Abstract
We consider the problem of estimating the direction-of-arrival (DoA) of a desired source located in a known region of interest in the presence of interfering sources and multipath. We propose an approach that precedes the DoA estimation and relies on generating a set of reference steering vectors. The steering vectors' generative model is a free space model, which is beneficial for many DoA estimation algorithms. The set of reference steering vectors is then used to compute a function that maps the received signals from the adverse environment to a reference domain free from interfering sources and multipath. We show theoretically and empirically that the proposed map, which is analogous to domain adaption, improves DoA estimation by mitigating interference and multipath effects. Specifically, we demonstrate a substantial improvement in accuracy when the proposed approach is applied before three commonly used beamformers: the delay-and-sum (DS), the minimum variance distortionless response (MVDR), and the Multiple Signal Classification (MUSIC).
Introduction
Estimating the Direction of Arrival (DoA) is a foundational task in signal processing, essential for applications ranging from 5G telecommunications to autonomous vehicle navigation. However, real-world environments present two significant challenges: multipath interference (echoes from surfaces) and interfering sources (competing signals). Classical signal processing methods often struggle when interference is stronger than the desired signal, while modern deep learning approaches require massive amounts of labeled data and high computational power.
This paper proposes a "plug-and-play" domain adaptation layer that cleans signal measurements before they reach DoA estimation algorithms. By mapping noisy, interference-heavy real-world signals to a clean, simulated "free-space" reference domain, we significantly boost the accuracy of classic beamformers without requiring labeled training data.
Methodology: Domain Adaptation via Correlation Alignment
The core innovation is a linear mapping function inspired by Domain Adaptation (DA) techniques, specifically Correlation Alignment (CORAL). The process follows four key steps:
- Reference Domain Generation: Before deployment, a "free-space" model analytically generates steering vectors for a specific Region of Interest (ROI), yielding a mean correlation matrix ($\Sigma_S$) for an ideal environment.
- Operational Domain Adaptation: During an initial "adaptation phase," the deployed array collects unlabeled signals from the actual environment to calculate a mean correlation matrix ($\Sigma_A$).
- The Mapping Function: A mapping matrix $E = \Sigma_S^{1/2} \Sigma_A^{-1/2}$ is computed. This matrix effectively "transports" signals from the messy operational domain to the clean reference domain. Geometrically, this is analogous to parallel transport on the manifold of Hermitian Positive Definite (HPD) matrices.
- Estimation: Standard DoA algorithms (DS, MVDR, or MUSIC) are applied to the mapped signals rather than the raw measurements.
One of the primary advantages of this mapping is implicit null steering. The mapping matrix $E$ learns to ignore environmental components that do not match the simulated free-space model, effectively creating "dead zones" in the directions of interfering sources.
Experimental Results
The proposed method was evaluated using high-fidelity simulations for both acoustic (9-microphone array in a reverberant room) and Radio Frequency (RF) scenarios (9-element antenna array at 2.4 GHz).
Robustness to Reverberation
In acoustic settings, the system was tested against varying reverberation times ($\beta$). As the room becomes more "echoic," standard beamformers see a sharp rise in estimation error. In contrast, the proposed adaptation maintains significantly higher accuracy.
Interference Rejection
In scenarios with heavy interference (Signal-to-Interference Ratio of -20dB), standard beamformers often failed, showing median errors of approximately 14°. With the proposed adaptation, the median error dropped to under 3°. In RF settings, as the Signal-to-Noise Ratio (SNR) improved, the proposed method's accuracy closely approached the Cramér-Rao Bound, the theoretical physical limit for estimator accuracy.
Conclusion
This unsupervised, low-complexity approach provides a powerful bridge between classical signal processing and machine learning. By aligning second-order statistics, the method allows existing DoA algorithms to function effectively in environments for which they were not originally designed. Future work will focus on extending this framework from narrowband to wideband signals and addressing scenarios with rapidly changing interference environments.
Summary updated