Cristian R. Rojas

Robust optimal experiment design for system identification

This paper develops the idea of min-max robust experiment design for dynamic system identification. The idea of min-max experiment design has been explored in the statistics literature. However, the technique is virtually unknown by the engineering community and, accordingly, there has been little prior work on examining its properties when applied to dynamic system identification. This paper initiates an exploration of these ideas. The paper considers linear systems with energy (or power) bounded inputs. We assume that the parameters lie in a given compact set and optimise the worst case over this set. We also provide a detailed analysis of the solution for an illustrative one parameter example and propose a convex optimisation algorithm that can be applied more generally to a discretised approximation to the design problem. We also examine the role played by different design criteria and present a simulation example illustrating the merits of the proposed approach.

A Note on the SPICE Method

In this article, we analyze the SPICE method developed in , and establish its connections with other standard sparse estimation methods such as the Lasso and the LAD-Lasso. This result positions SPICE as a computationally efficient technique for the calculation of Lasso-type estimators. Conversely, this connection is very useful for establishing the asymptotic properties of SPICE under several problem scenarios and for suggesting suitable modifications in cases where the naive version of SPICE would not work.

A Receding Horizon Algorithm to Generate Binary Signals with a Prescribed Autocovariance

Optimal test signals are frequently specified in terms of their second order properties, e.g. autocovariance or spectrum. However, to utilize these signals in practice, one needs to be able to produce realizations whose second order properties closely approximate the prescribed properties. Of particular interest are binary waveforms since they have the highest form-factor in the sense that they achieve maximal energy for a given amplitude. In this paper we utilize ideas from model predictive control to generate a binary waveform whose sampled autocovariance is as close as possible to some prescribed autocovariance. Several simulated examples are presented verifying the veracity of the algorithm. Also, a proof of convergence is given for the special case of bandlimited white noise. This proof is based on expressing the system in the form of a switched linear system.

Identification of Box-Jenkins models using structured ARX models and nuclear norm relaxation

In this contribution we present a method to estimate structured high order ARX models. By this we mean that the estimated model, despite its high order is close to a low order model. This is achieved by adding two terms to the least-squares cost function. These two terms correspond to nuclear norms of two Hankel matrices. These Hankel matrices are constructed from the impulse response coefficients of the inverse noise model, and the numerator polynomial of the model dynamics, respectively. In a simulation study the method is shown to be competitive as compared to the prediction error method. In particular, in the study the performance degrades more gracefully than for the Prediction Error Method when the signal to noise ratio decreases.

Sparse estimation based on a validation criterion

A sparse estimator with close ties with the LASSO (least absolute shrinkage and selection operator) is analysed. The basic idea of the estimator is to relax the least-squares cost function to what the least-squares method would achieve on validation data and then use this as a constraint in the minimization of the ℓ1-norm of the parameter vector. In a linear regression framework, exact conditions are established for when the estimator is consistent in probability and when it possesses sparseness. By adding a re-estimation step, where least-squares is used to re-estimate the non-zero elements of the parameter vector, the so called Oracle property can be obtained, i.e. the estimator achieves the asymptotic Cramér-Rao lower bound corresponding to when it is known which regressors are active. The method is shown to perform favourably compared to other methods on a simulation example.

On the calculation of the D-optimal multisine excitation power spectrum for broadband impedance spectroscopy measurements

The successful application of impedance spectroscopy in daily practice requires accurate measurements for modeling complex physiological or electrochemical phenomena in a single frequency or several frequencies at different (or simultaneous) time instants. Nowadays, two approaches are possible for frequency domain impedance spectroscopy measurements: (1) using the classical technique of frequency sweep and (2) using (non-)periodic broadband signals, i.e. multisine excitations. Both techniques share the common problem of how to design the experimental conditions, e.g. the excitation power spectrum, in order to achieve accuracy of maximum impedance model parameters from the impedance data modeling process. The original contribution of this paper is the calculation and design of the D-optimal multisine excitation power spectrum for measuring impedance systems modeled as 2R-1C equivalent electrical circuits. The extension of the results presented for more complex impedance models is also discussed. The influence of the multisine power spectrum on the accuracy of the impedance model parameters is analyzed based on the Fisher information matrix. Furthermore, the optimal measuring frequency range is given based on the properties of the covariance matrix. Finally, simulations and experimental results are provided to validate the theoretical aspects presented.

ROBUST IDENTIFICATION OF PROCESS MODELS FROM PLANT DATA

A precursor to any advanced control solution is the step of obtaining an accurate model of the process. Suitable models can be obtained from phenomenological reasoning, analysis of plant data or a combination of both. Here, we will focus on the problem of estimating (or calibrating) models from plant data. A key goal is to achieve robust identification. By robust we mean that small errors in the hypotheses should lead to small errors in the estimated models. We argue that, in some circumstances, it is essential that special precautions, including discarding some part of the data, be taken to ensure that robustness is preserved. We present several practical case studies to illustrate the results.

On optimal input design for nonlinear FIR-type systems

We consider optimal input design for system identification of nonlinear FIR-type systems in the prediction error (PEM) framework. The input sequences are designed in terms of their statistical properties and not directly in time domain. The starting point is the asymptotic properties of PEM estimates. The fact that the inverse covariance matrix of the estimated parameters is linear in the input probability density function is exploited to formulate convex optimization problems. The main issues considered are the parameterization of the input pdf, reduction of the number of free variables in the optimization and to some extent signal generation. Two special model classes where tractable problems are obtainable are studied in detail. Convex formulations of the input design problem are presented for the static nonlinear and nonlinear FIR cases. Numerical examples of the discussed ideas are also presented.

Iterative Data-Driven H-infinity Norm Estimation of Multivariable Systems With Application to Robust Active Vibration Isolation

This paper aims to develop a new data-driven H∞ norm estimation algorithm for model-error modeling of multivariable systems. An iterative approach is presented that requires significantly a fewer prior assumptions on the true system, hence it provides stronger guarantees in a robust control design. The iterative estimation algorithm is embedded in a robust control design framework with a judiciously selected uncertainty structure to facilitate high control performance. The approach is experimentally implemented on an industrial active vibration isolation system.

Sparse Estimation of Polynomial and Rational Dynamical Models

In many practical situations, it is highly desirable to estimate an accurate mathematical model of a real system using as few parameters as possible. At the same time, the need for an accurate description of the system behavior without knowing its complete dynamical structure often leads to model parameterizations describing a rich set of possible hypotheses; an unavoidable choice, which suggests sparsity of the desired parameter estimate. An elegant way to impose this expectation of sparsity is to estimate the parameters by penalizing the criterion with the ℓ0 “norm” of the parameters. Due to the non-convex nature of the ℓ0-norm, this penalization is often implemented as solving an optimization program based on a convex relaxation (e.g., ℓ1/LASSO, nuclear norm, . . .). Two difficulties arise when trying to apply these methods: (1) the need to use cross-validation or some related technique for choosing the values of regularization parameters associated with the ℓ1 penalty; and (2) the requirement that the (unpenalized) cost function must be convex. To address the first issue, we propose a new technique for sparse linear regression called SPARSEVA, with close ties with the LASSO (least absolute shrinkage and selection operator), which provides an automatic tuning of the amount of regularization. The second difficulty, which imposes a severe constraint on the types of model structures or estimation methods on which the ℓ1 relaxation can be applied, is addressed by combining SPARSEVA and the Steiglitz-McBride method. To demonstrate the advantages of the proposed approach, a solid theoretical analysis and an extensive simulation study are provided.