Tail-Aware Unsupervised Learning for Wideband Terahertz Quantized Beamforming in Sensor Data Transmission

Data-intensive sensor networks require high-capacity wireless links to transmit sensor measurements. However, in wideband analog beamforming, beam squint reduces array gain, while imperfect channel state information (CSI) and finite-resolution phase shifters further degrade spectral efficiency (SE). To address these challenges, this paper proposes a center-anchored reliability-aware and tail-aware quantized beamforming (CRTQ-BF) algorithm for wideband terahertz sensor-data transmission over single-user multiple-input single-output (MISO) links with one radio-frequency (RF) chain. CRTQ-BF comprises three modules: the center-anchored reliability-aware refinement module learns a wideband phase correction from CSI reliability and a center-subcarrier phase reference; the tail-aware scenario learning module increases the weight of low-SE CSI-perturbation scenarios while retaining the mean-SE objective; and the quantization-aware phase projection module applies hard finite-phase quantization before scenario SE evaluation. Simulation results show that with increased inference time, CRTQ-BF achieves a higher mean SE with fewer trainable parameters than the physically anchored beamforming neural network (BFNN) baseline and demonstrates good robustness to CSI estimation errors. Remote-estimation simulations further show that CRTQ-BF improves sensor-data delivery reliability and reduces remote state-estimation error.