Introduction
In the evolving landscape of millimeter-wave (mmWave) MIMO systems, lens antenna arrays have emerged as a compelling technology. Their ability to focus electromagnetic energy onto specific antenna elements—a property known as energy-focusing—allows for efficient path division. However, this hardware advantage introduces unique signal processing challenges, particularly when multiple signal sources have similar incident angles. This paper, published in IEEE Access, proposes a robust three-stage algorithm to achieve high-resolution Direction-of-Arrival (DoA) estimation in these complex scenarios.
The Challenge: Overlapping Signal Footprints
Lens antenna arrays differ fundamentally from traditional Uniform Linear Arrays (ULAs). While ULAs exhibit a "Vandermonde" structure in their array response, lens antennas follow a "sinc" function distribution. Ideally, each signal's DoA maps to a distinct group of antennas, as shown in the following diagram where signal paths are clearly separated.
However, when devices are spatially close, their incident angles become nearly identical, causing their energy "footprints" on the array to overlap. Standard subspace-based estimation methods like Root-MUSIC are not directly applicable because they rely on the Vandermonde structure. When footprints overlap, separating the signals becomes mathematically difficult using conventional tools. This scenario is illustrated below, where signals from different angles hit nearly the same region of the array.
The Proposed Solution: A Three-Stage Pipeline
The authors address this by combining specialized signal processing with machine learning techniques in a three-step pipeline:
- Modified Root-MUSIC: The researchers developed a mathematical transformation (Theorem 1) that re-expresses the lens array's sinc response as a polynomial. This effectively "tricks" the system into a form compatible with the Root-MUSIC framework, allowing for the generation of candidate roots representing potential DoAs.
- Outlier Detection: Unlike standard Root-MUSIC, which selects roots closest to the unit circle, lens arrays produce a more complex root distribution. The authors employ the Local Outlier Factor (LOF) algorithm—an unsupervised outlier detection method—to identify and prune "junk" roots caused by noise or mathematical artifacts.
- K-medoids Clustering: The remaining high-quality candidate roots are grouped into $L$ clusters (where $L$ is the number of signal sources). The center of each cluster, determined by the K-medoids algorithm, provides the final DoA estimate.
Performance and Results
The algorithm was validated through extensive numerical simulations involving a lens antenna array with 41 elements and up to 14 uncorrelated narrowband sources. Key findings include:
- High Accuracy with Few Snapshots: The method maintains a low Root Mean Square Error (RMSE) even with as few as 10 snapshots, a critical requirement for low-latency mmWave communications.
- Robustness in Dense Scenarios: In environments with many sources ($L=14$), the proposed method significantly outperformed simple energy-based selection and "fast" Root-MUSIC variants. It proved that "compressing" equations to save time (as fast variants do) loses critical information in dense environments.
- Resilience to Noise: By treating root selection as a data cleaning problem (using LOF and K-medoids), the system is far more resilient to noise than purely mathematical solvers.
Conclusion
By bridging the gap between non-standard hardware responses and powerful subspace estimation tools, this research provides a viable path for high-performance DoA estimation in mmWave systems. The integration of machine learning for outlier detection and clustering transforms a sensitive mathematical problem into a robust engineering solution, capable of handling the dense signal environments typical of next-generation wireless networks.
Summary updated