A Data-Driven Method to Discover Association Rules of Mode-Dependent Alarms

Abstract

Alarm systems are considered to be essential components of industrial control systems, and are used to communicate the indication of alarm states to operators, ensuring the safe and efficient operation of modern industrial plants. Ideally, alarms should only appear when there are abnormal conditions that require operator responses, and the alarm rate should be kept to a reasonable number. But realistically, that’s not always the case, and operators are often overloaded by nuisance alarms and alarm floods, distracting them from truly critical alarms, and potentially causing accidents to occur. A variety of methods to manage alarm systems and improve their performance have been studied. Among them, state-based alarming is an advanced technique to reduce nuisance alarms and alarm floods by suppressing alarms associated with certain conditions or operating modes. To implement this technique, the key is to find the associations between operating modes and alarms. In practice, this is mainly done based on the experience of plant operators and expert knowledge of process engineers, and thus is very time consuming. Therefore, this thesis proposes a data driven method to discover the association rules of mode-dependent alarms for both single and multiple operating modes from the historical Alarm and Event (A&E) logs in an efficient way based on a data mining approach named FP-Growth (Frequent Pattern-Growth). The effectiveness of the proposed method is demonstrated by two industrial case studies.