Neural Network-Based Channel Estimation for Physical Uplink Shared Channel in Over-the-Air Evaluation

Neural channel estimation has demonstrated promising link-level performance gains in simulation, yet its viability for real-time deployment in fifth generation (5G) base-station receivers under practical hardware constraints remains insufficiently explored. In particular, most existing work evaluates channel estimation offline using recorded datasets rather than under realistic over-the-air (OTA) conditions. This paper investigates the real-time OTA performance of neural channel estimation for 5G new radio physical uplink shared channel within a fully integrated hardware testbed. We develop two neural estimators: CNN-est, a 3D convolutional model, and DSCNN-est, a computationally efficient depthwise-separable variant. Both estimators employ a fused estimator-denoiser architecture and operate with a single-symbol Type-2 demodulation reference signal pilot pattern, achieving up to 5.1% uplink throughput gain. Simulation results show accuracy comparable to linear minimum mean square error across diverse clustered delay line channels, while real-time OTA measurements demonstrate up to 1 dB signal-to-noise-ratio reduction at a target uplink block error rate of 0.1 compared to the OpenAirInterface least-squares baseline under practical radio-frequency impairments. These findings provide concrete evidence of the feasibility and performance trade-offs of neural channel estimation in live 5G systems. © 2026 IEEE.