Model-based approach to generate training sequences for discrete event anomaly detection in manufacturing

In the field of process control, more alarms are generated than can be physically addressed by a single operator, which is a significant problem. This situation is called an "alarm flood". Alarm floods occur because of badly designed alarm management systems (AMS) or causal dependent disturbances that raise multiple alarms based on only a single error. Functional dependent discrete alarm sequences can be modeled using the "formalized process description". Based on this model, dependent events can be analyzed with "sequence-based anomaly detection". The disadvantage is that anomaly detection algorithms need a vast quantity of data to detect anomalous sequences based on training sequences. Furthermore, these training sequences have to contain a few anomalous sequences. In this publication, we present a model-based approach to generate training sequences based on engineering data and analysis of historical alarm data. In the manufacturing field, no existing approach integrates engineering documents to generate training sequences for anomaly detection. Furthermore, in this publication, we introduce a model-based approach to model the signal behavior of plants. This model can be used to extract rules for anomaly detection analysis. The rules are used as input for further anomaly detection analysis to recognize more true positive alarm sequences.