution model to generate image I SR With high-resolution real image I HR The differences in feature space, the feature maps extracted from the model-generated images and the real target images respectively, calculate the Euclidean distance difference between their specific layer outputs in the VGG network, and the inconsistency in perceptual dimensions such as texture and edges, thereby guiding the model to focus on key attention areas more accurately.

[0069] Step 5: Based on the fusion multi-scale feature strategy and the decoupling head detection module, the insulator defect detection model performs defect detection and classification on the insulator defect image. The fusion multi-scale feature strategy is used to enhance the recognition ability of targets of different sizes, and the decoupling detection head module is used to separate and optimize the classification and positioning tasks. Ultimately, the location coordinates and type of the defect are output to achieve accurate identification. Specifically:

[0070] Since insulator defects in actual images often vary in size and are affected by complex backgrounds, especially tiny cracks or discharge marks, they may appear in regions of different scales. Therefore, in order to effectively integrate image features from different scales and enhance the model's ability to recognize both large and small targets, a multi-scale feature fusion strategy is adopted, which is expressed as follows:

[0071]

[0072] Where, F fusedis the output feature map after multi-scale feature fusion, which serves as the input feature of the subsequent decoupling detection head module, and is used to take into account the information of large and small targets at the same time, thereby improving the detection accuracy of multi-scale defects in complex images; s is the number of scales involved in the fusion, which is 3, including small scale, medium scale, and large scale; W s is the weighting coefficient; is the feature map extracted from the s-th scale;

[0073] In order to optimize the classification accuracy and target positioning accuracy respectively, a decoupling loss function L is constructed in the decoupling head detection module. decouple , which is expressed as follows:

[0074] L decouple =α head L cls +β head L box

[0075] Where, L decouple is the decoupling loss function; α head and β head are all weight coefficients, and α head +β head =1,α head , β head ∈(0,1);

[0076] Finally, based on the fusion of multi-scale feature strategy and decoupling head detection module, the output results of the insulator defect detection model include location information, classification probability and defect type, which are expressed as follows:

[0077] DetectionResults={(x,y,w,h,p,label)}

[0078] Where DetectionResults is the detection result; p is the classification probability of the insulator defect; label is the type of the insulator defect.

[0079] Step 6: To further improve model accuracy, a real-time feedback mechanism is introduced to enable the detection process to have continuous self-adaptation capabilities. The real-time feedback mechanism includes: constructing a feedback-enhanced loss function and dynamically adjusting model parameters. Specifically:

[0080] By comparing the model output with the manual annotation results, the error term is calculated and the feedback enhancement loss function is constructed based on it. Its expression is as follows:

[0081]

[0082] Where N is the total number of samples participating in feedback optimization; y i The true labels are manually annotated; is the predicted label of the model; p i is the confidence of the model prediction; λ and μ are adjustment coefficients, and λ+μ=1; L feedback Enhance the loss value for the final feedback to drive the model for incremental updates;

[0083] Dynamically adjusting model parameters includes enhancing weight adjustment and detection threshold adjustment. Detection threshold adjustment refers to dynamically adjusting the threshold for defect classification when missed detection or false detection occurs in historical feedback. The expression for detection threshold adjustment is as follows:

[0084] Threshold new =α thr Threshold old +β thr Feedback error

[0085] Where, Threshold new is the updated detection threshold; Threshold old is the detection threshold before updating; Feedback error The error calculated from historical feedback is the error term calculated based on the difference between the model prediction results and the manual annotation results in the historical detection records. It is calculated using conventional methods such as mean square error or cross entropy loss, indicating the deviation of the model's classification accuracy on previous samples. error is the model prediction error term defined in step 6; α thr , β thr Both are dynamic adjustment coefficients of model thresholds, which are used to balance the importance of historical thresholds and feedback errors. They can be manually set empirical values, such as α thr =0.8, β thr =0.2; or adopt a dynamic adjustment method based on the error ratio, determine the adjustment range according to the current feedback error ratio, and satisfy α thr +β thr =1.

[0086] Step 7: Based on the error feedback information calculated in step 6, a weighted combination loss function is constructed and gradient update optimization is performed to perform incremental learning on the insulator defect detection model. This allows for continuous learning and continuously improves the ability to detect minor defects in complex scenarios, forming a closed-loop optimization process. Specifically:

[0087] In order to comprehensively consider the three factors of classification, positioning and historical error in the feedback learning mechanism, a weighted combination loss function is constructed as the final training target. The expression of the weighted combination loss function is as follows:

[0088] L feedback-enhanced =αL cls+βL box +γL feedback

[0089] Among them, the feedback loss L feedback It is used to measure the difference between the model prediction confidence and the human confirmation label, which is defined as follows:

[0090]

[0091] Where, L feedback-enhanced is a weighted combination loss function, which represents the total loss function used for gradient update during model training; L feedback is the error term loss between the predicted output of the insulator defect detection model and the manual feedback; α, β, and γ are the weighting coefficients of classification loss, border loss, and feedback loss, respectively, and α+β+γ=1, which is used to balance the contribution of each loss term; N is the total number of samples involved in the feedback error calculation; Confidence in model predictions; The correct label is manually confirmed;

[0092] Gradient update optimization refers to using feedback to enhance the loss function L feedback The weight parameter θ of the insulator defect detection model is optimized by gradient descent, and its expression is as follows:

[0093]

[0094] Where η is the learning rate; is the gradient of the total loss with respect to the model parameters; the gradient update optimization constitutes a continuous learning mechanism of the system, which can use real feedback to continuously improve the insulator defect detection model's ability to detect tiny defects and complex backgrounds in actual scenarios.

[0095] To achieve continuous optimization, the parameters in the insulator defect detection model are iteratively updated according to the gradient of the feedback enhancement loss function. The update rule is as follows:

[0096]

[0097] Where θ is the model weight parameter; η is the learning rate; It is the gradient of the feedback enhancement loss function with respect to the model parameters θ.

[0098] The beneficial effects of the present invention are:

[0099] Compared with the traditional distribution network transmission line insulator damage detection method, the improved distribution network transmission line insulator damage detection method based on YOLOv8 in the present invention has higher detection accuracy and robustness, can effectively improve the efficiency and accuracy of fault detection, and has the characteristics of automation, real-time and high adaptability, significantly reducing the burden and cost of manual inspections, improving the safety and reliability of the power system, and by adopting deep learning technology and data enhancement methods, it can work stably in complex environments, providing strong support for the operation and maintenance of power equipment; specifically:

[0100] 1. This paper introduces a depth-wise separable convolutional module and a MAP-CA attention mechanism based on an improved object detection algorithm based on YOLOv8, which greatly improves detection accuracy and robustness. In particular, the detection capability of small target defects is significantly enhanced under complex backgrounds and changing lighting conditions.

[0101] 2. This invention, by combining super-resolution reconstruction technology, facilitates deep learning to restore details in low-resolution images, especially for clearly restoring tiny defects such as cracks and discharge marks, thereby improving image resolution and clarity. In addition, by combining data enhancement techniques such as rotation, flipping, and cropping, the diversity of training data is increased, further improving the generalization ability of the model.

[0102] 3. By adding a real-time feedback mechanism, the present invention enables the detection model to have a dynamic update strategy, and can adjust the model processing method according to the detection results to improve the detection accuracy and optimize subsequent detection tasks. It can effectively improve the real-time and accuracy of detection, and is particularly suitable for application scenarios with high efficiency and high safety requirements such as power inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0103] Figure 1 Schematic diagram of the process of this embodiment;

[0104] Figure 2 is a typical insulator defect image sample collected in this embodiment;

[0105] Figure 3 Schematic diagram showing the comparison of the insulator defect image before and after reconstruction and enhancement using the super-resolution algorithm in this embodiment;

[0106] Figure 4 Schematic diagram of the YOLOv8 model structure;

[0107] Figure 5 Schematic diagram of the improved YOLOv8 model structure in this embodiment;

[0108] Figure 6 Schematic diagram of the detection result of the insulator defect image A according to this embodiment;

[0109] Figure 7 Schematic diagram of the detection result of the insulator defect image B according to this embodiment;

[0110] Figure 8 Schematic diagram of the detection result of the insulator defect image C according to this embodiment;

[0111] Figure 9 Schematic diagram of the detection result of the insulator defect image D according to this embodiment;

[0112] Figure 10 Schematic diagram comparing the detection effects of different detection models for insulator defect image E;

[0113] Figure 11 Schematic diagram comparing the detection effects of different detection models for insulator defect image F;

[0114] Figure 12 Schematic diagram comparing the detection effects of different detection models for insulator defect image G;

[0115] Figure 13 Schematic diagram comparing the detection effects of different detection models for the insulator defect image H. DETAILED DESCRIPTION

[0116] The technical solution of the present invention is described below in conjunction with the accompanying drawings and implementation methods.

[0117] Example

[0118] The present invention is applied to the insulator defect detection of a certain aerial transmission line image as an example. Figure 1 As shown, the detection process for insulator defect detection in this embodiment includes: image acquisition and preprocessing, data enhancement, super-resolution image restoration, model construction, defect detection, result feedback and parameter update, forming a closed-loop iterative mechanism of "training-detection-feedback-optimization" to enhance the continuous learning and adaptability of the defect detection model in practical applications.

[0119] This embodiment specifically includes the following steps:

[0120] Step 1: Collect several insulator defect images, such as Figure 2 As shown, the typical insulator defect images collected in this embodiment include: insulator damage, insulator pollution, insulator aging and insulator flashover, which serve as the basic data source for the detection model training and detection verification of this embodiment;

[0121] The collected insulator image is resized, denoised, and normalized to obtain a standardized preprocessed image; specifically:

[0122] The normalized expression is:

[0123]

[0124] Where, I norm (x, y) is the pixel value of the normalized image; I orig (x,y) is the pixel value of the original image; x,y is the spatial coordinate of the image pixel.

[0125] Step 2: To enhance the model's adaptability to defects in different environments and postures, perform data augmentation on the standardized preprocessed image; data augmentation includes at least one of rotation, flipping, or cropping.

[0126] Step 3: Based on data enhancement, the deep learning-based super-resolution image reconstruction algorithm ESRT is used to enhance the image details of the insulator defect area in the image, restoring the detailed texture lost due to low resolution or shooting noise. Specifically:

[0127] The expression of the super-resolution reconstruction algorithm is as follows:

[0128] I rec =ESRT(I lowres ;θ)

[0129] Where, I rec is the restored image; ESRT is the super-resolution reconstruction algorithm; I lowres is the standardized preprocessed image; θ is the parameter;

[0130] Next, a composite loss function is constructed to guide the image restoration process, so that the restored image has high quality in terms of pixel accuracy and semantic consistency. Specifically, the expression of the composite loss function is as follows:

[0131] L SR =λ1L reconstruction +λ2L perception

[0132] Where, L SR is the composite loss function; L reconstruction is the reconstruction loss function, which is used to calculate the mean square error between image pixels to measure the difference between the super-resolution restored image and the original high-resolution image at the pixel level; L perception is a perceptual loss function that calculates the difference value in high-dimensional feature space and is based on the intermediate layer output of the VGG pre-trained neural network to capture the difference in perceptual consistency of the image in semantics and texture; λ1 and λ2 are both weight coefficients, and λ1+λ2=1.

[0133] like Figure 3As shown in the figure, after the image data enhancement in step 2 and the super-resolution image reconstruction in step 3, the edges of insulator defects such as cracks are clearer and the texture coherence is enhanced, which helps to improve the subsequent detection accuracy, especially in images with complex backgrounds or blurred edges, showing better detection robustness.

[0134] Step 4: To improve detection accuracy and robustness, we continue to construct a multi-task joint loss function that includes the category loss function, bounding box regression loss function, IOU loss function, and attention mechanism loss function as the training target of the insulator defect detection model;

[0135] like Figure 4 As shown, the YOLOv8 model is known to have certain feature extraction and detection capabilities, but it has a problem of reduced detection accuracy when processing small targets or complex background images. Therefore, this embodiment optimizes the YOLOv8 basic model to improve its detection capabilities in high-voltage transmission scenarios.

[0136] like Figure 5 As shown in the figure, this embodiment introduces the MAP-CA attention mechanism and the GSConv efficient convolution module based on the structure of the YOLOv8 basic model to obtain the MAP-YOLOv8 detection model. The MAP-CA module combines maximum and average pooling to enhance the model's perception of key insulator areas such as fine cracks and discharge traces. The GSConv module can improve feature extraction efficiency and reduce computational redundancy. The improved YOLOv8 model significantly outperforms the original YOLOv8 in detection accuracy, robustness, and speed.

[0137] Specifically, the expression of this embodiment is as follows:

[0138] (x,y,w,h,p)=YOLOv8(F gsconv (F att (I aug )))

[0139] After image enhancement, it is first processed by the coordinate attention module MAP-CA to extract the semantic expression of key areas. The process is defined as follows:

[0140] F att (I aug )=MAP-CA(I aug )

[0141] Subsequently, in order to improve the efficiency of feature extraction and reduce computational overhead, the above feature map is input into the fused convolution module GSConv and processed as follows:

[0142]

[0143] Where: x, y represent the position coordinates of the target frame; w, h are the width and height of the target frame; p is the predicted probability that the target frame belongs to the defect category; I aug is the enhanced input image; F att (I aug ) is the input image I aug Image feature map processed by the MAP-CA attention mechanism module fused with maximum pooling and average pooling; CA(I aug ) is the process of modeling the salient regions of image channels and spatial structures by the attention mechanism; F input is the input feature map; F depthwise It is the feature map after depth convolution; F gsconv is the output feature map after the image feature map is input into the GSConv module; DepthwiseConv(·) is a depthwise separable convolution operation, which is used to independently extract spatial features for each channel of the image feature map; PointwiseConv(·) is a point-by-point convolution operation, which is used to achieve cross-channel fusion; ⊕ represents the feature fusion operation, which uses either channel splicing or weighted summation.

[0144] In order to balance classification accuracy, target positioning accuracy, and the effectiveness of the attention mechanism, a multi-task joint loss function is constructed as the training target of the insulator defect detection model. The multi-task joint loss function includes the category loss function, the bounding box regression loss function, the IOU loss function, and the attention mechanism loss function.

[0145] The expression of the multi-task joint loss function is as follows:

[0146] L total =α det L cls +β det L box +γL iou +δL att

[0147] in:

[0148]

[0149] L box =1-CIoU(B pred ,B true )

[0150] Where, L total is the multi-task joint loss function; L cls is the category loss function, and the cross entropy loss function is used to measure the deviation between the model's predicted category probability and the true category label; L boxis the bounding box regression loss function, which can comprehensively consider the factors of bounding box position overlap, center point distance and aspect ratio through CIoU loss measurement; L iou is the intersection-over-union loss function; L att is the attention mechanism loss function; α det , β det , γ, and δ are weight coefficients used to balance the contribution of different loss items; C is the number of categories; y c is the true label; p c is the predicted category probability; CIoU is the loss function for bounding box regression in target detection tasks; B pred is the predicted bounding box; B true The real border.

[0151] Step 5: Based on the fusion multi-scale feature strategy and the decoupling head detection module, the insulator defect detection model performs defect detection and classification on the insulator defect image; specifically:

[0152] The expression of the multi-scale feature fusion strategy is as follows:

[0153]

[0154] Where, F fused is the output feature map after multi-scale feature fusion, which serves as the input feature of the subsequent decoupling detection head module, and is used to take into account the information of large and small targets at the same time, thereby improving the detection accuracy of multi-scale defects in complex images; s is the number of scales involved in the fusion, which is 3, including small scale, medium scale, and large scale; W s is the weighting coefficient; is the feature map extracted from the s-th scale.

[0155] In order to optimize the classification accuracy and target positioning accuracy respectively, the following loss function structure is constructed in the decoupling head detection module to separate the classification loss and regression loss into models. The expression is as follows:

[0156] L decouple =α head L cls +β head L box

[0157] Where, L decouple is the decoupling loss function; α head and β head are all weight coefficients, and α head +β head =1,α head , β head ∈(0,1).

[0158] Finally, based on the fusion of multi-scale feature strategy and decoupling head detection module, the output results of the insulator defect detection model include location information, classification probability and defect type, which are expressed as follows:

[0159] DetectionResults={(x,y,w,h,p,label)}

[0160] Where DetectionResults is the detection result; p is the classification probability of the insulator defect; label is the type of the insulator defect.

[0161] Figure 6 、 7 , 8, and 9 respectively show the results of insulator defect detection performed on four real insulator images A, B, C, and D according to this embodiment. Specifically:

[0162] Figure 6 In Chinese: "Insulator 0.92" means that in this embodiment, the target in the detection frame is determined to be an insulator with a confidence level of 92%;

[0163] Figure 7 In the example, “flashover defect 0.84” means that in this embodiment, the target in the detection frame is determined to be an insulator with a flashover defect at a confidence level of 84%; “insulator 0.92” means that in this embodiment, the target in the detection frame is determined to be an insulator with a confidence level of 92%.

[0164] Figure 8 In the example, "shockproof hammer 0.78" means that the target in the detection frame is determined to be a shockproof hammer with a confidence level of 78% in this embodiment; "shockproof hammer 0.82" means that the target in the detection frame is determined to be a shockproof hammer with a confidence level of 82% in this embodiment; "insulator 0.91" means that the target in the detection frame is determined to be an insulator with a confidence level of 91% in this embodiment;

[0165] Figure 9 In the figure, “insulator 0.94” means that in this embodiment, the target in the detection frame is determined to be an insulator with a confidence level of 94%; “insulator 0.89” means that in this embodiment, the target in the detection frame is determined to be an insulator with a confidence level of 89%; “insulator 0.92” means that in this embodiment, the target in the detection frame is determined to be an insulator with a confidence level of 92%; and “insulator 0.87” means that in this embodiment, the target in the detection frame is determined to be an insulator with a confidence level of 87%.

[0166] Depend on Figures 6-9 It can be seen that this embodiment can relatively accurately detect various minor defects and identify power grid components such as insulators and shock absorbers, and the detection boundary box has a high degree of fit, indicating that this embodiment has good detection accuracy and generalization ability.

[0167] Step 6: To further improve model accuracy, a real-time feedback mechanism is introduced to enable the detection process to have continuous self-adaptation capabilities. The real-time feedback mechanism includes constructing a feedback-enhanced loss function and dynamically adjusting model parameters. Specifically:

[0168] By comparing the model output with the manual annotation results, the error term is calculated and the feedback enhancement loss function is constructed based on it. Its expression is as follows:

[0169]

[0170] Where N is the total number of samples participating in feedback optimization; y i The true labels are manually annotated; is the predicted label of the model; p i is the confidence of the model prediction; λ and μ are adjustment coefficients, and λ+μ=1; L feedback Enhance the loss value for the final feedback to drive the model for incremental updates;

[0171] Dynamically adjusting model parameters includes enhancing weight adjustment and detection threshold adjustment. Detection threshold adjustment refers to dynamically adjusting the threshold of defect classification when missed detection or false detection occurs in historical feedback.

[0172] The expression for detection threshold adjustment is as follows:

[0173] Threshold new =α thr Threshold old +β thr Feedback error

[0174] Where, Threshold new is the updated detection threshold; Threshold old is the detection threshold before updating; Feedback error The error calculated from historical feedback is the error term calculated based on the difference between the model prediction results and the manual annotation results in the historical detection records. It is calculated using conventional methods such as mean square error or cross entropy loss, indicating the deviation of the model's classification accuracy on previous samples. error is the model prediction error term defined in step 6; α thr , β thr are adjustment coefficients, and α thr +β thr =1,α thr , β thr ∈(0,1).

[0175] Step 7: Based on the error feedback information calculated in step 6, a weighted combination loss function is constructed and gradient update optimization is performed to perform incremental learning on the insulator defect detection model. This allows for continuous learning and continuously improves the ability to detect minor defects in complex scenarios, forming a closed-loop optimization process. Specifically:

[0176] In order to comprehensively consider the three factors of classification, positioning and historical error in the feedback learning mechanism, a weighted combination loss function is constructed as the final training target, which is expressed as follows:

[0177] L feedback-enhanced =αL cl s+βL box +γL feedback

[0178] Among them, the feedback loss L feedback It is used to measure the difference between the model prediction confidence and the human confirmation label, which is defined as follows:

[0179]

[0180] Where, L feedback-enhanced is a weighted combination loss function, which represents the total loss function used for gradient update during model training; L feedback is the error term loss between the predicted output of the insulator defect detection model and the manual feedback; α, β, and γ are the weighting coefficients of classification loss, border loss, and feedback loss, respectively, and α+β+γ=1, which is used to balance the contribution of each loss term; N is the total number of samples involved in the feedback error calculation; Confidence in model predictions; The correct label is manually confirmed.

[0181] Gradient update optimization refers to using feedback to enhance the loss function L feedback The weight parameter θ of the insulator defect detection model is optimized by gradient descent, and its expression is as follows:

[0182]

[0183] Where η is the learning rate; is the gradient of the total loss with respect to the model parameters; the gradient update optimization constitutes a continuous learning mechanism of the system, which can continuously adjust the insulator defect detection model using real feedback, and improve the adaptability to tiny defects and complex backgrounds in actual scenarios.

[0184] To achieve continuous optimization, the parameters in the insulator defect detection model are iteratively updated according to the gradient of the feedback enhancement loss function. The update rule is as follows:

[0185]

[0186] Where θ is the model weight parameter; η is the learning rate; It is the gradient of the feedback enhancement loss function with respect to the model parameters θ.

[0187] In order to verify the effectiveness of the insulator defect detection model proposed in this embodiment, a multi-model comparison experiment was carried out. The experiment selected 4 transmission line image samples and used YOLOv5, YOLOv7, YOLOX, YOLOv8 and the MAP-YOLOv8 detection model of this embodiment for detection tests. The detection results of each model for the same transmission line image sample are shown in the following figure. Figures 10 to 13 shown.

[0188] according to Figures 10-13 It can be seen that although YOLOv5, YOLOv7 and YOLOX have certain detection capabilities when identifying complete insulators, they are prone to problems such as detection frame offset, target omission or insufficient classification confidence when identifying edge targets, complex lighting or low-resolution areas. Especially in the insulator strings and damaged areas in the center or bottom of the image, the three often have incomplete detection and inaccurate category judgment.

[0189] The detection effect of the YOLOv8 model is improved compared with traditional models. It is more accurate in insulator integrity detection and bounding box positioning, and can also preliminarily identify some defective areas. However, when facing small-sized targets or areas with complex backgrounds, there are still problems with confidence fluctuations and boundary overlap.

[0190] In contrast, the MAP-YOLOv8 model proposed in this embodiment demonstrates higher detection accuracy and stability in multiple image regions: it can not only accurately identify multiple insulator components, damaged locations, and broken edges, but also provide high-confidence classification outputs for subtle defects such as flashover marks and hammer-shaped components. The detection frame is accurately positioned, the classification results are clear, the adaptability is stronger, and the robustness is significantly enhanced.

[0191] In summary, this embodiment demonstrates leading detection performance in different application scenarios. Due to the integration of attention mechanism and image enhancement structure, it has significant advantages in edge target recognition, abnormal component detection and multi-category differentiation of small targets. In particular, it has the best performance in dense target recognition and multi-target fine classification, and is suitable for the actual deployment needs of intelligent inspection of power transmission lines.

Claims

1. A method for detecting insulator damage in distribution network transmission lines based on the YOLO algorithm, characterized in that: The steps include: Collect several insulator defect images, perform size adjustment, denoising and normalization processing to obtain standardized pre-processed images; Perform data augmentation on standardized preprocessed images; The super-resolution reconstruction algorithm is used to enhance the image details of the insulator defect area in the image; Construct a composite loss function to guide the image detail enhancement process; An insulator defect detection model based on the YOLOv8 architecture was constructed. The MAP-CA attention mechanism module was introduced to strengthen the focus on image detail enhancement areas. At the same time, the Neck structure was embedded in the GSConv module in the YOLOv8 architecture. Construct a multi-task joint loss function as the training objective of the insulator defect detection model; Based on the fusion of multi-scale feature strategy and decoupling head detection module, the insulator defect detection model detects and classifies insulator defect images; Introducing a real-time feedback mechanism, which includes constructing a feedback enhancement loss function and dynamically adjusting model parameters; A weighted combination loss function is constructed and gradient update optimization is performed to improve the detection capability of the insulator defect detection model.

2. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The data enhancement includes at least one of rotation, flipping or cropping.

3. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The insulator defect includes at least one of insulator damage, discharge marks on the insulator surface, or damage to the shock-absorbing hammer structure.

4. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The composite loss function includes a reconstruction loss function and a perception loss function.

5. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The expression of the insulator defect detection model is as follows: (x,y,w,h,p)=YOLOv8(F gsconv (F att (I aug ))) in: F att (I aug )=MAP-CA(I aug ) Where: x, y represent the position coordinates of the target frame; w, h are the width and height of the target frame; p is the predicted probability that the target frame belongs to the defect category; I avg is the enhanced input image; F att (I aug ) is the image feature map after the input image is processed by the MAP-CA attention mechanism module with maximum pooling and average pooling; F gsconv The output feature map after the image feature map is input into the GSConv module; CA(I aug ) is the attention mechanism modeling process of salient regions of image channels and spatial structures; DepthwiseConv(·) is a depthwise separable convolution operation, which is used to independently extract spatial features from each channel of the image feature map; PointwiseConv(·) is a point-by-point convolution operation, which is used to achieve cross-channel fusion; ⊕ represents a feature fusion operation.

6. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The multi-task joint loss function includes a category loss function, a bounding box regression loss function, an IOU loss function, and an attention mechanism loss function; The expression of the multi-task joint loss function is as follows: L total =a det L cls +b det L box +γL iou +δL att in: L box =1-CIoU(B pred ,B true ) Where, L total is the multi-task joint loss function; L cls is the category loss function; L box is the bounding box regression loss function; L iou is the intersection-over-union loss function; l att is the attention mechanism loss function; α det , β det , γ, and δ are weight coefficients used to balance the contribution of different loss items; C is the number of categories; y c is the true label; p c is the predicted category probability; CIoU is the loss function for bounding box regression in target detection tasks; B pred is the predicted bounding box; B true The real border.

7. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The expression of the multi-scale feature fusion strategy is as follows: Where, F fused is the output feature map after multi-scale feature fusion; s is the number of scales involved in fusion; W s is the weighting coefficient; is the feature map extracted from the s-th scale.

8. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The expression of the decoupling head detection module is as follows: L decouple =a head L cls +b head L box Where, L decouple is the decoupling loss function; α head and β head are all weight coefficients, and α head +β head =1,α head , β head ∈(0,1).

9. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The expression of the feedback enhancement loss function is as follows: Where N is the total number of samples participating in feedback optimization; y i The true labels are manually annotated; is the predicted label of the model; p i is the confidence of the model prediction; λ and μ are adjustment coefficients, and they satisfy λ+μ=1; L feedback Enhance the loss value for feedback.

10. The method for detecting insulator damage in a power distribution network transmission line based on the YOLO algorithm according to claim 1, characterized in that: The expression of the weighted combination loss function is as follows: L feedback-enhanced =αL cls +βL box +γL feedback in: Where, L feedback-enhanced is the weighted combination loss function; α, β, γ are all weighted coefficients, and α+β+γ=1; L feedback is the error term loss between the predicted output of the insulator defect detection model and the manual feedback; N is the total number of samples involved in the feedback error calculation; Confidence in model predictions; The correct label is manually confirmed; The expression of the gradient update optimization is as follows: Where θ is the model weight parameter; η is the learning rate; It is the gradient of the feedback enhancement loss function with respect to the model parameters θ.

Images