An adaptive source-free unsupervised domain adaptation method for mechanical fault detection

Abstract

Cross-machine fault detection is crucial due to the challenges of data labeling. While domain adaptation methods facilitate diagnosis across rotating machines, they often require data sharing, which is impractical due to privacy concerns and large data transmission. Although domain generalization and source-free unsupervised domain adaptation (SFUDA) methods address privacy issues, most fail to consider dynamic distribution shifts within and between domains, limiting their effectiveness. To overcome this challenge, an adaptive SFUDA method named AI3M is proposed. The AI3M pre-trains a source model with intra- and inter-domain information maximization loss to reduce distribution shifts within and between domains, and then adapts the model with a target-guided adaptation strategy to minimize the dynamic gap between different machines. Experiments on datasets from 11 wind turbines across 8 wind farms show that the proposed method outperforms state-of-the-art DG and SFUDA approaches, achieving superior cross-machine fault detection performance.