Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
VOLUME XX, 2022 1
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number
Input Spectrum Design for Identification of a
Thermostat System
Md. Tanjil Sarker1, Ai Hui Tan2, and Timothy Tzen Vun Yap3
1,2Faculty of Engineering, Multimedia University, 63100 Cyberjaya, Malaysia
3Faculty of Computing and Informatics, Multimedia University, 63100 Cyberjaya, Malaysia
Corresponding author: Ai Hui Tan (e-mail: htai@mmu.edu.my).
The first author acknowledges financial sponsorship from the Multimedia University, Malaysia, under the Graduate Research Assistantship scheme and the
ICT Division, Bangladesh, under the Research Fellowship scheme.
ABSTRACT This paper considers the shaping of amplitude spectra of perturbation signals for the
identification of a thermostat system. The current approach in control engineering practice utilizes flat
spectrum signals, which may not result in the highest possible accuracy. This research aims to investigate
the effectiveness of optimal signals with amplitude spectra designed using two state-of-the-art software
approaches, namely the model-based optimal signal excitation 2 (MOOSE2) design and the optimal
excitation (optexcit) design, in improving estimation accuracy. Such a comparison on a real system is
currently lacking. In particular, there exists a research gap on how the combined choice of signal and model
structure affects performance measures. In this research, two model structures are used, which are the
autoregressive with exogenous input (ARX) and the output error (OE) model structures. Four performance
measures are compared, namely the determinant of the covariance matrix of the parameter estimates and the
minimum error, mean error and maximum error in the frequency response. Results show that the optimal
signals are effective in reducing the determinant of the covariance matrix and the maximum error in the
frequency response for the thermostat system, when applied in combination with the ARX model structure.
The flat spectrum signal remains very useful as a general broadband perturbation signal as it provides a
good overall fit of the frequency response. The findings from this work highlight the benefits of applying
optimal signals especially if the identification results are to be used for control, since these signals improve
key performance measures which have direct implications on controller design.
INDEX TERMS Estimation, perturbation signals, signal design, system identification, thermostat systems.
I. INTRODUCTION
System identification is widely applied in control
engineering to build models from input-output data [1].
These models are particularly useful in the design of model-
based controllers such as the model predictive controller
[2]. The accuracy of the model depends on the quality of
the input-output data which is, in turn, determined by the
input or perturbation signal used for the identification test.
Periodic perturbation signals have the advantages of
allowing the effects of transients to be removed and
enabling averaging to be performed [3]. They can be
categorized into fixed spectrum and computer-optimized
signals [4]. The former category is constructed based on
finite field arithmetic, whereas the latter category is
generated via optimization using computer programs [5].
One of the characteristics that can be optimized is the
amplitude spectra.
Flat spectrum signals are ideal when there is little prior
information available about the system under test. However,
they may not result in the highest possible estimation
accuracy. Fortunately, in some cases, such as in model
predictive control applications, an initial model of the system
is available at the point when an updated model is sought
after due to process aging and revamps. In many other
applications, it is possible to extract information about the
system under test from historical data or preliminary step
tests. It is useful to capitalize on the a priori information
when designing perturbation signals since it is known that for
the identification of linear systems, the asymptotic properties
of the model parameter estimates depend only on the input
spectrum [6]. Such signals are termed optimal signals. It is
important to investigate how their amplitude spectra affect
performance measures related to estimation accuracy since
this will determine the usability of the models obtained.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
2
Several signal designs are available in the literature, such
as [7]–[14]. In [7], signal design for non-iterative direct data-
driven techniques was considered for data-driven control,
where the amplitude spectrum was designed to minimize the
degradation caused by noise. An online method using past
input-output data was presented in [8], where it was shown to
be sample efficient. In [9], two signal design criteria
involving sensitivity functions were proposed and analyzed,
for the identification of systems with uncertainties. The
approach in [10] viewed the signal design from the
perspective of optimizing an input trajectory that maximizes
parameter identifiability. The signal-to-noise ratio (SNR) was
utilized to guide the amplitude spectrum shaping in [11].
Signal design for kernel-based identification was considered
in [12], where a two-step procedure using quadratic
transformation was applied. A Bayesian A-optimality
method was proposed in [13], which can be utilized for
online signal design, also in the context of kernel-based
identification. More recently, the direct spectrum shaping
technique [14] was proposed, with the advantage of having a
significantly lower computational load compared to that of
the Bayesian A-optimality approach.
Of the several designs available, those that come with
user-friendly software implementations are particularly
useful. Two of these software approaches represent the
current state-of-the-art, namely the model based optimal
input signal design 2 (MOOSE2) from the MOOSE2
program [15]–[17] and the optimal excitation (optexcit)
design from the Frequency Domain System Identification
(FDIDENT) Toolbox in MATLAB [18]–[21]. The
improvement achieved in the estimation accuracy across an
iteration for the optexcit signal was analyzed in [22] under
various SNRs. In [23], the combined use of MOOSE2 and
optexcit signals for the multivariable case was investigated
but the results are applicable only to the specific systems
tested.
Despite several signal designs being available, there is a
lack of comparison between them because different signals
were typically tested on different systems in the literature,
making comparison challenging. Additionally, the majority
of the existing works deal only with simulated systems. This
motivates the current work, where the input spectrum
design is implemented on a thermostat system. In
particular, the MOOSE2 and optexcit designs are compared
with the benchmark flat spectrum signal. The results reveal
how the combined choice of signal and model structure
affects performance measures related to estimation accuracy.
A thermostat system is used in this study for the following
reasons. Firstly, the regulatory function of a thermostat can
minimize energy consumption and provide cost savings of
greater than 40% [24]. Secondly, the thermostat is rather
ubiquitous, being used in air conditioners, water heaters,
refrigerators and ovens, thus ensuring that the results of this
research are very much applicable to control engineering
practice. Thirdly, many thermodynamic systems such as
furnaces [25] share similar characteristics in terms of the
smoothness of the dynamic responses, allowing the findings
from the current work to be generalized to a wide range of
systems.
The rest of the paper is organized as follows. Section II
provides the problem statement. Section III describes the
experimental set-up. Preliminary tests and the identification
of a benchmark model are described in Sections IV and V,
respectively. Comparison between various signal designs is
discussed in Section VI. Finally, concluding remarks and
suggestions for future work are presented in Section VII.
II. PROBLEM STATEMENT
The block diagram of a general thermostat system is
depicted in Figure 1. There are many types of thermostats
available [24], which are suited for different applications.
The controller may range from a simple on-off controller to
a more complicated proportional-integral-derivative class of
controller. The actuator depends on the heating and/or
cooling mechanisms used, and may range from a valve to a
pulse-width modulator.
FIGURE 1. Block diagram of a thermostat system.
In this paper, the identification of the system under test is
considered, where the input (denoted by u) is the desired
temperature scaled to its corresponding voltage value and the
output (denoted by y) is the temperature of the
thermodynamic process. The perturbation signal is fed into
the system as the input in order to excite the system. The
problem statement is formally stated as follows. It is required
to compare the estimation accuracy of the system transfer
function G(z) = Y(z)/U(z) for the flat, MOOSE2 and optexcit
amplitude spectra, where U and Y denote the z-transforms of
u and y, respectively, based on four performance measures:
(i) the determinant of the covariance matrix of the parameter
estimates defined by
}det{PD
, where P is the covariance
matrix of the estimated parameters,
(ii) the minimum error in the frequency response defined by
|))()(
ˆ
(|
min
min kGkGA k
, where G(k) is the actual system
Desired
temperature
input as
voltage, u
Thermostat
controller
Actuator
Process
Output
temperature, y
Thermal
sensor
+
_
Voltage feedback
Data acquisition system
System under test, G
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
3
frequency response and
)(
ˆkG
is the estimated system
frequency response at harmonic k,
(iii) the mean error in the frequency response defined by
F
k
kGkG
F
A
1
mean |)()(
ˆ
|
1
, where F is the highest
specified (excited) harmonic, and
(iv) the maximum error in the frequency response defined by
|))()(
ˆ
(|
max
max kGkGA k
.
The measure D provides an indication of the uncertainty of
the estimates. It is important in robust controller design
where decisions are guided by the size of the model
uncertainty. On the other hand, accuracy of the frequency
response is crucial in the design of controllers such as the
proportional-integral-derivative controller, where tuning is
frequently performed in the frequency domain. The measure
which is of prime importance depends on the application.
III. DESCRIPTION OF EXPERIMENTAL SET-UP
In this experiment, the ELWE LEHRSYSTEM thermostat
system [26] was utilized. The entire system (as shown in
Figure 2) has a length of 25.4cm, a width of 20.3cm and a
height of 30.5cm. It can accept manual input through an
adjustment knob, an internally programmed ramp input, and
a user-defined signal from a data acquisition system (DAQ).
The system has a proportional-integral controller.
FIGURE 2. Photograph of ELWE LEHRSYSTEM thermostat.
The ELWE LEHRSYSTEM thermostat can accept input
signals with two different voltage ranges, which are from
-10V to +10V and -15V to +15V. In this investigation, the
-10V to +10V range was utilized. The specifications of this
thermostat are given below:
• Operating temperature range: 0°C – 100°C
• Supply voltage: 230V
• Supply current: 0.5A
• Frequency: 50/60 Hz
Data input and output were performed through a National
Instruments (NI) myDAQ device, which is controlled
through the NI LabVIEW-based software appliance [27]. It
has the combined functions of digital multimeter,
oscilloscope, function generator, variable power supply,
Bode plot analyzer, dynamic signal analyzer, impedance
analyzer, two-wire current-voltage analyzer and three-wire
current-voltage analyzer.
In this investigation, the NI myDAQ was applied as an
arbitrary waveform (input signal) generator, which provided
input voltage to the ELWE LEHRSYSTEM thermostat.
Connections to the thermostat were made using probe wires.
The NI myDAQ device was connected via a universal serial
bus cable to a laptop with LabVIEW software installed. This
enabled experimental data to be displayed and stored in the
laptop for analysis. The photograph of the experimental set-
up is shown in Figure 3.
FIGURE 3. Photograph of the experimental set-up.
IV. PRELIMINARY TESTS
Preliminary tests were first conducted to check the
significance of nonlinear distortion as well as to obtain
suitable values of the sampling time ts and the measurement
time TN. A positive step test was applied, stepping the input
from 0V to +10V. The output was observed to change from
0C to around 100C. A negative step test was applied next,
stepping the input from +10V to 0V. The output decreased
back to 0C. In both cases, the time taken to reach 63% of the
total temperature change was found to be around 50s. The
output reached steady state after 250s. The positive and
negative step responses were largely symmetrical, indicating
that nonlinear distortion is negligible. Based on the
preliminary step tests, the sampling time ts was set to 5s
whereas the measurement time TN was set to 250s according
to recommendations in [5]. The signal period was calculated
using N = TN/ts = 50.
An initial model, Gi(z), was identified based on the step
response tests as
- . (1)
Only a low order model could be identified because the step
inputs lacked high frequency components and because the
system suffered from the presence of significant
disturbances. Gi(z) has a gain of 10. It will be subsequently
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
4
used in Section VI to design optimal signals, where the
parameters of Gi(z) will affect the shape of the amplitude
spectra of the optimal signals.
V. IDENTIFICATION OF BENCHMARK MODEL
A benchmark model was subsequently obtained by exciting
the system with a flat spectrum multisine signal of period N =
50 and 20 consecutive excited harmonics (from harmonics 1
to 20). The highest specified harmonic F was set to
approximately 0.4N, according to the recommendation in [5].
The sampling time ts was set to 5s based on the preliminary
tests. The signal amplitude ranged from 2.96V to 7.03V, with
a nominal value of 5V. The flat multisine signal u was fed
ten times into the system and the output data of periods 1 to
10 (y1, y2, …, y10) were measured in synchrony with the
input. The last five periods of the output data (y6 to y10) were
averaged to reduce the effects of noise. The SNR was
approximately 17dB, hence justifying the need for averaging
in the quest for a benchmark model. Figure 4 depicts the non-
averaged output and the averaged output signals as well as
the difference between them.
FIGURE 4. Five periods of the measured output.
FIGURE 5. DFT magnitudes of the measured output. Only the first 50
harmonics (out of 250 harmonics) are shown for better clarity.
The effects of averaging can be clearly observed in the
frequency domain, as depicted in Figure 5. Taking a 250-
point discrete Fourier transform (DFT), contributions of the
linear component of the system will appear at harmonics 5,
10, 15, 20, 25, … These are at multiples of five because five
steady state periods were taken. The power at the rest of the
harmonics such as harmonics 1, 2, 3, 4, 6, 7, 8, … can be
attributed to the effects of noise [3]. These are non-periodic
components and they can be removed via averaging. This
will ensure that the resulting benchmark model will have
high fidelity since it will be used as a basis for the
comparison of different input spectrum designs.
The average of the last five periods of the output data (y6
to y10) served as the training output for the identification of
the benchmark model. The second to fifth periods of the
output data (y2 to y5) were also averaged; this served as the
validation output. The training data and validation data were
fed into the System Identification Toolbox [28] in
MATLAB. The mean values of the signals were set to zero,
as is the typical practice in system identification. Transient
effects were removed by discarding the first period (y1).
The benchmark model order was selected using Akaike’s
Information Criterion [29]. The Rissanen’s Minimum
Description Length Criterion [30] also resulted in the same
order. The resulting benchmark model, Gb(z), is given by
. (2)
Gb(z) has a fit of
%
)
ˆ
(
1100 2
2
nn
nnn
y
yy
= 98.47%,
where n is the discrete time index and
y
ˆ
denotes the
estimated output. Gb(z) was identified using the
autoregressive with exogenous input (ARX) model structure;
a lower fit was obtained using the output error (OE) model
structure. The system gain of Gb(z) is 10.5.
FIGURE 6. Time domain signals with mean removed. Top: input;
bottom: output. For the bottom subplot - solid line: measured; dashed
line: benchmark model; dashed-dotted line: error.
050 100 150 200 250
-3
-2
-1
0
1
2
3
Time (s)
Voltage (V)
050 100 150 200 250
-20
-10
0
10
20
Time (s)
Temperature (deg C)
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
5
The measured output and the benchmark model output in
the time domain are compared in Figure 6. From Figure 6,
the benchmark model output matches the measured output
very well, with the error being almost zero. In Figure 7, the
DFT magnitudes are compared. The trend in the DFT
magnitude has been successfully captured by Gb(z). The high
accuracy was made possible by the averaging of multiple
periods of the input-output data.
FIGURE 7. DFT magnitudes. Top: input; bottom: output. For the bottom
subplot - circles: measured; asterisks: benchmark model.
VI. COMPARISON OF SIGNAL DESIGNS
Three different spectra were considered, namely the flat
spectrum, MOOSE2 spectrum and optexcit spectrum.
Perturbation signals with these spectra were implemented
using multisine signals of period N = 50 and 20 consecutive
excited harmonics (from harmonics 1 to 20). The sampling
time ts was set to 5s based on the preliminary tests. The flat
spectrum signal was the same as that used for identifying the
benchmark model and it is non-optimized. The MOOSE2
and optexcit signals are optimal signals and they were
designed based on Gi(z) given in (1).
To obtain the MOOSE2 signals, the MOOSE2 program
[15]–[17] was applied to generate the MOOSE2 spectrum.
The MOOSE2 program was set to minimize the determinant
of the covariance matrix of the estimated parameters based
on the D-optimality criterion. This made the objective
function of MOOSE2 as similar as possible to that of
optexcit. The MOOSE2 spectrum was parameterized as a
transfer function. The flat spectrum signal was passed
through this transfer function which works like a shaping
filter. The output of the filter gives the MOOSE2 signal.
The optexcit signal was generated by making use of the
optexcit algorithm in the FDIDENT Toolbox in MATLAB
[18]–[21]. The algorithm optimizes the power spectrum of
the input signal in the sense that it minimizes the volume of
the uncertainty ellipsoid of the estimated parameters. This
power spectrum was fed into a time-frequency swapping
algorithm [31] to generate the optexcit signal.
The three signals with different spectra were scaled to give
the same root-mean-square value of 2.68V, for fair
comparison. The signal and their DFT magnitudes are plotted
in Figures 8 and 9, respectively. The signals were fed into the
thermostat system in three separate experiments, after
shifting them by the nominal input voltage of 5V. In each
experiment, transient effects were removed by collecting two
periods of the output and discarding the first period. Noise at
the non-excited harmonics was filtered to improve the output
SNR. The input and output data, with mean values removed,
were fed into the System Identification Toolbox in
MATLAB. The training set was the same as the validation
set as only one period of steady state data was available. The
experiment was designed to test the accuracy of the
identification under stringent limits on the experiment time.
FIGURE 8. Perturbation signals in the time domain. Top: flat; middle:
MOOSE2; bottom: optexcit.
FIGURE 9. Perturbation signals in the frequency domain. Top: flat;
middle: MOOSE2; bottom: optexcit.
Two model structures were considered, namely the ARX
and OE model structures. Six different models were
estimated, which are GFA(z) and GFO(z) obtained using the
flat spectrum signal, GMA(z) and GMO(z) obtained using the
MOOSE2 signal and GOA(z) and GOO(z) obtained using the
0 5 10 15 20 25
0
5
10
15
Harmonic Number
DFT Magnitude
0 5 10 15 20 25
0
50
100
150
Harmonic Number
DFT Magnitude
050 100 150 200 250
-5
0
5
Time (s)
Voltage (V)
050 100 150 200 250
-5
0
5
Time (s)
Voltage (V)
050 100 150 200 250
-4
-2
0
2
4
Time (s)
Voltage (V)
0 5 10 15 20 25
0
10
20
Harmonic Number
DFT Magnitude
0 5 10 15 20 25
0
50
Harmonic Number
DFT Magnitude
0 5 10 15 20 25
0
50
Harmonic Number
DFT Magnitude
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
6
optexcit signal. GFA(z), GMA(z) and GOA(z) are ARX models,
whereas GFO(z), GMO(z) and GOO(z) are OE models. The
estimated models are
, (3)
, (4)
, (5)
, (6)
, (7)
. (8)
The results of the performance comparison are shown in
Table I. For the frequency response measures, the range of
the frequencies considered was from 0.025 rad/s to 0.503
rad/s, corresponding to the range covered by harmonics 1 to
20. The Bode plots are shown in Figures 10 and 11.
TABLE I
PERFORMANCE MEASURES FOR SIX DIFFERENT ESTIMATED MODELS. THE
SMALLEST VALUE IN EACH COLUMN IS SHOWN IN BOLD FONT.
Performance
measure
D
Amin
Amean
Amax
Flat ARX
1.47310-42
0.0010
0.0283
0.2894
Flat OE
9.92910-47
0.0422
0.0779
0.2459
MOOSE2 ARX
5.42510-94
0.0079
0.1155
0.2740
MOOSE2 OE
1.33610-61
0.0182
0.1161
0.2695
optexcit ARX
5.69210-94
0.0080
0.1145
0.2340
optexcit OE
5.32510-66
0.0121
0.1368
0.2732
In terms of D, the MOOSE2 signal with the ARX model
outperformed the other combinations, with optexcit ARX
following close behind. The flat spectrum signal resulted in
much larger values of D. This means that the uncertainties in
the model parameters are much larger than those obtained
using the MOOSE2 and optexcit signals. Nevertheless, all
the values of D are in fact very small, implying that all six
models are rather accurate. The ARX models generally
performed better than the OE ones on all performance
measures, because the ARX model can be solved using least
squares which is more robust compared to the recursive
solution for the OE model when the estimated model order is
quite high (in this case, 6).
FIGURE 10. Bode plots comparing the benchmark with the estimated
models. Top: GFA(z); middle: GMA(z); bottom: GOA(z).
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
7
FIGURE 11. Bode plots comparing the benchmark with the estimated
models. Top: GFO(z); middle: GMO(z); bottom: GOO(z).
The flat spectrum signal with the ARX model achieved the
lowest values of Amin and Amean. However, the optexcit ARX
combination led to the smallest Amax. This implies that the
distribution of the error is more uniform across the frequency
range of interest for the optexcit ARX combination, as is
evident from Figure 10. The different weightings in the cost
functions for the ARX and OE model structures led to a
higher estimation accuracy at the higher frequencies for the
ARX models, at the expense of a larger bias at the lower
frequencies.
Results from this study show that the combined choice of
signal and model structure affects the performance measures
in different ways. The best choice is application-dependent.
Optimal signals with specially designed spectra are effective
in reducing the covariance matrix of the parameter estimates
and making the error distribution more uniform. However,
the flat spectrum signal remains very effective as a general
broadband perturbation signal for system identification.
VII. CONCLUSION
Perturbation signals with three different amplitude spectra
corresponding to the flat, MOOSE2 and optexcit designs
were tested on a thermostat system. The optimal signals were
designed using a priori information from an initial model
identified through step tests. It was found that the combined
choice of signal and model structure can significantly affect
the performance measures and the optimal choice depends on
the measure which is of prime importance. The covariance
matrix of the parameter estimates and the maximum error in
the frequency response can be reduced using optimal signals.
The use of optimal signals will be beneficial when the
identification results are applied for controller design.
Nevertheless, the flat spectrum signal remains very useful as
a general broadband perturbation signal for system
identification as it provides a good overall fit of the
frequency response. For the thermostat system, the ARX
model structure outperformed the OE model structure
because the estimated model order was quite high. The
findings from the current work can be generalized to a wide
range of systems since the thermostat system has dynamics
which are similar to many practical systems. The insights
from this work are hence useful for control engineering
practice.
Suggestions for future work include amplitude spectrum
design of perturbation signals for multivariable and nonlinear
systems.
REFERENCES
[1] L. Ljung, “System Identification,” in Signal Analysis and Prediction,
A. Procházka, J. Uhlíř, P. W. J. Rayner, and N. G. Kingsbury, Eds.
Boston, MA, USA: Birkhäuser, 1998, pp. 163–173.
[2] L. Wang, Model Predictive Control System Design and
Implementation Using MATLAB. London, UK: Springer-Verlag,
2009.
[3] J. Schoukens, R. Pintelon, and Y. Rolain, Mastering System
Identification in 100 Exercises. Hoboken, NJ, USA: John Wiley &
Sons, Inc., 2012.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
8
[4] K. R. Godfrey, A. H. Tan, H. A. Barker, and B. Chong, “A survey of
readily accessible perturbation signals for system identification in the
frequency domain,” Control Eng. Pract., vol. 13, no. 11, pp. 1391–
1402, Nov. 2005.
[5] A. H. Tan and K. R. Godfrey, Industrial Process Identification:
Perturbation Signal Design and Applications. Cham, Switzerland:
Springer, 2019.
[6] L. Ljung, System Identification – Theory for the User. Upper Saddle
River, NJ, USA: Prentice Hall, 1999.
[7] S. Formentin, A. Karimi, and S. M. Savaresi, “Optimal input design
for direct data-driven tuning of model-reference controllers,”
Automatica, vol. 49, no. 6, pp. 1874–1882, Jun. 2013.
[8] H. J. van Waarde, “Beyond persistent excitation: Online experiment
design for data-driven modeling and control,” IEEE Control Syst.
Lett., vol. 6, pp. 319–324, 2022.
[9] C. Jauberthie, L. Denis-Vidal, Q. Li, and Z. Cherfi-Boulanger,
“Optimal input design for parameter estimation in a bounded-error
context for nonlinear dynamical systems,” Automatica, vol. 92, pp.
86–91, Jun. 2018.
[10] S. Park, D. Kato, Z. Gima, R. Klein, and S. Moura, “Optimal input
design for parameter identification in an electrochemical Li-ion
battery model,” in Proc. Annual American Control Conf.,
Milwaukee, WI, USA, 2018, pp. 2300–2305.
[11] H. Nian, M. Li, B. Hu, L. Chen, and Y. Xu, “Design method of
multisine signal for broadband impedance measurement,” IEEE J.
Emerging Selected Topics Power Elect., vol. 10, no. 3, pp. 2737–
2747, Jun. 2022.
[12] B. Mu and T. Chen, “On input design for regularized LTI system
identification: Power-constrained input,” Automatica, vol. 97, pp.
327–338, Nov. 2018.
[13] Y. Fujimoto, I. Maruta, and T. Sugie, “Input design for kernel-based
system identification from the viewpoint of frequency response,”
IEEE Trans. Autom. Control, vol. 63, no. 9, pp. 3075–3082, Sep.
2018.
[14] A. H. Tan, “Online input signal design for kernel-based impulse
response estimation,” IEEE Trans. Syst., Man, Cybern. Syst., vol. 52,
no. 11, pp. 7211–7222, Nov. 2022.
[15] M. Annergren and C. A. Larsson, “MOOSE2: Model Based Optimal
Input Design Toolbox for MATLAB (Version 2): User’s Guide,”
2016.
[16] M. Annergren and C. A. Larsson, “MOOSE2—A toolbox for least-
costly application-oriented input design,” SoftwareX, vol. 5, pp. 96–
100, Jun. 2016.
[17] M. Annergren, C. A. Larsson, H. Hjalmarsson, X. Bombois, and B.
Wahlberg, “Application-oriented input design in system
identification: Optimal input design for control,” IEEE Control Syst.
Mag., vol. 37, no. 2, pp. 31–56, Apr. 2017.
[18] J. Schoukens, P. Guillaume, and R. Pintelon, “Design of multisine
excitations,” in Proc. Int. Conf. Control, Edinburgh, UK, 1991, pp.
638–643.
[19] R. Pintelon and J. Schoukens, System Identification: A Frequency
Domain Approach. Hoboken, NJ, USA: John Wiley & Sons, Inc.,
2012.
[20] I. Kollár, Frequency Domain System Identification Toolbox for Use
with MATLAB. Natick, MA, USA: The MathWorks, Inc., 1994.
[21] I. Kollár, R. Pintelon, and J. Schoukens, “Frequency Domain System
Identification Toolbox for MATLAB: Improvements and new
possibilities,” IFAC Proceeding Volumes, vol. 30, pp. 943–946,
2017.
[22] M. T. Sarker, A. H. Tan, and T. T. V. Yap, “Performance evaluation
of iterative signal design for system identification,” in Proc. IEEE
Int. Conf. Automat. Control Intell. Syst., Shah Alam, Malaysia, 2022,
pp. 203–208.
[23] M. T. Sarker, A. H. Tan, and T. T. V. Yap, “Amplitude spectrum
design for multivariable system identification in open loop,” in Proc.
IEEE Int. Conf. Automat. Control Intell. Syst., Shah Alam, Malaysia,
2022, pp. 107–112.
[24] D. Bienvenido-Huertas, “Influence of the type of thermostat on the
energy saving obtained with adaptive setpoint temperatures: Analysis
in the current and future scenario,” Energ. Buildings, vol. 244, article
111024, Aug. 2021.
[25] K. C. Chook and A. H. Tan, “Identification of an electric resistance
furnace,” IEEE Trans. Instrum. Meas., vol. 56, no. 6, pp. 2262–2270,
Dec. 2007.
[26] ELWE LEHRSYSTEM Thermostat 10 13 002. ELWE, Cremlingen,
Germany, 2000.
[27] NI myDAQ User Guide and Specifications, National Instruments,
Austin, TX, USA, 2016.
[28] L. Ljung, System Identification Toolbox for Use with MATLAB.
Natick, MA, USA: The MathWorks, Inc., 1997.
[29] S. Portet, “A primer on model selection using the Akaike
Information Criterion,” Infect. Dis. Model., vol. 5, pp. 111–128,
2020.
[30] M. Kawakita and J. Takeuchi, “Minimum description length
principle in supervised learning with application to lasso,” IEEE
Trans. Inf. Theory, vol. 66, no. 7, pp. 4245–4269, Jul. 2020.
[31] E. Van Der Ouderaa, J. Schoukens, and J. Renneboog, “Peak factor
minimization using a time-frequency domain swapping algorithm,”
IEEE Trans. Instrum. Meas., vol. 37, no. 1, pp. 145–147, Mar. 1988.
Md. Tanjil Sarker received the B.Sc. degree in
Electrical and Electronics Engineering (EEE) and
Master of Business Administration (MBA)
degree in Human Resource Management (HRM)
from Bangladesh University, Dhaka, Bangladesh,
in 2013 and 2015, respectively. He obtained the
M.Sc. degree in Computer Science and
Engineering (CSE) from Jagannath University,
Dhaka, Bangladesh, in 2018. He is currently
pursuing his Ph.D. degree in Engineering in the
Faculty of Engineering, Multimedia University, Malaysia. Currently he is
an active graduate student member in the IEEE Student Branch of the
Malaysia Section. He has conducted many research works in relevant
fields. His research interests are in system identification, signal processing
and control, power system analysis and high voltage engineering.
Ai Hui Tan graduated with first class honours in
Electronic Engineering from the University of
Warwick, UK, in 1999. She was awarded a Ph.D.
in 2002 by the University of Warwick. Her
research interests are in the fields of system
identification and signal processing.
She joined the Faculty of Engineering,
Multimedia University, Cyberjaya, Malaysia, in
2002, where she is now an Associate Professor.
She was a consultant to Agilent Technologies
from 2012–2013. She authored the book
Industrial Process Identification: Perturbation Signal Design and
Applications, published by Springer in 2019.
Dr. Tan has served as a member of the International Federation of
Automatic Control (IFAC) Technical Committee on Modelling,
Identification and Signal Processing since December 2005. She is a
Chartered Engineer registered with the Engineering Council, UK.
Timothy Tzen Vun Yap received his B. Eng.
(Hons.) degree in Electronics majoring in
Computer from the Multimedia University in
2002, as well as his M.Eng.Sc. and Ph.D.
degrees from the System Identification and
Control Group, Multimedia University,
Malaysia, in 2006 and 2017, respectively. He is
currently a Senior Lecturer with the Faculty of
Computing and Informatics, Multimedia
University. His current research interests include
system identification, blockchain, data
engineering and machine learning. He currently chairs the Center of Big
Data and Blockchain Technologies at the Multimedia University.
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3234255
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
