采用深度迁移学习与自适应加权的滚动轴承故障诊断

Fault Diagnosis of Rolling Bearings Using Deep Transfer Learning and Adaptive Weighting

针对滚动轴承目标域数据中额外故障状态样本影响其故障诊断精度的问题,提出了采用深度迁移学习与自适应加权的滚动轴承故障诊断方法。建立特征提取模块,利用深度卷积神经网络将轴承样本映射到高维特征空间;利用迁移学习思想设计加权领域鉴别器,对样本进行自适应加权,并通过在特征空间的对抗训练,增大目标域与源域共有健康状态样本的领域相似性,抑制目标域额外故障状态样本与源域样本的领域相似性增强;依据样本权重,度量目标域与源域样本的相似性,设定阈值将目标域额外故障状态样本标记为未知故障;将源域故障诊断知识迁移到目标域共有健康状态样本的故障识别中。利用齿轮箱轴承数据、凯斯西储大学滚动轴承数据和机车轮对轴承数据对提出的方法进行验证,结果表明:所提方法在3个数据集上均达到89%以上的诊断精度,而对比方法的诊断精度均低于80%。所提方法能够克服额外故障状态样本的影响,有效实现滚动轴承故障诊断。

This paper proposes a method for fault diagnosis of rolling bearings using deep transfer learning and adaptive weighting considering the fact that the samples in the extra fault category of target domain dataset affects the fault diagnosis accuracy of rolling bearings.Firstly,a feature extraction module is established to map the bearing samples into high-dimensional feature space with deep convolution neural network.Secondly,a weighted domain discriminator is designed based on transfer learning and the samples are adaptively weighted.Through the adversarial training in the feature space,the domain similarity of the samples in the same health status is increased between the target and source domains,and the domain similarity of the samples in extra fault category are suppressed.Then,a threshold is set according to the sample weights and domain similarity so as to mark the samples in extra fault category as those with unknown faults.Finally,the fault diagnosis knowledge of source domain is transferred to the fault identification of samples in the same health status in target domain.The proposed method is verified with the gearbox bearing dataset bearing dataset of Case Western Reserve University and locomotive bearing dataset.The results show that the diagnosis accuracy of the proposed method is 89% higher for the three datasets,while that of the method compared is lower than 80%.This indicates that the proposed method can eliminate the influence of extra fault category and realize effective diagnosis of rolling bearing faults.