A time–frequency dual-branch network for fault diagnosis of wind turbine gearboxes under noise

Abstract As a critical component of wind turbines, planetary gearboxes are susceptible to interference from complex environments and transmission path effects, which often mask fault characteristics in vibration signals and compromise diagnostic accuracy. To address these challenges, this paper proposes a time–frequency dual-branch network with dual-level gated fusion (TFDN-DLGF). The model first employs a collaborative preprocessing mechanism that integrates dual-dimensional quality scoring with an improved complete ensemble empirical mode decomposition with adaptive noise signal-to-noise separation method to enhance signal quality and reinforce fault signatures. A dual-branch architecture then extracts complementary features: the time-domain branch utilizes impact-perceptive convolution, ResNext-based dense blocks, and BiGRU to effectively resist broadband noise interference while capturing multi-scale impact characteristics; the frequency-domain branch adopts convolutional neural network, frequency-channel dual attention, and BiGRU to extracts fault-sensitive frequency bands and their evolutionary patterns stably under noisy backgrounds. Finally, the DLGF module achieves adaptive integration at both feature and decision levels, significantly improving noise robustness. Experimental results demonstrate that the proposed method achieves accuracies of 93.75% and 94.50% under strong noise conditions and approaches near-perfect performance under noise-free conditions, substantially outperforming existing methods and providing a reliable solution for gearbox fault diagnosis in high-noise industrial environments.