After the course the student is able to:
- Describe the relations for signals and (dynamic) systems in time and frequency domain as well as conversions between continuous time and discrete time descriptions.
- Explain non-parametric and machine learning based system identification in time and frequency domain and comment on the validity of the obtained impulse response functions (IRF) and frequency response functions (FRF).
- Estimate parameters in models that are linear-in-the-parameters.
- Describe system identification with subspace identification techniques and use these techniques to obtain models.
- Explain system identification with prediction error identification methods (PEM), use these methods to obtain models, explain the approximate behaviour of these methods, evaluate and validate estimated models.
- Design an experiment to identify a set-up, collect the data and estimate a model of the system.
- Estimate parameters in more advanced models, explain the identifiability of the parameters and explain the sources for errors in the estimates.
- Explain the approaches for identification of closed-loop systems in time and frequency domain and implement an algorithm to estimate models from frequency domain data.
In everyday control or optimization practice, one requires the model of the corresponding system that is to be improved or controlled. As a system or its components cannot be always accurately described by physics-based principles, one aims at identifying a robust mathematical model, i.e. the parametrized linear/nonlinear function of the state, that can accurately describe the system dynamics. This is known as system identification. As both the system model and its parameters are assumed to be unknown, they are estimated given measured (input)-output data. In this manner, the discrepancy between the behaviour of the model and the real system is reduced. In general such an estimation can be achieved by a classical system identification/parameter estimation approaches, or by use of newly developed machine learning algorithms. In this course both of these approaches will be tackled, as well as their similarities and differences discussed.
Topics of this course are: The selection of the model structure, parameter estimation and the design of identification experiments for that purpose. One part of the course will be focusing on the system identification problem by considering a limited number of standard model structures for linear systems. In addition, attention will be paid to a more general parameter estimation in the time and frequency domain. Nonlinear systems will be also tackled such that their corresponding parametrization has a physical meaning.
The final grade is composed of two parts:
- An individual written examination of the course material. This contributes 50% to the final grade and a pass grade is required. The exact scope of the course material tested during this exam is communicated via Canvas. The test is scheduled once in the exam weeks at the end of block 2B and a resit is offered later.
- Answering standard assignments or solving an identification problem. This part may include a practical assignment and contributes the remaining 50% to the final grade. These assignments are available in block 2B such that students can complete the full course at the end of that block. A hand-in and grading schema is published on Canvas.