Defect Detection In Carbon Fiber Composite Core Conductors Based On Convolutional And Transformer Models

Carbon fiber composite core conductors are a widely used material in power transmission circuits,offering lighter weight and higher tensile strength compared to traditional power transmission materials.China has a vast land area,with a total power line length exceeding 1.6 million kilometers.Carbon fiber composite core conductors,due to their longer transmission distances and lower energy consumption,have broad application prospects in China.However,due to their relatively poor resistance to bending,carbon fiber composite core conductors are particularly susceptible to external compression damage during production and transportation.Especially during transportation and installation,the carbon fiber composite core conductors are long and lightweight,which can easily cause core rod damage or even serious issues such as conductor breakage.These problems can affect the normal power supply of transmission lines and increase the risk factor of power transmission.Therefore,it is crucial to promptly detect and address defects in carbon fiber composite core conductors.To ensure the safety of power supply in transmission lines,this thesis proposes two different conductor defect detection algorithms.These two algorithms can detect and determine the damaged areas and damage types of core rods in X-ray images of carbon fiber conductors.The main research content of thisthesis is as follows:1)Based on the fully convolutional network,We propose an improved YOLOv5 wire defect detection algorithm.This model introduces a Polarized Self-Attention Mechanism(PSA)in YOLOv5,which allows the network model to focus more on foreground targets and weaken attention to the background,thus reducing the negative impact of complex backgrounds on detection performance.To optimize the feature extraction of wire defects at different scales,we designed an Adaptive Multi-Stage Multi-Scale Feature Fusion Module,which reduces the loss of features at different scales and improves defect detection accuracy.Meanwhile,we adopted an area-based regularization loss,which improves the detection accuracy of small targets in wire defect detection by adding an Area Ratio Constraint(ABIOU loss).Through experimental comparative analysis,the performance indicators of the Fully Convolutional Network-based model designed in this paper have greatly improved compared to the original model.2)Based on the fusion model network,in order to solve the low global sensitivity of the fully convolutional network and improve the accuracy of target detection in large areas,this thesis designs a parallel cross-encoder-decoder detection network.The Transformer model is combined with the convolutional model to simultaneously extract features from the input image in parallel and interactively exchange feature information during the extraction process,enhancing the network’s perception of global and local features.At the same time,the calculation process of the Transformer model is optimized to group the input feature dimensions and remove redundant features using channel attention mechanisms.Then,the feature maps are divided into horizontal and vertical feature blocks for cross-fusion of features.A branch detector is also designed to improve the localization accuracy of defect regions and the classification accuracy of defect categories.Through experimental testing,the network designed in this thesis achieves a high localization accuracy of 83.5% for core rod damage areas,a classification accuracy of 92.6% for damage categories,and an overall detection accuracy of 92.0%,demonstrating satisfactory detection results.