Self-exploding Target Detection Of UHV Insulators In Complex Backgrounds Based On Deep Learning

In recent years,the development of China’s ultra-high voltage(UHV)transmission lines has been very rapid in order to meet the needs of energy resource development and transformation.UHV transmission technology can achieve long-distance and large-capacity transmission,which helps to solve the energy supply-demand imbalance between regions.Insulators are critical components used in UHV transmission lines,which support and insulate the conductors.Since the voltage of UHV transmission lines can reach several hundred kilovolts,the safety performance of insulators is crucial.Once an insulator explodes,it will instantly lose its insulation effect on the electricity,causing current short-circuits,voltage fluctuations,and possible damage to the power equipment.In addition,the fragments generated during the explosion may be ejected,causing personnel injury.In recent years,aerial inspections using helicopters or drones have become important methods for inspecting transmission lines.This dissertation focuses on detecting insulator explosion defects in UHV transmission lines using images obtained through these methods.A new convolutional neural network model based on the YOLOv5 algorithm was proposed in this dissertation.By improving the network structure,the training speed,detection accuracy,and model size have been improved compared to the original network.The main contributions of this dissertation are as follows:(1)To address the problem that traditional algorithms have difficulty accurately detecting defects in aerial images of ultra-high voltage insulators in complex power system inspection environments due to their multi-scale characteristics and small defect proportions,a YOLOv5 algorithm based on attention mechanisms and lightweight Ghost Net is proposed.Firstly,to suppress complex background interference,a spatial and channel convolution attention model is introduced to enhance the saliency of the target to be detected.Secondly,the Ghost Net network is used to replace the Conv structure of the original CSP3 structure,achieving the effect of network lightweight and reducing network parameters.Finally,in order to solve the problem of insufficient feature expression capability of the target to be detected,which may cause missed detection and false detection,the FPN+PAN structure in the original Neck layer is replaced by the Bi FPN structure,which effectively fuses the target’s multi-scale features through cross-scale feature fusion,reducing the loss of feature information caused by excessive network layering and significantly improving the detection accuracy.(2)To address the slow convergence speed of the original YOLOv5 network and the problem of imbalanced sample categories,the loss function is improved.The GIo U loss function in the network is improved to the SIo U loss function,which is more robust to changes in the shape and size of the target box in the loss calculation of the real box and the predicted box,better adapting to targets of shapes and sizes,reducing the error between the predicted box and the real box,and using the Sigmoid function,which makes it more differentiable,making it more convenient to perform gradient back propagation when using SIo U as the loss function,thus optimizing network parameters.(3)To address this issue,data augmentation techniques and transfer learning strategies were used to experiment and validate the improved network structure.First,the researchers used data augmentation methods such as Mosaic to construct a dataset of high-voltage insulator disconnection and annotated the data samples using Labelme software.Second,during the network training phase,transfer learning techniques such as parameter sharing and frozen-thawed training models were used to alleviate the problem of network generalization caused by small sample datasets,while significantly reducing training time and computational resources.Finally,the improved algorithm was compared quantitatively with related target detection algorithms,and the significance of the network’s treatment of the detected targets before and after improvement was visually explained using Grad-CAM heat maps.Experimental results show that the improved algorithm has more universal applications and higher training efficiency,effectively improving target detection accuracy in complex backgrounds,and is suitable for a wide range of working environments.