In this paper, we propose an online training framework for mmWave Massive MIMO channel estimation (CE) with limited pilots, where the training is based on real-time received pilot samples from the base station without requiring knowledge of the true channel. To realize this, we propose four axioms for a legitimate online loss function, based on which we develop a model-free online training algorithm with convergence analysis. Simulation shows that the proposed online deep neural network (DNN) achieves comparable CE accuracy to model-based compressive sensing (CS) algorithms, while enjoying much faster computation. In addition, the proposed method is robust to various model mismatches and can adapt to the change of the underlying propagation environment.

中文翻译：

---

具有任意阵列几何结构的MmWave大规模MIMO信道估计的在线深层神经网络

在本文中，我们提出了一种用于有限导频的毫米波大规模MIMO信道估计（CE）的在线训练框架，其中的训练基于从基站实时接收的导频样本，而无需了解真实信道。为了实现这一点，我们提出了一个合法的在线损失函数的四个公理，在此基础上，我们开发了一种具有收敛性分析的无模型在线训练算法。仿真表明，提出的在线深层神经网络（DNN）具有与基于模型的压缩感知（CS）算法相当的CE精度，同时享有更快的计算速度。另外，所提出的方法对于各种模型失配是鲁棒的，并且可以适应底层传播环境的变化。