Comparison of alarm flood classification methods
Mäkelä, Elias (2024)
2024
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-02-26
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202402222472
https://urn.fi/URN:NBN:fi:tuni-202402222472
Tiivistelmä
Operators in process industries grapple with challenges of alarm systems due to their increasing complexity. These alarm systems, intended to alert operators to faults and issues, can lead to ‘alarm floods’—periods of excessively high alarm rates that cause a heavy cognitive load. Traditional alarm flood mitigation techniques such as alarm filtering or delay timers fail to address situations where a fault is propagating through a system. To address this issue many machine learning solutions have been applied, which attempt to assist operators in decision making during an alarm flood or to analyse historical data for alarm system improvements. However, the current research has gaps in comparing the performance of these methods and applying them to nonsynthetic data.
This thesis addresses these gaps by selecting the top methods from state-of-the-art alarm flood research and applies them to a real-world dataset collected from metal factory automation systems. The methods are used to classify alarm floods found from 15 years of historical data based on their root cause. The methods are evaluated with several metrics measuring their ability to predict the root cause correctly and being able to distinguish a new type of root cause.
Results demonstrate the potential of neural network-based methods, particularly domain knowledge-fused Word2Vec, which reaches 91.8% accuracy in offline analysis. This thesis highlights the efficacy of early prediction during online alarm flood, the challenges of manually identifying alarm flood root causes and the lack of high-performing methods which utilize alarm time information to its full extend.
This thesis addresses these gaps by selecting the top methods from state-of-the-art alarm flood research and applies them to a real-world dataset collected from metal factory automation systems. The methods are used to classify alarm floods found from 15 years of historical data based on their root cause. The methods are evaluated with several metrics measuring their ability to predict the root cause correctly and being able to distinguish a new type of root cause.
Results demonstrate the potential of neural network-based methods, particularly domain knowledge-fused Word2Vec, which reaches 91.8% accuracy in offline analysis. This thesis highlights the efficacy of early prediction during online alarm flood, the challenges of manually identifying alarm flood root causes and the lack of high-performing methods which utilize alarm time information to its full extend.