As for the fault diagnosis of gear at early stage, the conventional methods of signal processing are significantly interfered by noise, blocking the fault feature extraction of gear. This paper proposes a PSO sparse decomposition combined with PSO (Particle swarm optimization) algorithm and sparse decomposition algorithm, lowering the computing complexity of sparse decomposition, and also proposes a ‘Matching index’ as the signal feature. The research result of the simulated signal indicates that PSO decomposition performs well under condition of strong noise and improves the SNR greatly. What’s more, the PSO sparse decomposition is proved efficiently in fault signal feature extraction of gear through the analysis of the signal from aeroengine gear hub. The ‘Matching index’ of fault signal is 04 higher equally than that of normal signal. This is superior obviously to the traditional methods.