Semi-supervised transfer learning preserving spatial homogeneity for gearbox diagnostics in extraneous transient noise

Planetary gearboxes are critical components in a wide range of applications, including electric motors, automotive systems, and wind turbines. However, current fault diagnosis methods face significant challenges in accurately detecting faults due to the prevalence of large amounts of unlabelled data and interference from transient events. To address these issues, this paper proposes a novel Semi-Supervised Transfer Learning (SSTL) approach. SSTL integrates semi-supervised techniques with transfer learning and introduces innovative normalisation strategies to overcome the limitations of traditional methods, particularly in environments characterised by transient interference and limited labelled data. The proposed approach offers several key contributions: (1) a semi-supervised transfer learning model that effectively leverages unlabelled data while mapping the source to the target domain, thereby enhancing fault detection accuracy, (2) a label migration and matching strategy that assigns pseudo labels to transient signals, addressing persistent challenges in signal processing, and (3) a limiting normalisation strategy designed to mitigate the effects of transient interference and stabilise vibration signals, thus improving the robustness of the model. Two case studies are presented to validate the effectiveness of SSTL, demonstrating its superiority over existing methods in terms of fault diagnosis accuracy and reliability, especially in scenarios with limited labelled data and frequent transient interference.