Domain Adaptation for DoA Estimation in Multipath Channels With Interferences

Amitay Bar, Joseph S. Picard, Israel Cohen, Ronen Talmon

TL;DR

This paper introduces a "plug-and-play" domain adaptation layer that cleans up signal measurements before they reach Direction of Arrival (DoA) estimation algorithms. By mapping noisy, interference-heavy real-world signals to a clean, simulated "free-space" reference domain, the authors significantly boost the accuracy of classic beamformers like MUSIC and MVDR without requiring any labeled training data.

What Problem Does This Paper Address?

Estimating where a signal is coming from (DoA) is a foundational task in everything from 5G networking to autonomous vehicles. However, two "villains" usually ruin the party: multipath interference (signals bouncing off walls and creating echoes) and interfering sources (other transmitters masking the desired signal).

Current solutions often fall into two camps:

  1. Classical Signal Processing: These methods are mathematically elegant but often struggle or fail entirely when interference is stronger than the desired signal.
  2. Deep Learning: While powerful, these models require massive amounts of "labeled" data (signals where the true DoA is already known) and are often computationally expensive.

The authors sought a middle ground: a low-complexity, unsupervised method that can "learn" the environment on the fly and suppress interference without needing a pre-recorded dataset.

Methods

The core innovation is a linear mapping function inspired by Domain Adaptation (DA) in machine learning—specifically a technique called CORAL (Correlation Alignment).

  1. The Reference Domain: Before deployment, the researchers use a simple "free-space" model to analytically generate "steering vectors" (how a signal should look) for a specific Region of Interest (ROI). They calculate a mean correlation matrix ($\Sigma_S$) for this ideal world.
  2. The Operational Domain: Once the array is deployed, it enters an "adaptation phase" where it collects unlabeled signals from the actual, messy environment. It calculates a mean correlation matrix ($\Sigma_A$) for this real world.
  3. The Map: The authors compute a mapping matrix $E = \Sigma_S^{1/2} \Sigma_A^{-1/2}$. This matrix effectively "transports" signals from the messy operational domain to the clean reference domain.
  4. Estimation: Any standard DoA algorithm (Delay-and-Sum, MVDR, or MUSIC) is then applied to the mapped signals rather than the raw ones.

From a geometric perspective, this process is analogous to parallel transport on the manifold of Hermitian Positive Definite (HPD) matrices, essentially sliding the signal's statistical properties along a curved mathematical surface to a "cleaner" location.

Datasets

The authors evaluated their approach using high-fidelity simulations for both acoustic and Radio Frequency (RF) scenarios:

  • Acoustic Simulation: A 9-microphone array in a reverberant room (simulated using the "image method"). They tested various reverberation times ($\beta$) to see how the system handled different levels of "echoiness."
  • RF Simulation: A 9-element antenna array operating at 2.4 GHz. They simulated complex Gaussian noise and narrowband signals to test the limits of interference rejection.
  • Benchmarks: The method was tested against standard implementations of DS, MVDR, and MUSIC beamformers.

Key Insights and Ideas

  • Unsupervised Learning: The "adaptation phase" doesn't need to know where the signals are coming from. It just needs to observe the environment. This makes it highly practical for real-world deployment where labeled data is non-existent.
  • Implicit Null Steering: One of the "aha" moments is that the mapping matrix $E$ implicitly creates "nulls" (dead zones) in the directions of interfering sources. It learns to ignore the parts of the environment that don't match the simulated free-space model.
  • Mathematical Bridge: The paper elegantly connects machine learning (Domain Adaptation) with differential geometry (Riemannian manifolds), showing that aligning second-order statistics is a powerful way to handle environmental mismatch.
  • Plug-and-Play: Because the output is just a transformed signal vector, this method can be bolted onto almost any existing DoA algorithm that relies on a sample correlation matrix.

Key Results

  • Massive Error Reduction: In scenarios with heavy interference (Signal-to-Interference Ratio of -20dB), standard beamformers often failed, showing median errors of ~14°. With the proposed adaptation, the median error dropped to under 3°.
  • Robustness to Echoes: As reverberation time increased (making the room "noisier"), the performance gap between the proposed method and standard methods widened, proving its effectiveness in highly multipath environments.
  • Approaching the Limit: 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 of how accurate an estimator can be).

Limitations and Future Work

  • Region of Interest (ROI): The method requires prior knowledge of the general area where the desired source is located. If the source moves outside this pre-defined ROI, the simulated steering vectors would need to be regenerated.
  • Narrowband vs. Wideband: The current paper focuses primarily on narrowband signals (single frequency). While they mention extensions to wideband (like speech), a full implementation across multiple frequency bins remains a task for future research.
  • Stationarity: The approach assumes the interference environment is relatively stable between the adaptation phase and the estimation phase. Rapidly changing environments (e.g., a fast-moving interferer) might require more frequent re-adaptation.

Summary updated

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 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, 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).
BibTeX
@article{bar2025domain,
  author = {Amitay Bar and Joseph S. Picard and Israel Cohen and Ronen Talmon},
  title = {{Domain Adaptation for DoA Estimation in Multipath Channels With Interferences}},
  journal = {IEEE Transactions on Signal Processing},
  year = {2025},
  doi = {10.1109/tsp.2025.3577257},
}

Loading related content...

Figure