Similarity Analysis of Industrial Alarm Floods Based on Word Embedding and Move-Split-Merge Distance

In industrial facilities, alarm systems are essential for process monitoring. However, due to the simplicity of alarm configuration and the poor performance of the alarm system, alarm floods often happen. Similarity analysis of alarm floods compares alarm flood sequences to look for common patterns. These patterns can offer information that is helpful in identifying root causes of alarm floods. Existing methods for alarm flood similarity analysis can only conduct similarity measures based on match operations for alarms represented by text strings; as a result, the match operations ignore the correlations between alarm occurrences. In this work, a new alarm flood similarity analysis method based on word embedding and Move-Split-Merge (MSM) distance is proposed, in order to reveal sequence similarity from a new perspective. The contributions are mainly twofold: 1) The correlations of alarm occurrences are considered in the alarm encoding via word embedding; 2) The MSM distance is calculated to analyze the similarity of alarm flood sequences of different lengths. Then, clustering results are compared with the original true labels of the alarm floods. The effectiveness of the proposed method is demonstrated by a case study with alarm data generated by a public industrial model of the Vinyl Acetate Monomer process.

I. Introduction

The presense of alarm floods impairs the effectiveness of the alarm system. In an alarm flood, a large number of alarms are generated, with the result that the alarm system is partially or completely inoperative [1]. In the ANSI/ISA-18.2 standard [2], an alarm flood is known as a situation that the amount of alarms exceeds the management capability of an operator. Therefore, it is crucial to identify and address the root cause of alarm floods as soon as possible. The alarm flood sequences are usually analyzed to solve this problem. This is because alarm flood sequences generally have some degree of recurrence and similar floods usually suggest that they have the same root cause. In the past decade, many sequence alignment methods have been proposed to compare alarm flood sequences; the results are helpful for root cause analysis and online prediction.