Gridless Variational Direction-of-Arrival Estimation in Heteroscedastic Noise Environment

Qi Zhang, Jiang Zhu, Yuantao Gu, Zhiwei XuIEEE Journal of Oceanic Engineering 2021 · OpenReview

TL;DR

This paper introduces a novel gridless variational Bayesian approach (MVHN) for estimating the direction of arrival (DOA) of signals in challenging underwater environments characterized by heteroscedastic noise. The proposed method effectively handles noise that varies across both time (snapshots) and space (antennas), a common issue in real-world scenarios, and automatically estimates crucial parameters like the number of sources and noise variance.

What Problem Does This Paper Address?

Estimating the direction of arrival (DOA) of signals is crucial in many applications, especially in underwater acoustics where signals can be weak and the environment is noisy. Traditional methods like conventional beamforming (CBF) are simple but suffer from poor resolution (wide beams and high sidelobes). High-resolution methods like MUSIC and MVDR offer better resolution but are sensitive to signal mismatches and require prior knowledge of the number of sources. A significant challenge arises from heteroscedastic noise, where the noise characteristics (variance) are not constant across sensors or time. This non-stationarity severely degrades the performance of existing DOA estimation techniques. This paper specifically tackles DOA estimation in environments where noise variance varies across both snapshots and antennas, a scenario not well-addressed by prior work.

Methods

The core contribution is the development of the Multi-Snapshot Variational Line Spectral Estimation dealing with Heteroscedastic Noise (MVHN) algorithm. This method frames the DOA estimation problem within a Bayesian framework, treating DOAs as random variables. Key aspects of the approach include:

  • Variational Bayesian Inference: Instead of intractable exact posterior calculations, MVHN uses variational inference to approximate the posterior distribution of the unknown parameters. This allows for efficient estimation.
  • Gridless Approach: Unlike methods that discretize the possible DOA space into grids, MVHN treats frequencies (which are directly related to DOAs) as continuous parameters, avoiding the "grid mismatch" problem and offering higher resolution.
  • Handling Heteroscedastic Noise: The algorithm explicitly models and estimates the noise variance that varies across both antennas and snapshots (Case IV).
  • Variants for Simpler Noise Models: The paper also derives specialized versions, MVHN-S (noise varying only across snapshots) and MVHN-A (noise varying only across antennas), which are more efficient for these specific scenarios.
  • Automatic Parameter Estimation: MVHN automatically estimates the number of sources, noise variance parameters, and prior distribution parameters, removing the need for manual tuning.
  • Posterior Distribution Output: Instead of just point estimates, MVHN provides the posterior probability distribution of the DOAs, offering a measure of uncertainty.

Datasets

The paper evaluates its proposed algorithms using simulated data generated according to four defined noise cases (Case I: wide-sense stationary, Case II: stationary across antennas, varying across snapshots, Case III: stationary across snapshots, varying across antennas, Case IV: heteroscedastic across both). Additionally, a real experimental acoustic dataset collected at sea is used for validation. This dataset involved horizontal linear arrays deployed on the seafloor, with a towed acoustic source.

Key Insights and Ideas

  • Heteroscedastic noise is a critical factor: The paper strongly demonstrates that ignoring heteroscedastic noise leads to significant performance degradation, particularly in terms of DOA estimation accuracy and bias.
  • MVHN offers high resolution and robustness: The proposed MVHN and its variants (MVHN-S, MVHN-A) consistently outperform traditional methods like CBF, especially in resolving closely spaced targets.
  • MVHN's bias issue: While MVHN performs well in estimating noise variances, its direct application to Case IV can lead to biased DOA estimates due to the absence of noise averaging over snapshots. This highlights the importance of using the variants (MVHN-S/A) when applicable.
  • Unified framework for noise models: The paper shows that MVHN-S and MVHN-A can be derived from the general MVHN, demonstrating a unified approach to handling different noise heteroscedasticity scenarios.
  • Uncertainty quantification is valuable: Providing the posterior distribution of DOAs is crucial for understanding the reliability of the estimates, especially in challenging noise conditions.

Key Results

  • High Resolution: MVHN-S and MVHN-A successfully resolve closely spaced targets where CBF fails.
  • Performance Improvement: In simulations, MVHN-S and MVHN-A generally outperform MVALSE (a related method for homogeneous noise) and other SBL-based approaches, especially in terms of signal reconstruction error and model order estimation probability.
  • MVHN-S and MVHN-A Preference: For Case II (noise varying across snapshots), MVHN-S is preferred. For Case III (noise varying across antennas), MVHN-A is preferred. For Case IV (heteroscedastic noise), MVHN-S or MVHN-A are recommended over MVHN due to MVHN's bias.
  • Computational Complexity: MVHN and its variants offer a good balance between performance and computational complexity, generally being faster than some SBL-based methods.
  • Real Data Validation: The real data analysis confirms the high resolution of MVHN-A, showing it can accurately estimate the DOA of a towed source and distinguish it from other acoustic signals.

Limitations and Future Work

  • MVHN Bias: The paper notes that MVHN can produce biased DOA estimates in Case IV due to the lack of noise averaging. This suggests that the variants (MVHN-S/A) are generally preferred.
  • Computational Cost: While more efficient than some SBL methods, MVHN still has a non-trivial computational complexity, especially for large numbers of antennas and snapshots.
  • Future Work: The paper implicitly suggests further exploration into optimizing the computational efficiency and potentially refining the handling of complex noise scenarios. The authors also mention that the Cramér-Rao Bound (CRB) is derived for performance comparison, indicating its importance in evaluating future algorithms.

Summary updated

Abstract

Horizontal line arrays are often employed in underwater environments to estimate the direction of arrival (DOA) of a weak signal. Conventional beamforming is robust but has wide beamwidths and high-level sidelobes. High-resolution methods, such as minimum-variance distortionless response and subspace-based MUSIC algorithm, produce low sidelobe levels and narrow beamwidths, but are sensitive to signal mismatch, and require many snapshots and the knowledge of number of sources. In addition, heteroscedastic noise (HN) where the variance varies across observations and sensors due to nonstationary environments degrades the conventional methods significantly. This article studies DOA in an HN environment, where the variance of noise is varied across the snapshots and the antennas. By treating the DOAs as random variables and the nuisance parameters of the noise variance different across the snapshots and the antennas, multisnapshot variational line spectral estimation dealing with HN (MVHN) is proposed, which automatically estimates the noise variance, nuisance parameters of the prior distribution, and number of sources, and provides the uncertain degrees of DOA estimates. When the noise variance only varies across the snapshots or the antennas, the variants of MVHN, i.e., MVHN-S and MVHN-A, can be naturally developed. Finally, substantial numerical experiments are conducted to illustrate the proposed algorithms’ performance, including a real data set in a DOA application.
BibTeX
@article{zhang2021gridless,
  author = {Qi Zhang and Jiang Zhu and Yuantao Gu and Zhiwei Xu},
  title = {{Gridless Variational Direction-of-Arrival Estimation in Heteroscedastic Noise Environment}},
  journal = {IEEE Journal of Oceanic Engineering},
  year = {2021},
  url = {https://openreview.net/forum?id=sjEHEjLZlx},
}
Figure