Machine Learning-Based Adaptive Receive Filtering: Proof-Of-Concept On An Sdr Platform

The constant demand for low latency and high data rates in a modern mobile communications network creates new scientific challenges in each new generation. An accurate reconstruction of transmission data of as many users as possible at the base station is a task that can become mathematically very complex in today's mobile networks. The scientific community and our group at Fraunhofer Institute for Telecommunications (Heinrich Hertz Institute, HHI) are therefore investigating and comparing new methods such as NOMA with conventional orthogonal multi-user recognition techniques such as OFDMA. In combination with a large number of antennas on the receiving base station, a practical implementation can become very complex. In addition, subsequent errors in estimation with methods such as successive interference cancellation (SIC) can make it impossible for some users to reconstruct the data reliably if a large number of users are affected by incorrectly estimated parameters (such as user channels, covariance matrix, noise variance, etc.). In our demonstrator, we show how a receiver based on machine learning (ML) can help with this task. With low complexity, the modulation symbols for each user are detected without the need to estimate the typical parameters. All users send their pilots and data alternately on the same resources (frequency and time). The regular pilot phases are used to relearn changes in the channel. The method has the potential to contribute significantly to latency reduction through parallel implementation.