Deep learning (DL)-based beam training schemes have enhanced the transmission capacity for intelligent reflecting surface (IRS)-assisted millimeter wave (mmWave) massive multiple-input multiple-output (MIMO) systems. Nevertheless, the effects of noise and model complexity are still bottlenecks for most of those approaches. In this paper, we propose a hybrid convolutional neural network (CNN) encoder-based Transformer (HCNT) DL model containing robust channel expression and a scoring-based optimal beam decision. First, the hybrid CNN encoder parallelly tackles the real and imaginary components of the sampled channel collected by active sensors of the adopted semi-passive IRS with the grouped and point-wise convolution as the alternative to the complicated serial process. Second, we convert the continuous channel expression into binary sequences by leaky integrated-and-fire (LIF) mechanism seeking a robust representation against the noise effects. Finally, feature attention mechanism determines the prediction of the optimal beam by the scores of the relationship between potential direction and quantized binary tokens. Experimental results show that the proposed HCNT outperforms the existing schemes achieving a higher spectral efficiency under different noise levels with lower complexity.
IRS-Assisted mmWave Massive MIMO Systems Beam Training with Hybrid CNN Encoder-based Transformer Deep Learning Model
2023-10-10
1704047 byte
Conference paper
Electronic Resource
English