University of Minnesota
Neural Differential Equation Based Modeling for Dynamic Systems
Abstract
dc:description.abstractModeling of complex dynamical systems allows for the prediction of and control over how the system will operate in the future. The control of the system allows for efficient and safe operations. Before a system is used, generally a controller must be designed for the system. How ever, after the deployment of the system, it is possible the system could degrade and evolve into an unstable state potentially causing harm to operators or damage to the system itself rendering it inoperable. To circumvent the possible harm and damage, it is desired to model an in-situ using a digital twin, a virtual copy of the system that behaves as the real system would. A digital twin is also helpful for real-time monitoring of the system, to show that the system is being properly controlled and all parts of the system are operable. The creation of a digital twin for an arbitrary system using machine learning architecture that can model the underlying differential equations is the motivation of this thesis. To this end, this thesis studies neural differential equation-based machine learning methods for the modeling of real-world dynamics systems.The initial method proposed, Recurrent Neural Ordinary Differential Equations (RNODE), is investigated using two case studies. The first case study examines a forced duffing oscillator using point-to-point mapping, while the second case study analyzes a real-world simulation of a machine tool spindle. The case studies investigated show that RNODE may not be a suitable methodology for real time prediction in systems with complex coupling and large amount of noise is present. These limitations prevent RNODE from serving as a suitable digital twin. Additionally, Nonlinear Continuous State Space Neural Network (NLCSNN) is introduced as a general approach for modeling a system’s internal differential relationships. Relevant background information is provided to further understand how physics can be embedded into the neural network’s architecture. Two case studies are investigated using simulated nonlinear duffing oscillator systems. These simulated case studies are used to show the proposed architecture functions as expected and show improvement over previously established architectures. Another more complex case study using a real-world continuous damping control damper shows the benefits of the NLCSNN are applicable beyond simulated systems.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Eischens, Reese
Subjects
dc:subject × 6Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Eischens, Reese. (2024). Neural Differential Equation Based Modeling for Dynamic Systems. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/269159.