604 Y. Liu et al.
/ Chemical Engineering Science 102 (2013) 602–612
S1 ¼ fX1 ; Y 1 g, where X1 ¼ fx1;i ¼ ½x1;i;1 ; x1;i;2 ; ⋯; x1;i;m1 T A Rm1 gi ¼ 1;⋯;N
with m1 input variables and Y 1 ¼ fy1;i A Rgi ¼ 1;⋯;N are the
input and output variables of the first reactor with N samples,
respectively. Generally, the product quality Y l ¼ fyl;i A Rgi ¼ 1;⋯;N in
the lth reactor can be mainly controlled by the related
manipulated variables X l ¼ fxl;i A Rml gi ¼ 1;⋯;N in this reactor. Addi-
tionally, it can also be affected by some process variables
f½xT1;i ; ⋯; xTl1;i T A R∑j ¼ 1 mj gi ¼ 1;⋯;N in the previous reactors (Lou
l1
et al., 2012; Zhang et al., 2006). Without enough process
knowledge, the generalized input variables can be formulated
as xl ¼ ½xT1 ; ⋯; xTl1 ; xTl T A R∑j ¼ 1 mj with mj input variables of the
l
jth reactor. Correspondingly, the modeling data set of the
lth reactor can be further represented as Sl ¼ fXl ; Y l g; l ¼ 1; ⋯; L,
Xl ¼ fxl;i ¼ ½xT1;i ; ⋯; xTl1;i ; xTl;i T A R∑j ¼ 1 mj gi ¼ 1;⋯;N
l
where and
Y l ¼ fyl;i A Rgi ¼ 1;⋯;N .
2.2. Basic LSSVR modeling method
Generally, the soft sensor model development of an SR process
based on the SVR/LSSVR framework can be described as a problem
aiming to learn a mapping f: X-Y using a modeling set S¼[X, Y],
mainly using the kernel learning method. For the lth reactor, a
general nonlinear model for process modeling can be formulated
as (Liu et al., 2009, 2012)
yl;i ¼ f l ðwl ; bl ; xl;i Þ þ el;i ¼ wTl ϕðxl;i Þ þ bl þ el;i ; i ¼ 1; ⋯; N ð1Þ
where fl is the wanted model of the lth reactor; ϕ is a feature map;
yl,i and el,i denote the output measurement and the approximation
error for the ith sample of the lth reactor; and xl,i is a general input
vector usually composed of several measured variables. The
symbols wl and bl are the model parameter vector and the bias
term of the lth reactor, respectively. When the philosophy of the
statistical learning theory and the LSSVR framework is applied to
Eq. (1) (Schölkopf and Smola, 2002; Suykens et al., 2002), the
following optimization problem is formulated:
( γ
minJ l ðwl ; bl ; el;N Þ ¼ 12jjwl jj2 þ 2l jjel;N jj2
ð2Þ
s:t: yl;i wl ϕðxl;i Þbl el;i ¼ 0; i ¼ 1; ⋯; N
T
where el;N ¼ ½el;1 ; el;2 ; ⋯; el;N T is the approximation error. The for-
mulation consists of equality constraints instead of inequality
constraints in the conventional SVR algorithm and takes into
account a squared error with the regularization term. Therefore,
this reformulation greatly simplifies a modeling problem and
makes it computationally efficient (Suykens et al., 2002). The
user-defined regularization parameter γl (γl 40) determines the
trade-off between the model's complexity and approximation
accuracy (Suykens et al., 2002). The Lagrangian can be constructed
to solve the optimization problem in Eq. (2) (Schölkopf and Smola,
2002), then the solution can be expressed as
" #
1l;N 1Tl;N Pl;N yl;N
αl ¼ Pl;N yl;N  T ð3Þ
1l;N yl;N 1l;N
1Tl;N Pl;N yl;N
bl ¼ ð4Þ
1Tl;N yl;N 1l;N
where αl ¼ ½αl;1 ; ⋯; αl;N T are Lagrange multipliers; yl;N ¼ ½yl;1 ; ⋯;
yl;N T ; Il;N A RNN is a unit matrix and 1l;N is a vector of ones;
Pl;N ¼ H1 l;N and Hl;N ¼ Kl;N þ Il;N =γ l with the kernel matrix Kl;N , with
its elements Κl;N ði; jÞ ¼ 〈ϕðxl;i Þ; ϕðxl;j Þ〉; 8 i; j ¼ 1; ⋯; N (Schölkopf
and Smola, 2002; Suykens et al., 2002).
Finally, for a new test input xl;t , the LSSVR model estimation of
the lth reactor, i.e., y^ l;t , can be obtained (Suykens et al., 2002)
y^ l;t ¼ f l ðwl ; bl ; xl;t Þ
N
¼ ∑ αl;i 〈ϕðxl;i Þ; ϕðxl;t Þ〉 þ bl
i¼1
¼ αTl kl;t þ bl ð5Þ
where kl;t ðiÞ ¼ 〈ϕðxl;i Þ; ϕðxl;t Þ〉; 8 i ¼ 1; ⋯; N is a kernel vector for
estimation of the test sample.
In summary, the development of a soft sensor model amounts
to solving a set of linear equations in the high dimensional feature
space introduced by the kernel transform. The LSSVR model for
the first reactor can be built in a simply way. However, as for the
Lth reactor, it is not a trivial task because the set of the variables
fxL;i A R∑j ¼ 1 mj gi ¼ 1;⋯;N , including the variables at the current reactor
L
L and the variables from the previous reactors, becomes more
complicated. Without enough process knowledge, it is difficult to
determine a suitable input variable for constructing a model.
2.3. Traditional two-step LSSVR-based soft sensors
Generally, in the construction of a data-driven model for a
chemical process with many input variables, two traditional
approaches have been investigated (Gustafsson, 2005; Kadlec
et al., 2009; Shi and Liu, 2005; Zamprogna et al., 2005; Zhao
et al., 2010). One can be considered as the variable or feature
selection method (Ojeda et al., 2008; Pan et al., 2012). In the
context of modeling for the lth reactor, variable selection aims at
finding a subset of variables Xsub;l ¼ fxsub;l;i gi ¼ 1;⋯;N that can result
in more accurate and compact predictors. The variable selection
method should capture the relevant information and filter out
those inputs that are irrelevant to the specific regression model
(Ojeda et al., 2008; Pan et al., 2012). Actually, there is no single
definition for the best subset as different algorithms will lead to
different subsets. The main disadvantage of variable selection
methods is the high computational cost because the search space
becomes bigger as the number of the variables in the lth reactor
increases to 2∑j ¼ 1 mj (Ojeda et al., 2008).
l
The other classical method can be considered as the variable or
feature extraction. Among them, LV-based methods in the form of
PCA and PLS have a dominating role in process modeling (Cao
et al., 2003; Gustafsson, 2005; Kadlec et al., 2009; Shi and Liu,
2005; Zamprogna et al., 2005; Zhao et al., 2010). In these methods,
correlation patterns among the variables can be modeled and
underlying features are used to build and explain a quantitative
relationship relevant to the quality properties. However, LV-based
analysis approaches cannot tell which type of LVs best fits the
model for the observed data, since there are different ways to
define LVs (Gustafsson, 2005; Zhao et al., 2010). Besides, for a
nonlinear modeling problem, most of the previous investigations
often used PCA or PLS methods to extract LVs as a preprocessing
step (Kadlec et al., 2009; Shi and Liu, 2005; Zamprogna et al.,
2005). However, as for an SR process, the extracted important LVs
XLVs;l ¼ fxLVs;l;i gi ¼ 1;⋯;N may capture most of the process variations,
but not necessarily explain quality properties. Furthermore, this
two-step strategy divides the soft sensor modeling task into two
separate sub-tasks, regardless its connection in a unified
framework.
Similar to the above approaches, the LSSVR-based modeling
method, which combines corresponding soft sensors with the
above variable or feature selection methods, is shown in Fig. 2,
where Xsub;l ¼ fxsub;l;i gi ¼ 1;⋯;N and XLVs;l ¼ fxLVs;l;i gi ¼ 1;⋯;N are the
transformed variables and extracted feature variables, respectively.
Correspondingly, two kinds of LSSVR-based soft sensors can be
constructed, respectively. However, both of them can be consid-
ered as indirect two-step modeling approaches. Additionally, the
sequential relationship in the whole SR process has not been
investigated before because it was difficult to characterize.
 Y. Liu et al. / Chemical Engineering Science 102 (2013) 602–612 605

Test sample xl ,t

Input and output data Basic soft sensor LSSVR model for the
for training related reactor Prediction yˆl ,t
Sl = {Xl , Yl } , l = 1,L , L (Eq. 5)
Test sample xl ,t

Traditional two-step
Yl and a subset of LSSVR model for the
soft sensor A
variables or features related reactor Prediction yˆl ,t
Xsub, l (Eq. 5)

Test sample xl ,t
Traditional two-step
Yl and extracted LVs LSSVR model for the
soft sensor B
based on PCA related reactor Prediction yˆl ,t
XLVs, l (Eq. 5)

Fig. 2. Flow diagram of the traditional LSSVR-based two-step soft sensor modeling methods for the lth reactor of a sequential process.

3. Just-in-time sequential LSSVR model for an SRMP process transformed to zl;t first using Eqs. (7) and (8).

3.1. Sequential LSSVR modeling framework and its efficient training zl;t ¼ ½y1;t ; ⋯; yl1;t ; xTl;t T A Rl1þml ð9Þ
for an SR process Finally, the prediction y^ l;t can be obtained using the obtained
SLSSVR model f~ l
This paper presents a novel soft sensor modeling method
which can integrate the variable extraction and quality prediction y^ l;t ¼ f~ l ðωl ; β l ; zl;t Þ
into a unified framework. Additionally, the sequential relationship N
in an SR process is investigated. For the lth reactor of an SR ¼ ∑ α~ l;i 〈ϕðzl;i Þ; ϕðzl;t Þ〉 þ β l
i¼1
process, the soft sensor modeling framework can be described as
below ~ Tl k~ l;t þ β l
¼α ð10Þ
(
minJ l ðωl ; β l ; ξl;N Þ ¼ 12jjjωl j2 þ γ2l jjξl;N jj2 where k~ l;t ðiÞ ¼ 〈ϕðzl;i Þ; ϕðzl;t Þ〉; 8 i ¼ 1; ⋯; N is a kernel vector and
ð6Þ α~ l ¼ ½α~ l;1 ; ⋯; α~ l;N T are Lagrange multipliers of the SLSSVR model.
s:t: yl;i ωTl ϕðzl;i Þβ l ξl;i ¼ 0; i ¼ 1; ⋯; N

 Generally, the development of a good LSSVR model depends on
where ωl ; β l are the model parameters of the final SLSSVR soft suitable selection of related model parameters. For an LSSVR model
sensor model f~ l ; and ξl;N ¼ ½ξl;1 ; ξl;2 ; ⋯; ξl;N T is the approximation with the common Gaussian kernel Kðxi ; xj Þ ¼ exp½jjxi xj jj=s,
error. The new input variable zl;i can be formulated as there are two parameters: the regularization parameter γ 40 and
the related kernel parameter s40 to be chosen. There exist some
zl;i ¼ ½y1;i ; ⋯; yl1;i ; xTl;i T A Rl1þml ; i ¼ 1; ⋯; N ð7Þ
efficient model selection criteria, e.g., cross-validation and Bayesian
where inference (Cawley and Talbot, 2004; Schölkopf and Smola, 2002;
8 Suykens et al., 2002). Here, the parameters selection procedure is
>
>
> y1;i ¼ g 1 ðw1 ; b1 ; x1;i Þ ¼ wT1 ϕðx1;i Þ þ b1 ; i ¼ 1; ⋯; N mainly based on the fast leave-one-out cross-validation (FLOO-CV)
>
>
< y ¼ g ðw ; b ; x Þ ¼ wT ϕðx Þ þ b ; i ¼ 1; ⋯; N criterion (Cawley and Talbot, 2004, 2007).
2;i 2 2 2 2;i 2 2;i 2
> The traditional leave-one-out cross-validation (LOO-CV)
>⋯
>
>
> method has been shown to give an approximately unbiased
: yl1;i ¼ g l1 ðwl1 ; bl1 ; xl1;i Þ ¼ wTl1 ϕðxl1;i Þ þ bl1 ; i ¼ 1; ⋯; N
estimation of the generalization properties of the statistical
ð8Þ models. It can provide a sensible criterion for model selection.
where fwj ; bj gj ¼ 1;⋯;l1 are the model parameters of transform Especially for the conventional SVR method, bounds on the
models fg j gj ¼ 1;⋯;l1 , respectively. For the jth ðj ¼ 1; ⋯; l1Þ reactor, LOO-CV estimator have been proved to be an effective criterion
the LSSVR model can be trained to obtain the nonlinear transform for model selection (Cawley and Talbot, 2004, 2007). However,
of xj;i to the “virtual” quality variables yj;i ; i ¼ 1; ⋯; N; j ¼ 1; ⋯; conventional LOO-CV for SVR/LSSVR is computationally expensive,
l1. Then the new input variable zl;i ¼ ½y1;i ; ⋯; yl1;i ; xTl;i T A Rl1þml so it is not suitable for online performing. Therefore, applying the
and the corresponding new training set Zl ¼ fzl;i gi ¼ 1;⋯;N are con- FLOO-CV criterion (Cawley and Talbot, 2004) as the foundation of
structed. Note that this variable zl;i is different from the previous the parameter selection is feasible.
xl;i in Eq. (1). Based on this transform, the dimension of the new For the lth reactor, the total FLOO-CV error of the LSSVR model
extracted variable zl;i is generally much less than the previous one EFLOOCV
l;N with a candidate parameter (γ, s) and N training samples
of xl;i for many SR processes because l1 þ ml o o ∑lj ¼ 1 mj , can be obtained (Cawley and Talbot, 2004)
 
especially when l is large. Moreover, the process information N N 
 αl;i 

contained in variables of the previous reactors can be represented EFLOOCV
l;N ¼ ∑ jjξ FLOOCV
l;i jj ¼ ∑   ð11Þ
i ¼ 1 P l;N;ii þ sl;i =ol
2
i¼1
by the corresponding LSSVR transform models fg j gj ¼ 1;⋯;l1 . Con-
sequently, zl;i can be considered as a new transformed input where P l;N;ii is the item at the ith row and ith column of Pl,N,
variable, including the extracted sequential relationship of the sl ¼ Pl;N 1l;N ¼ ½sl;1 ; ⋯; sl;N T and ol ¼ 1Tl;N Pl;N 1l;N . The related terms
previous reactors and information in the current reactor. (i.e., Pl,N and αl,i) are available. Additionally, the computational
After training, the aforementioned LSSVR models of the lth load of sl and ol is small. Consequently, compared to the conven-
FLOOCV
reactor can be obtained. Then, the new input xl;t can be tional LOO-CV method, the computation of El;N is much more
606 Y. Liu et al. / Chemical Engineering Science 102 (2013) 602–612
Fig. 3. Schematic diagram of the sequential LSSVR modeling method for an SR process.
efficient, which allows further selection of parameters of a
sequential of LSSVR models.
The proposed sequential modeling method is shown in Fig. 3.
For the lth (l4 1) reactor, since the transform of the first (l  1)
reactors yj;i ; i ¼ 1; ⋯; N; j ¼ 1; ⋯; l1 can be mainly determined
by the corresponding LSSVR models fg j gj ¼ 1;⋯;l1 , the prediction
model f~ l and the first (l  1) transformed LSSVR models
fg j gj ¼ 1;⋯;l1 are trained simultaneously. For online prediction of a
query samplexl;t , the transformed variable zl;t ¼ ½y1;t ; ⋯; yl1;t ;
xTl;t T A Rl1þml can describe the input variables in a sequential
way with an extracted form. Furthermore, this new transformed
variable contains quality information of the first (l  1) reactors.
Consequently, the online prediction with the model f~ l can be
achieved in a straightforward way.
As shown in Fig. 3, a forward FLOO-CV-based optimization
method for solving the sequential modeling framework is pro-
posed. As for the lth reactor, the modeling and prediction can be
implemented in a unified way instead of implementing on the
reactors one by one. First, the candidate parameter set can be
defined as Θl;K ¼ fθl;k ¼ ½γ 1;k ; s1;k ; ⋯; γ l1;k ; sl1;k ; γ l;k ; sl;k T gk ¼ 1;⋯;K
with K groups of parameters. Among the set Θl;K , ðγ 1;k ; s1;k ; ⋯;
γ l1;k ; sl1;k Þk ¼ 1;⋯;K are for the first (l  1) LSSVR models
fg j gj ¼ 1;⋯;l1 and ðγ l;k ; sl;k Þk ¼ 1;⋯;K are for the model f~ l , respectively.
Then, based on the FLOO-CV criterion, the proposed SLSSVR model
can be efficiently trained in a forward manner. The final SLSSVR
model can be obtained as follows:
8
< minEFLOOCV
l;N
: for Θl;K ¼ fθl;k ¼ ½γ 1;k ; s1;k ; ⋯; γ l1;k ; sl1;k ; γ l;k ; sl;k T gk ¼ 1;⋯;K
ð12Þ
On one hand, with the transformed LSSVR models fg j gj ¼ 1;⋯;l1 ,
the “virtual” quality variables are connected with the input
variables in the previous reactors. They can provide the related
quality information for the final reactor. On the other hand, note
that the main purpose is to achieve good quality prediction for the
final reactor. As can be seen in Eq. (12), the “virtual” quality
variables, yj;i ; i ¼ 1; ⋯; N; j ¼ 1; ⋯; l1, of the first (l  1) reac-
tors can be adjusted among Θl;K . That is, they can also be
considered as “adjustable” input variables (but not fixed) in
zl;t ¼ ½y1;t ; ⋯; yl1;t ; xTl;t T to make the final model f~ l more suitable
for quality prediction. Consequently, the “virtual” quality variables
provide a sequential link from the first reactor to the last one in an
SR process.
Additionally, it can be noted that, for N training data with the
candidate parameter set Θl;K , the complexity of FLOO-CV can be
reduced to about O(KN3) operations, compared to the traditional
LOO-CV with about O(KN4) operations for training an LSSVR
model. Consequently, the computational load can be greatly
reduced for this sequential modeling problem. This allows the
 Y. Liu et al. / Chemical Engineering Science 102 (2013) 602–612
proposed sequential modeling framework to be implemented in
an efficient manner.
3.2. Just-in-time LSSVR modeling approach for
multi-grade processes
As the market is highly competitive, time-to-market, quality
and differentiation are the main competitive factors in the market.
Many different grades have to be produced in a polymer manu-
facturing process. In many cases, there will not be sufficient data to
train the model throughout the entire input space. The distribu-
tion of modeling data is generally uneven. Sufficient modeling
data can be obtained in some steady-state operating conditions,
but few data are available beyond these areas. Consequently, it is
difficult to describe the whole multi-grade (MG) process with
multiple grades using only a single model (Kaneko et al., 2011; Kim
et al., 2005; Liu, 2007; Yu, 2012b).
This problem may be overcome by fixing models a priori in the
given areas, and applying learning methods only where the
samples are available and reliable. Local models can provide the
advantages of learning efficiency and generalization. However, it is
difficult to automatically find the suitable level of locality for a
given subspace of an arbitrary problem. Therefore, in this section,
a JITL modeling strategy combined with LSSVR is proposed for the
whole SRMG process. The JITL method, which is inspired by the
ideas from local modeling and database technology, has been
developed as an attractive alternative to nonlinear process model-
ing, monitoring and control (Atkeson et al., 1997; Bontempi et al.,
1999; Cheng and Chiu, 2004; Fujiwara et al., 2009; Ge and Song,
2010; Hu et al., 2013; Kano and Ogawa, 2010; Liu et al., 2007,
Fig. 4. Illustrations of the proposed JS-LSSVR and SLSSVR modeling methods and other traditional methods for modeling of an SRMG process (with sequential-reactor and
multi-grade characteristics).
607
2012). The just-in-time LSSVR (JLSSVR) model was first proposed
by Liu et al. (2007) for batch processes and then extended by Ge
and Song (2010) for continue processes. Generally, for a query
sample xt, there are three main steps to build a JLSSVR model (Liu
et al., 2007, 2012)
Step 1: Select the relevant samples to construct a similar set
Ssim in the database S based on some defined similarity criteria.
Step 2: Build a JLSSVR model fJLSSVR(xt) using the relevant
dataset Ssim.
Step 3: Predict the output y^ t online for the current query
sample xt and then discard the JLSSVR model fJLSSVR(xt).
With the same three-step procedure, a new JLSSVR model can
be built for the next query sample. Generally, the Euclidean
distance-based similarity is the most commonly utilized index
(Atkeson et al., 1997; Bontempi et al., 1999; Cheng and Chiu, 2004).
The similarity factor (SF) sti between the query sample xt and the
sample xi in the dataset is defined below (Atkeson et al., 1997;
Bontempi et al., 1999)
sti ¼ expðdti Þ ¼ expðjjxi xt jjÞ; i ¼ 1; ⋯; N ð13Þ
where dti is the distance similarity between xt and xi in X. The
value of sti is bounded between 0 and 1 and when sti approaches to
1, xt resembles xi closely. Although the distance-and-angle-based
SF can obtain better performance than only utilization of the
Euclidean distance (Cheng and Chiu, 2004; Liu et al., 2012), an
additional parameter for balance of the distance and the angle
should be determined. Additionally, some correlation-based simi-
larity criteria have been proposed recently (Fujiwara et al., 2009;
608 Y. Liu et al. / Chemical Engineering Science 102 (2013) 602–612
Kano and Ogawa, 2010). However, it is not the main scope in this
study, so it is not investigated here.
Based on the similarity criterion, the n (nmin rn rnmax) related
samples should be chosen for building a JLSSVR model. Generally,
two parameters nmin and nmax are chosen. Only the relevant data
sets formed by the nminth relevant data to the nmaxth relevant data
are used in the model (Cheng and Chiu, 2004). However, how to
determine the number of similar samples is not straightforward.
Cheng and Chiu (2004) adopted a linear search method with a
search step Δn to choose n. Nevertheless, the computational
burden may become large if the range of [nmin, nmax] is wide.
To solve the problem, an improved similarity criterion which
was recently proposed for batch processes is adopted (Liu et al.,
2012). First, the nmax most similar samples can be selected due to
the SF criterion in Eq. (13). Then, rank the nmax similar samples
according to the degree of similarity. Define a cumulative similar-
ity factor (CSF) Stn as below (Liu et al., 2012)
n yl1;t ; xTl;t T A Rl1þml (Eq. (9)) before online prediction.
∑ sti
i¼1
Stn ¼ nmax
; n r nmax
∑ sti
i¼1
which represents the cumulative similarity of n most similar
samples compared to the relevant dataset Ssim. The CSF index
can be utilized to access the cumulative similarity and then it can
determine the n most similar samples in a more reasonable way.
As an alternative, the search of [nmin, nmax] can be substituted by
the choice of Stn, e.g., Stn ¼0.9. This means 90% of the similar
samples have been selected. Especially for an MG process, the
relevant set of a query sample for different grades are generally
different. It is difficult to determine the range of [nmin, nmax]
beforehand. Therefore, compared to the method of searching n in
the range of [nmin, nmax] directly, the CSF index is more meaningful
and its computational burden can be reduced by simply selecting
Stn (Liu et al., 2012).
With a suitable choice of Stn, the similar samples n are
determined and the relevant dataset Ssim can be obtained. Then,
with a candidate parameter, the JLSSVR model can be constructed.
As aforementioned, the model parameter can also be selected
based on the FLOO-CV criterion (Cawley and Talbot, 2004). Finally,
if the SRMG process is in batch operations, the selection of similar
samples can be further implemented in a time-scale window due
to frequent repetition of batch runs. Additionally, the JITL method
can be combined with the moving-window strategy to save the
searching time of similar samples, especially for process monitor-
ing (Hu et al., 2013).
3.3. Implementation steps of just-in-time sequential LSSVR for an
SRMG process
As aforementioned, reliable online quality prediction of an
SRMG process often encounters different challenges, including
process nonlinearity, input variables selection/extraction, sequential
relationship in reactors, and shifting operating modes by multiple
grades. In this section, a just-in-time sequential LSSVR (JS-LSSVR)
Table 1
Comparisons of several soft sensors, including JS-LSSVR, SLSSVR, JLSSVR, PCA–LSSVR, and LSSVR, for an SRMG process (with sequential-reactor and multi-grade
characteristics).
Method Abbreviation
Just-in-time Sequential LSSVR (proposed) JS-LSSVR
Sequential LSSVR (proposed) SLSSVR
Just-in-time LSSVR (Liu et al., 2012) JLSSVR
Two-step PCA-based LSSVR (Cao et al., 2003) PCA–LSSVR
Basic LSSVR (Shi and Liu, 2006; Suykens et al., 2002) LSSVR
nonlinear modeling method is proposed for the whole SRMG
process. It can integrate sequential learning and local learning.
The step-by-step procedures of the JS-LSSVR approach for online
quality predictions of an SRMG process are summarized as below.
The flow diagram of the corresponding procedures of the JS-LSSVR
and SLSSVR methods is shown in Fig. 4.
Step 1: Collect the process input and output data, i.e.,
Sl ¼ fXl ; Y l g, for model training of the lth reactor.
Step 2: Train an SLSSVR model for the sequential relationship in
an SR process using an efficient FLOO-CV training strategy (Eqs.
(11) and (12)). The model data set, i.e., S~ l ¼ fZl ; Y l g, can be
obtained (Eqs. (6)–(9)), where the input variables in the
previous reactors are substituted with “virtual” quality vari-
ables ðZl Þ via LSSVR transform models.
Step 3: For a new input measurement in the test set, i.e., xl;t , it
should be transformed into a new input variable zl;t ¼ ½y1;t ; ⋯;
Step 4: Further search its similar set S~ l;sim ¼ fZl;sim ; Y l;sim g using
ð14Þ the similarity criterion (Eqs. (13) and (14)). A JS-LSSVR model
for zl;t is constructed and the prediction y^ l;t can be obtained (Eq.
(10)). Go to Step 3 and do the same procedure for the online
prediction of the next new input measurement.
In Steps 1–3, the SLSSVR modeling method can integrate the
variable extraction and quality prediction into a unified framework
for an SR process. The process variables (Xj ; j ¼ 1; 2; ⋯; l) are
transformed into the “virtual” quality variables Zj ; j ¼ 1; 2; ⋯; l
under the constraints of the sequential relationships. As shown
in Fig. 4, using the proposed transform strategy, the number of
input variables ðXj ; j ¼ 1; 2; ⋯; lÞ can be reduced from ∑lj ¼ 1 mj to
l1 þ ml . Step 4 is for the JS-LSSVR model. A similar set with n
rather than N samples (n o o N) is utilized to construct the model
to improve the local learning ability for the online prediction of a
query sample.
Additionally, the main properties of these modeling methods
for an SRMG process are listed in Table 1. For detailed compar-
isons, other modeling methods, including LSSVR, PCA–LSSVR, and
JLSSVR aforementioned, are also shown on the top of Fig. 4. As
shown in Fig. 4 and Table 1, none of LSSVR, PCA–LSSVR nor JLSSVR
methods can capture the sequential relationship in an SR process.
The main property of SLSSVR is that it can integrate the variable
extraction and quality prediction into a unified framework
through the “virtual” quality variables as the connected bridge.
This is the first contribution of this work and it is different from
the traditional LSSVR and two-step modeling methods. However,
as aforementioned, only SLSSVR is not enough for modeling of an
MG process although the sequential relationship is captured. Due
to the MG data, the distribution of “virtual” quality variables is
widely and non-uniformly located. On the other hand, as shown in
Table 1, compared with a single global model, the JLSSVR model
can offer a more accurate prediction for an MG process because of
its local learning ability, but the input variables and the sequential
relationship in an SR process are not explored. Naturally, the JS-
LSSVR method, which is the second main contribution of this
Input variables transformation Sequential relationship captured Suitable for multiple grades
Yes Yes Yes
Yes Yes No
No No Yes
Yes No No
No No No
 Y. Liu et al. / Chemical Engineering Science 102 (2013) 602–612 609

E-201

V-234
R-231 R-229 R-233

R-234

E-201

V-232

Fig. 5. A simplified flowchart of the industrial SRMG polymerization process in an industrial plant in Taiwan.

work, combines the advantages of both SLSSVR and JLSSVR. It has Table 2
Comparisons of prediction between MI2 and MI3 using different soft sensors
sequential, global–local, and quality-relevant properties, so it is
(The best results are bold and italics.).
more suitable for a process with both of sequential-reactor and
multi-grade characteristics. MI and its range Method Input variables description RMSE RE (%)
In this work, the LSSVR model is adopted as the basic nonlinear
model. In general, SVR, GPR, and other comparable nonlinear KL MI3 (0–0.3) JS-LSSVR 1+1+9¼ 11 0.020 21.1
SLSSVR 1+1+9¼ 11 0.026 29.1
models (Chen and Ren, 2009; Ge et al., 2011; Schölkopf and Smola,
JLSSVR 12+8+9 ¼29 0.033 44.0
2002; Tipping, 2001; Yu, 2012b; Zhang et al., 2010) can also be PCA–LSSVR 13 0.041 46.9
applied to this framework. As for these nonlinear methods, LSSVR 12+8+9 ¼29 0.041 50.0
efficient training of a related just-in-time nonlinear model (e.g., PLS 10 0.055 61.9
a just-in-time sequential GPR model) should be developed. Alter- MI2 (0–10) JS-LSSVR 1+8¼ 9 0.768 23.9
natively, the SVR model with time difference recently proposed by SLSSVR 1+8¼ 9 0.777 34.8
Kaneko and Funatsu (2011) can tackle the nonlinearity in process JLSSVR 12+8¼ 20 0.849 32.4
PCA–LSSVR 8 1.035 42.2
variables and the degradation of the model. These would be LSSVR 12+8¼ 20 0.943 37.8
interesting future studies when they are applied to modeling of PLS 8 1.069 45.3
SRMG processes.

PCA–LSSVR and LSSVR models are trained using the FLOO-CV

4. Online prediction of MI in an industrial process
In this section, the proposed JS-LSSVR and SLSSVR soft sensor
modeling methods are validated by predicting MI of an industrial
polyethylene production process in a plant in Taiwan. All the data
samples have been collected from daily process records and the
corresponding laboratory analysis. After simply preprocessing the
set with a simple 3-sigma criterion, about 300 samples collected
from a product line from July 2009 to June 2011 are investigated in
this study. Among them, the data from July 2009 to December
2010 are used for training and data in 2011 are for test. The data
are available on the web of http://wavenet.cycu.edu.tw/~cpse/
web/Resources.htm.
A simplified flowchart of the polyethylene production process
is shown in Fig. 5. There are three main reactors in this poly-
ethylene production process. Their product quality variables (melt
indices) for the sequential reactors are noted as MI1, MI2, MI3,
respectively. A total of 29 process variables correlated with the
product quality have been chosen. There are 12 variables in the 1st
reactor, 8 variables in the 2nd reactor, and 9 variables in the 3rd
reactor, respectively. Besides, there are 3 main product grades for
this production line. MI is sampled and analyzed in the lab once a
day, always in the morning. Owing to lack of online analyzers for
MI, the operating variables have to be manipulated until the assay
results are available. Consequently, off-grade products and materi-
als have been produced inevitably.
The JS-LSSVR and SLSSVR methods are applied to predict MI2
and MI3, compared with traditional JLSSVR, PCA–LSSVR, LSSVR,
and PLS methods. For PCA–LSSVR and PLS, the number of LVs for
each reactor is chosen by the common cumulative percentage of
variance criterion of 85% (Kadlec et al., 2009). Then, all SLSSVR,
criterion. To assess the prediction performance, two performance
indices, including root-mean-square error (RMSE) and relative
RMSE (RE), are both considered and defined respectively as
follows:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N
RMSEl ¼ ∑ ðy^ l;i yl;i Þ2 =N; l ¼ 2; 3 ð15Þ
i¼1
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
N
REl ¼ ∑ ½ðy^ l;i yl;i Þ=yl;i 2 =N; l ¼ 2; 3 ð16Þ
i¼1
Details about comparisons between MI2 and MI3 prediction
among JS-LSSVR, SLSSVR, JLSSVR, PCA–LSSVR, LSSVR and PLS
methods are tabulated in Table 2. First, the results show that
the conventional PLS method, represented as a multivariable
LV method, is insufficient for this case, especially for prediction
of MI3. Although some improved PLS approaches have been
utilized for MI prediction of polyethylene processes (Ahmed
et al., 2009), they may be unsuitable for modeling of complex
nonlinear SRMG problems. The basic LSSVR shows better predic-
tion performance than PLS mainly because of its nonlinear
modeling ability. PCA–LSSVR (Cao et al., 2003), represented as a
two-step modeling method, does not always perform better than
LSSVR because the extracted LVs may fail to explain quality
properties. Moreover, the sequential relationship in the SRMG
process has not been captured by PCA–LSSVR, LSSVR or PLS.
JLSSVR (Liu et al., 2012), as a local nonlinear modeling method,
can achieve higher prediction accuracy than PCA–LSSVR and
LSSVR approaches. As for prediction of MI2, JLSSVR is only a little
inferior to SLSSVR in terms of RMSE. Nevertheless, as for predic-
tion of MI3, JLSSVR shows much inferior to SLSSVR both in two