Sliding-mode learning of a neuro-adaptive robust control configuration

Performance and robustness are highly desirable characteristics for any control method. But they are not found in general simultaneously in the same configuration due to its opposite nature. Recently, the compromise between robustness and performance has motivated new studies, mixing adaptive and robust methods, which are complementary in dealing with uncertainties and parameter variation. Variable structure control is a very successful adaptive method which has attracted much attention recently due to its inherent robustness, and also because it is equally applied to linear and nonlinear systems. The basic idea is to restrict the state space of a given plant through a so-called sliding surface, whose dynamics is simpler than the original plant dynamics. Enforcing a state-space trajectory from the initial condition of the plant to reach the surface in finite time, once there the plant remains on the surface and its dynamics is substituted by the surface dynamics. For adequately designed surfaces, they present the invariance property, guaranteeing an intrinsic robustness because the new dynamics does not depend on the plant parameters. Associating sliding-mode algorithms to artificial neural networks, some of the proposed configurations may present simultaneously good performance and robustness. In this work, a new configuration is proposed, implementing a neuro-adaptive control method using the variable structure approach to adjust the neural network weights, and presenting also robustness. The main idea is to add to a regular controller signal, a second control signal generated by an artificial neural network, in order to compensate for perturbations of the plant. An adaptive online learning is adopted whose transient signals are expected do not disturb the main control loop. It is expected also that the controller performance will be maintained through a wide variation of operational conditions of the plant, independently of perturbations caused by structural and parametric variations or nonlinearities not considered in the model. The configuration is explored through two different cases, when there is an acceptable linear model and when such model is not available. Numerical simulations presenting good results justify the expectations for the configuration.