Multitarget domain adaptation with transferable hyperbolic prototypes for intelligent fault diagnosis

Intelligent fault diagnosis models always behave poorly in a new unannotated scenario owing to the domain shift problem. As a solution, deep transfer learning has led to frontier research in cross-domain intelligent fault diagnosis. However, the existing deep transfer learning is mainly oriented towards a single target domain. Domain adaptation for multitarget relies on multiple transfer processes and are dispersed into multiple target models. For one model to adapt to multiple target domains, inconsistent matching for each target and confusion of class information are significant hindrances. Following this consideration, we propose a blending target adversarial network with transferable hyperbolic prototypes (THPN). Contrapuntally, the proposed TPHN reconciles the discrepancy between the source domain and multiple target domains and enforces sample-level classification consistency. The source–target and target–target domain pairs are aligned by confusing the blending target domain discriminator. Optimizing the prototypical score discrepancy in hyperbolic space guarantees classification consistency at the sample level. The proposed and compared methods are extensively validated on two multiscene datasets to evaluate their effectiveness. The experimental results demonstrate the superiority and feasibility of the proposed method for cross-domain fault diagnosis in multitarget scenarios.