Deep transfer learning has been shown to be effective when performing intelligent fault diagnosis (IFD) because of its strong feature representation performance when characterizing vibration signals under variable working conditions. However, when target domain data is not available, the ability to train a model effectively could be very challenging since the feature distribution of the target domain does not contribute to the training procedure of the model. To deal with this scenario, a domain-invariant feature exploration (DIFE) method is proposed for IFD under unseen target working conditions. As the name suggests, domain-invariant features refer to the shared and common features that do not change among different working conditions when performing IFD. To further explore these transferable features, DIFE first divides the originally invariant features into two different groups: (1) the internally invariant features, which are embedded in an individual domain and obtained by using a Fourier transform, and (2) the mutually invariant features—features shared across multiple working conditions by aligning these domains. To increase computational efficiency, knowledge distillation (KD) is also used here to capture the internally domain invariant features, which also helps save on fast Fourier transform (FFT) operations for unseen target domain data. Feature fusion is used to formulate the final domain-invariant features since the originally invariant features are divided into two different groups. To ensure diversity of the extracted features, their differences should be maximized. Two experiments indicate that the proposed DIFE method could provide a better domain-invariant feature representation and successfully solve the cross-domain diagnosis problem under unseen working conditions.

中文翻译：

---

域不变特征探索，用于在看不见和时变工况下进行智能故障诊断


深度迁移学习已被证明在执行智能故障诊断 （IFD） 时是有效的，因为它在表征可变工作条件下的振动信号时具有强大的特征表示性能。但是，当目标域数据不可用时，有效训练模型的能力可能非常具有挑战性，因为目标域的特征分布对模型的训练过程没有贡献。针对这一场景，该文提出一种在看不见的目标工况下进行IFD的域不变特征探索（DIFE）方法。顾名思义，域不变特征是指在执行 IFD 时，在不同工况下不变的共享和公共特征。为了进一步探索这些可转移特征，DIFE 首先将原本不变的特征分为两组：（1） 内部不变特征，嵌入在单个域中并使用傅里叶变换获得，以及 （2） 互不变特征——通过对齐这些域在多个工作条件下共享的特征。为了提高计算效率，这里还使用了知识蒸馏 （KD） 来捕获内部域不变特征，这也有助于节省对看不见的目标域数据的快速傅里叶变换 （FFT） 操作。特征融合用于构建最终的域不变特征，因为最初的不变特征被分为两个不同的组。为了确保提取特征的多样性，应最大化它们的差异。 两个实验表明，所提出的 DIFE 方法可以提供更好的域不变特征表示，并成功解决了看不见的工作条件下的跨域诊断问题。