摘要
直线电机进给系统因长期在复杂工况环境下服役导致性能及状态转移随机性更强,对其运行状态进行健康诊断,对于及时发现状态退化趋势并进行预防性维护具有重要意义。针对直线电机进给系统健康诊断中缺乏故障负样本且运行数据时序性强等问题,在分析半监督异常检测生成对抗网络(GANomaly)与门控循环单元(Gated Recurrent Unit,GRU)网络原理的基础上,提出一种新的基于GANomaly-GRU的直线电机进给系统健康诊断方法,该方法在GANomaly网络框架基础上引入GRU设计生成器、判别器和重构编码器,以避免网络出现梯度消失和梯度爆炸等问题,第四子网络采用异常样本概率模型和解码重构(Decoding Reconstruction,DR)评分在无负样本的情况下对直线电机进给系统运行状态进行健康诊断。最后在直线电机进给系统健康诊断实验平台上测试和验证该方法的准确性和快速性,并与经典的诊断方法,如异常检测生成对抗网络(Anomaly Generative Adversarial Networks,AnoGAN)、高效异常检测生成对抗网络(Efficient GAN-Based Anomaly Detection,EGBAD)和无监督深度卷积生成对抗网络(Deep Convolutional Generative Adversarial Networks,DCGAN)进行对比,结果表明所提方法平均准确率分别提高了13.1%、24.0%和15.2%,诊断时间分别缩短了37、26和13 ms,表明所提方法较好地解决了直线电机进给系统故障样本数据不足和样本标记问题,对难以获取足量有效故障样本数据的复杂机械装备健康诊断具有一定的参考价值。
Since the linear motor feed system has been in service for a long time under complex working conditions,its performance and state transition are more random.It is of great significance to perform health diagnosis on its operating state and timely identify state degradation trends and perform preventive maintenance.In view of the lack of faulty negative samples in the health diagnosis of the linear motor feed system and the strong timing of the running data,the semi-supervised anomaly detection GANomaly and the Gated Recurrent Unit(GRU)network principle were analyzed,a new linear motor feed system health diagnosis method was proposed based on GANomaly-GRU.The GRU was introduced to design generator,discriminator and reconstruction encoder based on GANomaly network framework to avoid gradient disappearance and gradient explosion in the network.An abnormal sample probability model and Decoding Reconstruction(DR)score were adopted in the fourth sub-network to perform health diagnosis on the operating state of the system without negative samples.Finally,the accuracy and rapidity efficiency of the method were tested and verified on the health diagnosis experimental platform of linear motor feed system,and compared with the classic diagnostic methods such as Anomaly Generative Adversarial Networks(AnoGAN),Efficient GAN-Based Anomaly Detection(EGBAD)and Deep Convolutional Generative Adversarial Networks(DCGAN).The results show that the average diagnostic accuracy of the proposed method is increased by 13.1%,24.0%and 15.2%respectively,and the diagnosis time is shortened by 37,26 and 13 ms respectively.The experimental results show that the proposed method can solve the problem of insufficient fault sample data and sample labeling of the linear motor feed system,which has certain reference value for the complex mechanical equipment health diagnosis based on insufficient fault sample data.
作者
杨泽青
李月
陈英姝
刘丽冰
魏强
YANG Zeqing;LI Yue;CHEN Yingshu;LIU Libing;WEI Qiang(School of Mechanical Engineering,Hebei University of Technology,Tianjin 300130,China;Research Institute for Structure Technology of Advanced Equipment,Hebei University of Technology,Tianjin 300401,China)
出处
《现代制造工程》
CSCD
北大核心
2021年第11期128-135,115,共9页
Modern Manufacturing Engineering
基金
国家自然科学基金项目(52175461)
河北省自然科学基金和重点基础研究专项项目(E2017202294)
河北省青年拔尖人才项目(210014)
天津市自然科学基金项目(16JCYBJC19100)。
