摘要
电机磁瓦是一种主要用在永磁电机上的瓦状磁铁,其表面缺陷检测是确保产品质量和电机性能的一个重要环节。因此,基于标准化流和卷积自编码器,构建了两阶段表面缺陷检测方法,同时采用联邦学习实现了在保护电机缺陷数据隐私的同时,对电机磁瓦表面缺陷进行高效检测。首先,利用预训练的编码器提取特征,将其作为标准化流分类网络的输入,并共享给解码器部分;然后,利用重构误差实现缺陷分割,利用联邦学习实现对数据的隐私保护。相比于有监督学习对缺陷样本的依赖,该方法在训练时只需正常样本,实现了无监督检测,摆脱了对缺陷样本的依赖。最后,在电机磁瓦表面缺陷数据集上验证了该方法的优越性。
The motor magnetic tile is a tile-shaped magnet used in permanent magnet motors,and its surface defect detection is crucial for ensuring product quality and motor performance.Therefore,a two-stage surface defect detection method based on normalizing flow and convolutional autoencoder is constructed.Additionally,federated learning is employed to achieve efficient detection of motor magnetic tile surface defects while ensuring privacy protection for motor defect data.Firstly,features are extracted by using a pre-trained encoder,then input into the normalizing flow classification network and shared with the decoder.Subsequently,defect segmentation is performed by using reconstruction errors,and data privacy is protected by federated learning.Compared with the dependence of supervised learning on defective samples,the method only needs normal samples during training,realizing unsupervised detection and getting rid of the dependence on defective samples.Finally,the superiority of the proposed method is demonstrated by validation on a surface defect dataset of motor magnetic tile.
作者
朱文鹏
郭峰
平作为
梁英杰
兰儒恺
张永
ZHU Wenpeng;GUO Feng;PING Zuowei;LIANG Yingjie;LAN Rukai;ZHANG Yong(School of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China;National Key Laboratory of Science and Technology on Vessel Integrated Power System,Naval University of Engineering,Wuhan 430205,China)
出处
《控制工程》
CSCD
北大核心
2023年第7期1219-1225,共7页
Control Engineering of China
基金
国家自然科学基金资助项目(61873197,61701517)
关键词
缺陷检测
电机磁瓦
联邦学习
无监督
标准化流
Defect detection
motor magnetic tile
federated learning
unsupervised
normalizing flow
作者简介
朱文鹏(1998-),男,湖北黄冈人,研究生,主要研究方向为工业产品的表面缺陷检测等;通信作者:梁英杰(1985-),男,河北邢台人,博士,副教授,主要从事故障诊断与大数据分析等方面的教学与科研工作(Email:lian142536@126.com)。
