Open Collections

Abstract

Data-driven approaches have attracted a lot of interest from both industry and academia in recent years. In modern industrial sectors, the vast amounts of data generated by sensors, control systems, and other monitoring devices provide an opportunity to use data-driven approaches to address industrial challenges. Among these challenges, alarm flood occurrence is one of the most significant in industrial control systems. Alarm flood classification is crucial in industrial sectors as it helps manage the overwhelming number of alarms during abnormal conditions. Effective alarm flood classification enables human operators to quickly take a proper action in the abnormal conditions, reducing the risk of potential system failures. By analyzing alarm data, alarm flood classification improves decision-making, thereby enhancing overall safety and efficiency in complex industrial processes. This thesis presents an innovative approach for the clustering and classification of alarm floods in industrial systems, addressing the challenge of managing overwhelming alarm annunciations in complex processes. The research focuses on the development of a novel representation for alarms and alarm floods, which enables the detection and analysis of changes in alarm flood sequences over time due to varying operational conditions, environmental factors, or modifications in system settings. By leveraging information theory principles, such as the Jensen–Shannon divergence, a new metric is introduced to measure similarities between alarm floods. Additionally, the proposed methodology introduces dynamic clustering, which allows for the incremental creation of a pattern database and an alarm flood database, even in the absence of historical alarm data. The thesis also introduces a forgetting factor to update alarm flood patterns in the pattern database, ensuring accurate alarm flood clustering and classification. This approach is further enhanced by the implementation of a Long Short-Term Memory network for early classification of alarm floods using the created alarm flood database, facilitating timely and effective responses to abnormal conditions. The effectiveness of the proposed methods is demonstrated through a case study using the Tennessee Eastman process, highlighting significant improvements in alarm flood classification.