Investigation On Infrared Image Detection Method Of Power Equipment Based On Convolutional Neural Network

Safe and stable operation of power equipment is the key to ensuring reliable power supply from the power grid.By patrolling the equipment to detect the operating condition of the power equipment,it is possible to prevent accidents caused by equipment defects or malfunctions.Considering that infrared thermal imaging technology can provide a non-contact detection method to obtain the thermal status information of power equipment,the status detection of the power equipment can be performed under the condition of uninterrupted power.Such a technology is thus widely used in the live detection of power equipment.However,the current inspection and monitoring system lacks intelligent analysis of the acquired infrared image data of the equipment,and experienced power engineers are essential for analysis and further diagnosis.This inevitably consumes a lot of manpower and time costs and greatly reduces the efficiency of power equipment status detection and evaluation,thereby restricting the improvement of the intelligent operation and inspection level of power equipment.With the development of computer vision technology in recent years,the method of combining deep learning and image processing technology has become a study hotspot with huge development prospects in power equipment image detection.This paper focuses on investigating the detection methods of infrared images of power equipment in substations,combines the detection of different scenes in substations,and respectively proposes a high-precision detection method and a lightweight real-time detection method based on the improved YOLO algorithm.Its main works are as follows:(1)A dedicated database for infrared images of power equipment in substations is constructed.Collect the original infrared image of the power equipment of the power grid company’s substation,and preprocess the image.To meet the data-driven needs of deep learning,this paper further expands the database by performing image enhancement processing on the original image.Label Img is used to label the four power equipment labels of insulators,instrument transformers,breakers,and arresters.Besides,a dedicated database of infrared images of power equipment is created according to the standard format of the object detection data set in computer vision.Finally,K-means++ algorithm is used to cluster the database label data and select the anchor box required by the subsequent algorithm.(2)A high-precision power equipment detection method is designed and implemented.The detection performance of the original model is improved by improving the input terminal,backbone network,neck and border regression methods of the YOLOv3 model.Transfer learning is used to initialize the model weights,and then train and test on the PASCAL VOC dataset and the infrared image database of power equipment constructed in this paper,and compare and evaluate with the three mainstream methods of Faster R-CNN,SSD and YOLOv3.The experimental results prove that the proposed method greatly improves the detection accuracy while sacrificing a certain detection speed;when the Io U threshold is 0.5,the proposed method has a m AP value of 96.0% tested in the infrared database constructed in this paper,which is 1.4%,3.0%,and 3.6% higher than Faster R-CNN,SSD,and YOLOv3,respectively;when the Io U threshold is0.75,the m AP value of the proposed method is 89.6%,which is 6.4% 1.7%,and4.5% higher than Faster R-CNN,SSD,and YOLOv3,respectively;in addition,the test results on the PASCAL VOC dataset also further demonstrate the advantages of the proposed method.(3)An optimized lightweight infrared image detection method for power equipment is investigated and implemented.The proposed method is a new lightweight model generated by optimizing the backbone network and location loss function of YOLOv3-tiny.Experiment with the PASCAL VOC dataset and the infrared image database of power equipment constructed in this paper and compare with YOLOv3.The result proves that in the infrared database of power equipment,the memory occupied by the improved YOLOv3-tiny is only 22 MB,which is 11 MB lower than that of YOLOv3-tiny.Compared with the three mainstream algorithms of Faster R-CNN,SSD,and YOLOv3,the improved YOLOv3-tiny has an absolute advantage in terms of memory usage;in addition,without reducing the detection speed,the detection accuracy of the improved YOLOv3-tiny has been significantly improved.In summary,for the power equipment detection of different scenes in substations,this paper studies a high-precision detection method and a lightweight detection method respectively.Among them,the proposed improved YOLOv3 model can accurately locate power equipment at a faster speed,but the network structure of the model is more complex,so it is suitable for non-mobile terminal detection systems in substations(such as infrared cameras,etc.);the proposed YOLOv3-tiny optimization model has a very small memory footprint and the fastest detection speed,which is suitable for mobile-end detection systems for power equipment(such as drones,inspection robots,etc.).