You are on page 1of 12

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/308969667

Instrumentation and Process Control: Process Control

Chapter · January 2011


DOI: 10.1016/B978-0-12-374407-4.00413-1

CITATIONS READS

0 4,715

1 author:

Petia Georgieva
University of Aveiro
151 PUBLICATIONS 1,125 CITATIONS

SEE PROFILE

All content following this page was uploaded by Petia Georgieva on 10 October 2016.

The user has requested enhancement of the downloaded file.


Provided for non-commercial research and educational use only.
Not for reproduction, distribution or commercial use.

This article was originally published in Encyclopedia of Dairy Sciences, Second


Edition, published by Elsevier, and the attached copy is provided by Elsevier for the
author’s benefit and for the benefit of the author’s institution, for non-commercial
research and educational use including without limitation use in instruction at your
institution, sending it to specific colleagues who you know, and providing a copy to
your institution’s administrator.

All other uses, reproduction and distribution, including without limitation


commercial reprints, selling or licensing copies or access, or posting on open
internet sites, your personal or institution’s website or repository, are prohibited.
For exceptions, permission may be sought for such use through Elsevier’s
permissions site at:

http://www.elsevier.com/locate/permissionusematerial

Georgieva P (2011) Plant and Equipment | Instrumentation and Process Control:


Process Control In: Fuquay JW, Fox PF and McSweeney PLH (eds.), Encyclopedia
of Dairy Sciences, Second Edition, vol. 4, pp. 242–251. San Diego: Academic Press.

ª 2011 Elsevier Ltd. All rights reserved.


Author's personal copy

Instrumentation and Process Control: Process Control


P Georgieva, University of Aveiro, Aveiro, Portugal
ª 2011 Elsevier Ltd. All rights reserved.

This article is a revision of the previous edition article by


R. Oliveira, P. Georgieva, and S. Feyo de Azevedo, Volume 3,
pp 1392–1401, ª 2002, Elsevier Ltd.

Introduction – Dairy Industry, Aims, equipment, and new control algorithms are solutions
Trends implemented by an increasing number of companies.
For example, specialized processes such as ultrafiltration
Dairy processing industries worldwide have undergone and modern drying processes have increased the oppor-
rationalization, with a trend toward fewer but larger tunity for the recovery of milk solids that were formerly
plants specialized in a limited range of products and discharged.
operated by fewer people. Plants producing market milk ME provides a bridge between ERP systems and real-
and products with short shelf-life, such as yogurts, creams, time control supplied by PLC and SCADA systems, for
and soft cheeses, tend to be located on the fringe of urban functions such as batch process control and production
centers close to consumer markets. Plants manufacturing scheduling.
items with longer shelf-life, such as butter, milk powders,
cheese, and whey powders, tend to be located in rural
Batch and Semibatch Process Operation
areas closer to the milk supply.
Basic dairy processes have changed little in the past Batch or fed-batch mode of operation is a typical pro-
decade. However, changing global needs and the pressure duction scheme for a large group of dairy processes. It is
of globalization create contemporary challenges related to related to the formulation of control problems in terms
rapid development of new, high-value-added products, of the economic or performance objective at the end of
improvements of production efficiency, and more energy- the process; for example, milk powder production qual-
efficient, cost-saving, and environmentally friendly pro- ity is evaluated by the total quantity and concentration
cesses. Issues such as pollution prevention and reduction of solids at the end of the process. The main challenge of
of waste loads and product losses are critical for the dairy batch production is the large batch-to-batch variation
industry. To achieve these requirements, an integrated of the final characteristics. This lack of process repeat-
smart production policy is expected based on a holistic ability is an inherent consequence of the variability of
plant-wide representation of the production chain from raw materials and uncontrolled parameters and not
process control level, through plant management, to cor- necessarily due to improper control policies, though
porate management. Modern dairy technology consists of the latter would greatly enhance the problem. Over the
a few hierarchical stages: past 10 years, a number of control methods have been
Enterprise resource planning (ERP), which refers to a researched to cope with process constraints and different
system that sits at the top of an integrated dairy infra- objectives derived from economic or environmental con-
structure to manage the demands of business, in terms siderations with the objective to drive the process to its
of functions such as sales and distribution, accounting, optimal state of profit maximization and cost minimiza-
materials management, asset management, and plant tion. Some of the most typical control paradigms and
maintenance. new model-based control trends are presented in the rest
Manufacturing execution (ME) refers to a system that of this article.
centers on the product as it moves through the plant.
Supervisory control and data acquisition (SCADA) is a
system that provides the operator and other users with Classical Closed-Loop Process
access to direct regulatory process control, such as pro- Control – Established Control
grammable logic controllers (PLCs). The SCADA system Engineering Practice
works in real time and provides graphical status displays
and process monitoring. Though the PLCs form the basis Any control system in which the output is monitored
of most traditional process control systems in the dairy (measured), compared (subtracted) with the reference
industry, the use of more sophisticated control systems, (desired) input, and the difference (the error) used to
information technology (IT)-based modern control actuate the controller until the output equals the

242
Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251
Author's personal copy
Plant and Equipment | Instrumentation and Process Control: Process Control 243

+
e(t) u(t) Statistical Process Control
Reference Controller Process Controlled
input output
– Statistical process control (SPC) is one of the most power-
Measured ful tools for process improvement. Although SPC is new
Sensor
output to the dairy industry, it is not really a new approach. It has
Figure 1 Closed-loop process control. been used successfully in manufacturing businesses for
more than 60 years. SPC is a set of analytical tools of
which the control chart is one of the most important.
reference is called a closed-loop or feedback control sys- Control charts are helpful in signaling that a change has
tem (see Figure 1). occurred. The fundamental concept of control charts is to
The controller design consists of computing the PID distinguish between inherent random variation and real
parameters Kp ; KI ; Kd such that the closed-loop perfor- changes in output, quality, or measured performance.
mance and stability requirements are achieved. Each SPC methods can be used to signal emerging problems
parameter is related to one of the terms (proportional, and to evaluate the positive or negative impact of a
integral, or derivative) of the control law. change in a management practice or the implementation
Proportional action. The magnitude of the contribution of a new product. Next, some of the most commonly used
of the proportional term to the overall control action is online multivariate SPC techniques are presented briefly.
determined by proportional gain, Kp . A high proportional
gain results in a large change in the output for a given
change in the error. If the proportional gain is too high, Empirical Linear Techniques
the system can become unstable. In contrast, a small gain
results in a small output response to a large input error Batch process modeling and monitoring has been always a
and a less responsive (or sensitive) controller. In the challenging problem in dairy engineering due to the pre-
sence of nonlinear behavior and serial correlation,
absence of disturbances, pure proportional control will
correlated and/or collinear data, varying batch lengths,
not settle at its target value but will retain a steady-state
and multiproduct production. Current state of the art
error that is a function of the proportional gain and the
empirical techniques include the bilinear approaches of
process gain.
multiway principal component analysis (MPCA), the
Integral action. The contribution of the integral term is
multiway partial least squares (MPLS), and the trilinear
proportional to the error magnitude and error duration.
methodologies of parallel factor analysis (PARAFAC) and
Summing the instantaneous errors over time (integrating
PARAFAC II. Although the above bilinear and trilinear
the error) gives the accumulated offset. The magnitude of
techniques have been applied successfully to batch pro-
the contribution of the integral term to the overall control
cesses, they experience a number of limitations. For
action is determined by the integral gain, KI . The integral
example, they do not incorporate the process dynamics,
action (when added to the proportional action – PI con-
and with the exception of PARAFAC II, the duration of
troller) accelerates the movement of the process toward
the batches is assumed to be constant. Moreover, for
the setpoint and eliminates the residual steady-state error
online monitoring, it is required that the whole batch
that occurs with a proportional-only controller. However,
trajectory is known or is predictable. This requirement
because the integral term responds to accumulated errors
results in certain assumptions being made to in-fill the
from the past, it increases oscillations and overshoots of
unknown future values of the batch trajectory. Finally, all
the setpoint.
these techniques are linear and to a greater or lesser
Derivative action. The rate of change of the process
extent fail to capture the nonlinear nature of a batch
error is calculated by determining the slope of the error
process. Alternative approaches that overcome the issues
over time (its first derivative over time) and multiply-
of data in-filling and unequal batches are presented next.
ing this rate of change by the derivative gain, Kd . The
magnitude of the contribution of the derivative action
to the overall control action is termed the derivative
Moving Window Principal Component Analysis
gain, Kd . The derivative action slows the rate of
change of the controller output and thus decreases In moving window principal component analysis
oscillations and overshoots, an effect that is most (MWPCA), typically, measurements from a batch process
noticeable close to the controller setpoint. However, are arranged in a three-dimensional matrix X
differentiation of a signal amplifies noise, and thus (NB  NV  NT) where NB, NV, and NT are the number
this term in the controller is highly sensitive to noise of batches, variables, and time instants (Figure 2(a)). The
in the error term and can cause a process to become three-dimensional matrix X can be transformed to a bidi-
unstable if the noise and the derivative gain are mensional matrix by unfolding over the batch dimension
sufficiently large. (NB  (NV  NT)), as shown in Figure 2.

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
244 Plant and Equipment | Instrumentation and Process Control: Process Control

(a)
NT
Time
1

Batches x

NR Variable NV


(b)
t=1 → t = ti -1 t = ti → t = NT
1

NB
1 → NV 1 → NV 1 → NV 1 → NV 1 → NV


(c) 1→k
1
Pt

2 ∗NV

Figure 2 Moving window principal component analysis (MWPCA).

A scaling is usually applied to the unfolded matrix X T 2 ¼ tt St– 1 tTt ½3


before an ordinary principal components analysis (PCA).
where tk are the k retained PCA scores and S is their
The mean of each column of X is subtracted from each
covariance matrix. The SPE is then calculated as follows:
data element of this column. This way of mean centering
is very important as it results in the removal of the main SPE ¼ et eTt ½4
nonlinear component in the data. Furthermore, by scaling  
the variables in each column of X, the differences in the e ¼ xnew I – Pk PTk ½5
measurement units between variables can be handled to
allow equal weight to be given to each variable at each
time interval. A PCA model is then developed on a mov- Batch Dynamic Principal Component Analysis
ing window of data. Having selected the length of the
The MWPCA approach does not capture the dynamic
moving window (L), MWPCA then develops NTL þ 1
behavior within a batch process. The batch dynamic princi-
PCA models for each time interval by decomposing the pal component analysis (BDPCA) is an alternative method
(NB  NV) matrix X into a systematic and noisy part: that uses lagged variables to incorporate process dynamics.
X ¼ Tk PTk þ E ½1 More specifically in BDPCA, each batch is isolated from the
others (see Figure 3(a)). A matrix Xiv (NT  NV) is formed
where Tk and Pk are the matrices of the k retained for each iv batch. Then each of the NV variables is lagged d
principal component scores and loading respectively, times resulting in a lagged Xiv [(NT  d)  (NV ? (d þ 1))]
and E is the matrix of the residuals. The number of the matrix (see Figure 3(b)). The covariance matrix of the
retained principal components, k, is usually determined lagged Xiv matrix, Siv, is then calculated (see Figure 3(c)).
by the means of cross-validation. For the application The procedure is repeated for all NB batches, resulting in
exemplified, the order of the moving window was NB Siv covariance matrices. The elements in each of the Siv
selected to be L ¼ 2. For each PCA model, the loading matrices are a measure of the dynamic relationship between
matrix Pk is stored. Having performed a PCA analysis, a
variables in batch iv. Having calculated these dynamic cor-
set of online monitoring tools can then be developed.
relations for all NB batches, an average covariance matrix,
Typically, these tools are Hotelling’s T 2 and squared
Savg, is then calculated based on the NB Siv covariance
prediction error (SPE) control charts. For instance, con-
matrices (see Figure 3(d)):
sidering that a new batch xnew is to be monitored, the
Hotelling’s T 2 is calculated using the k retained PCA
P
NB
scores by: ðNT – d – 1Þ Siv
iv¼1
Savg ¼ ½6
tk ¼ xnew Pk ½2 NV ðNT – d Þ

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
Plant and Equipment | Instrumentation and Process Control: Process Control 245

Lagged
Variables Variables
1 → NV 1 → (K + 1) × NV
1
↓ Batch 1 ⇒ Xd, 1 ⇒ S1
NT

1
↓ Batch 2 ⇒ Xd, 2 ⇒ S2
NT

1
↓ Batch 3 ⇒ Xd, 3 ⇒ S3 ⇒ Savg ⇒ P
NT

1
↓ …. ⇒ …. ⇒ …
NT

1
↓ Batch NB ⇒ Xd, NV ⇒ SNB
NT
(a) (b) (c) (d) (e)

Figure 3 Batch dynamic principal component analysis (BDPCA).

The average covariance matrix, Savg, expresses the aver- noisy part as in eqn [1]. Partial least squares (PLS) ana-
age dynamic relationships between the process lysis is then performed between the unfolded matrix X
measurements. A PCA model is then developed based and the dummy y vector:
on Savg (see Figure 3(e)). The resulting BDPCA model
y ¼ Tk c þ f ½7
is finally used to calculate the T2 and SPE statistics for
monitoring purposes. where c is the regression vector of y onto the PLS scores
Tk (see Figure 4(b)) and f are the PLS model residuals.
The number of PLS latent variables to be retained are
Batch Observation Level
selected as those that provide an adequate description of
The batch observation level (BOL) method considers the both the X and y spaces. For setting up an online mon-
problem with unequal batch lengths. In BOL, the original itoring scheme, the PLS scores retained are then
three-way data is unfolded over the variable’s dimension rearranged over the batch dimension resulting in an
(see Figure 4(a)). A dummy y-variable that can be a time [NB  (NT?k)] matrix (see Figure 4(c)) and their mean
index or a batch maturity index is then specified. Data are and standard deviation calculated for each sample point
scaled and matrix X is transformed into a systematic and and stored (see Figure 4(d)). In an online situation,

Variables Scores
1 → NV 1→k
1
↓ Batch 1
NT (c) t=1 t=2 t=3 … t = NT
1 1→k 1→k 1→k 1→k 1→k
↓ Batch 2 ⇒ ⇒ 1
NT ↓
1 NT
↓ Batch 3

NT
1 (d) t=1 t=2 t=3 … t = NT
↓ …. 1→k 1→k 1→k 1→k 1→k
NT mean
1 std
↓ Batch NB
NT
(a) (b)
Figure 4 Batch observation level (BOL).

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
246 Plant and Equipment | Instrumentation and Process Control: Process Control

when a new sample is obtained, the scores are calculated [11]) of the process, new latent variables that provide a
initially and then scaled using the mean and the standard reliable approximation of the true system states can be
deviation of the corresponding sample point. These calculated. PCR scores capture the variability between
scaled scores are plotted against their control limits in process measurements, while PLS and CVA latent vari-
univariate score plot charts. Similar to the previous ables are those linear combinations of the past that
approaches, T2 and SPE charts can also be constructed. include the information required to predict process
future. The result of applying either a PCR, PLS, or
CVA analysis is a weighting matrix Jt , which is used to
Time-Varying State Space Modeling identify the system states through the past vector pt :

Time-varying state space (TVSS) modeling is an alter- tt ¼ Jt pt ½12


native approach for batch process modeling and Once the system states have been identified, the state
monitoring with the following state space model: space matrices can be computed using a least squares
tt þ1 ¼ Ct tt þ wt ½8 solution.

y t ¼ H t tt þ e t ½9

where t is the system states, y is the available process Intelligent Control


measurements, and w and e are the state and output
residuals with covariance matrices Q and R, respectively. The process modeling and monitoring techniques dis-
Finally, C and H are the state space model matrices, cussed in the section ‘Statistical Process Control’ are the
which are assumed to be time-varying as they aim to first step toward developing modern computer-based
describe a nonstationary process. To develop the model, control systems able to collect a large amount of process
the data are initially unfolded and scaled as in MWPCA. operational data, store it in databases, and display it to the
The procedure to compute the TVSS matrices C and H operator. The next step is the subsequent decision making
then proceeds through the identification of the system that still relies mainly on the human operators; however,
states. For a time interval t ¼ k, the past and the future of new concepts and methodologies for automatic analysis
the system are defined as shown in Figure 5(a). The past and developing a computationally intelligent control (IC)
(p) of the process is associated with the past process are slowly being introduced in dairy manufacturing. IC
measurements of all batches at time k up to a specific deals with the application of data mining, machine learn-
lag (in Figure 5(a), the time lag, K, was set to a value of ing, and knowledge discovery paradigms, artificial
two): intelligence, expert systems, fuzzy logic, and neural net-
 T works for controlling complex physical processes that are
pt ¼ yt –1 yt –2    yt –K ½10
difficult to control using conventional methods. The main
The future ( f ) of the process is the current and future modules of an IC system are discussed below.
process measurements of all batches (in Figure 5(a), the
future horizon, L, is set up to a value of one):
Perception Subsystem
 T
f t ¼ yt yt þ1    yt þL ½11 Information from the plant and the environment is col-
Now, by applying any one of either PLS, principal com- lected and processed into a form suitable for perception.
ponent regression (PCR), or canonical variate analysis Basic elements of a perception subsystem are
(CVA) between the past (eqn [10]) and the future (eqn
• Sensor
data
arrays – provide raw plant and environmental

(a)
t=1 t = ti -2 t = ti -1 t = ti t = NT
• Signal processing – transforms data into information
and knowledge
1

• Data fusion and pattern recognition – uses multidi-
mensional and varying nature data spaces to extract
NB underlying patterns describing the plant and the
1 → NV 1 → NV 1 → NV 1 → NV 1 → NV environment

(b)
Cognition Subsystem
Past ↔ Future → T In an IC framework, cognition is concerned with the
decision-making process under conditions of uncertainty.
Figure 5 Time-varying state space model (TVSS). Basic activities of a cognition subsystem are

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
Plant and Equipment | Instrumentation and Process Control: Process Control 247

• Reasoning
fuzzy logic
– using knowledge-based algorithms and membership functions, known as fuzzy sets. Fuzzy set
theory provides a means for representing uncertainty. In
• Strategic planning – using adaptive search and genetic
algorithms for optimum policy evaluation and path
general, probability theory is the primary tool for analyz-
ing uncertainty and assumes that uncertainty is a random
prediction process. However, not all uncertainty is random, and
• Learning – using adaptive supervised (teacher
supported) or unsupervised (self) learning paradigms
fuzzy set theory is used to model the kind of uncertainty
associated with imprecision, vagueness, and lack of
information.
Conventional set theory distinguishes elements that are
The Actuator Subsystem
members or not members of a set, with very clear, crisp
The actuators operate using signals from the cognition boundaries between them. For example, temperatures
subsystem to drive the plant to some desired state. In the between 20  C and 30  C belong to the crisp set ‘medium
event of actuator/sensor failure, the IC system has to be temperature’, and all temperatures between these bound-
able to reconfigure its control strategy. aries have a membership value of one ( ¼ 1). The central
Though milk production is a traditional dairy process concept of fuzzy set theory is that the membership function
for which conventional controllers have been used inten-  can have a value between 0 and 1. The shape of the
sively, IC appears to be quite plausible for such processes. membership function is also known as the universe of
If the temperature drifts during processing, safety can be discourse. Among the most typical fuzzy set shapes are
compromised or the product flavor or texture might be symmetrical triangles, trapezoids, and Gaussian or bell-
ruined. To perform the delicate balancing act that dairy shaped curves. Each set is given a linguistic label to identify
products demand, plant operators need complete control, it; for example, positive big (PB), positive small (PS), about
including the ability to know precisely what is happening zero (Z), negative big (NB), negative small (NS). The size
at every moment and to perform urgent changes on time. of the universe of discourse depends on the range of the
It is not easy, especially when the product is hidden from variable and the number of the sets. The number and shape
sight in tanks and pipes more than 90% of the time. of fuzzy sets are a trade-off between precision of control
Companies that process the same products day in and action and real-time computational complexity.
day out have experience on their side, but those that
change frequently have to be very agile and adaptable.
In this sense, fuzzy and model predictive controllers are
Fuzzy Rulebase Module
very good examples for building IC systems that are well
accepted generally in the process industry. The fuzzy rulebase consists of a set of antecedent-
consequent linguistic rules of the form
Example: OR IF e is PS AND ce is NS THEN u is PS
Fuzzy Logic Control System
½13
The basic structure of a fuzzy logic control system where e is the error (process reference – measured value),
(FLCS) is shown in Figure 6. deðt Þ
ce is the rate of change of the error ce ¼ , and the
dt
objective of the FLC (see Figure 6) is to minimize e
Fuzzification Module
and ce.
Fuzzification is a process of mapping the input variables In this example, the FLC input variables are e and ce,
to the fuzzy logic controller (FLC) into a set of the fuzzy logic controller (FLC) output variable is u, and

Data base & FL controller


Rule base

r(t) e(t) u(t)


+ y(t)
Fuzzy
Fuzzification Defuzzification Plant
Inference

Meaurement
system

Figure 6 Fuzzy logic (FL) control system.

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
248 Plant and Equipment | Instrumentation and Process Control: Process Control

they all have values defined as fuzzy sets. The fuzzy behavior of the modeled plant. The mechanistic
conditional statement eqn [13] is often called a model is a result of extensive, specially designed
Mamdani-type rule after Mamdani who first used it to experiments and domain knowledge on the physical
control a steam plant. The rule base is constructed using a laws that govern the process at hand. Analytical mod-
priori knowledge from the following sources: eling is time and resource consuming, but it has the
main advantage of permitting good generalizations and
• Physical laws that govern the plant dynamics.
scale-up.
• Imprecise heuristic
Data from other controllers.

• enced plant operatorsknowledge obtained from experi-
and experts.
Data-driven alternatives (black-box models) are based on
data mining and machine learning techniques and aim
at extracting process knowledge from databases col-
lected during the normal operation of the plant. The
Fuzzy Inference Module
development of data-driven models usually takes less
Fuzzy inference is the process of mapping membership time and resources; however, their generalization out-
values from the input variable(s) through the rulebase to side the data space used to build the model is poor. One
the output variable(s). of the most common black-box models is the artificial
Example (continued): If at a certain moment the mem- neural network (ANN) paradigm, which will be
bership values of the input variables are described in more detail later.
PS ðeÞ ¼ a; NS ðceÞ ¼ b ½14 • Hybrid modeling (gray-box model) is a combination of
the two previous approaches and is also known as
based on fuzzy set operations and the fuzzy rule eqn [13], knowledge-based hybrid modeling (KBHM). KBHM
the membership value of the control action u is offers a reasonable compromise between the exten-
computed as sive efforts to obtain a fully parameterized structure
PS ðuÞ ¼ max½minðPS ðeÞ; NS ðceÞÞ ½15 and the poor generalization of the data-driven
models.
Equation [15] is referred to as the max-min inference
process or max-min fuzzy reasoning. Process modeling for the dairy industry is strongly
influenced by the recent trends in building data-based
or KBHM models. For example, fouling, the unwanted
Defuzzification Module
formation of deposits on heated surfaces, is a major
Defuzzification is the process of mapping from a set of unsolved problem in the dairy industry. A direct con-
inferred fuzzy control signals belonging to fuzzy sets to a sequence of fouling is reduction in the processing
nonfuzzy (crisp) control signal. The center of area is the efficiency, because the material deposited disturbs
most well-known defuzzification technique, which can be both the fluid flow and the heat transfer, which in
expressed as turn may impair product quality. Additionally, the
P
n deposit removal shortens the running time between
C ðAðui ÞÞAðui Þ cleaning cycles and thus increases the costs. As a result,
u ðcrispÞ ¼ i¼1 P n ½16 daily cleaning is a common practice in the dairy indus-
Aðui Þ try and is necessary for hygienic and product quality
i¼1
requirements. The additional annual costs in the dairy
where Aðui Þ is the area of the ith membership function industry caused by fouling are estimated at US$260
computed at the previous step (the fuzzy inference), and million per year. A relevant model of fouling would
C ðAðui ÞÞ is the center of the respective area. be a powerful tool for the development of strategies to
avoid or reduce this unwanted process. However, due
to its very complex nature, milk fouling is only par-
Model-Based Predictive Control tially understood, and its modeling is a challenging
Modeling Approaches issue. A KBHM model combining parameterized equa-
tions (for fluid flow, heat and mass transfer) with
Process modeling strategies can be divided into the qualitative knowledge in fuzzy logic form (for protein
following main streams: and salt deposition) was developed a few years ago.
• with
The analytical approach (white-box models) is concerned
building the so-called mechanistic model, also
The model describes the fouling behavior with regard
to the temperature and the pressure drop in a time-
known as first-principles model, which is a set of dependent manner, and it is not restricted to a certain
mathematical expressions that reflect the dynamic (dif- dairy product or plant configuration. KBHM was suc-
ferential equations) or static (algebraic equations) cessfully used for testing technological improvements

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
Plant and Equipment | Instrumentation and Process Control: Process Control 249

in the heat treatment of milk in tubular heat ANNs have been applied to design robust neural con-
exchangers. trollers with guaranteed stability and reference tracking.
The neural control problem can be approached in a
direct or indirect control design framework. Direct
Model Predictive Control – General Formulation ANN control means that the controller has an ANN
structure, whereas in the indirect ANN control scheme,
The term model predictive control (MPC) does not
first an ANN is used to model the process to be con-
refer to a particular control method; instead, it cor-
trolled, and this model is then used in a more
responds with a general control approach. The MPC
conventional controller design. The implementation of
concept, introduced in the late 1970s, has evolved to
the first approach is simple, but the design and the
a mature level and has become an attractive control
tuning are rather challenging. The indirect design is
strategy implemented in a variety of process indus-
very flexible, the model is typically trained in advance,
tries and in the dairy industry in particular. The
and the controller is designed online. Moreover, the
main difference between the MPC configurations is
ANNs appear to be rather convenient numerical models
the model used to predict the future behavior of the
when dealing with nonlinear systems or in general with
process or the implemented optimization procedure.
systems for which data are available but little is known
First, the MPC based on linear models gained popu-
on the physical mechanisms that determine their
larity as an industrial alternative to PID control, and
dynamics.
later nonlinear cases such as tubular heat exchangers
The most popular ANN structures for modeling
and drying processes were reported as successfully
reasons are feedforward networks (FFNNs) and recur-
MPC-controlled processes. MPC is an optimization-
rent networks (RNNs). The RNNs are most suitable
based multivariable constrained control technique
for dynamic system modeling, due to the memory
that uses a dynamic model, from the types described
introduced by the recurrent (delayed and/or fed back)
in the previous section, for process output
signals (see Figure 7). Normally, the RNN has two
predictions.
vector inputs (r and p) formed by past values of the
At each sampling time, the model is updated on the
process input and the network output, respectively. A
basis of new measurements and state variables esti-
linear activation function is often located at the output
mates. Then, the open-loop optimal manipulated
node (layer 2), and S-shape functions are usually the
variable moves are computed over a finite (predefined)
hidden nodes (layer 1).
prediction horizon with respect to some performance
index, and the manipulated variables for the subsequent
prediction horizon are implemented. The prediction
horizon is shifted or shrunk by usually one sampling ANN Error Tolerant Model Predictive Control
time into the future, and the previous steps are MPC controllers proved to be the most promising alter-
repeated. native to the traditional PID control that has the
potential to overcome the problem of the lack of repeat-
Artificial Neural Networks ability and related product recycling or loss increase. It
has been applied successfully in batch dairy processes
Over the past 20 years, ANNs became a well-established assuming linear process dynamics. However, online
methodology not only as a reliable classifier with count- execution of MPC with predictions running on a large
less applications but also as a data-driven modeling number of empirical and analytical nonlinear algebraic
framework. The remarkable success of the ANN differential equations (the process model) make this
approach is in great part due to the following features: alternative computationally more demanding or even
1. ANNs are universal approximators. It has been proved
that any continuous nonlinear function can be approxi-
Input Layer 1 Layer 2 Output
mated arbitrarily well over a compact set by a
multilayer ANNs, which consists of one or more Two r (7×1)
W1_1
hidden layers. Delays Tansig
(7×2) n1 (1×1)
2. Learning and adaptation. The intelligence of ANNs (2×1) Σ w2_1 Purelin
comes from their generalization ability with respect to b1 n2
(7×1) (1×7) Σ
unknown data. Online adaptation of the weights is (7×1)
b2 (1×1)
possible. Two p
W1_2 (1×1)
Delays (7×1)
3. Multivariable systems. ANNs may have many inputs (7×2)
(2×1)
and outputs, which makes it easy to model multivari-
able systems. Figure 7 Recurrent neural network (RNN) architecture.

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
250 Plant and Equipment | Instrumentation and Process Control: Process Control

unfeasible for dairy processes with fast nonlinear NN MPC


dynamics. Even for processes where the standard control
approach (e.g., ladder logic) is not the best solution, the
implementation of MPC is impeded due to high compu- Optimization ANN
procedure process model
tational costs.
A recent modification of the classical MPC, termed
ANN Error Tolerant (ET) MPC, reduces considerably Process
the average duration of each optimization step and makes (KBHM)
the MPC computationally more efficient and attractive
Figure 8 ANN-based model predictive control (MPC).
for industrial applications (see Figure 8). The ANN
model is integrated into the controller, and the The discretized version of the modified performance
optimization procedure is executed only when the error index is
is above a predefined value.

8 ( Hp
>
> X X
Hc
>
> u: min F ¼ l ð eðt þ kÞ Þ 2
– l ðuðt þ kÞÞ2 ; if EP > 
< ½uðt þkÞ;uðt þkþ1;...uðt þHc ÞÞ
1 2
k¼1 k¼1
uðt þ kÞ ¼ ½17
>
> ; PRþ
>
>
:
u ; if EP < 

1 PHp  
 follow publications in the Journal of Food Engineering,
where EP ¼ k ¼ 1 e ðt þ kÞ ;eðt þ kÞ ¼ ref ðt þ kÞ –
Hp Journal of Biotechnology and Bioengineering, and the
yp ðt þ kÞ;uðt þ kÞ ¼ uðt þ k – 1Þ – uðt þ k – 2Þ; is the International Dairy Journal.
prediction model response. The prediction horizon Hp is
the number of time steps over which the prediction errors
are minimized, and the control horizon Hc is the number
of time steps over which the control increments are mini- Acknowledgment
mized. u ðt þ kÞ; u ðt þ k þ 1Þ; . . . u ðt þ Hc Þ are
tentative values of the future control signal, which are This work was financially supported by the Portuguese
limited by umin and umax . The controller is denoted as an Foundation for Science and Technology within the activ-
ET MPC formulation because the optimization is per- ity of the Institute of Electronic Engineering and
formed only when the error function EP is bigger than a Telematics of Aveiro (IEETA).
predefined real positive value . To reduce the computa-
tional burden when the error is less than , the control See also: Plant and Equipment: Instrumentation and
action is equal to u , which is the last value of u, computed Process Control: Instrumentation.
before the error enters the  strip. Note that EP in eqn [17]
is defined as the mean value of the future errors, between the
predicted output and its reference along the next Hp steps. Further Reading
Braha D (2001) Data Mining for Design and Manufacturing: Methods
and Applications (Massive Computing). Dordrecht, Netherlands:
Conclusion Kluwer Academic Publishers.
Burns RS (2001) Advanced Control Engineering. Linare House, Jordan
In new integrated dairy plants, each process is carried out in Hill, Oxford: Butterworth-Heinemann.
Camacho EF and Bordons C (2004) Model Predictive Control in the
multiple phases, and there exists strong nonlinear and Process Industry. London: Springer-Verlag.
dynamic effects between the variables. Therefore, modern Haykin S (1999) Neural Networks: A Comprehensive Foundation. Upper
process control systems have usually a hierarchical archi- Saddle River, NJ: Prentice Hall.
Nagy ZK and Braatz RD (2003) Robust nonlinear model predictive
tecture including decentralized controllers, remote input control of batch processes. AIChE Journal 49: 1776–1786.
and output (I/O) modules, fieldbus systems, local area net- Norgaard M, Ravn O, Poulsen NK, and Hansen LK (2000) Neural
works (LANs), and smart sensors and other devices. The Networks for Modelling and Control of Dynamic Systems. London:
Springer-Verlag.
huge amount of information flows are stored and used on- Oliveira C, Georgieva P, Rocha F, Ferreira A, and Feyo de Azevedo S
line or off-line for executing the IC alternatives like FLC, (2007) Dynamical model of brushite precipitation. Journal of Crystal
ANN MPC, or SPC, all methods described in this article. Growth 305: 201–210.
Oliveira C, Georgieva P, Rocha F, and Feyo de Azevedo S (2008)
Interested readers are advised to consult not only the Artificial neural networks for modeling in reaction process systems.
references in the ‘Further Reading’ section but also to Neural Computing & Applications 18: 15–24.

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251


Author's personal copy
Plant and Equipment | Instrumentation and Process Control: Process Control 251

Petermeier H, Benning R, Delgado A, and Becker T (2002) Simoglou A, Martin EB, and Morris AJ (2002) Statistical performance
Hybrid model of the fouling process in tubular heat monitoring of dynamic multivariate processes using state space
exchangers for the dairy industry. Journal of Food Engineering modelling. Computers & Chemical Engineering 26: 909–920.
55: 9–17. Suárez LAP, Georgieva P, and Feyo de Azevedo S (2009)
Rossiter JA (2003) Model Based Predictive Control. A Practical Computationally efficient process control with neural network-based
Approach. New York: CRC Press. predictive models. International Joint Conference on Neural
Simoglou A, Georgieva P, Martin EB, Morris J, and Networks (IJCNN), Atlanta, GA, USA, 14– 9 June.
Feyo de Azevedo S (2005) online monitoring of a sugar Wang XZ (1999) Data Mining and Knowledge Discovery for Process
crystallization process. Computers & Chemical Engineering Monitoring and Control (Advances in Industrial Control). London:
29: 1411–1422. Springer.

Encyclopedia of Dairy Sciences, Second Edition, 2011, Vol. 4, 242-251

View publication stats

You might also like