Abstract

In this paper, in order to overcome the channel distortion due to doubly selective channel, a new deep learning-based channel estimation scheme is proposed. In the proposed scheme, two neural networks are combined to learn both the temporal and spectral characteristics of the channel. One is the AE (autoencoder) which is trained to track the channel characteristic in the time domain and the other is the LSTM (long short term memory) which is trained to track the channel characteristic in the frequency domain. Firstly, the AE is trained and then the LSTM is sequentially trained based on the AE output. After that, the trained neural network is used to the conventional DPA (data pilot aided) process to estimate the channel values in frequency domain. It is noted that the proposed LSTM network consists of fixed number of LSTM units, so it enables to track the temporal variation of the channel reliably, regardless of the position of the OFDM (orthogonal frequency division multiplexing) symbol in a frame. Therefore, the proposed scheme can enhance the PER (packet error rate) performance more by reducing the error propagation due to the DPA process. Through numerical results, it is confirmed that the proposed scheme shows the best PER performance of the conventional schemes in doubly selective channel environment.