System identification of a steam distillation pilot scale using arx and narx approaches

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Issue: 01 | Jan-2014, Available @ http://www.ijret.org 144
SYSTEM IDENTIFICATION OF A STEAM DISTILLATION PILOT-
SCALE USING ARX AND NARX APPROACHES
Haslizamri Md Shariff1
, Mohd Hezri Fazalul Rahiman2
, Ihsan Yassin3
, Mazidah Tajjudin4
1
Asst. Lecturer, 2, 3, 4
Lecturer, Faculty of Electrical Engineering, Universiti Teknologi MARA, Shah Alam Malaysia
Abstract
This paper presents steam temperature models for steam distillation pilot-scale (SDPS) by comparing Pseudo Random Binary
Sequence (PRBS) versus Multi-Sine (M-Sine) perturbation signal Both perturbation signals were applied to nonlinear steam
distillation system to study the capability of these input signals in exciting nonlinearity of system dynamics. In this work, both linear
and nonlinear ARX model structures have been investigated. Five statistical approaches have been observed to evaluate the developed
steam temperature models, namely, coefficient of determination, R2
; auto-correlation function, ACF; cross-correlation function, CCF;
root mean square error, RMSE; and residual histogram. The results showed that the nonlinear ARX models are superior as compared
to the linear models when M-Sine perturbation applied to the steam distillation system. While, PRBS perturbation exhibit insufficient
to model nonlinear system dynamic
Keywords: SDPS, PRBS, M-Sine, ARX, R2
, ACF, CCF, RMSE
---------------------------------------------------------------------***-------------------------------------------------------------------------
1. INTRODUCTION
Steam distillation is one of the earlier and common separation
techniques in chemical manufacturing [1], [2]. In 1991,
estimated that 40,000 distillation columns operate in the
United State alone to produce the essential oil, comprising
40% of all energy usage in the refining and commodity
chemical manufacturing sector [3]. In Malaysia, the essential
oil is produced using steam distillation techniques and the
demand of essential oil is increasing every year, nevertheless
the production of steam distillation process has not been
explored widely [2], [4]. Only few number of research efforts
have been reported in improving the oil extraction techniques
since the past three decades [2].
In real time process, most of the chemical engineering
processes are nonlinear in their dynamics [1], [5], [6], [7],
including the behavior of the distillation column [1], [5], [7],
[8]. Until today, almost all the works related to identification
of steam distillation column still using linear model to
represent the process dynamic [1], [2], [9], [10], [11].
Unfortunately, the linear identification is limited for a given
input range [7], [12], [13], [14]. In the last decades, there has
been a tendency towards nonlinear modeling in various
application areas; encourage with technological innovations
has resulted less limitations on the computational, memory
and data-acquisition level, making nonlinear modeling a more
feasible and flexible choices [7], [12], [13], [14]
To be more feasible and flexible, a systematic modeling is
required to describe a phenomenon of interest with improved
understanding for the purposes of simulation, prediction and
control. In order to develop systematic model, system
identification is under control engineering field is offer how to
build the dynamics of a system as a set of mathematical model
based on the observation of input and output data [15], [16]. The
mathematical model created is capable of relating the system
output for any given input in such a way that it can even predict
the future of the system. The primary goal of system
identification is to reduce errors between model and true system
[17]. The model is capable of facilitating the controller and
optimizing the system in which the traditional control technique
find a difficult to achieve [12], [14].
2. SYSTEM IDENTIFICATION
System identification research and applications has been growth
up and it was recognized as important tool in numerous fields.
Nowadays, system identification is getting more attention owing
to widespread development of sophisticated and efficient
algorithms, coupled with the advancement of digital processing
and computing.
The derivation of a relevant „system description‟ from the
observed data is termed as system identification, and the
resultant system description as a „model‟ [1]. Scientifically,
system identification deals with the problem of building
mathematical models of dynamical systems to describe the
underlying mechanism of the observed data of the systems [1].
In order to design and implement high performance of control
system the dynamic model is required. The obtained model must
be validated to verify whether the model is preciseness before
implemented. This is done by comparing the output of the

__________________________________________________________________________________________
obtained model and the output of the plant using part of the
experimental data that has been reserved for this purpose [18].
There are two main types of empirical models: linear models
and nonlinear models [12]. Linear models provide an
appropriate representation of the process in a small
neighborhood of an operating point. However, when the
process is operated outside this constrained region, the model
predictions will not accurate. Other offer is nonlinear model,
where tend to capture more accurately the process behavior,
making the adequate for controlling a real process in a wide
region of operation.
The distillation column are widely used in chemical processes
and exhibit nonlinear dynamic behavior [6], [19], [20]. In
recent year, the has been an increasing interest in modeling of
heating process especially in distillation column to extract the
essential oil where the linear model become most common in
industrial application [21]. Unfortunately, linear
approximations are only valid for a given input range [12].
The nature of the chemical industrial itself, economics of
operation and the unit operation themselves, imposes
additional requirements on process models [7]. The need for
improved product quality while maintaining a safe and
economical operation requires that plants be operated over a
broad range about the nominal operating point [7]. This leads
to plant operation close to constraints and excites
nonlinearities in system behavior. In distillation column
exhibit symmetric output changes to symmetric input changes
(reflux ratio) [22]. In addition, reacting system often display
nonlinearities arising due to the reaction mechanism or due to
the non-isothermal nature of the rate constant [7]. This
nonlinear behavior presents a difficulty for linear controller
due to their limitations [6]. Thus nonlinearity is integral part of
chemical process operation and must be accounted for while
developing nonlinear models [7].
From logical perspective it would seem that a nonlinear
system would require a nonlinear model to fully exhibit its
characteristics [23]. However, common practice most that a
linear model will be the first choice with which to identify a
model of a nonlinear system process, and that this course of
action often leads to satisfactory model fit for its purpose [23].
In the case of approximation linear model of nonlinear system,
it can be validate by compare linear and nonlinear model to
identify which model display the best result. To ensure the
system process able to fully exhibit its characteristics,
persistently exciting perturbation signal is required.
Otherwise, comparison between both model obtained will not
exhibit significant improvement on this effort. The main
objective of this research is to reveal a solution to this
problem.
3. PERTURBATION SIGNAL
System identification deals with the problem of how to estimate
the model of a system from measured input and output signals
[24]. The system can be linear or nonlinear depending on type of
the system, linear or nonlinear model can be estimated. The most
important thing is perturbation signal injected to a system must
have sufficient excitation (enough fluctuating) of desired effect
in order to measure and describe some property of the system
dynamic.
In nonlinearity identification of system dynamic, perturbation
signal must be persistently exciting (or enough fluctuating) in
order to excite the full range of nonlinearity process dynamics
[7]. For linear system, deterministic PRBS have been commonly
used. However PRBS input is insufficient to exhibit the
nonlinearity behavior of dynamic system. It is in agreement with
the claimed by [14], [25], [26] that the PRBS consists of only
two levels, the resulting data may not provide sufficient
information to identify nonlinear behavior [1], [14], [26], [27].
These signal cannot excite certain nonlinearities, so that more
input levels in the sequence are necessary [7], [17], [26], [27]. In
addition, the magnitude of PRBS is too large may bias in
estimation of linear kernel.
In this research, NARX model will be developed with
persistently exciting required perturbation signal; M-level PRS.
The objective of using M-level PRS input is to study this signal
capability to excite the nonlinearity of system dynamic. The
advantages of these signal operate at many operating levels and
provide the possibility of identifying nonlinearities behavior, or
of identifying linear behavior in the presence of nonlinearities
[25], [26].
4. MODEL STRUCTURE SELECTION
4.1 ARX Model
An ARX model is one of the linear model in system
identification. The ARX model comprising of past output and
exogenous input variable is represented as past input data. The
ARX model is among the simple models for linear process and it
is easy to be implemented.
The ARX model is written as [15]:
)(
)(
1
)(
)(
)(
)( te
qA
tu
qA
qB
qty nk
 
(1)
Where the polynomial A(q) and B(q) are defined by:
na
naqaqaqA 
 ...1)( 1
1 (2)
nb
nbqbqbbqB 
 ...)( 1
10 (3)

__________________________________________________________________________________________
Where A(q) and B(q)to be estimated which represent the
overall system dynamic andq-1as time shift operator and this q
description is completely equivalent to the Z-transform form
i.e. q corresponds to z [15] and the signal flow can be realized
as :
Fig 1: The ARX model structure
4.2 NARX Model
Most dynamical systems can be better represented by
nonlinear models [12], [28], where the models are able to
describe the global behavior of the system over the whole
operating range. It contrast with linear models are only able
approximate the system around a given operating point [28].
One of the most frequent studied classes of nonlinear models
are called block-oriented models, which consist of the
interconnection of linear blocks and nonlinear block [29].
There are many approach of nonlinear system identification
methods, where the methods are different owing to varied
positioning of interconnection, types of linear and nonlinear
functions [28]. One of the method is nonlinear auto regressive
with exogenous input (NARX) model, where the nonlinearity
estimator block is combine with linear ARX and nonlinear
function in parallel, maps the regressor output to the model
output.
The structure of NARX shown in figure 2:
Fig 2: Structure Of Nonlinear Auto-Regressive With
Exogenous Input (NARX)
The equation of NARX model can be written as:
(4)
Where y(t) is output, r are the regressors, u is input and L is an
autoregressive with exogenous (ARX) linear function. D is a
scalar off-set and g(Q(u-r)) represents output of nonlinear
function and Q is projection matrix that makes the calculations
well-conditioned.
The NARX model is suitable for modeling both the stochastic
and deterministic components of a system and is capable of
describing wide variety of nonlinear system [2].
5. MODEL VALIDATION
Model validation is final stage, where the stage is mandatory
step to decide the identified model is accepted or not[30]. The
purpose of model validation is to verify whether the identified
model fulfill the modeling requirement for a particular
application. In achieving good estimated model, it is necessary
to distinguish between the lack of fit between model and data
due to random processes and that due to lack of model
complexity.In most statistical tools, a measure of model fit is
determine by coefficient of determination, R2 [31]. The R2
given by;
(5)
The RMSE is used to assess the forecasting performance of a
predictor. As the prediction accuracy increases, the RMSE
decreases [32]. The RMSE is given by;
(6)
The RMSE also used to evaluate the performance of a predictor
over another by calculating the improvement achieved. It is
useful especially when calculating the significant improvement
achieved by a nonlinear model over a linear model [33].
The improvement calculation is given as follows;
(7)
The ACF is a mathematical function that is used frequently in
signal processing for analyzing series of values such as time-
domains signal [2].Relative to PSD, ACF is its time-domain
counterpart [34]. The ACF reveal the strength of relationship
between two observation as a function of the time separation
between them, or in other words, the cross-correlation of a signal
with itself. ACF is useful in investigating repeating patterns in a
signal such as determining the presence of a periodic signal
which has been buried by noise. This capability makes ACF very
important tools in determining the whiteness of stochastic signal.
The ACF and CCF respectively given by;
+
+u y
e
Regressor
u(t), u(t-1),…
y(t), y(t-1),…
Nonlinear
g(.)
linear
(LTI)

__________________________________________________________________________________________
(8)
(9)
where the terms of and are the average residuals and
inputs respectively. is the lag and it is common to investigate
the ACF and CCF between lag ±20 [35].
The power spectral density (PSD) of a signal is a description
of the distribution of the signal power of the signal versus
frequency [34]. The PSD is capable of capturing the frequency
content of a stochastic process. The unit of PSD is commonly
expressed in power/frequency (dB/Hz). The PSD of a signal
can be estimated by using periodogram.
Residual which is also known as prediction error describes the
error in the fit of the model to the observation . The
residuals can be used to provide the information about the
adequacy of the fitted model[31]. Residual of prediction is
given
- ;i = 1,2,3,….,N
Where is the observed output, is the predicted output, iis
the sequence and N is the number the data. From a time –
series plot of residuals, if the residual is randomly distributed
around zero, it indicated that the estimated model describes
the observed data well[36].
Histogram is a method to summarize a data distribution into
several intervals and the number of data points in each interval
is represented as bar length[37]. It is sometimes referred to as
frequency distribution[38]. Based on the time-series of
residual ., a histogram can be plotted. The residual are
expected to be normally distributed because the normal
distribution often provides an adequate approximation to the
distribution of many measured quantities.
6. EXPERIMENTAL DESIGN
6.1 Experimental Set-Up
The steam distillation pilot plant using 1500W coil-type heater
to generate steam. The heater is immersed in 10liters of water
for 2000 seconds. Two (2) resistive temperature detectors
(RTD) PT-100 were installed. The primary RTD used to
monitor water temperature in the column, and secondary RTD
to monitor steam temperature that was installed 30cm from
steam outlet. The output from both RTDs is resistance are
converting to voltage by using signal converter that produced
output within 1V to 5V for temperature range varies from 0°C
to 100°C. Power controller used to control heater are
manipulated by providing control signal.
Fig 3: Prototype of Steam Distillation Pilot Plant
6.2 Methodology
Fig 4: Experimental Steps

__________________________________________________________________________________________
6.3 Data Analysis
6.3.1 PRBS Perturbation
Three sets of data collected perturb by PRBS perturbation.
The PRBS data with difference probability band, B; 0.08, 0.05
and 0.02 are shown in the Figures 5, 7 and 9 respectively. The
Figures 6, 8, 10 shows the output and input PSD for PRBS
data for B; 0.08, 0.05 and 0.02 respectively. From figure 6(a),
8(a) and 10(a), the PSD shows that the output signal is
concentrated at very low frequencies. It is obvious that as the
frequency increased, the signal power is drops. Meanwhile,
the input signals in figures 6(b), 8(b) and 10(b) are richer in
frequency contents. The bandwidth of the system is
approximately less than 0.05 Hz. This signifies that the system
is of very low frequency dynamics.
Measured Temperature
Sample number
PRBS (B=0.08)
Sample number
Fig 5: PRBS data set 1
Frequency (Hz)
(a)
Frequency (Hz)
(b)
Fig 6: The PSD of simulated for data set 1
(a) output (b) input
Sample number
PRBS (B=0.05)
Sample number
Frequency (Hz)
(a)
Power/Frequency(dB/Hz)y(t)u(t)
u(t)y(t)Power/Frequency(dB/Hz)
Power/Frequency(dB/Hz)

__________________________________________________________________________________________
Frequency (Hz)
(b)
(a) Output (b) input
Sample number
PRBS (B=0.02)
Sample number
Frequency (Hz)
(a)
Frequency (Hz)
(b)
6.3.2 M-Sine Perturbation
Three sets of data collected perturb by M-Sine perturbation.
Figures 11, 13, 15 show the measured output and input data
driven by M-Sine input signals. All the data set will be
estimated and validated; then proceed to compare with models
developed by PRBS perturbation.
Referring to figures 12(a), 14(a) and 16(a) PSD shows that the
output signal is concentrated at very low frequencies only.
Meanwhile, the input signal in figures 12(b), 14(b) and 16(b),
is richer in frequency content. Means each input PSD shows
that the input consists of signals at various frequencies. The
power signal is spread almost consistently throughout the
frequencies, which is almost same to white noise properties.
The bandwidth of the system is approximately less than 0.1Hz.
This signifies that the system is of very low frequency
dynamics. It is obvious that as the frequency is increased, the
signal power drops.
Sample number
M-Sine signal
Sample number
Fig 11: M-Sine Data set 1
y(t)u(t)
y(t)u(t)

__________________________________________________________________________________________
Frequency (Hz)
(a)
Frequency (Hz)
(b)
Sample number
M-Sine signal
Sample number
Fig 13: M-Sine PRS Data set 2
Frequency (Hz)
(a)
Frequency (Hz)
(b)
Measured temperature
Sample number
M-Sine signal
Sample number
Fig 15: M-Sine Data set 3
Frequency (Hz)
(a)
Frequency (Hz)
(b)
Power/Frequency
(dB/Hz)
Power/Frequency
(dB/Hz)
Power/Frequency
(dB/Hz)u(t)
y(t)u(t)Power/Frequency
(dB/Hz)
Power/Frequency
(dB/Hz)
Power/Frequency
(dB/Hz)y(t)

__________________________________________________________________________________________
7. RESULTS
7.1 Results for PRBS Data
Three sets of data have been collected perturbed by PRBS
perturbation; used for training and testing using the ARX and
NARX approaches. The summary of results as per in table 1
and 2, which is represent the ARX and NARX performance
respectively. The data set 1, 2 and 3 represents the ARX1-
PRBS, ARX2-PRBS and ARX3-PRBS respectively. For the
nonlinear models, suffix N is added and known as NARX1-
PRBS, NARX2-PRBS and NARX3-PRBS. The table 3 shows
the R2
and RMSE comparison in improvement achieved by
NARX models over the ARX models.
From the table 1 and 2, the ARX and NARX models exhibit
almost identical model fit and RMSE value.
Referring to table 3 the comparison of R2
and RMSE show no
significant improvement detected by the NARX over the ARX
models. The NARX models is failed to outperform the ARX
models; reveal the nonlinear model poor as a predictor of
nonlinear steam distillation system.
Figures 17 show the measured output, predicted output and
residual for the ARX1-PRBS. From the residual plot, the 1-SAP
prediction is in good agreement with the measured output.
Table 1: Summary of ARX results using PRBS perturbation
Model R2
(%) ACF CCF RMSE
ARX1-PRBS 97.95 - - 0.001428
ARX2-PRBS 97.00 lag 1,2,3 (3) - 0.001714
ARX3-PRBS 97.01 lag 1,2,8 (3) - 0.001729
Table 2: Summary of NARX result using PRBS perturbation
Model R2
(%) ACF CCF RMSE
NARX1-PRBS 97.96 - - 0.001422
NARX2-PRBS 97.77 lag 1,2,3 (3) lag 1,2,3 (3) 0.001710
NARX3-PRBS 97.01 lag 1,2,8 (3) - 0.001720
Table 3: R2
and RMSE comparison between ARX and NARX model for validation data.
______________________________________________________
Models % R2
RMSE
______________________________________________________
ARX1-PRBS 97.95 0.001428
NARX1-PRBS 97.96 0.001422
---------------------------------------------------------------------------------
% Improvement 0.01% 0.42%
---------------------------------------------------------------------------------
ARX2-PRBS 97.00 0.001714
NARX2-PRBS 97.77 0.001710
---------------------------------------------------------------------------------
% Improvement 0.79% 0.23%
---------------------------------------------------------------------------------
ARX3-PRBS 97.01 0.001729
NARX-PRBS 97.01 0.001720
---------------------------------------------------------------------------------
% Improvement 0% 0.52%
______________________________________________________
The figure 18 show the residual histogram for ARX1-PRBS,
which is indicate the data normally distributed and produced
zero means close to the zero.

__________________________________________________________________________________________
Figures 19(a) show the ACF of the residual. In both figures,
no significant correlation is detected outside the 95%
confidence limit, which indicates the residual has passed the
whiteness test. The cross-correlation test for ARX1-PRBS is
in figures 19(b). From the figure, no correlation detected
outside the 95% confidence limit of CCF, which signifies
that the information were fully utilized by the model
structure.
Measured, predicted & residual. % R2
=97.95%
Sample number
(a)
Sample number
(b)
Sample number
(c)
Fig 17: ARX1-PRBS
(a) Measured (b) predicted (c) residual.
ε(t)
Fig 18: The residual histogram of ARX1-PRBS for validation
data.
Lag
(a)
Lag
(b)
Fig 19: Correlation tests of ARX1-PRBS
(a) ACF (b) CCF
Measured, predicted & residual % R2
=97.96%
Sample number
(a)
Sample number
(b)
Sample number
(c)
Fig 20: Validation of NARX1-PRBS (a) simulated (b)
predicted (c) residual
Figures 20 show the measured output, predicted output and
residual for the NARX1-PRBS. From the residual plot, the 1-
SAP prediction is in good agreement with the measured output.
In figure 21 shows the residual histogram for NARX1-PRBS,
which is indicate the data normally distributed and produced
zero means close to the zero.
y-hat(˚C)
Rµε(t)
ε(t)
Rεε(t)
y(t)(˚C)y-hat(˚C)
Frequency
y(t)(˚C)
ε(t)

__________________________________________________________________________________________
Figures 22(a) show the ACF of the residual. In both figures, no
significant correlation is detected outside the 95% confidence
limit, which indicates the residual has passed the whiteness test.
The cross-correlation test for ARX1-PRBS is in figures 22(b).
From the figure, no correlation detected outside the 95%
confidence limit of CCF, which signifies that the information
were fully utilized by the model structure.
ε(t)
Fig 21: The residual histogram for NARX1-PRBS.
7.2 Result for M-Sine
(a)
(b)
Fig 22: Correlation tests of NARX1-PRBS (a) ACF (b) CCF
Table 4: Summary of ARX result using M-Sine perturbation
Model R2
(%) ACF CCF RMSE
___________________________________________________________________________________________
ARX1-MSine 65.07 - - 0.01133
ARX2-MSine 63.94 - - 0.01258
ARX3-MSine 68.65 lag 13(1) - 0.01262
Table 5: Summary of NARX result using M-Sine perturbation
Model R2
(%) ACF CCF RMSE
___________________________________________________________________________________________
NARX1-MSine 98.84 - lag -20~-10,-4~-1(25) 0.00525
NARX2-MSine 99.00 - - 0.00627
NARX3-MSine 99.17 - - 0.00514
___________________________________________________________________________________________
Legend: Shaded values indicate the best model.
Table 4 and 5 shows the result summary of ARX and NARX
models driven by M-Sine input respectively. The R2
for linear
ARX models is less than 70%.
However, NARX models exhibits very good R2
compared to
ARX models. The comparison of R2
and RMSE show the
significant improvement achieved by the NARX over the
ARX models. The NARX3-MSine is the best model where
highest model fit, passed correlation tests and lowest RMSE.
The second best is NARX2-MSine model, where the model
also sufficient to approximate the nonlinear system well, while
satisfying all validation criteria‟s.
Rεε(t)Rµε(t)
Frequency

__________________________________________________________________________________________
Table 6: R2
and RMSE result comparison between ARX and
NARX model for validation data.
____________________________________________
Models % R2
RMSE
____________________________________________
ARX1-MSine 65.07 0.01133
NARX1-MSine 98.84 0.00525
------------------------------------------------------------------
Improvement 51.90% 53.66%
------------------------------------------------------------------
ARX2-MSine 63.94 0.01258
NARX2-MSine 99.00 0.00627
------------------------------------------------------------------
------------------------------------------------------------------
ARX3-MSine 68.65 0.01262
NARX3-MSine 99.17 0.00514
------------------------------------------------------------------
------------------------------------------------------------------
Legend: Shaded value indicates best result.
The improvement achieved by the NARX models over the
ARX models driven by M-Sine input is clearly seen in table
6, which is the R2
and RMSE is improved more than 40% and
50% respectively. Claimed by [33], [39], the good model or
predictor is capable to reveal significant improvement of
RMSE, as the prediction accuracy increases, the RMSE
decreases. Means, it is apparent that the NARX models
perturbed by M-Sine capable to predict the nonlinear steam
distillation system behavior.
Figure 23 shows the measured output, predicted and residual
for NARX2-MSine models; second best model. The predicted
output shows good agreement with the measured output,
which can be confirmed by the residual plot. In figure 24,
residual histogram for NARX2-MSine is normally distributed
and indicates the model produced residuals mean close to the
zero. Figures 25(a) is ACF result for NARX2-MSine, show
the residual pass the whiteness test since all points lie within
the 95% confidence limit. The CCF results for NARX2-MSine
show in figures 25 (b), no significant correlation between
input and residual, thus information in the training data was
completely modeled.
Figure 26 shows the measured output, predicted output and
residual for the best model; NARX3-MSine. The predicted
output shows good agreement with the simulated output,
which can be confirmed by the residual plot. In figure 27, the
residual histogram of NARX3-MSine model is distributed as
per normal.
The ACF for NARX3-MSine show in figure 28(a), the
residual pass the whiteness test since all points lie within the
95% confidence limit.
The CCF results for NARX3-MSine show in figure 28 (b), no
significant correlation between input and residual, thus
information in the training data was completely modeled.
Measured, predicted & residual, %R2
= 99.00%
Sample number
(a)
Sample number
(b)
Sample number
(c)
Fig 23: Validation of NARX2-MSine (a) simulated (b)
ε(t)
Fig 24: Residual Histogram of NARX2-MSine
Lag
(a)
y-hat(˚C)ε(t)
Rµε(t)
ȳ(t)(˚C)
Rεε(t)
Frequency

__________________________________________________________________________________________
Lag
(b)
Fig 25: Correlation tests of NARX2-MSine
(a) ACF (b) CCF
Measured, predicted & residual, %R2
= 99.17%
Sample number
(a)
Sample number
(b)
Sample number
(c)
Fig 26: Validation of NARX3-MSine (a) simulated (b)
ε(t)
Fig 27: The residual histogram for NARX3-MSine.
Lag
(a)
Lag
(b)
Fig 28: Correlation tests of NARX3-MSine
(a) ACF (b) CCF
CONCLUSIONS
Linear and nonlinear ARX models for the nonlinear steam
distillation system have been successfully developed. Results
showed that NARX model were developed based on M-Sine
signal is the most accurate model in approximating nonlinear
system dynamic as compared to the model developed based on
PRBS perturbation. It can be evident that although PRBS
perturbation produced good model validation, it seemed that
the nonlinear model is almost as good as the linear model,
which is inaccurate. In addition, the nonlinear model driven by
PRBS input failed to reveal an actual nonlinear steam
distillation system behavior. On the other hand, the models
driven by M-Sine input revealed the unexplained dynamics
which can be translated in to significant improvements, shown
by the nonlinear models as compared to linear models.
REFERENCES
[1] D. E. Rivera, H. Lee, H. D. Mittelmann, and M. W.
Braun, “High-Purity Distillation,” IEEE Control
Systems Magazine, no. October, pp. 72–89, 2007.
[2] Mohd Hezri Fazalul Rahiman, “System Identification
Of Steam Distillation Essential Oil Extraction
System.” Phd Dissertation, Universiti Teknologi
MARA Malaysia, 2009.
[3] and R. A. K. Humphrey, J.L., A.F. Segibert,
“Separation Technologies: Advances and Priorities,”
DOE/ID/12920-1. Washington, D.C.: U.S.
Department of Energy., 1991.
[4] N. Ahmad, J. Kasim, S. Z. Mahmud, S. Abdul, K.
Yamani, A. Mokhtar, N. Yuziah, and M. Yunus,
“Manufacture and Properties of Oil Palm
Particleboard,” no. June, pp. 1–3, 2011.
y-hat(˚C)
Rµε(t)
Rεε(t)Frequency
ȳ(t)(˚C)
ε(t)

__________________________________________________________________________________________
[5] M. L. L. W.L.Luyben, “Essential of Process Control,”
Singapore;McGraw-Hills, 1997.
[6] A. A. Bachnas, “Linear Parameter-Varying Modeling
of a High-Purity Distillation Collumn,” Master
Science dissertation, no. System And Control At Delft
University Of Technology, 2012.
[7] A. S. . Soni, “CONTROL-RELEVANT SYSTEM
IDENTIFICATION USING NONLINEAR
VOLTERRA AND VOLTERRA-LAGUERRE
MODELS,” PhD Thesis, University of Pittsburgh ,
Submitted to the Graduate Faculty of the Science,
2006.
[8] Alina-Simona Baiesu, “Modeling a Nonlinear Binary
Distillation Column,” CEAI, vol. 13, no. 1, pp. 49–53,
2011.
[9] N. Ismail, N. Tajjudin, M. Hezri, F. Rahiman, M. N.
Taib, and U. T. Mara, “Modeling of Dynamic
Response of Essential Oil Extraction Process,” pp.
298–301, 2009.
[10] M. H. A. Jalil, Z. Yusuf, N. N. Nordin, N. Kasuan, M.
N. Taib, M. H. Fazalul, and U. T. Mara, “A
Simulation Study of Model Reference Adaptive
Control on Temperature Control of Glycerin
Bleaching Process,” pp. 8–10, 2013.
[11] Z. Muhammad, Z. M. Yusoff, M. H. F. Rahiman, and
M. N. Taib, “Steam temperature control for steam
distillation pot using model predictive control,” IEEE
8th International Colloquium on Signal Processing
and its Applications, pp. 474–479, Mar. 2012.
[12] J. Paduart, L. Lauwers, J. Swevers, K. Smolders, J.
Schoukens, and R. Pintelon, “Identification of
nonlinear systems using Polynomial Nonlinear State
Space models,” Automatica, vol. 46, no. 4, pp. 647–
656, Apr. 2010.
[13] Ir. Johan Paduart, “Identification of nonlinear systems
using Polynomial Nonlinear State Space models,”
Doctor of Engineering, Vrije Universiteit Brussel.,
2010.
[14] Oliver Nelles, “Nonlinear System Identification, From
Classical Approaches To Neural Networks and Fuzzy
Models.” Springer Verlag, Heidelberg, Germany,
2000.
[15] L. Ljung, System Identification Theory for User. 2nd
Edition,Prentice Hall, 1999.
[16] L. Ljung, “Perspectives on system identification,”
Annual Reviews in Control, vol. 34, no. 1, pp. 1–12,
Apr. 2010.
[17] L. Hyunjin, “A Plant-Friendly Multivariable System
Identification Framework Based on Identification Test
Monitoring,” PhD Dissertation, Arizona State
University USA, no. December, 2006.
[18] M. H. F. Rahiman, M. N. Taib, R. Adnan, and Y. M.
Salleh, “Analysis of weight decay regularisation in
NNARX nonlinear identification,” 2009 5th
International Colloquium on Signal Processing & Its
Applications, pp. 355–361, Mar. 2009.
[19] I. Yunan, I. M. Yassin, S. Farid, S. Adnan, M. Hezri,
and F. Rahiman, “Identification of Essential Oil
Extraction System using Radial Basis Function ( RBF
) Neural Network,” IEEE 8th International
Colloquium on Signal Processing And its Application,
pp. 495–499, 2012.
[20] F. Q. Elizabeth, K. Beauchard, F. Ammouri, and P.
Rouchon, “Stability and asymptotic observers of
binary distillation processes described by nonlinear
convection / diffusion models,” 2012 American
Control Conference, pp. 3352–3358, 2012.
[21] A. N.-R. C.Bordons, “Model Based predictive control
of an olive oil mill,” Journal of food Engineering, vol.
84, pp. 1–11, 2008.
[22] F. J. D. I. R.K.Pearson, B.A.Ogunnaike,
“Identification of Discrete Convolution Models For
Nonlinear Process,” IEEE Trans.,Acous.,Speech and
Signal Processing, 1995.
[23] K. R. Thompson, “Implementation of Gaussian
Process models for Nonlinear System Identification,”
PhD Dissertation, Glasgow University, 2009.
[24] E. Martin, Linear Models of Nonlinear Systems,
Linkoping., no. 985. Department of Electrical
Engineering, Linkoping University, Sweden, 2005.
[25] K. Godfrey, “Perturbation Signals For System
Identification.” Prentice Hall Internation (UK)
Limited Hertfordshire UK, 2000.
[26] M. W. Braun, D. E. Rivera, A. Stenman, W. Foslien,
and C. Hrenya, “Identification of a RTP Wafer
Reactor Multi-level Pseudo-Random Signal Design
and „ Model-on-Demand ‟ Estimation Applied to
Nonlinear Identification of a RTP Wafer Reactor,”
LiTH-ISH-R-2156, ACC’99, San Diego, 1999.
[27] M. Deflorian and S. Zaglauer, “Design of Experiments
for nonlinear dynamic system identification,” The 18th
IFAC Congress Milano (Italy), pp. 13179–13184,
2011.
[28] N. Patcharaprakiti, K. Kirtikara, D. Chenvidhya, V.
Monyakul, and B. Muenpinij, “Modeling of Single
Phase Inverter of Photovoltaic System Using System
Identification,” 2010 Second International Conference
on Computer and Network Technology, pp. 462–466,
2010.
[29] T. Hatanaka, K. Uosaki, and M. Kogat, “Block
Oriented Nonlinear Model Identification by
Evolutionary Computation Approach.”
[30] G. Z. Ioan D.Landau, Digital Control Systems.
Springer-Verlag London Limited, 2005.
[31] N. F. H. D.C.Montgomery, G.C. Runger,
“Engineering Statistics,” 3rd.Edition, John Wiley &
Sons, 2004.
[32] H.Demuth, “Neural Network Tollbox User Guides,”
4th.Edition Natick,USA, 2005.
[33] J.Sjoberg, “Non-linear Systen Identification with
Neural Networks,” Ph.D Dissertation, Linkoping
University, Linkoping, Sweden, 1995.

__________________________________________________________________________________________
[34] M.J.Roberts, “Signals ans systems: Analysis using
transform methods and MATLAB,” 1st. edition New
York; McGraw-Hill, 2013.
[35] L. K. H. Magnus Noorgard, O.Ravn, N.K.Poulsen,
“Neural Network for modeling and control of
dyanamics system,” London: Springer-Verlag London
Ltd, 2000.
[36] M. I. Natick, MA, “Curve Fitting Toolbox User‟s
Guide, Release 2007a,” 2007.
[37] A. B. D.Kundu, “Statisticl Computing -Existing
Methods And Recent Developments,” Harrow; U.K,
Alpha Science Internation Ltd., 2004.
[38] C. C. J.K. Taylor, “Statistical technologies for Data
Analysis,” 2nd Edition, Florida,U.S.A: Chapman and
Hall/CRC, 2004.
[39] M. H. H.Demuth, M.Beale, “Neural Network Toolbox
User‟s Guide.” 2005.

System identification of a steam distillation pilot scale using arx and narx approaches

More Related Content

What's hot (8)

Viewers also liked (9)

Similar to System identification of a steam distillation pilot scale using arx and narx approaches (20)

More from eSAT Journals (20)

Recently uploaded (20)

System identification of a steam distillation pilot scale using arx and narx approaches