Multi-source domain mechanical fault diagnosis method based on adversarial semantic decoupling and prototype space geometric constraints

Multi-source domain adaptation enhances cross-domain generalisation capability in fault diagnosis by integrating knowledge from multiple source domains. However, existing methods suffer from severe interference of working condition-related features. The fault features in vibration signals are highly coupled with working condition-specific components, leading to negative transfer phenomena when models are transferred across different working conditions. This paper proposes a multi-source fault diagnosis method using adversarial semantic decoupling and prototype space geometric constraints. First, a heterogeneous modal information interaction network is designed to extract highly discriminative cross-modal fusion features. Second, an adversarial semantic decoupling method is proposed, which achieves effective separation of fault semantic features from working condition-specific components through mask-guided dual-pathway decoupling, thereby alleviating the high coupling between them. Finally, an adaptive pseudo-label dual-screening mechanism with prototype space geometric constraints is developed. This mechanism constructs constraint boundaries for fault class prototypes using geometric similarity and filters high-confidence pseudo-labels through iterative updates of adaptive thresholds, further mitigating negative transfer phenomena. Extensive experimental results on public datasets and real industrial data validate the effectiveness of the proposed method.