Comparison of Different Probabilistic Graphical Models as Causal Models in Alarm Flood Reduction

The increasing amount of alarms and information for an operator in a modern plant becomes a significant safety risk. Although the notifications are a valuable support, they also lead to the curse of overloading with information for the operator. Due to the huge amount of alarms it is almost impossible to separate the crucial information from the insignificant ones. Therefore, new procedures are required to reduce these alarm floods and support the operator to minimize the safety risk. One approach is based on learning a causal model that represents the relationships between the alarms. This allows alarm sequences that are causally implied to be reduced to the root cause alarm. Fundamental element of this approach is the causal model. Therefore in this work, different probabilistic graphical models are considered and evaluated on the basis of appropriate criteria. A real use case of a bottle filling module serves as a benchmark for how well they are suitable as a causal model for the application in alarm flood reduction.