A learning anomaly detection algorithm for hybrid manufacturing systems

For complex and distributed technical systems, model-based anomaly detection solutions often show better results than approaches, which do not use explicit behavior models. But so-far, the creation and maintenance of such models turned out to be the major problem of such model-based approaches. One solution to this problem is provided by machine learning: Behavior models can be learned automatically based on system observations. Such a machine-learning-based solution would create an anomaly detection algorithm, which could learn and therefore could adapt itself to new situations. In this paper, such an algorithm is described: The ANODA algorithm for anomaly detection uses the HyBUTLA algorithm to learn behavior models in form of hybrid timed probabilistic automata. For the first time, a thorough theoretical analysis of this algorithm is presented. Practical results also underline the applicability of these algorithms.