Alarm data mining in complex industrial facilities using adaptive discretization based on time clustering and unified pattern fusion mining

Abstract Alarm systems in complex industrial facilities are crucial for ensuring operational efficiency and safety. However, the intricate interconnections and collaborations among facility components often lead to fault propagation, resulting in alarm redundancy and thereby increasing the difficulty of addressing genuine faults. One effective way to address this issue is to mine meaningful alarm patterns, which can be used to optimize alarm management. Existing alarm data mining techniques lack effective discretization methods for unlabeled continuous time-series alarm data during the data preprocessing stage, compromising the accuracy of mining results. Moreover, these methods typically identify only single association rules or sequential patterns, making it challenging to comprehensively reveal the interactions and dependencies among complex events—a critical factor for effective alarm management. To address these challenges, this paper proposes an adaptive discretization based on time clustering (ADTC) method, which transforms historical alarm data into transaction sets while effectively preserving the key features of the original data. Building on this, a unified pattern fusion mining (UPFM) method is developed to mine co-occurrence relationships and sequential dependency rules of alarm events, utilizing directed graphs for visualization. This approach enhances the accuracy of analyzing interactions and dependencies among complex equipment. To validate the effectiveness of the proposed methods, experiments were conducted on both synthetic alarm datasets and historical alarm data from a port transshipment system in northern China. Compared to discretization methods with different time window sizes and the density clustering discretization method based on DBSCAN, the ADTC method outperformed other methods in both support and confidence metrics. Additionally, the UPFM method successfully mined sequential dependencies and co-occurrence relationship rules of alarm events, demonstrating its effectiveness and accuracy in data mining within complex industrial environments.