Hierarchical grouping and visualization of correlated alarms using time-augmented word embedding

In industrial processes, a large number of alarms displayed on human–machine interface screens may overwhelm human operators. This prevents them from taking appropriate corrective actions in a timely manner. Therefore, this paper proposes a three-stage computational procedure for grouping and visualizing correlated alarms, such that root-cause abnormalities can be more easily identified by the human operators. In the first stage, using a word embedding-based approach, alarm tags are transformed into real-valued vectors, where time stamps of the alarms are used rather than their order of occurrence. In the second stage, a multi-level density-based clustering approach is utilized to group correlated alarms hierarchically. In the third stage, a hierarchical visualization approach is developed to display alarm groups to human operators, which depicts hierarchical and statistical information. The implementation and effectiveness of the three-stage computational procedure are demonstrated using an alarm dataset generated for the benchmark Tennessee Eastman process system.