Feedback linearization of discrete-time nonlinear uncertain plants via firstprinciples- based serial neuro-gray-box models

In this paper, discrete-time control schemes based on feedback linearization of serial gray-box models are considered for partially known nonlinear processes. These techniques combine the benefits of feedback linearization, neural networks, and serial gray-box modeling, which result in larger dynamic operating ranges, better extrapolation properties, and fewer data acquisition efforts in comparison with the corresponding black-box-based schemes. First-principles-based serial gray-box models are classified into invertible and non-invertible structures for training purposes, and an improved approximate feedback linearization scheme based on Taylor series terms of a non-affine gray-box model is proposed. Moreover, an affine gray-box model is developed for applying the exact feedback linearization scheme. Simulation results on a fermentation process show that the proposed methods yield significant improvement in modeling and control performance in comparison with that of the black-box feedback linearization schemes.