Intelligent model reference nonlinear friction compensation using neural networks and Lyapunov based adaptive control

Two approaches to eliminating nonlinear effects that would otherwise render the linear controllers designed for a plant ineffective are shown to work well experimentally. A neural network compensation scheme assumes no a priori knowledge as to the structure of the nonlinearity and, with suitable computational capability and sufficient training and data, allows `inversion´ of the undesirable nonlinear effects. Since the network is learning both a structure as well as parameter values, the computational load is high. A further problem is the lack of stability guarantees for the weight update procedure and the distinct possibility of the network converging to local minima in the error backpropagation algorithm, although these phenomena did not appear to be problematic in experiment. A Lyapunov-based strategy offers fast parameter estimation with vastly reduced computation loads and hence the capability of adapting to varying surface friction conditions in real time