Machine learning-enabled channel estimation for massive MIMO in high-speed railway communications

ABSTRACT With the rapid development of 5G mobile communications, achieving seamless wireless connectivity in high-speed rail scenarios remains challenging. Massive MIMO technology shows great promise, but channel estimation faces severe difficulties due to extreme Doppler effects and non-stationary propagation characteristics. Conventional LS and LMMSE algorithms suffer significant accuracy degradation under high mobility conditions and fail to exploit the temporal correlation of massive MIMO channels.This paper proposes a novel CNN-BiLSTM network architecture for channel estimation. The model employs a spherical coordinate system to represent multipath propagation, including line-of-sight, sea surface reflection, and single/double-bounced components. A DFT-based beam space representation reduces dimensionality and achieves channel sparsity. The architecture utilizes CNN layers for spatial feature extraction and BiLSTM layers for bidirectional temporal modeling. Additionally, a hybrid learning approach combining theoretical Rician and ray-traced channel models enhances generalization across various propagation conditions.Simulation results demonstrate that the proposed scheme achieves a 10.4 dB NMSE improvement over conventional LMMSE at 400 km/h, reduces pilot overhead by approximately 67%, while maintaining sum-rate performance within 5% of ideal CSI conditions.