Improving Machine-to-Machine (M2M) communication is essential for the development of Smart Factory as data can be exchanged and processed more efficiently. Herein this study, we employ the Deep Learning (DL) concepts aimed at improving end-to-end performance (E2E) M2M communication systems. Training the physical layers requires the explicit channel information to be fully known, which can be solved with generative adversarial network (GAN). Nonetheless, due to its deep neural network (DNN) structure, the GAN scheme is subjected to gradient vanishing and over-fitting, two major obstacles that can hinder the training process and limit the performance of the model. As a result, the system is significantly downgraded. To address these issues, we study a method known as Residual-Aided generative adversarial network (RA-GAN) learning scheme, in which the two problems are dealt with respectively by introducing a better propagation mechanism and a regularizer to the loss function. Herein this paper, the system model is described and the two problems are derived analytically. We also analyze the optimal learning scheme (where the channel-agnostic) and a Rayleigh-based learning scheme for comparison study. Through analyzing the block error rate (BLER), we can demonstrate that the RA-GAN approach achieves performance comparable to the optimal scheme, and significantly outperforms the conventional GAN method.