In vibration-based gearbox fault diagnosis, accurately extracting periodic transient impact signals caused by defects is critical for achieving fault diagnosis. However, the vibration signals measured in practice often contain various interference components, making the extraction of transient impacts quite challenging. To address the issue, this paper proposes a sparsity-enhancing regularization method based on harmonic features. Firstly, a weighted sparse optimization model for gearbox fault signals is established. Secondly, an indicator reflecting the strength of periodic transient impact signals is constructed based on harmonic features. Finally, reweighted regularization is implemented based on the indicator to penalize interference signals, thus enhancing the noise reduction capability. The analysis results of both simulated signals and practical cases verify that the proposed method outperforms other sparse decomposition methods and conventional signal processing methods in terms of the reconstruction accuracy for fault signals, thereby providing more accurate gearbox fault diagnosis results.