基于Point-Attention点云分类的激光雷达故障诊断方法研究

打开文本图片集

中图分类号：TN958.98-34 文献标识码：A 文章编号：1004-373X(2025)20-0010-08

MethodofLiDARfaultdiagnosisbasedonPoint-Attention pointcloudclassification

TAN Guangxing, CHENG Xing, CHEN Haifeng (School of Automation,Guangxi Universityof Science and Technology,Liuzhou 545ooo,China)

Abstract:Inthefieldof intellgentvehiclesandautonomousrobots,LiDARsensorsarewidelyusedforenvironmental perceptionandobjectdetectionduetotheirhighprecisionandreliability.Therefore,faultdiagnosisofLiDARisparticularly important.InternalfaultsofLiDARoftencomewithfirmwarealerts,butfaultdetectioncausedbyexternalenvironmentalfactors presentsgreater challenges.For example,occlusion faultsinLiDARpoint cloudscaused byvehicle deformationordirtare difcult torflectdirectlyatthefirmwarelevelandcanonlybediagnosedbymeansofexternaldetection.Onthisbasis,a methodofPoint-Atention-basedLiDARocclusionfaultdiagnosisisproposed.Themodel'sabilitytoextractkeyfeaturesfrom pointclouddataisenhancedbycombiningwithmulti-head geometricatentionmechanismmodule,CBAMmoduleandresidual conectionmechanism,therebyimproving clasficationaccuracyandrobustness.Onthereal ScanObjectNNdatasetandthe ModelNet4Obenchmarkdataset，theexperimentalverificationofthePoint-Attentionmodelisconducted.Itcanrealize classification accuracy of 93.7% and 82.5% ,respectively.The time feature capture mechanism is fused to make model better adapttothetemporalcorelationinrealscenarios,therebyhandlingLiDARocclusionfaultsmoreacurately.Theexperimental resultsdemonstrate thattheproposedmethodcanefectivelydiagnoseLiDARoccusion faults，andbestoverallaccuracycan reach over 99% ，providing an eficient and accurate solution forLiDAR fault diagnosis.

Keywords:LiDAR;fault diagnosis;pointcloud classfication;residualconnection;occlusiondetection;time featurecapture

0 引言

在智能车辆和自主机器人领域，激光雷达(LiDAR)

传感器因高精度和可靠性，被广泛应用于环境感知和物体检测。（剩余16969字）