NOMA Channel State Estimation: Deep Learning Approaches

The growing demand for ultra-reliable, high-capacity wireless networks necessitates efficient channel state estimation (CSE) in 6G-enabled non-orthogonal multiple access (NOMA) systems. Conventional estimators like least squares (LS) and minimum mean square error (MMSE) suffer from significant performance degradation under limited pilot symbols, creating a critical bottleneck for spectral efficiency. To address this, we propose a hybrid deep neural network (HyDNN) architecture that synergizes a 1D convolutional neural network (CNN) with bidirectional long short-term memory (BiLSTM) and gated recurrent units (GRU) to jointly exploit spatial-temporal channel correlations. The model is trained offline on multipath fading channel data and deployed online as a pilot-assisted estimator, requiring no prior channel knowledge. Extensive evaluations demonstrate that HyDNN reduces the symbol error rate (SER) by 34% (at 15 dB SNR) compared to MMSE with only 2 pilots, while achieving near-optimal performance with 64 pilots. The Adam-optimized HyDNN (CNN-GRU variant) outperforms standalone CNNs, LSTMs, and GRUs, showing robustness to reduced cyclic prefix lengths (e.g., CP= 12). Notably, with 2 pilots, HyDNN achieves an SER of 0.228 (vs. 0.465 for LS at 10 dB SNR), reducing pilot overhead by 75% without sacrificing accuracy. These advancements position HyDNN as a scalable, resource-efficient solution for next-generation NOMA systems, aligning with 6G goals of low-latency and high spectral efficiency.