Speaker
Description
System identification is a data-driven approach to model the behavior of dynamical systems. Most algorithms of this type have a black-box view of the system, i.e. they relies exclusively on the observed input/output measurements, rather than utilizing the underlying physics. The identification process can be done in both time and frequency domains having various trade-offs, such as computational complexity, accuracy, noise and data reduction abilities.
In this talk, we develop a method to identify single input single output (SISO) linear and time invariant (LTI) systems in the frequency domain. The criteria for modeling and identification are formulated in terms of mean square error (MSE) that results in a separable nonlinear least squares problem, where the transfer function of the SISO LTI system is to be estimated. The construction of the proposed numerical method is discussed in the framework of a general mathematical model called variable projection. In this approach, we use the rational orthonormal parameterization of the corresponding transfer function whose poles and zeros are to be identified. As a case study, we consider the identification of several SISO LTI systems, whose transfer functions are known a priori. In the experiments, the input/output signals are simulated by the state-space realization of the corresponding system. By means of comparison tests performed on several randomly initialized zero/pole configurations, we demonstrate that our method outperforms conventional identification techniques in this field.
