An Efficient Deep Neural Networks-Based Channel Estimation and Signal Detection in OFDM Systems

Abstract

This article represents an improvement in the efficiency of Deep Neural Networks (DNN) employed in OFDM systems for estimation of channel, detection of symbols by using Hyperparameter tuning. Deep learning is increasing its popularity in each field and it’s a set of Artificial Intelligence (AI). Deep Neural Network (DNN) model is completely different from traditional Orthogonal Frequency Division Multiplexing (OFDM) receiver in such a way that OFDM receiver has to estimate the channel state information separately then it is able to detect the transmitted signal, but DNN model will be trained offline by simulated data over various channel statistics then directly employed in online for detection of the transmitted signal directly. Deep learning gives a performance better than traditional methods, i.e., Minimum Mean Square Error (MMSE) of channel estimation and symbol detection in wireless communications. First, we have analyzed the performance of traditional methods, i.e., MMSE over Winner II channel without deep learning by observing BER with cyclic prefix and without Cyclic Prefix (CP) and with variation in number of pilot carriers. Thereafter we will deploy hyperparameter tuning with proposed DNN model in OFDM for estimation of channel and detection of symbols in the winner II channel mainly for the Urban power delay profile using QPSK modulation. Here, we have claimed that by reducing the no of epochs from 20,000 to 10,000, training period will be reduced from 52 to 26 h, i.e., 50% while maintaining the almost same BER performance.