Multi-sensor fusion meta-transfer learning based on target-guided cross-domain navigation for few-shot fault diagnosis

Fault prediction and health management of aero-engines are essential for ensuring flight safety. However, existing methods remain insufficient to address severe cross-domain feature distribution discrepancies under highly data-scarce scenarios. To overcome these challenges, this study proposes a multi-sensor fusion meta-transfer learning approach based on target-guided cross-domain navigation (MSFMTL). Firstly, we design a multi-sensor information fusion strategy based on cross-attention, which breaks through the physical constraints of single-signal representation and enables deep semantic integration of multi-source signals. Secondly, we develop an innovative target-guided cross-domain navigation mechanism that adaptively regulates the distances between “sample–prototype” pairs in the prototype network. By dynamically modulating the geometric distances between query samples and support prototypes, the method extracts domain-invariant discriminative features and effectively alleviates feature shifts across source and target domains. Finally, a meta-transfer learning framework is constructed to optimize the parameter space of the model, thereby enhancing diagnostic robustness under few-shot cross-domain scenarios. Comparative experiments with state-of-the-art methods demonstrate that the proposed approach achieves superior fault identification accuracy and diagnostic stability, validating its effectiveness in intelligent fault diagnosis of aero-engine inter-shaft bearings.