Adaptable Machine Learning for Nonlinear Dynamical Systems
Abstract
To predict catastrophic transitions leading to system collapse based solely on time series data is a challenging problem in nonlinear dynamics. The problem is also of high relevance to naval missions. For example, for an aircraft to land on an aircraft carrier, maneuvers such as reducing the fuel input to the gas turbine engine are necessary but this can lead to a flameout of the engine. The development of an effective method to detect flameout precursors so as to predict the possible occurrence of engine flameout must necessarily rely on time series data, as the equations of the underlying dynamical system are unknown. While sparse optimization such as compressivesensing has been developed to discover the system equations from data, the scope of its application is often limited as it requires the system equations to be simple by certain mathematical criterion, whereas real world systems such as those responsible for engine flameout are undoubtedly significantly more complicated. Thisproposal presents a plan to develop an adaptable machinelearning framework to predict critical transitions and collapse in general nonlinear and complex dynamical systems. The focus will be on reservoir computing machines with a recurrent neural network architecture, which have recently been demonstrated to be particularly suitable for predicting the state evolution of nonlinear dynamical and chaotic systems. The proposed research consists of threeresearch thrusts. Thrust 1 is to develop an adaptable reservoir computing paradigm to solve challenging prediction problems that were previously deemed extremely difficult or unsolvable in traditional nonlinear dynamics, with a focus on predicting critical transitions and system collapse.Thrust 2 is to establish the foundations of adaptable reservoir computing to gain an understanding of its inner working, contributing to the general field of explainable machine learning. Thrust 3 is to apply the adaptable reservoir computing scheme to the realworld problem of predicting flameout in gas turbine engines.Approved for Public Release
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 06, 2021
- Source ID
- N000142112323
Entities
People
- Ying-Cheng Lai
Organizations
- Arizona State University
- Office of Naval Research
- United States Navy