A Cascaded Neural Network for Robust Phase-Only Beamforming Under Covariance Matrix Mismatch

Area of Science:

• Signal Processing
• Machine Learning
• Array Signal Processing

Background:

• Covariance matrix mismatch degrades beamforming performance.
• Finite-snapshot estimation errors and signal contamination are key challenges.
• Conventional methods often require explicit uncertainty modeling.

Purpose of the Study:

• To develop a robust phase-only beamforming framework.
• To mitigate performance loss due to covariance matrix errors.
• To enable data-driven covariance reconstruction and phase prediction.

Main Methods:

• A cascaded neural network combining a denoising autoencoder (DAE) and a residual network (ResNet).
• DAE reconstructs ideal covariance representations and extracts features.
• ResNet maps features to phase-only excitation vectors, avoiding online optimization.

Main Results:

• The proposed framework demonstrates improved tolerance to covariance mismatch.
• Achieves competitive Signal-to-Interference-plus-Noise Ratio (SINR) performance.
• Effective under limited-snapshot and noisy covariance estimation conditions.

Conclusions:

• The cascaded network offers a data-driven approach to robust beamforming.
• Separates covariance denoising and phase excitation emulation effectively.
• Outperforms conventional methods in challenging covariance conditions.