Efficiency of machines has been an important topic in many areas of industry, at least since the introduction of the term "Industry 4.0". One of the challenges to improve the efficiency of industrial plants is the implementation of anomaly detection systems. These systems are designed to detect faults or wear and tear before a machine failure, for example, and to plan necessary maintenance cycles of machines (or parts) dynamically based on the current status.
A possible approach to solve this challenge by using hybrid time-dependent automata of normal behavior learned from sensor data to model continuous signals. This model reduces continuous signals (e.g. energy data) to single states. This allows an effective anomaly detection of continuous signals to be implemented. Finally, this methodology is compared to classical methods and a preview of other applications, e.g. predictive maintenance planning, is given.
- Process data modelling and monitoring of system states
In this approach, the normal behavior and the specific energy consumption of a system is determined by recording several training cycles of the system over a certain period of time. In this way, the typical behaviour of an industrial plant can be mapped. From these data a hybrid, time-based automaton is automatically learned in steps 2 - 4. Hereby the complexity of the plant functionality is displayed in a clear and compact way. In a next step, the learned model is used to detect anomalies in the discrete sensor signals in the industrial plant by comparing them with the process image.
- Anomaly detection in continuous signals
With the help of the hybrid time-based automaton, all relevant continuous data are additionally available as numerical vectors per state. State changes in the model describe a change in the dynamics of the system. This allows to separate even very complex continuous signals into vectors that are easy to model and to consider each one individually in the overall context. Here a method is worked out to model these signals using an adaptive regression or a Kalman filter. By a subsequent comparison of the learned vectors with the current measurement data, hypotheses about the system state can be established and quantified. This allows the detection of anomalies over the entire production cycle depending on the current system state.
By combining a learned hybrid automata of the normal state, the methodology allows not only the possibility of anomalies in the state space, but also the modeling of the continuous dynamics of a system state. The advantage of this method compared to a classical, static limit check is a model of the entire continuous dynamics, reduced to individually modelable state vectors. In addition, this allows to react flexibly to changes in the system, e.g. to learn and test new signal characteristics. Especially with regard to modular machines or systems that produce a very large number of different products, even with small batch sizes. Here the flexibility of the algorithms used is of decisive importance. Another application is the possibility to implement a predictive maintenance planning, which is based on the real system state that is continuously estimated based on all observations.
 Niggemann, Oliver; Stein, Benno; Vodencarevic, Asmir; Maier, Alexander; Kleine Büning, Hans: Learning Behavior Models for Hybrid Timed Systems. In: Twenty-Sixth Conference on Artificial Intelligence (AAAI-12) Jul 2012, Toronto, Canda.
 Kroll, Björn; Sebastian Schriegel; Schramm, Stefan; Niggemann, Oliver: A Software Architecture for the Analysis of Energy- and Process-Data. In: 18th International Conference on Emerging Technologies & Factory Automation (ETFA) Cagliari, Italy, Sep 2013