A Transferability-Enhanced and Structural Information Mining Network for Cross-Domain Fault Diagnosis

Traditional deep transfer learning methods usually perform global domain adaptation on Euclidean domains, focusing solely on the overall feature distributional differences while overlooking the domain-invariant and structural information between samples. To address this limitation, this paper proposes the transferability-enhanced structural information mining network (TESIMN). First, the multiscale domain adaptation module (MSDAM), is capable of adapting feature distributions at multiple scales, including global distribution differences and subdomain distribution differences. This helps to improve the generalization ability of the network so that it can better adapt to the differences between domains. To mine the structural information of vibration signals, a dual-channel graph convolutional network (DCGCN) is proposed. The representation of domain-invariant features as nodes in the graph samples helps to maintain the consistency of the data structure in cross-domain tasks, thus facilitating cross-domain fault diagnosis. By constructing distance and correlation graphs and feeding them into the DCGCN, the structural relationships among signals are learnt efficiently. The DCGCN utilizes ChebyNet to process the structural information of the non-Euclidean domain signals embedded in the graph data to better capture the similarities and interdependencies among the impassable faults. To assess the model's performance, experiments were conducted on three datasets, and the average accuracy on all tasks of the proposed method is 3.4%, 3.13%, and 0.08% higher than the second place on three datasets, respectively.