A privacy-preserving transfer network for universal fault diagnosis of the gearbox with dynamic threshold

In industrial applications, deep transfer learning (DTL) has become a promising paradigm for improving the generalisation of diagnostic models by leveraging labelled source domains to adapt to unlabelled target domains. However, existing DTL-based methods are often tailored to specific scenarios with fixed domain settings and rely on direct access to source data, raising privacy concerns. To overcome these limitations, a novel Privacy-preserving Universal Transfer Network (PUniTN) is proposed for intelligent fault diagnosis (IFD) of gearbox under universal domain adaptation (UniDA) settings. PUniTN adopts a source-free architecture to ensure data privacy and employs a feature decomposition strategy to orthogonally separate common and private fault features. Meanwhile, a decision boundary with dynamic threshold further enables accurate identification of private fault types in the target domain. Moreover, a multi-component objective combining supervised, decomposition, and consistency losses enhances model generalisation. Experiments on two public gearbox benchmarks demonstrate that PUniTN consistently outperforms state-of-the-art methods in both shared and private fault recognition, providing a robust and privacy-preserving solution for intelligent diagnostics in open, evolving industrial environments.