Anticipating Critical Transitions with Nonlinearity, Periodicity and Heterogeneity

[摘要] Many natural and engineering systems may switch abruptly from one stable state to another due to a small perturbation to the system;;s state or a small change in the underlining conditions. In ecosystems, for example, extinctions of species or desertification can occur rapidly. Therefore, critical transitions can be dangerous to a number of systems, and it could be very beneficial if monitoring or early warning methods were available while the system is still in the healthy regime. The approach of critical transitions in many natural and engineering systems is accompanied by a phenomenon called critical slowing down. Theoretical and experimental studies have suggested that responses to small perturbations become increasingly slow when these systems are near critical transitions. Statistics such as variance, autocorrelation calculated from time series data have been proposed as early warning signals to anticipate the system;;s approach to a transition point. The problem of anticipating critical transitions becomes more complicated when other factors come into play. Factors such as nonlinearity, periodicity and heterogeneity can alter the behavior of the system, and thus affect the applicability of generic early warning signals. This thesis examines the effect of these factors on the critical transition of a system, and develops new data-driven approaches accordingly. To deal with and exploit the existence of nonlinearity in the system, recoveries from large instead of small perturbations are used to calculate the recovery rates of the system versus amplitudes. Under the circumstances of periodicity, recovery rates are calculated discretely via the Poincare section. Using experimental and computational data, we show that a combination of using recoveries from large perturbations and calculating recovery rates using the Poincare section can be highly effective in terms of anticipating critical transitions for systems with parametric resonance. Moreover, this thesis develops new early warning signals for spatially extended systems based on the eigenvalues of the covariance matrix. We mathematically show that the dominanceof the largest eigenvalue of the covariance matrix can be used as an early warning signal by establishing the relationship between the eigenvalues of the covariance matrix and the eigenvalues of the force matrix. This new set of early warning signals are especially useful when the system has strong spatial heterogeneity. Lastly, this thesis investigates the influence of the choice of hyper-parameters, such as moving window size, sample rate, detrending methods, on the robustness of several early warning signals. General rules regarding data preparation and hypothesis testing are proposed.

[发布机构] University of Michigan

[效力级别] Mechanical Engineering

[关键词] Anticipating critical transitions;Mechanical Engineering;Engineering;Mechanical Engineering