A Machine Learning Approach to Blind Multi-Path Classification for Massive MIMO Systems

Abstract

This paper concerns the problem of the multi-path classification in a multi-user multi-input multi-output (MIMO) system. We propose a machine learning approach to achieve a blind multi-path classification in the uplink (UL) scheme of a multi-user massive MIMO system. Note that the “blind” term in our approach represents the achievement of the multi-path classification without different pilot sequences on different users, without prior channel state information (CSI) at each user, and without any exploiting the special properties of the received signal. Specifically, our approach consists of two phases. In the first phase, multiple users transmit communication-requests to the base station (BS) for message transmission. The BS only estimates the scaled large-scale path loss of each user, which is determined by the distance between the transmitter and the receiver and is independent of the number of multi-path. Then, the BS compares the difference of the scaled large-scale path loss between any two users. If the difference is sufficiently large, the BS notifies all users to permit their simultaneous message transmissions. However, if the difference is small, the BS notifies each user to slightly adjust their transmission power and then permits their simultaneous message transmissions as well. In the second phase, all users simultaneously transmit their messages using the same radio resource. Then the BS selects one predetermined constellation point from the received pilot symbols as the input of clustering algorithms. According to the clustering results, the BS classifies each multi-path into a specific user. The key intuition of our approach is that the clustering algorithms can generate multiple cluster centroids and each cluster centroid represents the average reception power of each user. Moreover, we use a weighted ensemble clustering algorithm to further improve the performance of our approach. We implemented our approach and conducted extensive performance comparison. Our experimental results show that, when the received signal-to-noise ratio (SNR) is more than 13 dB, our approach with the weighted ensemble clustering algorithm can correctly classify all multi-path to the corresponding user and the output SNR can be improved by 3.2 dB, where we consider three users in an 8PSK system and each user possess 50 multi-path.