A Data Augmentation-Based Channel Estimation Scheme for Intelligent Reflecting Surface-Assisted Wireless Communication System

The performance of intelligent reflecting surface-assisted wireless communication systems highly depends on the accuracy of channel estimation. To resolve the contradiction between existing data-driven channel estimation schemes that rely on ground truth (labels of the true channels) for network updates and the unavailability of ground truth in practical applications, an accurate and deep denoising convolutional neural network (ADDnet) is designed. Channel state information (CSI) estimates derived from the least squares method are utilized as a replacement for ground truth to train the network, which facilitates feasible data-driven channel estimation in practice. Additionally, a training dataset construction method based on the deep convolutional generative adversarial networks (DCGANs) data augmentation mechanism is proposed. New samples generated through the game between the generator and the discriminator are mixed with the original data to construct training dataset. It can help overcome the limitations imposed on network generalizability by training datasets generated through computer simulations and reduce the resource consumption and latency associated with online collection of training data. Simulation results demonstrate that, compared to existing channel estimation schemes, the proposed channel estimation scheme exhibits superior denoising capability and adaptability to variations in the number of users and channel conditions.