Mingrui Li

Mingrui Li

Oct. 30, 2026

Unsupervised domain adaptation, Machinery fault detection, Time-frequency correspondence, Representation learning, Dual-consistency learning.

Unsupervised domain adaptation works well for cross-domain mechanical fault diagnosis, but mainstream approaches overlook the intrinsic time-frequency relationships within vibration samples. Global distribution alignment will easily damage delicate local fault characteristics. To remedy these shortcomings, we construct a dual-consistency representation learning framework. It combines internal time-frequency physical matching of single samples with global domain alignment. Equipped with bidirectional token cross-attention and domain-decoupled contrastive regularization, the model can preserve robust diagnostic features against noise and large domain gaps. Under standardized practical UDA experimental settings, cross-dataset tests prove our method greatly reduces false alarms and delivers superior generalization performance compared with competing baselines, with no target-domain labels adopted for training.

© 2026, the Authors. Published by ISCCAC

This is an open access article distributed under the CC BY-NC license