YConvFormer: A Lightweight and Robust Transformer for Gearbox Fault Diagnosis with Time-Frequency Fusion

This paper addresses the core contradiction in fault diagnosis of gearboxes in heavy-duty equipment, where it is challenging to achieve both lightweight and robustness in dynamic industrial environments. Current diagnostic algorithms often struggle with balancing computational efficiency and diagnostic accuracy, particularly in noisy and variable operating conditions. Many existing methods either rely on complex architectures that are computationally expensive or oversimplified models that lack robustness to environmental interference. A novel, lightweight, and robust diagnostic network, YConvFormer, is proposed. Firstly, a time-frequency joint input channel is introduced, which integrates time-domain waveforms and frequency-domain spectrums at the input layer. It incorporates an Efficient Channel Attention mechanism with dynamic weighting to filter noise in specific frequency bands, suppressing high-frequency noise and enhancing the complementary relationship between time-frequency features. Secondly, an axial-enhanced broadcast attention mechanism is proposed. It models long-range temporal dependencies through spatial axial modeling, expanding the receptive field of shock features, while channel axial reinforcement strengthens the interaction of harmonics across frequency bands. This mechanism refines temporal modeling with minimal computation. Finally, the YConvFormer lightweight architecture is proposed, which combines shallow feature processing with global-local modeling, significantly reducing computational load. The experimental results on the XJTU and SEU gearbox datasets show that the proposed method improves the average accuracy by 6.55% and 19.58%, respectively, compared to the best baseline model, LiteFormer.