Using self-organizing maps to learn hybrid timed automata in absence of discrete events

Modern industrial plants become more complex and consequently monitoring them often exceeds the capabilities of human operators. Model-based diagnosis is a commonly used approach to identify anomalies and root causes within a system through the use of models, which are often times manually created by experts. However, manual modelling takes a lot of effort and is not suitable for today's fast-changing systems. Today, the large amount of sensor data provided by modern plants enables data-driven solutions and models can be learned from data, significantly reducing the manual modelling efforts. These data-driven solutions enable tasks such as condition monitoring: anomalies can be detected automatically, giving operators the chance to restore the plant to a working state before production losses occur. The choice of the model depends on a couple of factors, one of which is the type of the available signals. Hybrid timed automata are one type of model which separate the systems behaviour into different modes, e.g. 'valve open' or 'motor is running' through discrete events which are for example created from binary signals of the plant or through real-valued signal thresholds, defined by experts. The real-valued signals are then separated into the corresponding modes to improve the anomaly detection process in comparison to unseparated data. The anomaly detection for hybrid timed automata combines the detection of timing errors and sequence errors in the mode changes and the detection of anomalies in the real-valued signals. However, binary signals or expert knowledge to generate the much needed discrete events are not always available from the plant and automata can not be learned. The unsupervised, nonparametric approach presented and evaluated in this paper uses self-organizing maps and watershed transformations to allow the use of hybrid timed automata on data where learning of automata was not possible before.