基于数据生成与迁移学习的轴承小样本故障诊断  

Bearing Small Sample Fault Diagnosis based on Data Generation and Transfer Learning

作　　者：吴定会 方钦 吴楚宜 Wu Dinghui;Fang Qin;Wu Chuyi(Engineering Research Center of Internet of Things Application Technology,Ministry of Education,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学物联网应用技术教育部工程研究中心,江苏无锡214122

出　　处：《机械传动》2020年第11期139-144,共6页Journal of Mechanical Transmission

摘　　要：针对风电机轴承历史运行数据来源单一、数据量少,导致风电机轴承故障诊断性能受限问题,提出一种基于数据生成与迁移学习的轴承小样本故障诊断方法。首先,对于轴承数据集中存在类不平衡、数据稀缺的问题,提出一种基于门限机制的数据生成方法,采用与轴承驱动端同轴的桨叶端数据为模板产生足量的生成数据,结合真实数据作为源数据集;然后,根据数据的时序关联性和小样本的应用场景,提出一种基于一维卷积神经网络(One Dimensional Convolutional Neural Network,1DCNN)和双向门限单元(Bidirectional Gated Recurrent Unit,BiGRU)的迁移学习(Transfer Learning)方法,先用源数据集在训练网络上训练获得源模型,再用少量驱动端数据作为目标数据集对其进行微调(Fine-tuning)获得目标模型;最后,对目标模型全连接层的输出采用Softmax函数进行故障诊断。实验表明,提出的故障检测方法在目标集小样本数据的场景下平均精度达到99.67%,分类效果明显,泛化能力强。Aiming at the problem of limited fault diagnosis performance caused by a single source of historical operating data for wind turbine bearings and a small amount of data,a small sample fault diagnosis method of bearings based on data generation and transfer learning is proposed.First of all,for the problems of class imbalance and data scarcity in the wind turbine bearing dataset,a data generation method based on a gate mechanism is proposed.The blade-end data coaxial with the drive-end of the bearing is used as a template to generate a sufficient amount of generated data.The data is used as the source dataset.Then,according to the time series correlation of the data and the small sample application scenario,a transfer learning method based on one-dimensional convolutional neural network(1DCNN)and bidirectional gated recurrent unit(BiGRU)is proposed.First,the source dataset is trained on the training network to obtain the source model,and then a small amount of drive-end data is used as the target dataset to fine-tuning to obtain the target model.Finally,the Softmax function is used for fault diagnosis on the output of the fully-connected layer of the target model.Experiments show that the proposed fault detection method has an average accuracy of 99.67%in the scenario of small sample data of the target dataset,so the obvious classification effect and strong generalization ability could be seen.

关 键 词：风电机轴承 小样本 数据生成 门限机制 迁移学习 微调 故障诊断 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TM315[自动化与计算机技术—控制科学与工程]