一种面向旋转机械的基于 Transformer 特征提取的域自适应故障诊断

针对基于深度学习的旋转机械故障诊断方法在新工作条件下缺乏标注数据、跨域诊断精度较低的问题,提出了一种基 于 Transformer 的域自适应故障诊断方法。 采用 Transformer 的变体 VOLO 构造特征提取器以获取细粒度更佳的故障特征表示。 利用源域数据进行监督学习对源域和目标域数据的特征提取器进行预训练,并且冻结源域提取器参数以获取固定的源域特征。 利用域对抗自适应策略和局部最大平均差异结合目标域未标注数据训练目标域特征提取器,实现源域特征与目标域特征的边 缘分布、条件分布对齐。 通过两个多工况实验对所提出的故障诊断算法进行了验证,结果表明提出的基于 Transformer 特征提 取的域自适应故障诊断方法相比 5 种传统域自适应方法,在齿轮和轴承数据集上分别平均提升了 22. 15% 和 11. 67% 的诊断精 度,证明所提出方法对于跨域诊断精度具有提升作用。

Domain adaptive fault diagnosis based on Transformer feature extraction for rotating machinery

To address the problems of lack of labeled data and low cross-domain diagnosis accuracy in the fault diagnosis method of rotating machinery based on deep learning under new working conditions, a domain adaptive fault diagnosis method based on Transformer is proposed. A variant of Transformer, VOLO, is used to construct the feature extractor to obtain fine-grained and better fault feature representation. The supervised learning with source domain data pretrains feature extractors on source and target domain data, and freezes source domain extractor parameters to obtain fixed source domain features. Using domain adversarial adaptive strategy and local maximum mean difference combined with target domain unlabeled data to train target domain feature extractor, the edge distribution and conditional distribution of source domain features and target domain features are aligned. The proposed fault diagnosis algorithm is evaluated by two multi-condition experiments. Results show that the proposed domain adaptive fault diagnosis method based on Transformer feature extraction is more efficient than the five traditional domain adaptive methods on gear and bearing datasets. The average diagnostic accuracy is improved by 22. 15% and 11. 67% , respectively, which proves that the proposed method can improve the cross-domain diagnostic accuracy.