Research On Defects Detection Of Insulators Based On Improved YOLOv5s

With the development of China’s large power grid,transmission lines are increasing,especially the vigorous construction of extra-high voltage lines,which puts forward higher requirements for the safety of power lines.Insulators,as important equipment of the power system,are prone to pollution and aging due to long-term exposure during actual use,resulting in defects and failures of insulators,which in turn cause power system failures.Therefore,regular inspection of the power system is necessary to ensure the safe operation of the power system,but manual inspection is time-consuming,labor-intensive,dangerous and expensive,so the use of unmanned aerial vehicles for power line inspection has been the inevitable trend of power development.This paper combines the research related to UAV insulator inspection,makes full use of deep learning target detection algorithm,constructs a deep learning insulator fault recognition model,proposes an aerial insulator burst fault recognition algorithm for power system based on improved YOLOv5 s algorithm,and verifies the recognition accuracy and application feasibility of the algorithm.The experiments prove that it has good effect on the identification of missing faults of extra-high voltage insulators.Firstly,this paper describes the basic process of two-stage deep learning algorithms R-CNN,Faster R-CNN,Mask R-CNN based on the analysis of the basic composition and classification of deep learning algorithms.According to the actual insulator state detection of aerial photographs,the basic composition and basic principle of YOLOv5 algorithm are elaborated,as well as its shortcomings in practical applications.Secondly,according to the problems of cluttered background,unclear target and small target of UAV shooting samples,the Neck structure of the algorithm is optimized from the analysis of YOLOv5 structure,the attention mechanism is introduced to strengthen the key features,and the characteristic balance factor of the loss function is analyzed from the perspective of optimizing the IOU loss function,the levy factor is introduced,the balance factor of The calculation is fully optimized to improve the convergence of the loss function,and the experimental analysis method is applied to analyze the accuracy and feasibility of the algorithm model in terms of model recall,model average accuracy m AP,and loss function value.Finally,the main process based on the improved YOLOv5 s insulator missing fault algorithm is elaborated,and the construction of the insulator sample library is highlighted in terms of sample naming acquisition,sample labeling,and sample proportion adjustment based on the lack of realistic insulator datasets and other problems,and the correctness of the dataset is verified by applying the Python dataset testing program.On this basis,in order to strengthen the robustness of the relevant parameters of the model,the model training process and methods are studied in depth,and the Mosaic data enhancement method,mixed accuracy training and adaptive anchor frame calculation method are used to train the improved YOLOv5 s insulator defects detection model,and after obtaining the relevant model parameters,the rotation test,tensile test and other methods are applied to the The test results show that the improved YOLOv5 s insulator defects detection algorithm model has a high recognition accuracy in the actual detection application,so it has a good practical application feasibility.The three-point improvement mechanism for YOLOv5 s in the study can better achieve the expected convergence during model training with good global performance.The improved YOLOv5 s has a 39.3% improvement in the average model accuracy m AP value compared to the traditional Faster R-CNN;compared to the traditional Mask R-CNN,the improved YOLOv5 s has a26.3% improvement in m AP value compared to the conventional Mask R-CNN;while the improved YOLOv5 s has a 13.8% improvement in m AP value compared to the original unimproved YOLOv5 s and a 12.6% improvement in m AP value compared to the original YOLOv5 s with the CBAM attention model.Further,not only the original image detection,but also the improved YOLOv5 s can maintain a certain defect detection rate in the rotation test,stretch test and complex background test.