Online Classification of Alarm Floods Using a Word2vec Algorithm

Abstract

Alarm floods are periods of intense alarm activity that may hinder control room operators' ability to diagnose and respond to process abnormalities. In this context, a method to guide and assist operators during alarm floods would provide critical support in preventing abnormalities from escalating into serious accidents. Therefore, this study introduces a novel approach for the online classification of alarm floods based on their fault categories. Historical alarm data are used to train an ensemble of Natural Language Processing models, specifically word2vec, which learn contextual relationships between alarms under different fault conditions. As a new alarm flood appears, the models predict the most probable context alarms by exploiting the knowledge gained during training. Finally, a scoring system is proposed to reward the models that make correct predictions and eventually identify the most probable fault category. The efficacy of the method has been tested on simulated alarm data from the Tennessee Eastman Process benchmark. The results are encouraging, as the models achieved relatively high accuracy in most fault categories.