An error-compensated immersion and invariance adaptive estimator for hammerstein neural networks with binary-valued observations.

Designing a parameter recovery algorithm using binary-valued observations that can handle the quantized identification problem of a nonlinear system is essential for various engineering processes including spectrum sensing to network communication, automatic control, and cyber-physical systems. This study addresses the parameter estimation problem based on binary-valued measurements for Hammerstein neural networks by constructing an error-compensated immersion and invariance (I&I)-based estimation method. First, based on a filter operator and half-decomposition technique, a filtered identification regression model is established. Subsequently, an online verification of the regression matrix satisfying the persistent excitation condition is performed using a filtered vector and matrix. Second, using regression vectors and a compensation strategy, identification error information is extracted from quantized observations, thereby eliminating the dependence on unpredictable excitation levels in diverse regressor elements. Third, by constructing two auxiliary functions and identification error information, a novel I&I adaptive estimator, incorporating a recursive correction factor, is designed. Finally, the effectiveness and practicality of the designed algorithm are confirmed through an illustrative example and validation using a real-life motor-driving system.