A Denoising Autoencoder based wireless channel transfer function estimator for OFDM communication system Tomohisa Wada 1) , Takao Toma 1) , Mursal Dawodi 2) , Jawid Baktash 2) 1) Dept. of Information Engineering, University of the Ryukyus, Senbaru 1, Nishihara, Okinawa, Japan 2) Graduate School of Engineering and Science University of the Ryukyus, Senbaru 1, Nishihara, Okinawa, Japan Abstract— This paper proposes a channel estimation method for Orthogonal Frequency Division Multiple Access (OFDM) communication system by utilizing a Neural Network (NN) based a Machine Learning (ML). Especially, Autoencoder is utilized to estimate Channel Transfer Function (CTF) and to reduce a noise on the estimate. Japanese Digital TV broadcast system is assumed as target system. Then 8k FFT/IFFT is used and number of sub- carriers are 5617 such as mode3 in Integrated Services Digital Broadcasting-Terrestrial (ISDB-T) spec. 5617 complex CTF points must be estimated by limited number of scattered pilot sub- carriers. Assumed channel condition is 2 wave multipath channel with Additive White Gaussian Noise (AWGN). The multipath parameters are randomly generated. To train the autoencoder, 5000 CTFs are generated and pre-training was performed. System performance was evaluated by measuring Bit Error Rate (BER). The system with conventional frequency-domain interpolator and the system with autoencoder based were compared. According to BER simulation results, the autoencoder based system has shown lower BER than the conventional. At BER=10 -5 , autoencoder system shows roughly 2dB gain than conventional system. Keywords—Neural Network, Machine Learning, Deep Learninig, Autoencoder, OFDM, Channel Estimation, Denoise I. INTRODUCTION The machine learning (ML) and, in particular, deep learning (DL) applications has growing rapidly in the last decade. The application fields cover almost every industrial areas [1, 2, 3]. Digital communication related applications such as channel coding, channel decoding, detection, MIMO detection, deep learning communication system are also investigated by many researchers [4, 5, 6, 7]. For example, a paper [8] proposed a channel estimation application for OFDM communication system and it shows the advantage of DL systems using 64 sub- carries OFDM system. In this article, we introduce a channel estimation method utilizing NNs. In order to show the usefulness of NNs, one big real OFDM system is assumed such as Integrated Services Digital Broadcasting-Terrestrial (ISDB-T) system, which is used as Digital TV service in Japan, the Philippines, Latin America. Fig. 1 shows the simple diagram of the OFDM communication system. The upper side is the transmitter side and the lower side is the receiver side. Bit information are mapped to constellations through 64QAM mapper and 5617 constellations with 2575 zeros are 8192 points IFFTed as OFDM modulation [9]. In order to avoid a Inter Symbol Interference (ISI), Guard Interval (GI) is pre-attached at each OFDM symbol as Cyclic prefix (CP) manner. Transmitted radio wave go through multipath channel. Then the receiver basically performs reverse order operation. To mitigate a distortion caused by the multipath channel, channel transfer function (CTF) must be estimated and the FFT outputs are divided by the CTFs. By the scatter pilot (SP), the CTF at SP position can be measured. To estimate all CTF values, Time-domain and Frequency-domain interpolation is required as shown in the figure. The proposed system replaces the frequency-domain interpolator with autoencoder. The section II describes the system architecture including the system block diagram with and without autoencoder. The detail of the system operation will be described. The section III shows computer simulation results. Finally, the summary is concluded in section IV. II. SYSTEM AECHITECTURE Fig. 2 shows a block diagram of receiver’s equalizer. The upper (a) is a conventional method and the lower (b) is the proposed autoencoder channel estimator. Fig.3 show the Time- Frequency representation of OFDM signals. The blue circles are scatter pilots (SPs), which is BPSK modulated, and receiver knows the pilot value. Then receiver can estimated the CTF values at SP positions. In order to estimate all CTF values at all circle points, first time-domain interpolation is performed with 15 tap FIR filter. Then after, by performing frequency-domain interpolation, all CTF values can be estimated. An autoencoder is a neural network that is trained to attempt to copy its input to its output. In addition, autoencoder has capability of denoising which is reported [10, 11 Chapter14.2.2]. When some inputs are dropout, autoencoder tries to estimate the Fig. 1: OFDM communication system. 978-1-5386-7822-0/19/$31.00 ©2019 IEEE ICAIIC 2019 530