Process monitoring using causal graphical models, with application to clogging detection in steel continuous casting. (September 2021)

• Record Type: Journal Article
• Authors: Yang, Shu; Rebmann, Andreas; Tang, Ming; Moravec, Rudolf; Behrmann, Dylan; Baird, Morgan; Bequette, B. Wayne
• Is Part Of: Journal of process control. Volume 105(2021)
• Page Start: 259
• Page End: 266
• Publication Date: 2021-09
• Subjects: Process monitoring -- Multiple model -- Causal reasoning -- Smart manufacturing -- Steel continuous casting; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2021.08.006
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: The availability of manufacturing data is expected to grow exponentially due to the accelerating advancement in information technology, smart sensing, and industrial internet of things. To be able to efficiently leverage industrial "big data" to aid real-time decision making, smart manufacturing needs to incorporate field knowledge into the data-driven modeling process. When field knowledge of causality is available, causal graphical models are an effective way to incorporate it into a data-driven modeling process, leading to improved robustness and prediction power under distribution shifts. In this work, a process monitoring method based on causal graphical models and a multiple model framework is developed to detect clogging in a steel continuous casting process. By exploiting the statistical independence and invariance properties implied by the causal graphical model, this proposed method removes the effects of the confounding disturbance, leading to improved detection performance. Additionally, through a comparative experiment, it is shown that the causal graphical model is crucial to ensure that the disturbance is correctly removed while the information of interest is retained. This presented method is a good example of data-driven, knowledge-enabled models deployed in smart manufacturing. Highlights: A process monitoring method that combines causality with machine learning is proposed. An application of this methodology at a steelmaking plant is provided. Improvement by leveraging causality is demonstrated through a comparative experiment. This method is a data-driven, knowledge-enabled application of smart manufacturing.