深度多模态迁移学习在轴承故障诊断中的研究

Research on Deep Multimodal Transfer Learning in Bearing Fault Diagnosis

在工程实际中,设备数据样本往往以正常数据居多。故障样本稀缺且模态单一使得可用于训练的故障信息特征提取不足,同时训练和测试数据分布往往存在差异,从而导致模型迁移诊断能力较弱。针对该问题,提出将深度学习模型CNN与多模态融合迁移学习技术相结合(Deep Multimodal Fusion Transfer Learning,DMFTL)应用于轴承的故障诊断中。首先以CNN为基本学习框架,将原始一维振动信号的时域和频域进行多模态信息融合对模型预训练;然后以最大均值差异(MMD)为度量准则,通过域自适应来最小化源域和目标域的差异;最后引入构造的正则项到模型中,以完成跨域诊断。通过对CWRU轴承数据集的迁移诊断试验及对比分析,验证了该方法的有效性和优越性。

In engineering practice,the equipment data samples tend to be mostly normal data.The scarcity of fault samples and single modal make the fault information feature extraction that can be used for training insufficient.At the same time,there are often differences in the distribution of training and test data,which leads to weak model migration and diagnosis capabilities.In response to this problem,it is proposed to combine the deep learning model CNN with multimodal fusion transfer learning technol-ogy(Deep Multimodal Fusion Transfer Learning,DMFTL)to apply to the fault diagnosis of bearings.Firstly,using CNN as the basic learning framework,the time domain and frequency domain of the original one-dimensional vibration signal are fused with multi-modal information to pre-train the model.Then,using the maximum mean difference(MMD)as the measurement criterion,the difference between the source domain and the target domain is minimized through domain adaptation.Finally,the constructed regular items are introduced into the model to complete cross domain diagnosis.Through the transfer diagnosis test and comparative analysis of CWRU bearing data set,the effectiveness and superiority of the method for rolling bearing fault di-agnosis are verified.