Neural network augmentation of linear controllers with application to underwater vehicles

An approach to augment a linear compensator with an online neural network is presented. This scheme provides the benefits of adaptation with only minor modification to the existing control architecture, which is a substantial advantage over other approaches that require complete redesign. A neural network update law that guarantees bounded tracking for the augmented architecture is outlined. The advantages of the proposed technique are demonstrated through an application to an autonomous underwater vehicle. The design requirement is for attitude control such that robust trajectory following is achieved. A detailed nonlinear model of the AUV is given, and an operating point for nominal design is selected, about which a linear approximation is obtained. Structured uncertainties due to errors in the computation of hydrodynamic coefficients, linearization and truncation of plant dynamics, as well as effects of unknown disturbances are included in the control synthesis and compensated for by the neural network