Research On Inspection Method Of Substation Panel And Cabinet Based On Dynamic Machine Vision

In today’s information society,people are increasingly dependent on electricity,and the electricity demand is also growing.The substation is a crucial component of the electrical system that can efficiently transfer voltage and current to complete power dispatching more quickly.It also plays a crucial role in the safety of the electrical grid.Staff performs the patrol inspection of conventional substations,which is inefficient,labor-intensive,has a high error rate.The maintenance and operation staff cannot handle the tremendous workload due to the growing number of substations.Robotic automatic inspection systems are suggested as a potential solution to these issues.These inspection robots are equipped with location and path finding systems and equipment status recognition algorithms.However,most of the current equipment recognition algorithms use traditional image recognition technology,which has limited recognition effect on complex environment and multitype equipment status detection.This thesis presents an inspection method of transformer substation panel cabinet based on machine vision.The Automated Guided Vehicle(AGV)equipped with Simultaneous Localization and Mapping(SLAM)laser navigation technology and high-definition camera are used to collect the equipment switch status,and the target detection algorithm based on depth learning is studied to complete the status recognition task.The main work is as follows:(1)The YOLOv4 target detection algorithm is improved.According to the small target characteristics of substation switch,three improved strategies are proposed for YOLOv4 algorithm: decoupling detection head based on Anchor free,empty space pyramid module,and attention feature fusion.The decoupling detection head based on Anchor free cancels the anchor design of YOLO,and uses the characteristic graph grid as the anchor,which has better detection effect for the target with a ratio of close to 1:1 such as the substation switch.The detection head is decoupled,which greatly improves the detection accuracy.The void space pyramid module introduces void convolution based on SPP,which can effectively expand the receptive field and capture multi-scale context information.Attention feature fusion solves the problem of fusing different scale features through multi-scale attention module,making feature fusion obtain better results.Finally,the effectiveness of the improved YOLOv4 algorithm is verified by comparing the improved YOLOv4 algorithm with the original algorithm.(2)The identification of the switch state of the substation has been completed.Di fferent switches are identified with different recognition algorithms,part of the switch state recognition is converted into a classification task,a classification data set is mad e,and image classification is completed through a convolutional neural network,whil e other switches are judged by traditional image processing algorithms.The recognitio n algorithms can obtain better recognition results,and solve the problem of state recog nition of different switches.(3)Carried out the patrol inspection of the substation.The inspection process is d esigned,and the detection and status identification of the substation switch is complet ed through the pictures collected by the SLAM inspection robot.Finally,the identifica tion result is compared with the preset switch status,and the abnormal identification a nd diagnosis of the switch status of the substation equipment are realized.