Direction of arrival estimation with neural networks via test time self-supervised optimization and array manifold...

Area of Science:

• Signal Processing
• Array Signal Processing
• Machine Learning for Signal Processing

Background:

• Adaptive beamformers like Minimum Variance Distortionless Response (MVDR) are crucial for signal processing but sensitive to errors.
• Mismatches in the sample covariance matrix (SCM) and array steering vectors degrade performance.
• Existing robust MVDR methods often rely on specific assumptions or pre-trained models.

Purpose of the Study:

• To develop a robust direction-of-arrival (DOA) estimation framework using an adaptive beamformer.
• To introduce a closed-loop Neural-MVDR system that enforces distortionless constraints with a parameterized array model.
• To enable self-supervised, per-frame adaptation at test time, eliminating the need for offline pretraining.

Main Methods:

• Proposes a closed-loop Neural-MVDR framework for DOA estimation.
• Enforces a distortionless constraint aligned with a physically parameterized array model.
• Employs per-frame, self-supervised adaptation, alternating between steering-vector refinement, robust SCM reconstruction via a neural module, and analytical MVDR spectral estimation.

Main Results:

• Demonstrates improved robustness in DOA estimation compared to conventional beamformers and MVDR variants.
• Achieves enhanced performance on both synthetic data and real-world SWellEx-96 S5 event data.
• Outperforms existing robust MVDR techniques including eigenspace suppression, covariance matrix tapering, and shrinkage methods.

Conclusions:

• The proposed Neural-MVDR framework offers a robust and adaptive solution for DOA estimation.
• Self-supervised, real-time adaptation significantly enhances performance without requiring external labeled datasets.
• This approach provides a promising direction for improving adaptive beamforming in practical scenarios with model uncertainties.