基于多源域深度迁移学习的机械故障诊断

Mechanical fault diagnosis based on multi-source domain deep transfer learning

针对不同工况下的机械故障诊断问题,迁移学习方法相比于深度学习具有明显的成效,单源域迁移故障诊断仍会出现负迁移和模型泛化能力差的问题。因此,本文提出一种基于多源域深度迁移学习的机械故障诊断方法。首先,进行锚适配器的构建,获得多源域-目标域适配器数据对。其次,建立基于深度域适应的迁移学习网络模型获得每个数据对的分类器与预测结果。最后,采用加权集成的方式进行分类器集成,用于最终的故障诊断识别。所提方法充分集成多源域故障特征信息,提取域不变特征,避免负迁移的问题,提高模型的泛化能力。通过一个滚动轴承数据来验证提出方法的性能,结果表明,多工况迁移故障诊断分类精度明显高于其中任意单一工况迁移,最高可提高8.78%,与其他方法相比,所提方法具有较好的精度和泛化能力。

For mechanical fault diagnosis under different working conditions,the transfer learning method has obvious effect compared with deep learning,but the single source domain transfer fault diagnosis still has problems of negative transfer and poor model generalization ability.Here,a mechanical fault diagnosis method based on multi-source domain deep transfer learning was proposed.Firstly,an anchor adapter was constructed to obtain the multi-source domain-target domain adapter data pairs.Secondly,a transfer learning network model based on deep domain adaptation was established to obtain each data pair’s classifier and prediction results.Finally,classifiers were integrated by using the weighted integration for the final fault diagnosis and recognition.It was shown that the proposed method can fully integrate multi-source domain fault feature information,extract domain invariant features,avoid the problem of negative transfer,and improve the model generalization ability;the performance of the proposed method is verified with a rolling bearing data;the fault diagnosis classification accuracy of multi-condition transfer is obviously higher than that of any single condition transfer,the maximum improvement can be increased by 8.78%;compared with other methods,the proposed method has better accuracy and generalization ability.