A Robust Alternate Least Squares Algorithm for Approximate Joint Diagonalization in Overdetermined Blind Source Separation
Shinya Saito, Kunio Oishi · 2021 IEEE 10th Global Conference on Consumer Electronics (GCCE) · 2021
A new robust approach for solving the approximate joint diagonalization (AJD) of target matrices in overdetermined blind source separation (BSS) is proposed. This approach is based on alternate least-squares (ALS) algorithm and is called the fast diagonalization implemented by minimizing the direct and inverse indirect least-squares criteria (FDMDII) algorithm. Starting from the initialization of the rectangular unmixing matrix as signal subspace component, the unmixing matrix is estimated from the product of residual square unmixing matrix by alternately minimizing two least-squares criteria at every ALS iteration. Fast convergence of the ALS solution is achieved by employing the FDMDII algorithm in noisy environments.