The stochastic parallel gradient descent algorithm based on the generalized phase diversity wavefront sensor is presented for co-phasing of segmented mirrors. Cost functions of the optimization algorithm were built up in different circular zones for intensity images of the sensor. In order to achieve high accuracy for co-phasing, four phase diversity functions with increasing amplitudes were applied to the sensor for improving the strength of output signal from the wavefront sensor during the aberrations of the segmented mirror decreasing with the co-phasing process. A simulated segmented mirror was used to test the feasibility of this method. The numerical experiments show that the co-phasing accuracy is very high for the aberrations of the segmented mirrors less than 1.5 wavelengths. And the algorithm is very robust and noise tolerant.