EAGER: Predictive Entropy-based State-Space Methodologies For Early-Warning of Critical System Transitions

Michigan State University, East Lansing MI

Investigators

Abstract

The project aims at developing a potentially transformative approach and methodologies for early-warning indicators for impending critical transitions or catastrophic failure in engineered and biological systems. The methodologies will employ uncertainty measures, such as various forms of entropy ( a measure of disorder in systems), and independent signal components to predict the impending onset of critical transitions to failure. The availability of early-warning system indicators would profoundly impact a broad range of engineering and natural systems. Relevant examples of engineered systems include power and communication networks, aircraft engines, industrial robots, electronics, machines and bridge infrastructures; examples of natural systems include epileptic seizures, organ failures, heart attacks and brain strokes. Specifically, predictive indicators for the impending onset of power system failures, based on sensory measurements, is crucial for taking steps to prevent sudden failures, including blackouts. The focus of the research is on entropy-based dynamically predictive indicators that give warning signals prior to the onset of catastrophic critical transitions. In the case of engineered systems, e.g., power generators and power systems, the approach would exploit the existence of reliable compact models for its predictive inference. For natural systems, e.g., predicting the impending epileptic seizures from EEG measurements, the approach would adaptively identify a viable model while simultaneously extracting indicators for its prediction. In both domains, the goal is to extract ensembles of the state signal's independent components and their statistical indicators from measurement profiles to infer the impending catastrophic transition to failure. The approach is expected to develop a new transformative framework of entropy-based core state independent components and fuse it into the established realm of Kalman and nonlinear prediction ideas as well as the dynamical systems concepts of critical transitions to bifurcations and chaos. It thus advances and integrates the disparate disciplines of (i) statistical independent component analysis, (ii) dynamical systems bifurcations, and (iii) recursive prediction system theory including the Kalman and nonlinear predictors. The goal is to develop the approach and methodologies for early-warning predictive systems, and validate their performance on specific prototype engineering models as well as natural (biological) systems.