Frequency Domain Learning Scheme for Massive MIMO Using Deep Neural Network

Abstract

Massive MIMO is one of the cornerstones of 5G technology. MIMO scaled up to hundreds or even thousands of antenna terminals can result in an extensive increase in the capacity at reduced computational complexity. Channel State Information (CSI) estimation has an indispensable role in the deployment of massive MIMO. Since the spatial information is important for the massive MIMO phase component to have higher significance as compared to the magnitude component in CSI. If the phase estimation of the channel can be made accurate, we can ensure efficient estimation of channel gains as well. Thereby ensuring the error-free transmission of massive data. The proposed multi-layer perceptron model for massive MIMO takes the beamformed signal with higher directivity as its input and learns the features of different channel conditions and predict the direction of arrival (DoA) or Angle of Arrival (AoA) of the received signal. This accurate prediction of DoA helps in the estimation of channel conditions much better than the time domain counterpart especially with a reduced number of iterations. The proposed system has better metrics about the accuracy, mean squared error (MSE) performance, and bit error rate (BER) performance. The number of epochs required for training is less implies computational complexity is less, which is a significant improvement comparing with other data-driven techniques. Such a scheme that can make predictions on the channel at a very lesser time helps to adapt the transmission parameters according to the channel thereby ensuring in building a communication network that can handle the transmission of the huge volume of data that are free from any transmission errors or distortions.