To solve the instability of conditional deep convolutional generative adversarial network （C-DCGAN） in training process， we propose a two time-scale update rule （TTUR） for C-DCGAN with stochastic gradient descent in model training for mechanical fault diagnosis. The model stability is improved when the discriminator and generator have learning rates for their own. Firstly， the convergence of TTUR in C-DCGAN model is proved. Secondly， the validity of the method is verified on the bearing data set of Case Western Reserve University （CWUR） and planetary gearbox data set of laboratories. Finally， the Jensen-Shannon divergence （JSD） is introduced to capture the similarity of generated data to real ones. Experiments suggest that TTUR improves the learning for C-DCGAN and outperforms conventional C-DCGAN.