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

This paper presents a cascaded neural network framework for phase-only beamforming under covariance matrix mismatch. The proposed architecture combines a denoising autoencoder (DAE) with a residual network (ResNet) to address performance degradation caused by finite-snapshot covariance estimation errors and signal-of-interest contamination. The DAE reconstructs an ideal covariance representation from mismatched covariance inputs and provides compact covariance features for subsequent phase prediction. The ResNet then maps the denoised covariance features to phase-only excitation vectors, thereby avoiding repeated online optimization. Unlike conventional robust adaptive beamforming methods that rely on explicit uncertainty modeling or iterative covariance reconstruction, the proposed framework separates covariance feature denoising from phase excitation emulation in a data-driven manner. Numerical results demonstrate that the cascaded network improves covariance-mismatch tolerance and achieves competitive output SINR performance under limited-snapshot and noisy covariance-estimation conditions.