This paper presents a machine learning–driven localization framework that leverages Angle of Arrival (AoA) as a robust feature extracted from massive MIMO OFDM channel state information (CSI). The goal is to enable accurate user or device localization in complex outdoor environments, where multipath propagation, obstacles, and varying line-of-sight (LoS) conditions challenge conventional methods. The proposed system integrates high-resolution AoA estimation techniques—MUSIC and ESPRIT—into a two-stage hierarchical classifier. As illustrated in Figure 1, the first stage distinguishes between LoS and NLoS regions using a binary classifier. Once the region is identified, a second-stage multi-class classifier performs fine-grained trajectory (track) identification within that region. This structure allows each classifier to operate on more homogeneous data, improving both robustness and accuracy. AoA features are extracted from a real-world 64-antenna massive MIMO dataset collected at the Nokia campus in Stuttgart. The system processes CSI through sliding windows, producing AoA vectors of length 200 (4 rows × 50 subcarriers), which serve as model inputs. MUSIC and ESPRIT generate nearly identical AoA estimates (e.g., −35° for Track 11, Figure 4), but ESPRIT demonstrates far lower computation time, making it appealing for real-time operation.
A variety of machine learning models—including logistic regression, KNN, Random Forest, GBM, XGBoost, LightGBM, and stacking ensembles—are evaluated. The first-stage LoS/NLoS classifier reaches 100% accuracy across all models, underscoring the discriminative power of AoA features. In the second stage, the hierarchical approach outperforms a single multi-class baseline, with the top-performing stacking ensemble achieving 98% accuracy in LoS regions and 95% in NLoS regions. The study highlights that AoA, when combined with hierarchical machine learning, is a strong candidate for high-reliability outdoor localization, even under challenging propagation conditions. The authors also note potential extensions toward joint azimuth/elevation AoA processing, deep learning approaches, and integration with physical-layer authentication frameworks.
High-accuracy AoA-based Localization using Hierarchical ML Classifiers in Outdoor Environments