Learnable Topology Enhanced Heterogeneous Network for Unbalanced Cross-Domain Fault Diagnosis

The emerging graph convolutional networks (GCN) have led to their widespread application in fault diagnosis due to their advantages in feature extraction. However, there are still some limitations of the GCN-based methods in the face of cross-domain diagnostic tasks. First, in most of the existing studies, the graph topology is constructed according to certain criteria. However, the fixed topology may contain domain-specific information, which could hinder the extraction of domain-invariant features. Second, the class unbalanced data would lead to training bias, degrading the diagnostic performance for the minority classes of data. In order to address the challenges posed by these limitations to cross-domain diagnostics, this paper proposed a learnable topology enhanced heterogeneous network (LTHN). First, a learnable topology-aware module is designed to update the multi-sensor graph topology and augment the unbalanced data in a non-parametric manner. Then, a fault detection induced binary classifier and a fine-grained classification induced multi-class classifier form a heterogeneous network, which is able to avoid biased training by reweighting the data according to the pairwise heterogeneous inferences. Finally, the multi-class classifier output the diagnostic result. Cross-domain diagnostic experiments were conducted on the constructed sliding bearing test rig. Comparative studies with the state-of-the-art methods showed the effectiveness and the superiority of the proposed method.