Hybrid beamforming (HBF) design for wideband millimeter-wave (mmWave) systems has two challenges: 1) design complexity due to frequency selectivity; 2) large pilot overhead and feedback overhead due to heavy dependence on the channel information. This letter proposes a deep learning-based HBF method for wideband mmWave systems. First, we simplify the complex HBF design problem into a network optimization problem by designing the loss function with spectral efficiency as an optimization objective. Then, we replace the instantaneous channel state information with channel statistics information of multiple time slots to achieve low system overhead. Specifically, we develop an HBF network (HBFNet) based on convolutional neural network with the channel covariance matrix (CCM) as input and the hybrid beamformer as output. Meanwhile, we design a constraint layer to satisfy the constant modulus constraint and power constraint in HBF design. In addition, we adopt the imperfect CCM for training to make the proposed method more robust.

中文翻译：

---

用于宽带毫米波系统的基于深度学习的开销最小化混合波束成形

宽带毫米波 (mmWave) 系统的混合波束成形 (HBF) 设计面临两个挑战：1) 频率选择性导致的设计复杂性；2) 由于对信道信息的严重依赖，导频开销和反馈开销很大。这封信提出了一种用于宽带毫米波系统的基于深度学习的 HBF 方法。首先，我们通过设计以频谱效率为优化目标的损失函数，将复杂的 HBF 设计问题简化为网络优化问题。然后，我们将瞬时信道状态信息替换为多个时隙的信道统计信息，以实现低系统开销。具体来说，我们开发了一个基于卷积神经网络的 HBF 网络 (HBFNet)，其中通道协方差矩阵 (CCM) 作为输入，混合波束形成器作为输出。同时，我们设计了一个约束层来满足HBF设计中的恒模约束和功率约束。此外，我们采用不完美的 CCM 进行训练，以使所提出的方法更加稳健。