An improved atom pursuit method based on the asymmetric Gaussian chirplet model (AGCM) was developed to achieve better performance in low signal-to-noise ratio environment in order to address the issue of poor optimal atom pursuit performance in traditional sparse representation algorithms under strong interference conditions in gearbox. Wavelet atoms were constructed by using AGCM with multi-parameter and variable shapes. Then a feature-enhanced signal was obtained through an impulse enhancement method. Atom parameters most compatible with fault impulses were identified by maximizing a statistical index under Gaussian hypothesis. Then fault transient components were separated through a multiple enhancement sparse representation algorithm. The effectiveness of the proposed method was validated by using both public dataset and gearbox fault simulation dataset. Comparative analysis with original methods and other existing approaches demonstrates that the proposed method can construct superior sparse representation atoms under strong interference conditions in gearbox signals.