Online passive learning of timed automata for cyber-physical production systems

Model-based approaches are very often used for diagnosis in production systems. And since the manual creation of behavior models is a tough task, many learning algorithms have been constructed for the automatic model identification. Most of them are tested and evaluated on artificial datasets on personal computers only. However, the implementation on cyber-physical production systems puts additional requirements on learning algorithms, for instance the real-time aspect or the usage of memory space. This paper analyzes the requirements on learning algorithms for cyber-physical production systems and presents an appropriate online learning algorithm, the Online Timed Automaton Learning Algorithm, OTALA. It is the first online passive learning algorithm for timed automata which in addition copes without negative learning examples. An analysis of the algorithm and comparison with offline learning algorithms completes this contribution.