Dual Attention Smoothing Adaptation Networks for Aeroengine Multisource Cross-Domain Fault Diagnosis under Category Shift

Abstract

In aeroengine fault diagnosis, domain adaptation-based methods have effectively handled the data distribution shift caused by variations in operating conditions. However, current research mainly focuses on knowledge transfer from a single source domain. In real engineering scenarios, data are collected from multiple conditions, and each domain contains only a subset of the total health states, leading to category shift. Using data from multiple source domains can improve diagnostic capabilities in the target domain, but aligning them with the target domain directly may result in negative transfer caused by category shift. To address these challenges, we design a novel network architecture, the dual attention–based smooth adaptation network (DASAN), with three distinctive characteristics: (1) a dual-domain interdomain attention mechanism that identifies shared health states across multiple source domains while mitigating the impact of noisy target domain samples, (2) a CORAL distance–based domain level attention (CDDLA) that mitigates the negative effects of source domains that differ significantly from the target domain, and (3) a weighted environment label smoothing (WELS) strategy built upon CDDLA to enhance the stability of adversarial training. Extensive experimentation on turboshaft engine and Case Western Reserve University (CWRU) bearing data sets validated the superior performance of our proposed approach compared to advanced methods.