基于迁移学习的电气元件识别方法研究李殿明，于正林（长春理工大学机电工程学院，长春130022）摘要：电气元件识别是电控柜自动布线机的关键一步。由于采用普通深度学习训练小样本集可能出现过拟合，所以将VGG16、ResNet50模型迁移到电气元件识别领域，再利用网络手术技术分别将基于VGG16、ResNet50迁移模型的分类器替换成极限学习机。普通深度学习模型VGG16、ResNet50识别准确率分别为68.27%和72.94%；迁移学习模型识别准确率对比普通深度学习分别调高了18.89%、17.28%；使用迁移学习和ELM结合方法，准确率比对应迁移模型分别提高了5.56%、7.16%。结果表明，迁移学习更适合小样本电气元件的识别，并且迁移学习和ELM相结合的方法可进一步提高准确率。关键词：电气元件；识别；普通深度学习；迁移学习；极限学习机中图分类号：TP391.4文献标志码：A文章编号：1672-9870（2023）01-0073-08ResearchonElectricalComponentIdentificationMethodBasedonTransferLearningLIDianming，YUZhenglin（SchoolofMechatronicEngineering，ChangchunUniversityofScienceandTechnology，Changchun130022）Abstract：Identificationofelectricalcomponentsisthekeystepofautomaticwiringmachineforelectriccontrolcabinet.Sincethesmallsamplesettrainedbyordinarydeeplearningmayoverfit，theVGG16andResNet50modelsaretransferredtothefieldofelectricalcomponentidentification.Usingnetworksurgerytechnology，theclassifierbasedonVGG16andResNet50migrationmodelwasreplacedwithextremelearningmachine.TherecognitionaccuracyofcommondeeplearningmodelsVGG16andResNet50is68.27%and72.94%respectively.Comparedwithcommondeeplearning，therecognitionaccuracyoftransferlearningmodelwasincreasedby18.89%and17.28%respectively.Comparedwiththecorrespondingmigrationmodel，theaccuracyofthecombinedELMandtransferlearningmethodisimprovedby5.56%and7.16%respec-tively.Theresultsshowthattransferlearningismoresuitablefortheidentificationofsmallsampleelectricalcomponents，andthemethodcombinedwithtransferlearningandELMcanfurtherimprovetheaccuracy.Keywords：electricalcomponents；identification；generaldeeplearning；transferlearning；extremelearningmachine智能制造是“中国制造2025”的主攻方向，也是推动制造业向智能化、数字化...
