Seismic data noise suppression method based on deep learning is an important field of seismic data denoising processing. Deep learning method breaks through the limitation of traditional filter processing, and shows high efficiency and good signal- to- noise separation effect in noise suppression of conventional seismic data. However, for deep weak reflection data, the current deep learning methods have limited feature extraction ability, and it is difficult to achieve good denoising effect. Combined with the characteristics of deep reflection seismic data, and aiming at the limitations of current deep learning noise suppression methods in feature extraction and data set dependence, we propose a stochastic noise suppression method for deep reflection seismic data based on Attention Cycle- Consistent Generative Adversarial Networks (A- CGAN). Using the domain mapping idea of Cycle- Consistent Generative Adversarial Networks (Cycle- GAN), the requirements on data sets are reduced. In order to build a denoising network suitable for deep reflection seismic data, improvements are made to Cycle- GAN from three aspects: adding residual structure and attention mechanism to Cycle- GAN generator (denoising device) to deepen the depth of the network and improve its feature extraction capability; the block decision is used in the discriminator of Cycle- GAN to improve the precision and accuracy of the discriminator; adding the perception consistency loss function to the loss function improves the ability of the network model to recover texture details. The numerical model data and the actual deep reflection seismic data are tested to verify the effectiveness of the optimization algorithm, which shows good application value.