Transformer-based Hybrid Beamforming with Dynamic Subarray for Near-Space Airship-Borne Communications

This paper proposes a hybrid beamforming framework for massive multiple-input multiple-output (MIMO) in near-space airship-borne communications. To achieve high energy efficiency (EE) in energy-constraint airships, a dynamic subarray structure is introduced, where each radio frequency chain (RFC) is connected to a disjoint subset of the antennas according to channel state information (CSI). The proposed joint dynamic hybrid beamforming network (DyHBFNet) comprises three key components: 1) An analog beamforming network (ABFNet) that optimizes the analog beamforming matrices and provides auxiliary information for the antenna selection network (ASNet) design, 2) an ASNet that dynamically optimizes the connections between antennas and RFCs, and 3) a digital beamforming network (DBFNet) that optimizes digital beamforming matrices by employing a model-driven weighted minimum mean square error algorithm for improving beamforming performance and convergence speed. The proposed ABFNet, ASNet, and DBFNet are all designed based on advanced Transformer encoders. Simulation results demonstrate that the proposed framework significantly enhances spectral efficiency and EE compared to baseline schemes. Additionally, its robust performance under imperfect CSI makes it a scalable solution for practical implementations.

💡 Research Summary

The paper addresses the challenge of implementing massive MIMO on near‑space airships, which are power‑constrained platforms operating at altitudes of 20–100 km. To reconcile the limited number of RF chains with the large antenna arrays required for high array gain, the authors introduce a dynamic subarray architecture: each RF chain is connected to a disjoint subset of antennas, and the subset is reconfigured according to instantaneous channel state information (CSI).

A joint deep‑learning framework called DyHBFNet is proposed, consisting of three Transformer‑based modules.

1. Analog Beamforming Network (ABFNet) converts the complex‑valued CSI tensor into a real‑valued matrix, embeds it into a 256‑dimensional space, and processes it through three identical Transformer encoder layers (multi‑head self‑attention + feed‑forward). The output yields a normalized analog beamforming matrix that satisfies the unit‑modulus constraint, as well as auxiliary real‑valued features (ˆF_RF) that will be used by the antenna selection module.
2. Antenna Selection Network (ASNet) treats the mapping from antennas to RF chains as an N_t‑class classification problem. It concatenates the auxiliary features from ABFNet with the CSI, feeds the combined sequence into a second Transformer encoder, and finally applies a soft‑max followed by an arg‑max per antenna to generate a binary selection matrix X_sel. This matrix enforces the “one‑RF‑per‑antenna” hardware rule while dynamically adapting to channel conditions.
3. Digital Beamforming Network (DBFNet) replaces the conventional iterative Weighted Minimum Mean Square Error (WMMSE) algorithm with a model‑driven version. By algebraically reformulating the WMMSE solution in terms of a small set of parameters {a, b, c}, a third Transformer encoder predicts these parameters directly from the equivalent baseband channel H_equ = H^T F_RF. The digital precoder F_BB is then obtained in closed form, eliminating the need for costly iterations and accelerating convergence.

The overall optimization problem maximizes spectral efficiency (SE) under three constraints: unit‑modulus for analog phases, binary antenna‑RF assignment, and total transmit power for the digital precoder.

Simulation settings use a 256‑antenna planar array, 8 RF chains (equal to the number of users), and 64‑subcarrier OFDM with a Rician channel model (one LoS path plus several NLoS components). Compared with fixed subarray HBF, conventional block‑coordinate descent, and purely data‑driven deep‑learning HBF, DyHBFNet achieves 10–15 % higher SE and more than 20 % improvement in energy efficiency (EE). Moreover, performance degrades only marginally when CSI error reaches 10 %, demonstrating robustness to imperfect channel knowledge. The model‑driven WMMSE reduces computational complexity by a factor of five relative to the classic iterative WMMSE while preserving or surpassing its achievable SE.

The authors conclude that the synergy of dynamic subarrays, Transformer‑based feature extraction, and model‑driven digital precoding provides a practical pathway to deploy massive MIMO on power‑limited near‑space platforms. Limitations include the need for sizable training data and computational resources for the Transformers, and the current focus on a single‑airship, single‑cell scenario. Future work may explore multi‑airship coordination, mobility‑aware online adaptation, and hardware‑friendly lightweight Transformer variants. Overall, the paper offers a compelling solution for 6G‑era high‑altitude communications, balancing hardware constraints with advanced AI‑driven beamforming performance.