An impact feature extraction method, based on twodimensional kernel density estimation for S transform spectrum, is proposed to analyze the vibration signal for gear fault diagnosis. In this approach, S transform is used to process the vibration signal, firstly. Secondly, the obtained Stransform spectrum is multiplied by a factor and then rounded to obtain an integer matrix. Finally, the time and the frequency of the Stransform spectrum are used to construct a twodimensional random variable, and the elements in the integer matrix are taken as the corresponding sample number of the twodimensional random variable. The kernel density of the twodimensional random variable is consequently estimated and a twodimensional kernel density function is obtained. Specifically, the kernel density function is acquired by the smoothing and denoising procedure of the S transform spectrum, in which the noise is effectively suppressed while the impulse signature is enhanced. By means of the processing of the simulated vibration signal and the gearbox fault vibration signals, results show that the proposed method can extract the periodic impact characteristics from the vibration signal effectively, which means the proposed method can be used for gearbox fault diagnosis.