Defect detection method for vibration damper of power transmission line

By constructing and training deep learning models, identifying and positioning the defects of the anti-vibration hammer of the transmission line, the problems of low identification efficiency and poor accuracy in the existing technology are solved, and the rapid and accurate identification and positioning of the anti-vibration hammer of the transmission line are achieved. Improve the safety and stability of the power system.

CN118918072APending Publication Date: 2024-11-08NANCHANG UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410954102.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing technology is difficult to quickly and accurately identify and locate different defects of the anti-vibration hammer of the transmission line, which leads to the threat of the safe and stable operation of the power system, and the manual defect check efficiency is low, which can easily lead to false detection and missed inspection.

Method used

Build an image data set of anti-vibration hammer defects in transmission lines, train an anti-vibration hammer defect detection model based on deep learning, use the backbone feature extraction network, feature fusion network and classification prediction network, combine the image data to collect image data, perform sample expansion and data Strengthen to achieve identification of defects such as defects, rust, collisions, etc.

Benefits of technology

It realizes the rapid and accurate identification and positioning of anti-vibration hammer defects in transmission lines, reduces manual interference, improves the accuracy of defect inspections, and reduces false detection and missed inspections. It assists the inspection personnel in the line to accurately identify and regional maintenance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918072A_ABST
    Figure CN118918072A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of power transmission lines, in particular to a power transmission line damper defect detection method, which specifically comprises the following steps: constructing a power transmission line damper defect image data set; constructing a vibration damper defect detection model; training a vibration damper defect detection model based on the defect image data set, and obtaining an optimal weight of the vibration damper defect detection model; and detecting a to-be-detected image by using the trained vibration damper defect detection model to obtain a detection result. According to the power transmission line damper defect detection method provided by the invention, the defect can be accurately and quickly identified and positioned only by inputting the power transmission line damper image obtained by inspection, and reference can be provided for intelligent inspection of power equipment in other application fields of the power industry.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of power transmission lines, and in particular to a method for detecting defects of anti-vibration hammers in power transmission lines. Background Art

[0002] As an important hardware in power transmission lines, the anti-vibration hammer is used to absorb or weaken the kinetic energy transmitted by wind to the overhead power transmission wires, so as to cut off the supply line of the wire vibration energy, thereby reducing the vibration damage of the overhead power transmission wires. However, since the anti-vibration hammer is exposed to the natural environment for a long time, it is very easy to produce metal cracks, loose bolts and other problems, which in turn cause the anti-vibration hammer to be damaged, rusted, and collided. The morphological and structural defects make the anti-vibration hammer unable to effectively play the anti-vibration role. If the defects of the anti-vibration hammer are not discovered and eliminated in time, it is very likely to cause the transmission wire to vibrate and dance in the breeze, which will seriously threaten the safe and stable operation of the power system.

[0003] A large amount of image data will be generated during the power grid inspection. In the subsequent image processing, it is necessary to manually observe and check the massive images. Long-term work can easily cause fatigue to the staff, which reduces the accuracy of the defect detection. With its powerful feature expression ability and end-to-end network structure, deep learning technology perfectly meets the needs of massive image defect detection, which can not only save labor costs, but also reduce false detection and missed detection. Most of the existing research on the anti-vibration hammer of the transmission line only focuses on the identification of the anti-vibration hammer or the fault classification of the anti-vibration hammer has only one defect. Different defects have different impacts on the safety of the transmission line, and the inspection department needs to have different response strategies, so it is necessary to refine the defects of the anti-vibration hammer. With the lightweight of the deep model and the enhancement of the computing power of small computers, machine vision has made rapid progress in the field of defect detection and classification of power equipment. The combination of machine vision and anti-vibration hammer defect images can not only avoid the interference of human factors, but also make quick and accurate judgments on the defects of the anti-vibration hammer, and then assist the line inspection personnel to accurately identify the defects of the anti-vibration hammer, which is of great significance for regional line maintenance. Summary of the invention

[0004] The object of the present invention is to provide a method for detecting defects of a transmission line anti-vibration hammer, so as to improve the problems existing in the prior art.

[0005] A method for detecting defects of anti-vibration hammers for power transmission lines comprises the following steps: constructing an image data set of defects of anti-vibration hammers for power transmission lines; constructing an anti-vibration hammer defect detection model; training the anti-vibration hammer defect detection model based on the defect image data set and obtaining the optimal weight of the anti-vibration hammer defect detection model; and detecting an image to be detected using the trained anti-vibration hammer defect detection model to obtain a detection result.

[0006] More preferably, the image data of the anti-vibration hammer of the transmission line is obtained, specifically including: collecting the image data of the anti-vibration hammer of the transmission line through a drone, and constructing an anti-vibration hammer defect image data set containing four categories of "defect", "rust", "collision" and "normal"; and expanding the sample of the anti-vibration hammer defect image data set, and dividing the anti-vibration hammer defect image data set into a training set, a verification set and a test set at a ratio of 8:1:1.

[0007] More preferably, the transmission line anti-vibration hammer image data includes images under backgrounds including the sky, forests, fields, sand, and houses; the image sample expansion methods include adding noise, adjusting brightness, rotating, cropping images, translating, and mirroring. The training set images are expanded by five times, and each picture is expanded using at least two methods. The expanded images are screened to eliminate images with poor quality.

[0008] More preferably, the anti-vibration hammer defect detection model is constructed by a backbone feature extraction network, a feature fusion network and a classification prediction network; the backbone feature extraction network is composed of a CBS module, a C2f-MLCAi module and a focus modulation module, where i=1, 2, 3, 4; three preliminary feature layers P1, preliminary feature layer P2 and preliminary feature layer P3 of different sizes are respectively derived from the C2f-MLCA2 module, the C2f-MLCA3 module and the focus modulation module, and input into the feature fusion network for feature fusion; the enhanced features S2, S3 and S5 after feature fusion are input into the classification prediction network for classification prediction, so as to realize the detection of image defects of the anti-vibration hammer of the transmission line.

[0009] More preferably, the construction of the backbone feature extraction network specifically includes: constructing a CBS module using ordinary convolution Conv, batch normalization BN and SiLU activation function; operating the CBS module through a split layer and performing skip-layer connection of n residual layers Bottleneck-MLCA to form a C2f-MLCAi module, wherein a lightweight hybrid local channel attention mechanism MLCA is introduced in the residual layer Bottleneck; the processing process of the focus modulation module is to perform focus modulation on the input image, slice the input features according to the focus window size and focus level, and perform affine transformation on the sliced ​​features and then normalize them as output.

[0010] More preferably, three preliminary feature layers P1, P2, and P3 of different scales are respectively derived from the C2f-MLCA2 module, the C2f-MLCA3 module, and the focusing modulation module, and input into the feature fusion network for feature fusion, specifically including: upsampling the feature P3 and fusing it with the feature P2 to obtain a fused feature F1; at the same time, the C2F module is used to extract the features of F1 to obtain an enhanced feature S1; S1 is further upsampled and fused with the feature P1 to obtain a fused feature F2; at this time, the C2F module is continued to extract the features of F2 to obtain an enhanced feature S2, and the enhanced feature S2 is combined with the enhanced feature S2 after the feature extraction operation is enhanced by the CBS module. 1 is fused to obtain fused feature F3. After the fused feature F3 is extracted by the C2F module, the enhanced feature S3 is obtained. After the enhanced feature S3 is extracted by the CBS module, it is fused with the feature P3 to obtain fused feature S4. After the fused feature S4 is extracted by the C2F module, the enhanced feature S5 is obtained. Finally, the enhanced features S2, S3 and S5 are used as the output of the feature fusion network.

[0011] More preferably, the classification prediction network consists of two convolution branches, one convolution branch consists of three convolution Convs and a prediction box loss module with a Slideloss sample weighting function added, and the other convolution branch consists of three convolution Convs and a classification loss module. The enhanced features S2, S3 and S5 are respectively input into three SHeads of different sizes of 80×80×256, 40×40×512 and 20×20×512 for classification prediction, so as to realize the detection of small, medium and large defects in the anti-vibration hammer image of the transmission line.

[0012] More preferably, the prediction box loss module of the Slideloss sample weighted function is specifically composed of the Slideloss sample weighted function, the CIoU loss function and the distribution focus loss DFL.

[0013] More preferably, the anti-vibration hammer defect detection model is trained based on the defect image data set, specifically including: inputting the divided transmission line anti-vibration hammer image training set into the anti-vibration hammer defect detection model for pre-training, a total of 200 rounds of training, and the training is divided into two stages: freezing and thawing. In the first stage, the MFS-YOLO network model backbone is frozen and the Mosaic and Mixup data enhancement methods are used cross-used to iterate the training for 50 times; in the second stage, the MFS-YOLO network model backbone is thawed, and then iterative training is performed for 150 rounds; after the training is completed, the training weight with the smallest loss value is taken, which is the optimal training weight of the anti-vibration hammer defect detection model.

[0014] More preferably, the trained anti-vibration hammer defect detection model is used to detect the image to be detected to obtain the detection result, which specifically includes: inputting the optimal training weight into the anti-vibration hammer defect detection model to perform defect detection, setting the intersection over union (IOU) threshold of the prediction box to 0.5 when detecting the defects of the anti-vibration hammer image of the transmission line, and setting the confidence threshold for determining whether it is a defective target to 0.01; using a non-maximum suppression (NMS) to eliminate redundant prediction boxes, and obtain the transmission line anti-vibration hammer defect detection result.

[0015] Compared with the prior art, the present invention has the following advantages:

[0016] The present invention provides a method for detecting defects of a transmission line anti-vibration hammer. Defects can be accurately and quickly identified and located by simply inputting an image of the transmission line anti-vibration hammer obtained during inspection. This method can provide a reference for other application fields of intelligent inspection of power equipment in the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of the method of the present invention;

[0018] Figure 2 It is the MFS-YOLO model structure used in the present invention;

[0019] Figure 3 The SHead structure of the classification prediction network of the MFS-YOLO model used in the present invention;

[0020] Figure 4 It is the C2f-MLCAi module structure of the MFS-YOLO model used in the present invention;

[0021] Figure 5 This is the defect detection result of the MFS-YOLO model.

[0022] Figure 6 This is the defect detection result of the MFS-YOLO model. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are 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 creative work are within the scope of protection of the present invention. Unless otherwise defined, the technical terms or scientific terms used herein should be understood by people with general skills in the field to which the present invention belongs. "Including" and similar words used in this article mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0024] The embodiment of the present invention provides a method for detecting defects of a transmission line anti-vibration hammer, comprising the following steps:

[0025] Step 1: construct a transmission line anti-vibration hammer defect image dataset and perform preprocessing;

[0026] The transmission line anti-vibration hammer image data is collected by drones. The transmission line anti-vibration hammer image data includes images under the background of sky, forest, field, sand, and house, which contains 1246 images in total. The original images are saved in GIF and TIF formats. For the convenience of subsequent research, they are uniformly converted into .JPG format; a data set of anti-vibration hammer defect images containing four categories of "defect", "rust", "collision", and "normal" is constructed; and the anti-vibration hammer defect image data set is sampled. The image sample expansion methods include adding noise, adjusting brightness, rotating, cropping images, translating, and mirroring. The training set images are expanded by five times. Each image uses at least two methods for sample expansion. The expanded images are screened and the images with poor quality are eliminated. The total number of expanded images is 2785. The number of defects of various types is shown in Table 1:

[0027] Table 1

[0028] Defect Type normal Rust damaged Collision Number of tags 4825 3310 841 398

[0029] The preprocessed images are labeled with the anti-vibration hammers and their defects using the LabelImg software: the anti-vibration hammers with defects are labeled defective; the anti-vibration hammers with rust are labeled rusty; the anti-vibration hammers with collisions are labeled collide; the anti-vibration hammers with normal collisions are labeled normal, and the positions of the label names are stored. The anti-vibration hammer defect image dataset is divided into training set, validation set and test set according to the ratio of 8:1:1. The training set is input into the MFS-YOLO network model for training. After each round of training, the validation set is used to verify the training effect of this round. If the loss value of the verification result is gradually decreasing, it proves that the training is effective. Otherwise, the training is invalid and the training is stopped.

[0030] Step 2: Construct the MFS-YOLO anti-vibration hammer defect detection model. The anti-vibration hammer defect detection model consists of three parts: the backbone feature extraction network MFNet, the feature fusion network PNeck, and the classification prediction network SHead. Among them, the backbone feature extraction network consists of the CBS module, the C2f-MLCAi module, and the focus modulation module, i = 1, 2, 3, 4; Figure 2 As shown, the specific operations are as follows:

[0031] The construction of the backbone feature extraction network specifically includes: constructing a CBS module using ordinary convolution Conv, batch normalization BN and SiLU activation function; the CBS module is operated through a split layer and connected to n residual layers Bottleneck-MLCA by skipping layers to form a C2f-MLCAi module, wherein a lightweight hybrid local channel attention mechanism MLCA is introduced in the residual layer Bottleneck; the processing process of the focus modulation module is to perform focus modulation on the input image, slice the input features according to the focus window size and focus level, and perform affine transformation on the sliced ​​features and then normalize them as output;

[0032] Specifically, the input image size is set to 640×640×3, and first input into the backbone feature extraction network MFNet. After two CBS module feature extractions, the image size becomes 160×160×128, and then the image size remains unchanged after the C2f-MLCA1 module feature extraction. After the CBS module operation, the image size becomes 80×80×256. The image size remains unchanged after the C2f-MLCA2 module feature extraction. After the CBS module operation, the image size becomes 40×40×512. The image size remains unchanged after the C2f-MLCA3 module feature extraction. After the CBS module operation, the image size becomes 40×40×512. After the C2f-MLCA4 module and the focus modulation module FocalModulation feature extraction, the image size remains unchanged, and the C2f-MLCA2 module, C2f-MLCA3 module and the focus modulation network FocalModulation are respectively connected. The three preliminary feature layers P1, P2, and P3 of different scales output by the Modulation module are used as the input of the feature fusion network PNeck, and the feature layer sizes are 80×80×128, 40×40×512, and 20×20×512 respectively; the C2f-MLCA module introduces hybrid local channel attention MLCA to consider channel information and spatial information while ensuring lightweight, and combines local information with global information to improve the expression effect of the model. The focus modulation module FocalModulation can process input images of different sizes, and accurately identify and locate information in the image, so that the model can detect defect targets more accurately;

[0033] Among them, the CBS module includes ordinary convolution Conv, batch normalization Batch Normalization and activation function SiLU.

[0034] Among them, the C2f-MLCAi structure is as follows Figure 4 As shown, the CBS module is constructed by performing a split layer operation and performing a skip-layer connection of n residual layers Bottleneck-MLCA.

[0035] Among them, the C2F module is composed of CBS, split module and n Bottleneck modules.

[0036] The processing process of the feature fusion network PNeck is as follows: feature P3 is upsampled and then fused with feature P2 to obtain fused feature F1; at the same time, the C2F module is used to extract the features of F1 to obtain the enhanced feature S1; S1 is further upsampled and fused with feature P1 to obtain the fused feature F2; at this time, the C2F module is continued to extract the features of F2 to obtain the enhanced feature S2, and the enhanced feature S2 is fused with the enhanced feature S1 after the enhanced feature extraction operation of the CBS module to obtain the fused feature F3, and the fused feature F3 is extracted by the C2F module to obtain the enhanced feature S3, and the enhanced feature S3 is fused with the feature P3 after the feature extraction operation of the CBS module to obtain the fused feature S4, and the fused feature S4 is extracted by the C2F module to obtain the enhanced feature S5, and finally the enhanced features S2, S3 and S5 are used as the output of the feature fusion network PNeck:

[0037] It also includes an upsampling module for enlarging low-resolution feature maps to high resolution for more detailed target detection;

[0038] The prediction network SHead consists of two convolutional branches, such as Figure 3 As shown in the figure, one convolution branch is composed of 3 convolution Convs and a prediction box loss module with SlideloSS sample weighting function added. The Slideloss function can alleviate the imbalance problem between samples. The other convolution branch is composed of 3 convolution Convs and a classification loss module. The enhanced features S2, S3 and S5 are respectively input into three different sizes of SHead of 80×80×256, 40×40×512 and 20×20×512 for anti-vibration hammer defect detection.

[0039] Among them, the prediction box loss module of the Slideloss sample weighted function is specifically composed of the SlidelosS sample weighted function, the CIoU loss function and the distribution focus loss DFL.

[0040] In short, three preliminary feature layers P1, P2 and P3 of different sizes are respectively derived from the C2f-MLCA2 module, the C2f-MLCA3 module and the focusing modulation module, and input into the feature fusion network for feature fusion; the enhanced features S2, S3 and S5 after feature fusion are input into the classification prediction network for classification prediction, so as to realize the detection of defects in the anti-vibration hammer image of the transmission line, and realize the detection of small, medium and large defects in the anti-vibration hammer image of the transmission line.

[0041] Compared with the single attention mechanism, the hybrid local channel attention mechanism MLCA can focus on local areas and channel features at the same time, so as to capture the information in the image more flexibly and consider the information in the spatial and channel dimensions, which is conducive to improving the model's comprehensive understanding of image features. The weights of different local areas and channel features can be dynamically adjusted according to actual conditions, so as to more accurately capture the importance of different areas and features in the image; the focus modulation module can dynamically modulate the feature map based on each region of interest, so as to adapt to the importance of features in different regions and modulate them accurately. Compared with the SPPF module, it has stronger dynamics, accuracy and adaptability; the Slideloss function can alleviate the imbalance problem between samples compared to ordinary loss functions, so the constructed MFS-YOLO anti-vibration hammer defect detection model has good detection and generalization capabilities.

[0042] Step 3: training the anti-vibration hammer defect detection model based on the defect image data set and obtaining the optimal weight of the anti-vibration hammer defect detection model; specifically including:

[0043] The divided transmission line anti-vibration hammer image training set is input into the anti-vibration hammer defect detection model for pre-training. A total of 200 rounds of training are carried out. The training is divided into two stages: freezing and thawing. In the first stage, the backbone of the MFS-YOLO network model is frozen and the Mosaic and Mixup data enhancement methods are used for cross-iteration training for 50 times, with a batch-size of 8 and a learning rate of 0.001. In the second stage, the backbone of the MFS-YOLO network model is thawed and iteratively trained for 150 rounds, with a batch-size of 4 and a learning rate of 0.0005. Multi-threaded training is enabled, and the number of threads num-workers is set to 2. The pre-processed transmission line anti-vibration hammer defect image dataset is used as input, and the Mosaic and Mixup data enhancement methods are used for training to obtain the training weight with the smallest loss value, which is the optimal training weight of the anti-vibration hammer defect detection model.

[0044] Step 4: Use the trained anti-vibration hammer defect detection model to detect the image to be detected and obtain the detection result;

[0045] The optimal training weights are input into the anti-vibration hammer defect detection model for defect detection. When detecting the anti-vibration hammer image defects of the transmission line, the intersection over union (IOU) threshold of the prediction box is set to 0.5, and the confidence threshold for determining whether it is a defect target is set to 0.01. The redundant prediction boxes are eliminated using a non-maximum suppression (NMS) method to obtain the transmission line anti-vibration hammer defect detection results.

[0046] The experimental results show that the mAP value of the anti-vibration hammer defect test image is as high as 95.23%. The visualization results of the anti-vibration hammer defect detection are as follows: Figure 5 As shown, the effectiveness of the method proposed in this embodiment is verified, which can assist line inspection personnel in accurately detecting anti-vibration hammer defects, thereby promoting differentiated prevention and control of transmission line faults.

[0047] Although the embodiments of the present invention are described in detail above, it is obvious to those skilled in the art that various modifications and variations can be made to these embodiments. However, it should be understood that such modifications and variations are within the scope and spirit of the present invention as described in the claims. Moreover, the present invention described herein may have other embodiments and may be implemented or realized in a variety of ways.

Claims

1. A method for detecting defects of anti-vibration hammers in power transmission lines, It is characterized in that The steps include: Construct a transmission line anti-vibration hammer defect image dataset; construct an anti-vibration hammer defect detection model; train the anti-vibration hammer defect detection model based on the defect image dataset and obtain the optimal weight of the anti-vibration hammer defect detection model; use the trained anti-vibration hammer defect detection model to detect the image to be detected and obtain the detection result.

2. A method for detecting defects of a transmission line anti-vibration hammer according to claim 1, It is characterized in that Acquire transmission line anti-vibration hammer image data, including: The image data of the anti-vibration hammer of transmission line is collected by unmanned aerial vehicle, and a dataset of anti-vibration hammer defect images containing four categories of "defect", "rust", "collision" and "normal" is constructed; the sample of the anti-vibration hammer defect image dataset is expanded, and the anti-vibration hammer defect image dataset is divided into training set, verification set and test set according to 8:1:

1.

3. A method for detecting defects of a transmission line anti-vibration hammer according to claim 2, It is characterized in that The image data of the anti-vibration hammer for transmission lines include images under backgrounds including the sky, forests, fields, sand, and houses; the image sample expansion methods include adding noise, adjusting brightness, rotating, cropping images, translating, and mirroring. The training set images are expanded by five times. Each picture is expanded using at least two methods. The expanded images are screened to eliminate images with poor quality.

4. A method for detecting defects of a transmission line anti-vibration hammer according to claim 2, It is characterized in that The anti-vibration hammer defect detection model is constructed by a backbone feature extraction network, a feature fusion network and a classification prediction network; the backbone feature extraction network is composed of a CBS module, a C2f-MLCAi module and a focus modulation module, where i=1, 2, 3, 4; three preliminary feature layers P1, P2 and P3 of different sizes are respectively derived from the C2f-MLCA2 module, the C2f-MLCA3 module and the focus modulation module, and input into the feature fusion network for feature fusion; the enhanced features S2, S3 and S5 after feature fusion are input into the classification prediction network for classification prediction, so as to realize the detection of image defects of the anti-vibration hammer of the transmission line.

5. A method for detecting defects of a transmission line anti-vibration hammer according to claim 4, It is characterized in that The construction of the backbone feature extraction network specifically includes: The CBS module is constructed using ordinary convolution Conv, batch normalization BN and SiLU activation function; the CBS module is operated through the split layer and skip-layer connection of n residual layers Bottleneck-MLCA to form a C2f-MLCAi module, in which a lightweight hybrid local channel attention mechanism MLCA is introduced in the residual layer Bottleneck; the processing process of the focus modulation module is to perform focus modulation on the input image, slice the input features according to the focus window size and focus level, and the sliced ​​features are affine transformed and normalized as output.

6. A method for detecting defects of a transmission line anti-vibration hammer according to claim 5, It is characterized in that Three preliminary feature layers P1, P2, and P3 of different scales are derived from the C2f-MLCA2 module, the C2f-MLCA3 module, and the focus modulation module, and input into the feature fusion network for feature fusion, including: After upsampling feature P3, it is fused with feature P2 to obtain fused feature F1; at the same time, the C2F module is used to extract features from F1 to obtain enhanced feature S1; S1 is further upsampled and fused with feature P1 to obtain fused feature F2; at this time, the C2F module is continued to be used to extract features from F2 to obtain enhanced feature S2, and the enhanced feature S2 is fused with the enhanced feature S1 after the enhanced feature extraction operation of the CBS module to obtain fused feature F3, and the fused feature F3 is extracted by the C2F module to obtain enhanced feature S3, and the enhanced feature S3 is fused with feature P3 after the feature extraction operation of the CBS module to obtain fused feature S4, and the fused feature S4 is extracted by the C2F module to obtain enhanced feature S5. Finally, the enhanced features S2, S3 and S5 are used as the output of the feature fusion network.

7. A method for detecting defects of a transmission line anti-vibration hammer according to claim 6, It is characterized in that The classification prediction network consists of two convolution branches. One convolution branch consists of three convolution Convs and a prediction box loss module with a Slideloss sample weighted function added. The other convolution branch consists of three convolution Convs and a classification loss module. The enhanced features S2, S3 and S5 are respectively input into three SHeads of different sizes (80×80×256, 40×40×512 and 20×20×512) for classification prediction, thereby realizing the detection of small, medium and large defects in the anti-vibration hammer image of the transmission line.

8. A method for detecting defects of a transmission line anti-vibration hammer according to claim 7, It is characterized in that The prediction box loss module of the Slideloss sample weighted function is specifically composed of the Slideloss sample weighted function, the CIoU loss function and the distribution focus loss DFL.

9. A method for detecting defects of a transmission line anti-vibration hammer according to claim 8, It is characterized in that The anti-vibration hammer defect detection model is trained based on the defect image dataset, including: The divided transmission line anti-vibration hammer image training set is input into the anti-vibration hammer defect detection model for pre-training. A total of 200 rounds of training are conducted. The training is divided into two stages: freezing and thawing. In the first stage, the backbone of the MFS-YOLO network model is frozen and the Mosaic and Mixup data enhancement methods are used cross-iterated for 50 times. In the second stage, the backbone of the MFS-YOLO network model is thawed and iteratively trained for 150 rounds. After the training is completed, the training weight with the smallest loss value is taken, which is the optimal training weight of the anti-vibration hammer defect detection model.

10. A method for detecting defects of a transmission line anti-vibration hammer according to claim 9, It is characterized in that The trained anti-vibration hammer defect detection model is used to detect the image to be detected and obtain the detection results, including: The optimal training weights are input into the anti-vibration hammer defect detection model for defect detection. When detecting the anti-vibration hammer image defects of the transmission line, the intersection over union (IOU) threshold of the prediction box is set to 0.5, and the confidence threshold for determining whether it is a defect target is set to 0.

01. The redundant prediction boxes are eliminated using a non-maximum suppression (NMS) method to obtain the transmission line anti-vibration hammer defect detection results.

Citation Information

Cited By

  • PCB small target defect detection method based on improved YOLOv8x

    CN119722669A