Synergy-driven time-frequency adversarial network for characterizing fault characteristic frequencies in rotating machinery.

Rotating machinery often operates under variable-speed conditions, which generate non-stationary signals containing dense and crossing instantaneous frequencies (IFs), posing significant challenges to existing time-frequency analysis (TFA). Accordingly, the synergy-driven time-frequency adversarial network (SDTFAN) is proposed. First, a learnable time-frequency characterizing module (LTFCM) is designed, where the frequency and chirp-rate of the Chirplet kernel are adaptively learned to optimize the time-frequency mapping process and generate standardized multi-channel time-frequency representations (TFRs). Second, a synergistic time-frequency enhancement framework (STFEF) is constructed with the proposed time-frequency distillation attention module (TFDAM) and time-frequency axial enhancement module (TFAEM). These modules refine local time-frequency features, suppress background interference and enhance global structural dependencies to improve the TFR energy concentration. Finally, a semi-supervised adversarial training strategy is proposed to alleviate the dependence on paired time-frequency labels, namely the ideal TFR and penalty TFR, during model training by introducing adversarial constraints from the unlabeled branch to exploit potential time-frequency features. The results obtained from both simulated signals and planetary gearbox and bearing data demonstrate that the SDTFAN can effectively represent dense and crossing IFs. The Rényi entropy further confirms that the SDTFAN achieves better energy concentration than several TFA methods.