Dual Driven Leaning for Joint Activity Detection and Channel Estimation in Multibeam LEO Satellite Communications

This paper investigates the uplink massive connectivity by grant-free random access in intelligent reflecting surface (IRS) assisted low earth orbit satellite communications. By leveraging sporadic activity of the ground devices (GDs), the joint device activity detection and channel estimation (JADCE) problem can be addressed by compressive sensing (CS) algorithms, which either fail to satisfy estimation accuracy or suffer from high computation complexities. Consequently, we propose a general data and model dual driven architecture to efficiently solve the JADCE problem through an unfolded iterative network. Specifically, we improve the original multiple-measurement-vectors (MMV) orthogonal approximate message passing (OAMP) algorithm with an unrolled model driven neural network to exploit the sparse beamspace channel. Moreover, we incorporate the data driven in each iteration, termed model and data dual driven OAMP network (DOAMPNet), which adaptively learns channel sparsity and improves channel estimation performance with model guarantees. Extensive simulations are provided to demonstrate the superiority of the proposed model and data dual driven networks compared with existing methods in terms of estimation accuracy. Remarkably, the proposed DOAMPNet reduces pilot overhead by about 40%, and achieves a normalized mean-square error improvement of about 4 dB when signal-to-noise ratio is 10 dB.