Clustering analysis of process alarms using word embedding

Abstract

Industrial alarm systems have evolved significantly over the recent years both in terms of the number of observed alarms and the complexity of their presentation and management, seriously challenging the decision making abilities of the operators. These management challenges often arise from the presence of poorly configured alarms, as well as many nuisance and even flooding alarms. This necessitates better presentation tools to help operators understand the relations among various alarm events so that they can make more informed decisions. Instead of transforming process data into a binary alarm sequence for analysis, as often done in the literature, this paper proposes an alarm clustering method that takes advantage of the information contained in the alarm logs themselves. Using a word embedding technique, a novel clustering scheme and a multi-dimensional scaling method, the new method facilitates the grouping of correlated alarms. Such an approach is expected to further provide insight towards the removal of redundant alarms, and offer a sound basis for a subsequent causality analysis and identification of the alarm root cause. To demonstrate its merits, the proposed method is applied to the alarm events observed in a central heating and cooling plant located at a university campus.