| With the wide application of transmission line monitoring device,more and more attention has been paid to the monitoring device’s detection performance of potential environmental hazards around transmission lines.However,at present,in the test of AI capability of transmission line monitoring device,complete test system is lacking,and the test efficiency is low.And the original evaluation index is difficult to directly and effectively analyze the test results;At the same time,the construction process of test samples is completely manual,which not only consumes a lot of manpower,but also may result in mismarking and missing marking of samples due to manual mistakes,resulting in inaccurate test results of AI capability of monitoring device.To solve the above problems,it is of great significance to design a monitoring device test system which can realize the ability of constructing test samples and automatically testing AI.As the monitoring device belongs to embedded device,the evaluation index of AI ability and the overall design scheme of AI ability test system are proposed by analyzing the test method,test process and test sample design process of embedded system.The test system mainly includes the construction of test samples and the automatic cross-testing process.The construction process of test samples is completed by the automatic labeling,grading and manual correction of test samples.Automatic cross test is realized through the coordination of target detection capability test unit,current acquisition unit and comprehensive processing unit.In order to meet the need of automatic labeling of test samples,an improved Center Net object detection model was proposed to label test samples.The main network of the original Center Net model has few layers,insufficient feature extraction,and inaccurate predicted locations.The network with ECA Efficient Net-B0 is first designed for feature extraction,and the AF-Bi FPN module is proposed for feature fusion to enrich the feature information of hidden targets.Then MIOU loss is introduced to jointly predict the location and size of the target center point,reduce the location deviation of the target detection process,and then realize the automatic labeling of test samples.In order to further qualitatively judge the AI ability of the monitoring device,the test samples were graded according to the completeness of the external characteristics of the hidden danger target,and then the AI ability level of the monitoring device was evaluated.The classification of test samples is realized by an improved YOLOv4 target detection model.Aiming at the problems of large scale difference of hidden targets and insufficient feature extraction of small targets by YOLOv4,a multi-scale void convolution module was designed,CBAM attention mechanism and FRM feature refinement module were introduced to improve the detection ability of the model for targets at different scales,and then the classification of test samples was realized.By analyzing the construction process of test samples and the principle of automated cross testing,a test sample construction tool was designed,and AI capability testing software was developed.In addition,AI capability testing system was built by combining industrial computer,current monitoring equipment,switch and other hardware equipment,and AI capability testing of monitoring device was realized.The AI capability test system of transmission line monitoring device has been put into use in Hebei Electric Power Research Institute of State Grid.This system can meet the requirements of the construction of test samples and the automatic testing of AI capability,greatly reducing labor consumption,saving test time and effectively testing the AI capability of monitoring device. |
PDF Full Text Request