Computers and Chemical Engineering 71 (2014) 77–93

Contents lists available at ScienceDirect

Computers and Chemical Engineering

journal homepage: www.elsevier.com/locate/compchemeng

Adaptive soft sensor modeling framework based on just-in-time

learning and kernel partial least squares regression for nonlinear
multiphase batch processes
Huaiping Jin, Xiangguang Chen ∗ , Jianwen Yang, Lei Wu
Department of Chemical Engineering, Beijing Institute of Technology, Beijing 100081, People’s Republic of China

a r t i c l e i n f o a b s t r a c t

Article history: Batch processes are characterized by inherent nonlinearity, multiple phases and time-varying behavior
Received 20 December 2013 that pose great challenges for accurate state estimation. A multiphase just-in-time (MJIT) learning based
Received in revised form 3 April 2014 kernel partial least squares (KPLS) method is proposed for multiphase batch processes. Gaussian mix-
Accepted 19 July 2014
ture model is estimated to identify different operating phases where various JIT-KPLS frameworks are
Available online 28 July 2014
built. By applying Bayesian inference strategy, the query data is classified into a particular phase with
the maximal posterior probability, and thus the corresponding JIT-KPLS framework is chosen for online
Keywords:
prediction. To further improve the predictive accuracy of the MJIT-KPLS algorithm, a hybrid similarity
Adaptive soft sensor
Batch process
measure and an adaptive selection strategy are proposed for selecting local modeling samples. Moreover,
Kernel partial least squares maximal similarity replacement rule is proposed to update database. A procedure of input variable selec-
Just-in-time learning tion based on partial mutual information is also presented. The effectiveness of the MJIT-KPLS algorithm
Partial mutual information is demonstrated through application to industrial fed-batch chlortetracycline fermentation process.
Chlortetracycline fermentation process © 2014 Elsevier Ltd. All rights reserved.

1. Introduction
Batch or semibatch processes have been widely used to produce
special chemicals, materials for microelectronics, pharmaceutical
and agricultural products (Cinar et al., 2003). As prior require-
ments, the reliable real-time measurements play a crucial role in
process automation, monitoring and optimization (Alford, 2006).
However, the lack of reliable online sensors, which can accurately
detect the important state variables, has become one of the major
challenges of controlling batch processes accurately, automatically
and optimally (Chen et al., 2006; Dochain, 2008; Nicoletti and Jain,
2009).
Over the past two decades, soft sensors have received increas-
ing attention in both academia and industry due to their inferential
estimation capability. Although soft sensors have been applied in
broad fields, the online prediction remains the dominant applica-
tion (Kadlec et al., 2009). When online analyzers are not available,
soft sensing technology aims to provide estimations of difficult-
to-measure variables based on some easy-to-measure variables.
Generally, soft sensors can be classified into two groups: first-
∗ Corresponding author. Tel.: +86 13601333018; fax: +86 010 68914662.
E-mail addresses: jinhuaiping@gmail.com (H. Jin), xgc1@bit.edu.cn (X. Chen),
yangjianwen@bit.edu.cn (J. Yang), wulei@bit.edu.cn (L. Wu).
principle models and data-driven models (Kadlec et al., 2009). The
focus of this work is on the data-driven soft sensor modeling.
Reviews of this type of soft sensor application have been pub-
lished in references Fortuna et al. (2007), Haimi et al. (2013), Kadlec
et al. (2009), Kano and Nakagawa (2008), Kano and Fujiwara (2013),
Sliskovic et al. (2011).
The most common data-driven modeling techniques for devel-
oping soft sensors are multivariate statistical techniques such as
multivariate linear regression (MLR) (Kano and Ogawa, 2010), prin-
ciple component regression (PCR) (Jolliffe, 2002) and partial least
squares (PLS) (Lin et al., 2007; Kano and Fujiwara, 2013; Sharmin
et al., 2006). These linear modeling methods account for 90% of
the soft sensors used in industry (Kano and Ogawa, 2010) due to
their statistical background, ease of interpretability, and because
they can deal efficiently with data collinearity, which is common
among industrial datasets (Kadlec and Gabrys, 2011). Nevertheless,
these linear methods cannot function well when applied to highly
nonlinear processes like batch processes. Thus many efforts have
been attempted to nonlinear approaches, such as artificial neu-
ral networks (ANN) (Cui et al., 2012; Gonzaga et al., 2009), KPLS
(Jia et al., 2013; Yu, 2012a), support vector regression (Yu, 2012b),
neuro-fuzzy system (Jang, 1993; Jassar et al., 2009) and Gaussian
process regression (GPR) (Grbić et al., 2013). However, there exist
some issues remained to be solved to develop soft sensors for batch
processes.

http://dx.doi.org/10.1016/j.compchemeng.2014.07.014
0098-1354/© 2014 Elsevier Ltd. All rights reserved.
78 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93
One particular drawback of many current soft sensors is their
non-adaptive nature. Traditionally, the predictive models are not
adaptive, and once deployed into the real-life operation, the mod-
els will not change, whereas the operation environment is often
changing. To cope with changes in process characteristics, model
maintenance is essential to maintain high estimation performance
for a long time. Without process or expert knowledge, the soft
sensor model has to be updated automatically. Thus many kinds
of recursive modeling methods, which update models by priori-
tizing newer samples, have been proposed (Kadlec et al., 2011).
Although these methods can adapt the soft sensor to a new oper-
ation condition recursively, they cannot cope with abrupt changes
in process characteristics, which are caused by replacement of a
catalyst, cleaning of equipment, etc., because a query sample just
after an abrupt change becomes significantly different from the
prioritized samples. It is also a common practice to enhance the
model adaptation by the ensemble learning framework (Grbić et al.,
2013; Kadlec et al., 2011; Kadlec and Gabrys, 2011; Yu, 2012c).
In the ensemble framework, data are divided into different sub-
domains and local sub-models are constructed over each domain.
In this way, instead of using a single global model, multiple sim-
pler models are developed and then combined to obtain the final
prediction. Some other adaptive soft sensors were developed by
partitioning the process data into multiple clusters corresponding
to different operating phases where local predictive model is built
for each phase (Yu, 2012a; Yu and Qin, 2008, 2009). When imple-
mented online, the query sample, for which an output estimation
is required, is firstly classified into a particular phase, and then
the corresponding local model representing the identical phase is
adaptively chosen for prediction. Although such phase based multi-
model methods outperforms the single models, they are unable to
effectively capture the between-phase transient dynamics. Thus, a
Bayesian model averaging (BMA) based multi-model method was
proposed to tackle this issue (Yu et al., 2013). In practice, however,
it is difficult to determine the partition number due to the lack of
the quantitative and precise information of phase divisions. More
importantly, the local models used in the BMA based multi-model
methods are built offline and not updated once deployed online,
thus the changes in process characteristics cannot be well dealt
with due to the time-varying nature of the real-life processes.
Recently, just-in-time (JIT) learning has attracted increasing
attention in process modeling and soft sensor development (Cheng
and Chiu, 2004; Fujiwara et al., 2009). By applying JIT learning, a
local model is constructed from the samples similar to the query
sample. Thus, on one hand, JIT based model can cope with abrupt
changes as well as gradual ones. On the other hand, it can deal with
nonlinearity since it builds a local model repeatedly. Compared
to the traditional modeling methods which can be considered as
global modeling, JIT based method exhibits a local model structure
where a local model is built from the historical dataset selected
by some similarity measure to the query data when the estima-
tion is required. Once the estimated output is obtained, the built
local model is discarded. Nevertheless, the estimation accuracy of
JIT based models is expected to be further improved by selecting
the optimal combination of local regression function, input variable
selection, similarity measure, and database updating scheme, etc.,
simultaneously.
To enhance the predictive performance and adaptability of JIT
based models, there are some problems remain to be addressed.
Although linear local model based JIT methods (Cheng and Chiu,
2004; Kim et al., 2013) can successfully address the process
nonlinearity, high nonlinearity variable relationships are widely
existed in batch processes where a local linear model may not
always function well. Thus a nonlinear modeling technique with
high computational efficiency is preferable. In addition, the com-
monly used similarity measures rarely take into account process
characteristics. In real applications, the most frequently used
similarity measures are based on the Euclidean distance or the
Mahalanobis distance (Cheng and Chiu, 2004; Ge and Song, 2010;
Schaal et al., 2002), most of which are defined only from the per-
spective of sample algebraic space irrespective of specific process
knowledge. Moreover, the optimal local modeling size is usually
determined offline as reported in (Ge and Song, 2010). An adaptive
strategy is required to choose local modeling samples adaptively
for each query data. Besides, a reliable database updating scheme is
also crucial for JIT based models to tackle changes in process charac-
teristics. Usually, the database is updated only by simply removing
the oldest samples and adding the new samples (Shigemori et al.,
2011).
The second problem encountered in soft sensor modeling is
the lack of a systematic guideline for input variable selection. As
reported in literature work (Cui et al., 2012; Kadlec and Gabrys,
2011; Kim et al., 2013; Pani et al., 2013), input variables are often
selected based on engineers’ personal experience and prior pro-
cess knowledge. However, it is time-consuming for the engineers
to select the input variables since trial and error is inevitable. Addi-
tionally, the selected variables may not be optimal. Also, it becomes
very difficult even for experienced engineers to properly select
input variables when a large number of variables are measured and
physical and chemical phenomena are not sufficiently understood.
Consequently, various data-based methods have been proposed for
selecting proper input variables.
One popular approach for reducing the input variable dimen-
sion is by projecting the original input space into an adequate lower
dimensional space. The most popular methods for achieving such
projection task are based on linear projection of input space, such
as the widely known principle component analysis (PCA) and PLS.
However, the new variables resulting from such methods are diffi-
cult to interpret in terms of actual process variables (Delgado et al.,
2009). More importantly, the underlying assumption of linearly
structured dependence contradicts to the development of statis-
tical model of nonlinear processes.
Another approach for reducing input space dimension is based
on selecting the most important variables from all potential vari-
ables according to some criteria. As a nonparametric and nonlinear
measure of relevance derived from information theory, mutual
information (MI) has recently been applied to nonlinear processes
modeling (May et al., 2008a), process monitoring (Chen et al., 2013;
Rashid and Yu, 2012a, 2012b) and soft sensor design (Grbić et al.,
2013). Unlike those linear methods that only consider linear rela-
tionships between variables, MI is theoretically able to identify
relations of any type. It furthermore makes no assumption about
the distribution of the data. However, several issues have arisen
in the formulation of MI-based selection algorithms, which are:
the ability of handling the inter-dependencies between candidates
and the lack of an appropriate principle for determining when to
halt the selection procedure (Chow and Huang, 2005). To tackle
this issue, the partial mutual information (PMI) criterion has been
developed by considering the effect of the already selected input
variables when evaluating the relevance between one plausible
input variable and the output variable (May et al., 2008a, 2008b;
Sharma, 2000).
Apart from the model adaptation and input variable selection,
during data-driven soft sensor modeling, much attention is only
paid to the plant data, whereas the process characteristics are usu-
ally ignored. In practice, batch processes are often characterized by
their multiphase characteristics where multiple operating phases
are involved. Thus the accuracy and reliability of quality variable
prediction can heavily degrade as the operating phase and pro-
cess dynamics change. Multiphase modeling strategy with phase
identification has been reported more efficient than the conven-
tional single-model based methods (Yu, 2012a; Yu and Qin, 2008,
 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93 79

2009; Yu et al., 2013). These results indicate that multiphase nature
essentially needs to be considered when designing data-driven soft
sensors for batch processes.
To address the above-mentioned issues, a novel adaptive soft
sensor modeling framework is proposed for nonlinear multiphase
batch processes. This soft sensing algorithm is outlined as follows:
Due to the curse of dimensionality, it is not feasible to calculate
the nonlinear mapping of each original input and then conduct the
PLS regression in the feature space. To tackle this issue, by apply-
ing the so-called ‘kernel trick’, the inner product of two mapped
samples can be given by
Kij = K(xi , xj ) = ˚(xi )T ˚(xj ), (2)

(i) JIT learning framework is adopted due to its capability of where K(·, ·) denotes a nonlinear kernel function. Denote  as an
dealing with changes in process characteristics as well as (n × m) matrix whose ith row is the vector ˚(xi ) in a m-dimensional
T
nonlinearity. Within this learning framework, a new hybrid feature space F, we can see that  represents the (n × n) ker-
similarity measure is defined by integrating Euclidean distance nel Gram matrix K of the cross dot products between all mapped
n
based similarity with process phase similarity. Further, the input data points {˚(xi )}i=1 . Depending on the nonlinear trans-
optimal local modeling samples are adaptively determined by formation ˚(·), the feature space can be high-dimensional, even
online cross-validation optimization. Besides, a maximal simi- infinite dimensional when a kernel function is used. In this study,
larity replacement rule is proposed to update sample database. the following radial basis function (RBF) is selected as the kernel
(ii) KPLS is chosen as the local modeling technique for two reasons. function
One reason is that it is more effective to capture the nonlinear  
||xi − xj ||2
characteristics of batch processes than linear methods such as K(xi , xj ) = exp − , (3)
2 2
PLS regression. Another reason is that KPLS essentially requires
only linear algebra, making it as simple as a regular linear PLS where  stands for the width of RBF kernel.
regression. The KPLS algorithm is outlined as follows. Let A be the desired
(iii) Gaussian mixture model (GMM) is estimated to identify the number of latent variables (LV). Repeat for i = 1 to A
operating phases of batch process, and then various JIT-KPLS
modeling frameworks are constructed for different phases. Step 1. ti = Ki ui , ti ← ti /||ti ||.
(iv) The input variables are selected based on PMI criterion Step 2. ci = YTi ti .
between potential input variables and output variable by Step 3. ui = Yi ci , ui ← ui /||ui ||.
performing a stepwise selection procedure, which effectively Step 4. Ki+1 = (I − ti ti T )Ki (I − ti ti T ), where I is an n-dimensional
alleviates the effect of the redundancy between input variables. identity matrix.
Step 5. Yi+1 = Yi − ti ti T Yi .
The proposed MJIT-KPLS soft sensing framework allows per-
forming model adaptation in four aspects. First, a query sample After extracting the desired latent variables, the corresponding
can be automatically classified into a particular operating phase regression coefficient matrix is expressed as
based on the estimated GMM model through Bayesian inference −1 T
T
strategy. Second, in the JIT learning framework, a local KPLS model B =  U(TT KU) T Y (4)
is constructed online when the estimation for the query sample is
and to make prediction on training data we can write
required. Third, an adaptive strategy is proposed to select the opti-
−1 T
mal local modeling size adaptively for local KPLS modeling. Finally, Ŷ = B = KU(TT KU) T Y = TTT Y, (5)
the proposed framework can adapt to new process state by adding
T T −1
new samples into the database from which the samples for local where T may be expressed as T = R with R =  U(T KU) .
modeling are selected. For predictions made on testing points {xi }n+nt
i=n+1
the matrix of
The rest of this paper proceeds as follows. Section 2 briefly out- regression coefficients B have to be used, i.e.
lines the theories about KPLS regression, JIT learning, PMI criterion, −1 T
and GMM algorithm. The proposed adaptive soft sensing algorithm, Ŷt = t B = Kt U(TT KU) T Y, (6)
MJIT-KPLS, is discussed in detail in Section 3. Subsequently, a case where t is the matrix of the mapped testing samples and the
study of industrial fed-batch chlortetracycline fermentation pro- resulting Kt is the (nt × n) “test” matrix whose elements are Kij =
cess is used to evaluate the proposed algorithm in Section 4. Finally, n+nt
K(xi , xj ) where {xi }i=n+1 and {xj }nj=1 are the testing and training
this research is concluded in Section 5.
samples, respectively.
It is worth noting that before nonlinear kernel projection, the
2. Preliminaries mean centering should be conducted on the above kernel matrix
by the following procedures
In this section, KPLS regression, JIT learning, PMI criterion, and  1
  1

GMM algorithm are briefly introduced. K= I− 1n 1Tn K I− 1n 1Tn (7)
n n
 1
 1

2.1. Kernel partial least squares Kt = Kt − 1n 1T K I− 1n 1Tn , (8)
n t n n
As an extension of the standard PLS proposed by Wold (1966), where I is again an n-dimensional identity matrix, and 1n , 1nt rep-
the KPLS model is constructed by mapping the original input into resent the vectors whose elements are ones, with length n and nt ,
a feature space F where a linear PLS model is created (Rosipal and respectively.
Trejo, 2001). Assume a nonlinear transformation of the input vari-
ables {xi }ni=1 into a high-dimensional feature space F, i.e. mapping 2.2. Just-in-time learning
given by

˚ : xi ∈ Rd → ˚(xi ) ∈ F (1) Just-in-time learning (Cybenko, 1996) is also called instance-

based learning, lazing learning or model-on-demand. This
Effectively it means that we can obtain a nonlinear regression approach can cope with changes in process characteristics as well
model in the space of the original input variables. as nonlinearity, thus has been used for nonlinear process modeling,
80 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93

Global modeling Just-in-time modeling the PMI between X and Y, for a set of already selected inputs Z, is
therefore given by
Modeling Query Query   
data data   fX  ,Y  (x , y )
tools
x new PMI(X, Y |Z) = fX  ,Y  (x , y ) ln dx dy , (9)
x new fX  (x )fY  (y )

Offline
Relevant Local
Database global Database dataset model
model
Predicted Predicted
Historical output Historical Modeling output
data ŷnew data tools ŷnew
Once the model is built, the Once the predicted output is obtained,
database is discarded. the local model is discarded.
Fig. 1. Comparison between global and JIT modeling methods.
monitoring as well as soft sensing (Cheng and Chiu, 2004; Fujiwara
et al., 2009; Kim et al., 2013).
In comparison with the conventional global modeling methods,
JIT modeling has the following features: (i) the available input and
output data are stored in a database; (ii) only when estimation is
required, a local model is constructed from relevant samples deter-
mined through some similarity measure to query data, and then the
output variable is estimated; (iii) the built local model is discarded
after its use for estimation; and (iv) JIT based model is inherently
adaptive by simply adding the new available input and output data
into the database.
The difference between conventional global modeling and JIT
modeling is illustrated in Fig. 1. The advantages of global models are
that they enjoy a mature mathematical and computational theory
so that the resulting models are easy to interpret in the frame-
work of the underlying applications (as physical laws for example).
However, global methods have two major drawbacks. First, the
optimizations are difficult to perform and, with the data essentially
replaced by the model, there are no good methods to update models
if new data become available (Cybenko, 1996). Second, the global
modeling needs to build a single global model which in practice
is extremely compute intensive and rarely yields good approxima-
tions on large problems. Alternatively, JIT learning uses relevant
data similar to query data to build models dynamically as the need
arises.
In the JIT learning framework, given a query data xnew , there
are three main steps to predict its output: (i) relevant samples that
are most relevant to xnew are selected from the database by a sim-
ilarity measure; (ii) a local model is built based on the selected
dataset; and (iii) the new output ynew is predicted using the built
where
x = x − E(x|Z), y = y − E(y|Z), (10)
where E(·) denotes the expectation operation. Use of the condi-
tional expectations ensures that the resulting variables x and y
stand for the residual information in variables x and y, once the
effect of the already selected variable(s) Z has been taken into con-
sideration.
PMI is symmetric under the same condition Z, PMI(X, Y |Z) =
PMI(Y, X|Z). We have 0 ≤ PMI(X, Y |Z) where zero is obtained only
if X and Y are independent under condition Z. In the case of Z = ∅,
PMI(X, Y |Z) = MI(X, Y ).
A sample estimate of the PMI criterion can be formulated as
 
1
n
fX  ,Y  (xi , yi )
PMI(X, Y |Z) = ln , (11)
n fX  (xi )fY  (yi )
i=1
where xi and yi are the ith residuals in the sample data set of size n;
fY  (yi ), fX  (xi ), and fX  ,Y  (xi , yi ) are the marginal and joint probability
densities, respectively.
The computation of PMI is similar to that of MI, where both
require estimation of probability densities. In MI, the probability
densities are estimated for the original inputs and output, whereas
in PMI, the densities are estimated for the residual information in
variables X and Y after considering the effect of the already selected
inputs Z. Various methods have been proposed for PMI estima-
tion, such as box-counting algorithm (Paluš et al., 2001), histograms
(Fernando et al., 2009), kernel estimator (May et al., 2008a, 2008b;
Scott, 1992; Sharma, 2000), and k-nearest neighbor (k-NN) statis-
tics (Frenzel and Pompe, 2007).
2.4. Gaussian mixture model
Gaussian mixture model (GMM) is a probabilistic model that
assumes all the data points are generated from a mixture of a
finite number of Gaussian distributions with unknown parameters.
As a kind of statistical inference based clustering method, GMM
has been widely used in pattern recognition and machine learning
(Bishop, 2006) as well as industrial applications such as process
monitoring (Yu and Qin, 2008, 2009) and soft sensor modeling (Yu,
2012a). For an arbitrary data x ∈ Rd , we suppose that it comes from
a Gaussian mixture distribution given by
C

local model, then the local model is discarded. When a new query
data comes, a new local model will be built based on the same
procedure.
2.3. Partial mutual information
p(x|GM ) = i N(x|␮i , i ), (12)
i=1
where C denotes the number of Gaussian components; GM =
{␮1 , . . ., ␮C , 1 , . . ., C , ␲1 , . . ., ␲C } is the vector of parameters of
the mixture model; N(x|␮i , i ) represents a multivariate Gaussian
distribution with the mean vector ␮i and the covariance matrix i ,
and the corresponding probability density function is given by
Mutual information (MI) is a nonparametric and nonlinear mea-  
sure of relevance derived from information theory. It can be used 1 1 −1
p(x|GM,i ) =
exp − (x − ␮i )T i (x − ␮i ) (13)
to identify nonlinear dependence between input and output vari- 2
(2)d det(i )
ables. However, MI cannot deal with dependencies between input
variables. To overcome this problem, Sharma (2000) introduced with the mixing coefficients {i }Ci=1 which satisfy
the partial mutual information (PMI) criterion, which addresses
the issue of redundant variables. This is achieved by quantifying
C

the nonlinear dependence of candidate input X on output Y that is i = 1, 0 ≤ i ≤ 1. (14)

not accounted for by the already selected input variables Z. Thus, i=1
 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93
Suppose we have a dataset of observations X = {xi }ni=1 that are inde-
pendently drawn from the Gaussian mixture distribution, then the
parameters of mixture can be estimated by maximizing the log of
the likelihood function given by
C  However, two difficulties are encountered in the PMI based
n
ln p(X|GM ) = ln c N (xi |␮c , c )
i c=1
The expectation-maximization (EM) algorithm is commonly used
for GMM estimation (McLachlan and Peel, 2000).
According to the Bayesian inference strategy, the posterior
probability of x belonging to the ith Gaussian component can be
calculated by
i N(x|␮i , i )
P(GM,i |x) =
C
c N(x|␮c , c )
c=1
3. Proposed adaptive soft sensor
The proposed MJIT-KPLS soft sensing algorithm builds soft
sensors online by performing KPLS regression in multiphase JIT
modeling framework. The JIT learning framework for each operat-
ing phase is similarly determined by a group of offline parameters
which are different from phase to phase. Thus the development
of MJIT-KPLS soft sensors consists of building multiple similar
JIT-KPLS modeling frameworks. Without considering a particular
phase, the development of JIT-KPLS soft sensors can be split into
the following steps: (i) outlier detection of multiway data; (ii) input
variable selection; (iii) similarity measure definition; (iv) adaptive
sample selection, online training and prediction; and (v) database
updating. The above steps will be discussed in detail in the fol-
lowing sections. The selection of multiphase offline parameters is
finally presented.
3.1. Outlier detection of multiway data
To develop robust soft sensors, 3 edit rule is firstly used to
clean outliers in modeling samples. The presence of outliers may
lead to model misspecification, biased parameter estimation and
incorrect analysis results (Liu et al., 2004). However, the outlier
detection of batch process data is different from other processes
due to their unique three-dimension matrix X - (IB × JB × KB ) with
IB , JB , and KB representing the number of batches, process vari-
ables, and sampling instants, respectively. Firstly, X - are converted
into a two-dimension matrix X (IB × KB JB ) via batch-wise unfolding
(Westerhuis et al., 1999). Then 3 edit rule is performed on the data
block within the same sampling instant to remove the outliers.
Finally, the clean samples are further rearranged to X (IB KB × JB )
through variable-wise unfolding (Westerhuis et al., 1999) as illus-
trated in Fig. 2.
3.2. Input variable selection
Selecting appropriate input variables is crucial to enhance the
estimation accuracy and maintain the reliability of soft sensors, but
the variable selection is one of the most difficult tasks concern-
ing soft sensor development (George, 2000). Presence of irrelevant
(or less relevant) variable data in the input data will leads to
performance deterioration of soft sensors. A soft sensor works sat-
isfactorily if only those secondary variables that are most relevant
to the primary variable are employed (Pani and Mohanta, 2011).
To select input variables most relevant to output variable while
at the same time considering the redundancy between input vari-
ables, PMI criterion is applied in our work. In practice, the true
81
values of PMI between variables are not known, and thus have to
be estimated from the available data {X,y}. For this reason, a k-NN
based estimator proposed by Frenzel and Pompe (2007) is used to
calculate the PMI criterion.
selection of input variables. The first one concerns the choice of
(15)
neighbor number used in k-NN based estimator. The number k of
neighbors taken into account in the PMI estimation must be chosen
carefully, especially in the case of a small sample and noisy data.
Often, the result of the selection highly depends on the choice of
this parameter. If the number of neighbors is too small, the esti-
mation will have a large variance; if the number of neighbors is
too large, the variance of the estimator will be small, but all esti-
mations will converge to zero, even for highly-dependent variables.
(16) To address this issue, a combined K-fold/permutation test (François
et al., 2007) is used for determining the optimal number of neigh-
bors.
The second difficulty lies in deciding when to halt the selection
procedure. As necessary part of any stepwise predictor selection
method is a criterion that indicates whether the selected input
variable is significant to the output variable. In this work, a con-
fidence limit is calculated assuming independence between inputs
and the output variable. This independence is formed by random-
izing samples from the original variable, and this can be achieved
by bootstrapping for which the bootstrapping size is chosen as
100 in our work. Then, we can obtain a 95th percentile confidence
limit (referred to as the 95th percentile randomized sample PMI
score). That is, if the PMI score of the selected input is greater than
the estimated 95th percentile randomized sample PMI score, one
can conclude that there would be less than a 5% chance that the
variables are truly independent.
A stepwise procedure of input variable selection based PMI cri-
terion can now be summarized as follows:
Step 1. Initialize the potential input variables denoted by {Xi }di=1 ,
and set the selected inputs as the vector Zin .
Step 2. Determine the optimal number of neighbors for k-NN esti-
mator using the combined K-fold/permutation test.
Step 3. Compute the PMI values between the output variable Y and
each of the plausible new input variables in Zin , conditional to the
pre-existing input variables Z.
Step 4. Identify the variable with the highest PMI score in the pre-
vious step. If this PMI score is greater than the 95th percentile
randomized sample PMI score, add the new input variable into Z;
otherwise terminate the selection procedure.
Step 5. Repeat Steps 3 and 4 until all significant input variables
have been selected.
3.3. Similarity measure definition
In JIT learning framework, those samples with high similarity to
query data are selected for building local model when estimation is
required. The similarity measure between samples has significant
influence on the estimation accuracy of JIT based models. The most
frequently used similarity is defined on the basis of the Euclidean
distance or the Mahalanobis distance (Ge and Song, 2010; Kim
et al., 2013). Some other similarity measures are based on the angle
based (Cheng and Chiu, 2004) and correlation based (Fujiwara et al.,
2009) methods. More similarity measures were surveyed by (Kano
and Fujiwara, 2013; Kim et al., 2013). However, these similarity
measures are defined only from the perspective of sample algebraic
space while no process knowledge is concerned. Actually, the aim of
defining similarity measure is to select those samples that describe
similar process characteristics. A drawback of evaluating sample
similarity from algebraic space alone is that by using such simi-
larity measure, some samples with high similarity exhibit greatly
82 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93

Fig. 2. Illustration of outlier detection for batch processes data.

different process behavior since they do not belong to similar by dividing their standard deviations, and (ii) center them by sub-
process phases. Hence, for batch processes, we aim to enhance tracting their averages.
the predictive capability of JIT based soft sensors by integrating
Euclidean distance based sample similarity with phase similarity. 3.4. Adaptive sample selection, online training and prediction
Subsequently, it is necessary to define a measure for describing
phase similarity. In batch processes each run has the same duration After defining the similarity measure, the similarities between
and the variables follow closely some predetermined trajectories the query data and the historical samples can be calculated. For JIT
(Westerhuis et al., 1999). In general, batch processes go through based soft sensors, only those samples that are most relevant to
multiple phases during the batch run. Among process variables, query data are selected. Thus, the number of local modeling sam-
process duration plays a significant role to indicate the progress of ples Nlocal is usually less than the number of samples (n) in database.
the batch. Thus in this work, process duration is chosen as an indi- In practice, how to determine the local modeling size is difficult.
cator variable to evaluate the phase similarity. Although the phase The conventional method is to determine this parameter offline
similarity can be roughly defined based on process duration, it is not by trial and error, as reported in (Ge and Song, 2010). Although
enough to select local modeling samples by adopting phase simi- the local modeling size determined offline give rise to high accu-
larity alone, because batch-to-batch variation in batch processes is racy for training data, such empirical method is time-consuming
inevitable. Therefore, a hybrid similarity measure is defined by and the obtained parameters are no longer optimal for new pro-
samp phase
cess data. Thus, the predictive performance may deteriorate if using
si = si + (1 − )si , (17) fixed number of local modeling size. Moreover, different query data
 samp  may correspond to different optimal number of local modeling size.
samp di To enhance the predictive accuracy and adaptability of JIT based
si = exp − samp , (18)
d ϕsamp soft sensors, an adaptive sample strategy based on online cross-
min and N max
 validation is proposed. Specially, two parameters Nlocal local
phase are chosen such that only the relevant datasets formed by the
phase di
si = exp − , (19) min max
Nlocal th relevant data to the Nlocal th data are used in the regression.
phase
d ϕphase min , N max ](l = 1, 2, . . ., L) with the search step N
For Nl ∈ [Nlocal local step , the
samp total number of potential local models involved in the optimization
di = ||xnew − xi ||, (20)
procedure is
phase  
di = ||tnew − ti ||, (21) max − N min
Nlocal local
L = int +1 (22)
Nstep
samp phase
where si , si
and (i = 1, 2, · · ·, n) denote hybrid similarity,
si
sample similarity and phase similarity, respectively;  is a weight The detailed algorithm of the JIT-KPLS soft sensor is described as
parameter between 0 and 1; d
samp
and d
phase
are the standard devi- follows. Given a database D = {X, y} = {xi , yi }ni=1 , the offline param-
min , N max , N
eters Nlocal
ations of di
samp
and di
phase
, respectively; tnew and ti represent process local step , weight parameter , the number of latent

duration corresponding to query data xnew and historical samples
in database, respectively; ϕsamp and ϕphase are two localization
parameters. The proposed hybrid similarity measure is equivalent
to the conventional single similarity measure when  = 1.
Usually, the two similarity measures, si
samp
and si
phase
should be
normalized before weighted calculation to eliminate the influence
resulted from magnitude differences. In the absence of knowledge
about the relative importance of the variables, the standard mul-
tivariate approach is use to (i) scale each variable to unit variance
variables LV, RBF kernel width , and a query data xnew with time
variable tnew :
Step 1. Set l = 1, and calculate the similarity si (i = 1, 2, . . ., n)
between xnew and each data xi by using Eq. (17).
Step 2. Arrange all the si in the descending order. The relevant
dataset Dl = (Xl , yl ) is constructed by selecting Nl most relevant
samples, corresponding to the largest si to the Nl th largest si , where
min + (l − 1)N
Nl = Nlocal
step .
 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93
Step 3. Compute the prediction performance of local KPLS model
on training dataset Dl via K-fold cross-validation.
We split the selected dataset Dl into K roughly equal-sized parts.
For the kth part, we fit the local KPLS model to the other K − 1
parts of the data and the prediction error of the fitted model when
predicting the kth part of samples is given by
 Most data-driven soft sensor modeling techniques rely on
nk
y
i=1 i
− fKPLS (xi , k )
RMSEk = , (23)
nk
where fKPLS (x, k ) represents the local KPLS model built from K − 1
parts of the dataset Dl , k means parameters of local PLS model,
and nk denotes the number of samples in the kth part of samples.
Repeat the above training and prediction for K times, and then the
cross-validation estimation of prediction error is calculated by
1
K
CVl = RMSEk
K new operation condition may not give rise to similar output as
k=1
Step 4. Set l = l + 1, and go to Step 2 until l = L.
Step 5. According to the cross-validation errors, the optimal Nl is
determined by
opt
Nl = arg min(CVl )
l replacement (MSR) rule is proposed to update database in MJIT-
opt
Step 6. Train local KPLS model using the most Nl relevant sam-
opt
ples, corresponding to the largest si to the Nl th largest si , and
then the output variable is predicted by
ŷnew = fKPLS (xnew , opt ),
where opt denotes the optimal model parameters of the local
KPLS model.
Step 7. When the next query data comes, go to Step 1.
In the present work, five-fold cross-validation is adopted. The
KPLS regression detailed in Section 2.1 servers as the local modeling
method in the proposed adaptive soft sensing algorithm.
3.5. Database updating scheme
JIT based models assume that all available observations are
stored in a database, which distinguishes JIT modeling from other
global modeling techniques. Obviously, database technology is cru-
cial to JIT based models. JIT modeling updates models in a different
way from traditional global methods. For example, neural networks
need to be retrained to adjust the network parameters according
to the new operation. In the extreme cases, the network structure
needs to be re-determined to achieve better prediction. Clearly,
this procedure is not desirable from computational point of view.
In contrast, JIT based model is inherently adaptive by simply adding
the current available input and output data to the database.
As Kim et al. (2013) stated, for nonlinear time-varying pro-
cesses, both the age and the density of samples are important
indexes to evaluate the goodness of database. The estimation per-
formance may deteriorate when the samples in the database are
old and sparse. The samples in the database should be carefully
selected to assure the estimation performance. Assume a database
with unchanged size has been built in advance, our focus shifts
to updating database when implementing online. To maintain the
database size, adding a new sample means removing an old one.
A popular updating strategy is the moving time-window method,
in which a newly obtained sample is added to the database and
the oldest sample is removed from the database. This method has
been successfully used in the steel processes for more than 7 years
(Shigemori et al., 2011). Nevertheless, this updating method should
be improved for applying in batch processes. The main drawback
83
of the conventional updating method is that the samples in some
process regions may become sparse because the oldest samples
deleted may describe different characteristics from that described
by the new samples. For batch processes, new points should be used
to replace those old points that describe similar process behaviors
to assure that the samples in every process region is new and dense.
supervised learning where the model is trained on available his-
torical data. This learning method is an inductive process – given a
finite training set we wish to infer a function that makes predictions
for all possible input values. For all the variety, supervised learn-
ing algorithms are based on the idea that similar input patterns
will usually give rise to similar outputs (or output distributions)
(Rasmussen and Williams, 2006). This idea is the essence of data-
driven modeling that is based on supervised learning. Therefore,
changes in process characteristics will result in the changes of rela-
tions between input and output data. Exactly, similar input under
(24)
before. Consequently, to capture the newest process characteris-
tics, the task of updating database shifts to finding the old samples
whose inputs are most similar to inputs of new samples, and then
replace them.
Following the idea mentioned above, maximum similarity
(25)
KPLS soft sensing algorithm. In our proposed updating strategy, the
sample updating is only carried out in sample datasets with similar
process characteristics. That is, when a new sample is added into
the database, an old sample that is most similar to the new sample
will be removed.
(26)
The modeling framework of the JIT-KPLS soft sensor is depicted
in Fig. 3. Given a new sample (xnew , ynew ) with time variable tnew ,
and similarity weight update . The database updating algorithm is
explained as follows:
Step 1. Check whether the new sample is outlier; if yes, discard it;
otherwise go to Step 2.
Step 2. Compute the similarity between the new sample and his-
torical samples based on the proposed hybrid similarity measure
given by
si = update si
samp phase
+ (1 − update )si , (27)
where the similarity weight update may have different value from
that used in adaptive sample selection.
Step 3. Find out the historical sample (xold , yold ) whose input has
the highest similarity to that of the new sample, then remove it
and add the new sample (xnew , ynew ) into the database.
Step 4. When a new sample is available, go to Step 1.
3.6. Selection of offline parameters
Although local model is built online in JIT learning framework,
there are still some critical offline parameters need to be optimized
offline using the modeling data. These parameters are as follows:
(i) The number of latent variables for KPLS regression, LV.
(ii) The width of RBF kernel function for KPLS regression, .
(iii) The similarity weights,  and update , which are used for
selecting local modeling samples and updating database,
respectively.
(iv) The localization parameters, ϕsamp and ϕphase , which are used
for sample similarity and phase similarity, respectively.
min ,
(v) The search parameters of the optimal local modeling size, Nlocal
max , and N
Nlocal
step , which represent the minimum, the maximum
and the search step.
84 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93

Fig. 3. Modeling framework of the JIT-KPLS soft sensing algorithm.

These parameters construct a vector given below The schematic diagram of the proposed MJIT-KPLS algorithm
is depicted in Fig. 4, and the step-by-step procedure is listed as
min max
off = {LV, , , update , Nlocal , Nstep , Nlocal , ϕsamp , ϕphase } (28) follows:

Since multiple phases are often involved in batch processes, multi-
phase modeling is preferable to the conventional single modeling
framework. That is, multiple groups of offline parameters, i.e.
off,i (i = 1, 2, . . ., C) are required for different operating phases.
Thus, GMM is estimated to identify the operating phases through-
out batch run, resulting C Gaussian components. For the C identified
phases, the corresponding JIT-KPLS modeling frameworks are
expressed as
{JIT-KPLS(x|off,1 ), JIT-KPLS(x|off,2 ), . . ., JIT-KPLS(x|off,C )}
(29)
Step 1. Collect the input and output data of the batch processes
and remove the outliers using multiway 3 edit rule.
Step 2. Perform GMM clustering to identify C operating phases
where various JIT-KPLS modeling frameworks are determined for
different phases.
Step 3. When a query data comes, firstly compute its posterior
probabilities with respect to different operating phases and choose
the JIT-KPLS modeling framework corresponding to the maximal
posterior probabilities.
Step 4. Select the dataset most relevant to query data from the
database and then build a local KPLS model.

Within each phase, the optimal parameters off,i are optimized

offline for the ith operating phase. When implemented online, the
Bayesian inference based posterior probabilities for query data xnew
with respect to different phases are estimated, and then the JIT-
KPLS modeling framework corresponding to the batch phase with
the maximal posterior probability is adaptively chosen for online
prediction.
It is worth noting that the sample database is equally shared
by all JIT-KPLS modeling frameworks. That means that for every
query data, the local modeling samples are searched throughout
the whole database irrespective of considering which phase the
query data belongs to. By contrast, only the offline parameters cor-
responding to the same phase as query data are adaptively chosen
for local KPLS training and prediction.
Owing to the lack of a systematic guideline for determining
the optimal values of offline parameters, two methods are recom-
mended: (i) selection by optimization on validation set; and (ii)
cross-validation on training data. Ideally, if we had enough data,
we would set aside a validation set and use it to assess the perfor-
mance of our predictive model that is characterized by potential
offline parameters. If data are scarce, to solve the problem, K-fold
cross-validation is preferable, in which part of the available data in
training data is used to fit the model, and a different part is used
to test it. The combination of offline parameters resulting in the
smallest validation error is finally chosen. Fig. 4. Schematic diagram of the proposed MJIT-KPLS soft sensing algorithm.
 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93 85

Fig. 5. Schematic diagram of the industrial fed-batch CTC fermentation process.

Step 5. Estimate the output of the query data, and then discard the 4.1. Industrial chlortetracycline fed-batch fermentation process
model.
Step 6. When a new query data comes, return to Step 3. Chlortetracycline (CTC) is a kind of broad-spectrum antibiotics,
which is included in the tetracycline family and produced by the
4. Case study cultivation of Streptomyces aureofaciens. It has been widely used
in pharmaceutical, agriculture and animal husbandry. Especially,
In this section, the proposed soft sensing algorithm is tested on CTC is a perfect feed additive due to the ability of promoting the
industrial fed-batch chlortetracycline fermentation process. Two growth of animals. In addition, CTC is characterized by low residual
indexes are used to evaluate the model performance, including amount, low production cost and mature production technology.
root-mean-square error (RMSE) and the coefficient of determina- Thus, CTC has become a bacteriostatic growth additive with the
tion (R2 ) given by largest consumption in feed industry for a long time and will hold
 this trend in the future.

 1
Ntest The CTC fermentation process under study is an industrial
RMSE = 
2
(ŷi − yi ) , (30) process practiced in a biochemical factory belonging to Charoen
Ntest
i=1 Pokphand Group. The schematic diagram of the fed-batch CTC fer-
  mentation process is shown in Fig. 5.
Ntest 2 The cultivation of S. aureofaciens is carried out in an air-lift
i=1
(ŷi − yi )
R2 = 1 −  , (31) stirred fermentor with a volume of 120 m3 . This process mainly
Ntest
i=1
(yi − ȳi )2 goes through two stages: the cultivation process begins with a short
period of batch operation to grow the microorganisms and maxi-
where ŷi is the estimated output, yi is the actual output, and ȳi is mize the cell density; and then it is switched to fed-batch mode in
the mean of the actual output; Ntest denotes the number of testing order to boost the synthesis of CTC as the target metabolite while
samples. the cell growth continues at a slower rate. The fermentor serves
RMSE is used as an indicator of the predictive accuracy in online as the major equipment in the process with the substrate and air
prediction. R2 represents a squared correlation between the actual continuously fed in to supply the raw materials of cell culture and
output and the predicted output, it measures how much of the total maintain the necessary oxygen consumption for microorganisms.
variance in the output variable data can be explained by the model. The industrial plant and product are shown in Fig. 6. Meanwhile,
If the model accounts for a small percentage of the total variation the trend plots of variables are given in Fig. 7.
of the output variable (R2 → 0), i.e. if the unexplained variation is A distributed control system is applied to allow the auto-
high, then the model is either not good or there is a weak correlation matic supervision and control of the main process variables. The
between the input and output variables. supervisory system is built by using Ethernet interconnecting pro-
The configuration of the computer for soft sensor modeling is grammable logical controllers, which actuate directly on process
listed as follows: OS: Windows 7 Ultimate (32 bit); CPU: Pentium equipments, with human–machine user interfaces. Five control
(R) Dual-Core E6600 (3.06 GHz, 3.07 GHz); RAM: 2 GB; MATLAB ver- loops are involved: temperature, pH, antifoam feeding, DO, and
sion: 2010a. substrate feeding. The temperature is controlled at 32 ◦ C in batch
86 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93

on their experience. This situation results from the fact that no

reliable online sensor is available for measurements of substrate
concentration.
Although there are some attempts being done on online mea-
surements, no reliable sensors, up to now, are suitable for industrial
use for fermentations. Bioprocesses are remarkably harsh envi-
ronments for online sensors, as the growing culture can infiltrate
sensors and this invalidates their results. In real production process,
substrate concentration has to be analyzed offline every 6 h. In addi-
tion, large time delay is caused since a large number of fermentors
need to be analyzed at each analysis period. To provide online pre-
diction of substrate concentration, a data-driven soft sensor based
on the proposed soft sensing algorithm is developed in next section.

4.2. Online prediction of substrate concentration

Three industrial datasets were collected for soft sensor modeling

and evaluation: (i) training dataset (from September 3 to November
18, 2011) collected from 50 batches consists of 755 samples with
Fig. 6. The industrial plant and product of CTC fermentation process. the sampling interval of 6 h; (ii) testing dataset I (from November
19, 2011, to January 2, 2012) collected from 25 batches consists
of 388 samples with the sampling interval of 6 h; and (iii) test-
culture period and 29 ◦ C in fed-batch period by manipulating the ing dataset II (from March 27 to April 22, 2013) collected from 30
flow rate of cold water. Ammonia water is automatically fed in to batches consists of 355 samples with the sampling interval of 8 h.
control the pH value at about 5.95. Antifoam feeding is triggered The training dataset is used for model learning while the testing
once the foam has reached the foam detection electrode. DO is con- dataset I is used for model evaluation. Besides, the testing dataset
trolled by manipulating air flow rate using a cascade PID controller. II is used to assess the adaptation capability of the proposed MJIT-
Different from the four variables mentioned above, the feeding rate KPLS and other soft sensors because it describes a new process
of substrate remains to be determined by skilled operators based state whose characteristics has significantly changed. Nine input

34 6.4 150
oxygen (%)
Dissolved

32 6.2 100
pH

30 6 50

28 0
0 50 100 0 50 100 0 50 100

5
x 10
Air flow rate (m3/h)

4000 4 800
feeding rate (L/h)
onsumption (m3)

600
Substrate

3000
2 400
Air

2000
200
1000 0 0
0 50 100 0 50 100 0 50 100

4
x 10
Culture volume (m3)

4 3000 100
consumption (L)
consumption (L)

Ammonia
Substrate

2000 90
2
1000 80

0 0 70
0 50 100 0 50 100 0 50 100

6
concentration (%)
Substrate

2
0 50 100
Time (h)

Fig. 7. Trend plots of variables in fed-batch CTC fermentation process.

 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93 87

Table 1 Table 2
Results of input variable selection based on PMI criterion. The 95th percentile PMI Average prediction results from MJIT-KPLS A-D models (testing dataset I).
corresponds to the 95th percentile randomized sample PMI score. Note that the last
row with numbers in bold represents the case where the PMI score is lower than the Model RMSE R2
95th percentile randomized sample PMI score, because of which pH is not included MJIT-KPLS A 0.2483 0.9442
in the final input variable set. MJIT-KPLS B 0.2416 0.9472
MJIT-KPLS C 0.2461 0.9452
No. Potential input variable PMI 95th percentile PMI
MJIT-KPLS D 0.2393 0.9482
1 Cultivation time (min) 0.7105 0.1557
2 Substrate feeding rate (L/h) 0.6347 0.1188
3 Ammonia consumption (L) 0.4967 0.1120 MJIT-KPLS C: MJIT-KPLS with single similarity measure and adap-
4 Substrate consumption (L) 0.4296 0.1321 tive sample selection.
5 Air consumption (m3 ) 0.3946 0.1205
6 Temperature (◦ C) 0.3209 0.0953
MJIT-KPLS D: MJIT-KPLS with hybrid similarity measure and adap-
7 Culture volume (m3 ) 0.2905 0.1258 tive sample selection.
8 Dissolved oxygen (%) 0.2737 0.1192
9 Air flow rate (m3 /h) 0.2524 0.1023 To build MJIT-KPLS soft sensors, the vector of offline parame-
10 pH 0.1026 0.1225
ters off discussed in Section 3.6 must be determined before online
implementation. The localization parameters, ϕsamp and ϕphase are
set to 2.5. And the similarity weight for database updating is deter-
variables were finally selected for soft sensor development based mined by trial and error in the range of [0, 0.1, . . ., 1]. It is noted
on PMI criterion (see Table 1). that the fixed local modeling size Nlocal is used for MJIT-KPLS A and B
min , N
soft sensors, while a group of search parameters {Nlocal max
To enable multiphase modeling, the operating phases of indus- step , Nlocal }
trial fed-batch CTC fermentation process are identified using are employed in MJIT-KPLS C and D soft sensors. The remaining
Gaussian mixture model and the phase identification results are undetermined offline parameters are optimized by five-fold cross-
shown in Fig. 8. A three-stage cultivation is applied to the industrial validation in the following ranges:
CTC production process, where S. aureofaciens firstly goes through
LV = [1, 2, . . ., 15]; (i)
progressive scale-up cultivation during the former two stages until
the microorganisms concentration reaches a desired level, and then  = [1, 2, . . ., 15]; (ii)
the seed liquid is inoculated to the fermentor corresponding to the
third stage. Thus, for the third stage on which we focus, the entire  = [0, 0.1, . . ., 1]; (iii)
CTC cultivation process is identified with three distinct phases that Nlocal = [40, 45, . . ., 150]; (iv)
correspond to the exponential phase, stationary phase, and death
phase throughout the microbial culture cycle. Subsequently, vari- min max
Nlocal = [40, 45, . . ., 90], Nlocal = [110, 115, . . ., 150],
ous JIT-KPLS modeling frameworks are built for different operating
phases. and Nstep = 5. (v)
The first goal of our experiments is to evaluate the effectiveness The offline parameters for different phases are determined in the
of the proposed hybrid similarity measure and the adaptive sample same way. The above settings result in building a large number
selection strategy. Thus, the following four models with different of JIT modeling frameworks with different parameter combination
combination of similarity measure and sample selection strategy off . For every type of MJIT-KPLS models, ten groups of the optimal
are compared. offline parameters determined using training data are selected to
obtain reliable evaluation of the model performance.
MJIT-KPLS A: MJIT-KPLS with single similarity measure and fixed The performance comparison of MJIT-KPLS A-D models on the
local modeling size. testing dataset I is shown in Fig. 9. Meanwhile, the average pre-
MJIT-KPLS B: MJIT-KPLS with hybrid similarity measure and fixed diction results are listed in Table 2. Comparing model A with B
local modeling size. reveals that the MJIT-KPLS soft sensors using the proposed hybrid
similarity measure outperforms those using the conventional sin-
gle similarity measure. The comparison results between model C
4 and D also support the efficiency of the new hybrid similarity mea-
sure. This fact indicates that, for batch processes, phase similarity

Death phase
0.25
3
Phase No.

Stationary phase
0.245
2
RMSE

Exponential
phase

1 0.24

MJIT-KPLS A
MJIT-KPLS B
MJIT-KPLS C
MJIT-KPLS D
0 0.235
0 10 20 30 40 50 60 70 80 90 100 1 2 3 4 5 6 7 8 9 10
Time (h) Model number

Fig. 8. Phase identification of industrial fed-batch CTC fermentation process. Fig. 9. Performance comparison between MJIT-KPLS A-D models (testing dataset I).
88 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93

measure plays a critical role of evaluating process similarity. It also Without update
shows that process priori knowledge or expert knowledge has to 0.255 λ update=1

be taken into account especially when developing data-driven soft λ update=0.8

λ update=0.6
sensors. Moreover, the adaptive sample selection strategy is also
analyzed. A direct comparison of the results of models B and D 0.25

indicates that the prediction accuracy is improved by using the

RMSE
proposed adaptive sample selection strategy instead of the con-
0.245
ventional fixed modeling size. According to the above comparison
results, it can be concluded that MJIT-KPLS D performs best owing
to the hybrid similarity measure and the adaptive sample selection 0.24
method.
In the next step, the effectiveness of the proposed database
updating scheme, i.e. maximum similarity replacement (MSR) rule 0.235
1 2 3 4 5 6 7 8 9 10
is evaluated. For traditional data-based modeling methods, it is not Model number
(a) MJIT-KPLS A
a trivial task to update models online. In contrast, MJIT-KPLS is
0.25
inherently adaptive by simply adding available new input-output Without update
0.248
pairs into database, which makes the soft sensor capable of cap- λ update=1

turing the new process characteristics. In the present soft sensing 0.246 λ update=0.8
λ update=0.6
algorithm, MSR rule is applied to update the database. Two sce- 0.244
narios are compared: online prediction without updating and with
0.242
updating under various update similarity weights. In the former,

RMSE
0.24
the original database remains unchanged, whereas in the latter the
database is constantly updated by using MSR rule once a new input- 0.238

output pair is available. Compared to the prediction results without 0.236

database updating, without exception, the prediction errors of 0.234
MJIT-KPLS A-D models have all been clearly reduced by using MSR
0.232
database updating scheme (see Fig. 10).
To illustrate the predictive capability of the proposed soft 0.23
1 2 3 4 5 6 7 8 9 10
sensing algorithm, the prediction results from MJIT-KPLS D model Model number
(b) MJIT-KPLS B
for four testing runs are depicted in Fig. 11. It is readily observed that
0.255
the proposed soft sensor can be used to predict substrate concen- Without update
tration with small RMSE and high R2 . It should be emphasized that λ update=1
λ update=0.8
the adaptive sample selection method is used to determine local
0.25 λ update=0.6
modeling size for MJIT-KPLS D model. The optimal local modeling
sizes corresponding to the four testing batches are shown in Fig. 12.
It is shown that the obtained optimal local modeling size changes
RMSE

0.245
along sampling time. This may be due to the fact that batch process
shows different nonlinearity within different phases, resulting in
the variations of the optimal number of samples required to build
0.24
local models.
To further assess the predictive accuracy and the adaptation
capability of the proposed MJIT-KPLS soft sensing algorithm, the
0.235
following soft sensors are developed for comparison: 1 2 3 4 5 6 7 8 9 10
Model number
(i) Single-model based soft sensors. Three popular global model- (c) MJIT-KPLS C

ing techniques are selected: PLS (de Jong, 1993), KPLS (Rosipal 0.245
and Trejo, 2001), and GPR (Rasmussen and Williams, 2006).
(ii) Conventional multi-model based soft sensors. The adaptive
multi-model soft sensor modeling framework (Yu, 2012a) is 0.24
adopted to handle multiphase characteristics of batch pro-
cesses by constructing local predictive models for various
RMSE

operating phases. First, Gaussian mixture model is estimated 0.235

from training data to identify different operating phases. Fur-
ther, various localized models are built to characterize the
shifting dynamics across different phases. Using the Bayesian 0.23
Without update
λ update=1
inference strategy, the new sample is classified into a partic-
λ update=0.8
ular phase with the maximal posterior probability, and thus,
λ update=0.6
the local model representing the identical phase is chosen for 0.225
1 2 3 4 5 6 7 8 9 10
prediction. By applying the adaptive multi-model modeling
Model number
framework, multi-kernel KPLS (MKPLS) and multi-kernel GPR (d) MJIT-KPLS D

(MGPR) soft sensors are constructed.

Fig. 10. Prediction performance of MJIT-KPLS A-D models before and after database
(iii) Bayesian model averaging (BMA) based multi-model soft sen-
updating (testing dataset I).
sors. Though multiple localized models can handle multiphase
operation, they may not effectively characterize the transient
dynamics during the transitional stages. Thus, the BMA based
multi-model soft sensing framework (Yu et al., 2013) is used to
 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93 89

6.5 160
Offline analysis

Number of local modeling samples

6
MJIT-KPLS D prediction 150
Substrate concentration (%)

5.5
140
5
4.5 130
4
120
3.5 Phase I Phase II Phase III
3 110

2.5 100
2 Phase I Phase II Phase III
90
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100
Time (h)
Time (h)
(a) Testing batc h 1
(a) Testing batch 1, RMSE = 0.1599, R2 = 0.9647

6.5 160

Number of local modeling samples

6 Offline analysis
150
MJIT-KPLS D prediction
Substrate concentration (%)

5.5
140
5
4.5 130

4 120
3.5 Phase I Phase II Phase III
110
3
2.5 100
Phase I Phase II Phase III
2 90
0 10 20 30 40 50 60 70 80 90 100
0 10 20 30 40 50 60 70 80 90 100
Time (h)
Time (h) (b) Test ing batch 2
(b) Testing batch 2, RMSE = 0.1576, R2 = 0.9822
160
6.5
Number of local modeling samples

150
6 Offline analys is
MJIT-KPLS D prediction
Substrate concentration (%)

5.5 140

5 130
4.5
120
4
3.5 Phase I Phase II Phase III 110

3 100
2.5 Phase I Phase II Phase III
90
2 0 10 20 30 40 50 60 70 80 90 100
Time (h)
0 10 20 30 40 50 60 70 80 90 100 (c) Test ing batc h 3
Time (h)
(c) Testing batch 3, RMSE = 0.1475, R2 = 0.9865 160
Number of local modeling samples

6.5 150

6 Offline analys is
140
MJIT-KPLS D prediction
Substrate concentration (%)

5.5
130
5
4.5 120

4 110
3.5 Phase I Phase II Phase III
100
3 Phase I Phase II Phase III
2.5 90
0 10 20 30 40 50 60 70 80 90 100
2 Time (h)
(d) Testing batch 4
0 10 20 30 40 50 60 70 80 90 100
Time (h) Fig. 12. The adaptive optimal number of local modeling samples from MJIT-KPLS D
(d) Testing batch 4, RMSE = 0.1484, R2 = 0.9861 soft sensor, update = 0.8.

Fig. 11. Prediction results of four testing batches by using MJIT-KPLS D model with
database updating, and update = 0.8.
90 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93

0.5 GPR soft sensors result in predictions that are by a large margin
Method 1: PLS more accurate than the predictions of the single PLS soft sensor. The
Method 2: KPLS
Method 3: GPR
great enhancement of prediction accuracy indicates that the non-
0.45
Method 4: MKPLS linear modeling techniques are more suitable for nonlinear batch
Method 5: MGPR processes.
Method 6: BMA-MKPLS
0.4 Method 7: BMA-MGPR
Though single nonlinear soft sensors are capable of dealing
Method 8: JIT-PLS with process nonlinearity, they may be ill-suited for batch pro-
Method 9: JIT-KPLS cesses with multiple operating phases. The multi-model based KPLS
RMSE

Method 10: MJIT-KPLS

0.35
BMA based
0.3 Single-model Multi-model JIT learning
multi-model
0.25
Testing dataset I
Testing dataset II
0.2
1 2 3 4 5 6 7 8
Method
Fig. 13. Prediction RMSE from different soft sensor modeling methods.
built BMA-MKPLS and BMA-MGPR soft sensors. Compared to
the multi-model modeling methods, the local models for BMA
based methods are constructed in the same way, whereas the
prediction for samples within the transitional stage performs
differently. In BMA framework, the between-phase transitional
stage is determined by the posterior probabilities of measure-
ment samples with respect to the local models. Further, the
corresponding two consecutive phases are set as the adap-
tive weightings to integrate the two adjacent local modes for
prediction.
(iv) JIT learning based soft sensors. JIT-PLS and JIT-KPLS soft sensors
are developed within the newly proposed JIT learning frame-
work. That means, the proposed hybrid similarity measure,
the adaptive sample selection strategy, and the MSR database
updating scheme are applied to the two soft sensors. Unlike
MJIT-KPLS approach, JIT-PLS and JIT-KPLS only build a global
JIT learning framework irrespective of the multiphase charac-
teristics of batch processes. It should be noted that the JIT-PLS
method is the same as the JIT-KPLS approach except that the
local PLS model instead of KPLS model is built.
The prediction results on testing dataset I and II from differ-
ent soft sensors are given in Fig. 13. Meanwhile, the quantitative
model comparison is listed in Table 3. The prediction performance
on the testing dataset I are firstly analyzed. One can see that sin-
gle PLS based soft sensor leads to fairly poor prediction in terms
of much higher RMSE and lower R2 . This is due to the fact that
PLS is essentially linear modeling technique, and thus cannot well
handle process nonlinearity. In contrast, both the single KPLS and
and GPR are adopted so that the multiphase operation of batch
processes can be well characterized and thus give rise to better
prediction performance than the regular nonlinear single KPLS and
GPR soft sensors (see Fig. 13 and Table 3).
Although the multiple localized models in MKPLS and MGPR
performs better than the single models, such methods may not
effectively characterize the transient dynamics during the transi-
tional stages between consecutive phases. Compared to MKPLS and
MGPR soft sensors, one can see that the BMA-MKPLS and BMA-
9 10
MGPR soft sensors perform better because the Bayesian model
averaging strategy is designed to adaptively integrate the local
model predictions within the transitional stages between two adja-
cent operating phases (see Fig. 13 and Table 3).
In contrast with the above-mentioned single or multi-model
based soft sensors which build global or local model offline on train-
ing data, the JIT based soft sensors build local models by selecting
samples similar to the query data. Thus JIT learning based models
can effectively cope with abrupt changes in process characteristics
as well as gradual ones. Furthermore, it can cope with nonlin-
earity by building local models repeatedly. Based on the present
results in Fig. 13 and Table 3, it is obvious that the JIT-PLS, JIT-KPLS,
MJIT-KPLS are superior to other methods. Even though linear PLS
is adopted in JIT-PLS method, it still obtains satisfactory prediction
results due to the use of the JIT learning framework. The JIT-KPLS
method shows better performance than JIT-PLS method since the
nonlinear KPLS instead of PLS is built. However, the JIT-PLS and
JIT-KPLS soft sensors construct a single JIT modeling framework
without considering the multiplicity of operating phases of batch
processes, which comprises the prediction performance. Among
all soft sensors, MJIT-KPLS performs best due to the integration of
multiphase modeling, hybrid similarity measure, adaptive sample
selection, and MSR database updating.
All soft sensors are also evaluated on testing dataset II which
describes a new process state different from the training data and
testing dataset I. Similar to the comparison results from testing
dataset I, nonlinear modeling techniques is more capable of hand-
ling process nonlinearity than linear method. Compared to the
single KPLS and GPR models, however, the prediction accuracies
of MKPLS and MGPR are not improved by applying multi-model
modeling framework. This performance degradation may be due to
the changes of process characteristics. In contrast, the BMA based
MKPLS and MGPR approaches give lower RMSE and higher R2 than

Table 3
Quantitative comparison of prediction results from different soft sensors. Testing dataset I (from November 19, 2011, to January 2, 2012). Testing dataset II (from March 27
to April 22, 2013). The similarity weight used for database updating is update = 0.8.

No. Method Type Testing dataset I Testing dataset II

2
RMSE R RMSE R2

1 PLS Single-model 0.3837 0.8668 0.4780 0.8315

2 KPLS Single-model 0.2669 0.9356 0.3595 0.9047
3 GPR Single-model 0.2691 0.9345 0.3518 0.9088
4 MKPLS Multi-model 0.2587 0.9395 0.3786 0.8943
5 MGPR Multi-model 0.2539 0.9417 0.3585 0.9052
6 BMA-MKPLS BMA multi-model 0.2432 0.9465 0.3421 0.9137
7 BMA-MGPR BMA multi-model 0.2471 0.9448 0.3424 0.9135
8 JIT-PLS JIT learning 0.2366 0.9494 0.3147 0.9270
9 JIT-KPLS JIT learning 0.2342 0.9506 0.3129 0.9282
10 MJIT-KPLS JIT learning 0.2294 0.9524 0.3088 0.9297
 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93 91

0.05 0.08

0.07
0.045 mean = 0.0372, std = 0.0022 mean = 0.0411, std = 0.0075
CPU time (s)

0.06

CPU time (s)

0.04
0.05

0.035 0.04

0.03
0.03
1 2 3 4 5 6 7 8 9 10 0.02
Model number 0 50 100 150 200 250 300
(a) MJIT-KPLS A Testing sample
(a) MJIT-KPLS A
0.05
0.07
0.045 mean = 0.0360, std = 0.0049 mean = 0.0394, std = 0.0054
0.06
CPU time (s)

0.04

CPU time (s)

0.05
0.035

0.04
0.03

0.03
0.025
1 2 3 4 5 6 7 8 9 10
Model number
0.02
(b) MJIT-KPLS B 0 50 100 150 200 250 300
Test ing sample
(b) MJIT-KPLS B
0.8
1.2 mean = 0.8733, std = 0.2096
mean = 0.5483, std = 0.0931
1.1
0.7
CPU time (s)

1
CPU time (s)

0.9
0.6
0.8
0.7
0.5
0.6
0.5
1 2 3 4 5 6 7 8 9 10 0.4
Model number
(c) MJIT-KPLS C 0 50 100 150 200 250 300
Testing sample
1.1
(c) MJIT-KPLS C
1 mean = 0.7841, std = 0.1212
1.1
mean = 0.7040, std = 0.1108
0.9
CPU time (s)

1
0.8
0.9
CPU time (s)

0.7
0.8
0.6
0.7
0.5
1 2 3 4 5 6 7 8 9 10 0.6
Model number
(d) MJIT-KPLS D 0.5
0 50 100 150 200 250 300
Fig. 14. Average CPU running time for all testing samples from ten optimal models. Testing sample
(d) MJIT-KPLS D

other single-model or multi-model based methods. In addition, the
prediction results of JIT based soft sensors once again outperform
other soft sensors. However, it is obvious that the model perfor-
mance of all soft sensors deteriorate due to the great changes in
process characteristics. Nevertheless, the JIT based soft sensors gain
much better results than other methods due to their capability of
adapting to new process environment. In practice, the conventional
or BMA based multi-model methods cannot well deal with the
changes in process characteristics because such models are built
offline and remain unchanged once implemented online. By con-
trast, JIT based soft sensors are continuously updated by adding
new input-output data into the database.
Fig. 15. An example of CPU time for each testing sample using MJIT-KPLS D soft
sensor.
Apart from the prediction accuracy, the real-time performance
is also very important for JIT learning. Unlike the traditional model-
ing methods which construct models offline, the JIT based model is
built only when the estimation of a query data is required. Thus the
computational load of the proposed MJIT-KPLS method is analyzed.
The average CPU running time for all testing samples is shown in
Fig. 14, where ten optimal models are selected for each type of
MJIT-KPLS method. Meanwhile, the CPU time for each testing sam-
ple from an example prediction is given in Fig. 15. Clearly, models
92 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93
C–D spend more time than models A–B due to an optimization pro-
cess which is used to determine the local modeling size adaptively
by cross-validation. The statistical mean values and standard devia-
tions of the CPU running time show that the real-time performance
of the proposed MJIT-KPLS methods is satisfactory for most batch
processes where reliable online sensors for key quality variables
are usually not available. For the batch fermentation application
under study, such short running time is negligible compared to the
offline analysis of 6 h interval.
5. Conclusions
The MJIT-KPLS soft sensing algorithm, the main contribution of
this work, provides a novel adaptive soft sensor modeling frame-
work for multiphase batch processes. By exploiting the JIT learning
framework, MJIT-KPLS allows to cope with changes in process
characteristics as well as process nonlinearity. Moreover, Gaussian
mixture model enables automatic phase identification and mul-
tiphase modeling. Through the Bayesian inference strategy, the
JIT-KPLS modeling framework representing the identical phase as
the query data is adaptively chosen for online prediction, which is in
favor of capturing shifting dynamics throughout multiple operating
phases of batch processes. Besides, the estimation accuracy of MJIT-
KPLS approach is further improved by using the combination of a
hybrid similarity measure, an adaptive sample selection strategy, a
maximal similarity replacement based database updating scheme,
and the PMI based input variable selection.
The MJIT-KPLS soft sensor support model adaptation possibili-
ties in four aspects: (i) automatically identifying operating phase
based on Gaussian mixture model and Bayesian inference strategy;
(ii) performing JIT learning to build local KPLS model online from
the samples similar to the query data; (iii) adaptively determining
the optimal local modeling size based on cross-validation; and (iv)
adapting to new process environment by updating database.
The application to the industrial fed-batch CTC fermentation
process has demonstrated that, in terms of the predictive accu-
racy and model adaptation capability, the proposed MJIT-KPLS soft
sensing framework is superior to the single-model based methods,
the conventional multi-model based methods, and the Bayesian
model averaging based multi-model methods.
Acknowledgements
We thank Charoen Pokphand Group for their financial support
and for providing the industrial datasets of fed-batch CTC fermen-
tation process. We also appreciate the valuable comments and
suggestions of the anonymous reviewers.
References
Alford JS. Bioprocess control: advances and challenges. Comput Chem Eng
2006;30(10):1464–75.
Bishop CM. Pattern recognition and machine learning. New York: Springer; 2006.
Chen J, Yu J, Mori J, Rashid MM, Hu G, Yu H, et al. A non-Gaussian pattern matching
based dynamic process monitoring approach and its application to cryogenic air
separation process. Comput Chem Eng 2013;58:40–53.
Chen LZ, Nguang SK, Chen XD. Modelling and optimization of biotechnological pro-
cesses: artificial intelligence approaches. Berlin/Heidelberg: Springer; 2006.
Cheng C, Chiu M-S. A new data-based methodology for nonlinear process modeling.
Chem Eng Sci 2004;59(13):2801–10.
Chow TW, Huang D. Estimating optimal feature subsets using efficient esti-
mation of high-dimensional mutual information. IEEE Trans Neural Netw
2005;16(1):213–24.
Cinar A, Parulekar SJ, Undey C, Birol G. Batch fermentation: modeling, monitoring,
and control. New York: CRC Press; 2003.
Cui L, Xie P, Sun J, Yu T, Yuan J. Data-driven prediction of the product for-
mation in industrial 2-keto-l-gulonic acid fermentation. Comput Chem Eng
2012;36:386–91.
Cybenko G. Just-in-time learning and estimation. In: Bittanti S, Picci G, editors. Iden-
tification, adaptation, learning: the science of learning models from data. Berlin:
Springer; 1996. p. 423–34.
de Jong S. SIMPLS: an alternative approach to partial least squares regression.
Chemom Intell Lab Syst 1993;18(3):251–63.
Delgado MR, Nagai EY, de Arruda LVR. A neuro-coevolutionary genetic fuzzy system
to design soft sensors. Soft Comput 2009;13(5):481–95.
Dochain D. Bioprocess control. London: ISE Ltd; 2008.
Fernando TMKG, Maier HR, Dandy GC. Selection of input variables for data driven
models: an average shifted histogram partial mutual information estimator
approach. J Hydrol 2009;367(3):165–76.
Fortuna L, Graziani S, Rizzo A, Xibilia MG. Soft sensors for monitoring and control of
industrial processes. London: Springer; 2007.
Fujiwara K, Kano M, Hasebe S, Takinami A. Soft-sensor development using
correlation-based just-in-time modeling. AIChE J 2009;55(7):1754–65.
François D, Rossi F, Wertz V, Verleysen M. Resampling methods for parameter-
free and robust feature selection with mutual information. Neurocomputing
2007;70(7):1276–88.
Frenzel S, Pompe B. Partial mutual information for coupling analysis of multivariate
time series. Phys Rev Lett 2007;99(20):204101.
Ge Z, Song Z. A comparative study of just-in-time-learning based methods for online
soft sensor modeling. Chemom Intell Lab Syst 2010;104(2):306–17.
George EI. The variable selection problem. J Am Stat Assoc 2000;95(452):1304–8.
Gonzaga J, Meleiro L, Kiang C, Maciel Filho R. ANN-based soft-sensor for real-time
process monitoring and control of an industrial polymerization process. Comput
Chem Eng 2009;33(1):43–9.
Grbić R, Slišković D, Kadlec P. Adaptive soft sensor for online prediction and process
monitoring based on a mixture of Gaussian process models. Comput Chem Eng
2013;58:84–97.
Haimi H, Mulas M, Corona F, Vahala R. Data-derived soft-sensors for bio-
logical wastewater treatment plants: an overview. Environ Model Softw
2013;47:88–107.
Jang J-S. ANFIS: adaptive-network-based fuzzy inference system. IEEE Trans Syst
Man Cybern 1993;23(3):665–85.
Jassar S, Liao Z, Zhao L. Adaptive neuro-fuzzy based inferential sensor model for
estimating the average air temperature in space heating systems. Build Environ
2009;44(8):1609–16.
Jia R, Mao Z, Wang F. KPLS model based product quality control for batch processes.
CIESC J 2013;64(4):1332–9.
Jolliffe I. Principal component analysis. second ed. New York: Springer; 2002.
Kadlec P, Gabrys B, Strandt S. Data-driven soft sensors in the process industry.
Comput Chem Eng 2009;33(4):795–814.
Kadlec P, Gabrys B. Local learning-based adaptive soft sensor for catalyst activation
prediction. AIChE J 2011;57(5):1288–301.
Kadlec P, Grbić R, Gabrys B. Review of adaptation mechanisms for data-driven soft
sensors. Comput Chem Eng 2011;35(1):1–24.
Kano M, Nakagawa Y. Data-based process monitoring, process control, and quality
improvement: recent developments and applications in steel industry. Comput
Chem Eng 2008;32(1):12–24.
Kano M, Ogawa M. The state of the art in chemical process control in Japan: good
practice and questionnaire survey. J Process Control 2010;20(9):969–82.
Kano M, Fujiwara K. Virtual sensing technology in process industries: trends
and challenges revealed by recent industrial applications. J Chem Eng Jpn
2013;46(1):1–17.
Kim S, Kano M, Hasebe S, Takinami A, Seki T. Long-term industrial applications of
inferential control based on just-in-time soft-sensors: economical impact and
challenges. Ind Eng Chem Res 2013;52(35):12346–56.
Lin B, Recke B, Knudsen JK, Jørgensen SB. A systematic approach for soft sensor
development. Comput Chem Eng 2007;31(5):419–25.
Liu H, Shah S, Jiang W. On-line outlier detection and data cleaning. Comput Chem
Eng 2004;28(9):1635–47.
May RJ, Maier HR, Dandy GC, Fernando T. Non-linear variable selection for arti-
ficial neural networks using partial mutual information. Environ Mod Softw
2008a;23(10):1312–26.
May RJ, Dandy GC, Maier HR, Nixon JB. Application of partial mutual information
variable selection to ANN forecasting of water quality in water distribution
systems. Environ Model Softw 2008b;23(10):1289–99.
McLachlan G, Peel D. Finite mixture models. New York: John Wiley & Sons; 2000.
Nicoletti MC, Jain LC. Computational intelligence techniques for bioprocess mod-
elling, supervision and control. Berlin/Heidelberg: Springer; 2009.
Paluš M, Komárek V, Hrnčíř Z, Štěrbová K. Synchronization as adjustment of informa-
tion rates: detection from bivariate time series. Phys Rev E 2001;63(4):046211.
Pani AK, Mohanta HK. A survey of data treatment techniques for soft sensor design.
Chem Prod Process Model 2011;6(1), Article 2.
Pani AK, Vadlamudi VK, Mohanta HK. Development and comparison of neural net-
work based soft sensors for online estimation of cement clinker quality. ISA
Trans 2013;52(1):19–29.
Rashid MM, Yu J. A new dissimilarity method integrating multidimensional mutual
information and independent component analysis for non-Gaussian dynamic
process monitoring. Chemom Intell Lab Syst 2012a;115:44–58.
Rashid MM, Yu J. Nonlinear and non-Gaussian dynamic batch process monitoring
using a new multiway kernel independent component analysis and multidi-
mensional mutual information based dissimilarity approach. Ind Eng Chem Res
2012b;51(33):10910–20.
Rasmussen C, Williams C. Gaussian processes for machine learning. Cambridge, MA:
MIT Press; 2006.
 H. Jin et al. / Computers and Chemical Engineering 71 (2014) 77–93
Rosipal R, Trejo LJ. Kernel partial least squares regression in reproducing kernel
hilbert space. J Mach Learn Res 2001;2:97–123.
Schaal S, Atkeson CG, Vijayakumar S. Scalable techniques from nonparametric statis-
tics for real time robot learning. Appl Intell 2002;17(1):49–60.
Scott DW. Multivariate density estimation: theory, practice, and visualization. New
York: John Wiley & Sons; 1992.
Sharma A. Seasonal to interannual rainfall probabilistic forecasts for improved water
supply management: part 1—a strategy for system predictor identification. J
Hydrol 2000;239(1):232–9.
Sharmin R, Sundararaj U, Shah S, Griend LV, Sun Y-J. Inferential sensors for estima-
tion of polymer quality parameters: industrial application of a PLS-based soft
sensor for a LDPE plant. Chem Eng Sci 2006;61(19):6372–84.
Shigemori H, Kano M, Hasebe S. Optimum quality design system for steel
products through locally weighted regression model. J Process Control
2011;21(2):293–301.
Sliskovic D, Grbic R, Hocenski Z. Methods for plant data-based process modeling in
soft-sensor development. Automatika 2011;52(4):306–18.
Westerhuis JA, Kourti T, MacGregor JF. Comparing alternative approaches
for multivariate statistical analysis of batch process data. J Chemom
1999;13(3–4):397–413.
93
Wold H. Nonlinear estimation by iterative least squares procedures. In: Research
papers in statistics. New York: Wiley; 1966. p. 411–44.
Yu J. Multiway Gaussian mixture model based adaptive kernel partial least
squares regression method for soft sensor estimation and reliable qual-
ity prediction of nonlinear multiphase batch processes. Ind Eng Chem Res
2012a;51(40):13227–37.
Yu J. A Bayesian inference based two-stage support vector regression frame-
work for soft sensor development in batch bioprocesses. Comput Chem Eng
2012b;41:134–44.
Yu J. Online quality prediction of nonlinear and non-Gaussian chemical processes
with shifting dynamics using finite mixture model based Gaussian process
regression approach. Chem Eng Sci 2012c;82:22–30.
Yu J, Chen K, Rashid MM. A Bayesian model averaging based multi-kernel Gaussian
process regression framework for nonlinear state estimation and quality pre-
diction of multiphase batch processes with transient dynamics and uncertainty.
Chem Eng Sci 2013;93:96–109.
Yu J, Qin SJ. Multimode process monitoring with Bayesian inference-based finite
Gaussian mixture models. AIChE J 2008;54(7):1811–29.
Yu J, Qin SJ. Multiway Gaussian mixture model based multiphase batch process
monitoring. Ind Eng Chem Res 2009;48(18):8585–94.