Investigation of Deep Learning Based NOMA System Over Time Varying Fading

This study investigates the deep learning (DL)-based uplink power domain non-orthogonal multiple access (PD-NOMA) system, with a focus on channel state information (CSI) estimators. The effects of channel outages and symbol error rates (SER) caused by time-selective Nakagami-m fading with Doppler spreading are looked at in this study. To figure out CSI, the long-short-term memory (LSTM) method is used. To get the best performance from the deep neural network (DNN), a variety of datasets are used to train it. When DL-based channel estimation is added to the NOMA system, SER goes down by a lot, especially when power allocation stays the same. In particular, when SER values like 10
⁻²
are taken into account, the uplink NOMA technique reduces power by a lot, from about 5–8 dBm for users far and medium distances away to up to 4 dBm for users closer by. The results of the simulation show that using a DL-based channel estimate results in a significant power loss of about 6 dBm for users located both far away and close by in order to achieve a certain OP of 10
⁻⁴
, compared to the standard root mean square error (RMSE)-based method. Simulations show that DL-based channel estimation works better than NOMA over different distances and node velocities. The outage probability (OP) curve for the distance coupling between 1000 and 400 meters shows a bigger difference in user channel gains than the curve for the distance pairing between 1000 and 600 meters.