Data-Driven Modeling and Pattern Recognition of Dynamical Systems
Author | : Pritthi Chattopadhyay |
Publisher | : |
Total Pages | : |
Release | : 2018 |
ISBN-10 | : OCLC:1050724247 |
ISBN-13 | : |
Rating | : 4/5 ( Downloads) |
Download or read book Data-Driven Modeling and Pattern Recognition of Dynamical Systems written by Pritthi Chattopadhyay and published by . This book was released on 2018 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: Human-engineered complex systems need to be monitored consistently to ensuretheir safety and efficiency, which might be affected due to degradation over timeor unanticipated disturbances. For systems that change at a fast time scale, insteadof active health monitoring, preventative system design is more feasible andeffective. Both active health monitoring and preventative system design can bedone using physics-based or data-driven models. In comparison to physics-basedmodels, data-driven models do not require knowledge of the underlying systemdynamics; they determine the relation between the relevant input and output variablesfrom a training data set. This is useful when there is lack of understandingof the system dynamics or the developed models are inadequate. One such scenariois combustion, where the difficulties include nonlinear dynamics involvingseveral input parameters; existence of bifurcations in the dynamic behavior andextremely high sensitivity of the combustor behavior to even small changes insome of the design parameters. Similarly, for batteries, sufficient knowledge of theelectrochemical characteristics is necessary to develop models for parameter identification at different operating points of the nonlinear battery dynamics. Thisdissertation develops dynamic data-driven models for combustor design and batteryhealth monitoring, using concepts of machine learning and statistics, whichdo not require much knowledge of the underlying system dynamics.But the performance of a data-driven algorithm depends on many factors namely:1. Availability of training data which covers all events of interest. For applicationsinvolving time series data, each individual time series must also besufficiently long, to encompass the dynamics of the underlying system foreach event.2. The quality of extracted features, i.e. whether they capture all the informationabout the system.3. The relation between the relevant input and output variables remaining constantduring the time the algorithm is being trained.Hence, the second part of the dissertation develops an unsupervised algorithm forscenarios where condition (iii) might not hold; quanties the eect of the nonconformityof condition (i) on the performance of an algorithm and proposes afeature extraction algorithm to ensure conformity of condition (ii).