System Identification Toolbox™ Getting Started Guide

System Identification Toolbox™
Getting Started Guide
R2014a
Lennart Ljung
How to Contact MathWorks
www.mathworks.com Web
comp.soft-sys.matlab Newsgroup
www.mathworks.com/contact_TS.html Technical Support
[email protected] Product enhancement suggestions

[email protected] Bug reports
[email protected] Documentation error reports
[email protected] Order status, license renewals, passcodes
[email protected] Sales, pricing, and general information
508-647-7000 (Phone)
508-647-7001 (Fax)
The MathWorks, Inc.

3 Apple Hill Drive
Natick, MA 01760-2098
For contact information about worldwide offices, see the MathWorks Web site.
System Identification Toolbox™ Getting Started Guide
© COPYRIGHT 1988–2014 by The MathWorks, Inc.
The software described in this document is furnished under a license agreement. The software may be used
or copied only under the terms of the license agreement. No part of this manual may be photocopied or
reproduced in any form without prior written consent from The MathWorks, Inc.
FEDERAL ACQUISITION: This provision applies to all acquisitions of the Program and Documentation
by, for, or through the federal government of the United States. By accepting delivery of the Program
or Documentation, the government hereby agrees that this software or documentation qualifies as
commercial computer software or commercial computer software documentation as such terms are used
or defined in FAR 12.212, DFARS Part 227.72, and DFARS 252.227-7014. Accordingly, the terms and
conditions of this Agreement and only those rights specified in this Agreement, shall pertain to and govern
the use, modification, reproduction, release, performance, display, and disclosure of the Program and
Documentation by the federal government (or other entity acquiring for or through the federal government)
and shall supersede any conflicting contractual terms or conditions. If this License fails to meet the
government’s needs or is inconsistent in any respect with federal procurement law, the government agrees
to return the Program and Documentation, unused, to The MathWorks, Inc.
Trademarks
MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See
www.mathworks.com/trademarks for a list of additional trademarks. Other product or brand
names may be trademarks or registered trademarks of their respective holders.
Patents
MathWorks products are protected by one or more U.S. patents. Please see
www.mathworks.com/patents for more information.
Revision History
March 2007 First printing New for Version 7.0 (Release 2007a)
September 2007 Second printing Revised for Version 7.1 (Release 2007b)
March 2008 Third printing Revised for Version 7.2 (Release 2008a)
October 2008 Online only Revised for Version 7.2.1 (Release 2008b)
March 2009 Online only Revised for Version 7.3 (Release 2009a)
September 2009 Online only Revised for Version 7.3.1 (Release 2009b)
April 2011 Online only Revised for Version 7.4.2 (Release 2011a)
September 2012 Online only Revised for Version 8.1 (Release 2012b)
September 2013 Online only Revised for Version 8.3 (Release 2013b)
Contents
Product Overview
1
System Identification Toolbox Product Description . . 1-2
Key Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-2
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-4
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
What Is System Identification? . . . . . . . . . . . . . . . . . . . . . . 1-5
About Dynamic Systems and Models . . . . . . . . . . . . . . . . . . 1-5
System Identification Requires Measured Data . . . . . . . . . 1-9
Building Models from Data . . . . . . . . . . . . . . . . . . . . . . . . . 1-11
Black-Box Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-13
Grey-Box Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-18
Evaluating Model Quality . . . . . . . . . . . . . . . . . . . . . . . . . . 1-19
Learn More . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-23
Related Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-26
Using This Product

2
When to Use the GUI vs. the Command Line . . . . . . . . . 2-2
System Identification Workflow . . . . . . . . . . . . . . . . . . . . . 2-3
Commands for Model Estimation . . . . . . . . . . . . . . . . . . . . 2-6
v
Linear Model Identification
3
Identify Linear Models Using System Identification
Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
Preparing Data for System Identification . . . . . . . . . . . . . . 3-3
Saving the Session . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-19
Estimating Linear Models Using Quick Start . . . . . . . . . . . 3-21
Estimating Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . 3-28
Viewing Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 3-48
Exporting the Model to the MATLAB Workspace . . . . . . . . 3-52
Exporting the Model to the LTI Viewer . . . . . . . . . . . . . . . . 3-54
Identify Linear Models Using the Command Line . . . . . 3-55

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-55
Preparing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-56
Estimating Impulse Response Models . . . . . . . . . . . . . . . . . 3-67
Estimating Delays in the Multiple-Input System . . . . . . . . 3-71
Estimating Model Orders Using an ARX Model
Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-73
Estimating Transfer Functions . . . . . . . . . . . . . . . . . . . . . . 3-81
Estimating Process Models . . . . . . . . . . . . . . . . . . . . . . . . . . 3-84
Estimating Black-Box Polynomial Models . . . . . . . . . . . . . . 3-95
Simulating and Predicting Model Output . . . . . . . . . . . . . . 3-108
Identify Low-Order Transfer Functions (Process

Models) Using System Identification Tool . . . . . . . . . . 3-114
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-114
What Is a Continuous-Time Process Model? . . . . . . . . . . . . 3-115
Preparing Data for System Identification . . . . . . . . . . . . . . 3-116
Estimating a Second-Order Transfer Function (Process
Model) with Complex Poles . . . . . . . . . . . . . . . . . . . . . . . 3-124
Estimating a Process Model with a Noise Component . . . . 3-132
Viewing Model Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 3-139
Exporting the Model to the MATLAB Workspace . . . . . . . . 3-141
Simulating a System Identification Toolbox Model in
Simulink Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-142
vi Contents
Nonlinear Model Identification
4
Identify Nonlinear Black-Box Models Using System
Identification Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-2
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-2
What Are Nonlinear Black-Box Models? . . . . . . . . . . . . . . . 4-3
Preparing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-7
Estimating Nonlinear ARX Models . . . . . . . . . . . . . . . . . . . 4-13
Estimating Hammerstein-Wiener Models . . . . . . . . . . . . . . 4-26
vii
viii Contents
1
Product Overview
• “System Identification Toolbox Product Description” on page 1-2

• “Acknowledgments” on page 1-4
• “Overview” on page 1-5
• “Related Products” on page 1-26
1 Product Overview
System Identification Toolbox Product Description

Create linear and nonlinear dynamic system models from measured
input-output data
System Identification Toolbox™ provides MATLAB® functions, Simulink®
blocks, and an app for constructing mathematical models of dynamic systems
from measured input-output data. It lets you create and use models of
dynamic systems not easily modeled from first principles or specifications.
You can use time-domain and frequency-domain input-output data to identify
continuous-time and discrete-time transfer functions, process models, and
state-space models. The toolbox also provides algorithms for embedded online
parameter estimation.
The toolbox provides identification techniques such as maximum likelihood,

prediction-error minimization (PEM), and subspace system identification. To
represent nonlinear system dynamics, you can estimate Hammerstein-Weiner
models and nonlinear ARX models with wavelet network, tree-partition,
and sigmoid network nonlinearities. The toolbox performs grey-box system
identification for estimating parameters of a user-defined model. You can
use the identified model for system response prediction and plant modeling
in Simulink. The toolbox also supports time-series data modeling and
time-series forecasting.
Key Features
• Transfer function, process model, and state-space model identification
using time-domain and frequency-domain response data
• Autoregressive (ARX, ARMAX), Box-Jenkins, and Output-Error model
estimation using maximum likelihood, prediction-error minimization
(PEM), and subspace system identification techniques
• Online model parameter estimation
• Time-series modeling (AR, ARMA) and forecasting
• Identification of nonlinear ARX models and Hammerstein-Weiner models
with input-output nonlinearities such as saturation and dead zone
• Linear and nonlinear grey-box system identification for estimation of
user-defined models
1-2
System Identification Toolbox™ Product Description
• Delay estimation, detrending, filtering, resampling, and reconstruction

of missing data
1-3
1 Product Overview
Acknowledgments
System Identification Toolbox software is developed in association with the
following leading researchers in the system identification field:
Lennart Ljung. Professor Lennart Ljung is with the Department of

Electrical Engineering at Linköping University in Sweden. He is a recognized
leader in system identification and has published numerous papers and books
in this area.
Qinghua Zhang. Dr. Qinghua Zhang is a researcher at Institut National

de Recherche en Informatique et en Automatique (INRIA) and at Institut de
Recherche en Informatique et Systèmes Aléatoires (IRISA), both in Rennes,
France. He conducts research in the areas of nonlinear system identification,
fault diagnosis, and signal processing with applications in the fields of energy,
automotive, and biomedical systems.
Peter Lindskog. Dr. Peter Lindskog is employed by NIRA Dynamics

AB, Sweden. He conducts research in the areas of system identification,
signal processing, and automatic control with a focus on vehicle industry
applications.
Anatoli Juditsky. Professor Anatoli Juditsky is with the Laboratoire Jean

Kuntzmann at the Université Joseph Fourier, Grenoble, France. He conducts
research in the areas of nonparametric statistics, system identification, and
stochastic optimization.
1-4
Overview
Overview
In this section...
“What Is System Identification?” on page 1-5
“About Dynamic Systems and Models” on page 1-5
“System Identification Requires Measured Data” on page 1-9
“Building Models from Data” on page 1-11
“Black-Box Modeling” on page 1-13
“Grey-Box Modeling” on page 1-18
“Evaluating Model Quality” on page 1-19
“Learn More” on page 1-23
What Is System Identification?

System identification is a methodology for building mathematical models
of dynamic systems using measurements of the system’s input and output
signals.
The process of system identification requires that you:
• Measure the input and output signals from your system in time or
frequency domain.
• Select a model structure.
• Apply an estimation method to estimate value for the adjustable
parameters in the candidate model structure.
• Evaluate the estimated model to see if the model is adequate for your
application needs.
About Dynamic Systems and Models

• “What Is a Dynamic Model?” on page 1-6
• “Continuous-Time Dynamic Model Example” on page 1-6
• “Discrete-Time Dynamic Model Example” on page 1-8
1-5
1 Product Overview
What Is a Dynamic Model?

In a dynamic system, the values of the output signals depend on both the
instantaneous values of its input signals and also on the past behavior of the
system. For example, a car seat is a dynamic system—the seat shape (settling
position) depends on both the current weight of the passenger (instantaneous
value) and how long this passenger has been riding in the car (past behavior).
A model is a mathematical relationship between a system’s input and output

variables. Models of dynamic systems are typically described by differential
or difference equations, transfer functions, state-space equations, and
pole-zero-gain models.
You can represent dynamic models both in continuous-time and discrete-time

form.
An often-used example of a dynamic model is the equation of motion of a

spring-mass-damper system. As shown in the next figure, the mass moves in
response to the force F(t) applied on the base to which the mass is attached.
The input and output of this system are the force F(t) and displacement y(t)
respectively.
F(t)
y(t)
m
C
Mass-Spring-Damper System Excited by Force F(t)
Continuous-Time Dynamic Model Example

You can represent the same physical system as several equivalent models.
For example, you can represent the mass-spring-damper system in continuous
time as a second order differential equation:
1-6
Overview
d2 y dy
m +c + ky(t) = F (t)
2 dt
dt
where m is the mass, k the spring’s stiffness constant, and c the damping
coefficient. The solution to this differential equation lets you determine the
displacement of the mass, y(t), as a function of external force F(t) at any time t
for known values of constant m, c and k.
dy(t)
Consider the displacement, y(t), and velocity, v(t) = , as state variables:
dt
 y(t) 
x(t)   
 v(t) 
You can express the previous equation of motion as a state-space model of
the system:
dx
 Ax(t)  BF (t)
dt
y(t)  Cx(t)
The matrices A, B, and C are related to the constants m, c and k as follows:
 0 1 
A k c
  
 m m 
 1 
B  0
 m 
C  1 0
You can also obtain a transfer function model of the spring-mass-damper

system by taking the Laplace transform of the differential equation:
Y (s) 1
G(s) = =
2
F (s) (ms + cs + k)
1-7
1 Product Overview
where s is the Laplace variable.
Discrete-Time Dynamic Model Example

Suppose you can only observe the input and output variables F(t) and y(t)
of the mass-spring-damper system at discrete time instants t = nTs, where
Ts is a fixed time interval and n = 0, 1 , 2, .... The variables are said to be
sampled with sampling interval Ts. Then, you can represent the relationship
between the sampled input-output variables as a second order difference
equation, such as:
y(t) + a1 y(t − Ts ) + a2 y(t − 2Ts ) = bF (t − Ts )
Often, for simplicity, Ts is taken as one time unit, and the equation can be
written as:
y(t) + a1 y(t − 1) + a2 y(t − 2) = bF (t − 1)
where a1 and a2 are the model parameters. The model parameters are related
to the system constants m, c, and k, and the sampling interval Ts.
This difference equation shows the dynamic nature of the model. The
displacement value at the time instant t depends not only on the value of
force F at a previous time instant, but also on the displacement values at the
previous two time instants y(t–1) and y(t–2).
You can use this equation to compute the displacement at a specific time.
The displacement is represented as a weighted sum of the past input and
output values:
y(t) = bF (t − 1) − a1 y(t − 1) − a2 y(t − 2)
This equation shows an iterative way of generating values of output y(t)

starting from initial conditions (y(0) and y(1)) and measurements of input F(t).
This computation is called simulation.
Alternatively, the output value at a given time t can be computed using the
measured values of output at previous two time instants and the input value
at a previous time instant. This computation is called prediction. For more
1-8
Overview
information on simulation and prediction using a model, see “Simulating and

Predicting Model Output”.
You can also represent a discrete-time equation of motion in state-space and

transfer-function forms by performing the transformations similar to those
described in “Continuous-Time Dynamic Model Example” on page 1-6.
System Identification Requires Measured Data

• “Why Does System Identification Require Data?” on page 1-9
• “Time Domain Data” on page 1-9
• “Frequency Domain Data” on page 1-10
• “Data Quality Requirements” on page 1-11
Why Does System Identification Require Data?

System identification uses the input and output signals you measure from
a system to estimate the values of adjustable parameters in a given model
structure.
Obtaining a good model of your system depends on how well your measured
data reflects the behavior of the system. See “Data Quality Requirements”
on page 1-11.
Using this toolbox, you build models using time-domain input-output signals,
frequency response data, time series signals, and time-series spectra.
Time Domain Data

Time-domain data consists of the input and output variables of the system
that you record at a uniform sampling interval over a period of time.
For example, if you measure the input force, F(t), and mass displacement, y(t),
of the spring-mass-damper system at a uniform sampling frequency of 10 Hz,
you obtain the following vectors of measured values:
1-9
1 Product Overview
umeas  [ F (Ts ), F (2Ts ), F (3Ts ),..., F ( NTs )]

ymeas  [ y(Ts ), y(2Ts ), y(3Ts ),..., y( NTs )]
where Ts = 0.1 seconds and NTs is time of the last measurement.
If you want to build a discrete-time model from this data, the data vectors
umeas and ymeas and the sampling interval Ts provide sufficient information for
creating such a model.
If you want to build a continuous-time model, you should also know the
intersample behavior of the input signals during the experiment. For
example, the input may be piecewise constant (zero-order hold) or piecewise
linear (first-order hold) between samples.
Frequency Domain Data

Frequency domain data represents measurements of the system input and
output variables that you record or store in the frequency domain. The
frequency domain signals are Fourier transforms of the corresponding time
domain signals.
Frequency domain data can also represent the frequency response of the
system, represented by the set of complex response values over a given
frequency range. The frequency response describes the outputs to sinusoidal
inputs. If the input is a sine wave with frequency ω, then the output is also
a sine wave of the same frequency, whose amplitude is A(ω) times the input
signal amplitude and a phase shift of Φ(ω) with respect to the input signal.
The frequency response is A(ω)e(iΦ(ω)).
In the case of the mass-spring-damper system, you can obtain the frequency
response data by using a sinusoidal input force and measuring the
corresponding amplitude gain and phase shift of the response, over a range
of input frequencies.
You can use frequency-domain data to build both discrete-time and

continuous-time models of your system.
1-10
Overview
Data Quality Requirements

System identification requires that your data capture the important dynamics
of your system. Good experimental design ensures that you measure the right
variables with sufficient accuracy and duration to capture the dynamics you
want to model. In general, your experiment must:
• Use inputs that excite the system dynamics adequately. For example, a
single step is seldom enough excitation.
• Measure data long enough to capture the important time constants.
• Set up data acquisition system to have good signal-to-noise ratio.
• Measure data at appropriate sampling intervals or frequency resolution.
You can analyze the data quality before building the model using techniques
available in the Signal Processing Toolbox™ software. For example, analyze
the input spectra to determine if the input signals have sufficient power over
the bandwidth of the system.
You can also analyze your data to determine peak frequencies, input
delays, important time constants, and indication of nonlinearities using
non-parametric analysis tools in this toolbox. You can use this information
for configuring model structures for building models from data. See the
following topics:
• “Identifying Impulse-Response Models”

• “Identifying Frequency-Response Models”
Building Models from Data

• “System Identification Requires a Model Structure” on page 1-11
• “How the Toolbox Computes Model Parameters” on page 1-12
• “Configuring the Parameter Estimation Algorithm” on page 1-13
System Identification Requires a Model Structure

A model structure is a mathematical relationship between input and output
variables that contains unknown parameters. Examples of model structures
1-11
1 Product Overview
are transfer functions with adjustable poles and zeros, state space equations
with unknown system matrices, and nonlinear parameterized functions.
The following difference equation represents a simple model structure:
y(k) + ay(k − 1) = bu(k)
where a and b are adjustable parameters.
The system identification process requires that you choose a model structure
and apply the estimation methods to determine the numerical values of the
model parameters.
You can use one of the following approaches to choose the model structure:
• You want a model that is able to reproduce your measured data and is as
simple as possible. You can try various mathematical structures available
in the toolbox. This modeling approach is called black-box modeling.
• You want a specific structure for your model, which you may have derived
from first principles, but do not know numerical values of its parameters.
You can then represent the model structure as a set of equations or
state-space system in MATLAB and estimate the values of its parameters
from data. This approach is known as grey-box modeling.
How the Toolbox Computes Model Parameters

The System Identification Toolbox software estimates model parameters by
minimizing the error between the model output and the measured response.
The output ymodel of the linear model is given by:
ymodel(t) = Gu(t)
where G is the transfer function.
To determine G, the toolbox minimizes the difference between the model

output ymodel(t) and the measured output ymeas(t). The minimization criterion
is a weighted norm of the error, v(t), where:
v(t) = ymeas(t) – ymodel(t) = ymeas(t) – Gu(t).
1-12
Overview
ymodel(t) is one of the following:
• Simulated response of the model for a given input u(t).

• Predicted response of the model for a given input u(t) and past
measurements of output (ymeas(t-1), ymeas(t-2),...).
Accordingly, the error v(t) is called simulation error or prediction error. The
estimation algorithms adjust parameters in the model structure G such that
the norm of this error is as small as possible.
Configuring the Parameter Estimation Algorithm

You can configure the estimation algorithm by:
• Configuring the minimization criterion to focus the estimation in a desired

frequency range, such as put more emphasis at lower frequencies and
deemphasize higher frequency noise contributions. You can also configure
the criterion to target the intended application needs for the model such
as simulation or prediction.
• Specifying optimization options for iterative estimation algorithms.
The majority of estimation algorithms in this toolbox are iterative. You can
configure an iterative estimation algorithm by specifying options, such as
the optimization method and the maximum number of iterations.
For more information about configuring the estimation algorithm, see the
topics for estimating specific model structures.
Black-Box Modeling
• “Selecting Black-Box Model Structure and Order” on page 1-13
• “When to Use Nonlinear Model Structures?” on page 1-15
• “Black-Box Estimation Example” on page 1-16
Selecting Black-Box Model Structure and Order

Black-box modeling is useful when your primary interest is in fitting the data
regardless of a particular mathematical structure of the model. The toolbox
1-13
1 Product Overview
provides several linear and nonlinear black-box model structures, which have
traditionally been useful for representing dynamic systems. These model
structures vary in complexity depending on the flexibility you need to account
for the dynamics and noise in your system. You can choose one of these
structures and compute its parameters to fit the measured response data.
Black-box modeling is usually a trial-and-error process, where you estimate

the parameters of various structures and compare the results. Typically, you
start with the simple linear model structure and progress to more complex
structures. You might also choose a model structure because you are more
familiar with this structure or because you have specific application needs.
The simplest linear black-box structures require the fewest options to

configure:
• Transfer function, with a given number of poles and zeros.

• Linear ARX model, which is the simplest input-output polynomial model.
• State-space model, which you can estimate by specifying the number of
model states
Estimation of some of these structures also uses noniterative estimation

algorithms, which further reduces complexity.
You can configure a model structure using the model order. The definition of
model order varies depending on the type of model you select. For example, if
you choose a transfer function representation, the model order is related to the
number of poles and zeros. For state-space representation, the model order
corresponds to the number of states. In some cases, such as for linear ARX and
state-space model structures, you can estimate the model order from the data.
If the simple model structures do not produce good models, you can select
more complex model structures by:
• Specifying a higher model order for the same linear model structure.
Higher model order increases the model flexibility for capturing complex
phenomena. However, unnecessarily high orders can make the model less
reliable.
• Explicitly modeling the noise:
1-14
Overview
y(t)=Gu(t)+He(t)
where H models the additive disturbance by treating the disturbance as
the output of a linear system driven by a white noise source e(t).
Using a model structure that explicitly models the additive disturbance can
help to improve the accuracy of the measured component G. Furthermore,
such a model structure is useful when your main interest is using the model
for predicting future response values.
• Using a different linear model structure.
See “Linear Model Structures” in the User’s Guide.
• Using a nonlinear model structure.
Nonlinear models have more flexibility in capturing complex phenomena
than linear models of similar orders. See “Available Nonlinear Models”
in User’s Guide.
Ultimately, you choose the simplest model structure that provides the best
fit to your measured data. For more information, see “Estimating Linear
Models Using Quick Start” on page 3-21.
Regardless of the structure you choose for estimation, you can simplify the
model for your application needs. For example, you can separate out the
measured dynamics (G) from the noise dynamics (H) to obtain a simpler model
that represents just the relationship between y and u. You can also linearize
a nonlinear model about an operating point.
When to Use Nonlinear Model Structures?

A linear model is often sufficient to accurately describe the system dynamics
and, in most cases, you should first try to fit linear models. If the linear model
output does not adequately reproduce the measured output, you might need
to use a nonlinear model.
You can assess the need to use a nonlinear model structure by plotting the
response of the system to an input. If you notice that the responses differ
depending on the input level or input sign, try using a nonlinear model. For
example, if the output response to an input step up is faster than the response
to a step down, you might need a nonlinear model.
1-15
1 Product Overview
Before building a nonlinear model of a system that you know is nonlinear,

try transforming the input and output variables such that the relationship
between the transformed variables is linear. For example, consider a system
that has current and voltage as inputs to an immersion heater, and the
temperature of the heated liquid as an output. The output depends on the
inputs via the power of the heater, which is equal to the product of current
and voltage. Instead of building a nonlinear model for this two-input and
one-output system, you can create a new input variable by taking the product
of current and voltage and then build a linear model that describes the
relationship between power and temperature.
If you cannot determine variable transformations that yield a linear

relationship between input and output variables, you can use nonlinear
structures such as Nonlinear ARX or Hammerstein-Wiener models. For a list
of supported nonlinear model structures and when to use them, see “Available
Nonlinear Models” in User’s Guide.
Black-Box Estimation Example

You can use the GUI or commands to estimate linear and nonlinear models of
various structures. In most cases, you choose a model structure and estimate
the model parameters using a single command.
Consider the mass-spring-damper system, described in “About Dynamic

Systems and Models” on page 1-5. If you do not know the equation of motion
of this system, you can use a black-box modeling approach to build a model.
For example, you can estimate transfer functions or state-space models by
specifying the orders of these model structures.
A transfer function is a ratio of polynomials:
G(s) =
( b0 + b1 s + b2 s2 + ...)
(1 + f1 s + f2 s2 + ...)
For the mass-spring damper system, this transfer function is:
1
G(s) 
 ms 2
 cs  k 
1-16
Overview
which is a system with no zeros and 2 poles.
In discrete-time, the transfer function of the mass-spring-damper system

can be:
bz−1
G ( z−1 ) =
(1 + f1 z−1 + f2 z−2 )
where the model orders correspond to the number of coefficients of the
numerator and the denominator (nb = 1 and nf = 2) and the input-output
delay equals the lowest order exponent of z–1 in the numerator (nk = 1).
In continuous-time, you can build a linear transfer function model using the
tfest command:
m = tfest(data, 2, 0)
where data is your measured input-output data, represented as an iddata

object and the model order is the set of number of poles (2) and the number
of zeros (0).
Similarly, you can build a discrete-time model Output Error structure using
the following command:
m = oe(data, [1 2 1])
The model order is [nb nf nk] = [1 2 1]. Usually, you do not know the model
orders in advance. You should try several model order values until you find
the orders that produce an acceptable model.
Alternatively, you can choose a state-space structure to represent the

mass-spring-damper system and estimate the model parameters using the
ssest or the n4sid command:
m = ssest(data, 2)
where order = 2 represents the number of states in the model.
In black-box modeling, you do not need the system’s equation of motion—only

a guess of the model orders.
1-17
1 Product Overview
For more information about building models, see “Steps for Using the System
Identification Tool GUI” and “Model Estimation Commands” in the User’s
Guide.
Grey-Box Modeling
In some situations, you can deduce the model structure from physical
principles. For example, the mathematical relationship between the input
force and the resulting mass displacement in the mass-spring-damper system
is well known. In state-space form, the model is given by:
dx
 Ax(t)  BF (t)
dt
y(t)  Cx(t)
where x(t) = [y(t);v(t)] is the state vector. The coefficients A, B, and C are
functions of the model parameters:
A = [0 1; –k/m –c/m]
B = [0; 1/m]
C = [1 0]
Here, you fully know the model structure but do not know the values of its
parameters—m, c and k.
In the grey-box approach, you use the data to estimate the values of the
unknown parameters of your model structure. You specify the model structure
by a set of differential or difference equations in MATLAB and provide some
initial guess for the unknown parameters specified.
In general, you build grey-box models by:
1 Creating a template model structure.
2 Configuring the model parameters with initial values and constraints (if
any).
1-18
Overview
3 Applying an estimation method to the model structure and computing the

model parameter values.
The following table summarizes the ways you can specify a grey-box model
structure.
Grey-Box Structure Learn More

Representation
Represent the state-space model • “Estimate State-Space Models
structure as a structured idss model with Canonical Parameterization”
object and estimate the state-space
• “Estimate State-Space
matrices A, B and C.
Models with Structured
You can compute the parameter Parameterization”
values, such as m, c, and k, from
the state space matrices A and B.
For example, m = 1/B(2) and k =
–A(2,1)m.
Represent the state-space model “Grey-Box Model Estimation”
structure as an idgrey model object.
You can directly estimate the values
of parameters m, c and k.
Evaluating Model Quality

• “How to Evaluate and Improve Model Quality” on page 1-19
• “Comparing Model Response to Measured Response” on page 1-20
• “Analyzing Residuals” on page 1-21
• “Analyzing Model Uncertainty” on page 1-22
How to Evaluate and Improve Model Quality

After you estimate the model, you can evaluate the model quality by:
• “Comparing Model Response to Measured Response” on page 1-20

• “Analyzing Residuals” on page 1-21
1-19
1 Product Overview
• “Analyzing Model Uncertainty” on page 1-22
Ultimately, you must assess the quality of your model based on whether the
model adequately addresses the needs of your application. For information
about other available model analysis techniques, see “Model Analysis”.
If you do not get a satisfactory model, you can iteratively improve your results
by trying a different model structure, changing the estimation algorithm
settings, or performing additional data processing. For more information
about estimating each type of model structure, see the User’s Guide. If
these changes do not improve your results, you might need to revisit your
experimental design and data gathering procedures.
Comparing Model Response to Measured Response

Typically, you evaluate the quality of a model by comparing the model
response to the measured output for the same input signal.
Suppose you use a black-box modeling approach to create dynamic models

of the spring-mass damper system. You try various model structures and
orders, such as:
model1 = arx(data, [2 1 1]);

model2 = n4sid(data, 3)
You can simulate these models with a particular input and compare their
responses against the measured values of the displacement for the same input
applied to the real system. The following figure compares the simulated and
measured responses for a step input.
1-20
Overview
The previous figure indicates that model2 is better than model1 because
model2 better fits the data (65% vs. 83%).
The % fit indicates the agreement between the model response and the
measured output: 100 means a perfect fit, and 0 indicates a poor fit (that is,
the model output has the same fit to the measured output as the mean of
the measured output).
For more information, see “Simulating and Predicting Model Output” in the
User’s Guide.
Analyzing Residuals
The System Identification Toolbox software lets you perform residual analysis
to assess the model quality. Residuals represent the portion of the output
1-21
1 Product Overview
data not explained by the estimated model. A good model has residuals
uncorrelated with past inputs.
For more information, see “Residual Analysis” in the User’s Guide.
Analyzing Model Uncertainty

When you estimate the model parameters from data, you obtain their nominal
values that are accurate within a confidence region. The size of this region is
determined by the values of the parameter uncertainties computed during
estimation. The magnitude of the uncertainties provide a measure of the
reliability of the model. Large uncertainties in parameters can result from
unnecessarily high model orders, inadequate excitation levels in the input
data, and poor signal-to-noise ratio in measured data.
You can compute and visualize the effect of parameter uncertainties on

the model response in time and frequency domains using pole-zero maps,
Bode response, and step response plots. For example, in the following Bode
plot of an estimated model, the shaded regions represent the uncertainty
in amplitude and phase of model’s frequency response, computed using the
uncertainty in the parameters. The plot shows that the uncertainty is low
only in the 5 to 50 rad/s frequency range, which indicates that the model is
reliable only in this frequency range.
1-22
Overview
For more information, see “Computing Model Uncertainty” in the User’s

Guide.
Learn More
The System Identification Toolbox documentation provides you with the
necessary information to use this product. Additional resources are available
to help you learn more about specific aspects of system identification theory
and applications.
The following book describes methods for system identification and physical
modeling:
Ljung, L., and T. Glad. Modeling of Dynamic Systems. PTR Prentice Hall,
Upper Saddle River, NJ, 1994.
1-23
1 Product Overview
These books provide detailed information about system identification theory

and algorithms:
• Ljung, L. System Identification: Theory for the User. Second edition. PTR
Prentice Hall, Upper Saddle River, NJ, 1999.
• Söderström, T., and P. Stoica. System Identification. Prentice Hall
International, London, 1989.
For information about working with frequency-domain data, see the following
book:
Pintelon, R., and J. Schoukens. System Identification. A Frequency Domain

Approach. Wiley-IEEE Press, New York, 2001.
For information on nonlinear identification, see the following references:
• Sjöberg, J., Q. Zhang, L. Ljung, A. Benveniste, B. Deylon, P. Glorennec, H.

Hjalmarsson, and A. Juditsky, “Nonlinear Black-Box Modeling in System
Identification: a Unified Overview.” Automatica. Vol. 31, Issue 12, 1995,
pp. 1691–1724.
• Juditsky, A., H. Hjalmarsson, A. Benveniste, B. Delyon, L. Ljung,
J. Sjöberg, and Q. Zhang, “Nonlinear Black-Box Models in System
Identification: Mathematical Foundations.” Automatica. Vol. 31, Issue 12,
1995, pp. 1725–1750.
• Zhang, Q., and A. Benveniste, “Wavelet networks.” IEEE Transactions on
Neural Networks. Vol. 3, Issue 6, 1992, pp. 889–898.
• Zhang, Q., “Using Wavelet Network in Nonparametric Estimation.” IEEE
Transactions on Neural Networks. Vol. 8, Issue 2, 1997, pp. 227–236.
For more information about systems and signals, see the following book:
Oppenheim, J., and Willsky, A.S. Signals and Systems. PTR Prentice Hall,
Upper Saddle River, NJ, 1985.
The following textbook describes numerical techniques for parameter

estimation using criterion minimization:
1-24
Overview
Dennis, J.E., Jr., and R.B. Schnabel. Numerical Methods for Unconstrained
Optimization and Nonlinear Equations. PTR Prentice Hall, Upper Saddle
River, NJ, 1983.
1-25
1 Product Overview
Related Products
The following table summarizes MathWorks® products that extend and
complement the System Identification Toolbox software. For current
information about these and other MathWorks products, point your Web
browser to:
www.mathworks.com
Product Description
“Control System Toolbox™” Provides extensive tools to analyze
plant models created in the System
Identification Toolbox software and
to tune control systems based on
these plant models. You can use
the identified models directly for
advanced linear analysis and control
design tasks — no conversion of the
format required.
“Model Predictive Control Toolbox” Uses the linear plant models
created in the System Identification
Toolbox software for predicting plant
behavior that is optimized by the
model-predictive controller.
“Neural Network Toolbox™” Provides flexible neural-network
structures for estimating nonlinear
models using the System
Identification Toolbox software.
“Optimization Toolbox™” When this toolbox is installed,
you have the option of using
the lsqnonlin optimization
algorithm for linear and nonlinear
identification.
1-26
Related Products
Product Description
“Robust Control Toolbox™” Provides tools to design
multiple-input and multiple-output
(MIMO) control systems based on
plant models created in the System
Identification Toolbox software.
Helps you assess robustness based
on confidence bounds for the
identified plant model.
“Signal Processing Toolbox” Provides additional options for:
• Filtering
(The System Identification
Toolbox software provides only
the fifth-order Butterworth filter.)
• Spectral analysis
After using the advanced data

processing capabilities of the Signal
Processing Toolbox software, you
can import the data into the System
Identification Toolbox software for
modeling.
“Simulink” Provides System Identification
blocks for simulating the models
you identified using the System
Identification Toolbox software. Also
provides blocks for model estimation.
1-27
1 Product Overview
1-28
2
Using This Product
• “When to Use the GUI vs. the Command Line” on page 2-2
• “System Identification Workflow” on page 2-3
• “Commands for Model Estimation” on page 2-6
2 Using This Product
When to Use the GUI vs. the Command Line

After installing the System Identification Toolbox product, you can start the
System Identification Tool or work at the command line.
You can work either in the GUI or at the command line to preprocess data,
and estimate, validate, and compare models.
The following operations are available only at the command line:
• Generating input and output data (see idinput).

• Estimating coefficients of linear and nonlinear ordinary differential or
difference equations (grey-box models).
• Using recursive online estimation methods. For more information, see
topics about estimating linear models recursively in “Input-Output
Polynomial Models”.
• Converting between continuous-time and discrete-time models (see c2d
and d2c reference pages).
• Converting models to Control System Toolbox LTI objects (see ss, tf, and
zpk).
Note Conversions to LTI objects require the Control System Toolbox

software.
New users should start by using the GUI to become familiar with the product.
To open the GUI, on the Apps tab of MATLAB desktop, in the Apps section,
click System Identification. Alternatively, type ident in the MATLAB
Command Window.
To work at the command line, type the commands directly in the MATLAB
Command Window. For more information about a command, type doc
command_name in the MATLAB Command Window.
2-2
System Identification Workflow

System identification is an iterative process, where you identify models with
different structures from data and compare model performance. Ultimately,
you choose the simplest model that best describes the dynamics of your
system.
Because this toolbox lets you estimate different model structures quickly, you
should try as many different structures as possible to see which one produces
the best results.
A system identification workflow might include the following tasks:
1 Process data for system identification by:
• Importing data into the MATLAB workspace.

• Representing data in the System Identification Tool or as an iddata or
idfrd object in the MATLAB workspace.
• Plotting data to examine both time- and frequency-domain behavior.
To analyze the data for the presence of constant offsets and trends,
delay, feedback, and signal excitation levels, you can also use the advice
command.
• Preprocessing data by removing offsets and linear trends, interpolating
missing values, filtering to emphasize a specific frequency range, or
resampling (interpolating or decimating) using a different time interval.
2-3
2 Identify linear or nonlinear models:
• Frequency-response models
• Impulse-response models
• Low-order transfer functions (process models)
• Input-output polynomial models
• State-space models
• Transfer function models
• Nonlinear black-box models
• Ordinary difference or differential equations (grey-box models)
3 Validate models.
When you do not achieve a satisfactory model, try a different model

structure and order or try another identification algorithm. In some cases,
you can improve results by including a noise model.
You might need to preprocess your data before doing further estimation.
For example, if there is too much high-frequency noise in your data, you
might need to filter or decimate (resample) the data before modeling.
4 Postprocess models:
• Transform between continuous- and discrete-time domains

• Transform between model representations
• Extract numerical model data
• Subreference, concatenate and merge models
• Linearize nonlinear models
5 Use identified models for:
• Simulation and prediction

• Control design for the estimated plant using other MathWorks products.
You can import an estimated linear model into Control System Toolbox,
Model Predictive Control Toolbox™, Robust Control Toolbox, or Simulink
software.
2-4
• As dynamic blocks in Simulink
For online applications, you can perform online estimation.
2-5
Commands for Model Estimation

The following table summarizes System Identification Toolbox estimation
commands. For detailed information about using each command, see the
corresponding reference page.
Commands for Constructing and Estimating Models
Model Type Estimation Commands

Transfer function models tfest
Process models (low procest
order transfer functions
expressed in time-constant
form)
Linear input-output armax (ARMAX and ARIMAX models)
polynomial models arx (ARX and ARIX models)
bj (BJ only)
iv4 (ARX only)
oe (OE only)
polyest (for all models)
State-space models n4sid
ssest
Linear time-series models ar
arx (for multiple outputs)
ivar
Nonlinear ARX models nlarx
Hammerstein-Wiener nlhw
models
2-6
3
Linear Model Identification
• “Identify Linear Models Using System Identification Tool” on page 3-2

• “Identify Linear Models Using the Command Line” on page 3-55
• “Identify Low-Order Transfer Functions (Process Models) Using System
Identification Tool” on page 3-114
3 Linear Model Identification
Identify Linear Models Using System Identification Tool

In this section...
“Introduction” on page 3-2
“Preparing Data for System Identification” on page 3-3
“Saving the Session” on page 3-19
“Estimating Linear Models Using Quick Start” on page 3-21
“Estimating Linear Models” on page 3-28
“Viewing Model Parameters” on page 3-48
“Exporting the Model to the MATLAB Workspace” on page 3-52
“Exporting the Model to the LTI Viewer” on page 3-54
Introduction
• “Objectives” on page 3-2
• “Data Description” on page 3-3
Objectives
Estimate and validate linear models from single-input/single-output (SISO)
data to find the one that best describes the system dynamics.
After completing this tutorial, you will be able to accomplish the following
tasks using the System Identification Tool :
• Import data arrays from the MATLAB workspace into the GUI.
• Plot the data.
• Process data by removing offsets from the input and output signals.
• Estimate, validate, and compare linear models.
• Export models to the MATLAB workspace.
3-2
Note The tutorial uses time-domain data to demonstrate how you

can estimate linear models. The same workflow applies to fitting
frequency-domain data.
This tutorial is based on the example in section 17.3 of System Identification:

Theory for the User, Second Edition, by Lennart Ljung, Prentice Hall PTR,
1999.
Data Description
This tutorial uses the data file dryer2.mat, which contains
single-input/single-output (SISO) time-domain data from Feedback Process
Trainer PT326. The input and output signals each contain 1000 data samples.
This system heats the air at the inlet using a mesh of resistor wire, similar to
a hair dryer. The input is the power supplied to the resistor wires, and the
output is the air temperature at the outlet.
Preparing Data for System Identification

• “Loading Data into the MATLAB Workspace” on page 3-3
• “Opening the System Identification Tool ” on page 3-4
• “Importing Data Arrays into the System Identification Tool” on page 3-4
• “Plotting and Processing Data” on page 3-9
Loading Data into the MATLAB Workspace

Load the data in dryer2.mat by typing the following command in the
MATLAB Command Window:
load dryer2
This command loads the data into the MATLAB workspace as two column
vectors, u2 and y2, respectively. The variable u2 is the input data and y2 is
the output data.
3-3
Opening the System Identification Tool

To open the System Identification Tool, type the following command in the
ident
The default session name, Untitled, appears in the title bar.
Importing Data Arrays into the System Identification Tool

You can import the single-input/single-output (SISO) data from a sample data
file dryer2.mat into the GUI from the MATLAB workspace.
You must have already loaded the sample data into MATLAB, as described in
“Loading Data into the MATLAB Workspace” on page 3-3, and opened the
System Identification Tool, as described in “Opening the System Identification
Tool ” on page 3-4.
3-4
To import data arrays into the System Identification Tool:
1 In the System Identification Tool, select Import data > Time domain
data. This action opens the Import Data dialog box.
2 In the Import Data dialog box, specify the following options:
• Input — Enter u2 as the name of the input variable.

• Output — Enter y2 as the name of the output variable.
• Data name — Change the default name to data. This name labels the
data in the System Identification Tool after the import operation is
completed.
• Starting time — Enter 0 as the starting time. This value designates
the starting value of the time axis on time plots.
• Sampling interval — Enter 0.08 as the time between successive
samples in seconds. This value is the actual sampling interval in the
experiment.
3-5
The Import Data dialog box now resembles the following figure.
3 In the Data Information area, click More to expand the dialog box and
specify the following options:
3-6
Input Properties
• InterSample — Accept the default zoh (zero-order hold) to indicate
that the input signal was piecewise-constant between samples during
data acquisition. This setting specifies the behavior of the input signals
between samples when you transform the resulting models between
discrete-time and continuous-time representations.
• Period — Accept the default inf to specify a nonperiodic input.
Note For a periodic input, enter the whole number of periods of the
input signal in your experiment.
Channel Names
• Input — Enter power.
Tip Naming channels helps you to identify data in plots. For

multivariable input and output signals, you can specify the names of
individual Input and Output channels, separated by commas.
• Output — Enter temperature.
Physical Units of Variables

• Input — Enter W for power units.
Tip When you have multiple inputs and outputs, enter a

comma-separated list of Input and Output units corresponding to each
channel.
• Output — Enter ^oC for temperature units.
Notes — Enter comments about the experiment or the data. For

example, you might enter the experiment name, date, and a description of
3-7
experimental conditions. When you estimate models from this data, these
models inherit your notes.
The expanded Import Data dialog box now resembles the following figure.
3-8
4 Click Import to add the data to the System Identification Tool. The tool
displays an icon to represent the data.
5 Click Close to close the Import Data dialog box.
Plotting and Processing Data

In this portion of the tutorial, you evaluate the data and process it for system
identification. You learn how to:
• Plot the data.

• Remove offsets from the data by subtracting the mean values of the input
and the output.
• Split the data into two parts to use one part model estimation and the other
part for model validation.
The reason you subtract the mean values from each signal is because,
typically, you build linear models that describe the responses for deviations
from a physical equilibrium. With steady-state data, it is reasonable to
assume that the mean levels of the signals correspond to such an equilibrium.
Thus, you can seek models around zero without modeling the absolute
equilibrium levels in physical units.
3-9
You must have already imported data into the System Identification Tool, as
described in “Importing Data Arrays into the System Identification Tool”
on page 3-4.
To plot and process the data:
1 In the System Identification Tool, select the Time plot check box to open
the Time Plot. If the plot window is empty, click the data icon in the
System Identification Tool.
The top axes show the output data (temperature), and the bottom axes
show the input data (power). Both the input and the output data have
nonzero mean values.
3-10
2 Subtract the mean input value from the input data and the mean output
value from the output data. In the System Identification Tool, select
<--Preprocess > Remove means.
3-11
This action adds a new data set to the System Identification Tool with
the default name datad (the suffix d means detrend), and updates the
Time Plot window to display both the original and the detrended data. The
detrended data has a zero mean value.
3-12
3 Specify the detrended data to be used for estimating models. Drag the data
set datad to the Working Data rectangle.
Drag and drop data set.
4 Split the data into two parts and specify the first part for model estimation,
and the second part for model validation, as described in the following steps.
a Select <--Preprocess > Select range to open the Select Range window.
3-13
b In the Select Range window, create a data set containing the first 500
samples. In the Samples field, specify 1 500.
Tip You can also select data samples using the mouse by clicking and
dragging a rectangular region on the plot. If you select samples on the
input-channel axes, the corresponding region is also selected on the
output-channel axes.
c In the Data name field, type the name data_est.
3-14
d Click Insert to add this new data set to the System Identification Tool
to be used for model estimation.
e Repeat this process to create a second data set containing a subset of
the data to use for validation. In the Select Range window, specify the
last 500 samples in the Samples field. Type the name data_val in the
Data name field. Click Insert to add this new data set to the System
Identification Tool.
3-15
f Click Close to close the Select Range window.
3-16
5 In the System Identification Tool, drag and drop data_est to the Working
Data rectangle, and drag and drop data_val to the Validation Data
rectangle.
Drag and drop estimation data set
Drag and drop validation data set
3-17
6 To get information about a data set, right-click its icon. For example,
right-click data_est to open the Data/model Info dialog box.
3-18
You can also change certain values in the Data/model Info dialog box,
including:
• Changing the name of the data set in the Data name field.
• Changing the color of the data icon in the Color field. You specify colors
using RGB values (relative amounts of red, green, and blue). Each value
is between 0 and 1. For example, [1,0,0] indicates that only red is
present, and no green and blue are mixed into the overall color.
• Viewing or editing the commands executed on this data set in the Diary
And Notes area. This area contains the command-line equivalent to
the processing you performed using the System Identification Tool.
For example, as shown in the Data/model Info: estimate window, the
data_est data set is a result of importing the data, detrending the mean
values, and selecting the first 500 samples of the data.
% Import data
datad = detrend(data,0)
data_est = datad([1:500])
For more information about these and other toolbox commands, see the
corresponding reference pages.
The Data/model Info dialog box also displays the total number of samples, the
sampling interval, and the output and input channel names and units. This
information is not editable.
Tip As an alternative shortcut, you can select Preprocess > Quick start
from the System Identification Tool to perform all of the data processing
steps in this tutorial.
Learn More. For information about supported data processing operations,

such as resampling and filtering the data, see “Preprocess Data”.
Saving the Session

After you process the data, as described in “Plotting and Processing Data” on
page 3-9, you can delete any data sets in the window that you do not need for
estimation and validation, and save your session. You can open this session
3-19
later and use it as a starting point for model estimation and validation
without repeating these preparatory steps.
You must have already processed the data into the System Identification Tool,
as described in “Plotting and Processing Data” on page 3-9.
To delete specific data sets from a session and save the session:
1 In the System Identification Tool:
a Drag and drop the data data set into Trash.
b Drag and drop the datad data set into Trash.
Note Moving items to the Trash does not delete them. To permanently
delete items, select Options > Empty trash in the System Identification
Tool.
The following figure shows the System Identification Tool after moving
the items to Trash.
3-20
2 Drag and drop the data_est and data_val data sets to fill the empty
rectangles, as shown in the following figure.
3 Select File > Save session as to open the Save Session dialog box, and
browse to the folder where you want to save the session file.
4 In the File name field, type the name of the session

dryer2_processed_data, and click Save. The resulting file has
a .sid extension.
Tip You can open a saved session when starting the System Identification
Tool by typing the following command at the MATLAB prompt:
ident('dryer2_processed_data')
For more information about managing sessions, see “Starting and Managing
GUI Sessions”.
Estimating Linear Models Using Quick Start

• “How to Estimate Linear Models Using Quick Start” on page 3-22
3-21
• “Types of Quick Start Linear Models” on page 3-23

• “Validating the Quick Start Models” on page 3-24
How to Estimate Linear Models Using Quick Start

You can use the Quick Start feature in the System Identification Toolbox to
estimate linear models. Quick Start might produce the final linear models
you decide to use, or provide you with information required to configure the
estimation of accurate parametric models, such as time constants, input
delays, and resonant frequencies.
You must have already processed the data for estimation, as described in
“Plotting and Processing Data” on page 3-9.
In the System Identification Tool , select Estimate > Quick start.
This action generates plots of step response, frequency-response, and the

output of state-space and polynomial models. For more information about
these plots, see “Validating the Quick Start Models” on page 3-24.
3-22
Types of Quick Start Linear Models

Quick Start estimates the following four types of models and adds the
following to the System Identification Tool with default names:
• imp — Step response over a period of time using the impulse algorithm.
• spad — Frequency response over a range of frequencies using the spa
algorithm. The frequency response is the Fourier transform of the impulse
response of a linear system.
By default, the model is evaluated at 128 frequency values, ranging from
0 to the Nyquist frequency.
• arxqs — Fourth-order autoregressive (ARX) model using the arx algorithm.
This model is parametric and has the following structure:
y(t) + a1 y(t − 1) +  + ana y(t − na ) =

b1u(t − nk ) +  + bnbu(t − nk − nb + 1) + e(t)
y(t) represents the output at time t, u(t) represents the input at time t, na is
the number of poles, nb is the number of b parameters (equal to the number
of zeros plus 1), nk is the number of samples before the input affects output
of the system (called the delay or dead time of the model), and e(t) is the
white-noise disturbance. System Identification Toolbox software estimates
the parameters a1  an and b1  bn using the input and output data from
the estimation data set.
In arxqs, na=nb=4, and nk is estimated from the step response model imp.
3-23
• n4s3 — State-space model calculated using n4sid. The algorithm

automatically selects the model order (in this case, 3).
This model is parametric and has the following structure:
dx
 Ax(t)  Bu(t)  Ke(t)
dt
y(t)  Cx(t)  Du(t)  e(t)
y(t) represents the output at time t, u(t) represents the input at time t, x
is the state vector, and e(t) is the white-noise disturbance. The System
Identification Toolbox product estimates the state-space matrices A, B,
C, D, and K.
Note The Quick Start option does not create a transfer function model or a
process model which can also be good starting model types.
Validating the Quick Start Models

Quick Start generates the following plots during model estimation to help you
validate the quality of the models:
• Step-response plot
• Frequency-response plot
• Model-output plot
You must have already estimated models using Quick Start to generate
these plots, as described in “How to Estimate Linear Models Using Quick
Start” on page 3-22.
Step-Response Plot. The following step-response plot shows agreement

among the different model structures and the measured data, which means
that all of these structures have similar dynamics.
3-24
Tip If you closed the plot window, select the Transient resp check box to
reopen this window. If the plot is empty, click the model icons in the System
Identification Tool window to display the models on the plot.
Step Response for imp, arxqs, and n4s3
Tip You can use the step-response plot to estimate the dead time of linear
systems. For example, the previous step-response plot shows a time delay of
about 0.25 s before the system responds to the input. This response delay,
or dead time, is approximately equal to about three samples because the
sampling interval is 0.08 s for this data set.
Frequency-Response Plot. The following frequency-response plot shows

agreement among the different model structures and the measured data,
which means that all of these structures have similar dynamics.
3-25
Tip If you closed this plot window, select the Frequency resp check box to
reopen this window. If the plot is empty, click the model icons in the System
Identification Tool window to display the models on the plot.
Frequency Response for Models spad, arxqs, and n4s3
Model-Output Plot. The Model Output window shows agreement among the
different model structures and the measured output in the validation data.
Tip If you closed the Model Output window, select the Model output check
box to reopen this window. If the plot is empty, click the model icons in the
System Identification Tool window to display the models on the plot.
3-26
Measured Output and Model Output for Models imp, arxqs, and n4s3
The model-output plot shows the model response to the input in the validation
data. The fit values for each model are summarized in the Best Fits area
of the Model Output window. The models in the Best Fits list are ordered
from best at the top to worst at the bottom. The fit between the two curves is
computed such that 100 means a perfect fit, and 0 indicates a poor fit (that
is, the model output has the same fit to the measured output as the mean of
the measured output).
In this example, the output of the models matches the validation data output,
which indicates that the models seem to capture the main system dynamics
and that linear modeling is sufficient.
Tip To compare predicted model output instead of simulated output, select

this option from the Options menu in the Model Output window.
3-27
Estimating Linear Models

• “Strategy for Estimating Accurate Models” on page 3-28
• “Estimating Possible Model Orders” on page 3-28
• “Identifying Transfer Function Models” on page 3-33
• “Identifying State-Space Models” on page 3-36
• “Identifying ARMAX Models” on page 3-41
• “Choosing the Best Model” on page 3-45
Strategy for Estimating Accurate Models

The linear models you estimated in “Estimating Linear Models Using Quick
Start” on page 3-21 showed that a linear model sufficiently represents the
dynamics of the system.
In this portion of the tutorial, you get accurate parametric models by

performing the following tasks:
1 Identifying initial model orders and delays from your data using a simple,
polynomial model structure (ARX).
2 Exploring more complex model structures with orders and delays close to
the initial values you found.
The resulting models are discrete-time models.
Estimating Possible Model Orders

To identify black-box models, you must specify the model order. However,
how can you tell what model orders to specify for your black-box models? To
answer this question, you can estimate simple polynomial (ARX) models for
a range of orders and delays and compare the performance of these models.
You choose the orders and delays that correspond to the best model fit as an
initial guess for more accurate modeling using various model structures such
as transfer function and state-space models.
About ARX Models. For a single-input/single-output system (SISO), the

ARX model structure is:
3-28
y(t) + a1 y(t − 1) +  + ana y(t − na ) =

y(t) represents the output at time t, u(t) represents the input at time t, na is the
number of poles, nb is the number of zeros plus 1, nk is the input delay—the
number of samples before the input affects the system output (called delay or
dead time of the model), and e(t) is the white-noise disturbance.
You specify the model orders na, nb, and nk to estimate ARX models. The
System Identification Toolbox product estimates the parameters a1  an
and b1  bn from the data.
How to Estimate Model Orders.
1 In the System Identification Tool, select Estimate > Polynomial Models

to open the Polynomial Models dialog box.
2 From the Structure list, select ARX: [na nb nk]. By default, this is
already selected.
3-29
3 Edit the Orders field to try all combinations of poles, zeros, and delays,
where each value is from 1 to 10:
[1:10 1:10 1:10]
4 Click Estimate to open the ARX Model Structure Selection window, which
displays the model performance for each combination of model parameters.
You use this plot to select the best-fit model.
3-30
• The horizontal axis is the total number of parameters — na + nb.

• The vertical axis, called Unexplained output variance (in %), is
the portion of the output not explained by the model—the ARX model
prediction error for the number of parameters shown on the horizontal
axis.
The prediction error is the sum of the squares of the differences between
the validation data output and the model one-step-ahead predicted
output.
• nk is the delay.
Three rectangles are highlighted on the plot in green, blue, and red. Each
color indicates a type of best-fit criterion, as follows:
• Red — Best fit minimizes the sum of the squares of the difference
between the validation data output and the model output. This rectangle
indicates the overall best fit.
• Green — Best fit minimizes Rissanen MDL criterion.
• Blue — Best fit minimizes Akaike AIC criterion.
3-31
In this tutorial, the Unexplained output variance (in %) value remains

approximately constant for the combined number of parameters from 4 to
20. Such constancy indicates that model performance does not improve at
higher orders. Thus, low-order models might fit the data equally well.
Note When you use the same data set for estimation and validation, use
the MDL and AIC criteria to select model orders. These criteria compensate
for overfitting that results from using too many parameters. For more
information about these criteria, see the selstruc reference page.
5 In the ARX Model Structure Selection window, click the red bar
(corresponding to 15 on the horizontal axis), and click Insert. This
selection inserts na=6, nb=9, and nk=2 into the Polynomial Models dialog
box and performs the estimation.
This action adds the model arx692 to the System Identification Tool and
updates the plots to include the response of the model.
Note The default name of the parametric model contains the model type
and the number of poles, zeros, and delays. For example, arx692 is an ARX
model with na=6, nb=9, and a delay of two samples.
6 In the ARX Model Structure Selection window, click the third bar
corresponding to 4 parameters on the horizontal axis (the lowest order that
still gives a good fit), and click Insert.
• This selection inserts na=2, nb=2, and nk=3 (a delay of three samples)
into the Polynomial Models dialog box and performs the estimation.
• The model arx223 is added to the System Identification Tool and the
plots are updated to include its response and output.
7 Click Close to close the ARX Model Structure Selection window.
8 Click Close to close the Polynomial Models dialog box.
3-32
Identifying Transfer Function Models

By estimating ARX models for different order combinations, as described in
“Estimating Possible Model Orders” on page 3-28, you identified the number
of poles, zeros, and delays that provide a good starting point for systematically
exploring different models.
The overall best fit for this system corresponds to a model with six poles, nine
zeros, and a delay of two samples. It also showed that a low-order model
with na = 2 (two poles), nb = 2 (one zero), and nk = 3 (input-output delay) also
provides a good fit. Thus, you should explore model orders close to these
values.
In this portion of the tutorial, you estimate a transfer function model.
• “About Transfer Function Models” on page 3-33

• “How to Estimate Transfer Function Models” on page 3-34
• “Learn More” on page 3-36
About Transfer Function Models. The general transfer function model

structure is:
num(s)
Y (s)  U (s)  E(s)
den(s)
Y(s), U(s) and E(s) represent the Laplace transforms of the output, input
and error, respectively. num(s) and den(s) represent the numerator and
denominator polynomials that define the relationship between the input and
the output. The roots of the denominator polynomial are referred to as the
model poles. The roots of the numerator polynomial are referred to as the
model zeros.
You must specify the number of poles and zeros to estimate a transfer function
model. The System Identification Toolbox product estimates the numerator
and denominator polynomials, and input/output delays from the data.
The transfer function model structure is a good choice for quick estimation
because it requires that you specify only 2 parameters to get started: np is
the number of poles and nz is the number of zeros.
3-33
How to Estimate Transfer Function Models.
1 In the System Identification Tool, select Estimate > Transfer Function

Models to open the Transfer Functions dialog box.
2 In the Transfer Functions dialog box, specify the following options:
• Number of poles — Leave the default value 2 to specify a second order

function, for two poles.
• Number of zeros — Leave the default value 1.
• Continuous-time — Leave this checked.
3 Click I/O Delay to expand the input/output delay specification area.
By estimating ARX models for different order combinations, as described

in “Estimating Possible Model Orders” on page 3-28, you identified a 3
sample delay (nk = 3). This delay translates to a continuous-time delay
of (nk-1)*Ts , which is equal to 0.16 seconds.
Specify Delay as 0.16 seconds. Leave Fixed checked.
Use the default Estimation Options. By default, the tool assigns the name
tf1 to the model. The dialog box should look like this.
3-34
4 Click Estimate to add a transfer function model called tf1 to the System
Identification Tool. You can view the output of the estimation of the
transfer function model in comparison with the estimations of other
models, in the Model output window.
3-35
Tip If you closed the Model Output window, you can regenerate it by
selecting the Model output check box in the System Identification Tool.
If the new model does not appear on the plot, click the model icon in the
System Identification Tool to make the model active.
5 Click Close to close the Transfer Functions dialog box.
Learn More. To learn more about identifying transfer function models, see
“Transfer Function Models”.
Identifying State-Space Models

3-36
zeros, and a delay of two samples. It also showed that a low-order model with
na=2 (two poles), nb=2 (one zero), and nk=3 (input-output delay) also provides
a good fit. Thus, you should explore model orders close to these values.
In this portion of the tutorial, you estimate a state-space model.
• “About State-Space Models” on page 3-37

• “How to Estimate State-Space Models” on page 3-37
About State-Space Models. The general state-space model structure

(innovation form) is:
dx
 Ax(t)  Bu(t)  Ke(t)
dt
y(t)  Cx(t)  Du(t)  e(t)
y(t) represents the output at time t, u(t) represents the input at time t, x(t) is
the state vector at time t, and e(t) is the white-noise disturbance.
You must specify a single integer as the model order (dimension of the state
vector) to estimate a state-space model. The System Identification Toolbox
product estimates the state-space matrices A, B, C, D, and K from the data.
The state-space model structure is a good choice for quick estimation because
it requires that you specify only the number of states (which equals the
number of poles). You can optionally also specify the delays and feedthrough
behavior.
How to Estimate State-Space Models.
1 In the System Identification Tool , select Estimate > State Space Models
to open the State Space Models dialog box.
3-37
2 In the Specify value field, specify the model order. Type 6 to create a
sixth-order state-space model.
This choice is based on the fact that the best-fit ARX model has six poles.
Tip Although this tutorial estimates a sixth-order state-space model, you

might want to explore whether a lower-order model adequately represents
the system dynamics.
3 Click Estimation Options to expand the estimation options area.
4 Change Focus to Simulation to optimize the model to use for output

simulation.
The State Space Models dialog box looks like the following figure.
3-38
5 Click Estimate to add a state-space model called ss1 to the System

You can view the output of the estimation of the state-space model in
comparison with the estimations of other models, in the Model output
window.
3-39
6 Click Close to close the State Space Models dialog box.
Learn More. To learn more about identifying state-space models, see

“State-Space Models”.
3-40
Identifying ARMAX Models

zeros, and a delay of two samples. It also showed that a low-order model with
na=2 (two poles), nb=2 (one zero), and nk=3 also provides a good fit. Thus, you
should explore model orders close to these values.
In this portion of the tutorial, you estimate an ARMAX input-output

polynomial model.
• “About ARMAX Models” on page 3-41

• “How to Estimate ARMAX Models” on page 3-42
About ARMAX Models. For a single-input/single-output system (SISO), the

ARMAX polynomial model structure is:
y(t) + a1 y(t − 1) +  + ana y(t − na ) =

b1u(t − nk ) +  + bnbu(t − nk − nb + 1) +
e(t) + c1 e(t − 1) +  + cnc e(t − nc )
the number of poles for the dynamic model, nb is the number of zeros plus 1, nc
is the number of poles for the disturbance model, nk is the number of samples
before the input affects output of the system (called the delay or dead time of
the model), and e(t) is the white-noise disturbance.
Note The ARMAX model is more flexible than the ARX model because
the ARMAX structure contains an extra polynomial to model the additive
disturbance.
3-41
You must specify the model orders to estimate ARMAX models. The System
Identification Toolbox product estimates the model parameters a1  an ,
b1  bn , and c1  cn from the data.
How to Estimate ARMAX Models.
1 In the System Identification Tool , select Estimate > Polynomial Models

to open the Polynomial Models dialog box.
2 From the Structure list, select ARMAX: [na nb nc nk] to estimate an

ARMAX model.
3-42
3 In the Orders field, set the orders na, nb, nc, and nk to the following values:
[2 2 2 2]
The tool assigns the name to the model amx2222, by default, visible in the
Name field.
4 Click Estimate to add the ARMAX model to the System Identification Tool.
3-43
5 Repeat steps 3 and 4 using higher Orders 3 3 2 2. These orders result in

a model that fits the data almost as well as the higher order ARX model
arx692.
6 Click Close to close the Polynomial Models dialog box.
Learn More. To learn more about identifying input-output polynomial

models, such as ARMAX, see “Input-Output Polynomial Models”.
3-44
Choosing the Best Model

You can compare models to choose the model with the best performance.
You must have already estimated the models, as described in “Estimating

Linear Models” on page 3-28.
Summary of Models. The following figure shows the System Identification

Tool, which includes all the estimated models in “Estimating Linear Models”
on page 3-28.
Examining the Model Output. Examine the model output plot to see how
well the model output matches the measured output in the validation data
set. A good model is the simplest model that best describes the dynamics and
successfully simulates or predicts the output for different inputs. Models are
listed by name in the Best Fits area of the Model Output plot. Note that one
of the simpler models, amx3322, produced a similar fit as the highest-order
model you created, arx692.
3-45
Tip If you closed the Model Output window, you can regenerate it by selecting
the Model output check box in the System Identification Tool. If the new
model does not appear on the plot, click the model icon in the System
Identification Tool to make the model active.
To validate your models using a different data set, you can drag and drop this
data set into the Validation Data rectangle in the System Identification
Tool. If you transform validation data into the frequency domain, the Model
Output plot updates to show the model comparison in the frequency domain.
3-46
To get a closer look at how well these models fit the data, magnify a portion of
the plot by clicking and dragging a rectangle around the region of interest, as
shown in the following figure.
3-47
Releasing the mouse magnifies this region and shows that the output of all
models matches the validation data well.
Viewing Model Parameters

• “Viewing Model Parameter Values” on page 3-48
• “Viewing Parameter Uncertainties” on page 3-50
Viewing Model Parameter Values

You can view the numerical parameter values for each estimated model.

To view the parameter values of the model amx3322, right-click the model icon
in the System Identification Tool. The Data/model Info dialog box opens.
3-48
3-49
The noneditable area of the Data/model Info dialog box lists the parameter
values correspond to the following difference equation for your system:
y(t)  1.508 y(t  1)  0.7291 y(t  2)  0.1219 y(t  3) 

0.004257u(t  2)  0.06201u(t  3)  0.02643u(t  4)  e(t)  0.5835e(t  1)  0.258
88 e(t  2)
Note The coefficient of u(t-2) is not significantly different from zero. This
lack of difference explains why delay values of both 2 and 3 give good results.
Parameter values appear in the following format:
A( z)  1  a1 z1    ana z na
B( z)  b1 z nk    bnb z nb nk1
C( z)  1  c1 z1    cnc z nc
The parameters appear in the ARMAX model structure, as follows:
A(q) y(t) = B(q)u(t) + C(q) e(t)
which corresponds to this general difference equation:
y(t) + a1 y(t − 1) +  + ana y(t − na ) =

b1u(t − nk ) +  + bnbu(t − nk − nb + 1) +
e(t) + c1 e(t − 1) +  + cnc e(t − nc )
the number of poles for the dynamic model, nb is the number of zeros plus 1, nc
is the number of poles for the disturbance model, nk is the number of samples
before the input affects output of the system (called the delay or dead time of
the model), and e(t) is the white-noise disturbance.
Viewing Parameter Uncertainties

You can view parameter uncertainties of estimated models.
3-50

To view parameter uncertainties, click Present in the Data/model Info dialog

box, and view the model information at the MATLAB prompt.
amx3322 =
Discrete-time ARMAX model: A(z)y(t) = B(z)u(t) + C(z)e(t)
A(z) = 1 - 1.508 (+/- 0.05919) z^-1 + 0.7293 (+/- 0.08734) z^-2

- 0.1219 (+/- 0.03424) z^-3
B(z) = 0.004256 (+/- 0.001563) z^-2 + 0.06201 (+/- 0.002409) z^-3

+ 0.02642 (+/- 0.005633) z^-4
C(z) = 1 - 0.5837 (+/- 0.07189) z^-1 + 0.2588 (+/- 0.05253) z^-2
Name: amx3322
Sample time: 0.08 seconds
Parameterization:
Polynomial orders: na=3 nb=3 nc=2 nk=2
Number of free coefficients: 8
Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.
Status:
Termination condition: Near (local) minimum, (norm(g) < tol).
Number of iterations: 5, Number of function evaluations: 11
Estimated using POLYEST on time domain data "data_est".

Fit to estimation data: 95.3% (prediction focus)
FPE: 0.00163, MSE: 0.00155
More information in model's "Report" property.
The 1-standard deviation uncertainty for the model parameters is in

parentheses next to each parameter value.
3-51
Exporting the Model to the MATLAB Workspace

The models you create in the System Identification Tool are not automatically
available in the MATLAB workspace. To make a model available to other
toolboxes, Simulink, and System Identification Toolbox commands, you must
export your model from the System Identification Tool to the MATLAB
workspace. For example, if the model is a plant that requires a controller, you
can import the model from the MATLAB workspace into the Control System
Toolbox product.

3-52
To export the amx3322 model, drag it to the To Workspace rectangle in the

System Identification Tool.
Drag and drop model to Workspace
The model appears in the MATLAB Workspace browser.
Note This model is an idpoly model object.
3-53
After the model is in the MATLAB workspace, you can perform other
operations on the model. For example, if you have the Control System Toolbox
product installed, you might transform the model to a state-space object using:
ss_model=ss(amx3322)
Exporting the Model to the LTI Viewer

If you have the Control System Toolbox product installed, the To LTI Viewer
rectangle appears in the System Identification Tool.
The LTI Viewer is a graphical user interface for viewing and manipulating
the response plots of linear models. It displays the following plots:
• Step- and impulse-response

• Bode, Nyquist, and Nichols
• Frequency-response singular values
• Pole/zero
• Response to general input signals
• Unforced response starting from given initial states (only for state-space
models)
To plot a model in the LTI Viewer, drag and drop the model icon to the To
LTI Viewer rectangle in the System Identification Tool.
For more information about working with plots in the LTI Viewer, see “LTI
Viewer Overview”.
3-54
Identify Linear Models Using the Command Line

In this section...
“Preparing Data” on page 3-56
“Estimating Impulse Response Models” on page 3-67
“Estimating Delays in the Multiple-Input System” on page 3-71
“Estimating Model Orders Using an ARX Model Structure” on page 3-73
“Estimating Transfer Functions” on page 3-81
“Estimating Process Models” on page 3-84
“Estimating Black-Box Polynomial Models” on page 3-95
“Simulating and Predicting Model Output” on page 3-108
Introduction
Objectives
Estimate and validate linear models from multiple-input/single-output
(MISO) data to find the one that best describes the system dynamics.
tasks using the command line:
• Create data objects to represent data.

• Plot the data.
• Process data by removing offsets from the input and output signals.
• Estimate and validate linear models from the data.
• Simulate and predict model output.
3-55
Note This tutorial uses time-domain data to demonstrate how you

Data Description
This tutorial uses the data file co2data.mat, which contains two experiments
of two-input and single-output (MISO) time-domain data from a steady-state
that the operator perturbed from equilibrium values.
In the first experiment, the operator introduced a pulse wave to both inputs.
In the second experiment, the operator introduced a pulse wave to the first
input and a step signal to the second input.
Preparing Data
• “Plotting the Input/Output Data” on page 3-57
• “Removing Equilibrium Values from the Data” on page 3-60
• “Using Objects to Represent Data for System Identification” on page 3-60
• “Creating iddata Objects” on page 3-61
• “Plotting the Data in a Data Object” on page 3-64
3-56
• “Selecting a Subset of the Data” on page 3-66

Load the data.
load co2data;
This command loads the following five variables into the MATLAB Workspace:
• Input_exp1 and Output_exp1 are the input and output data from the first
experiment, respectively.
• Input_exp2 and Output_exp2 are the input and output data from the
second experiment, respectively.
• Time is the time vector from 0 to 1000 minutes, increasing in equal
increments of 0.5 min.
For both experiments, the input data consists of two columns of values. The
first column of values is the rate of chemical consumption (in kilograms per
minute), and the second column of values is the current (in amperes). The
output data is a single column of the rate of carbon-dioxide production (in
milligrams per minute).
Plotting the Input/Output Data

Plot the input and output data from both experiments.
plot(Time,Input_exp1,Time,Output_exp1)
legend('Input 1','Input 2','Output 1')
figure
plot(Time,Input_exp2,Time,Output_exp2)
legend('Input 1','Input 2','Output 1')
3-57
3-58
The first plot shows the first experiment, where the operator applies a pulse
wave to each input to perturb it from its steady-state equilibrium.
The second plot shows the second experiment, where the operator applies a
pulse wave to the first input and a step signal to the second input.
3-59
Removing Equilibrium Values from the Data

Plotting the data, as described in “Plotting the Input/Output Data” on page
3-57, shows that the inputs and the outputs have nonzero equilibrium values.
In this portion of the tutorial, you subtract equilibrium values from the data.
Zoom in on the plots to see that the earliest moment when the operator
applies a disturbance to the inputs occurs after 25 minutes of steady-state
conditions (or after the first 50 samples). Thus, the average value of the first
50 samples represents the equilibrium conditions.
Use the following commands to remove the equilibrium values from inputs
and outputs in both experiments:
Input_exp1 = Input_exp1-...
ones(size(Input_exp1,1),1)*mean(Input_exp1(1:50,:));
Output_exp1 = Output_exp1-...
mean(Output_exp1(1:50,:));
Input_exp2 = Input_exp2-...
ones(size(Input_exp2,1),1)*mean(Input_exp2(1:50,:));
Output_exp2 = Output_exp2-...
mean(Output_exp2(1:50,:));
Using Objects to Represent Data for System Identification

The System Identification Toolbox data objects, iddata and idfrd,
encapsulate data values and data properties into a single entity. You can use
the System Identification Toolbox commands to conveniently manipulate
these data objects as single entities.
In this portion of the tutorial, you create two iddata objects, one for each
of the two experiments. You use the data from Experiment 1 for model
estimation, and the data from Experiment 2 for model validation. You work
3-60
with two independent data sets because you use one data set for model
estimation and the other for model validation.
Note When two independent data sets are not available, you can split one
data set into two parts, assuming that each part contains enough information
to adequately represent the system dynamics.
Creating iddata Objects

You must have already loaded the sample data into the MATLAB workspace,
as described in “Loading Data into the MATLAB Workspace” on page 3-57.
Use these commands to create two data objects, ze and zv :
Ts = 0.5; % Sampling interval is 0.5 min

ze = iddata(Output_exp1,Input_exp1,Ts);
zv = iddata(Output_exp2,Input_exp2,Ts);
ze contains data from Experiment 1 and zv contains data from Experiment 2.

Ts is the sampling interval.
The iddata constructor requires three arguments for time-domain data and
has the following syntax:
data_obj = iddata(output,input,sampling_interval);
3-61
To view the properties of an iddata object, use the get command. For
example, type this command to get the properties of the estimation data:
get(ze)
get(ze)
ans =
Domain: 'Time'
Name: ''
OutputData: [2001x1 double]
y: 'Same as OutputData'
OutputName: {'y1'}
OutputUnit: {''}
InputData: [2001x2 double]
u: 'Same as InputData'
InputName: {2x1 cell}
InputUnit: {2x1 cell}
Period: [2x1 double]
InterSample: {2x1 cell}
Ts: 0.5000
Tstart: []
SamplingInstants: [2001x0 double]
TimeUnit: 'seconds'
ExperimentName: 'Exp1'
Notes: {}
UserData: []
To learn more about the properties of this data object, see the iddata
reference page.
3-62
To modify data properties, you can use dot notation or the set command. For
example, to assign channel names and units that label plot axes, type the
following syntax in the MATLAB Command Window:
% Set time units to minutes ze.TimeUnit = ’min’; % Set names of input

channels ze.InputName = {’ConsumptionRate’,’Current’}; % Set units for
input variables ze.InputUnit = {’kg/min’,’A’}; % Set name of output channel
ze.OutputName = ’Production’; % Set unit of output channel ze.OutputUnit =
’mg/min’; % Set validation data properties zv.TimeUnit = ’min’; zv.InputName
= {’ConsumptionRate’,’Current’}; zv.InputUnit = {’kg/min’,’A’}; zv.OutputName
= ’Production’; zv.OutputUnit = ’mg/min’;
% Set time units to minutes

ze.TimeUnit = 'min';
% Set names of input channels
ze.InputName = {'ConsumptionRate','Current'};
% Set units for input variables
ze.InputUnit = {'kg/min','A'};
% Set name of output channel
ze.OutputName = 'Production';
% Set unit of output channel
ze.OutputUnit = 'mg/min';
% Set validation data properties

zv.TimeUnit = 'min';
zv.InputName = {'ConsumptionRate','Current'};
zv.InputUnit = {'kg/min','A'};
zv.OutputName = 'Production';
zv.OutputUnit = 'mg/min';
3-63
You can verify that the InputName property of ze is changed, or "index" into
this property:
ze.inputname;
Tip Property names, such as InputUnit, are not case sensitive. You can also
abbreviate property names that start with Input or Output by substituting u
for Input and y for Output in the property name. For example, OutputUnit is
equivalent to yunit.
Plotting the Data in a Data Object

You can plot iddata objects using the plot command.
Plot the estimation data.
plot(ze)
3-64
The bottom axes show inputs ConsumptionRate and Current, and the top
axes show the output ProductionRate .
Plot the validation data in a new figure window.
figure % Open a new MATLAB Figure window

plot(zv) % Plot the validation data
3-65
Zoom in on the plots to see that the experiment process amplifies the first
input ( ConsumptionRate ) by a factor of 2, and amplifies the second input
(Current ) by a factor of 10.
Selecting a Subset of the Data

Before you begin, create a new data set that contains only the first 1000
samples of the original estimation and validation data sets to speed up the
calculations.
Ze1 = ze(1:1000);
3-66
Zv1 = zv(1:1000);
For more information about indexing into iddata objects, see the
Estimating Impulse Response Models

• “Why Estimate Step- and Frequency-Response Models?” on page 3-67
• “Estimating the Frequency Response” on page 3-67
• “Estimating the Empirical Step Response” on page 3-69
Why Estimate Step- and Frequency-Response Models?

Frequency-response and step-response are nonparametric models that can
help you understand the dynamic characteristics of your system. These
models are not represented by a compact mathematical formula with
adjustable parameters. Instead, they consist of data tables.
In this portion of the tutorial, you estimate these models using the data set
ze. You must have already created ze, as described in “Creating iddata
Objects” on page 3-61.
The response plots from these models show the following characteristics of
the system:
• The response from the first input to the output might be a second-order
function.
• The response from the second input to the output might be a first-order
or an overdamped function.
Estimating the Frequency Response

The System Identification Toolbox product provides three functions for
estimating the frequency response:
• etfe computes the empirical transfer function using Fourier analysis.

• spa estimates the transfer function using spectral analysis for a fixed
frequency resolution.
3-67
• spafdr lets you specify a variable frequency resolution for estimating the
frequency response.
Use spa to estimate the frequency response.
Ge=spa(ze);
Plot the frequency response as a Bode plot.
bode(Ge)
3-68
The amplitude peaks at the frequency of about 0.7 rad/s, which suggests
a possible resonant behavior (complex poles) for the first input-to-output
combination - ConsumptionRate to Production .
In both plots, the phase rolls off rapidly, which suggests a time delay for both
input/output combinations.
Estimating the Empirical Step Response

To estimate the step response from the data, first estimate a non-parametric
impulse response model (FIR filter) from data and then plot its step response.
% model estimation
Mimp = impulseest(Ze1,60);
% step response
step(Mimp)
3-69
The step response for the first input/output combination suggests an

overshoot, which indicates the presence of an underdamped mode (complex
poles) in the physical system.
The step response from the second input to the output shows no overshoot,
which indicates either a first-order response or a higher-order response with
real poles (overdamped response).
3-70
The step-response plot indicates a nonzero delay in the system, which is

consistent with the rapid phase roll-off you got in the Bode plot you created in
“Estimating the Empirical Step Response” on page 3-69.
Estimating Delays in the Multiple-Input System

• “Why Estimate Delays?” on page 3-71
• “Estimating Delays Using the ARX Model Structure” on page 3-71
• “Estimating Delays Using Alternative Methods” on page 3-72
Why Estimate Delays?

To identify parametric black-box models, you must specify the input/output
delay as part of the model order.
If you do not know the input/output delays for your system from the
experiment, you can use the System Identification Toolbox software to
estimate the delay.
Estimating Delays Using the ARX Model Structure

In the case of single-input systems, you can read the delay on the
impulse-response plot. However, in the case of multiple-input systems, such
as the one in this tutorial, you might be unable to tell which input caused the
initial change in the output and you can use the delayest command instead.
The delayest command estimates the time delay in a dynamic system by

estimating a low-order, discrete-time ARX model with a range of delays, and
then choosing the delay that corresponding to the best fit.
3-71
The ARX model structure is one of the simplest black-box parametric

structures. In discrete-time, the ARX structure is a difference equation with
the following form:
y(t) + a1 y(t − 1) +  + ana y(t − na ) =

the number of poles, nb is the number of b parameters (equal to the number
of zeros plus 1), nk is the number of samples before the input affects output
of the system (called the delay or dead time of the model), and e(t) is the
white-noise disturbance.
delayest assumes that na=nb=2 and that the noise e is white or insignificant,
and estimates nk.
To estimate the delay in this system, type:
delayest(ze)
ans =
5 10
This result includes two numbers because there are two inputs: the estimated
delay for the first input is 5 data samples, and the estimated delay for the
second input is 10 data samples. Because the sampling interval for the
experiments is 0.5 min, this corresponds to a 2.5 -min delay before the first
input affects the output, and a 5.0 -min delay before the second input affects
the output.
Estimating Delays Using Alternative Methods

There are two alternative methods for estimating the time delay in the system:
• Plot the time plot of the input and output data and read the time difference
between the first change in the input and the first change in the output.
This method is practical only for single-input/single-output system; in the
3-72
case of multiple-input systems, you might be unable to tell which input

caused the initial change in the output.
• Plot the impulse response of the data with a 1-standard-deviation
confidence region. You can estimate the time delay using the time when
the impulse response is first outside the confidence region.
Estimating Model Orders Using an ARX Model

Structure
• “Why Estimate Model Order?” on page 3-73
• “Commands for Estimating the Model Order” on page 3-74
• “Model Order for the First Input-Output Combination” on page 3-76
• “Model Order for the Second Input-Output Combination” on page 3-79
Why Estimate Model Order?

Model order is one or more integers that define the complexity of the model.
In general, model order is related to the number of poles, the number of zeros,
and the response delay (time in terms of the number of samples before the
output responds to the input). The specific meaning of model order depends
on the model structure.
To compute parametric black-box models, you must provide the model order
as an input. If you do not know the order of your system, you can estimate it.
After completing the steps in this section, you get the following results:
• For the first input/output combination: na=2, nb=2, and the delay nk=5.
• For the second input/output combination: na=1, nb=1, and the delay nk=10.
Later, you explore different model structures by specifying model-order values

that are slight variations around these initial estimate.
3-73
Commands for Estimating the Model Order

In this portion of the tutorial, you use struc, arxstruc, and selstruc to
estimate and compare simple polynomial (ARX) models for a range of model
orders and delays, and select the best orders based on the quality of the model.
The following list describes the results of using each command:
• struc creates a matrix of model-order combinations for a specified range of

na, nb, and nk values.
• arxstruc takes the output from struc, systematically estimates an
ARX model for each model order, and compares the model output to the
measured output. arxstruc returns the loss function for each model, which
is the normalized sum of squared prediction errors.
• selstruc takes the output from arxstruc and opens the ARX Model
Structure Selection window, which resembles the following figure, to help
you choose the model order.
You use this plot to select the best-fit model.
3-74
- The horizontal axis is the total number of parameters — na + nb.

- The vertical axis, called Unexplained output variance (in %), is
the portion of the output not explained by the model—the ARX model
prediction error for the number of parameters shown on the horizontal
axis.
The prediction error is the sum of the squares of the differences between
the validation data output and the model one-step-ahead predicted
output.
- nk is the delay.
Three rectangles are highlighted on the plot in green, blue, and red. Each
color indicates a type of best-fit criterion, as follows:
- Red — Best fit minimizes the sum of the squares of the difference
between the validation data output and the model output. This rectangle
indicates the overall best fit.
- Green — Best fit minimizes Rissanen MDL criterion.
- Blue — Best fit minimizes Akaike AIC criterion.
3-75
In this tutorial, the Unexplained output variance (in %) value remains

approximately constant for the combined number of parameters from 4 to
20. Such constancy indicates that model performance does not improve at
higher orders. Thus, low-order models might fit the data equally well.
Note When you use the same data set for estimation and validation, use
the MDL and AIC criteria to select model orders. These criteria compensate
for overfitting that results from using too many parameters. For more
information about these criteria, see the selstruc reference page.
Model Order for the First Input-Output Combination

In this tutorial, there are two inputs to the system and one output and you
estimate model orders for each input/output combination independently. You
can either estimate the delays from the two inputs simultaneously or one
input at a time.
It makes sense to try the following order combinations for the first
input/output combination:
• na=2:5
• nb=1:5
• n k= 5
This is because the nonparametric models you estimated in “Estimating

Impulse Response Models” on page 3-67 show that the response for the first
input/output combination might have a second-order response. Similarly, in
“Estimating Delays in the Multiple-Input System” on page 3-71, the delay for
this input/output combination was estimated to be 5.
To estimate model order for the first input/output combination:
1 Use struc to create a matrix of possible model orders.
NN1 = struc(2:5,1:5,5);
3-76
2 Use selstruc to compute the loss functions for the ARX models with the
orders in NN1.
selstruc(arxstruc(ze(:,:,1),zv(:,:,1),NN1))
Note (ze(:,:,1) selects the first input in the data.
3-77
This command opens the interactive ARX Model Structure Selection

window.
Note The Rissanen MDL and Akaike AIC criteria produces equivalent
results and are both indicated by a blue rectangle on the plot.
The red rectangle represents the best overall fit, which occurs for na=2,
nb=3, and nk=5. The height difference between the red and blue rectangles
is insignificant. Therefore, you can choose the parameter combination that
corresponds to the lowest model order and the simplest model.
3-78
3 Click the blue rectangle, and then click Select to choose that combination
of orders:
n a= 2
n b= 2
n k= 5
4 To continue, press any key while in the MATLAB Command Window.
Model Order for the Second Input-Output Combination

It makes sense to try the following order combinations for the second
• na=1:3
• nb=1:3
• nk=10
This is because the nonparametric models you estimated in “Estimating

Impulse Response Models” on page 3-67 show that the response for the second
input/output combination might have a first-order response. Similarly, in
“Estimating Delays in the Multiple-Input System” on page 3-71, the delay for
this input/output combination was estimated to be 10.
To estimate model order for the second input/output combination:
1 Use struc to create a matrix of possible model orders.
NN2 = struc(1:3,1:3,10);
3-79
2 Use selstruc to compute the loss functions for the ARX models with the
orders in NN2.
selstruc(arxstruc(ze(:,:,2),zv(:,:,2),NN2))
This command opens the interactive ARX Model Structure Selection

window.
Note The Akaike AIC and the overall best fit criteria produces equivalent
results. Both are indicated by the same red rectangle on the plot.
The height difference between all of the rectangles is insignificant and

all of these model orders result in similar model performance. Therefore,
you can choose the parameter combination that corresponds to the lowest
model order and the simplest model.
3-80
3 Click the yellow rectangle on the far left, and then click Select to choose
the lowest order: na=1, nb=1, and nk=10.
4 To continue, press any key while in the MATLAB Command Window.
Estimating Transfer Functions

• “Specifying the Structure of the Transfer Function” on page 3-81
• “Validating the Model” on page 3-83
• “Residual Analysis” on page 3-83
Specifying the Structure of the Transfer Function

In this portion of the tutorial, you estimate a continuous-time transfer
function. You must have already prepared your data, as described in
“Preparing Data” on page 3-56. You can use the following results of estimated
model orders to specify the orders of the model:
• For the first input/output combination, use:

- Two poles, corresponding to na=2 in the ARX model.
- Delay of 5, corresponding to nk=5 samples (or 2.5 minutes) in the ARX
model.
• For the second input/output combination, use:
- One pole, corresponding to na=1 in the ARX model
- Delay of 10, corresponding to nk=10 samples (or 5 minutes) in the ARX
model.
You can estimate a transfer function of these orders using the tfest
command. For the estimation, you can also choose to view a progress
report by setting the Display option to on in the option set created by the
tfestOptions command.
Opt = tfestOptions('Display', 'on');
Collect the model orders and delays into variables to pass to tfest.
np = [2 1];
3-81
ioDelay = [2.5 5];
Estimate the transfer function.
mtf = tfest(Ze1, np, [], ioDelay, Opt);
View the model’s coefficients.
mtf
mtf =
From input "ConsumptionRate" to output "Production":

91.75 s - 1.272
exp(-2.5*s) * ---------------------
s^2 + 37.76 s + 1.009
From input "Current" to output "Production":

5.186
exp(-5*s) * ----------
s + 0.5066
Continuous-time identified transfer function.
Parameterization:
Number of poles: [2 1] Number of zeros: [1 0]
Use "tfdata", "getpvec", "getcov" for parameters and their uncertainties
Status:
Estimated using TFEST on time domain data "Ze1".
Fit to estimation data: 85.89% (simulation focus)
FPE: 6.424, MSE: 6.265
The model’s display shows more than 85% fit to estimation data.
3-82
Validating the Model

In this portion of the tutorial, you create a plot that compares the actual
output and the model output using the compare command.
compare(Zv1, mtf)
The comparison shows about 60% fit.
Residual Analysis
Use the resid command to evaluate the characteristics of the residuals.
3-83
resid(Zv1, mtf)
The residuals show high degree of auto-correlation. This is not unexpected

since the model mtf does not have any components to describe the nature of
the noise separately. To model both the measured input-output dynamics
and the disturbance dynamics you will need to use a model structure that
contains elements to describe the noise component. You can use bj, ssest
and procest commands, which create models of polynomial, state-space and
process structures respectively. These structures, among others, contain
elements to capture the noise behavior.
Estimating Process Models

• “Specifying the Structure of the Process Model” on page 3-85
• “Viewing the Model Structure and Parameter Values” on page 3-86
• “Specifying Initial Guesses for Time Delays” on page 3-87
• “Estimating Model Parameters Using procest” on page 3-87
3-84

• “Estimating a Process Model with Noise Model” on page 3-93
Specifying the Structure of the Process Model

In this portion of the tutorial, you estimate a low-order, continuous-time
transfer function (process model). the System Identification Toolbox product
supports continuous-time models with at most three poles (which might
contain underdamped poles), one zero, a delay element, and an integrator.
You must have already prepared your data, as described in “Preparing Data”
on page 3-56.
You can use the following results of estimated model orders to specify the
orders of the model:

model.
- One pole, corresponding to na=1 in the ARX model.
model.
Note Because there is no relationship between the number of zeros estimated

by the discrete-time ARX model and its continuous-time counterpart, you
do not have an estimate for the number of zeros. In this tutorial, you can
specify one zero for the first input/output combination, and no zero for the
second-output combination.
Use the idproc command to create two model structures, one for each
midproc0 = idproc({'P2ZUD','P1D'}, 'TimeUnit', 'minutes');
3-85
idproc accepts a cell array that contains two strings that specify the model
structure for each input/output combination:
• 'P2ZUD' represents a transfer function with two poles ( P2 ), one zero ( Z ),

underdamped (complex-conjugate) poles ( U ) and a delay ( D ).
• 'P1D' represents a transfer function with one pole ( P1 ) and a delay ( D ).
The example also uses the TimeUnit parameter to specify the unit of time
used.
Viewing the Model Structure and Parameter Values

View the two resulting models.
midproc0
midproc0 =
Process model with 2 inputs: y = G11(s)u1 + G12(s)u2
From input 1 to output 1:
1+Tz*s
G11(s) = Kp * ---------------------- * exp(-Td*s)
1+2*Zeta*Tw*s+(Tw*s)^2
Kp = NaN
Tw = NaN
Zeta = NaN
Td = NaN
Tz = NaN
From input 2 to output 1:

Kp
G12(s) = ---------- * exp(-Td*s)
1+Tp1*s
Kp = NaN
Tp1 = NaN
Td = NaN
3-86
Parameterization:
'P2DUZ' 'P1D'
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Created by direct construction or transformation. Not estimated.
The parameter values are set to NaN because they are not yet estimated.
Specifying Initial Guesses for Time Delays

Set the time delay property of the model object to 2.5 min and 5 min for each
input/output pair as initial guesses. Also, set an upper limit on the delay
because good initial guesses are available.
midproc0.Structure(1,1).Td.Value=2.5;
midproc0.Structure(1,2).Td.Value=5;
midproc0.Structure(1,1).Td.Maximum=3;
midproc0.Structure(1,2).Td.Maximum=7;
Note When setting the delay constraints, you must specify the delays in
terms of actual time units (minutes, in this case) and not the number of
samples.
Estimating Model Parameters Using procest

procest is an iterative estimator of process models, which means that it uses
an iterative nonlinear least-squares algorithm to minimize a cost function.
The cost function is the weighted sum of the squares of the errors.
Depending on its arguments, procest estimates different black-box

polynomial models. You can use procest, for example, to estimate parameters
for linear continuous-time transfer-function, state-space, or polynomial
model structures. To use procest, you must provide a model structure with
unknown parameters and the estimation data as input arguments.
For this portion of the tutorial, you must have already defined the model
structure, as described in “Specifying the Structure of the Process Model” on
3-87
page 3-85. Use midproc0 as the model structure and Ze1 as the estimation
data:
midproc = procest(Ze1, midproc0);

present(midproc)
midproc =
Process model with 2 inputs: y = G11(s)u1 + G12(s)u2
From input "ConsumptionRate" to output "Production":
1+Tz*s
G11(s) = Kp * ---------------------- * exp(-Td*s)
1+2*Zeta*Tw*s+(Tw*s)^2
Kp = -1.1804 +/- 0.29988

Tw = 1.6566 +/- 629.69
Zeta = 15.917 +/- 6039.2
Td = 2.425 +/- 56.972
Tz = -109.2 +/- 57.371
From input "Current" to output "Production":

Kp
G12(s) = ---------- * exp(-Td*s)
1+Tp1*s
Kp = 10.264 +/- 0.048403

Tp1 = 2.0488 +/- 0.054899
Td = 4.9175 +/- 0.034373
Parameterization:
'P2DUZ' 'P1D'
Use "getpvec", "getcov" for parameters and their uncertainties.
Status:
Termination condition: Maximum number of iterations reached.
Estimated using PROCEST on time domain data "Ze1".

Fit to estimation data: 86.2%
3-88
FPE: 6.08, MSE: 5.99

Unlike discrete-time polynomial models, continuous-time models let you

estimate the delays. In this case, the estimated delay values are different
from the initial guesses you specified of 2.5 and 5 , respectively. The large
uncertainties in the estimated values of the parameters of G_1(s) indicate
that the dynamics from input 1 to the output are not captured well.
To learn more about estimating models, see “Process Models”.

In this portion of the tutorial, you create a plot that compares the actual
output and the model output.
compare(Zv1,midproc)
3-89
The plot shows that the model output reasonably agrees with the measured
output: there is an agreement of 60% between the model and the validation
data.
Use resid to perform residual analysis:
resid(Zv1,midproc)
3-90
Because the sample system has two inputs, there are two cross-correlation
plots of the residuals with each input, as shown in the following figure.
Autocorrelation and Cross-Correlations of Residuals with the First Input
3-91
The cross-correlation between the second input and residual errors is

significant.
After MATLAB displays the first plot, press Enter to view the
cross-correlation with the second input, as shown in the following figure.
Cross-Correlations of Residuals with the Second Input
In the preceding figure, the autocorrelation plot shows values outside the
confidence region and indicates that the residuals are correlated.
Change the algorithm for iterative parameter estimation to

Levenberg-Marquardt.
Opt = procestOptions;
Opt.SearchMethod = 'lm';
3-92
midproc1 = procest(Ze1, midproc, Opt);
Tweaking the algorithm properties or specifying initial parameter guesses

and rerunning the estimation may improve the simulation results. Adding a
noise model may improve prediction results but not necessarily the simulation
results.
Estimating a Process Model with Noise Model

This portion of the tutorial shows how to estimate a process model and include
a noise model in the estimation. Including a noise model typically improves
model prediction results but not necessarily the simulation results.
Use the following commands to specify a first-order ARMA noise:
Opt = procestOptions;
Opt.DisturbanceModel = 'ARMA1';
midproc2 = procest(Ze1, midproc0, Opt);
You can type 'dist' instead of 'DisturbanceModel'. Property names are

not case sensitive, and you only need to include the portion of the name that
uniquely identifies the property.
Compare the resulting model to the measured data.
compare(Zv1,midproc2)
3-93
The plot shows that the model output maintains reasonable agreement with
the validation-data output.
Perform residual analysis.
figure
resid(Zv1,midproc2)
Press Enter to view the cross-correlation of the residuals with the second
input.
3-94
The next plot shows that adding a noise model produces uncorrelated
residuals: the top set of axes show that the autocorrelation values are inside
the confidence bounds. This indicates a more accurate model.
Estimating Black-Box Polynomial Models

• “Model Orders for Estimating Polynomial Models” on page 3-95
• “Estimating a Linear ARX Model” on page 3-96
• “Estimating State-Space Models” on page 3-99
• “Estimating a Box-Jenkins Model” on page 3-103
• “Comparing Model Output to Measured Output” on page 3-106
Model Orders for Estimating Polynomial Models

In this portion of the tutorial, you estimate several different types of
black-box, input-output polynomial models.
3-95
You must have already prepared your data, as described in “Preparing Data”
on page 3-56.
You can use the following previous results of estimated model orders to specify
the orders of the polynomial model:

- One zero, corresponding to nb=2 in the ARX model.
model.
- One pole, corresponding to na=1 in the ARX model.
- No zeros, corresponding to nb=1 in the ARX model.
model.
Estimating a Linear ARX Model
• “About ARX Models” on page 3-96

• “Estimating ARX Models Using arx” on page 3-97
• “Accessing Model Data” on page 3-98
About ARX Models. For a single-input/single-output system (SISO), the

ARX model structure is:
y(t) + a1 y(t − 1) +  + ana y(t − na ) =

the number of poles, nb is the number of zeros plus 1, nk is the number of
samples before the input affects output of the system (called the delay or dead
time of the model), and e(t) is the white-noise disturbance.
3-96
The ARX model structure does not distinguish between the poles for
individual input/output paths: dividing the ARX equation by A, which
contains the poles, shows that A appears in the denominator for both inputs.
Therefore, you can set na to the sum of the poles for each input/output pair,
which is equal to 3 in this case.
The System Identification Toolbox product estimates the parameters a1  an

and b1  bn using the data and the model orders you specify.
Estimating ARX Models Using arx. Use arx to compute the polynomial
coefficients using the fast, noniterative method arx:
marx = arx(Ze1,'na',3,'nb',[2 1],'nk',[5 10]);

present(marx) % Displays model parameters
% with uncertainty information
marx =
Discrete-time ARX model: A(z)y(t) = B(z)u(t) + e(t)
A(z) = 1 - 1.027 (+/- 0.02917) z^-1 + 0.1675 (+/- 0.04214) z^-2

+ 0.01307 (+/- 0.02591) z^-3
B1(z) = 1.86 (+/- 0.1896) z^-5 - 1.608 (+/- 0.1894) z^-6
B2(z) = 1.612 (+/- 0.07417) z^-10
Sample time: 0.5 minutes
Parameterization:
Polynomial orders: na=3 nb=[2 1] nk=[5 10]
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using ARX on time domain data "Ze1".
FPE: 2.75, MSE: 2.722
3-97
MATLAB estimates the polynomials A , B1 , and B2. The uncertainty for each
of the model parameters is computed to 1 standard deviation and appears in
Alternatively, you can use the following shorthand syntax and specify model
orders as a single vector:
marx = arx(Ze1,[3 2 1 5 10]);
Accessing Model Data. The model you estimated, marx, is a discrete-time

idpoly object. To get the properties of this model object, you can use the
get function:
get(marx)
a: [1 -1.0266 0.1675 0.0131]

b: {[0 0 0 0 0 1.8600 -1.6085] [0 0 0 0 0 0 0 0 0 0 1.611
c: 1
d: 1
f: {[1] [1]}
IntegrateNoise: 0
Variable: 'z^-1'
ioDelay: [0 0]
Structure: [1x1 pmodel.polynomial]
NoiseVariance: 2.7611
Report: [1x1 idresults.arx]
InputDelay: [2x1 double]
OutputDelay: 0
Ts: 0.5000
TimeUnit: 'minutes'
InputName: {2x1 cell}
InputUnit: {2x1 cell}
InputGroup: [1x1 struct]
OutputName: {'Production'}
OutputUnit: {'mg/min'}
OutputGroup: [1x1 struct]
Name: ''
Notes: {}
UserData: []
3-98
SamplingGrid: [1x1 struct]
You can access the information stored by these properties using dot notation.
For example, you can compute the discrete poles of the model by finding the
roots of the A polynomial.
marx_poles=roots(marx.a)
marx_poles =
0.7953
0.2883
-0.0570
In this case, you access the A polynomial using marx.a.
The model marx describes system dynamics using three discrete poles.
Tip You can also use pole to compute the poles of a model directly.
Learn More. To learn more about estimating polynomial models, see

“Input-Output Polynomial Models”.
For more information about accessing model data, see “Data Extraction”.
Estimating State-Space Models
• “About State-Space Models” on page 3-100

• “Estimating State-Space Models Using n4sid” on page 3-100
3-99
About State-Space Models. The general state-space model structure is:
dx
 Ax(t)  Bu(t)  Ke(t)
dt
y(t)  Cx(t)  Du(t)  e(t)
y(t) represents the output at time t, u(t) represents the input at time t, x(t) is
the state vector at time t, and e(t) is the white-noise disturbance.
You must specify a single integer as the model order (dimension of the state
vector) to estimate a state-space model. By default, the delay equals 1.
The System Identification Toolbox product estimates the state-space matrices

A, B, C, D, and K using the model order and the data you specify.
The state-space model structure is a good choice for quick estimation because
it contains only two parameters: n is the number of poles (the size of the A
matrix) and nk is the delay.
Estimating State-Space Models Using n4sid. Use the n4sid command

to specify a range of model orders and evaluate the performance of several
state-space models (orders 2 to 8):
mn4sid = n4sid(Ze1,2:8,'InputDelay',[4 9]);
3-100
This command uses the fast, noniterative (subspace) method and opens the
following plot. You use this plot to decide which states provide a significant
relative contribution to the input/output behavior, and which states provide
the smallest contribution.
The vertical axis is a relative measure of how much each state contributes to
the input/output behavior of the model (log of singular values of the covariance
matrix). The horizontal axis corresponds to the model order n. This plot
recommends n=4, indicated by a red rectangle.
To select this model order, select 4 from the Model Order drop-down list
and click Apply.
By default, n4sid uses a free parameterization of the state-space form. To

estimate a canonical form instead, set the value of the SSParameterization
property to 'Canonical' . You can also specify the input-to-output delay (in
samples) using the 'InputDelay' property.
mCanonical = n4sid(Ze1, 3,'SSParameterization', 'canonical','InputDelay', [

present(mCanonical); % Display model properties
3-101
mCanonical =
Discrete-time identified state-space model:
x(t+Ts) = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)
A =
x1 x2 x3
x1 0 1 0
x2 0 0 1
x3 0.0738 +/- 0.05922 -0.6083 +/- 0.1624 1.445 +/- 0.1284
B =
ConsumptionR Current
x1 1.844 +/- 0.1756 0.5631 +/- 0.1225
x2 1.064 +/- 0.1679 2.309 +/- 0.1226
x3 0.2769 +/- 0.0966 1.878 +/- 0.1061
C =
x1 x2 x3
Production 1 0 0
D =
ConsumptionR Current
Production 0 0
K =
Production
x1 0.8676 +/- 0.03183
x2 0.6864 +/- 0.04166
x3 0.5113 +/- 0.04379
Input delays (sampling periods): 4 9
Parameterization:
CANONICAL form with indices: 3.
Feedthrough: none
Disturbance component: estimate
3-102
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Estimated using N4SID on time domain data "Ze1".
FPE: 2.456, MSE: 2.341
Note mn4sid and mCanonical are discrete-time models. To estimate a

continuous-time model, set the 'Ts' property to 0 in the estimation command,
or use the ssest command:
mCT1 = n4sid(Ze1, 3, 'Ts', 0, 'InputDelay', [2.5 5])

mCT2 = ssest(Ze1, 3,'InputDelay', [2.5 5])
Learn More. To learn more about estimating state-space models, see

“State-Space Models”.
Estimating a Box-Jenkins Model
• “About Box-Jenkins Models” on page 3-103

• “Estimating a BJ Model Using pem” on page 3-104
About Box-Jenkins Models. The general Box-Jenkins (BJ) structure is:
nu
Bi (q) C(q)
y(t) = ∑ ui ( t − nki ) + e(t)
F
i=1 i
( q) D(q)
To estimate a BJ model, you need to specify the parameters nb, nf, nc, nd,
and nk.
Whereas the ARX model structure does not distinguish between the poles
for individual input/output paths, the BJ model provides more flexibility in
3-103
modeling the poles and zeros of the disturbance separately from the poles
and zeros of the system dynamics.
Estimating a BJ Model Using pem. You can use pem to estimate the BJ
model. pem is an iterative method and has the following general syntax:
pem(data,'na',na,'nb',nb,'nc',nc,'nd',nd,'nf',nf,'nk',nk)
3-104
To estimate the BJ model, type:
na = 0;
nb = [ 2 1 ];
nc = 1;
nd = 1;
nf = [ 2 1 ];
nk = [ 5 10];
mbj = polyest(Ze1,[na nb nc nd nf nk]);
This command specifies nf=2 , nb=2 , nk=5 for the first input, and nf=nb=1
and nk=10 for the second input.
Display the model information.
present(mbj)
mbj =
Discrete-time BJ model: y(t) = [B(z)/F(z)]u(t) + [C(z)/D(z)]e(t)
B1(z) = 1.823 (+/- 0.1792) z^-5 - 1.315 (+/- 0.2368) z^-6
B2(z) = 1.791 (+/- 0.06435) z^-10
C(z) = 1 + 0.1066 (+/- 0.04015) z^-1
D(z) = 1 - 0.7453 (+/- 0.027) z^-1
F1(z) = 1 - 1.321 (+/- 0.06939) z^-1 + 0.5911 (+/- 0.05516) z^-2
F2(z) = 1 - 0.8314 (+/- 0.006445) z^-1
Parameterization:
Polynomial orders: nb=[2 1] nc=1 nd=1 nf=[2 1]
nk=[5 10]
Use "polydata", "getpvec", "getcov" for parameters and their uncertainti
Status:
Termination condition: Near (local) minimum, (norm(g) < tol).
3-105
Estimated using POLYEST on time domain data "Ze1".
FPE: 2.757, MSE: 2.692
Tip Alternatively, you can use the following shorthand syntax that specifies
the orders as a single vector:
mbj = bj(Ze1,[2 1 1 1 2 1 5 10]);
bj is a version of pem that specifically estimates the BJ model structure.
Learn More. To learn more about identifying input-output polynomial

models, such as BJ, see “Input-Output Polynomial Models”.
Comparing Model Output to Measured Output

Compare the output of the ARX, state-space, and Box-Jenkins models to the
measured output.
compare(Zv1,marx,mn4sid,mbj)
3-106
compare plots the measured output in the validation data set against the
simulated output from the models. The input data from the validation data
set serves as input to the models.
To perform residual analysis on the ARX model, type the following command:
resid(Zv1,marx)
Because the sample system has two inputs, there are two plots showing the
cross-correlation of the residuals with each input. Press Enter to view the
cross-correlation with the second input.
3-107
To perform residual analysis on the state-space model, type the following

command:
resid(Zv1,mn4sid)
Finally, to perform residual analysis on the BJ model, type the following

command:
resid(Zv1,mbj)
All three models simulate the output equally well and have uncorrelated
residuals. Therefore, choose the ARX model because it is the simplest of the
three input-output polynomial models and adequately captures the process
dynamics.
Simulating and Predicting Model Output

• “Simulating the Model Output” on page 3-108
• “Predicting the Future Output” on page 3-110
Simulating the Model Output

In this portion of the tutorial, you simulate the model output. You must
have already created the continuous-time model midproc2, as described in
“Estimating Process Models” on page 3-84.
Simulating the model output requires the following information:
• Input values to the model

• Initial conditions for the simulation (also called initial states)
For example, the following commands use the iddata and idinput commands
to construct an input data set, and use sim to simulate the model output:
% Create input for simulation

U = iddata([],idinput([200 2]),'Ts',0.5,'TimeUnit','min');
% Simulate the response setting initial conditions equal to zero
ysim_1 = sim(midproc2,U);
3-108
To maximize the fit between the simulated response of a model to the

measured output for the same input, you can compute the initial conditions
corresponding to the measured data. The best fitting initial conditions can
be obtained by using findstates(idParametric) on the state-space version
of the estimated model. The following commands estimate the initial states
X0est from the data set Zv1:
% State-space version of the model needed for computing initial states

midproc2_ss = idss(midproc2);
X0est = findstates(midproc2_ss, Zv1);
Next, simulate the model using the initial states estimated from the data.
% Simulation input
Usim = Zv1(:,[],:);
Opt = simOptions('InitialCondition',X0est);
ysim_2 = sim(midproc2_ss, Usim, Opt);
Compare the simulated and the measured output on a plot.
figure
plot([ysim_2.y, Zv1.y])
legend({'model output','measured'})
xlabel('time'), ylabel('Output')
3-109
Predicting the Future Output

Many control-design applications require you to predict the future outputs of
a dynamic system using the past input/output data.
3-110
For example, use predict to predict the model response five steps ahead:
predict(midproc2,Ze1,5)
3-111
Compare the predicted output values with the measured output values. The
third argument of compare specifies a five-step-ahead prediction. When you
do not specify a third argument, compare assumes an infinite prediction
horizon and simulates the model output instead.
compare(Ze1,midproc2,5)
3-112
Use pe to compute the prediction error Err between the predicted output of
midproc2 and the measured output. Then, plot the error spectrum using the
spectrum command.
[Err] = pe(midproc2,Zv1);
spectrum(spa(Err,[],logspace(-2,2,200)))
The prediction errors are plotted with a 1-standard-deviation confidence

interval. The errors are greater at high frequencies because of the
high-frequency nature of the disturbance.
3-113
Identify Low-Order Transfer Functions (Process Models)

Using System Identification Tool
In this section...
“What Is a Continuous-Time Process Model?” on page 3-115
“Preparing Data for System Identification” on page 3-116
“Estimating a Second-Order Transfer Function (Process Model) with
Complex Poles” on page 3-124
“Estimating a Process Model with a Noise Component” on page 3-132
“Viewing Model Parameters” on page 3-139
“Exporting the Model to the MATLAB Workspace” on page 3-141
“Simulating a System Identification Toolbox Model in Simulink Software”
on page 3-142
Introduction
Objectives
Estimate and validate simple, continuous-time transfer functions from
single-input/single-output (SISO) data to find the one that best describes
the system dynamics.
tasks using the System Identification Tool :
• Import data objects from the MATLAB workspace into the GUI.
• Plot and process the data.
• Estimate and validate low-order, continuous-time models from the data.
• Export models to the MATLAB workspace.
3-114
Identify Low-Order Transfer Functions (Process Models) Using System Identification Tool
• Simulate the model using Simulink software.
Note This tutorial uses time-domain data to demonstrate how you

Data Description
This tutorial uses the data file proc_data.mat, which contains 200 samples
of simulated single-input/single-output (SISO) time-domain data. The input
is a random binary signal that oscillates between -1 and 1. White noise
(corresponding to a load disturbance) is added to the input with a standard
deviation of 0.2, which results in a signal-to-noise ratio of about 20 dB. This
data is simulated using a second-order system with underdamped modes
(complex poles) and a peak response at 1 rad/s:
1
G(s) = e−2 s
2
1 + 0 .2 s + s
The sampling interval of the simulation is 1 second.
What Is a Continuous-Time Process Model?

Continuous-time process models are low-order transfer functions that describe
the system dynamics using static gain, a time delay before the system output
responds to the input, and characteristic time constants associated with poles
and zeros. Such models are popular in the industry and are often used for
tuning PID controllers, for example. Process model parameters have physical
significance.
You can specify different process model structures by varying the number of
poles, adding an integrator, or including a time delay or a zero. The highest
process model order you can specify in this toolbox is three, and the poles can
be real or complex (underdamped modes).
3-115
In general, a linear system is characterized by a transfer function G, which is

an operator that takes the input u to the output y:
y = Gu
For a continuous-time system, G relates the Laplace transforms of the input

U(s) and the output Y(s), as follows:
Y (s) = G(s)U (s)
In this tutorial, you estimate G using different process-model structures.
For example, the following model structure is a first-order, continuous-time

model, where K is the static gain, Tp1 is a time constant, and Td is the
input-to-output delay:
K
G(s) = e− sTd
1 + sTp1
Preparing Data for System Identification

• “Opening the System Identification Tool ” on page 3-117
• “Importing Data Objects into the System Identification Tool” on page 3-117
• “Plotting and Processing Data” on page 3-120

Load the data in proc_data.mat by typing the following command in the
load proc_data
This command loads the data into the MATLAB workspace as the data
object z. For more information about iddata objects, see the corresponding
reference page.
3-116
Opening the System Identification Tool

To open the System Identification Tool , type the following command
ident
Importing Data Objects into the System Identification Tool

You can import data object into the GUI from the MATLAB workspace.
You must have already loaded the sample data into MATLAB, as described in
“Loading Data into the MATLAB Workspace” on page 3-116, and opened the
GUI, as described in “Opening the System Identification Tool ” on page 3-117.
To import a data object into the System Identification Tool :
3-117
1 In the System Identification Tool , select Import data > Data object.
This action opens the Import Data dialog box.
2 In the Import Data dialog box, specify the following options:
• Object — Enter z as the name of the MATLAB variable that is the

time-domain data object. Press Enter.
• Data name — Use the default name z, which is the same as the name
of the data object you are importing. This name labels the data in the
System Identification Tool after the import operation is completed.
• Starting time — Enter 0 as the starting time. This value designates
the starting value of the time axis on time plots.
• Sampling interval — Enter 1 as the time between successive samples
in seconds. This value represents the actual sampling interval in the
experiment.
The Import Data dialog box now resembles the following figure.
3-118
3-119
3 Click Import to add the data to the System Identification Tool. The tool
adds an icon to represent the data.
4 Click Close to close the Import Data dialog box.
Plotting and Processing Data

In this portion of the tutorial, you evaluate the data and process it for system
identification. You learn how to:
• Plot the data.

• Remove offsets by subtracting the mean values of the input and the output.
• Split the data into two parts. You use one part of the data for model
estimation, and the other part of the data for model validation.
You must have already imported data into the System Identification Tool, as
described in “Importing Data Objects into the System Identification Tool” on
page 3-117.
3-120
To plot and process the data:
1 In the System Identification Tool, select the Time plot check box to open
the Time Plot window.
The bottom axes show the input data—a random binary sequence, and
the top axes show the output data.
The next two steps demonstrate how to modify the axis limits in the plot.
2 To modify the vertical-axis limits for the input data, select Options > Set
axes limits in the Time Plot figure window.
3-121
3 In the Limits for Time Plot dialog box, set the new vertical axis limit of the
input data channel u1 to [-1.5 1.5]. Click Apply and Close.
Note The other two fields in the Limits for Time Plot dialog box, Time
and y1, let you set the axis limits for the time axis and the output channel
axis, respectively. You can also specify each axis to be logarithmic or linear
by selecting the corresponding option.
3-122
The following figure shows the updated time plot.
4 In the System Identification Tool , select <--Preprocess > Quick start to

perform the following four actions:
• Subtract the mean value from each channel.
• Split the data into two parts.
• Specify the first part of the data as estimation data (or Working Data).
• Specify the second part of the data as Validation Data.
Learn More. For information about supported data processing operations,

such as resampling and filtering the data, see “Preprocess Data”.
3-123
Estimating a Second-Order Transfer Function (Process

Model) with Complex Poles
• “Estimating a Second-Order Transfer Function Using Default Settings” on
page 3-124
• “Tips for Specifying Known Parameters” on page 3-128
Estimating a Second-Order Transfer Function Using Default

Settings
In this portion of the tutorial, you estimate models with this structure:
K
G(s) = e−Td s
( 1 + 2ξTw s + Tw2 s2 )
You must have already processed the data for estimation, as described in
“Plotting and Processing Data” on page 3-120.
To identify a second-order transfer function:
1 In the System Identification Tool, select Estimate > Process models to

open the Process Models dialog box.
3-124
2 In the Model Transfer Function area of the Process Models dialog box,
specify the following options:
• Under Poles, select 2 and Underdamped.
This selection updates the Model Transfer Function to a second-order
model structure that can contain complex poles.
• Make sure that the Zero and Integrator check boxes are cleared to
exclude a zero and an integrator (self-regulating ) from the model.
3 The Parameter area of the Process Models dialog box now shows four
active parameters: K, Tw, Zeta, and Td. In the Initial Guess area, keep
the default Auto-selected option to calculate the initial parameter values
during the estimation. The Initial Guess column in the Parameter table
displays Auto.
4 Keep the default Bounds values, which specify the minimum and
maximum values of each parameter.
3-125
Tip If you know the range of possible values for a parameter, you can type
these values into the corresponding Bounds fields to help the estimation
algorithm.
5 Keep the default settings for the estimation algorithm:
• Disturbance Model — None means that the algorithm does not

estimate the noise model. This option also sets the Focus to Simulation.
• Focus — Simulation means that the estimation algorithm does not use
the noise model to weigh the relative importance of how closely to fit
the data in various frequency ranges. Instead, the algorithm uses the
input spectrum in a particular frequency range to weigh the relative
importance of the fit in that frequency range.
Tip The Simulation setting is optimized for identifying models that

you plan to use for output simulation. If you plan to use your model
for output prediction or control applications, or to improve parameter
estimates using a noise model, select Prediction.
• Initial condition — Auto means that the algorithm analyzes the data
and chooses the optimum method for handling the initial state of the
system. If you get poor results, you might try setting a specific method
for handling initial states, rather than choosing it automatically.
• Covariance — Estimate means that the algorithm computes parameter
uncertainties that display as model confidence regions on plots.
The tool assigns a name to the model, shown in the Name field (located at
the bottom of the dialog box). By default, the name is the acronym P2DU,
which indicates two poles (P2), a delay (D), and underdamped modes (U).
3-126
6 Click Estimate to add the model P2DU to the System Identification Tool.
3-127
Tips for Specifying Known Parameters

If you know a parameter value exactly, you can type this value in the Initial
Guess column of the Process Models dialog box.
If you know the approximate value of a parameter, you can help the estimation
algorithm by entering an initial value in the Initial Guess column. In this
case, keep the Known check box cleared to allow the estimation to fine-tune
this initial guess.
For example, to fix the time-delay value Td at 2s, you can type this value into
Value field of the Parameter table in the Process Models dialog box and select
the corresponding Known check box.
Known checkboxes.

You can analyze the following plots to evaluate the quality of the model:
3-128
• Comparison of the model output and the measured output on a time plot
• Autocorrelation of the output residuals, and cross-correlation of the input
and the output residuals
You must have already estimated the model, as described in “Estimating a

Second-Order Transfer Function Using Default Settings” on page 3-124.
Examining Model Output. You can use the model-output plot to check how
well the model output matches the measured output in the validation data
set. A good model is the simplest model that best describes the dynamics and
successfully simulates or predicts the output for different inputs.
To generate the model-output plot, select the Model output check box in the
System Identification Tool. If the plot is empty, click the model icon in the
System Identification Tool window to display the model on the plot.
The System Identification Toolbox product uses input validation data as input
to the model, and plots the simulated output on top of the output validation
3-129
data. The preceding plot shows that the model output agrees well with the
validation-data output.
The Best Fits area of the Model Output plot shows the agreement (in percent)
between the model output and the validation-data output.
Recall that the data was simulated using the following second-order system
with underdamped modes (complex poles), as described in “Data Description”
on page 3-115, and has a peak response at 1 rad/s:
1
G(s) = e−2 s
1 + 0.2s + s2
Because the data includes noise at the input during the simulation, the
estimated model cannot exactly reproduce the model used to simulate the
data.
Examining Model Residuals. You can validate a model by checking the

behavior of its residuals.
To generate a Residual Analysis plot, select the Model resids check box in
the System Identification Tool.
3-130
The top axes show the autocorrelation of residuals for the output (whiteness
test). The horizontal scale is the number of lags, which is the time difference
(in samples) between the signals at which the correlation is estimated. Any
fluctuations within the confidence interval are considered to be insignificant.
A good model should have a residual autocorrelation function within the
confidence interval, indicating that the residuals are uncorrelated. However,
in this example, the residuals appear to be correlated, which is natural
because the noise model is used to make the residuals white.
The bottom axes show the cross-correlation of the residuals with the
input. A good model should have residuals uncorrelated with past inputs
(independence test). Evidence of correlation indicates that the model does not
describe how a portion of the output relates to the corresponding input. For
example, when there is a peak outside the confidence interval for lag k, this
means that the contribution to the output y(t) that originates from the input
u(t-k) is not properly described by the model. In this example, there is no
correlation between the residuals and the inputs.
3-131
Thus, residual analysis indicates that this model is good, but that there might
be a need for a noise model.
Estimating a Process Model with a Noise Component

• “Estimating a Second-Order Process Model with Complex Poles” on page
3-132
• “Validating the Models” on page 3-134
Estimating a Second-Order Process Model with Complex Poles

In this portion of the tutorial, you estimate a second-order transfer function
and include a noise model. By including a noise model, you optimize the
estimation results for prediction application.
You must have already estimated the model, as described in “Estimating a

Second-Order Transfer Function Using Default Settings” on page 3-124.
To estimate a second-order transfer function with noise:
1 If the Process Models dialog box is not open, select Estimate > Process
Models in the System Identification Tool. This action opens the Process
Models dialog box.
3-132
2 In the Model Transfer Function area, specify the following options:
• Under Poles, select 2 and Underdamped. This selection updates the

Model Transfer Function to a second-order model structure that can
contain complex poles. Make sure that the Zero and Integrator check
boxes are cleared to exclude a zero and an integrator (self-regulating )
from the model.
• Disturbance Model — Set to Order 1 to estimate a noise model H as a

continuous-time, first-order ARMA model:
C
H= e
D
where and D are first-order polynomials, and e is white noise.

This action specifies the Focus as Prediction, which improves accuracy
in the frequency range where the noise level is low. For example, if there
is more noise at high frequencies, the algorithm assigns less importance
to accurately fitting the high-frequency portions of the data.
• Name — Edit the model name to P2DUe1 to generate a model with a
unique name in the System Identification Tool.
3-133
3 Click Estimate.
4 In the Process Models dialog box, set the Disturbance Model to Order 2
to estimate a second-order noise model.
5 Edit the Name field to P2DUe2 to generate a model with a unique name
in the System Identification Tool.
6 Click Estimate.
Validating the Models

In this portion of the tutorial, you evaluate model performance using the
Model Output and the Residual Analysis plots.
You must have already estimated the models, as described in “Estimating a

Second-Order Transfer Function Using Default Settings” on page 3-124 and
“Estimating a Second-Order Process Model with Complex Poles” on page
3-132.
Comparing the Model Output Plots. To generate the Model Output

plot, select the Model output check box in the System Identification Tool.
If the plot is empty or a model output does not appear on the plot, click the
model icons in the System Identification Tool window to display these models
on the plot.
The following Model Output plot shows the simulated model output, by
default. The simulated response of the models is approximately the same for
models with and without noise. Thus, including the noise model does not
affect the simulated output.
3-134
To view the predicted model output, select Options > 5 step ahead
predicted output in the Model Output plot window.
The following Model Output plot shows that the predicted model output of
P2DUe2 (with a second-order noise model) is better than the predicted output
of the other two models (without noise and with a first-order noise model,
respectively).
3-135
3-136
Comparing the Residual Analysis Plots. To generate the Residual

Analysis plot, select the Model resids check box in the System Identification
Tool. If the plot is empty, click the model icons in the System Identification
Tool window to display these models on the plot.
P2DUe2 falls well within the confidence bounds on the Residual Analysis plot.
To view residuals for P2DUe2 only, remove models P2DU and P2DUe1 from the
Residual Analysis plot by clicking the corresponding icons in the System
3-137
The Residual Analysis plot updates, as shown in the following figure.
The whiteness test for P2DUe2 shows that the residuals are uncorrelated, and
the independence test shows no correlation between the residuals and the
inputs. These tests indicate that P2DUe2 is a good model.
3-138
Viewing Model Parameters

• “Viewing Model Parameter Values” on page 3-139
• “Viewing Parameter Uncertainties” on page 3-140
Viewing Model Parameter Values

You can view the numerical parameter values and other information about
the model P2DUe2 by right-clicking the model icon in the System Identification
Tool . The Data/model Info dialog box opens.
3-139
The noneditable area of the dialog box lists the model coefficients that
correspond to the following model structure:
K
G(s) = e−Td s
( 1 + 2ξTw s + Tw2 s2 )
The coefficients agree with the model used to simulate the data:
1
G(s) = e−2 s
2
1 + 0 .2 s + s
Viewing Parameter Uncertainties

To view parameter uncertainties for the system transfer function, click
Present in the Data/model Info dialog box, and view the information in the
MATLAB Command Window.
Kp = 0.99821 +/- 0.019982

Tw = 0.99987 +/- 0.0037697
Zeta = 0.10828 +/- 0.0042304
Td = 2.004 +/- 0.0029717
The 1-standard-deviation uncertainty for each model parameter follows the

+/- symbol.
P2DUe2 also includes an additive noise term, where H is a second-order ARMA

model and e is white noise:
C
H= e
D
The software displays the noise model H as a ratio of two polynomials,

C(s)/D(s), where:
C(s) = s^2 + 2.186 (+/- 0.08467) s + 1.089 (+/- 0.07951)

D(s) = s^2 + 0.2561 (+/- 0.09044) s + 0.5969 (+/- 0.3046)
The 1-standard deviation uncertainty for the model parameters is in

3-140
Exporting the Model to the MATLAB Workspace

You can perform further analysis on your estimated models from the
MATLAB workspace. For example, if the model is a plant that requires a
controller, you can import the model from the MATLAB workspace into the
Control System Toolbox product. Furthermore, to simulate your model in the
Simulink software (perhaps as part of a larger dynamic system), you can
import this model as a Simulink block.
The models you create in the System Identification Tool are not automatically
available in the MATLAB workspace. To make a model available to other
toolboxes, Simulink, and the System Identification Toolbox commands, you
must export your model from the System Identification Tool to the MATLAB
workspace.
To export the P2DUe2 model, drag the model icon to the To Workspace
rectangle in the System Identification Tool,. The model now appears in the
MATLAB Workspace browser.
Note This model is an idproc model object. Model objects encapsulate

all properties of the model. To learn more about this model object, see the
3-141
Simulating a System Identification Toolbox Model in

Simulink Software
• “Prerequisites for This Tutorial” on page 3-142
• “Preparing Input Data” on page 3-142
• “Building the Simulink Model” on page 3-143
• “Configuring Blocks and Simulation Parameters” on page 3-144
• “Running the Simulation” on page 3-148
Prerequisites for This Tutorial

In this tutorial, you create a simple Simulink model that uses blocks from
the System Identification Toolbox library to bring the data z and the model
P2DUe2 into Simulink.
To perform the steps in this tutorial, Simulink must be installed on your

computer.
Furthermore, you must have already performed the following steps:
• Load the data set, as described in “Loading Data into the MATLAB
Workspace” on page 3-116.
• Estimate the second-order process model, as described in “Estimating a
Second-Order Process Model with Complex Poles” on page 3-132.
• Export the model to the MATLAB workspace, as described in “Exporting
the Model to the MATLAB Workspace” on page 3-141.
Preparing Input Data

Use the input channel of the data set z as input for simulating the model
output by typing the following in the MATLAB Command Window:
z_input = z; % Creates a new iddata object.

z_input.y = []; % Sets the output channel
% to empty.
Alternatively, you can specify any input signal.
3-142
Learn More. For more information about representing data signals for
system identification, see “Representing Data in MATLAB Workspace”.
Building the Simulink Model

To add blocks to a Simulink model:
1 In the MATLAB Command Window, type simulink.
2 Select File > New > Model to open a new model window.
3 In the Simulink Library Browser, select the System Identification

Toolbox library. The right side of the window displays blocks specific to
the System Identification Toolbox product.
Tip Alternatively, to access the System Identification block library, type

slident in the MATLAB Command Window.
4 Drag the following System Identification Toolbox blocks to the new model
window:
• IDDATA Sink block
• IDDATA Source block
• IDMODEL model block
5 In the Simulink Library Browser, select the Simulink > Sinks library,
and drag the Scope block to the new model window.
3-143
6 In the Simulink model window, connect the blocks to resembles the

following figure.
Next, you configure these blocks to get data from the MATLAB workspace
and set the simulation time interval and duration.
Configuring Blocks and Simulation Parameters

This procedure guides you through the following tasks to configure the model
blocks:
• Getting data from the MATLAB workspace.

• Setting the simulation time interval and duration.
1 In the Simulink Editor, select Simulation > Model Configuration

Parameters.
2 In the Configuration Parameters dialog box, type 200 in the Stop time
field. Click OK.
This value sets the duration of the simulation to 200 seconds.
3-144
3 Double-click the Iddata Source block to open the Source Block Parameters:
Iddata Source dialog box. Then, type the following variable name in the
IDDATA object field:
z_input
This variable is the data object in the MATLAB workspace that contains
the input data.
Tip As a shortcut, you can drag and drop the variable name from the
MATLAB Workspace browser to the IDDATA object field.
Click OK.
4 Double-click the Idmodel block to open the Function Block Parameters:

Idmodel dialog box.
a Type the following variable name in the Model variable field:
P2DUe2
This variable represents the name of the model in the MATLAB

workspace.
3-145
b Clear the Add noise check box to exclude noise from the simulation.
Click OK.
When Add noise is selected, Simulink derives the noise amplitude from
the NoiseVariance property of the model and adds noise to the model
accordingly. The simulation propagates this noise according to the noise
model H that was estimated with the system dynamics:
C
H= e
D
3-146
Click OK.
5 Double-click the Iddata Sink block to open the Sink Block Parameters:
Iddata Sink dialog box. Type the following variable name in the IDDATA
Name field:
z_sim_out
6 Type 1 in the Sample Time (sec.) field to set the sampling time of the
output data to match the sampling time of the input data.
Click OK.
3-147
The resulting change to the Simulink model is shown in the following figure.
Running the Simulation
1 In the Simulink Editor, select Simulation > Run.
2 Double-click the Scope block to display the time plot of the model output.
3-148
3 In the MATLAB Workspace browser, notice the variable z_sim_out that

stores the model output as an iddata object. You specified this variable
name when you configured the Iddata Sink block.
This variable stores the simulated output of the model, and it is now
available for further processing and exploration.
3-149
3-150
4
Nonlinear Model
Identification
4 Nonlinear Model Identification
Identify Nonlinear Black-Box Models Using System

Identification Tool
In this section...
“What Are Nonlinear Black-Box Models?” on page 4-3
“Preparing Data” on page 4-7
“Estimating Nonlinear ARX Models” on page 4-13
“Estimating Hammerstein-Wiener Models” on page 4-26
Introduction
Objectives
Estimate and validate nonlinear models from single-input/single-output
(SISO) data to find the one that best represents your system dynamics.
tasks using the System Identification Tool GUI:
• Import data objects from the MATLAB workspace into the GUI.
• Estimate and validate nonlinear models from the data.
• Plot and analyze the behavior of the nonlinearities.
Data Description
This tutorial uses the data file twotankdata.mat, which contains SISO
time-domain data for a two-tank system, shown in the following figure.
4-2
Identify Nonlinear Black-Box Models Using System Identification Tool
Tank 1
Tank 2
Two-Tank System
In the two-tank system, water pours through a pipe into Tank 1, drains into
Tank 2, and leaves the system through a small hole at the bottom of Tank
2. The measured input u(t) to the system is the voltage applied to the pump
that feeds the water into Tank 1 (in volts). The measured output y(t) is the
height of the water in the lower tank (in meters).
Based on Bernoulli’s law, which states that water flowing through a small
hole at the bottom of a tank depends nonlinearly on the level of the water in
the tank, you expect the relationship between the input and the output data
to be nonlinear.
twotankdata.mat includes 3000 samples with a sampling interval of 0.2 s.
What Are Nonlinear Black-Box Models?

• “Types of Nonlinear Black-Box Models” on page 4-4
• “What Is a Nonlinear ARX Model?” on page 4-4
• “What Is a Hammerstein-Wiener Model?” on page 4-5
4-3
Types of Nonlinear Black-Box Models

You can estimate nonlinear discrete-time black-box models for both
single-output and multiple-output time-domain data. You can choose from
two types of nonlinear, black-box model structures:
• Nonlinear ARX models

• Hammerstein-Wiener models
Note You can estimate Hammerstein-Wiener black-box models from

input/output data only. These models do not support time-series data, where
there is no input.
For more information on estimating nonlinear black-box models, see

“Nonlinear Model Identification”.
What Is a Nonlinear ARX Model?

This block diagram represents the structure of a nonlinear ARX model in a
simulation scenario:
Nonlinearity Estimator
u Nonlinear
Regressors Function y
u(t),u(t-1),y(t-1), ...
Linear
Function
The nonlinear ARX model computes the output y in two stages:
1 Computes regressors from the current and past input values and past
output data.
In the simplest case, regressors are delayed inputs and outputs, such as
u(t-1) and y(t-3)—called standard regressors. You can also specify custom
4-4
regressors, which are nonlinear functions of delayed inputs and outputs.

For example, tan(u(t-1)) or u(t-1)*y(t-3).
By default, all regressors are inputs to both the linear and the nonlinear
function blocks of the nonlinearity estimator. You can choose a subset of
regressors as inputs to the nonlinear function block.
2 The nonlinearity estimator block maps the regressors to the model output
using a combination of nonlinear and linear functions. You can select from
available nonlinearity estimators, such as tree-partition networks, wavelet
networks, and multi-layer neural networks. You can also exclude either
the linear or the nonlinear function block from the nonlinearity estimator.
The nonlinearity estimator block can include linear and nonlinear blocks in
parallel. For example:
F ( x) = LT ( x − r) + d + g ( Q( x − r) )
x is a vector of the regressors. LT ( x) + d is the output of the linear function

block and is affine when d≠0. d is a scalar offset. g ( Q( x − r) ) represents the
output of the nonlinear function block. r is the mean of the regressors x. Q is
a projection matrix that makes the calculations well conditioned. The exact
form of F(x) depends on your choice of the nonlinearity estimator.
Estimating a nonlinear ARX model computes the model parameter values,

such as L, r, d, Q, and other parameters specifying g. Resulting models are
idnlarx objects that store all model data, including model regressors and
parameters of the nonlinearity estimator. See the idnlarx reference page
for more information.
What Is a Hammerstein-Wiener Model?

This block diagram represents the structure of a Hammerstein-Wiener model:
u(t) Input w(t) Linear x(t) Output y(t)

Nonlinearity Block Nonlinearity
f B/F h
4-5
where:
• w(t) = f(u(t)) is a nonlinear function transforming input data u(t). w(t) has
the same dimension as u(t).
• x(t) = (B/F)w(t) is a linear transfer function. x(t) has the same dimension
as y(t).
where B and F are similar to polynomials in the linear Output-Error model,
as described in “What Are Polynomial Models?”.
For ny outputs and nu inputs, the linear block is a transfer function matrix
containing entries:
B j,i (q)
F j,i (q)
where j = 1,2,...,ny and i = 1,2,...,nu.

• y(t) = h(x(t)) is a nonlinear function that maps the output of the linear
block to the system output.
w(t) and x(t) are internal variables that define the input and output of the
linear block, respectively.
Because f acts on the input port of the linear block, this function is called
the input nonlinearity. Similarly, because h acts on the output port of the
linear block, this function is called the output nonlinearity. If system contains
several inputs and outputs, you must define the functions f and h for each
input and output signal.
You do not have to include both the input and the output nonlinearity in the
model structure. When a model contains only the input nonlinearity f, it is
called a Hammerstein model. Similarly, when the model contains only the
output nonlinearity h), it is called a Wiener model.
The nonlinearities f and h are scalar functions, one nonlinear function for
each input and output channel.
The Hammerstein-Wiener model computes the output y in three stages:
1 Computes w(t) = f(u(t)) from the input data.
4-6
w(t) is an input to the linear transfer function B/F.
The input nonlinearity is a static (memoryless) function, where the value of

the output a given time t depends only on the input value at time t.
You can configure the input nonlinearity as a sigmoid network, wavelet

network, saturation, dead zone, piecewise linear function, one-dimensional
polynomial, or a custom network. You can also remove the input
nonlinearity.
2 Computes the output of the linear block using w(t) and initial conditions:
x(t) = (B/F)w(t).
You can configure the linear block by specifying the numerator B and
denominator F orders.
3 Compute the model output by transforming the output of the linear block
x(t) using the nonlinear function h: y(t) = h(x(t)).
Similar to the input nonlinearity, the output nonlinearity is a static

function. Configure the output nonlinearity in the same way as the input
nonlinearity. You can also remove the output nonlinearity, such that y(t) =
x(t).
Resulting models are idnlhw objects that store all model data, including
model parameters and nonlinearity estimator. See the idnlhw reference page
for more information.
Preparing Data
• “Creating iddata Objects” on page 4-8
• “Starting the System Identification Tool” on page 4-10
• “Importing Data Objects into the System Identification Tool” on page 4-11

Load sample data in twotankdata.mat by typing the following command
in the MATLAB Command Window:
4-7
load twotankdata
This command loads the following two variables into the MATLAB Workspace
browser:
• u is the input data, which is the voltage applied to the pump that feeds the
water into Tank 1 (in volts).
• y is the output data, which is the water height in Tank 2 (in meters).
Creating iddata Objects

System Identification Toolbox data objects encapsulate both data values and
data properties into a single entity. You can use the System Identification
Toolbox commands to conveniently manipulate these data objects as single
entities.
You must have already loaded the sample data into the MATLAB workspace,
as described in “Loading Data into the MATLAB Workspace” on page 4-7.
Use the following commands to create two iddata data objects, ze and zv,
where ze contains data for model estimation and zv contains data for model
validation. Ts is the sampling interval.
Ts = 0.2; % Sampling interval is 0.2 sec

z = iddata(y,u,Ts);
% First 1000 samples used for estimation
ze = z(1:1000);
% Remaining samples used for validation
zv = z(1001:3000);
To view the properties of the iddata object, use the get command. For
example, type this command to get the properties of the estimation data:
get(ze)
4-8
MATLAB software returns the following data properties and values:
Domain: 'Time'
Name: ''
OutputData: [1000x1 double]
y: 'Same as OutputData'
OutputName: {'y1'}
OutputUnit: {''}
InputData: [1000x1 double]
u: 'Same as InputData'
InputName: {'u1'}
InputUnit: {''}
Period: Inf
InterSample: 'zoh'
Ts: 0.2000
Tstart: 0.2000
SamplingInstants: [1000x0 double]
TimeUnit: 'seconds'
ExperimentName: 'Exp1'
Notes: {}
UserData: []
4-9
To modify data properties, use dot notation or the set command. For example,
to assign channel names and units that label plot axes, type the following
syntax in the MATLAB Command Window:
% Set time units to minutes

ze.TimeUnit = 'sec';
% Set names of input channels
ze.InputName = 'Voltage';
% Set units for input variables
ze.InputUnit = 'V';
% Set name of output channel
ze.OutputName = 'Height';
% Set unit of output channel
ze.OutputUnit = 'm';
% Set validation data properties

zv.TimeUnit = 'sec';
zv.InputName = 'Voltage';
zv.InputUnit = 'V';
zv.OutputName = 'Height';
zv.OutputUnit = 'm';
To verify that the InputName property of ze is changed, type the following

command:
ze.inputname
Tip Property names, such as InputName, are not case sensitive. You can also
abbreviate property names that start with Input or Output by substituting u
for Input and y for Output in the property name. For example, OutputUnit is
equivalent to yunit.
Starting the System Identification Tool

To open the System Identification Tool GUI, type the following command
in the MATLAB Command Window:
ident
4-10
Importing Data Objects into the System Identification Tool

You can import the data objects into the GUI from the MATLAB workspace.
You must have already created the data objects, as described in “Creating
iddata Objects” on page 4-8, and opened the GUI, as described in “Starting
the System Identification Tool” on page 4-10.
To import data objects:
1 In the System Identification Tool, select Import data > Data object.
4-11
This action opens the Import Data dialog box.
2 Enter ze in the Object field to import the estimation data. Press Enter.
This action enters the object information into the Import Data fields.
Click More to view additional information about this data, including

channel names and units.
3 Click Import to add the icon named ze to the System Identification Tool
GUI.
4 In the Import Data dialog box, type zv in the Object field to import the
validation data. Press Enter.
5 Click Import to add the icon named zv to the System Identification Tool
GUI.
6 In the Import Data dialog box, click Close.
4-12
7 In the System Identification Tool GUI, drag the validation data zv icon to
the Validation Data rectangle. The estimation data ze icon is already
designated in the Working Data rectangle.
The System Identification Tool now resembles the following figure.
Estimating Nonlinear ARX Models

• “Estimating a Nonlinear ARX Model with Default Settings” on page 4-14
• “Plotting Nonlinearity Cross-Sections for Nonlinear ARX Models” on page
4-18
• “Changing the Nonlinear ARX Model Structure” on page 4-21
• “Selecting a Subset of Regressors in the Nonlinear Block” on page 4-24
• “Specifying a Previously-Estimated Model with Different Nonlinearity”
on page 4-25
4-13
• “Selecting the Best Model” on page 4-26
Estimating a Nonlinear ARX Model with Default Settings

In this portion of the tutorial, you estimate a nonlinear ARX model using
default model structure and estimation options.
You must have already prepared the data, as described in “Preparing Data”
on page 4-7. For more information about nonlinear ARX models, see “What Is
a Nonlinear ARX Model?” on page 4-4
4-14
1 In the System Identification Tool, select Estimate > Nonlinear models.
This action opens the Nonlinear Models dialog box.
The Configure tab is already open and the default Model type is
Nonlinear ARX.
4-15
In the Regressors tab, the Input Channels and Output Channels have
Delay set to 1 and No. of Terms set to 2. The model output y(t) is related
to the input u(t) via the following nonlinear autoregressive equation:
y(t) = f ( y(t − 1), y(t − 2), u(t − 1), u(t − 2) )
f is the nonlinearity estimator selected in the Nonlinearity drop-down

list of the Model Properties tab, and is Wavelet Network by default.
The number of units for the nonlinearity estimator is set to Select
automatically and controls the flexibility of the nonlinearity—more units
correspond to a more flexible nonlinearity.
2 Click Estimate.
This action adds the model nlarx1 to the System Identification Tool, as
shown in the following figure.
4-16
The Nonlinear Models dialog box displays a summary of the estimation

information in the Estimate tab. The Fit (%) is the mean square error
between the measured data and the simulated output of the model: 100%
corresponds to a perfect fit (no error) and 0% to a model that is not capable
of explaining any of the variation of the output and only the mean level.
Note Fit (%) is computed using the estimation data set, and not the
validation data set. However, the model output plot in the next step
compares the fit to the validation data set.
3 In the System Identification Tool, select the Model output check box.
This action simulates the model using the input validation data as input to
the model and plots the simulated output on top of the output validation
data.
4-17
The Best Fits area of the Model Output plot shows that the agreement
Plotting Nonlinearity Cross-Sections for Nonlinear ARX Models

Perform the following procedure to view the shape of the nonlinearity as a
function of regressors on a Nonlinear ARX Model plot.
4-18
1 In the System Identification Tool, select the Nonlinear ARX check box to
view the nonlinearity cross-sections.
By default, the plot shows the relationship between the output regressors
Height(t-1) and Height(t-2). This plot shows a regular plane in the
following figure. Thus, the relationship between the regressors and the
output is approximately a linear plane.
4-19
2 In the Nonlinear ARX Model Plot window, set Regressor 1 to

Voltage(t-1). Set Regressor 2 to Voltage(t-2). Click Apply.
The relationship between these regressors and the output is nonlinear, as

shown in the following plot.
3 To rotate the nonlinearity surface, select Style > Rotate 3D and drag
the plot to a new orientation.
4 To display a 1-D cross-section for Regressor 1, set Regressor 2 to none,

and click Apply. The following figure shows the resulting nonlinearity
4-20
magnitude for Regressor 1, which represents the time-shifted voltage

signal, Voltage(t-1).
Changing the Nonlinear ARX Model Structure

In this portion of the tutorial, you estimate a nonlinear ARX model with
specific input delay and nonlinearity settings. Typically, you select model
orders by trial and error until you get a model that produces an accurate fit
to the data.
You must have already estimated the nonlinear ARX model with default
settings, as described in “Estimating a Nonlinear ARX Model with Default
Settings” on page 4-14.
1 In the Nonlinear Models dialog box, click the Configure tab, and click
the Regressors tab.
2 For the Voltage input channel, double-click the corresponding Delay cell,
enter 3, and press Enter.
4-21
This action updates the Resulting Regressors list to show Voltage(t-3)

and Voltage(t-4) — terms with a minimum input delay of three samples.
3 Click Estimate.
This action adds the model nlarx2 to the System Identification Tool
GUI and updates the Model Output window to include this model. The
Nonlinear Models dialog box displays the new estimation information in
the Estimate tab.
4 In the Nonlinear Models dialog box, click the Configure tab, and select
the Model Properties tab.
5 In the Number of units in nonlinear block area, select Enter, and type
6. This number controls the flexibility of the nonlinearity.
4-22
6 Click Estimate.
This action adds the model nlarx3 to the System Identification Tool GUI.
It also updates the Model Output window, as shown in the following figure.
4-23
Selecting a Subset of Regressors in the Nonlinear Block

You can estimate a nonlinear ARX model that includes only a subset of
standard regressors that enter as inputs to the nonlinear block. By default,
all standard and custom regressors are used in the nonlinear block. In this
portion of the tutorial, you only include standard regressors.
You must have already specified the model structure, as described in

“Changing the Nonlinear ARX Model Structure” on page 4-21.
the Regressors tab.
2 Click the Edit Regressors button to open the Model Regressors dialog box.
3 Clear the following check boxes:
• Height(t-2)
• Voltage(t-3)
Click OK.
This action excludes the time-shifted Height(t-2) and Voltage(t-3) from

the list of inputs to the nonlinear block.
4 Click Estimate.
4-24
This action adds the model nlarx4 to the System Identification Tool. It also
updates the Model Output window.
Specifying a Previously-Estimated Model with Different

Nonlinearity
You can estimate a series of nonlinear ARX models by making systematic
variations to the model structure and base each new model on the
configuration of a previously estimated model. In this portion of the tutorial,
you estimate a nonlinear ARX model that is similar to an existing model
(nlarx3), but with a different nonlinearity.
1 In the Nonlinear Models dialog box, select the Configure tab. Click
Initialize. This action opens the Initial Model Specification dialog box.
2 In the Initial Model list, select nlarx3. Click OK.
3 Click the Model Properties tab.
4 In the Nonlinearity list, select Sigmoid Network.
5 In the Number of units in nonlinear block field, type 6.
4-25
6 Click Estimate.
This action adds the model nlarx5 to the System Identification Tool. It also
updates the Model Output plot, as shown in the following figure.
Selecting the Best Model

The best model is the simplest model that accurately describes the dynamics.
To view information about the best model, including the model order,
nonlinearity, and list of regressors, right-click the model icon in the System
Estimating Hammerstein-Wiener Models

• “Estimating Hammerstein-Wiener Models with Default Settings” on page
4-27
• “Plotting the Nonlinearities and Linear Transfer Function” on page 4-30
• “Changing the Hammerstein-Wiener Model Input Delay” on page 4-34
4-26
• “Changing the Nonlinearity Estimator in a Hammerstein-Wiener Model”

on page 4-35
• “Selecting the Best Model” on page 4-38
Estimating Hammerstein-Wiener Models with Default Settings

In this portion of the tutorial, you estimate nonlinear Hammerstein-Wiener
models using default model structure and estimation options.
You must have already prepared the data, as described in “Preparing Data”
on page 4-7. For more information about nonlinear ARX models, see “What Is
a Hammerstein-Wiener Model?” on page 4-5
1 In the System Identification Tool, select Estimate > Nonlinear models to

open the Nonlinear Models dialog box.
2 In the Configure tab, select Hammerstein-Wiener in the Model type list.
4-27
The I/O Nonlinearity tab is open. The default nonlinearity estimator

is Piecewise Linear with 10 units for Input Channels and Output
Channels, which corresponds to 10 breakpoints for the piecewise linear
function.
4-28
3 Select the Linear Block tab to view the model orders and input delay.
By default, the model orders and delay of the linear output-error (OE)
model are nb=2, nf=3, and nk=1.
4 Click Estimate.
This action adds the model nlhw1 to the System Identification Tool.
4-29
5 In the System Identification Tool, select the Model output check box.
This action simulates the model using the input validation data as input to
the model and plots the simulated output on top of the output validation
data.
The Best Fits area of the Model Output window shows the agreement
Plotting the Nonlinearities and Linear Transfer Function

You can plot the input/output nonlinearities and the linear transfer function
of the model on a Hammerstein-Wiener plot.
4-30
1 In the System Identification Tool, select the Hamm-Wiener check box to

view the Hammerstein-Wiener model plot.
The plot displays the input nonlinearity, as shown in the following figure.
4-31
2 Click the yNL rectangle in the top portion of the Hammerstein-Wiener

Model Plot window.
The plot updates to display the output nonlinearity.
4-32
3 Click the Linear Block rectangle in the top portion of the

Hammerstein-Wiener Model Plot window.
The plot updates to display the step response of the linear transfer function.
4-33
4 In the Choose plot type list, select Bode. This action displays a Bode
plot of the linear transfer function.
Changing the Hammerstein-Wiener Model Input Delay

In this portion of the tutorial, you estimate a Hammerstein-Wiener model
with a specific model order and nonlinearity settings. Typically, you select
model orders and delays by trial and error until you get a model that produces
a satisfactory fit to the data.
You must have already estimated the Hammerstein-Wiener model with

default settings, as described in “Estimating Hammerstein-Wiener Models
with Default Settings” on page 4-27.
4-34
the Linear Block tab.
2 For the Voltage input channel, double-click the corresponding Input

Delay (nk) cell, change the value to 3, and press Enter.
3 Click Estimate.
This action adds the model nlhw2 to the System Identification Tool GUI
and the Model Output window is updated to include this model, as shown
in the following figure.
The Best Fits area of the Model Output window shows the quality of the
nlhw2 fit.
Changing the Nonlinearity Estimator in a Hammerstein-Wiener

Model
In this portion of the example, you modify the default Hammerstein-Wiener
model structure by changing its nonlinearity estimator.
4-35
Tip If you know that your system includes saturation or dead-zone

nonlinearities, you can specify these specialized nonlinearity estimators
in your model. Piecewise Linear and Sigmoid Network are nonlinearity
estimators for general nonlinearity approximation.
1 In the Nonlinear Models dialog box, click the Configure tab.
2 In the I/O Nonlinearity tab, for the Voltage input, click the Nonlinearity
cell, and select Sigmoid Network from the list. Click the corresponding No.
of Units cell and set the value to 20.
3 Click Estimate.
This action adds the model nlhw3 to the System Identification Tool GUI. It
also updates the Model Output window, as shown in the following figure.
4-36
4 In the Nonlinear Models dialog box, click the Configure tab.
5 In the I/O Nonlinearity tab, set the Voltage input Nonlinearity to

Wavelet Network. This action sets the No. of Units to be determined
automatically, by default.
6 Set the Height output Nonlinearity to One-dimensional Polynomial.
4-37
7 Click Estimate.
This action adds the model nlhw4 to the System Identification Tool GUI. It
also updates the Model Output window, as shown in the following figure.
Selecting the Best Model

The best model is the simplest model that accurately describes the dynamics.
In this example, the best model fit was produced in “Changing the
Nonlinearity Estimator in a Hammerstein-Wiener Model” on page 4-35.
4-38

System Identification Toolbox™ Getting Started Guide

Uploaded by

Copyright:

Available Formats

System Identification Toolbox™ Getting Started Guide

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

System Identification Toolbox™ Getting Started Guide

Uploaded by

Copyright:

Available Formats

System Identification Toolbox™

Getting Started Guide

[email protected] Product enhancement suggestions

The MathWorks, Inc.

Related Products . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-26

Using This Product

System Identification Workflow . . . . . . . . . . . . . . . . . . . . . 2-3

Commands for Model Estimation . . . . . . . . . . . . . . . . . . . . 2-6

Identify Linear Models Using the Command Line . . . . . 3-55

Identify Low-Order Transfer Functions (Process

• “System Identification Toolbox Product Description” on page 1-2

System Identification Toolbox Product Description

The toolbox provides identification techniques such as maximum likelihood,

• Delay estimation, detrending, filtering, resampling, and reconstruction

Lennart Ljung. Professor Lennart Ljung is with the Department of

Qinghua Zhang. Dr. Qinghua Zhang is a researcher at Institut National

Peter Lindskog. Dr. Peter Lindskog is employed by NIRA Dynamics

Anatoli Juditsky. Professor Anatoli Juditsky is with the Laboratoire Jean

What Is System Identification?

The process of system identification requires that you:

About Dynamic Systems and Models

What Is a Dynamic Model?

A model is a mathematical relationship between a system’s input and output

You can represent dynamic models both in continuous-time and discrete-time

An often-used example of a dynamic model is the equation of motion of a

Mass-Spring-Damper System Excited by Force F(t)

Continuous-Time Dynamic Model Example

The matrices A, B, and C are related to the constants m, c and k as follows:

You can also obtain a transfer function model of the spring-mass-damper

where s is the Laplace variable.

Discrete-Time Dynamic Model Example

y(t) + a1 y(t − Ts ) + a2 y(t − 2Ts ) = bF (t − Ts )

y(t) + a1 y(t − 1) + a2 y(t − 2) = bF (t − 1)

y(t) = bF (t − 1) − a1 y(t − 1) − a2 y(t − 2)

This equation shows an iterative way of generating values of output y(t)

information on simulation and prediction using a model, see “Simulating and

You can also represent a discrete-time equation of motion in state-space and

System Identification Requires Measured Data

Why Does System Identification Require Data?

Time Domain Data

umeas  [ F (Ts ), F (2Ts ), F (3Ts ),..., F ( NTs )]

where Ts = 0.1 seconds and NTs is time of the last measurement.

Frequency Domain Data

You can use frequency-domain data to build both discrete-time and

Data Quality Requirements

• “Identifying Impulse-Response Models”

Building Models from Data

System Identification Requires a Model Structure

The following difference equation represents a simple model structure:

y(k) + ay(k − 1) = bu(k)

where a and b are adjustable parameters.

How the Toolbox Computes Model Parameters

where G is the transfer function.

To determine G, the toolbox minimizes the difference between the model

v(t) = ymeas(t) – ymodel(t) = ymeas(t) – Gu(t).

ymodel(t) is one of the following:

• Simulated response of the model for a given input u(t).

Configuring the Parameter Estimation Algorithm

• Configuring the minimization criterion to focus the estimation in a desired

Selecting Black-Box Model Structure and Order