Time-frequency reconstruction discrepancy-guided domain adaptation for open-set fault diagnosis of train bogie bearing.

Open-set domain adaptation (OSDA) fault diagnosis requires addressing both domain shifts and unknown fault identification. High speed train (HST) vibration signals exhibit highly coupled time-frequency non-stationarity, and domain shifts do not manifest uniformly across the time and frequency domains. Feature extraction from either the time or frequency domain alone therefore fails to capture the comprehensive cross-domain discrepancy. Furthermore, prevailing methods perform alignment and separation within the same feature space, where strong initial alignment makes the subsequent separation of unknown samples particularly challenging. To address these challenges, this paper proposes a time-frequency feature reconstruction discrepancy guided domain adaptation framework (TFRDA) for HST bogie fault diagnosis. A time-frequency fusion encoder is developed that extracts frequency-domain features through explicit amplitude-phase fusion, and integrates them with time-domain features to form a time-frequency representation with enhanced fault discriminability. Furthermore, an align-reconstruct domain adaptation network (ARDA) constructs two distinct feature spaces through separate source-domain and target-domain reconstruction constraints, and identifies unknown faults by measuring the reconstruction discrepancy between the two spaces. This dual-representation mechanism decouples alignment from detection, avoiding the boundary ambiguity inherent in single-space methods. Experiments on public bearing and HST bogie datasets demonstrate that TFRDA achieves superior diagnostic accuracy compared to existing OSDA methods, providing an effective solution for trustworthy railway maintenance.