《自动化技术与应用》2024年第43卷第1期控制理论与应用ControlTheoryandApplicationsTechniquesofAutomation&Applications卷积神经网络的神华关键配件状态自动跟踪研究*卓卉（国家能源投资集团有限责任公司，北京100120）摘要:为避免关键配件异常状态带来的神华铁路货车运行安全隐患，提出基于卷积神经网络的神华关键配件状态自动跟踪方法。神华关键配件状态图像作为卷积神经网络的输入数据，经过卷积和池化操作后获得神华关键配件状态检测结果，并将其作为图像第一帧的状态，然后利用核相关滤波训练获得的回归模型估计图像下一帧的状态，实现神华关键配件状态自动跟踪。实验结果表明：该方法能够获得较为完整、清晰的神华关键配件状态图像；不同神华关键配件状态检测的MCC值均在0.8以上，且能够在异常状态发生之前得到状态检测结果；各时刻的神华关键配件状态跟踪结果与实际结果完全相同。关键词:卷积神经网络；神华关键配件；状态自动跟踪；CCD相机；核相关滤波；回归模型中图分类号：TP183文献标识码:A文章编号:1003-7241(2024)01-0071-04ResearchonAutomaticStateTrackingofShenhuaKeyPartsBasedonConvolutionalNeuralNetworkZHOUHui(NationalEnergyInvestmentCorporationLimited,Beijing100120China)Abstract:InordertoavoidthehiddendangerofShenhuarailwayfreighttrainoperationcausedbytheabnormalstateofkeyparts,theauto-maticstatetrackingmethodofShenhuakeypartsbasedonconvolutionalneuralnetworkisstudied.TheCCDcameraisusedtocollectthestateimageofShenhuakeyaccessories,whichisusedastheinputdataofconvolutionneuralnetwork.Afterconvolu-tionandpooling,thestatedetectionresultofShenhuakeyaccessoriesisobtained,anditisusedasthestateofthefirstframeoftheimage.Onthisbasis,theregressionmodelobtainedbynuclearcorrelationfiltertrainingisusedtoestimatethestateofthenextframeoftheimage,realizetheautomatictrackingofthestatusofShenhuakeyaccessories.TheexperimentalresultsshowthatthismethodcanobtainarelativelycompleteandclearstateimageofShenhuakeyparts.TheMCCvaluesofdifferentShen-huakeyaccessoriesareabove0.8,andthestatedetectionresultscanbeobtainedbeforetheabnormalstateoccurs.ThestatustrackingresultsofShenhuakeyaccessoriesateachtimeareexactlythesameastheactualresults.Keywords:convo...
