Massive MIMO is considered as promising technology enabling 5G and beyond 5G cellular communication networks. High-resolution gain and angle estimation of the channel are significant challenges in the design of massive MIMO. The existing signal subspace-based estimation techniques lack the resolution for the detection of slight angles. The severity increases when both azimuth and elevation angle are jointly estimating for the 3D channel. The existing techniques rely on the combination of signal subspace-based approaches to estimate both azimuth and elevation angles. The resulting estimate’s accuracy is less due to its low resolution and angular coupling. Data-driven techniques can address this issue. This work is the first-ever attempt to apply data-driven techniques for 3D channel estimation of massive MIMO to the best of our knowledge. This paper considers two approaches for the same, one based on K nearest neighbour (KNN) and the other based on deep neural network (DNN) with restricted Boltzmann machine (RBM). We also investigated data generation and feature extraction for data-driven communication technologies in three ways. An intensive performance analysis of both architectures using these three feature vectors is carried out. The simulation reveals deep learning (DL) model’s superiority compared to the machine learning (ML)-based counterpart and other signal subspace-based estimation techniques. SNR of 10 dB lesser than other signal subspace estimation is required for ML-based approach, whereas the DL-based estimator needs 20 dB lesser SNR for the received signal to achieve the same BER value. In the comparative study on the feature vectors for data-driven estimation techniques, the data processing based on the Pearson correlation feature vector (PCFV) performs the best. Performance comparison of the DNN model with the KNN model is based on tenfold cross-validation showing an average AUC of 0.915 for DL based estimation and a coefficient of 0.904 for ML-based counterpart.

大规模MIMO被认为是支持5G和5G蜂窝通信网络以外的有前途的技术。信道的高分辨率增益和角度估计是大规模MIMO设计中的重大挑战。现有的基于信号子空间的估计技术缺乏用于检测微小角度的分辨率。当共同估计3D通道的方位角和仰角时，严重性增加。现有技术依赖于基于信号子空间的方法的组合来估计方位角和仰角。由于其低分辨率和角度耦合，因此得出的估算值的准确性较低。数据驱动技术可以解决此问题。这项工作是我们首次尝试将数据驱动技术用于大规模MIMO的3D信道估计。本文考虑了两种相同的方法，一种是基于K最近邻（KNN），另一种是基于具有约束Boltzmann机（RBM）的深度神经网络（DNN）。我们还以三种方式研究了数据驱动的通信技术的数据生成和特征提取。使用这三个特征向量对两种体系结构进行了深入的性能分析。与基于机器学习（ML）的对应方法和其他基于信号子空间的估计技术相比，该仿真揭示了深度学习（DL）模型的优越性。基于ML的方法需要比其他信号子空间估计低10 dB的SNR，而基于DL的估计器需要比接收到的信号低20 dB的SNR才能达到相同的BER值。在对数据驱动的估计技术的特征向量的比较研究中，基于Pearson相关特征向量（PCFV）的数据处理表现最好。DNN模型与KNN模型的性能比较基于十倍交叉验证，对于基于DL的估计，平均AUC为0.915，对于基于ML的对应物，系数为0.904。