基于数据增强与领域泛化的轴承跨域故障诊断

Cross-Domain Fault Diagnosis of Bearing Based on Data Augmentation and Domain Generalization

在实际故障诊断任务中,待诊断任务往往不可预知,现有的一些迁移学习方法在构建迁移模型时,大多只集中在单一数据来源的学习上,并且极大依赖于目标域数据的样本数量等。针对此问题,提出一种基于数据增强与领域泛化的故障诊断方法。提出一种将一维振动信号转换为二维特征指标灰度图的数据预处理方法;利用带有梯度惩罚的深度条件Wasserstein对抗网络对多源域数据进行数据增强;最后,采取多域对抗学习策略,缩小多域间的分布差异,从而实现各域的特征域自适应。在轴承数据集上对所提方法的有效性和可靠性进行了充分的实验验证。实验结果表明:所提方法具有较高的稳定性和泛化性能,并且诊断精度优于其他方法。

In the practical fault diagnosis tasks,the target task is often unknown in advance,existing transfer learning methods mostly focus on learning from a single data source when constructing transfer models,they heavily rely on the quantity of samples in the target domain.In view of this problem,a fault diagnosis method was proposed based on data augmentation and domain generalization.A data preprocessing method was introduced to transform 1D vibration signals into a 2D feature indicator grayscale image.A deep conditional Wasserstein generative adversarial network with gradient penalty was proposed to augment the data from multiple source domains.Finally,a multi-source domain adversarial learning strategy was adopted to reduce the distribution differences among the multiple source domains,achieving feature domain adaptation for each source domain.The effectiveness and reliability of the proposed method were thoroughly validated on a bearing dataset.Experimental results demonstrate the proposed method has high stability and generalization performance,and better diagnostic accuracy than other methods.