Dynamic Condition-Matching Network for End-to-End Machinery Health Monitoring Under Time-Varying Operating Conditions

Dynamic monitoring of critical machinery under time-varying operating conditions remains a challenging problem. Existing methods often fail to efficiently and conveniently generate reliable health indicators. To address this issue, a Dynamic Condition-Matching Network (DCMN) is proposed to end-to-end generate machinery health indicators under time-varying operating conditions. First, a self-supervised matching-pair-based sample generation method is developed. Real and virtual vibration–speed pairs are constructed to produce pseudo-regression labels for exploring latent cross-modal relationships. Second, cross-attention is introduced into DCMN to model the vibration–speed cross-modal representation association. Fused features that reflect deviation from the reference baseline are thereby generated. Finally, a cross-modal matching-rate prediction head is established in DCMN. The health indicators under given vibration and speed inputs can be computed in an end-to-end manner. The effectiveness of DCMN is validated through a case study under time-varying operating conditions. Experimental results demonstrate that DCMN can effectively generate condition-invariant health indicators. It outperforms the compared methods and provides an elegant approach for machinery health monitoring under time-varying conditions.