Deep Learning Based Fingerprint Positioning for Multi-Cell Massive MIMO-OFDM Systems

In this paper, we investigate user terminal (UT) fingerprint positioning for multi-cell massive multiple-input multiple-output orthogonal frequency division multiplexing (MIMO-OFDM) systems in non-line-of-sight scenario. We first introduce a refined double beam-based channel model to the positioning problem and extract the energy coupling matrix in the refined beam domain as location-related fingerprint. By taking advantage of refined spatial and frequency beams, the new fingerprint (i.e., energy coupling matrix) contains plentiful and stationary multi-path information, such as power, angle of arrival (AoA), and delay of arrival (DoA), which are favorable to positioning. We then propose a novel deep learning-based fingerprint positioning method to locate the 2-dimension (2D) position of UTs, utilizing the multi-cell fingerprint as the input. In particular, we propose two deep neural network (DNN) architectures in this paper. The first DNN consists of convolutional neural networks (CNN) and multilayer perceptron, refining the elementary features from the multi-cell fingerprints and dividing the UTs into their belonging categories. In the second DNN, a new network architecture, Transformer, is applied to the positioning problem. The Transformer is directly based solely on self-attention mechanisms to sequences of fingerprint patches, which can perform outstandingly on the 2D position coordinates regression. Simulation results show that the proposed positioning methods can outperform the existing methods in terms of positioning error.