Design Of Foreign Object Detection Algorithm For Transmission Lines And Its Robustness Study

With the development of intelligent information technology,deep learning-based intelligent inspection equipment is widely used for the safety detection of transmission lines.However,deep learning network models have the disadvantage of being vulnerable to adversarial attacks,which can cause them to fail.Therefore,this thesis proposes an efficient and accurate transmission line foreign object detection algorithm based on deep learning,and conducts robustness research on it.To address the problem of low detection accuracy of foreign objects on transmission lines using the YOLOv5 algorithm,this thesis proposes a transmission line foreign object detection algorithm based on MS-YOLOv5.The CSP-Dark Net53 backbone network is replaced with an improved lightweight Mobilenetv2 structure to reduce the computation of training the network and improve detection speed.The SENet attention mechanism model is integrated,and deeper downsampling is added to the feature extraction module to improve the recognition accuracy of small targets in transmission line foreign objects.Experimental results show that the proposed MS-YOLOv5 algorithm effectively improves the accuracy of foreign object recognition on transmission lines.In response to the problem of low efficiency and non-robust attack effects of existing gradient-based target recognition attacks,this thesis proposes an HRs-PGD attack algorithm based on the hot restart mechanism.Combining with the hot restart mechanism,this algorithm introduces the cosine annealing idea into the PGD attack method to change the learning rate regularly in a cosine period during iteration,and the interval of each restart is increased progressively to make the learning rate gradually decrease and eventually converge to the optimum.Experimental results show that the HRs-PGD attack algorithm can make the iteration process converge faster.A self-supervised learning method based on multi-scale feature fusion is proposed to solve the overfitting problem of the foreign object detection algorithm for transmission lines.By building a transfer learning network model to extract features from the source domain and target domain to reduce the difference between the two domains,the method then transfers the label samples of the categories with larger distributions to those with smaller ones based on transfer learning.Finally,a transmission line foreign object detection system based on MS-YOLOv5 was built using Python and Qt software.The system includes data augmentation interface,adversarial sample generation interface,data transfer learning interface,and model training and testing interface.It can intuitively display the experimental test results of the transmission line foreign object detection algorithm based on MS-YOLOv5.