IoT/INTR Joint Seminar | Online Training of DNN-Assisted Channel Estimation for Massive MIMO Systems

Massive MIMO systems can achieve high spectral efficiency as well as interference mitigation through spatial multiplexing and spatial filtering. Channel estimation (CSIR) and feedback (CSIT) of Massive MIMO systems is a critical but challenging component to realize these potential gains. For instance, conventional channel estimation solution requires pilot overhead, which scales with number of transmit antennas. In order to decrease the pilot overheads, we will need to exploit the underlying sparsity structures of massive MIMO channels. There are many works that consider exploiting different levels of sparsity (such as link level sparsity and multi-user sparsity) in the channel estimation and channel feedbacks. However, these solutions are iterative and cannot be implemented on a real-time basis. Recently, there are some works that utilize DNN to achieve real-time inferencing of these iterative CE algorithms. However, the training of these DNN solutions require labeled data (knowledge of true channel) and hence, they are trained in an offline manner based on the channel samples generated from simulations. As such, the offline training will suffer from potential model mismatch because the actual scattering environment the mobile sees will surely be different from that assumed in the simulation model. In this talk, we introduce an online training framework for real-time implementation of the DNN for channel estimation and feedbacks in massive MIMO systems. We first consider a point-to-point system and extend the framework to multi-user systems. The proposed online DNN can track the actual propagation environment on a real-time basis without prior knowledge of the true channel matrix and therefore, is robust to different propagation models, antenna geometry as well as the underlying non-linearity.