In the sixth-generation (6G) cellular networks, hybrid beamforming would be a real-time optimization problem that is becoming progressively more challenging. Although numerical computation-based iterative methods such as the minimal mean square error (MMSE) and the alternative manifold-optimization (Alt-Min) can already attain near-optimal performance, their computational cost renders them unsuitable for real-time applications. However, recent studies have demonstrated that machine learning techniques like deep neural networks (DNN) can learn the mapping done by those algorithms between channel state information (CSI) and near-optimal resource allocation, and then approximate this mapping in near real-time. In light of this, we investigate various DNN architectures for beamforming challenges in the terahertz (THz) band for ultra-massive multiple-input multiple-output (UM-MIMO) and explore their contextual mathematical modeling. Specifically, we design a sophisticated 1D convolutional neural network and long short-term memory (1D CNN-LSTM) based fusion-separation scheme, which can approach the performance of the Alt-Min algorithm in terms of spectral efficiency (SE) and, at the same time, use significantly less computational effort. Simulation results indicate that the proposed system can attain almost the same level of SE as that of the numerical iterative algorithms, while incurring a substantial reduction in computational cost. Our DNN-based approach also exhibits exceptional adaptability to diverse network setups and high scalability. Although the current model only addresses the fully connected hybrid architecture, our approach can also be expanded to address a variety of other network topologies. INDEX TERMS 6G, CNN, Hybrid Beamforming, LSTM, UM-MIMO

中文翻译：

---

用于 6G 超大规模 MIMO 混合波束成形的基于 CNN-LSTM 的融合分离深度神经网络

在第六代 (6G) 蜂窝网络中，混合波束成形将是一个实时优化问题，并且变得越来越具有挑战性。尽管基于数值计算的迭代方法，如最小均方误差 (MMSE) 和替代流形优化 (Alt-Min) 已经可以达到接近最佳的性能，但它们的计算成本使它们不适合实时应用。然而，最近的研究表明，深度神经网络 (DNN) 等机器学习技术可以学习由这些算法在通道状态信息 (CSI) 和接近最优资源分配之间完成的映射，然后近乎实时地逼近这种映射。有鉴于此，我们研究了各种 DNN 架构，以应对用于超大规模多输入多输出 (UM-MIMO) 的太赫兹 (THz) 频段的波束成形挑战，并探索它们的上下文数学建模。具体来说，我们设计了一个复杂的 1D 卷积神经网络和基于长短期记忆 (1D CNN-LSTM) 的融合分离方案，它可以在频谱效率 (SE) 方面接近 Alt-Min 算法的性能，并且在同时，显着减少计算工作量。仿真结果表明，所提出的系统可以达到与数值迭代算法几乎相同的 SE 水平，同时大大降低了计算成本。我们基于 DNN 的方法还表现出对各种网络设置的出色适应性和高可扩展性。尽管当前模型仅解决完全连接的混合架构，但我们的方法也可以扩展到解决各种其他网络拓扑。索引术语 6G、CNN、混合波束成形、LSTM、UM-MIMO