Deep Channel Estimator for Superimposed Pilot-Aided MIMO-OFDM Systems

Abstract

Pilot overhead is a bottleneck for improving spectral efficiency in multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems. To alleviate this issue, superimposed pilot (SIP)-aided transmissions have been widely studied. However, channel estimation under SIP remains challenging due to the strong interference from multi-stream data. In this letter, we propose a deep learning (DL)-based channel estimator, namely DeepCE, for SIP-aided MIMO-OFDM systems. The proposed DeepCE adopts a standard transformer encoder as its backbone for non-iterative channel estimation. To capture the inherent time-frequency correlations of OFDM channels, we propose a novel two-dimensional (2D) positional encoding (PE) scheme that guarantees 2D shift invariance. Simulation results show that the proposed DeepCE outperforms the least squares, linear minimum mean squared error, and DL-based channel estimators. Moreover, compared to the one-dimensional PE scheme, the proposed 2D-PE scheme further improves channel estimation accuracy while preserving superior generalization capability.