Dynamic Meta-Decoupler-inspired Single-Universal Domain Generalization for Intelligent Fault Diagnosis

Abstract

Rotating machinery in industry operates under complex conditions, with monitoring data influenced by irregular load fluctuations. Traditional domain generalization methods address distribution shifts using data from multi-source domains. However, it is time-consuming and expensive to collect data that covers all operating conditions and fault types. To overcome these limitations, this paper considers a more realistic yet challenging scenario called Single-Universal Domain Generalization (Single-UDG). It utilizes only single-source domain data to address the difficulties of unknown target domain data and unknown class recognition. We propose a novel learning framework called Dynamic Meta-Decoupler by decoupling domain-dynamic parameters. By adding Meta-Perturb and Parameters-Perturb strategies, Dynamic Meta-Decoupler is enforced to learn more robust shared features. Additionally, to fully tackle the challenges posed by Single-UDG, we propose a novel training strategy called Meta Generative Adversarial Network (MetaGAN). By utilizing Meta-Perturb-enhanced instances, our model is enhanced to generalize to unknown target domains and reject unknown faults. Extensive experiments conducted on two machinery datasets demonstrate that our model effectively addresses Single-UDG fault diagnosis under unknown working conditions.