Classification Algorithms and Deep Embedded Learning for Solving Sparse Inverse Linear Problem in Massive MIMO Systems

摘要

In massive MIMO systems, base stations (BSs) may deploy a large number of antennas. When the system operates in a frequency-division-duplex (FDD) mode, the extensive number of antennas poses a significant challenge to the channel estimation in the downlink of the communication system. One option is to explore the possible underlying channel structure whereby the high-dimensional channel vector admits a low-dimensional (sparse) representation in some basis or dictionary. In this paper, we investigate the possibility to provide compressed (sparse) channel representation by performing clustering techniques for complex data elements. For this purpose we evaluate the performance of two algorithms: The first one is an adaptation of the classic k-means algorithm. The second one is a non-parametric method based on deep embedded clustering. The compressed channel representation may be seen as an alternative to methods based on compressed channel feedback, with the difference that the channel compression takes place directly in the domain of the low-dimensional channel representation.