AI-Based Channel Estimation towards 6G Non-Terrestrial Networks

The adoption of non-terrestrial networks in 6G introduces challenging propagation and channel-estimation conditions, including large propagation delays and residual Doppler effects. Furthermore, even harsher conditions will be met due to the 6G-proposed DMRS configurations with a limited pilot availability. This thesis investigates the use of deep learning for channel estimation in such scenarios and presents TransformEs, a compact convolutional-attention architecture specifically designed to operate independently of the size of the resource grid. The proposed model combines a convolutional front-end with a compact self-attention mechanism and is trained end-to-end using synthetic NTN-TDL-D channel realizations generated under different DMRS configurations. Its performance is assessed against conventional LS and LMMSE estimators through both numerical simulations and a hardware-in-the-loop emulator implementing the NTN channel and Doppler characteristics. The proposed approach provides lower BLER than the considered benchmarks for most of the investigated pilot configurations. Its benefit becomes particularly pronounced under reduced pilot availability: on the hardware emulator, the SNR improvement over LMMSE at a BLER of 10% increases from 0.14 dB with the full pilot set to 1.01 dB when two thirds of the pilots are removed. In addition, the computational analysis reveals a linear dependence of TransformEs on the number of allocated resource blocks, compared with the cubic complexity of LMMSE, resulting in more than 54% fewer multiply-accumulate operations at N_PRB=10. Overall, the results demonstrate the potential of compact attention-based architectures to provide robust channel estimates under realistic NTN conditions while offering substantially lower computational requirements than conventional LMMSE estimation for increasing resource allocations.