Abstract

This contribution investigates a feed-forward neural network approach for event detection in hybrid dynamical models. Machine learning algorithms are commonly used in software development. In recent years these approaches have also been increasingly applied in modelling and simulation of physical systems. A significant amount of these models use artificial neural networks. However, hybrid dynamical systems describe a combination of different methods to describe a continuous process, which experiences behavioural changes at discrete events. Accordingly, the models of such systems are based on a combination of discrete and continuous methods and are often illustrated as automaton. Based on these two areas an approach, to predict the event time of the discrete processes, is presented. The different required elements are defined and a general approach is outlined. The feasibility of this concept is examined on the basis of one examples. If the given imbalanced data is resampled, training can be successful. Unfortunately, even then, the generalised classification of events often does not work sufficiently. The evaluation of the approximation results of the discrete events in hybrid systems suggests that neural networks are not suitable to classify the system states with regard to the occurrence of an event. In the outlook we suggest an alternative approach to predict the event with neural networks.