Implementation of KDML-Based Channel Estimator on GNU Radio Platform

Abstract

Deep learning (DL) technology has been applied to a number of thorny problems in the field of communication, among which channel estimation is always attracting interests. However, DL is often used as an independent channel estimator which have ignored the effective role of traditional estimation methods for practical communication system. Recently a KDML-based channel estimator is proposed which combines traditional channel estimation and DL methods and obtains improved performance in simulation experiments. This paper verifies KDML-based channel estimator on a software-defined radio (SDR) platform. To do this, we design a KDML-based channel estimator framework on GNU radio platform and build an orthogonal frequency division multiplexing (OFDM) system containing learning modules. This framework adopts the long short time memory (LSTM) network to fine-tune the results of the least squares (LS) estimator according to KDML. We collect the channel state information (CSI) on GNU radio platform to train the LSTM network offline and reload the LSTM network back to the GNU radio platform to perform the online channel estimation. The simulation results and air-interface tests both validate the performance of the KDML and effectiveness of our proposed framework.