Investigating the Adaptability of Alarm Root-Cause Analysis Methods for Discrete Process Types

In modern automated production systems, quick reactions to faults are necessary to minimize production down-times. Alarm floods make this task difficult for operators by impairing situational awareness and decision making. For continuous process types, promising approaches exist to manage alarm floods by applying root-cause analysis methods to identify a root-cause of alarms, thereby reducing an alarm flood to only the most relevant messages. So far, attempts to adapt these methods to discrete processes have been limited. Differing boundary conditions like activity periods, the timescale of lingering faults in the technical process, and the different ways in which an operator interacts with a discrete process necessitate additional adaption efforts for machine and plant manufacturers. The main contribution is threefold. First, boundary conditions for discrete processes are examined; specifically how they affect established root-cause analysis methods. Next, Bayesian Network structure learning, a common root-cause analysis method used in continuous processes, is adapted to a packaging machine, which represents a typical discrete process. Finally, means to leverage the structure learning approach for assistance systems are investigated to support both human operators and alarm system engineers in the discrete process industry. Machine operators may gain insight into which alarms floods are particularly problematic for operators and which operator interactions cause further problems. Machine operators may learn which strategies have successfully resolved alarm floods in the past.