Deep learning based channel estimation in massive MIMO systems with phase noise

Abstract

In this thesis, channel estimation in Massive MIMO system with phase noise is discussed. As required data traffic increases, the technology with higher throughput and wider bandwidth were needed. Massive MIMO which can achieve high spectral density became one of the promising solutions and mmWave was adopted to increase the bandwidth. To efficiently use Massive MIMO, acquiring accurate channel state information is needed. So, channel estimation methods with high accuracy and low complexity have been widely researched. Phase noise comes from an imperfect oscillator, it has time-varying variance unlike white noise and becomes severe as operating frequency increases. In a modern wireless communication system, the importance of compensating phase noise is rising as the number of antennas and operating frequency increase. To be specific, the importance of intercarrier interference (ICI) from the MIMO system is growing. Unlike common phase error, ICI has complex problems which make it challenging to cancel. Thus, many types of research have been done to effectively remove ICI. Both massive MIMO channel estimation and phase noise cancellation are analytically intractable, so there are no general solution and existing solutions are sub-optimal. In this thesis, to deal with both problems, we use deep learning. Our proposed method uses deep learning model called Pix2Pix which is based on conditional generative adversarial model (GAN) and widely used in image-to-image translation problems. Since the channel is similar to image due to the fact that it can be interpreted as 2 dimensions, we used Pix2Pix which is widely used in image processing. As a result, we confirmed that the performance of our method in massive MIMO with phase noise outperforms the conventional method, and we analyzed the performance of our model by changing parameters.