Development of Data-Driven Methods for Alarm Flood Monitoring and Analysis

Abstract

Alarm floods present substantial challenges to industrial process safety, given their diverse causes and potential for severe consequences. Modern process industries involve sophisticated networks of devices that are interconnected in both upstream and downstream directions. The interconnected nature of these units and devices is a result of complex industrial production processes and the necessity to effectively manage numerous variables. Therefore, when a fault or abnormal condition occurs in an upstream or downstream component, it can lead to fault propagation due to the interconnected nature of the units and the feedback mechanisms that exist to control and regulate the overall system. Thus, this phenomenon leads to an increased number of alarms being generated across the system, causing an alarm flood. As industrial processes become more complex, plant operators often find it increasingly difficult to respond effectively, particularly during an alarm flood. The increased rate of alarms during an alarm flood overwhelms plant operators, resulting in delayed response times and further deterioration of the situation. Consequently, decision supports in alarm flood situations are in great demand to assist plant operators in assessing the root causes and taking corrective actions in time. Therefore, this thesis focuses on developing data-driven methods to efficiently manage alarm flood situations and minimize their impacts. Three research topics are considered. Firstly, early prediction of an incoming alarm flood sequence can provide valuable information to industrial operators, facilitating them to take corrective actions in time. A real-time pattern matching and ranking approach is proposed in this work to conduct similarity analysis under an online alarm flood situation and to export the results as a ranking list of historical alarm flood sequences. Unit and setbased pre-matching mechanisms are proposed to remove irrelevant sequences, and a set-based indexing and extension strategy is applied to further avoid unnecessary computation. Real-time decision supports in the form of ranking of similar historical alarm flood sequences are presented to the industrial operators. Secondly, a novel association rule mining approach is proposed for real-time prediction of alarm events during an alarm flood situation. This approach integrates a modified compact prediction tree model with new features, namely, the timetable and co-occurrence matrix, and is constructed based on historical alarm sequences. An alarm relevancy detection strategy is designed to identify and eliminate irrelevant alarms from the ongoing alarm flood. Furthermore, the proposed approach provides confidence intervals of the time differences between the subsequent predicted alarm events for time prediction. Such real-time assistance during alarm flood situations can greatly simplify the decision-making process for industrial operators. Finally, a reinforcement learning (RL) approach is proposed for early prediction of industrial alarm floods and to provide real-time guidance to plant operators in prompt mitigation of such situations. Based on various association rule metrics, irrelevant alarms are identified and eliminated to avoid inaccurate recommendations. A sequence reconstruction strategy is adopted to generate potential online scenarios by exploiting the alarm relations that exist in the historical sequences. Additionally, several criteria are introduced and implemented in the existing historical sequences to reformulate the training set for improved learning. To ensure accuracy and early recommendations, a double deep Q-network (DDQN) algorithm is incorporated into the proposed method. The effectiveness of these proposed methods is demonstrated by industrial case studies based on real industrial data from an oil refinery plant. By adopting these proposed approaches, plant operators could handle alarm floods proactively, resulting in an improvement in operational efficiency and safety.