Abstract
For high-mobility wireless communications in Orthogonal Time Frequency Space (OTFS) and Non-Orthogonal Multiple Access (NOMA) systems, accurate Channel Estimation (CE) is mandatory. Conventional pilot-based methods, such as Least Squares (LS) and Minimum Mean Square Error (MMSE) estimators, have poor scalability and estimation accuracy in vehicular environments with severe Doppler shifts and multipath propagation. This research study recommended a Deep Bayesian Gaussian Process-Compressive Sensing (DBGP-CS) model that combines probabilistic modelling with sparse signal recovery for CE in OTFS-NOMA. The model uses Deep Neural Networks (DNNs) to learn non-linear delay-Doppler (DD) domain features, Gaussian Processes (GP) for uncertainty quantification, and compressive sensing to exploit channel sparsity, resulting in a 50% reduction in pilot overhead. The network simulations were tested with 100 users at 120 km/h mobility (1,112 Hz Doppler shift) using the Extended Typical Urban (ETU) channel model. The proposed DBGP-CS reported Normalized Mean Squared Error (NMSE) of 0.01447 at 12 dB, which is 90% lower than MMSE-CE. Bit Error Rate (BER) of 0.021159 at 12 dB was also reported. The model was tested across varying mobility speeds (60-120 km/h), multipath complexity (3-9 paths), and energy allocations.