Differential geometry offers a variety of analysis and design methodologies for nonlinear Systems [7]. One of the most prominent is input-output-linearization via state feedback. A severe drawback of this method is the need of an accurate model of the plant, since the output of interest has to be differentiated successively. Extensions of the method like adaptive linearization [11], robust linearization [13]. and asymptotically exact linearization [2] account for 'small' model-plant-mismatch.Nevertheless, even the development of a rough mechanistic model is often costly and time consuming. Data-based modeling of dynamic systems using neural networks offers a cost-effective alternative. In [10], a system of relative degree one was linearizedby state feedback, using a neural model of the plant dynamics. Here, it is shown that this method can he extended and successfully applied to systems of higher relative degree. This can be done using only data of the normal operation of the plant, i.e. no additional excitation of the plant for identification is needed.Section 2 gives an overview of input-output-linearization using neural process models. In section 3, the method is applied to trajectory tracking of a batch polymerization reactor. The highly nonlinear and complex dynamics of the batch polymerizationprocess are approximated by neural networks, using only measured state variables of one batch for training. Based on the learned model, an input-output-linearization is designed which significantly improves the control performance compared toconventional control.
展开▼
