Register now After registration you will be able to apply for this opportunity online.
Online identification of time-varying systems using data-driven models - a case study
We consider a novel model for dynamical systems that gave rise to new and highly performant data-driven control methods recently. Even though these methods are typically limited to linear time-invariant systems, recent work has proposed adapting the model online using tools for subspace tracking. This project aims to demonstrate the method’s effectiveness through case studies, highlighting its potential for superior performance in practical applications.
Keywords: system identification, data-driven control, time-varying systems
System identification is a crucial task in engineering and science, enabling the development of models for control, prediction, and analysis of dynamic systems. While conventional methods that identify parametric models (such as state-space matrices or ARX coefficients) are well-established [1], their performance can be limited, particularly in scenarios involving complex dynamics. Non-parametric approaches have emerged as a powerful alternative in the data-driven control literature [2], often demonstrating superior performance by directly using raw data as system representations. However, these methods are typically restricted to linear time-invariant systems, limiting their utility in environments where system properties evolve over time. To address this challenge, a novel approach has been developed that adapts the underlying data-driven model online using subspace tracking techniques (also known as streaming principal component analysis in machine learning). This project aims at demonstrating the performance of this new method on simulated case studies, such as the online identification of a rocket with time-varying mass.
[1] L. Ljung, System identification theory for the user, 1999.
[2] I. Markovsky, L. Huang, and F. Dörfler, "Data-driven control based on the behavioral approach: From theory to applications in power systems" IEEE Control Systems
[3] A. Sasfi, A. Padoan, I. Markovsky, F. Dörfler, "Subspace tracking for online system identification" arXiv preprint arXiv:2412.09052, 2024
System identification is a crucial task in engineering and science, enabling the development of models for control, prediction, and analysis of dynamic systems. While conventional methods that identify parametric models (such as state-space matrices or ARX coefficients) are well-established [1], their performance can be limited, particularly in scenarios involving complex dynamics. Non-parametric approaches have emerged as a powerful alternative in the data-driven control literature [2], often demonstrating superior performance by directly using raw data as system representations. However, these methods are typically restricted to linear time-invariant systems, limiting their utility in environments where system properties evolve over time. To address this challenge, a novel approach has been developed that adapts the underlying data-driven model online using subspace tracking techniques (also known as streaming principal component analysis in machine learning). This project aims at demonstrating the performance of this new method on simulated case studies, such as the online identification of a rocket with time-varying mass.
[1] L. Ljung, System identification theory for the user, 1999.
[2] I. Markovsky, L. Huang, and F. Dörfler, "Data-driven control based on the behavioral approach: From theory to applications in power systems" IEEE Control Systems
[3] A. Sasfi, A. Padoan, I. Markovsky, F. Dörfler, "Subspace tracking for online system identification" arXiv preprint arXiv:2412.09052, 2024
- Learn about data-driven, non-parametric system models, and novel methods for their adaptation
- Set up a simulation environment in python or Matlab, and simulate time-varying systems such as a rocket with changing mass
- Implement methods for the online identification of the system relying on subspace tracking techniques. Benchmark against conventional parametric identification methods on a prediction or fault detection task
- Draw conclusions from the results and document them in a final report
- Learn about data-driven, non-parametric system models, and novel methods for their adaptation - Set up a simulation environment in python or Matlab, and simulate time-varying systems such as a rocket with changing mass - Implement methods for the online identification of the system relying on subspace tracking techniques. Benchmark against conventional parametric identification methods on a prediction or fault detection task - Draw conclusions from the results and document them in a final report
Please send your cover letter and CV including transcript of records in PDF format via email to Andras Sasfi (asasfi@ethz.ch). We look forward to your application
Please send your cover letter and CV including transcript of records in PDF format via email to Andras Sasfi (asasfi@ethz.ch). We look forward to your application