Model-driven Deep Learning Based Turbo-MIMO Receiver

This paper considers a multiple-input multiple-output (MIMO) receiver with insufficient pilots in fast fading channel environment. Previous studies demonstrated that the pilot sequences should be relatively sufficient to obtain acceptable channel state information. To address this requirement, we investigate the model-driven deep learning based Turbo-MIMO receiver that includes joint channel estimation, signal detection and channel decoding (JCDD) modules. First, we use a short pilot sequence to produce a preliminary estimate of the channel matrix by linear minimum mean-squared error algorithm. Sub-sequently, we re-estimate the channel matrix with the assistance of more reliably estimated symbols and re-detect the data symbols utilizing the soft statistics from the channel decoder. Signal detection is realized in the receiver by representing the expectation propagation (EP) algorithm as multi-layer deep feed-forward networks to optimize the necessary damping factors, which can effectively compensate for the channel estimation error. Numerical results show that the proposed model-driven Turbo-MIMO receiver significantly outperforms the existing algorithms and is effective for the channel estimation with insufficient pilot sequences.