WorksEngineering
ANNA: Attention-Free Neural Neighborhood Aggregation for Channel Estimation on Regenerative NTN Payloads
Abstract
Accurate deep learning (DL)-based channel estimation (CE) for non-terrestrial network (NTN) uplinks must meet the processing, power, and latency constraints of regenerative payloads. We propose Attention-free Neural Neighborhood Aggregation (ANNA), a lightweight architecture combining per-resource block (RB) pilot projection, convolutional neighborhood aggregation, and shared reconstruction, with dense-grid expansion only at the output. Evaluation against learned and classical estimators uses paired waveform-generated datasets with receiver frequency correction enabled and disabled. With correction enabled, 16-bit floating point (FP16) deployment on a 10 W Jetson Orin NX achieves 0.372 ms 99th-percentile (P99) latency and 2.15 dB lower mean normalized mean-squared error (NMSE) than the fastest learned baseline at 29.6% higher latency. Relative to the most accurate learned baseline, ANNA is 4.08 times as fast and consumes 80.5% less module energy per sample, with a 1.14 dB accuracy penalty. It meets the adopted 0.5 ms P99 target in all 100 benchmark runs, demonstrating a favorable accuracy-latency-energy trade-off for embedded uplink CE.