Geometry-Informed Neural Network for Pilot-Limited Channel Estimation via Sensing from Uplink Traffic

Abstract

Reliable wireless systems depend on accurate uplink channel estimation for precoding, scheduling, and link adaptation at the base station. In the upper mid-band, wide bandwidths and large antenna arrays increase the channel dimension and make pilot-efficient estimation particularly challenging. While uplink data is available as every user transmits it, sounding resources are shared across all users in the cell, leaving only a few pilots for each user. Thus, channel estimation heavily depends on prior information. Existing approaches obtain such priors in different ways, each with limitations. Ray tracing requires a site model, channel knowledge maps rely on measurements at or near the locations they represent, and learned channel distributions do not explicitly capture propagation geometry. We therefore propose TRaffic-Aided Channel Estimation (TRACE), which builds the prior from uplink traffic without additional sounding. TRACE learns a geometry-aware channel prior from routine multi-UE uplink traffic and combines it with the available pilots for channel estimation. In a ray-traced street canyon at 15 GHz, TRACE achieves -7.48 dB NMSE with 8 pilots, outperforming three state-of-the-art priors and two geometric ablations by more than 3 dB, while remaining effective with only 2 pilots. Within a deployment, TRACE remains robust across SNRs, carrier frequencies, and reduced mapping traffic without retraining.