Sei sulla pagina 1di 14

Available online at www.sciencedirect.

com

Computers and Chemical Engineering 32 (2008) 230–243

Online monitoring of multi-phase batch processes using phase-based


multivariate statistical process control
Xuan-Tien Doan a , Rajagopalan Srinivasan a,b,∗
a Institute of Chemical and Engineering Sciences, 1 Pesek Road, Jurong Island, Singapore 627833, Singapore
b Department of Chemical and Biomolecular Engineering, National University of Singapore,

10 Kent Ridge Crescent, Singapore 119260, Singapore


Received 22 December 2006; received in revised form 8 May 2007; accepted 12 May 2007
Available online 21 May 2007

Abstract
Online monitoring of batch processes using multivariate statistical methods has attracted enormous research interests due to its practical
importance. In this paper, we focus on an important issue that continues to confound online batch process monitoring—run-to-run variations
that do not confirm to a normal distribution around a reference trajectory. Here, we show that a phase-based decomposition of the trajectory
offers a systematic way to overcome this challenge. In our approach, phase changes are detected online using Singular points in key variables.
Run-to-run variations among different instances of a phase are synchronized by using time warping. Finally, phased-based multivariate statistical
process control models are used to monitor the execution of the batch and detect abnormalities. This phase-based monitoring approach is robust
to run-to-run variations arising from changes in initial conditions and event timings as is illustrated using a well-known fermentation process
simulation.
© 2007 Elsevier Ltd. All rights reserved.

Keywords: Multi-stage process; Feature synchronization; Dynamic time warping; Dynamic PCA; Singular point

1. Introduction batch completes is obtained. A critical difficulty in this approach


is that since information about the future evolution of the tra-
Batch processing is the preferred mode in a number of jectory is not available, simplifying (and sometimes simplistic)
industries including chemical, specialty, pharmaceutical, and assumptions have to be made. The other strategy adopts a com-
semiconductor manufacturing. It has received increasing atten- pliance perspective and aims to ascertain that the ongoing batch
tion recently due to the need to cope with the uncertainty in a is progressing within a pre-determined normal operation zone.
globalized raw material supply network as well as high fluc- In doing so, it is expected (without a guarantee) that all product
tuations in market demand. For any batch process, successful quality-, environment-, and safety-related constraints would be
operation means producing quality products consistently, at min- met in an economical way at the end of the batch. This strat-
imum cost, in a safe manner while complying with ever stringent egy of minimizing deviations from a “golden batch” obviates
safety and environment constraints. Realtime process monitor- the problem associated with the lack of information about the
ing is essential to achieve this. Two different batch process future evolution of the process. The compliance perspective has
monitoring strategies can be broadly differentiated. The first therefore received the most attention in literature; however, the
strategy adopts a closed-loop quality-control perspective and two approaches can be integrated.
aims to estimate the end-of-batch quality in realtime. Through Owing to rapid advancements in sensor development and
this, a direct indication of what would be achieved when the application, hundreds if not thousands, of common process
variables can be measured online. While batch processes are
typically less instrumented than continuous ones, a substantial
∗ Corresponding author at: Departmentof Chemical and Biomolecular Engi- amount of processing condition related measurements are usu-
neering, National University of Singapore, 10 Kent Ridge Crescent, Singapore
119260, Singapore. Tel.: +65 65168041; fax: +65 67791936.
ally available. Further, since each product is typically produced
E-mail addresses: doan xuan tien@ices.a-star.edu.sg (X.-T. Doan), many times, operations data from multiple similar batches are
chergs@nus.edu.sg (R. Srinivasan). also commonly available in many industries. Process historian

0098-1354/$ – see front matter © 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compchemeng.2007.05.010
X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243 231

databases undoubtedly contain a great amount of information Table 1


about the process at hand, and can therefore serve as a valu- Online measurements in injection molding process
able source of knowledge and information for real-time process No. Variables
monitoring and product quality control, if suitably “mined”. 1 Cavity temperature (◦ C)
Multivariate statistical process control (MSPC) techniques in 2 Nozzle pressure (bar)
general, and principal component analysis (PCA) in particular, 3 Stroke (mm)
have been widely regarded as a promising approach for min- 4 Hydraulic pressure (bar)
ing control-relevant information from process history data. Its 5 Plastication pressure (bar)
6 Back pressure (bar)
successful applications have been reported in numerous process 7 Cavity pressure (bar)
industries (Kourti & MacGregor, 1995; Li, Yue, Valle-Cervantes,
& Qin, 2000; Lin et al., 2006; Zhao, Zhang, & Xu, 2006). Tra-
ditionally, MSPC has been applied to processes operating in a stage, the screw rotates forward and fills the mold cavity with
“steady-state” wherein the originating signals can be considered melt. After this, the packing-holding stage starts immediately,
as statistically stationary—the signal properties do not change and packs additional melt into the cavity under pressure to com-
with time. This requirement is not usually satisfied by batch pensate for the contraction due to cooling and solidification of
and fedbatch processes. Therefore, although MSPC studies date the polymer. This stage continues until eventually the gate at the
back to the late 1980s (Qin, 2003), its application to batch pro- mold freezes, isolating the mold from the injection unit. Then
cesses only began in the mid-90s after a series of landmark the cooling stage begins where the polymer inside the mold con-
papers by MacGregor and co-workers (Nomikos & MacGregor, tinues to solidify until suitable for discharge. At the same time
1995a,b,c), in which the multiway principal component analysis the screw moves backward and more melt is generated for the
(MPCA) approach was developed. Since then, the application next cycle. Lu and colleagues conducted experiments on a lab-
of MSPC approaches to batch process monitoring and quality scale injection-molding process. Readers are referred to (Lu,
control has been studied extensively (e.g. Kourti, 2003b, 2005; Gao, Yang, & Wang, 2004; Lu & Gao, 2005) for further details
Ündey & Çinar, 2002; Ündey, Ertunç, & Çinar, 2003). However, of the experimental setup. Here, we use data from their exper-
most of these approaches make a critical assumption—that devi- iments for illustration. Table 1 shows the seven variables that
ations in the batch profile are small. We illustrate this assumption were measured online in their process.
using a case study.
1.2. Challenges in applying MSPC to batch processes
1.1. Injection molding process
Fig. 2 shows the profile of one important variable – nozzle
Injection molding is a manufacturing process for making pressure – in two batches (Run A and Run B) of the injection
parts from thermoplastic material. It is often used in mass- molding process. Certain key characteristics of batch operation
production and prototyping for production of a wide variety of are evident from this, including (1) significant batch-to-batch
parts ranging from small items such as bottle caps to bigger variation, (2) non-stationarity of the signals, (3) landmarks in
parts such as entire body panels of cars. The process consists the variable profile, and (4) multi-phase behavior.
of injection (or filling), packing-holding, and cooling stages. The first two characteristics have received the most attention
Fig. 1 shows an illustration of an injection-molding machine. in literature. The statistical stationarity requirement in MSPC
Plastic material, usually in the form of pellets, are loaded into
the hopper on top of the injection unit and melted. During filling

Fig. 2. Typical injection molding profiles showing location of Singular points


Fig. 1. Schematic of an injection-molding machine (Lu et al., 2004). and run-to-run variation in timings.
232 X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243

control). Three major physical phases can be differentiated


in the injection molding process—filling (in Run A, 0–90th
samples), holding-packing (91st–400th samples), and cooling
(401st–end of batch). The segment of the trajectory during each
phase could have different statistical properties as is evident in
Fig. 2.
The phase-based decomposition of the trajectory allows a
new perspective. In the multi-phase point-of-view, batch-to-
batch variation essentially means that there are variations in
the magnitude and duration of the phases. Dynamic events that
mark the beginning and end of the phase would not be time-
synchronized with the reference batch if there are significant
run-to-run variations. In Fig. 2, although the filling phase starts
at t = 0 in Run A and Run B, it completes at different times.
Similar observations can be made for other landmarks as well.
These variations could be due to endogenous or exogenous con-
Fig. 3. (a) Mean normal trajectory of nozzle pressure profiles and (b) deviation ditions such as changes in raw materials and process operator’s
of nozzle pressure from the mean profile during Run A and Run B. actions and are inherent to process operations. Evidently, a time-
based comparison of the current run with the reference trajectory
approaches is usually satisfied by first defining a reference (aver- can lead to false-positives and false-negatives during monitor-
age) trajectory, and considering the deviations in the trajectory ing. Instead, the deviations in a phase should be considered in the
of the ongoing batch with-respect to the reference. It is typically right context, by comparing it with the corresponding phase in
expected that the deviation from the reference trajectory would the reference trajectory. This is the approach propounded here.
follow a normal distribution. While this assumption is adequate A number of alternate approaches to this issue have been dis-
when there are small batch-to-batch variations, it is often not sat- cussed in literature (Nomikos & MacGregor, 1995b; Ündey &
isfied when there are large variations, especially in the timings of Çinar, 2002; Ündey et al., 2003). One solution is to artificially
key events. Fig. 3 shows the reference trajectory calculated from cut of batch data, or add zeros, or repeat the last observations
11 normal batches of the injection-molding process and the devi- to make batch trajectories equal (Gregersen & Jorensen, 1999;
ations of Run A and Run B (both normal) from this reference. Lennox, Hiden, Montague, Kornfeld, & Goulding, 2002; Zheng,
As can be seen there, there are substantial, systematic deviations McAvoy, Huang, & Chen, 2001). This could lead to signifi-
from normality especially around key landmarks in the process, cant loss of information (due to the discarded data) (Ündey &
which would lead to false positives during monitoring. Çinar, 2002) and the underlying correlations would be broken
These deviations from normality stem from the fact that cor- (by the data manipulation) (Kourti, 2003a). More importantly,
responding events in the different batches occur at distinctly they do not in anyway address the underlying reason for the
different times. If the corresponding events can be synchronized, mismatch in timing of other dynamic features. Another solution
the trajectories can be warped onto one another to make their is the use of an indicator variable, which serves as a surro-
lengths equal, using dynamic time warping (DTW), a method to gate for time and can be used for synchronizing the trajectory.
locally translate, compress, and expand a pair of signals in such a The selected variable must be smooth, progress monotonically
way that the difference between them is minimized. More details in time, and have the same characteristics (starting-, ending-
of DTW are described in Section 2.2. DTW is a versatile tech- values and other dynamic features) for all batches (Nomikos &
nique, however, it requires significant computational resources MacGregor, 1995b). Examples of possible indicator variables
and is inherently not suited for online application. Further, the are reaction extent (Neogi & Schlags, 2001), percent of reac-
mapping of the corresponding events has been attempted pre- tor volume decrease, and percent of substrate addition (Ündey,
viously only after the batch has completed. Monitoring can, in Tatara, & Çinar, 2004). Ündey and Çinar (2002) discussed online
such cases, therefore be done only in an offline sense. A robust monitoring of multi-phase processes using PCA coupled with
and systematic approach for dealing with batch-to-batch vari- an indicator variable. A shortcoming of this approach is that a
ation and non-stationarity in the context of MSPC online is suitable indicator variable is not available in many processes,
therefore required and is the key issue addressed in this paper. including the ones considered here. Also, when indicator vari-
Our approach is based on the fourth characteristic mentioned ables are used along with phase-based monitoring, the number
above—multi-phase behavior. of phases that a batch can be divided into, is equal to the num-
More often than not, batch processes progress through ber of unique indicator variables that can be identified in the
multiple phases which correspond to steps occurring in a process. Since usually very few, if any, indicator variables can
processing unit as a succession of operational and/or phe- be identified in any process—the batch can at best be very
nomenological events such as physical operations, chemical coarsely defined in this approach. The method proposed in this
reactions, and microbial activities (Ündey & Çinar, 2002). work shares the same philosophy as the indicator variable-based
Each phase usually has its own characteristic dynamics and approach, but uses the more general concept of landmarks in key
hence deserves individual treatment (modeling, monitoring, variables to synchronize the trajectory.
X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243 233

Because of their high information content, SPs can be


used for signal representation and segmentation (Bakshi &
Stephanopoulos, 1996; Mallat & Hwang, 1992), as well as sig-
Fig. 4. Proposed  MSPC framework. nal synchronization and comparison (Srinivasan & Qian, 2005).
In this paper, SPs are used to decompose batch trajectories into
phases with homogenous properties.
Other variants that share some similarity to the phase-based A signal is characterized by an ordered list of SPs as is any
monitoring approach proposed here have been reported by segment of the signal. Since a phase corresponds to a segment of
Kosanovich, Piovoso, Dahl, MacGregor, and Nomikos (1994) the batch profile, SPs can be used to robustly characterize phases.
and Dong, McAvoy, and Wang (1996). However, both of The starting and ending of a phase is associated with observable
these studies only considered phases with fixed time length. events, as discussed earlier. Observable events usually result in
Muthuswamy and Srinivasan (2003) recognized that each phase SPs; thus the start of a phase can be flagged in realtime when the
in a multi-phase fermentation process is physiologically differ- corresponding SP is detected. Table 2 shows the SPs in the con-
ent from another. Consequently, they proposed a phase-based text of the different phases in the injection molding process. As
supervisory control scheme including rule-based identification can be seen there, the occurrence of SP5 flags the end of the fill-
of the current fermentation phase and phase-cognizant control ing phase and the beginning of the packing-holding phase. When
action. Taking a different approach, Lu and Gao (2004) reasoned SP8 is detected, it is evident that this phase has completed and
that process correlation remains more or less the same within an the process is moving into the next stage of operation—cooling
operation stage and used K-means clustering to identify process phase. This linkage between phases and SPs can be used to detect
stages and build a sub-PCA model for each. phase changes and switch phase-specific models.
SPs can be detected offline when the complete signal
2. Phase-based batch monitoring framework trajectory is available (Srinivasan & Qian, 2005) or online
as the signal evolves (Srinivasan & Qian, 2007). A num-
In this paper, we develop a phase-based MSPC (MSPC) ber of algorithms can be used for SP identification. For
framework for realtime batch process monitoring. The term example, Kaistha and Moore (2001) developed a method
“phase” here refers to a distinguishable segment of a batch (or based on template matching and used it to synchronize
fedbatch) trajectory with clearly defined endpoints. Monitoring simulated as well as industrial batch data. Srinivasan and
is performed by comparing an online batch with a reference that Qian (2005) used the robust first- and second-derivatives
has been suitably annotated. Details of the reference generation of the signal to identify its SPs for signal synchronization
is described in Section 3.1. The proposed framework consists of and comparison. Methods based on qualitative trend analy-
three elements, as shown in Fig. 4. sis (Mauryaa, Rengaswamyb, & Venkatasubramanian, 2007;
Venkatasubramanian, Rengaswamy, Kavuri, & Yin, 2003),
(1) Phase identification: through Singular point-based dynamic curve registration (Ramsay & Silverman, 1997), as well as fuzzy
feature synchronization in realtime. clustering (Abonyi, Feil, Nemeth, & Arva, 2005) have been com-
(2) Trajectory alignment: to align the samples from the online monly used identifying landmarks in time series. In this paper,
batch with the corresponding phase in the reference. we detect SPs based on non-conformity of a sample to a linear
(3) Dynamic phase modeling: to represent the relationships model of the current phase.
among the variables during a phase and detect deviations
if any. 2.2. Trajectory alignment

Singular points enable the identification of the current phase


Each of these elements is described in detail next. of an ongoing batch in real-time. For monitoring, further res-
olution of the time-scale within the ongoing phase is needed.
2.1. Phase identification Specifically, it is necessary to locate the point in the reference
trajectory that corresponds to the last sample of the ongoing
The first step to phase-based monitoring is to detect phases batch, so that the current point can be scaled accurately. In this
from online process measurements. In this work, we use Sin- work, trajectory alignment is performed using time warping.
gular points for this purpose. Srinivasan and Qian (2005) DTW originates from speech recognition and is capable of
noted that information content is not homogenously distributed translating, compressing, and expanding two time series signals
throughout a signal. Some landmark points, termed Singular (one called the reference and the other the test) in such a way that
points, in a signal contain more information about the dynamic the difference between them is minimized. Generally, there are
behavior. Examples of Singular points (SPs) include points of two classes of DTW methods—symmetric DTW and asymmetric
discontinuities, inflection, trend changes, and local extrema. DTW. While the former treats the two signals equally and warps
Mathematically, a SP is a triplet (Γ, Ω, Ψ ), where Γ is the time both their time axes (to a new scale) to obtain the minimum
of occurrence of the SP, Ω the magnitude, and Ψ its type. Fig. 2 distance, the latter maps the time axis of the test signal onto that
shows examples of SPs in nozzle pressure during two runs of of the reference. In this paper, the asymmetric DTW is used to
the injection molding process. map the online signals to the reference trajectory.
234 X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243

Let T(M × J) and R(K × J) denote two time-sampled tra- can then be transformed to the following dynamic programming
jectories of length M and K, which correspond to the test and problem:
reference signals, respectively. J is the number of variables. Let ⎧
i and j be the time indices of R and T trajectories, respectively. ⎨ DA (i − 1, j) + d(i, j)

1 ≤ i ≤ K. 1 ≤ j ≤ M. DTW maps T onto R by searching for DA (i, j) = min DA (i − 1, j − 1) + 2d(i, j) (2)


an optimal sequence F∗ of P points on the i–j plane such that a DA (i, j − 1) + d(i, j)
distance measure between T and R is minimized.
Mathematically, let where DA (i, j) is the minimum accumulated distance between
point (1, 1) and point (i, j).
F = {c(1), c(2), . . . , c(p), . . . , c(P)} The Itakura local constraint defines a different set of prede-
cessors (i − 1, j), (i − 1, j − 1), and (i − 1, j − 2) and results
be the series of mappings from T to R in a local slope in [ 1/2 2 ] (Itakura, 1975). The optimization
c(p) = [ i(p) j(p) ], 1≤p≤K problem in Eq. (2) then changes to:
⎧ ∗
The local distance between specific points in T and R is given ⎨ DA (i − 1, j) + d(i, j) or [∞ if condition A ]

by DA (i, j) = min DA (i − 1, j − 1) + d(i, j)


DA (i − 1, j − 2) + d(i, j)
d(i, j) = R(i) − T(j) (3)
and the global distance between signals T and R is where, DA (1, 1) = d(1, 1) and condition A∗ indicates that the

P predecessor of point (i − 1, j) is the point (i − 2, j).
d(c(p)) · w(p) Global constraints define the subset of the total search space
p=1 for finding the optimal path. These are motivated by the fact
D(T, R) = that a wide search space is expensive, in terms of both com-

P
w(p) putation time and storage space. Band global constraint is a
p=1 typical global constraint and limits the maximum deviation of
the optimal path from the linear one from (1, 1) to (K, M) to a
where w(p) are weighting coefficients, which are set to 1 in this pre-specified amount, B.
work. The optimal warping is then given by Endpoint constraints ensure that the endpoints of T and R
match
F∗ = argminF [D(T, R)] (1)
c(1) = [ 1 1 ], c(P) = [ K M] (4)
Constraints are needed to reduce the search space of the align-
ment. These are motivated by physical considerations, to avoid More details of the DTW algorithm can be found in Sankoff and
excessive compression or expansion, accelerate the calculation, Kruskal (1983).
or other problem-specific limits on the alignment. Local con- DTW results in a globally optimal mapping of T to R. It has
straints determine local features for each point. For example, the been widely applied for trajectory alignment of batch processes
Sakoe–Chiba local constraint allows a point (i, j) in the grid to (e.g. Gollmer & Posten, 1996; Kassidas, MacGregor, & Taylor,
be reached from points (i − 1, j), (i − 1, j − 1), and (i, j − 1) 1998). However, it is not suitable for online application because
(Sakoe & Chiba, 1978). The optimization problem in Eq. (1) (1) the endpoint constraint in Eq. (4) cannot be specified for

Table 2
Singular points in nozzle pressure profiles
SP Run A (20 ms, bar, e/s/t) Run B (20 ms, bar, e/s/t) Stage Observable event

SP1 (1, −5, s) (1, −3, s) Filling Batch starts and filling begins with abrupt increase
in pressure
SP2 (35, 267, t) (30, 312, t) Filling Nozzle pressure attains designed value for filling
phase
SP3 (79, 276, t) (50, 334, t) Filling Pressure increase to inject additional melt into the
cavity
SP4 (85, 394, e) (54, 473, e) Filling Pressure reaches maximum
SP5 (108, 151, s) (75, 150, s) Filling → packing-holding Filling phase ends and packing-holding phase starts
SP6 (384, 150, s) (354, 151, s) Packing-holding Packing-holding phase completes and nozzle starts
being retracted (into the molding)
SP7 (404, 13, e) (371, 13, e) Packing-holding Pressure reaches minimum
SP8 (418, 30, s) (388, 33, s) Packing-holding → cooling Nozzle retraction is done and cooling phase and
plastication begin
SP9 (748, 25, s) (713, 27, s) Cooling Plastication phase completes
SP10 (1194, 1, s) (1164, 1, s) Cooling Cooling phase ends and batch finishes

e: extrema; s: discontinuity; t: trend change.


X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243 235

the ongoing batch since the corresponding point in the reference length. d is a time lag that is used to capture the serial correlation
batch is not know a priori, and (2) the dynamic program has to be in the process.
solved for every sample which makes its computationally very The corresponding covariance matrix Si for the time-lagged
expensive in both memory and time. In the current context of data is
synchronizing a phase of the ongoing batch with that of the ref-
erence, the global optimality criteria is not a critical requirement (Xd )T (Xd )
Si = (7)
since the optimal assignment at each time point is not physically K−d−1
significant nor necessary in practise. An approximate identifi- The average covariance matrix Savg for I batches is given by
cation of the corresponding point in the reference trajectory is
sufficient and can be performed efficiently using the extrapola- (K − d − 1)Ii=1 Si
tive time warping (XTW) algorithm proposed by Srinivasan and S avg
= (8)
I(K − d)
Qian (2005). The XTW algorithm is a greedy search modifi-
cation of DTW that optimizes each point locally. In contrast to Solving the eigen-decomposition of the covariance matrix Savg
DTW’s backward search (see Eq. (3)), XTW proceeds based on a and retaining a principal components results in the DPCA model
forward search and considers three possible successors for point of the batch process.
(i, j)— (i + 1, j), (i + 1, j + 1), (i + 1, j + 2). This transforms
Eq. (3) to:
⎧ ∗ ∗
⎨ DA (i, j) + d(i + 1, j) or [∞ if condition B ] j = j


DA (i + 1, j ) = min DA (i, j) + d(i + 1, j + 1) j∗ = j + 1 (5)

⎩ ∗
DA (i, j) + d(i + 1, j + 2) j =j+2

with initial condition DA (1, 1) = d(1, 1). Condition B∗ indi- Like in PCA, Hotelling’s T 2 is commonly employed as a
cates that the predecessor of point j ∗ is j. Thus, the decision in monitoring index in DPCA as well
real-time to locate the corresponding point for i is based only on
three comparisons: to increase j by 0, 1, or 2. Further, XTW also (Tk )2 = (X(k))T Pa −1 T
a Pa X(k) (9)
obviates the end-point constraint of DTW (Eq. (4)) and makes
it well-suited for online time warping. We therefore use DTW where X(k) = [ (x(k))T (x(k − 1))T · · ·(x(k − d))T ] is the
for offline trajectory synchronization and XTW online. time-lagged vector of the current measurement x(k), Pa the load-
Once the corresponding point in the reference that should be ing matrix containing a loading vectors, and  a diagonal matrix
compared with the current sample from the ongoing batch has containing the a principal components.
been located, the online is normalized based on this. The phase- The control limit Tα2 can be approximated by means of the
based synchronization ensures that the variation of the ongoing F-distribution
batch from the reference would follow a normal distribution a(K − d − 1)(K − d + 1)
when the batch proceeds normally. To detect abnormalities then, Tα2 = Fa,K−d−a,α (10)
(K − d)(K − d − a)
any MPSC approach can be used. In this work, we have used
dynamic PCA. where α is the confidence limit. We have used α = 95% in this
study.
2.3. Dynamic phase modeling In the following section, we summarize how the phase identi-
fication, trajectory alignment, and dynamic phase modeling are
Dynamic PCA (DPCA) is an extension of PCA to the realm brought together for online monitoring in the  MSPC approach.
of dynamic processes with serial correlation. Ku, Storer, and
Georgakis (1995) proposed the DPCA and Chen and Liu (2002) 3.  MSPC implementation
first applied it to monitor batch processes. Mathematically,
DPCA starts with forming a time-lagged window for each of Like any MSPC approach,  MSPC requires the develop-
the batches in the reference database ment of statistical models offline which are then used for online
⎛ T T T ⎞ monitoring. We first explain how this reference trajectory is
(xi (d + 1)) (xi (d)) ... (xi (1)) constructed.
⎜ ⎟
⎜ (xi (d + 2))T (xi (d + 1))T . . . (xi (2))
T ⎟
⎜ ⎟
⎜ ⎟ 3.1. Generation of phase-based reference model
⎜ .
.. .
.. .
.. ⎟
Xid = ⎜ ⎟(6)
⎜ ⎟
⎜ . . . ⎟ Fig. 5 illustrates the offline modeling in the proposed 
⎜ .. .. .. ⎟
⎝ ⎠ MSPC approach. The following are the key steps:
T T T
(xi (K)) (xi (K − 1)) . . . (xi (K − d))
(1) Let there be I normal batches in the historical database.
where x(k) = [ x1,k x2,k · · · xJ,k
T
] is the J-dimensional Denote one of them, that is representative of the average
observation vector at time k from the ith batch, and K the batch run, as the golden batch.
236 X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243

Fig. 5. Reference model development in  MSPC approach.

(2) Select one are more key variables that reflect the important (3) If x is not a SP, using the reference trajectory of the cur-
phases and phase changes. rent phase, identify the point that corresponds to x using
(3) Identify SPs in the profiles of the key variables in the golden XTW. Scale x using the mean and standard deviation at the
batch. Shortlist the SPs that correspond to phase changes. corresponding point in the reference trajectory.
(4) Segment all the J variables in the golden batch into phases (4) Construct a time-lagged data matrix using the previous d
based on the time of the occurrence of the shortlisted SPs. scaled observations.
(5) Identify corresponding SPs for the remaining I − 1 nor- (5) Evaluate the T 2 statistic using the scaled measurement on
mal batches, and segment the training batches into phases. I the current MSPC model. Compare the obtained T 2 statistic
training sets are now available for each phase, but these sets with its control limit and announce a fault if exceeded.
would be of different durations. (6) Go back to step 1 and process the next measurement.
(6) For each phase, synchronize the length of the I − 1 training
sets with that of the golden batch by asymmetric DTW, using Next, we illustrate the above algorithms using the injection
the latter as the reference. molding process described in Section 1.1.
(7) For each phase, develop a MSPC model using the I equi-
length training sets. Calculate the control limit Tα2 for the 3.3. Illustration: online monitoring of injection molding
phase as well. process using MSPC

The golden batch, the shortlisted SPs, and the MSPC models The injection modeling process has well defined phases as
for the different phases together form the aggregate phase-based described earlier and is therefore well suited for MSPC. We use
reference model. This aggregate model can be used for online data from nine normal batches to develop the aggregate model.
monitoring as described below. Their batch lengths varied from 1164 to 1194 samples.

3.2. Online implementation (1) Run A, shown in Fig. 2 with a length of 1194 samples was
selected as the golden batch.
The procedure for online monitoring in the  MSPC (2) Of the seven online measurement available, nozzle pressure
approach is outlined in Fig. 6. was selected as the key variable to detect phase changes.
(3) A total of 10 SPs were identified in nozzle pressure during
(1) Obtain new measurement x. Run A as shown in Fig. 2 and Table 2. Of these SP5 and SP8
(2) Check if x is a SP using the SP identification algorithm. If are used to flag phase changes.
x is a SP, compare with the SPs in the reference model. (4) All the seven variables during Run A were segmented into
(a) If a matching SP is found in the reference model, flag a three phases, filling, packing-holding, and cooling, during
phase change and retrieve the MSPC model for the new [1 108], [109 418], and [419 1194] samples, respectively.
phase. (5) The other eight batches in the historical database were sim-
(b) If a matching SP is not found in the reference model, ilarly analyzed. The time of occurrence of the SPs had a
calculate T 2 statistic for the key variable and compare variance of about 13 samples between the training batches.
with Tα2 . Announce a fault if the control limit has been One training run, Run B, shown in Fig. 2 had a duration of
exceeded. 1164 samples. Its SPs were identified as tabulated in Table 2.
X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243 237

Fig. 6. Algorithm for online supervision by  MSPC approach.

From this, the duration of the three phases in this case were These three DPCA models along with Run A and its SPs are
calculated as [ 1 75 ], [ 76 388 ], and [ 389 1164 ] sam- used as the aggregate reference model to monitor the injection
ples, respectively. molding process. Results from one such monitoring run, hence-
(6) Each phase was individually synchronized using asymmet- forth called the test, is described next. Fig. 8 shows the profile
ric DTW with the segment from Run A as the reference and of variable 6, back pressure, during this run.
that from the other historical batches as the test. Fig. 7 shows Online samples for the seven variables were collected. A
the effect of time warping Run B. As can be seen there, as a check for Singular points is conducted on the key variable,
result, the key events in Run A (golden batch) and Run B are nozzle pressure. Initially the process starts in the filling phase.
now synchronized. Similar time warping was performed for At t = 353 sample, a SP is detected and found to match SP5
the other seven normal batches in the historical database. in Run A. This indicates that the process has moved to the
(7) Using the equi-length phase data obtained after time warp- packing-holding phase and the corresponding DPCA model is
ing, DPCA models were built for each phase. Three PCs subsequently used for monitoring. At t = 371 sample, the T 2
were retained in each case and captured 58%, 70%, and statistic jumps beyond its control limit and a fault is flagged as
61% of the variance for the three phases, respectively. T 2 shown in Fig. 9b. This corresponds to an abnormal spike in the
statistics and 95% control limit were also determined. back pressure at t = 371 sample that can be observed in Fig. 8.

Fig. 7. Phase-based time warping of Run B to golden batch Run A during


reference model generation. Fig. 8. Profile of back pressure during test run.
238 X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243

Table 3
Variables used in monitoring PenSim
No. Variables

1 Aeration rate (L/h)


2 Agitator power (W)
3 Substrate feed rate (L/h)
4 Substrate feed temperature (K)
5 Dissolved oxygen concentration (% saturation)
6 Culture volume (L)
7 Carbon dioxide concentration (mmol/L)
8 pH
9 Temperature (K)
10 Generated heat (kcal/h)

fermentation process was selected to demonstrate the advantages


of the proposed  MSPC framework.
A process simulator for the fed-batch penicillin fermentation
Fig. 9. Monitoring of test run in injection molding process by (a) DPCA, and process, known as PenSim, was developed by Birol, Ündey,
(b)  MSPC approaches.
and Çinar (2002) to provide a testbed for several applications
including batch process monitoring methods (Lee, Yoo, & Lee,
The process returns to normal subsequently and is reflected by 2004a,b; Ündey et al., 2004). The simulator, which is based
the T 2 statistic falling below the control limit at t = 380 sam- on a mechanistic model proposed by Bajpai and Reuss (1980),
ple. The cooling phase is flagged at t = 388 sample (SP8 ) and captures the effect of a number of key variables such as pH, tem-
the corresponding DPCA model is used then on. Spikes in back perature, concentrations of biomass, penicillin, carbon substrate
pressure occur again at t = 713, 1119, and 1162 samples and under various operating conditions. As pH and temperature
are flagged as abnormalities by the model. The batch then ends are critical to penicillin quality and quantity, a closed-loop
at t = 1164 sample. proportional-integral-derivative (PID) control system for these
The classical monolithic DPCA approach was also tested on variables are implemented in PenSim. Alternatively, pH control
the same run. During model development, the duration varia- can be implemented by adding concentrated acid (3 M) or base
tions among the historical batches was resolved by cutting-off solution in an on/off fashion when the PID controller is turned
all the batches to the minimum common length—1164 samples. off. As mentioned previously, the process typically consists of
A time lag of d = 3 was selected for the DPCA, and a three two stages including a biomass growth stage and a penicillin pro-
PC model that captured 52% of the variance was developed. duction stage. To optimize the amount of penicillin produced,
T 2 statistics and 95% control limit are used for monitoring. As the first stage is carried out in batch mode and the second in fed-
shown in Fig. 9a, a false positive is flagged at t = 354 s after the batch mode with glucose being added continuously. PenSim can
process has moved to the packing-holding phase. This is due to also simulate a number of faults including step and ramp changes
the mismatch of the phase switch timings between the golden in aeration rate, agitator power and substrate feed flow rate. In
batch and the test run. No other alarms are flagged by this model this study, we monitor the 10 variables shown in Table 3. The
even though disturbances occur in the back pressure as described four faults shown in Table 4 are studied.
above. These false positive and false negatives highlight the need
for phase-based models. 4.1. Reference model building

4. PenSim case study To generate the reference model that will be used subsequent
for model development, 14 batches of data were generated. Ini-
A fed-batch penicillin cultivation process, is used to illus- tial conditions and set points for each batch were randomly
trate the proposed  MSPC framework. Penicillin together with
other antibiotics have had major beneficial effects on human Table 4
and animal health (Demain, 2000). As a result, the production Faults in PenSim
of such secondary metabolites has received extensive attention Scenario Description Occurrence time (sample)
from both academia and industry (Atkinson & Mavituna, 1991). 1 Normal operation with varying NA
In industry, penicillin is commonly produced by fermentation of initial conditions
filamentous microorganisms. It is known that the target product, 2 15% step increase in substrate 160–end
the antibiotic is produced usually after cell growth has slowed feed flow rate
down (Demain, 2000). Hence, the cultivation process is com- 3 15% step decrease in substrate 160–end
feed flow rate
monly carried out in two stages: growing the cells in batch 4 15% step decrease in agitation 20–40
mode followed by penicillin synthesis in fed-batch mode. For its power
multi-phase, non-linear and discontinuous nature, the penicillin
X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243 239

selected in the range specified by Lee et al. (2004a). The dura- data for each phase also has unequal length (duration). There-
tion of the batches varied between 760 and 840 samples with a fore, one of the batches with batch length = 800 samples was
standard deviation of 25 samples. The variables were sampled selected as the golden batch and other normal batches mapped
every 0.5 h. It was observed that the occurrence of the phase shift to it. The Singular points in the golden batch occur at t = 1,
in each batch is around 86th–92nd samples. These 14 batches 77, 88, 90, 108, 800 samples. Asymmetric DTW with Itakura
were used to form the training set for both DPCA as well as  local constraint and band global constraint B = 60 was applied
MSPC approaches. to equalize the time lengths of the 3D matrices for each phase.
For DPCA modeling, all training batches need to be of the For each length-equalized data matrix, phase mean and stan-
same length. For this, the training data were cut-off to the longest dard deviation were evaluated and used for scaling. The scaled
common length of 760 samples. Consequently, the training data data was used for developing phase-based statistical models. As
forms a 3D (batch × time × variable) matrix, from which the explained earlier, any PCA variant can be used in the proposed
mean and standard deviation along the batch dimension were method. In this work, the type of model for each phase was
evaluated. These are subsequently used for scaling of the training decided based on the length of the phase. Multiway PCA mod-
data. DPCA algorithm as presented in Chen and Liu (2002) els were used for sharp transitions whose duration is less than 5
was then applied with time lag d = 2 samples. Six PCs were samples, PCA models for phases that are less than 50 samples
retained, which together captured 90% of the total variance of long, and DPCA models with time lag d = 2 samples for longer
the training data. Ninety nine percent control limit for T 2 statistic phases. In each model, six PCs were retained, capturing more
was also evaluated. The resulting monolithic model was used for than 90% of the total variance. 99% T 2 control limits were also
monitoring. evaluated for each of the phases.
For  MSPC modeling, the substrate feed rate and pH were
selected as key variables since their dynamic characteristic was 4.2. Process monitoring
deemed to be representative of the process. Fig. 10 shows the
SPs that were detected in one of the training batches. As can 4.2.1. Scenario 1: normal operation
be seen, six SPs were identified for each training data set. The A new batch representing normal operation was generated for
first and the last ones corresponding to the start and the end of testing purpose. The batch duration was 760 samples, initial sub-
the batches, were identified in both pH and substrate feed rate strate concentration 18 g/L, and all other conditions randomly
measurements. The SP at t = 92nd sample marks the transition selected in the same range as the training batches.
from the batch to the fed-batch mode, can be only detected in Fig. 11 shows the results of monitoring this batch using the
the substrate feed rate. Other SPs, which characterize signifi- DPCA and  MSPC approaches. In DPCA monitoring, online
cant process dynamics, can be seen in the pH profile. These SPs measurements were first scaled using the mean and standard
correspond to biological and operational changes in the process deviation of the reference batch. After scaling, the current mea-
(Doan, Srinivasan, Bapat, & Wangikar, 2007). Due to batch-to- surement and two most recent ones (d = 2 samples) were used
batch variations in initial conditions as well as batch duration, to form a lagged measurement, reflecting the ∼ 1 h time constant
the corresponding SPs occur at different times, as much as 80 of the process. T 2 statistic was then evaluated using the DPCA
samples for the last SP (i.e. batch duration) and 8 samples for model built perviously. The DPCA monitoring result shows that
the other SPs. Using the SPs identified, each training data set (Fig. 11a), initially the batch progresses normally until the 88th
was divided into five phases as shown in Fig. 10. The length of sample. At that time, the T 2 statistic exceeds its 99% control
the five phases varies from run to run. Consequently, the training limit, indicating a fault. It remains in alarm status till t = 93

Fig. 10. SPs detected in pH (top) and in substrate feed rate (bottom) of a typical Fig. 11. T 2 statistic during online monitoring of Scenario 1 using (a) DPCA,
training batch. and (b)  MSPC models.
240 X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243

sample when it returns to normal (i.e. below the control limit)


and remains till the batch completes at 760 samples. Clearly,
this is a false positive because the batch proceeds as per normal
operation.
In comparison, the  MSPC (Fig. 11b) showed no such
abnormality. Six SPs were detected in this scenario at t = 1,
82, 94, 96, 116, 760 samples. These differences in the time of
the SPs in this scenario vis-a-vis the reference batch is due to
differences in the initial conditions. Specifically, an increase in
the initial substrate concentration for the test batch means that
the microorganism needs more time to utilize the available glu-
cose. This in turn delayed the transition from one phase to the
next and hence results in an overall elongation of all the phases
(except the last SP, which corresponds to the batch duration pre-
specified by the user) as observed. Despite these differences, the
proposed method synchronizes automatically for the variation
in the phase lengths via online SP identification, and correctly Fig. 12. T 2 statistic during online monitoring of Scenario 2 using (a) DPCA,
switches to the right phase model at the right times.Then for each and (b)  MSPC models.
online sample, XTW determines the optimal matching point in
the reference trajectory of that phase. This matching point is SP mismatch is calculated. Since the corresponding T 2 statistic
used for scaling the online measurement and T 2 statistic eval- exceeds the 99% control limit, a fault is announced. This demon-
uated using the corresponding phase’s  MSPC model. The strates the advantage of characterizing a process trajectory using
proposed approach does not result in any excursions outside SPs. A Monte Carlo study was also performed by varying the
the 99% control limit, indicating normal operation from start magnitude of the fault (between 5% to 25%) and the time of
to end of the batch. Particularly, during t = 88–93 samples, the its occurrence (160–560 sample). In 10 different runs,  MSPC
batch is correctly identified to be in the second phase. Hence, detected the fault in every case; however DPCA was successful
the second  MSPC model, obtained previously by modeling in only 2 cases where the fault magnitude was at least 20%.
samples from t = 77 to 88 in the training batches, is used for
monitoring. This illustrates the batch-to-batch variation problem 4.2.3. Scenario 3: step decrease in glucose feed rate
and highlights the need for the SP-based phase synchronization. In Scenario 3, another test batch with nominal initial condi-
The conventional DPCA approach, which does not use such tions was generated. A 15% step decrease in glucose feed rate
synchronization inadvertently compares the online data during was introduced from t = 160 sample till the end of the batch.
t = 88–93 samples with the training data at the same time loca- The result from DPCA and  MSPC are shown in Fig. 14. Both
tion (t = 88–93 samples), which maps to the third and fourth DPCA and  MSPC approaches detect the 15% step decrease
phases in the reference batch, and results in false positives. in glucose feed rate without any significant delay (two samples).
However, the DPCA also flags another fault previously, around
4.2.2. Scenario 2: step increase in glucose feed rate t = 80 sample. This false positive is a minor excursion where
In Scenario 2, a test batch was generated starting with nominal the T 2 statistic exceeds the 99% control limit for five samples
initial conditions as in Birol et al. (2002). At t = 160 sample till
the end of the batch, a 15% step increase in glucose feed rate was
introduced. Both DPCA and  MSPC approaches were applied
for online monitoring. The T 2 statistic from the DPCA model
is shown in Fig. 12a and that from the  MSPC approach in
Fig. 12b.
Fig. 12 shows that the conventional DPCA approach could
not detect the 15% step increase in the glucose feed rate which
was introduced from t = 160 sample. Fig. 13 shows the profile
of the pH and substrate feed rate during this run. Note that the
15% step increase in the glucose feed rate is significant and
visible in both the profiles, yet the DPCA completely misses the
fault.
On the other hand,  MSPC detects the fault at t = 160 sam-
ple based on the unexpected SP at that location. Fig. 13 shows the
normal SPs as well as the unexpected SP identified for this batch.
This SP (160 sample, 4.995, 0.0493 L/h, sharp change) is clas-
sified as abnormal because a sharp change SP is not expected
at that stage of operation. Therefore, a residual based on the Fig. 13. pH (top) and substrate feed rate (bottom) profiles during Scenario 2.
X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243 241

Fig. 14. T 2 statistic during online monitoring of Scenario 3 using (a) DPCA, Fig. 15. T 2 statistic during online monitoring of Scenario 4 using (a) DPCA,
and (b)  MSPC models. and (b)  MSPC models.

from t = 80 to 85 samples due to batch-to-batch variation, as in


not explicitly account for changes in process dynamics from
Scenario 1. However, such nuisance alarms can deter the oper-
one phase to another during the batch run. This issue is accen-
ator from taking corrective actions when a real fault occurs, as
tuated when run-to-run variations occur between batches as is
is the case here.
common in many batch processes including fermentation. When
In contrast, the  MSPC detects only the 15% step decrease
batch dynamics are not accurately captured in a MSPC model,
in glucose feed rate at t = 160 sample; no other faults are
inaccuracies result during monitoring—either false alarms (false
reported. Lee et al. (2004a) also studied this fault using multiway
positive) or missed detection (false negatives).
PCA (MPCA) with variable-wise unfolding and time-varying
In this work, we have proposed a unified solution to this
covariances and compared their modified MPCA model with the
problem. Our approach recognizes that a batch can be decom-
conventional MPCA. Although both conventional and modified
posed into different phases—each having different dynamics.
MPCA in that study detected the fault, the detection delay var-
Each phase in a batch could be of different length and start
ied ranged from 40 to 200 samples, which is significantly longer
and end at a priori unknown times. Also, the duration of the
than that from either DPCA or  MSPC approach. In addition,
same phase could be different in different runs due to run-to-run
both the MPCA approaches in Lee et al. (2004a) seem to have a
variations for instance in the initial conditions. The  MSPC
false positive around the t = 80th sample as in our study. This
approach decomposes online batch monitoring into a phase iden-
shows that the false positive issue due to batch-to-batch varia-
tification problem and a phase monitoring one. The first is solved
tion is a recurring problem and one that can be addressed by our
using Singular points – landmarks in key variables – that presage
 MSPC approach.
the beginning of any phase. Since the occurrence of a phase is
detected by a SP, the proposed method is inherently robust to
4.2.4. Scenario 4: step decrease in agitation power run-to-run variations in duration of a phase— whenever a Sin-
Another test batch with duration of 830 samples (and other gular point occurs, the corresponding phase will be detected. The
nominal initial conditions) was generated. A 15% step decrease list of SPs in the reference batch map to the normal sequence of
in agitation power was introduced from t = 20 sample to t = 40 phases. Any deviation in the list of SPs observed online thus flags
sample. The reduction in agitation lowers mass transfer activ- a potential abnormality which can be confirmed using univariate
ity in the fermenter leading to a decrease in dissolved oxygen hypothesis testing on the SP. The use of SPs on key variables for
level and consequently reduced biomass growth. Fig. 15 shows synchronizing trajectories is thus a generalization of the indi-
the online monitoring results from the DPCA and  MSPC cator variable technique (Nomikos & MacGregor, 1995b) used
approaches. As can be seen, the batch having duration of 830 with DTW. In contrast to the indicator variable, the proposed
samples started and progressed under normal operation. The approach does not require the key variable to be monotonic.
fault is detected by both DPCA and  MSPC as soon as it Any variable that significantly correlates to phase changes can
occurs. These results are also in line with those of the modified be used as a key variable. Further, multiple variables can be used
MPCA of Lee et al. (2004a). if a single variable does not map to all phase changes. Thus,
the process can be divided into as many stages as is physically
5. Discussion meaningful, without any constraint on the phase length, number
of phases, or the availability of a suitable indicator variable.
Conventional MSPC approaches for online monitoring of Abnormality during the execution of a known phase is iden-
batch processes suffer from a major limitation since they do tified by phase monitoring. MSPC models specific to the phase
242 X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243

are developed from historical data. The dynamics during the Kaistha, N., & Moore, C. F. (2001). Extraction of event times in batch profiles
phase are captured by the time-varying mean and standard devi- for time synchronization and quality predictions. Industrial and Engineering
ation of the phase’s reference trajectory. When a new online Chemistry Research, 40, 252–260.
Kassidas, A., MacGregor, J. F., & Taylor, P. A. (1998). Synchronization of batch
sample becomes available, it is synchronized with the reference trajectories using dynamic time warping. AIChE Journal, 44, 864–875.
through time warping. This ensures that the sample is compared Kosanovich, K. A., Piovoso, M. J., Dahl, K. S., MacGregor, J. F., & Nomikos, P.
with the corresponding point in the reference, thus accounting (1994). Multiway PCA applied to an industrial batch process. In Proceedings
for run-to-run variations in phase execution. The online data of the American Control Conference, June (pp. 1294–1298).
after such normalization is used for phase monitoring. The  Kourti, T., & MacGregor, J. F. (1995). Process analysis, monitoring and diagno-
sis, using multivariate projection methods—A tutorial. Chemometrics and
MSPC approach does not place a limitation on the actual mod- Intelligent Laboratory Systems, 28, 3–21.
eling scheme—any MPSC approach can be used. In the case Kourti, T. (2003a). Abnormal situation detection, three-way data and projection
study in Section 4, PCA, multiway PCA, and DPCA were used. methods; robust data archiving and modeling for industrial applications.
However, other variants of PCA and PLS can be used as well. Annual Reviews in Control, 27, 131–139.
The latter is particulary useful for monitoring product quality Kourti, T. (2003b). Multivariate dynamic data modeling for analysis and sta-
tistical process control of batch processes, start-ups and grade transitions.
and is the focus of our current research. Journal of Chemometrics, 17, 93–109.
The key benefits of the  MSPC approach arise from the Kourti, T. (2005). Application of latent variable methods to process control and
individual models deployed for each phase, rather than a mono- multivariate statistical process control in industry. International Journal of
lithic one for the whole batch. The control limits for detecting Adaptive Control and Signal Processing, 19, 213–246.
deviations within a phase can be tighter since only dynamics Ku, W., Storer, R. H., & Georgakis, C. (1995). Disturbance detection and isola-
tion by dynamic principal component analysis. Chemometrics and Intelligent
appropriate to a phase is represented. This improve sensitivity Laboratory Systems, 30, 179–196.
to faults—both false positives and false negatives are reduced Lee, J. M., Yoo, C. K., & Lee, I. B. (2004a). Enhanced process monitoring of fed-
and justifies the extra effort required in developing the multiple batch penicillin cultivation using time-varying and multivariate statistical
models. analysis. Journal of Biotechnology, 110, 119–136.
Lee, J. M., Yoo, C. K., & Lee, I. B. (2004b). Fault detection of batch pro-
cesses using multiway kernel principal component analysis. Computers and
Acknowledgements Chemical Engineering, 28, 1837–1847.
Lennox, B., Hiden, H. G., Montague, A., Kornfeld, G., & Goulding, P. R. (2000).
Application of multivariate statistical process control to batch operations.
The data on injection molding experiments used in this study Computers and Chemical Engineering, 24, 291–296.
were kindly provided by Lu and colleagues. Lin, Z., Zhou, L., Mahajan, A., Song, S., Wang, T., Ge, Z., & Ellison, D.
(2006). Real-time endpoint monitoring and determination for a pharma-
ceutical salt formation process with in-line FT-IR spectroscopy. Journal of
References Pharmaceutical and Biomedical Analysis, 41, 99–104.
Li, W., Yue, H. H., Valle-Cervantes, S., & Qin, S. J. (2000). Recursive PCA for
Abonyi, J., Feil, B., Nemeth, S., & Arva, P. (2005). Modified gath–geva clus- adaptive process monitoring. Journal of Process Control, 10, 471–486.
tering for fuzzy segmentation of multivariate time-series. Fuzzy Sets and Lu, N., Gao, F., Yang, Y., & Wang, F. (2004). PCA-based modeling and on-
Systems, 149, 39–56. line monitoring strategy for uneven-length batch processes. Industrial and
Atkinson, B., & Mavituna, F. (1991). Biochemical engineering and biotechnol- Engineering Chemistry Research, 43, 3343–3352.
ogy handbook. New York: Stockton Press. Lu, N., & Gao, F. (2004). Sub-PCA modeling and on-line monitoring strategy
Bajpai, R. K., & Reuss, M. (1980). A mechanistic model for penicillin pro- for batch processes. AIChE Journal, 50, 255–259.
duction. Journal of Chemical Technology and Biotechnology, 30, 332– Lu, N., & Gao, F. (2005). Stage-based process analysis and quality prediction
344. for batch processes. Industrial and Engineering Chemistry Research, 44,
Bakshi, B. R., & Stephanopoulos, G. (1996). Intelligent systems in process 3547–3555.
engineering—Paradigms from design to operations. San Diego, CA: Aca- Mallat, S., & Hwang, W. L. (1992). Singularity detection and processing with
demic Press. wavelets. IEEE Transactions on Information Theory, 38, 617–643.
Birol, G., Ündey, C., & Çinar, A. (2002). A modular simulation package for Mauryaa, M. R., Rengaswamyb, R., & Venkatasubramanian, V. (2007). Fault
fed-batch fermentation: Penicillin production. Computers and Chemical diagnosis using dynamic trend analysis: A review and recent developments.
Engineering, 26, 1553–1565. Engineering Applications of Artificial Intelligence, 20, 133–146.
Chen, J., & Liu, K. C. (2002). On-line batch process monitoring using dynamic Muthuswamy, K., & Srinivasan, R. (2003). Phase-based supervisory control for
PCA and dynamic PLS. Chemical Engineering Science, 57, 63–75. fermentation process development. Journal of Process Control, 13, 367–382.
Demain, A. L. (2000). Small bus, big business: The economic power of the Neogi, D., & Schlags, C. E. (2001). Applications of multivariate statistical
microbe. Biotechnology Advances, 18, 499–514. techniques for monitoring emulsion batch processes. In Proceedings of the
Doan, X. -T., Srinivasan, R., Bapat, P., & Wangikar, P. (2007). Detection of American Control Conference (pp. 1177–1181).
phase shifts in batch fermentation via statistical analysis of the online Nomikos, P., & MacGregor, J. F. (1995a). Analysis, monitoring and fault diag-
measurements: A case study with rifamycin b fermentation. Journal of nosis of batch processes using multiblock and multiway PLS. Journal of
Biotechnology, doi:10.1016/j.jbiotec.2007.06.013, in press. Process Control, 5, 277–284.
Dong, D., McAvoy, T. J., & Wang, F. (1996). Batch tracking via non-linear Nomikos, P., & MacGregor, J. F. (1995b). Multi-way partial least squares in mon-
principal component analysis. AIChE Journal, 42, 2199–2208. itoring batch processes. Chemometrics and Intelligent Laboratory Systems,
Gollmer, K., & Posten, C. (1996). Supervision of bioprocesses using a dynamic 30, 97–108.
time warping algorithm. Control Engineering Practice, 4, 1287–1295. Nomikos, P., & MacGregor, J. F. (1995c). Multivariate SPC charts for monitoring
Gregersen, L., & Jorensen, S. B. (1999). Supervision of fed-batch fermentations. batch processes. Technometrics, 37, 41–59.
Chemical Engineering Journal, 75, 69–76. Qin, S. J. (2003). Statistical process monitoring: Basics and beyond. Journal of
Itakura, F. (1975). Minimum prediction residual principle applied to speech Chemometrics, 17, 480–502.
recognition. IEEE Transactions on Acoustics, Speech, and Signal Process- Ramsay, J. O., & Silverman, B. W. (1997). Functional Data Analysis. New York:
ing, 23, 52–72. Springer-Verlag.
X.-T. Doan, R. Srinivasan / Computers and Chemical Engineering 32 (2008) 230–243 243

Sakoe, H., & Chiba, S. (1978). Dynamic programming algorithm optimization for diagnosis. Industrial and Engineering Chemistry Research, 42, 4645–
for spoken word recognition. IEEE Transactions on Acoustics, Speech, and 4658.
Signal Processing, 26, 43–49. Ündey, C., Tatara, E., & Çinar, A. (2004). Intelligent real-time performance
Sankoff, D., & Kruskal, J. (1983). Time warps, string edits, and macromolecules: monitoring and quality prediction for batch/fed-batch cultivations. Journal
The theory and practice of sequence comparison. Addison-Wesley. of Biotechnology, 108, 61–77.
Srinivasan, R., & Qian, M. S. (2005). Offline temporal signal comparison using Venkatasubramanian, V., Rengaswamy, R., Kavuri, S. N., & Yin, K. (2003).
singular points augmented time warping. Industrial and Engineering Chem- A review of process fault detection and diagnosis. Part III. Process his-
istry Research, 44, 4697–4716. tory based methods. Computers and Chemical Engineering, 27, 327–
Srinivasan, R., & Qian, M. S. (2007). Online temporal signal comparison using 346.
singular points augmented time warping. Industrial and Engineering Chem- Zhao, S. J., Zhang, J., & Xu, Y. M. (2006). Performance monitoring of processes
istry Research, 46(13), 4531–4548. with multiple operating modes through multiple PLS models. Journal of
Ündey, C., & Çinar, A. (2002). Statistical monitoring of multistage, multiphase Process Control, 16, 763–772.
batch processes. IEEE Control Systems Magazine, 10, 40–52. Zheng, L., McAvoy, T. J., Huang, Y., & Chen, G. (2001). Application of mul-
Ündey, C., Ertunç, S., & Çinar, A. (2003). Online batch/fed-batch process per- tivariate statistical analysis in batch processes. Industrial and Engineering
formance monitoring, quality prediction, and variable-contribution analysis Chemistry Research, 40, 1641–1649.

Potrebbero piacerti anche