Balancing information and predictability: A pan latent feature model for plant-wide oscillations root cause analysis

Abstract

Analyzing the root cause for plant-wide oscillations is critical for maintaining the reliability and safety of complex systems with control loops. Oscillations in complex systems display varying degrees of predictability and information content. However, existing methods typically focus on a single aspect, which inherently restricts their comprehensiveness, flexibility, and accuracy of diagnosis. To address these challenges, this paper presents a novel pan-latent feature (PLF) modeling-based root cause analysis approach for plant-wide oscillations. PLF flexibly explores both predictability and information content within a unified model to extract informative, predictable, and a novel type of intermediate LFs that balance both attributes, enabling the comprehensive and flexible extraction of multi-type oscillations. By establishing explicit relationships between the extracted features and the original variables, PLF diagnoses the root cause variables of the extracted multi-type oscillations, providing multi-perspective diagnosis results. Through a numerical case study and a real-world plant-wide oscillation application, the proposed method demonstrates superior comprehensiveness, flexibility, and accuracy in finding the root variables of multi-type oscillations compared to existing approaches.