Aiming at the difficulty of weak fault feature extraction of planetary gear, a method of gear fault feature extraction based on the kurtosis of harmonic square envelope spectrum is proposed in this paper. By optimizing the global kurtosis of the time domain signal to the narrow-band kurtosis near the fault harmonic of the signal frequency domain, the influence of fault components on kurtosis is highlighted, and the interference of other components is weakened, and the kurtosis index of harmonic square envelope spectrum is proposed. Taking the proposed index as the standard to determine the optimal resonance frequency band, the frequency band with the most fault information is selected by using the method of 1/3 binary tree. After the analysis of square envelope spectrum, the gear fault feature is extracted. The experimental results show that the method can reduce the interference of other components and effectively extract the weak fault features of planetary gear.