Robust Neural Network Revived for Adaptive Blind Separation of Complex Wireless Signals
Marek Klemes · 2025
A robust adaptive algorithm for the blind separation of statistically-independent signals based on a modified analogue real-valued variation of the seminal Jutten-Herault (JH) network is developed and demonstrated in the context of separating complex-valued multi-level signals such as found in modern multi-input multi-output (MIMO) wireless networks. Development of the JH network is briefly outlined, with specific attention to its recursive non-causal structure and its development into a causal feed-forward modification with self-normalizing feature and robustness with respect to highly ill-conditioned signal mixtures. Additional aspects required to convert it to handle complex-valued signals in discrete time, over-sampling rates and input amplitude scaling, variable step-sizes, regularization and stopping criteria are also developed in this paper. Application to recovery of multiple quadrature-amplitude modulated (QAM) wireless signals is illustrated.