Neural-net based observers for sensorless drives

We demonstrate two applications of general regression neural networks to control. The first uses the network to learn the global linearization of a nonlinear plant, which transforms it into a simple linear plant. The concept is validated through numerical simulations. Conventional control methods are not applicable because they cannot deal with the nonlinearity and need exact information about the plant structure and parameters. In contrast to this the application of the proposed concept yields (after a learning phase) a nearly time optimal control behaviour of the system. The second approach uses the network as a tool for learning friction torque of an electrical drive, which is assumed to be an unknown nonlinear function with respect to the speed of the drive. In the simulations we have shown that this learning scheme converges very fast. Furthermore no persistently exciting system signals are needed. After the learning phase the network provides an accurate approximation of the friction torque and can be used, for example, for friction compensation in the speed controller. In both concepts the learning law for the general regression neural networks is derived based on a Lyapunov stability approach. This ensures convergence of the learning scheme and contrasts to gradient like learning laws used with many neural network models (e.g. the backpropagation network) where in principle convergence to the desired solution can not be guaranteed