Available on line at

Association of the Chemical Engineers AChE

www.ache.org.rs/CICEQ
Chemical Industry & Chemical Engineering Quarterly 15 (3) 175−187 (2009) CI&CEQ

S.K. LAHIRI SOFT SENSOR DEVELOPMENT AND OPTI-

NADEEM M. KHALFE MIZATION OF THE COMMERCIAL PETRO-
National institute of technology, CHEMICAL PLANT INTEGRATING
Durgapur, India
SUPPORT VECTOR REGRESSION
SCIENTIFIC PAPER AND GENETIC ALGORITHM
UDC 519:66:665.6
Soft sensors have been widely used in the industrial process control to im-
DOI: 10.2298/CICEQ0903175L prove the quality of the product and assure safety in the production. The core
of a soft sensor is to construct a soft sensing model. This paper introduces
support vector regression (SVR), a new powerful machine learning method
based on a statistical learning theory (SLT) into soft sensor modeling and pro-
poses a new soft sensing modeling method based on SVR. This paper pre-
sents an artificial intelligence based hybrid soft sensormodeling and optimaza-
tion strategies, namely support vector regression – genetic algorithm (SVR-GA)
for modeling and optimization of mono ethylene glycol (MEG) quality variable
in a commercial glycol plant. In the SVR-GA approach, a support vector re-
gression model is constructed for correlating the process data comprising va-
lues of operating and performance variables. Next, model inputs describing the
process operating variables are optimized using genetic algorithm with a view
to maximize the process performance. The SVR-GA is a new strategy for soft
sensor modeling and optimization. The major advantage of the strategies is
that modeling and optimization can be conducted exclusively from the historic
process data wherein the detailed knowledge of process phenomenology (re-
action mechanism, kinetics etc.) is not required. Using SVR-GA strategy, a
number of sets of optimized operating conditions were found. The optimized
solutions, when verified in an actual plant, resulted in a significant improve-
ment in the quality.
Key words: SVR; GA; modeling and optimization.

Cut throat competition, open global market and tween operating conditions and product quality, the
shrinking profit margin forced the process industries problem remains unsolved. No industry acceptable
to monitor and improve the product quality through a first-principle model is available for the product quality
faster and more systamatic way. Although the most improvement. Another option to solve this difficult si-
reliable approach to quality improvement will be the tuation is the use of historic operation data. In the last
use of precise first-principle models, such models are decade or so, data-based approaches have been wi-
not available in most newly developed processes and dely accepted for process control and monitoring in
modeling of a complex industrial process is very diffi- various industries. In the petrochemical industry, for
cult and time-consuming. In the glycol industry, for ex- example, lots of operating and QC variable data are
ample, the relationship of operating conditions to pro- generated in every few seconds from multitude of
duct quality is not clear. The product qualities have sensors. To achieve the product quality improvement,
been usually maintained by skilled operators on the we need to develop a system having at least the
basis of their experience and intuition. Although much following functions: 1) to predict the product quality
effort has been devoted to clarify the relationship be- from operating conditions, 2) to derive better opera-
ting conditions that can improve the product quality
Corresponding author: S.K. Lahiri, National institute of technology, and 3) to detect faults or malfunctions for preventing
Durgapur, India. an undesirable operation. The first function is realized
E-mail: sk_lahiri@hotmail.com
Paper received: 14 November, 2008
by developing a soft-sensor, which is a mathematical
Paper revised: 2 February, 2009 model to relate operating conditions to the product
Paper accepted: 3 August, 2009

175
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

quality. On the basis of the model, the second func- tistics, artificial neural network and support vector re-
tion is realized by formulating and solving an optima- gressions (SVR). In recent years, the support vector
zation problem. The third function is realized by a regression has been widely used (Vapnik [1,2] and
multivariate statistical process control (MSPC). Vapnik, Smola and Golowich [3]) as a useful tool to
When hardware sensors are not available, soft- the nonlinear soft sensing modeling. SVR is a com-
-sensors are key technologies for producing high puter modeling approach that learns from examples
quality products. Even when hardware sensors can through iterations without requiring a prior knowledge
be used, operators and engineers have found the pro- of the relationships of process parameters and QC
blems listed in Table 1. These problems with hard- variable. It is also capable of dealing with uncertain-
ware sensors were identified as the results of a ques- ties, noisy data, and non-linear relationships. SVR
tionnaire to 26 companies in Japan (PSE 143 Com- modeling has been known as “effortless computation”
mittee, 2004). Soft-sensors are judged to be useful for and readily used extensively due to their model-free
addressing these problems. approximation capabilities of complex decision-ma-
king processes. Once an SVR based process model
Table 1. Problems with hardware sensors; the results of a ques- is developed, it can be used for predicting the impor-
tionnaire to 26 companies in Japan tant quality variable. Also, it can be interfaced with
Percentage Recognized problem online DCS and continuous monitoring can be achiev-
27 Time consuming maintenance ed to yield the better process control. This SVR based
21 Need for calibration
process model can also be used for the process op-
timization to obtain the optimal values of the process
15 Aged deterioration
input variables that maximize the quality of the pro-
13 Insufficient accuracy
duct. In such situations, an efficient optimization for-
10 Long dead time, slow dynamics
malism known as Genetic Algorithm can be used. In
8 Large noise
the recent years, Genetic Algorithms (GAs) that are
2 Low reproducibility
members of the stochastic optimization formalisms
4 Others
have been used (Cartwright and Long. [5], Polifke,
Geng and Dobbeling [4], Garrard and Fraga [6],
For successful monitoring and control of che- Goldberg [7] and Hanagandi, Ploehn and M. Nikolaou
mical plants, there are important quality variables that [8]) with a great success in solving problems involving
are difficult to measure on-line, due to limitations such very large search spaces.
as cost, reliability, and long dead time. These measu- In the present paper, SVR formalism is inte-gra-
rement limitations may cause important problems ted with Genetic Algorithms to arrive at soft sensor
such as product loss, energy loss, toxic byproduct ge- modeling and process optimization strategies. The
neration, and safety problem. A soft sensor, an infe- strategy (henceforth referred to as “SVR-GA”) use an
rential model, can estimate the qualities of interest SVR as the nonlinear process modeling paradigm for
on-line using other available on-line measurements development of soft sensor, and the GA for optimizing
such as temperatures and pressures. An inferential the input space of the SVR model so that an improved
sensor provides valuable real-time information that is process performance is realized. This paper descri-
necessary for effective quality control. The major pur- bes a systematic approach to the development of in-
pose of using soft-sensors is to 1) stabilize the ferential measurements of ultraviolet (UV) transmit-
product quality via its online estimation, 2) reduce the tance (QC variable of monoethylene glycol (MEG)
energy and material consumption via an effective product in glycol plant) using SVR regression ana-
operation close to specifications/constraints and 3) lysis. After predicting the UV accurately, model inputs
validate online analyzers by comparison with the soft describing process operating variables are optimized
sensors. The soft sensor can be derived from the first using GAs with a view to maximize the UV. The SVR-
principal model when the model offers the sufficient -GA is a new strategy for the chemical process mo-
accuracy within the reasonable computation time. deling and optimization. The major advantage of the
However, due to complexity in industrial processes strategies is that modeling and optimization can be
there are cases when the first principle model is not conducted exclusively from the historic process data
available, or sometimes it takes too much time to wherein the detailed knowledge of the process phe-
compute. As a result, empirical data driven models nomenology (reaction mechanism, kinetics etc.) is not
are the most popular ones to develop soft sensors. required. The optimized solutions when verified in ac-
Empirical models are usually obtained based on va-
rious modeling techniques such as multivariate sta-

176
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

tual commercial plant resulted in a significant impro- alternative in this scenario. Making of a UV soft sen-
vement in the MEG quality. sor is not an easy task as a rigorous mathematical
model for the MEG product UV is still not available in
Development of the soft sensor for the product quality
literature which can predict UV transmittance to mini-
in a monoethylene glycol plant
mize the dependency on a lab analysis. The compre-
Recently, mono-ethylene glycol (MEG) has em- hensive process model is expected to take into ac-
erged as the most important petrochemical product as count various subjects, such as chemistry, chemical
its demand and price have been considerably rising reaction, UV deteriorating compound generation and
all over the world in the last few years. It is exten- accumulation which consequently become very com-
sively used as a main feed for the polyester fiber and plex. Industry needs this mathematical model to pre-
polyethylene tere-phthalate plastics production. UV is dict MEG UV on real time basis so that process para-
one of the most important quality parameters of MEG meters can be adjusted before the product goes off
and it indirectly represents the impurities level such specification.
as aldehyde, nitrogeneous compound and iron in the In this study, the SVR-GA strategy has been
MEG product. In laboratory, MEG product sample is used to model and optimize the MEG product UV for
exposed to UV light of different wavelengths (220, a commercial plant The optimized operating condi-
250,275 and 350 nm) and percentage of the UV light tions led to maximized UV of the product (MEG). The
transmitted through the MEG sample is measured. best sets of operating conditions obtained thereby
UV transmittance measures the presence of impu- when subjected to actual plant validation indeed re-
rities in MEG that absorb light in the ultraviolet region sulted in significant enhancements in UVs.
of the spectrum. These undesirable compounds are in
trace quantities in the parts per billion (ppb) ranges HYBRID SVR AND GA BASED MODELING
and primarily unknown in a chemical structure. Sam-
ples showing higher transmittance are considered to Process modeling and optimization formalisms
be of a greater quality grade. In Glycol plant the MEG The process optimization objective under consi-
is drawn off from MEG distillation column as a pro- deration is expressed as:
duct, its UV transmittance is affected by many things Given the process data comprising values of the
such as impurity formation in an upstream ethylene multiple process inputs and the corresponding values
oxide reactor, impurity formation and accumulation in of the process outputs ( MEG UV in this case), find
MEG column bottoms due to thermal degradation of the optimal values of the process inputs such that the
glycol, non removal and accumulation of aldehyde in prespecified measures of the process performance is
the system etc. Since these UV deteriorating impuri- maximized.
ties are in ppb ranges, they are very difficult to detect The SVR-GA strategy fulfills the above-stated
during the MEG production process and they have objective in two steps. In the first step, an SVR-based
hardly any effect on process parameters. That is why process model is developed. This model has the in-
it is very difficult for any phenomenological model for puts describing process operating parameters and
UV prediction to succeed in an industrial scenario. variables such as reflux ratio, reflux flow, MEG co-
Normally, online UV analyzers are not available to lumn top pressure, MEG column condenser pressure,
monitor the product MEG UV analysis in an ethylene MEG column control temperature, MEG column feed
glycol plant. Off-line methods for MEG quality control flow, upstream drying column control temperature,
is a common practice among the manufacturers, drying column bottom temperature, crude glycol re-
where a sample is withdrawn from the process and processing flow and its outputs represent the process
product stream for laboratory analysis several times a output variable mono-ethylene glycol UV. In the se-
day and analyzed by time consuming laboratory ana- cond step of the SVR-GA procedure, the input space
lysis. In the event of a process malfunction or ope- of the SVR model is optimized using a GA algorithm
rating under a suboptimal condition, the plant will con- such that the optimized process inputs result in the
tinue to produce an off-spec product until lab results enhanced values of the output variables.
become available. For a big world class capacity plant This optimization problem can be formulated as:
this represents a huge amount of offspec production Maximize UV = f (reflux ratio, reflux flow, MEG
results in enormous financial losses. This necessi- column top pressure, MEG column condenser pres-
tates the online UV sensors or analyzers which can sure, MEG column control temperature, MEG column
give UV continuously on a real time basis. Accurate, feed flow, drying column control temperature, drying
reliable and robust UV soft sensors can be a viable

177
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

column bottom temperature, crude glycol reprocess- At the same time, we would like this function to
sing flow); (1) be as flat as possible. Flatness in this sense means a
small w. This second aim is not as immediately intuit-
SVR-based modelling
tive as the first, but still important in the formulation of
i) Linear regression using SVR: Industrial data the optimization problem used to construct the SVR
contains noise. Normally different transmitter, signal approximation:
transmissions etc. add these noises with process pa-
rameters. Normal regression techniques try to reduce 1
minimize w ;
the prediction error on noisy training data. This empi- 2 2
rical risk minimization (ERM) principle is generally  y − < w .x i > −b ≤ ε
employed in the classical methods such as the least- subject to  i
square methods, the maximum likelihood methods < w .x i > +b − y i ≤ ε
and traditional ANN. The formulation of SVR embo- A key assumption in this formulation is that there
dies the structural risk minimization (SRM) principle, is a function f(x) that can approximate all input pairs
which has been shown to be superior, to traditional (xi,yi) with ε precession; however, this may not be the
empirical risk minimization (ERM) principle, employed case or perhaps some error allowance is desired.
by conventional neural networks. SRM minimizes an Thus the slack variable, ξi and ξi*, can be incorpo-
upper bound on the expected risk, as opposed to rated into the optimization problem to yield the follow-
ERM that minimizes the error on the training data. It is ing formulation:
this difference which equips SVM with a greater ability
l

 (ξ
to generalize.The SVR (Cherkassky and Ma [9]) al- 1
minimize w 2
+C i + ξ i∗ ) ;
gorithm attempts to position a tube around the data as 2 i =1
shown in Figure 1. ε is a precision parameter repre-
senting the radius of the tube located around the  y i − < w .x i > −b ≤ ε + ξ i 
 
regression function (see Figure 1); the region en- subject to < w .x i > +b − y i ≤ ε + ξ i∗ 
closed by the tube is known as “e-intensive zone”.  ξ i ,ξ i∗ ≥ 0 
 
The diameter of the tube should ideally be the amount
of noise in the data. The optimization criterion in SVR In Fig. 1, the sizes of the stated excess positive
penalizes those data points the y values of which lie and negative deviations are depicted by ξi and ξi*,
more than ε distance away from the fitted function, which are termed “slack” variables. Outside the [-ε,ε]
f(x). There are two basic aims in SVR. The first is to region, the slack variables assume nonzero values.
find a function f(x) that has at most ε deviation from The SVR fits f(x) to the data in such a manner that: i)
each of the targets of the training inputs. For the li- the training error is minimized by minimizing ξi and ξi*
near case, f is given by: and ii) w2 is minimized to increase the flatness of f(x)
or to penalize over complexity of the fitting function.
f(x) = <w.x> + b
The constant C > 0 determines the tradeoff be-
where <a.b> is the dot product between a and b. tween flatness (small w) and the degree to which de-

Figure 1. Schematic of SVR using an e-insensitive loss function.

178
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

viation larger than ξ are tolerated, and l is the number Techniques used in GA
of samples. This is reffered to as the ε insensitive loss For the implementation of GAs, there are certain
function proposed by Vapnik et al. which enables as a well defined steps.
sparse set of support vectors to be obtained for the
regression. Coding
ii) Nonlinear regression using SVR. Nonlinear Coding is the method by which the variables xi
function approximation can be achieved by replacing are coded into string structures. A linear mapping rule
the dot product of input vectors with a nonlinear trans- is used for the purpose of coding:
formation on the input vectors. This transformation is
x i(u ) − x i(l )
referred to as the kernel function and is represented x i = x i(l ) + (decoded value)
2l i − 1
by k(x,x’), where x and x’ are each input vectors.
Applying the kernel function to the dot product of The decoded value (of si, the binary digit in the
input vectors, we obtain: coding) is given by the rule:
 1 l  l i −1
−
 2 i , j =1

(α i − α i∗ )(α j − α ∗j )k ( x i , x j )

Decoded value=  2 s , where s ∈(0,1)
i
i i
maximize  l l  i =0

 −ε

 i =1

(α i + α i
∗
) + 
i =1
y i (α i − α i
∗ 
) 

For example, for a binary code(0111), the de-
coded value will be:
 l  (0111) = (1)20+(1)21+(1)22+(0)23 = 7
subject to  i  ∗
 (α i − α i ) = 0
 It is also evident that for a code of length l, there
 α ,α ∗ ∈ [0,C ]  are 2l combinations or codes possible. The length of
 i i 
the code depends upon the required accuracy for that
By using the Kernel function and corresponding variable.
kernel matrix, nonlinear functions approximations can
be achieved with SVR while mantaining the simplicity x i(u ) − x i(l )
Accuracy =
and computational efficiency of linear SVR approxi- 2l i
mations.
So, adhering to the above mentioned rules, it is
Training and testing possible to generate a number of guesses, or, in other
Training a support vector machine consists of an words, an initial population of coded points that lie in
iterative process in which the SVR is given the de- the given range of the function. Next step is cal-
sired inputs along with the correct outputs for those culation of fitness.
inputs. It then seeks to alter its margin (w) and bias Fitness
(b) to try and produce the correct output (within a As already been mentioned, GAs work on the
reasonable error margin). If it succeeds, it has learn- principles of “survival of the fittest”. This means that
ed the training set and is ready to perform upon pre- the good points or the points that yield maximum va-
viously unseen data. If it fails to produce the correct lues for the function are allowed to continue in the
output it re-reads the input and again tries to produce next generation, while the less profitable points are
the correct output. The margins and bias are slightly discarded from calculations. Depending upon whether
adjusted during each iteration through the training set the initial objective function needs to be maximized or
(known as a training cycle) until the appropriate mar- minimized, the fitness function is hence defined in the
gins and bias has been established. Depending upon following ways:
the complexity of the task to be learned, many thous-
ands of training cycles may be needed for the SVR to
F(x) = f(x) for a maximization problem
correctly identify the training set. Once the output is F(x)=1/(1+f(x)) for a minimization problem
correct, the margins (w) and bias (b) can be used with
The fitness function value for a particular coded
the same SVM on unseen data to examine how well it
string is known as the string’s fitness. This fitness va-
performs. SVM learning is considered successful only
lue is used to decide whether a particular string car-
if the system can perform well on test data on which
ries to the next generation or not.
the system has not been trained. This capability of a
GA operation begins with a population of ran-
SVM is called generalizability.
dom strings. These strings are selected from a given
range of the function and represent the design or de-

179
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

cision variables. To implement our optimization rou- Crossover

tine, three operations are carried out: After the selection operator has been implemen-
– reproduction; ted, there is a need to introduce some amount of ran-
– crossover; domness into the population in order to avoid getting
– mutation. trapped in local searches. To achieve this, a cross
Operators in GA over operation was performed. In the crossover ope-
ration, new strings are formed by the exchange of the
Reproduction
information among strings of the mating pool. For
The reproduction operator is also called the se- example:
lection operator. This is because it is this operator
that decides the strings to be selected for the next 00|000 → 00|111
generation. The end results of this operation are the 11|111 11|000 ← Crossover point
formation of a “mating pool”, where the above avera-
Parents Children
ge strings are copied in a probabilistic manner. The
rule can be represented as: The strings are chosen at random and a random
crossover point is decided, and crossover is per-
(Probability of selection into mating pool) ∝ (Fitness
formed in the method shown above. It is evident that,
of string)
using this method, better strings or worse strings may
The probability of the selection of the ith string be formed. If worse strings are formed, then they will
into the mating pool is given by: not survive for long, since the reproduction will elimi-
nate them. But what if the majority of the new strings
Fi
pi = n formed are worse? This undermines the purpose of
F
i =1
j reproduction. To avoid this situation, we do not select
all the strings in population for crossover. We intro-
duce a crossover probability (pc). Therefore, (100pc),
where Fi is the fitness of the ith string. Fj is the fitness
%, of the strings are used in crossover. (1-pc), %, of
of the jth string and n is the population size.
the strings is not used in crossover. Through this, we
The average fitness of all the strings is calcul-
have ensured that some of the good strings from the
ated by summing the fitness of individual strings and
mating pool remain unchanged. The procedure can
dividing by the population size, and is represented by
be summarized as follows:
the symbol F :
Step 1: select (100pc) of the strings out of the
n

F
mating pool at random.
i Step 2: select pairs of strings at random.
i =1
F = Step 3: decide a crossover point in each pair of
n
strings (again this done by a random number gene-
It is obvious that the string with the maximum ration over the length of the string and the appropriate
fitness will have the most number of copies in the position is decided according to the value of the ran-
mating pool. This is implemented using “roulette whe- dom number).
el selection”. The algorithm of this procedure is as Step 4: perform the crossover on the pairs of the
follows: strings by exchanging the appropriate bits.
Step 1: using Fi, calculate pi.
Mutation
Step 2: calculate the cumulative probability Pi.
Step 3: generate n random numbers (between 0 Mutation involves making changes in the popu-
and 1). lation members directly, that is, by flipping randomly
Step 4: copy the string that represents the selected bits in certain strings. The aim of the mu-
chosen random number in the cumulative probability tation is to change the population members by a small
range into the mating pool. A string with higher fitness amount to promote local searches when the optimum
will have a larger range in cumulative probability and is nearby. Mutation is performed by deciding a muta-
so has more probability of getting copied into the mat- tion probability, pm, and selecting strings, on which
ing pool. mutation is performed, at random. The procedure can
At the end of this implementation, all the strings be summarized as follows:
that are fit enough would have been copied into the Step 1: calculate the approximate number of
mating pool and this marks the end of the repro- mutations to be performed by:
duction operation.

180
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

Approximate number of mutations = n/pm dely as possible while at the same time trying to lo-
Step 2: generate random numbers to decide cate promising regions for further exploration.
whether mutation is to be performed on a particular Principal features possessed by the GAs are: i)
population member or not. This is decided by a “coin they are zeroth order optimization methods requiring
toss”. If the generated random number of a particular only the scalar values of the objective function, ii)
population member is lower than the prespecified capability to handle nonlinear, complex and noisy ob-
number (again chosen randomly) then that represent jective functions, iii) they perform global search and
the “true” case. If it is not lower than the pre specified thus are more likely to arrive at or near the global
number, then that will be considered as “false” case. optimum and iv) their search procedure being sto-
If the outcome is true, perform mutation, if false, do chastic, GAs do not impose pre-conditions, such as
not. smoothness, differentiability and continuity on the ob-
Step 3: if the outcome found in step 2 is true for jective function form.
a particular population member, then generate ano- The optimization objective underlying the GA-
ther random number to decide the mutation point over based optimization of an SVR model is defined as:
the length of the string. Once decided, flip the bit cor- Find the N-dimensional optimal decision variable
responding to the mutation point. vector, x* = [x1*,x2*,…,xN*]T representing optimal pro-
At the end of mutation, the strings obtained re- cess conditions so that it maximizes process outputs,
present the next generation. The same operations are yk ; k = 1,2,…,K. The corresponding single objective
carried out on this generation until the optimum value function fˆ to be maximized by the GA is defined in
is encountered. Eq. (1). In the GA procedure, the search for an op-
timal solution (decision) vector, x*, begins from a ran-
GA-based optimization of SVR models domly initialized population of probable (candidate)
solutions. The solutions, usually coded in the form of
Having developed an SVR-based process mo- binary strings (chromosomes), are then tested to
del, a GA algorithm is used to optimize the N- dimen- measure their fitness in fulfilling the optimization ob-
sional input space of the SVR model. Conventionally, jective. Subsequently, a main loop comprising follow-
various deterministic gradient-based methods are ing operations is performed: i) selection of better (fit-
used for performing optimization of the phenomeno- ter) parent chromosomes, ii) production of an off-
logical models. Most of these methods require that spring solution population by crossing over the gene-
the objective function should simultaneously satisfy tic material between pairs of the fitter parent chromo-
the smoothness, continuity, and differentiability crite- somes and iii) mutation (bit-flipping) of the offspring
ria. Although the nonlinear relationships approxima- strings. The implementation of this loop generates a
ted by an SVR model can be expressed in the form of new population of candidate solutions, which when
generic closed-form expressions, the objective func- compared to the previous population, usually fares
tion(s) cannot be guaranteed to satisfy the smooth- better at fulfilling the optimization objective. The best
ness criteria. Thus, the gradient-based methods can- string that evolves after repeating the above descri-
not be efficiently used for optimizing the input space bed loop till convergence, forms the solution to the
of an SVR model and, therefore, it becomes neces- optimization problem (refer Appendix 1). The step-
sary to explore alternative optimization formalisms, wise procedure for the GA-based optimization of an
which are lenient towards the form of the objective SVR model is provided in a flowchart in Figure 2.
function.
Genetic algorithms combine the “survival of the CASE STUDY OF MONO-ETHYLENE GLYCOL
fittest” principle of natural evolution with the genetic PRODUCT UV TRANSMITTANCE
propagation of characteristics, to arrive at a robust
Figure 3 describes a brief process description of
search and optimization technique. The principal dif-
glycol section of mono-ethylene glycol plant (MEG)
ference between the widely used deterministic gra-
where glycol (90%) and water solution (10%) fed to
dient-based optimization schemes and the stochastic
the drying column to remove the water from the drying
ones such as GA is that the latter class of methodo-
column top. The bottom of the drying column fed to
logies involves a random component at some stage in
the MEG column to distil MEG from heavier glycols
their implementation. For instance, GAs manipulate a
(namely diethylene glycol and triethylene glycol).
set of candidate solutions at random with the ob-
MEG product (99.9 mass% purity) is withdrawn from
jective of sampling the search (solution) space as wi-
the MEG column below the top packing bed. An over-

181
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

Figure 2. Flowchart for GA based optimization of SVR model.

head vapor purge of up to 10% of the product is taken data. Hourly average of actual plant operating data at
out to purge the light compounds. Figure 3 shows the the steady state was collected for approximately one
location of input parameters from the drying column year. Data were checked and cleaned for obvious in-
and MEG column which were used to build the model accuracy and retain those data when plant operation
of UV. was in a steady state and smooth. Finally 6273 re-
cords are qualified for neural regression. This wide
Development of the SVR-based correlation
range of database includes plant operation data at
The development of the SVR-based correlation various capacities starting from 75% capacity to
had been started with the collection of a large data- 110% of design capacity.
bank. The next step was to perform a neural regres-
sion, and to validate it statistically. Identification of input and output parameters

Collection of data Based on the operating experience in a glycol

plant, all physical parameters that influence UV are
The quality and quantity of data is very crucial in
put in a so-called “wish-list”. Out of the number of in-
SVR modeling as learning is primarily based on these
puts in a “wish list”, several sets of inputs were made

182
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

Figure 3. Process flow diagram of drying and MEG column.

and tested via rigorous trial-and-error on the SVR. Fi- luated to design a successful regression model. These
nally, the nine input variables (Table 2) have been fi- five parameters are 1) kernel type, 2) type of loss
nalized to predict UV. function, 3) kernel parameter, i.e. degree of polyno-
Support vector regression mial etc., 4) C (represents the trade-off between the
model-complexity and the approximation error) and 5)
For modeling purposes, the column operating
ε (signifies the width of the ε-insensitive zone used to
conditions data (see table 2) can be viewed as an
fit the training data).
example input matrix (X) of size (6273×9), and the
Since the prior knowledge is not there regarding
corresponding UV data as the example output matrix
the suitability of a particular value of any of the above
(Y) of size (6273×1). For SVR training, each row of X
five parameters, the strategy adopted here is holistic
represents a nine-dimensional input vector x =
and summarized in Figure 4. The SVR performance
= [x1,x2,…,x9]T, and the corresponding row of matrix Y
was evaluated exhaustively for all combinations of
denotes the one-dimensional desired (target) output
above parameters. All the kernel types available in
vector y = [y1]T. As the magnitude of inputs and out-
literature (namely linear, polynomial, Gaussian radial
puts greatly differ from each other, they are normali-
basis function, exponential radial basis function, spli-
zed in -1 to +1 scale. 80% of total dataset was chosen
nes, B-splines) is tested with all combinations of the
randomly for training and the rest 20% was selected
loss function (namely ε – insensitive loss function,
for validation and testing.
quadratic loss function). The degree of kernel was
There are five different parameters to be eva-
varied from 1 to 6, capacity control varied from 10000

Table 2. Input and output variable for SVR model

Input Variables Output variables

Reflux ratio (Product flow / Reflux flow) Mono-ethylene glycol UV
Reflux flow (Mt/h)
MEG Column top pressure (mmHg)
MEG Column condenser pressure (Barg)
MEG column control temperature (°C)
MEG column feed flow (Mt/h)
Drying column control temperature (°C)
Drying column bottom temperature (°C)
Crude glycol reprocessing flow

183
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

Figure 4. Flow chart of SVR algorithm implementation.

to 0.1 (typically six values: 10000, 1000, 100, 10, 1, 1. The average absolute relative error (AARE)
0.1) and epsilon varies from 0 to 25 (typically six on a test set should be minimum:
values: 0, 0.1, 1, 10, 25). Each run was exposed with N
y predicted − y experimental
N
the same training and testing data and AARE and 1
AARE = |( )|
sigma was calculated for each run. 1
y experimental
The statistical analysis of SVR prediction is ba-
2. The standard deviation of error( σ) should be
sed on the following performance criteria:
minimum:

184
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

 N − 1(| (y
1
σ = predicted (i ) − y experimental (i ))/y experimental (i ) | − AARE )2
1

RESULTS AND DISCUSSIONS ponding to both the outputs of training and test sets.
To validate the reliability of the model, the actual plant
SVR model developments for UV soft sensor data were taken from DCS at different plant load at
While the training set was utilized for the ite- different point of time and actual lab measured UV
rative updation of the SVR parameters (b and w), the was compared with the model predicted UV.
test set was used for simultaneously monitoring the Figure 5 describes a comparison of the outputs
generalization ability of the SVR model. For develop- as predicted by the SVR model and their target va-
ing an optimal SVR model, its five-structural parame- lues. Considering the fact that all the input output data
ter described above was varied systematically. For are from a real plant with their inherent noise, the very
choosing an overall optimal model, the criterion used low prediction error can be considered as an excellent
was the least AARE for the test set. The optimal mo- SVR model. Once developed, this SVR model can be
del that satisfied this criterion has Exponential radial used to quantitatively predict the effects of all input
basis function(erb f), ε insensitive loss function, width parameters on the MEG product UV transmittance.
of erb f = 2,C = 1000, and ε = 0.1.The average error GA-based optimization of the SVR model
(AARE) for training and test set is calculated as 0.04
and 0.042% and corresponding cross correlation co- After development of a successful SVR model of
efficient (R) calculated as 0.84 and 0.83, respectively. glycol column, the next step is to find out the best set
The low and comparable training and test error AARE of operating conditions which lead to maximum UV.
values indicate good prediction and generalization GA based hybrid model was run and optimum para-
ability of the trained SVR model. Good prediction and meters were evaluated (within their permissible ope-
generalization performance of the model is also evi- rating limit). Figure 6 describes the actual versus the
dent from the high and comparable R values corres- optimum UV. From Figure 6 it is clear that by making

Figure 5. Actual vs. predicted UV.

Figure 6. Actual vs. optimum UV.

185
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

a small change in the nine input parameters, the action mechanism, kinetics etc.) is not required. Ef-
1 to 2% rise in UV can be made. Refer to Table 3 for fective results indicate that SVR modeling method
the optimum value of input variables calculated by GA provides a new tool for soft sensing modeling and has
algorithm. Drying column control temperature was a promising application in the industrial process appli-
found to have a significant effect on MEG product UV. cations. Using SVR-GA strategy, a number of sets of
This temperature will help to strip out UV deteriorating optimized operating conditions leading to the maxi-
compounds from the drying column itself before they mized product UV was obtained. The optimized solu-
enter MEG column. Three cases were run with three tions, when verified in an actual plant, resulted in a si-
different limits of this temperature. The optimum value gnificant improvement in the product UV.
is shown in Table 3.The program was made online
where it gives the operator what should be the nine REFERENCES
input parameters at different time to maximize the UV
[1] V. Vapnik, The Nature of Statistical Learning Theory,
in real time basis. After verifying all the calculations,
Springer Verlag, New York, 1995
the optimum input parameters were maintained in an
[2] V. Vapnik, Statistical Learning Theory, John Wiley, New
actual plant and benefit found was exactly the same York, 1998
as calculated. This ensures the validation and accu- [3] V. Vapnik, A. Smola, ,S. Golowich, Adv. Neural Inform.
racy of this calculation. Proc. Syst. 9 (1996) 281-287

Table 3. Optimum value of input variables calculated by GA algorithm

Case 1 Case 2 Case 3

Parameters Min Optimum Max Min Optimum Max Min Optimum Max
value value value value value value value value value
Input Reflux ratio (product flow/reflux flow) 0.70 0.78 0.78 0.70 0.78 0.78 0.77 0.77 0.78
Reflux flow (Mt/h) 110.0 114.52 115.0 110 110.00 115 112 112 115
MEG column top pressure (mmHg) 92.0 94.75 97.0 92 92 97 92 92 97
MEG column condenser pressure (Barg) 1.60 1.66 1.67 1.60 1.60 1.68 1.67 1.67 1.70
MEG column control temperature (°C) 168.80 169.0 169.0 168.80 168.8 169.0 168.80 168.80 169.0
MEG column feed flow (Mt/h) 98.0 98.0 98.0 100 100 101 99 99 101
Drying column control temperature (°C) 85.0 90.92 100.0 82 82 90 90 90 95
Drying column bottom temperature (°C) 165.0 166.0 166.0 165 166 166 165 165 166
Crude glycol reprocessing flow (Output) 0.0 0.0 8.0 0.0 0.05 8.0 0.0 0.0 8.0
Output Optimized UV – 97.60 – – 95.0 – – 95.40 –
Actual UV – 96.0 – – 93.41 – – 94.78 –

CONCLUSION [4] W. Polifke, W. Geng, K. Dobbeling, Combus. Flame 113

(1988) 119-120
This paper introduces SVR into soft sensing [5] H.M. Cartwright, R. A. Long, Ind. Eng. Chem. Res. 32
modeling and proposes a new soft sensing modeling (1993) 2706-2713
method based on SVR. In the strategy, a soft sensor [6] A. Garrard, E. S. Fraga, Comput. Chem. Eng. 22 (1998)
model is developed using SVR method followed by 1837-1850
[7] D.E. Goldberg, Genetic Algorithms in Search Optimiza-
the input space of that model being optimized using
tion and Machine Learning, Addison–Wesley, New York,
GAs so that the process performance is maximized.
1989
The major advantage of the SVR-GA strategy is that [8] V. Hanagandi, H. Ploehn, M. Nikolaou, Chem. Eng. Sci.
modeling and optimization can be conducted exclusi- 51 (1996) 1071-1078.
vely from the historic process data wherein the de- [9] V. Cherkassky, Y. Ma, Neural Networks 17(1) (2004)
tailed knowledge of the process phenomenology (re- 113-126.

186
S.K. LAHIRI and N.M. KHALFE: SOFT SENSOR DEVELOPMENT AND OPTIMIZATION… CI&CEQ 15 (3) 175−187 (2009)

APPENDIX 1 -Wheel (RW) method or its more stable variant known

as the stochastic remainder selection (SRS).
Steps performed in GA Step 4 (crossover). Randomly select Npop/2
Step 1 (initialization). Set generation index (Ngen) number of parent pairs from the mating pool and
to zero and generate a population of Npop binary strings perform a crossover operation on each pair with the
(chromosomes) randomly; each string consisting of probability equal to pcross (0 < pcross ≤ 1). In cros-
lchr bits is divided into N segments equal to the num- sover, parent strings are cut at the same randomly
ber of decision (input) variables to be optimized. chosencrossover point to obtain two substrings per
Step 2 (fitness computation). Decode jth binary- parent. The substrings are then mutually exchanged
-coded chromosome (j = 1,2,…,Npop) to obtain its equi- between the parents and combined to form two off
valent decimal-valued solution vector (xj) using: spring chromosomes. This crossover operation is per-
N
formed on all the parent pairs in the mating pool to
(x nU − x nL ) × S n
x j ,n = x nL + ln
2 −1
; l
n =1
n = l chr [x1*,x2*,…,xN*]T obtain an offspring population of the size of the mat-
ing pool.
Step 5 (mutation). Flip (mutate) the bits of the
where x nL and x nU refer to the lower and upper offspring strings where the probability of a bit getting
bounds on xn, respectively; ln is the length of nth binary flipped (zero to one or vice versa) is equal to pmut. The
segment and Sn denotes the decimal equivalent of the recommended range of pmut is [0.01-0.05].
nth binary segment. Next, depending upon the model Step 6. Increment the generation index:
to be optimized, vector xj is used to compute the
output of an SVR model; this output is subsequently Ngen = Ngen + 1
used to calculate the fitness value (jξˆ) of the jth can- Step 7. Repeat Steps 2-6 on the newly gene-
didate solution. Upon computing the fitness scores of rated offspring strings until convergence is achieved.
Npop candidate solutions in the current population, the The criteria for the convergence are: Ngen exceeds its
solutions are ranked in the decreasing order of their maximum limit (Nmaxgen), or the fitness score of the
fitness scores. best (fittest) string in the offspring population under-
Step 3 (parent selection). From the current po- goes a very small or no change over successive ge-
pulation, choose Npop number of parent chromosomes nerations. After convergence, the string possessing
to form the mating pool. The members of this pool, the highest fitness value is decoded to obtain the op-
which are used to produce offspring population, pos- timized decision variable vector, x*.
sess relatively high fitness scores. The commonly
used parent selection techniques are the Roulette-

187