Interpretable Transfer Unrolling Network for Intelligent Cross-Domain Condition Monitoring of Rotating Machinery

In the domain of mechanical condition monitoring, deep neural networks (DNNs) have proven their excellent performance. Meanwhile, some interpretable networks have been proposed to improve interpretability and have shown promising results in interpretable intelligent condition monitoring. However, when the working condition becomes variable and the tasks of transfer learning need to be carried out, most existing interpretable networks no longer work effectively. Therefore, this article proposes an interpretable transfer-unrolling network (ITUN) to fulfill interpretable transfer tasks in the domain of condition monitoring. In ITUN, to realize the interpretability of feature extraction, a multilayer sparse coding model is established, and its iterative solving algorithm is deduced. Meanwhile, the equivalent form of iterative solving algorithm is attained by unrolling fast iterative soft thresholding algorithm (FISTA), which serves as the feature extractor of ITUN. Moreover, to facilitate cross-domain fault diagnosis, a double-dictionary is embedded into the sparse coding model to extract the shared and private features from source and target domains, respectively, and the loss function of ITUN for interpretable transfer diagnosis is designed. The effectiveness of the proposed ITUN is verified through two bearing datasets. In each transfer task of two bearing datasets, the diagnosis accuracy of the proposed ITUN is above 98% and its robustness is obviously superior to other comparison methods, which prove that the proposed ITUN can accomplish the task of interpretable transfer learning outstandingly in the domain of mechanical condition monitoring.