Design and Implementation of Deep Neural Network (DNN) using Bayesian Regularization for OFDM Channel Estimation

Authors: Sachin Rathor, Kapil Sahu

Certificate: View Certificate

Abstract

Orthogonal Frequency Division Multiplexing is a multi carrier system which owing to its spectral efficiency has evolved as the primary solution to high speed data networks. The fundamental problem still lies in the fact that wireless channels exhibit frequency selectivity thereby rendering high bit error rate (BER) to the system. The present work presents a technique used for deep learning based on the Bayesian Regularized Deep Neural Network (BRDNN) for channel estimation of an OFDM based network. The performance is evaluated based on the mean square error found in channel estimation. Moreover the number of epochs has also been considered as an evaluation parameter for judging the performance of the system. It is found that the proposed system attains a mean squared error of 0.25% and a BER of 10-4. It has been observed that the variation in the number of pilots results in a variation in BER of the system.

Introduction

Artificial Neural Networks have made their presence felt in several applications where finding relations and patterns is complex in nature. One such application is that of finding channel nature for computer and wireless networks. For such a time varying channel response, an artificial neural network consisting of multiple hidden layers needs to be used. These days, Orthogonal Frequency division multiplexing (OFDM) has become an effective multiplexing technique for several applications. The need for channel estimation lies in the fact that data transmission is computer networks are prone to errors arising out of the nature of the wireless channels. Networks such as LANs, MANs or WANs tend to have multiple devices connected to each other, which may move with time. If the nature of the channel is known, then corrective measures for circumventing the errors can be deigned.[2] However, this is a challenging task as wireless channels keep changing their nature over time. Moreover, spectrally efficient techniques such as OFDM tend to transmit data in narrowly spaced channels. Hence it is extremely difficult to find out the exact nature of the channel by analyzing the patterns existing in the input and output data streams. [3]The match between the input and output data streams need to be done. Let the input data stream be X and the output data stream be Y. The relation between input X to the channel and output Y of the channel decides the channel response ‘H’ Deep neural network (DNN) is a special category of artificial neural networks specially designed for analyzing complex data patterns. Still it is challenging to keep the error low and accuracy high for the OFDM based channel estimation.